]>
Commit | Line | Data |
---|---|---|
72865317 | 1 | /* |
4ea96698 | 2 | * Copyright (c) 2009-2014, 2016-2018 Nicira, Inc. |
72865317 BP |
3 | * |
4 | * Licensed under the Apache License, Version 2.0 (the "License"); | |
5 | * you may not use this file except in compliance with the License. | |
6 | * You may obtain a copy of the License at: | |
7 | * | |
8 | * http://www.apache.org/licenses/LICENSE-2.0 | |
9 | * | |
10 | * Unless required by applicable law or agreed to in writing, software | |
11 | * distributed under the License is distributed on an "AS IS" BASIS, | |
12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | |
13 | * See the License for the specific language governing permissions and | |
14 | * limitations under the License. | |
15 | */ | |
16 | ||
17 | #include <config.h> | |
db73f716 | 18 | #include "dpif-netdev.h" |
f5ace7cd | 19 | #include "dpif-netdev-private.h" |
72865317 | 20 | |
72865317 BP |
21 | #include <ctype.h> |
22 | #include <errno.h> | |
23 | #include <fcntl.h> | |
24 | #include <inttypes.h> | |
7f3adc00 | 25 | #include <net/if.h> |
b2befd5b | 26 | #include <sys/types.h> |
7daedce4 | 27 | #include <netinet/in.h> |
cdee00fd | 28 | #include <stdint.h> |
72865317 BP |
29 | #include <stdlib.h> |
30 | #include <string.h> | |
31 | #include <sys/ioctl.h> | |
7daedce4 | 32 | #include <sys/socket.h> |
72865317 | 33 | #include <sys/stat.h> |
72865317 BP |
34 | #include <unistd.h> |
35 | ||
9f861c91 | 36 | #include "bitmap.h" |
59e6d833 | 37 | #include "cmap.h" |
5cf3edb3 | 38 | #include "conntrack.h" |
2078901a | 39 | #include "conntrack-tp.h" |
7daedce4 | 40 | #include "coverage.h" |
4d4e68ed | 41 | #include "ct-dpif.h" |
72865317 | 42 | #include "csum.h" |
e14deea0 | 43 | #include "dp-packet.h" |
614c4892 | 44 | #include "dpif.h" |
82a48ead | 45 | #include "dpif-netdev-perf.h" |
72865317 | 46 | #include "dpif-provider.h" |
614c4892 | 47 | #include "dummy.h" |
afae68b1 | 48 | #include "fat-rwlock.h" |
72865317 | 49 | #include "flow.h" |
762d146a | 50 | #include "hmapx.h" |
140dd699 | 51 | #include "id-pool.h" |
4ea96698 | 52 | #include "ipf.h" |
72865317 | 53 | #include "netdev.h" |
b6cabb8f | 54 | #include "netdev-offload.h" |
79f36875 | 55 | #include "netdev-provider.h" |
de281153 | 56 | #include "netdev-vport.h" |
cdee00fd | 57 | #include "netlink.h" |
f094af7b | 58 | #include "odp-execute.h" |
72865317 | 59 | #include "odp-util.h" |
25d436fb BW |
60 | #include "openvswitch/dynamic-string.h" |
61 | #include "openvswitch/list.h" | |
62 | #include "openvswitch/match.h" | |
0d71302e | 63 | #include "openvswitch/ofp-parse.h" |
25d436fb | 64 | #include "openvswitch/ofp-print.h" |
64c96779 | 65 | #include "openvswitch/ofpbuf.h" |
3eb67853 | 66 | #include "openvswitch/shash.h" |
25d436fb | 67 | #include "openvswitch/vlog.h" |
5a034064 | 68 | #include "ovs-numa.h" |
61e7deb1 | 69 | #include "ovs-rcu.h" |
72865317 | 70 | #include "packets.h" |
fd016ae3 | 71 | #include "openvswitch/poll-loop.h" |
0de8783a | 72 | #include "pvector.h" |
26c6b6cd | 73 | #include "random.h" |
d33ed218 | 74 | #include "seq.h" |
3eb67853 | 75 | #include "smap.h" |
0cbfe35d | 76 | #include "sset.h" |
72865317 | 77 | #include "timeval.h" |
53902038 | 78 | #include "tnl-neigh-cache.h" |
7f9b8504 | 79 | #include "tnl-ports.h" |
74cc3969 | 80 | #include "unixctl.h" |
72865317 | 81 | #include "util.h" |
241bad15 | 82 | #include "uuid.h" |
7daedce4 | 83 | |
d98e6007 | 84 | VLOG_DEFINE_THIS_MODULE(dpif_netdev); |
72865317 | 85 | |
5bf84282 NK |
86 | /* Auto Load Balancing Defaults */ |
87 | #define ALB_ACCEPTABLE_IMPROVEMENT 25 | |
88 | #define ALB_PMD_LOAD_THRESHOLD 95 | |
89 | #define ALB_PMD_REBALANCE_POLL_INTERVAL 1 /* 1 Min */ | |
90 | #define MIN_TO_MSEC 60000 | |
91 | ||
8bb113da | 92 | #define FLOW_DUMP_MAX_BATCH 50 |
adcf00ba | 93 | /* Use per thread recirc_depth to prevent recirculation loop. */ |
3f9d3836 | 94 | #define MAX_RECIRC_DEPTH 6 |
adcf00ba | 95 | DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth, 0) |
e4cfed38 | 96 | |
c71ea3c4 IM |
97 | /* Use instant packet send by default. */ |
98 | #define DEFAULT_TX_FLUSH_INTERVAL 0 | |
99 | ||
72865317 | 100 | /* Configuration parameters. */ |
72865317 | 101 | enum { MAX_FLOWS = 65536 }; /* Maximum number of flows in flow table. */ |
4b27db64 JR |
102 | enum { MAX_METERS = 65536 }; /* Maximum number of meters. */ |
103 | enum { MAX_BANDS = 8 }; /* Maximum number of bands / meter. */ | |
104 | enum { N_METER_LOCKS = 64 }; /* Maximum number of meters. */ | |
72865317 | 105 | |
a13a0209 AT |
106 | COVERAGE_DEFINE(datapath_drop_meter); |
107 | COVERAGE_DEFINE(datapath_drop_upcall_error); | |
108 | COVERAGE_DEFINE(datapath_drop_lock_error); | |
109 | COVERAGE_DEFINE(datapath_drop_userspace_action_error); | |
110 | COVERAGE_DEFINE(datapath_drop_tunnel_push_error); | |
111 | COVERAGE_DEFINE(datapath_drop_tunnel_pop_error); | |
112 | COVERAGE_DEFINE(datapath_drop_recirc_error); | |
113 | COVERAGE_DEFINE(datapath_drop_invalid_port); | |
9df65060 | 114 | COVERAGE_DEFINE(datapath_drop_invalid_bond); |
a13a0209 AT |
115 | COVERAGE_DEFINE(datapath_drop_invalid_tnl_port); |
116 | COVERAGE_DEFINE(datapath_drop_rx_invalid_packet); | |
117 | ||
8a4e3a85 BP |
118 | /* Protects against changes to 'dp_netdevs'. */ |
119 | static struct ovs_mutex dp_netdev_mutex = OVS_MUTEX_INITIALIZER; | |
120 | ||
121 | /* Contains all 'struct dp_netdev's. */ | |
122 | static struct shash dp_netdevs OVS_GUARDED_BY(dp_netdev_mutex) | |
123 | = SHASH_INITIALIZER(&dp_netdevs); | |
124 | ||
623540e4 | 125 | static struct vlog_rate_limit upcall_rl = VLOG_RATE_LIMIT_INIT(600, 600); |
6b31e073 | 126 | |
5cf3edb3 | 127 | #define DP_NETDEV_CS_SUPPORTED_MASK (CS_NEW | CS_ESTABLISHED | CS_RELATED \ |
4cddb1f0 DB |
128 | | CS_INVALID | CS_REPLY_DIR | CS_TRACKED \ |
129 | | CS_SRC_NAT | CS_DST_NAT) | |
5cf3edb3 DDP |
130 | #define DP_NETDEV_CS_UNSUPPORTED_MASK (~(uint32_t)DP_NETDEV_CS_SUPPORTED_MASK) |
131 | ||
2494ccd7 | 132 | static struct odp_support dp_netdev_support = { |
f0fb825a | 133 | .max_vlan_headers = SIZE_MAX, |
2494ccd7 JS |
134 | .max_mpls_depth = SIZE_MAX, |
135 | .recirc = true, | |
5cf3edb3 DDP |
136 | .ct_state = true, |
137 | .ct_zone = true, | |
138 | .ct_mark = true, | |
139 | .ct_label = true, | |
2575df07 JP |
140 | .ct_state_nat = true, |
141 | .ct_orig_tuple = true, | |
142 | .ct_orig_tuple6 = true, | |
2494ccd7 JS |
143 | }; |
144 | ||
60d8ccae YW |
145 | /* EMC cache and SMC cache compose the datapath flow cache (DFC) |
146 | * | |
147 | * Exact match cache for frequently used flows | |
9bbf1c3d DDP |
148 | * |
149 | * The cache uses a 32-bit hash of the packet (which can be the RSS hash) to | |
150 | * search its entries for a miniflow that matches exactly the miniflow of the | |
0de8783a | 151 | * packet. It stores the 'dpcls_rule' (rule) that matches the miniflow. |
9bbf1c3d DDP |
152 | * |
153 | * A cache entry holds a reference to its 'dp_netdev_flow'. | |
154 | * | |
155 | * A miniflow with a given hash can be in one of EM_FLOW_HASH_SEGS different | |
156 | * entries. The 32-bit hash is split into EM_FLOW_HASH_SEGS values (each of | |
157 | * them is EM_FLOW_HASH_SHIFT bits wide and the remainder is thrown away). Each | |
158 | * value is the index of a cache entry where the miniflow could be. | |
159 | * | |
160 | * | |
60d8ccae YW |
161 | * Signature match cache (SMC) |
162 | * | |
163 | * This cache stores a 16-bit signature for each flow without storing keys, and | |
164 | * stores the corresponding 16-bit flow_table index to the 'dp_netdev_flow'. | |
165 | * Each flow thus occupies 32bit which is much more memory efficient than EMC. | |
166 | * SMC uses a set-associative design that each bucket contains | |
167 | * SMC_ENTRY_PER_BUCKET number of entries. | |
168 | * Since 16-bit flow_table index is used, if there are more than 2^16 | |
169 | * dp_netdev_flow, SMC will miss them that cannot be indexed by a 16-bit value. | |
170 | * | |
171 | * | |
9bbf1c3d DDP |
172 | * Thread-safety |
173 | * ============= | |
174 | * | |
175 | * Each pmd_thread has its own private exact match cache. | |
176 | * If dp_netdev_input is not called from a pmd thread, a mutex is used. | |
177 | */ | |
178 | ||
fc82e877 | 179 | #define EM_FLOW_HASH_SHIFT 13 |
9bbf1c3d DDP |
180 | #define EM_FLOW_HASH_ENTRIES (1u << EM_FLOW_HASH_SHIFT) |
181 | #define EM_FLOW_HASH_MASK (EM_FLOW_HASH_ENTRIES - 1) | |
182 | #define EM_FLOW_HASH_SEGS 2 | |
183 | ||
60d8ccae YW |
184 | /* SMC uses a set-associative design. A bucket contains a set of entries that |
185 | * a flow item can occupy. For now, it uses one hash function rather than two | |
186 | * as for the EMC design. */ | |
187 | #define SMC_ENTRY_PER_BUCKET 4 | |
188 | #define SMC_ENTRIES (1u << 20) | |
189 | #define SMC_BUCKET_CNT (SMC_ENTRIES / SMC_ENTRY_PER_BUCKET) | |
190 | #define SMC_MASK (SMC_BUCKET_CNT - 1) | |
191 | ||
4c30b246 CL |
192 | /* Default EMC insert probability is 1 / DEFAULT_EM_FLOW_INSERT_INV_PROB */ |
193 | #define DEFAULT_EM_FLOW_INSERT_INV_PROB 100 | |
194 | #define DEFAULT_EM_FLOW_INSERT_MIN (UINT32_MAX / \ | |
195 | DEFAULT_EM_FLOW_INSERT_INV_PROB) | |
196 | ||
9bbf1c3d | 197 | struct emc_entry { |
9bbf1c3d | 198 | struct dp_netdev_flow *flow; |
0de8783a | 199 | struct netdev_flow_key key; /* key.hash used for emc hash value. */ |
9bbf1c3d DDP |
200 | }; |
201 | ||
202 | struct emc_cache { | |
203 | struct emc_entry entries[EM_FLOW_HASH_ENTRIES]; | |
67ad54cb | 204 | int sweep_idx; /* For emc_cache_slow_sweep(). */ |
9bbf1c3d DDP |
205 | }; |
206 | ||
60d8ccae YW |
207 | struct smc_bucket { |
208 | uint16_t sig[SMC_ENTRY_PER_BUCKET]; | |
209 | uint16_t flow_idx[SMC_ENTRY_PER_BUCKET]; | |
210 | }; | |
211 | ||
212 | /* Signature match cache, differentiate from EMC cache */ | |
213 | struct smc_cache { | |
214 | struct smc_bucket buckets[SMC_BUCKET_CNT]; | |
215 | }; | |
216 | ||
217 | struct dfc_cache { | |
218 | struct emc_cache emc_cache; | |
219 | struct smc_cache smc_cache; | |
220 | }; | |
221 | ||
9bbf1c3d DDP |
222 | /* Iterate in the exact match cache through every entry that might contain a |
223 | * miniflow with hash 'HASH'. */ | |
224 | #define EMC_FOR_EACH_POS_WITH_HASH(EMC, CURRENT_ENTRY, HASH) \ | |
225 | for (uint32_t i__ = 0, srch_hash__ = (HASH); \ | |
226 | (CURRENT_ENTRY) = &(EMC)->entries[srch_hash__ & EM_FLOW_HASH_MASK], \ | |
227 | i__ < EM_FLOW_HASH_SEGS; \ | |
228 | i__++, srch_hash__ >>= EM_FLOW_HASH_SHIFT) | |
0de8783a JR |
229 | \f |
230 | /* Simple non-wildcarding single-priority classifier. */ | |
231 | ||
05f9e707 IM |
232 | /* Time in microseconds between successive optimizations of the dpcls |
233 | * subtable vector */ | |
234 | #define DPCLS_OPTIMIZATION_INTERVAL 1000000LL | |
3453b4d6 | 235 | |
05f9e707 IM |
236 | /* Time in microseconds of the interval in which rxq processing cycles used |
237 | * in rxq to pmd assignments is measured and stored. */ | |
238 | #define PMD_RXQ_INTERVAL_LEN 10000000LL | |
4809891b | 239 | |
c59e759f KT |
240 | /* Number of intervals for which cycles are stored |
241 | * and used during rxq to pmd assignment. */ | |
242 | #define PMD_RXQ_INTERVAL_MAX 6 | |
243 | ||
0de8783a | 244 | struct dpcls { |
3453b4d6 JS |
245 | struct cmap_node node; /* Within dp_netdev_pmd_thread.classifiers */ |
246 | odp_port_t in_port; | |
0de8783a | 247 | struct cmap subtables_map; |
da9cfca6 | 248 | struct pvector subtables; |
0de8783a | 249 | }; |
9bbf1c3d | 250 | |
9b4f08cd VDA |
251 | /* Data structure to keep packet order till fastpath processing. */ |
252 | struct dp_packet_flow_map { | |
253 | struct dp_packet *packet; | |
254 | struct dp_netdev_flow *flow; | |
255 | uint16_t tcp_flags; | |
256 | }; | |
257 | ||
0de8783a JR |
258 | static void dpcls_init(struct dpcls *); |
259 | static void dpcls_destroy(struct dpcls *); | |
3453b4d6 | 260 | static void dpcls_sort_subtable_vector(struct dpcls *); |
0de8783a JR |
261 | static void dpcls_insert(struct dpcls *, struct dpcls_rule *, |
262 | const struct netdev_flow_key *mask); | |
263 | static void dpcls_remove(struct dpcls *, struct dpcls_rule *); | |
3453b4d6 | 264 | static bool dpcls_lookup(struct dpcls *cls, |
60d8ccae | 265 | const struct netdev_flow_key *keys[], |
3453b4d6 JS |
266 | struct dpcls_rule **rules, size_t cnt, |
267 | int *num_lookups_p); | |
92c7c870 | 268 | |
4b27db64 JR |
269 | /* Set of supported meter flags */ |
270 | #define DP_SUPPORTED_METER_FLAGS_MASK \ | |
271 | (OFPMF13_STATS | OFPMF13_PKTPS | OFPMF13_KBPS | OFPMF13_BURST) | |
272 | ||
273 | /* Set of supported meter band types */ | |
274 | #define DP_SUPPORTED_METER_BAND_TYPES \ | |
275 | ( 1 << OFPMBT13_DROP ) | |
276 | ||
277 | struct dp_meter_band { | |
278 | struct ofputil_meter_band up; /* type, prec_level, pad, rate, burst_size */ | |
279 | uint32_t bucket; /* In 1/1000 packets (for PKTPS), or in bits (for KBPS) */ | |
280 | uint64_t packet_count; | |
281 | uint64_t byte_count; | |
282 | }; | |
283 | ||
284 | struct dp_meter { | |
285 | uint16_t flags; | |
286 | uint16_t n_bands; | |
287 | uint32_t max_delta_t; | |
288 | uint64_t used; | |
289 | uint64_t packet_count; | |
290 | uint64_t byte_count; | |
291 | struct dp_meter_band bands[]; | |
292 | }; | |
293 | ||
5bf84282 NK |
294 | struct pmd_auto_lb { |
295 | bool auto_lb_requested; /* Auto load balancing requested by user. */ | |
296 | bool is_enabled; /* Current status of Auto load balancing. */ | |
297 | uint64_t rebalance_intvl; | |
298 | uint64_t rebalance_poll_timer; | |
299 | }; | |
300 | ||
8a4e3a85 BP |
301 | /* Datapath based on the network device interface from netdev.h. |
302 | * | |
303 | * | |
304 | * Thread-safety | |
305 | * ============= | |
306 | * | |
307 | * Some members, marked 'const', are immutable. Accessing other members | |
308 | * requires synchronization, as noted in more detail below. | |
309 | * | |
310 | * Acquisition order is, from outermost to innermost: | |
311 | * | |
312 | * dp_netdev_mutex (global) | |
59e6d833 | 313 | * port_mutex |
9df65060 | 314 | * bond_mutex |
d0cca6c3 | 315 | * non_pmd_mutex |
8a4e3a85 | 316 | */ |
72865317 | 317 | struct dp_netdev { |
8a4e3a85 BP |
318 | const struct dpif_class *const class; |
319 | const char *const name; | |
6a8267c5 BP |
320 | struct ovs_refcount ref_cnt; |
321 | atomic_flag destroyed; | |
72865317 | 322 | |
8a4e3a85 BP |
323 | /* Ports. |
324 | * | |
e9985d6a DDP |
325 | * Any lookup into 'ports' or any access to the dp_netdev_ports found |
326 | * through 'ports' requires taking 'port_mutex'. */ | |
59e6d833 | 327 | struct ovs_mutex port_mutex; |
e9985d6a | 328 | struct hmap ports; |
d33ed218 | 329 | struct seq *port_seq; /* Incremented whenever a port changes. */ |
6c3eee82 | 330 | |
c71ea3c4 IM |
331 | /* The time that a packet can wait in output batch for sending. */ |
332 | atomic_uint32_t tx_flush_interval; | |
333 | ||
4b27db64 JR |
334 | /* Meters. */ |
335 | struct ovs_mutex meter_locks[N_METER_LOCKS]; | |
336 | struct dp_meter *meters[MAX_METERS]; /* Meter bands. */ | |
4b27db64 | 337 | |
65dcf3da BB |
338 | /* Probability of EMC insertions is a factor of 'emc_insert_min'.*/ |
339 | OVS_ALIGNED_VAR(CACHE_LINE_SIZE) atomic_uint32_t emc_insert_min; | |
79f36875 JS |
340 | /* Enable collection of PMD performance metrics. */ |
341 | atomic_bool pmd_perf_metrics; | |
60d8ccae YW |
342 | /* Enable the SMC cache from ovsdb config */ |
343 | atomic_bool smc_enable_db; | |
65dcf3da | 344 | |
6b31e073 RW |
345 | /* Protects access to ofproto-dpif-upcall interface during revalidator |
346 | * thread synchronization. */ | |
347 | struct fat_rwlock upcall_rwlock; | |
623540e4 EJ |
348 | upcall_callback *upcall_cb; /* Callback function for executing upcalls. */ |
349 | void *upcall_aux; | |
6b31e073 | 350 | |
e4e74c3a AW |
351 | /* Callback function for notifying the purging of dp flows (during |
352 | * reseting pmd deletion). */ | |
353 | dp_purge_callback *dp_purge_cb; | |
354 | void *dp_purge_aux; | |
355 | ||
65f13b50 AW |
356 | /* Stores all 'struct dp_netdev_pmd_thread's. */ |
357 | struct cmap poll_threads; | |
140dd699 IM |
358 | /* id pool for per thread static_tx_qid. */ |
359 | struct id_pool *tx_qid_pool; | |
360 | struct ovs_mutex tx_qid_pool_mutex; | |
e77c97b9 KT |
361 | /* Use measured cycles for rxq to pmd assignment. */ |
362 | bool pmd_rxq_assign_cyc; | |
65f13b50 AW |
363 | |
364 | /* Protects the access of the 'struct dp_netdev_pmd_thread' | |
365 | * instance for non-pmd thread. */ | |
366 | struct ovs_mutex non_pmd_mutex; | |
367 | ||
368 | /* Each pmd thread will store its pointer to | |
369 | * 'struct dp_netdev_pmd_thread' in 'per_pmd_key'. */ | |
370 | ovsthread_key_t per_pmd_key; | |
f2eee189 | 371 | |
a6a426d6 IM |
372 | struct seq *reconfigure_seq; |
373 | uint64_t last_reconfigure_seq; | |
374 | ||
a14b8947 | 375 | /* Cpu mask for pin of pmd threads. */ |
f2eee189 | 376 | char *pmd_cmask; |
6e3c6fa4 | 377 | |
a36de779 | 378 | uint64_t last_tnl_conf_seq; |
5cf3edb3 | 379 | |
57593fd2 | 380 | struct conntrack *conntrack; |
5bf84282 | 381 | struct pmd_auto_lb pmd_alb; |
9df65060 VDA |
382 | |
383 | /* Bonds. */ | |
384 | struct ovs_mutex bond_mutex; /* Protects updates of 'tx_bonds'. */ | |
385 | struct cmap tx_bonds; /* Contains 'struct tx_bond'. */ | |
72865317 BP |
386 | }; |
387 | ||
4b27db64 JR |
388 | static void meter_lock(const struct dp_netdev *dp, uint32_t meter_id) |
389 | OVS_ACQUIRES(dp->meter_locks[meter_id % N_METER_LOCKS]) | |
390 | { | |
391 | ovs_mutex_lock(&dp->meter_locks[meter_id % N_METER_LOCKS]); | |
392 | } | |
393 | ||
394 | static void meter_unlock(const struct dp_netdev *dp, uint32_t meter_id) | |
395 | OVS_RELEASES(dp->meter_locks[meter_id % N_METER_LOCKS]) | |
396 | { | |
397 | ovs_mutex_unlock(&dp->meter_locks[meter_id % N_METER_LOCKS]); | |
398 | } | |
399 | ||
400 | ||
8a4e3a85 | 401 | static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp, |
e9985d6a DDP |
402 | odp_port_t) |
403 | OVS_REQUIRES(dp->port_mutex); | |
ff073a71 | 404 | |
c59e759f KT |
405 | enum rxq_cycles_counter_type { |
406 | RXQ_CYCLES_PROC_CURR, /* Cycles spent successfully polling and | |
407 | processing packets during the current | |
408 | interval. */ | |
409 | RXQ_CYCLES_PROC_HIST, /* Total cycles of all intervals that are used | |
410 | during rxq to pmd assignment. */ | |
411 | RXQ_N_CYCLES | |
412 | }; | |
413 | ||
02bb2824 YL |
414 | enum { |
415 | DP_NETDEV_FLOW_OFFLOAD_OP_ADD, | |
416 | DP_NETDEV_FLOW_OFFLOAD_OP_MOD, | |
417 | DP_NETDEV_FLOW_OFFLOAD_OP_DEL, | |
418 | }; | |
419 | ||
420 | struct dp_flow_offload_item { | |
421 | struct dp_netdev_pmd_thread *pmd; | |
422 | struct dp_netdev_flow *flow; | |
423 | int op; | |
424 | struct match match; | |
425 | struct nlattr *actions; | |
426 | size_t actions_len; | |
427 | ||
428 | struct ovs_list node; | |
429 | }; | |
430 | ||
431 | struct dp_flow_offload { | |
432 | struct ovs_mutex mutex; | |
433 | struct ovs_list list; | |
434 | pthread_cond_t cond; | |
435 | }; | |
436 | ||
437 | static struct dp_flow_offload dp_flow_offload = { | |
438 | .mutex = OVS_MUTEX_INITIALIZER, | |
439 | .list = OVS_LIST_INITIALIZER(&dp_flow_offload.list), | |
440 | }; | |
441 | ||
442 | static struct ovsthread_once offload_thread_once | |
443 | = OVSTHREAD_ONCE_INITIALIZER; | |
444 | ||
05f9e707 | 445 | #define XPS_TIMEOUT 500000LL /* In microseconds. */ |
324c8374 | 446 | |
3eb67853 IM |
447 | /* Contained by struct dp_netdev_port's 'rxqs' member. */ |
448 | struct dp_netdev_rxq { | |
947dc567 DDP |
449 | struct dp_netdev_port *port; |
450 | struct netdev_rxq *rx; | |
451 | unsigned core_id; /* Core to which this queue should be | |
452 | pinned. OVS_CORE_UNSPEC if the | |
453 | queue doesn't need to be pinned to a | |
454 | particular core. */ | |
ee42dd70 | 455 | unsigned intrvl_idx; /* Write index for 'cycles_intrvl'. */ |
47a45d86 | 456 | struct dp_netdev_pmd_thread *pmd; /* pmd thread that polls this queue. */ |
79f36875 | 457 | bool is_vhost; /* Is rxq of a vhost port. */ |
c59e759f KT |
458 | |
459 | /* Counters of cycles spent successfully polling and processing pkts. */ | |
460 | atomic_ullong cycles[RXQ_N_CYCLES]; | |
461 | /* We store PMD_RXQ_INTERVAL_MAX intervals of data for an rxq and then | |
462 | sum them to yield the cycles used for an rxq. */ | |
463 | atomic_ullong cycles_intrvl[PMD_RXQ_INTERVAL_MAX]; | |
3eb67853 IM |
464 | }; |
465 | ||
72865317 BP |
466 | /* A port in a netdev-based datapath. */ |
467 | struct dp_netdev_port { | |
35303d71 | 468 | odp_port_t port_no; |
ca62bb16 BB |
469 | bool dynamic_txqs; /* If true XPS will be used. */ |
470 | bool need_reconfigure; /* True if we should reconfigure netdev. */ | |
72865317 | 471 | struct netdev *netdev; |
e9985d6a | 472 | struct hmap_node node; /* Node in dp_netdev's 'ports'. */ |
4b609110 | 473 | struct netdev_saved_flags *sf; |
3eb67853 | 474 | struct dp_netdev_rxq *rxqs; |
85a4f238 | 475 | unsigned n_rxq; /* Number of elements in 'rxqs' */ |
47a45d86 | 476 | unsigned *txq_used; /* Number of threads that use each tx queue. */ |
324c8374 | 477 | struct ovs_mutex txq_used_mutex; |
2fbadeb6 | 478 | bool emc_enabled; /* If true EMC will be used. */ |
0cbfe35d | 479 | char *type; /* Port type as requested by user. */ |
3eb67853 | 480 | char *rxq_affinity_list; /* Requested affinity of rx queues. */ |
72865317 BP |
481 | }; |
482 | ||
1c1e46ed AW |
483 | /* Contained by struct dp_netdev_flow's 'stats' member. */ |
484 | struct dp_netdev_flow_stats { | |
eb94da30 DDP |
485 | atomic_llong used; /* Last used time, in monotonic msecs. */ |
486 | atomic_ullong packet_count; /* Number of packets matched. */ | |
487 | atomic_ullong byte_count; /* Number of bytes matched. */ | |
488 | atomic_uint16_t tcp_flags; /* Bitwise-OR of seen tcp_flags values. */ | |
1c1e46ed AW |
489 | }; |
490 | ||
491 | /* A flow in 'dp_netdev_pmd_thread's 'flow_table'. | |
8a4e3a85 BP |
492 | * |
493 | * | |
494 | * Thread-safety | |
495 | * ============= | |
496 | * | |
497 | * Except near the beginning or ending of its lifespan, rule 'rule' belongs to | |
1c1e46ed | 498 | * its pmd thread's classifier. The text below calls this classifier 'cls'. |
8a4e3a85 BP |
499 | * |
500 | * Motivation | |
501 | * ---------- | |
502 | * | |
503 | * The thread safety rules described here for "struct dp_netdev_flow" are | |
504 | * motivated by two goals: | |
505 | * | |
506 | * - Prevent threads that read members of "struct dp_netdev_flow" from | |
507 | * reading bad data due to changes by some thread concurrently modifying | |
508 | * those members. | |
509 | * | |
510 | * - Prevent two threads making changes to members of a given "struct | |
511 | * dp_netdev_flow" from interfering with each other. | |
512 | * | |
513 | * | |
514 | * Rules | |
515 | * ----- | |
516 | * | |
ed79f89a DDP |
517 | * A flow 'flow' may be accessed without a risk of being freed during an RCU |
518 | * grace period. Code that needs to hold onto a flow for a while | |
519 | * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref(). | |
8a4e3a85 BP |
520 | * |
521 | * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the | |
ed79f89a DDP |
522 | * flow from being deleted from 'cls' and it doesn't protect members of 'flow' |
523 | * from modification. | |
8a4e3a85 BP |
524 | * |
525 | * Some members, marked 'const', are immutable. Accessing other members | |
526 | * requires synchronization, as noted in more detail below. | |
527 | */ | |
72865317 | 528 | struct dp_netdev_flow { |
11e5cf1f | 529 | const struct flow flow; /* Unmasked flow that created this entry. */ |
8a4e3a85 | 530 | /* Hash table index by unmasked flow. */ |
1c1e46ed AW |
531 | const struct cmap_node node; /* In owning dp_netdev_pmd_thread's */ |
532 | /* 'flow_table'. */ | |
241bad15 | 533 | const struct cmap_node mark_node; /* In owning flow_mark's mark_to_flow */ |
70e5ed6f | 534 | const ovs_u128 ufid; /* Unique flow identifier. */ |
241bad15 | 535 | const ovs_u128 mega_ufid; /* Unique mega flow identifier. */ |
bd5131ba | 536 | const unsigned pmd_id; /* The 'core_id' of pmd thread owning this */ |
1c1e46ed | 537 | /* flow. */ |
72865317 | 538 | |
ed79f89a DDP |
539 | /* Number of references. |
540 | * The classifier owns one reference. | |
541 | * Any thread trying to keep a rule from being freed should hold its own | |
542 | * reference. */ | |
543 | struct ovs_refcount ref_cnt; | |
544 | ||
11e5cf1f | 545 | bool dead; |
241bad15 | 546 | uint32_t mark; /* Unique flow mark assigned to a flow */ |
11e5cf1f | 547 | |
1c1e46ed AW |
548 | /* Statistics. */ |
549 | struct dp_netdev_flow_stats stats; | |
8a4e3a85 | 550 | |
45c626a3 | 551 | /* Actions. */ |
61e7deb1 | 552 | OVSRCU_TYPE(struct dp_netdev_actions *) actions; |
0de8783a | 553 | |
11e5cf1f DDP |
554 | /* While processing a group of input packets, the datapath uses the next |
555 | * member to store a pointer to the output batch for the flow. It is | |
556 | * reset after the batch has been sent out (See dp_netdev_queue_batches(), | |
f7ce4811 PS |
557 | * packet_batch_per_flow_init() and packet_batch_per_flow_execute()). */ |
558 | struct packet_batch_per_flow *batch; | |
11e5cf1f | 559 | |
0de8783a | 560 | /* Packet classification. */ |
342b8904 | 561 | char *dp_extra_info; /* String to return in a flow dump/get. */ |
0de8783a JR |
562 | struct dpcls_rule cr; /* In owning dp_netdev's 'cls'. */ |
563 | /* 'cr' must be the last member. */ | |
72865317 BP |
564 | }; |
565 | ||
ed79f89a | 566 | static void dp_netdev_flow_unref(struct dp_netdev_flow *); |
9bbf1c3d | 567 | static bool dp_netdev_flow_ref(struct dp_netdev_flow *); |
70e5ed6f | 568 | static int dpif_netdev_flow_from_nlattrs(const struct nlattr *, uint32_t, |
f0fb825a | 569 | struct flow *, bool); |
8a4e3a85 | 570 | |
a84cb64a BP |
571 | /* A set of datapath actions within a "struct dp_netdev_flow". |
572 | * | |
573 | * | |
574 | * Thread-safety | |
575 | * ============= | |
576 | * | |
45c626a3 | 577 | * A struct dp_netdev_actions 'actions' is protected with RCU. */ |
a84cb64a | 578 | struct dp_netdev_actions { |
a84cb64a BP |
579 | /* These members are immutable: they do not change during the struct's |
580 | * lifetime. */ | |
a84cb64a | 581 | unsigned int size; /* Size of 'actions', in bytes. */ |
9ff55ae2 | 582 | struct nlattr actions[]; /* Sequence of OVS_ACTION_ATTR_* attributes. */ |
a84cb64a BP |
583 | }; |
584 | ||
585 | struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *, | |
586 | size_t); | |
61e7deb1 BP |
587 | struct dp_netdev_actions *dp_netdev_flow_get_actions( |
588 | const struct dp_netdev_flow *); | |
589 | static void dp_netdev_actions_free(struct dp_netdev_actions *); | |
a84cb64a | 590 | |
947dc567 | 591 | struct polled_queue { |
922b28d4 | 592 | struct dp_netdev_rxq *rxq; |
947dc567 | 593 | odp_port_t port_no; |
2fbadeb6 | 594 | bool emc_enabled; |
35c91567 DM |
595 | bool rxq_enabled; |
596 | uint64_t change_seq; | |
947dc567 DDP |
597 | }; |
598 | ||
ae7ad0a1 IM |
599 | /* Contained by struct dp_netdev_pmd_thread's 'poll_list' member. */ |
600 | struct rxq_poll { | |
947dc567 DDP |
601 | struct dp_netdev_rxq *rxq; |
602 | struct hmap_node node; | |
ae7ad0a1 IM |
603 | }; |
604 | ||
57eebbb4 DDP |
605 | /* Contained by struct dp_netdev_pmd_thread's 'send_port_cache', |
606 | * 'tnl_port_cache' or 'tx_ports'. */ | |
d0cca6c3 | 607 | struct tx_port { |
324c8374 IM |
608 | struct dp_netdev_port *port; |
609 | int qid; | |
610 | long long last_used; | |
d0cca6c3 | 611 | struct hmap_node node; |
c71ea3c4 | 612 | long long flush_time; |
009e0033 | 613 | struct dp_packet_batch output_pkts; |
58ed6df0 | 614 | struct dp_netdev_rxq *output_pkts_rxqs[NETDEV_MAX_BURST]; |
d0cca6c3 DDP |
615 | }; |
616 | ||
9df65060 VDA |
617 | /* Contained by struct tx_bond 'slave_buckets'. */ |
618 | struct slave_entry { | |
619 | odp_port_t slave_id; | |
620 | atomic_ullong n_packets; | |
621 | atomic_ullong n_bytes; | |
622 | }; | |
623 | ||
624 | /* Contained by struct dp_netdev_pmd_thread's 'tx_bonds'. */ | |
625 | struct tx_bond { | |
626 | struct cmap_node node; | |
627 | uint32_t bond_id; | |
628 | struct slave_entry slave_buckets[BOND_BUCKETS]; | |
629 | }; | |
630 | ||
b010be17 IM |
631 | /* A set of properties for the current processing loop that is not directly |
632 | * associated with the pmd thread itself, but with the packets being | |
633 | * processed or the short-term system configuration (for example, time). | |
634 | * Contained by struct dp_netdev_pmd_thread's 'ctx' member. */ | |
635 | struct dp_netdev_pmd_thread_ctx { | |
636 | /* Latest measured time. See 'pmd_thread_ctx_time_update()'. */ | |
637 | long long now; | |
58ed6df0 IM |
638 | /* RX queue from which last packet was received. */ |
639 | struct dp_netdev_rxq *last_rxq; | |
2fbadeb6 IM |
640 | /* EMC insertion probability context for the current processing cycle. */ |
641 | uint32_t emc_insert_min; | |
d0cca6c3 DDP |
642 | }; |
643 | ||
e4cfed38 PS |
644 | /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate |
645 | * the performance overhead of interrupt processing. Therefore netdev can | |
646 | * not implement rx-wait for these devices. dpif-netdev needs to poll | |
647 | * these device to check for recv buffer. pmd-thread does polling for | |
1c1e46ed | 648 | * devices assigned to itself. |
e4cfed38 PS |
649 | * |
650 | * DPDK used PMD for accessing NIC. | |
651 | * | |
65f13b50 AW |
652 | * Note, instance with cpu core id NON_PMD_CORE_ID will be reserved for |
653 | * I/O of all non-pmd threads. There will be no actual thread created | |
654 | * for the instance. | |
1c1e46ed | 655 | * |
1859876c BB |
656 | * Each struct has its own flow cache and classifier per managed ingress port. |
657 | * For packets received on ingress port, a look up is done on corresponding PMD | |
658 | * thread's flow cache and in case of a miss, lookup is performed in the | |
659 | * corresponding classifier of port. Packets are executed with the found | |
660 | * actions in either case. | |
1c1e46ed | 661 | * */ |
65f13b50 | 662 | struct dp_netdev_pmd_thread { |
d9d73f84 IM |
663 | struct dp_netdev *dp; |
664 | struct ovs_refcount ref_cnt; /* Every reference must be refcount'ed. */ | |
665 | struct cmap_node node; /* In 'dp->poll_threads'. */ | |
666 | ||
65f13b50 AW |
667 | /* Per thread exact-match cache. Note, the instance for cpu core |
668 | * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly | |
d0cca6c3 DDP |
669 | * need to be protected by 'non_pmd_mutex'. Every other instance |
670 | * will only be accessed by its own pmd thread. */ | |
60d8ccae | 671 | OVS_ALIGNED_VAR(CACHE_LINE_SIZE) struct dfc_cache flow_cache; |
1c1e46ed | 672 | |
3453b4d6 | 673 | /* Flow-Table and classifiers |
1c1e46ed AW |
674 | * |
675 | * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding | |
3453b4d6 JS |
676 | * changes to 'classifiers' must be made while still holding the |
677 | * 'flow_mutex'. | |
1c1e46ed AW |
678 | */ |
679 | struct ovs_mutex flow_mutex; | |
d9d73f84 IM |
680 | struct cmap flow_table OVS_GUARDED; /* Flow table. */ |
681 | ||
682 | /* One classifier per in_port polled by the pmd */ | |
683 | struct cmap classifiers; | |
684 | /* Periodically sort subtable vectors according to hit frequencies */ | |
685 | long long int next_optimization; | |
686 | /* End of the next time interval for which processing cycles | |
687 | are stored for each polled rxq. */ | |
688 | long long int rxq_next_cycle_store; | |
689 | ||
2a2c67b4 KT |
690 | /* Last interval timestamp. */ |
691 | uint64_t intrvl_tsc_prev; | |
692 | /* Last interval cycles. */ | |
693 | atomic_ullong intrvl_cycles; | |
694 | ||
b010be17 IM |
695 | /* Current context of the PMD thread. */ |
696 | struct dp_netdev_pmd_thread_ctx ctx; | |
d9d73f84 | 697 | |
d9d73f84 IM |
698 | struct seq *reload_seq; |
699 | uint64_t last_reload_seq; | |
ec61d470 IM |
700 | |
701 | /* These are atomic variables used as a synchronization and configuration | |
702 | * points for thread reload/exit. | |
703 | * | |
704 | * 'reload' atomic is the main one and it's used as a memory | |
705 | * synchronization point for all other knobs and data. | |
706 | * | |
707 | * For a thread that requests PMD reload: | |
708 | * | |
709 | * * All changes that should be visible to the PMD thread must be made | |
710 | * before setting the 'reload'. These changes could use any memory | |
711 | * ordering model including 'relaxed'. | |
712 | * * Setting the 'reload' atomic should occur in the same thread where | |
713 | * all other PMD configuration options updated. | |
714 | * * Setting the 'reload' atomic should be done with 'release' memory | |
715 | * ordering model or stricter. This will guarantee that all previous | |
716 | * changes (including non-atomic and 'relaxed') will be visible to | |
717 | * the PMD thread. | |
718 | * * To check that reload is done, thread should poll the 'reload' atomic | |
719 | * to become 'false'. Polling should be done with 'acquire' memory | |
720 | * ordering model or stricter. This ensures that PMD thread completed | |
721 | * the reload process. | |
722 | * | |
723 | * For the PMD thread: | |
724 | * | |
725 | * * PMD thread should read 'reload' atomic with 'acquire' memory | |
726 | * ordering model or stricter. This will guarantee that all changes | |
727 | * made before setting the 'reload' in the requesting thread will be | |
728 | * visible to the PMD thread. | |
729 | * * All other configuration data could be read with any memory | |
730 | * ordering model (including non-atomic and 'relaxed') but *only after* | |
731 | * reading the 'reload' atomic set to 'true'. | |
732 | * * When the PMD reload done, PMD should (optionally) set all the below | |
733 | * knobs except the 'reload' to their default ('false') values and | |
734 | * (mandatory), as the last step, set the 'reload' to 'false' using | |
735 | * 'release' memory ordering model or stricter. This will inform the | |
736 | * requesting thread that PMD has completed a reload cycle. | |
737 | */ | |
d9d73f84 | 738 | atomic_bool reload; /* Do we need to reload ports? */ |
6d9fead1 | 739 | atomic_bool wait_for_reload; /* Can we busy wait for the next reload? */ |
e2cafa86 | 740 | atomic_bool reload_tx_qid; /* Do we need to reload static_tx_qid? */ |
299c8d61 | 741 | atomic_bool exit; /* For terminating the pmd thread. */ |
ec61d470 | 742 | |
d9d73f84 IM |
743 | pthread_t thread; |
744 | unsigned core_id; /* CPU core id of this pmd thread. */ | |
745 | int numa_id; /* numa node id of this pmd thread. */ | |
746 | bool isolated; | |
747 | ||
748 | /* Queue id used by this pmd thread to send packets on all netdevs if | |
749 | * XPS disabled for this netdev. All static_tx_qid's are unique and less | |
750 | * than 'cmap_count(dp->poll_threads)'. */ | |
751 | uint32_t static_tx_qid; | |
752 | ||
c71ea3c4 IM |
753 | /* Number of filled output batches. */ |
754 | int n_output_batches; | |
755 | ||
d9d73f84 IM |
756 | struct ovs_mutex port_mutex; /* Mutex for 'poll_list' and 'tx_ports'. */ |
757 | /* List of rx queues to poll. */ | |
758 | struct hmap poll_list OVS_GUARDED; | |
759 | /* Map of 'tx_port's used for transmission. Written by the main thread, | |
760 | * read by the pmd thread. */ | |
761 | struct hmap tx_ports OVS_GUARDED; | |
762 | ||
9df65060 VDA |
763 | struct ovs_mutex bond_mutex; /* Protects updates of 'tx_bonds'. */ |
764 | /* Map of 'tx_bond's used for transmission. Written by the main thread | |
765 | * and read by the pmd thread. */ | |
766 | struct cmap tx_bonds; | |
767 | ||
d9d73f84 IM |
768 | /* These are thread-local copies of 'tx_ports'. One contains only tunnel |
769 | * ports (that support push_tunnel/pop_tunnel), the other contains ports | |
770 | * with at least one txq (that support send). A port can be in both. | |
771 | * | |
772 | * There are two separate maps to make sure that we don't try to execute | |
773 | * OUTPUT on a device which has 0 txqs or PUSH/POP on a non-tunnel device. | |
774 | * | |
775 | * The instances for cpu core NON_PMD_CORE_ID can be accessed by multiple | |
776 | * threads, and thusly need to be protected by 'non_pmd_mutex'. Every | |
777 | * other instance will only be accessed by its own pmd thread. */ | |
778 | struct hmap tnl_port_cache; | |
779 | struct hmap send_port_cache; | |
780 | ||
82a48ead JS |
781 | /* Keep track of detailed PMD performance statistics. */ |
782 | struct pmd_perf_stats perf_stats; | |
d9d73f84 | 783 | |
5bf84282 NK |
784 | /* Stats from previous iteration used by automatic pmd |
785 | * load balance logic. */ | |
786 | uint64_t prev_stats[PMD_N_STATS]; | |
787 | atomic_count pmd_overloaded; | |
788 | ||
d9d73f84 IM |
789 | /* Set to true if the pmd thread needs to be reloaded. */ |
790 | bool need_reload; | |
6c3eee82 BP |
791 | }; |
792 | ||
72865317 BP |
793 | /* Interface to netdev-based datapath. */ |
794 | struct dpif_netdev { | |
795 | struct dpif dpif; | |
796 | struct dp_netdev *dp; | |
d33ed218 | 797 | uint64_t last_port_seq; |
72865317 BP |
798 | }; |
799 | ||
8a4e3a85 | 800 | static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no, |
e9985d6a DDP |
801 | struct dp_netdev_port **portp) |
802 | OVS_REQUIRES(dp->port_mutex); | |
8a4e3a85 | 803 | static int get_port_by_name(struct dp_netdev *dp, const char *devname, |
e9985d6a DDP |
804 | struct dp_netdev_port **portp) |
805 | OVS_REQUIRES(dp->port_mutex); | |
8a4e3a85 BP |
806 | static void dp_netdev_free(struct dp_netdev *) |
807 | OVS_REQUIRES(dp_netdev_mutex); | |
8a4e3a85 BP |
808 | static int do_add_port(struct dp_netdev *dp, const char *devname, |
809 | const char *type, odp_port_t port_no) | |
59e6d833 | 810 | OVS_REQUIRES(dp->port_mutex); |
c40b890f | 811 | static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *) |
59e6d833 | 812 | OVS_REQUIRES(dp->port_mutex); |
614c4892 BP |
813 | static int dpif_netdev_open(const struct dpif_class *, const char *name, |
814 | bool create, struct dpif **); | |
65f13b50 | 815 | static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd, |
1895cc8d | 816 | struct dp_packet_batch *, |
7d7ded7a DB |
817 | bool should_steal, |
818 | const struct flow *flow, | |
4edb9ae9 | 819 | const struct nlattr *actions, |
b010be17 | 820 | size_t actions_len); |
65f13b50 | 821 | static void dp_netdev_input(struct dp_netdev_pmd_thread *, |
1895cc8d | 822 | struct dp_packet_batch *, odp_port_t port_no); |
a90ed026 | 823 | static void dp_netdev_recirculate(struct dp_netdev_pmd_thread *, |
1895cc8d | 824 | struct dp_packet_batch *); |
41ccaa24 | 825 | |
6b31e073 | 826 | static void dp_netdev_disable_upcall(struct dp_netdev *); |
ae7ad0a1 | 827 | static void dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd); |
65f13b50 | 828 | static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, |
00873463 DDP |
829 | struct dp_netdev *dp, unsigned core_id, |
830 | int numa_id); | |
1c1e46ed | 831 | static void dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd); |
e9985d6a DDP |
832 | static void dp_netdev_set_nonpmd(struct dp_netdev *dp) |
833 | OVS_REQUIRES(dp->port_mutex); | |
834 | ||
e32971b8 | 835 | static void *pmd_thread_main(void *); |
b19befae | 836 | static struct dp_netdev_pmd_thread *dp_netdev_get_pmd(struct dp_netdev *dp, |
bd5131ba | 837 | unsigned core_id); |
1c1e46ed AW |
838 | static struct dp_netdev_pmd_thread * |
839 | dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos); | |
140dd699 IM |
840 | static void dp_netdev_del_pmd(struct dp_netdev *dp, |
841 | struct dp_netdev_pmd_thread *pmd); | |
e32971b8 | 842 | static void dp_netdev_destroy_all_pmds(struct dp_netdev *dp, bool non_pmd); |
d0cca6c3 | 843 | static void dp_netdev_pmd_clear_ports(struct dp_netdev_pmd_thread *pmd); |
d0cca6c3 | 844 | static void dp_netdev_add_port_tx_to_pmd(struct dp_netdev_pmd_thread *pmd, |
e32971b8 DDP |
845 | struct dp_netdev_port *port) |
846 | OVS_REQUIRES(pmd->port_mutex); | |
847 | static void dp_netdev_del_port_tx_from_pmd(struct dp_netdev_pmd_thread *pmd, | |
848 | struct tx_port *tx) | |
849 | OVS_REQUIRES(pmd->port_mutex); | |
d0cca6c3 | 850 | static void dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread *pmd, |
947dc567 DDP |
851 | struct dp_netdev_rxq *rxq) |
852 | OVS_REQUIRES(pmd->port_mutex); | |
e32971b8 DDP |
853 | static void dp_netdev_del_rxq_from_pmd(struct dp_netdev_pmd_thread *pmd, |
854 | struct rxq_poll *poll) | |
855 | OVS_REQUIRES(pmd->port_mutex); | |
c71ea3c4 IM |
856 | static int |
857 | dp_netdev_pmd_flush_output_packets(struct dp_netdev_pmd_thread *pmd, | |
858 | bool force); | |
9df65060 VDA |
859 | static void dp_netdev_add_bond_tx_to_pmd(struct dp_netdev_pmd_thread *pmd, |
860 | struct tx_bond *bond, bool update) | |
861 | OVS_EXCLUDED(pmd->bond_mutex); | |
862 | static void dp_netdev_del_bond_tx_from_pmd(struct dp_netdev_pmd_thread *pmd, | |
863 | uint32_t bond_id) | |
864 | OVS_EXCLUDED(pmd->bond_mutex); | |
009e0033 | 865 | |
e32971b8 | 866 | static void reconfigure_datapath(struct dp_netdev *dp) |
3eb67853 | 867 | OVS_REQUIRES(dp->port_mutex); |
1c1e46ed AW |
868 | static bool dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd); |
869 | static void dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd); | |
870 | static void dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd); | |
d0cca6c3 DDP |
871 | static void pmd_load_cached_ports(struct dp_netdev_pmd_thread *pmd) |
872 | OVS_REQUIRES(pmd->port_mutex); | |
3453b4d6 | 873 | static inline void |
4809891b KT |
874 | dp_netdev_pmd_try_optimize(struct dp_netdev_pmd_thread *pmd, |
875 | struct polled_queue *poll_list, int poll_cnt); | |
876 | static void | |
877 | dp_netdev_rxq_set_cycles(struct dp_netdev_rxq *rx, | |
878 | enum rxq_cycles_counter_type type, | |
879 | unsigned long long cycles); | |
880 | static uint64_t | |
881 | dp_netdev_rxq_get_cycles(struct dp_netdev_rxq *rx, | |
882 | enum rxq_cycles_counter_type type); | |
883 | static void | |
884 | dp_netdev_rxq_set_intrvl_cycles(struct dp_netdev_rxq *rx, | |
885 | unsigned long long cycles); | |
655856ef KT |
886 | static uint64_t |
887 | dp_netdev_rxq_get_intrvl_cycles(struct dp_netdev_rxq *rx, unsigned idx); | |
324c8374 IM |
888 | static void |
889 | dpif_netdev_xps_revalidate_pmd(const struct dp_netdev_pmd_thread *pmd, | |
b010be17 | 890 | bool purge); |
324c8374 | 891 | static int dpif_netdev_xps_get_tx_qid(const struct dp_netdev_pmd_thread *pmd, |
b010be17 | 892 | struct tx_port *tx); |
324c8374 | 893 | |
67ad54cb | 894 | static inline bool emc_entry_alive(struct emc_entry *ce); |
9bbf1c3d | 895 | static void emc_clear_entry(struct emc_entry *ce); |
60d8ccae | 896 | static void smc_clear_entry(struct smc_bucket *b, int idx); |
9bbf1c3d | 897 | |
cd995c73 | 898 | static void dp_netdev_request_reconfigure(struct dp_netdev *dp); |
79f36875 JS |
899 | static inline bool |
900 | pmd_perf_metrics_enabled(const struct dp_netdev_pmd_thread *pmd); | |
02bb2824 YL |
901 | static void queue_netdev_flow_del(struct dp_netdev_pmd_thread *pmd, |
902 | struct dp_netdev_flow *flow); | |
cd995c73 | 903 | |
9bbf1c3d DDP |
904 | static void |
905 | emc_cache_init(struct emc_cache *flow_cache) | |
906 | { | |
907 | int i; | |
908 | ||
67ad54cb | 909 | flow_cache->sweep_idx = 0; |
9bbf1c3d DDP |
910 | for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) { |
911 | flow_cache->entries[i].flow = NULL; | |
0de8783a | 912 | flow_cache->entries[i].key.hash = 0; |
09b0fa9c | 913 | flow_cache->entries[i].key.len = sizeof(struct miniflow); |
5fcff47b | 914 | flowmap_init(&flow_cache->entries[i].key.mf.map); |
9bbf1c3d DDP |
915 | } |
916 | } | |
917 | ||
60d8ccae YW |
918 | static void |
919 | smc_cache_init(struct smc_cache *smc_cache) | |
920 | { | |
921 | int i, j; | |
922 | for (i = 0; i < SMC_BUCKET_CNT; i++) { | |
923 | for (j = 0; j < SMC_ENTRY_PER_BUCKET; j++) { | |
924 | smc_cache->buckets[i].flow_idx[j] = UINT16_MAX; | |
925 | } | |
926 | } | |
927 | } | |
928 | ||
929 | static void | |
930 | dfc_cache_init(struct dfc_cache *flow_cache) | |
931 | { | |
932 | emc_cache_init(&flow_cache->emc_cache); | |
933 | smc_cache_init(&flow_cache->smc_cache); | |
934 | } | |
935 | ||
9bbf1c3d DDP |
936 | static void |
937 | emc_cache_uninit(struct emc_cache *flow_cache) | |
938 | { | |
939 | int i; | |
940 | ||
941 | for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) { | |
942 | emc_clear_entry(&flow_cache->entries[i]); | |
943 | } | |
944 | } | |
945 | ||
60d8ccae YW |
946 | static void |
947 | smc_cache_uninit(struct smc_cache *smc) | |
948 | { | |
949 | int i, j; | |
950 | ||
951 | for (i = 0; i < SMC_BUCKET_CNT; i++) { | |
952 | for (j = 0; j < SMC_ENTRY_PER_BUCKET; j++) { | |
953 | smc_clear_entry(&(smc->buckets[i]), j); | |
954 | } | |
955 | } | |
956 | } | |
957 | ||
958 | static void | |
959 | dfc_cache_uninit(struct dfc_cache *flow_cache) | |
960 | { | |
961 | smc_cache_uninit(&flow_cache->smc_cache); | |
962 | emc_cache_uninit(&flow_cache->emc_cache); | |
963 | } | |
964 | ||
67ad54cb AW |
965 | /* Check and clear dead flow references slowly (one entry at each |
966 | * invocation). */ | |
967 | static void | |
968 | emc_cache_slow_sweep(struct emc_cache *flow_cache) | |
969 | { | |
970 | struct emc_entry *entry = &flow_cache->entries[flow_cache->sweep_idx]; | |
971 | ||
972 | if (!emc_entry_alive(entry)) { | |
973 | emc_clear_entry(entry); | |
974 | } | |
975 | flow_cache->sweep_idx = (flow_cache->sweep_idx + 1) & EM_FLOW_HASH_MASK; | |
976 | } | |
977 | ||
b010be17 IM |
978 | /* Updates the time in PMD threads context and should be called in three cases: |
979 | * | |
980 | * 1. PMD structure initialization: | |
981 | * - dp_netdev_configure_pmd() | |
982 | * | |
983 | * 2. Before processing of the new packet batch: | |
984 | * - dpif_netdev_execute() | |
009e0033 | 985 | * - dp_netdev_process_rxq_port() |
b010be17 IM |
986 | * |
987 | * 3. At least once per polling iteration in main polling threads if no | |
988 | * packets received on current iteration: | |
989 | * - dpif_netdev_run() | |
990 | * - pmd_thread_main() | |
991 | * | |
992 | * 'pmd->ctx.now' should be used without update in all other cases if possible. | |
993 | */ | |
994 | static inline void | |
995 | pmd_thread_ctx_time_update(struct dp_netdev_pmd_thread *pmd) | |
996 | { | |
05f9e707 | 997 | pmd->ctx.now = time_usec(); |
b010be17 IM |
998 | } |
999 | ||
c4ea7529 BP |
1000 | /* Returns true if 'dpif' is a netdev or dummy dpif, false otherwise. */ |
1001 | bool | |
1002 | dpif_is_netdev(const struct dpif *dpif) | |
1003 | { | |
1004 | return dpif->dpif_class->open == dpif_netdev_open; | |
1005 | } | |
1006 | ||
72865317 BP |
1007 | static struct dpif_netdev * |
1008 | dpif_netdev_cast(const struct dpif *dpif) | |
1009 | { | |
c4ea7529 | 1010 | ovs_assert(dpif_is_netdev(dpif)); |
72865317 BP |
1011 | return CONTAINER_OF(dpif, struct dpif_netdev, dpif); |
1012 | } | |
1013 | ||
1014 | static struct dp_netdev * | |
1015 | get_dp_netdev(const struct dpif *dpif) | |
1016 | { | |
1017 | return dpif_netdev_cast(dpif)->dp; | |
1018 | } | |
6553d06b DDP |
1019 | \f |
1020 | enum pmd_info_type { | |
ce179f11 IM |
1021 | PMD_INFO_SHOW_STATS, /* Show how cpu cycles are spent. */ |
1022 | PMD_INFO_CLEAR_STATS, /* Set the cycles count to 0. */ | |
79f36875 JS |
1023 | PMD_INFO_SHOW_RXQ, /* Show poll lists of pmd threads. */ |
1024 | PMD_INFO_PERF_SHOW, /* Show pmd performance details. */ | |
6553d06b DDP |
1025 | }; |
1026 | ||
1027 | static void | |
82a48ead | 1028 | format_pmd_thread(struct ds *reply, struct dp_netdev_pmd_thread *pmd) |
6553d06b | 1029 | { |
6553d06b DDP |
1030 | ds_put_cstr(reply, (pmd->core_id == NON_PMD_CORE_ID) |
1031 | ? "main thread" : "pmd thread"); | |
6553d06b DDP |
1032 | if (pmd->numa_id != OVS_NUMA_UNSPEC) { |
1033 | ds_put_format(reply, " numa_id %d", pmd->numa_id); | |
1034 | } | |
d5c199ea | 1035 | if (pmd->core_id != OVS_CORE_UNSPEC && pmd->core_id != NON_PMD_CORE_ID) { |
bd5131ba | 1036 | ds_put_format(reply, " core_id %u", pmd->core_id); |
6553d06b DDP |
1037 | } |
1038 | ds_put_cstr(reply, ":\n"); | |
82a48ead JS |
1039 | } |
1040 | ||
1041 | static void | |
1042 | pmd_info_show_stats(struct ds *reply, | |
1043 | struct dp_netdev_pmd_thread *pmd) | |
1044 | { | |
1045 | uint64_t stats[PMD_N_STATS]; | |
1046 | uint64_t total_cycles, total_packets; | |
1047 | double passes_per_pkt = 0; | |
1048 | double lookups_per_hit = 0; | |
1049 | double packets_per_batch = 0; | |
1050 | ||
1051 | pmd_perf_read_counters(&pmd->perf_stats, stats); | |
1052 | total_cycles = stats[PMD_CYCLES_ITER_IDLE] | |
1053 | + stats[PMD_CYCLES_ITER_BUSY]; | |
1054 | total_packets = stats[PMD_STAT_RECV]; | |
1055 | ||
1056 | format_pmd_thread(reply, pmd); | |
6553d06b | 1057 | |
82a48ead JS |
1058 | if (total_packets > 0) { |
1059 | passes_per_pkt = (total_packets + stats[PMD_STAT_RECIRC]) | |
1060 | / (double) total_packets; | |
cc4891f3 | 1061 | } |
82a48ead JS |
1062 | if (stats[PMD_STAT_MASKED_HIT] > 0) { |
1063 | lookups_per_hit = stats[PMD_STAT_MASKED_LOOKUP] | |
1064 | / (double) stats[PMD_STAT_MASKED_HIT]; | |
1065 | } | |
1066 | if (stats[PMD_STAT_SENT_BATCHES] > 0) { | |
1067 | packets_per_batch = stats[PMD_STAT_SENT_PKTS] | |
1068 | / (double) stats[PMD_STAT_SENT_BATCHES]; | |
cc4891f3 IM |
1069 | } |
1070 | ||
6553d06b | 1071 | ds_put_format(reply, |
5a0e4aec BP |
1072 | " packets received: %"PRIu64"\n" |
1073 | " packet recirculations: %"PRIu64"\n" | |
1074 | " avg. datapath passes per packet: %.02f\n" | |
1075 | " emc hits: %"PRIu64"\n" | |
60d8ccae | 1076 | " smc hits: %"PRIu64"\n" |
5a0e4aec BP |
1077 | " megaflow hits: %"PRIu64"\n" |
1078 | " avg. subtable lookups per megaflow hit: %.02f\n" | |
1079 | " miss with success upcall: %"PRIu64"\n" | |
1080 | " miss with failed upcall: %"PRIu64"\n" | |
1081 | " avg. packets per output batch: %.02f\n", | |
82a48ead JS |
1082 | total_packets, stats[PMD_STAT_RECIRC], |
1083 | passes_per_pkt, stats[PMD_STAT_EXACT_HIT], | |
60d8ccae | 1084 | stats[PMD_STAT_SMC_HIT], |
82a48ead JS |
1085 | stats[PMD_STAT_MASKED_HIT], lookups_per_hit, |
1086 | stats[PMD_STAT_MISS], stats[PMD_STAT_LOST], | |
cc4891f3 | 1087 | packets_per_batch); |
6553d06b DDP |
1088 | |
1089 | if (total_cycles == 0) { | |
1090 | return; | |
1091 | } | |
1092 | ||
1093 | ds_put_format(reply, | |
5a0e4aec BP |
1094 | " idle cycles: %"PRIu64" (%.02f%%)\n" |
1095 | " processing cycles: %"PRIu64" (%.02f%%)\n", | |
82a48ead JS |
1096 | stats[PMD_CYCLES_ITER_IDLE], |
1097 | stats[PMD_CYCLES_ITER_IDLE] / (double) total_cycles * 100, | |
1098 | stats[PMD_CYCLES_ITER_BUSY], | |
1099 | stats[PMD_CYCLES_ITER_BUSY] / (double) total_cycles * 100); | |
6553d06b DDP |
1100 | |
1101 | if (total_packets == 0) { | |
1102 | return; | |
1103 | } | |
1104 | ||
1105 | ds_put_format(reply, | |
5a0e4aec | 1106 | " avg cycles per packet: %.02f (%"PRIu64"/%"PRIu64")\n", |
82a48ead | 1107 | total_cycles / (double) total_packets, |
6553d06b DDP |
1108 | total_cycles, total_packets); |
1109 | ||
1110 | ds_put_format(reply, | |
5a0e4aec | 1111 | " avg processing cycles per packet: " |
82a48ead JS |
1112 | "%.02f (%"PRIu64"/%"PRIu64")\n", |
1113 | stats[PMD_CYCLES_ITER_BUSY] / (double) total_packets, | |
1114 | stats[PMD_CYCLES_ITER_BUSY], total_packets); | |
6553d06b DDP |
1115 | } |
1116 | ||
79f36875 JS |
1117 | static void |
1118 | pmd_info_show_perf(struct ds *reply, | |
1119 | struct dp_netdev_pmd_thread *pmd, | |
1120 | struct pmd_perf_params *par) | |
1121 | { | |
1122 | if (pmd->core_id != NON_PMD_CORE_ID) { | |
1123 | char *time_str = | |
1124 | xastrftime_msec("%H:%M:%S.###", time_wall_msec(), true); | |
1125 | long long now = time_msec(); | |
1126 | double duration = (now - pmd->perf_stats.start_ms) / 1000.0; | |
1127 | ||
1128 | ds_put_cstr(reply, "\n"); | |
1129 | ds_put_format(reply, "Time: %s\n", time_str); | |
1130 | ds_put_format(reply, "Measurement duration: %.3f s\n", duration); | |
1131 | ds_put_cstr(reply, "\n"); | |
1132 | format_pmd_thread(reply, pmd); | |
1133 | ds_put_cstr(reply, "\n"); | |
1134 | pmd_perf_format_overall_stats(reply, &pmd->perf_stats, duration); | |
1135 | if (pmd_perf_metrics_enabled(pmd)) { | |
1136 | /* Prevent parallel clearing of perf metrics. */ | |
1137 | ovs_mutex_lock(&pmd->perf_stats.clear_mutex); | |
1138 | if (par->histograms) { | |
1139 | ds_put_cstr(reply, "\n"); | |
1140 | pmd_perf_format_histograms(reply, &pmd->perf_stats); | |
1141 | } | |
1142 | if (par->iter_hist_len > 0) { | |
1143 | ds_put_cstr(reply, "\n"); | |
1144 | pmd_perf_format_iteration_history(reply, &pmd->perf_stats, | |
1145 | par->iter_hist_len); | |
1146 | } | |
1147 | if (par->ms_hist_len > 0) { | |
1148 | ds_put_cstr(reply, "\n"); | |
1149 | pmd_perf_format_ms_history(reply, &pmd->perf_stats, | |
1150 | par->ms_hist_len); | |
1151 | } | |
1152 | ovs_mutex_unlock(&pmd->perf_stats.clear_mutex); | |
1153 | } | |
1154 | free(time_str); | |
1155 | } | |
1156 | } | |
1157 | ||
947dc567 DDP |
1158 | static int |
1159 | compare_poll_list(const void *a_, const void *b_) | |
1160 | { | |
1161 | const struct rxq_poll *a = a_; | |
1162 | const struct rxq_poll *b = b_; | |
1163 | ||
1164 | const char *namea = netdev_rxq_get_name(a->rxq->rx); | |
1165 | const char *nameb = netdev_rxq_get_name(b->rxq->rx); | |
1166 | ||
1167 | int cmp = strcmp(namea, nameb); | |
1168 | if (!cmp) { | |
1169 | return netdev_rxq_get_queue_id(a->rxq->rx) | |
1170 | - netdev_rxq_get_queue_id(b->rxq->rx); | |
1171 | } else { | |
1172 | return cmp; | |
1173 | } | |
1174 | } | |
1175 | ||
1176 | static void | |
1177 | sorted_poll_list(struct dp_netdev_pmd_thread *pmd, struct rxq_poll **list, | |
1178 | size_t *n) | |
216abd28 | 1179 | OVS_REQUIRES(pmd->port_mutex) |
947dc567 DDP |
1180 | { |
1181 | struct rxq_poll *ret, *poll; | |
1182 | size_t i; | |
1183 | ||
1184 | *n = hmap_count(&pmd->poll_list); | |
1185 | if (!*n) { | |
1186 | ret = NULL; | |
1187 | } else { | |
1188 | ret = xcalloc(*n, sizeof *ret); | |
1189 | i = 0; | |
1190 | HMAP_FOR_EACH (poll, node, &pmd->poll_list) { | |
1191 | ret[i] = *poll; | |
1192 | i++; | |
1193 | } | |
1194 | ovs_assert(i == *n); | |
1cc1b5f6 | 1195 | qsort(ret, *n, sizeof *ret, compare_poll_list); |
947dc567 DDP |
1196 | } |
1197 | ||
947dc567 DDP |
1198 | *list = ret; |
1199 | } | |
1200 | ||
ce179f11 IM |
1201 | static void |
1202 | pmd_info_show_rxq(struct ds *reply, struct dp_netdev_pmd_thread *pmd) | |
1203 | { | |
1204 | if (pmd->core_id != NON_PMD_CORE_ID) { | |
947dc567 | 1205 | struct rxq_poll *list; |
2a2c67b4 KT |
1206 | size_t n_rxq; |
1207 | uint64_t total_cycles = 0; | |
ce179f11 | 1208 | |
3eb67853 | 1209 | ds_put_format(reply, |
5a0e4aec | 1210 | "pmd thread numa_id %d core_id %u:\n isolated : %s\n", |
3eb67853 IM |
1211 | pmd->numa_id, pmd->core_id, (pmd->isolated) |
1212 | ? "true" : "false"); | |
ce179f11 | 1213 | |
d0cca6c3 | 1214 | ovs_mutex_lock(&pmd->port_mutex); |
2a2c67b4 | 1215 | sorted_poll_list(pmd, &list, &n_rxq); |
ce179f11 | 1216 | |
2a2c67b4 KT |
1217 | /* Get the total pmd cycles for an interval. */ |
1218 | atomic_read_relaxed(&pmd->intrvl_cycles, &total_cycles); | |
1219 | /* Estimate the cycles to cover all intervals. */ | |
1220 | total_cycles *= PMD_RXQ_INTERVAL_MAX; | |
1221 | ||
1222 | for (int i = 0; i < n_rxq; i++) { | |
1223 | struct dp_netdev_rxq *rxq = list[i].rxq; | |
1224 | const char *name = netdev_rxq_get_name(rxq->rx); | |
1225 | uint64_t proc_cycles = 0; | |
1226 | ||
1227 | for (int j = 0; j < PMD_RXQ_INTERVAL_MAX; j++) { | |
1228 | proc_cycles += dp_netdev_rxq_get_intrvl_cycles(rxq, j); | |
ce179f11 | 1229 | } |
5a0e4aec | 1230 | ds_put_format(reply, " port: %-16s queue-id: %2d", name, |
947dc567 | 1231 | netdev_rxq_get_queue_id(list[i].rxq->rx)); |
35c91567 DM |
1232 | ds_put_format(reply, " %s", netdev_rxq_enabled(list[i].rxq->rx) |
1233 | ? "(enabled) " : "(disabled)"); | |
5a0e4aec | 1234 | ds_put_format(reply, " pmd usage: "); |
2a2c67b4 KT |
1235 | if (total_cycles) { |
1236 | ds_put_format(reply, "%2"PRIu64"", | |
1237 | proc_cycles * 100 / total_cycles); | |
1238 | ds_put_cstr(reply, " %"); | |
1239 | } else { | |
1240 | ds_put_format(reply, "%s", "NOT AVAIL"); | |
1241 | } | |
1242 | ds_put_cstr(reply, "\n"); | |
ce179f11 | 1243 | } |
d0cca6c3 | 1244 | ovs_mutex_unlock(&pmd->port_mutex); |
947dc567 | 1245 | free(list); |
ce179f11 IM |
1246 | } |
1247 | } | |
1248 | ||
34d8e04b EC |
1249 | static int |
1250 | compare_poll_thread_list(const void *a_, const void *b_) | |
1251 | { | |
1252 | const struct dp_netdev_pmd_thread *a, *b; | |
1253 | ||
1254 | a = *(struct dp_netdev_pmd_thread **)a_; | |
1255 | b = *(struct dp_netdev_pmd_thread **)b_; | |
1256 | ||
1257 | if (a->core_id < b->core_id) { | |
1258 | return -1; | |
1259 | } | |
1260 | if (a->core_id > b->core_id) { | |
1261 | return 1; | |
1262 | } | |
1263 | return 0; | |
1264 | } | |
1265 | ||
1266 | /* Create a sorted list of pmd's from the dp->poll_threads cmap. We can use | |
1267 | * this list, as long as we do not go to quiescent state. */ | |
1268 | static void | |
1269 | sorted_poll_thread_list(struct dp_netdev *dp, | |
1270 | struct dp_netdev_pmd_thread ***list, | |
1271 | size_t *n) | |
1272 | { | |
1273 | struct dp_netdev_pmd_thread *pmd; | |
1274 | struct dp_netdev_pmd_thread **pmd_list; | |
1275 | size_t k = 0, n_pmds; | |
1276 | ||
1277 | n_pmds = cmap_count(&dp->poll_threads); | |
1278 | pmd_list = xcalloc(n_pmds, sizeof *pmd_list); | |
1279 | ||
1280 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
1281 | if (k >= n_pmds) { | |
1282 | break; | |
1283 | } | |
1284 | pmd_list[k++] = pmd; | |
1285 | } | |
1286 | ||
1287 | qsort(pmd_list, k, sizeof *pmd_list, compare_poll_thread_list); | |
1288 | ||
1289 | *list = pmd_list; | |
1290 | *n = k; | |
1291 | } | |
1292 | ||
cd995c73 KT |
1293 | static void |
1294 | dpif_netdev_pmd_rebalance(struct unixctl_conn *conn, int argc, | |
1295 | const char *argv[], void *aux OVS_UNUSED) | |
1296 | { | |
1297 | struct ds reply = DS_EMPTY_INITIALIZER; | |
1298 | struct dp_netdev *dp = NULL; | |
1299 | ||
1300 | ovs_mutex_lock(&dp_netdev_mutex); | |
1301 | ||
1302 | if (argc == 2) { | |
1303 | dp = shash_find_data(&dp_netdevs, argv[1]); | |
1304 | } else if (shash_count(&dp_netdevs) == 1) { | |
1305 | /* There's only one datapath */ | |
1306 | dp = shash_first(&dp_netdevs)->data; | |
1307 | } | |
1308 | ||
1309 | if (!dp) { | |
1310 | ovs_mutex_unlock(&dp_netdev_mutex); | |
1311 | unixctl_command_reply_error(conn, | |
1312 | "please specify an existing datapath"); | |
1313 | return; | |
1314 | } | |
1315 | ||
1316 | dp_netdev_request_reconfigure(dp); | |
1317 | ovs_mutex_unlock(&dp_netdev_mutex); | |
1318 | ds_put_cstr(&reply, "pmd rxq rebalance requested.\n"); | |
1319 | unixctl_command_reply(conn, ds_cstr(&reply)); | |
1320 | ds_destroy(&reply); | |
1321 | } | |
1322 | ||
6553d06b DDP |
1323 | static void |
1324 | dpif_netdev_pmd_info(struct unixctl_conn *conn, int argc, const char *argv[], | |
1325 | void *aux) | |
1326 | { | |
1327 | struct ds reply = DS_EMPTY_INITIALIZER; | |
34d8e04b | 1328 | struct dp_netdev_pmd_thread **pmd_list; |
6553d06b DDP |
1329 | struct dp_netdev *dp = NULL; |
1330 | enum pmd_info_type type = *(enum pmd_info_type *) aux; | |
82a48ead JS |
1331 | unsigned int core_id; |
1332 | bool filter_on_pmd = false; | |
1333 | size_t n; | |
6553d06b DDP |
1334 | |
1335 | ovs_mutex_lock(&dp_netdev_mutex); | |
1336 | ||
82a48ead | 1337 | while (argc > 1) { |
79f36875 | 1338 | if (!strcmp(argv[1], "-pmd") && argc > 2) { |
82a48ead JS |
1339 | if (str_to_uint(argv[2], 10, &core_id)) { |
1340 | filter_on_pmd = true; | |
1341 | } | |
1342 | argc -= 2; | |
1343 | argv += 2; | |
1344 | } else { | |
1345 | dp = shash_find_data(&dp_netdevs, argv[1]); | |
1346 | argc -= 1; | |
1347 | argv += 1; | |
1348 | } | |
6553d06b DDP |
1349 | } |
1350 | ||
1351 | if (!dp) { | |
82a48ead JS |
1352 | if (shash_count(&dp_netdevs) == 1) { |
1353 | /* There's only one datapath */ | |
1354 | dp = shash_first(&dp_netdevs)->data; | |
1355 | } else { | |
1356 | ovs_mutex_unlock(&dp_netdev_mutex); | |
1357 | unixctl_command_reply_error(conn, | |
1358 | "please specify an existing datapath"); | |
1359 | return; | |
1360 | } | |
6553d06b DDP |
1361 | } |
1362 | ||
34d8e04b EC |
1363 | sorted_poll_thread_list(dp, &pmd_list, &n); |
1364 | for (size_t i = 0; i < n; i++) { | |
1365 | struct dp_netdev_pmd_thread *pmd = pmd_list[i]; | |
1366 | if (!pmd) { | |
1367 | break; | |
1368 | } | |
82a48ead JS |
1369 | if (filter_on_pmd && pmd->core_id != core_id) { |
1370 | continue; | |
1371 | } | |
ce179f11 IM |
1372 | if (type == PMD_INFO_SHOW_RXQ) { |
1373 | pmd_info_show_rxq(&reply, pmd); | |
82a48ead JS |
1374 | } else if (type == PMD_INFO_CLEAR_STATS) { |
1375 | pmd_perf_stats_clear(&pmd->perf_stats); | |
1376 | } else if (type == PMD_INFO_SHOW_STATS) { | |
1377 | pmd_info_show_stats(&reply, pmd); | |
79f36875 JS |
1378 | } else if (type == PMD_INFO_PERF_SHOW) { |
1379 | pmd_info_show_perf(&reply, pmd, (struct pmd_perf_params *)aux); | |
6553d06b DDP |
1380 | } |
1381 | } | |
34d8e04b | 1382 | free(pmd_list); |
6553d06b DDP |
1383 | |
1384 | ovs_mutex_unlock(&dp_netdev_mutex); | |
1385 | ||
1386 | unixctl_command_reply(conn, ds_cstr(&reply)); | |
1387 | ds_destroy(&reply); | |
1388 | } | |
79f36875 JS |
1389 | |
1390 | static void | |
1391 | pmd_perf_show_cmd(struct unixctl_conn *conn, int argc, | |
1392 | const char *argv[], | |
1393 | void *aux OVS_UNUSED) | |
1394 | { | |
1395 | struct pmd_perf_params par; | |
1396 | long int it_hist = 0, ms_hist = 0; | |
1397 | par.histograms = true; | |
1398 | ||
1399 | while (argc > 1) { | |
1400 | if (!strcmp(argv[1], "-nh")) { | |
1401 | par.histograms = false; | |
1402 | argc -= 1; | |
1403 | argv += 1; | |
1404 | } else if (!strcmp(argv[1], "-it") && argc > 2) { | |
1405 | it_hist = strtol(argv[2], NULL, 10); | |
1406 | if (it_hist < 0) { | |
1407 | it_hist = 0; | |
1408 | } else if (it_hist > HISTORY_LEN) { | |
1409 | it_hist = HISTORY_LEN; | |
1410 | } | |
1411 | argc -= 2; | |
1412 | argv += 2; | |
1413 | } else if (!strcmp(argv[1], "-ms") && argc > 2) { | |
1414 | ms_hist = strtol(argv[2], NULL, 10); | |
1415 | if (ms_hist < 0) { | |
1416 | ms_hist = 0; | |
1417 | } else if (ms_hist > HISTORY_LEN) { | |
1418 | ms_hist = HISTORY_LEN; | |
1419 | } | |
1420 | argc -= 2; | |
1421 | argv += 2; | |
1422 | } else { | |
1423 | break; | |
1424 | } | |
1425 | } | |
1426 | par.iter_hist_len = it_hist; | |
1427 | par.ms_hist_len = ms_hist; | |
1428 | par.command_type = PMD_INFO_PERF_SHOW; | |
1429 | dpif_netdev_pmd_info(conn, argc, argv, &par); | |
1430 | } | |
9df65060 VDA |
1431 | |
1432 | static void | |
1433 | dpif_netdev_bond_show(struct unixctl_conn *conn, int argc, | |
1434 | const char *argv[], void *aux OVS_UNUSED) | |
1435 | { | |
1436 | struct ds reply = DS_EMPTY_INITIALIZER; | |
1437 | struct dp_netdev *dp = NULL; | |
1438 | ||
1439 | ovs_mutex_lock(&dp_netdev_mutex); | |
1440 | if (argc == 2) { | |
1441 | dp = shash_find_data(&dp_netdevs, argv[1]); | |
1442 | } else if (shash_count(&dp_netdevs) == 1) { | |
1443 | /* There's only one datapath. */ | |
1444 | dp = shash_first(&dp_netdevs)->data; | |
1445 | } | |
1446 | if (!dp) { | |
1447 | ovs_mutex_unlock(&dp_netdev_mutex); | |
1448 | unixctl_command_reply_error(conn, | |
1449 | "please specify an existing datapath"); | |
1450 | return; | |
1451 | } | |
1452 | ||
1453 | if (cmap_count(&dp->tx_bonds) > 0) { | |
1454 | struct tx_bond *dp_bond_entry; | |
1455 | uint32_t slave_id; | |
1456 | ||
1457 | ds_put_cstr(&reply, "Bonds:\n"); | |
1458 | CMAP_FOR_EACH (dp_bond_entry, node, &dp->tx_bonds) { | |
1459 | ds_put_format(&reply, " bond-id %"PRIu32":\n", | |
1460 | dp_bond_entry->bond_id); | |
1461 | for (int bucket = 0; bucket < BOND_BUCKETS; bucket++) { | |
1462 | slave_id = | |
1463 | odp_to_u32(dp_bond_entry->slave_buckets[bucket].slave_id); | |
1464 | ds_put_format(&reply, " bucket %d - slave %"PRIu32"\n", | |
1465 | bucket, slave_id); | |
1466 | } | |
1467 | } | |
1468 | } | |
1469 | ovs_mutex_unlock(&dp_netdev_mutex); | |
1470 | unixctl_command_reply(conn, ds_cstr(&reply)); | |
1471 | ds_destroy(&reply); | |
1472 | } | |
1473 | ||
6553d06b DDP |
1474 | \f |
1475 | static int | |
1476 | dpif_netdev_init(void) | |
1477 | { | |
1478 | static enum pmd_info_type show_aux = PMD_INFO_SHOW_STATS, | |
ce179f11 IM |
1479 | clear_aux = PMD_INFO_CLEAR_STATS, |
1480 | poll_aux = PMD_INFO_SHOW_RXQ; | |
6553d06b | 1481 | |
82a48ead JS |
1482 | unixctl_command_register("dpif-netdev/pmd-stats-show", "[-pmd core] [dp]", |
1483 | 0, 3, dpif_netdev_pmd_info, | |
6553d06b | 1484 | (void *)&show_aux); |
82a48ead JS |
1485 | unixctl_command_register("dpif-netdev/pmd-stats-clear", "[-pmd core] [dp]", |
1486 | 0, 3, dpif_netdev_pmd_info, | |
6553d06b | 1487 | (void *)&clear_aux); |
82a48ead JS |
1488 | unixctl_command_register("dpif-netdev/pmd-rxq-show", "[-pmd core] [dp]", |
1489 | 0, 3, dpif_netdev_pmd_info, | |
ce179f11 | 1490 | (void *)&poll_aux); |
79f36875 JS |
1491 | unixctl_command_register("dpif-netdev/pmd-perf-show", |
1492 | "[-nh] [-it iter-history-len]" | |
1493 | " [-ms ms-history-len]" | |
1494 | " [-pmd core] [dp]", | |
1495 | 0, 8, pmd_perf_show_cmd, | |
1496 | NULL); | |
cd995c73 KT |
1497 | unixctl_command_register("dpif-netdev/pmd-rxq-rebalance", "[dp]", |
1498 | 0, 1, dpif_netdev_pmd_rebalance, | |
1499 | NULL); | |
7178fefb JS |
1500 | unixctl_command_register("dpif-netdev/pmd-perf-log-set", |
1501 | "on|off [-b before] [-a after] [-e|-ne] " | |
1502 | "[-us usec] [-q qlen]", | |
1503 | 0, 10, pmd_perf_log_set_cmd, | |
1504 | NULL); | |
9df65060 VDA |
1505 | unixctl_command_register("dpif-netdev/bond-show", "[dp]", |
1506 | 0, 1, dpif_netdev_bond_show, | |
1507 | NULL); | |
6553d06b DDP |
1508 | return 0; |
1509 | } | |
72865317 | 1510 | |
2197d7ab | 1511 | static int |
2240af25 DDP |
1512 | dpif_netdev_enumerate(struct sset *all_dps, |
1513 | const struct dpif_class *dpif_class) | |
2197d7ab GL |
1514 | { |
1515 | struct shash_node *node; | |
1516 | ||
97be1538 | 1517 | ovs_mutex_lock(&dp_netdev_mutex); |
2197d7ab | 1518 | SHASH_FOR_EACH(node, &dp_netdevs) { |
2240af25 DDP |
1519 | struct dp_netdev *dp = node->data; |
1520 | if (dpif_class != dp->class) { | |
1521 | /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs. | |
1522 | * If the class doesn't match, skip this dpif. */ | |
1523 | continue; | |
1524 | } | |
2197d7ab GL |
1525 | sset_add(all_dps, node->name); |
1526 | } | |
97be1538 | 1527 | ovs_mutex_unlock(&dp_netdev_mutex); |
5279f8fd | 1528 | |
2197d7ab GL |
1529 | return 0; |
1530 | } | |
1531 | ||
add90f6f EJ |
1532 | static bool |
1533 | dpif_netdev_class_is_dummy(const struct dpif_class *class) | |
1534 | { | |
1535 | return class != &dpif_netdev_class; | |
1536 | } | |
1537 | ||
0aeaabc8 JP |
1538 | static const char * |
1539 | dpif_netdev_port_open_type(const struct dpif_class *class, const char *type) | |
1540 | { | |
1541 | return strcmp(type, "internal") ? type | |
e98d0cb3 | 1542 | : dpif_netdev_class_is_dummy(class) ? "dummy-internal" |
0aeaabc8 JP |
1543 | : "tap"; |
1544 | } | |
1545 | ||
72865317 BP |
1546 | static struct dpif * |
1547 | create_dpif_netdev(struct dp_netdev *dp) | |
1548 | { | |
462278db | 1549 | uint16_t netflow_id = hash_string(dp->name, 0); |
72865317 | 1550 | struct dpif_netdev *dpif; |
72865317 | 1551 | |
6a8267c5 | 1552 | ovs_refcount_ref(&dp->ref_cnt); |
72865317 | 1553 | |
72865317 | 1554 | dpif = xmalloc(sizeof *dpif); |
614c4892 | 1555 | dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id); |
72865317 | 1556 | dpif->dp = dp; |
d33ed218 | 1557 | dpif->last_port_seq = seq_read(dp->port_seq); |
72865317 BP |
1558 | |
1559 | return &dpif->dpif; | |
1560 | } | |
1561 | ||
4e022ec0 AW |
1562 | /* Choose an unused, non-zero port number and return it on success. |
1563 | * Return ODPP_NONE on failure. */ | |
1564 | static odp_port_t | |
e44768b7 | 1565 | choose_port(struct dp_netdev *dp, const char *name) |
59e6d833 | 1566 | OVS_REQUIRES(dp->port_mutex) |
e44768b7 | 1567 | { |
4e022ec0 | 1568 | uint32_t port_no; |
e44768b7 JP |
1569 | |
1570 | if (dp->class != &dpif_netdev_class) { | |
1571 | const char *p; | |
1572 | int start_no = 0; | |
1573 | ||
1574 | /* If the port name begins with "br", start the number search at | |
1575 | * 100 to make writing tests easier. */ | |
1576 | if (!strncmp(name, "br", 2)) { | |
1577 | start_no = 100; | |
1578 | } | |
1579 | ||
1580 | /* If the port name contains a number, try to assign that port number. | |
1581 | * This can make writing unit tests easier because port numbers are | |
1582 | * predictable. */ | |
1583 | for (p = name; *p != '\0'; p++) { | |
1584 | if (isdigit((unsigned char) *p)) { | |
1585 | port_no = start_no + strtol(p, NULL, 10); | |
ff073a71 BP |
1586 | if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE) |
1587 | && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) { | |
4e022ec0 | 1588 | return u32_to_odp(port_no); |
e44768b7 JP |
1589 | } |
1590 | break; | |
1591 | } | |
1592 | } | |
1593 | } | |
1594 | ||
ff073a71 BP |
1595 | for (port_no = 1; port_no <= UINT16_MAX; port_no++) { |
1596 | if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) { | |
4e022ec0 | 1597 | return u32_to_odp(port_no); |
e44768b7 JP |
1598 | } |
1599 | } | |
1600 | ||
4e022ec0 | 1601 | return ODPP_NONE; |
e44768b7 JP |
1602 | } |
1603 | ||
72865317 | 1604 | static int |
614c4892 BP |
1605 | create_dp_netdev(const char *name, const struct dpif_class *class, |
1606 | struct dp_netdev **dpp) | |
8a4e3a85 | 1607 | OVS_REQUIRES(dp_netdev_mutex) |
72865317 | 1608 | { |
1276e3db | 1609 | static struct ovsthread_once tsc_freq_check = OVSTHREAD_ONCE_INITIALIZER; |
72865317 BP |
1610 | struct dp_netdev *dp; |
1611 | int error; | |
72865317 | 1612 | |
1276e3db IM |
1613 | /* Avoid estimating TSC frequency for dummy datapath to not slow down |
1614 | * unit tests. */ | |
1615 | if (!dpif_netdev_class_is_dummy(class) | |
1616 | && ovsthread_once_start(&tsc_freq_check)) { | |
1617 | pmd_perf_estimate_tsc_frequency(); | |
1618 | ovsthread_once_done(&tsc_freq_check); | |
1619 | } | |
1620 | ||
462278db | 1621 | dp = xzalloc(sizeof *dp); |
8a4e3a85 BP |
1622 | shash_add(&dp_netdevs, name, dp); |
1623 | ||
1624 | *CONST_CAST(const struct dpif_class **, &dp->class) = class; | |
1625 | *CONST_CAST(const char **, &dp->name) = xstrdup(name); | |
6a8267c5 | 1626 | ovs_refcount_init(&dp->ref_cnt); |
1a65ba85 | 1627 | atomic_flag_clear(&dp->destroyed); |
8a4e3a85 | 1628 | |
81e89d5c | 1629 | ovs_mutex_init_recursive(&dp->port_mutex); |
e9985d6a | 1630 | hmap_init(&dp->ports); |
d33ed218 | 1631 | dp->port_seq = seq_create(); |
9df65060 VDA |
1632 | ovs_mutex_init(&dp->bond_mutex); |
1633 | cmap_init(&dp->tx_bonds); | |
1634 | ||
6b31e073 RW |
1635 | fat_rwlock_init(&dp->upcall_rwlock); |
1636 | ||
a6a426d6 IM |
1637 | dp->reconfigure_seq = seq_create(); |
1638 | dp->last_reconfigure_seq = seq_read(dp->reconfigure_seq); | |
1639 | ||
4b27db64 JR |
1640 | for (int i = 0; i < N_METER_LOCKS; ++i) { |
1641 | ovs_mutex_init_adaptive(&dp->meter_locks[i]); | |
1642 | } | |
1643 | ||
6b31e073 RW |
1644 | /* Disable upcalls by default. */ |
1645 | dp_netdev_disable_upcall(dp); | |
623540e4 | 1646 | dp->upcall_aux = NULL; |
6b31e073 | 1647 | dp->upcall_cb = NULL; |
e44768b7 | 1648 | |
57593fd2 | 1649 | dp->conntrack = conntrack_init(); |
5cf3edb3 | 1650 | |
4c30b246 | 1651 | atomic_init(&dp->emc_insert_min, DEFAULT_EM_FLOW_INSERT_MIN); |
c71ea3c4 | 1652 | atomic_init(&dp->tx_flush_interval, DEFAULT_TX_FLUSH_INTERVAL); |
4c30b246 | 1653 | |
65f13b50 | 1654 | cmap_init(&dp->poll_threads); |
e77c97b9 | 1655 | dp->pmd_rxq_assign_cyc = true; |
140dd699 IM |
1656 | |
1657 | ovs_mutex_init(&dp->tx_qid_pool_mutex); | |
1658 | /* We need 1 Tx queue for each possible core + 1 for non-PMD threads. */ | |
1659 | dp->tx_qid_pool = id_pool_create(0, ovs_numa_get_n_cores() + 1); | |
1660 | ||
65f13b50 AW |
1661 | ovs_mutex_init_recursive(&dp->non_pmd_mutex); |
1662 | ovsthread_key_create(&dp->per_pmd_key, NULL); | |
1663 | ||
e9985d6a | 1664 | ovs_mutex_lock(&dp->port_mutex); |
140dd699 IM |
1665 | /* non-PMD will be created before all other threads and will |
1666 | * allocate static_tx_qid = 0. */ | |
f2eee189 | 1667 | dp_netdev_set_nonpmd(dp); |
65f13b50 | 1668 | |
a3e8437a TLSC |
1669 | error = do_add_port(dp, name, dpif_netdev_port_open_type(dp->class, |
1670 | "internal"), | |
1671 | ODPP_LOCAL); | |
59e6d833 | 1672 | ovs_mutex_unlock(&dp->port_mutex); |
72865317 BP |
1673 | if (error) { |
1674 | dp_netdev_free(dp); | |
462278db | 1675 | return error; |
72865317 BP |
1676 | } |
1677 | ||
a36de779 | 1678 | dp->last_tnl_conf_seq = seq_read(tnl_conf_seq); |
462278db | 1679 | *dpp = dp; |
72865317 BP |
1680 | return 0; |
1681 | } | |
1682 | ||
a6a426d6 IM |
1683 | static void |
1684 | dp_netdev_request_reconfigure(struct dp_netdev *dp) | |
1685 | { | |
1686 | seq_change(dp->reconfigure_seq); | |
1687 | } | |
1688 | ||
1689 | static bool | |
1690 | dp_netdev_is_reconf_required(struct dp_netdev *dp) | |
1691 | { | |
1692 | return seq_read(dp->reconfigure_seq) != dp->last_reconfigure_seq; | |
1693 | } | |
1694 | ||
72865317 | 1695 | static int |
614c4892 | 1696 | dpif_netdev_open(const struct dpif_class *class, const char *name, |
4a387741 | 1697 | bool create, struct dpif **dpifp) |
72865317 | 1698 | { |
462278db | 1699 | struct dp_netdev *dp; |
5279f8fd | 1700 | int error; |
462278db | 1701 | |
97be1538 | 1702 | ovs_mutex_lock(&dp_netdev_mutex); |
462278db BP |
1703 | dp = shash_find_data(&dp_netdevs, name); |
1704 | if (!dp) { | |
5279f8fd | 1705 | error = create ? create_dp_netdev(name, class, &dp) : ENODEV; |
72865317 | 1706 | } else { |
5279f8fd BP |
1707 | error = (dp->class != class ? EINVAL |
1708 | : create ? EEXIST | |
1709 | : 0); | |
1710 | } | |
1711 | if (!error) { | |
1712 | *dpifp = create_dpif_netdev(dp); | |
72865317 | 1713 | } |
97be1538 | 1714 | ovs_mutex_unlock(&dp_netdev_mutex); |
462278db | 1715 | |
5279f8fd | 1716 | return error; |
72865317 BP |
1717 | } |
1718 | ||
88ace79b DDP |
1719 | static void |
1720 | dp_netdev_destroy_upcall_lock(struct dp_netdev *dp) | |
1721 | OVS_NO_THREAD_SAFETY_ANALYSIS | |
1722 | { | |
1723 | /* Check that upcalls are disabled, i.e. that the rwlock is taken */ | |
1724 | ovs_assert(fat_rwlock_tryrdlock(&dp->upcall_rwlock)); | |
1725 | ||
1726 | /* Before freeing a lock we should release it */ | |
1727 | fat_rwlock_unlock(&dp->upcall_rwlock); | |
1728 | fat_rwlock_destroy(&dp->upcall_rwlock); | |
1729 | } | |
1730 | ||
4b27db64 JR |
1731 | static void |
1732 | dp_delete_meter(struct dp_netdev *dp, uint32_t meter_id) | |
1733 | OVS_REQUIRES(dp->meter_locks[meter_id % N_METER_LOCKS]) | |
1734 | { | |
1735 | if (dp->meters[meter_id]) { | |
1736 | free(dp->meters[meter_id]); | |
1737 | dp->meters[meter_id] = NULL; | |
1738 | } | |
1739 | } | |
1740 | ||
9df65060 VDA |
1741 | static uint32_t |
1742 | hash_bond_id(uint32_t bond_id) | |
1743 | { | |
1744 | return hash_int(bond_id, 0); | |
1745 | } | |
1746 | ||
8a4e3a85 BP |
1747 | /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp' |
1748 | * through the 'dp_netdevs' shash while freeing 'dp'. */ | |
1ba530f4 BP |
1749 | static void |
1750 | dp_netdev_free(struct dp_netdev *dp) | |
8a4e3a85 | 1751 | OVS_REQUIRES(dp_netdev_mutex) |
1ba530f4 | 1752 | { |
e9985d6a | 1753 | struct dp_netdev_port *port, *next; |
9df65060 | 1754 | struct tx_bond *bond; |
4ad28026 | 1755 | |
8a4e3a85 BP |
1756 | shash_find_and_delete(&dp_netdevs, dp->name); |
1757 | ||
59e6d833 | 1758 | ovs_mutex_lock(&dp->port_mutex); |
e9985d6a | 1759 | HMAP_FOR_EACH_SAFE (port, next, node, &dp->ports) { |
c40b890f | 1760 | do_del_port(dp, port); |
1ba530f4 | 1761 | } |
59e6d833 | 1762 | ovs_mutex_unlock(&dp->port_mutex); |
4b27db64 | 1763 | |
9df65060 VDA |
1764 | ovs_mutex_lock(&dp->bond_mutex); |
1765 | CMAP_FOR_EACH (bond, node, &dp->tx_bonds) { | |
1766 | cmap_remove(&dp->tx_bonds, &bond->node, hash_bond_id(bond->bond_id)); | |
1767 | ovsrcu_postpone(free, bond); | |
1768 | } | |
1769 | ovs_mutex_unlock(&dp->bond_mutex); | |
1770 | ||
e32971b8 | 1771 | dp_netdev_destroy_all_pmds(dp, true); |
d916785c | 1772 | cmap_destroy(&dp->poll_threads); |
51852a57 | 1773 | |
140dd699 IM |
1774 | ovs_mutex_destroy(&dp->tx_qid_pool_mutex); |
1775 | id_pool_destroy(dp->tx_qid_pool); | |
1776 | ||
b9584f21 DDP |
1777 | ovs_mutex_destroy(&dp->non_pmd_mutex); |
1778 | ovsthread_key_delete(dp->per_pmd_key); | |
1779 | ||
57593fd2 | 1780 | conntrack_destroy(dp->conntrack); |
b9584f21 DDP |
1781 | |
1782 | ||
a6a426d6 IM |
1783 | seq_destroy(dp->reconfigure_seq); |
1784 | ||
d33ed218 | 1785 | seq_destroy(dp->port_seq); |
e9985d6a | 1786 | hmap_destroy(&dp->ports); |
3186ea46 | 1787 | ovs_mutex_destroy(&dp->port_mutex); |
88ace79b | 1788 | |
9df65060 VDA |
1789 | cmap_destroy(&dp->tx_bonds); |
1790 | ovs_mutex_destroy(&dp->bond_mutex); | |
1791 | ||
88ace79b DDP |
1792 | /* Upcalls must be disabled at this point */ |
1793 | dp_netdev_destroy_upcall_lock(dp); | |
9bbf1c3d | 1794 | |
4b27db64 JR |
1795 | int i; |
1796 | ||
1797 | for (i = 0; i < MAX_METERS; ++i) { | |
1798 | meter_lock(dp, i); | |
1799 | dp_delete_meter(dp, i); | |
1800 | meter_unlock(dp, i); | |
1801 | } | |
1802 | for (i = 0; i < N_METER_LOCKS; ++i) { | |
1803 | ovs_mutex_destroy(&dp->meter_locks[i]); | |
1804 | } | |
1805 | ||
f2eee189 | 1806 | free(dp->pmd_cmask); |
8a4e3a85 | 1807 | free(CONST_CAST(char *, dp->name)); |
72865317 BP |
1808 | free(dp); |
1809 | } | |
1810 | ||
8a4e3a85 BP |
1811 | static void |
1812 | dp_netdev_unref(struct dp_netdev *dp) | |
1813 | { | |
1814 | if (dp) { | |
1815 | /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't | |
1816 | * get a new reference to 'dp' through the 'dp_netdevs' shash. */ | |
1817 | ovs_mutex_lock(&dp_netdev_mutex); | |
24f83812 | 1818 | if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) { |
8a4e3a85 BP |
1819 | dp_netdev_free(dp); |
1820 | } | |
1821 | ovs_mutex_unlock(&dp_netdev_mutex); | |
1822 | } | |
1823 | } | |
1824 | ||
72865317 BP |
1825 | static void |
1826 | dpif_netdev_close(struct dpif *dpif) | |
1827 | { | |
1828 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
5279f8fd | 1829 | |
8a4e3a85 | 1830 | dp_netdev_unref(dp); |
72865317 BP |
1831 | free(dpif); |
1832 | } | |
1833 | ||
1834 | static int | |
7dab847a | 1835 | dpif_netdev_destroy(struct dpif *dpif) |
72865317 BP |
1836 | { |
1837 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
5279f8fd | 1838 | |
6a8267c5 | 1839 | if (!atomic_flag_test_and_set(&dp->destroyed)) { |
24f83812 | 1840 | if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) { |
6a8267c5 BP |
1841 | /* Can't happen: 'dpif' still owns a reference to 'dp'. */ |
1842 | OVS_NOT_REACHED(); | |
1843 | } | |
1844 | } | |
5279f8fd | 1845 | |
72865317 BP |
1846 | return 0; |
1847 | } | |
1848 | ||
eb94da30 DDP |
1849 | /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed |
1850 | * load/store semantics. While the increment is not atomic, the load and | |
1851 | * store operations are, making it impossible to read inconsistent values. | |
1852 | * | |
1853 | * This is used to update thread local stats counters. */ | |
1854 | static void | |
1855 | non_atomic_ullong_add(atomic_ullong *var, unsigned long long n) | |
1856 | { | |
1857 | unsigned long long tmp; | |
1858 | ||
1859 | atomic_read_relaxed(var, &tmp); | |
1860 | tmp += n; | |
1861 | atomic_store_relaxed(var, tmp); | |
1862 | } | |
1863 | ||
72865317 | 1864 | static int |
a8d9304d | 1865 | dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats) |
72865317 BP |
1866 | { |
1867 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
1c1e46ed | 1868 | struct dp_netdev_pmd_thread *pmd; |
82a48ead | 1869 | uint64_t pmd_stats[PMD_N_STATS]; |
8a4e3a85 | 1870 | |
1c1e46ed AW |
1871 | stats->n_flows = stats->n_hit = stats->n_missed = stats->n_lost = 0; |
1872 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
1873 | stats->n_flows += cmap_count(&pmd->flow_table); | |
82a48ead JS |
1874 | pmd_perf_read_counters(&pmd->perf_stats, pmd_stats); |
1875 | stats->n_hit += pmd_stats[PMD_STAT_EXACT_HIT]; | |
60d8ccae | 1876 | stats->n_hit += pmd_stats[PMD_STAT_SMC_HIT]; |
82a48ead JS |
1877 | stats->n_hit += pmd_stats[PMD_STAT_MASKED_HIT]; |
1878 | stats->n_missed += pmd_stats[PMD_STAT_MISS]; | |
1879 | stats->n_lost += pmd_stats[PMD_STAT_LOST]; | |
51852a57 | 1880 | } |
1ce3fa06 | 1881 | stats->n_masks = UINT32_MAX; |
847108dc | 1882 | stats->n_mask_hit = UINT64_MAX; |
5279f8fd | 1883 | |
72865317 BP |
1884 | return 0; |
1885 | } | |
1886 | ||
e4cfed38 | 1887 | static void |
65f13b50 | 1888 | dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread *pmd) |
e4cfed38 | 1889 | { |
accf8626 | 1890 | if (pmd->core_id == NON_PMD_CORE_ID) { |
d0cca6c3 DDP |
1891 | ovs_mutex_lock(&pmd->dp->non_pmd_mutex); |
1892 | ovs_mutex_lock(&pmd->port_mutex); | |
1893 | pmd_load_cached_ports(pmd); | |
1894 | ovs_mutex_unlock(&pmd->port_mutex); | |
1895 | ovs_mutex_unlock(&pmd->dp->non_pmd_mutex); | |
accf8626 AW |
1896 | return; |
1897 | } | |
1898 | ||
2788a1b1 | 1899 | seq_change(pmd->reload_seq); |
299c8d61 | 1900 | atomic_store_explicit(&pmd->reload, true, memory_order_release); |
65f13b50 | 1901 | } |
e4cfed38 | 1902 | |
59e6d833 BP |
1903 | static uint32_t |
1904 | hash_port_no(odp_port_t port_no) | |
1905 | { | |
1906 | return hash_int(odp_to_u32(port_no), 0); | |
1907 | } | |
1908 | ||
72865317 | 1909 | static int |
a3e8437a | 1910 | port_create(const char *devname, const char *type, |
b8d29252 | 1911 | odp_port_t port_no, struct dp_netdev_port **portp) |
72865317 | 1912 | { |
72865317 | 1913 | struct dp_netdev_port *port; |
2499a8ce | 1914 | enum netdev_flags flags; |
b8d29252 | 1915 | struct netdev *netdev; |
e32971b8 | 1916 | int error; |
72865317 | 1917 | |
b8d29252 | 1918 | *portp = NULL; |
72865317 BP |
1919 | |
1920 | /* Open and validate network device. */ | |
a3e8437a | 1921 | error = netdev_open(devname, type, &netdev); |
72865317 | 1922 | if (error) { |
b8d29252 | 1923 | return error; |
72865317 | 1924 | } |
72865317 BP |
1925 | /* XXX reject non-Ethernet devices */ |
1926 | ||
2499a8ce AC |
1927 | netdev_get_flags(netdev, &flags); |
1928 | if (flags & NETDEV_LOOPBACK) { | |
1929 | VLOG_ERR("%s: cannot add a loopback device", devname); | |
d17f4f08 | 1930 | error = EINVAL; |
b8d29252 | 1931 | goto out; |
2499a8ce AC |
1932 | } |
1933 | ||
e4cfed38 | 1934 | port = xzalloc(sizeof *port); |
35303d71 | 1935 | port->port_no = port_no; |
e4cfed38 PS |
1936 | port->netdev = netdev; |
1937 | port->type = xstrdup(type); | |
96e74404 | 1938 | port->sf = NULL; |
2fbadeb6 | 1939 | port->emc_enabled = true; |
e32971b8 DDP |
1940 | port->need_reconfigure = true; |
1941 | ovs_mutex_init(&port->txq_used_mutex); | |
e4cfed38 | 1942 | |
b8d29252 | 1943 | *portp = port; |
72865317 BP |
1944 | |
1945 | return 0; | |
d17f4f08 | 1946 | |
d17f4f08 | 1947 | out: |
b8d29252 | 1948 | netdev_close(netdev); |
d17f4f08 | 1949 | return error; |
72865317 BP |
1950 | } |
1951 | ||
b8d29252 DDP |
1952 | static int |
1953 | do_add_port(struct dp_netdev *dp, const char *devname, const char *type, | |
1954 | odp_port_t port_no) | |
1955 | OVS_REQUIRES(dp->port_mutex) | |
1956 | { | |
96e74404 | 1957 | struct netdev_saved_flags *sf; |
b8d29252 DDP |
1958 | struct dp_netdev_port *port; |
1959 | int error; | |
1960 | ||
1961 | /* Reject devices already in 'dp'. */ | |
1962 | if (!get_port_by_name(dp, devname, &port)) { | |
1963 | return EEXIST; | |
1964 | } | |
1965 | ||
a3e8437a | 1966 | error = port_create(devname, type, port_no, &port); |
b8d29252 DDP |
1967 | if (error) { |
1968 | return error; | |
1969 | } | |
1970 | ||
e9985d6a | 1971 | hmap_insert(&dp->ports, &port->node, hash_port_no(port_no)); |
b8d29252 DDP |
1972 | seq_change(dp->port_seq); |
1973 | ||
e32971b8 DDP |
1974 | reconfigure_datapath(dp); |
1975 | ||
3f51ea18 | 1976 | /* Check that port was successfully configured. */ |
96e74404 IM |
1977 | if (!dp_netdev_lookup_port(dp, port_no)) { |
1978 | return EINVAL; | |
1979 | } | |
1980 | ||
1981 | /* Updating device flags triggers an if_notifier, which triggers a bridge | |
1982 | * reconfiguration and another attempt to add this port, leading to an | |
1983 | * infinite loop if the device is configured incorrectly and cannot be | |
1984 | * added. Setting the promisc mode after a successful reconfiguration, | |
1985 | * since we already know that the device is somehow properly configured. */ | |
1986 | error = netdev_turn_flags_on(port->netdev, NETDEV_PROMISC, &sf); | |
1987 | if (error) { | |
1988 | VLOG_ERR("%s: cannot set promisc flag", devname); | |
1989 | do_del_port(dp, port); | |
1990 | return error; | |
1991 | } | |
1992 | port->sf = sf; | |
1993 | ||
1994 | return 0; | |
b8d29252 DDP |
1995 | } |
1996 | ||
247527db BP |
1997 | static int |
1998 | dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev, | |
4e022ec0 | 1999 | odp_port_t *port_nop) |
247527db BP |
2000 | { |
2001 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
3aa30359 BP |
2002 | char namebuf[NETDEV_VPORT_NAME_BUFSIZE]; |
2003 | const char *dpif_port; | |
4e022ec0 | 2004 | odp_port_t port_no; |
5279f8fd | 2005 | int error; |
247527db | 2006 | |
59e6d833 | 2007 | ovs_mutex_lock(&dp->port_mutex); |
3aa30359 | 2008 | dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf); |
4e022ec0 | 2009 | if (*port_nop != ODPP_NONE) { |
ff073a71 BP |
2010 | port_no = *port_nop; |
2011 | error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0; | |
232dfa4a | 2012 | } else { |
3aa30359 | 2013 | port_no = choose_port(dp, dpif_port); |
5279f8fd | 2014 | error = port_no == ODPP_NONE ? EFBIG : 0; |
232dfa4a | 2015 | } |
5279f8fd | 2016 | if (!error) { |
247527db | 2017 | *port_nop = port_no; |
5279f8fd | 2018 | error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no); |
247527db | 2019 | } |
59e6d833 | 2020 | ovs_mutex_unlock(&dp->port_mutex); |
5279f8fd BP |
2021 | |
2022 | return error; | |
72865317 BP |
2023 | } |
2024 | ||
2025 | static int | |
4e022ec0 | 2026 | dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no) |
72865317 BP |
2027 | { |
2028 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
5279f8fd BP |
2029 | int error; |
2030 | ||
59e6d833 | 2031 | ovs_mutex_lock(&dp->port_mutex); |
c40b890f BP |
2032 | if (port_no == ODPP_LOCAL) { |
2033 | error = EINVAL; | |
2034 | } else { | |
2035 | struct dp_netdev_port *port; | |
2036 | ||
2037 | error = get_port_by_number(dp, port_no, &port); | |
2038 | if (!error) { | |
2039 | do_del_port(dp, port); | |
2040 | } | |
2041 | } | |
59e6d833 | 2042 | ovs_mutex_unlock(&dp->port_mutex); |
5279f8fd BP |
2043 | |
2044 | return error; | |
72865317 BP |
2045 | } |
2046 | ||
2047 | static bool | |
4e022ec0 | 2048 | is_valid_port_number(odp_port_t port_no) |
72865317 | 2049 | { |
ff073a71 BP |
2050 | return port_no != ODPP_NONE; |
2051 | } | |
2052 | ||
2053 | static struct dp_netdev_port * | |
2054 | dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no) | |
e9985d6a | 2055 | OVS_REQUIRES(dp->port_mutex) |
ff073a71 BP |
2056 | { |
2057 | struct dp_netdev_port *port; | |
2058 | ||
e9985d6a | 2059 | HMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) { |
35303d71 | 2060 | if (port->port_no == port_no) { |
ff073a71 BP |
2061 | return port; |
2062 | } | |
2063 | } | |
2064 | return NULL; | |
72865317 BP |
2065 | } |
2066 | ||
2067 | static int | |
2068 | get_port_by_number(struct dp_netdev *dp, | |
4e022ec0 | 2069 | odp_port_t port_no, struct dp_netdev_port **portp) |
e9985d6a | 2070 | OVS_REQUIRES(dp->port_mutex) |
72865317 BP |
2071 | { |
2072 | if (!is_valid_port_number(port_no)) { | |
2073 | *portp = NULL; | |
2074 | return EINVAL; | |
2075 | } else { | |
ff073a71 | 2076 | *portp = dp_netdev_lookup_port(dp, port_no); |
0f6a066f | 2077 | return *portp ? 0 : ENODEV; |
72865317 BP |
2078 | } |
2079 | } | |
2080 | ||
b284085e | 2081 | static void |
62453dad | 2082 | port_destroy(struct dp_netdev_port *port) |
b284085e | 2083 | { |
62453dad DDP |
2084 | if (!port) { |
2085 | return; | |
b284085e | 2086 | } |
b284085e | 2087 | |
62453dad DDP |
2088 | netdev_close(port->netdev); |
2089 | netdev_restore_flags(port->sf); | |
accf8626 | 2090 | |
62453dad | 2091 | for (unsigned i = 0; i < port->n_rxq; i++) { |
947dc567 | 2092 | netdev_rxq_close(port->rxqs[i].rx); |
b284085e | 2093 | } |
324c8374 | 2094 | ovs_mutex_destroy(&port->txq_used_mutex); |
3eb67853 | 2095 | free(port->rxq_affinity_list); |
324c8374 | 2096 | free(port->txq_used); |
3eb67853 | 2097 | free(port->rxqs); |
62453dad DDP |
2098 | free(port->type); |
2099 | free(port); | |
b284085e PS |
2100 | } |
2101 | ||
72865317 BP |
2102 | static int |
2103 | get_port_by_name(struct dp_netdev *dp, | |
2104 | const char *devname, struct dp_netdev_port **portp) | |
59e6d833 | 2105 | OVS_REQUIRES(dp->port_mutex) |
72865317 BP |
2106 | { |
2107 | struct dp_netdev_port *port; | |
2108 | ||
e9985d6a | 2109 | HMAP_FOR_EACH (port, node, &dp->ports) { |
3efb6063 | 2110 | if (!strcmp(netdev_get_name(port->netdev), devname)) { |
72865317 BP |
2111 | *portp = port; |
2112 | return 0; | |
2113 | } | |
2114 | } | |
0f6a066f DDP |
2115 | |
2116 | /* Callers of dpif_netdev_port_query_by_name() expect ENODEV for a non | |
2117 | * existing port. */ | |
2118 | return ENODEV; | |
72865317 BP |
2119 | } |
2120 | ||
b9584f21 | 2121 | /* Returns 'true' if there is a port with pmd netdev. */ |
65f13b50 | 2122 | static bool |
b9584f21 | 2123 | has_pmd_port(struct dp_netdev *dp) |
e9985d6a | 2124 | OVS_REQUIRES(dp->port_mutex) |
65f13b50 AW |
2125 | { |
2126 | struct dp_netdev_port *port; | |
2127 | ||
e9985d6a | 2128 | HMAP_FOR_EACH (port, node, &dp->ports) { |
5dd57e80 | 2129 | if (netdev_is_pmd(port->netdev)) { |
b9584f21 | 2130 | return true; |
65f13b50 AW |
2131 | } |
2132 | } | |
2133 | ||
2134 | return false; | |
2135 | } | |
2136 | ||
c40b890f BP |
2137 | static void |
2138 | do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port) | |
59e6d833 | 2139 | OVS_REQUIRES(dp->port_mutex) |
72865317 | 2140 | { |
e9985d6a | 2141 | hmap_remove(&dp->ports, &port->node); |
d33ed218 | 2142 | seq_change(dp->port_seq); |
d0cca6c3 | 2143 | |
e32971b8 | 2144 | reconfigure_datapath(dp); |
72865317 | 2145 | |
62453dad | 2146 | port_destroy(port); |
72865317 BP |
2147 | } |
2148 | ||
2149 | static void | |
4c738a8d BP |
2150 | answer_port_query(const struct dp_netdev_port *port, |
2151 | struct dpif_port *dpif_port) | |
72865317 | 2152 | { |
3efb6063 | 2153 | dpif_port->name = xstrdup(netdev_get_name(port->netdev)); |
0cbfe35d | 2154 | dpif_port->type = xstrdup(port->type); |
35303d71 | 2155 | dpif_port->port_no = port->port_no; |
72865317 BP |
2156 | } |
2157 | ||
2158 | static int | |
4e022ec0 | 2159 | dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no, |
4c738a8d | 2160 | struct dpif_port *dpif_port) |
72865317 BP |
2161 | { |
2162 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
2163 | struct dp_netdev_port *port; | |
2164 | int error; | |
2165 | ||
e9985d6a | 2166 | ovs_mutex_lock(&dp->port_mutex); |
72865317 | 2167 | error = get_port_by_number(dp, port_no, &port); |
4afba28d | 2168 | if (!error && dpif_port) { |
4c738a8d | 2169 | answer_port_query(port, dpif_port); |
72865317 | 2170 | } |
e9985d6a | 2171 | ovs_mutex_unlock(&dp->port_mutex); |
5279f8fd | 2172 | |
72865317 BP |
2173 | return error; |
2174 | } | |
2175 | ||
2176 | static int | |
2177 | dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname, | |
4c738a8d | 2178 | struct dpif_port *dpif_port) |
72865317 BP |
2179 | { |
2180 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
2181 | struct dp_netdev_port *port; | |
2182 | int error; | |
2183 | ||
59e6d833 | 2184 | ovs_mutex_lock(&dp->port_mutex); |
72865317 | 2185 | error = get_port_by_name(dp, devname, &port); |
4afba28d | 2186 | if (!error && dpif_port) { |
4c738a8d | 2187 | answer_port_query(port, dpif_port); |
72865317 | 2188 | } |
59e6d833 | 2189 | ovs_mutex_unlock(&dp->port_mutex); |
5279f8fd | 2190 | |
72865317 BP |
2191 | return error; |
2192 | } | |
2193 | ||
61e7deb1 BP |
2194 | static void |
2195 | dp_netdev_flow_free(struct dp_netdev_flow *flow) | |
2196 | { | |
61e7deb1 | 2197 | dp_netdev_actions_free(dp_netdev_flow_get_actions(flow)); |
342b8904 | 2198 | free(flow->dp_extra_info); |
61e7deb1 BP |
2199 | free(flow); |
2200 | } | |
2201 | ||
ed79f89a DDP |
2202 | static void dp_netdev_flow_unref(struct dp_netdev_flow *flow) |
2203 | { | |
2204 | if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) { | |
2205 | ovsrcu_postpone(dp_netdev_flow_free, flow); | |
2206 | } | |
2207 | } | |
2208 | ||
70e5ed6f JS |
2209 | static uint32_t |
2210 | dp_netdev_flow_hash(const ovs_u128 *ufid) | |
2211 | { | |
2212 | return ufid->u32[0]; | |
2213 | } | |
2214 | ||
3453b4d6 JS |
2215 | static inline struct dpcls * |
2216 | dp_netdev_pmd_lookup_dpcls(struct dp_netdev_pmd_thread *pmd, | |
2217 | odp_port_t in_port) | |
2218 | { | |
2219 | struct dpcls *cls; | |
2220 | uint32_t hash = hash_port_no(in_port); | |
2221 | CMAP_FOR_EACH_WITH_HASH (cls, node, hash, &pmd->classifiers) { | |
2222 | if (cls->in_port == in_port) { | |
2223 | /* Port classifier exists already */ | |
2224 | return cls; | |
2225 | } | |
2226 | } | |
2227 | return NULL; | |
2228 | } | |
2229 | ||
2230 | static inline struct dpcls * | |
2231 | dp_netdev_pmd_find_dpcls(struct dp_netdev_pmd_thread *pmd, | |
2232 | odp_port_t in_port) | |
2233 | OVS_REQUIRES(pmd->flow_mutex) | |
2234 | { | |
2235 | struct dpcls *cls = dp_netdev_pmd_lookup_dpcls(pmd, in_port); | |
2236 | uint32_t hash = hash_port_no(in_port); | |
2237 | ||
2238 | if (!cls) { | |
2239 | /* Create new classifier for in_port */ | |
2240 | cls = xmalloc(sizeof(*cls)); | |
2241 | dpcls_init(cls); | |
2242 | cls->in_port = in_port; | |
2243 | cmap_insert(&pmd->classifiers, &cls->node, hash); | |
2244 | VLOG_DBG("Creating dpcls %p for in_port %d", cls, in_port); | |
2245 | } | |
2246 | return cls; | |
2247 | } | |
2248 | ||
241bad15 YL |
2249 | #define MAX_FLOW_MARK (UINT32_MAX - 1) |
2250 | #define INVALID_FLOW_MARK (UINT32_MAX) | |
2251 | ||
2252 | struct megaflow_to_mark_data { | |
2253 | const struct cmap_node node; | |
2254 | ovs_u128 mega_ufid; | |
2255 | uint32_t mark; | |
2256 | }; | |
2257 | ||
2258 | struct flow_mark { | |
2259 | struct cmap megaflow_to_mark; | |
2260 | struct cmap mark_to_flow; | |
2261 | struct id_pool *pool; | |
241bad15 YL |
2262 | }; |
2263 | ||
2264 | static struct flow_mark flow_mark = { | |
2265 | .megaflow_to_mark = CMAP_INITIALIZER, | |
2266 | .mark_to_flow = CMAP_INITIALIZER, | |
241bad15 YL |
2267 | }; |
2268 | ||
2269 | static uint32_t | |
2270 | flow_mark_alloc(void) | |
2271 | { | |
2272 | uint32_t mark; | |
2273 | ||
2274 | if (!flow_mark.pool) { | |
2275 | /* Haven't initiated yet, do it here */ | |
2276 | flow_mark.pool = id_pool_create(0, MAX_FLOW_MARK); | |
2277 | } | |
2278 | ||
2279 | if (id_pool_alloc_id(flow_mark.pool, &mark)) { | |
2280 | return mark; | |
2281 | } | |
2282 | ||
2283 | return INVALID_FLOW_MARK; | |
2284 | } | |
2285 | ||
2286 | static void | |
2287 | flow_mark_free(uint32_t mark) | |
2288 | { | |
2289 | id_pool_free_id(flow_mark.pool, mark); | |
2290 | } | |
2291 | ||
2292 | /* associate megaflow with a mark, which is a 1:1 mapping */ | |
2293 | static void | |
2294 | megaflow_to_mark_associate(const ovs_u128 *mega_ufid, uint32_t mark) | |
2295 | { | |
2296 | size_t hash = dp_netdev_flow_hash(mega_ufid); | |
2297 | struct megaflow_to_mark_data *data = xzalloc(sizeof(*data)); | |
2298 | ||
2299 | data->mega_ufid = *mega_ufid; | |
2300 | data->mark = mark; | |
2301 | ||
2302 | cmap_insert(&flow_mark.megaflow_to_mark, | |
2303 | CONST_CAST(struct cmap_node *, &data->node), hash); | |
2304 | } | |
2305 | ||
2306 | /* disassociate meagaflow with a mark */ | |
2307 | static void | |
2308 | megaflow_to_mark_disassociate(const ovs_u128 *mega_ufid) | |
2309 | { | |
2310 | size_t hash = dp_netdev_flow_hash(mega_ufid); | |
2311 | struct megaflow_to_mark_data *data; | |
2312 | ||
2313 | CMAP_FOR_EACH_WITH_HASH (data, node, hash, &flow_mark.megaflow_to_mark) { | |
2314 | if (ovs_u128_equals(*mega_ufid, data->mega_ufid)) { | |
2315 | cmap_remove(&flow_mark.megaflow_to_mark, | |
2316 | CONST_CAST(struct cmap_node *, &data->node), hash); | |
5752eae4 | 2317 | ovsrcu_postpone(free, data); |
241bad15 YL |
2318 | return; |
2319 | } | |
2320 | } | |
2321 | ||
2322 | VLOG_WARN("Masked ufid "UUID_FMT" is not associated with a mark?\n", | |
2323 | UUID_ARGS((struct uuid *)mega_ufid)); | |
2324 | } | |
2325 | ||
2326 | static inline uint32_t | |
2327 | megaflow_to_mark_find(const ovs_u128 *mega_ufid) | |
2328 | { | |
2329 | size_t hash = dp_netdev_flow_hash(mega_ufid); | |
2330 | struct megaflow_to_mark_data *data; | |
2331 | ||
2332 | CMAP_FOR_EACH_WITH_HASH (data, node, hash, &flow_mark.megaflow_to_mark) { | |
2333 | if (ovs_u128_equals(*mega_ufid, data->mega_ufid)) { | |
2334 | return data->mark; | |
2335 | } | |
2336 | } | |
2337 | ||
5d1765d3 IM |
2338 | VLOG_DBG("Mark id for ufid "UUID_FMT" was not found\n", |
2339 | UUID_ARGS((struct uuid *)mega_ufid)); | |
241bad15 YL |
2340 | return INVALID_FLOW_MARK; |
2341 | } | |
2342 | ||
2343 | /* associate mark with a flow, which is 1:N mapping */ | |
2344 | static void | |
2345 | mark_to_flow_associate(const uint32_t mark, struct dp_netdev_flow *flow) | |
2346 | { | |
2347 | dp_netdev_flow_ref(flow); | |
2348 | ||
2349 | cmap_insert(&flow_mark.mark_to_flow, | |
2350 | CONST_CAST(struct cmap_node *, &flow->mark_node), | |
2351 | hash_int(mark, 0)); | |
2352 | flow->mark = mark; | |
2353 | ||
2354 | VLOG_DBG("Associated dp_netdev flow %p with mark %u\n", flow, mark); | |
2355 | } | |
2356 | ||
2357 | static bool | |
2358 | flow_mark_has_no_ref(uint32_t mark) | |
2359 | { | |
2360 | struct dp_netdev_flow *flow; | |
2361 | ||
2362 | CMAP_FOR_EACH_WITH_HASH (flow, mark_node, hash_int(mark, 0), | |
2363 | &flow_mark.mark_to_flow) { | |
2364 | if (flow->mark == mark) { | |
2365 | return false; | |
2366 | } | |
2367 | } | |
2368 | ||
2369 | return true; | |
2370 | } | |
2371 | ||
2372 | static int | |
2373 | mark_to_flow_disassociate(struct dp_netdev_pmd_thread *pmd, | |
2374 | struct dp_netdev_flow *flow) | |
2375 | { | |
2376 | int ret = 0; | |
2377 | uint32_t mark = flow->mark; | |
2378 | struct cmap_node *mark_node = CONST_CAST(struct cmap_node *, | |
2379 | &flow->mark_node); | |
2380 | ||
2381 | cmap_remove(&flow_mark.mark_to_flow, mark_node, hash_int(mark, 0)); | |
2382 | flow->mark = INVALID_FLOW_MARK; | |
2383 | ||
2384 | /* | |
2385 | * no flow is referencing the mark any more? If so, let's | |
2386 | * remove the flow from hardware and free the mark. | |
2387 | */ | |
2388 | if (flow_mark_has_no_ref(mark)) { | |
30115809 | 2389 | struct netdev *port; |
241bad15 YL |
2390 | odp_port_t in_port = flow->flow.in_port.odp_port; |
2391 | ||
1061dc7c | 2392 | port = netdev_ports_get(in_port, pmd->dp->class); |
241bad15 | 2393 | if (port) { |
e7cb123f IM |
2394 | /* Taking a global 'port_mutex' to fulfill thread safety |
2395 | * restrictions for the netdev-offload-dpdk module. */ | |
2396 | ovs_mutex_lock(&pmd->dp->port_mutex); | |
30115809 | 2397 | ret = netdev_flow_del(port, &flow->mega_ufid, NULL); |
e7cb123f | 2398 | ovs_mutex_unlock(&pmd->dp->port_mutex); |
30115809 | 2399 | netdev_close(port); |
241bad15 | 2400 | } |
241bad15 YL |
2401 | |
2402 | flow_mark_free(mark); | |
2403 | VLOG_DBG("Freed flow mark %u\n", mark); | |
2404 | ||
2405 | megaflow_to_mark_disassociate(&flow->mega_ufid); | |
2406 | } | |
2407 | dp_netdev_flow_unref(flow); | |
2408 | ||
2409 | return ret; | |
2410 | } | |
2411 | ||
2412 | static void | |
2413 | flow_mark_flush(struct dp_netdev_pmd_thread *pmd) | |
2414 | { | |
2415 | struct dp_netdev_flow *flow; | |
2416 | ||
2417 | CMAP_FOR_EACH (flow, mark_node, &flow_mark.mark_to_flow) { | |
2418 | if (flow->pmd_id == pmd->core_id) { | |
02bb2824 | 2419 | queue_netdev_flow_del(pmd, flow); |
241bad15 YL |
2420 | } |
2421 | } | |
2422 | } | |
2423 | ||
aab96ec4 YL |
2424 | static struct dp_netdev_flow * |
2425 | mark_to_flow_find(const struct dp_netdev_pmd_thread *pmd, | |
2426 | const uint32_t mark) | |
2427 | { | |
2428 | struct dp_netdev_flow *flow; | |
2429 | ||
2430 | CMAP_FOR_EACH_WITH_HASH (flow, mark_node, hash_int(mark, 0), | |
2431 | &flow_mark.mark_to_flow) { | |
2432 | if (flow->mark == mark && flow->pmd_id == pmd->core_id && | |
2433 | flow->dead == false) { | |
2434 | return flow; | |
2435 | } | |
2436 | } | |
2437 | ||
2438 | return NULL; | |
2439 | } | |
2440 | ||
02bb2824 YL |
2441 | static struct dp_flow_offload_item * |
2442 | dp_netdev_alloc_flow_offload(struct dp_netdev_pmd_thread *pmd, | |
2443 | struct dp_netdev_flow *flow, | |
2444 | int op) | |
2445 | { | |
2446 | struct dp_flow_offload_item *offload; | |
2447 | ||
2448 | offload = xzalloc(sizeof(*offload)); | |
2449 | offload->pmd = pmd; | |
2450 | offload->flow = flow; | |
2451 | offload->op = op; | |
2452 | ||
2453 | dp_netdev_flow_ref(flow); | |
2454 | dp_netdev_pmd_try_ref(pmd); | |
2455 | ||
2456 | return offload; | |
2457 | } | |
2458 | ||
2459 | static void | |
2460 | dp_netdev_free_flow_offload(struct dp_flow_offload_item *offload) | |
2461 | { | |
2462 | dp_netdev_pmd_unref(offload->pmd); | |
2463 | dp_netdev_flow_unref(offload->flow); | |
2464 | ||
2465 | free(offload->actions); | |
2466 | free(offload); | |
2467 | } | |
2468 | ||
2469 | static void | |
2470 | dp_netdev_append_flow_offload(struct dp_flow_offload_item *offload) | |
2471 | { | |
2472 | ovs_mutex_lock(&dp_flow_offload.mutex); | |
2473 | ovs_list_push_back(&dp_flow_offload.list, &offload->node); | |
2474 | xpthread_cond_signal(&dp_flow_offload.cond); | |
2475 | ovs_mutex_unlock(&dp_flow_offload.mutex); | |
2476 | } | |
2477 | ||
2478 | static int | |
2479 | dp_netdev_flow_offload_del(struct dp_flow_offload_item *offload) | |
2480 | { | |
2481 | return mark_to_flow_disassociate(offload->pmd, offload->flow); | |
2482 | } | |
2483 | ||
2484 | /* | |
2485 | * There are two flow offload operations here: addition and modification. | |
2486 | * | |
2487 | * For flow addition, this function does: | |
2488 | * - allocate a new flow mark id | |
2489 | * - perform hardware flow offload | |
2490 | * - associate the flow mark with flow and mega flow | |
2491 | * | |
2492 | * For flow modification, both flow mark and the associations are still | |
2493 | * valid, thus only item 2 needed. | |
2494 | */ | |
2495 | static int | |
2496 | dp_netdev_flow_offload_put(struct dp_flow_offload_item *offload) | |
2497 | { | |
02bb2824 | 2498 | struct dp_netdev_pmd_thread *pmd = offload->pmd; |
319a9bb3 | 2499 | const struct dpif_class *dpif_class = pmd->dp->class; |
02bb2824 YL |
2500 | struct dp_netdev_flow *flow = offload->flow; |
2501 | odp_port_t in_port = flow->flow.in_port.odp_port; | |
2502 | bool modification = offload->op == DP_NETDEV_FLOW_OFFLOAD_OP_MOD; | |
2503 | struct offload_info info; | |
30115809 | 2504 | struct netdev *port; |
02bb2824 YL |
2505 | uint32_t mark; |
2506 | int ret; | |
2507 | ||
2508 | if (flow->dead) { | |
2509 | return -1; | |
2510 | } | |
2511 | ||
2512 | if (modification) { | |
2513 | mark = flow->mark; | |
2514 | ovs_assert(mark != INVALID_FLOW_MARK); | |
2515 | } else { | |
2516 | /* | |
2517 | * If a mega flow has already been offloaded (from other PMD | |
2518 | * instances), do not offload it again. | |
2519 | */ | |
2520 | mark = megaflow_to_mark_find(&flow->mega_ufid); | |
2521 | if (mark != INVALID_FLOW_MARK) { | |
2522 | VLOG_DBG("Flow has already been offloaded with mark %u\n", mark); | |
2523 | if (flow->mark != INVALID_FLOW_MARK) { | |
2524 | ovs_assert(flow->mark == mark); | |
2525 | } else { | |
2526 | mark_to_flow_associate(mark, flow); | |
2527 | } | |
2528 | return 0; | |
2529 | } | |
2530 | ||
2531 | mark = flow_mark_alloc(); | |
2532 | if (mark == INVALID_FLOW_MARK) { | |
2533 | VLOG_ERR("Failed to allocate flow mark!\n"); | |
2534 | } | |
2535 | } | |
2536 | info.flow_mark = mark; | |
319a9bb3 | 2537 | info.dpif_class = dpif_class; |
02bb2824 | 2538 | |
1061dc7c | 2539 | port = netdev_ports_get(in_port, pmd->dp->class); |
30115809 IM |
2540 | if (!port || netdev_vport_is_vport_class(port->netdev_class)) { |
2541 | netdev_close(port); | |
0a5cba65 | 2542 | goto err_free; |
02bb2824 | 2543 | } |
e7cb123f IM |
2544 | /* Taking a global 'port_mutex' to fulfill thread safety restrictions for |
2545 | * the netdev-offload-dpdk module. */ | |
2546 | ovs_mutex_lock(&pmd->dp->port_mutex); | |
30115809 | 2547 | ret = netdev_flow_put(port, &offload->match, |
02bb2824 YL |
2548 | CONST_CAST(struct nlattr *, offload->actions), |
2549 | offload->actions_len, &flow->mega_ufid, &info, | |
2550 | NULL); | |
e7cb123f | 2551 | ovs_mutex_unlock(&pmd->dp->port_mutex); |
30115809 | 2552 | netdev_close(port); |
02bb2824 YL |
2553 | |
2554 | if (ret) { | |
0a5cba65 | 2555 | goto err_free; |
02bb2824 YL |
2556 | } |
2557 | ||
2558 | if (!modification) { | |
2559 | megaflow_to_mark_associate(&flow->mega_ufid, mark); | |
2560 | mark_to_flow_associate(mark, flow); | |
2561 | } | |
02bb2824 | 2562 | return 0; |
0a5cba65 IM |
2563 | |
2564 | err_free: | |
2565 | if (!modification) { | |
2566 | flow_mark_free(mark); | |
2567 | } else { | |
2568 | mark_to_flow_disassociate(pmd, flow); | |
2569 | } | |
2570 | return -1; | |
02bb2824 YL |
2571 | } |
2572 | ||
2573 | static void * | |
2574 | dp_netdev_flow_offload_main(void *data OVS_UNUSED) | |
2575 | { | |
2576 | struct dp_flow_offload_item *offload; | |
2577 | struct ovs_list *list; | |
2578 | const char *op; | |
2579 | int ret; | |
2580 | ||
2581 | for (;;) { | |
2582 | ovs_mutex_lock(&dp_flow_offload.mutex); | |
2583 | if (ovs_list_is_empty(&dp_flow_offload.list)) { | |
2584 | ovsrcu_quiesce_start(); | |
2585 | ovs_mutex_cond_wait(&dp_flow_offload.cond, | |
2586 | &dp_flow_offload.mutex); | |
6c95dbf9 | 2587 | ovsrcu_quiesce_end(); |
02bb2824 YL |
2588 | } |
2589 | list = ovs_list_pop_front(&dp_flow_offload.list); | |
2590 | offload = CONTAINER_OF(list, struct dp_flow_offload_item, node); | |
2591 | ovs_mutex_unlock(&dp_flow_offload.mutex); | |
2592 | ||
2593 | switch (offload->op) { | |
2594 | case DP_NETDEV_FLOW_OFFLOAD_OP_ADD: | |
2595 | op = "add"; | |
2596 | ret = dp_netdev_flow_offload_put(offload); | |
2597 | break; | |
2598 | case DP_NETDEV_FLOW_OFFLOAD_OP_MOD: | |
2599 | op = "modify"; | |
2600 | ret = dp_netdev_flow_offload_put(offload); | |
2601 | break; | |
2602 | case DP_NETDEV_FLOW_OFFLOAD_OP_DEL: | |
2603 | op = "delete"; | |
2604 | ret = dp_netdev_flow_offload_del(offload); | |
2605 | break; | |
2606 | default: | |
2607 | OVS_NOT_REACHED(); | |
2608 | } | |
2609 | ||
2610 | VLOG_DBG("%s to %s netdev flow\n", | |
2611 | ret == 0 ? "succeed" : "failed", op); | |
2612 | dp_netdev_free_flow_offload(offload); | |
ef32a1a3 | 2613 | ovsrcu_quiesce(); |
02bb2824 YL |
2614 | } |
2615 | ||
2616 | return NULL; | |
2617 | } | |
2618 | ||
2619 | static void | |
2620 | queue_netdev_flow_del(struct dp_netdev_pmd_thread *pmd, | |
2621 | struct dp_netdev_flow *flow) | |
2622 | { | |
2623 | struct dp_flow_offload_item *offload; | |
2624 | ||
2625 | if (ovsthread_once_start(&offload_thread_once)) { | |
2626 | xpthread_cond_init(&dp_flow_offload.cond, NULL); | |
2627 | ovs_thread_create("dp_netdev_flow_offload", | |
2628 | dp_netdev_flow_offload_main, NULL); | |
2629 | ovsthread_once_done(&offload_thread_once); | |
2630 | } | |
2631 | ||
2632 | offload = dp_netdev_alloc_flow_offload(pmd, flow, | |
2633 | DP_NETDEV_FLOW_OFFLOAD_OP_DEL); | |
2634 | dp_netdev_append_flow_offload(offload); | |
2635 | } | |
2636 | ||
2637 | static void | |
2638 | queue_netdev_flow_put(struct dp_netdev_pmd_thread *pmd, | |
2639 | struct dp_netdev_flow *flow, struct match *match, | |
2640 | const struct nlattr *actions, size_t actions_len) | |
2641 | { | |
2642 | struct dp_flow_offload_item *offload; | |
2643 | int op; | |
2644 | ||
2645 | if (!netdev_is_flow_api_enabled()) { | |
2646 | return; | |
2647 | } | |
2648 | ||
2649 | if (ovsthread_once_start(&offload_thread_once)) { | |
2650 | xpthread_cond_init(&dp_flow_offload.cond, NULL); | |
2651 | ovs_thread_create("dp_netdev_flow_offload", | |
2652 | dp_netdev_flow_offload_main, NULL); | |
2653 | ovsthread_once_done(&offload_thread_once); | |
2654 | } | |
2655 | ||
2656 | if (flow->mark != INVALID_FLOW_MARK) { | |
2657 | op = DP_NETDEV_FLOW_OFFLOAD_OP_MOD; | |
2658 | } else { | |
2659 | op = DP_NETDEV_FLOW_OFFLOAD_OP_ADD; | |
2660 | } | |
2661 | offload = dp_netdev_alloc_flow_offload(pmd, flow, op); | |
2662 | offload->match = *match; | |
2663 | offload->actions = xmalloc(actions_len); | |
2664 | memcpy(offload->actions, actions, actions_len); | |
2665 | offload->actions_len = actions_len; | |
2666 | ||
2667 | dp_netdev_append_flow_offload(offload); | |
2668 | } | |
2669 | ||
72865317 | 2670 | static void |
1c1e46ed AW |
2671 | dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread *pmd, |
2672 | struct dp_netdev_flow *flow) | |
2673 | OVS_REQUIRES(pmd->flow_mutex) | |
72865317 | 2674 | { |
9f361d6b | 2675 | struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node); |
3453b4d6 JS |
2676 | struct dpcls *cls; |
2677 | odp_port_t in_port = flow->flow.in_port.odp_port; | |
2c0ea78f | 2678 | |
3453b4d6 JS |
2679 | cls = dp_netdev_pmd_lookup_dpcls(pmd, in_port); |
2680 | ovs_assert(cls != NULL); | |
2681 | dpcls_remove(cls, &flow->cr); | |
1c1e46ed | 2682 | cmap_remove(&pmd->flow_table, node, dp_netdev_flow_hash(&flow->ufid)); |
241bad15 | 2683 | if (flow->mark != INVALID_FLOW_MARK) { |
02bb2824 | 2684 | queue_netdev_flow_del(pmd, flow); |
241bad15 | 2685 | } |
9bbf1c3d | 2686 | flow->dead = true; |
ed79f89a DDP |
2687 | |
2688 | dp_netdev_flow_unref(flow); | |
72865317 BP |
2689 | } |
2690 | ||
2691 | static void | |
1c1e46ed | 2692 | dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd) |
72865317 | 2693 | { |
78c8df12 | 2694 | struct dp_netdev_flow *netdev_flow; |
72865317 | 2695 | |
1c1e46ed AW |
2696 | ovs_mutex_lock(&pmd->flow_mutex); |
2697 | CMAP_FOR_EACH (netdev_flow, node, &pmd->flow_table) { | |
2698 | dp_netdev_pmd_remove_flow(pmd, netdev_flow); | |
72865317 | 2699 | } |
1c1e46ed | 2700 | ovs_mutex_unlock(&pmd->flow_mutex); |
72865317 BP |
2701 | } |
2702 | ||
2703 | static int | |
2704 | dpif_netdev_flow_flush(struct dpif *dpif) | |
2705 | { | |
2706 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
1c1e46ed AW |
2707 | struct dp_netdev_pmd_thread *pmd; |
2708 | ||
2709 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
2710 | dp_netdev_pmd_flow_flush(pmd); | |
2711 | } | |
5279f8fd | 2712 | |
72865317 BP |
2713 | return 0; |
2714 | } | |
2715 | ||
b0ec0f27 | 2716 | struct dp_netdev_port_state { |
e9985d6a | 2717 | struct hmap_position position; |
4c738a8d | 2718 | char *name; |
b0ec0f27 BP |
2719 | }; |
2720 | ||
2721 | static int | |
2722 | dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep) | |
2723 | { | |
2724 | *statep = xzalloc(sizeof(struct dp_netdev_port_state)); | |
2725 | return 0; | |
2726 | } | |
2727 | ||
72865317 | 2728 | static int |
b0ec0f27 | 2729 | dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_, |
4c738a8d | 2730 | struct dpif_port *dpif_port) |
72865317 | 2731 | { |
b0ec0f27 | 2732 | struct dp_netdev_port_state *state = state_; |
72865317 | 2733 | struct dp_netdev *dp = get_dp_netdev(dpif); |
e9985d6a | 2734 | struct hmap_node *node; |
ff073a71 | 2735 | int retval; |
72865317 | 2736 | |
e9985d6a DDP |
2737 | ovs_mutex_lock(&dp->port_mutex); |
2738 | node = hmap_at_position(&dp->ports, &state->position); | |
ff073a71 BP |
2739 | if (node) { |
2740 | struct dp_netdev_port *port; | |
5279f8fd | 2741 | |
ff073a71 BP |
2742 | port = CONTAINER_OF(node, struct dp_netdev_port, node); |
2743 | ||
2744 | free(state->name); | |
2745 | state->name = xstrdup(netdev_get_name(port->netdev)); | |
2746 | dpif_port->name = state->name; | |
2747 | dpif_port->type = port->type; | |
35303d71 | 2748 | dpif_port->port_no = port->port_no; |
ff073a71 BP |
2749 | |
2750 | retval = 0; | |
2751 | } else { | |
2752 | retval = EOF; | |
72865317 | 2753 | } |
e9985d6a | 2754 | ovs_mutex_unlock(&dp->port_mutex); |
5279f8fd | 2755 | |
ff073a71 | 2756 | return retval; |
b0ec0f27 BP |
2757 | } |
2758 | ||
2759 | static int | |
4c738a8d | 2760 | dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_) |
b0ec0f27 | 2761 | { |
4c738a8d BP |
2762 | struct dp_netdev_port_state *state = state_; |
2763 | free(state->name); | |
b0ec0f27 BP |
2764 | free(state); |
2765 | return 0; | |
72865317 BP |
2766 | } |
2767 | ||
2768 | static int | |
67a4917b | 2769 | dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED) |
72865317 BP |
2770 | { |
2771 | struct dpif_netdev *dpif = dpif_netdev_cast(dpif_); | |
d33ed218 | 2772 | uint64_t new_port_seq; |
5279f8fd BP |
2773 | int error; |
2774 | ||
d33ed218 BP |
2775 | new_port_seq = seq_read(dpif->dp->port_seq); |
2776 | if (dpif->last_port_seq != new_port_seq) { | |
2777 | dpif->last_port_seq = new_port_seq; | |
5279f8fd | 2778 | error = ENOBUFS; |
72865317 | 2779 | } else { |
5279f8fd | 2780 | error = EAGAIN; |
72865317 | 2781 | } |
5279f8fd BP |
2782 | |
2783 | return error; | |
72865317 BP |
2784 | } |
2785 | ||
2786 | static void | |
2787 | dpif_netdev_port_poll_wait(const struct dpif *dpif_) | |
2788 | { | |
2789 | struct dpif_netdev *dpif = dpif_netdev_cast(dpif_); | |
5279f8fd | 2790 | |
d33ed218 | 2791 | seq_wait(dpif->dp->port_seq, dpif->last_port_seq); |
8a4e3a85 BP |
2792 | } |
2793 | ||
2794 | static struct dp_netdev_flow * | |
0de8783a | 2795 | dp_netdev_flow_cast(const struct dpcls_rule *cr) |
8a4e3a85 BP |
2796 | { |
2797 | return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL; | |
72865317 BP |
2798 | } |
2799 | ||
9bbf1c3d DDP |
2800 | static bool dp_netdev_flow_ref(struct dp_netdev_flow *flow) |
2801 | { | |
2802 | return ovs_refcount_try_ref_rcu(&flow->ref_cnt); | |
2803 | } | |
2804 | ||
79df317f DDP |
2805 | /* netdev_flow_key utilities. |
2806 | * | |
2807 | * netdev_flow_key is basically a miniflow. We use these functions | |
2808 | * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow | |
2809 | * functions (miniflow_clone_inline, miniflow_equal, ...), because: | |
2810 | * | |
2811 | * - Since we are dealing exclusively with miniflows created by | |
2812 | * miniflow_extract(), if the map is different the miniflow is different. | |
2813 | * Therefore we can be faster by comparing the map and the miniflow in a | |
2814 | * single memcmp(). | |
5fcff47b | 2815 | * - These functions can be inlined by the compiler. */ |
79df317f | 2816 | |
361d808d | 2817 | /* Given the number of bits set in miniflow's maps, returns the size of the |
caeb4906 | 2818 | * 'netdev_flow_key.mf' */ |
361d808d JR |
2819 | static inline size_t |
2820 | netdev_flow_key_size(size_t flow_u64s) | |
79df317f | 2821 | { |
361d808d | 2822 | return sizeof(struct miniflow) + MINIFLOW_VALUES_SIZE(flow_u64s); |
79df317f DDP |
2823 | } |
2824 | ||
79df317f DDP |
2825 | static inline bool |
2826 | netdev_flow_key_equal(const struct netdev_flow_key *a, | |
0de8783a JR |
2827 | const struct netdev_flow_key *b) |
2828 | { | |
caeb4906 JR |
2829 | /* 'b->len' may be not set yet. */ |
2830 | return a->hash == b->hash && !memcmp(&a->mf, &b->mf, a->len); | |
0de8783a JR |
2831 | } |
2832 | ||
2833 | /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow. | |
d79a39fe | 2834 | * The maps are compared bitwise, so both 'key->mf' and 'mf' must have been |
0de8783a JR |
2835 | * generated by miniflow_extract. */ |
2836 | static inline bool | |
2837 | netdev_flow_key_equal_mf(const struct netdev_flow_key *key, | |
2838 | const struct miniflow *mf) | |
79df317f | 2839 | { |
caeb4906 | 2840 | return !memcmp(&key->mf, mf, key->len); |
79df317f DDP |
2841 | } |
2842 | ||
2843 | static inline void | |
2844 | netdev_flow_key_clone(struct netdev_flow_key *dst, | |
0de8783a JR |
2845 | const struct netdev_flow_key *src) |
2846 | { | |
caeb4906 JR |
2847 | memcpy(dst, src, |
2848 | offsetof(struct netdev_flow_key, mf) + src->len); | |
0de8783a JR |
2849 | } |
2850 | ||
0de8783a JR |
2851 | /* Initialize a netdev_flow_key 'mask' from 'match'. */ |
2852 | static inline void | |
2853 | netdev_flow_mask_init(struct netdev_flow_key *mask, | |
2854 | const struct match *match) | |
2855 | { | |
09b0fa9c | 2856 | uint64_t *dst = miniflow_values(&mask->mf); |
5fcff47b | 2857 | struct flowmap fmap; |
0de8783a | 2858 | uint32_t hash = 0; |
5fcff47b | 2859 | size_t idx; |
0de8783a JR |
2860 | |
2861 | /* Only check masks that make sense for the flow. */ | |
5fcff47b JR |
2862 | flow_wc_map(&match->flow, &fmap); |
2863 | flowmap_init(&mask->mf.map); | |
0de8783a | 2864 | |
5fcff47b JR |
2865 | FLOWMAP_FOR_EACH_INDEX(idx, fmap) { |
2866 | uint64_t mask_u64 = flow_u64_value(&match->wc.masks, idx); | |
0de8783a | 2867 | |
5fcff47b JR |
2868 | if (mask_u64) { |
2869 | flowmap_set(&mask->mf.map, idx, 1); | |
2870 | *dst++ = mask_u64; | |
2871 | hash = hash_add64(hash, mask_u64); | |
0de8783a | 2872 | } |
0de8783a JR |
2873 | } |
2874 | ||
5fcff47b | 2875 | map_t map; |
0de8783a | 2876 | |
5fcff47b JR |
2877 | FLOWMAP_FOR_EACH_MAP (map, mask->mf.map) { |
2878 | hash = hash_add64(hash, map); | |
2879 | } | |
0de8783a | 2880 | |
5fcff47b | 2881 | size_t n = dst - miniflow_get_values(&mask->mf); |
0de8783a | 2882 | |
d70e8c28 | 2883 | mask->hash = hash_finish(hash, n * 8); |
0de8783a JR |
2884 | mask->len = netdev_flow_key_size(n); |
2885 | } | |
2886 | ||
361d808d | 2887 | /* Initializes 'dst' as a copy of 'flow' masked with 'mask'. */ |
0de8783a JR |
2888 | static inline void |
2889 | netdev_flow_key_init_masked(struct netdev_flow_key *dst, | |
2890 | const struct flow *flow, | |
2891 | const struct netdev_flow_key *mask) | |
79df317f | 2892 | { |
09b0fa9c JR |
2893 | uint64_t *dst_u64 = miniflow_values(&dst->mf); |
2894 | const uint64_t *mask_u64 = miniflow_get_values(&mask->mf); | |
0de8783a | 2895 | uint32_t hash = 0; |
d70e8c28 | 2896 | uint64_t value; |
0de8783a JR |
2897 | |
2898 | dst->len = mask->len; | |
361d808d | 2899 | dst->mf = mask->mf; /* Copy maps. */ |
0de8783a | 2900 | |
5fcff47b | 2901 | FLOW_FOR_EACH_IN_MAPS(value, flow, mask->mf.map) { |
d70e8c28 JR |
2902 | *dst_u64 = value & *mask_u64++; |
2903 | hash = hash_add64(hash, *dst_u64++); | |
0de8783a | 2904 | } |
09b0fa9c JR |
2905 | dst->hash = hash_finish(hash, |
2906 | (dst_u64 - miniflow_get_values(&dst->mf)) * 8); | |
0de8783a JR |
2907 | } |
2908 | ||
9bbf1c3d DDP |
2909 | static inline bool |
2910 | emc_entry_alive(struct emc_entry *ce) | |
2911 | { | |
2912 | return ce->flow && !ce->flow->dead; | |
2913 | } | |
2914 | ||
2915 | static void | |
2916 | emc_clear_entry(struct emc_entry *ce) | |
2917 | { | |
2918 | if (ce->flow) { | |
2919 | dp_netdev_flow_unref(ce->flow); | |
2920 | ce->flow = NULL; | |
2921 | } | |
2922 | } | |
2923 | ||
2924 | static inline void | |
2925 | emc_change_entry(struct emc_entry *ce, struct dp_netdev_flow *flow, | |
0de8783a | 2926 | const struct netdev_flow_key *key) |
9bbf1c3d DDP |
2927 | { |
2928 | if (ce->flow != flow) { | |
2929 | if (ce->flow) { | |
2930 | dp_netdev_flow_unref(ce->flow); | |
2931 | } | |
2932 | ||
2933 | if (dp_netdev_flow_ref(flow)) { | |
2934 | ce->flow = flow; | |
2935 | } else { | |
2936 | ce->flow = NULL; | |
2937 | } | |
2938 | } | |
0de8783a JR |
2939 | if (key) { |
2940 | netdev_flow_key_clone(&ce->key, key); | |
9bbf1c3d DDP |
2941 | } |
2942 | } | |
2943 | ||
2944 | static inline void | |
0de8783a | 2945 | emc_insert(struct emc_cache *cache, const struct netdev_flow_key *key, |
9bbf1c3d DDP |
2946 | struct dp_netdev_flow *flow) |
2947 | { | |
2948 | struct emc_entry *to_be_replaced = NULL; | |
2949 | struct emc_entry *current_entry; | |
2950 | ||
0de8783a JR |
2951 | EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) { |
2952 | if (netdev_flow_key_equal(¤t_entry->key, key)) { | |
9bbf1c3d | 2953 | /* We found the entry with the 'mf' miniflow */ |
0de8783a | 2954 | emc_change_entry(current_entry, flow, NULL); |
9bbf1c3d DDP |
2955 | return; |
2956 | } | |
2957 | ||
2958 | /* Replacement policy: put the flow in an empty (not alive) entry, or | |
2959 | * in the first entry where it can be */ | |
2960 | if (!to_be_replaced | |
2961 | || (emc_entry_alive(to_be_replaced) | |
2962 | && !emc_entry_alive(current_entry)) | |
0de8783a | 2963 | || current_entry->key.hash < to_be_replaced->key.hash) { |
9bbf1c3d DDP |
2964 | to_be_replaced = current_entry; |
2965 | } | |
2966 | } | |
2967 | /* We didn't find the miniflow in the cache. | |
2968 | * The 'to_be_replaced' entry is where the new flow will be stored */ | |
2969 | ||
0de8783a | 2970 | emc_change_entry(to_be_replaced, flow, key); |
9bbf1c3d DDP |
2971 | } |
2972 | ||
4c30b246 CL |
2973 | static inline void |
2974 | emc_probabilistic_insert(struct dp_netdev_pmd_thread *pmd, | |
2975 | const struct netdev_flow_key *key, | |
2976 | struct dp_netdev_flow *flow) | |
2977 | { | |
2978 | /* Insert an entry into the EMC based on probability value 'min'. By | |
2979 | * default the value is UINT32_MAX / 100 which yields an insertion | |
2980 | * probability of 1/100 ie. 1% */ | |
2981 | ||
2fbadeb6 | 2982 | uint32_t min = pmd->ctx.emc_insert_min; |
4c30b246 | 2983 | |
656238ee | 2984 | if (min && random_uint32() <= min) { |
60d8ccae | 2985 | emc_insert(&(pmd->flow_cache).emc_cache, key, flow); |
4c30b246 CL |
2986 | } |
2987 | } | |
2988 | ||
9bbf1c3d | 2989 | static inline struct dp_netdev_flow * |
0de8783a | 2990 | emc_lookup(struct emc_cache *cache, const struct netdev_flow_key *key) |
9bbf1c3d DDP |
2991 | { |
2992 | struct emc_entry *current_entry; | |
2993 | ||
0de8783a JR |
2994 | EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) { |
2995 | if (current_entry->key.hash == key->hash | |
2996 | && emc_entry_alive(current_entry) | |
2997 | && netdev_flow_key_equal_mf(¤t_entry->key, &key->mf)) { | |
9bbf1c3d | 2998 | |
0de8783a | 2999 | /* We found the entry with the 'key->mf' miniflow */ |
9bbf1c3d DDP |
3000 | return current_entry->flow; |
3001 | } | |
3002 | } | |
3003 | ||
3004 | return NULL; | |
3005 | } | |
3006 | ||
60d8ccae YW |
3007 | static inline const struct cmap_node * |
3008 | smc_entry_get(struct dp_netdev_pmd_thread *pmd, const uint32_t hash) | |
3009 | { | |
3010 | struct smc_cache *cache = &(pmd->flow_cache).smc_cache; | |
3011 | struct smc_bucket *bucket = &cache->buckets[hash & SMC_MASK]; | |
3012 | uint16_t sig = hash >> 16; | |
3013 | uint16_t index = UINT16_MAX; | |
3014 | ||
3015 | for (int i = 0; i < SMC_ENTRY_PER_BUCKET; i++) { | |
3016 | if (bucket->sig[i] == sig) { | |
3017 | index = bucket->flow_idx[i]; | |
3018 | break; | |
3019 | } | |
3020 | } | |
3021 | if (index != UINT16_MAX) { | |
3022 | return cmap_find_by_index(&pmd->flow_table, index); | |
3023 | } | |
3024 | return NULL; | |
3025 | } | |
3026 | ||
3027 | static void | |
3028 | smc_clear_entry(struct smc_bucket *b, int idx) | |
3029 | { | |
3030 | b->flow_idx[idx] = UINT16_MAX; | |
3031 | } | |
3032 | ||
3033 | /* Insert the flow_table index into SMC. Insertion may fail when 1) SMC is | |
3034 | * turned off, 2) the flow_table index is larger than uint16_t can handle. | |
3035 | * If there is already an SMC entry having same signature, the index will be | |
3036 | * updated. If there is no existing entry, but an empty entry is available, | |
3037 | * the empty entry will be taken. If no empty entry or existing same signature, | |
3038 | * a random entry from the hashed bucket will be picked. */ | |
3039 | static inline void | |
3040 | smc_insert(struct dp_netdev_pmd_thread *pmd, | |
3041 | const struct netdev_flow_key *key, | |
3042 | uint32_t hash) | |
3043 | { | |
3044 | struct smc_cache *smc_cache = &(pmd->flow_cache).smc_cache; | |
3045 | struct smc_bucket *bucket = &smc_cache->buckets[key->hash & SMC_MASK]; | |
3046 | uint16_t index; | |
3047 | uint32_t cmap_index; | |
3048 | bool smc_enable_db; | |
3049 | int i; | |
3050 | ||
3051 | atomic_read_relaxed(&pmd->dp->smc_enable_db, &smc_enable_db); | |
3052 | if (!smc_enable_db) { | |
3053 | return; | |
3054 | } | |
3055 | ||
3056 | cmap_index = cmap_find_index(&pmd->flow_table, hash); | |
3057 | index = (cmap_index >= UINT16_MAX) ? UINT16_MAX : (uint16_t)cmap_index; | |
3058 | ||
3059 | /* If the index is larger than SMC can handle (uint16_t), we don't | |
3060 | * insert */ | |
3061 | if (index == UINT16_MAX) { | |
3062 | return; | |
3063 | } | |
3064 | ||
3065 | /* If an entry with same signature already exists, update the index */ | |
3066 | uint16_t sig = key->hash >> 16; | |
3067 | for (i = 0; i < SMC_ENTRY_PER_BUCKET; i++) { | |
3068 | if (bucket->sig[i] == sig) { | |
3069 | bucket->flow_idx[i] = index; | |
3070 | return; | |
3071 | } | |
3072 | } | |
3073 | /* If there is an empty entry, occupy it. */ | |
3074 | for (i = 0; i < SMC_ENTRY_PER_BUCKET; i++) { | |
3075 | if (bucket->flow_idx[i] == UINT16_MAX) { | |
3076 | bucket->sig[i] = sig; | |
3077 | bucket->flow_idx[i] = index; | |
3078 | return; | |
3079 | } | |
3080 | } | |
3081 | /* Otherwise, pick a random entry. */ | |
3082 | i = random_uint32() % SMC_ENTRY_PER_BUCKET; | |
3083 | bucket->sig[i] = sig; | |
3084 | bucket->flow_idx[i] = index; | |
3085 | } | |
3086 | ||
72865317 | 3087 | static struct dp_netdev_flow * |
3453b4d6 JS |
3088 | dp_netdev_pmd_lookup_flow(struct dp_netdev_pmd_thread *pmd, |
3089 | const struct netdev_flow_key *key, | |
3090 | int *lookup_num_p) | |
2c0ea78f | 3091 | { |
3453b4d6 | 3092 | struct dpcls *cls; |
0de8783a | 3093 | struct dpcls_rule *rule; |
f825fdd4 BP |
3094 | odp_port_t in_port = u32_to_odp(MINIFLOW_GET_U32(&key->mf, |
3095 | in_port.odp_port)); | |
3453b4d6 | 3096 | struct dp_netdev_flow *netdev_flow = NULL; |
2c0ea78f | 3097 | |
3453b4d6 JS |
3098 | cls = dp_netdev_pmd_lookup_dpcls(pmd, in_port); |
3099 | if (OVS_LIKELY(cls)) { | |
60d8ccae | 3100 | dpcls_lookup(cls, &key, &rule, 1, lookup_num_p); |
3453b4d6 JS |
3101 | netdev_flow = dp_netdev_flow_cast(rule); |
3102 | } | |
8a4e3a85 | 3103 | return netdev_flow; |
2c0ea78f GS |
3104 | } |
3105 | ||
3106 | static struct dp_netdev_flow * | |
1c1e46ed AW |
3107 | dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread *pmd, |
3108 | const ovs_u128 *ufidp, const struct nlattr *key, | |
3109 | size_t key_len) | |
72865317 | 3110 | { |
1763b4b8 | 3111 | struct dp_netdev_flow *netdev_flow; |
70e5ed6f JS |
3112 | struct flow flow; |
3113 | ovs_u128 ufid; | |
3114 | ||
3115 | /* If a UFID is not provided, determine one based on the key. */ | |
3116 | if (!ufidp && key && key_len | |
f0fb825a | 3117 | && !dpif_netdev_flow_from_nlattrs(key, key_len, &flow, false)) { |
7a5e0ee7 | 3118 | odp_flow_key_hash(&flow, sizeof flow, &ufid); |
70e5ed6f JS |
3119 | ufidp = &ufid; |
3120 | } | |
72865317 | 3121 | |
70e5ed6f JS |
3122 | if (ufidp) { |
3123 | CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, dp_netdev_flow_hash(ufidp), | |
1c1e46ed | 3124 | &pmd->flow_table) { |
2ff8484b | 3125 | if (ovs_u128_equals(netdev_flow->ufid, *ufidp)) { |
70e5ed6f JS |
3126 | return netdev_flow; |
3127 | } | |
72865317 BP |
3128 | } |
3129 | } | |
8a4e3a85 | 3130 | |
72865317 BP |
3131 | return NULL; |
3132 | } | |
3133 | ||
a309e4f5 OM |
3134 | static bool |
3135 | dpif_netdev_get_flow_offload_status(const struct dp_netdev *dp, | |
3136 | const struct dp_netdev_flow *netdev_flow, | |
3137 | struct dpif_flow_stats *stats, | |
3138 | struct dpif_flow_attrs *attrs) | |
3139 | { | |
3140 | uint64_t act_buf[1024 / 8]; | |
3141 | struct nlattr *actions; | |
3142 | struct netdev *netdev; | |
3143 | struct match match; | |
3144 | struct ofpbuf buf; | |
3145 | ||
3146 | int ret = 0; | |
3147 | ||
3148 | if (!netdev_is_flow_api_enabled()) { | |
3149 | return false; | |
3150 | } | |
3151 | ||
3152 | netdev = netdev_ports_get(netdev_flow->flow.in_port.odp_port, dp->class); | |
3153 | if (!netdev) { | |
3154 | return false; | |
3155 | } | |
3156 | ofpbuf_use_stack(&buf, &act_buf, sizeof act_buf); | |
3157 | /* Taking a global 'port_mutex' to fulfill thread safety | |
3158 | * restrictions for the netdev-offload-dpdk module. */ | |
3159 | ovs_mutex_lock(&dp->port_mutex); | |
3160 | ret = netdev_flow_get(netdev, &match, &actions, &netdev_flow->mega_ufid, | |
3161 | stats, attrs, &buf); | |
3162 | ovs_mutex_unlock(&dp->port_mutex); | |
3163 | netdev_close(netdev); | |
3164 | if (ret) { | |
3165 | return false; | |
3166 | } | |
3167 | ||
3168 | return true; | |
3169 | } | |
3170 | ||
72865317 | 3171 | static void |
a309e4f5 OM |
3172 | get_dpif_flow_status(const struct dp_netdev *dp, |
3173 | const struct dp_netdev_flow *netdev_flow_, | |
3174 | struct dpif_flow_stats *stats, | |
3175 | struct dpif_flow_attrs *attrs) | |
feebdea2 | 3176 | { |
a309e4f5 OM |
3177 | struct dpif_flow_stats offload_stats; |
3178 | struct dpif_flow_attrs offload_attrs; | |
eb94da30 DDP |
3179 | struct dp_netdev_flow *netdev_flow; |
3180 | unsigned long long n; | |
3181 | long long used; | |
3182 | uint16_t flags; | |
3183 | ||
3184 | netdev_flow = CONST_CAST(struct dp_netdev_flow *, netdev_flow_); | |
3185 | ||
3186 | atomic_read_relaxed(&netdev_flow->stats.packet_count, &n); | |
3187 | stats->n_packets = n; | |
3188 | atomic_read_relaxed(&netdev_flow->stats.byte_count, &n); | |
3189 | stats->n_bytes = n; | |
3190 | atomic_read_relaxed(&netdev_flow->stats.used, &used); | |
3191 | stats->used = used; | |
3192 | atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags); | |
3193 | stats->tcp_flags = flags; | |
a309e4f5 OM |
3194 | |
3195 | if (dpif_netdev_get_flow_offload_status(dp, netdev_flow, | |
3196 | &offload_stats, &offload_attrs)) { | |
3197 | stats->n_packets += offload_stats.n_packets; | |
3198 | stats->n_bytes += offload_stats.n_bytes; | |
3199 | stats->used = MAX(stats->used, offload_stats.used); | |
3200 | stats->tcp_flags |= offload_stats.tcp_flags; | |
3201 | if (attrs) { | |
3202 | attrs->offloaded = offload_attrs.offloaded; | |
3203 | attrs->dp_layer = offload_attrs.dp_layer; | |
3204 | } | |
3205 | } else if (attrs) { | |
3206 | attrs->offloaded = false; | |
3207 | attrs->dp_layer = "ovs"; | |
3208 | } | |
72865317 BP |
3209 | } |
3210 | ||
7af12bd7 JS |
3211 | /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for |
3212 | * storing the netlink-formatted key/mask. 'key_buf' may be the same as | |
3213 | * 'mask_buf'. Actions will be returned without copying, by relying on RCU to | |
3214 | * protect them. */ | |
6fe09f8c | 3215 | static void |
a309e4f5 OM |
3216 | dp_netdev_flow_to_dpif_flow(const struct dp_netdev *dp, |
3217 | const struct dp_netdev_flow *netdev_flow, | |
7af12bd7 | 3218 | struct ofpbuf *key_buf, struct ofpbuf *mask_buf, |
64bb477f | 3219 | struct dpif_flow *flow, bool terse) |
6fe09f8c | 3220 | { |
64bb477f JS |
3221 | if (terse) { |
3222 | memset(flow, 0, sizeof *flow); | |
3223 | } else { | |
3224 | struct flow_wildcards wc; | |
3225 | struct dp_netdev_actions *actions; | |
3226 | size_t offset; | |
5262eea1 JG |
3227 | struct odp_flow_key_parms odp_parms = { |
3228 | .flow = &netdev_flow->flow, | |
3229 | .mask = &wc.masks, | |
2494ccd7 | 3230 | .support = dp_netdev_support, |
5262eea1 | 3231 | }; |
64bb477f JS |
3232 | |
3233 | miniflow_expand(&netdev_flow->cr.mask->mf, &wc.masks); | |
f4b835bb JR |
3234 | /* in_port is exact matched, but we have left it out from the mask for |
3235 | * optimnization reasons. Add in_port back to the mask. */ | |
3236 | wc.masks.in_port.odp_port = ODPP_NONE; | |
64bb477f JS |
3237 | |
3238 | /* Key */ | |
6fd6ed71 | 3239 | offset = key_buf->size; |
64bb477f | 3240 | flow->key = ofpbuf_tail(key_buf); |
5262eea1 | 3241 | odp_flow_key_from_flow(&odp_parms, key_buf); |
6fd6ed71 | 3242 | flow->key_len = key_buf->size - offset; |
64bb477f JS |
3243 | |
3244 | /* Mask */ | |
6fd6ed71 | 3245 | offset = mask_buf->size; |
64bb477f | 3246 | flow->mask = ofpbuf_tail(mask_buf); |
ec1f6f32 | 3247 | odp_parms.key_buf = key_buf; |
5262eea1 | 3248 | odp_flow_key_from_mask(&odp_parms, mask_buf); |
6fd6ed71 | 3249 | flow->mask_len = mask_buf->size - offset; |
64bb477f JS |
3250 | |
3251 | /* Actions */ | |
3252 | actions = dp_netdev_flow_get_actions(netdev_flow); | |
3253 | flow->actions = actions->actions; | |
3254 | flow->actions_len = actions->size; | |
3255 | } | |
6fe09f8c | 3256 | |
70e5ed6f JS |
3257 | flow->ufid = netdev_flow->ufid; |
3258 | flow->ufid_present = true; | |
1c1e46ed | 3259 | flow->pmd_id = netdev_flow->pmd_id; |
0d6b401c | 3260 | |
a309e4f5 | 3261 | get_dpif_flow_status(dp, netdev_flow, &flow->stats, &flow->attrs); |
342b8904 | 3262 | flow->attrs.dp_extra_info = netdev_flow->dp_extra_info; |
6fe09f8c JS |
3263 | } |
3264 | ||
36956a7d | 3265 | static int |
8c301900 JR |
3266 | dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len, |
3267 | const struct nlattr *mask_key, | |
3268 | uint32_t mask_key_len, const struct flow *flow, | |
f0fb825a | 3269 | struct flow_wildcards *wc, bool probe) |
8c301900 | 3270 | { |
ca8d3442 DDP |
3271 | enum odp_key_fitness fitness; |
3272 | ||
d40533fc | 3273 | fitness = odp_flow_key_to_mask(mask_key, mask_key_len, wc, flow, NULL); |
ca8d3442 | 3274 | if (fitness) { |
f0fb825a EG |
3275 | if (!probe) { |
3276 | /* This should not happen: it indicates that | |
3277 | * odp_flow_key_from_mask() and odp_flow_key_to_mask() | |
3278 | * disagree on the acceptable form of a mask. Log the problem | |
3279 | * as an error, with enough details to enable debugging. */ | |
3280 | static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5); | |
3281 | ||
3282 | if (!VLOG_DROP_ERR(&rl)) { | |
3283 | struct ds s; | |
3284 | ||
3285 | ds_init(&s); | |
3286 | odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s, | |
3287 | true); | |
3288 | VLOG_ERR("internal error parsing flow mask %s (%s)", | |
3289 | ds_cstr(&s), odp_key_fitness_to_string(fitness)); | |
3290 | ds_destroy(&s); | |
3291 | } | |
8c301900 | 3292 | } |
ca8d3442 DDP |
3293 | |
3294 | return EINVAL; | |
8c301900 JR |
3295 | } |
3296 | ||
3297 | return 0; | |
3298 | } | |
3299 | ||
3300 | static int | |
3301 | dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len, | |
f0fb825a | 3302 | struct flow *flow, bool probe) |
36956a7d | 3303 | { |
d40533fc | 3304 | if (odp_flow_key_to_flow(key, key_len, flow, NULL)) { |
f0fb825a EG |
3305 | if (!probe) { |
3306 | /* This should not happen: it indicates that | |
3307 | * odp_flow_key_from_flow() and odp_flow_key_to_flow() disagree on | |
3308 | * the acceptable form of a flow. Log the problem as an error, | |
3309 | * with enough details to enable debugging. */ | |
3310 | static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5); | |
3311 | ||
3312 | if (!VLOG_DROP_ERR(&rl)) { | |
3313 | struct ds s; | |
3314 | ||
3315 | ds_init(&s); | |
3316 | odp_flow_format(key, key_len, NULL, 0, NULL, &s, true); | |
3317 | VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s)); | |
3318 | ds_destroy(&s); | |
3319 | } | |
36956a7d BP |
3320 | } |
3321 | ||
3322 | return EINVAL; | |
3323 | } | |
3324 | ||
5cf3edb3 | 3325 | if (flow->ct_state & DP_NETDEV_CS_UNSUPPORTED_MASK) { |
07659514 JS |
3326 | return EINVAL; |
3327 | } | |
3328 | ||
36956a7d BP |
3329 | return 0; |
3330 | } | |
3331 | ||
72865317 | 3332 | static int |
6fe09f8c | 3333 | dpif_netdev_flow_get(const struct dpif *dpif, const struct dpif_flow_get *get) |
72865317 BP |
3334 | { |
3335 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
1763b4b8 | 3336 | struct dp_netdev_flow *netdev_flow; |
1c1e46ed | 3337 | struct dp_netdev_pmd_thread *pmd; |
c673049c IM |
3338 | struct hmapx to_find = HMAPX_INITIALIZER(&to_find); |
3339 | struct hmapx_node *node; | |
3340 | int error = EINVAL; | |
3341 | ||
3342 | if (get->pmd_id == PMD_ID_NULL) { | |
3343 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
3344 | if (dp_netdev_pmd_try_ref(pmd) && !hmapx_add(&to_find, pmd)) { | |
3345 | dp_netdev_pmd_unref(pmd); | |
3346 | } | |
3347 | } | |
3348 | } else { | |
3349 | pmd = dp_netdev_get_pmd(dp, get->pmd_id); | |
3350 | if (!pmd) { | |
3351 | goto out; | |
3352 | } | |
3353 | hmapx_add(&to_find, pmd); | |
1c1e46ed AW |
3354 | } |
3355 | ||
c673049c IM |
3356 | if (!hmapx_count(&to_find)) { |
3357 | goto out; | |
72865317 | 3358 | } |
1c1e46ed | 3359 | |
c673049c IM |
3360 | HMAPX_FOR_EACH (node, &to_find) { |
3361 | pmd = (struct dp_netdev_pmd_thread *) node->data; | |
3362 | netdev_flow = dp_netdev_pmd_find_flow(pmd, get->ufid, get->key, | |
3363 | get->key_len); | |
3364 | if (netdev_flow) { | |
a309e4f5 OM |
3365 | dp_netdev_flow_to_dpif_flow(dp, netdev_flow, get->buffer, |
3366 | get->buffer, get->flow, false); | |
c673049c IM |
3367 | error = 0; |
3368 | break; | |
3369 | } else { | |
3370 | error = ENOENT; | |
3371 | } | |
3372 | } | |
bc4a05c6 | 3373 | |
c673049c IM |
3374 | HMAPX_FOR_EACH (node, &to_find) { |
3375 | pmd = (struct dp_netdev_pmd_thread *) node->data; | |
3376 | dp_netdev_pmd_unref(pmd); | |
3377 | } | |
3378 | out: | |
3379 | hmapx_destroy(&to_find); | |
5279f8fd | 3380 | return error; |
72865317 BP |
3381 | } |
3382 | ||
241bad15 YL |
3383 | static void |
3384 | dp_netdev_get_mega_ufid(const struct match *match, ovs_u128 *mega_ufid) | |
3385 | { | |
3386 | struct flow masked_flow; | |
3387 | size_t i; | |
3388 | ||
3389 | for (i = 0; i < sizeof(struct flow); i++) { | |
3390 | ((uint8_t *)&masked_flow)[i] = ((uint8_t *)&match->flow)[i] & | |
3391 | ((uint8_t *)&match->wc)[i]; | |
3392 | } | |
7a5e0ee7 | 3393 | odp_flow_key_hash(&masked_flow, sizeof masked_flow, mega_ufid); |
241bad15 YL |
3394 | } |
3395 | ||
0de8783a | 3396 | static struct dp_netdev_flow * |
1c1e46ed AW |
3397 | dp_netdev_flow_add(struct dp_netdev_pmd_thread *pmd, |
3398 | struct match *match, const ovs_u128 *ufid, | |
ae2ceebd | 3399 | const struct nlattr *actions, size_t actions_len) |
1c1e46ed | 3400 | OVS_REQUIRES(pmd->flow_mutex) |
72865317 | 3401 | { |
342b8904 | 3402 | struct ds extra_info = DS_EMPTY_INITIALIZER; |
0de8783a JR |
3403 | struct dp_netdev_flow *flow; |
3404 | struct netdev_flow_key mask; | |
3453b4d6 | 3405 | struct dpcls *cls; |
342b8904 | 3406 | size_t unit; |
f4b835bb JR |
3407 | |
3408 | /* Make sure in_port is exact matched before we read it. */ | |
3409 | ovs_assert(match->wc.masks.in_port.odp_port == ODPP_NONE); | |
3453b4d6 | 3410 | odp_port_t in_port = match->flow.in_port.odp_port; |
ed79f89a | 3411 | |
f4b835bb JR |
3412 | /* As we select the dpcls based on the port number, each netdev flow |
3413 | * belonging to the same dpcls will have the same odp_port value. | |
3414 | * For performance reasons we wildcard odp_port here in the mask. In the | |
3415 | * typical case dp_hash is also wildcarded, and the resulting 8-byte | |
3416 | * chunk {dp_hash, in_port} will be ignored by netdev_flow_mask_init() and | |
3417 | * will not be part of the subtable mask. | |
3418 | * This will speed up the hash computation during dpcls_lookup() because | |
3419 | * there is one less call to hash_add64() in this case. */ | |
3420 | match->wc.masks.in_port.odp_port = 0; | |
0de8783a | 3421 | netdev_flow_mask_init(&mask, match); |
f4b835bb JR |
3422 | match->wc.masks.in_port.odp_port = ODPP_NONE; |
3423 | ||
0de8783a | 3424 | /* Make sure wc does not have metadata. */ |
5fcff47b JR |
3425 | ovs_assert(!FLOWMAP_HAS_FIELD(&mask.mf.map, metadata) |
3426 | && !FLOWMAP_HAS_FIELD(&mask.mf.map, regs)); | |
679ba04c | 3427 | |
0de8783a | 3428 | /* Do not allocate extra space. */ |
caeb4906 | 3429 | flow = xmalloc(sizeof *flow - sizeof flow->cr.flow.mf + mask.len); |
1c1e46ed | 3430 | memset(&flow->stats, 0, sizeof flow->stats); |
0de8783a | 3431 | flow->dead = false; |
11e5cf1f | 3432 | flow->batch = NULL; |
241bad15 | 3433 | flow->mark = INVALID_FLOW_MARK; |
bd5131ba | 3434 | *CONST_CAST(unsigned *, &flow->pmd_id) = pmd->core_id; |
0de8783a | 3435 | *CONST_CAST(struct flow *, &flow->flow) = match->flow; |
70e5ed6f | 3436 | *CONST_CAST(ovs_u128 *, &flow->ufid) = *ufid; |
0de8783a | 3437 | ovs_refcount_init(&flow->ref_cnt); |
0de8783a | 3438 | ovsrcu_set(&flow->actions, dp_netdev_actions_create(actions, actions_len)); |
2c0ea78f | 3439 | |
241bad15 | 3440 | dp_netdev_get_mega_ufid(match, CONST_CAST(ovs_u128 *, &flow->mega_ufid)); |
0de8783a | 3441 | netdev_flow_key_init_masked(&flow->cr.flow, &match->flow, &mask); |
3453b4d6 | 3442 | |
f4b835bb | 3443 | /* Select dpcls for in_port. Relies on in_port to be exact match. */ |
3453b4d6 JS |
3444 | cls = dp_netdev_pmd_find_dpcls(pmd, in_port); |
3445 | dpcls_insert(cls, &flow->cr, &mask); | |
72865317 | 3446 | |
342b8904 IM |
3447 | ds_put_cstr(&extra_info, "miniflow_bits("); |
3448 | FLOWMAP_FOR_EACH_UNIT (unit) { | |
3449 | if (unit) { | |
3450 | ds_put_char(&extra_info, ','); | |
3451 | } | |
3452 | ds_put_format(&extra_info, "%d", | |
3453 | count_1bits(flow->cr.mask->mf.map.bits[unit])); | |
3454 | } | |
3455 | ds_put_char(&extra_info, ')'); | |
3456 | flow->dp_extra_info = ds_steal_cstr(&extra_info); | |
3457 | ds_destroy(&extra_info); | |
3458 | ||
4c75aaab EJ |
3459 | cmap_insert(&pmd->flow_table, CONST_CAST(struct cmap_node *, &flow->node), |
3460 | dp_netdev_flow_hash(&flow->ufid)); | |
3461 | ||
02bb2824 | 3462 | queue_netdev_flow_put(pmd, flow, match, actions, actions_len); |
241bad15 | 3463 | |
beb75a40 | 3464 | if (OVS_UNLIKELY(!VLOG_DROP_DBG((&upcall_rl)))) { |
623540e4 | 3465 | struct ds ds = DS_EMPTY_INITIALIZER; |
9044f2c1 JG |
3466 | struct ofpbuf key_buf, mask_buf; |
3467 | struct odp_flow_key_parms odp_parms = { | |
3468 | .flow = &match->flow, | |
3469 | .mask = &match->wc.masks, | |
3470 | .support = dp_netdev_support, | |
3471 | }; | |
3472 | ||
3473 | ofpbuf_init(&key_buf, 0); | |
3474 | ofpbuf_init(&mask_buf, 0); | |
623540e4 | 3475 | |
9044f2c1 JG |
3476 | odp_flow_key_from_flow(&odp_parms, &key_buf); |
3477 | odp_parms.key_buf = &key_buf; | |
3478 | odp_flow_key_from_mask(&odp_parms, &mask_buf); | |
0de8783a | 3479 | |
623540e4 | 3480 | ds_put_cstr(&ds, "flow_add: "); |
70e5ed6f JS |
3481 | odp_format_ufid(ufid, &ds); |
3482 | ds_put_cstr(&ds, " "); | |
9044f2c1 JG |
3483 | odp_flow_format(key_buf.data, key_buf.size, |
3484 | mask_buf.data, mask_buf.size, | |
3485 | NULL, &ds, false); | |
623540e4 | 3486 | ds_put_cstr(&ds, ", actions:"); |
0722f341 | 3487 | format_odp_actions(&ds, actions, actions_len, NULL); |
623540e4 | 3488 | |
beb75a40 | 3489 | VLOG_DBG("%s", ds_cstr(&ds)); |
623540e4 | 3490 | |
9044f2c1 JG |
3491 | ofpbuf_uninit(&key_buf); |
3492 | ofpbuf_uninit(&mask_buf); | |
beb75a40 JS |
3493 | |
3494 | /* Add a printout of the actual match installed. */ | |
3495 | struct match m; | |
3496 | ds_clear(&ds); | |
3497 | ds_put_cstr(&ds, "flow match: "); | |
3498 | miniflow_expand(&flow->cr.flow.mf, &m.flow); | |
3499 | miniflow_expand(&flow->cr.mask->mf, &m.wc.masks); | |
b2f4b622 | 3500 | memset(&m.tun_md, 0, sizeof m.tun_md); |
beb75a40 JS |
3501 | match_format(&m, NULL, &ds, OFP_DEFAULT_PRIORITY); |
3502 | ||
3503 | VLOG_DBG("%s", ds_cstr(&ds)); | |
3504 | ||
623540e4 EJ |
3505 | ds_destroy(&ds); |
3506 | } | |
3507 | ||
0de8783a | 3508 | return flow; |
72865317 BP |
3509 | } |
3510 | ||
72865317 | 3511 | static int |
f5d317a1 DDP |
3512 | flow_put_on_pmd(struct dp_netdev_pmd_thread *pmd, |
3513 | struct netdev_flow_key *key, | |
3514 | struct match *match, | |
3515 | ovs_u128 *ufid, | |
3516 | const struct dpif_flow_put *put, | |
3517 | struct dpif_flow_stats *stats) | |
72865317 | 3518 | { |
1763b4b8 | 3519 | struct dp_netdev_flow *netdev_flow; |
f5d317a1 | 3520 | int error = 0; |
72865317 | 3521 | |
f5d317a1 DDP |
3522 | if (stats) { |
3523 | memset(stats, 0, sizeof *stats); | |
70e5ed6f JS |
3524 | } |
3525 | ||
1c1e46ed | 3526 | ovs_mutex_lock(&pmd->flow_mutex); |
f5d317a1 | 3527 | netdev_flow = dp_netdev_pmd_lookup_flow(pmd, key, NULL); |
1763b4b8 | 3528 | if (!netdev_flow) { |
89625d1e | 3529 | if (put->flags & DPIF_FP_CREATE) { |
1c1e46ed | 3530 | if (cmap_count(&pmd->flow_table) < MAX_FLOWS) { |
f5d317a1 | 3531 | dp_netdev_flow_add(pmd, match, ufid, put->actions, |
70e5ed6f | 3532 | put->actions_len); |
0de8783a | 3533 | error = 0; |
72865317 | 3534 | } else { |
5279f8fd | 3535 | error = EFBIG; |
72865317 BP |
3536 | } |
3537 | } else { | |
5279f8fd | 3538 | error = ENOENT; |
72865317 BP |
3539 | } |
3540 | } else { | |
beb75a40 | 3541 | if (put->flags & DPIF_FP_MODIFY) { |
8a4e3a85 BP |
3542 | struct dp_netdev_actions *new_actions; |
3543 | struct dp_netdev_actions *old_actions; | |
3544 | ||
3545 | new_actions = dp_netdev_actions_create(put->actions, | |
3546 | put->actions_len); | |
3547 | ||
61e7deb1 BP |
3548 | old_actions = dp_netdev_flow_get_actions(netdev_flow); |
3549 | ovsrcu_set(&netdev_flow->actions, new_actions); | |
679ba04c | 3550 | |
02bb2824 YL |
3551 | queue_netdev_flow_put(pmd, netdev_flow, match, |
3552 | put->actions, put->actions_len); | |
241bad15 | 3553 | |
f5d317a1 | 3554 | if (stats) { |
a309e4f5 | 3555 | get_dpif_flow_status(pmd->dp, netdev_flow, stats, NULL); |
a84cb64a BP |
3556 | } |
3557 | if (put->flags & DPIF_FP_ZERO_STATS) { | |
97447f55 DDP |
3558 | /* XXX: The userspace datapath uses thread local statistics |
3559 | * (for flows), which should be updated only by the owning | |
3560 | * thread. Since we cannot write on stats memory here, | |
3561 | * we choose not to support this flag. Please note: | |
3562 | * - This feature is currently used only by dpctl commands with | |
3563 | * option --clear. | |
3564 | * - Should the need arise, this operation can be implemented | |
3565 | * by keeping a base value (to be update here) for each | |
3566 | * counter, and subtracting it before outputting the stats */ | |
3567 | error = EOPNOTSUPP; | |
72865317 | 3568 | } |
8a4e3a85 | 3569 | |
61e7deb1 | 3570 | ovsrcu_postpone(dp_netdev_actions_free, old_actions); |
2c0ea78f | 3571 | } else if (put->flags & DPIF_FP_CREATE) { |
5279f8fd | 3572 | error = EEXIST; |
2c0ea78f GS |
3573 | } else { |
3574 | /* Overlapping flow. */ | |
3575 | error = EINVAL; | |
72865317 BP |
3576 | } |
3577 | } | |
1c1e46ed | 3578 | ovs_mutex_unlock(&pmd->flow_mutex); |
5279f8fd | 3579 | return error; |
72865317 BP |
3580 | } |
3581 | ||
72865317 | 3582 | static int |
f5d317a1 | 3583 | dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put) |
72865317 BP |
3584 | { |
3585 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
beb75a40 | 3586 | struct netdev_flow_key key, mask; |
1c1e46ed | 3587 | struct dp_netdev_pmd_thread *pmd; |
f5d317a1 DDP |
3588 | struct match match; |
3589 | ovs_u128 ufid; | |
3590 | int error; | |
f0fb825a | 3591 | bool probe = put->flags & DPIF_FP_PROBE; |
72865317 | 3592 | |
f5d317a1 DDP |
3593 | if (put->stats) { |
3594 | memset(put->stats, 0, sizeof *put->stats); | |
3595 | } | |
f0fb825a EG |
3596 | error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow, |
3597 | probe); | |
f5d317a1 DDP |
3598 | if (error) { |
3599 | return error; | |
3600 | } | |
3601 | error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len, | |
3602 | put->mask, put->mask_len, | |
f0fb825a | 3603 | &match.flow, &match.wc, probe); |
f5d317a1 DDP |
3604 | if (error) { |
3605 | return error; | |
1c1e46ed AW |
3606 | } |
3607 | ||
f5d317a1 DDP |
3608 | if (put->ufid) { |
3609 | ufid = *put->ufid; | |
3610 | } else { | |
7a5e0ee7 | 3611 | odp_flow_key_hash(&match.flow, sizeof match.flow, &ufid); |
f5d317a1 DDP |
3612 | } |
3613 | ||
35fe9efb IM |
3614 | /* The Netlink encoding of datapath flow keys cannot express |
3615 | * wildcarding the presence of a VLAN tag. Instead, a missing VLAN | |
3616 | * tag is interpreted as exact match on the fact that there is no | |
3617 | * VLAN. Unless we refactor a lot of code that translates between | |
3618 | * Netlink and struct flow representations, we have to do the same | |
3619 | * here. This must be in sync with 'match' in handle_packet_upcall(). */ | |
3620 | if (!match.wc.masks.vlans[0].tci) { | |
3621 | match.wc.masks.vlans[0].tci = htons(0xffff); | |
3622 | } | |
3623 | ||
f5d317a1 | 3624 | /* Must produce a netdev_flow_key for lookup. |
beb75a40 JS |
3625 | * Use the same method as employed to create the key when adding |
3626 | * the flow to the dplcs to make sure they match. */ | |
3627 | netdev_flow_mask_init(&mask, &match); | |
3628 | netdev_flow_key_init_masked(&key, &match.flow, &mask); | |
f5d317a1 DDP |
3629 | |
3630 | if (put->pmd_id == PMD_ID_NULL) { | |
3631 | if (cmap_count(&dp->poll_threads) == 0) { | |
3632 | return EINVAL; | |
3633 | } | |
3634 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
3635 | struct dpif_flow_stats pmd_stats; | |
3636 | int pmd_error; | |
3637 | ||
3638 | pmd_error = flow_put_on_pmd(pmd, &key, &match, &ufid, put, | |
3639 | &pmd_stats); | |
3640 | if (pmd_error) { | |
3641 | error = pmd_error; | |
3642 | } else if (put->stats) { | |
3643 | put->stats->n_packets += pmd_stats.n_packets; | |
3644 | put->stats->n_bytes += pmd_stats.n_bytes; | |
3645 | put->stats->used = MAX(put->stats->used, pmd_stats.used); | |
3646 | put->stats->tcp_flags |= pmd_stats.tcp_flags; | |
3647 | } | |
3648 | } | |
3649 | } else { | |
3650 | pmd = dp_netdev_get_pmd(dp, put->pmd_id); | |
3651 | if (!pmd) { | |
3652 | return EINVAL; | |
3653 | } | |
3654 | error = flow_put_on_pmd(pmd, &key, &match, &ufid, put, put->stats); | |
3655 | dp_netdev_pmd_unref(pmd); | |
3656 | } | |
3657 | ||
3658 | return error; | |
3659 | } | |
3660 | ||
3661 | static int | |
3662 | flow_del_on_pmd(struct dp_netdev_pmd_thread *pmd, | |
3663 | struct dpif_flow_stats *stats, | |
3664 | const struct dpif_flow_del *del) | |
3665 | { | |
3666 | struct dp_netdev_flow *netdev_flow; | |
3667 | int error = 0; | |
3668 | ||
1c1e46ed AW |
3669 | ovs_mutex_lock(&pmd->flow_mutex); |
3670 | netdev_flow = dp_netdev_pmd_find_flow(pmd, del->ufid, del->key, | |
3671 | del->key_len); | |
1763b4b8 | 3672 | if (netdev_flow) { |
f5d317a1 | 3673 | if (stats) { |
a309e4f5 | 3674 | get_dpif_flow_status(pmd->dp, netdev_flow, stats, NULL); |
feebdea2 | 3675 | } |
1c1e46ed | 3676 | dp_netdev_pmd_remove_flow(pmd, netdev_flow); |
72865317 | 3677 | } else { |
5279f8fd | 3678 | error = ENOENT; |
72865317 | 3679 | } |
1c1e46ed | 3680 | ovs_mutex_unlock(&pmd->flow_mutex); |
f5d317a1 DDP |
3681 | |
3682 | return error; | |
3683 | } | |
3684 | ||
3685 | static int | |
3686 | dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del) | |
3687 | { | |
3688 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
3689 | struct dp_netdev_pmd_thread *pmd; | |
3690 | int error = 0; | |
3691 | ||
3692 | if (del->stats) { | |
3693 | memset(del->stats, 0, sizeof *del->stats); | |
3694 | } | |
3695 | ||
3696 | if (del->pmd_id == PMD_ID_NULL) { | |
3697 | if (cmap_count(&dp->poll_threads) == 0) { | |
3698 | return EINVAL; | |
3699 | } | |
3700 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
3701 | struct dpif_flow_stats pmd_stats; | |
3702 | int pmd_error; | |
3703 | ||
3704 | pmd_error = flow_del_on_pmd(pmd, &pmd_stats, del); | |
3705 | if (pmd_error) { | |
3706 | error = pmd_error; | |
3707 | } else if (del->stats) { | |
3708 | del->stats->n_packets += pmd_stats.n_packets; | |
3709 | del->stats->n_bytes += pmd_stats.n_bytes; | |
3710 | del->stats->used = MAX(del->stats->used, pmd_stats.used); | |
3711 | del->stats->tcp_flags |= pmd_stats.tcp_flags; | |
3712 | } | |
3713 | } | |
3714 | } else { | |
3715 | pmd = dp_netdev_get_pmd(dp, del->pmd_id); | |
3716 | if (!pmd) { | |
3717 | return EINVAL; | |
3718 | } | |
3719 | error = flow_del_on_pmd(pmd, del->stats, del); | |
3720 | dp_netdev_pmd_unref(pmd); | |
3721 | } | |
3722 | ||
5279f8fd BP |
3723 | |
3724 | return error; | |
72865317 BP |
3725 | } |
3726 | ||
ac64794a BP |
3727 | struct dpif_netdev_flow_dump { |
3728 | struct dpif_flow_dump up; | |
1c1e46ed AW |
3729 | struct cmap_position poll_thread_pos; |
3730 | struct cmap_position flow_pos; | |
3731 | struct dp_netdev_pmd_thread *cur_pmd; | |
d2ad7ef1 JS |
3732 | int status; |
3733 | struct ovs_mutex mutex; | |
e723fd32 JS |
3734 | }; |
3735 | ||
ac64794a BP |
3736 | static struct dpif_netdev_flow_dump * |
3737 | dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump) | |
72865317 | 3738 | { |
ac64794a | 3739 | return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up); |
e723fd32 JS |
3740 | } |
3741 | ||
ac64794a | 3742 | static struct dpif_flow_dump * |
7e8b7199 | 3743 | dpif_netdev_flow_dump_create(const struct dpif *dpif_, bool terse, |
a692410a | 3744 | struct dpif_flow_dump_types *types OVS_UNUSED) |
e723fd32 | 3745 | { |
ac64794a | 3746 | struct dpif_netdev_flow_dump *dump; |
e723fd32 | 3747 | |
1c1e46ed | 3748 | dump = xzalloc(sizeof *dump); |
ac64794a | 3749 | dpif_flow_dump_init(&dump->up, dpif_); |
64bb477f | 3750 | dump->up.terse = terse; |
ac64794a BP |
3751 | ovs_mutex_init(&dump->mutex); |
3752 | ||
3753 | return &dump->up; | |
e723fd32 JS |
3754 | } |
3755 | ||
3756 | static int | |
ac64794a | 3757 | dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_) |
e723fd32 | 3758 | { |
ac64794a | 3759 | struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_); |
e723fd32 | 3760 | |
ac64794a BP |
3761 | ovs_mutex_destroy(&dump->mutex); |
3762 | free(dump); | |
704a1e09 BP |
3763 | return 0; |
3764 | } | |
3765 | ||
ac64794a BP |
3766 | struct dpif_netdev_flow_dump_thread { |
3767 | struct dpif_flow_dump_thread up; | |
3768 | struct dpif_netdev_flow_dump *dump; | |
8bb113da RW |
3769 | struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH]; |
3770 | struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH]; | |
ac64794a BP |
3771 | }; |
3772 | ||
3773 | static struct dpif_netdev_flow_dump_thread * | |
3774 | dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread) | |
3775 | { | |
3776 | return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up); | |
3777 | } | |
3778 | ||
3779 | static struct dpif_flow_dump_thread * | |
3780 | dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_) | |
3781 | { | |
3782 | struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_); | |
3783 | struct dpif_netdev_flow_dump_thread *thread; | |
3784 | ||
3785 | thread = xmalloc(sizeof *thread); | |
3786 | dpif_flow_dump_thread_init(&thread->up, &dump->up); | |
3787 | thread->dump = dump; | |
3788 | return &thread->up; | |
3789 | } | |
3790 | ||
3791 | static void | |
3792 | dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_) | |
3793 | { | |
3794 | struct dpif_netdev_flow_dump_thread *thread | |
3795 | = dpif_netdev_flow_dump_thread_cast(thread_); | |
3796 | ||
3797 | free(thread); | |
3798 | } | |
3799 | ||
704a1e09 | 3800 | static int |
ac64794a | 3801 | dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_, |
8bb113da | 3802 | struct dpif_flow *flows, int max_flows) |
ac64794a BP |
3803 | { |
3804 | struct dpif_netdev_flow_dump_thread *thread | |
3805 | = dpif_netdev_flow_dump_thread_cast(thread_); | |
3806 | struct dpif_netdev_flow_dump *dump = thread->dump; | |
8bb113da | 3807 | struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH]; |
a309e4f5 OM |
3808 | struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif); |
3809 | struct dp_netdev *dp = get_dp_netdev(&dpif->dpif); | |
8bb113da RW |
3810 | int n_flows = 0; |
3811 | int i; | |
14608a15 | 3812 | |
ac64794a | 3813 | ovs_mutex_lock(&dump->mutex); |
8bb113da | 3814 | if (!dump->status) { |
1c1e46ed AW |
3815 | struct dp_netdev_pmd_thread *pmd = dump->cur_pmd; |
3816 | int flow_limit = MIN(max_flows, FLOW_DUMP_MAX_BATCH); | |
3817 | ||
3818 | /* First call to dump_next(), extracts the first pmd thread. | |
3819 | * If there is no pmd thread, returns immediately. */ | |
3820 | if (!pmd) { | |
3821 | pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos); | |
3822 | if (!pmd) { | |
3823 | ovs_mutex_unlock(&dump->mutex); | |
3824 | return n_flows; | |
8bb113da | 3825 | |
8bb113da | 3826 | } |
d2ad7ef1 | 3827 | } |
1c1e46ed AW |
3828 | |
3829 | do { | |
3830 | for (n_flows = 0; n_flows < flow_limit; n_flows++) { | |
3831 | struct cmap_node *node; | |
3832 | ||
3833 | node = cmap_next_position(&pmd->flow_table, &dump->flow_pos); | |
3834 | if (!node) { | |
3835 | break; | |
3836 | } | |
3837 | netdev_flows[n_flows] = CONTAINER_OF(node, | |
3838 | struct dp_netdev_flow, | |
3839 | node); | |
3840 | } | |
3841 | /* When finishing dumping the current pmd thread, moves to | |
3842 | * the next. */ | |
3843 | if (n_flows < flow_limit) { | |
3844 | memset(&dump->flow_pos, 0, sizeof dump->flow_pos); | |
3845 | dp_netdev_pmd_unref(pmd); | |
3846 | pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos); | |
3847 | if (!pmd) { | |
3848 | dump->status = EOF; | |
3849 | break; | |
3850 | } | |
3851 | } | |
3852 | /* Keeps the reference to next caller. */ | |
3853 | dump->cur_pmd = pmd; | |
3854 | ||
3855 | /* If the current dump is empty, do not exit the loop, since the | |
3856 | * remaining pmds could have flows to be dumped. Just dumps again | |
3857 | * on the new 'pmd'. */ | |
3858 | } while (!n_flows); | |
8a4e3a85 | 3859 | } |
ac64794a | 3860 | ovs_mutex_unlock(&dump->mutex); |
ac64794a | 3861 | |
8bb113da RW |
3862 | for (i = 0; i < n_flows; i++) { |
3863 | struct odputil_keybuf *maskbuf = &thread->maskbuf[i]; | |
3864 | struct odputil_keybuf *keybuf = &thread->keybuf[i]; | |
3865 | struct dp_netdev_flow *netdev_flow = netdev_flows[i]; | |
3866 | struct dpif_flow *f = &flows[i]; | |
7af12bd7 | 3867 | struct ofpbuf key, mask; |
8bb113da | 3868 | |
7af12bd7 JS |
3869 | ofpbuf_use_stack(&key, keybuf, sizeof *keybuf); |
3870 | ofpbuf_use_stack(&mask, maskbuf, sizeof *maskbuf); | |
a309e4f5 | 3871 | dp_netdev_flow_to_dpif_flow(dp, netdev_flow, &key, &mask, f, |
64bb477f | 3872 | dump->up.terse); |
8bb113da | 3873 | } |
feebdea2 | 3874 | |
8bb113da | 3875 | return n_flows; |
72865317 BP |
3876 | } |
3877 | ||
3878 | static int | |
758c456d | 3879 | dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute) |
65f13b50 | 3880 | OVS_NO_THREAD_SAFETY_ANALYSIS |
72865317 BP |
3881 | { |
3882 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
65f13b50 | 3883 | struct dp_netdev_pmd_thread *pmd; |
1895cc8d | 3884 | struct dp_packet_batch pp; |
72865317 | 3885 | |
cf62fa4c PS |
3886 | if (dp_packet_size(execute->packet) < ETH_HEADER_LEN || |
3887 | dp_packet_size(execute->packet) > UINT16_MAX) { | |
72865317 BP |
3888 | return EINVAL; |
3889 | } | |
3890 | ||
65f13b50 AW |
3891 | /* Tries finding the 'pmd'. If NULL is returned, that means |
3892 | * the current thread is a non-pmd thread and should use | |
b19befae | 3893 | * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */ |
65f13b50 AW |
3894 | pmd = ovsthread_getspecific(dp->per_pmd_key); |
3895 | if (!pmd) { | |
b19befae | 3896 | pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID); |
546e57d4 DDP |
3897 | if (!pmd) { |
3898 | return EBUSY; | |
3899 | } | |
65f13b50 AW |
3900 | } |
3901 | ||
05267613 AZ |
3902 | if (execute->probe) { |
3903 | /* If this is part of a probe, Drop the packet, since executing | |
3904 | * the action may actually cause spurious packets be sent into | |
3905 | * the network. */ | |
d1ce9c20 YS |
3906 | if (pmd->core_id == NON_PMD_CORE_ID) { |
3907 | dp_netdev_pmd_unref(pmd); | |
3908 | } | |
05267613 AZ |
3909 | return 0; |
3910 | } | |
3911 | ||
65f13b50 AW |
3912 | /* If the current thread is non-pmd thread, acquires |
3913 | * the 'non_pmd_mutex'. */ | |
3914 | if (pmd->core_id == NON_PMD_CORE_ID) { | |
3915 | ovs_mutex_lock(&dp->non_pmd_mutex); | |
3916 | } | |
1c1e46ed | 3917 | |
2fbadeb6 IM |
3918 | /* Update current time in PMD context. We don't care about EMC insertion |
3919 | * probability, because we are on a slow path. */ | |
b010be17 IM |
3920 | pmd_thread_ctx_time_update(pmd); |
3921 | ||
36d8de17 DDP |
3922 | /* The action processing expects the RSS hash to be valid, because |
3923 | * it's always initialized at the beginning of datapath processing. | |
3924 | * In this case, though, 'execute->packet' may not have gone through | |
3925 | * the datapath at all, it may have been generated by the upper layer | |
3926 | * (OpenFlow packet-out, BFD frame, ...). */ | |
3927 | if (!dp_packet_rss_valid(execute->packet)) { | |
3928 | dp_packet_set_rss_hash(execute->packet, | |
3929 | flow_hash_5tuple(execute->flow, 0)); | |
3930 | } | |
3931 | ||
72c84bc2 | 3932 | dp_packet_batch_init_packet(&pp, execute->packet); |
9f17f104 | 3933 | pp.do_not_steal = true; |
66e4ad8a | 3934 | dp_netdev_execute_actions(pmd, &pp, false, execute->flow, |
b010be17 | 3935 | execute->actions, execute->actions_len); |
c71ea3c4 | 3936 | dp_netdev_pmd_flush_output_packets(pmd, true); |
36d8de17 | 3937 | |
65f13b50 AW |
3938 | if (pmd->core_id == NON_PMD_CORE_ID) { |
3939 | ovs_mutex_unlock(&dp->non_pmd_mutex); | |
e9985d6a | 3940 | dp_netdev_pmd_unref(pmd); |
65f13b50 | 3941 | } |
8a4e3a85 | 3942 | |
758c456d | 3943 | return 0; |
72865317 BP |
3944 | } |
3945 | ||
1a0c894a | 3946 | static void |
57924fc9 SB |
3947 | dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops, |
3948 | enum dpif_offload_type offload_type OVS_UNUSED) | |
1a0c894a BP |
3949 | { |
3950 | size_t i; | |
3951 | ||
3952 | for (i = 0; i < n_ops; i++) { | |
3953 | struct dpif_op *op = ops[i]; | |
3954 | ||
3955 | switch (op->type) { | |
3956 | case DPIF_OP_FLOW_PUT: | |
fa37affa | 3957 | op->error = dpif_netdev_flow_put(dpif, &op->flow_put); |
1a0c894a BP |
3958 | break; |
3959 | ||
3960 | case DPIF_OP_FLOW_DEL: | |
fa37affa | 3961 | op->error = dpif_netdev_flow_del(dpif, &op->flow_del); |
1a0c894a BP |
3962 | break; |
3963 | ||
3964 | case DPIF_OP_EXECUTE: | |
fa37affa | 3965 | op->error = dpif_netdev_execute(dpif, &op->execute); |
1a0c894a | 3966 | break; |
6fe09f8c JS |
3967 | |
3968 | case DPIF_OP_FLOW_GET: | |
fa37affa | 3969 | op->error = dpif_netdev_flow_get(dpif, &op->flow_get); |
6fe09f8c | 3970 | break; |
1a0c894a BP |
3971 | } |
3972 | } | |
3973 | } | |
3974 | ||
5bf84282 NK |
3975 | /* Enable or Disable PMD auto load balancing. */ |
3976 | static void | |
3977 | set_pmd_auto_lb(struct dp_netdev *dp) | |
3978 | { | |
3979 | unsigned int cnt = 0; | |
3980 | struct dp_netdev_pmd_thread *pmd; | |
3981 | struct pmd_auto_lb *pmd_alb = &dp->pmd_alb; | |
3982 | ||
3983 | bool enable_alb = false; | |
3984 | bool multi_rxq = false; | |
3985 | bool pmd_rxq_assign_cyc = dp->pmd_rxq_assign_cyc; | |
3986 | ||
3987 | /* Ensure that there is at least 2 non-isolated PMDs and | |
3988 | * one of them is polling more than one rxq. */ | |
3989 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
3990 | if (pmd->core_id == NON_PMD_CORE_ID || pmd->isolated) { | |
3991 | continue; | |
3992 | } | |
3993 | ||
3994 | if (hmap_count(&pmd->poll_list) > 1) { | |
3995 | multi_rxq = true; | |
3996 | } | |
3997 | if (cnt && multi_rxq) { | |
3998 | enable_alb = true; | |
3999 | break; | |
4000 | } | |
4001 | cnt++; | |
4002 | } | |
4003 | ||
4004 | /* Enable auto LB if it is requested and cycle based assignment is true. */ | |
4005 | enable_alb = enable_alb && pmd_rxq_assign_cyc && | |
4006 | pmd_alb->auto_lb_requested; | |
4007 | ||
4008 | if (pmd_alb->is_enabled != enable_alb) { | |
4009 | pmd_alb->is_enabled = enable_alb; | |
4010 | if (pmd_alb->is_enabled) { | |
4011 | VLOG_INFO("PMD auto load balance is enabled " | |
4012 | "(with rebalance interval:%"PRIu64" msec)", | |
4013 | pmd_alb->rebalance_intvl); | |
4014 | } else { | |
4015 | pmd_alb->rebalance_poll_timer = 0; | |
4016 | VLOG_INFO("PMD auto load balance is disabled"); | |
4017 | } | |
4018 | } | |
4019 | ||
4020 | } | |
4021 | ||
d4f6865c DDP |
4022 | /* Applies datapath configuration from the database. Some of the changes are |
4023 | * actually applied in dpif_netdev_run(). */ | |
f2eee189 | 4024 | static int |
d4f6865c | 4025 | dpif_netdev_set_config(struct dpif *dpif, const struct smap *other_config) |
f2eee189 AW |
4026 | { |
4027 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
d4f6865c | 4028 | const char *cmask = smap_get(other_config, "pmd-cpu-mask"); |
e77c97b9 KT |
4029 | const char *pmd_rxq_assign = smap_get_def(other_config, "pmd-rxq-assign", |
4030 | "cycles"); | |
4c30b246 CL |
4031 | unsigned long long insert_prob = |
4032 | smap_get_ullong(other_config, "emc-insert-inv-prob", | |
4033 | DEFAULT_EM_FLOW_INSERT_INV_PROB); | |
4034 | uint32_t insert_min, cur_min; | |
c71ea3c4 | 4035 | uint32_t tx_flush_interval, cur_tx_flush_interval; |
5bf84282 | 4036 | uint64_t rebalance_intvl; |
c71ea3c4 IM |
4037 | |
4038 | tx_flush_interval = smap_get_int(other_config, "tx-flush-interval", | |
4039 | DEFAULT_TX_FLUSH_INTERVAL); | |
4040 | atomic_read_relaxed(&dp->tx_flush_interval, &cur_tx_flush_interval); | |
4041 | if (tx_flush_interval != cur_tx_flush_interval) { | |
4042 | atomic_store_relaxed(&dp->tx_flush_interval, tx_flush_interval); | |
4043 | VLOG_INFO("Flushing interval for tx queues set to %"PRIu32" us", | |
4044 | tx_flush_interval); | |
4045 | } | |
f2eee189 | 4046 | |
a6a426d6 IM |
4047 | if (!nullable_string_is_equal(dp->pmd_cmask, cmask)) { |
4048 | free(dp->pmd_cmask); | |
4049 | dp->pmd_cmask = nullable_xstrdup(cmask); | |
4050 | dp_netdev_request_reconfigure(dp); | |
f2eee189 AW |
4051 | } |
4052 | ||
4c30b246 CL |
4053 | atomic_read_relaxed(&dp->emc_insert_min, &cur_min); |
4054 | if (insert_prob <= UINT32_MAX) { | |
4055 | insert_min = insert_prob == 0 ? 0 : UINT32_MAX / insert_prob; | |
4056 | } else { | |
4057 | insert_min = DEFAULT_EM_FLOW_INSERT_MIN; | |
4058 | insert_prob = DEFAULT_EM_FLOW_INSERT_INV_PROB; | |
4059 | } | |
4060 | ||
4061 | if (insert_min != cur_min) { | |
4062 | atomic_store_relaxed(&dp->emc_insert_min, insert_min); | |
4063 | if (insert_min == 0) { | |
2fbadeb6 | 4064 | VLOG_INFO("EMC insertion probability changed to zero"); |
4c30b246 CL |
4065 | } else { |
4066 | VLOG_INFO("EMC insertion probability changed to 1/%llu (~%.2f%%)", | |
4067 | insert_prob, (100 / (float)insert_prob)); | |
4068 | } | |
4069 | } | |
4070 | ||
79f36875 JS |
4071 | bool perf_enabled = smap_get_bool(other_config, "pmd-perf-metrics", false); |
4072 | bool cur_perf_enabled; | |
4073 | atomic_read_relaxed(&dp->pmd_perf_metrics, &cur_perf_enabled); | |
4074 | if (perf_enabled != cur_perf_enabled) { | |
4075 | atomic_store_relaxed(&dp->pmd_perf_metrics, perf_enabled); | |
4076 | if (perf_enabled) { | |
4077 | VLOG_INFO("PMD performance metrics collection enabled"); | |
4078 | } else { | |
4079 | VLOG_INFO("PMD performance metrics collection disabled"); | |
4080 | } | |
4081 | } | |
4082 | ||
60d8ccae YW |
4083 | bool smc_enable = smap_get_bool(other_config, "smc-enable", false); |
4084 | bool cur_smc; | |
4085 | atomic_read_relaxed(&dp->smc_enable_db, &cur_smc); | |
4086 | if (smc_enable != cur_smc) { | |
4087 | atomic_store_relaxed(&dp->smc_enable_db, smc_enable); | |
4088 | if (smc_enable) { | |
4089 | VLOG_INFO("SMC cache is enabled"); | |
4090 | } else { | |
4091 | VLOG_INFO("SMC cache is disabled"); | |
4092 | } | |
4093 | } | |
e77c97b9 KT |
4094 | |
4095 | bool pmd_rxq_assign_cyc = !strcmp(pmd_rxq_assign, "cycles"); | |
4096 | if (!pmd_rxq_assign_cyc && strcmp(pmd_rxq_assign, "roundrobin")) { | |
4097 | VLOG_WARN("Unsupported Rxq to PMD assignment mode in pmd-rxq-assign. " | |
4098 | "Defaulting to 'cycles'."); | |
4099 | pmd_rxq_assign_cyc = true; | |
4100 | pmd_rxq_assign = "cycles"; | |
4101 | } | |
4102 | if (dp->pmd_rxq_assign_cyc != pmd_rxq_assign_cyc) { | |
4103 | dp->pmd_rxq_assign_cyc = pmd_rxq_assign_cyc; | |
4104 | VLOG_INFO("Rxq to PMD assignment mode changed to: \'%s\'.", | |
4105 | pmd_rxq_assign); | |
4106 | dp_netdev_request_reconfigure(dp); | |
4107 | } | |
5bf84282 NK |
4108 | |
4109 | struct pmd_auto_lb *pmd_alb = &dp->pmd_alb; | |
4110 | pmd_alb->auto_lb_requested = smap_get_bool(other_config, "pmd-auto-lb", | |
4111 | false); | |
4112 | ||
4113 | rebalance_intvl = smap_get_int(other_config, "pmd-auto-lb-rebal-interval", | |
4114 | ALB_PMD_REBALANCE_POLL_INTERVAL); | |
4115 | ||
4116 | /* Input is in min, convert it to msec. */ | |
4117 | rebalance_intvl = | |
4118 | rebalance_intvl ? rebalance_intvl * MIN_TO_MSEC : MIN_TO_MSEC; | |
4119 | ||
4120 | if (pmd_alb->rebalance_intvl != rebalance_intvl) { | |
4121 | pmd_alb->rebalance_intvl = rebalance_intvl; | |
4122 | } | |
4123 | ||
4124 | set_pmd_auto_lb(dp); | |
f2eee189 AW |
4125 | return 0; |
4126 | } | |
4127 | ||
3eb67853 IM |
4128 | /* Parses affinity list and returns result in 'core_ids'. */ |
4129 | static int | |
4130 | parse_affinity_list(const char *affinity_list, unsigned *core_ids, int n_rxq) | |
4131 | { | |
4132 | unsigned i; | |
4133 | char *list, *copy, *key, *value; | |
4134 | int error = 0; | |
4135 | ||
4136 | for (i = 0; i < n_rxq; i++) { | |
51c37a56 | 4137 | core_ids[i] = OVS_CORE_UNSPEC; |
3eb67853 IM |
4138 | } |
4139 | ||
4140 | if (!affinity_list) { | |
4141 | return 0; | |
4142 | } | |
4143 | ||
4144 | list = copy = xstrdup(affinity_list); | |
4145 | ||
4146 | while (ofputil_parse_key_value(&list, &key, &value)) { | |
4147 | int rxq_id, core_id; | |
4148 | ||
4149 | if (!str_to_int(key, 0, &rxq_id) || rxq_id < 0 | |
4150 | || !str_to_int(value, 0, &core_id) || core_id < 0) { | |
4151 | error = EINVAL; | |
4152 | break; | |
4153 | } | |
4154 | ||
4155 | if (rxq_id < n_rxq) { | |
4156 | core_ids[rxq_id] = core_id; | |
4157 | } | |
4158 | } | |
4159 | ||
4160 | free(copy); | |
4161 | return error; | |
4162 | } | |
4163 | ||
4164 | /* Parses 'affinity_list' and applies configuration if it is valid. */ | |
4165 | static int | |
4166 | dpif_netdev_port_set_rxq_affinity(struct dp_netdev_port *port, | |
4167 | const char *affinity_list) | |
4168 | { | |
4169 | unsigned *core_ids, i; | |
4170 | int error = 0; | |
4171 | ||
4172 | core_ids = xmalloc(port->n_rxq * sizeof *core_ids); | |
4173 | if (parse_affinity_list(affinity_list, core_ids, port->n_rxq)) { | |
4174 | error = EINVAL; | |
4175 | goto exit; | |
4176 | } | |
4177 | ||
4178 | for (i = 0; i < port->n_rxq; i++) { | |
4179 | port->rxqs[i].core_id = core_ids[i]; | |
4180 | } | |
4181 | ||
4182 | exit: | |
4183 | free(core_ids); | |
4184 | return error; | |
4185 | } | |
4186 | ||
2fbadeb6 IM |
4187 | /* Returns 'true' if one of the 'port's RX queues exists in 'poll_list' |
4188 | * of given PMD thread. */ | |
4189 | static bool | |
4190 | dpif_netdev_pmd_polls_port(struct dp_netdev_pmd_thread *pmd, | |
4191 | struct dp_netdev_port *port) | |
4192 | OVS_EXCLUDED(pmd->port_mutex) | |
4193 | { | |
4194 | struct rxq_poll *poll; | |
4195 | bool found = false; | |
4196 | ||
4197 | ovs_mutex_lock(&pmd->port_mutex); | |
4198 | HMAP_FOR_EACH (poll, node, &pmd->poll_list) { | |
4199 | if (port == poll->rxq->port) { | |
4200 | found = true; | |
4201 | break; | |
4202 | } | |
4203 | } | |
4204 | ovs_mutex_unlock(&pmd->port_mutex); | |
4205 | return found; | |
4206 | } | |
4207 | ||
4208 | /* Updates port configuration from the database. The changes are actually | |
4209 | * applied in dpif_netdev_run(). */ | |
3eb67853 IM |
4210 | static int |
4211 | dpif_netdev_port_set_config(struct dpif *dpif, odp_port_t port_no, | |
4212 | const struct smap *cfg) | |
4213 | { | |
4214 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
4215 | struct dp_netdev_port *port; | |
4216 | int error = 0; | |
4217 | const char *affinity_list = smap_get(cfg, "pmd-rxq-affinity"); | |
2fbadeb6 | 4218 | bool emc_enabled = smap_get_bool(cfg, "emc-enable", true); |
3eb67853 IM |
4219 | |
4220 | ovs_mutex_lock(&dp->port_mutex); | |
4221 | error = get_port_by_number(dp, port_no, &port); | |
2fbadeb6 IM |
4222 | if (error) { |
4223 | goto unlock; | |
4224 | } | |
4225 | ||
4226 | if (emc_enabled != port->emc_enabled) { | |
4227 | struct dp_netdev_pmd_thread *pmd; | |
4228 | struct ds ds = DS_EMPTY_INITIALIZER; | |
4229 | uint32_t cur_min, insert_prob; | |
4230 | ||
4231 | port->emc_enabled = emc_enabled; | |
4232 | /* Mark for reload all the threads that polls this port and request | |
4233 | * for reconfiguration for the actual reloading of threads. */ | |
4234 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
4235 | if (dpif_netdev_pmd_polls_port(pmd, port)) { | |
4236 | pmd->need_reload = true; | |
4237 | } | |
4238 | } | |
4239 | dp_netdev_request_reconfigure(dp); | |
4240 | ||
4241 | ds_put_format(&ds, "%s: EMC has been %s.", | |
4242 | netdev_get_name(port->netdev), | |
4243 | (emc_enabled) ? "enabled" : "disabled"); | |
4244 | if (emc_enabled) { | |
4245 | ds_put_cstr(&ds, " Current insertion probability is "); | |
4246 | atomic_read_relaxed(&dp->emc_insert_min, &cur_min); | |
4247 | if (!cur_min) { | |
4248 | ds_put_cstr(&ds, "zero."); | |
4249 | } else { | |
4250 | insert_prob = UINT32_MAX / cur_min; | |
4251 | ds_put_format(&ds, "1/%"PRIu32" (~%.2f%%).", | |
4252 | insert_prob, 100 / (float) insert_prob); | |
4253 | } | |
4254 | } | |
4255 | VLOG_INFO("%s", ds_cstr(&ds)); | |
4256 | ds_destroy(&ds); | |
4257 | } | |
4258 | ||
4259 | /* Checking for RXq affinity changes. */ | |
4260 | if (!netdev_is_pmd(port->netdev) | |
3eb67853 IM |
4261 | || nullable_string_is_equal(affinity_list, port->rxq_affinity_list)) { |
4262 | goto unlock; | |
4263 | } | |
4264 | ||
4265 | error = dpif_netdev_port_set_rxq_affinity(port, affinity_list); | |
4266 | if (error) { | |
4267 | goto unlock; | |
4268 | } | |
4269 | free(port->rxq_affinity_list); | |
4270 | port->rxq_affinity_list = nullable_xstrdup(affinity_list); | |
4271 | ||
4272 | dp_netdev_request_reconfigure(dp); | |
4273 | unlock: | |
4274 | ovs_mutex_unlock(&dp->port_mutex); | |
4275 | return error; | |
4276 | } | |
4277 | ||
5bf93d67 EJ |
4278 | static int |
4279 | dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED, | |
4280 | uint32_t queue_id, uint32_t *priority) | |
4281 | { | |
4282 | *priority = queue_id; | |
4283 | return 0; | |
4284 | } | |
4285 | ||
72865317 | 4286 | \f |
9ff55ae2 | 4287 | /* Creates and returns a new 'struct dp_netdev_actions', whose actions are |
1401f6de | 4288 | * a copy of the 'size' bytes of 'actions' input parameters. */ |
a84cb64a BP |
4289 | struct dp_netdev_actions * |
4290 | dp_netdev_actions_create(const struct nlattr *actions, size_t size) | |
4291 | { | |
4292 | struct dp_netdev_actions *netdev_actions; | |
4293 | ||
9ff55ae2 DDP |
4294 | netdev_actions = xmalloc(sizeof *netdev_actions + size); |
4295 | memcpy(netdev_actions->actions, actions, size); | |
a84cb64a BP |
4296 | netdev_actions->size = size; |
4297 | ||
4298 | return netdev_actions; | |
4299 | } | |
4300 | ||
a84cb64a | 4301 | struct dp_netdev_actions * |
61e7deb1 | 4302 | dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow) |
a84cb64a | 4303 | { |
61e7deb1 | 4304 | return ovsrcu_get(struct dp_netdev_actions *, &flow->actions); |
a84cb64a BP |
4305 | } |
4306 | ||
61e7deb1 BP |
4307 | static void |
4308 | dp_netdev_actions_free(struct dp_netdev_actions *actions) | |
a84cb64a | 4309 | { |
61e7deb1 | 4310 | free(actions); |
a84cb64a BP |
4311 | } |
4312 | \f | |
a19896ab JS |
4313 | static void |
4314 | dp_netdev_rxq_set_cycles(struct dp_netdev_rxq *rx, | |
4315 | enum rxq_cycles_counter_type type, | |
4316 | unsigned long long cycles) | |
a2ac666d | 4317 | { |
a19896ab | 4318 | atomic_store_relaxed(&rx->cycles[type], cycles); |
a2ac666d CL |
4319 | } |
4320 | ||
4809891b | 4321 | static void |
a19896ab | 4322 | dp_netdev_rxq_add_cycles(struct dp_netdev_rxq *rx, |
4809891b KT |
4323 | enum rxq_cycles_counter_type type, |
4324 | unsigned long long cycles) | |
4325 | { | |
a19896ab | 4326 | non_atomic_ullong_add(&rx->cycles[type], cycles); |
4809891b KT |
4327 | } |
4328 | ||
4329 | static uint64_t | |
4330 | dp_netdev_rxq_get_cycles(struct dp_netdev_rxq *rx, | |
4331 | enum rxq_cycles_counter_type type) | |
4332 | { | |
4333 | unsigned long long processing_cycles; | |
4334 | atomic_read_relaxed(&rx->cycles[type], &processing_cycles); | |
4335 | return processing_cycles; | |
4336 | } | |
4337 | ||
4338 | static void | |
4339 | dp_netdev_rxq_set_intrvl_cycles(struct dp_netdev_rxq *rx, | |
4340 | unsigned long long cycles) | |
4341 | { | |
4ee87ad3 BP |
4342 | unsigned int idx = rx->intrvl_idx++ % PMD_RXQ_INTERVAL_MAX; |
4343 | atomic_store_relaxed(&rx->cycles_intrvl[idx], cycles); | |
4809891b KT |
4344 | } |
4345 | ||
655856ef KT |
4346 | static uint64_t |
4347 | dp_netdev_rxq_get_intrvl_cycles(struct dp_netdev_rxq *rx, unsigned idx) | |
4348 | { | |
4349 | unsigned long long processing_cycles; | |
4350 | atomic_read_relaxed(&rx->cycles_intrvl[idx], &processing_cycles); | |
4351 | return processing_cycles; | |
4352 | } | |
4353 | ||
79f36875 JS |
4354 | #if ATOMIC_ALWAYS_LOCK_FREE_8B |
4355 | static inline bool | |
4356 | pmd_perf_metrics_enabled(const struct dp_netdev_pmd_thread *pmd) | |
4357 | { | |
4358 | bool pmd_perf_enabled; | |
4359 | atomic_read_relaxed(&pmd->dp->pmd_perf_metrics, &pmd_perf_enabled); | |
4360 | return pmd_perf_enabled; | |
4361 | } | |
4362 | #else | |
4363 | /* If stores and reads of 64-bit integers are not atomic, the full PMD | |
4364 | * performance metrics are not available as locked access to 64 bit | |
4365 | * integers would be prohibitively expensive. */ | |
4366 | static inline bool | |
4367 | pmd_perf_metrics_enabled(const struct dp_netdev_pmd_thread *pmd OVS_UNUSED) | |
4368 | { | |
4369 | return false; | |
4370 | } | |
4371 | #endif | |
4372 | ||
c71ea3c4 | 4373 | static int |
009e0033 IM |
4374 | dp_netdev_pmd_flush_output_on_port(struct dp_netdev_pmd_thread *pmd, |
4375 | struct tx_port *p) | |
4376 | { | |
58ed6df0 | 4377 | int i; |
009e0033 | 4378 | int tx_qid; |
cc4891f3 | 4379 | int output_cnt; |
009e0033 | 4380 | bool dynamic_txqs; |
58ed6df0 IM |
4381 | struct cycle_timer timer; |
4382 | uint64_t cycles; | |
c71ea3c4 | 4383 | uint32_t tx_flush_interval; |
58ed6df0 IM |
4384 | |
4385 | cycle_timer_start(&pmd->perf_stats, &timer); | |
009e0033 IM |
4386 | |
4387 | dynamic_txqs = p->port->dynamic_txqs; | |
4388 | if (dynamic_txqs) { | |
4389 | tx_qid = dpif_netdev_xps_get_tx_qid(pmd, p); | |
4390 | } else { | |
4391 | tx_qid = pmd->static_tx_qid; | |
4392 | } | |
4393 | ||
cc4891f3 | 4394 | output_cnt = dp_packet_batch_size(&p->output_pkts); |
58ed6df0 | 4395 | ovs_assert(output_cnt > 0); |
cc4891f3 | 4396 | |
b30896c9 | 4397 | netdev_send(p->port->netdev, tx_qid, &p->output_pkts, dynamic_txqs); |
009e0033 | 4398 | dp_packet_batch_init(&p->output_pkts); |
cc4891f3 | 4399 | |
c71ea3c4 IM |
4400 | /* Update time of the next flush. */ |
4401 | atomic_read_relaxed(&pmd->dp->tx_flush_interval, &tx_flush_interval); | |
4402 | p->flush_time = pmd->ctx.now + tx_flush_interval; | |
4403 | ||
4404 | ovs_assert(pmd->n_output_batches > 0); | |
4405 | pmd->n_output_batches--; | |
4406 | ||
82a48ead JS |
4407 | pmd_perf_update_counter(&pmd->perf_stats, PMD_STAT_SENT_PKTS, output_cnt); |
4408 | pmd_perf_update_counter(&pmd->perf_stats, PMD_STAT_SENT_BATCHES, 1); | |
58ed6df0 IM |
4409 | |
4410 | /* Distribute send cycles evenly among transmitted packets and assign to | |
4411 | * their respective rx queues. */ | |
4412 | cycles = cycle_timer_stop(&pmd->perf_stats, &timer) / output_cnt; | |
4413 | for (i = 0; i < output_cnt; i++) { | |
4414 | if (p->output_pkts_rxqs[i]) { | |
4415 | dp_netdev_rxq_add_cycles(p->output_pkts_rxqs[i], | |
4416 | RXQ_CYCLES_PROC_CURR, cycles); | |
4417 | } | |
4418 | } | |
c71ea3c4 IM |
4419 | |
4420 | return output_cnt; | |
009e0033 IM |
4421 | } |
4422 | ||
c71ea3c4 IM |
4423 | static int |
4424 | dp_netdev_pmd_flush_output_packets(struct dp_netdev_pmd_thread *pmd, | |
4425 | bool force) | |
009e0033 IM |
4426 | { |
4427 | struct tx_port *p; | |
c71ea3c4 IM |
4428 | int output_cnt = 0; |
4429 | ||
4430 | if (!pmd->n_output_batches) { | |
4431 | return 0; | |
4432 | } | |
009e0033 IM |
4433 | |
4434 | HMAP_FOR_EACH (p, node, &pmd->send_port_cache) { | |
c71ea3c4 IM |
4435 | if (!dp_packet_batch_is_empty(&p->output_pkts) |
4436 | && (force || pmd->ctx.now >= p->flush_time)) { | |
4437 | output_cnt += dp_netdev_pmd_flush_output_on_port(pmd, p); | |
009e0033 IM |
4438 | } |
4439 | } | |
c71ea3c4 | 4440 | return output_cnt; |
009e0033 IM |
4441 | } |
4442 | ||
a2ac666d | 4443 | static int |
65f13b50 | 4444 | dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd, |
a19896ab | 4445 | struct dp_netdev_rxq *rxq, |
947dc567 | 4446 | odp_port_t port_no) |
e4cfed38 | 4447 | { |
79f36875 | 4448 | struct pmd_perf_stats *s = &pmd->perf_stats; |
1895cc8d | 4449 | struct dp_packet_batch batch; |
a19896ab | 4450 | struct cycle_timer timer; |
1895cc8d | 4451 | int error; |
79f36875 JS |
4452 | int batch_cnt = 0; |
4453 | int rem_qlen = 0, *qlen_p = NULL; | |
58ed6df0 | 4454 | uint64_t cycles; |
e4cfed38 | 4455 | |
a19896ab JS |
4456 | /* Measure duration for polling and processing rx burst. */ |
4457 | cycle_timer_start(&pmd->perf_stats, &timer); | |
58ed6df0 IM |
4458 | |
4459 | pmd->ctx.last_rxq = rxq; | |
1895cc8d | 4460 | dp_packet_batch_init(&batch); |
58ed6df0 | 4461 | |
79f36875 JS |
4462 | /* Fetch the rx queue length only for vhostuser ports. */ |
4463 | if (pmd_perf_metrics_enabled(pmd) && rxq->is_vhost) { | |
4464 | qlen_p = &rem_qlen; | |
4465 | } | |
4466 | ||
4467 | error = netdev_rxq_recv(rxq->rx, &batch, qlen_p); | |
e4cfed38 | 4468 | if (!error) { |
a19896ab | 4469 | /* At least one packet received. */ |
3c33f0ff | 4470 | *recirc_depth_get() = 0; |
009e0033 | 4471 | pmd_thread_ctx_time_update(pmd); |
940ac2ce | 4472 | batch_cnt = dp_packet_batch_size(&batch); |
79f36875 JS |
4473 | if (pmd_perf_metrics_enabled(pmd)) { |
4474 | /* Update batch histogram. */ | |
4475 | s->current.batches++; | |
4476 | histogram_add_sample(&s->pkts_per_batch, batch_cnt); | |
4477 | /* Update the maximum vhost rx queue fill level. */ | |
4478 | if (rxq->is_vhost && rem_qlen >= 0) { | |
4479 | uint32_t qfill = batch_cnt + rem_qlen; | |
4480 | if (qfill > s->current.max_vhost_qfill) { | |
4481 | s->current.max_vhost_qfill = qfill; | |
4482 | } | |
4483 | } | |
4484 | } | |
4485 | /* Process packet batch. */ | |
947dc567 | 4486 | dp_netdev_input(pmd, &batch, port_no); |
e4cfed38 | 4487 | |
a19896ab | 4488 | /* Assign processing cycles to rx queue. */ |
58ed6df0 | 4489 | cycles = cycle_timer_stop(&pmd->perf_stats, &timer); |
a19896ab JS |
4490 | dp_netdev_rxq_add_cycles(rxq, RXQ_CYCLES_PROC_CURR, cycles); |
4491 | ||
79f36875 | 4492 | dp_netdev_pmd_flush_output_packets(pmd, false); |
a19896ab JS |
4493 | } else { |
4494 | /* Discard cycles. */ | |
4495 | cycle_timer_stop(&pmd->perf_stats, &timer); | |
4496 | if (error != EAGAIN && error != EOPNOTSUPP) { | |
4497 | static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5); | |
4498 | ||
4499 | VLOG_ERR_RL(&rl, "error receiving data from %s: %s", | |
4500 | netdev_rxq_get_name(rxq->rx), ovs_strerror(error)); | |
4501 | } | |
e4cfed38 | 4502 | } |
a2ac666d | 4503 | |
58ed6df0 IM |
4504 | pmd->ctx.last_rxq = NULL; |
4505 | ||
79f36875 | 4506 | return batch_cnt; |
e4cfed38 PS |
4507 | } |
4508 | ||
e32971b8 DDP |
4509 | static struct tx_port * |
4510 | tx_port_lookup(const struct hmap *hmap, odp_port_t port_no) | |
4511 | { | |
4512 | struct tx_port *tx; | |
4513 | ||
4514 | HMAP_FOR_EACH_IN_BUCKET (tx, node, hash_port_no(port_no), hmap) { | |
4515 | if (tx->port->port_no == port_no) { | |
4516 | return tx; | |
4517 | } | |
4518 | } | |
4519 | ||
4520 | return NULL; | |
4521 | } | |
4522 | ||
9df65060 VDA |
4523 | static struct tx_bond * |
4524 | tx_bond_lookup(const struct cmap *tx_bonds, uint32_t bond_id) | |
4525 | { | |
4526 | uint32_t hash = hash_bond_id(bond_id); | |
4527 | struct tx_bond *tx; | |
4528 | ||
4529 | CMAP_FOR_EACH_WITH_HASH (tx, node, hash, tx_bonds) { | |
4530 | if (tx->bond_id == bond_id) { | |
4531 | return tx; | |
4532 | } | |
4533 | } | |
4534 | return NULL; | |
4535 | } | |
4536 | ||
dc36593c DDP |
4537 | static int |
4538 | port_reconfigure(struct dp_netdev_port *port) | |
4539 | { | |
4540 | struct netdev *netdev = port->netdev; | |
dc36593c DDP |
4541 | int i, err; |
4542 | ||
dc36593c DDP |
4543 | /* Closes the existing 'rxq's. */ |
4544 | for (i = 0; i < port->n_rxq; i++) { | |
947dc567 DDP |
4545 | netdev_rxq_close(port->rxqs[i].rx); |
4546 | port->rxqs[i].rx = NULL; | |
dc36593c | 4547 | } |
4809891b | 4548 | unsigned last_nrxq = port->n_rxq; |
dc36593c DDP |
4549 | port->n_rxq = 0; |
4550 | ||
050c60bf | 4551 | /* Allows 'netdev' to apply the pending configuration changes. */ |
606f6650 | 4552 | if (netdev_is_reconf_required(netdev) || port->need_reconfigure) { |
e32971b8 DDP |
4553 | err = netdev_reconfigure(netdev); |
4554 | if (err && (err != EOPNOTSUPP)) { | |
4555 | VLOG_ERR("Failed to set interface %s new configuration", | |
4556 | netdev_get_name(netdev)); | |
4557 | return err; | |
4558 | } | |
dc36593c | 4559 | } |
050c60bf | 4560 | /* If the netdev_reconfigure() above succeeds, reopens the 'rxq's. */ |
3eb67853 IM |
4561 | port->rxqs = xrealloc(port->rxqs, |
4562 | sizeof *port->rxqs * netdev_n_rxq(netdev)); | |
324c8374 IM |
4563 | /* Realloc 'used' counters for tx queues. */ |
4564 | free(port->txq_used); | |
4565 | port->txq_used = xcalloc(netdev_n_txq(netdev), sizeof *port->txq_used); | |
4566 | ||
dc36593c | 4567 | for (i = 0; i < netdev_n_rxq(netdev); i++) { |
38259bd7 BP |
4568 | bool new_queue = i >= last_nrxq; |
4569 | if (new_queue) { | |
4570 | memset(&port->rxqs[i], 0, sizeof port->rxqs[i]); | |
4571 | } | |
4572 | ||
947dc567 | 4573 | port->rxqs[i].port = port; |
79f36875 | 4574 | port->rxqs[i].is_vhost = !strncmp(port->type, "dpdkvhost", 9); |
38259bd7 | 4575 | |
947dc567 | 4576 | err = netdev_rxq_open(netdev, &port->rxqs[i].rx, i); |
dc36593c DDP |
4577 | if (err) { |
4578 | return err; | |
4579 | } | |
4580 | port->n_rxq++; | |
4581 | } | |
4582 | ||
3eb67853 IM |
4583 | /* Parse affinity list to apply configuration for new queues. */ |
4584 | dpif_netdev_port_set_rxq_affinity(port, port->rxq_affinity_list); | |
4585 | ||
606f6650 EC |
4586 | /* If reconfiguration was successful mark it as such, so we can use it */ |
4587 | port->need_reconfigure = false; | |
4588 | ||
dc36593c DDP |
4589 | return 0; |
4590 | } | |
4591 | ||
e32971b8 DDP |
4592 | struct rr_numa_list { |
4593 | struct hmap numas; /* Contains 'struct rr_numa' */ | |
4594 | }; | |
4595 | ||
4596 | struct rr_numa { | |
4597 | struct hmap_node node; | |
4598 | ||
4599 | int numa_id; | |
4600 | ||
4601 | /* Non isolated pmds on numa node 'numa_id' */ | |
4602 | struct dp_netdev_pmd_thread **pmds; | |
4603 | int n_pmds; | |
4604 | ||
4605 | int cur_index; | |
79da1e41 | 4606 | bool idx_inc; |
e32971b8 DDP |
4607 | }; |
4608 | ||
4609 | static struct rr_numa * | |
4610 | rr_numa_list_lookup(struct rr_numa_list *rr, int numa_id) | |
4611 | { | |
4612 | struct rr_numa *numa; | |
4613 | ||
4614 | HMAP_FOR_EACH_WITH_HASH (numa, node, hash_int(numa_id, 0), &rr->numas) { | |
4615 | if (numa->numa_id == numa_id) { | |
4616 | return numa; | |
4617 | } | |
4618 | } | |
4619 | ||
4620 | return NULL; | |
4621 | } | |
4622 | ||
c37813fd BM |
4623 | /* Returns the next node in numa list following 'numa' in round-robin fashion. |
4624 | * Returns first node if 'numa' is a null pointer or the last node in 'rr'. | |
4625 | * Returns NULL if 'rr' numa list is empty. */ | |
4626 | static struct rr_numa * | |
4627 | rr_numa_list_next(struct rr_numa_list *rr, const struct rr_numa *numa) | |
4628 | { | |
4629 | struct hmap_node *node = NULL; | |
4630 | ||
4631 | if (numa) { | |
4632 | node = hmap_next(&rr->numas, &numa->node); | |
4633 | } | |
4634 | if (!node) { | |
4635 | node = hmap_first(&rr->numas); | |
4636 | } | |
4637 | ||
4638 | return (node) ? CONTAINER_OF(node, struct rr_numa, node) : NULL; | |
4639 | } | |
4640 | ||
e32971b8 DDP |
4641 | static void |
4642 | rr_numa_list_populate(struct dp_netdev *dp, struct rr_numa_list *rr) | |
4643 | { | |
4644 | struct dp_netdev_pmd_thread *pmd; | |
4645 | struct rr_numa *numa; | |
4646 | ||
4647 | hmap_init(&rr->numas); | |
4648 | ||
4649 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
4650 | if (pmd->core_id == NON_PMD_CORE_ID || pmd->isolated) { | |
4651 | continue; | |
4652 | } | |
4653 | ||
4654 | numa = rr_numa_list_lookup(rr, pmd->numa_id); | |
4655 | if (!numa) { | |
4656 | numa = xzalloc(sizeof *numa); | |
4657 | numa->numa_id = pmd->numa_id; | |
4658 | hmap_insert(&rr->numas, &numa->node, hash_int(pmd->numa_id, 0)); | |
4659 | } | |
4660 | numa->n_pmds++; | |
4661 | numa->pmds = xrealloc(numa->pmds, numa->n_pmds * sizeof *numa->pmds); | |
4662 | numa->pmds[numa->n_pmds - 1] = pmd; | |
79da1e41 KT |
4663 | /* At least one pmd so initialise curr_idx and idx_inc. */ |
4664 | numa->cur_index = 0; | |
4665 | numa->idx_inc = true; | |
e32971b8 DDP |
4666 | } |
4667 | } | |
4668 | ||
e77c97b9 KT |
4669 | /* |
4670 | * Returns the next pmd from the numa node. | |
4671 | * | |
4672 | * If 'updown' is 'true' it will alternate between selecting the next pmd in | |
4673 | * either an up or down walk, switching between up/down when the first or last | |
4674 | * core is reached. e.g. 1,2,3,3,2,1,1,2... | |
4675 | * | |
4676 | * If 'updown' is 'false' it will select the next pmd wrapping around when last | |
4677 | * core reached. e.g. 1,2,3,1,2,3,1,2... | |
4678 | */ | |
e32971b8 | 4679 | static struct dp_netdev_pmd_thread * |
e77c97b9 | 4680 | rr_numa_get_pmd(struct rr_numa *numa, bool updown) |
e32971b8 | 4681 | { |
79da1e41 KT |
4682 | int numa_idx = numa->cur_index; |
4683 | ||
4684 | if (numa->idx_inc == true) { | |
4685 | /* Incrementing through list of pmds. */ | |
4686 | if (numa->cur_index == numa->n_pmds-1) { | |
4687 | /* Reached the last pmd. */ | |
e77c97b9 KT |
4688 | if (updown) { |
4689 | numa->idx_inc = false; | |
4690 | } else { | |
4691 | numa->cur_index = 0; | |
4692 | } | |
79da1e41 KT |
4693 | } else { |
4694 | numa->cur_index++; | |
4695 | } | |
4696 | } else { | |
4697 | /* Decrementing through list of pmds. */ | |
4698 | if (numa->cur_index == 0) { | |
4699 | /* Reached the first pmd. */ | |
4700 | numa->idx_inc = true; | |
4701 | } else { | |
4702 | numa->cur_index--; | |
4703 | } | |
4704 | } | |
4705 | return numa->pmds[numa_idx]; | |
e32971b8 DDP |
4706 | } |
4707 | ||
4708 | static void | |
4709 | rr_numa_list_destroy(struct rr_numa_list *rr) | |
4710 | { | |
4711 | struct rr_numa *numa; | |
4712 | ||
4713 | HMAP_FOR_EACH_POP (numa, node, &rr->numas) { | |
4714 | free(numa->pmds); | |
4715 | free(numa); | |
4716 | } | |
4717 | hmap_destroy(&rr->numas); | |
4718 | } | |
4719 | ||
655856ef KT |
4720 | /* Sort Rx Queues by the processing cycles they are consuming. */ |
4721 | static int | |
cc131ac1 | 4722 | compare_rxq_cycles(const void *a, const void *b) |
655856ef | 4723 | { |
28080276 KT |
4724 | struct dp_netdev_rxq *qa; |
4725 | struct dp_netdev_rxq *qb; | |
8368866e | 4726 | uint64_t cycles_qa, cycles_qb; |
655856ef KT |
4727 | |
4728 | qa = *(struct dp_netdev_rxq **) a; | |
4729 | qb = *(struct dp_netdev_rxq **) b; | |
4730 | ||
8368866e KT |
4731 | cycles_qa = dp_netdev_rxq_get_cycles(qa, RXQ_CYCLES_PROC_HIST); |
4732 | cycles_qb = dp_netdev_rxq_get_cycles(qb, RXQ_CYCLES_PROC_HIST); | |
655856ef | 4733 | |
8368866e KT |
4734 | if (cycles_qa != cycles_qb) { |
4735 | return (cycles_qa < cycles_qb) ? 1 : -1; | |
a130f1a8 KT |
4736 | } else { |
4737 | /* Cycles are the same so tiebreak on port/queue id. | |
4738 | * Tiebreaking (as opposed to return 0) ensures consistent | |
4739 | * sort results across multiple OS's. */ | |
f0aa3801 BP |
4740 | uint32_t port_qa = odp_to_u32(qa->port->port_no); |
4741 | uint32_t port_qb = odp_to_u32(qb->port->port_no); | |
4742 | if (port_qa != port_qb) { | |
4743 | return port_qa > port_qb ? 1 : -1; | |
a130f1a8 KT |
4744 | } else { |
4745 | return netdev_rxq_get_queue_id(qa->rx) | |
4746 | - netdev_rxq_get_queue_id(qb->rx); | |
4747 | } | |
655856ef | 4748 | } |
655856ef KT |
4749 | } |
4750 | ||
e32971b8 DDP |
4751 | /* Assign pmds to queues. If 'pinned' is true, assign pmds to pinned |
4752 | * queues and marks the pmds as isolated. Otherwise, assign non isolated | |
4753 | * pmds to unpinned queues. | |
4754 | * | |
4755 | * The function doesn't touch the pmd threads, it just stores the assignment | |
4756 | * in the 'pmd' member of each rxq. */ | |
4757 | static void | |
4758 | rxq_scheduling(struct dp_netdev *dp, bool pinned) OVS_REQUIRES(dp->port_mutex) | |
4759 | { | |
4760 | struct dp_netdev_port *port; | |
4761 | struct rr_numa_list rr; | |
c37813fd | 4762 | struct rr_numa *non_local_numa = NULL; |
655856ef | 4763 | struct dp_netdev_rxq ** rxqs = NULL; |
97bf8f47 | 4764 | int n_rxqs = 0; |
655856ef KT |
4765 | struct rr_numa *numa = NULL; |
4766 | int numa_id; | |
e77c97b9 | 4767 | bool assign_cyc = dp->pmd_rxq_assign_cyc; |
e32971b8 DDP |
4768 | |
4769 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
e32971b8 DDP |
4770 | if (!netdev_is_pmd(port->netdev)) { |
4771 | continue; | |
4772 | } | |
4773 | ||
e32971b8 DDP |
4774 | for (int qid = 0; qid < port->n_rxq; qid++) { |
4775 | struct dp_netdev_rxq *q = &port->rxqs[qid]; | |
4776 | ||
4777 | if (pinned && q->core_id != OVS_CORE_UNSPEC) { | |
4778 | struct dp_netdev_pmd_thread *pmd; | |
4779 | ||
4780 | pmd = dp_netdev_get_pmd(dp, q->core_id); | |
4781 | if (!pmd) { | |
4782 | VLOG_WARN("There is no PMD thread on core %d. Queue " | |
4783 | "%d on port \'%s\' will not be polled.", | |
4784 | q->core_id, qid, netdev_get_name(port->netdev)); | |
4785 | } else { | |
4786 | q->pmd = pmd; | |
4787 | pmd->isolated = true; | |
433a3fa5 GM |
4788 | VLOG_INFO("Core %d on numa node %d assigned port \'%s\' " |
4789 | "rx queue %d.", pmd->core_id, pmd->numa_id, | |
4790 | netdev_rxq_get_name(q->rx), | |
4791 | netdev_rxq_get_queue_id(q->rx)); | |
e32971b8 DDP |
4792 | dp_netdev_pmd_unref(pmd); |
4793 | } | |
4794 | } else if (!pinned && q->core_id == OVS_CORE_UNSPEC) { | |
8368866e KT |
4795 | uint64_t cycle_hist = 0; |
4796 | ||
655856ef KT |
4797 | if (n_rxqs == 0) { |
4798 | rxqs = xmalloc(sizeof *rxqs); | |
e32971b8 | 4799 | } else { |
655856ef | 4800 | rxqs = xrealloc(rxqs, sizeof *rxqs * (n_rxqs + 1)); |
e32971b8 | 4801 | } |
8368866e | 4802 | |
e77c97b9 KT |
4803 | if (assign_cyc) { |
4804 | /* Sum the queue intervals and store the cycle history. */ | |
4805 | for (unsigned i = 0; i < PMD_RXQ_INTERVAL_MAX; i++) { | |
4806 | cycle_hist += dp_netdev_rxq_get_intrvl_cycles(q, i); | |
4807 | } | |
4808 | dp_netdev_rxq_set_cycles(q, RXQ_CYCLES_PROC_HIST, | |
4809 | cycle_hist); | |
4810 | } | |
655856ef KT |
4811 | /* Store the queue. */ |
4812 | rxqs[n_rxqs++] = q; | |
e32971b8 DDP |
4813 | } |
4814 | } | |
4815 | } | |
4816 | ||
e77c97b9 | 4817 | if (n_rxqs > 1 && assign_cyc) { |
655856ef KT |
4818 | /* Sort the queues in order of the processing cycles |
4819 | * they consumed during their last pmd interval. */ | |
cc131ac1 | 4820 | qsort(rxqs, n_rxqs, sizeof *rxqs, compare_rxq_cycles); |
655856ef KT |
4821 | } |
4822 | ||
4823 | rr_numa_list_populate(dp, &rr); | |
4824 | /* Assign the sorted queues to pmds in round robin. */ | |
97bf8f47 | 4825 | for (int i = 0; i < n_rxqs; i++) { |
655856ef KT |
4826 | numa_id = netdev_get_numa_id(rxqs[i]->port->netdev); |
4827 | numa = rr_numa_list_lookup(&rr, numa_id); | |
4828 | if (!numa) { | |
4829 | /* There are no pmds on the queue's local NUMA node. | |
4830 | Round robin on the NUMA nodes that do have pmds. */ | |
4831 | non_local_numa = rr_numa_list_next(&rr, non_local_numa); | |
4832 | if (!non_local_numa) { | |
4833 | VLOG_ERR("There is no available (non-isolated) pmd " | |
4834 | "thread for port \'%s\' queue %d. This queue " | |
4835 | "will not be polled. Is pmd-cpu-mask set to " | |
4836 | "zero? Or are all PMDs isolated to other " | |
4837 | "queues?", netdev_rxq_get_name(rxqs[i]->rx), | |
4838 | netdev_rxq_get_queue_id(rxqs[i]->rx)); | |
4839 | continue; | |
4840 | } | |
e77c97b9 | 4841 | rxqs[i]->pmd = rr_numa_get_pmd(non_local_numa, assign_cyc); |
655856ef KT |
4842 | VLOG_WARN("There's no available (non-isolated) pmd thread " |
4843 | "on numa node %d. Queue %d on port \'%s\' will " | |
4844 | "be assigned to the pmd on core %d " | |
4845 | "(numa node %d). Expect reduced performance.", | |
4846 | numa_id, netdev_rxq_get_queue_id(rxqs[i]->rx), | |
4847 | netdev_rxq_get_name(rxqs[i]->rx), | |
4848 | rxqs[i]->pmd->core_id, rxqs[i]->pmd->numa_id); | |
4849 | } else { | |
e77c97b9 KT |
4850 | rxqs[i]->pmd = rr_numa_get_pmd(numa, assign_cyc); |
4851 | if (assign_cyc) { | |
4852 | VLOG_INFO("Core %d on numa node %d assigned port \'%s\' " | |
4853 | "rx queue %d " | |
4854 | "(measured processing cycles %"PRIu64").", | |
4855 | rxqs[i]->pmd->core_id, numa_id, | |
4856 | netdev_rxq_get_name(rxqs[i]->rx), | |
4857 | netdev_rxq_get_queue_id(rxqs[i]->rx), | |
4858 | dp_netdev_rxq_get_cycles(rxqs[i], | |
4859 | RXQ_CYCLES_PROC_HIST)); | |
4860 | } else { | |
4861 | VLOG_INFO("Core %d on numa node %d assigned port \'%s\' " | |
4862 | "rx queue %d.", rxqs[i]->pmd->core_id, numa_id, | |
4863 | netdev_rxq_get_name(rxqs[i]->rx), | |
4864 | netdev_rxq_get_queue_id(rxqs[i]->rx)); | |
4865 | } | |
655856ef KT |
4866 | } |
4867 | } | |
4868 | ||
e32971b8 | 4869 | rr_numa_list_destroy(&rr); |
655856ef | 4870 | free(rxqs); |
e32971b8 DDP |
4871 | } |
4872 | ||
140dd699 IM |
4873 | static void |
4874 | reload_affected_pmds(struct dp_netdev *dp) | |
4875 | { | |
4876 | struct dp_netdev_pmd_thread *pmd; | |
4877 | ||
4878 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
4879 | if (pmd->need_reload) { | |
241bad15 | 4880 | flow_mark_flush(pmd); |
140dd699 | 4881 | dp_netdev_reload_pmd__(pmd); |
8f077b31 DM |
4882 | } |
4883 | } | |
4884 | ||
4885 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
4886 | if (pmd->need_reload) { | |
4887 | if (pmd->core_id != NON_PMD_CORE_ID) { | |
4888 | bool reload; | |
4889 | ||
4890 | do { | |
4891 | atomic_read_explicit(&pmd->reload, &reload, | |
4892 | memory_order_acquire); | |
4893 | } while (reload); | |
4894 | } | |
140dd699 IM |
4895 | pmd->need_reload = false; |
4896 | } | |
4897 | } | |
4898 | } | |
4899 | ||
6e3c6fa4 DDP |
4900 | static void |
4901 | reconfigure_pmd_threads(struct dp_netdev *dp) | |
4902 | OVS_REQUIRES(dp->port_mutex) | |
4903 | { | |
e32971b8 DDP |
4904 | struct dp_netdev_pmd_thread *pmd; |
4905 | struct ovs_numa_dump *pmd_cores; | |
140dd699 IM |
4906 | struct ovs_numa_info_core *core; |
4907 | struct hmapx to_delete = HMAPX_INITIALIZER(&to_delete); | |
4908 | struct hmapx_node *node; | |
e32971b8 | 4909 | bool changed = false; |
140dd699 | 4910 | bool need_to_adjust_static_tx_qids = false; |
e32971b8 DDP |
4911 | |
4912 | /* The pmd threads should be started only if there's a pmd port in the | |
4913 | * datapath. If the user didn't provide any "pmd-cpu-mask", we start | |
4914 | * NR_PMD_THREADS per numa node. */ | |
4915 | if (!has_pmd_port(dp)) { | |
4916 | pmd_cores = ovs_numa_dump_n_cores_per_numa(0); | |
4917 | } else if (dp->pmd_cmask && dp->pmd_cmask[0]) { | |
4918 | pmd_cores = ovs_numa_dump_cores_with_cmask(dp->pmd_cmask); | |
4919 | } else { | |
4920 | pmd_cores = ovs_numa_dump_n_cores_per_numa(NR_PMD_THREADS); | |
4921 | } | |
4922 | ||
140dd699 IM |
4923 | /* We need to adjust 'static_tx_qid's only if we're reducing number of |
4924 | * PMD threads. Otherwise, new threads will allocate all the freed ids. */ | |
4925 | if (ovs_numa_dump_count(pmd_cores) < cmap_count(&dp->poll_threads) - 1) { | |
4926 | /* Adjustment is required to keep 'static_tx_qid's sequential and | |
4927 | * avoid possible issues, for example, imbalanced tx queue usage | |
4928 | * and unnecessary locking caused by remapping on netdev level. */ | |
4929 | need_to_adjust_static_tx_qids = true; | |
4930 | } | |
4931 | ||
4932 | /* Check for unwanted pmd threads */ | |
4933 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
4934 | if (pmd->core_id == NON_PMD_CORE_ID) { | |
4935 | continue; | |
4936 | } | |
4937 | if (!ovs_numa_dump_contains_core(pmd_cores, pmd->numa_id, | |
4938 | pmd->core_id)) { | |
4939 | hmapx_add(&to_delete, pmd); | |
4940 | } else if (need_to_adjust_static_tx_qids) { | |
e2cafa86 | 4941 | atomic_store_relaxed(&pmd->reload_tx_qid, true); |
140dd699 | 4942 | pmd->need_reload = true; |
e32971b8 DDP |
4943 | } |
4944 | } | |
4945 | ||
140dd699 IM |
4946 | HMAPX_FOR_EACH (node, &to_delete) { |
4947 | pmd = (struct dp_netdev_pmd_thread *) node->data; | |
4948 | VLOG_INFO("PMD thread on numa_id: %d, core id: %2d destroyed.", | |
4949 | pmd->numa_id, pmd->core_id); | |
4950 | dp_netdev_del_pmd(dp, pmd); | |
4951 | } | |
4952 | changed = !hmapx_is_empty(&to_delete); | |
4953 | hmapx_destroy(&to_delete); | |
e32971b8 | 4954 | |
140dd699 IM |
4955 | if (need_to_adjust_static_tx_qids) { |
4956 | /* 'static_tx_qid's are not sequential now. | |
4957 | * Reload remaining threads to fix this. */ | |
4958 | reload_affected_pmds(dp); | |
4959 | } | |
e32971b8 | 4960 | |
140dd699 IM |
4961 | /* Check for required new pmd threads */ |
4962 | FOR_EACH_CORE_ON_DUMP(core, pmd_cores) { | |
4963 | pmd = dp_netdev_get_pmd(dp, core->core_id); | |
4964 | if (!pmd) { | |
8afbf2fa IM |
4965 | struct ds name = DS_EMPTY_INITIALIZER; |
4966 | ||
140dd699 | 4967 | pmd = xzalloc(sizeof *pmd); |
e32971b8 | 4968 | dp_netdev_configure_pmd(pmd, dp, core->core_id, core->numa_id); |
8afbf2fa IM |
4969 | |
4970 | ds_put_format(&name, "pmd-c%02d/id:", core->core_id); | |
4971 | pmd->thread = ovs_thread_create(ds_cstr(&name), | |
4972 | pmd_thread_main, pmd); | |
4973 | ds_destroy(&name); | |
4974 | ||
140dd699 IM |
4975 | VLOG_INFO("PMD thread on numa_id: %d, core id: %2d created.", |
4976 | pmd->numa_id, pmd->core_id); | |
4977 | changed = true; | |
4978 | } else { | |
4979 | dp_netdev_pmd_unref(pmd); | |
e32971b8 | 4980 | } |
140dd699 IM |
4981 | } |
4982 | ||
4983 | if (changed) { | |
4984 | struct ovs_numa_info_numa *numa; | |
e32971b8 DDP |
4985 | |
4986 | /* Log the number of pmd threads per numa node. */ | |
4987 | FOR_EACH_NUMA_ON_DUMP (numa, pmd_cores) { | |
140dd699 | 4988 | VLOG_INFO("There are %"PRIuSIZE" pmd threads on numa node %d", |
e32971b8 DDP |
4989 | numa->n_cores, numa->numa_id); |
4990 | } | |
4991 | } | |
4992 | ||
4993 | ovs_numa_dump_destroy(pmd_cores); | |
4994 | } | |
4995 | ||
e32971b8 DDP |
4996 | static void |
4997 | pmd_remove_stale_ports(struct dp_netdev *dp, | |
4998 | struct dp_netdev_pmd_thread *pmd) | |
4999 | OVS_EXCLUDED(pmd->port_mutex) | |
5000 | OVS_REQUIRES(dp->port_mutex) | |
5001 | { | |
5002 | struct rxq_poll *poll, *poll_next; | |
5003 | struct tx_port *tx, *tx_next; | |
5004 | ||
5005 | ovs_mutex_lock(&pmd->port_mutex); | |
5006 | HMAP_FOR_EACH_SAFE (poll, poll_next, node, &pmd->poll_list) { | |
5007 | struct dp_netdev_port *port = poll->rxq->port; | |
5008 | ||
5009 | if (port->need_reconfigure | |
5010 | || !hmap_contains(&dp->ports, &port->node)) { | |
5011 | dp_netdev_del_rxq_from_pmd(pmd, poll); | |
5012 | } | |
5013 | } | |
5014 | HMAP_FOR_EACH_SAFE (tx, tx_next, node, &pmd->tx_ports) { | |
5015 | struct dp_netdev_port *port = tx->port; | |
5016 | ||
5017 | if (port->need_reconfigure | |
5018 | || !hmap_contains(&dp->ports, &port->node)) { | |
5019 | dp_netdev_del_port_tx_from_pmd(pmd, tx); | |
5020 | } | |
5021 | } | |
5022 | ovs_mutex_unlock(&pmd->port_mutex); | |
5023 | } | |
5024 | ||
5025 | /* Must be called each time a port is added/removed or the cmask changes. | |
5026 | * This creates and destroys pmd threads, reconfigures ports, opens their | |
5027 | * rxqs and assigns all rxqs/txqs to pmd threads. */ | |
5028 | static void | |
5029 | reconfigure_datapath(struct dp_netdev *dp) | |
5030 | OVS_REQUIRES(dp->port_mutex) | |
5031 | { | |
6d9fead1 | 5032 | struct hmapx busy_threads = HMAPX_INITIALIZER(&busy_threads); |
e32971b8 DDP |
5033 | struct dp_netdev_pmd_thread *pmd; |
5034 | struct dp_netdev_port *port; | |
5035 | int wanted_txqs; | |
6e3c6fa4 | 5036 | |
a6a426d6 IM |
5037 | dp->last_reconfigure_seq = seq_read(dp->reconfigure_seq); |
5038 | ||
e32971b8 DDP |
5039 | /* Step 1: Adjust the pmd threads based on the datapath ports, the cores |
5040 | * on the system and the user configuration. */ | |
5041 | reconfigure_pmd_threads(dp); | |
6e3c6fa4 | 5042 | |
e32971b8 | 5043 | wanted_txqs = cmap_count(&dp->poll_threads); |
324c8374 | 5044 | |
e32971b8 DDP |
5045 | /* The number of pmd threads might have changed, or a port can be new: |
5046 | * adjust the txqs. */ | |
5047 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
5048 | netdev_set_tx_multiq(port->netdev, wanted_txqs); | |
324c8374 IM |
5049 | } |
5050 | ||
e32971b8 DDP |
5051 | /* Step 2: Remove from the pmd threads ports that have been removed or |
5052 | * need reconfiguration. */ | |
5053 | ||
5054 | /* Check for all the ports that need reconfiguration. We cache this in | |
85a4f238 | 5055 | * 'port->need_reconfigure', because netdev_is_reconf_required() can |
f598f462 IM |
5056 | * change at any time. |
5057 | * Also mark for reconfiguration all ports which will likely change their | |
5058 | * 'dynamic_txqs' parameter. It's required to stop using them before | |
5059 | * changing this setting and it's simpler to mark ports here and allow | |
5060 | * 'pmd_remove_stale_ports' to remove them from threads. There will be | |
5061 | * no actual reconfiguration in 'port_reconfigure' because it's | |
5062 | * unnecessary. */ | |
e32971b8 | 5063 | HMAP_FOR_EACH (port, node, &dp->ports) { |
f598f462 IM |
5064 | if (netdev_is_reconf_required(port->netdev) |
5065 | || (port->dynamic_txqs | |
5066 | != (netdev_n_txq(port->netdev) < wanted_txqs))) { | |
e32971b8 DDP |
5067 | port->need_reconfigure = true; |
5068 | } | |
5069 | } | |
5070 | ||
5071 | /* Remove from the pmd threads all the ports that have been deleted or | |
5072 | * need reconfiguration. */ | |
5073 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
5074 | pmd_remove_stale_ports(dp, pmd); | |
5075 | } | |
5076 | ||
5077 | /* Reload affected pmd threads. We must wait for the pmd threads before | |
5078 | * reconfiguring the ports, because a port cannot be reconfigured while | |
5079 | * it's being used. */ | |
5080 | reload_affected_pmds(dp); | |
5081 | ||
5082 | /* Step 3: Reconfigure ports. */ | |
5083 | ||
5084 | /* We only reconfigure the ports that we determined above, because they're | |
5085 | * not being used by any pmd thread at the moment. If a port fails to | |
5086 | * reconfigure we remove it from the datapath. */ | |
f582b6df BP |
5087 | struct dp_netdev_port *next_port; |
5088 | HMAP_FOR_EACH_SAFE (port, next_port, node, &dp->ports) { | |
dc36593c | 5089 | int err; |
6e3c6fa4 | 5090 | |
e32971b8 DDP |
5091 | if (!port->need_reconfigure) { |
5092 | continue; | |
5093 | } | |
5094 | ||
dc36593c DDP |
5095 | err = port_reconfigure(port); |
5096 | if (err) { | |
5097 | hmap_remove(&dp->ports, &port->node); | |
5098 | seq_change(dp->port_seq); | |
5099 | port_destroy(port); | |
324c8374 | 5100 | } else { |
e32971b8 | 5101 | port->dynamic_txqs = netdev_n_txq(port->netdev) < wanted_txqs; |
6e3c6fa4 DDP |
5102 | } |
5103 | } | |
e32971b8 DDP |
5104 | |
5105 | /* Step 4: Compute new rxq scheduling. We don't touch the pmd threads | |
5106 | * for now, we just update the 'pmd' pointer in each rxq to point to the | |
5107 | * wanted thread according to the scheduling policy. */ | |
5108 | ||
5109 | /* Reset all the pmd threads to non isolated. */ | |
5110 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
5111 | pmd->isolated = false; | |
5112 | } | |
5113 | ||
5114 | /* Reset all the queues to unassigned */ | |
5115 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
5116 | for (int i = 0; i < port->n_rxq; i++) { | |
5117 | port->rxqs[i].pmd = NULL; | |
5118 | } | |
5119 | } | |
5120 | ||
5121 | /* Add pinned queues and mark pmd threads isolated. */ | |
5122 | rxq_scheduling(dp, true); | |
5123 | ||
5124 | /* Add non-pinned queues. */ | |
5125 | rxq_scheduling(dp, false); | |
5126 | ||
5127 | /* Step 5: Remove queues not compliant with new scheduling. */ | |
6d9fead1 DM |
5128 | |
5129 | /* Count all the threads that will have at least one queue to poll. */ | |
5130 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
5131 | for (int qid = 0; qid < port->n_rxq; qid++) { | |
5132 | struct dp_netdev_rxq *q = &port->rxqs[qid]; | |
5133 | ||
5134 | if (q->pmd) { | |
5135 | hmapx_add(&busy_threads, q->pmd); | |
5136 | } | |
5137 | } | |
5138 | } | |
5139 | ||
e32971b8 DDP |
5140 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { |
5141 | struct rxq_poll *poll, *poll_next; | |
5142 | ||
5143 | ovs_mutex_lock(&pmd->port_mutex); | |
5144 | HMAP_FOR_EACH_SAFE (poll, poll_next, node, &pmd->poll_list) { | |
5145 | if (poll->rxq->pmd != pmd) { | |
5146 | dp_netdev_del_rxq_from_pmd(pmd, poll); | |
6d9fead1 DM |
5147 | |
5148 | /* This pmd might sleep after this step if it has no rxq | |
5149 | * remaining. Tell it to busy wait for new assignment if it | |
5150 | * has at least one scheduled queue. */ | |
5151 | if (hmap_count(&pmd->poll_list) == 0 && | |
5152 | hmapx_contains(&busy_threads, pmd)) { | |
5153 | atomic_store_relaxed(&pmd->wait_for_reload, true); | |
5154 | } | |
e32971b8 DDP |
5155 | } |
5156 | } | |
5157 | ovs_mutex_unlock(&pmd->port_mutex); | |
5158 | } | |
5159 | ||
6d9fead1 DM |
5160 | hmapx_destroy(&busy_threads); |
5161 | ||
e32971b8 DDP |
5162 | /* Reload affected pmd threads. We must wait for the pmd threads to remove |
5163 | * the old queues before readding them, otherwise a queue can be polled by | |
5164 | * two threads at the same time. */ | |
5165 | reload_affected_pmds(dp); | |
5166 | ||
5167 | /* Step 6: Add queues from scheduling, if they're not there already. */ | |
5168 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
5169 | if (!netdev_is_pmd(port->netdev)) { | |
5170 | continue; | |
5171 | } | |
5172 | ||
5173 | for (int qid = 0; qid < port->n_rxq; qid++) { | |
5174 | struct dp_netdev_rxq *q = &port->rxqs[qid]; | |
5175 | ||
5176 | if (q->pmd) { | |
5177 | ovs_mutex_lock(&q->pmd->port_mutex); | |
5178 | dp_netdev_add_rxq_to_pmd(q->pmd, q); | |
5179 | ovs_mutex_unlock(&q->pmd->port_mutex); | |
5180 | } | |
5181 | } | |
5182 | } | |
5183 | ||
9df65060 VDA |
5184 | /* Add every port and bond to the tx port and bond caches of |
5185 | * every pmd thread, if it's not there already and if this pmd | |
5186 | * has at least one rxq to poll. | |
5187 | */ | |
e32971b8 DDP |
5188 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { |
5189 | ovs_mutex_lock(&pmd->port_mutex); | |
5190 | if (hmap_count(&pmd->poll_list) || pmd->core_id == NON_PMD_CORE_ID) { | |
9df65060 VDA |
5191 | struct tx_bond *bond; |
5192 | ||
e32971b8 DDP |
5193 | HMAP_FOR_EACH (port, node, &dp->ports) { |
5194 | dp_netdev_add_port_tx_to_pmd(pmd, port); | |
5195 | } | |
9df65060 VDA |
5196 | |
5197 | CMAP_FOR_EACH (bond, node, &dp->tx_bonds) { | |
5198 | dp_netdev_add_bond_tx_to_pmd(pmd, bond, false); | |
5199 | } | |
e32971b8 DDP |
5200 | } |
5201 | ovs_mutex_unlock(&pmd->port_mutex); | |
5202 | } | |
5203 | ||
5204 | /* Reload affected pmd threads. */ | |
5205 | reload_affected_pmds(dp); | |
5bf84282 NK |
5206 | |
5207 | /* Check if PMD Auto LB is to be enabled */ | |
5208 | set_pmd_auto_lb(dp); | |
6e3c6fa4 DDP |
5209 | } |
5210 | ||
050c60bf DDP |
5211 | /* Returns true if one of the netdevs in 'dp' requires a reconfiguration */ |
5212 | static bool | |
5213 | ports_require_restart(const struct dp_netdev *dp) | |
5214 | OVS_REQUIRES(dp->port_mutex) | |
5215 | { | |
5216 | struct dp_netdev_port *port; | |
5217 | ||
5218 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
5219 | if (netdev_is_reconf_required(port->netdev)) { | |
5220 | return true; | |
5221 | } | |
5222 | } | |
5223 | ||
5224 | return false; | |
5225 | } | |
5226 | ||
5bf84282 NK |
5227 | /* Calculates variance in the values stored in array 'a'. 'n' is the number |
5228 | * of elements in array to be considered for calculating vairance. | |
5229 | * Usage example: data array 'a' contains the processing load of each pmd and | |
5230 | * 'n' is the number of PMDs. It returns the variance in processing load of | |
5231 | * PMDs*/ | |
5232 | static uint64_t | |
5233 | variance(uint64_t a[], int n) | |
5234 | { | |
5235 | /* Compute mean (average of elements). */ | |
5236 | uint64_t sum = 0; | |
5237 | uint64_t mean = 0; | |
5238 | uint64_t sqDiff = 0; | |
5239 | ||
5240 | if (!n) { | |
5241 | return 0; | |
5242 | } | |
5243 | ||
5244 | for (int i = 0; i < n; i++) { | |
5245 | sum += a[i]; | |
5246 | } | |
5247 | ||
5248 | if (sum) { | |
5249 | mean = sum / n; | |
5250 | ||
5251 | /* Compute sum squared differences with mean. */ | |
5252 | for (int i = 0; i < n; i++) { | |
5253 | sqDiff += (a[i] - mean)*(a[i] - mean); | |
5254 | } | |
5255 | } | |
5256 | return (sqDiff ? (sqDiff / n) : 0); | |
5257 | } | |
5258 | ||
5259 | ||
5260 | /* Returns the variance in the PMDs usage as part of dry run of rxqs | |
5261 | * assignment to PMDs. */ | |
5262 | static bool | |
5263 | get_dry_run_variance(struct dp_netdev *dp, uint32_t *core_list, | |
5264 | uint32_t num_pmds, uint64_t *predicted_variance) | |
5265 | OVS_REQUIRES(dp->port_mutex) | |
5266 | { | |
5267 | struct dp_netdev_port *port; | |
5268 | struct dp_netdev_pmd_thread *pmd; | |
5269 | struct dp_netdev_rxq **rxqs = NULL; | |
5270 | struct rr_numa *numa = NULL; | |
5271 | struct rr_numa_list rr; | |
5272 | int n_rxqs = 0; | |
5273 | bool ret = false; | |
5274 | uint64_t *pmd_usage; | |
5275 | ||
5276 | if (!predicted_variance) { | |
5277 | return ret; | |
5278 | } | |
5279 | ||
5280 | pmd_usage = xcalloc(num_pmds, sizeof(uint64_t)); | |
5281 | ||
5282 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
5283 | if (!netdev_is_pmd(port->netdev)) { | |
5284 | continue; | |
5285 | } | |
5286 | ||
5287 | for (int qid = 0; qid < port->n_rxq; qid++) { | |
5288 | struct dp_netdev_rxq *q = &port->rxqs[qid]; | |
5289 | uint64_t cycle_hist = 0; | |
5290 | ||
5291 | if (q->pmd->isolated) { | |
5292 | continue; | |
5293 | } | |
5294 | ||
5295 | if (n_rxqs == 0) { | |
5296 | rxqs = xmalloc(sizeof *rxqs); | |
5297 | } else { | |
5298 | rxqs = xrealloc(rxqs, sizeof *rxqs * (n_rxqs + 1)); | |
5299 | } | |
5300 | ||
5301 | /* Sum the queue intervals and store the cycle history. */ | |
5302 | for (unsigned i = 0; i < PMD_RXQ_INTERVAL_MAX; i++) { | |
5303 | cycle_hist += dp_netdev_rxq_get_intrvl_cycles(q, i); | |
5304 | } | |
5305 | dp_netdev_rxq_set_cycles(q, RXQ_CYCLES_PROC_HIST, | |
5306 | cycle_hist); | |
5307 | /* Store the queue. */ | |
5308 | rxqs[n_rxqs++] = q; | |
5309 | } | |
5310 | } | |
5311 | if (n_rxqs > 1) { | |
5312 | /* Sort the queues in order of the processing cycles | |
5313 | * they consumed during their last pmd interval. */ | |
5314 | qsort(rxqs, n_rxqs, sizeof *rxqs, compare_rxq_cycles); | |
5315 | } | |
5316 | rr_numa_list_populate(dp, &rr); | |
5317 | ||
5318 | for (int i = 0; i < n_rxqs; i++) { | |
5319 | int numa_id = netdev_get_numa_id(rxqs[i]->port->netdev); | |
5320 | numa = rr_numa_list_lookup(&rr, numa_id); | |
5321 | if (!numa) { | |
5322 | /* Abort if cross NUMA polling. */ | |
5323 | VLOG_DBG("PMD auto lb dry run." | |
5324 | " Aborting due to cross-numa polling."); | |
5325 | goto cleanup; | |
5326 | } | |
5327 | ||
5328 | pmd = rr_numa_get_pmd(numa, true); | |
5329 | VLOG_DBG("PMD auto lb dry run. Predicted: Core %d on numa node %d " | |
5330 | "to be assigned port \'%s\' rx queue %d " | |
5331 | "(measured processing cycles %"PRIu64").", | |
5332 | pmd->core_id, numa_id, | |
5333 | netdev_rxq_get_name(rxqs[i]->rx), | |
5334 | netdev_rxq_get_queue_id(rxqs[i]->rx), | |
5335 | dp_netdev_rxq_get_cycles(rxqs[i], RXQ_CYCLES_PROC_HIST)); | |
5336 | ||
5337 | for (int id = 0; id < num_pmds; id++) { | |
5338 | if (pmd->core_id == core_list[id]) { | |
5339 | /* Add the processing cycles of rxq to pmd polling it. */ | |
5340 | pmd_usage[id] += dp_netdev_rxq_get_cycles(rxqs[i], | |
5341 | RXQ_CYCLES_PROC_HIST); | |
5342 | } | |
5343 | } | |
5344 | } | |
5345 | ||
5346 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
5347 | uint64_t total_cycles = 0; | |
5348 | ||
5349 | if ((pmd->core_id == NON_PMD_CORE_ID) || pmd->isolated) { | |
5350 | continue; | |
5351 | } | |
5352 | ||
5353 | /* Get the total pmd cycles for an interval. */ | |
5354 | atomic_read_relaxed(&pmd->intrvl_cycles, &total_cycles); | |
5355 | /* Estimate the cycles to cover all intervals. */ | |
5356 | total_cycles *= PMD_RXQ_INTERVAL_MAX; | |
5357 | for (int id = 0; id < num_pmds; id++) { | |
5358 | if (pmd->core_id == core_list[id]) { | |
5359 | if (pmd_usage[id]) { | |
5360 | pmd_usage[id] = (pmd_usage[id] * 100) / total_cycles; | |
5361 | } | |
5362 | VLOG_DBG("PMD auto lb dry run. Predicted: Core %d, " | |
5363 | "usage %"PRIu64"", pmd->core_id, pmd_usage[id]); | |
5364 | } | |
5365 | } | |
5366 | } | |
5367 | *predicted_variance = variance(pmd_usage, num_pmds); | |
5368 | ret = true; | |
5369 | ||
5370 | cleanup: | |
5371 | rr_numa_list_destroy(&rr); | |
5372 | free(rxqs); | |
5373 | free(pmd_usage); | |
5374 | return ret; | |
5375 | } | |
5376 | ||
5377 | /* Does the dry run of Rxq assignment to PMDs and returns true if it gives | |
5378 | * better distribution of load on PMDs. */ | |
5379 | static bool | |
5380 | pmd_rebalance_dry_run(struct dp_netdev *dp) | |
5381 | OVS_REQUIRES(dp->port_mutex) | |
5382 | { | |
5383 | struct dp_netdev_pmd_thread *pmd; | |
5384 | uint64_t *curr_pmd_usage; | |
5385 | ||
5386 | uint64_t curr_variance; | |
5387 | uint64_t new_variance; | |
5388 | uint64_t improvement = 0; | |
5389 | uint32_t num_pmds; | |
5390 | uint32_t *pmd_corelist; | |
eef85380 | 5391 | struct rxq_poll *poll; |
5bf84282 NK |
5392 | bool ret; |
5393 | ||
5394 | num_pmds = cmap_count(&dp->poll_threads); | |
5395 | ||
5396 | if (num_pmds > 1) { | |
5397 | curr_pmd_usage = xcalloc(num_pmds, sizeof(uint64_t)); | |
5398 | pmd_corelist = xcalloc(num_pmds, sizeof(uint32_t)); | |
5399 | } else { | |
5400 | return false; | |
5401 | } | |
5402 | ||
5403 | num_pmds = 0; | |
5404 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
5405 | uint64_t total_cycles = 0; | |
5406 | uint64_t total_proc = 0; | |
5407 | ||
5408 | if ((pmd->core_id == NON_PMD_CORE_ID) || pmd->isolated) { | |
5409 | continue; | |
5410 | } | |
5411 | ||
5412 | /* Get the total pmd cycles for an interval. */ | |
5413 | atomic_read_relaxed(&pmd->intrvl_cycles, &total_cycles); | |
5414 | /* Estimate the cycles to cover all intervals. */ | |
5415 | total_cycles *= PMD_RXQ_INTERVAL_MAX; | |
5416 | ||
eef85380 IM |
5417 | ovs_mutex_lock(&pmd->port_mutex); |
5418 | HMAP_FOR_EACH (poll, node, &pmd->poll_list) { | |
5bf84282 | 5419 | for (unsigned i = 0; i < PMD_RXQ_INTERVAL_MAX; i++) { |
eef85380 | 5420 | total_proc += dp_netdev_rxq_get_intrvl_cycles(poll->rxq, i); |
5bf84282 | 5421 | } |
5bf84282 | 5422 | } |
eef85380 IM |
5423 | ovs_mutex_unlock(&pmd->port_mutex); |
5424 | ||
5bf84282 NK |
5425 | if (total_proc) { |
5426 | curr_pmd_usage[num_pmds] = (total_proc * 100) / total_cycles; | |
5427 | } | |
5428 | ||
5429 | VLOG_DBG("PMD auto lb dry run. Current: Core %d, usage %"PRIu64"", | |
5430 | pmd->core_id, curr_pmd_usage[num_pmds]); | |
5431 | ||
5432 | if (atomic_count_get(&pmd->pmd_overloaded)) { | |
5433 | atomic_count_set(&pmd->pmd_overloaded, 0); | |
5434 | } | |
5435 | ||
5436 | pmd_corelist[num_pmds] = pmd->core_id; | |
5437 | num_pmds++; | |
5438 | } | |
5439 | ||
5440 | curr_variance = variance(curr_pmd_usage, num_pmds); | |
5441 | ret = get_dry_run_variance(dp, pmd_corelist, num_pmds, &new_variance); | |
5442 | ||
5443 | if (ret) { | |
5444 | VLOG_DBG("PMD auto lb dry run. Current PMD variance: %"PRIu64"," | |
5445 | " Predicted PMD variance: %"PRIu64"", | |
5446 | curr_variance, new_variance); | |
5447 | ||
5448 | if (new_variance < curr_variance) { | |
5449 | improvement = | |
5450 | ((curr_variance - new_variance) * 100) / curr_variance; | |
5451 | } | |
5452 | if (improvement < ALB_ACCEPTABLE_IMPROVEMENT) { | |
5453 | ret = false; | |
5454 | } | |
5455 | } | |
5456 | ||
5457 | free(curr_pmd_usage); | |
5458 | free(pmd_corelist); | |
5459 | return ret; | |
5460 | } | |
5461 | ||
5462 | ||
a36de779 PS |
5463 | /* Return true if needs to revalidate datapath flows. */ |
5464 | static bool | |
e4cfed38 PS |
5465 | dpif_netdev_run(struct dpif *dpif) |
5466 | { | |
5467 | struct dp_netdev_port *port; | |
5468 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
546e57d4 | 5469 | struct dp_netdev_pmd_thread *non_pmd; |
a36de779 | 5470 | uint64_t new_tnl_seq; |
c71ea3c4 | 5471 | bool need_to_flush = true; |
5bf84282 NK |
5472 | bool pmd_rebalance = false; |
5473 | long long int now = time_msec(); | |
5474 | struct dp_netdev_pmd_thread *pmd; | |
e4cfed38 | 5475 | |
e9985d6a | 5476 | ovs_mutex_lock(&dp->port_mutex); |
546e57d4 DDP |
5477 | non_pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID); |
5478 | if (non_pmd) { | |
5479 | ovs_mutex_lock(&dp->non_pmd_mutex); | |
5480 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
5481 | if (!netdev_is_pmd(port->netdev)) { | |
5482 | int i; | |
55c955bd | 5483 | |
2fbadeb6 IM |
5484 | if (port->emc_enabled) { |
5485 | atomic_read_relaxed(&dp->emc_insert_min, | |
5486 | &non_pmd->ctx.emc_insert_min); | |
5487 | } else { | |
5488 | non_pmd->ctx.emc_insert_min = 0; | |
5489 | } | |
5490 | ||
546e57d4 | 5491 | for (i = 0; i < port->n_rxq; i++) { |
35c91567 DM |
5492 | |
5493 | if (!netdev_rxq_enabled(port->rxqs[i].rx)) { | |
5494 | continue; | |
5495 | } | |
5496 | ||
c71ea3c4 IM |
5497 | if (dp_netdev_process_rxq_port(non_pmd, |
5498 | &port->rxqs[i], | |
5499 | port->port_no)) { | |
5500 | need_to_flush = false; | |
5501 | } | |
546e57d4 | 5502 | } |
55c955bd | 5503 | } |
e4cfed38 | 5504 | } |
c71ea3c4 IM |
5505 | if (need_to_flush) { |
5506 | /* We didn't receive anything in the process loop. | |
5507 | * Check if we need to send something. | |
5508 | * There was no time updates on current iteration. */ | |
5509 | pmd_thread_ctx_time_update(non_pmd); | |
5510 | dp_netdev_pmd_flush_output_packets(non_pmd, false); | |
5511 | } | |
5512 | ||
b010be17 | 5513 | dpif_netdev_xps_revalidate_pmd(non_pmd, false); |
546e57d4 | 5514 | ovs_mutex_unlock(&dp->non_pmd_mutex); |
6e3c6fa4 | 5515 | |
546e57d4 DDP |
5516 | dp_netdev_pmd_unref(non_pmd); |
5517 | } | |
1c1e46ed | 5518 | |
5bf84282 NK |
5519 | struct pmd_auto_lb *pmd_alb = &dp->pmd_alb; |
5520 | if (pmd_alb->is_enabled) { | |
5521 | if (!pmd_alb->rebalance_poll_timer) { | |
5522 | pmd_alb->rebalance_poll_timer = now; | |
5523 | } else if ((pmd_alb->rebalance_poll_timer + | |
5524 | pmd_alb->rebalance_intvl) < now) { | |
5525 | pmd_alb->rebalance_poll_timer = now; | |
5526 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
5527 | if (atomic_count_get(&pmd->pmd_overloaded) >= | |
5528 | PMD_RXQ_INTERVAL_MAX) { | |
5529 | pmd_rebalance = true; | |
5530 | break; | |
5531 | } | |
5532 | } | |
5533 | ||
5534 | if (pmd_rebalance && | |
5535 | !dp_netdev_is_reconf_required(dp) && | |
5536 | !ports_require_restart(dp) && | |
5537 | pmd_rebalance_dry_run(dp)) { | |
5538 | VLOG_INFO("PMD auto lb dry run." | |
5539 | " requesting datapath reconfigure."); | |
5540 | dp_netdev_request_reconfigure(dp); | |
5541 | } | |
5542 | } | |
5543 | } | |
5544 | ||
a6a426d6 | 5545 | if (dp_netdev_is_reconf_required(dp) || ports_require_restart(dp)) { |
e32971b8 | 5546 | reconfigure_datapath(dp); |
6e3c6fa4 DDP |
5547 | } |
5548 | ovs_mutex_unlock(&dp->port_mutex); | |
5549 | ||
53902038 | 5550 | tnl_neigh_cache_run(); |
7f9b8504 | 5551 | tnl_port_map_run(); |
a36de779 PS |
5552 | new_tnl_seq = seq_read(tnl_conf_seq); |
5553 | ||
5554 | if (dp->last_tnl_conf_seq != new_tnl_seq) { | |
5555 | dp->last_tnl_conf_seq = new_tnl_seq; | |
5556 | return true; | |
5557 | } | |
5558 | return false; | |
e4cfed38 PS |
5559 | } |
5560 | ||
5561 | static void | |
5562 | dpif_netdev_wait(struct dpif *dpif) | |
5563 | { | |
5564 | struct dp_netdev_port *port; | |
5565 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
5566 | ||
59e6d833 | 5567 | ovs_mutex_lock(&dp_netdev_mutex); |
e9985d6a DDP |
5568 | ovs_mutex_lock(&dp->port_mutex); |
5569 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
050c60bf | 5570 | netdev_wait_reconf_required(port->netdev); |
55c955bd PS |
5571 | if (!netdev_is_pmd(port->netdev)) { |
5572 | int i; | |
5573 | ||
490e82af | 5574 | for (i = 0; i < port->n_rxq; i++) { |
947dc567 | 5575 | netdev_rxq_wait(port->rxqs[i].rx); |
55c955bd | 5576 | } |
e4cfed38 PS |
5577 | } |
5578 | } | |
e9985d6a | 5579 | ovs_mutex_unlock(&dp->port_mutex); |
59e6d833 | 5580 | ovs_mutex_unlock(&dp_netdev_mutex); |
a36de779 | 5581 | seq_wait(tnl_conf_seq, dp->last_tnl_conf_seq); |
e4cfed38 PS |
5582 | } |
5583 | ||
d0cca6c3 DDP |
5584 | static void |
5585 | pmd_free_cached_ports(struct dp_netdev_pmd_thread *pmd) | |
5586 | { | |
5587 | struct tx_port *tx_port_cached; | |
5588 | ||
c71ea3c4 IM |
5589 | /* Flush all the queued packets. */ |
5590 | dp_netdev_pmd_flush_output_packets(pmd, true); | |
324c8374 | 5591 | /* Free all used tx queue ids. */ |
b010be17 | 5592 | dpif_netdev_xps_revalidate_pmd(pmd, true); |
324c8374 | 5593 | |
57eebbb4 DDP |
5594 | HMAP_FOR_EACH_POP (tx_port_cached, node, &pmd->tnl_port_cache) { |
5595 | free(tx_port_cached); | |
5596 | } | |
5597 | HMAP_FOR_EACH_POP (tx_port_cached, node, &pmd->send_port_cache) { | |
d0cca6c3 DDP |
5598 | free(tx_port_cached); |
5599 | } | |
5600 | } | |
5601 | ||
5602 | /* Copies ports from 'pmd->tx_ports' (shared with the main thread) to | |
899363ed BB |
5603 | * thread-local copies. Copy to 'pmd->tnl_port_cache' if it is a tunnel |
5604 | * device, otherwise to 'pmd->send_port_cache' if the port has at least | |
5605 | * one txq. */ | |
d0cca6c3 DDP |
5606 | static void |
5607 | pmd_load_cached_ports(struct dp_netdev_pmd_thread *pmd) | |
5608 | OVS_REQUIRES(pmd->port_mutex) | |
5609 | { | |
5610 | struct tx_port *tx_port, *tx_port_cached; | |
5611 | ||
5612 | pmd_free_cached_ports(pmd); | |
57eebbb4 DDP |
5613 | hmap_shrink(&pmd->send_port_cache); |
5614 | hmap_shrink(&pmd->tnl_port_cache); | |
d0cca6c3 DDP |
5615 | |
5616 | HMAP_FOR_EACH (tx_port, node, &pmd->tx_ports) { | |
57eebbb4 DDP |
5617 | if (netdev_has_tunnel_push_pop(tx_port->port->netdev)) { |
5618 | tx_port_cached = xmemdup(tx_port, sizeof *tx_port_cached); | |
5619 | hmap_insert(&pmd->tnl_port_cache, &tx_port_cached->node, | |
5620 | hash_port_no(tx_port_cached->port->port_no)); | |
5621 | } | |
5622 | ||
5623 | if (netdev_n_txq(tx_port->port->netdev)) { | |
5624 | tx_port_cached = xmemdup(tx_port, sizeof *tx_port_cached); | |
5625 | hmap_insert(&pmd->send_port_cache, &tx_port_cached->node, | |
5626 | hash_port_no(tx_port_cached->port->port_no)); | |
5627 | } | |
d0cca6c3 DDP |
5628 | } |
5629 | } | |
5630 | ||
140dd699 IM |
5631 | static void |
5632 | pmd_alloc_static_tx_qid(struct dp_netdev_pmd_thread *pmd) | |
5633 | { | |
5634 | ovs_mutex_lock(&pmd->dp->tx_qid_pool_mutex); | |
5635 | if (!id_pool_alloc_id(pmd->dp->tx_qid_pool, &pmd->static_tx_qid)) { | |
5636 | VLOG_ABORT("static_tx_qid allocation failed for PMD on core %2d" | |
5637 | ", numa_id %d.", pmd->core_id, pmd->numa_id); | |
5638 | } | |
5639 | ovs_mutex_unlock(&pmd->dp->tx_qid_pool_mutex); | |
5640 | ||
5641 | VLOG_DBG("static_tx_qid = %d allocated for PMD thread on core %2d" | |
5642 | ", numa_id %d.", pmd->static_tx_qid, pmd->core_id, pmd->numa_id); | |
5643 | } | |
5644 | ||
5645 | static void | |
5646 | pmd_free_static_tx_qid(struct dp_netdev_pmd_thread *pmd) | |
5647 | { | |
5648 | ovs_mutex_lock(&pmd->dp->tx_qid_pool_mutex); | |
5649 | id_pool_free_id(pmd->dp->tx_qid_pool, pmd->static_tx_qid); | |
5650 | ovs_mutex_unlock(&pmd->dp->tx_qid_pool_mutex); | |
5651 | } | |
5652 | ||
e4cfed38 | 5653 | static int |
d0cca6c3 | 5654 | pmd_load_queues_and_ports(struct dp_netdev_pmd_thread *pmd, |
947dc567 | 5655 | struct polled_queue **ppoll_list) |
e4cfed38 | 5656 | { |
947dc567 | 5657 | struct polled_queue *poll_list = *ppoll_list; |
ae7ad0a1 IM |
5658 | struct rxq_poll *poll; |
5659 | int i; | |
e4cfed38 | 5660 | |
d0cca6c3 | 5661 | ovs_mutex_lock(&pmd->port_mutex); |
947dc567 DDP |
5662 | poll_list = xrealloc(poll_list, hmap_count(&pmd->poll_list) |
5663 | * sizeof *poll_list); | |
a1fdee13 | 5664 | |
ae7ad0a1 | 5665 | i = 0; |
947dc567 | 5666 | HMAP_FOR_EACH (poll, node, &pmd->poll_list) { |
922b28d4 | 5667 | poll_list[i].rxq = poll->rxq; |
947dc567 | 5668 | poll_list[i].port_no = poll->rxq->port->port_no; |
2fbadeb6 | 5669 | poll_list[i].emc_enabled = poll->rxq->port->emc_enabled; |
35c91567 DM |
5670 | poll_list[i].rxq_enabled = netdev_rxq_enabled(poll->rxq->rx); |
5671 | poll_list[i].change_seq = | |
5672 | netdev_get_change_seq(poll->rxq->port->netdev); | |
947dc567 | 5673 | i++; |
e4cfed38 | 5674 | } |
d0cca6c3 DDP |
5675 | |
5676 | pmd_load_cached_ports(pmd); | |
5677 | ||
5678 | ovs_mutex_unlock(&pmd->port_mutex); | |
e4cfed38 | 5679 | |
e4cfed38 | 5680 | *ppoll_list = poll_list; |
d42f9307 | 5681 | return i; |
e4cfed38 PS |
5682 | } |
5683 | ||
6c3eee82 | 5684 | static void * |
e4cfed38 | 5685 | pmd_thread_main(void *f_) |
6c3eee82 | 5686 | { |
65f13b50 | 5687 | struct dp_netdev_pmd_thread *pmd = f_; |
82a48ead | 5688 | struct pmd_perf_stats *s = &pmd->perf_stats; |
e4cfed38 | 5689 | unsigned int lc = 0; |
947dc567 | 5690 | struct polled_queue *poll_list; |
6d9fead1 | 5691 | bool wait_for_reload = false; |
e2cafa86 | 5692 | bool reload_tx_qid; |
d42f9307 | 5693 | bool exiting; |
6d9fead1 | 5694 | bool reload; |
e4cfed38 PS |
5695 | int poll_cnt; |
5696 | int i; | |
a2ac666d | 5697 | int process_packets = 0; |
6c3eee82 | 5698 | |
e4cfed38 PS |
5699 | poll_list = NULL; |
5700 | ||
65f13b50 AW |
5701 | /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */ |
5702 | ovsthread_setspecific(pmd->dp->per_pmd_key, pmd); | |
6930c7e0 DDP |
5703 | ovs_numa_thread_setaffinity_core(pmd->core_id); |
5704 | dpdk_set_lcore_id(pmd->core_id); | |
d0cca6c3 | 5705 | poll_cnt = pmd_load_queues_and_ports(pmd, &poll_list); |
60d8ccae | 5706 | dfc_cache_init(&pmd->flow_cache); |
140dd699 | 5707 | pmd_alloc_static_tx_qid(pmd); |
ae7ad0a1 | 5708 | |
e2cafa86 | 5709 | reload: |
5bf84282 NK |
5710 | atomic_count_init(&pmd->pmd_overloaded, 0); |
5711 | ||
7dd671f0 MK |
5712 | /* List port/core affinity */ |
5713 | for (i = 0; i < poll_cnt; i++) { | |
ce179f11 | 5714 | VLOG_DBG("Core %d processing port \'%s\' with queue-id %d\n", |
922b28d4 KT |
5715 | pmd->core_id, netdev_rxq_get_name(poll_list[i].rxq->rx), |
5716 | netdev_rxq_get_queue_id(poll_list[i].rxq->rx)); | |
4f5d13e2 KT |
5717 | /* Reset the rxq current cycles counter. */ |
5718 | dp_netdev_rxq_set_cycles(poll_list[i].rxq, RXQ_CYCLES_PROC_CURR, 0); | |
7dd671f0 MK |
5719 | } |
5720 | ||
2788a1b1 | 5721 | if (!poll_cnt) { |
6d9fead1 DM |
5722 | if (wait_for_reload) { |
5723 | /* Don't sleep, control thread will ask for a reload shortly. */ | |
5724 | do { | |
5725 | atomic_read_explicit(&pmd->reload, &reload, | |
5726 | memory_order_acquire); | |
5727 | } while (!reload); | |
5728 | } else { | |
5729 | while (seq_read(pmd->reload_seq) == pmd->last_reload_seq) { | |
5730 | seq_wait(pmd->reload_seq, pmd->last_reload_seq); | |
5731 | poll_block(); | |
5732 | } | |
2788a1b1 | 5733 | } |
2788a1b1 DDP |
5734 | } |
5735 | ||
2a2c67b4 KT |
5736 | pmd->intrvl_tsc_prev = 0; |
5737 | atomic_store_relaxed(&pmd->intrvl_cycles, 0); | |
a19896ab | 5738 | cycles_counter_update(s); |
79f36875 JS |
5739 | /* Protect pmd stats from external clearing while polling. */ |
5740 | ovs_mutex_lock(&pmd->perf_stats.stats_mutex); | |
e4cfed38 | 5741 | for (;;) { |
79f36875 | 5742 | uint64_t rx_packets = 0, tx_packets = 0; |
c71ea3c4 | 5743 | |
a19896ab | 5744 | pmd_perf_start_iteration(s); |
79f36875 | 5745 | |
e4cfed38 | 5746 | for (i = 0; i < poll_cnt; i++) { |
2fbadeb6 | 5747 | |
35c91567 DM |
5748 | if (!poll_list[i].rxq_enabled) { |
5749 | continue; | |
5750 | } | |
5751 | ||
2fbadeb6 IM |
5752 | if (poll_list[i].emc_enabled) { |
5753 | atomic_read_relaxed(&pmd->dp->emc_insert_min, | |
5754 | &pmd->ctx.emc_insert_min); | |
5755 | } else { | |
5756 | pmd->ctx.emc_insert_min = 0; | |
5757 | } | |
5758 | ||
a2ac666d | 5759 | process_packets = |
a19896ab | 5760 | dp_netdev_process_rxq_port(pmd, poll_list[i].rxq, |
a2ac666d | 5761 | poll_list[i].port_no); |
79f36875 | 5762 | rx_packets += process_packets; |
e4cfed38 PS |
5763 | } |
5764 | ||
79f36875 | 5765 | if (!rx_packets) { |
c71ea3c4 IM |
5766 | /* We didn't receive anything in the process loop. |
5767 | * Check if we need to send something. | |
5768 | * There was no time updates on current iteration. */ | |
5769 | pmd_thread_ctx_time_update(pmd); | |
79f36875 | 5770 | tx_packets = dp_netdev_pmd_flush_output_packets(pmd, false); |
c71ea3c4 IM |
5771 | } |
5772 | ||
e4cfed38 | 5773 | if (lc++ > 1024) { |
e4cfed38 | 5774 | lc = 0; |
84067a4c | 5775 | |
fbe0962b | 5776 | coverage_try_clear(); |
4809891b | 5777 | dp_netdev_pmd_try_optimize(pmd, poll_list, poll_cnt); |
9dede5cf | 5778 | if (!ovsrcu_try_quiesce()) { |
60d8ccae | 5779 | emc_cache_slow_sweep(&((pmd->flow_cache).emc_cache)); |
9dede5cf | 5780 | } |
84067a4c | 5781 | |
35c91567 DM |
5782 | for (i = 0; i < poll_cnt; i++) { |
5783 | uint64_t current_seq = | |
5784 | netdev_get_change_seq(poll_list[i].rxq->port->netdev); | |
5785 | if (poll_list[i].change_seq != current_seq) { | |
5786 | poll_list[i].change_seq = current_seq; | |
5787 | poll_list[i].rxq_enabled = | |
5788 | netdev_rxq_enabled(poll_list[i].rxq->rx); | |
5789 | } | |
5790 | } | |
6c3eee82 | 5791 | } |
68a0625b DM |
5792 | |
5793 | atomic_read_explicit(&pmd->reload, &reload, memory_order_acquire); | |
5794 | if (OVS_UNLIKELY(reload)) { | |
5795 | break; | |
5796 | } | |
5797 | ||
79f36875 JS |
5798 | pmd_perf_end_iteration(s, rx_packets, tx_packets, |
5799 | pmd_perf_metrics_enabled(pmd)); | |
e4cfed38 | 5800 | } |
79f36875 | 5801 | ovs_mutex_unlock(&pmd->perf_stats.stats_mutex); |
6c3eee82 | 5802 | |
d0cca6c3 | 5803 | poll_cnt = pmd_load_queues_and_ports(pmd, &poll_list); |
6d9fead1 | 5804 | atomic_read_relaxed(&pmd->wait_for_reload, &wait_for_reload); |
e2cafa86 | 5805 | atomic_read_relaxed(&pmd->reload_tx_qid, &reload_tx_qid); |
299c8d61 | 5806 | atomic_read_relaxed(&pmd->exit, &exiting); |
d42f9307 DDP |
5807 | /* Signal here to make sure the pmd finishes |
5808 | * reloading the updated configuration. */ | |
5809 | dp_netdev_pmd_reload_done(pmd); | |
5810 | ||
e2cafa86 DM |
5811 | if (reload_tx_qid) { |
5812 | pmd_free_static_tx_qid(pmd); | |
5813 | pmd_alloc_static_tx_qid(pmd); | |
5814 | } | |
9bbf1c3d | 5815 | |
d42f9307 | 5816 | if (!exiting) { |
e4cfed38 PS |
5817 | goto reload; |
5818 | } | |
6c3eee82 | 5819 | |
e2cafa86 | 5820 | pmd_free_static_tx_qid(pmd); |
60d8ccae | 5821 | dfc_cache_uninit(&pmd->flow_cache); |
e4cfed38 | 5822 | free(poll_list); |
d0cca6c3 | 5823 | pmd_free_cached_ports(pmd); |
6c3eee82 BP |
5824 | return NULL; |
5825 | } | |
5826 | ||
6b31e073 RW |
5827 | static void |
5828 | dp_netdev_disable_upcall(struct dp_netdev *dp) | |
5829 | OVS_ACQUIRES(dp->upcall_rwlock) | |
5830 | { | |
5831 | fat_rwlock_wrlock(&dp->upcall_rwlock); | |
5832 | } | |
5833 | ||
5dddf960 JR |
5834 | \f |
5835 | /* Meters */ | |
5836 | static void | |
5837 | dpif_netdev_meter_get_features(const struct dpif * dpif OVS_UNUSED, | |
5838 | struct ofputil_meter_features *features) | |
5839 | { | |
4b27db64 JR |
5840 | features->max_meters = MAX_METERS; |
5841 | features->band_types = DP_SUPPORTED_METER_BAND_TYPES; | |
5842 | features->capabilities = DP_SUPPORTED_METER_FLAGS_MASK; | |
5843 | features->max_bands = MAX_BANDS; | |
5dddf960 JR |
5844 | features->max_color = 0; |
5845 | } | |
5846 | ||
425a7b9e JP |
5847 | /* Applies the meter identified by 'meter_id' to 'packets_'. Packets |
5848 | * that exceed a band are dropped in-place. */ | |
4b27db64 JR |
5849 | static void |
5850 | dp_netdev_run_meter(struct dp_netdev *dp, struct dp_packet_batch *packets_, | |
5851 | uint32_t meter_id, long long int now) | |
5852 | { | |
5853 | struct dp_meter *meter; | |
5854 | struct dp_meter_band *band; | |
79c81260 | 5855 | struct dp_packet *packet; |
4b27db64 JR |
5856 | long long int long_delta_t; /* msec */ |
5857 | uint32_t delta_t; /* msec */ | |
5c41c31e | 5858 | uint32_t delta_in_us; /* usec */ |
79c81260 | 5859 | const size_t cnt = dp_packet_batch_size(packets_); |
4b27db64 JR |
5860 | uint32_t bytes, volume; |
5861 | int exceeded_band[NETDEV_MAX_BURST]; | |
5862 | uint32_t exceeded_rate[NETDEV_MAX_BURST]; | |
5863 | int exceeded_pkt = cnt; /* First packet that exceeded a band rate. */ | |
5864 | ||
5865 | if (meter_id >= MAX_METERS) { | |
5866 | return; | |
5867 | } | |
5868 | ||
5869 | meter_lock(dp, meter_id); | |
5870 | meter = dp->meters[meter_id]; | |
5871 | if (!meter) { | |
5872 | goto out; | |
5873 | } | |
5874 | ||
5875 | /* Initialize as negative values. */ | |
5876 | memset(exceeded_band, 0xff, cnt * sizeof *exceeded_band); | |
5877 | /* Initialize as zeroes. */ | |
5878 | memset(exceeded_rate, 0, cnt * sizeof *exceeded_rate); | |
5879 | ||
5880 | /* All packets will hit the meter at the same time. */ | |
42697ca7 | 5881 | long_delta_t = now / 1000 - meter->used / 1000; /* msec */ |
4b27db64 | 5882 | |
acc5df0e IM |
5883 | if (long_delta_t < 0) { |
5884 | /* This condition means that we have several threads fighting for a | |
5885 | meter lock, and the one who received the packets a bit later wins. | |
5886 | Assuming that all racing threads received packets at the same time | |
5887 | to avoid overflow. */ | |
5888 | long_delta_t = 0; | |
5c41c31e JL |
5889 | delta_in_us = 0; |
5890 | } else { | |
5891 | delta_in_us = (now - meter->used) % 1000; | |
acc5df0e IM |
5892 | } |
5893 | ||
4b27db64 JR |
5894 | /* Make sure delta_t will not be too large, so that bucket will not |
5895 | * wrap around below. */ | |
5896 | delta_t = (long_delta_t > (long long int)meter->max_delta_t) | |
5897 | ? meter->max_delta_t : (uint32_t)long_delta_t; | |
5898 | ||
5899 | /* Update meter stats. */ | |
5900 | meter->used = now; | |
5901 | meter->packet_count += cnt; | |
5902 | bytes = 0; | |
e883448e | 5903 | DP_PACKET_BATCH_FOR_EACH (i, packet, packets_) { |
79c81260 | 5904 | bytes += dp_packet_size(packet); |
4b27db64 JR |
5905 | } |
5906 | meter->byte_count += bytes; | |
5907 | ||
5908 | /* Meters can operate in terms of packets per second or kilobits per | |
5909 | * second. */ | |
5910 | if (meter->flags & OFPMF13_PKTPS) { | |
5911 | /* Rate in packets/second, bucket 1/1000 packets. */ | |
5912 | /* msec * packets/sec = 1/1000 packets. */ | |
5913 | volume = cnt * 1000; /* Take 'cnt' packets from the bucket. */ | |
5914 | } else { | |
5915 | /* Rate in kbps, bucket in bits. */ | |
5916 | /* msec * kbps = bits */ | |
5917 | volume = bytes * 8; | |
5918 | } | |
5919 | ||
5920 | /* Update all bands and find the one hit with the highest rate for each | |
5921 | * packet (if any). */ | |
5922 | for (int m = 0; m < meter->n_bands; ++m) { | |
5923 | band = &meter->bands[m]; | |
5924 | ||
5925 | /* Update band's bucket. */ | |
5926 | band->bucket += delta_t * band->up.rate; | |
5c41c31e | 5927 | band->bucket += delta_in_us * band->up.rate / 1000; |
4b27db64 JR |
5928 | if (band->bucket > band->up.burst_size) { |
5929 | band->bucket = band->up.burst_size; | |
5930 | } | |
5931 | ||
5932 | /* Drain the bucket for all the packets, if possible. */ | |
5933 | if (band->bucket >= volume) { | |
5934 | band->bucket -= volume; | |
5935 | } else { | |
5936 | int band_exceeded_pkt; | |
5937 | ||
5938 | /* Band limit hit, must process packet-by-packet. */ | |
5939 | if (meter->flags & OFPMF13_PKTPS) { | |
5940 | band_exceeded_pkt = band->bucket / 1000; | |
5941 | band->bucket %= 1000; /* Remainder stays in bucket. */ | |
5942 | ||
5943 | /* Update the exceeding band for each exceeding packet. | |
5944 | * (Only one band will be fired by a packet, and that | |
5945 | * can be different for each packet.) */ | |
e883448e | 5946 | for (int i = band_exceeded_pkt; i < cnt; i++) { |
4b27db64 JR |
5947 | if (band->up.rate > exceeded_rate[i]) { |
5948 | exceeded_rate[i] = band->up.rate; | |
5949 | exceeded_band[i] = m; | |
5950 | } | |
5951 | } | |
5952 | } else { | |
5953 | /* Packet sizes differ, must process one-by-one. */ | |
5954 | band_exceeded_pkt = cnt; | |
e883448e | 5955 | DP_PACKET_BATCH_FOR_EACH (i, packet, packets_) { |
79c81260 | 5956 | uint32_t bits = dp_packet_size(packet) * 8; |
4b27db64 JR |
5957 | |
5958 | if (band->bucket >= bits) { | |
5959 | band->bucket -= bits; | |
5960 | } else { | |
5961 | if (i < band_exceeded_pkt) { | |
5962 | band_exceeded_pkt = i; | |
5963 | } | |
5964 | /* Update the exceeding band for the exceeding packet. | |
5965 | * (Only one band will be fired by a packet, and that | |
5966 | * can be different for each packet.) */ | |
5967 | if (band->up.rate > exceeded_rate[i]) { | |
5968 | exceeded_rate[i] = band->up.rate; | |
5969 | exceeded_band[i] = m; | |
5970 | } | |
5971 | } | |
5972 | } | |
5973 | } | |
5974 | /* Remember the first exceeding packet. */ | |
5975 | if (exceeded_pkt > band_exceeded_pkt) { | |
5976 | exceeded_pkt = band_exceeded_pkt; | |
5977 | } | |
5978 | } | |
5979 | } | |
5980 | ||
425a7b9e JP |
5981 | /* Fire the highest rate band exceeded by each packet, and drop |
5982 | * packets if needed. */ | |
4b27db64 | 5983 | size_t j; |
79c81260 | 5984 | DP_PACKET_BATCH_REFILL_FOR_EACH (j, cnt, packet, packets_) { |
4b27db64 JR |
5985 | if (exceeded_band[j] >= 0) { |
5986 | /* Meter drop packet. */ | |
5987 | band = &meter->bands[exceeded_band[j]]; | |
5988 | band->packet_count += 1; | |
5989 | band->byte_count += dp_packet_size(packet); | |
a13a0209 | 5990 | COVERAGE_INC(datapath_drop_meter); |
4b27db64 JR |
5991 | dp_packet_delete(packet); |
5992 | } else { | |
5993 | /* Meter accepts packet. */ | |
5994 | dp_packet_batch_refill(packets_, packet, j); | |
5995 | } | |
5996 | } | |
5997 | out: | |
5998 | meter_unlock(dp, meter_id); | |
5999 | } | |
6000 | ||
6001 | /* Meter set/get/del processing is still single-threaded. */ | |
5dddf960 | 6002 | static int |
8101f03f | 6003 | dpif_netdev_meter_set(struct dpif *dpif, ofproto_meter_id meter_id, |
4b27db64 | 6004 | struct ofputil_meter_config *config) |
5dddf960 | 6005 | { |
4b27db64 | 6006 | struct dp_netdev *dp = get_dp_netdev(dpif); |
8101f03f | 6007 | uint32_t mid = meter_id.uint32; |
4b27db64 JR |
6008 | struct dp_meter *meter; |
6009 | int i; | |
6010 | ||
4b27db64 JR |
6011 | if (mid >= MAX_METERS) { |
6012 | return EFBIG; /* Meter_id out of range. */ | |
6013 | } | |
6014 | ||
6508c845 | 6015 | if (config->flags & ~DP_SUPPORTED_METER_FLAGS_MASK) { |
4b27db64 JR |
6016 | return EBADF; /* Unsupported flags set */ |
6017 | } | |
2029ce9a | 6018 | |
6508c845 JP |
6019 | if (config->n_bands > MAX_BANDS) { |
6020 | return EINVAL; | |
2029ce9a AVA |
6021 | } |
6022 | ||
4b27db64 JR |
6023 | for (i = 0; i < config->n_bands; ++i) { |
6024 | switch (config->bands[i].type) { | |
6025 | case OFPMBT13_DROP: | |
6026 | break; | |
6027 | default: | |
6028 | return ENODEV; /* Unsupported band type */ | |
6029 | } | |
6030 | } | |
6031 | ||
6032 | /* Allocate meter */ | |
6033 | meter = xzalloc(sizeof *meter | |
6034 | + config->n_bands * sizeof(struct dp_meter_band)); | |
4b27db64 | 6035 | |
d0db81ea JP |
6036 | meter->flags = config->flags; |
6037 | meter->n_bands = config->n_bands; | |
6038 | meter->max_delta_t = 0; | |
6039 | meter->used = time_usec(); | |
4b27db64 | 6040 | |
d0db81ea JP |
6041 | /* set up bands */ |
6042 | for (i = 0; i < config->n_bands; ++i) { | |
6043 | uint32_t band_max_delta_t; | |
4b27db64 | 6044 | |
d0db81ea JP |
6045 | /* Set burst size to a workable value if none specified. */ |
6046 | if (config->bands[i].burst_size == 0) { | |
6047 | config->bands[i].burst_size = config->bands[i].rate; | |
6048 | } | |
6049 | ||
6050 | meter->bands[i].up = config->bands[i]; | |
6051 | /* Convert burst size to the bucket units: */ | |
6052 | /* pkts => 1/1000 packets, kilobits => bits. */ | |
6053 | meter->bands[i].up.burst_size *= 1000; | |
6054 | /* Initialize bucket to empty. */ | |
6055 | meter->bands[i].bucket = 0; | |
6056 | ||
6057 | /* Figure out max delta_t that is enough to fill any bucket. */ | |
6058 | band_max_delta_t | |
6059 | = meter->bands[i].up.burst_size / meter->bands[i].up.rate; | |
6060 | if (band_max_delta_t > meter->max_delta_t) { | |
6061 | meter->max_delta_t = band_max_delta_t; | |
6062 | } | |
4b27db64 | 6063 | } |
d0db81ea JP |
6064 | |
6065 | meter_lock(dp, mid); | |
6066 | dp_delete_meter(dp, mid); /* Free existing meter, if any */ | |
6067 | dp->meters[mid] = meter; | |
6068 | meter_unlock(dp, mid); | |
6069 | ||
6070 | return 0; | |
5dddf960 JR |
6071 | } |
6072 | ||
6073 | static int | |
4b27db64 JR |
6074 | dpif_netdev_meter_get(const struct dpif *dpif, |
6075 | ofproto_meter_id meter_id_, | |
6076 | struct ofputil_meter_stats *stats, uint16_t n_bands) | |
5dddf960 | 6077 | { |
4b27db64 | 6078 | const struct dp_netdev *dp = get_dp_netdev(dpif); |
4b27db64 | 6079 | uint32_t meter_id = meter_id_.uint32; |
866bc756 | 6080 | int retval = 0; |
4b27db64 JR |
6081 | |
6082 | if (meter_id >= MAX_METERS) { | |
6083 | return EFBIG; | |
6084 | } | |
866bc756 JP |
6085 | |
6086 | meter_lock(dp, meter_id); | |
6087 | const struct dp_meter *meter = dp->meters[meter_id]; | |
4b27db64 | 6088 | if (!meter) { |
866bc756 JP |
6089 | retval = ENOENT; |
6090 | goto done; | |
4b27db64 JR |
6091 | } |
6092 | if (stats) { | |
6093 | int i = 0; | |
6094 | ||
4b27db64 JR |
6095 | stats->packet_in_count = meter->packet_count; |
6096 | stats->byte_in_count = meter->byte_count; | |
6097 | ||
6098 | for (i = 0; i < n_bands && i < meter->n_bands; ++i) { | |
6099 | stats->bands[i].packet_count = meter->bands[i].packet_count; | |
6100 | stats->bands[i].byte_count = meter->bands[i].byte_count; | |
6101 | } | |
4b27db64 JR |
6102 | |
6103 | stats->n_bands = i; | |
6104 | } | |
866bc756 JP |
6105 | |
6106 | done: | |
6107 | meter_unlock(dp, meter_id); | |
6108 | return retval; | |
5dddf960 JR |
6109 | } |
6110 | ||
6111 | static int | |
4b27db64 JR |
6112 | dpif_netdev_meter_del(struct dpif *dpif, |
6113 | ofproto_meter_id meter_id_, | |
6114 | struct ofputil_meter_stats *stats, uint16_t n_bands) | |
5dddf960 | 6115 | { |
4b27db64 JR |
6116 | struct dp_netdev *dp = get_dp_netdev(dpif); |
6117 | int error; | |
6118 | ||
6119 | error = dpif_netdev_meter_get(dpif, meter_id_, stats, n_bands); | |
6120 | if (!error) { | |
6121 | uint32_t meter_id = meter_id_.uint32; | |
6122 | ||
6123 | meter_lock(dp, meter_id); | |
6124 | dp_delete_meter(dp, meter_id); | |
6125 | meter_unlock(dp, meter_id); | |
4b27db64 JR |
6126 | } |
6127 | return error; | |
5dddf960 JR |
6128 | } |
6129 | ||
6130 | \f | |
6b31e073 RW |
6131 | static void |
6132 | dpif_netdev_disable_upcall(struct dpif *dpif) | |
6133 | OVS_NO_THREAD_SAFETY_ANALYSIS | |
6134 | { | |
6135 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
6136 | dp_netdev_disable_upcall(dp); | |
6137 | } | |
6138 | ||
6139 | static void | |
6140 | dp_netdev_enable_upcall(struct dp_netdev *dp) | |
6141 | OVS_RELEASES(dp->upcall_rwlock) | |
6142 | { | |
6143 | fat_rwlock_unlock(&dp->upcall_rwlock); | |
6144 | } | |
6145 | ||
6146 | static void | |
6147 | dpif_netdev_enable_upcall(struct dpif *dpif) | |
6148 | OVS_NO_THREAD_SAFETY_ANALYSIS | |
6149 | { | |
6150 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
6151 | dp_netdev_enable_upcall(dp); | |
6152 | } | |
6153 | ||
ae7ad0a1 | 6154 | static void |
accf8626 AW |
6155 | dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd) |
6156 | { | |
6d9fead1 | 6157 | atomic_store_relaxed(&pmd->wait_for_reload, false); |
e2cafa86 | 6158 | atomic_store_relaxed(&pmd->reload_tx_qid, false); |
2788a1b1 | 6159 | pmd->last_reload_seq = seq_read(pmd->reload_seq); |
8f077b31 | 6160 | atomic_store_explicit(&pmd->reload, false, memory_order_release); |
accf8626 AW |
6161 | } |
6162 | ||
1c1e46ed | 6163 | /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns |
546e57d4 DDP |
6164 | * the pointer if succeeds, otherwise, NULL (it can return NULL even if |
6165 | * 'core_id' is NON_PMD_CORE_ID). | |
1c1e46ed AW |
6166 | * |
6167 | * Caller must unrefs the returned reference. */ | |
65f13b50 | 6168 | static struct dp_netdev_pmd_thread * |
bd5131ba | 6169 | dp_netdev_get_pmd(struct dp_netdev *dp, unsigned core_id) |
65f13b50 AW |
6170 | { |
6171 | struct dp_netdev_pmd_thread *pmd; | |
55847abe | 6172 | const struct cmap_node *pnode; |
65f13b50 | 6173 | |
b19befae | 6174 | pnode = cmap_find(&dp->poll_threads, hash_int(core_id, 0)); |
1c1e46ed AW |
6175 | if (!pnode) { |
6176 | return NULL; | |
6177 | } | |
65f13b50 AW |
6178 | pmd = CONTAINER_OF(pnode, struct dp_netdev_pmd_thread, node); |
6179 | ||
1c1e46ed | 6180 | return dp_netdev_pmd_try_ref(pmd) ? pmd : NULL; |
65f13b50 AW |
6181 | } |
6182 | ||
f2eee189 AW |
6183 | /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */ |
6184 | static void | |
6185 | dp_netdev_set_nonpmd(struct dp_netdev *dp) | |
e9985d6a | 6186 | OVS_REQUIRES(dp->port_mutex) |
f2eee189 AW |
6187 | { |
6188 | struct dp_netdev_pmd_thread *non_pmd; | |
6189 | ||
6190 | non_pmd = xzalloc(sizeof *non_pmd); | |
00873463 | 6191 | dp_netdev_configure_pmd(non_pmd, dp, NON_PMD_CORE_ID, OVS_NUMA_UNSPEC); |
f2eee189 AW |
6192 | } |
6193 | ||
1c1e46ed AW |
6194 | /* Caller must have valid pointer to 'pmd'. */ |
6195 | static bool | |
6196 | dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd) | |
6197 | { | |
6198 | return ovs_refcount_try_ref_rcu(&pmd->ref_cnt); | |
6199 | } | |
6200 | ||
6201 | static void | |
6202 | dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd) | |
6203 | { | |
6204 | if (pmd && ovs_refcount_unref(&pmd->ref_cnt) == 1) { | |
6205 | ovsrcu_postpone(dp_netdev_destroy_pmd, pmd); | |
6206 | } | |
6207 | } | |
6208 | ||
6209 | /* Given cmap position 'pos', tries to ref the next node. If try_ref() | |
6210 | * fails, keeps checking for next node until reaching the end of cmap. | |
6211 | * | |
6212 | * Caller must unrefs the returned reference. */ | |
6213 | static struct dp_netdev_pmd_thread * | |
6214 | dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos) | |
6215 | { | |
6216 | struct dp_netdev_pmd_thread *next; | |
6217 | ||
6218 | do { | |
6219 | struct cmap_node *node; | |
6220 | ||
6221 | node = cmap_next_position(&dp->poll_threads, pos); | |
6222 | next = node ? CONTAINER_OF(node, struct dp_netdev_pmd_thread, node) | |
6223 | : NULL; | |
6224 | } while (next && !dp_netdev_pmd_try_ref(next)); | |
6225 | ||
6226 | return next; | |
6227 | } | |
6228 | ||
65f13b50 | 6229 | /* Configures the 'pmd' based on the input argument. */ |
6c3eee82 | 6230 | static void |
65f13b50 | 6231 | dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp, |
00873463 | 6232 | unsigned core_id, int numa_id) |
65f13b50 AW |
6233 | { |
6234 | pmd->dp = dp; | |
65f13b50 AW |
6235 | pmd->core_id = core_id; |
6236 | pmd->numa_id = numa_id; | |
e32971b8 | 6237 | pmd->need_reload = false; |
c71ea3c4 | 6238 | pmd->n_output_batches = 0; |
1c1e46ed AW |
6239 | |
6240 | ovs_refcount_init(&pmd->ref_cnt); | |
299c8d61 | 6241 | atomic_init(&pmd->exit, false); |
2788a1b1 DDP |
6242 | pmd->reload_seq = seq_create(); |
6243 | pmd->last_reload_seq = seq_read(pmd->reload_seq); | |
14e3e12a | 6244 | atomic_init(&pmd->reload, false); |
1c1e46ed | 6245 | ovs_mutex_init(&pmd->flow_mutex); |
d0cca6c3 | 6246 | ovs_mutex_init(&pmd->port_mutex); |
9df65060 | 6247 | ovs_mutex_init(&pmd->bond_mutex); |
1c1e46ed | 6248 | cmap_init(&pmd->flow_table); |
3453b4d6 | 6249 | cmap_init(&pmd->classifiers); |
58ed6df0 | 6250 | pmd->ctx.last_rxq = NULL; |
b010be17 IM |
6251 | pmd_thread_ctx_time_update(pmd); |
6252 | pmd->next_optimization = pmd->ctx.now + DPCLS_OPTIMIZATION_INTERVAL; | |
6253 | pmd->rxq_next_cycle_store = pmd->ctx.now + PMD_RXQ_INTERVAL_LEN; | |
947dc567 | 6254 | hmap_init(&pmd->poll_list); |
d0cca6c3 | 6255 | hmap_init(&pmd->tx_ports); |
57eebbb4 DDP |
6256 | hmap_init(&pmd->tnl_port_cache); |
6257 | hmap_init(&pmd->send_port_cache); | |
9df65060 | 6258 | cmap_init(&pmd->tx_bonds); |
65f13b50 AW |
6259 | /* init the 'flow_cache' since there is no |
6260 | * actual thread created for NON_PMD_CORE_ID. */ | |
6261 | if (core_id == NON_PMD_CORE_ID) { | |
60d8ccae | 6262 | dfc_cache_init(&pmd->flow_cache); |
140dd699 | 6263 | pmd_alloc_static_tx_qid(pmd); |
65f13b50 | 6264 | } |
82a48ead | 6265 | pmd_perf_stats_init(&pmd->perf_stats); |
65f13b50 AW |
6266 | cmap_insert(&dp->poll_threads, CONST_CAST(struct cmap_node *, &pmd->node), |
6267 | hash_int(core_id, 0)); | |
6268 | } | |
6269 | ||
1c1e46ed AW |
6270 | static void |
6271 | dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd) | |
6272 | { | |
3453b4d6 JS |
6273 | struct dpcls *cls; |
6274 | ||
1c1e46ed | 6275 | dp_netdev_pmd_flow_flush(pmd); |
57eebbb4 DDP |
6276 | hmap_destroy(&pmd->send_port_cache); |
6277 | hmap_destroy(&pmd->tnl_port_cache); | |
d0cca6c3 | 6278 | hmap_destroy(&pmd->tx_ports); |
9df65060 | 6279 | cmap_destroy(&pmd->tx_bonds); |
947dc567 | 6280 | hmap_destroy(&pmd->poll_list); |
3453b4d6 JS |
6281 | /* All flows (including their dpcls_rules) have been deleted already */ |
6282 | CMAP_FOR_EACH (cls, node, &pmd->classifiers) { | |
6283 | dpcls_destroy(cls); | |
7c269972 | 6284 | ovsrcu_postpone(free, cls); |
3453b4d6 JS |
6285 | } |
6286 | cmap_destroy(&pmd->classifiers); | |
1c1e46ed AW |
6287 | cmap_destroy(&pmd->flow_table); |
6288 | ovs_mutex_destroy(&pmd->flow_mutex); | |
2788a1b1 | 6289 | seq_destroy(pmd->reload_seq); |
d0cca6c3 | 6290 | ovs_mutex_destroy(&pmd->port_mutex); |
9df65060 | 6291 | ovs_mutex_destroy(&pmd->bond_mutex); |
1c1e46ed AW |
6292 | free(pmd); |
6293 | } | |
6294 | ||
6295 | /* Stops the pmd thread, removes it from the 'dp->poll_threads', | |
6296 | * and unrefs the struct. */ | |
65f13b50 | 6297 | static void |
e4e74c3a | 6298 | dp_netdev_del_pmd(struct dp_netdev *dp, struct dp_netdev_pmd_thread *pmd) |
6c3eee82 | 6299 | { |
d0cca6c3 DDP |
6300 | /* NON_PMD_CORE_ID doesn't have a thread, so we don't have to synchronize, |
6301 | * but extra cleanup is necessary */ | |
65f13b50 | 6302 | if (pmd->core_id == NON_PMD_CORE_ID) { |
febf4a7a | 6303 | ovs_mutex_lock(&dp->non_pmd_mutex); |
60d8ccae | 6304 | dfc_cache_uninit(&pmd->flow_cache); |
d0cca6c3 | 6305 | pmd_free_cached_ports(pmd); |
140dd699 | 6306 | pmd_free_static_tx_qid(pmd); |
febf4a7a | 6307 | ovs_mutex_unlock(&dp->non_pmd_mutex); |
65f13b50 | 6308 | } else { |
299c8d61 | 6309 | atomic_store_relaxed(&pmd->exit, true); |
65f13b50 | 6310 | dp_netdev_reload_pmd__(pmd); |
65f13b50 AW |
6311 | xpthread_join(pmd->thread, NULL); |
6312 | } | |
ae7ad0a1 | 6313 | |
d0cca6c3 | 6314 | dp_netdev_pmd_clear_ports(pmd); |
ae7ad0a1 | 6315 | |
e4e74c3a AW |
6316 | /* Purges the 'pmd''s flows after stopping the thread, but before |
6317 | * destroying the flows, so that the flow stats can be collected. */ | |
6318 | if (dp->dp_purge_cb) { | |
6319 | dp->dp_purge_cb(dp->dp_purge_aux, pmd->core_id); | |
6320 | } | |
65f13b50 | 6321 | cmap_remove(&pmd->dp->poll_threads, &pmd->node, hash_int(pmd->core_id, 0)); |
1c1e46ed | 6322 | dp_netdev_pmd_unref(pmd); |
65f13b50 | 6323 | } |
6c3eee82 | 6324 | |
e32971b8 DDP |
6325 | /* Destroys all pmd threads. If 'non_pmd' is true it also destroys the non pmd |
6326 | * thread. */ | |
65f13b50 | 6327 | static void |
e32971b8 | 6328 | dp_netdev_destroy_all_pmds(struct dp_netdev *dp, bool non_pmd) |
65f13b50 AW |
6329 | { |
6330 | struct dp_netdev_pmd_thread *pmd; | |
d916785c DDP |
6331 | struct dp_netdev_pmd_thread **pmd_list; |
6332 | size_t k = 0, n_pmds; | |
6333 | ||
e32971b8 | 6334 | n_pmds = cmap_count(&dp->poll_threads); |
d916785c | 6335 | pmd_list = xcalloc(n_pmds, sizeof *pmd_list); |
65f13b50 AW |
6336 | |
6337 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
e32971b8 | 6338 | if (!non_pmd && pmd->core_id == NON_PMD_CORE_ID) { |
b9584f21 DDP |
6339 | continue; |
6340 | } | |
d916785c DDP |
6341 | /* We cannot call dp_netdev_del_pmd(), since it alters |
6342 | * 'dp->poll_threads' (while we're iterating it) and it | |
6343 | * might quiesce. */ | |
6344 | ovs_assert(k < n_pmds); | |
6345 | pmd_list[k++] = pmd; | |
6c3eee82 | 6346 | } |
d916785c DDP |
6347 | |
6348 | for (size_t i = 0; i < k; i++) { | |
6349 | dp_netdev_del_pmd(dp, pmd_list[i]); | |
6350 | } | |
6351 | free(pmd_list); | |
65f13b50 | 6352 | } |
6c3eee82 | 6353 | |
d0cca6c3 DDP |
6354 | /* Deletes all rx queues from pmd->poll_list and all the ports from |
6355 | * pmd->tx_ports. */ | |
cc245ce8 | 6356 | static void |
d0cca6c3 | 6357 | dp_netdev_pmd_clear_ports(struct dp_netdev_pmd_thread *pmd) |
cc245ce8 IM |
6358 | { |
6359 | struct rxq_poll *poll; | |
d0cca6c3 | 6360 | struct tx_port *port; |
9df65060 | 6361 | struct tx_bond *tx; |
cc245ce8 | 6362 | |
d0cca6c3 | 6363 | ovs_mutex_lock(&pmd->port_mutex); |
947dc567 | 6364 | HMAP_FOR_EACH_POP (poll, node, &pmd->poll_list) { |
cc245ce8 IM |
6365 | free(poll); |
6366 | } | |
d0cca6c3 DDP |
6367 | HMAP_FOR_EACH_POP (port, node, &pmd->tx_ports) { |
6368 | free(port); | |
6369 | } | |
6370 | ovs_mutex_unlock(&pmd->port_mutex); | |
9df65060 VDA |
6371 | |
6372 | ovs_mutex_lock(&pmd->bond_mutex); | |
6373 | CMAP_FOR_EACH (tx, node, &pmd->tx_bonds) { | |
6374 | cmap_remove(&pmd->tx_bonds, &tx->node, hash_bond_id(tx->bond_id)); | |
6375 | ovsrcu_postpone(free, tx); | |
6376 | } | |
6377 | ovs_mutex_unlock(&pmd->bond_mutex); | |
cc245ce8 IM |
6378 | } |
6379 | ||
e32971b8 | 6380 | /* Adds rx queue to poll_list of PMD thread, if it's not there already. */ |
b68872d8 | 6381 | static void |
e32971b8 DDP |
6382 | dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread *pmd, |
6383 | struct dp_netdev_rxq *rxq) | |
6384 | OVS_REQUIRES(pmd->port_mutex) | |
b68872d8 | 6385 | { |
e32971b8 DDP |
6386 | int qid = netdev_rxq_get_queue_id(rxq->rx); |
6387 | uint32_t hash = hash_2words(odp_to_u32(rxq->port->port_no), qid); | |
6388 | struct rxq_poll *poll; | |
b68872d8 | 6389 | |
e32971b8 DDP |
6390 | HMAP_FOR_EACH_WITH_HASH (poll, node, hash, &pmd->poll_list) { |
6391 | if (poll->rxq == rxq) { | |
6392 | /* 'rxq' is already polled by this thread. Do nothing. */ | |
6393 | return; | |
d0cca6c3 | 6394 | } |
cc245ce8 | 6395 | } |
cc245ce8 | 6396 | |
e32971b8 DDP |
6397 | poll = xmalloc(sizeof *poll); |
6398 | poll->rxq = rxq; | |
6399 | hmap_insert(&pmd->poll_list, &poll->node, hash); | |
b68872d8 | 6400 | |
e32971b8 | 6401 | pmd->need_reload = true; |
ae7ad0a1 IM |
6402 | } |
6403 | ||
e32971b8 | 6404 | /* Delete 'poll' from poll_list of PMD thread. */ |
ae7ad0a1 | 6405 | static void |
e32971b8 DDP |
6406 | dp_netdev_del_rxq_from_pmd(struct dp_netdev_pmd_thread *pmd, |
6407 | struct rxq_poll *poll) | |
d0cca6c3 | 6408 | OVS_REQUIRES(pmd->port_mutex) |
ae7ad0a1 | 6409 | { |
e32971b8 DDP |
6410 | hmap_remove(&pmd->poll_list, &poll->node); |
6411 | free(poll); | |
ae7ad0a1 | 6412 | |
e32971b8 | 6413 | pmd->need_reload = true; |
ae7ad0a1 IM |
6414 | } |
6415 | ||
d0cca6c3 DDP |
6416 | /* Add 'port' to the tx port cache of 'pmd', which must be reloaded for the |
6417 | * changes to take effect. */ | |
cc245ce8 | 6418 | static void |
d0cca6c3 DDP |
6419 | dp_netdev_add_port_tx_to_pmd(struct dp_netdev_pmd_thread *pmd, |
6420 | struct dp_netdev_port *port) | |
e32971b8 | 6421 | OVS_REQUIRES(pmd->port_mutex) |
d0cca6c3 | 6422 | { |
57eebbb4 DDP |
6423 | struct tx_port *tx; |
6424 | ||
e32971b8 DDP |
6425 | tx = tx_port_lookup(&pmd->tx_ports, port->port_no); |
6426 | if (tx) { | |
6427 | /* 'port' is already on this thread tx cache. Do nothing. */ | |
6428 | return; | |
6429 | } | |
6430 | ||
57eebbb4 | 6431 | tx = xzalloc(sizeof *tx); |
d0cca6c3 | 6432 | |
324c8374 IM |
6433 | tx->port = port; |
6434 | tx->qid = -1; | |
c71ea3c4 | 6435 | tx->flush_time = 0LL; |
009e0033 | 6436 | dp_packet_batch_init(&tx->output_pkts); |
d0cca6c3 | 6437 | |
324c8374 | 6438 | hmap_insert(&pmd->tx_ports, &tx->node, hash_port_no(tx->port->port_no)); |
e32971b8 | 6439 | pmd->need_reload = true; |
d0cca6c3 DDP |
6440 | } |
6441 | ||
e32971b8 DDP |
6442 | /* Del 'tx' from the tx port cache of 'pmd', which must be reloaded for the |
6443 | * changes to take effect. */ | |
b9584f21 | 6444 | static void |
e32971b8 DDP |
6445 | dp_netdev_del_port_tx_from_pmd(struct dp_netdev_pmd_thread *pmd, |
6446 | struct tx_port *tx) | |
6447 | OVS_REQUIRES(pmd->port_mutex) | |
b9584f21 | 6448 | { |
e32971b8 DDP |
6449 | hmap_remove(&pmd->tx_ports, &tx->node); |
6450 | free(tx); | |
6451 | pmd->need_reload = true; | |
6c3eee82 | 6452 | } |
9df65060 VDA |
6453 | |
6454 | /* Add bond to the tx bond cmap of 'pmd'. */ | |
6455 | static void | |
6456 | dp_netdev_add_bond_tx_to_pmd(struct dp_netdev_pmd_thread *pmd, | |
6457 | struct tx_bond *bond, bool update) | |
6458 | OVS_EXCLUDED(pmd->bond_mutex) | |
6459 | { | |
6460 | struct tx_bond *tx; | |
6461 | ||
6462 | ovs_mutex_lock(&pmd->bond_mutex); | |
6463 | tx = tx_bond_lookup(&pmd->tx_bonds, bond->bond_id); | |
6464 | ||
6465 | if (tx && !update) { | |
6466 | /* It's not an update and the entry already exists. Do nothing. */ | |
6467 | goto unlock; | |
6468 | } | |
6469 | ||
6470 | if (tx) { | |
6471 | struct tx_bond *new_tx = xmemdup(bond, sizeof *bond); | |
6472 | ||
6473 | /* Copy the stats for each bucket. */ | |
6474 | for (int i = 0; i < BOND_BUCKETS; i++) { | |
6475 | uint64_t n_packets, n_bytes; | |
6476 | ||
6477 | atomic_read_relaxed(&tx->slave_buckets[i].n_packets, &n_packets); | |
6478 | atomic_read_relaxed(&tx->slave_buckets[i].n_bytes, &n_bytes); | |
6479 | atomic_init(&new_tx->slave_buckets[i].n_packets, n_packets); | |
6480 | atomic_init(&new_tx->slave_buckets[i].n_bytes, n_bytes); | |
6481 | } | |
6482 | cmap_replace(&pmd->tx_bonds, &tx->node, &new_tx->node, | |
6483 | hash_bond_id(bond->bond_id)); | |
6484 | ovsrcu_postpone(free, tx); | |
6485 | } else { | |
6486 | tx = xmemdup(bond, sizeof *bond); | |
6487 | cmap_insert(&pmd->tx_bonds, &tx->node, hash_bond_id(bond->bond_id)); | |
6488 | } | |
6489 | unlock: | |
6490 | ovs_mutex_unlock(&pmd->bond_mutex); | |
6491 | } | |
6492 | ||
6493 | /* Delete bond from the tx bond cmap of 'pmd'. */ | |
6494 | static void | |
6495 | dp_netdev_del_bond_tx_from_pmd(struct dp_netdev_pmd_thread *pmd, | |
6496 | uint32_t bond_id) | |
6497 | OVS_EXCLUDED(pmd->bond_mutex) | |
6498 | { | |
6499 | struct tx_bond *tx; | |
6500 | ||
6501 | ovs_mutex_lock(&pmd->bond_mutex); | |
6502 | tx = tx_bond_lookup(&pmd->tx_bonds, bond_id); | |
6503 | if (tx) { | |
6504 | cmap_remove(&pmd->tx_bonds, &tx->node, hash_bond_id(tx->bond_id)); | |
6505 | ovsrcu_postpone(free, tx); | |
6506 | } | |
6507 | ovs_mutex_unlock(&pmd->bond_mutex); | |
6508 | } | |
6c3eee82 | 6509 | \f |
b5cbbcf6 AZ |
6510 | static char * |
6511 | dpif_netdev_get_datapath_version(void) | |
6512 | { | |
6513 | return xstrdup("<built-in>"); | |
6514 | } | |
6515 | ||
72865317 | 6516 | static void |
1c1e46ed | 6517 | dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow, int cnt, int size, |
11bfdadd | 6518 | uint16_t tcp_flags, long long now) |
72865317 | 6519 | { |
eb94da30 | 6520 | uint16_t flags; |
72865317 | 6521 | |
eb94da30 DDP |
6522 | atomic_store_relaxed(&netdev_flow->stats.used, now); |
6523 | non_atomic_ullong_add(&netdev_flow->stats.packet_count, cnt); | |
6524 | non_atomic_ullong_add(&netdev_flow->stats.byte_count, size); | |
6525 | atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags); | |
6526 | flags |= tcp_flags; | |
6527 | atomic_store_relaxed(&netdev_flow->stats.tcp_flags, flags); | |
51852a57 BP |
6528 | } |
6529 | ||
623540e4 | 6530 | static int |
e14deea0 | 6531 | dp_netdev_upcall(struct dp_netdev_pmd_thread *pmd, struct dp_packet *packet_, |
7af12bd7 | 6532 | struct flow *flow, struct flow_wildcards *wc, ovs_u128 *ufid, |
623540e4 EJ |
6533 | enum dpif_upcall_type type, const struct nlattr *userdata, |
6534 | struct ofpbuf *actions, struct ofpbuf *put_actions) | |
6535 | { | |
1c1e46ed | 6536 | struct dp_netdev *dp = pmd->dp; |
623540e4 | 6537 | |
623540e4 EJ |
6538 | if (OVS_UNLIKELY(!dp->upcall_cb)) { |
6539 | return ENODEV; | |
6540 | } | |
6541 | ||
6542 | if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl))) { | |
6543 | struct ds ds = DS_EMPTY_INITIALIZER; | |
623540e4 | 6544 | char *packet_str; |
cf62fa4c | 6545 | struct ofpbuf key; |
5262eea1 JG |
6546 | struct odp_flow_key_parms odp_parms = { |
6547 | .flow = flow, | |
1dea1435 | 6548 | .mask = wc ? &wc->masks : NULL, |
2494ccd7 | 6549 | .support = dp_netdev_support, |
5262eea1 | 6550 | }; |
623540e4 EJ |
6551 | |
6552 | ofpbuf_init(&key, 0); | |
5262eea1 | 6553 | odp_flow_key_from_flow(&odp_parms, &key); |
2482b0b0 | 6554 | packet_str = ofp_dp_packet_to_string(packet_); |
623540e4 | 6555 | |
6fd6ed71 | 6556 | odp_flow_key_format(key.data, key.size, &ds); |
623540e4 EJ |
6557 | |
6558 | VLOG_DBG("%s: %s upcall:\n%s\n%s", dp->name, | |
6559 | dpif_upcall_type_to_string(type), ds_cstr(&ds), packet_str); | |
6560 | ||
6561 | ofpbuf_uninit(&key); | |
6562 | free(packet_str); | |
6fd6ed71 | 6563 | |
623540e4 EJ |
6564 | ds_destroy(&ds); |
6565 | } | |
6566 | ||
8d8ab6c2 JG |
6567 | return dp->upcall_cb(packet_, flow, ufid, pmd->core_id, type, userdata, |
6568 | actions, wc, put_actions, dp->upcall_aux); | |
623540e4 EJ |
6569 | } |
6570 | ||
bde94613 FA |
6571 | static inline uint32_t |
6572 | dpif_netdev_packet_get_rss_hash_orig_pkt(struct dp_packet *packet, | |
6573 | const struct miniflow *mf) | |
6574 | { | |
6575 | uint32_t hash; | |
6576 | ||
6577 | if (OVS_LIKELY(dp_packet_rss_valid(packet))) { | |
6578 | hash = dp_packet_get_rss_hash(packet); | |
6579 | } else { | |
6580 | hash = miniflow_hash_5tuple(mf, 0); | |
6581 | dp_packet_set_rss_hash(packet, hash); | |
6582 | } | |
6583 | ||
6584 | return hash; | |
6585 | } | |
6586 | ||
9bbf1c3d | 6587 | static inline uint32_t |
048963aa DDP |
6588 | dpif_netdev_packet_get_rss_hash(struct dp_packet *packet, |
6589 | const struct miniflow *mf) | |
9bbf1c3d | 6590 | { |
048963aa | 6591 | uint32_t hash, recirc_depth; |
9bbf1c3d | 6592 | |
f2f44f5d DDP |
6593 | if (OVS_LIKELY(dp_packet_rss_valid(packet))) { |
6594 | hash = dp_packet_get_rss_hash(packet); | |
6595 | } else { | |
9bbf1c3d | 6596 | hash = miniflow_hash_5tuple(mf, 0); |
2bc1bbd2 | 6597 | dp_packet_set_rss_hash(packet, hash); |
9bbf1c3d | 6598 | } |
048963aa DDP |
6599 | |
6600 | /* The RSS hash must account for the recirculation depth to avoid | |
6601 | * collisions in the exact match cache */ | |
6602 | recirc_depth = *recirc_depth_get_unsafe(); | |
6603 | if (OVS_UNLIKELY(recirc_depth)) { | |
6604 | hash = hash_finish(hash, recirc_depth); | |
048963aa | 6605 | } |
9bbf1c3d DDP |
6606 | return hash; |
6607 | } | |
6608 | ||
f7ce4811 | 6609 | struct packet_batch_per_flow { |
8cbf4f47 DDP |
6610 | unsigned int byte_count; |
6611 | uint16_t tcp_flags; | |
8cbf4f47 DDP |
6612 | struct dp_netdev_flow *flow; |
6613 | ||
1895cc8d | 6614 | struct dp_packet_batch array; |
8cbf4f47 DDP |
6615 | }; |
6616 | ||
6617 | static inline void | |
f7ce4811 PS |
6618 | packet_batch_per_flow_update(struct packet_batch_per_flow *batch, |
6619 | struct dp_packet *packet, | |
aab96ec4 | 6620 | uint16_t tcp_flags) |
8cbf4f47 | 6621 | { |
cf62fa4c | 6622 | batch->byte_count += dp_packet_size(packet); |
aab96ec4 | 6623 | batch->tcp_flags |= tcp_flags; |
940ac2ce | 6624 | dp_packet_batch_add(&batch->array, packet); |
8cbf4f47 DDP |
6625 | } |
6626 | ||
6627 | static inline void | |
f7ce4811 PS |
6628 | packet_batch_per_flow_init(struct packet_batch_per_flow *batch, |
6629 | struct dp_netdev_flow *flow) | |
8cbf4f47 | 6630 | { |
11e5cf1f | 6631 | flow->batch = batch; |
8cbf4f47 | 6632 | |
11e5cf1f | 6633 | batch->flow = flow; |
1895cc8d | 6634 | dp_packet_batch_init(&batch->array); |
8cbf4f47 DDP |
6635 | batch->byte_count = 0; |
6636 | batch->tcp_flags = 0; | |
8cbf4f47 DDP |
6637 | } |
6638 | ||
6639 | static inline void | |
f7ce4811 | 6640 | packet_batch_per_flow_execute(struct packet_batch_per_flow *batch, |
b010be17 | 6641 | struct dp_netdev_pmd_thread *pmd) |
8cbf4f47 DDP |
6642 | { |
6643 | struct dp_netdev_actions *actions; | |
6644 | struct dp_netdev_flow *flow = batch->flow; | |
6645 | ||
940ac2ce PC |
6646 | dp_netdev_flow_used(flow, dp_packet_batch_size(&batch->array), |
6647 | batch->byte_count, | |
05f9e707 | 6648 | batch->tcp_flags, pmd->ctx.now / 1000); |
8cbf4f47 DDP |
6649 | |
6650 | actions = dp_netdev_flow_get_actions(flow); | |
6651 | ||
66e4ad8a | 6652 | dp_netdev_execute_actions(pmd, &batch->array, true, &flow->flow, |
b010be17 | 6653 | actions->actions, actions->size); |
8cbf4f47 DDP |
6654 | } |
6655 | ||
8aaa125d | 6656 | static inline void |
e14deea0 | 6657 | dp_netdev_queue_batches(struct dp_packet *pkt, |
aab96ec4 | 6658 | struct dp_netdev_flow *flow, uint16_t tcp_flags, |
47a45d86 KT |
6659 | struct packet_batch_per_flow *batches, |
6660 | size_t *n_batches) | |
9bbf1c3d | 6661 | { |
f7ce4811 | 6662 | struct packet_batch_per_flow *batch = flow->batch; |
11e5cf1f | 6663 | |
f9fe365b AZ |
6664 | if (OVS_UNLIKELY(!batch)) { |
6665 | batch = &batches[(*n_batches)++]; | |
f7ce4811 | 6666 | packet_batch_per_flow_init(batch, flow); |
9bbf1c3d DDP |
6667 | } |
6668 | ||
aab96ec4 | 6669 | packet_batch_per_flow_update(batch, pkt, tcp_flags); |
9bbf1c3d DDP |
6670 | } |
6671 | ||
9b4f08cd VDA |
6672 | static inline void |
6673 | packet_enqueue_to_flow_map(struct dp_packet *packet, | |
6674 | struct dp_netdev_flow *flow, | |
6675 | uint16_t tcp_flags, | |
6676 | struct dp_packet_flow_map *flow_map, | |
6677 | size_t index) | |
6678 | { | |
6679 | struct dp_packet_flow_map *map = &flow_map[index]; | |
6680 | map->flow = flow; | |
6681 | map->packet = packet; | |
6682 | map->tcp_flags = tcp_flags; | |
6683 | } | |
6684 | ||
60d8ccae YW |
6685 | /* SMC lookup function for a batch of packets. |
6686 | * By doing batching SMC lookup, we can use prefetch | |
6687 | * to hide memory access latency. | |
6688 | */ | |
6689 | static inline void | |
6690 | smc_lookup_batch(struct dp_netdev_pmd_thread *pmd, | |
6691 | struct netdev_flow_key *keys, | |
6692 | struct netdev_flow_key **missed_keys, | |
6693 | struct dp_packet_batch *packets_, | |
9b4f08cd VDA |
6694 | const int cnt, |
6695 | struct dp_packet_flow_map *flow_map, | |
6696 | uint8_t *index_map) | |
60d8ccae YW |
6697 | { |
6698 | int i; | |
6699 | struct dp_packet *packet; | |
6700 | size_t n_smc_hit = 0, n_missed = 0; | |
6701 | struct dfc_cache *cache = &pmd->flow_cache; | |
6702 | struct smc_cache *smc_cache = &cache->smc_cache; | |
6703 | const struct cmap_node *flow_node; | |
9b4f08cd VDA |
6704 | int recv_idx; |
6705 | uint16_t tcp_flags; | |
60d8ccae YW |
6706 | |
6707 | /* Prefetch buckets for all packets */ | |
6708 | for (i = 0; i < cnt; i++) { | |
6709 | OVS_PREFETCH(&smc_cache->buckets[keys[i].hash & SMC_MASK]); | |
6710 | } | |
6711 | ||
6712 | DP_PACKET_BATCH_REFILL_FOR_EACH (i, cnt, packet, packets_) { | |
6713 | struct dp_netdev_flow *flow = NULL; | |
6714 | flow_node = smc_entry_get(pmd, keys[i].hash); | |
6715 | bool hit = false; | |
9b4f08cd VDA |
6716 | /* Get the original order of this packet in received batch. */ |
6717 | recv_idx = index_map[i]; | |
60d8ccae YW |
6718 | |
6719 | if (OVS_LIKELY(flow_node != NULL)) { | |
6720 | CMAP_NODE_FOR_EACH (flow, node, flow_node) { | |
6721 | /* Since we dont have per-port megaflow to check the port | |
6722 | * number, we need to verify that the input ports match. */ | |
6723 | if (OVS_LIKELY(dpcls_rule_matches_key(&flow->cr, &keys[i]) && | |
6724 | flow->flow.in_port.odp_port == packet->md.in_port.odp_port)) { | |
9b4f08cd VDA |
6725 | tcp_flags = miniflow_get_tcp_flags(&keys[i].mf); |
6726 | ||
60d8ccae | 6727 | /* SMC hit and emc miss, we insert into EMC */ |
60d8ccae YW |
6728 | keys[i].len = |
6729 | netdev_flow_key_size(miniflow_n_values(&keys[i].mf)); | |
18e08953 | 6730 | emc_probabilistic_insert(pmd, &keys[i], flow); |
9b4f08cd VDA |
6731 | /* Add these packets into the flow map in the same order |
6732 | * as received. | |
6733 | */ | |
6734 | packet_enqueue_to_flow_map(packet, flow, tcp_flags, | |
6735 | flow_map, recv_idx); | |
60d8ccae YW |
6736 | n_smc_hit++; |
6737 | hit = true; | |
6738 | break; | |
6739 | } | |
6740 | } | |
6741 | if (hit) { | |
6742 | continue; | |
6743 | } | |
6744 | } | |
6745 | ||
6746 | /* SMC missed. Group missed packets together at | |
6747 | * the beginning of the 'packets' array. */ | |
6748 | dp_packet_batch_refill(packets_, packet, i); | |
9b4f08cd VDA |
6749 | |
6750 | /* Preserve the order of packet for flow batching. */ | |
6751 | index_map[n_missed] = recv_idx; | |
6752 | ||
60d8ccae YW |
6753 | /* Put missed keys to the pointer arrays return to the caller */ |
6754 | missed_keys[n_missed++] = &keys[i]; | |
6755 | } | |
6756 | ||
6757 | pmd_perf_update_counter(&pmd->perf_stats, PMD_STAT_SMC_HIT, n_smc_hit); | |
6758 | } | |
6759 | ||
6760 | /* Try to process all ('cnt') the 'packets' using only the datapath flow cache | |
a90ed026 | 6761 | * 'pmd->flow_cache'. If a flow is not found for a packet 'packets[i]', the |
8aaa125d | 6762 | * miniflow is copied into 'keys' and the packet pointer is moved at the |
60d8ccae YW |
6763 | * beginning of the 'packets' array. The pointers of missed keys are put in the |
6764 | * missed_keys pointer array for future processing. | |
9bbf1c3d DDP |
6765 | * |
6766 | * The function returns the number of packets that needs to be processed in the | |
6767 | * 'packets' array (they have been moved to the beginning of the vector). | |
a90ed026 | 6768 | * |
02305520 FA |
6769 | * For performance reasons a caller may choose not to initialize the metadata |
6770 | * in 'packets_'. If 'md_is_valid' is false, the metadata in 'packets' | |
6771 | * is not valid and must be initialized by this function using 'port_no'. | |
6772 | * If 'md_is_valid' is true, the metadata is already valid and 'port_no' | |
6773 | * will be ignored. | |
9bbf1c3d DDP |
6774 | */ |
6775 | static inline size_t | |
60d8ccae | 6776 | dfc_processing(struct dp_netdev_pmd_thread *pmd, |
72c84bc2 | 6777 | struct dp_packet_batch *packets_, |
1895cc8d | 6778 | struct netdev_flow_key *keys, |
60d8ccae | 6779 | struct netdev_flow_key **missed_keys, |
f7ce4811 | 6780 | struct packet_batch_per_flow batches[], size_t *n_batches, |
9b4f08cd VDA |
6781 | struct dp_packet_flow_map *flow_map, |
6782 | size_t *n_flows, uint8_t *index_map, | |
a90ed026 | 6783 | bool md_is_valid, odp_port_t port_no) |
72865317 | 6784 | { |
b89c678b | 6785 | struct netdev_flow_key *key = &keys[0]; |
60d8ccae YW |
6786 | size_t n_missed = 0, n_emc_hit = 0; |
6787 | struct dfc_cache *cache = &pmd->flow_cache; | |
72c84bc2 | 6788 | struct dp_packet *packet; |
45df9fef | 6789 | const size_t cnt = dp_packet_batch_size(packets_); |
2fbadeb6 | 6790 | uint32_t cur_min = pmd->ctx.emc_insert_min; |
72c84bc2 | 6791 | int i; |
aab96ec4 | 6792 | uint16_t tcp_flags; |
60d8ccae | 6793 | bool smc_enable_db; |
9b4f08cd VDA |
6794 | size_t map_cnt = 0; |
6795 | bool batch_enable = true; | |
8cbf4f47 | 6796 | |
60d8ccae | 6797 | atomic_read_relaxed(&pmd->dp->smc_enable_db, &smc_enable_db); |
82a48ead JS |
6798 | pmd_perf_update_counter(&pmd->perf_stats, |
6799 | md_is_valid ? PMD_STAT_RECIRC : PMD_STAT_RECV, | |
6800 | cnt); | |
f79b1ddb | 6801 | |
45df9fef | 6802 | DP_PACKET_BATCH_REFILL_FOR_EACH (i, cnt, packet, packets_) { |
9bbf1c3d | 6803 | struct dp_netdev_flow *flow; |
aab96ec4 | 6804 | uint32_t mark; |
9bbf1c3d | 6805 | |
5a2fed48 AZ |
6806 | if (OVS_UNLIKELY(dp_packet_size(packet) < ETH_HEADER_LEN)) { |
6807 | dp_packet_delete(packet); | |
a13a0209 | 6808 | COVERAGE_INC(datapath_drop_rx_invalid_packet); |
84d6d5eb EJ |
6809 | continue; |
6810 | } | |
8cbf4f47 | 6811 | |
45df9fef | 6812 | if (i != cnt - 1) { |
72c84bc2 | 6813 | struct dp_packet **packets = packets_->packets; |
a90ed026 | 6814 | /* Prefetch next packet data and metadata. */ |
72a5e2b8 | 6815 | OVS_PREFETCH(dp_packet_data(packets[i+1])); |
a90ed026 | 6816 | pkt_metadata_prefetch_init(&packets[i+1]->md); |
72a5e2b8 DDP |
6817 | } |
6818 | ||
a90ed026 DDP |
6819 | if (!md_is_valid) { |
6820 | pkt_metadata_init(&packet->md, port_no); | |
6821 | } | |
aab96ec4 YL |
6822 | |
6823 | if ((*recirc_depth_get() == 0) && | |
6824 | dp_packet_has_flow_mark(packet, &mark)) { | |
6825 | flow = mark_to_flow_find(pmd, mark); | |
9b4f08cd | 6826 | if (OVS_LIKELY(flow)) { |
aab96ec4 | 6827 | tcp_flags = parse_tcp_flags(packet); |
9b4f08cd VDA |
6828 | if (OVS_LIKELY(batch_enable)) { |
6829 | dp_netdev_queue_batches(packet, flow, tcp_flags, batches, | |
6830 | n_batches); | |
6831 | } else { | |
6832 | /* Flow batching should be performed only after fast-path | |
6833 | * processing is also completed for packets with emc miss | |
6834 | * or else it will result in reordering of packets with | |
6835 | * same datapath flows. */ | |
6836 | packet_enqueue_to_flow_map(packet, flow, tcp_flags, | |
6837 | flow_map, map_cnt++); | |
6838 | } | |
aab96ec4 YL |
6839 | continue; |
6840 | } | |
6841 | } | |
6842 | ||
5a2fed48 | 6843 | miniflow_extract(packet, &key->mf); |
d262ac2c | 6844 | key->len = 0; /* Not computed yet. */ |
b137383e IM |
6845 | key->hash = |
6846 | (md_is_valid == false) | |
6847 | ? dpif_netdev_packet_get_rss_hash_orig_pkt(packet, &key->mf) | |
6848 | : dpif_netdev_packet_get_rss_hash(packet, &key->mf); | |
6849 | ||
6850 | /* If EMC is disabled skip emc_lookup */ | |
6851 | flow = (cur_min != 0) ? emc_lookup(&cache->emc_cache, key) : NULL; | |
8aaa125d | 6852 | if (OVS_LIKELY(flow)) { |
aab96ec4 | 6853 | tcp_flags = miniflow_get_tcp_flags(&key->mf); |
60d8ccae | 6854 | n_emc_hit++; |
9b4f08cd VDA |
6855 | if (OVS_LIKELY(batch_enable)) { |
6856 | dp_netdev_queue_batches(packet, flow, tcp_flags, batches, | |
6857 | n_batches); | |
6858 | } else { | |
6859 | /* Flow batching should be performed only after fast-path | |
6860 | * processing is also completed for packets with emc miss | |
6861 | * or else it will result in reordering of packets with | |
6862 | * same datapath flows. */ | |
6863 | packet_enqueue_to_flow_map(packet, flow, tcp_flags, | |
6864 | flow_map, map_cnt++); | |
6865 | } | |
8aaa125d | 6866 | } else { |
d1aa0b94 | 6867 | /* Exact match cache missed. Group missed packets together at |
72c84bc2 AZ |
6868 | * the beginning of the 'packets' array. */ |
6869 | dp_packet_batch_refill(packets_, packet, i); | |
9b4f08cd VDA |
6870 | |
6871 | /* Preserve the order of packet for flow batching. */ | |
6872 | index_map[n_missed] = map_cnt; | |
6873 | flow_map[map_cnt++].flow = NULL; | |
6874 | ||
400486f7 | 6875 | /* 'key[n_missed]' contains the key of the current packet and it |
60d8ccae YW |
6876 | * will be passed to SMC lookup. The next key should be extracted |
6877 | * to 'keys[n_missed + 1]'. | |
6878 | * We also maintain a pointer array to keys missed both SMC and EMC | |
6879 | * which will be returned to the caller for future processing. */ | |
6880 | missed_keys[n_missed] = key; | |
400486f7 | 6881 | key = &keys[++n_missed]; |
9b4f08cd VDA |
6882 | |
6883 | /* Skip batching for subsequent packets to avoid reordering. */ | |
6884 | batch_enable = false; | |
9bbf1c3d DDP |
6885 | } |
6886 | } | |
9b4f08cd VDA |
6887 | /* Count of packets which are not flow batched. */ |
6888 | *n_flows = map_cnt; | |
9bbf1c3d | 6889 | |
60d8ccae YW |
6890 | pmd_perf_update_counter(&pmd->perf_stats, PMD_STAT_EXACT_HIT, n_emc_hit); |
6891 | ||
6892 | if (!smc_enable_db) { | |
6893 | return dp_packet_batch_size(packets_); | |
6894 | } | |
6895 | ||
6896 | /* Packets miss EMC will do a batch lookup in SMC if enabled */ | |
9b4f08cd VDA |
6897 | smc_lookup_batch(pmd, keys, missed_keys, packets_, |
6898 | n_missed, flow_map, index_map); | |
4f150744 | 6899 | |
72c84bc2 | 6900 | return dp_packet_batch_size(packets_); |
9bbf1c3d DDP |
6901 | } |
6902 | ||
82a48ead | 6903 | static inline int |
47a45d86 KT |
6904 | handle_packet_upcall(struct dp_netdev_pmd_thread *pmd, |
6905 | struct dp_packet *packet, | |
a260d966 | 6906 | const struct netdev_flow_key *key, |
82a48ead | 6907 | struct ofpbuf *actions, struct ofpbuf *put_actions) |
a260d966 PS |
6908 | { |
6909 | struct ofpbuf *add_actions; | |
6910 | struct dp_packet_batch b; | |
6911 | struct match match; | |
6912 | ovs_u128 ufid; | |
6913 | int error; | |
79f36875 | 6914 | uint64_t cycles = cycles_counter_update(&pmd->perf_stats); |
a260d966 PS |
6915 | |
6916 | match.tun_md.valid = false; | |
6917 | miniflow_expand(&key->mf, &match.flow); | |
c98eedf9 | 6918 | memset(&match.wc, 0, sizeof match.wc); |
a260d966 PS |
6919 | |
6920 | ofpbuf_clear(actions); | |
6921 | ofpbuf_clear(put_actions); | |
6922 | ||
7a5e0ee7 | 6923 | odp_flow_key_hash(&match.flow, sizeof match.flow, &ufid); |
a260d966 PS |
6924 | error = dp_netdev_upcall(pmd, packet, &match.flow, &match.wc, |
6925 | &ufid, DPIF_UC_MISS, NULL, actions, | |
6926 | put_actions); | |
6927 | if (OVS_UNLIKELY(error && error != ENOSPC)) { | |
6928 | dp_packet_delete(packet); | |
a13a0209 | 6929 | COVERAGE_INC(datapath_drop_upcall_error); |
82a48ead | 6930 | return error; |
a260d966 PS |
6931 | } |
6932 | ||
6933 | /* The Netlink encoding of datapath flow keys cannot express | |
6934 | * wildcarding the presence of a VLAN tag. Instead, a missing VLAN | |
6935 | * tag is interpreted as exact match on the fact that there is no | |
6936 | * VLAN. Unless we refactor a lot of code that translates between | |
6937 | * Netlink and struct flow representations, we have to do the same | |
35fe9efb | 6938 | * here. This must be in sync with 'match' in dpif_netdev_flow_put(). */ |
f0fb825a EG |
6939 | if (!match.wc.masks.vlans[0].tci) { |
6940 | match.wc.masks.vlans[0].tci = htons(0xffff); | |
a260d966 PS |
6941 | } |
6942 | ||
6943 | /* We can't allow the packet batching in the next loop to execute | |
6944 | * the actions. Otherwise, if there are any slow path actions, | |
6945 | * we'll send the packet up twice. */ | |
72c84bc2 | 6946 | dp_packet_batch_init_packet(&b, packet); |
66e4ad8a | 6947 | dp_netdev_execute_actions(pmd, &b, true, &match.flow, |
b010be17 | 6948 | actions->data, actions->size); |
a260d966 PS |
6949 | |
6950 | add_actions = put_actions->size ? put_actions : actions; | |
6951 | if (OVS_LIKELY(error != ENOSPC)) { | |
6952 | struct dp_netdev_flow *netdev_flow; | |
6953 | ||
6954 | /* XXX: There's a race window where a flow covering this packet | |
6955 | * could have already been installed since we last did the flow | |
6956 | * lookup before upcall. This could be solved by moving the | |
6957 | * mutex lock outside the loop, but that's an awful long time | |
af741ca3 | 6958 | * to be locking revalidators out of making flow modifications. */ |
a260d966 | 6959 | ovs_mutex_lock(&pmd->flow_mutex); |
3453b4d6 | 6960 | netdev_flow = dp_netdev_pmd_lookup_flow(pmd, key, NULL); |
a260d966 PS |
6961 | if (OVS_LIKELY(!netdev_flow)) { |
6962 | netdev_flow = dp_netdev_flow_add(pmd, &match, &ufid, | |
6963 | add_actions->data, | |
6964 | add_actions->size); | |
6965 | } | |
6966 | ovs_mutex_unlock(&pmd->flow_mutex); | |
60d8ccae YW |
6967 | uint32_t hash = dp_netdev_flow_hash(&netdev_flow->ufid); |
6968 | smc_insert(pmd, key, hash); | |
4c30b246 | 6969 | emc_probabilistic_insert(pmd, key, netdev_flow); |
a260d966 | 6970 | } |
79f36875 JS |
6971 | if (pmd_perf_metrics_enabled(pmd)) { |
6972 | /* Update upcall stats. */ | |
6973 | cycles = cycles_counter_update(&pmd->perf_stats) - cycles; | |
6974 | struct pmd_perf_stats *s = &pmd->perf_stats; | |
6975 | s->current.upcalls++; | |
6976 | s->current.upcall_cycles += cycles; | |
6977 | histogram_add_sample(&s->cycles_per_upcall, cycles); | |
6978 | } | |
82a48ead | 6979 | return error; |
a260d966 PS |
6980 | } |
6981 | ||
9bbf1c3d | 6982 | static inline void |
65f13b50 | 6983 | fast_path_processing(struct dp_netdev_pmd_thread *pmd, |
1895cc8d | 6984 | struct dp_packet_batch *packets_, |
60d8ccae | 6985 | struct netdev_flow_key **keys, |
9b4f08cd VDA |
6986 | struct dp_packet_flow_map *flow_map, |
6987 | uint8_t *index_map, | |
b010be17 | 6988 | odp_port_t in_port) |
9bbf1c3d | 6989 | { |
31c82130 | 6990 | const size_t cnt = dp_packet_batch_size(packets_); |
1a0d5831 | 6991 | #if !defined(__CHECKER__) && !defined(_WIN32) |
9bbf1c3d DDP |
6992 | const size_t PKT_ARRAY_SIZE = cnt; |
6993 | #else | |
1a0d5831 | 6994 | /* Sparse or MSVC doesn't like variable length array. */ |
cd159f1a | 6995 | enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST }; |
9bbf1c3d | 6996 | #endif |
31c82130 | 6997 | struct dp_packet *packet; |
3453b4d6 | 6998 | struct dpcls *cls; |
0de8783a | 6999 | struct dpcls_rule *rules[PKT_ARRAY_SIZE]; |
65f13b50 | 7000 | struct dp_netdev *dp = pmd->dp; |
82a48ead | 7001 | int upcall_ok_cnt = 0, upcall_fail_cnt = 0; |
3453b4d6 | 7002 | int lookup_cnt = 0, add_lookup_cnt; |
9bbf1c3d DDP |
7003 | bool any_miss; |
7004 | ||
e883448e | 7005 | for (size_t i = 0; i < cnt; i++) { |
0de8783a | 7006 | /* Key length is needed in all the cases, hash computed on demand. */ |
60d8ccae | 7007 | keys[i]->len = netdev_flow_key_size(miniflow_n_values(&keys[i]->mf)); |
9bbf1c3d | 7008 | } |
3453b4d6 JS |
7009 | /* Get the classifier for the in_port */ |
7010 | cls = dp_netdev_pmd_lookup_dpcls(pmd, in_port); | |
7011 | if (OVS_LIKELY(cls)) { | |
60d8ccae YW |
7012 | any_miss = !dpcls_lookup(cls, (const struct netdev_flow_key **)keys, |
7013 | rules, cnt, &lookup_cnt); | |
3453b4d6 JS |
7014 | } else { |
7015 | any_miss = true; | |
7016 | memset(rules, 0, sizeof(rules)); | |
7017 | } | |
623540e4 EJ |
7018 | if (OVS_UNLIKELY(any_miss) && !fat_rwlock_tryrdlock(&dp->upcall_rwlock)) { |
7019 | uint64_t actions_stub[512 / 8], slow_stub[512 / 8]; | |
7020 | struct ofpbuf actions, put_actions; | |
623540e4 EJ |
7021 | |
7022 | ofpbuf_use_stub(&actions, actions_stub, sizeof actions_stub); | |
7023 | ofpbuf_use_stub(&put_actions, slow_stub, sizeof slow_stub); | |
7024 | ||
e883448e | 7025 | DP_PACKET_BATCH_FOR_EACH (i, packet, packets_) { |
0de8783a | 7026 | struct dp_netdev_flow *netdev_flow; |
623540e4 | 7027 | |
0de8783a | 7028 | if (OVS_LIKELY(rules[i])) { |
623540e4 EJ |
7029 | continue; |
7030 | } | |
7031 | ||
7032 | /* It's possible that an earlier slow path execution installed | |
0de8783a | 7033 | * a rule covering this flow. In this case, it's a lot cheaper |
623540e4 | 7034 | * to catch it here than execute a miss. */ |
60d8ccae | 7035 | netdev_flow = dp_netdev_pmd_lookup_flow(pmd, keys[i], |
3453b4d6 | 7036 | &add_lookup_cnt); |
623540e4 | 7037 | if (netdev_flow) { |
3453b4d6 | 7038 | lookup_cnt += add_lookup_cnt; |
0de8783a | 7039 | rules[i] = &netdev_flow->cr; |
623540e4 EJ |
7040 | continue; |
7041 | } | |
7042 | ||
60d8ccae | 7043 | int error = handle_packet_upcall(pmd, packet, keys[i], |
82a48ead JS |
7044 | &actions, &put_actions); |
7045 | ||
7046 | if (OVS_UNLIKELY(error)) { | |
7047 | upcall_fail_cnt++; | |
7048 | } else { | |
7049 | upcall_ok_cnt++; | |
7050 | } | |
623540e4 EJ |
7051 | } |
7052 | ||
7053 | ofpbuf_uninit(&actions); | |
7054 | ofpbuf_uninit(&put_actions); | |
7055 | fat_rwlock_unlock(&dp->upcall_rwlock); | |
ac8c2081 | 7056 | } else if (OVS_UNLIKELY(any_miss)) { |
e883448e | 7057 | DP_PACKET_BATCH_FOR_EACH (i, packet, packets_) { |
0de8783a | 7058 | if (OVS_UNLIKELY(!rules[i])) { |
31c82130 | 7059 | dp_packet_delete(packet); |
a13a0209 | 7060 | COVERAGE_INC(datapath_drop_lock_error); |
82a48ead | 7061 | upcall_fail_cnt++; |
ac8c2081 DDP |
7062 | } |
7063 | } | |
623540e4 | 7064 | } |
84d6d5eb | 7065 | |
e883448e | 7066 | DP_PACKET_BATCH_FOR_EACH (i, packet, packets_) { |
84d6d5eb | 7067 | struct dp_netdev_flow *flow; |
9b4f08cd VDA |
7068 | /* Get the original order of this packet in received batch. */ |
7069 | int recv_idx = index_map[i]; | |
7070 | uint16_t tcp_flags; | |
8cbf4f47 | 7071 | |
0de8783a | 7072 | if (OVS_UNLIKELY(!rules[i])) { |
84d6d5eb EJ |
7073 | continue; |
7074 | } | |
7075 | ||
84d6d5eb | 7076 | flow = dp_netdev_flow_cast(rules[i]); |
60d8ccae YW |
7077 | uint32_t hash = dp_netdev_flow_hash(&flow->ufid); |
7078 | smc_insert(pmd, keys[i], hash); | |
0de8783a | 7079 | |
60d8ccae | 7080 | emc_probabilistic_insert(pmd, keys[i], flow); |
9b4f08cd VDA |
7081 | /* Add these packets into the flow map in the same order |
7082 | * as received. | |
7083 | */ | |
7084 | tcp_flags = miniflow_get_tcp_flags(&keys[i]->mf); | |
7085 | packet_enqueue_to_flow_map(packet, flow, tcp_flags, | |
7086 | flow_map, recv_idx); | |
8cbf4f47 DDP |
7087 | } |
7088 | ||
82a48ead JS |
7089 | pmd_perf_update_counter(&pmd->perf_stats, PMD_STAT_MASKED_HIT, |
7090 | cnt - upcall_ok_cnt - upcall_fail_cnt); | |
7091 | pmd_perf_update_counter(&pmd->perf_stats, PMD_STAT_MASKED_LOOKUP, | |
7092 | lookup_cnt); | |
7093 | pmd_perf_update_counter(&pmd->perf_stats, PMD_STAT_MISS, | |
7094 | upcall_ok_cnt); | |
7095 | pmd_perf_update_counter(&pmd->perf_stats, PMD_STAT_LOST, | |
7096 | upcall_fail_cnt); | |
72865317 BP |
7097 | } |
7098 | ||
a90ed026 DDP |
7099 | /* Packets enter the datapath from a port (or from recirculation) here. |
7100 | * | |
02305520 FA |
7101 | * When 'md_is_valid' is true the metadata in 'packets' are already valid. |
7102 | * When false the metadata in 'packets' need to be initialized. */ | |
adcf00ba | 7103 | static void |
a90ed026 | 7104 | dp_netdev_input__(struct dp_netdev_pmd_thread *pmd, |
1895cc8d | 7105 | struct dp_packet_batch *packets, |
a90ed026 | 7106 | bool md_is_valid, odp_port_t port_no) |
9bbf1c3d | 7107 | { |
1a0d5831 | 7108 | #if !defined(__CHECKER__) && !defined(_WIN32) |
37eabc70 | 7109 | const size_t PKT_ARRAY_SIZE = dp_packet_batch_size(packets); |
9bbf1c3d | 7110 | #else |
1a0d5831 | 7111 | /* Sparse or MSVC doesn't like variable length array. */ |
cd159f1a | 7112 | enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST }; |
9bbf1c3d | 7113 | #endif |
47a45d86 KT |
7114 | OVS_ALIGNED_VAR(CACHE_LINE_SIZE) |
7115 | struct netdev_flow_key keys[PKT_ARRAY_SIZE]; | |
60d8ccae | 7116 | struct netdev_flow_key *missed_keys[PKT_ARRAY_SIZE]; |
f7ce4811 | 7117 | struct packet_batch_per_flow batches[PKT_ARRAY_SIZE]; |
72c84bc2 | 7118 | size_t n_batches; |
9b4f08cd VDA |
7119 | struct dp_packet_flow_map flow_map[PKT_ARRAY_SIZE]; |
7120 | uint8_t index_map[PKT_ARRAY_SIZE]; | |
7121 | size_t n_flows, i; | |
7122 | ||
3453b4d6 | 7123 | odp_port_t in_port; |
9bbf1c3d | 7124 | |
8aaa125d | 7125 | n_batches = 0; |
60d8ccae | 7126 | dfc_processing(pmd, packets, keys, missed_keys, batches, &n_batches, |
9b4f08cd VDA |
7127 | flow_map, &n_flows, index_map, md_is_valid, port_no); |
7128 | ||
72c84bc2 | 7129 | if (!dp_packet_batch_is_empty(packets)) { |
3453b4d6 JS |
7130 | /* Get ingress port from first packet's metadata. */ |
7131 | in_port = packets->packets[0]->md.in_port.odp_port; | |
60d8ccae | 7132 | fast_path_processing(pmd, packets, missed_keys, |
9b4f08cd | 7133 | flow_map, index_map, in_port); |
8aaa125d DDP |
7134 | } |
7135 | ||
9b4f08cd VDA |
7136 | /* Batch rest of packets which are in flow map. */ |
7137 | for (i = 0; i < n_flows; i++) { | |
7138 | struct dp_packet_flow_map *map = &flow_map[i]; | |
7139 | ||
7140 | if (OVS_UNLIKELY(!map->flow)) { | |
7141 | continue; | |
7142 | } | |
7143 | dp_netdev_queue_batches(map->packet, map->flow, map->tcp_flags, | |
7144 | batches, &n_batches); | |
7145 | } | |
7146 | ||
ad9f0581 BB |
7147 | /* All the flow batches need to be reset before any call to |
7148 | * packet_batch_per_flow_execute() as it could potentially trigger | |
7149 | * recirculation. When a packet matching flow ‘j’ happens to be | |
7150 | * recirculated, the nested call to dp_netdev_input__() could potentially | |
7151 | * classify the packet as matching another flow - say 'k'. It could happen | |
7152 | * that in the previous call to dp_netdev_input__() that same flow 'k' had | |
7153 | * already its own batches[k] still waiting to be served. So if its | |
7154 | * ‘batch’ member is not reset, the recirculated packet would be wrongly | |
7155 | * appended to batches[k] of the 1st call to dp_netdev_input__(). */ | |
603f2ce0 EJ |
7156 | for (i = 0; i < n_batches; i++) { |
7157 | batches[i].flow->batch = NULL; | |
7158 | } | |
7159 | ||
8aaa125d | 7160 | for (i = 0; i < n_batches; i++) { |
b010be17 | 7161 | packet_batch_per_flow_execute(&batches[i], pmd); |
9bbf1c3d DDP |
7162 | } |
7163 | } | |
7164 | ||
a90ed026 DDP |
7165 | static void |
7166 | dp_netdev_input(struct dp_netdev_pmd_thread *pmd, | |
1895cc8d | 7167 | struct dp_packet_batch *packets, |
a90ed026 DDP |
7168 | odp_port_t port_no) |
7169 | { | |
3453b4d6 | 7170 | dp_netdev_input__(pmd, packets, false, port_no); |
a90ed026 DDP |
7171 | } |
7172 | ||
7173 | static void | |
7174 | dp_netdev_recirculate(struct dp_netdev_pmd_thread *pmd, | |
1895cc8d | 7175 | struct dp_packet_batch *packets) |
a90ed026 | 7176 | { |
3453b4d6 | 7177 | dp_netdev_input__(pmd, packets, true, 0); |
a90ed026 DDP |
7178 | } |
7179 | ||
9080a111 | 7180 | struct dp_netdev_execute_aux { |
65f13b50 | 7181 | struct dp_netdev_pmd_thread *pmd; |
66e4ad8a | 7182 | const struct flow *flow; |
9080a111 JR |
7183 | }; |
7184 | ||
e4e74c3a AW |
7185 | static void |
7186 | dpif_netdev_register_dp_purge_cb(struct dpif *dpif, dp_purge_callback *cb, | |
7187 | void *aux) | |
7188 | { | |
7189 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
7190 | dp->dp_purge_aux = aux; | |
7191 | dp->dp_purge_cb = cb; | |
7192 | } | |
7193 | ||
6b31e073 | 7194 | static void |
623540e4 EJ |
7195 | dpif_netdev_register_upcall_cb(struct dpif *dpif, upcall_callback *cb, |
7196 | void *aux) | |
6b31e073 RW |
7197 | { |
7198 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
623540e4 | 7199 | dp->upcall_aux = aux; |
6b31e073 RW |
7200 | dp->upcall_cb = cb; |
7201 | } | |
7202 | ||
324c8374 IM |
7203 | static void |
7204 | dpif_netdev_xps_revalidate_pmd(const struct dp_netdev_pmd_thread *pmd, | |
b010be17 | 7205 | bool purge) |
324c8374 IM |
7206 | { |
7207 | struct tx_port *tx; | |
7208 | struct dp_netdev_port *port; | |
7209 | long long interval; | |
7210 | ||
57eebbb4 | 7211 | HMAP_FOR_EACH (tx, node, &pmd->send_port_cache) { |
9f7a3035 | 7212 | if (!tx->port->dynamic_txqs) { |
324c8374 IM |
7213 | continue; |
7214 | } | |
b010be17 | 7215 | interval = pmd->ctx.now - tx->last_used; |
05f9e707 | 7216 | if (tx->qid >= 0 && (purge || interval >= XPS_TIMEOUT)) { |
324c8374 IM |
7217 | port = tx->port; |
7218 | ovs_mutex_lock(&port->txq_used_mutex); | |
7219 | port->txq_used[tx->qid]--; | |
7220 | ovs_mutex_unlock(&port->txq_used_mutex); | |
7221 | tx->qid = -1; | |
7222 | } | |
7223 | } | |
7224 | } | |
7225 | ||
7226 | static int | |
7227 | dpif_netdev_xps_get_tx_qid(const struct dp_netdev_pmd_thread *pmd, | |
b010be17 | 7228 | struct tx_port *tx) |
324c8374 IM |
7229 | { |
7230 | struct dp_netdev_port *port; | |
7231 | long long interval; | |
7232 | int i, min_cnt, min_qid; | |
7233 | ||
b010be17 IM |
7234 | interval = pmd->ctx.now - tx->last_used; |
7235 | tx->last_used = pmd->ctx.now; | |
324c8374 | 7236 | |
05f9e707 | 7237 | if (OVS_LIKELY(tx->qid >= 0 && interval < XPS_TIMEOUT)) { |
324c8374 IM |
7238 | return tx->qid; |
7239 | } | |
7240 | ||
7241 | port = tx->port; | |
7242 | ||
7243 | ovs_mutex_lock(&port->txq_used_mutex); | |
7244 | if (tx->qid >= 0) { | |
7245 | port->txq_used[tx->qid]--; | |
7246 | tx->qid = -1; | |
7247 | } | |
7248 | ||
7249 | min_cnt = -1; | |
7250 | min_qid = 0; | |
7251 | for (i = 0; i < netdev_n_txq(port->netdev); i++) { | |
7252 | if (port->txq_used[i] < min_cnt || min_cnt == -1) { | |
7253 | min_cnt = port->txq_used[i]; | |
7254 | min_qid = i; | |
7255 | } | |
7256 | } | |
7257 | ||
7258 | port->txq_used[min_qid]++; | |
7259 | tx->qid = min_qid; | |
7260 | ||
7261 | ovs_mutex_unlock(&port->txq_used_mutex); | |
7262 | ||
b010be17 | 7263 | dpif_netdev_xps_revalidate_pmd(pmd, false); |
324c8374 IM |
7264 | |
7265 | VLOG_DBG("Core %d: New TX queue ID %d for port \'%s\'.", | |
7266 | pmd->core_id, tx->qid, netdev_get_name(tx->port->netdev)); | |
7267 | return min_qid; | |
7268 | } | |
7269 | ||
d0cca6c3 | 7270 | static struct tx_port * |
57eebbb4 DDP |
7271 | pmd_tnl_port_cache_lookup(const struct dp_netdev_pmd_thread *pmd, |
7272 | odp_port_t port_no) | |
7273 | { | |
7274 | return tx_port_lookup(&pmd->tnl_port_cache, port_no); | |
7275 | } | |
7276 | ||
7277 | static struct tx_port * | |
7278 | pmd_send_port_cache_lookup(const struct dp_netdev_pmd_thread *pmd, | |
7279 | odp_port_t port_no) | |
d0cca6c3 | 7280 | { |
57eebbb4 | 7281 | return tx_port_lookup(&pmd->send_port_cache, port_no); |
d0cca6c3 DDP |
7282 | } |
7283 | ||
a36de779 | 7284 | static int |
d0cca6c3 | 7285 | push_tnl_action(const struct dp_netdev_pmd_thread *pmd, |
1895cc8d PS |
7286 | const struct nlattr *attr, |
7287 | struct dp_packet_batch *batch) | |
a36de779 | 7288 | { |
d0cca6c3 | 7289 | struct tx_port *tun_port; |
a36de779 | 7290 | const struct ovs_action_push_tnl *data; |
4c742796 | 7291 | int err; |
a36de779 PS |
7292 | |
7293 | data = nl_attr_get(attr); | |
7294 | ||
81765c00 | 7295 | tun_port = pmd_tnl_port_cache_lookup(pmd, data->tnl_port); |
a36de779 | 7296 | if (!tun_port) { |
4c742796 PS |
7297 | err = -EINVAL; |
7298 | goto error; | |
a36de779 | 7299 | } |
324c8374 | 7300 | err = netdev_push_header(tun_port->port->netdev, batch, data); |
4c742796 PS |
7301 | if (!err) { |
7302 | return 0; | |
7303 | } | |
7304 | error: | |
7305 | dp_packet_delete_batch(batch, true); | |
7306 | return err; | |
a36de779 PS |
7307 | } |
7308 | ||
66525ef3 PS |
7309 | static void |
7310 | dp_execute_userspace_action(struct dp_netdev_pmd_thread *pmd, | |
7d7ded7a | 7311 | struct dp_packet *packet, bool should_steal, |
66525ef3 PS |
7312 | struct flow *flow, ovs_u128 *ufid, |
7313 | struct ofpbuf *actions, | |
b010be17 | 7314 | const struct nlattr *userdata) |
66525ef3 PS |
7315 | { |
7316 | struct dp_packet_batch b; | |
7317 | int error; | |
7318 | ||
7319 | ofpbuf_clear(actions); | |
7320 | ||
7321 | error = dp_netdev_upcall(pmd, packet, flow, NULL, ufid, | |
7322 | DPIF_UC_ACTION, userdata, actions, | |
7323 | NULL); | |
7324 | if (!error || error == ENOSPC) { | |
72c84bc2 | 7325 | dp_packet_batch_init_packet(&b, packet); |
7d7ded7a | 7326 | dp_netdev_execute_actions(pmd, &b, should_steal, flow, |
b010be17 | 7327 | actions->data, actions->size); |
7d7ded7a | 7328 | } else if (should_steal) { |
66525ef3 | 7329 | dp_packet_delete(packet); |
a13a0209 | 7330 | COVERAGE_INC(datapath_drop_userspace_action_error); |
66525ef3 PS |
7331 | } |
7332 | } | |
7333 | ||
9df65060 VDA |
7334 | static bool |
7335 | dp_execute_output_action(struct dp_netdev_pmd_thread *pmd, | |
7336 | struct dp_packet_batch *packets_, | |
7337 | bool should_steal, odp_port_t port_no) | |
7338 | { | |
7339 | struct tx_port *p = pmd_send_port_cache_lookup(pmd, port_no); | |
7340 | struct dp_packet_batch out; | |
7341 | ||
7342 | if (!OVS_LIKELY(p)) { | |
7343 | COVERAGE_ADD(datapath_drop_invalid_port, | |
7344 | dp_packet_batch_size(packets_)); | |
7345 | dp_packet_delete_batch(packets_, should_steal); | |
7346 | return false; | |
7347 | } | |
7348 | if (!should_steal) { | |
7349 | dp_packet_batch_clone(&out, packets_); | |
7350 | dp_packet_batch_reset_cutlen(packets_); | |
7351 | packets_ = &out; | |
7352 | } | |
7353 | dp_packet_batch_apply_cutlen(packets_); | |
7354 | #ifdef DPDK_NETDEV | |
7355 | if (OVS_UNLIKELY(!dp_packet_batch_is_empty(&p->output_pkts) | |
7356 | && packets_->packets[0]->source | |
7357 | != p->output_pkts.packets[0]->source)) { | |
7358 | /* XXX: netdev-dpdk assumes that all packets in a single | |
7359 | * output batch has the same source. Flush here to | |
7360 | * avoid memory access issues. */ | |
7361 | dp_netdev_pmd_flush_output_on_port(pmd, p); | |
7362 | } | |
7363 | #endif | |
7364 | if (dp_packet_batch_size(&p->output_pkts) | |
7365 | + dp_packet_batch_size(packets_) > NETDEV_MAX_BURST) { | |
7366 | /* Flush here to avoid overflow. */ | |
7367 | dp_netdev_pmd_flush_output_on_port(pmd, p); | |
7368 | } | |
7369 | if (dp_packet_batch_is_empty(&p->output_pkts)) { | |
7370 | pmd->n_output_batches++; | |
7371 | } | |
7372 | ||
7373 | struct dp_packet *packet; | |
7374 | DP_PACKET_BATCH_FOR_EACH (i, packet, packets_) { | |
7375 | p->output_pkts_rxqs[dp_packet_batch_size(&p->output_pkts)] = | |
7376 | pmd->ctx.last_rxq; | |
7377 | dp_packet_batch_add(&p->output_pkts, packet); | |
7378 | } | |
7379 | return true; | |
7380 | } | |
7381 | ||
7382 | static void | |
7383 | dp_execute_lb_output_action(struct dp_netdev_pmd_thread *pmd, | |
7384 | struct dp_packet_batch *packets_, | |
7385 | bool should_steal, uint32_t bond) | |
7386 | { | |
7387 | struct tx_bond *p_bond = tx_bond_lookup(&pmd->tx_bonds, bond); | |
7388 | struct dp_packet_batch out; | |
7389 | struct dp_packet *packet; | |
7390 | ||
7391 | if (!p_bond) { | |
7392 | COVERAGE_ADD(datapath_drop_invalid_bond, | |
7393 | dp_packet_batch_size(packets_)); | |
7394 | dp_packet_delete_batch(packets_, should_steal); | |
7395 | return; | |
7396 | } | |
7397 | if (!should_steal) { | |
7398 | dp_packet_batch_clone(&out, packets_); | |
7399 | dp_packet_batch_reset_cutlen(packets_); | |
7400 | packets_ = &out; | |
7401 | } | |
7402 | dp_packet_batch_apply_cutlen(packets_); | |
7403 | ||
7404 | DP_PACKET_BATCH_FOR_EACH (i, packet, packets_) { | |
7405 | /* | |
7406 | * Lookup the bond-hash table using hash to get the slave. | |
7407 | */ | |
7408 | uint32_t hash = dp_packet_get_rss_hash(packet); | |
7409 | struct slave_entry *s_entry = &p_bond->slave_buckets[hash & BOND_MASK]; | |
7410 | odp_port_t bond_member = s_entry->slave_id; | |
7411 | uint32_t size = dp_packet_size(packet); | |
7412 | struct dp_packet_batch output_pkt; | |
7413 | ||
7414 | dp_packet_batch_init_packet(&output_pkt, packet); | |
7415 | if (OVS_LIKELY(dp_execute_output_action(pmd, &output_pkt, true, | |
7416 | bond_member))) { | |
7417 | /* Update slave stats. */ | |
7418 | non_atomic_ullong_add(&s_entry->n_packets, 1); | |
7419 | non_atomic_ullong_add(&s_entry->n_bytes, size); | |
7420 | } | |
7421 | } | |
7422 | } | |
7423 | ||
a36de779 | 7424 | static void |
1895cc8d | 7425 | dp_execute_cb(void *aux_, struct dp_packet_batch *packets_, |
7d7ded7a | 7426 | const struct nlattr *a, bool should_steal) |
4b27db64 | 7427 | OVS_NO_THREAD_SAFETY_ANALYSIS |
9080a111 JR |
7428 | { |
7429 | struct dp_netdev_execute_aux *aux = aux_; | |
623540e4 | 7430 | uint32_t *depth = recirc_depth_get(); |
28e2fa02 DDP |
7431 | struct dp_netdev_pmd_thread *pmd = aux->pmd; |
7432 | struct dp_netdev *dp = pmd->dp; | |
09f9da0b | 7433 | int type = nl_attr_type(a); |
d0cca6c3 | 7434 | struct tx_port *p; |
a13a0209 | 7435 | uint32_t packet_count, packets_dropped; |
9080a111 | 7436 | |
09f9da0b JR |
7437 | switch ((enum ovs_action_attr)type) { |
7438 | case OVS_ACTION_ATTR_OUTPUT: | |
9df65060 VDA |
7439 | dp_execute_output_action(pmd, packets_, should_steal, |
7440 | nl_attr_get_odp_port(a)); | |
7441 | return; | |
c71ea3c4 | 7442 | |
9df65060 VDA |
7443 | case OVS_ACTION_ATTR_LB_OUTPUT: |
7444 | dp_execute_lb_output_action(pmd, packets_, should_steal, | |
7445 | nl_attr_get_u32(a)); | |
7446 | return; | |
09f9da0b | 7447 | |
a36de779 | 7448 | case OVS_ACTION_ATTR_TUNNEL_PUSH: |
47e1b3b6 IM |
7449 | if (should_steal) { |
7450 | /* We're requested to push tunnel header, but also we need to take | |
7451 | * the ownership of these packets. Thus, we can avoid performing | |
7452 | * the action, because the caller will not use the result anyway. | |
7453 | * Just break to free the batch. */ | |
7454 | break; | |
a36de779 | 7455 | } |
47e1b3b6 | 7456 | dp_packet_batch_apply_cutlen(packets_); |
a13a0209 AT |
7457 | packet_count = dp_packet_batch_size(packets_); |
7458 | if (push_tnl_action(pmd, a, packets_)) { | |
7459 | COVERAGE_ADD(datapath_drop_tunnel_push_error, | |
7460 | packet_count); | |
7461 | } | |
47e1b3b6 | 7462 | return; |
a36de779 PS |
7463 | |
7464 | case OVS_ACTION_ATTR_TUNNEL_POP: | |
7465 | if (*depth < MAX_RECIRC_DEPTH) { | |
aaca4fe0 | 7466 | struct dp_packet_batch *orig_packets_ = packets_; |
8611f9a4 | 7467 | odp_port_t portno = nl_attr_get_odp_port(a); |
a36de779 | 7468 | |
57eebbb4 | 7469 | p = pmd_tnl_port_cache_lookup(pmd, portno); |
a36de779 | 7470 | if (p) { |
1895cc8d | 7471 | struct dp_packet_batch tnl_pkt; |
a36de779 | 7472 | |
7d7ded7a | 7473 | if (!should_steal) { |
aaca4fe0 WT |
7474 | dp_packet_batch_clone(&tnl_pkt, packets_); |
7475 | packets_ = &tnl_pkt; | |
7476 | dp_packet_batch_reset_cutlen(orig_packets_); | |
a36de779 PS |
7477 | } |
7478 | ||
aaca4fe0 WT |
7479 | dp_packet_batch_apply_cutlen(packets_); |
7480 | ||
a13a0209 | 7481 | packet_count = dp_packet_batch_size(packets_); |
324c8374 | 7482 | netdev_pop_header(p->port->netdev, packets_); |
a13a0209 AT |
7483 | packets_dropped = |
7484 | packet_count - dp_packet_batch_size(packets_); | |
7485 | if (packets_dropped) { | |
7486 | COVERAGE_ADD(datapath_drop_tunnel_pop_error, | |
7487 | packets_dropped); | |
7488 | } | |
72c84bc2 | 7489 | if (dp_packet_batch_is_empty(packets_)) { |
1c8f98d9 PS |
7490 | return; |
7491 | } | |
9235b479 | 7492 | |
72c84bc2 | 7493 | struct dp_packet *packet; |
e883448e | 7494 | DP_PACKET_BATCH_FOR_EACH (i, packet, packets_) { |
72c84bc2 | 7495 | packet->md.in_port.odp_port = portno; |
a36de779 | 7496 | } |
9235b479 PS |
7497 | |
7498 | (*depth)++; | |
7499 | dp_netdev_recirculate(pmd, packets_); | |
7500 | (*depth)--; | |
a36de779 PS |
7501 | return; |
7502 | } | |
a13a0209 AT |
7503 | COVERAGE_ADD(datapath_drop_invalid_tnl_port, |
7504 | dp_packet_batch_size(packets_)); | |
7505 | } else { | |
7506 | COVERAGE_ADD(datapath_drop_recirc_error, | |
7507 | dp_packet_batch_size(packets_)); | |
a36de779 PS |
7508 | } |
7509 | break; | |
7510 | ||
623540e4 EJ |
7511 | case OVS_ACTION_ATTR_USERSPACE: |
7512 | if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) { | |
aaca4fe0 | 7513 | struct dp_packet_batch *orig_packets_ = packets_; |
623540e4 | 7514 | const struct nlattr *userdata; |
aaca4fe0 | 7515 | struct dp_packet_batch usr_pkt; |
623540e4 EJ |
7516 | struct ofpbuf actions; |
7517 | struct flow flow; | |
7af12bd7 | 7518 | ovs_u128 ufid; |
aaca4fe0 | 7519 | bool clone = false; |
4fc65926 | 7520 | |
623540e4 EJ |
7521 | userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA); |
7522 | ofpbuf_init(&actions, 0); | |
8cbf4f47 | 7523 | |
aaca4fe0 | 7524 | if (packets_->trunc) { |
7d7ded7a | 7525 | if (!should_steal) { |
aaca4fe0 WT |
7526 | dp_packet_batch_clone(&usr_pkt, packets_); |
7527 | packets_ = &usr_pkt; | |
aaca4fe0 WT |
7528 | clone = true; |
7529 | dp_packet_batch_reset_cutlen(orig_packets_); | |
7530 | } | |
7531 | ||
7532 | dp_packet_batch_apply_cutlen(packets_); | |
7533 | } | |
7534 | ||
72c84bc2 | 7535 | struct dp_packet *packet; |
e883448e | 7536 | DP_PACKET_BATCH_FOR_EACH (i, packet, packets_) { |
72c84bc2 | 7537 | flow_extract(packet, &flow); |
7a5e0ee7 | 7538 | odp_flow_key_hash(&flow, sizeof flow, &ufid); |
7d7ded7a | 7539 | dp_execute_userspace_action(pmd, packet, should_steal, &flow, |
b010be17 | 7540 | &ufid, &actions, userdata); |
db73f716 | 7541 | } |
aaca4fe0 WT |
7542 | |
7543 | if (clone) { | |
7544 | dp_packet_delete_batch(packets_, true); | |
7545 | } | |
7546 | ||
623540e4 EJ |
7547 | ofpbuf_uninit(&actions); |
7548 | fat_rwlock_unlock(&dp->upcall_rwlock); | |
6b31e073 | 7549 | |
ac8c2081 DDP |
7550 | return; |
7551 | } | |
a13a0209 AT |
7552 | COVERAGE_ADD(datapath_drop_lock_error, |
7553 | dp_packet_batch_size(packets_)); | |
09f9da0b | 7554 | break; |
572f732a | 7555 | |
adcf00ba AZ |
7556 | case OVS_ACTION_ATTR_RECIRC: |
7557 | if (*depth < MAX_RECIRC_DEPTH) { | |
1895cc8d | 7558 | struct dp_packet_batch recirc_pkts; |
572f732a | 7559 | |
7d7ded7a | 7560 | if (!should_steal) { |
1895cc8d PS |
7561 | dp_packet_batch_clone(&recirc_pkts, packets_); |
7562 | packets_ = &recirc_pkts; | |
28e2fa02 | 7563 | } |
8cbf4f47 | 7564 | |
72c84bc2 | 7565 | struct dp_packet *packet; |
e883448e | 7566 | DP_PACKET_BATCH_FOR_EACH (i, packet, packets_) { |
72c84bc2 | 7567 | packet->md.recirc_id = nl_attr_get_u32(a); |
8cbf4f47 | 7568 | } |
28e2fa02 DDP |
7569 | |
7570 | (*depth)++; | |
1895cc8d | 7571 | dp_netdev_recirculate(pmd, packets_); |
adcf00ba AZ |
7572 | (*depth)--; |
7573 | ||
ac8c2081 | 7574 | return; |
adcf00ba | 7575 | } |
ac8c2081 | 7576 | |
a13a0209 AT |
7577 | COVERAGE_ADD(datapath_drop_recirc_error, |
7578 | dp_packet_batch_size(packets_)); | |
ac8c2081 | 7579 | VLOG_WARN("Packet dropped. Max recirculation depth exceeded."); |
572f732a | 7580 | break; |
572f732a | 7581 | |
5cf3edb3 DDP |
7582 | case OVS_ACTION_ATTR_CT: { |
7583 | const struct nlattr *b; | |
a76a37ef | 7584 | bool force = false; |
5cf3edb3 DDP |
7585 | bool commit = false; |
7586 | unsigned int left; | |
7587 | uint16_t zone = 0; | |
2078901a | 7588 | uint32_t tp_id = 0; |
5cf3edb3 DDP |
7589 | const char *helper = NULL; |
7590 | const uint32_t *setmark = NULL; | |
7591 | const struct ovs_key_ct_labels *setlabel = NULL; | |
4cddb1f0 DB |
7592 | struct nat_action_info_t nat_action_info; |
7593 | struct nat_action_info_t *nat_action_info_ref = NULL; | |
7594 | bool nat_config = false; | |
5cf3edb3 DDP |
7595 | |
7596 | NL_ATTR_FOR_EACH_UNSAFE (b, left, nl_attr_get(a), | |
7597 | nl_attr_get_size(a)) { | |
7598 | enum ovs_ct_attr sub_type = nl_attr_type(b); | |
7599 | ||
7600 | switch(sub_type) { | |
b80e259f | 7601 | case OVS_CT_ATTR_FORCE_COMMIT: |
a76a37ef JR |
7602 | force = true; |
7603 | /* fall through. */ | |
5cf3edb3 DDP |
7604 | case OVS_CT_ATTR_COMMIT: |
7605 | commit = true; | |
7606 | break; | |
7607 | case OVS_CT_ATTR_ZONE: | |
7608 | zone = nl_attr_get_u16(b); | |
7609 | break; | |
7610 | case OVS_CT_ATTR_HELPER: | |
7611 | helper = nl_attr_get_string(b); | |
7612 | break; | |
7613 | case OVS_CT_ATTR_MARK: | |
7614 | setmark = nl_attr_get(b); | |
7615 | break; | |
7616 | case OVS_CT_ATTR_LABELS: | |
7617 | setlabel = nl_attr_get(b); | |
7618 | break; | |
8e83854c JR |
7619 | case OVS_CT_ATTR_EVENTMASK: |
7620 | /* Silently ignored, as userspace datapath does not generate | |
7621 | * netlink events. */ | |
7622 | break; | |
ebe62ec1 | 7623 | case OVS_CT_ATTR_TIMEOUT: |
2078901a WT |
7624 | if (!str_to_uint(nl_attr_get_string(b), 10, &tp_id)) { |
7625 | VLOG_WARN("Invalid Timeout Policy ID: %s.", | |
7626 | nl_attr_get_string(b)); | |
7627 | tp_id = DEFAULT_TP_ID; | |
7628 | } | |
ebe62ec1 | 7629 | break; |
4cddb1f0 DB |
7630 | case OVS_CT_ATTR_NAT: { |
7631 | const struct nlattr *b_nest; | |
7632 | unsigned int left_nest; | |
7633 | bool ip_min_specified = false; | |
7634 | bool proto_num_min_specified = false; | |
7635 | bool ip_max_specified = false; | |
7636 | bool proto_num_max_specified = false; | |
7637 | memset(&nat_action_info, 0, sizeof nat_action_info); | |
7638 | nat_action_info_ref = &nat_action_info; | |
7639 | ||
7640 | NL_NESTED_FOR_EACH_UNSAFE (b_nest, left_nest, b) { | |
7641 | enum ovs_nat_attr sub_type_nest = nl_attr_type(b_nest); | |
7642 | ||
7643 | switch (sub_type_nest) { | |
7644 | case OVS_NAT_ATTR_SRC: | |
7645 | case OVS_NAT_ATTR_DST: | |
7646 | nat_config = true; | |
7647 | nat_action_info.nat_action |= | |
7648 | ((sub_type_nest == OVS_NAT_ATTR_SRC) | |
7649 | ? NAT_ACTION_SRC : NAT_ACTION_DST); | |
7650 | break; | |
7651 | case OVS_NAT_ATTR_IP_MIN: | |
7652 | memcpy(&nat_action_info.min_addr, | |
7653 | nl_attr_get(b_nest), | |
7654 | nl_attr_get_size(b_nest)); | |
7655 | ip_min_specified = true; | |
7656 | break; | |
7657 | case OVS_NAT_ATTR_IP_MAX: | |
7658 | memcpy(&nat_action_info.max_addr, | |
7659 | nl_attr_get(b_nest), | |
7660 | nl_attr_get_size(b_nest)); | |
7661 | ip_max_specified = true; | |
7662 | break; | |
7663 | case OVS_NAT_ATTR_PROTO_MIN: | |
7664 | nat_action_info.min_port = | |
7665 | nl_attr_get_u16(b_nest); | |
7666 | proto_num_min_specified = true; | |
7667 | break; | |
7668 | case OVS_NAT_ATTR_PROTO_MAX: | |
7669 | nat_action_info.max_port = | |
7670 | nl_attr_get_u16(b_nest); | |
7671 | proto_num_max_specified = true; | |
7672 | break; | |
7673 | case OVS_NAT_ATTR_PERSISTENT: | |
7674 | case OVS_NAT_ATTR_PROTO_HASH: | |
7675 | case OVS_NAT_ATTR_PROTO_RANDOM: | |
7676 | break; | |
7677 | case OVS_NAT_ATTR_UNSPEC: | |
7678 | case __OVS_NAT_ATTR_MAX: | |
7679 | OVS_NOT_REACHED(); | |
7680 | } | |
7681 | } | |
7682 | ||
7683 | if (ip_min_specified && !ip_max_specified) { | |
7684 | nat_action_info.max_addr = nat_action_info.min_addr; | |
7685 | } | |
7686 | if (proto_num_min_specified && !proto_num_max_specified) { | |
7687 | nat_action_info.max_port = nat_action_info.min_port; | |
7688 | } | |
7689 | if (proto_num_min_specified || proto_num_max_specified) { | |
7690 | if (nat_action_info.nat_action & NAT_ACTION_SRC) { | |
7691 | nat_action_info.nat_action |= NAT_ACTION_SRC_PORT; | |
7692 | } else if (nat_action_info.nat_action & NAT_ACTION_DST) { | |
7693 | nat_action_info.nat_action |= NAT_ACTION_DST_PORT; | |
7694 | } | |
7695 | } | |
7696 | break; | |
7697 | } | |
5cf3edb3 DDP |
7698 | case OVS_CT_ATTR_UNSPEC: |
7699 | case __OVS_CT_ATTR_MAX: | |
7700 | OVS_NOT_REACHED(); | |
7701 | } | |
7702 | } | |
7703 | ||
4cddb1f0 DB |
7704 | /* We won't be able to function properly in this case, hence |
7705 | * complain loudly. */ | |
7706 | if (nat_config && !commit) { | |
7707 | static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 5); | |
7708 | VLOG_WARN_RL(&rl, "NAT specified without commit."); | |
7709 | } | |
7710 | ||
57593fd2 | 7711 | conntrack_execute(dp->conntrack, packets_, aux->flow->dl_type, force, |
bd7d93f8 | 7712 | commit, zone, setmark, setlabel, aux->flow->tp_src, |
b010be17 | 7713 | aux->flow->tp_dst, helper, nat_action_info_ref, |
2078901a | 7714 | pmd->ctx.now / 1000, tp_id); |
07659514 | 7715 | break; |
5cf3edb3 | 7716 | } |
07659514 | 7717 | |
5dddf960 | 7718 | case OVS_ACTION_ATTR_METER: |
4b27db64 | 7719 | dp_netdev_run_meter(pmd->dp, packets_, nl_attr_get_u32(a), |
b010be17 | 7720 | pmd->ctx.now); |
4b27db64 JR |
7721 | break; |
7722 | ||
09f9da0b JR |
7723 | case OVS_ACTION_ATTR_PUSH_VLAN: |
7724 | case OVS_ACTION_ATTR_POP_VLAN: | |
7725 | case OVS_ACTION_ATTR_PUSH_MPLS: | |
7726 | case OVS_ACTION_ATTR_POP_MPLS: | |
7727 | case OVS_ACTION_ATTR_SET: | |
6d670e7f | 7728 | case OVS_ACTION_ATTR_SET_MASKED: |
09f9da0b | 7729 | case OVS_ACTION_ATTR_SAMPLE: |
53e1d6f1 | 7730 | case OVS_ACTION_ATTR_HASH: |
09f9da0b | 7731 | case OVS_ACTION_ATTR_UNSPEC: |
aaca4fe0 | 7732 | case OVS_ACTION_ATTR_TRUNC: |
6fcecb85 YY |
7733 | case OVS_ACTION_ATTR_PUSH_ETH: |
7734 | case OVS_ACTION_ATTR_POP_ETH: | |
535e3acf | 7735 | case OVS_ACTION_ATTR_CLONE: |
f59cb331 YY |
7736 | case OVS_ACTION_ATTR_PUSH_NSH: |
7737 | case OVS_ACTION_ATTR_POP_NSH: | |
1fe178d2 | 7738 | case OVS_ACTION_ATTR_CT_CLEAR: |
5b34f8fc | 7739 | case OVS_ACTION_ATTR_CHECK_PKT_LEN: |
a13a0209 | 7740 | case OVS_ACTION_ATTR_DROP: |
09f9da0b JR |
7741 | case __OVS_ACTION_ATTR_MAX: |
7742 | OVS_NOT_REACHED(); | |
da546e07 | 7743 | } |
ac8c2081 | 7744 | |
7d7ded7a | 7745 | dp_packet_delete_batch(packets_, should_steal); |
98403001 BP |
7746 | } |
7747 | ||
4edb9ae9 | 7748 | static void |
65f13b50 | 7749 | dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd, |
1895cc8d | 7750 | struct dp_packet_batch *packets, |
7d7ded7a | 7751 | bool should_steal, const struct flow *flow, |
b010be17 | 7752 | const struct nlattr *actions, size_t actions_len) |
72865317 | 7753 | { |
b010be17 | 7754 | struct dp_netdev_execute_aux aux = { pmd, flow }; |
9080a111 | 7755 | |
7d7ded7a | 7756 | odp_execute_actions(&aux, packets, should_steal, actions, |
8cbf4f47 | 7757 | actions_len, dp_execute_cb); |
72865317 BP |
7758 | } |
7759 | ||
4d4e68ed DDP |
7760 | struct dp_netdev_ct_dump { |
7761 | struct ct_dpif_dump_state up; | |
7762 | struct conntrack_dump dump; | |
7763 | struct conntrack *ct; | |
7764 | struct dp_netdev *dp; | |
7765 | }; | |
7766 | ||
7767 | static int | |
7768 | dpif_netdev_ct_dump_start(struct dpif *dpif, struct ct_dpif_dump_state **dump_, | |
ded30c74 | 7769 | const uint16_t *pzone, int *ptot_bkts) |
4d4e68ed DDP |
7770 | { |
7771 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
7772 | struct dp_netdev_ct_dump *dump; | |
7773 | ||
7774 | dump = xzalloc(sizeof *dump); | |
7775 | dump->dp = dp; | |
57593fd2 | 7776 | dump->ct = dp->conntrack; |
4d4e68ed | 7777 | |
57593fd2 | 7778 | conntrack_dump_start(dp->conntrack, &dump->dump, pzone, ptot_bkts); |
4d4e68ed DDP |
7779 | |
7780 | *dump_ = &dump->up; | |
7781 | ||
7782 | return 0; | |
7783 | } | |
7784 | ||
7785 | static int | |
7786 | dpif_netdev_ct_dump_next(struct dpif *dpif OVS_UNUSED, | |
7787 | struct ct_dpif_dump_state *dump_, | |
7788 | struct ct_dpif_entry *entry) | |
7789 | { | |
7790 | struct dp_netdev_ct_dump *dump; | |
7791 | ||
7792 | INIT_CONTAINER(dump, dump_, up); | |
7793 | ||
7794 | return conntrack_dump_next(&dump->dump, entry); | |
7795 | } | |
7796 | ||
7797 | static int | |
7798 | dpif_netdev_ct_dump_done(struct dpif *dpif OVS_UNUSED, | |
7799 | struct ct_dpif_dump_state *dump_) | |
7800 | { | |
7801 | struct dp_netdev_ct_dump *dump; | |
7802 | int err; | |
7803 | ||
7804 | INIT_CONTAINER(dump, dump_, up); | |
7805 | ||
7806 | err = conntrack_dump_done(&dump->dump); | |
7807 | ||
7808 | free(dump); | |
7809 | ||
7810 | return err; | |
7811 | } | |
7812 | ||
5d9cbb4c | 7813 | static int |
817a7657 YHW |
7814 | dpif_netdev_ct_flush(struct dpif *dpif, const uint16_t *zone, |
7815 | const struct ct_dpif_tuple *tuple) | |
5d9cbb4c DDP |
7816 | { |
7817 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
7818 | ||
817a7657 | 7819 | if (tuple) { |
57593fd2 | 7820 | return conntrack_flush_tuple(dp->conntrack, tuple, zone ? *zone : 0); |
817a7657 | 7821 | } |
57593fd2 | 7822 | return conntrack_flush(dp->conntrack, zone); |
5d9cbb4c DDP |
7823 | } |
7824 | ||
c92339ad DB |
7825 | static int |
7826 | dpif_netdev_ct_set_maxconns(struct dpif *dpif, uint32_t maxconns) | |
7827 | { | |
7828 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
7829 | ||
57593fd2 | 7830 | return conntrack_set_maxconns(dp->conntrack, maxconns); |
c92339ad DB |
7831 | } |
7832 | ||
7833 | static int | |
7834 | dpif_netdev_ct_get_maxconns(struct dpif *dpif, uint32_t *maxconns) | |
7835 | { | |
7836 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
7837 | ||
57593fd2 | 7838 | return conntrack_get_maxconns(dp->conntrack, maxconns); |
c92339ad DB |
7839 | } |
7840 | ||
875075b3 DB |
7841 | static int |
7842 | dpif_netdev_ct_get_nconns(struct dpif *dpif, uint32_t *nconns) | |
7843 | { | |
7844 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
7845 | ||
57593fd2 | 7846 | return conntrack_get_nconns(dp->conntrack, nconns); |
875075b3 DB |
7847 | } |
7848 | ||
64207120 DB |
7849 | static int |
7850 | dpif_netdev_ct_set_tcp_seq_chk(struct dpif *dpif, bool enabled) | |
7851 | { | |
7852 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
7853 | ||
7854 | return conntrack_set_tcp_seq_chk(dp->conntrack, enabled); | |
7855 | } | |
7856 | ||
7857 | static int | |
7858 | dpif_netdev_ct_get_tcp_seq_chk(struct dpif *dpif, bool *enabled) | |
7859 | { | |
7860 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
7861 | *enabled = conntrack_get_tcp_seq_chk(dp->conntrack); | |
7862 | return 0; | |
7863 | } | |
7864 | ||
a7f33fdb DB |
7865 | static int |
7866 | dpif_netdev_ct_set_limits(struct dpif *dpif OVS_UNUSED, | |
7867 | const uint32_t *default_limits, | |
7868 | const struct ovs_list *zone_limits) | |
7869 | { | |
7870 | int err = 0; | |
7871 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
7872 | if (default_limits) { | |
7873 | err = zone_limit_update(dp->conntrack, DEFAULT_ZONE, *default_limits); | |
7874 | if (err != 0) { | |
7875 | return err; | |
7876 | } | |
7877 | } | |
7878 | ||
7879 | struct ct_dpif_zone_limit *zone_limit; | |
7880 | LIST_FOR_EACH (zone_limit, node, zone_limits) { | |
7881 | err = zone_limit_update(dp->conntrack, zone_limit->zone, | |
7882 | zone_limit->limit); | |
7883 | if (err != 0) { | |
7884 | break; | |
7885 | } | |
7886 | } | |
7887 | return err; | |
7888 | } | |
7889 | ||
7890 | static int | |
7891 | dpif_netdev_ct_get_limits(struct dpif *dpif OVS_UNUSED, | |
7892 | uint32_t *default_limit, | |
7893 | const struct ovs_list *zone_limits_request, | |
7894 | struct ovs_list *zone_limits_reply) | |
7895 | { | |
7896 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
7897 | struct conntrack_zone_limit czl; | |
7898 | ||
7899 | czl = zone_limit_get(dp->conntrack, DEFAULT_ZONE); | |
7900 | if (czl.zone == DEFAULT_ZONE) { | |
7901 | *default_limit = czl.limit; | |
7902 | } else { | |
7903 | return EINVAL; | |
7904 | } | |
7905 | ||
7906 | if (!ovs_list_is_empty(zone_limits_request)) { | |
7907 | struct ct_dpif_zone_limit *zone_limit; | |
7908 | LIST_FOR_EACH (zone_limit, node, zone_limits_request) { | |
7909 | czl = zone_limit_get(dp->conntrack, zone_limit->zone); | |
7910 | if (czl.zone == zone_limit->zone || czl.zone == DEFAULT_ZONE) { | |
7911 | ct_dpif_push_zone_limit(zone_limits_reply, zone_limit->zone, | |
7912 | czl.limit, czl.count); | |
7913 | } else { | |
7914 | return EINVAL; | |
7915 | } | |
7916 | } | |
7917 | } else { | |
7918 | for (int z = MIN_ZONE; z <= MAX_ZONE; z++) { | |
7919 | czl = zone_limit_get(dp->conntrack, z); | |
7920 | if (czl.zone == z) { | |
7921 | ct_dpif_push_zone_limit(zone_limits_reply, z, czl.limit, | |
7922 | czl.count); | |
7923 | } | |
7924 | } | |
7925 | } | |
7926 | ||
7927 | return 0; | |
7928 | } | |
7929 | ||
7930 | static int | |
7931 | dpif_netdev_ct_del_limits(struct dpif *dpif OVS_UNUSED, | |
7932 | const struct ovs_list *zone_limits) | |
7933 | { | |
7934 | int err = 0; | |
7935 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
7936 | struct ct_dpif_zone_limit *zone_limit; | |
7937 | LIST_FOR_EACH (zone_limit, node, zone_limits) { | |
7938 | err = zone_limit_delete(dp->conntrack, zone_limit->zone); | |
7939 | if (err != 0) { | |
7940 | break; | |
7941 | } | |
7942 | } | |
7943 | ||
7944 | return err; | |
7945 | } | |
7946 | ||
2078901a WT |
7947 | static int |
7948 | dpif_netdev_ct_set_timeout_policy(struct dpif *dpif, | |
7949 | const struct ct_dpif_timeout_policy *dpif_tp) | |
7950 | { | |
7951 | struct timeout_policy tp; | |
7952 | struct dp_netdev *dp; | |
7953 | ||
7954 | dp = get_dp_netdev(dpif); | |
7955 | memcpy(&tp.policy, dpif_tp, sizeof tp.policy); | |
7956 | return timeout_policy_update(dp->conntrack, &tp); | |
7957 | } | |
7958 | ||
7959 | static int | |
7960 | dpif_netdev_ct_get_timeout_policy(struct dpif *dpif, uint32_t tp_id, | |
7961 | struct ct_dpif_timeout_policy *dpif_tp) | |
7962 | { | |
7963 | struct timeout_policy *tp; | |
7964 | struct dp_netdev *dp; | |
7965 | int err = 0; | |
7966 | ||
7967 | dp = get_dp_netdev(dpif); | |
7968 | tp = timeout_policy_get(dp->conntrack, tp_id); | |
7969 | if (!tp) { | |
7970 | return ENOENT; | |
7971 | } | |
7972 | memcpy(dpif_tp, &tp->policy, sizeof tp->policy); | |
7973 | return err; | |
7974 | } | |
7975 | ||
7976 | static int | |
7977 | dpif_netdev_ct_del_timeout_policy(struct dpif *dpif, | |
7978 | uint32_t tp_id) | |
7979 | { | |
7980 | struct dp_netdev *dp; | |
7981 | int err = 0; | |
7982 | ||
7983 | dp = get_dp_netdev(dpif); | |
7984 | err = timeout_policy_delete(dp->conntrack, tp_id); | |
7985 | return err; | |
7986 | } | |
7987 | ||
7988 | static int | |
7989 | dpif_netdev_ct_get_timeout_policy_name(struct dpif *dpif OVS_UNUSED, | |
7990 | uint32_t tp_id, | |
7991 | uint16_t dl_type OVS_UNUSED, | |
7992 | uint8_t nw_proto OVS_UNUSED, | |
7993 | char **tp_name, bool *is_generic) | |
7994 | { | |
7995 | struct ds ds = DS_EMPTY_INITIALIZER; | |
7996 | ||
7997 | ds_put_format(&ds, "%"PRIu32, tp_id); | |
7998 | *tp_name = ds_steal_cstr(&ds); | |
7999 | *is_generic = true; | |
8000 | return 0; | |
8001 | } | |
8002 | ||
4ea96698 DB |
8003 | static int |
8004 | dpif_netdev_ipf_set_enabled(struct dpif *dpif, bool v6, bool enable) | |
8005 | { | |
8006 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
57593fd2 | 8007 | return ipf_set_enabled(conntrack_ipf_ctx(dp->conntrack), v6, enable); |
4ea96698 DB |
8008 | } |
8009 | ||
8010 | static int | |
8011 | dpif_netdev_ipf_set_min_frag(struct dpif *dpif, bool v6, uint32_t min_frag) | |
8012 | { | |
8013 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
57593fd2 | 8014 | return ipf_set_min_frag(conntrack_ipf_ctx(dp->conntrack), v6, min_frag); |
4ea96698 DB |
8015 | } |
8016 | ||
8017 | static int | |
8018 | dpif_netdev_ipf_set_max_nfrags(struct dpif *dpif, uint32_t max_frags) | |
8019 | { | |
8020 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
57593fd2 | 8021 | return ipf_set_max_nfrags(conntrack_ipf_ctx(dp->conntrack), max_frags); |
4ea96698 DB |
8022 | } |
8023 | ||
8024 | /* Adjust this function if 'dpif_ipf_status' and 'ipf_status' were to | |
8025 | * diverge. */ | |
8026 | static int | |
8027 | dpif_netdev_ipf_get_status(struct dpif *dpif, | |
8028 | struct dpif_ipf_status *dpif_ipf_status) | |
8029 | { | |
8030 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
57593fd2 | 8031 | ipf_get_status(conntrack_ipf_ctx(dp->conntrack), |
4ea96698 DB |
8032 | (struct ipf_status *) dpif_ipf_status); |
8033 | return 0; | |
8034 | } | |
8035 | ||
8036 | static int | |
8037 | dpif_netdev_ipf_dump_start(struct dpif *dpif OVS_UNUSED, | |
8038 | struct ipf_dump_ctx **ipf_dump_ctx) | |
8039 | { | |
8040 | return ipf_dump_start(ipf_dump_ctx); | |
8041 | } | |
8042 | ||
8043 | static int | |
8044 | dpif_netdev_ipf_dump_next(struct dpif *dpif, void *ipf_dump_ctx, char **dump) | |
8045 | { | |
8046 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
57593fd2 | 8047 | return ipf_dump_next(conntrack_ipf_ctx(dp->conntrack), ipf_dump_ctx, |
4ea96698 DB |
8048 | dump); |
8049 | } | |
8050 | ||
8051 | static int | |
8052 | dpif_netdev_ipf_dump_done(struct dpif *dpif OVS_UNUSED, void *ipf_dump_ctx) | |
8053 | { | |
8054 | return ipf_dump_done(ipf_dump_ctx); | |
8055 | ||
8056 | } | |
8057 | ||
9df65060 VDA |
8058 | static int |
8059 | dpif_netdev_bond_add(struct dpif *dpif, uint32_t bond_id, | |
8060 | odp_port_t *slave_map) | |
8061 | { | |
8062 | struct tx_bond *new_tx = xzalloc(sizeof *new_tx); | |
8063 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
8064 | struct dp_netdev_pmd_thread *pmd; | |
8065 | ||
8066 | /* Prepare new bond mapping. */ | |
8067 | new_tx->bond_id = bond_id; | |
8068 | for (int bucket = 0; bucket < BOND_BUCKETS; bucket++) { | |
8069 | new_tx->slave_buckets[bucket].slave_id = slave_map[bucket]; | |
8070 | } | |
8071 | ||
8072 | ovs_mutex_lock(&dp->bond_mutex); | |
8073 | /* Check if bond already existed. */ | |
8074 | struct tx_bond *old_tx = tx_bond_lookup(&dp->tx_bonds, bond_id); | |
8075 | if (old_tx) { | |
8076 | cmap_replace(&dp->tx_bonds, &old_tx->node, &new_tx->node, | |
8077 | hash_bond_id(bond_id)); | |
8078 | ovsrcu_postpone(free, old_tx); | |
8079 | } else { | |
8080 | cmap_insert(&dp->tx_bonds, &new_tx->node, hash_bond_id(bond_id)); | |
8081 | } | |
8082 | ovs_mutex_unlock(&dp->bond_mutex); | |
8083 | ||
8084 | /* Update all PMDs with new bond mapping. */ | |
8085 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
8086 | dp_netdev_add_bond_tx_to_pmd(pmd, new_tx, true); | |
8087 | } | |
8088 | return 0; | |
8089 | } | |
8090 | ||
8091 | static int | |
8092 | dpif_netdev_bond_del(struct dpif *dpif, uint32_t bond_id) | |
8093 | { | |
8094 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
8095 | struct dp_netdev_pmd_thread *pmd; | |
8096 | struct tx_bond *tx; | |
8097 | ||
8098 | ovs_mutex_lock(&dp->bond_mutex); | |
8099 | /* Check if bond existed. */ | |
8100 | tx = tx_bond_lookup(&dp->tx_bonds, bond_id); | |
8101 | if (tx) { | |
8102 | cmap_remove(&dp->tx_bonds, &tx->node, hash_bond_id(bond_id)); | |
8103 | ovsrcu_postpone(free, tx); | |
8104 | } else { | |
8105 | /* Bond is not present. */ | |
8106 | ovs_mutex_unlock(&dp->bond_mutex); | |
8107 | return ENOENT; | |
8108 | } | |
8109 | ovs_mutex_unlock(&dp->bond_mutex); | |
8110 | ||
8111 | /* Remove the bond map in all pmds. */ | |
8112 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
8113 | dp_netdev_del_bond_tx_from_pmd(pmd, bond_id); | |
8114 | } | |
8115 | return 0; | |
8116 | } | |
8117 | ||
8118 | static int | |
8119 | dpif_netdev_bond_stats_get(struct dpif *dpif, uint32_t bond_id, | |
8120 | uint64_t *n_bytes) | |
8121 | { | |
8122 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
8123 | struct dp_netdev_pmd_thread *pmd; | |
8124 | ||
8125 | if (!tx_bond_lookup(&dp->tx_bonds, bond_id)) { | |
8126 | return ENOENT; | |
8127 | } | |
8128 | ||
8129 | /* Search the bond in all PMDs. */ | |
8130 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
8131 | struct tx_bond *pmd_bond_entry | |
8132 | = tx_bond_lookup(&pmd->tx_bonds, bond_id); | |
8133 | ||
8134 | if (!pmd_bond_entry) { | |
8135 | continue; | |
8136 | } | |
8137 | ||
8138 | /* Read bond stats. */ | |
8139 | for (int i = 0; i < BOND_BUCKETS; i++) { | |
8140 | uint64_t pmd_n_bytes; | |
8141 | ||
8142 | atomic_read_relaxed(&pmd_bond_entry->slave_buckets[i].n_bytes, | |
8143 | &pmd_n_bytes); | |
8144 | n_bytes[i] += pmd_n_bytes; | |
8145 | } | |
8146 | } | |
8147 | return 0; | |
8148 | } | |
8149 | ||
72865317 | 8150 | const struct dpif_class dpif_netdev_class = { |
72865317 | 8151 | "netdev", |
f87c1357 | 8152 | true, /* cleanup_required */ |
6553d06b | 8153 | dpif_netdev_init, |
2197d7ab | 8154 | dpif_netdev_enumerate, |
0aeaabc8 | 8155 | dpif_netdev_port_open_type, |
72865317 BP |
8156 | dpif_netdev_open, |
8157 | dpif_netdev_close, | |
7dab847a | 8158 | dpif_netdev_destroy, |
e4cfed38 PS |
8159 | dpif_netdev_run, |
8160 | dpif_netdev_wait, | |
72865317 | 8161 | dpif_netdev_get_stats, |
dcdcad68 | 8162 | NULL, /* set_features */ |
72865317 BP |
8163 | dpif_netdev_port_add, |
8164 | dpif_netdev_port_del, | |
3eb67853 | 8165 | dpif_netdev_port_set_config, |
72865317 BP |
8166 | dpif_netdev_port_query_by_number, |
8167 | dpif_netdev_port_query_by_name, | |
98403001 | 8168 | NULL, /* port_get_pid */ |
b0ec0f27 BP |
8169 | dpif_netdev_port_dump_start, |
8170 | dpif_netdev_port_dump_next, | |
8171 | dpif_netdev_port_dump_done, | |
72865317 BP |
8172 | dpif_netdev_port_poll, |
8173 | dpif_netdev_port_poll_wait, | |
72865317 | 8174 | dpif_netdev_flow_flush, |
ac64794a BP |
8175 | dpif_netdev_flow_dump_create, |
8176 | dpif_netdev_flow_dump_destroy, | |
8177 | dpif_netdev_flow_dump_thread_create, | |
8178 | dpif_netdev_flow_dump_thread_destroy, | |
704a1e09 | 8179 | dpif_netdev_flow_dump_next, |
1a0c894a | 8180 | dpif_netdev_operate, |
6b31e073 RW |
8181 | NULL, /* recv_set */ |
8182 | NULL, /* handlers_set */ | |
d4f6865c | 8183 | dpif_netdev_set_config, |
5bf93d67 | 8184 | dpif_netdev_queue_to_priority, |
6b31e073 RW |
8185 | NULL, /* recv */ |
8186 | NULL, /* recv_wait */ | |
8187 | NULL, /* recv_purge */ | |
e4e74c3a | 8188 | dpif_netdev_register_dp_purge_cb, |
6b31e073 RW |
8189 | dpif_netdev_register_upcall_cb, |
8190 | dpif_netdev_enable_upcall, | |
8191 | dpif_netdev_disable_upcall, | |
b5cbbcf6 | 8192 | dpif_netdev_get_datapath_version, |
4d4e68ed DDP |
8193 | dpif_netdev_ct_dump_start, |
8194 | dpif_netdev_ct_dump_next, | |
8195 | dpif_netdev_ct_dump_done, | |
5d9cbb4c | 8196 | dpif_netdev_ct_flush, |
c92339ad DB |
8197 | dpif_netdev_ct_set_maxconns, |
8198 | dpif_netdev_ct_get_maxconns, | |
875075b3 | 8199 | dpif_netdev_ct_get_nconns, |
64207120 DB |
8200 | dpif_netdev_ct_set_tcp_seq_chk, |
8201 | dpif_netdev_ct_get_tcp_seq_chk, | |
a7f33fdb DB |
8202 | dpif_netdev_ct_set_limits, |
8203 | dpif_netdev_ct_get_limits, | |
8204 | dpif_netdev_ct_del_limits, | |
2078901a WT |
8205 | dpif_netdev_ct_set_timeout_policy, |
8206 | dpif_netdev_ct_get_timeout_policy, | |
8207 | dpif_netdev_ct_del_timeout_policy, | |
1f161318 YHW |
8208 | NULL, /* ct_timeout_policy_dump_start */ |
8209 | NULL, /* ct_timeout_policy_dump_next */ | |
8210 | NULL, /* ct_timeout_policy_dump_done */ | |
2078901a | 8211 | dpif_netdev_ct_get_timeout_policy_name, |
4ea96698 DB |
8212 | dpif_netdev_ipf_set_enabled, |
8213 | dpif_netdev_ipf_set_min_frag, | |
8214 | dpif_netdev_ipf_set_max_nfrags, | |
8215 | dpif_netdev_ipf_get_status, | |
8216 | dpif_netdev_ipf_dump_start, | |
8217 | dpif_netdev_ipf_dump_next, | |
8218 | dpif_netdev_ipf_dump_done, | |
5dddf960 JR |
8219 | dpif_netdev_meter_get_features, |
8220 | dpif_netdev_meter_set, | |
8221 | dpif_netdev_meter_get, | |
8222 | dpif_netdev_meter_del, | |
9df65060 VDA |
8223 | dpif_netdev_bond_add, |
8224 | dpif_netdev_bond_del, | |
8225 | dpif_netdev_bond_stats_get, | |
72865317 | 8226 | }; |
614c4892 | 8227 | |
74cc3969 BP |
8228 | static void |
8229 | dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED, | |
8230 | const char *argv[], void *aux OVS_UNUSED) | |
8231 | { | |
e9985d6a | 8232 | struct dp_netdev_port *port; |
74cc3969 | 8233 | struct dp_netdev *dp; |
ff073a71 | 8234 | odp_port_t port_no; |
74cc3969 | 8235 | |
8a4e3a85 | 8236 | ovs_mutex_lock(&dp_netdev_mutex); |
74cc3969 BP |
8237 | dp = shash_find_data(&dp_netdevs, argv[1]); |
8238 | if (!dp || !dpif_netdev_class_is_dummy(dp->class)) { | |
8a4e3a85 | 8239 | ovs_mutex_unlock(&dp_netdev_mutex); |
74cc3969 BP |
8240 | unixctl_command_reply_error(conn, "unknown datapath or not a dummy"); |
8241 | return; | |
8242 | } | |
8a4e3a85 BP |
8243 | ovs_refcount_ref(&dp->ref_cnt); |
8244 | ovs_mutex_unlock(&dp_netdev_mutex); | |
74cc3969 | 8245 | |
59e6d833 | 8246 | ovs_mutex_lock(&dp->port_mutex); |
e9985d6a | 8247 | if (get_port_by_name(dp, argv[2], &port)) { |
74cc3969 | 8248 | unixctl_command_reply_error(conn, "unknown port"); |
8a4e3a85 | 8249 | goto exit; |
74cc3969 BP |
8250 | } |
8251 | ||
ff073a71 BP |
8252 | port_no = u32_to_odp(atoi(argv[3])); |
8253 | if (!port_no || port_no == ODPP_NONE) { | |
74cc3969 | 8254 | unixctl_command_reply_error(conn, "bad port number"); |
8a4e3a85 | 8255 | goto exit; |
74cc3969 | 8256 | } |
ff073a71 | 8257 | if (dp_netdev_lookup_port(dp, port_no)) { |
74cc3969 | 8258 | unixctl_command_reply_error(conn, "port number already in use"); |
8a4e3a85 | 8259 | goto exit; |
74cc3969 | 8260 | } |
59e6d833 | 8261 | |
e9985d6a DDP |
8262 | /* Remove port. */ |
8263 | hmap_remove(&dp->ports, &port->node); | |
e32971b8 | 8264 | reconfigure_datapath(dp); |
59e6d833 | 8265 | |
e9985d6a DDP |
8266 | /* Reinsert with new port number. */ |
8267 | port->port_no = port_no; | |
8268 | hmap_insert(&dp->ports, &port->node, hash_port_no(port_no)); | |
e32971b8 | 8269 | reconfigure_datapath(dp); |
59e6d833 | 8270 | |
d33ed218 | 8271 | seq_change(dp->port_seq); |
74cc3969 | 8272 | unixctl_command_reply(conn, NULL); |
8a4e3a85 BP |
8273 | |
8274 | exit: | |
59e6d833 | 8275 | ovs_mutex_unlock(&dp->port_mutex); |
8a4e3a85 | 8276 | dp_netdev_unref(dp); |
74cc3969 BP |
8277 | } |
8278 | ||
0cbfe35d BP |
8279 | static void |
8280 | dpif_dummy_register__(const char *type) | |
8281 | { | |
8282 | struct dpif_class *class; | |
8283 | ||
8284 | class = xmalloc(sizeof *class); | |
8285 | *class = dpif_netdev_class; | |
8286 | class->type = xstrdup(type); | |
8287 | dp_register_provider(class); | |
8288 | } | |
8289 | ||
8420c7ad BP |
8290 | static void |
8291 | dpif_dummy_override(const char *type) | |
8292 | { | |
65d43fdc YT |
8293 | int error; |
8294 | ||
8295 | /* | |
8296 | * Ignore EAFNOSUPPORT to allow --enable-dummy=system with | |
8297 | * a userland-only build. It's useful for testsuite. | |
8298 | */ | |
8299 | error = dp_unregister_provider(type); | |
8300 | if (error == 0 || error == EAFNOSUPPORT) { | |
8420c7ad BP |
8301 | dpif_dummy_register__(type); |
8302 | } | |
8303 | } | |
8304 | ||
614c4892 | 8305 | void |
8420c7ad | 8306 | dpif_dummy_register(enum dummy_level level) |
614c4892 | 8307 | { |
8420c7ad | 8308 | if (level == DUMMY_OVERRIDE_ALL) { |
0cbfe35d BP |
8309 | struct sset types; |
8310 | const char *type; | |
8311 | ||
8312 | sset_init(&types); | |
8313 | dp_enumerate_types(&types); | |
8314 | SSET_FOR_EACH (type, &types) { | |
8420c7ad | 8315 | dpif_dummy_override(type); |
0cbfe35d BP |
8316 | } |
8317 | sset_destroy(&types); | |
8420c7ad BP |
8318 | } else if (level == DUMMY_OVERRIDE_SYSTEM) { |
8319 | dpif_dummy_override("system"); | |
614c4892 | 8320 | } |
0cbfe35d BP |
8321 | |
8322 | dpif_dummy_register__("dummy"); | |
74cc3969 BP |
8323 | |
8324 | unixctl_command_register("dpif-dummy/change-port-number", | |
74467d5c | 8325 | "dp port new-number", |
74cc3969 | 8326 | 3, 3, dpif_dummy_change_port_number, NULL); |
614c4892 | 8327 | } |
0de8783a JR |
8328 | \f |
8329 | /* Datapath Classifier. */ | |
8330 | ||
0fcf0776 ZF |
8331 | static void |
8332 | dpcls_subtable_destroy_cb(struct dpcls_subtable *subtable) | |
8333 | { | |
8334 | cmap_destroy(&subtable->rules); | |
a0b36b39 | 8335 | ovsrcu_postpone(free, subtable->mf_masks); |
0fcf0776 ZF |
8336 | ovsrcu_postpone(free, subtable); |
8337 | } | |
8338 | ||
0de8783a JR |
8339 | /* Initializes 'cls' as a classifier that initially contains no classification |
8340 | * rules. */ | |
8341 | static void | |
8342 | dpcls_init(struct dpcls *cls) | |
8343 | { | |
8344 | cmap_init(&cls->subtables_map); | |
da9cfca6 | 8345 | pvector_init(&cls->subtables); |
0de8783a JR |
8346 | } |
8347 | ||
8348 | static void | |
8349 | dpcls_destroy_subtable(struct dpcls *cls, struct dpcls_subtable *subtable) | |
8350 | { | |
3453b4d6 | 8351 | VLOG_DBG("Destroying subtable %p for in_port %d", subtable, cls->in_port); |
da9cfca6 | 8352 | pvector_remove(&cls->subtables, subtable); |
0de8783a JR |
8353 | cmap_remove(&cls->subtables_map, &subtable->cmap_node, |
8354 | subtable->mask.hash); | |
0fcf0776 | 8355 | ovsrcu_postpone(dpcls_subtable_destroy_cb, subtable); |
0de8783a JR |
8356 | } |
8357 | ||
8358 | /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the | |
8359 | * caller's responsibility. | |
8360 | * May only be called after all the readers have been terminated. */ | |
8361 | static void | |
8362 | dpcls_destroy(struct dpcls *cls) | |
8363 | { | |
8364 | if (cls) { | |
8365 | struct dpcls_subtable *subtable; | |
8366 | ||
8367 | CMAP_FOR_EACH (subtable, cmap_node, &cls->subtables_map) { | |
361d808d | 8368 | ovs_assert(cmap_count(&subtable->rules) == 0); |
0de8783a JR |
8369 | dpcls_destroy_subtable(cls, subtable); |
8370 | } | |
8371 | cmap_destroy(&cls->subtables_map); | |
da9cfca6 | 8372 | pvector_destroy(&cls->subtables); |
0de8783a JR |
8373 | } |
8374 | } | |
8375 | ||
8376 | static struct dpcls_subtable * | |
8377 | dpcls_create_subtable(struct dpcls *cls, const struct netdev_flow_key *mask) | |
8378 | { | |
8379 | struct dpcls_subtable *subtable; | |
8380 | ||
8381 | /* Need to add one. */ | |
caeb4906 JR |
8382 | subtable = xmalloc(sizeof *subtable |
8383 | - sizeof subtable->mask.mf + mask->len); | |
0de8783a | 8384 | cmap_init(&subtable->rules); |
3453b4d6 | 8385 | subtable->hit_cnt = 0; |
0de8783a | 8386 | netdev_flow_key_clone(&subtable->mask, mask); |
aadede3d | 8387 | |
a0b36b39 HH |
8388 | /* The count of bits in the mask defines the space required for masks. |
8389 | * Then call gen_masks() to create the appropriate masks, avoiding the cost | |
8390 | * of doing runtime calculations. */ | |
8391 | uint32_t unit0 = count_1bits(mask->mf.map.bits[0]); | |
8392 | uint32_t unit1 = count_1bits(mask->mf.map.bits[1]); | |
8393 | subtable->mf_bits_set_unit0 = unit0; | |
8394 | subtable->mf_bits_set_unit1 = unit1; | |
8395 | subtable->mf_masks = xmalloc(sizeof(uint64_t) * (unit0 + unit1)); | |
8396 | netdev_flow_key_gen_masks(mask, subtable->mf_masks, unit0, unit1); | |
8397 | ||
f54d8f00 HH |
8398 | /* Probe for a specialized generic lookup function. */ |
8399 | subtable->lookup_func = dpcls_subtable_generic_probe(unit0, unit1); | |
8400 | ||
8401 | /* If not set, assign generic lookup. Generic works for any miniflow. */ | |
8402 | if (!subtable->lookup_func) { | |
8403 | subtable->lookup_func = dpcls_subtable_lookup_generic; | |
8404 | } | |
aadede3d | 8405 | |
0de8783a | 8406 | cmap_insert(&cls->subtables_map, &subtable->cmap_node, mask->hash); |
3453b4d6 | 8407 | /* Add the new subtable at the end of the pvector (with no hits yet) */ |
da9cfca6 | 8408 | pvector_insert(&cls->subtables, subtable, 0); |
84dbfb2b | 8409 | VLOG_DBG("Creating %"PRIuSIZE". subtable %p for in_port %d", |
3453b4d6 | 8410 | cmap_count(&cls->subtables_map), subtable, cls->in_port); |
da9cfca6 | 8411 | pvector_publish(&cls->subtables); |
0de8783a JR |
8412 | |
8413 | return subtable; | |
8414 | } | |
8415 | ||
8416 | static inline struct dpcls_subtable * | |
8417 | dpcls_find_subtable(struct dpcls *cls, const struct netdev_flow_key *mask) | |
8418 | { | |
8419 | struct dpcls_subtable *subtable; | |
8420 | ||
8421 | CMAP_FOR_EACH_WITH_HASH (subtable, cmap_node, mask->hash, | |
8422 | &cls->subtables_map) { | |
8423 | if (netdev_flow_key_equal(&subtable->mask, mask)) { | |
8424 | return subtable; | |
8425 | } | |
8426 | } | |
8427 | return dpcls_create_subtable(cls, mask); | |
8428 | } | |
8429 | ||
3453b4d6 JS |
8430 | |
8431 | /* Periodically sort the dpcls subtable vectors according to hit counts */ | |
8432 | static void | |
8433 | dpcls_sort_subtable_vector(struct dpcls *cls) | |
8434 | { | |
8435 | struct pvector *pvec = &cls->subtables; | |
8436 | struct dpcls_subtable *subtable; | |
8437 | ||
8438 | PVECTOR_FOR_EACH (subtable, pvec) { | |
8439 | pvector_change_priority(pvec, subtable, subtable->hit_cnt); | |
8440 | subtable->hit_cnt = 0; | |
8441 | } | |
8442 | pvector_publish(pvec); | |
8443 | } | |
8444 | ||
8445 | static inline void | |
4809891b KT |
8446 | dp_netdev_pmd_try_optimize(struct dp_netdev_pmd_thread *pmd, |
8447 | struct polled_queue *poll_list, int poll_cnt) | |
3453b4d6 JS |
8448 | { |
8449 | struct dpcls *cls; | |
5bf84282 NK |
8450 | uint64_t tot_idle = 0, tot_proc = 0; |
8451 | unsigned int pmd_load = 0; | |
3453b4d6 | 8452 | |
b010be17 | 8453 | if (pmd->ctx.now > pmd->rxq_next_cycle_store) { |
2a2c67b4 | 8454 | uint64_t curr_tsc; |
5bf84282 NK |
8455 | struct pmd_auto_lb *pmd_alb = &pmd->dp->pmd_alb; |
8456 | if (pmd_alb->is_enabled && !pmd->isolated | |
8457 | && (pmd->perf_stats.counters.n[PMD_CYCLES_ITER_IDLE] >= | |
8458 | pmd->prev_stats[PMD_CYCLES_ITER_IDLE]) | |
8459 | && (pmd->perf_stats.counters.n[PMD_CYCLES_ITER_BUSY] >= | |
8460 | pmd->prev_stats[PMD_CYCLES_ITER_BUSY])) | |
8461 | { | |
8462 | tot_idle = pmd->perf_stats.counters.n[PMD_CYCLES_ITER_IDLE] - | |
8463 | pmd->prev_stats[PMD_CYCLES_ITER_IDLE]; | |
8464 | tot_proc = pmd->perf_stats.counters.n[PMD_CYCLES_ITER_BUSY] - | |
8465 | pmd->prev_stats[PMD_CYCLES_ITER_BUSY]; | |
8466 | ||
8467 | if (tot_proc) { | |
8468 | pmd_load = ((tot_proc * 100) / (tot_idle + tot_proc)); | |
8469 | } | |
8470 | ||
8471 | if (pmd_load >= ALB_PMD_LOAD_THRESHOLD) { | |
8472 | atomic_count_inc(&pmd->pmd_overloaded); | |
8473 | } else { | |
8474 | atomic_count_set(&pmd->pmd_overloaded, 0); | |
8475 | } | |
8476 | } | |
8477 | ||
8478 | pmd->prev_stats[PMD_CYCLES_ITER_IDLE] = | |
8479 | pmd->perf_stats.counters.n[PMD_CYCLES_ITER_IDLE]; | |
8480 | pmd->prev_stats[PMD_CYCLES_ITER_BUSY] = | |
8481 | pmd->perf_stats.counters.n[PMD_CYCLES_ITER_BUSY]; | |
8482 | ||
4809891b KT |
8483 | /* Get the cycles that were used to process each queue and store. */ |
8484 | for (unsigned i = 0; i < poll_cnt; i++) { | |
8485 | uint64_t rxq_cyc_curr = dp_netdev_rxq_get_cycles(poll_list[i].rxq, | |
8486 | RXQ_CYCLES_PROC_CURR); | |
8487 | dp_netdev_rxq_set_intrvl_cycles(poll_list[i].rxq, rxq_cyc_curr); | |
8488 | dp_netdev_rxq_set_cycles(poll_list[i].rxq, RXQ_CYCLES_PROC_CURR, | |
8489 | 0); | |
8490 | } | |
2a2c67b4 KT |
8491 | curr_tsc = cycles_counter_update(&pmd->perf_stats); |
8492 | if (pmd->intrvl_tsc_prev) { | |
8493 | /* There is a prev timestamp, store a new intrvl cycle count. */ | |
8494 | atomic_store_relaxed(&pmd->intrvl_cycles, | |
8495 | curr_tsc - pmd->intrvl_tsc_prev); | |
8496 | } | |
8497 | pmd->intrvl_tsc_prev = curr_tsc; | |
4809891b | 8498 | /* Start new measuring interval */ |
b010be17 | 8499 | pmd->rxq_next_cycle_store = pmd->ctx.now + PMD_RXQ_INTERVAL_LEN; |
4809891b KT |
8500 | } |
8501 | ||
b010be17 | 8502 | if (pmd->ctx.now > pmd->next_optimization) { |
3453b4d6 JS |
8503 | /* Try to obtain the flow lock to block out revalidator threads. |
8504 | * If not possible, just try next time. */ | |
8505 | if (!ovs_mutex_trylock(&pmd->flow_mutex)) { | |
8506 | /* Optimize each classifier */ | |
8507 | CMAP_FOR_EACH (cls, node, &pmd->classifiers) { | |
8508 | dpcls_sort_subtable_vector(cls); | |
8509 | } | |
8510 | ovs_mutex_unlock(&pmd->flow_mutex); | |
8511 | /* Start new measuring interval */ | |
b010be17 IM |
8512 | pmd->next_optimization = pmd->ctx.now |
8513 | + DPCLS_OPTIMIZATION_INTERVAL; | |
3453b4d6 JS |
8514 | } |
8515 | } | |
8516 | } | |
8517 | ||
0de8783a JR |
8518 | /* Insert 'rule' into 'cls'. */ |
8519 | static void | |
8520 | dpcls_insert(struct dpcls *cls, struct dpcls_rule *rule, | |
8521 | const struct netdev_flow_key *mask) | |
8522 | { | |
8523 | struct dpcls_subtable *subtable = dpcls_find_subtable(cls, mask); | |
8524 | ||
3453b4d6 | 8525 | /* Refer to subtable's mask, also for later removal. */ |
0de8783a JR |
8526 | rule->mask = &subtable->mask; |
8527 | cmap_insert(&subtable->rules, &rule->cmap_node, rule->flow.hash); | |
8528 | } | |
8529 | ||
8530 | /* Removes 'rule' from 'cls', also destructing the 'rule'. */ | |
8531 | static void | |
8532 | dpcls_remove(struct dpcls *cls, struct dpcls_rule *rule) | |
8533 | { | |
8534 | struct dpcls_subtable *subtable; | |
8535 | ||
8536 | ovs_assert(rule->mask); | |
8537 | ||
3453b4d6 | 8538 | /* Get subtable from reference in rule->mask. */ |
0de8783a | 8539 | INIT_CONTAINER(subtable, rule->mask, mask); |
0de8783a JR |
8540 | if (cmap_remove(&subtable->rules, &rule->cmap_node, rule->flow.hash) |
8541 | == 0) { | |
3453b4d6 | 8542 | /* Delete empty subtable. */ |
0de8783a | 8543 | dpcls_destroy_subtable(cls, subtable); |
da9cfca6 | 8544 | pvector_publish(&cls->subtables); |
0de8783a JR |
8545 | } |
8546 | } | |
8547 | ||
a0b36b39 HH |
8548 | /* Inner loop for mask generation of a unit, see netdev_flow_key_gen_masks. */ |
8549 | static inline void | |
8550 | netdev_flow_key_gen_mask_unit(uint64_t iter, | |
8551 | const uint64_t count, | |
8552 | uint64_t *mf_masks) | |
8553 | { | |
8554 | int i; | |
8555 | for (i = 0; i < count; i++) { | |
8556 | uint64_t lowest_bit = (iter & -iter); | |
8557 | iter &= ~lowest_bit; | |
8558 | mf_masks[i] = (lowest_bit - 1); | |
8559 | } | |
8560 | /* Checks that count has covered all bits in the iter bitmap. */ | |
8561 | ovs_assert(iter == 0); | |
8562 | } | |
8563 | ||
8564 | /* Generate a mask for each block in the miniflow, based on the bits set. This | |
8565 | * allows easily masking packets with the generated array here, without | |
8566 | * calculations. This replaces runtime-calculating the masks. | |
8567 | * @param key The table to generate the mf_masks for | |
8568 | * @param mf_masks Pointer to a u64 array of at least *mf_bits* in size | |
8569 | * @param mf_bits_total Number of bits set in the whole miniflow (both units) | |
8570 | * @param mf_bits_unit0 Number of bits set in unit0 of the miniflow | |
8571 | */ | |
8572 | void | |
8573 | netdev_flow_key_gen_masks(const struct netdev_flow_key *tbl, | |
8574 | uint64_t *mf_masks, | |
8575 | const uint32_t mf_bits_u0, | |
8576 | const uint32_t mf_bits_u1) | |
8577 | { | |
8578 | uint64_t iter_u0 = tbl->mf.map.bits[0]; | |
8579 | uint64_t iter_u1 = tbl->mf.map.bits[1]; | |
8580 | ||
8581 | netdev_flow_key_gen_mask_unit(iter_u0, mf_bits_u0, &mf_masks[0]); | |
8582 | netdev_flow_key_gen_mask_unit(iter_u1, mf_bits_u1, &mf_masks[mf_bits_u0]); | |
8583 | } | |
8584 | ||
361d808d JR |
8585 | /* Returns true if 'target' satisfies 'key' in 'mask', that is, if each 1-bit |
8586 | * in 'mask' the values in 'key' and 'target' are the same. */ | |
f5ace7cd | 8587 | bool |
0de8783a JR |
8588 | dpcls_rule_matches_key(const struct dpcls_rule *rule, |
8589 | const struct netdev_flow_key *target) | |
8590 | { | |
09b0fa9c JR |
8591 | const uint64_t *keyp = miniflow_get_values(&rule->flow.mf); |
8592 | const uint64_t *maskp = miniflow_get_values(&rule->mask->mf); | |
5fcff47b | 8593 | uint64_t value; |
0de8783a | 8594 | |
5fcff47b JR |
8595 | NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value, target, rule->flow.mf.map) { |
8596 | if (OVS_UNLIKELY((value & *maskp++) != *keyp++)) { | |
0de8783a JR |
8597 | return false; |
8598 | } | |
8599 | } | |
8600 | return true; | |
8601 | } | |
8602 | ||
5b1c9c78 FA |
8603 | /* For each miniflow in 'keys' performs a classifier lookup writing the result |
8604 | * into the corresponding slot in 'rules'. If a particular entry in 'keys' is | |
0de8783a JR |
8605 | * NULL it is skipped. |
8606 | * | |
8607 | * This function is optimized for use in the userspace datapath and therefore | |
8608 | * does not implement a lot of features available in the standard | |
8609 | * classifier_lookup() function. Specifically, it does not implement | |
8610 | * priorities, instead returning any rule which matches the flow. | |
8611 | * | |
5b1c9c78 | 8612 | * Returns true if all miniflows found a corresponding rule. */ |
0de8783a | 8613 | static bool |
60d8ccae | 8614 | dpcls_lookup(struct dpcls *cls, const struct netdev_flow_key *keys[], |
3453b4d6 JS |
8615 | struct dpcls_rule **rules, const size_t cnt, |
8616 | int *num_lookups_p) | |
0de8783a | 8617 | { |
5b1c9c78 | 8618 | /* The received 'cnt' miniflows are the search-keys that will be processed |
63906f18 BB |
8619 | * to find a matching entry into the available subtables. |
8620 | * The number of bits in map_type is equal to NETDEV_MAX_BURST. */ | |
aadede3d | 8621 | #define MAP_BITS (sizeof(uint32_t) * CHAR_BIT) |
63906f18 | 8622 | BUILD_ASSERT_DECL(MAP_BITS >= NETDEV_MAX_BURST); |
0de8783a | 8623 | |
0de8783a | 8624 | struct dpcls_subtable *subtable; |
aadede3d | 8625 | uint32_t keys_map = TYPE_MAXIMUM(uint32_t); /* Set all bits. */ |
63906f18 BB |
8626 | |
8627 | if (cnt != MAP_BITS) { | |
8628 | keys_map >>= MAP_BITS - cnt; /* Clear extra bits. */ | |
0de8783a JR |
8629 | } |
8630 | memset(rules, 0, cnt * sizeof *rules); | |
8631 | ||
3453b4d6 | 8632 | int lookups_match = 0, subtable_pos = 1; |
aadede3d | 8633 | uint32_t found_map; |
3453b4d6 | 8634 | |
5b1c9c78 FA |
8635 | /* The Datapath classifier - aka dpcls - is composed of subtables. |
8636 | * Subtables are dynamically created as needed when new rules are inserted. | |
8637 | * Each subtable collects rules with matches on a specific subset of packet | |
8638 | * fields as defined by the subtable's mask. We proceed to process every | |
8639 | * search-key against each subtable, but when a match is found for a | |
8640 | * search-key, the search for that key can stop because the rules are | |
8641 | * non-overlapping. */ | |
da9cfca6 | 8642 | PVECTOR_FOR_EACH (subtable, &cls->subtables) { |
aadede3d HH |
8643 | /* Call the subtable specific lookup function. */ |
8644 | found_map = subtable->lookup_func(subtable, keys_map, keys, rules); | |
63906f18 | 8645 | |
aadede3d HH |
8646 | /* Count the number of subtables searched for this packet match. This |
8647 | * estimates the "spread" of subtables looked at per matched packet. */ | |
8648 | uint32_t pkts_matched = count_1bits(found_map); | |
8649 | lookups_match += pkts_matched * subtable_pos; | |
63906f18 | 8650 | |
aadede3d HH |
8651 | /* Clear the found rules, and return early if all packets are found. */ |
8652 | keys_map &= ~found_map; | |
63906f18 | 8653 | if (!keys_map) { |
3453b4d6 JS |
8654 | if (num_lookups_p) { |
8655 | *num_lookups_p = lookups_match; | |
8656 | } | |
aadede3d | 8657 | return true; |
0de8783a | 8658 | } |
3453b4d6 JS |
8659 | subtable_pos++; |
8660 | } | |
aadede3d | 8661 | |
3453b4d6 JS |
8662 | if (num_lookups_p) { |
8663 | *num_lookups_p = lookups_match; | |
0de8783a | 8664 | } |
aadede3d | 8665 | return false; |
0de8783a | 8666 | } |