]>
Commit | Line | Data |
---|---|---|
72865317 | 1 | /* |
f582b6df | 2 | * Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014, 2016, 2017 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" |
72865317 | 19 | |
72865317 BP |
20 | #include <ctype.h> |
21 | #include <errno.h> | |
22 | #include <fcntl.h> | |
23 | #include <inttypes.h> | |
7f3adc00 | 24 | #include <net/if.h> |
b2befd5b | 25 | #include <sys/types.h> |
7daedce4 | 26 | #include <netinet/in.h> |
cdee00fd | 27 | #include <stdint.h> |
72865317 BP |
28 | #include <stdlib.h> |
29 | #include <string.h> | |
30 | #include <sys/ioctl.h> | |
7daedce4 | 31 | #include <sys/socket.h> |
72865317 | 32 | #include <sys/stat.h> |
72865317 BP |
33 | #include <unistd.h> |
34 | ||
9f861c91 | 35 | #include "bitmap.h" |
59e6d833 | 36 | #include "cmap.h" |
5cf3edb3 | 37 | #include "conntrack.h" |
7daedce4 | 38 | #include "coverage.h" |
4d4e68ed | 39 | #include "ct-dpif.h" |
72865317 | 40 | #include "csum.h" |
e14deea0 | 41 | #include "dp-packet.h" |
614c4892 | 42 | #include "dpif.h" |
82a48ead | 43 | #include "dpif-netdev-perf.h" |
72865317 | 44 | #include "dpif-provider.h" |
614c4892 | 45 | #include "dummy.h" |
afae68b1 | 46 | #include "fat-rwlock.h" |
72865317 | 47 | #include "flow.h" |
762d146a | 48 | #include "hmapx.h" |
140dd699 | 49 | #include "id-pool.h" |
6c3eee82 | 50 | #include "latch.h" |
72865317 | 51 | #include "netdev.h" |
79f36875 | 52 | #include "netdev-provider.h" |
de281153 | 53 | #include "netdev-vport.h" |
cdee00fd | 54 | #include "netlink.h" |
f094af7b | 55 | #include "odp-execute.h" |
72865317 | 56 | #include "odp-util.h" |
25d436fb BW |
57 | #include "openvswitch/dynamic-string.h" |
58 | #include "openvswitch/list.h" | |
59 | #include "openvswitch/match.h" | |
0d71302e | 60 | #include "openvswitch/ofp-parse.h" |
25d436fb | 61 | #include "openvswitch/ofp-print.h" |
64c96779 | 62 | #include "openvswitch/ofpbuf.h" |
3eb67853 | 63 | #include "openvswitch/shash.h" |
25d436fb | 64 | #include "openvswitch/vlog.h" |
5a034064 | 65 | #include "ovs-numa.h" |
61e7deb1 | 66 | #include "ovs-rcu.h" |
72865317 | 67 | #include "packets.h" |
fd016ae3 | 68 | #include "openvswitch/poll-loop.h" |
0de8783a | 69 | #include "pvector.h" |
26c6b6cd | 70 | #include "random.h" |
d33ed218 | 71 | #include "seq.h" |
3eb67853 | 72 | #include "smap.h" |
0cbfe35d | 73 | #include "sset.h" |
72865317 | 74 | #include "timeval.h" |
53902038 | 75 | #include "tnl-neigh-cache.h" |
7f9b8504 | 76 | #include "tnl-ports.h" |
74cc3969 | 77 | #include "unixctl.h" |
72865317 | 78 | #include "util.h" |
7daedce4 | 79 | |
d98e6007 | 80 | VLOG_DEFINE_THIS_MODULE(dpif_netdev); |
72865317 | 81 | |
8bb113da | 82 | #define FLOW_DUMP_MAX_BATCH 50 |
adcf00ba | 83 | /* Use per thread recirc_depth to prevent recirculation loop. */ |
3f9d3836 | 84 | #define MAX_RECIRC_DEPTH 6 |
adcf00ba | 85 | DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth, 0) |
e4cfed38 | 86 | |
c71ea3c4 IM |
87 | /* Use instant packet send by default. */ |
88 | #define DEFAULT_TX_FLUSH_INTERVAL 0 | |
89 | ||
72865317 | 90 | /* Configuration parameters. */ |
72865317 | 91 | enum { MAX_FLOWS = 65536 }; /* Maximum number of flows in flow table. */ |
4b27db64 JR |
92 | enum { MAX_METERS = 65536 }; /* Maximum number of meters. */ |
93 | enum { MAX_BANDS = 8 }; /* Maximum number of bands / meter. */ | |
94 | enum { N_METER_LOCKS = 64 }; /* Maximum number of meters. */ | |
72865317 | 95 | |
8a4e3a85 BP |
96 | /* Protects against changes to 'dp_netdevs'. */ |
97 | static struct ovs_mutex dp_netdev_mutex = OVS_MUTEX_INITIALIZER; | |
98 | ||
99 | /* Contains all 'struct dp_netdev's. */ | |
100 | static struct shash dp_netdevs OVS_GUARDED_BY(dp_netdev_mutex) | |
101 | = SHASH_INITIALIZER(&dp_netdevs); | |
102 | ||
623540e4 | 103 | static struct vlog_rate_limit upcall_rl = VLOG_RATE_LIMIT_INIT(600, 600); |
6b31e073 | 104 | |
5cf3edb3 | 105 | #define DP_NETDEV_CS_SUPPORTED_MASK (CS_NEW | CS_ESTABLISHED | CS_RELATED \ |
4cddb1f0 DB |
106 | | CS_INVALID | CS_REPLY_DIR | CS_TRACKED \ |
107 | | CS_SRC_NAT | CS_DST_NAT) | |
5cf3edb3 DDP |
108 | #define DP_NETDEV_CS_UNSUPPORTED_MASK (~(uint32_t)DP_NETDEV_CS_SUPPORTED_MASK) |
109 | ||
2494ccd7 | 110 | static struct odp_support dp_netdev_support = { |
f0fb825a | 111 | .max_vlan_headers = SIZE_MAX, |
2494ccd7 JS |
112 | .max_mpls_depth = SIZE_MAX, |
113 | .recirc = true, | |
5cf3edb3 DDP |
114 | .ct_state = true, |
115 | .ct_zone = true, | |
116 | .ct_mark = true, | |
117 | .ct_label = true, | |
2575df07 JP |
118 | .ct_state_nat = true, |
119 | .ct_orig_tuple = true, | |
120 | .ct_orig_tuple6 = true, | |
2494ccd7 JS |
121 | }; |
122 | ||
79df317f | 123 | /* Stores a miniflow with inline values */ |
9bbf1c3d | 124 | |
9bbf1c3d | 125 | struct netdev_flow_key { |
caeb4906 JR |
126 | uint32_t hash; /* Hash function differs for different users. */ |
127 | uint32_t len; /* Length of the following miniflow (incl. map). */ | |
0de8783a | 128 | struct miniflow mf; |
8fd47924 | 129 | uint64_t buf[FLOW_MAX_PACKET_U64S]; |
9bbf1c3d DDP |
130 | }; |
131 | ||
132 | /* Exact match cache for frequently used flows | |
133 | * | |
134 | * The cache uses a 32-bit hash of the packet (which can be the RSS hash) to | |
135 | * search its entries for a miniflow that matches exactly the miniflow of the | |
0de8783a | 136 | * packet. It stores the 'dpcls_rule' (rule) that matches the miniflow. |
9bbf1c3d DDP |
137 | * |
138 | * A cache entry holds a reference to its 'dp_netdev_flow'. | |
139 | * | |
140 | * A miniflow with a given hash can be in one of EM_FLOW_HASH_SEGS different | |
141 | * entries. The 32-bit hash is split into EM_FLOW_HASH_SEGS values (each of | |
142 | * them is EM_FLOW_HASH_SHIFT bits wide and the remainder is thrown away). Each | |
143 | * value is the index of a cache entry where the miniflow could be. | |
144 | * | |
145 | * | |
146 | * Thread-safety | |
147 | * ============= | |
148 | * | |
149 | * Each pmd_thread has its own private exact match cache. | |
150 | * If dp_netdev_input is not called from a pmd thread, a mutex is used. | |
151 | */ | |
152 | ||
fc82e877 | 153 | #define EM_FLOW_HASH_SHIFT 13 |
9bbf1c3d DDP |
154 | #define EM_FLOW_HASH_ENTRIES (1u << EM_FLOW_HASH_SHIFT) |
155 | #define EM_FLOW_HASH_MASK (EM_FLOW_HASH_ENTRIES - 1) | |
156 | #define EM_FLOW_HASH_SEGS 2 | |
157 | ||
4c30b246 CL |
158 | /* Default EMC insert probability is 1 / DEFAULT_EM_FLOW_INSERT_INV_PROB */ |
159 | #define DEFAULT_EM_FLOW_INSERT_INV_PROB 100 | |
160 | #define DEFAULT_EM_FLOW_INSERT_MIN (UINT32_MAX / \ | |
161 | DEFAULT_EM_FLOW_INSERT_INV_PROB) | |
162 | ||
9bbf1c3d | 163 | struct emc_entry { |
9bbf1c3d | 164 | struct dp_netdev_flow *flow; |
0de8783a | 165 | struct netdev_flow_key key; /* key.hash used for emc hash value. */ |
9bbf1c3d DDP |
166 | }; |
167 | ||
168 | struct emc_cache { | |
169 | struct emc_entry entries[EM_FLOW_HASH_ENTRIES]; | |
67ad54cb | 170 | int sweep_idx; /* For emc_cache_slow_sweep(). */ |
9bbf1c3d DDP |
171 | }; |
172 | ||
173 | /* Iterate in the exact match cache through every entry that might contain a | |
174 | * miniflow with hash 'HASH'. */ | |
175 | #define EMC_FOR_EACH_POS_WITH_HASH(EMC, CURRENT_ENTRY, HASH) \ | |
176 | for (uint32_t i__ = 0, srch_hash__ = (HASH); \ | |
177 | (CURRENT_ENTRY) = &(EMC)->entries[srch_hash__ & EM_FLOW_HASH_MASK], \ | |
178 | i__ < EM_FLOW_HASH_SEGS; \ | |
179 | i__++, srch_hash__ >>= EM_FLOW_HASH_SHIFT) | |
0de8783a JR |
180 | \f |
181 | /* Simple non-wildcarding single-priority classifier. */ | |
182 | ||
05f9e707 IM |
183 | /* Time in microseconds between successive optimizations of the dpcls |
184 | * subtable vector */ | |
185 | #define DPCLS_OPTIMIZATION_INTERVAL 1000000LL | |
3453b4d6 | 186 | |
05f9e707 IM |
187 | /* Time in microseconds of the interval in which rxq processing cycles used |
188 | * in rxq to pmd assignments is measured and stored. */ | |
189 | #define PMD_RXQ_INTERVAL_LEN 10000000LL | |
4809891b | 190 | |
c59e759f KT |
191 | /* Number of intervals for which cycles are stored |
192 | * and used during rxq to pmd assignment. */ | |
193 | #define PMD_RXQ_INTERVAL_MAX 6 | |
194 | ||
0de8783a | 195 | struct dpcls { |
3453b4d6 JS |
196 | struct cmap_node node; /* Within dp_netdev_pmd_thread.classifiers */ |
197 | odp_port_t in_port; | |
0de8783a | 198 | struct cmap subtables_map; |
da9cfca6 | 199 | struct pvector subtables; |
0de8783a | 200 | }; |
9bbf1c3d | 201 | |
0de8783a JR |
202 | /* A rule to be inserted to the classifier. */ |
203 | struct dpcls_rule { | |
204 | struct cmap_node cmap_node; /* Within struct dpcls_subtable 'rules'. */ | |
205 | struct netdev_flow_key *mask; /* Subtable's mask. */ | |
206 | struct netdev_flow_key flow; /* Matching key. */ | |
207 | /* 'flow' must be the last field, additional space is allocated here. */ | |
208 | }; | |
209 | ||
210 | static void dpcls_init(struct dpcls *); | |
211 | static void dpcls_destroy(struct dpcls *); | |
3453b4d6 | 212 | static void dpcls_sort_subtable_vector(struct dpcls *); |
0de8783a JR |
213 | static void dpcls_insert(struct dpcls *, struct dpcls_rule *, |
214 | const struct netdev_flow_key *mask); | |
215 | static void dpcls_remove(struct dpcls *, struct dpcls_rule *); | |
3453b4d6 | 216 | static bool dpcls_lookup(struct dpcls *cls, |
0de8783a | 217 | const struct netdev_flow_key keys[], |
3453b4d6 JS |
218 | struct dpcls_rule **rules, size_t cnt, |
219 | int *num_lookups_p); | |
0de8783a | 220 | \f |
4b27db64 JR |
221 | /* Set of supported meter flags */ |
222 | #define DP_SUPPORTED_METER_FLAGS_MASK \ | |
223 | (OFPMF13_STATS | OFPMF13_PKTPS | OFPMF13_KBPS | OFPMF13_BURST) | |
224 | ||
225 | /* Set of supported meter band types */ | |
226 | #define DP_SUPPORTED_METER_BAND_TYPES \ | |
227 | ( 1 << OFPMBT13_DROP ) | |
228 | ||
229 | struct dp_meter_band { | |
230 | struct ofputil_meter_band up; /* type, prec_level, pad, rate, burst_size */ | |
231 | uint32_t bucket; /* In 1/1000 packets (for PKTPS), or in bits (for KBPS) */ | |
232 | uint64_t packet_count; | |
233 | uint64_t byte_count; | |
234 | }; | |
235 | ||
236 | struct dp_meter { | |
237 | uint16_t flags; | |
238 | uint16_t n_bands; | |
239 | uint32_t max_delta_t; | |
240 | uint64_t used; | |
241 | uint64_t packet_count; | |
242 | uint64_t byte_count; | |
243 | struct dp_meter_band bands[]; | |
244 | }; | |
245 | ||
8a4e3a85 BP |
246 | /* Datapath based on the network device interface from netdev.h. |
247 | * | |
248 | * | |
249 | * Thread-safety | |
250 | * ============= | |
251 | * | |
252 | * Some members, marked 'const', are immutable. Accessing other members | |
253 | * requires synchronization, as noted in more detail below. | |
254 | * | |
255 | * Acquisition order is, from outermost to innermost: | |
256 | * | |
257 | * dp_netdev_mutex (global) | |
59e6d833 | 258 | * port_mutex |
d0cca6c3 | 259 | * non_pmd_mutex |
8a4e3a85 | 260 | */ |
72865317 | 261 | struct dp_netdev { |
8a4e3a85 BP |
262 | const struct dpif_class *const class; |
263 | const char *const name; | |
6b31e073 | 264 | struct dpif *dpif; |
6a8267c5 BP |
265 | struct ovs_refcount ref_cnt; |
266 | atomic_flag destroyed; | |
72865317 | 267 | |
8a4e3a85 BP |
268 | /* Ports. |
269 | * | |
e9985d6a DDP |
270 | * Any lookup into 'ports' or any access to the dp_netdev_ports found |
271 | * through 'ports' requires taking 'port_mutex'. */ | |
59e6d833 | 272 | struct ovs_mutex port_mutex; |
e9985d6a | 273 | struct hmap ports; |
d33ed218 | 274 | struct seq *port_seq; /* Incremented whenever a port changes. */ |
6c3eee82 | 275 | |
c71ea3c4 IM |
276 | /* The time that a packet can wait in output batch for sending. */ |
277 | atomic_uint32_t tx_flush_interval; | |
278 | ||
4b27db64 JR |
279 | /* Meters. */ |
280 | struct ovs_mutex meter_locks[N_METER_LOCKS]; | |
281 | struct dp_meter *meters[MAX_METERS]; /* Meter bands. */ | |
4b27db64 | 282 | |
65dcf3da BB |
283 | /* Probability of EMC insertions is a factor of 'emc_insert_min'.*/ |
284 | OVS_ALIGNED_VAR(CACHE_LINE_SIZE) atomic_uint32_t emc_insert_min; | |
79f36875 JS |
285 | /* Enable collection of PMD performance metrics. */ |
286 | atomic_bool pmd_perf_metrics; | |
65dcf3da | 287 | |
6b31e073 RW |
288 | /* Protects access to ofproto-dpif-upcall interface during revalidator |
289 | * thread synchronization. */ | |
290 | struct fat_rwlock upcall_rwlock; | |
623540e4 EJ |
291 | upcall_callback *upcall_cb; /* Callback function for executing upcalls. */ |
292 | void *upcall_aux; | |
6b31e073 | 293 | |
e4e74c3a AW |
294 | /* Callback function for notifying the purging of dp flows (during |
295 | * reseting pmd deletion). */ | |
296 | dp_purge_callback *dp_purge_cb; | |
297 | void *dp_purge_aux; | |
298 | ||
65f13b50 AW |
299 | /* Stores all 'struct dp_netdev_pmd_thread's. */ |
300 | struct cmap poll_threads; | |
140dd699 IM |
301 | /* id pool for per thread static_tx_qid. */ |
302 | struct id_pool *tx_qid_pool; | |
303 | struct ovs_mutex tx_qid_pool_mutex; | |
65f13b50 AW |
304 | |
305 | /* Protects the access of the 'struct dp_netdev_pmd_thread' | |
306 | * instance for non-pmd thread. */ | |
307 | struct ovs_mutex non_pmd_mutex; | |
308 | ||
309 | /* Each pmd thread will store its pointer to | |
310 | * 'struct dp_netdev_pmd_thread' in 'per_pmd_key'. */ | |
311 | ovsthread_key_t per_pmd_key; | |
f2eee189 | 312 | |
a6a426d6 IM |
313 | struct seq *reconfigure_seq; |
314 | uint64_t last_reconfigure_seq; | |
315 | ||
a14b8947 | 316 | /* Cpu mask for pin of pmd threads. */ |
f2eee189 | 317 | char *pmd_cmask; |
6e3c6fa4 | 318 | |
a36de779 | 319 | uint64_t last_tnl_conf_seq; |
5cf3edb3 DDP |
320 | |
321 | struct conntrack conntrack; | |
72865317 BP |
322 | }; |
323 | ||
4b27db64 JR |
324 | static void meter_lock(const struct dp_netdev *dp, uint32_t meter_id) |
325 | OVS_ACQUIRES(dp->meter_locks[meter_id % N_METER_LOCKS]) | |
326 | { | |
327 | ovs_mutex_lock(&dp->meter_locks[meter_id % N_METER_LOCKS]); | |
328 | } | |
329 | ||
330 | static void meter_unlock(const struct dp_netdev *dp, uint32_t meter_id) | |
331 | OVS_RELEASES(dp->meter_locks[meter_id % N_METER_LOCKS]) | |
332 | { | |
333 | ovs_mutex_unlock(&dp->meter_locks[meter_id % N_METER_LOCKS]); | |
334 | } | |
335 | ||
336 | ||
8a4e3a85 | 337 | static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp, |
e9985d6a DDP |
338 | odp_port_t) |
339 | OVS_REQUIRES(dp->port_mutex); | |
ff073a71 | 340 | |
c59e759f KT |
341 | enum rxq_cycles_counter_type { |
342 | RXQ_CYCLES_PROC_CURR, /* Cycles spent successfully polling and | |
343 | processing packets during the current | |
344 | interval. */ | |
345 | RXQ_CYCLES_PROC_HIST, /* Total cycles of all intervals that are used | |
346 | during rxq to pmd assignment. */ | |
347 | RXQ_N_CYCLES | |
348 | }; | |
349 | ||
05f9e707 | 350 | #define XPS_TIMEOUT 500000LL /* In microseconds. */ |
324c8374 | 351 | |
3eb67853 IM |
352 | /* Contained by struct dp_netdev_port's 'rxqs' member. */ |
353 | struct dp_netdev_rxq { | |
947dc567 DDP |
354 | struct dp_netdev_port *port; |
355 | struct netdev_rxq *rx; | |
356 | unsigned core_id; /* Core to which this queue should be | |
357 | pinned. OVS_CORE_UNSPEC if the | |
358 | queue doesn't need to be pinned to a | |
359 | particular core. */ | |
ee42dd70 | 360 | unsigned intrvl_idx; /* Write index for 'cycles_intrvl'. */ |
47a45d86 | 361 | struct dp_netdev_pmd_thread *pmd; /* pmd thread that polls this queue. */ |
79f36875 | 362 | bool is_vhost; /* Is rxq of a vhost port. */ |
c59e759f KT |
363 | |
364 | /* Counters of cycles spent successfully polling and processing pkts. */ | |
365 | atomic_ullong cycles[RXQ_N_CYCLES]; | |
366 | /* We store PMD_RXQ_INTERVAL_MAX intervals of data for an rxq and then | |
367 | sum them to yield the cycles used for an rxq. */ | |
368 | atomic_ullong cycles_intrvl[PMD_RXQ_INTERVAL_MAX]; | |
3eb67853 IM |
369 | }; |
370 | ||
72865317 BP |
371 | /* A port in a netdev-based datapath. */ |
372 | struct dp_netdev_port { | |
35303d71 | 373 | odp_port_t port_no; |
ca62bb16 BB |
374 | bool dynamic_txqs; /* If true XPS will be used. */ |
375 | bool need_reconfigure; /* True if we should reconfigure netdev. */ | |
72865317 | 376 | struct netdev *netdev; |
e9985d6a | 377 | struct hmap_node node; /* Node in dp_netdev's 'ports'. */ |
4b609110 | 378 | struct netdev_saved_flags *sf; |
3eb67853 | 379 | struct dp_netdev_rxq *rxqs; |
85a4f238 | 380 | unsigned n_rxq; /* Number of elements in 'rxqs' */ |
47a45d86 | 381 | unsigned *txq_used; /* Number of threads that use each tx queue. */ |
324c8374 | 382 | struct ovs_mutex txq_used_mutex; |
0cbfe35d | 383 | char *type; /* Port type as requested by user. */ |
3eb67853 | 384 | char *rxq_affinity_list; /* Requested affinity of rx queues. */ |
72865317 BP |
385 | }; |
386 | ||
1c1e46ed AW |
387 | /* Contained by struct dp_netdev_flow's 'stats' member. */ |
388 | struct dp_netdev_flow_stats { | |
eb94da30 DDP |
389 | atomic_llong used; /* Last used time, in monotonic msecs. */ |
390 | atomic_ullong packet_count; /* Number of packets matched. */ | |
391 | atomic_ullong byte_count; /* Number of bytes matched. */ | |
392 | atomic_uint16_t tcp_flags; /* Bitwise-OR of seen tcp_flags values. */ | |
1c1e46ed AW |
393 | }; |
394 | ||
395 | /* A flow in 'dp_netdev_pmd_thread's 'flow_table'. | |
8a4e3a85 BP |
396 | * |
397 | * | |
398 | * Thread-safety | |
399 | * ============= | |
400 | * | |
401 | * Except near the beginning or ending of its lifespan, rule 'rule' belongs to | |
1c1e46ed | 402 | * its pmd thread's classifier. The text below calls this classifier 'cls'. |
8a4e3a85 BP |
403 | * |
404 | * Motivation | |
405 | * ---------- | |
406 | * | |
407 | * The thread safety rules described here for "struct dp_netdev_flow" are | |
408 | * motivated by two goals: | |
409 | * | |
410 | * - Prevent threads that read members of "struct dp_netdev_flow" from | |
411 | * reading bad data due to changes by some thread concurrently modifying | |
412 | * those members. | |
413 | * | |
414 | * - Prevent two threads making changes to members of a given "struct | |
415 | * dp_netdev_flow" from interfering with each other. | |
416 | * | |
417 | * | |
418 | * Rules | |
419 | * ----- | |
420 | * | |
ed79f89a DDP |
421 | * A flow 'flow' may be accessed without a risk of being freed during an RCU |
422 | * grace period. Code that needs to hold onto a flow for a while | |
423 | * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref(). | |
8a4e3a85 BP |
424 | * |
425 | * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the | |
ed79f89a DDP |
426 | * flow from being deleted from 'cls' and it doesn't protect members of 'flow' |
427 | * from modification. | |
8a4e3a85 BP |
428 | * |
429 | * Some members, marked 'const', are immutable. Accessing other members | |
430 | * requires synchronization, as noted in more detail below. | |
431 | */ | |
72865317 | 432 | struct dp_netdev_flow { |
11e5cf1f | 433 | const struct flow flow; /* Unmasked flow that created this entry. */ |
8a4e3a85 | 434 | /* Hash table index by unmasked flow. */ |
1c1e46ed AW |
435 | const struct cmap_node node; /* In owning dp_netdev_pmd_thread's */ |
436 | /* 'flow_table'. */ | |
70e5ed6f | 437 | const ovs_u128 ufid; /* Unique flow identifier. */ |
bd5131ba | 438 | const unsigned pmd_id; /* The 'core_id' of pmd thread owning this */ |
1c1e46ed | 439 | /* flow. */ |
72865317 | 440 | |
ed79f89a DDP |
441 | /* Number of references. |
442 | * The classifier owns one reference. | |
443 | * Any thread trying to keep a rule from being freed should hold its own | |
444 | * reference. */ | |
445 | struct ovs_refcount ref_cnt; | |
446 | ||
11e5cf1f DDP |
447 | bool dead; |
448 | ||
1c1e46ed AW |
449 | /* Statistics. */ |
450 | struct dp_netdev_flow_stats stats; | |
8a4e3a85 | 451 | |
45c626a3 | 452 | /* Actions. */ |
61e7deb1 | 453 | OVSRCU_TYPE(struct dp_netdev_actions *) actions; |
0de8783a | 454 | |
11e5cf1f DDP |
455 | /* While processing a group of input packets, the datapath uses the next |
456 | * member to store a pointer to the output batch for the flow. It is | |
457 | * reset after the batch has been sent out (See dp_netdev_queue_batches(), | |
f7ce4811 PS |
458 | * packet_batch_per_flow_init() and packet_batch_per_flow_execute()). */ |
459 | struct packet_batch_per_flow *batch; | |
11e5cf1f | 460 | |
0de8783a JR |
461 | /* Packet classification. */ |
462 | struct dpcls_rule cr; /* In owning dp_netdev's 'cls'. */ | |
463 | /* 'cr' must be the last member. */ | |
72865317 BP |
464 | }; |
465 | ||
ed79f89a | 466 | static void dp_netdev_flow_unref(struct dp_netdev_flow *); |
9bbf1c3d | 467 | static bool dp_netdev_flow_ref(struct dp_netdev_flow *); |
70e5ed6f | 468 | static int dpif_netdev_flow_from_nlattrs(const struct nlattr *, uint32_t, |
f0fb825a | 469 | struct flow *, bool); |
8a4e3a85 | 470 | |
a84cb64a BP |
471 | /* A set of datapath actions within a "struct dp_netdev_flow". |
472 | * | |
473 | * | |
474 | * Thread-safety | |
475 | * ============= | |
476 | * | |
45c626a3 | 477 | * A struct dp_netdev_actions 'actions' is protected with RCU. */ |
a84cb64a | 478 | struct dp_netdev_actions { |
a84cb64a BP |
479 | /* These members are immutable: they do not change during the struct's |
480 | * lifetime. */ | |
a84cb64a | 481 | unsigned int size; /* Size of 'actions', in bytes. */ |
9ff55ae2 | 482 | struct nlattr actions[]; /* Sequence of OVS_ACTION_ATTR_* attributes. */ |
a84cb64a BP |
483 | }; |
484 | ||
485 | struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *, | |
486 | size_t); | |
61e7deb1 BP |
487 | struct dp_netdev_actions *dp_netdev_flow_get_actions( |
488 | const struct dp_netdev_flow *); | |
489 | static void dp_netdev_actions_free(struct dp_netdev_actions *); | |
a84cb64a | 490 | |
947dc567 | 491 | struct polled_queue { |
922b28d4 | 492 | struct dp_netdev_rxq *rxq; |
947dc567 DDP |
493 | odp_port_t port_no; |
494 | }; | |
495 | ||
ae7ad0a1 IM |
496 | /* Contained by struct dp_netdev_pmd_thread's 'poll_list' member. */ |
497 | struct rxq_poll { | |
947dc567 DDP |
498 | struct dp_netdev_rxq *rxq; |
499 | struct hmap_node node; | |
ae7ad0a1 IM |
500 | }; |
501 | ||
57eebbb4 DDP |
502 | /* Contained by struct dp_netdev_pmd_thread's 'send_port_cache', |
503 | * 'tnl_port_cache' or 'tx_ports'. */ | |
d0cca6c3 | 504 | struct tx_port { |
324c8374 IM |
505 | struct dp_netdev_port *port; |
506 | int qid; | |
507 | long long last_used; | |
d0cca6c3 | 508 | struct hmap_node node; |
c71ea3c4 | 509 | long long flush_time; |
009e0033 | 510 | struct dp_packet_batch output_pkts; |
58ed6df0 | 511 | struct dp_netdev_rxq *output_pkts_rxqs[NETDEV_MAX_BURST]; |
d0cca6c3 DDP |
512 | }; |
513 | ||
b010be17 IM |
514 | /* A set of properties for the current processing loop that is not directly |
515 | * associated with the pmd thread itself, but with the packets being | |
516 | * processed or the short-term system configuration (for example, time). | |
517 | * Contained by struct dp_netdev_pmd_thread's 'ctx' member. */ | |
518 | struct dp_netdev_pmd_thread_ctx { | |
519 | /* Latest measured time. See 'pmd_thread_ctx_time_update()'. */ | |
520 | long long now; | |
58ed6df0 IM |
521 | /* RX queue from which last packet was received. */ |
522 | struct dp_netdev_rxq *last_rxq; | |
d0cca6c3 DDP |
523 | }; |
524 | ||
e4cfed38 PS |
525 | /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate |
526 | * the performance overhead of interrupt processing. Therefore netdev can | |
527 | * not implement rx-wait for these devices. dpif-netdev needs to poll | |
528 | * these device to check for recv buffer. pmd-thread does polling for | |
1c1e46ed | 529 | * devices assigned to itself. |
e4cfed38 PS |
530 | * |
531 | * DPDK used PMD for accessing NIC. | |
532 | * | |
65f13b50 AW |
533 | * Note, instance with cpu core id NON_PMD_CORE_ID will be reserved for |
534 | * I/O of all non-pmd threads. There will be no actual thread created | |
535 | * for the instance. | |
1c1e46ed | 536 | * |
1859876c BB |
537 | * Each struct has its own flow cache and classifier per managed ingress port. |
538 | * For packets received on ingress port, a look up is done on corresponding PMD | |
539 | * thread's flow cache and in case of a miss, lookup is performed in the | |
540 | * corresponding classifier of port. Packets are executed with the found | |
541 | * actions in either case. | |
1c1e46ed | 542 | * */ |
65f13b50 | 543 | struct dp_netdev_pmd_thread { |
d9d73f84 IM |
544 | struct dp_netdev *dp; |
545 | struct ovs_refcount ref_cnt; /* Every reference must be refcount'ed. */ | |
546 | struct cmap_node node; /* In 'dp->poll_threads'. */ | |
547 | ||
548 | pthread_cond_t cond; /* For synchronizing pmd thread reload. */ | |
549 | struct ovs_mutex cond_mutex; /* Mutex for condition variable. */ | |
accf8626 | 550 | |
65f13b50 AW |
551 | /* Per thread exact-match cache. Note, the instance for cpu core |
552 | * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly | |
d0cca6c3 DDP |
553 | * need to be protected by 'non_pmd_mutex'. Every other instance |
554 | * will only be accessed by its own pmd thread. */ | |
d9d73f84 | 555 | struct emc_cache flow_cache; |
1c1e46ed | 556 | |
3453b4d6 | 557 | /* Flow-Table and classifiers |
1c1e46ed AW |
558 | * |
559 | * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding | |
3453b4d6 JS |
560 | * changes to 'classifiers' must be made while still holding the |
561 | * 'flow_mutex'. | |
1c1e46ed AW |
562 | */ |
563 | struct ovs_mutex flow_mutex; | |
d9d73f84 IM |
564 | struct cmap flow_table OVS_GUARDED; /* Flow table. */ |
565 | ||
566 | /* One classifier per in_port polled by the pmd */ | |
567 | struct cmap classifiers; | |
568 | /* Periodically sort subtable vectors according to hit frequencies */ | |
569 | long long int next_optimization; | |
570 | /* End of the next time interval for which processing cycles | |
571 | are stored for each polled rxq. */ | |
572 | long long int rxq_next_cycle_store; | |
573 | ||
2a2c67b4 KT |
574 | /* Last interval timestamp. */ |
575 | uint64_t intrvl_tsc_prev; | |
576 | /* Last interval cycles. */ | |
577 | atomic_ullong intrvl_cycles; | |
578 | ||
b010be17 IM |
579 | /* Current context of the PMD thread. */ |
580 | struct dp_netdev_pmd_thread_ctx ctx; | |
d9d73f84 IM |
581 | |
582 | struct latch exit_latch; /* For terminating the pmd thread. */ | |
583 | struct seq *reload_seq; | |
584 | uint64_t last_reload_seq; | |
585 | atomic_bool reload; /* Do we need to reload ports? */ | |
586 | pthread_t thread; | |
587 | unsigned core_id; /* CPU core id of this pmd thread. */ | |
588 | int numa_id; /* numa node id of this pmd thread. */ | |
589 | bool isolated; | |
590 | ||
591 | /* Queue id used by this pmd thread to send packets on all netdevs if | |
592 | * XPS disabled for this netdev. All static_tx_qid's are unique and less | |
593 | * than 'cmap_count(dp->poll_threads)'. */ | |
594 | uint32_t static_tx_qid; | |
595 | ||
c71ea3c4 IM |
596 | /* Number of filled output batches. */ |
597 | int n_output_batches; | |
598 | ||
d9d73f84 IM |
599 | struct ovs_mutex port_mutex; /* Mutex for 'poll_list' and 'tx_ports'. */ |
600 | /* List of rx queues to poll. */ | |
601 | struct hmap poll_list OVS_GUARDED; | |
602 | /* Map of 'tx_port's used for transmission. Written by the main thread, | |
603 | * read by the pmd thread. */ | |
604 | struct hmap tx_ports OVS_GUARDED; | |
605 | ||
606 | /* These are thread-local copies of 'tx_ports'. One contains only tunnel | |
607 | * ports (that support push_tunnel/pop_tunnel), the other contains ports | |
608 | * with at least one txq (that support send). A port can be in both. | |
609 | * | |
610 | * There are two separate maps to make sure that we don't try to execute | |
611 | * OUTPUT on a device which has 0 txqs or PUSH/POP on a non-tunnel device. | |
612 | * | |
613 | * The instances for cpu core NON_PMD_CORE_ID can be accessed by multiple | |
614 | * threads, and thusly need to be protected by 'non_pmd_mutex'. Every | |
615 | * other instance will only be accessed by its own pmd thread. */ | |
616 | struct hmap tnl_port_cache; | |
617 | struct hmap send_port_cache; | |
618 | ||
82a48ead JS |
619 | /* Keep track of detailed PMD performance statistics. */ |
620 | struct pmd_perf_stats perf_stats; | |
d9d73f84 IM |
621 | |
622 | /* Set to true if the pmd thread needs to be reloaded. */ | |
623 | bool need_reload; | |
6c3eee82 BP |
624 | }; |
625 | ||
72865317 BP |
626 | /* Interface to netdev-based datapath. */ |
627 | struct dpif_netdev { | |
628 | struct dpif dpif; | |
629 | struct dp_netdev *dp; | |
d33ed218 | 630 | uint64_t last_port_seq; |
72865317 BP |
631 | }; |
632 | ||
8a4e3a85 | 633 | static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no, |
e9985d6a DDP |
634 | struct dp_netdev_port **portp) |
635 | OVS_REQUIRES(dp->port_mutex); | |
8a4e3a85 | 636 | static int get_port_by_name(struct dp_netdev *dp, const char *devname, |
e9985d6a DDP |
637 | struct dp_netdev_port **portp) |
638 | OVS_REQUIRES(dp->port_mutex); | |
8a4e3a85 BP |
639 | static void dp_netdev_free(struct dp_netdev *) |
640 | OVS_REQUIRES(dp_netdev_mutex); | |
8a4e3a85 BP |
641 | static int do_add_port(struct dp_netdev *dp, const char *devname, |
642 | const char *type, odp_port_t port_no) | |
59e6d833 | 643 | OVS_REQUIRES(dp->port_mutex); |
c40b890f | 644 | static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *) |
59e6d833 | 645 | OVS_REQUIRES(dp->port_mutex); |
614c4892 BP |
646 | static int dpif_netdev_open(const struct dpif_class *, const char *name, |
647 | bool create, struct dpif **); | |
65f13b50 | 648 | static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd, |
1895cc8d | 649 | struct dp_packet_batch *, |
7d7ded7a DB |
650 | bool should_steal, |
651 | const struct flow *flow, | |
4edb9ae9 | 652 | const struct nlattr *actions, |
b010be17 | 653 | size_t actions_len); |
65f13b50 | 654 | static void dp_netdev_input(struct dp_netdev_pmd_thread *, |
1895cc8d | 655 | struct dp_packet_batch *, odp_port_t port_no); |
a90ed026 | 656 | static void dp_netdev_recirculate(struct dp_netdev_pmd_thread *, |
1895cc8d | 657 | struct dp_packet_batch *); |
41ccaa24 | 658 | |
6b31e073 | 659 | static void dp_netdev_disable_upcall(struct dp_netdev *); |
ae7ad0a1 | 660 | static void dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd); |
65f13b50 | 661 | static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, |
00873463 DDP |
662 | struct dp_netdev *dp, unsigned core_id, |
663 | int numa_id); | |
1c1e46ed | 664 | static void dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd); |
e9985d6a DDP |
665 | static void dp_netdev_set_nonpmd(struct dp_netdev *dp) |
666 | OVS_REQUIRES(dp->port_mutex); | |
667 | ||
e32971b8 | 668 | static void *pmd_thread_main(void *); |
b19befae | 669 | static struct dp_netdev_pmd_thread *dp_netdev_get_pmd(struct dp_netdev *dp, |
bd5131ba | 670 | unsigned core_id); |
1c1e46ed AW |
671 | static struct dp_netdev_pmd_thread * |
672 | dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos); | |
140dd699 IM |
673 | static void dp_netdev_del_pmd(struct dp_netdev *dp, |
674 | struct dp_netdev_pmd_thread *pmd); | |
e32971b8 | 675 | static void dp_netdev_destroy_all_pmds(struct dp_netdev *dp, bool non_pmd); |
d0cca6c3 | 676 | static void dp_netdev_pmd_clear_ports(struct dp_netdev_pmd_thread *pmd); |
d0cca6c3 | 677 | static void dp_netdev_add_port_tx_to_pmd(struct dp_netdev_pmd_thread *pmd, |
e32971b8 DDP |
678 | struct dp_netdev_port *port) |
679 | OVS_REQUIRES(pmd->port_mutex); | |
680 | static void dp_netdev_del_port_tx_from_pmd(struct dp_netdev_pmd_thread *pmd, | |
681 | struct tx_port *tx) | |
682 | OVS_REQUIRES(pmd->port_mutex); | |
d0cca6c3 | 683 | static void dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread *pmd, |
947dc567 DDP |
684 | struct dp_netdev_rxq *rxq) |
685 | OVS_REQUIRES(pmd->port_mutex); | |
e32971b8 DDP |
686 | static void dp_netdev_del_rxq_from_pmd(struct dp_netdev_pmd_thread *pmd, |
687 | struct rxq_poll *poll) | |
688 | OVS_REQUIRES(pmd->port_mutex); | |
c71ea3c4 IM |
689 | static int |
690 | dp_netdev_pmd_flush_output_packets(struct dp_netdev_pmd_thread *pmd, | |
691 | bool force); | |
009e0033 | 692 | |
e32971b8 | 693 | static void reconfigure_datapath(struct dp_netdev *dp) |
3eb67853 | 694 | OVS_REQUIRES(dp->port_mutex); |
1c1e46ed AW |
695 | static bool dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd); |
696 | static void dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd); | |
697 | static void dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd); | |
d0cca6c3 DDP |
698 | static void pmd_load_cached_ports(struct dp_netdev_pmd_thread *pmd) |
699 | OVS_REQUIRES(pmd->port_mutex); | |
3453b4d6 | 700 | static inline void |
4809891b KT |
701 | dp_netdev_pmd_try_optimize(struct dp_netdev_pmd_thread *pmd, |
702 | struct polled_queue *poll_list, int poll_cnt); | |
703 | static void | |
704 | dp_netdev_rxq_set_cycles(struct dp_netdev_rxq *rx, | |
705 | enum rxq_cycles_counter_type type, | |
706 | unsigned long long cycles); | |
707 | static uint64_t | |
708 | dp_netdev_rxq_get_cycles(struct dp_netdev_rxq *rx, | |
709 | enum rxq_cycles_counter_type type); | |
710 | static void | |
711 | dp_netdev_rxq_set_intrvl_cycles(struct dp_netdev_rxq *rx, | |
712 | unsigned long long cycles); | |
655856ef KT |
713 | static uint64_t |
714 | dp_netdev_rxq_get_intrvl_cycles(struct dp_netdev_rxq *rx, unsigned idx); | |
324c8374 IM |
715 | static void |
716 | dpif_netdev_xps_revalidate_pmd(const struct dp_netdev_pmd_thread *pmd, | |
b010be17 | 717 | bool purge); |
324c8374 | 718 | static int dpif_netdev_xps_get_tx_qid(const struct dp_netdev_pmd_thread *pmd, |
b010be17 | 719 | struct tx_port *tx); |
324c8374 | 720 | |
67ad54cb | 721 | static inline bool emc_entry_alive(struct emc_entry *ce); |
9bbf1c3d DDP |
722 | static void emc_clear_entry(struct emc_entry *ce); |
723 | ||
cd995c73 | 724 | static void dp_netdev_request_reconfigure(struct dp_netdev *dp); |
79f36875 JS |
725 | static inline bool |
726 | pmd_perf_metrics_enabled(const struct dp_netdev_pmd_thread *pmd); | |
cd995c73 | 727 | |
9bbf1c3d DDP |
728 | static void |
729 | emc_cache_init(struct emc_cache *flow_cache) | |
730 | { | |
731 | int i; | |
732 | ||
67ad54cb | 733 | flow_cache->sweep_idx = 0; |
9bbf1c3d DDP |
734 | for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) { |
735 | flow_cache->entries[i].flow = NULL; | |
0de8783a | 736 | flow_cache->entries[i].key.hash = 0; |
09b0fa9c | 737 | flow_cache->entries[i].key.len = sizeof(struct miniflow); |
5fcff47b | 738 | flowmap_init(&flow_cache->entries[i].key.mf.map); |
9bbf1c3d DDP |
739 | } |
740 | } | |
741 | ||
742 | static void | |
743 | emc_cache_uninit(struct emc_cache *flow_cache) | |
744 | { | |
745 | int i; | |
746 | ||
747 | for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) { | |
748 | emc_clear_entry(&flow_cache->entries[i]); | |
749 | } | |
750 | } | |
751 | ||
67ad54cb AW |
752 | /* Check and clear dead flow references slowly (one entry at each |
753 | * invocation). */ | |
754 | static void | |
755 | emc_cache_slow_sweep(struct emc_cache *flow_cache) | |
756 | { | |
757 | struct emc_entry *entry = &flow_cache->entries[flow_cache->sweep_idx]; | |
758 | ||
759 | if (!emc_entry_alive(entry)) { | |
760 | emc_clear_entry(entry); | |
761 | } | |
762 | flow_cache->sweep_idx = (flow_cache->sweep_idx + 1) & EM_FLOW_HASH_MASK; | |
763 | } | |
764 | ||
b010be17 IM |
765 | /* Updates the time in PMD threads context and should be called in three cases: |
766 | * | |
767 | * 1. PMD structure initialization: | |
768 | * - dp_netdev_configure_pmd() | |
769 | * | |
770 | * 2. Before processing of the new packet batch: | |
771 | * - dpif_netdev_execute() | |
009e0033 | 772 | * - dp_netdev_process_rxq_port() |
b010be17 IM |
773 | * |
774 | * 3. At least once per polling iteration in main polling threads if no | |
775 | * packets received on current iteration: | |
776 | * - dpif_netdev_run() | |
777 | * - pmd_thread_main() | |
778 | * | |
779 | * 'pmd->ctx.now' should be used without update in all other cases if possible. | |
780 | */ | |
781 | static inline void | |
782 | pmd_thread_ctx_time_update(struct dp_netdev_pmd_thread *pmd) | |
783 | { | |
05f9e707 | 784 | pmd->ctx.now = time_usec(); |
b010be17 IM |
785 | } |
786 | ||
c4ea7529 BP |
787 | /* Returns true if 'dpif' is a netdev or dummy dpif, false otherwise. */ |
788 | bool | |
789 | dpif_is_netdev(const struct dpif *dpif) | |
790 | { | |
791 | return dpif->dpif_class->open == dpif_netdev_open; | |
792 | } | |
793 | ||
72865317 BP |
794 | static struct dpif_netdev * |
795 | dpif_netdev_cast(const struct dpif *dpif) | |
796 | { | |
c4ea7529 | 797 | ovs_assert(dpif_is_netdev(dpif)); |
72865317 BP |
798 | return CONTAINER_OF(dpif, struct dpif_netdev, dpif); |
799 | } | |
800 | ||
801 | static struct dp_netdev * | |
802 | get_dp_netdev(const struct dpif *dpif) | |
803 | { | |
804 | return dpif_netdev_cast(dpif)->dp; | |
805 | } | |
6553d06b DDP |
806 | \f |
807 | enum pmd_info_type { | |
ce179f11 IM |
808 | PMD_INFO_SHOW_STATS, /* Show how cpu cycles are spent. */ |
809 | PMD_INFO_CLEAR_STATS, /* Set the cycles count to 0. */ | |
79f36875 JS |
810 | PMD_INFO_SHOW_RXQ, /* Show poll lists of pmd threads. */ |
811 | PMD_INFO_PERF_SHOW, /* Show pmd performance details. */ | |
6553d06b DDP |
812 | }; |
813 | ||
814 | static void | |
82a48ead | 815 | format_pmd_thread(struct ds *reply, struct dp_netdev_pmd_thread *pmd) |
6553d06b | 816 | { |
6553d06b DDP |
817 | ds_put_cstr(reply, (pmd->core_id == NON_PMD_CORE_ID) |
818 | ? "main thread" : "pmd thread"); | |
6553d06b DDP |
819 | if (pmd->numa_id != OVS_NUMA_UNSPEC) { |
820 | ds_put_format(reply, " numa_id %d", pmd->numa_id); | |
821 | } | |
d5c199ea | 822 | if (pmd->core_id != OVS_CORE_UNSPEC && pmd->core_id != NON_PMD_CORE_ID) { |
bd5131ba | 823 | ds_put_format(reply, " core_id %u", pmd->core_id); |
6553d06b DDP |
824 | } |
825 | ds_put_cstr(reply, ":\n"); | |
82a48ead JS |
826 | } |
827 | ||
828 | static void | |
829 | pmd_info_show_stats(struct ds *reply, | |
830 | struct dp_netdev_pmd_thread *pmd) | |
831 | { | |
832 | uint64_t stats[PMD_N_STATS]; | |
833 | uint64_t total_cycles, total_packets; | |
834 | double passes_per_pkt = 0; | |
835 | double lookups_per_hit = 0; | |
836 | double packets_per_batch = 0; | |
837 | ||
838 | pmd_perf_read_counters(&pmd->perf_stats, stats); | |
839 | total_cycles = stats[PMD_CYCLES_ITER_IDLE] | |
840 | + stats[PMD_CYCLES_ITER_BUSY]; | |
841 | total_packets = stats[PMD_STAT_RECV]; | |
842 | ||
843 | format_pmd_thread(reply, pmd); | |
6553d06b | 844 | |
82a48ead JS |
845 | if (total_packets > 0) { |
846 | passes_per_pkt = (total_packets + stats[PMD_STAT_RECIRC]) | |
847 | / (double) total_packets; | |
cc4891f3 | 848 | } |
82a48ead JS |
849 | if (stats[PMD_STAT_MASKED_HIT] > 0) { |
850 | lookups_per_hit = stats[PMD_STAT_MASKED_LOOKUP] | |
851 | / (double) stats[PMD_STAT_MASKED_HIT]; | |
852 | } | |
853 | if (stats[PMD_STAT_SENT_BATCHES] > 0) { | |
854 | packets_per_batch = stats[PMD_STAT_SENT_PKTS] | |
855 | / (double) stats[PMD_STAT_SENT_BATCHES]; | |
cc4891f3 IM |
856 | } |
857 | ||
6553d06b | 858 | ds_put_format(reply, |
82a48ead JS |
859 | "\tpackets received: %"PRIu64"\n" |
860 | "\tpacket recirculations: %"PRIu64"\n" | |
861 | "\tavg. datapath passes per packet: %.02f\n" | |
862 | "\temc hits: %"PRIu64"\n" | |
863 | "\tmegaflow hits: %"PRIu64"\n" | |
864 | "\tavg. subtable lookups per megaflow hit: %.02f\n" | |
865 | "\tmiss with success upcall: %"PRIu64"\n" | |
866 | "\tmiss with failed upcall: %"PRIu64"\n" | |
867 | "\tavg. packets per output batch: %.02f\n", | |
868 | total_packets, stats[PMD_STAT_RECIRC], | |
869 | passes_per_pkt, stats[PMD_STAT_EXACT_HIT], | |
870 | stats[PMD_STAT_MASKED_HIT], lookups_per_hit, | |
871 | stats[PMD_STAT_MISS], stats[PMD_STAT_LOST], | |
cc4891f3 | 872 | packets_per_batch); |
6553d06b DDP |
873 | |
874 | if (total_cycles == 0) { | |
875 | return; | |
876 | } | |
877 | ||
878 | ds_put_format(reply, | |
82a48ead JS |
879 | "\tidle cycles: %"PRIu64" (%.02f%%)\n" |
880 | "\tprocessing cycles: %"PRIu64" (%.02f%%)\n", | |
881 | stats[PMD_CYCLES_ITER_IDLE], | |
882 | stats[PMD_CYCLES_ITER_IDLE] / (double) total_cycles * 100, | |
883 | stats[PMD_CYCLES_ITER_BUSY], | |
884 | stats[PMD_CYCLES_ITER_BUSY] / (double) total_cycles * 100); | |
6553d06b DDP |
885 | |
886 | if (total_packets == 0) { | |
887 | return; | |
888 | } | |
889 | ||
890 | ds_put_format(reply, | |
82a48ead JS |
891 | "\tavg cycles per packet: %.02f (%"PRIu64"/%"PRIu64")\n", |
892 | total_cycles / (double) total_packets, | |
6553d06b DDP |
893 | total_cycles, total_packets); |
894 | ||
895 | ds_put_format(reply, | |
896 | "\tavg processing cycles per packet: " | |
82a48ead JS |
897 | "%.02f (%"PRIu64"/%"PRIu64")\n", |
898 | stats[PMD_CYCLES_ITER_BUSY] / (double) total_packets, | |
899 | stats[PMD_CYCLES_ITER_BUSY], total_packets); | |
6553d06b DDP |
900 | } |
901 | ||
79f36875 JS |
902 | static void |
903 | pmd_info_show_perf(struct ds *reply, | |
904 | struct dp_netdev_pmd_thread *pmd, | |
905 | struct pmd_perf_params *par) | |
906 | { | |
907 | if (pmd->core_id != NON_PMD_CORE_ID) { | |
908 | char *time_str = | |
909 | xastrftime_msec("%H:%M:%S.###", time_wall_msec(), true); | |
910 | long long now = time_msec(); | |
911 | double duration = (now - pmd->perf_stats.start_ms) / 1000.0; | |
912 | ||
913 | ds_put_cstr(reply, "\n"); | |
914 | ds_put_format(reply, "Time: %s\n", time_str); | |
915 | ds_put_format(reply, "Measurement duration: %.3f s\n", duration); | |
916 | ds_put_cstr(reply, "\n"); | |
917 | format_pmd_thread(reply, pmd); | |
918 | ds_put_cstr(reply, "\n"); | |
919 | pmd_perf_format_overall_stats(reply, &pmd->perf_stats, duration); | |
920 | if (pmd_perf_metrics_enabled(pmd)) { | |
921 | /* Prevent parallel clearing of perf metrics. */ | |
922 | ovs_mutex_lock(&pmd->perf_stats.clear_mutex); | |
923 | if (par->histograms) { | |
924 | ds_put_cstr(reply, "\n"); | |
925 | pmd_perf_format_histograms(reply, &pmd->perf_stats); | |
926 | } | |
927 | if (par->iter_hist_len > 0) { | |
928 | ds_put_cstr(reply, "\n"); | |
929 | pmd_perf_format_iteration_history(reply, &pmd->perf_stats, | |
930 | par->iter_hist_len); | |
931 | } | |
932 | if (par->ms_hist_len > 0) { | |
933 | ds_put_cstr(reply, "\n"); | |
934 | pmd_perf_format_ms_history(reply, &pmd->perf_stats, | |
935 | par->ms_hist_len); | |
936 | } | |
937 | ovs_mutex_unlock(&pmd->perf_stats.clear_mutex); | |
938 | } | |
939 | free(time_str); | |
940 | } | |
941 | } | |
942 | ||
947dc567 DDP |
943 | static int |
944 | compare_poll_list(const void *a_, const void *b_) | |
945 | { | |
946 | const struct rxq_poll *a = a_; | |
947 | const struct rxq_poll *b = b_; | |
948 | ||
949 | const char *namea = netdev_rxq_get_name(a->rxq->rx); | |
950 | const char *nameb = netdev_rxq_get_name(b->rxq->rx); | |
951 | ||
952 | int cmp = strcmp(namea, nameb); | |
953 | if (!cmp) { | |
954 | return netdev_rxq_get_queue_id(a->rxq->rx) | |
955 | - netdev_rxq_get_queue_id(b->rxq->rx); | |
956 | } else { | |
957 | return cmp; | |
958 | } | |
959 | } | |
960 | ||
961 | static void | |
962 | sorted_poll_list(struct dp_netdev_pmd_thread *pmd, struct rxq_poll **list, | |
963 | size_t *n) | |
964 | { | |
965 | struct rxq_poll *ret, *poll; | |
966 | size_t i; | |
967 | ||
968 | *n = hmap_count(&pmd->poll_list); | |
969 | if (!*n) { | |
970 | ret = NULL; | |
971 | } else { | |
972 | ret = xcalloc(*n, sizeof *ret); | |
973 | i = 0; | |
974 | HMAP_FOR_EACH (poll, node, &pmd->poll_list) { | |
975 | ret[i] = *poll; | |
976 | i++; | |
977 | } | |
978 | ovs_assert(i == *n); | |
1cc1b5f6 | 979 | qsort(ret, *n, sizeof *ret, compare_poll_list); |
947dc567 DDP |
980 | } |
981 | ||
947dc567 DDP |
982 | *list = ret; |
983 | } | |
984 | ||
ce179f11 IM |
985 | static void |
986 | pmd_info_show_rxq(struct ds *reply, struct dp_netdev_pmd_thread *pmd) | |
987 | { | |
988 | if (pmd->core_id != NON_PMD_CORE_ID) { | |
947dc567 | 989 | struct rxq_poll *list; |
2a2c67b4 KT |
990 | size_t n_rxq; |
991 | uint64_t total_cycles = 0; | |
ce179f11 | 992 | |
3eb67853 IM |
993 | ds_put_format(reply, |
994 | "pmd thread numa_id %d core_id %u:\n\tisolated : %s\n", | |
995 | pmd->numa_id, pmd->core_id, (pmd->isolated) | |
996 | ? "true" : "false"); | |
ce179f11 | 997 | |
d0cca6c3 | 998 | ovs_mutex_lock(&pmd->port_mutex); |
2a2c67b4 | 999 | sorted_poll_list(pmd, &list, &n_rxq); |
ce179f11 | 1000 | |
2a2c67b4 KT |
1001 | /* Get the total pmd cycles for an interval. */ |
1002 | atomic_read_relaxed(&pmd->intrvl_cycles, &total_cycles); | |
1003 | /* Estimate the cycles to cover all intervals. */ | |
1004 | total_cycles *= PMD_RXQ_INTERVAL_MAX; | |
1005 | ||
1006 | for (int i = 0; i < n_rxq; i++) { | |
1007 | struct dp_netdev_rxq *rxq = list[i].rxq; | |
1008 | const char *name = netdev_rxq_get_name(rxq->rx); | |
1009 | uint64_t proc_cycles = 0; | |
1010 | ||
1011 | for (int j = 0; j < PMD_RXQ_INTERVAL_MAX; j++) { | |
1012 | proc_cycles += dp_netdev_rxq_get_intrvl_cycles(rxq, j); | |
ce179f11 | 1013 | } |
2a2c67b4 | 1014 | ds_put_format(reply, "\tport: %-16s\tqueue-id: %2d", name, |
947dc567 | 1015 | netdev_rxq_get_queue_id(list[i].rxq->rx)); |
2a2c67b4 KT |
1016 | ds_put_format(reply, "\tpmd usage: "); |
1017 | if (total_cycles) { | |
1018 | ds_put_format(reply, "%2"PRIu64"", | |
1019 | proc_cycles * 100 / total_cycles); | |
1020 | ds_put_cstr(reply, " %"); | |
1021 | } else { | |
1022 | ds_put_format(reply, "%s", "NOT AVAIL"); | |
1023 | } | |
1024 | ds_put_cstr(reply, "\n"); | |
ce179f11 | 1025 | } |
d0cca6c3 | 1026 | ovs_mutex_unlock(&pmd->port_mutex); |
947dc567 | 1027 | free(list); |
ce179f11 IM |
1028 | } |
1029 | } | |
1030 | ||
34d8e04b EC |
1031 | static int |
1032 | compare_poll_thread_list(const void *a_, const void *b_) | |
1033 | { | |
1034 | const struct dp_netdev_pmd_thread *a, *b; | |
1035 | ||
1036 | a = *(struct dp_netdev_pmd_thread **)a_; | |
1037 | b = *(struct dp_netdev_pmd_thread **)b_; | |
1038 | ||
1039 | if (a->core_id < b->core_id) { | |
1040 | return -1; | |
1041 | } | |
1042 | if (a->core_id > b->core_id) { | |
1043 | return 1; | |
1044 | } | |
1045 | return 0; | |
1046 | } | |
1047 | ||
1048 | /* Create a sorted list of pmd's from the dp->poll_threads cmap. We can use | |
1049 | * this list, as long as we do not go to quiescent state. */ | |
1050 | static void | |
1051 | sorted_poll_thread_list(struct dp_netdev *dp, | |
1052 | struct dp_netdev_pmd_thread ***list, | |
1053 | size_t *n) | |
1054 | { | |
1055 | struct dp_netdev_pmd_thread *pmd; | |
1056 | struct dp_netdev_pmd_thread **pmd_list; | |
1057 | size_t k = 0, n_pmds; | |
1058 | ||
1059 | n_pmds = cmap_count(&dp->poll_threads); | |
1060 | pmd_list = xcalloc(n_pmds, sizeof *pmd_list); | |
1061 | ||
1062 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
1063 | if (k >= n_pmds) { | |
1064 | break; | |
1065 | } | |
1066 | pmd_list[k++] = pmd; | |
1067 | } | |
1068 | ||
1069 | qsort(pmd_list, k, sizeof *pmd_list, compare_poll_thread_list); | |
1070 | ||
1071 | *list = pmd_list; | |
1072 | *n = k; | |
1073 | } | |
1074 | ||
cd995c73 KT |
1075 | static void |
1076 | dpif_netdev_pmd_rebalance(struct unixctl_conn *conn, int argc, | |
1077 | const char *argv[], void *aux OVS_UNUSED) | |
1078 | { | |
1079 | struct ds reply = DS_EMPTY_INITIALIZER; | |
1080 | struct dp_netdev *dp = NULL; | |
1081 | ||
1082 | ovs_mutex_lock(&dp_netdev_mutex); | |
1083 | ||
1084 | if (argc == 2) { | |
1085 | dp = shash_find_data(&dp_netdevs, argv[1]); | |
1086 | } else if (shash_count(&dp_netdevs) == 1) { | |
1087 | /* There's only one datapath */ | |
1088 | dp = shash_first(&dp_netdevs)->data; | |
1089 | } | |
1090 | ||
1091 | if (!dp) { | |
1092 | ovs_mutex_unlock(&dp_netdev_mutex); | |
1093 | unixctl_command_reply_error(conn, | |
1094 | "please specify an existing datapath"); | |
1095 | return; | |
1096 | } | |
1097 | ||
1098 | dp_netdev_request_reconfigure(dp); | |
1099 | ovs_mutex_unlock(&dp_netdev_mutex); | |
1100 | ds_put_cstr(&reply, "pmd rxq rebalance requested.\n"); | |
1101 | unixctl_command_reply(conn, ds_cstr(&reply)); | |
1102 | ds_destroy(&reply); | |
1103 | } | |
1104 | ||
6553d06b DDP |
1105 | static void |
1106 | dpif_netdev_pmd_info(struct unixctl_conn *conn, int argc, const char *argv[], | |
1107 | void *aux) | |
1108 | { | |
1109 | struct ds reply = DS_EMPTY_INITIALIZER; | |
34d8e04b | 1110 | struct dp_netdev_pmd_thread **pmd_list; |
6553d06b DDP |
1111 | struct dp_netdev *dp = NULL; |
1112 | enum pmd_info_type type = *(enum pmd_info_type *) aux; | |
82a48ead JS |
1113 | unsigned int core_id; |
1114 | bool filter_on_pmd = false; | |
1115 | size_t n; | |
6553d06b DDP |
1116 | |
1117 | ovs_mutex_lock(&dp_netdev_mutex); | |
1118 | ||
82a48ead | 1119 | while (argc > 1) { |
79f36875 | 1120 | if (!strcmp(argv[1], "-pmd") && argc > 2) { |
82a48ead JS |
1121 | if (str_to_uint(argv[2], 10, &core_id)) { |
1122 | filter_on_pmd = true; | |
1123 | } | |
1124 | argc -= 2; | |
1125 | argv += 2; | |
1126 | } else { | |
1127 | dp = shash_find_data(&dp_netdevs, argv[1]); | |
1128 | argc -= 1; | |
1129 | argv += 1; | |
1130 | } | |
6553d06b DDP |
1131 | } |
1132 | ||
1133 | if (!dp) { | |
82a48ead JS |
1134 | if (shash_count(&dp_netdevs) == 1) { |
1135 | /* There's only one datapath */ | |
1136 | dp = shash_first(&dp_netdevs)->data; | |
1137 | } else { | |
1138 | ovs_mutex_unlock(&dp_netdev_mutex); | |
1139 | unixctl_command_reply_error(conn, | |
1140 | "please specify an existing datapath"); | |
1141 | return; | |
1142 | } | |
6553d06b DDP |
1143 | } |
1144 | ||
34d8e04b EC |
1145 | sorted_poll_thread_list(dp, &pmd_list, &n); |
1146 | for (size_t i = 0; i < n; i++) { | |
1147 | struct dp_netdev_pmd_thread *pmd = pmd_list[i]; | |
1148 | if (!pmd) { | |
1149 | break; | |
1150 | } | |
82a48ead JS |
1151 | if (filter_on_pmd && pmd->core_id != core_id) { |
1152 | continue; | |
1153 | } | |
ce179f11 IM |
1154 | if (type == PMD_INFO_SHOW_RXQ) { |
1155 | pmd_info_show_rxq(&reply, pmd); | |
82a48ead JS |
1156 | } else if (type == PMD_INFO_CLEAR_STATS) { |
1157 | pmd_perf_stats_clear(&pmd->perf_stats); | |
1158 | } else if (type == PMD_INFO_SHOW_STATS) { | |
1159 | pmd_info_show_stats(&reply, pmd); | |
79f36875 JS |
1160 | } else if (type == PMD_INFO_PERF_SHOW) { |
1161 | pmd_info_show_perf(&reply, pmd, (struct pmd_perf_params *)aux); | |
6553d06b DDP |
1162 | } |
1163 | } | |
34d8e04b | 1164 | free(pmd_list); |
6553d06b DDP |
1165 | |
1166 | ovs_mutex_unlock(&dp_netdev_mutex); | |
1167 | ||
1168 | unixctl_command_reply(conn, ds_cstr(&reply)); | |
1169 | ds_destroy(&reply); | |
1170 | } | |
79f36875 JS |
1171 | |
1172 | static void | |
1173 | pmd_perf_show_cmd(struct unixctl_conn *conn, int argc, | |
1174 | const char *argv[], | |
1175 | void *aux OVS_UNUSED) | |
1176 | { | |
1177 | struct pmd_perf_params par; | |
1178 | long int it_hist = 0, ms_hist = 0; | |
1179 | par.histograms = true; | |
1180 | ||
1181 | while (argc > 1) { | |
1182 | if (!strcmp(argv[1], "-nh")) { | |
1183 | par.histograms = false; | |
1184 | argc -= 1; | |
1185 | argv += 1; | |
1186 | } else if (!strcmp(argv[1], "-it") && argc > 2) { | |
1187 | it_hist = strtol(argv[2], NULL, 10); | |
1188 | if (it_hist < 0) { | |
1189 | it_hist = 0; | |
1190 | } else if (it_hist > HISTORY_LEN) { | |
1191 | it_hist = HISTORY_LEN; | |
1192 | } | |
1193 | argc -= 2; | |
1194 | argv += 2; | |
1195 | } else if (!strcmp(argv[1], "-ms") && argc > 2) { | |
1196 | ms_hist = strtol(argv[2], NULL, 10); | |
1197 | if (ms_hist < 0) { | |
1198 | ms_hist = 0; | |
1199 | } else if (ms_hist > HISTORY_LEN) { | |
1200 | ms_hist = HISTORY_LEN; | |
1201 | } | |
1202 | argc -= 2; | |
1203 | argv += 2; | |
1204 | } else { | |
1205 | break; | |
1206 | } | |
1207 | } | |
1208 | par.iter_hist_len = it_hist; | |
1209 | par.ms_hist_len = ms_hist; | |
1210 | par.command_type = PMD_INFO_PERF_SHOW; | |
1211 | dpif_netdev_pmd_info(conn, argc, argv, &par); | |
1212 | } | |
6553d06b DDP |
1213 | \f |
1214 | static int | |
1215 | dpif_netdev_init(void) | |
1216 | { | |
1217 | static enum pmd_info_type show_aux = PMD_INFO_SHOW_STATS, | |
ce179f11 IM |
1218 | clear_aux = PMD_INFO_CLEAR_STATS, |
1219 | poll_aux = PMD_INFO_SHOW_RXQ; | |
6553d06b | 1220 | |
82a48ead JS |
1221 | unixctl_command_register("dpif-netdev/pmd-stats-show", "[-pmd core] [dp]", |
1222 | 0, 3, dpif_netdev_pmd_info, | |
6553d06b | 1223 | (void *)&show_aux); |
82a48ead JS |
1224 | unixctl_command_register("dpif-netdev/pmd-stats-clear", "[-pmd core] [dp]", |
1225 | 0, 3, dpif_netdev_pmd_info, | |
6553d06b | 1226 | (void *)&clear_aux); |
82a48ead JS |
1227 | unixctl_command_register("dpif-netdev/pmd-rxq-show", "[-pmd core] [dp]", |
1228 | 0, 3, dpif_netdev_pmd_info, | |
ce179f11 | 1229 | (void *)&poll_aux); |
79f36875 JS |
1230 | unixctl_command_register("dpif-netdev/pmd-perf-show", |
1231 | "[-nh] [-it iter-history-len]" | |
1232 | " [-ms ms-history-len]" | |
1233 | " [-pmd core] [dp]", | |
1234 | 0, 8, pmd_perf_show_cmd, | |
1235 | NULL); | |
cd995c73 KT |
1236 | unixctl_command_register("dpif-netdev/pmd-rxq-rebalance", "[dp]", |
1237 | 0, 1, dpif_netdev_pmd_rebalance, | |
1238 | NULL); | |
7178fefb JS |
1239 | unixctl_command_register("dpif-netdev/pmd-perf-log-set", |
1240 | "on|off [-b before] [-a after] [-e|-ne] " | |
1241 | "[-us usec] [-q qlen]", | |
1242 | 0, 10, pmd_perf_log_set_cmd, | |
1243 | NULL); | |
6553d06b DDP |
1244 | return 0; |
1245 | } | |
72865317 | 1246 | |
2197d7ab | 1247 | static int |
2240af25 DDP |
1248 | dpif_netdev_enumerate(struct sset *all_dps, |
1249 | const struct dpif_class *dpif_class) | |
2197d7ab GL |
1250 | { |
1251 | struct shash_node *node; | |
1252 | ||
97be1538 | 1253 | ovs_mutex_lock(&dp_netdev_mutex); |
2197d7ab | 1254 | SHASH_FOR_EACH(node, &dp_netdevs) { |
2240af25 DDP |
1255 | struct dp_netdev *dp = node->data; |
1256 | if (dpif_class != dp->class) { | |
1257 | /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs. | |
1258 | * If the class doesn't match, skip this dpif. */ | |
1259 | continue; | |
1260 | } | |
2197d7ab GL |
1261 | sset_add(all_dps, node->name); |
1262 | } | |
97be1538 | 1263 | ovs_mutex_unlock(&dp_netdev_mutex); |
5279f8fd | 1264 | |
2197d7ab GL |
1265 | return 0; |
1266 | } | |
1267 | ||
add90f6f EJ |
1268 | static bool |
1269 | dpif_netdev_class_is_dummy(const struct dpif_class *class) | |
1270 | { | |
1271 | return class != &dpif_netdev_class; | |
1272 | } | |
1273 | ||
0aeaabc8 JP |
1274 | static const char * |
1275 | dpif_netdev_port_open_type(const struct dpif_class *class, const char *type) | |
1276 | { | |
1277 | return strcmp(type, "internal") ? type | |
e98d0cb3 | 1278 | : dpif_netdev_class_is_dummy(class) ? "dummy-internal" |
0aeaabc8 JP |
1279 | : "tap"; |
1280 | } | |
1281 | ||
72865317 BP |
1282 | static struct dpif * |
1283 | create_dpif_netdev(struct dp_netdev *dp) | |
1284 | { | |
462278db | 1285 | uint16_t netflow_id = hash_string(dp->name, 0); |
72865317 | 1286 | struct dpif_netdev *dpif; |
72865317 | 1287 | |
6a8267c5 | 1288 | ovs_refcount_ref(&dp->ref_cnt); |
72865317 | 1289 | |
72865317 | 1290 | dpif = xmalloc(sizeof *dpif); |
614c4892 | 1291 | dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id); |
72865317 | 1292 | dpif->dp = dp; |
d33ed218 | 1293 | dpif->last_port_seq = seq_read(dp->port_seq); |
72865317 BP |
1294 | |
1295 | return &dpif->dpif; | |
1296 | } | |
1297 | ||
4e022ec0 AW |
1298 | /* Choose an unused, non-zero port number and return it on success. |
1299 | * Return ODPP_NONE on failure. */ | |
1300 | static odp_port_t | |
e44768b7 | 1301 | choose_port(struct dp_netdev *dp, const char *name) |
59e6d833 | 1302 | OVS_REQUIRES(dp->port_mutex) |
e44768b7 | 1303 | { |
4e022ec0 | 1304 | uint32_t port_no; |
e44768b7 JP |
1305 | |
1306 | if (dp->class != &dpif_netdev_class) { | |
1307 | const char *p; | |
1308 | int start_no = 0; | |
1309 | ||
1310 | /* If the port name begins with "br", start the number search at | |
1311 | * 100 to make writing tests easier. */ | |
1312 | if (!strncmp(name, "br", 2)) { | |
1313 | start_no = 100; | |
1314 | } | |
1315 | ||
1316 | /* If the port name contains a number, try to assign that port number. | |
1317 | * This can make writing unit tests easier because port numbers are | |
1318 | * predictable. */ | |
1319 | for (p = name; *p != '\0'; p++) { | |
1320 | if (isdigit((unsigned char) *p)) { | |
1321 | port_no = start_no + strtol(p, NULL, 10); | |
ff073a71 BP |
1322 | if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE) |
1323 | && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) { | |
4e022ec0 | 1324 | return u32_to_odp(port_no); |
e44768b7 JP |
1325 | } |
1326 | break; | |
1327 | } | |
1328 | } | |
1329 | } | |
1330 | ||
ff073a71 BP |
1331 | for (port_no = 1; port_no <= UINT16_MAX; port_no++) { |
1332 | if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) { | |
4e022ec0 | 1333 | return u32_to_odp(port_no); |
e44768b7 JP |
1334 | } |
1335 | } | |
1336 | ||
4e022ec0 | 1337 | return ODPP_NONE; |
e44768b7 JP |
1338 | } |
1339 | ||
72865317 | 1340 | static int |
614c4892 BP |
1341 | create_dp_netdev(const char *name, const struct dpif_class *class, |
1342 | struct dp_netdev **dpp) | |
8a4e3a85 | 1343 | OVS_REQUIRES(dp_netdev_mutex) |
72865317 BP |
1344 | { |
1345 | struct dp_netdev *dp; | |
1346 | int error; | |
72865317 | 1347 | |
462278db | 1348 | dp = xzalloc(sizeof *dp); |
8a4e3a85 BP |
1349 | shash_add(&dp_netdevs, name, dp); |
1350 | ||
1351 | *CONST_CAST(const struct dpif_class **, &dp->class) = class; | |
1352 | *CONST_CAST(const char **, &dp->name) = xstrdup(name); | |
6a8267c5 | 1353 | ovs_refcount_init(&dp->ref_cnt); |
1a65ba85 | 1354 | atomic_flag_clear(&dp->destroyed); |
8a4e3a85 | 1355 | |
59e6d833 | 1356 | ovs_mutex_init(&dp->port_mutex); |
e9985d6a | 1357 | hmap_init(&dp->ports); |
d33ed218 | 1358 | dp->port_seq = seq_create(); |
6b31e073 RW |
1359 | fat_rwlock_init(&dp->upcall_rwlock); |
1360 | ||
a6a426d6 IM |
1361 | dp->reconfigure_seq = seq_create(); |
1362 | dp->last_reconfigure_seq = seq_read(dp->reconfigure_seq); | |
1363 | ||
4b27db64 JR |
1364 | for (int i = 0; i < N_METER_LOCKS; ++i) { |
1365 | ovs_mutex_init_adaptive(&dp->meter_locks[i]); | |
1366 | } | |
1367 | ||
6b31e073 RW |
1368 | /* Disable upcalls by default. */ |
1369 | dp_netdev_disable_upcall(dp); | |
623540e4 | 1370 | dp->upcall_aux = NULL; |
6b31e073 | 1371 | dp->upcall_cb = NULL; |
e44768b7 | 1372 | |
5cf3edb3 DDP |
1373 | conntrack_init(&dp->conntrack); |
1374 | ||
4c30b246 | 1375 | atomic_init(&dp->emc_insert_min, DEFAULT_EM_FLOW_INSERT_MIN); |
c71ea3c4 | 1376 | atomic_init(&dp->tx_flush_interval, DEFAULT_TX_FLUSH_INTERVAL); |
4c30b246 | 1377 | |
65f13b50 | 1378 | cmap_init(&dp->poll_threads); |
140dd699 IM |
1379 | |
1380 | ovs_mutex_init(&dp->tx_qid_pool_mutex); | |
1381 | /* We need 1 Tx queue for each possible core + 1 for non-PMD threads. */ | |
1382 | dp->tx_qid_pool = id_pool_create(0, ovs_numa_get_n_cores() + 1); | |
1383 | ||
65f13b50 AW |
1384 | ovs_mutex_init_recursive(&dp->non_pmd_mutex); |
1385 | ovsthread_key_create(&dp->per_pmd_key, NULL); | |
1386 | ||
e9985d6a | 1387 | ovs_mutex_lock(&dp->port_mutex); |
140dd699 IM |
1388 | /* non-PMD will be created before all other threads and will |
1389 | * allocate static_tx_qid = 0. */ | |
f2eee189 | 1390 | dp_netdev_set_nonpmd(dp); |
65f13b50 | 1391 | |
a3e8437a TLSC |
1392 | error = do_add_port(dp, name, dpif_netdev_port_open_type(dp->class, |
1393 | "internal"), | |
1394 | ODPP_LOCAL); | |
59e6d833 | 1395 | ovs_mutex_unlock(&dp->port_mutex); |
72865317 BP |
1396 | if (error) { |
1397 | dp_netdev_free(dp); | |
462278db | 1398 | return error; |
72865317 BP |
1399 | } |
1400 | ||
a36de779 | 1401 | dp->last_tnl_conf_seq = seq_read(tnl_conf_seq); |
462278db | 1402 | *dpp = dp; |
72865317 BP |
1403 | return 0; |
1404 | } | |
1405 | ||
a6a426d6 IM |
1406 | static void |
1407 | dp_netdev_request_reconfigure(struct dp_netdev *dp) | |
1408 | { | |
1409 | seq_change(dp->reconfigure_seq); | |
1410 | } | |
1411 | ||
1412 | static bool | |
1413 | dp_netdev_is_reconf_required(struct dp_netdev *dp) | |
1414 | { | |
1415 | return seq_read(dp->reconfigure_seq) != dp->last_reconfigure_seq; | |
1416 | } | |
1417 | ||
72865317 | 1418 | static int |
614c4892 | 1419 | dpif_netdev_open(const struct dpif_class *class, const char *name, |
4a387741 | 1420 | bool create, struct dpif **dpifp) |
72865317 | 1421 | { |
462278db | 1422 | struct dp_netdev *dp; |
5279f8fd | 1423 | int error; |
462278db | 1424 | |
97be1538 | 1425 | ovs_mutex_lock(&dp_netdev_mutex); |
462278db BP |
1426 | dp = shash_find_data(&dp_netdevs, name); |
1427 | if (!dp) { | |
5279f8fd | 1428 | error = create ? create_dp_netdev(name, class, &dp) : ENODEV; |
72865317 | 1429 | } else { |
5279f8fd BP |
1430 | error = (dp->class != class ? EINVAL |
1431 | : create ? EEXIST | |
1432 | : 0); | |
1433 | } | |
1434 | if (!error) { | |
1435 | *dpifp = create_dpif_netdev(dp); | |
6b31e073 | 1436 | dp->dpif = *dpifp; |
72865317 | 1437 | } |
97be1538 | 1438 | ovs_mutex_unlock(&dp_netdev_mutex); |
462278db | 1439 | |
5279f8fd | 1440 | return error; |
72865317 BP |
1441 | } |
1442 | ||
88ace79b DDP |
1443 | static void |
1444 | dp_netdev_destroy_upcall_lock(struct dp_netdev *dp) | |
1445 | OVS_NO_THREAD_SAFETY_ANALYSIS | |
1446 | { | |
1447 | /* Check that upcalls are disabled, i.e. that the rwlock is taken */ | |
1448 | ovs_assert(fat_rwlock_tryrdlock(&dp->upcall_rwlock)); | |
1449 | ||
1450 | /* Before freeing a lock we should release it */ | |
1451 | fat_rwlock_unlock(&dp->upcall_rwlock); | |
1452 | fat_rwlock_destroy(&dp->upcall_rwlock); | |
1453 | } | |
1454 | ||
4b27db64 JR |
1455 | static void |
1456 | dp_delete_meter(struct dp_netdev *dp, uint32_t meter_id) | |
1457 | OVS_REQUIRES(dp->meter_locks[meter_id % N_METER_LOCKS]) | |
1458 | { | |
1459 | if (dp->meters[meter_id]) { | |
1460 | free(dp->meters[meter_id]); | |
1461 | dp->meters[meter_id] = NULL; | |
1462 | } | |
1463 | } | |
1464 | ||
8a4e3a85 BP |
1465 | /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp' |
1466 | * through the 'dp_netdevs' shash while freeing 'dp'. */ | |
1ba530f4 BP |
1467 | static void |
1468 | dp_netdev_free(struct dp_netdev *dp) | |
8a4e3a85 | 1469 | OVS_REQUIRES(dp_netdev_mutex) |
1ba530f4 | 1470 | { |
e9985d6a | 1471 | struct dp_netdev_port *port, *next; |
4ad28026 | 1472 | |
8a4e3a85 BP |
1473 | shash_find_and_delete(&dp_netdevs, dp->name); |
1474 | ||
59e6d833 | 1475 | ovs_mutex_lock(&dp->port_mutex); |
e9985d6a | 1476 | HMAP_FOR_EACH_SAFE (port, next, node, &dp->ports) { |
c40b890f | 1477 | do_del_port(dp, port); |
1ba530f4 | 1478 | } |
59e6d833 | 1479 | ovs_mutex_unlock(&dp->port_mutex); |
4b27db64 | 1480 | |
e32971b8 | 1481 | dp_netdev_destroy_all_pmds(dp, true); |
d916785c | 1482 | cmap_destroy(&dp->poll_threads); |
51852a57 | 1483 | |
140dd699 IM |
1484 | ovs_mutex_destroy(&dp->tx_qid_pool_mutex); |
1485 | id_pool_destroy(dp->tx_qid_pool); | |
1486 | ||
b9584f21 DDP |
1487 | ovs_mutex_destroy(&dp->non_pmd_mutex); |
1488 | ovsthread_key_delete(dp->per_pmd_key); | |
1489 | ||
1490 | conntrack_destroy(&dp->conntrack); | |
1491 | ||
1492 | ||
a6a426d6 IM |
1493 | seq_destroy(dp->reconfigure_seq); |
1494 | ||
d33ed218 | 1495 | seq_destroy(dp->port_seq); |
e9985d6a | 1496 | hmap_destroy(&dp->ports); |
3186ea46 | 1497 | ovs_mutex_destroy(&dp->port_mutex); |
88ace79b DDP |
1498 | |
1499 | /* Upcalls must be disabled at this point */ | |
1500 | dp_netdev_destroy_upcall_lock(dp); | |
9bbf1c3d | 1501 | |
4b27db64 JR |
1502 | int i; |
1503 | ||
1504 | for (i = 0; i < MAX_METERS; ++i) { | |
1505 | meter_lock(dp, i); | |
1506 | dp_delete_meter(dp, i); | |
1507 | meter_unlock(dp, i); | |
1508 | } | |
1509 | for (i = 0; i < N_METER_LOCKS; ++i) { | |
1510 | ovs_mutex_destroy(&dp->meter_locks[i]); | |
1511 | } | |
1512 | ||
f2eee189 | 1513 | free(dp->pmd_cmask); |
8a4e3a85 | 1514 | free(CONST_CAST(char *, dp->name)); |
72865317 BP |
1515 | free(dp); |
1516 | } | |
1517 | ||
8a4e3a85 BP |
1518 | static void |
1519 | dp_netdev_unref(struct dp_netdev *dp) | |
1520 | { | |
1521 | if (dp) { | |
1522 | /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't | |
1523 | * get a new reference to 'dp' through the 'dp_netdevs' shash. */ | |
1524 | ovs_mutex_lock(&dp_netdev_mutex); | |
24f83812 | 1525 | if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) { |
8a4e3a85 BP |
1526 | dp_netdev_free(dp); |
1527 | } | |
1528 | ovs_mutex_unlock(&dp_netdev_mutex); | |
1529 | } | |
1530 | } | |
1531 | ||
72865317 BP |
1532 | static void |
1533 | dpif_netdev_close(struct dpif *dpif) | |
1534 | { | |
1535 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
5279f8fd | 1536 | |
8a4e3a85 | 1537 | dp_netdev_unref(dp); |
72865317 BP |
1538 | free(dpif); |
1539 | } | |
1540 | ||
1541 | static int | |
7dab847a | 1542 | dpif_netdev_destroy(struct dpif *dpif) |
72865317 BP |
1543 | { |
1544 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
5279f8fd | 1545 | |
6a8267c5 | 1546 | if (!atomic_flag_test_and_set(&dp->destroyed)) { |
24f83812 | 1547 | if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) { |
6a8267c5 BP |
1548 | /* Can't happen: 'dpif' still owns a reference to 'dp'. */ |
1549 | OVS_NOT_REACHED(); | |
1550 | } | |
1551 | } | |
5279f8fd | 1552 | |
72865317 BP |
1553 | return 0; |
1554 | } | |
1555 | ||
eb94da30 DDP |
1556 | /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed |
1557 | * load/store semantics. While the increment is not atomic, the load and | |
1558 | * store operations are, making it impossible to read inconsistent values. | |
1559 | * | |
1560 | * This is used to update thread local stats counters. */ | |
1561 | static void | |
1562 | non_atomic_ullong_add(atomic_ullong *var, unsigned long long n) | |
1563 | { | |
1564 | unsigned long long tmp; | |
1565 | ||
1566 | atomic_read_relaxed(var, &tmp); | |
1567 | tmp += n; | |
1568 | atomic_store_relaxed(var, tmp); | |
1569 | } | |
1570 | ||
72865317 | 1571 | static int |
a8d9304d | 1572 | dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats) |
72865317 BP |
1573 | { |
1574 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
1c1e46ed | 1575 | struct dp_netdev_pmd_thread *pmd; |
82a48ead | 1576 | uint64_t pmd_stats[PMD_N_STATS]; |
8a4e3a85 | 1577 | |
1c1e46ed AW |
1578 | stats->n_flows = stats->n_hit = stats->n_missed = stats->n_lost = 0; |
1579 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
1580 | stats->n_flows += cmap_count(&pmd->flow_table); | |
82a48ead JS |
1581 | pmd_perf_read_counters(&pmd->perf_stats, pmd_stats); |
1582 | stats->n_hit += pmd_stats[PMD_STAT_EXACT_HIT]; | |
1583 | stats->n_hit += pmd_stats[PMD_STAT_MASKED_HIT]; | |
1584 | stats->n_missed += pmd_stats[PMD_STAT_MISS]; | |
1585 | stats->n_lost += pmd_stats[PMD_STAT_LOST]; | |
51852a57 | 1586 | } |
1ce3fa06 | 1587 | stats->n_masks = UINT32_MAX; |
847108dc | 1588 | stats->n_mask_hit = UINT64_MAX; |
5279f8fd | 1589 | |
72865317 BP |
1590 | return 0; |
1591 | } | |
1592 | ||
e4cfed38 | 1593 | static void |
65f13b50 | 1594 | dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread *pmd) |
e4cfed38 | 1595 | { |
accf8626 | 1596 | if (pmd->core_id == NON_PMD_CORE_ID) { |
d0cca6c3 DDP |
1597 | ovs_mutex_lock(&pmd->dp->non_pmd_mutex); |
1598 | ovs_mutex_lock(&pmd->port_mutex); | |
1599 | pmd_load_cached_ports(pmd); | |
1600 | ovs_mutex_unlock(&pmd->port_mutex); | |
1601 | ovs_mutex_unlock(&pmd->dp->non_pmd_mutex); | |
accf8626 AW |
1602 | return; |
1603 | } | |
1604 | ||
1605 | ovs_mutex_lock(&pmd->cond_mutex); | |
2788a1b1 | 1606 | seq_change(pmd->reload_seq); |
14e3e12a | 1607 | atomic_store_relaxed(&pmd->reload, true); |
accf8626 AW |
1608 | ovs_mutex_cond_wait(&pmd->cond, &pmd->cond_mutex); |
1609 | ovs_mutex_unlock(&pmd->cond_mutex); | |
65f13b50 | 1610 | } |
e4cfed38 | 1611 | |
59e6d833 BP |
1612 | static uint32_t |
1613 | hash_port_no(odp_port_t port_no) | |
1614 | { | |
1615 | return hash_int(odp_to_u32(port_no), 0); | |
1616 | } | |
1617 | ||
72865317 | 1618 | static int |
a3e8437a | 1619 | port_create(const char *devname, const char *type, |
b8d29252 | 1620 | odp_port_t port_no, struct dp_netdev_port **portp) |
72865317 | 1621 | { |
4b609110 | 1622 | struct netdev_saved_flags *sf; |
72865317 | 1623 | struct dp_netdev_port *port; |
2499a8ce | 1624 | enum netdev_flags flags; |
b8d29252 | 1625 | struct netdev *netdev; |
e32971b8 | 1626 | int error; |
72865317 | 1627 | |
b8d29252 | 1628 | *portp = NULL; |
72865317 BP |
1629 | |
1630 | /* Open and validate network device. */ | |
a3e8437a | 1631 | error = netdev_open(devname, type, &netdev); |
72865317 | 1632 | if (error) { |
b8d29252 | 1633 | return error; |
72865317 | 1634 | } |
72865317 BP |
1635 | /* XXX reject non-Ethernet devices */ |
1636 | ||
2499a8ce AC |
1637 | netdev_get_flags(netdev, &flags); |
1638 | if (flags & NETDEV_LOOPBACK) { | |
1639 | VLOG_ERR("%s: cannot add a loopback device", devname); | |
d17f4f08 | 1640 | error = EINVAL; |
b8d29252 | 1641 | goto out; |
2499a8ce AC |
1642 | } |
1643 | ||
e32971b8 DDP |
1644 | error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf); |
1645 | if (error) { | |
1646 | VLOG_ERR("%s: cannot set promisc flag", devname); | |
1647 | goto out; | |
324c8374 IM |
1648 | } |
1649 | ||
e4cfed38 | 1650 | port = xzalloc(sizeof *port); |
35303d71 | 1651 | port->port_no = port_no; |
e4cfed38 PS |
1652 | port->netdev = netdev; |
1653 | port->type = xstrdup(type); | |
4b609110 | 1654 | port->sf = sf; |
e32971b8 DDP |
1655 | port->need_reconfigure = true; |
1656 | ovs_mutex_init(&port->txq_used_mutex); | |
e4cfed38 | 1657 | |
b8d29252 | 1658 | *portp = port; |
72865317 BP |
1659 | |
1660 | return 0; | |
d17f4f08 | 1661 | |
d17f4f08 | 1662 | out: |
b8d29252 | 1663 | netdev_close(netdev); |
d17f4f08 | 1664 | return error; |
72865317 BP |
1665 | } |
1666 | ||
b8d29252 DDP |
1667 | static int |
1668 | do_add_port(struct dp_netdev *dp, const char *devname, const char *type, | |
1669 | odp_port_t port_no) | |
1670 | OVS_REQUIRES(dp->port_mutex) | |
1671 | { | |
1672 | struct dp_netdev_port *port; | |
1673 | int error; | |
1674 | ||
1675 | /* Reject devices already in 'dp'. */ | |
1676 | if (!get_port_by_name(dp, devname, &port)) { | |
1677 | return EEXIST; | |
1678 | } | |
1679 | ||
a3e8437a | 1680 | error = port_create(devname, type, port_no, &port); |
b8d29252 DDP |
1681 | if (error) { |
1682 | return error; | |
1683 | } | |
1684 | ||
e9985d6a | 1685 | hmap_insert(&dp->ports, &port->node, hash_port_no(port_no)); |
b8d29252 DDP |
1686 | seq_change(dp->port_seq); |
1687 | ||
e32971b8 DDP |
1688 | reconfigure_datapath(dp); |
1689 | ||
b8d29252 DDP |
1690 | return 0; |
1691 | } | |
1692 | ||
247527db BP |
1693 | static int |
1694 | dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev, | |
4e022ec0 | 1695 | odp_port_t *port_nop) |
247527db BP |
1696 | { |
1697 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
3aa30359 BP |
1698 | char namebuf[NETDEV_VPORT_NAME_BUFSIZE]; |
1699 | const char *dpif_port; | |
4e022ec0 | 1700 | odp_port_t port_no; |
5279f8fd | 1701 | int error; |
247527db | 1702 | |
59e6d833 | 1703 | ovs_mutex_lock(&dp->port_mutex); |
3aa30359 | 1704 | dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf); |
4e022ec0 | 1705 | if (*port_nop != ODPP_NONE) { |
ff073a71 BP |
1706 | port_no = *port_nop; |
1707 | error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0; | |
232dfa4a | 1708 | } else { |
3aa30359 | 1709 | port_no = choose_port(dp, dpif_port); |
5279f8fd | 1710 | error = port_no == ODPP_NONE ? EFBIG : 0; |
232dfa4a | 1711 | } |
5279f8fd | 1712 | if (!error) { |
247527db | 1713 | *port_nop = port_no; |
5279f8fd | 1714 | error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no); |
247527db | 1715 | } |
59e6d833 | 1716 | ovs_mutex_unlock(&dp->port_mutex); |
5279f8fd BP |
1717 | |
1718 | return error; | |
72865317 BP |
1719 | } |
1720 | ||
1721 | static int | |
4e022ec0 | 1722 | dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no) |
72865317 BP |
1723 | { |
1724 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
5279f8fd BP |
1725 | int error; |
1726 | ||
59e6d833 | 1727 | ovs_mutex_lock(&dp->port_mutex); |
c40b890f BP |
1728 | if (port_no == ODPP_LOCAL) { |
1729 | error = EINVAL; | |
1730 | } else { | |
1731 | struct dp_netdev_port *port; | |
1732 | ||
1733 | error = get_port_by_number(dp, port_no, &port); | |
1734 | if (!error) { | |
1735 | do_del_port(dp, port); | |
1736 | } | |
1737 | } | |
59e6d833 | 1738 | ovs_mutex_unlock(&dp->port_mutex); |
5279f8fd BP |
1739 | |
1740 | return error; | |
72865317 BP |
1741 | } |
1742 | ||
1743 | static bool | |
4e022ec0 | 1744 | is_valid_port_number(odp_port_t port_no) |
72865317 | 1745 | { |
ff073a71 BP |
1746 | return port_no != ODPP_NONE; |
1747 | } | |
1748 | ||
1749 | static struct dp_netdev_port * | |
1750 | dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no) | |
e9985d6a | 1751 | OVS_REQUIRES(dp->port_mutex) |
ff073a71 BP |
1752 | { |
1753 | struct dp_netdev_port *port; | |
1754 | ||
e9985d6a | 1755 | HMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) { |
35303d71 | 1756 | if (port->port_no == port_no) { |
ff073a71 BP |
1757 | return port; |
1758 | } | |
1759 | } | |
1760 | return NULL; | |
72865317 BP |
1761 | } |
1762 | ||
1763 | static int | |
1764 | get_port_by_number(struct dp_netdev *dp, | |
4e022ec0 | 1765 | odp_port_t port_no, struct dp_netdev_port **portp) |
e9985d6a | 1766 | OVS_REQUIRES(dp->port_mutex) |
72865317 BP |
1767 | { |
1768 | if (!is_valid_port_number(port_no)) { | |
1769 | *portp = NULL; | |
1770 | return EINVAL; | |
1771 | } else { | |
ff073a71 | 1772 | *portp = dp_netdev_lookup_port(dp, port_no); |
0f6a066f | 1773 | return *portp ? 0 : ENODEV; |
72865317 BP |
1774 | } |
1775 | } | |
1776 | ||
b284085e | 1777 | static void |
62453dad | 1778 | port_destroy(struct dp_netdev_port *port) |
b284085e | 1779 | { |
62453dad DDP |
1780 | if (!port) { |
1781 | return; | |
b284085e | 1782 | } |
b284085e | 1783 | |
62453dad DDP |
1784 | netdev_close(port->netdev); |
1785 | netdev_restore_flags(port->sf); | |
accf8626 | 1786 | |
62453dad | 1787 | for (unsigned i = 0; i < port->n_rxq; i++) { |
947dc567 | 1788 | netdev_rxq_close(port->rxqs[i].rx); |
b284085e | 1789 | } |
324c8374 | 1790 | ovs_mutex_destroy(&port->txq_used_mutex); |
3eb67853 | 1791 | free(port->rxq_affinity_list); |
324c8374 | 1792 | free(port->txq_used); |
3eb67853 | 1793 | free(port->rxqs); |
62453dad DDP |
1794 | free(port->type); |
1795 | free(port); | |
b284085e PS |
1796 | } |
1797 | ||
72865317 BP |
1798 | static int |
1799 | get_port_by_name(struct dp_netdev *dp, | |
1800 | const char *devname, struct dp_netdev_port **portp) | |
59e6d833 | 1801 | OVS_REQUIRES(dp->port_mutex) |
72865317 BP |
1802 | { |
1803 | struct dp_netdev_port *port; | |
1804 | ||
e9985d6a | 1805 | HMAP_FOR_EACH (port, node, &dp->ports) { |
3efb6063 | 1806 | if (!strcmp(netdev_get_name(port->netdev), devname)) { |
72865317 BP |
1807 | *portp = port; |
1808 | return 0; | |
1809 | } | |
1810 | } | |
0f6a066f DDP |
1811 | |
1812 | /* Callers of dpif_netdev_port_query_by_name() expect ENODEV for a non | |
1813 | * existing port. */ | |
1814 | return ENODEV; | |
72865317 BP |
1815 | } |
1816 | ||
b9584f21 | 1817 | /* Returns 'true' if there is a port with pmd netdev. */ |
65f13b50 | 1818 | static bool |
b9584f21 | 1819 | has_pmd_port(struct dp_netdev *dp) |
e9985d6a | 1820 | OVS_REQUIRES(dp->port_mutex) |
65f13b50 AW |
1821 | { |
1822 | struct dp_netdev_port *port; | |
1823 | ||
e9985d6a | 1824 | HMAP_FOR_EACH (port, node, &dp->ports) { |
5dd57e80 | 1825 | if (netdev_is_pmd(port->netdev)) { |
b9584f21 | 1826 | return true; |
65f13b50 AW |
1827 | } |
1828 | } | |
1829 | ||
1830 | return false; | |
1831 | } | |
1832 | ||
c40b890f BP |
1833 | static void |
1834 | do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port) | |
59e6d833 | 1835 | OVS_REQUIRES(dp->port_mutex) |
72865317 | 1836 | { |
e9985d6a | 1837 | hmap_remove(&dp->ports, &port->node); |
d33ed218 | 1838 | seq_change(dp->port_seq); |
d0cca6c3 | 1839 | |
e32971b8 | 1840 | reconfigure_datapath(dp); |
72865317 | 1841 | |
62453dad | 1842 | port_destroy(port); |
72865317 BP |
1843 | } |
1844 | ||
1845 | static void | |
4c738a8d BP |
1846 | answer_port_query(const struct dp_netdev_port *port, |
1847 | struct dpif_port *dpif_port) | |
72865317 | 1848 | { |
3efb6063 | 1849 | dpif_port->name = xstrdup(netdev_get_name(port->netdev)); |
0cbfe35d | 1850 | dpif_port->type = xstrdup(port->type); |
35303d71 | 1851 | dpif_port->port_no = port->port_no; |
72865317 BP |
1852 | } |
1853 | ||
1854 | static int | |
4e022ec0 | 1855 | dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no, |
4c738a8d | 1856 | struct dpif_port *dpif_port) |
72865317 BP |
1857 | { |
1858 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
1859 | struct dp_netdev_port *port; | |
1860 | int error; | |
1861 | ||
e9985d6a | 1862 | ovs_mutex_lock(&dp->port_mutex); |
72865317 | 1863 | error = get_port_by_number(dp, port_no, &port); |
4afba28d | 1864 | if (!error && dpif_port) { |
4c738a8d | 1865 | answer_port_query(port, dpif_port); |
72865317 | 1866 | } |
e9985d6a | 1867 | ovs_mutex_unlock(&dp->port_mutex); |
5279f8fd | 1868 | |
72865317 BP |
1869 | return error; |
1870 | } | |
1871 | ||
1872 | static int | |
1873 | dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname, | |
4c738a8d | 1874 | struct dpif_port *dpif_port) |
72865317 BP |
1875 | { |
1876 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
1877 | struct dp_netdev_port *port; | |
1878 | int error; | |
1879 | ||
59e6d833 | 1880 | ovs_mutex_lock(&dp->port_mutex); |
72865317 | 1881 | error = get_port_by_name(dp, devname, &port); |
4afba28d | 1882 | if (!error && dpif_port) { |
4c738a8d | 1883 | answer_port_query(port, dpif_port); |
72865317 | 1884 | } |
59e6d833 | 1885 | ovs_mutex_unlock(&dp->port_mutex); |
5279f8fd | 1886 | |
72865317 BP |
1887 | return error; |
1888 | } | |
1889 | ||
61e7deb1 BP |
1890 | static void |
1891 | dp_netdev_flow_free(struct dp_netdev_flow *flow) | |
1892 | { | |
61e7deb1 | 1893 | dp_netdev_actions_free(dp_netdev_flow_get_actions(flow)); |
61e7deb1 BP |
1894 | free(flow); |
1895 | } | |
1896 | ||
ed79f89a DDP |
1897 | static void dp_netdev_flow_unref(struct dp_netdev_flow *flow) |
1898 | { | |
1899 | if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) { | |
1900 | ovsrcu_postpone(dp_netdev_flow_free, flow); | |
1901 | } | |
1902 | } | |
1903 | ||
70e5ed6f JS |
1904 | static uint32_t |
1905 | dp_netdev_flow_hash(const ovs_u128 *ufid) | |
1906 | { | |
1907 | return ufid->u32[0]; | |
1908 | } | |
1909 | ||
3453b4d6 JS |
1910 | static inline struct dpcls * |
1911 | dp_netdev_pmd_lookup_dpcls(struct dp_netdev_pmd_thread *pmd, | |
1912 | odp_port_t in_port) | |
1913 | { | |
1914 | struct dpcls *cls; | |
1915 | uint32_t hash = hash_port_no(in_port); | |
1916 | CMAP_FOR_EACH_WITH_HASH (cls, node, hash, &pmd->classifiers) { | |
1917 | if (cls->in_port == in_port) { | |
1918 | /* Port classifier exists already */ | |
1919 | return cls; | |
1920 | } | |
1921 | } | |
1922 | return NULL; | |
1923 | } | |
1924 | ||
1925 | static inline struct dpcls * | |
1926 | dp_netdev_pmd_find_dpcls(struct dp_netdev_pmd_thread *pmd, | |
1927 | odp_port_t in_port) | |
1928 | OVS_REQUIRES(pmd->flow_mutex) | |
1929 | { | |
1930 | struct dpcls *cls = dp_netdev_pmd_lookup_dpcls(pmd, in_port); | |
1931 | uint32_t hash = hash_port_no(in_port); | |
1932 | ||
1933 | if (!cls) { | |
1934 | /* Create new classifier for in_port */ | |
1935 | cls = xmalloc(sizeof(*cls)); | |
1936 | dpcls_init(cls); | |
1937 | cls->in_port = in_port; | |
1938 | cmap_insert(&pmd->classifiers, &cls->node, hash); | |
1939 | VLOG_DBG("Creating dpcls %p for in_port %d", cls, in_port); | |
1940 | } | |
1941 | return cls; | |
1942 | } | |
1943 | ||
72865317 | 1944 | static void |
1c1e46ed AW |
1945 | dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread *pmd, |
1946 | struct dp_netdev_flow *flow) | |
1947 | OVS_REQUIRES(pmd->flow_mutex) | |
72865317 | 1948 | { |
9f361d6b | 1949 | struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node); |
3453b4d6 JS |
1950 | struct dpcls *cls; |
1951 | odp_port_t in_port = flow->flow.in_port.odp_port; | |
2c0ea78f | 1952 | |
3453b4d6 JS |
1953 | cls = dp_netdev_pmd_lookup_dpcls(pmd, in_port); |
1954 | ovs_assert(cls != NULL); | |
1955 | dpcls_remove(cls, &flow->cr); | |
1c1e46ed | 1956 | cmap_remove(&pmd->flow_table, node, dp_netdev_flow_hash(&flow->ufid)); |
9bbf1c3d | 1957 | flow->dead = true; |
ed79f89a DDP |
1958 | |
1959 | dp_netdev_flow_unref(flow); | |
72865317 BP |
1960 | } |
1961 | ||
1962 | static void | |
1c1e46ed | 1963 | dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd) |
72865317 | 1964 | { |
78c8df12 | 1965 | struct dp_netdev_flow *netdev_flow; |
72865317 | 1966 | |
1c1e46ed AW |
1967 | ovs_mutex_lock(&pmd->flow_mutex); |
1968 | CMAP_FOR_EACH (netdev_flow, node, &pmd->flow_table) { | |
1969 | dp_netdev_pmd_remove_flow(pmd, netdev_flow); | |
72865317 | 1970 | } |
1c1e46ed | 1971 | ovs_mutex_unlock(&pmd->flow_mutex); |
72865317 BP |
1972 | } |
1973 | ||
1974 | static int | |
1975 | dpif_netdev_flow_flush(struct dpif *dpif) | |
1976 | { | |
1977 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
1c1e46ed AW |
1978 | struct dp_netdev_pmd_thread *pmd; |
1979 | ||
1980 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
1981 | dp_netdev_pmd_flow_flush(pmd); | |
1982 | } | |
5279f8fd | 1983 | |
72865317 BP |
1984 | return 0; |
1985 | } | |
1986 | ||
b0ec0f27 | 1987 | struct dp_netdev_port_state { |
e9985d6a | 1988 | struct hmap_position position; |
4c738a8d | 1989 | char *name; |
b0ec0f27 BP |
1990 | }; |
1991 | ||
1992 | static int | |
1993 | dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep) | |
1994 | { | |
1995 | *statep = xzalloc(sizeof(struct dp_netdev_port_state)); | |
1996 | return 0; | |
1997 | } | |
1998 | ||
72865317 | 1999 | static int |
b0ec0f27 | 2000 | dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_, |
4c738a8d | 2001 | struct dpif_port *dpif_port) |
72865317 | 2002 | { |
b0ec0f27 | 2003 | struct dp_netdev_port_state *state = state_; |
72865317 | 2004 | struct dp_netdev *dp = get_dp_netdev(dpif); |
e9985d6a | 2005 | struct hmap_node *node; |
ff073a71 | 2006 | int retval; |
72865317 | 2007 | |
e9985d6a DDP |
2008 | ovs_mutex_lock(&dp->port_mutex); |
2009 | node = hmap_at_position(&dp->ports, &state->position); | |
ff073a71 BP |
2010 | if (node) { |
2011 | struct dp_netdev_port *port; | |
5279f8fd | 2012 | |
ff073a71 BP |
2013 | port = CONTAINER_OF(node, struct dp_netdev_port, node); |
2014 | ||
2015 | free(state->name); | |
2016 | state->name = xstrdup(netdev_get_name(port->netdev)); | |
2017 | dpif_port->name = state->name; | |
2018 | dpif_port->type = port->type; | |
35303d71 | 2019 | dpif_port->port_no = port->port_no; |
ff073a71 BP |
2020 | |
2021 | retval = 0; | |
2022 | } else { | |
2023 | retval = EOF; | |
72865317 | 2024 | } |
e9985d6a | 2025 | ovs_mutex_unlock(&dp->port_mutex); |
5279f8fd | 2026 | |
ff073a71 | 2027 | return retval; |
b0ec0f27 BP |
2028 | } |
2029 | ||
2030 | static int | |
4c738a8d | 2031 | dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_) |
b0ec0f27 | 2032 | { |
4c738a8d BP |
2033 | struct dp_netdev_port_state *state = state_; |
2034 | free(state->name); | |
b0ec0f27 BP |
2035 | free(state); |
2036 | return 0; | |
72865317 BP |
2037 | } |
2038 | ||
2039 | static int | |
67a4917b | 2040 | dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED) |
72865317 BP |
2041 | { |
2042 | struct dpif_netdev *dpif = dpif_netdev_cast(dpif_); | |
d33ed218 | 2043 | uint64_t new_port_seq; |
5279f8fd BP |
2044 | int error; |
2045 | ||
d33ed218 BP |
2046 | new_port_seq = seq_read(dpif->dp->port_seq); |
2047 | if (dpif->last_port_seq != new_port_seq) { | |
2048 | dpif->last_port_seq = new_port_seq; | |
5279f8fd | 2049 | error = ENOBUFS; |
72865317 | 2050 | } else { |
5279f8fd | 2051 | error = EAGAIN; |
72865317 | 2052 | } |
5279f8fd BP |
2053 | |
2054 | return error; | |
72865317 BP |
2055 | } |
2056 | ||
2057 | static void | |
2058 | dpif_netdev_port_poll_wait(const struct dpif *dpif_) | |
2059 | { | |
2060 | struct dpif_netdev *dpif = dpif_netdev_cast(dpif_); | |
5279f8fd | 2061 | |
d33ed218 | 2062 | seq_wait(dpif->dp->port_seq, dpif->last_port_seq); |
8a4e3a85 BP |
2063 | } |
2064 | ||
2065 | static struct dp_netdev_flow * | |
0de8783a | 2066 | dp_netdev_flow_cast(const struct dpcls_rule *cr) |
8a4e3a85 BP |
2067 | { |
2068 | return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL; | |
72865317 BP |
2069 | } |
2070 | ||
9bbf1c3d DDP |
2071 | static bool dp_netdev_flow_ref(struct dp_netdev_flow *flow) |
2072 | { | |
2073 | return ovs_refcount_try_ref_rcu(&flow->ref_cnt); | |
2074 | } | |
2075 | ||
79df317f DDP |
2076 | /* netdev_flow_key utilities. |
2077 | * | |
2078 | * netdev_flow_key is basically a miniflow. We use these functions | |
2079 | * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow | |
2080 | * functions (miniflow_clone_inline, miniflow_equal, ...), because: | |
2081 | * | |
2082 | * - Since we are dealing exclusively with miniflows created by | |
2083 | * miniflow_extract(), if the map is different the miniflow is different. | |
2084 | * Therefore we can be faster by comparing the map and the miniflow in a | |
2085 | * single memcmp(). | |
5fcff47b | 2086 | * - These functions can be inlined by the compiler. */ |
79df317f | 2087 | |
361d808d | 2088 | /* Given the number of bits set in miniflow's maps, returns the size of the |
caeb4906 | 2089 | * 'netdev_flow_key.mf' */ |
361d808d JR |
2090 | static inline size_t |
2091 | netdev_flow_key_size(size_t flow_u64s) | |
79df317f | 2092 | { |
361d808d | 2093 | return sizeof(struct miniflow) + MINIFLOW_VALUES_SIZE(flow_u64s); |
79df317f DDP |
2094 | } |
2095 | ||
79df317f DDP |
2096 | static inline bool |
2097 | netdev_flow_key_equal(const struct netdev_flow_key *a, | |
0de8783a JR |
2098 | const struct netdev_flow_key *b) |
2099 | { | |
caeb4906 JR |
2100 | /* 'b->len' may be not set yet. */ |
2101 | return a->hash == b->hash && !memcmp(&a->mf, &b->mf, a->len); | |
0de8783a JR |
2102 | } |
2103 | ||
2104 | /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow. | |
d79a39fe | 2105 | * The maps are compared bitwise, so both 'key->mf' and 'mf' must have been |
0de8783a JR |
2106 | * generated by miniflow_extract. */ |
2107 | static inline bool | |
2108 | netdev_flow_key_equal_mf(const struct netdev_flow_key *key, | |
2109 | const struct miniflow *mf) | |
79df317f | 2110 | { |
caeb4906 | 2111 | return !memcmp(&key->mf, mf, key->len); |
79df317f DDP |
2112 | } |
2113 | ||
2114 | static inline void | |
2115 | netdev_flow_key_clone(struct netdev_flow_key *dst, | |
0de8783a JR |
2116 | const struct netdev_flow_key *src) |
2117 | { | |
caeb4906 JR |
2118 | memcpy(dst, src, |
2119 | offsetof(struct netdev_flow_key, mf) + src->len); | |
0de8783a JR |
2120 | } |
2121 | ||
0de8783a JR |
2122 | /* Initialize a netdev_flow_key 'mask' from 'match'. */ |
2123 | static inline void | |
2124 | netdev_flow_mask_init(struct netdev_flow_key *mask, | |
2125 | const struct match *match) | |
2126 | { | |
09b0fa9c | 2127 | uint64_t *dst = miniflow_values(&mask->mf); |
5fcff47b | 2128 | struct flowmap fmap; |
0de8783a | 2129 | uint32_t hash = 0; |
5fcff47b | 2130 | size_t idx; |
0de8783a JR |
2131 | |
2132 | /* Only check masks that make sense for the flow. */ | |
5fcff47b JR |
2133 | flow_wc_map(&match->flow, &fmap); |
2134 | flowmap_init(&mask->mf.map); | |
0de8783a | 2135 | |
5fcff47b JR |
2136 | FLOWMAP_FOR_EACH_INDEX(idx, fmap) { |
2137 | uint64_t mask_u64 = flow_u64_value(&match->wc.masks, idx); | |
0de8783a | 2138 | |
5fcff47b JR |
2139 | if (mask_u64) { |
2140 | flowmap_set(&mask->mf.map, idx, 1); | |
2141 | *dst++ = mask_u64; | |
2142 | hash = hash_add64(hash, mask_u64); | |
0de8783a | 2143 | } |
0de8783a JR |
2144 | } |
2145 | ||
5fcff47b | 2146 | map_t map; |
0de8783a | 2147 | |
5fcff47b JR |
2148 | FLOWMAP_FOR_EACH_MAP (map, mask->mf.map) { |
2149 | hash = hash_add64(hash, map); | |
2150 | } | |
0de8783a | 2151 | |
5fcff47b | 2152 | size_t n = dst - miniflow_get_values(&mask->mf); |
0de8783a | 2153 | |
d70e8c28 | 2154 | mask->hash = hash_finish(hash, n * 8); |
0de8783a JR |
2155 | mask->len = netdev_flow_key_size(n); |
2156 | } | |
2157 | ||
361d808d | 2158 | /* Initializes 'dst' as a copy of 'flow' masked with 'mask'. */ |
0de8783a JR |
2159 | static inline void |
2160 | netdev_flow_key_init_masked(struct netdev_flow_key *dst, | |
2161 | const struct flow *flow, | |
2162 | const struct netdev_flow_key *mask) | |
79df317f | 2163 | { |
09b0fa9c JR |
2164 | uint64_t *dst_u64 = miniflow_values(&dst->mf); |
2165 | const uint64_t *mask_u64 = miniflow_get_values(&mask->mf); | |
0de8783a | 2166 | uint32_t hash = 0; |
d70e8c28 | 2167 | uint64_t value; |
0de8783a JR |
2168 | |
2169 | dst->len = mask->len; | |
361d808d | 2170 | dst->mf = mask->mf; /* Copy maps. */ |
0de8783a | 2171 | |
5fcff47b | 2172 | FLOW_FOR_EACH_IN_MAPS(value, flow, mask->mf.map) { |
d70e8c28 JR |
2173 | *dst_u64 = value & *mask_u64++; |
2174 | hash = hash_add64(hash, *dst_u64++); | |
0de8783a | 2175 | } |
09b0fa9c JR |
2176 | dst->hash = hash_finish(hash, |
2177 | (dst_u64 - miniflow_get_values(&dst->mf)) * 8); | |
0de8783a JR |
2178 | } |
2179 | ||
5fcff47b JR |
2180 | /* Iterate through netdev_flow_key TNL u64 values specified by 'FLOWMAP'. */ |
2181 | #define NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(VALUE, KEY, FLOWMAP) \ | |
2182 | MINIFLOW_FOR_EACH_IN_FLOWMAP(VALUE, &(KEY)->mf, FLOWMAP) | |
0de8783a JR |
2183 | |
2184 | /* Returns a hash value for the bits of 'key' where there are 1-bits in | |
2185 | * 'mask'. */ | |
2186 | static inline uint32_t | |
2187 | netdev_flow_key_hash_in_mask(const struct netdev_flow_key *key, | |
2188 | const struct netdev_flow_key *mask) | |
2189 | { | |
09b0fa9c | 2190 | const uint64_t *p = miniflow_get_values(&mask->mf); |
0de8783a | 2191 | uint32_t hash = 0; |
5fcff47b | 2192 | uint64_t value; |
0de8783a | 2193 | |
5fcff47b JR |
2194 | NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value, key, mask->mf.map) { |
2195 | hash = hash_add64(hash, value & *p++); | |
0de8783a JR |
2196 | } |
2197 | ||
09b0fa9c | 2198 | return hash_finish(hash, (p - miniflow_get_values(&mask->mf)) * 8); |
79df317f DDP |
2199 | } |
2200 | ||
9bbf1c3d DDP |
2201 | static inline bool |
2202 | emc_entry_alive(struct emc_entry *ce) | |
2203 | { | |
2204 | return ce->flow && !ce->flow->dead; | |
2205 | } | |
2206 | ||
2207 | static void | |
2208 | emc_clear_entry(struct emc_entry *ce) | |
2209 | { | |
2210 | if (ce->flow) { | |
2211 | dp_netdev_flow_unref(ce->flow); | |
2212 | ce->flow = NULL; | |
2213 | } | |
2214 | } | |
2215 | ||
2216 | static inline void | |
2217 | emc_change_entry(struct emc_entry *ce, struct dp_netdev_flow *flow, | |
0de8783a | 2218 | const struct netdev_flow_key *key) |
9bbf1c3d DDP |
2219 | { |
2220 | if (ce->flow != flow) { | |
2221 | if (ce->flow) { | |
2222 | dp_netdev_flow_unref(ce->flow); | |
2223 | } | |
2224 | ||
2225 | if (dp_netdev_flow_ref(flow)) { | |
2226 | ce->flow = flow; | |
2227 | } else { | |
2228 | ce->flow = NULL; | |
2229 | } | |
2230 | } | |
0de8783a JR |
2231 | if (key) { |
2232 | netdev_flow_key_clone(&ce->key, key); | |
9bbf1c3d DDP |
2233 | } |
2234 | } | |
2235 | ||
2236 | static inline void | |
0de8783a | 2237 | emc_insert(struct emc_cache *cache, const struct netdev_flow_key *key, |
9bbf1c3d DDP |
2238 | struct dp_netdev_flow *flow) |
2239 | { | |
2240 | struct emc_entry *to_be_replaced = NULL; | |
2241 | struct emc_entry *current_entry; | |
2242 | ||
0de8783a JR |
2243 | EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) { |
2244 | if (netdev_flow_key_equal(¤t_entry->key, key)) { | |
9bbf1c3d | 2245 | /* We found the entry with the 'mf' miniflow */ |
0de8783a | 2246 | emc_change_entry(current_entry, flow, NULL); |
9bbf1c3d DDP |
2247 | return; |
2248 | } | |
2249 | ||
2250 | /* Replacement policy: put the flow in an empty (not alive) entry, or | |
2251 | * in the first entry where it can be */ | |
2252 | if (!to_be_replaced | |
2253 | || (emc_entry_alive(to_be_replaced) | |
2254 | && !emc_entry_alive(current_entry)) | |
0de8783a | 2255 | || current_entry->key.hash < to_be_replaced->key.hash) { |
9bbf1c3d DDP |
2256 | to_be_replaced = current_entry; |
2257 | } | |
2258 | } | |
2259 | /* We didn't find the miniflow in the cache. | |
2260 | * The 'to_be_replaced' entry is where the new flow will be stored */ | |
2261 | ||
0de8783a | 2262 | emc_change_entry(to_be_replaced, flow, key); |
9bbf1c3d DDP |
2263 | } |
2264 | ||
4c30b246 CL |
2265 | static inline void |
2266 | emc_probabilistic_insert(struct dp_netdev_pmd_thread *pmd, | |
2267 | const struct netdev_flow_key *key, | |
2268 | struct dp_netdev_flow *flow) | |
2269 | { | |
2270 | /* Insert an entry into the EMC based on probability value 'min'. By | |
2271 | * default the value is UINT32_MAX / 100 which yields an insertion | |
2272 | * probability of 1/100 ie. 1% */ | |
2273 | ||
2274 | uint32_t min; | |
2275 | atomic_read_relaxed(&pmd->dp->emc_insert_min, &min); | |
2276 | ||
656238ee | 2277 | if (min && random_uint32() <= min) { |
4c30b246 CL |
2278 | emc_insert(&pmd->flow_cache, key, flow); |
2279 | } | |
2280 | } | |
2281 | ||
9bbf1c3d | 2282 | static inline struct dp_netdev_flow * |
0de8783a | 2283 | emc_lookup(struct emc_cache *cache, const struct netdev_flow_key *key) |
9bbf1c3d DDP |
2284 | { |
2285 | struct emc_entry *current_entry; | |
2286 | ||
0de8783a JR |
2287 | EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) { |
2288 | if (current_entry->key.hash == key->hash | |
2289 | && emc_entry_alive(current_entry) | |
2290 | && netdev_flow_key_equal_mf(¤t_entry->key, &key->mf)) { | |
9bbf1c3d | 2291 | |
0de8783a | 2292 | /* We found the entry with the 'key->mf' miniflow */ |
9bbf1c3d DDP |
2293 | return current_entry->flow; |
2294 | } | |
2295 | } | |
2296 | ||
2297 | return NULL; | |
2298 | } | |
2299 | ||
72865317 | 2300 | static struct dp_netdev_flow * |
3453b4d6 JS |
2301 | dp_netdev_pmd_lookup_flow(struct dp_netdev_pmd_thread *pmd, |
2302 | const struct netdev_flow_key *key, | |
2303 | int *lookup_num_p) | |
2c0ea78f | 2304 | { |
3453b4d6 | 2305 | struct dpcls *cls; |
0de8783a | 2306 | struct dpcls_rule *rule; |
f825fdd4 BP |
2307 | odp_port_t in_port = u32_to_odp(MINIFLOW_GET_U32(&key->mf, |
2308 | in_port.odp_port)); | |
3453b4d6 | 2309 | struct dp_netdev_flow *netdev_flow = NULL; |
2c0ea78f | 2310 | |
3453b4d6 JS |
2311 | cls = dp_netdev_pmd_lookup_dpcls(pmd, in_port); |
2312 | if (OVS_LIKELY(cls)) { | |
2313 | dpcls_lookup(cls, key, &rule, 1, lookup_num_p); | |
2314 | netdev_flow = dp_netdev_flow_cast(rule); | |
2315 | } | |
8a4e3a85 | 2316 | return netdev_flow; |
2c0ea78f GS |
2317 | } |
2318 | ||
2319 | static struct dp_netdev_flow * | |
1c1e46ed AW |
2320 | dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread *pmd, |
2321 | const ovs_u128 *ufidp, const struct nlattr *key, | |
2322 | size_t key_len) | |
72865317 | 2323 | { |
1763b4b8 | 2324 | struct dp_netdev_flow *netdev_flow; |
70e5ed6f JS |
2325 | struct flow flow; |
2326 | ovs_u128 ufid; | |
2327 | ||
2328 | /* If a UFID is not provided, determine one based on the key. */ | |
2329 | if (!ufidp && key && key_len | |
f0fb825a | 2330 | && !dpif_netdev_flow_from_nlattrs(key, key_len, &flow, false)) { |
1c1e46ed | 2331 | dpif_flow_hash(pmd->dp->dpif, &flow, sizeof flow, &ufid); |
70e5ed6f JS |
2332 | ufidp = &ufid; |
2333 | } | |
72865317 | 2334 | |
70e5ed6f JS |
2335 | if (ufidp) { |
2336 | CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, dp_netdev_flow_hash(ufidp), | |
1c1e46ed | 2337 | &pmd->flow_table) { |
2ff8484b | 2338 | if (ovs_u128_equals(netdev_flow->ufid, *ufidp)) { |
70e5ed6f JS |
2339 | return netdev_flow; |
2340 | } | |
72865317 BP |
2341 | } |
2342 | } | |
8a4e3a85 | 2343 | |
72865317 BP |
2344 | return NULL; |
2345 | } | |
2346 | ||
2347 | static void | |
eb94da30 | 2348 | get_dpif_flow_stats(const struct dp_netdev_flow *netdev_flow_, |
1763b4b8 | 2349 | struct dpif_flow_stats *stats) |
feebdea2 | 2350 | { |
eb94da30 DDP |
2351 | struct dp_netdev_flow *netdev_flow; |
2352 | unsigned long long n; | |
2353 | long long used; | |
2354 | uint16_t flags; | |
2355 | ||
2356 | netdev_flow = CONST_CAST(struct dp_netdev_flow *, netdev_flow_); | |
2357 | ||
2358 | atomic_read_relaxed(&netdev_flow->stats.packet_count, &n); | |
2359 | stats->n_packets = n; | |
2360 | atomic_read_relaxed(&netdev_flow->stats.byte_count, &n); | |
2361 | stats->n_bytes = n; | |
2362 | atomic_read_relaxed(&netdev_flow->stats.used, &used); | |
2363 | stats->used = used; | |
2364 | atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags); | |
2365 | stats->tcp_flags = flags; | |
72865317 BP |
2366 | } |
2367 | ||
7af12bd7 JS |
2368 | /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for |
2369 | * storing the netlink-formatted key/mask. 'key_buf' may be the same as | |
2370 | * 'mask_buf'. Actions will be returned without copying, by relying on RCU to | |
2371 | * protect them. */ | |
6fe09f8c | 2372 | static void |
70e5ed6f | 2373 | dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow *netdev_flow, |
7af12bd7 | 2374 | struct ofpbuf *key_buf, struct ofpbuf *mask_buf, |
64bb477f | 2375 | struct dpif_flow *flow, bool terse) |
6fe09f8c | 2376 | { |
64bb477f JS |
2377 | if (terse) { |
2378 | memset(flow, 0, sizeof *flow); | |
2379 | } else { | |
2380 | struct flow_wildcards wc; | |
2381 | struct dp_netdev_actions *actions; | |
2382 | size_t offset; | |
5262eea1 JG |
2383 | struct odp_flow_key_parms odp_parms = { |
2384 | .flow = &netdev_flow->flow, | |
2385 | .mask = &wc.masks, | |
2494ccd7 | 2386 | .support = dp_netdev_support, |
5262eea1 | 2387 | }; |
64bb477f JS |
2388 | |
2389 | miniflow_expand(&netdev_flow->cr.mask->mf, &wc.masks); | |
f4b835bb JR |
2390 | /* in_port is exact matched, but we have left it out from the mask for |
2391 | * optimnization reasons. Add in_port back to the mask. */ | |
2392 | wc.masks.in_port.odp_port = ODPP_NONE; | |
64bb477f JS |
2393 | |
2394 | /* Key */ | |
6fd6ed71 | 2395 | offset = key_buf->size; |
64bb477f | 2396 | flow->key = ofpbuf_tail(key_buf); |
5262eea1 | 2397 | odp_flow_key_from_flow(&odp_parms, key_buf); |
6fd6ed71 | 2398 | flow->key_len = key_buf->size - offset; |
64bb477f JS |
2399 | |
2400 | /* Mask */ | |
6fd6ed71 | 2401 | offset = mask_buf->size; |
64bb477f | 2402 | flow->mask = ofpbuf_tail(mask_buf); |
ec1f6f32 | 2403 | odp_parms.key_buf = key_buf; |
5262eea1 | 2404 | odp_flow_key_from_mask(&odp_parms, mask_buf); |
6fd6ed71 | 2405 | flow->mask_len = mask_buf->size - offset; |
64bb477f JS |
2406 | |
2407 | /* Actions */ | |
2408 | actions = dp_netdev_flow_get_actions(netdev_flow); | |
2409 | flow->actions = actions->actions; | |
2410 | flow->actions_len = actions->size; | |
2411 | } | |
6fe09f8c | 2412 | |
70e5ed6f JS |
2413 | flow->ufid = netdev_flow->ufid; |
2414 | flow->ufid_present = true; | |
1c1e46ed | 2415 | flow->pmd_id = netdev_flow->pmd_id; |
6fe09f8c JS |
2416 | get_dpif_flow_stats(netdev_flow, &flow->stats); |
2417 | } | |
2418 | ||
36956a7d | 2419 | static int |
8c301900 JR |
2420 | dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len, |
2421 | const struct nlattr *mask_key, | |
2422 | uint32_t mask_key_len, const struct flow *flow, | |
f0fb825a | 2423 | struct flow_wildcards *wc, bool probe) |
8c301900 | 2424 | { |
ca8d3442 DDP |
2425 | enum odp_key_fitness fitness; |
2426 | ||
8d8ab6c2 | 2427 | fitness = odp_flow_key_to_mask(mask_key, mask_key_len, wc, flow); |
ca8d3442 | 2428 | if (fitness) { |
f0fb825a EG |
2429 | if (!probe) { |
2430 | /* This should not happen: it indicates that | |
2431 | * odp_flow_key_from_mask() and odp_flow_key_to_mask() | |
2432 | * disagree on the acceptable form of a mask. Log the problem | |
2433 | * as an error, with enough details to enable debugging. */ | |
2434 | static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5); | |
2435 | ||
2436 | if (!VLOG_DROP_ERR(&rl)) { | |
2437 | struct ds s; | |
2438 | ||
2439 | ds_init(&s); | |
2440 | odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s, | |
2441 | true); | |
2442 | VLOG_ERR("internal error parsing flow mask %s (%s)", | |
2443 | ds_cstr(&s), odp_key_fitness_to_string(fitness)); | |
2444 | ds_destroy(&s); | |
2445 | } | |
8c301900 | 2446 | } |
ca8d3442 DDP |
2447 | |
2448 | return EINVAL; | |
8c301900 JR |
2449 | } |
2450 | ||
2451 | return 0; | |
2452 | } | |
2453 | ||
2454 | static int | |
2455 | dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len, | |
f0fb825a | 2456 | struct flow *flow, bool probe) |
36956a7d | 2457 | { |
8d8ab6c2 | 2458 | if (odp_flow_key_to_flow(key, key_len, flow)) { |
f0fb825a EG |
2459 | if (!probe) { |
2460 | /* This should not happen: it indicates that | |
2461 | * odp_flow_key_from_flow() and odp_flow_key_to_flow() disagree on | |
2462 | * the acceptable form of a flow. Log the problem as an error, | |
2463 | * with enough details to enable debugging. */ | |
2464 | static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5); | |
2465 | ||
2466 | if (!VLOG_DROP_ERR(&rl)) { | |
2467 | struct ds s; | |
2468 | ||
2469 | ds_init(&s); | |
2470 | odp_flow_format(key, key_len, NULL, 0, NULL, &s, true); | |
2471 | VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s)); | |
2472 | ds_destroy(&s); | |
2473 | } | |
36956a7d BP |
2474 | } |
2475 | ||
2476 | return EINVAL; | |
2477 | } | |
2478 | ||
5cf3edb3 | 2479 | if (flow->ct_state & DP_NETDEV_CS_UNSUPPORTED_MASK) { |
07659514 JS |
2480 | return EINVAL; |
2481 | } | |
2482 | ||
36956a7d BP |
2483 | return 0; |
2484 | } | |
2485 | ||
72865317 | 2486 | static int |
6fe09f8c | 2487 | dpif_netdev_flow_get(const struct dpif *dpif, const struct dpif_flow_get *get) |
72865317 BP |
2488 | { |
2489 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
1763b4b8 | 2490 | struct dp_netdev_flow *netdev_flow; |
1c1e46ed | 2491 | struct dp_netdev_pmd_thread *pmd; |
c673049c IM |
2492 | struct hmapx to_find = HMAPX_INITIALIZER(&to_find); |
2493 | struct hmapx_node *node; | |
2494 | int error = EINVAL; | |
2495 | ||
2496 | if (get->pmd_id == PMD_ID_NULL) { | |
2497 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
2498 | if (dp_netdev_pmd_try_ref(pmd) && !hmapx_add(&to_find, pmd)) { | |
2499 | dp_netdev_pmd_unref(pmd); | |
2500 | } | |
2501 | } | |
2502 | } else { | |
2503 | pmd = dp_netdev_get_pmd(dp, get->pmd_id); | |
2504 | if (!pmd) { | |
2505 | goto out; | |
2506 | } | |
2507 | hmapx_add(&to_find, pmd); | |
1c1e46ed AW |
2508 | } |
2509 | ||
c673049c IM |
2510 | if (!hmapx_count(&to_find)) { |
2511 | goto out; | |
72865317 | 2512 | } |
1c1e46ed | 2513 | |
c673049c IM |
2514 | HMAPX_FOR_EACH (node, &to_find) { |
2515 | pmd = (struct dp_netdev_pmd_thread *) node->data; | |
2516 | netdev_flow = dp_netdev_pmd_find_flow(pmd, get->ufid, get->key, | |
2517 | get->key_len); | |
2518 | if (netdev_flow) { | |
2519 | dp_netdev_flow_to_dpif_flow(netdev_flow, get->buffer, get->buffer, | |
2520 | get->flow, false); | |
2521 | error = 0; | |
2522 | break; | |
2523 | } else { | |
2524 | error = ENOENT; | |
2525 | } | |
2526 | } | |
bc4a05c6 | 2527 | |
c673049c IM |
2528 | HMAPX_FOR_EACH (node, &to_find) { |
2529 | pmd = (struct dp_netdev_pmd_thread *) node->data; | |
2530 | dp_netdev_pmd_unref(pmd); | |
2531 | } | |
2532 | out: | |
2533 | hmapx_destroy(&to_find); | |
5279f8fd | 2534 | return error; |
72865317 BP |
2535 | } |
2536 | ||
0de8783a | 2537 | static struct dp_netdev_flow * |
1c1e46ed AW |
2538 | dp_netdev_flow_add(struct dp_netdev_pmd_thread *pmd, |
2539 | struct match *match, const ovs_u128 *ufid, | |
ae2ceebd | 2540 | const struct nlattr *actions, size_t actions_len) |
1c1e46ed | 2541 | OVS_REQUIRES(pmd->flow_mutex) |
72865317 | 2542 | { |
0de8783a JR |
2543 | struct dp_netdev_flow *flow; |
2544 | struct netdev_flow_key mask; | |
3453b4d6 | 2545 | struct dpcls *cls; |
f4b835bb JR |
2546 | |
2547 | /* Make sure in_port is exact matched before we read it. */ | |
2548 | ovs_assert(match->wc.masks.in_port.odp_port == ODPP_NONE); | |
3453b4d6 | 2549 | odp_port_t in_port = match->flow.in_port.odp_port; |
ed79f89a | 2550 | |
f4b835bb JR |
2551 | /* As we select the dpcls based on the port number, each netdev flow |
2552 | * belonging to the same dpcls will have the same odp_port value. | |
2553 | * For performance reasons we wildcard odp_port here in the mask. In the | |
2554 | * typical case dp_hash is also wildcarded, and the resulting 8-byte | |
2555 | * chunk {dp_hash, in_port} will be ignored by netdev_flow_mask_init() and | |
2556 | * will not be part of the subtable mask. | |
2557 | * This will speed up the hash computation during dpcls_lookup() because | |
2558 | * there is one less call to hash_add64() in this case. */ | |
2559 | match->wc.masks.in_port.odp_port = 0; | |
0de8783a | 2560 | netdev_flow_mask_init(&mask, match); |
f4b835bb JR |
2561 | match->wc.masks.in_port.odp_port = ODPP_NONE; |
2562 | ||
0de8783a | 2563 | /* Make sure wc does not have metadata. */ |
5fcff47b JR |
2564 | ovs_assert(!FLOWMAP_HAS_FIELD(&mask.mf.map, metadata) |
2565 | && !FLOWMAP_HAS_FIELD(&mask.mf.map, regs)); | |
679ba04c | 2566 | |
0de8783a | 2567 | /* Do not allocate extra space. */ |
caeb4906 | 2568 | flow = xmalloc(sizeof *flow - sizeof flow->cr.flow.mf + mask.len); |
1c1e46ed | 2569 | memset(&flow->stats, 0, sizeof flow->stats); |
0de8783a | 2570 | flow->dead = false; |
11e5cf1f | 2571 | flow->batch = NULL; |
bd5131ba | 2572 | *CONST_CAST(unsigned *, &flow->pmd_id) = pmd->core_id; |
0de8783a | 2573 | *CONST_CAST(struct flow *, &flow->flow) = match->flow; |
70e5ed6f | 2574 | *CONST_CAST(ovs_u128 *, &flow->ufid) = *ufid; |
0de8783a | 2575 | ovs_refcount_init(&flow->ref_cnt); |
0de8783a | 2576 | ovsrcu_set(&flow->actions, dp_netdev_actions_create(actions, actions_len)); |
2c0ea78f | 2577 | |
0de8783a | 2578 | netdev_flow_key_init_masked(&flow->cr.flow, &match->flow, &mask); |
3453b4d6 | 2579 | |
f4b835bb | 2580 | /* Select dpcls for in_port. Relies on in_port to be exact match. */ |
3453b4d6 JS |
2581 | cls = dp_netdev_pmd_find_dpcls(pmd, in_port); |
2582 | dpcls_insert(cls, &flow->cr, &mask); | |
72865317 | 2583 | |
4c75aaab EJ |
2584 | cmap_insert(&pmd->flow_table, CONST_CAST(struct cmap_node *, &flow->node), |
2585 | dp_netdev_flow_hash(&flow->ufid)); | |
2586 | ||
beb75a40 | 2587 | if (OVS_UNLIKELY(!VLOG_DROP_DBG((&upcall_rl)))) { |
623540e4 | 2588 | struct ds ds = DS_EMPTY_INITIALIZER; |
9044f2c1 JG |
2589 | struct ofpbuf key_buf, mask_buf; |
2590 | struct odp_flow_key_parms odp_parms = { | |
2591 | .flow = &match->flow, | |
2592 | .mask = &match->wc.masks, | |
2593 | .support = dp_netdev_support, | |
2594 | }; | |
2595 | ||
2596 | ofpbuf_init(&key_buf, 0); | |
2597 | ofpbuf_init(&mask_buf, 0); | |
623540e4 | 2598 | |
9044f2c1 JG |
2599 | odp_flow_key_from_flow(&odp_parms, &key_buf); |
2600 | odp_parms.key_buf = &key_buf; | |
2601 | odp_flow_key_from_mask(&odp_parms, &mask_buf); | |
0de8783a | 2602 | |
623540e4 | 2603 | ds_put_cstr(&ds, "flow_add: "); |
70e5ed6f JS |
2604 | odp_format_ufid(ufid, &ds); |
2605 | ds_put_cstr(&ds, " "); | |
9044f2c1 JG |
2606 | odp_flow_format(key_buf.data, key_buf.size, |
2607 | mask_buf.data, mask_buf.size, | |
2608 | NULL, &ds, false); | |
623540e4 | 2609 | ds_put_cstr(&ds, ", actions:"); |
0722f341 | 2610 | format_odp_actions(&ds, actions, actions_len, NULL); |
623540e4 | 2611 | |
beb75a40 | 2612 | VLOG_DBG("%s", ds_cstr(&ds)); |
623540e4 | 2613 | |
9044f2c1 JG |
2614 | ofpbuf_uninit(&key_buf); |
2615 | ofpbuf_uninit(&mask_buf); | |
beb75a40 JS |
2616 | |
2617 | /* Add a printout of the actual match installed. */ | |
2618 | struct match m; | |
2619 | ds_clear(&ds); | |
2620 | ds_put_cstr(&ds, "flow match: "); | |
2621 | miniflow_expand(&flow->cr.flow.mf, &m.flow); | |
2622 | miniflow_expand(&flow->cr.mask->mf, &m.wc.masks); | |
b2f4b622 | 2623 | memset(&m.tun_md, 0, sizeof m.tun_md); |
beb75a40 JS |
2624 | match_format(&m, NULL, &ds, OFP_DEFAULT_PRIORITY); |
2625 | ||
2626 | VLOG_DBG("%s", ds_cstr(&ds)); | |
2627 | ||
623540e4 EJ |
2628 | ds_destroy(&ds); |
2629 | } | |
2630 | ||
0de8783a | 2631 | return flow; |
72865317 BP |
2632 | } |
2633 | ||
72865317 | 2634 | static int |
f5d317a1 DDP |
2635 | flow_put_on_pmd(struct dp_netdev_pmd_thread *pmd, |
2636 | struct netdev_flow_key *key, | |
2637 | struct match *match, | |
2638 | ovs_u128 *ufid, | |
2639 | const struct dpif_flow_put *put, | |
2640 | struct dpif_flow_stats *stats) | |
72865317 | 2641 | { |
1763b4b8 | 2642 | struct dp_netdev_flow *netdev_flow; |
f5d317a1 | 2643 | int error = 0; |
72865317 | 2644 | |
f5d317a1 DDP |
2645 | if (stats) { |
2646 | memset(stats, 0, sizeof *stats); | |
70e5ed6f JS |
2647 | } |
2648 | ||
1c1e46ed | 2649 | ovs_mutex_lock(&pmd->flow_mutex); |
f5d317a1 | 2650 | netdev_flow = dp_netdev_pmd_lookup_flow(pmd, key, NULL); |
1763b4b8 | 2651 | if (!netdev_flow) { |
89625d1e | 2652 | if (put->flags & DPIF_FP_CREATE) { |
1c1e46ed | 2653 | if (cmap_count(&pmd->flow_table) < MAX_FLOWS) { |
f5d317a1 | 2654 | dp_netdev_flow_add(pmd, match, ufid, put->actions, |
70e5ed6f | 2655 | put->actions_len); |
0de8783a | 2656 | error = 0; |
72865317 | 2657 | } else { |
5279f8fd | 2658 | error = EFBIG; |
72865317 BP |
2659 | } |
2660 | } else { | |
5279f8fd | 2661 | error = ENOENT; |
72865317 BP |
2662 | } |
2663 | } else { | |
beb75a40 | 2664 | if (put->flags & DPIF_FP_MODIFY) { |
8a4e3a85 BP |
2665 | struct dp_netdev_actions *new_actions; |
2666 | struct dp_netdev_actions *old_actions; | |
2667 | ||
2668 | new_actions = dp_netdev_actions_create(put->actions, | |
2669 | put->actions_len); | |
2670 | ||
61e7deb1 BP |
2671 | old_actions = dp_netdev_flow_get_actions(netdev_flow); |
2672 | ovsrcu_set(&netdev_flow->actions, new_actions); | |
679ba04c | 2673 | |
f5d317a1 DDP |
2674 | if (stats) { |
2675 | get_dpif_flow_stats(netdev_flow, stats); | |
a84cb64a BP |
2676 | } |
2677 | if (put->flags & DPIF_FP_ZERO_STATS) { | |
97447f55 DDP |
2678 | /* XXX: The userspace datapath uses thread local statistics |
2679 | * (for flows), which should be updated only by the owning | |
2680 | * thread. Since we cannot write on stats memory here, | |
2681 | * we choose not to support this flag. Please note: | |
2682 | * - This feature is currently used only by dpctl commands with | |
2683 | * option --clear. | |
2684 | * - Should the need arise, this operation can be implemented | |
2685 | * by keeping a base value (to be update here) for each | |
2686 | * counter, and subtracting it before outputting the stats */ | |
2687 | error = EOPNOTSUPP; | |
72865317 | 2688 | } |
8a4e3a85 | 2689 | |
61e7deb1 | 2690 | ovsrcu_postpone(dp_netdev_actions_free, old_actions); |
2c0ea78f | 2691 | } else if (put->flags & DPIF_FP_CREATE) { |
5279f8fd | 2692 | error = EEXIST; |
2c0ea78f GS |
2693 | } else { |
2694 | /* Overlapping flow. */ | |
2695 | error = EINVAL; | |
72865317 BP |
2696 | } |
2697 | } | |
1c1e46ed | 2698 | ovs_mutex_unlock(&pmd->flow_mutex); |
5279f8fd | 2699 | return error; |
72865317 BP |
2700 | } |
2701 | ||
72865317 | 2702 | static int |
f5d317a1 | 2703 | dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put) |
72865317 BP |
2704 | { |
2705 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
beb75a40 | 2706 | struct netdev_flow_key key, mask; |
1c1e46ed | 2707 | struct dp_netdev_pmd_thread *pmd; |
f5d317a1 DDP |
2708 | struct match match; |
2709 | ovs_u128 ufid; | |
2710 | int error; | |
f0fb825a | 2711 | bool probe = put->flags & DPIF_FP_PROBE; |
72865317 | 2712 | |
f5d317a1 DDP |
2713 | if (put->stats) { |
2714 | memset(put->stats, 0, sizeof *put->stats); | |
2715 | } | |
f0fb825a EG |
2716 | error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow, |
2717 | probe); | |
f5d317a1 DDP |
2718 | if (error) { |
2719 | return error; | |
2720 | } | |
2721 | error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len, | |
2722 | put->mask, put->mask_len, | |
f0fb825a | 2723 | &match.flow, &match.wc, probe); |
f5d317a1 DDP |
2724 | if (error) { |
2725 | return error; | |
1c1e46ed AW |
2726 | } |
2727 | ||
f5d317a1 DDP |
2728 | if (put->ufid) { |
2729 | ufid = *put->ufid; | |
2730 | } else { | |
2731 | dpif_flow_hash(dpif, &match.flow, sizeof match.flow, &ufid); | |
2732 | } | |
2733 | ||
2734 | /* Must produce a netdev_flow_key for lookup. | |
beb75a40 JS |
2735 | * Use the same method as employed to create the key when adding |
2736 | * the flow to the dplcs to make sure they match. */ | |
2737 | netdev_flow_mask_init(&mask, &match); | |
2738 | netdev_flow_key_init_masked(&key, &match.flow, &mask); | |
f5d317a1 DDP |
2739 | |
2740 | if (put->pmd_id == PMD_ID_NULL) { | |
2741 | if (cmap_count(&dp->poll_threads) == 0) { | |
2742 | return EINVAL; | |
2743 | } | |
2744 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
2745 | struct dpif_flow_stats pmd_stats; | |
2746 | int pmd_error; | |
2747 | ||
2748 | pmd_error = flow_put_on_pmd(pmd, &key, &match, &ufid, put, | |
2749 | &pmd_stats); | |
2750 | if (pmd_error) { | |
2751 | error = pmd_error; | |
2752 | } else if (put->stats) { | |
2753 | put->stats->n_packets += pmd_stats.n_packets; | |
2754 | put->stats->n_bytes += pmd_stats.n_bytes; | |
2755 | put->stats->used = MAX(put->stats->used, pmd_stats.used); | |
2756 | put->stats->tcp_flags |= pmd_stats.tcp_flags; | |
2757 | } | |
2758 | } | |
2759 | } else { | |
2760 | pmd = dp_netdev_get_pmd(dp, put->pmd_id); | |
2761 | if (!pmd) { | |
2762 | return EINVAL; | |
2763 | } | |
2764 | error = flow_put_on_pmd(pmd, &key, &match, &ufid, put, put->stats); | |
2765 | dp_netdev_pmd_unref(pmd); | |
2766 | } | |
2767 | ||
2768 | return error; | |
2769 | } | |
2770 | ||
2771 | static int | |
2772 | flow_del_on_pmd(struct dp_netdev_pmd_thread *pmd, | |
2773 | struct dpif_flow_stats *stats, | |
2774 | const struct dpif_flow_del *del) | |
2775 | { | |
2776 | struct dp_netdev_flow *netdev_flow; | |
2777 | int error = 0; | |
2778 | ||
1c1e46ed AW |
2779 | ovs_mutex_lock(&pmd->flow_mutex); |
2780 | netdev_flow = dp_netdev_pmd_find_flow(pmd, del->ufid, del->key, | |
2781 | del->key_len); | |
1763b4b8 | 2782 | if (netdev_flow) { |
f5d317a1 DDP |
2783 | if (stats) { |
2784 | get_dpif_flow_stats(netdev_flow, stats); | |
feebdea2 | 2785 | } |
1c1e46ed | 2786 | dp_netdev_pmd_remove_flow(pmd, netdev_flow); |
72865317 | 2787 | } else { |
5279f8fd | 2788 | error = ENOENT; |
72865317 | 2789 | } |
1c1e46ed | 2790 | ovs_mutex_unlock(&pmd->flow_mutex); |
f5d317a1 DDP |
2791 | |
2792 | return error; | |
2793 | } | |
2794 | ||
2795 | static int | |
2796 | dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del) | |
2797 | { | |
2798 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
2799 | struct dp_netdev_pmd_thread *pmd; | |
2800 | int error = 0; | |
2801 | ||
2802 | if (del->stats) { | |
2803 | memset(del->stats, 0, sizeof *del->stats); | |
2804 | } | |
2805 | ||
2806 | if (del->pmd_id == PMD_ID_NULL) { | |
2807 | if (cmap_count(&dp->poll_threads) == 0) { | |
2808 | return EINVAL; | |
2809 | } | |
2810 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
2811 | struct dpif_flow_stats pmd_stats; | |
2812 | int pmd_error; | |
2813 | ||
2814 | pmd_error = flow_del_on_pmd(pmd, &pmd_stats, del); | |
2815 | if (pmd_error) { | |
2816 | error = pmd_error; | |
2817 | } else if (del->stats) { | |
2818 | del->stats->n_packets += pmd_stats.n_packets; | |
2819 | del->stats->n_bytes += pmd_stats.n_bytes; | |
2820 | del->stats->used = MAX(del->stats->used, pmd_stats.used); | |
2821 | del->stats->tcp_flags |= pmd_stats.tcp_flags; | |
2822 | } | |
2823 | } | |
2824 | } else { | |
2825 | pmd = dp_netdev_get_pmd(dp, del->pmd_id); | |
2826 | if (!pmd) { | |
2827 | return EINVAL; | |
2828 | } | |
2829 | error = flow_del_on_pmd(pmd, del->stats, del); | |
2830 | dp_netdev_pmd_unref(pmd); | |
2831 | } | |
2832 | ||
5279f8fd BP |
2833 | |
2834 | return error; | |
72865317 BP |
2835 | } |
2836 | ||
ac64794a BP |
2837 | struct dpif_netdev_flow_dump { |
2838 | struct dpif_flow_dump up; | |
1c1e46ed AW |
2839 | struct cmap_position poll_thread_pos; |
2840 | struct cmap_position flow_pos; | |
2841 | struct dp_netdev_pmd_thread *cur_pmd; | |
d2ad7ef1 JS |
2842 | int status; |
2843 | struct ovs_mutex mutex; | |
e723fd32 JS |
2844 | }; |
2845 | ||
ac64794a BP |
2846 | static struct dpif_netdev_flow_dump * |
2847 | dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump) | |
72865317 | 2848 | { |
ac64794a | 2849 | return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up); |
e723fd32 JS |
2850 | } |
2851 | ||
ac64794a | 2852 | static struct dpif_flow_dump * |
7e8b7199 PB |
2853 | dpif_netdev_flow_dump_create(const struct dpif *dpif_, bool terse, |
2854 | char *type OVS_UNUSED) | |
e723fd32 | 2855 | { |
ac64794a | 2856 | struct dpif_netdev_flow_dump *dump; |
e723fd32 | 2857 | |
1c1e46ed | 2858 | dump = xzalloc(sizeof *dump); |
ac64794a | 2859 | dpif_flow_dump_init(&dump->up, dpif_); |
64bb477f | 2860 | dump->up.terse = terse; |
ac64794a BP |
2861 | ovs_mutex_init(&dump->mutex); |
2862 | ||
2863 | return &dump->up; | |
e723fd32 JS |
2864 | } |
2865 | ||
2866 | static int | |
ac64794a | 2867 | dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_) |
e723fd32 | 2868 | { |
ac64794a | 2869 | struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_); |
e723fd32 | 2870 | |
ac64794a BP |
2871 | ovs_mutex_destroy(&dump->mutex); |
2872 | free(dump); | |
704a1e09 BP |
2873 | return 0; |
2874 | } | |
2875 | ||
ac64794a BP |
2876 | struct dpif_netdev_flow_dump_thread { |
2877 | struct dpif_flow_dump_thread up; | |
2878 | struct dpif_netdev_flow_dump *dump; | |
8bb113da RW |
2879 | struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH]; |
2880 | struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH]; | |
ac64794a BP |
2881 | }; |
2882 | ||
2883 | static struct dpif_netdev_flow_dump_thread * | |
2884 | dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread) | |
2885 | { | |
2886 | return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up); | |
2887 | } | |
2888 | ||
2889 | static struct dpif_flow_dump_thread * | |
2890 | dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_) | |
2891 | { | |
2892 | struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_); | |
2893 | struct dpif_netdev_flow_dump_thread *thread; | |
2894 | ||
2895 | thread = xmalloc(sizeof *thread); | |
2896 | dpif_flow_dump_thread_init(&thread->up, &dump->up); | |
2897 | thread->dump = dump; | |
2898 | return &thread->up; | |
2899 | } | |
2900 | ||
2901 | static void | |
2902 | dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_) | |
2903 | { | |
2904 | struct dpif_netdev_flow_dump_thread *thread | |
2905 | = dpif_netdev_flow_dump_thread_cast(thread_); | |
2906 | ||
2907 | free(thread); | |
2908 | } | |
2909 | ||
704a1e09 | 2910 | static int |
ac64794a | 2911 | dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_, |
8bb113da | 2912 | struct dpif_flow *flows, int max_flows) |
ac64794a BP |
2913 | { |
2914 | struct dpif_netdev_flow_dump_thread *thread | |
2915 | = dpif_netdev_flow_dump_thread_cast(thread_); | |
2916 | struct dpif_netdev_flow_dump *dump = thread->dump; | |
8bb113da | 2917 | struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH]; |
8bb113da RW |
2918 | int n_flows = 0; |
2919 | int i; | |
14608a15 | 2920 | |
ac64794a | 2921 | ovs_mutex_lock(&dump->mutex); |
8bb113da | 2922 | if (!dump->status) { |
1c1e46ed AW |
2923 | struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif); |
2924 | struct dp_netdev *dp = get_dp_netdev(&dpif->dpif); | |
2925 | struct dp_netdev_pmd_thread *pmd = dump->cur_pmd; | |
2926 | int flow_limit = MIN(max_flows, FLOW_DUMP_MAX_BATCH); | |
2927 | ||
2928 | /* First call to dump_next(), extracts the first pmd thread. | |
2929 | * If there is no pmd thread, returns immediately. */ | |
2930 | if (!pmd) { | |
2931 | pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos); | |
2932 | if (!pmd) { | |
2933 | ovs_mutex_unlock(&dump->mutex); | |
2934 | return n_flows; | |
8bb113da | 2935 | |
8bb113da | 2936 | } |
d2ad7ef1 | 2937 | } |
1c1e46ed AW |
2938 | |
2939 | do { | |
2940 | for (n_flows = 0; n_flows < flow_limit; n_flows++) { | |
2941 | struct cmap_node *node; | |
2942 | ||
2943 | node = cmap_next_position(&pmd->flow_table, &dump->flow_pos); | |
2944 | if (!node) { | |
2945 | break; | |
2946 | } | |
2947 | netdev_flows[n_flows] = CONTAINER_OF(node, | |
2948 | struct dp_netdev_flow, | |
2949 | node); | |
2950 | } | |
2951 | /* When finishing dumping the current pmd thread, moves to | |
2952 | * the next. */ | |
2953 | if (n_flows < flow_limit) { | |
2954 | memset(&dump->flow_pos, 0, sizeof dump->flow_pos); | |
2955 | dp_netdev_pmd_unref(pmd); | |
2956 | pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos); | |
2957 | if (!pmd) { | |
2958 | dump->status = EOF; | |
2959 | break; | |
2960 | } | |
2961 | } | |
2962 | /* Keeps the reference to next caller. */ | |
2963 | dump->cur_pmd = pmd; | |
2964 | ||
2965 | /* If the current dump is empty, do not exit the loop, since the | |
2966 | * remaining pmds could have flows to be dumped. Just dumps again | |
2967 | * on the new 'pmd'. */ | |
2968 | } while (!n_flows); | |
8a4e3a85 | 2969 | } |
ac64794a | 2970 | ovs_mutex_unlock(&dump->mutex); |
ac64794a | 2971 | |
8bb113da RW |
2972 | for (i = 0; i < n_flows; i++) { |
2973 | struct odputil_keybuf *maskbuf = &thread->maskbuf[i]; | |
2974 | struct odputil_keybuf *keybuf = &thread->keybuf[i]; | |
2975 | struct dp_netdev_flow *netdev_flow = netdev_flows[i]; | |
2976 | struct dpif_flow *f = &flows[i]; | |
7af12bd7 | 2977 | struct ofpbuf key, mask; |
8bb113da | 2978 | |
7af12bd7 JS |
2979 | ofpbuf_use_stack(&key, keybuf, sizeof *keybuf); |
2980 | ofpbuf_use_stack(&mask, maskbuf, sizeof *maskbuf); | |
64bb477f JS |
2981 | dp_netdev_flow_to_dpif_flow(netdev_flow, &key, &mask, f, |
2982 | dump->up.terse); | |
8bb113da | 2983 | } |
feebdea2 | 2984 | |
8bb113da | 2985 | return n_flows; |
72865317 BP |
2986 | } |
2987 | ||
2988 | static int | |
758c456d | 2989 | dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute) |
65f13b50 | 2990 | OVS_NO_THREAD_SAFETY_ANALYSIS |
72865317 BP |
2991 | { |
2992 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
65f13b50 | 2993 | struct dp_netdev_pmd_thread *pmd; |
1895cc8d | 2994 | struct dp_packet_batch pp; |
72865317 | 2995 | |
cf62fa4c PS |
2996 | if (dp_packet_size(execute->packet) < ETH_HEADER_LEN || |
2997 | dp_packet_size(execute->packet) > UINT16_MAX) { | |
72865317 BP |
2998 | return EINVAL; |
2999 | } | |
3000 | ||
65f13b50 AW |
3001 | /* Tries finding the 'pmd'. If NULL is returned, that means |
3002 | * the current thread is a non-pmd thread and should use | |
b19befae | 3003 | * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */ |
65f13b50 AW |
3004 | pmd = ovsthread_getspecific(dp->per_pmd_key); |
3005 | if (!pmd) { | |
b19befae | 3006 | pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID); |
546e57d4 DDP |
3007 | if (!pmd) { |
3008 | return EBUSY; | |
3009 | } | |
65f13b50 AW |
3010 | } |
3011 | ||
05267613 AZ |
3012 | if (execute->probe) { |
3013 | /* If this is part of a probe, Drop the packet, since executing | |
3014 | * the action may actually cause spurious packets be sent into | |
3015 | * the network. */ | |
d1ce9c20 YS |
3016 | if (pmd->core_id == NON_PMD_CORE_ID) { |
3017 | dp_netdev_pmd_unref(pmd); | |
3018 | } | |
05267613 AZ |
3019 | return 0; |
3020 | } | |
3021 | ||
65f13b50 AW |
3022 | /* If the current thread is non-pmd thread, acquires |
3023 | * the 'non_pmd_mutex'. */ | |
3024 | if (pmd->core_id == NON_PMD_CORE_ID) { | |
3025 | ovs_mutex_lock(&dp->non_pmd_mutex); | |
3026 | } | |
1c1e46ed | 3027 | |
b010be17 IM |
3028 | /* Update current time in PMD context. */ |
3029 | pmd_thread_ctx_time_update(pmd); | |
3030 | ||
36d8de17 DDP |
3031 | /* The action processing expects the RSS hash to be valid, because |
3032 | * it's always initialized at the beginning of datapath processing. | |
3033 | * In this case, though, 'execute->packet' may not have gone through | |
3034 | * the datapath at all, it may have been generated by the upper layer | |
3035 | * (OpenFlow packet-out, BFD frame, ...). */ | |
3036 | if (!dp_packet_rss_valid(execute->packet)) { | |
3037 | dp_packet_set_rss_hash(execute->packet, | |
3038 | flow_hash_5tuple(execute->flow, 0)); | |
3039 | } | |
3040 | ||
72c84bc2 | 3041 | dp_packet_batch_init_packet(&pp, execute->packet); |
66e4ad8a | 3042 | dp_netdev_execute_actions(pmd, &pp, false, execute->flow, |
b010be17 | 3043 | execute->actions, execute->actions_len); |
c71ea3c4 | 3044 | dp_netdev_pmd_flush_output_packets(pmd, true); |
36d8de17 | 3045 | |
65f13b50 AW |
3046 | if (pmd->core_id == NON_PMD_CORE_ID) { |
3047 | ovs_mutex_unlock(&dp->non_pmd_mutex); | |
e9985d6a | 3048 | dp_netdev_pmd_unref(pmd); |
65f13b50 | 3049 | } |
8a4e3a85 | 3050 | |
758c456d | 3051 | return 0; |
72865317 BP |
3052 | } |
3053 | ||
1a0c894a BP |
3054 | static void |
3055 | dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops) | |
3056 | { | |
3057 | size_t i; | |
3058 | ||
3059 | for (i = 0; i < n_ops; i++) { | |
3060 | struct dpif_op *op = ops[i]; | |
3061 | ||
3062 | switch (op->type) { | |
3063 | case DPIF_OP_FLOW_PUT: | |
3064 | op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put); | |
3065 | break; | |
3066 | ||
3067 | case DPIF_OP_FLOW_DEL: | |
3068 | op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del); | |
3069 | break; | |
3070 | ||
3071 | case DPIF_OP_EXECUTE: | |
3072 | op->error = dpif_netdev_execute(dpif, &op->u.execute); | |
3073 | break; | |
6fe09f8c JS |
3074 | |
3075 | case DPIF_OP_FLOW_GET: | |
3076 | op->error = dpif_netdev_flow_get(dpif, &op->u.flow_get); | |
3077 | break; | |
1a0c894a BP |
3078 | } |
3079 | } | |
3080 | } | |
3081 | ||
d4f6865c DDP |
3082 | /* Applies datapath configuration from the database. Some of the changes are |
3083 | * actually applied in dpif_netdev_run(). */ | |
f2eee189 | 3084 | static int |
d4f6865c | 3085 | dpif_netdev_set_config(struct dpif *dpif, const struct smap *other_config) |
f2eee189 AW |
3086 | { |
3087 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
d4f6865c | 3088 | const char *cmask = smap_get(other_config, "pmd-cpu-mask"); |
4c30b246 CL |
3089 | unsigned long long insert_prob = |
3090 | smap_get_ullong(other_config, "emc-insert-inv-prob", | |
3091 | DEFAULT_EM_FLOW_INSERT_INV_PROB); | |
3092 | uint32_t insert_min, cur_min; | |
c71ea3c4 IM |
3093 | uint32_t tx_flush_interval, cur_tx_flush_interval; |
3094 | ||
3095 | tx_flush_interval = smap_get_int(other_config, "tx-flush-interval", | |
3096 | DEFAULT_TX_FLUSH_INTERVAL); | |
3097 | atomic_read_relaxed(&dp->tx_flush_interval, &cur_tx_flush_interval); | |
3098 | if (tx_flush_interval != cur_tx_flush_interval) { | |
3099 | atomic_store_relaxed(&dp->tx_flush_interval, tx_flush_interval); | |
3100 | VLOG_INFO("Flushing interval for tx queues set to %"PRIu32" us", | |
3101 | tx_flush_interval); | |
3102 | } | |
f2eee189 | 3103 | |
a6a426d6 IM |
3104 | if (!nullable_string_is_equal(dp->pmd_cmask, cmask)) { |
3105 | free(dp->pmd_cmask); | |
3106 | dp->pmd_cmask = nullable_xstrdup(cmask); | |
3107 | dp_netdev_request_reconfigure(dp); | |
f2eee189 AW |
3108 | } |
3109 | ||
4c30b246 CL |
3110 | atomic_read_relaxed(&dp->emc_insert_min, &cur_min); |
3111 | if (insert_prob <= UINT32_MAX) { | |
3112 | insert_min = insert_prob == 0 ? 0 : UINT32_MAX / insert_prob; | |
3113 | } else { | |
3114 | insert_min = DEFAULT_EM_FLOW_INSERT_MIN; | |
3115 | insert_prob = DEFAULT_EM_FLOW_INSERT_INV_PROB; | |
3116 | } | |
3117 | ||
3118 | if (insert_min != cur_min) { | |
3119 | atomic_store_relaxed(&dp->emc_insert_min, insert_min); | |
3120 | if (insert_min == 0) { | |
3121 | VLOG_INFO("EMC has been disabled"); | |
3122 | } else { | |
3123 | VLOG_INFO("EMC insertion probability changed to 1/%llu (~%.2f%%)", | |
3124 | insert_prob, (100 / (float)insert_prob)); | |
3125 | } | |
3126 | } | |
3127 | ||
79f36875 JS |
3128 | bool perf_enabled = smap_get_bool(other_config, "pmd-perf-metrics", false); |
3129 | bool cur_perf_enabled; | |
3130 | atomic_read_relaxed(&dp->pmd_perf_metrics, &cur_perf_enabled); | |
3131 | if (perf_enabled != cur_perf_enabled) { | |
3132 | atomic_store_relaxed(&dp->pmd_perf_metrics, perf_enabled); | |
3133 | if (perf_enabled) { | |
3134 | VLOG_INFO("PMD performance metrics collection enabled"); | |
3135 | } else { | |
3136 | VLOG_INFO("PMD performance metrics collection disabled"); | |
3137 | } | |
3138 | } | |
3139 | ||
f2eee189 AW |
3140 | return 0; |
3141 | } | |
3142 | ||
3eb67853 IM |
3143 | /* Parses affinity list and returns result in 'core_ids'. */ |
3144 | static int | |
3145 | parse_affinity_list(const char *affinity_list, unsigned *core_ids, int n_rxq) | |
3146 | { | |
3147 | unsigned i; | |
3148 | char *list, *copy, *key, *value; | |
3149 | int error = 0; | |
3150 | ||
3151 | for (i = 0; i < n_rxq; i++) { | |
51c37a56 | 3152 | core_ids[i] = OVS_CORE_UNSPEC; |
3eb67853 IM |
3153 | } |
3154 | ||
3155 | if (!affinity_list) { | |
3156 | return 0; | |
3157 | } | |
3158 | ||
3159 | list = copy = xstrdup(affinity_list); | |
3160 | ||
3161 | while (ofputil_parse_key_value(&list, &key, &value)) { | |
3162 | int rxq_id, core_id; | |
3163 | ||
3164 | if (!str_to_int(key, 0, &rxq_id) || rxq_id < 0 | |
3165 | || !str_to_int(value, 0, &core_id) || core_id < 0) { | |
3166 | error = EINVAL; | |
3167 | break; | |
3168 | } | |
3169 | ||
3170 | if (rxq_id < n_rxq) { | |
3171 | core_ids[rxq_id] = core_id; | |
3172 | } | |
3173 | } | |
3174 | ||
3175 | free(copy); | |
3176 | return error; | |
3177 | } | |
3178 | ||
3179 | /* Parses 'affinity_list' and applies configuration if it is valid. */ | |
3180 | static int | |
3181 | dpif_netdev_port_set_rxq_affinity(struct dp_netdev_port *port, | |
3182 | const char *affinity_list) | |
3183 | { | |
3184 | unsigned *core_ids, i; | |
3185 | int error = 0; | |
3186 | ||
3187 | core_ids = xmalloc(port->n_rxq * sizeof *core_ids); | |
3188 | if (parse_affinity_list(affinity_list, core_ids, port->n_rxq)) { | |
3189 | error = EINVAL; | |
3190 | goto exit; | |
3191 | } | |
3192 | ||
3193 | for (i = 0; i < port->n_rxq; i++) { | |
3194 | port->rxqs[i].core_id = core_ids[i]; | |
3195 | } | |
3196 | ||
3197 | exit: | |
3198 | free(core_ids); | |
3199 | return error; | |
3200 | } | |
3201 | ||
3202 | /* Changes the affinity of port's rx queues. The changes are actually applied | |
3203 | * in dpif_netdev_run(). */ | |
3204 | static int | |
3205 | dpif_netdev_port_set_config(struct dpif *dpif, odp_port_t port_no, | |
3206 | const struct smap *cfg) | |
3207 | { | |
3208 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
3209 | struct dp_netdev_port *port; | |
3210 | int error = 0; | |
3211 | const char *affinity_list = smap_get(cfg, "pmd-rxq-affinity"); | |
3212 | ||
3213 | ovs_mutex_lock(&dp->port_mutex); | |
3214 | error = get_port_by_number(dp, port_no, &port); | |
3215 | if (error || !netdev_is_pmd(port->netdev) | |
3216 | || nullable_string_is_equal(affinity_list, port->rxq_affinity_list)) { | |
3217 | goto unlock; | |
3218 | } | |
3219 | ||
3220 | error = dpif_netdev_port_set_rxq_affinity(port, affinity_list); | |
3221 | if (error) { | |
3222 | goto unlock; | |
3223 | } | |
3224 | free(port->rxq_affinity_list); | |
3225 | port->rxq_affinity_list = nullable_xstrdup(affinity_list); | |
3226 | ||
3227 | dp_netdev_request_reconfigure(dp); | |
3228 | unlock: | |
3229 | ovs_mutex_unlock(&dp->port_mutex); | |
3230 | return error; | |
3231 | } | |
3232 | ||
5bf93d67 EJ |
3233 | static int |
3234 | dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED, | |
3235 | uint32_t queue_id, uint32_t *priority) | |
3236 | { | |
3237 | *priority = queue_id; | |
3238 | return 0; | |
3239 | } | |
3240 | ||
72865317 | 3241 | \f |
9ff55ae2 | 3242 | /* Creates and returns a new 'struct dp_netdev_actions', whose actions are |
1401f6de | 3243 | * a copy of the 'size' bytes of 'actions' input parameters. */ |
a84cb64a BP |
3244 | struct dp_netdev_actions * |
3245 | dp_netdev_actions_create(const struct nlattr *actions, size_t size) | |
3246 | { | |
3247 | struct dp_netdev_actions *netdev_actions; | |
3248 | ||
9ff55ae2 DDP |
3249 | netdev_actions = xmalloc(sizeof *netdev_actions + size); |
3250 | memcpy(netdev_actions->actions, actions, size); | |
a84cb64a BP |
3251 | netdev_actions->size = size; |
3252 | ||
3253 | return netdev_actions; | |
3254 | } | |
3255 | ||
a84cb64a | 3256 | struct dp_netdev_actions * |
61e7deb1 | 3257 | dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow) |
a84cb64a | 3258 | { |
61e7deb1 | 3259 | return ovsrcu_get(struct dp_netdev_actions *, &flow->actions); |
a84cb64a BP |
3260 | } |
3261 | ||
61e7deb1 BP |
3262 | static void |
3263 | dp_netdev_actions_free(struct dp_netdev_actions *actions) | |
a84cb64a | 3264 | { |
61e7deb1 | 3265 | free(actions); |
a84cb64a BP |
3266 | } |
3267 | \f | |
a19896ab JS |
3268 | static void |
3269 | dp_netdev_rxq_set_cycles(struct dp_netdev_rxq *rx, | |
3270 | enum rxq_cycles_counter_type type, | |
3271 | unsigned long long cycles) | |
a2ac666d | 3272 | { |
a19896ab | 3273 | atomic_store_relaxed(&rx->cycles[type], cycles); |
a2ac666d CL |
3274 | } |
3275 | ||
4809891b | 3276 | static void |
a19896ab | 3277 | dp_netdev_rxq_add_cycles(struct dp_netdev_rxq *rx, |
4809891b KT |
3278 | enum rxq_cycles_counter_type type, |
3279 | unsigned long long cycles) | |
3280 | { | |
a19896ab | 3281 | non_atomic_ullong_add(&rx->cycles[type], cycles); |
4809891b KT |
3282 | } |
3283 | ||
3284 | static uint64_t | |
3285 | dp_netdev_rxq_get_cycles(struct dp_netdev_rxq *rx, | |
3286 | enum rxq_cycles_counter_type type) | |
3287 | { | |
3288 | unsigned long long processing_cycles; | |
3289 | atomic_read_relaxed(&rx->cycles[type], &processing_cycles); | |
3290 | return processing_cycles; | |
3291 | } | |
3292 | ||
3293 | static void | |
3294 | dp_netdev_rxq_set_intrvl_cycles(struct dp_netdev_rxq *rx, | |
3295 | unsigned long long cycles) | |
3296 | { | |
4ee87ad3 BP |
3297 | unsigned int idx = rx->intrvl_idx++ % PMD_RXQ_INTERVAL_MAX; |
3298 | atomic_store_relaxed(&rx->cycles_intrvl[idx], cycles); | |
4809891b KT |
3299 | } |
3300 | ||
655856ef KT |
3301 | static uint64_t |
3302 | dp_netdev_rxq_get_intrvl_cycles(struct dp_netdev_rxq *rx, unsigned idx) | |
3303 | { | |
3304 | unsigned long long processing_cycles; | |
3305 | atomic_read_relaxed(&rx->cycles_intrvl[idx], &processing_cycles); | |
3306 | return processing_cycles; | |
3307 | } | |
3308 | ||
79f36875 JS |
3309 | #if ATOMIC_ALWAYS_LOCK_FREE_8B |
3310 | static inline bool | |
3311 | pmd_perf_metrics_enabled(const struct dp_netdev_pmd_thread *pmd) | |
3312 | { | |
3313 | bool pmd_perf_enabled; | |
3314 | atomic_read_relaxed(&pmd->dp->pmd_perf_metrics, &pmd_perf_enabled); | |
3315 | return pmd_perf_enabled; | |
3316 | } | |
3317 | #else | |
3318 | /* If stores and reads of 64-bit integers are not atomic, the full PMD | |
3319 | * performance metrics are not available as locked access to 64 bit | |
3320 | * integers would be prohibitively expensive. */ | |
3321 | static inline bool | |
3322 | pmd_perf_metrics_enabled(const struct dp_netdev_pmd_thread *pmd OVS_UNUSED) | |
3323 | { | |
3324 | return false; | |
3325 | } | |
3326 | #endif | |
3327 | ||
c71ea3c4 | 3328 | static int |
009e0033 IM |
3329 | dp_netdev_pmd_flush_output_on_port(struct dp_netdev_pmd_thread *pmd, |
3330 | struct tx_port *p) | |
3331 | { | |
58ed6df0 | 3332 | int i; |
009e0033 | 3333 | int tx_qid; |
cc4891f3 | 3334 | int output_cnt; |
009e0033 | 3335 | bool dynamic_txqs; |
58ed6df0 IM |
3336 | struct cycle_timer timer; |
3337 | uint64_t cycles; | |
c71ea3c4 | 3338 | uint32_t tx_flush_interval; |
58ed6df0 IM |
3339 | |
3340 | cycle_timer_start(&pmd->perf_stats, &timer); | |
009e0033 IM |
3341 | |
3342 | dynamic_txqs = p->port->dynamic_txqs; | |
3343 | if (dynamic_txqs) { | |
3344 | tx_qid = dpif_netdev_xps_get_tx_qid(pmd, p); | |
3345 | } else { | |
3346 | tx_qid = pmd->static_tx_qid; | |
3347 | } | |
3348 | ||
cc4891f3 | 3349 | output_cnt = dp_packet_batch_size(&p->output_pkts); |
58ed6df0 | 3350 | ovs_assert(output_cnt > 0); |
cc4891f3 | 3351 | |
b30896c9 | 3352 | netdev_send(p->port->netdev, tx_qid, &p->output_pkts, dynamic_txqs); |
009e0033 | 3353 | dp_packet_batch_init(&p->output_pkts); |
cc4891f3 | 3354 | |
c71ea3c4 IM |
3355 | /* Update time of the next flush. */ |
3356 | atomic_read_relaxed(&pmd->dp->tx_flush_interval, &tx_flush_interval); | |
3357 | p->flush_time = pmd->ctx.now + tx_flush_interval; | |
3358 | ||
3359 | ovs_assert(pmd->n_output_batches > 0); | |
3360 | pmd->n_output_batches--; | |
3361 | ||
82a48ead JS |
3362 | pmd_perf_update_counter(&pmd->perf_stats, PMD_STAT_SENT_PKTS, output_cnt); |
3363 | pmd_perf_update_counter(&pmd->perf_stats, PMD_STAT_SENT_BATCHES, 1); | |
58ed6df0 IM |
3364 | |
3365 | /* Distribute send cycles evenly among transmitted packets and assign to | |
3366 | * their respective rx queues. */ | |
3367 | cycles = cycle_timer_stop(&pmd->perf_stats, &timer) / output_cnt; | |
3368 | for (i = 0; i < output_cnt; i++) { | |
3369 | if (p->output_pkts_rxqs[i]) { | |
3370 | dp_netdev_rxq_add_cycles(p->output_pkts_rxqs[i], | |
3371 | RXQ_CYCLES_PROC_CURR, cycles); | |
3372 | } | |
3373 | } | |
c71ea3c4 IM |
3374 | |
3375 | return output_cnt; | |
009e0033 IM |
3376 | } |
3377 | ||
c71ea3c4 IM |
3378 | static int |
3379 | dp_netdev_pmd_flush_output_packets(struct dp_netdev_pmd_thread *pmd, | |
3380 | bool force) | |
009e0033 IM |
3381 | { |
3382 | struct tx_port *p; | |
c71ea3c4 IM |
3383 | int output_cnt = 0; |
3384 | ||
3385 | if (!pmd->n_output_batches) { | |
3386 | return 0; | |
3387 | } | |
009e0033 IM |
3388 | |
3389 | HMAP_FOR_EACH (p, node, &pmd->send_port_cache) { | |
c71ea3c4 IM |
3390 | if (!dp_packet_batch_is_empty(&p->output_pkts) |
3391 | && (force || pmd->ctx.now >= p->flush_time)) { | |
3392 | output_cnt += dp_netdev_pmd_flush_output_on_port(pmd, p); | |
009e0033 IM |
3393 | } |
3394 | } | |
c71ea3c4 | 3395 | return output_cnt; |
009e0033 IM |
3396 | } |
3397 | ||
a2ac666d | 3398 | static int |
65f13b50 | 3399 | dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd, |
a19896ab | 3400 | struct dp_netdev_rxq *rxq, |
947dc567 | 3401 | odp_port_t port_no) |
e4cfed38 | 3402 | { |
79f36875 | 3403 | struct pmd_perf_stats *s = &pmd->perf_stats; |
1895cc8d | 3404 | struct dp_packet_batch batch; |
a19896ab | 3405 | struct cycle_timer timer; |
1895cc8d | 3406 | int error; |
79f36875 JS |
3407 | int batch_cnt = 0; |
3408 | int rem_qlen = 0, *qlen_p = NULL; | |
58ed6df0 | 3409 | uint64_t cycles; |
e4cfed38 | 3410 | |
a19896ab JS |
3411 | /* Measure duration for polling and processing rx burst. */ |
3412 | cycle_timer_start(&pmd->perf_stats, &timer); | |
58ed6df0 IM |
3413 | |
3414 | pmd->ctx.last_rxq = rxq; | |
1895cc8d | 3415 | dp_packet_batch_init(&batch); |
58ed6df0 | 3416 | |
79f36875 JS |
3417 | /* Fetch the rx queue length only for vhostuser ports. */ |
3418 | if (pmd_perf_metrics_enabled(pmd) && rxq->is_vhost) { | |
3419 | qlen_p = &rem_qlen; | |
3420 | } | |
3421 | ||
3422 | error = netdev_rxq_recv(rxq->rx, &batch, qlen_p); | |
e4cfed38 | 3423 | if (!error) { |
a19896ab | 3424 | /* At least one packet received. */ |
3c33f0ff | 3425 | *recirc_depth_get() = 0; |
009e0033 | 3426 | pmd_thread_ctx_time_update(pmd); |
a2ac666d | 3427 | batch_cnt = batch.count; |
79f36875 JS |
3428 | if (pmd_perf_metrics_enabled(pmd)) { |
3429 | /* Update batch histogram. */ | |
3430 | s->current.batches++; | |
3431 | histogram_add_sample(&s->pkts_per_batch, batch_cnt); | |
3432 | /* Update the maximum vhost rx queue fill level. */ | |
3433 | if (rxq->is_vhost && rem_qlen >= 0) { | |
3434 | uint32_t qfill = batch_cnt + rem_qlen; | |
3435 | if (qfill > s->current.max_vhost_qfill) { | |
3436 | s->current.max_vhost_qfill = qfill; | |
3437 | } | |
3438 | } | |
3439 | } | |
3440 | /* Process packet batch. */ | |
947dc567 | 3441 | dp_netdev_input(pmd, &batch, port_no); |
e4cfed38 | 3442 | |
a19896ab | 3443 | /* Assign processing cycles to rx queue. */ |
58ed6df0 | 3444 | cycles = cycle_timer_stop(&pmd->perf_stats, &timer); |
a19896ab JS |
3445 | dp_netdev_rxq_add_cycles(rxq, RXQ_CYCLES_PROC_CURR, cycles); |
3446 | ||
79f36875 | 3447 | dp_netdev_pmd_flush_output_packets(pmd, false); |
a19896ab JS |
3448 | } else { |
3449 | /* Discard cycles. */ | |
3450 | cycle_timer_stop(&pmd->perf_stats, &timer); | |
3451 | if (error != EAGAIN && error != EOPNOTSUPP) { | |
3452 | static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5); | |
3453 | ||
3454 | VLOG_ERR_RL(&rl, "error receiving data from %s: %s", | |
3455 | netdev_rxq_get_name(rxq->rx), ovs_strerror(error)); | |
3456 | } | |
e4cfed38 | 3457 | } |
a2ac666d | 3458 | |
58ed6df0 IM |
3459 | pmd->ctx.last_rxq = NULL; |
3460 | ||
79f36875 | 3461 | return batch_cnt; |
e4cfed38 PS |
3462 | } |
3463 | ||
e32971b8 DDP |
3464 | static struct tx_port * |
3465 | tx_port_lookup(const struct hmap *hmap, odp_port_t port_no) | |
3466 | { | |
3467 | struct tx_port *tx; | |
3468 | ||
3469 | HMAP_FOR_EACH_IN_BUCKET (tx, node, hash_port_no(port_no), hmap) { | |
3470 | if (tx->port->port_no == port_no) { | |
3471 | return tx; | |
3472 | } | |
3473 | } | |
3474 | ||
3475 | return NULL; | |
3476 | } | |
3477 | ||
dc36593c DDP |
3478 | static int |
3479 | port_reconfigure(struct dp_netdev_port *port) | |
3480 | { | |
3481 | struct netdev *netdev = port->netdev; | |
dc36593c DDP |
3482 | int i, err; |
3483 | ||
dc36593c DDP |
3484 | /* Closes the existing 'rxq's. */ |
3485 | for (i = 0; i < port->n_rxq; i++) { | |
947dc567 DDP |
3486 | netdev_rxq_close(port->rxqs[i].rx); |
3487 | port->rxqs[i].rx = NULL; | |
dc36593c | 3488 | } |
4809891b | 3489 | unsigned last_nrxq = port->n_rxq; |
dc36593c DDP |
3490 | port->n_rxq = 0; |
3491 | ||
050c60bf | 3492 | /* Allows 'netdev' to apply the pending configuration changes. */ |
606f6650 | 3493 | if (netdev_is_reconf_required(netdev) || port->need_reconfigure) { |
e32971b8 DDP |
3494 | err = netdev_reconfigure(netdev); |
3495 | if (err && (err != EOPNOTSUPP)) { | |
3496 | VLOG_ERR("Failed to set interface %s new configuration", | |
3497 | netdev_get_name(netdev)); | |
3498 | return err; | |
3499 | } | |
dc36593c | 3500 | } |
050c60bf | 3501 | /* If the netdev_reconfigure() above succeeds, reopens the 'rxq's. */ |
3eb67853 IM |
3502 | port->rxqs = xrealloc(port->rxqs, |
3503 | sizeof *port->rxqs * netdev_n_rxq(netdev)); | |
324c8374 IM |
3504 | /* Realloc 'used' counters for tx queues. */ |
3505 | free(port->txq_used); | |
3506 | port->txq_used = xcalloc(netdev_n_txq(netdev), sizeof *port->txq_used); | |
3507 | ||
dc36593c | 3508 | for (i = 0; i < netdev_n_rxq(netdev); i++) { |
38259bd7 BP |
3509 | bool new_queue = i >= last_nrxq; |
3510 | if (new_queue) { | |
3511 | memset(&port->rxqs[i], 0, sizeof port->rxqs[i]); | |
3512 | } | |
3513 | ||
947dc567 | 3514 | port->rxqs[i].port = port; |
79f36875 | 3515 | port->rxqs[i].is_vhost = !strncmp(port->type, "dpdkvhost", 9); |
38259bd7 | 3516 | |
947dc567 | 3517 | err = netdev_rxq_open(netdev, &port->rxqs[i].rx, i); |
dc36593c DDP |
3518 | if (err) { |
3519 | return err; | |
3520 | } | |
3521 | port->n_rxq++; | |
3522 | } | |
3523 | ||
3eb67853 IM |
3524 | /* Parse affinity list to apply configuration for new queues. */ |
3525 | dpif_netdev_port_set_rxq_affinity(port, port->rxq_affinity_list); | |
3526 | ||
606f6650 EC |
3527 | /* If reconfiguration was successful mark it as such, so we can use it */ |
3528 | port->need_reconfigure = false; | |
3529 | ||
dc36593c DDP |
3530 | return 0; |
3531 | } | |
3532 | ||
e32971b8 DDP |
3533 | struct rr_numa_list { |
3534 | struct hmap numas; /* Contains 'struct rr_numa' */ | |
3535 | }; | |
3536 | ||
3537 | struct rr_numa { | |
3538 | struct hmap_node node; | |
3539 | ||
3540 | int numa_id; | |
3541 | ||
3542 | /* Non isolated pmds on numa node 'numa_id' */ | |
3543 | struct dp_netdev_pmd_thread **pmds; | |
3544 | int n_pmds; | |
3545 | ||
3546 | int cur_index; | |
79da1e41 | 3547 | bool idx_inc; |
e32971b8 DDP |
3548 | }; |
3549 | ||
3550 | static struct rr_numa * | |
3551 | rr_numa_list_lookup(struct rr_numa_list *rr, int numa_id) | |
3552 | { | |
3553 | struct rr_numa *numa; | |
3554 | ||
3555 | HMAP_FOR_EACH_WITH_HASH (numa, node, hash_int(numa_id, 0), &rr->numas) { | |
3556 | if (numa->numa_id == numa_id) { | |
3557 | return numa; | |
3558 | } | |
3559 | } | |
3560 | ||
3561 | return NULL; | |
3562 | } | |
3563 | ||
c37813fd BM |
3564 | /* Returns the next node in numa list following 'numa' in round-robin fashion. |
3565 | * Returns first node if 'numa' is a null pointer or the last node in 'rr'. | |
3566 | * Returns NULL if 'rr' numa list is empty. */ | |
3567 | static struct rr_numa * | |
3568 | rr_numa_list_next(struct rr_numa_list *rr, const struct rr_numa *numa) | |
3569 | { | |
3570 | struct hmap_node *node = NULL; | |
3571 | ||
3572 | if (numa) { | |
3573 | node = hmap_next(&rr->numas, &numa->node); | |
3574 | } | |
3575 | if (!node) { | |
3576 | node = hmap_first(&rr->numas); | |
3577 | } | |
3578 | ||
3579 | return (node) ? CONTAINER_OF(node, struct rr_numa, node) : NULL; | |
3580 | } | |
3581 | ||
e32971b8 DDP |
3582 | static void |
3583 | rr_numa_list_populate(struct dp_netdev *dp, struct rr_numa_list *rr) | |
3584 | { | |
3585 | struct dp_netdev_pmd_thread *pmd; | |
3586 | struct rr_numa *numa; | |
3587 | ||
3588 | hmap_init(&rr->numas); | |
3589 | ||
3590 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
3591 | if (pmd->core_id == NON_PMD_CORE_ID || pmd->isolated) { | |
3592 | continue; | |
3593 | } | |
3594 | ||
3595 | numa = rr_numa_list_lookup(rr, pmd->numa_id); | |
3596 | if (!numa) { | |
3597 | numa = xzalloc(sizeof *numa); | |
3598 | numa->numa_id = pmd->numa_id; | |
3599 | hmap_insert(&rr->numas, &numa->node, hash_int(pmd->numa_id, 0)); | |
3600 | } | |
3601 | numa->n_pmds++; | |
3602 | numa->pmds = xrealloc(numa->pmds, numa->n_pmds * sizeof *numa->pmds); | |
3603 | numa->pmds[numa->n_pmds - 1] = pmd; | |
79da1e41 KT |
3604 | /* At least one pmd so initialise curr_idx and idx_inc. */ |
3605 | numa->cur_index = 0; | |
3606 | numa->idx_inc = true; | |
e32971b8 DDP |
3607 | } |
3608 | } | |
3609 | ||
79da1e41 KT |
3610 | /* Returns the next pmd from the numa node in |
3611 | * incrementing or decrementing order. */ | |
e32971b8 DDP |
3612 | static struct dp_netdev_pmd_thread * |
3613 | rr_numa_get_pmd(struct rr_numa *numa) | |
3614 | { | |
79da1e41 KT |
3615 | int numa_idx = numa->cur_index; |
3616 | ||
3617 | if (numa->idx_inc == true) { | |
3618 | /* Incrementing through list of pmds. */ | |
3619 | if (numa->cur_index == numa->n_pmds-1) { | |
3620 | /* Reached the last pmd. */ | |
3621 | numa->idx_inc = false; | |
3622 | } else { | |
3623 | numa->cur_index++; | |
3624 | } | |
3625 | } else { | |
3626 | /* Decrementing through list of pmds. */ | |
3627 | if (numa->cur_index == 0) { | |
3628 | /* Reached the first pmd. */ | |
3629 | numa->idx_inc = true; | |
3630 | } else { | |
3631 | numa->cur_index--; | |
3632 | } | |
3633 | } | |
3634 | return numa->pmds[numa_idx]; | |
e32971b8 DDP |
3635 | } |
3636 | ||
3637 | static void | |
3638 | rr_numa_list_destroy(struct rr_numa_list *rr) | |
3639 | { | |
3640 | struct rr_numa *numa; | |
3641 | ||
3642 | HMAP_FOR_EACH_POP (numa, node, &rr->numas) { | |
3643 | free(numa->pmds); | |
3644 | free(numa); | |
3645 | } | |
3646 | hmap_destroy(&rr->numas); | |
3647 | } | |
3648 | ||
655856ef KT |
3649 | /* Sort Rx Queues by the processing cycles they are consuming. */ |
3650 | static int | |
cc131ac1 | 3651 | compare_rxq_cycles(const void *a, const void *b) |
655856ef | 3652 | { |
28080276 KT |
3653 | struct dp_netdev_rxq *qa; |
3654 | struct dp_netdev_rxq *qb; | |
8368866e | 3655 | uint64_t cycles_qa, cycles_qb; |
655856ef KT |
3656 | |
3657 | qa = *(struct dp_netdev_rxq **) a; | |
3658 | qb = *(struct dp_netdev_rxq **) b; | |
3659 | ||
8368866e KT |
3660 | cycles_qa = dp_netdev_rxq_get_cycles(qa, RXQ_CYCLES_PROC_HIST); |
3661 | cycles_qb = dp_netdev_rxq_get_cycles(qb, RXQ_CYCLES_PROC_HIST); | |
655856ef | 3662 | |
8368866e KT |
3663 | if (cycles_qa != cycles_qb) { |
3664 | return (cycles_qa < cycles_qb) ? 1 : -1; | |
a130f1a8 KT |
3665 | } else { |
3666 | /* Cycles are the same so tiebreak on port/queue id. | |
3667 | * Tiebreaking (as opposed to return 0) ensures consistent | |
3668 | * sort results across multiple OS's. */ | |
f0aa3801 BP |
3669 | uint32_t port_qa = odp_to_u32(qa->port->port_no); |
3670 | uint32_t port_qb = odp_to_u32(qb->port->port_no); | |
3671 | if (port_qa != port_qb) { | |
3672 | return port_qa > port_qb ? 1 : -1; | |
a130f1a8 KT |
3673 | } else { |
3674 | return netdev_rxq_get_queue_id(qa->rx) | |
3675 | - netdev_rxq_get_queue_id(qb->rx); | |
3676 | } | |
655856ef | 3677 | } |
655856ef KT |
3678 | } |
3679 | ||
e32971b8 DDP |
3680 | /* Assign pmds to queues. If 'pinned' is true, assign pmds to pinned |
3681 | * queues and marks the pmds as isolated. Otherwise, assign non isolated | |
3682 | * pmds to unpinned queues. | |
3683 | * | |
655856ef KT |
3684 | * If 'pinned' is false queues will be sorted by processing cycles they are |
3685 | * consuming and then assigned to pmds in round robin order. | |
3686 | * | |
e32971b8 DDP |
3687 | * The function doesn't touch the pmd threads, it just stores the assignment |
3688 | * in the 'pmd' member of each rxq. */ | |
3689 | static void | |
3690 | rxq_scheduling(struct dp_netdev *dp, bool pinned) OVS_REQUIRES(dp->port_mutex) | |
3691 | { | |
3692 | struct dp_netdev_port *port; | |
3693 | struct rr_numa_list rr; | |
c37813fd | 3694 | struct rr_numa *non_local_numa = NULL; |
655856ef | 3695 | struct dp_netdev_rxq ** rxqs = NULL; |
97bf8f47 | 3696 | int n_rxqs = 0; |
655856ef KT |
3697 | struct rr_numa *numa = NULL; |
3698 | int numa_id; | |
e32971b8 DDP |
3699 | |
3700 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
e32971b8 DDP |
3701 | if (!netdev_is_pmd(port->netdev)) { |
3702 | continue; | |
3703 | } | |
3704 | ||
e32971b8 DDP |
3705 | for (int qid = 0; qid < port->n_rxq; qid++) { |
3706 | struct dp_netdev_rxq *q = &port->rxqs[qid]; | |
3707 | ||
3708 | if (pinned && q->core_id != OVS_CORE_UNSPEC) { | |
3709 | struct dp_netdev_pmd_thread *pmd; | |
3710 | ||
3711 | pmd = dp_netdev_get_pmd(dp, q->core_id); | |
3712 | if (!pmd) { | |
3713 | VLOG_WARN("There is no PMD thread on core %d. Queue " | |
3714 | "%d on port \'%s\' will not be polled.", | |
3715 | q->core_id, qid, netdev_get_name(port->netdev)); | |
3716 | } else { | |
3717 | q->pmd = pmd; | |
3718 | pmd->isolated = true; | |
3719 | dp_netdev_pmd_unref(pmd); | |
3720 | } | |
3721 | } else if (!pinned && q->core_id == OVS_CORE_UNSPEC) { | |
8368866e KT |
3722 | uint64_t cycle_hist = 0; |
3723 | ||
655856ef KT |
3724 | if (n_rxqs == 0) { |
3725 | rxqs = xmalloc(sizeof *rxqs); | |
e32971b8 | 3726 | } else { |
655856ef | 3727 | rxqs = xrealloc(rxqs, sizeof *rxqs * (n_rxqs + 1)); |
e32971b8 | 3728 | } |
8368866e KT |
3729 | /* Sum the queue intervals and store the cycle history. */ |
3730 | for (unsigned i = 0; i < PMD_RXQ_INTERVAL_MAX; i++) { | |
3731 | cycle_hist += dp_netdev_rxq_get_intrvl_cycles(q, i); | |
3732 | } | |
3733 | dp_netdev_rxq_set_cycles(q, RXQ_CYCLES_PROC_HIST, cycle_hist); | |
3734 | ||
655856ef KT |
3735 | /* Store the queue. */ |
3736 | rxqs[n_rxqs++] = q; | |
e32971b8 DDP |
3737 | } |
3738 | } | |
3739 | } | |
3740 | ||
655856ef KT |
3741 | if (n_rxqs > 1) { |
3742 | /* Sort the queues in order of the processing cycles | |
3743 | * they consumed during their last pmd interval. */ | |
cc131ac1 | 3744 | qsort(rxqs, n_rxqs, sizeof *rxqs, compare_rxq_cycles); |
655856ef KT |
3745 | } |
3746 | ||
3747 | rr_numa_list_populate(dp, &rr); | |
3748 | /* Assign the sorted queues to pmds in round robin. */ | |
97bf8f47 | 3749 | for (int i = 0; i < n_rxqs; i++) { |
655856ef KT |
3750 | numa_id = netdev_get_numa_id(rxqs[i]->port->netdev); |
3751 | numa = rr_numa_list_lookup(&rr, numa_id); | |
3752 | if (!numa) { | |
3753 | /* There are no pmds on the queue's local NUMA node. | |
3754 | Round robin on the NUMA nodes that do have pmds. */ | |
3755 | non_local_numa = rr_numa_list_next(&rr, non_local_numa); | |
3756 | if (!non_local_numa) { | |
3757 | VLOG_ERR("There is no available (non-isolated) pmd " | |
3758 | "thread for port \'%s\' queue %d. This queue " | |
3759 | "will not be polled. Is pmd-cpu-mask set to " | |
3760 | "zero? Or are all PMDs isolated to other " | |
3761 | "queues?", netdev_rxq_get_name(rxqs[i]->rx), | |
3762 | netdev_rxq_get_queue_id(rxqs[i]->rx)); | |
3763 | continue; | |
3764 | } | |
3765 | rxqs[i]->pmd = rr_numa_get_pmd(non_local_numa); | |
3766 | VLOG_WARN("There's no available (non-isolated) pmd thread " | |
3767 | "on numa node %d. Queue %d on port \'%s\' will " | |
3768 | "be assigned to the pmd on core %d " | |
3769 | "(numa node %d). Expect reduced performance.", | |
3770 | numa_id, netdev_rxq_get_queue_id(rxqs[i]->rx), | |
3771 | netdev_rxq_get_name(rxqs[i]->rx), | |
3772 | rxqs[i]->pmd->core_id, rxqs[i]->pmd->numa_id); | |
3773 | } else { | |
3774 | rxqs[i]->pmd = rr_numa_get_pmd(numa); | |
3775 | VLOG_INFO("Core %d on numa node %d assigned port \'%s\' " | |
3776 | "rx queue %d (measured processing cycles %"PRIu64").", | |
3777 | rxqs[i]->pmd->core_id, numa_id, | |
3778 | netdev_rxq_get_name(rxqs[i]->rx), | |
3779 | netdev_rxq_get_queue_id(rxqs[i]->rx), | |
3780 | dp_netdev_rxq_get_cycles(rxqs[i], RXQ_CYCLES_PROC_HIST)); | |
3781 | } | |
3782 | } | |
3783 | ||
e32971b8 | 3784 | rr_numa_list_destroy(&rr); |
655856ef | 3785 | free(rxqs); |
e32971b8 DDP |
3786 | } |
3787 | ||
140dd699 IM |
3788 | static void |
3789 | reload_affected_pmds(struct dp_netdev *dp) | |
3790 | { | |
3791 | struct dp_netdev_pmd_thread *pmd; | |
3792 | ||
3793 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
3794 | if (pmd->need_reload) { | |
3795 | dp_netdev_reload_pmd__(pmd); | |
3796 | pmd->need_reload = false; | |
3797 | } | |
3798 | } | |
3799 | } | |
3800 | ||
6e3c6fa4 DDP |
3801 | static void |
3802 | reconfigure_pmd_threads(struct dp_netdev *dp) | |
3803 | OVS_REQUIRES(dp->port_mutex) | |
3804 | { | |
e32971b8 DDP |
3805 | struct dp_netdev_pmd_thread *pmd; |
3806 | struct ovs_numa_dump *pmd_cores; | |
140dd699 IM |
3807 | struct ovs_numa_info_core *core; |
3808 | struct hmapx to_delete = HMAPX_INITIALIZER(&to_delete); | |
3809 | struct hmapx_node *node; | |
e32971b8 | 3810 | bool changed = false; |
140dd699 | 3811 | bool need_to_adjust_static_tx_qids = false; |
e32971b8 DDP |
3812 | |
3813 | /* The pmd threads should be started only if there's a pmd port in the | |
3814 | * datapath. If the user didn't provide any "pmd-cpu-mask", we start | |
3815 | * NR_PMD_THREADS per numa node. */ | |
3816 | if (!has_pmd_port(dp)) { | |
3817 | pmd_cores = ovs_numa_dump_n_cores_per_numa(0); | |
3818 | } else if (dp->pmd_cmask && dp->pmd_cmask[0]) { | |
3819 | pmd_cores = ovs_numa_dump_cores_with_cmask(dp->pmd_cmask); | |
3820 | } else { | |
3821 | pmd_cores = ovs_numa_dump_n_cores_per_numa(NR_PMD_THREADS); | |
3822 | } | |
3823 | ||
140dd699 IM |
3824 | /* We need to adjust 'static_tx_qid's only if we're reducing number of |
3825 | * PMD threads. Otherwise, new threads will allocate all the freed ids. */ | |
3826 | if (ovs_numa_dump_count(pmd_cores) < cmap_count(&dp->poll_threads) - 1) { | |
3827 | /* Adjustment is required to keep 'static_tx_qid's sequential and | |
3828 | * avoid possible issues, for example, imbalanced tx queue usage | |
3829 | * and unnecessary locking caused by remapping on netdev level. */ | |
3830 | need_to_adjust_static_tx_qids = true; | |
3831 | } | |
3832 | ||
3833 | /* Check for unwanted pmd threads */ | |
3834 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
3835 | if (pmd->core_id == NON_PMD_CORE_ID) { | |
3836 | continue; | |
3837 | } | |
3838 | if (!ovs_numa_dump_contains_core(pmd_cores, pmd->numa_id, | |
3839 | pmd->core_id)) { | |
3840 | hmapx_add(&to_delete, pmd); | |
3841 | } else if (need_to_adjust_static_tx_qids) { | |
3842 | pmd->need_reload = true; | |
e32971b8 DDP |
3843 | } |
3844 | } | |
3845 | ||
140dd699 IM |
3846 | HMAPX_FOR_EACH (node, &to_delete) { |
3847 | pmd = (struct dp_netdev_pmd_thread *) node->data; | |
3848 | VLOG_INFO("PMD thread on numa_id: %d, core id: %2d destroyed.", | |
3849 | pmd->numa_id, pmd->core_id); | |
3850 | dp_netdev_del_pmd(dp, pmd); | |
3851 | } | |
3852 | changed = !hmapx_is_empty(&to_delete); | |
3853 | hmapx_destroy(&to_delete); | |
e32971b8 | 3854 | |
140dd699 IM |
3855 | if (need_to_adjust_static_tx_qids) { |
3856 | /* 'static_tx_qid's are not sequential now. | |
3857 | * Reload remaining threads to fix this. */ | |
3858 | reload_affected_pmds(dp); | |
3859 | } | |
e32971b8 | 3860 | |
140dd699 IM |
3861 | /* Check for required new pmd threads */ |
3862 | FOR_EACH_CORE_ON_DUMP(core, pmd_cores) { | |
3863 | pmd = dp_netdev_get_pmd(dp, core->core_id); | |
3864 | if (!pmd) { | |
3865 | pmd = xzalloc(sizeof *pmd); | |
e32971b8 | 3866 | dp_netdev_configure_pmd(pmd, dp, core->core_id, core->numa_id); |
e32971b8 | 3867 | pmd->thread = ovs_thread_create("pmd", pmd_thread_main, pmd); |
140dd699 IM |
3868 | VLOG_INFO("PMD thread on numa_id: %d, core id: %2d created.", |
3869 | pmd->numa_id, pmd->core_id); | |
3870 | changed = true; | |
3871 | } else { | |
3872 | dp_netdev_pmd_unref(pmd); | |
e32971b8 | 3873 | } |
140dd699 IM |
3874 | } |
3875 | ||
3876 | if (changed) { | |
3877 | struct ovs_numa_info_numa *numa; | |
e32971b8 DDP |
3878 | |
3879 | /* Log the number of pmd threads per numa node. */ | |
3880 | FOR_EACH_NUMA_ON_DUMP (numa, pmd_cores) { | |
140dd699 | 3881 | VLOG_INFO("There are %"PRIuSIZE" pmd threads on numa node %d", |
e32971b8 DDP |
3882 | numa->n_cores, numa->numa_id); |
3883 | } | |
3884 | } | |
3885 | ||
3886 | ovs_numa_dump_destroy(pmd_cores); | |
3887 | } | |
3888 | ||
e32971b8 DDP |
3889 | static void |
3890 | pmd_remove_stale_ports(struct dp_netdev *dp, | |
3891 | struct dp_netdev_pmd_thread *pmd) | |
3892 | OVS_EXCLUDED(pmd->port_mutex) | |
3893 | OVS_REQUIRES(dp->port_mutex) | |
3894 | { | |
3895 | struct rxq_poll *poll, *poll_next; | |
3896 | struct tx_port *tx, *tx_next; | |
3897 | ||
3898 | ovs_mutex_lock(&pmd->port_mutex); | |
3899 | HMAP_FOR_EACH_SAFE (poll, poll_next, node, &pmd->poll_list) { | |
3900 | struct dp_netdev_port *port = poll->rxq->port; | |
3901 | ||
3902 | if (port->need_reconfigure | |
3903 | || !hmap_contains(&dp->ports, &port->node)) { | |
3904 | dp_netdev_del_rxq_from_pmd(pmd, poll); | |
3905 | } | |
3906 | } | |
3907 | HMAP_FOR_EACH_SAFE (tx, tx_next, node, &pmd->tx_ports) { | |
3908 | struct dp_netdev_port *port = tx->port; | |
3909 | ||
3910 | if (port->need_reconfigure | |
3911 | || !hmap_contains(&dp->ports, &port->node)) { | |
3912 | dp_netdev_del_port_tx_from_pmd(pmd, tx); | |
3913 | } | |
3914 | } | |
3915 | ovs_mutex_unlock(&pmd->port_mutex); | |
3916 | } | |
3917 | ||
3918 | /* Must be called each time a port is added/removed or the cmask changes. | |
3919 | * This creates and destroys pmd threads, reconfigures ports, opens their | |
3920 | * rxqs and assigns all rxqs/txqs to pmd threads. */ | |
3921 | static void | |
3922 | reconfigure_datapath(struct dp_netdev *dp) | |
3923 | OVS_REQUIRES(dp->port_mutex) | |
3924 | { | |
3925 | struct dp_netdev_pmd_thread *pmd; | |
3926 | struct dp_netdev_port *port; | |
3927 | int wanted_txqs; | |
6e3c6fa4 | 3928 | |
a6a426d6 IM |
3929 | dp->last_reconfigure_seq = seq_read(dp->reconfigure_seq); |
3930 | ||
e32971b8 DDP |
3931 | /* Step 1: Adjust the pmd threads based on the datapath ports, the cores |
3932 | * on the system and the user configuration. */ | |
3933 | reconfigure_pmd_threads(dp); | |
6e3c6fa4 | 3934 | |
e32971b8 | 3935 | wanted_txqs = cmap_count(&dp->poll_threads); |
324c8374 | 3936 | |
e32971b8 DDP |
3937 | /* The number of pmd threads might have changed, or a port can be new: |
3938 | * adjust the txqs. */ | |
3939 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
3940 | netdev_set_tx_multiq(port->netdev, wanted_txqs); | |
324c8374 IM |
3941 | } |
3942 | ||
e32971b8 DDP |
3943 | /* Step 2: Remove from the pmd threads ports that have been removed or |
3944 | * need reconfiguration. */ | |
3945 | ||
3946 | /* Check for all the ports that need reconfiguration. We cache this in | |
85a4f238 IM |
3947 | * 'port->need_reconfigure', because netdev_is_reconf_required() can |
3948 | * change at any time. */ | |
e32971b8 DDP |
3949 | HMAP_FOR_EACH (port, node, &dp->ports) { |
3950 | if (netdev_is_reconf_required(port->netdev)) { | |
3951 | port->need_reconfigure = true; | |
3952 | } | |
3953 | } | |
3954 | ||
3955 | /* Remove from the pmd threads all the ports that have been deleted or | |
3956 | * need reconfiguration. */ | |
3957 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
3958 | pmd_remove_stale_ports(dp, pmd); | |
3959 | } | |
3960 | ||
3961 | /* Reload affected pmd threads. We must wait for the pmd threads before | |
3962 | * reconfiguring the ports, because a port cannot be reconfigured while | |
3963 | * it's being used. */ | |
3964 | reload_affected_pmds(dp); | |
3965 | ||
3966 | /* Step 3: Reconfigure ports. */ | |
3967 | ||
3968 | /* We only reconfigure the ports that we determined above, because they're | |
3969 | * not being used by any pmd thread at the moment. If a port fails to | |
3970 | * reconfigure we remove it from the datapath. */ | |
f582b6df BP |
3971 | struct dp_netdev_port *next_port; |
3972 | HMAP_FOR_EACH_SAFE (port, next_port, node, &dp->ports) { | |
dc36593c | 3973 | int err; |
6e3c6fa4 | 3974 | |
e32971b8 DDP |
3975 | if (!port->need_reconfigure) { |
3976 | continue; | |
3977 | } | |
3978 | ||
dc36593c DDP |
3979 | err = port_reconfigure(port); |
3980 | if (err) { | |
3981 | hmap_remove(&dp->ports, &port->node); | |
3982 | seq_change(dp->port_seq); | |
3983 | port_destroy(port); | |
324c8374 | 3984 | } else { |
e32971b8 | 3985 | port->dynamic_txqs = netdev_n_txq(port->netdev) < wanted_txqs; |
6e3c6fa4 DDP |
3986 | } |
3987 | } | |
e32971b8 DDP |
3988 | |
3989 | /* Step 4: Compute new rxq scheduling. We don't touch the pmd threads | |
3990 | * for now, we just update the 'pmd' pointer in each rxq to point to the | |
3991 | * wanted thread according to the scheduling policy. */ | |
3992 | ||
3993 | /* Reset all the pmd threads to non isolated. */ | |
3994 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
3995 | pmd->isolated = false; | |
3996 | } | |
3997 | ||
3998 | /* Reset all the queues to unassigned */ | |
3999 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
4000 | for (int i = 0; i < port->n_rxq; i++) { | |
4001 | port->rxqs[i].pmd = NULL; | |
4002 | } | |
4003 | } | |
4004 | ||
4005 | /* Add pinned queues and mark pmd threads isolated. */ | |
4006 | rxq_scheduling(dp, true); | |
4007 | ||
4008 | /* Add non-pinned queues. */ | |
4009 | rxq_scheduling(dp, false); | |
4010 | ||
4011 | /* Step 5: Remove queues not compliant with new scheduling. */ | |
4012 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
4013 | struct rxq_poll *poll, *poll_next; | |
4014 | ||
4015 | ovs_mutex_lock(&pmd->port_mutex); | |
4016 | HMAP_FOR_EACH_SAFE (poll, poll_next, node, &pmd->poll_list) { | |
4017 | if (poll->rxq->pmd != pmd) { | |
4018 | dp_netdev_del_rxq_from_pmd(pmd, poll); | |
4019 | } | |
4020 | } | |
4021 | ovs_mutex_unlock(&pmd->port_mutex); | |
4022 | } | |
4023 | ||
4024 | /* Reload affected pmd threads. We must wait for the pmd threads to remove | |
4025 | * the old queues before readding them, otherwise a queue can be polled by | |
4026 | * two threads at the same time. */ | |
4027 | reload_affected_pmds(dp); | |
4028 | ||
4029 | /* Step 6: Add queues from scheduling, if they're not there already. */ | |
4030 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
4031 | if (!netdev_is_pmd(port->netdev)) { | |
4032 | continue; | |
4033 | } | |
4034 | ||
4035 | for (int qid = 0; qid < port->n_rxq; qid++) { | |
4036 | struct dp_netdev_rxq *q = &port->rxqs[qid]; | |
4037 | ||
4038 | if (q->pmd) { | |
4039 | ovs_mutex_lock(&q->pmd->port_mutex); | |
4040 | dp_netdev_add_rxq_to_pmd(q->pmd, q); | |
4041 | ovs_mutex_unlock(&q->pmd->port_mutex); | |
4042 | } | |
4043 | } | |
4044 | } | |
4045 | ||
4046 | /* Add every port to the tx cache of every pmd thread, if it's not | |
4047 | * there already and if this pmd has at least one rxq to poll. */ | |
4048 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
4049 | ovs_mutex_lock(&pmd->port_mutex); | |
4050 | if (hmap_count(&pmd->poll_list) || pmd->core_id == NON_PMD_CORE_ID) { | |
4051 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
4052 | dp_netdev_add_port_tx_to_pmd(pmd, port); | |
4053 | } | |
4054 | } | |
4055 | ovs_mutex_unlock(&pmd->port_mutex); | |
4056 | } | |
4057 | ||
4058 | /* Reload affected pmd threads. */ | |
4059 | reload_affected_pmds(dp); | |
6e3c6fa4 DDP |
4060 | } |
4061 | ||
050c60bf DDP |
4062 | /* Returns true if one of the netdevs in 'dp' requires a reconfiguration */ |
4063 | static bool | |
4064 | ports_require_restart(const struct dp_netdev *dp) | |
4065 | OVS_REQUIRES(dp->port_mutex) | |
4066 | { | |
4067 | struct dp_netdev_port *port; | |
4068 | ||
4069 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
4070 | if (netdev_is_reconf_required(port->netdev)) { | |
4071 | return true; | |
4072 | } | |
4073 | } | |
4074 | ||
4075 | return false; | |
4076 | } | |
4077 | ||
a36de779 PS |
4078 | /* Return true if needs to revalidate datapath flows. */ |
4079 | static bool | |
e4cfed38 PS |
4080 | dpif_netdev_run(struct dpif *dpif) |
4081 | { | |
4082 | struct dp_netdev_port *port; | |
4083 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
546e57d4 | 4084 | struct dp_netdev_pmd_thread *non_pmd; |
a36de779 | 4085 | uint64_t new_tnl_seq; |
c71ea3c4 | 4086 | bool need_to_flush = true; |
e4cfed38 | 4087 | |
e9985d6a | 4088 | ovs_mutex_lock(&dp->port_mutex); |
546e57d4 DDP |
4089 | non_pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID); |
4090 | if (non_pmd) { | |
4091 | ovs_mutex_lock(&dp->non_pmd_mutex); | |
4092 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
4093 | if (!netdev_is_pmd(port->netdev)) { | |
4094 | int i; | |
55c955bd | 4095 | |
546e57d4 | 4096 | for (i = 0; i < port->n_rxq; i++) { |
c71ea3c4 IM |
4097 | if (dp_netdev_process_rxq_port(non_pmd, |
4098 | &port->rxqs[i], | |
4099 | port->port_no)) { | |
4100 | need_to_flush = false; | |
4101 | } | |
546e57d4 | 4102 | } |
55c955bd | 4103 | } |
e4cfed38 | 4104 | } |
c71ea3c4 IM |
4105 | if (need_to_flush) { |
4106 | /* We didn't receive anything in the process loop. | |
4107 | * Check if we need to send something. | |
4108 | * There was no time updates on current iteration. */ | |
4109 | pmd_thread_ctx_time_update(non_pmd); | |
4110 | dp_netdev_pmd_flush_output_packets(non_pmd, false); | |
4111 | } | |
4112 | ||
b010be17 | 4113 | dpif_netdev_xps_revalidate_pmd(non_pmd, false); |
546e57d4 | 4114 | ovs_mutex_unlock(&dp->non_pmd_mutex); |
6e3c6fa4 | 4115 | |
546e57d4 DDP |
4116 | dp_netdev_pmd_unref(non_pmd); |
4117 | } | |
1c1e46ed | 4118 | |
a6a426d6 | 4119 | if (dp_netdev_is_reconf_required(dp) || ports_require_restart(dp)) { |
e32971b8 | 4120 | reconfigure_datapath(dp); |
6e3c6fa4 DDP |
4121 | } |
4122 | ovs_mutex_unlock(&dp->port_mutex); | |
4123 | ||
53902038 | 4124 | tnl_neigh_cache_run(); |
7f9b8504 | 4125 | tnl_port_map_run(); |
a36de779 PS |
4126 | new_tnl_seq = seq_read(tnl_conf_seq); |
4127 | ||
4128 | if (dp->last_tnl_conf_seq != new_tnl_seq) { | |
4129 | dp->last_tnl_conf_seq = new_tnl_seq; | |
4130 | return true; | |
4131 | } | |
4132 | return false; | |
e4cfed38 PS |
4133 | } |
4134 | ||
4135 | static void | |
4136 | dpif_netdev_wait(struct dpif *dpif) | |
4137 | { | |
4138 | struct dp_netdev_port *port; | |
4139 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
4140 | ||
59e6d833 | 4141 | ovs_mutex_lock(&dp_netdev_mutex); |
e9985d6a DDP |
4142 | ovs_mutex_lock(&dp->port_mutex); |
4143 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
050c60bf | 4144 | netdev_wait_reconf_required(port->netdev); |
55c955bd PS |
4145 | if (!netdev_is_pmd(port->netdev)) { |
4146 | int i; | |
4147 | ||
490e82af | 4148 | for (i = 0; i < port->n_rxq; i++) { |
947dc567 | 4149 | netdev_rxq_wait(port->rxqs[i].rx); |
55c955bd | 4150 | } |
e4cfed38 PS |
4151 | } |
4152 | } | |
e9985d6a | 4153 | ovs_mutex_unlock(&dp->port_mutex); |
59e6d833 | 4154 | ovs_mutex_unlock(&dp_netdev_mutex); |
a36de779 | 4155 | seq_wait(tnl_conf_seq, dp->last_tnl_conf_seq); |
e4cfed38 PS |
4156 | } |
4157 | ||
d0cca6c3 DDP |
4158 | static void |
4159 | pmd_free_cached_ports(struct dp_netdev_pmd_thread *pmd) | |
4160 | { | |
4161 | struct tx_port *tx_port_cached; | |
4162 | ||
c71ea3c4 IM |
4163 | /* Flush all the queued packets. */ |
4164 | dp_netdev_pmd_flush_output_packets(pmd, true); | |
324c8374 | 4165 | /* Free all used tx queue ids. */ |
b010be17 | 4166 | dpif_netdev_xps_revalidate_pmd(pmd, true); |
324c8374 | 4167 | |
57eebbb4 DDP |
4168 | HMAP_FOR_EACH_POP (tx_port_cached, node, &pmd->tnl_port_cache) { |
4169 | free(tx_port_cached); | |
4170 | } | |
4171 | HMAP_FOR_EACH_POP (tx_port_cached, node, &pmd->send_port_cache) { | |
d0cca6c3 DDP |
4172 | free(tx_port_cached); |
4173 | } | |
4174 | } | |
4175 | ||
4176 | /* Copies ports from 'pmd->tx_ports' (shared with the main thread) to | |
899363ed BB |
4177 | * thread-local copies. Copy to 'pmd->tnl_port_cache' if it is a tunnel |
4178 | * device, otherwise to 'pmd->send_port_cache' if the port has at least | |
4179 | * one txq. */ | |
d0cca6c3 DDP |
4180 | static void |
4181 | pmd_load_cached_ports(struct dp_netdev_pmd_thread *pmd) | |
4182 | OVS_REQUIRES(pmd->port_mutex) | |
4183 | { | |
4184 | struct tx_port *tx_port, *tx_port_cached; | |
4185 | ||
4186 | pmd_free_cached_ports(pmd); | |
57eebbb4 DDP |
4187 | hmap_shrink(&pmd->send_port_cache); |
4188 | hmap_shrink(&pmd->tnl_port_cache); | |
d0cca6c3 DDP |
4189 | |
4190 | HMAP_FOR_EACH (tx_port, node, &pmd->tx_ports) { | |
57eebbb4 DDP |
4191 | if (netdev_has_tunnel_push_pop(tx_port->port->netdev)) { |
4192 | tx_port_cached = xmemdup(tx_port, sizeof *tx_port_cached); | |
4193 | hmap_insert(&pmd->tnl_port_cache, &tx_port_cached->node, | |
4194 | hash_port_no(tx_port_cached->port->port_no)); | |
4195 | } | |
4196 | ||
4197 | if (netdev_n_txq(tx_port->port->netdev)) { | |
4198 | tx_port_cached = xmemdup(tx_port, sizeof *tx_port_cached); | |
4199 | hmap_insert(&pmd->send_port_cache, &tx_port_cached->node, | |
4200 | hash_port_no(tx_port_cached->port->port_no)); | |
4201 | } | |
d0cca6c3 DDP |
4202 | } |
4203 | } | |
4204 | ||
140dd699 IM |
4205 | static void |
4206 | pmd_alloc_static_tx_qid(struct dp_netdev_pmd_thread *pmd) | |
4207 | { | |
4208 | ovs_mutex_lock(&pmd->dp->tx_qid_pool_mutex); | |
4209 | if (!id_pool_alloc_id(pmd->dp->tx_qid_pool, &pmd->static_tx_qid)) { | |
4210 | VLOG_ABORT("static_tx_qid allocation failed for PMD on core %2d" | |
4211 | ", numa_id %d.", pmd->core_id, pmd->numa_id); | |
4212 | } | |
4213 | ovs_mutex_unlock(&pmd->dp->tx_qid_pool_mutex); | |
4214 | ||
4215 | VLOG_DBG("static_tx_qid = %d allocated for PMD thread on core %2d" | |
4216 | ", numa_id %d.", pmd->static_tx_qid, pmd->core_id, pmd->numa_id); | |
4217 | } | |
4218 | ||
4219 | static void | |
4220 | pmd_free_static_tx_qid(struct dp_netdev_pmd_thread *pmd) | |
4221 | { | |
4222 | ovs_mutex_lock(&pmd->dp->tx_qid_pool_mutex); | |
4223 | id_pool_free_id(pmd->dp->tx_qid_pool, pmd->static_tx_qid); | |
4224 | ovs_mutex_unlock(&pmd->dp->tx_qid_pool_mutex); | |
4225 | } | |
4226 | ||
e4cfed38 | 4227 | static int |
d0cca6c3 | 4228 | pmd_load_queues_and_ports(struct dp_netdev_pmd_thread *pmd, |
947dc567 | 4229 | struct polled_queue **ppoll_list) |
e4cfed38 | 4230 | { |
947dc567 | 4231 | struct polled_queue *poll_list = *ppoll_list; |
ae7ad0a1 IM |
4232 | struct rxq_poll *poll; |
4233 | int i; | |
e4cfed38 | 4234 | |
d0cca6c3 | 4235 | ovs_mutex_lock(&pmd->port_mutex); |
947dc567 DDP |
4236 | poll_list = xrealloc(poll_list, hmap_count(&pmd->poll_list) |
4237 | * sizeof *poll_list); | |
a1fdee13 | 4238 | |
ae7ad0a1 | 4239 | i = 0; |
947dc567 | 4240 | HMAP_FOR_EACH (poll, node, &pmd->poll_list) { |
922b28d4 | 4241 | poll_list[i].rxq = poll->rxq; |
947dc567 DDP |
4242 | poll_list[i].port_no = poll->rxq->port->port_no; |
4243 | i++; | |
e4cfed38 | 4244 | } |
d0cca6c3 DDP |
4245 | |
4246 | pmd_load_cached_ports(pmd); | |
4247 | ||
4248 | ovs_mutex_unlock(&pmd->port_mutex); | |
e4cfed38 | 4249 | |
e4cfed38 | 4250 | *ppoll_list = poll_list; |
d42f9307 | 4251 | return i; |
e4cfed38 PS |
4252 | } |
4253 | ||
6c3eee82 | 4254 | static void * |
e4cfed38 | 4255 | pmd_thread_main(void *f_) |
6c3eee82 | 4256 | { |
65f13b50 | 4257 | struct dp_netdev_pmd_thread *pmd = f_; |
82a48ead | 4258 | struct pmd_perf_stats *s = &pmd->perf_stats; |
e4cfed38 | 4259 | unsigned int lc = 0; |
947dc567 | 4260 | struct polled_queue *poll_list; |
d42f9307 | 4261 | bool exiting; |
e4cfed38 PS |
4262 | int poll_cnt; |
4263 | int i; | |
a2ac666d | 4264 | int process_packets = 0; |
6c3eee82 | 4265 | |
e4cfed38 PS |
4266 | poll_list = NULL; |
4267 | ||
65f13b50 AW |
4268 | /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */ |
4269 | ovsthread_setspecific(pmd->dp->per_pmd_key, pmd); | |
6930c7e0 DDP |
4270 | ovs_numa_thread_setaffinity_core(pmd->core_id); |
4271 | dpdk_set_lcore_id(pmd->core_id); | |
d0cca6c3 | 4272 | poll_cnt = pmd_load_queues_and_ports(pmd, &poll_list); |
e215018b | 4273 | emc_cache_init(&pmd->flow_cache); |
e4cfed38 | 4274 | reload: |
140dd699 | 4275 | pmd_alloc_static_tx_qid(pmd); |
ae7ad0a1 | 4276 | |
7dd671f0 MK |
4277 | /* List port/core affinity */ |
4278 | for (i = 0; i < poll_cnt; i++) { | |
ce179f11 | 4279 | VLOG_DBG("Core %d processing port \'%s\' with queue-id %d\n", |
922b28d4 KT |
4280 | pmd->core_id, netdev_rxq_get_name(poll_list[i].rxq->rx), |
4281 | netdev_rxq_get_queue_id(poll_list[i].rxq->rx)); | |
4f5d13e2 KT |
4282 | /* Reset the rxq current cycles counter. */ |
4283 | dp_netdev_rxq_set_cycles(poll_list[i].rxq, RXQ_CYCLES_PROC_CURR, 0); | |
7dd671f0 MK |
4284 | } |
4285 | ||
2788a1b1 DDP |
4286 | if (!poll_cnt) { |
4287 | while (seq_read(pmd->reload_seq) == pmd->last_reload_seq) { | |
4288 | seq_wait(pmd->reload_seq, pmd->last_reload_seq); | |
4289 | poll_block(); | |
4290 | } | |
4291 | lc = UINT_MAX; | |
4292 | } | |
4293 | ||
2a2c67b4 KT |
4294 | pmd->intrvl_tsc_prev = 0; |
4295 | atomic_store_relaxed(&pmd->intrvl_cycles, 0); | |
a19896ab | 4296 | cycles_counter_update(s); |
79f36875 JS |
4297 | /* Protect pmd stats from external clearing while polling. */ |
4298 | ovs_mutex_lock(&pmd->perf_stats.stats_mutex); | |
e4cfed38 | 4299 | for (;;) { |
79f36875 | 4300 | uint64_t rx_packets = 0, tx_packets = 0; |
c71ea3c4 | 4301 | |
a19896ab | 4302 | pmd_perf_start_iteration(s); |
79f36875 | 4303 | |
e4cfed38 | 4304 | for (i = 0; i < poll_cnt; i++) { |
a2ac666d | 4305 | process_packets = |
a19896ab | 4306 | dp_netdev_process_rxq_port(pmd, poll_list[i].rxq, |
a2ac666d | 4307 | poll_list[i].port_no); |
79f36875 | 4308 | rx_packets += process_packets; |
e4cfed38 PS |
4309 | } |
4310 | ||
79f36875 | 4311 | if (!rx_packets) { |
c71ea3c4 IM |
4312 | /* We didn't receive anything in the process loop. |
4313 | * Check if we need to send something. | |
4314 | * There was no time updates on current iteration. */ | |
4315 | pmd_thread_ctx_time_update(pmd); | |
79f36875 | 4316 | tx_packets = dp_netdev_pmd_flush_output_packets(pmd, false); |
c71ea3c4 IM |
4317 | } |
4318 | ||
e4cfed38 | 4319 | if (lc++ > 1024) { |
14e3e12a | 4320 | bool reload; |
6c3eee82 | 4321 | |
e4cfed38 | 4322 | lc = 0; |
84067a4c | 4323 | |
fbe0962b | 4324 | coverage_try_clear(); |
4809891b | 4325 | dp_netdev_pmd_try_optimize(pmd, poll_list, poll_cnt); |
9dede5cf FL |
4326 | if (!ovsrcu_try_quiesce()) { |
4327 | emc_cache_slow_sweep(&pmd->flow_cache); | |
4328 | } | |
84067a4c | 4329 | |
14e3e12a DDP |
4330 | atomic_read_relaxed(&pmd->reload, &reload); |
4331 | if (reload) { | |
6c3eee82 BP |
4332 | break; |
4333 | } | |
4334 | } | |
79f36875 JS |
4335 | pmd_perf_end_iteration(s, rx_packets, tx_packets, |
4336 | pmd_perf_metrics_enabled(pmd)); | |
e4cfed38 | 4337 | } |
79f36875 | 4338 | ovs_mutex_unlock(&pmd->perf_stats.stats_mutex); |
6c3eee82 | 4339 | |
d0cca6c3 | 4340 | poll_cnt = pmd_load_queues_and_ports(pmd, &poll_list); |
d42f9307 DDP |
4341 | exiting = latch_is_set(&pmd->exit_latch); |
4342 | /* Signal here to make sure the pmd finishes | |
4343 | * reloading the updated configuration. */ | |
4344 | dp_netdev_pmd_reload_done(pmd); | |
4345 | ||
140dd699 | 4346 | pmd_free_static_tx_qid(pmd); |
9bbf1c3d | 4347 | |
d42f9307 | 4348 | if (!exiting) { |
e4cfed38 PS |
4349 | goto reload; |
4350 | } | |
6c3eee82 | 4351 | |
e215018b | 4352 | emc_cache_uninit(&pmd->flow_cache); |
e4cfed38 | 4353 | free(poll_list); |
d0cca6c3 | 4354 | pmd_free_cached_ports(pmd); |
6c3eee82 BP |
4355 | return NULL; |
4356 | } | |
4357 | ||
6b31e073 RW |
4358 | static void |
4359 | dp_netdev_disable_upcall(struct dp_netdev *dp) | |
4360 | OVS_ACQUIRES(dp->upcall_rwlock) | |
4361 | { | |
4362 | fat_rwlock_wrlock(&dp->upcall_rwlock); | |
4363 | } | |
4364 | ||
5dddf960 JR |
4365 | \f |
4366 | /* Meters */ | |
4367 | static void | |
4368 | dpif_netdev_meter_get_features(const struct dpif * dpif OVS_UNUSED, | |
4369 | struct ofputil_meter_features *features) | |
4370 | { | |
4b27db64 JR |
4371 | features->max_meters = MAX_METERS; |
4372 | features->band_types = DP_SUPPORTED_METER_BAND_TYPES; | |
4373 | features->capabilities = DP_SUPPORTED_METER_FLAGS_MASK; | |
4374 | features->max_bands = MAX_BANDS; | |
5dddf960 JR |
4375 | features->max_color = 0; |
4376 | } | |
4377 | ||
4b27db64 JR |
4378 | /* Returns false when packet needs to be dropped. */ |
4379 | static void | |
4380 | dp_netdev_run_meter(struct dp_netdev *dp, struct dp_packet_batch *packets_, | |
4381 | uint32_t meter_id, long long int now) | |
4382 | { | |
4383 | struct dp_meter *meter; | |
4384 | struct dp_meter_band *band; | |
79c81260 | 4385 | struct dp_packet *packet; |
4b27db64 JR |
4386 | long long int long_delta_t; /* msec */ |
4387 | uint32_t delta_t; /* msec */ | |
79c81260 | 4388 | const size_t cnt = dp_packet_batch_size(packets_); |
4b27db64 JR |
4389 | uint32_t bytes, volume; |
4390 | int exceeded_band[NETDEV_MAX_BURST]; | |
4391 | uint32_t exceeded_rate[NETDEV_MAX_BURST]; | |
4392 | int exceeded_pkt = cnt; /* First packet that exceeded a band rate. */ | |
4393 | ||
4394 | if (meter_id >= MAX_METERS) { | |
4395 | return; | |
4396 | } | |
4397 | ||
4398 | meter_lock(dp, meter_id); | |
4399 | meter = dp->meters[meter_id]; | |
4400 | if (!meter) { | |
4401 | goto out; | |
4402 | } | |
4403 | ||
4404 | /* Initialize as negative values. */ | |
4405 | memset(exceeded_band, 0xff, cnt * sizeof *exceeded_band); | |
4406 | /* Initialize as zeroes. */ | |
4407 | memset(exceeded_rate, 0, cnt * sizeof *exceeded_rate); | |
4408 | ||
4409 | /* All packets will hit the meter at the same time. */ | |
05f9e707 | 4410 | long_delta_t = (now - meter->used) / 1000; /* msec */ |
4b27db64 JR |
4411 | |
4412 | /* Make sure delta_t will not be too large, so that bucket will not | |
4413 | * wrap around below. */ | |
4414 | delta_t = (long_delta_t > (long long int)meter->max_delta_t) | |
4415 | ? meter->max_delta_t : (uint32_t)long_delta_t; | |
4416 | ||
4417 | /* Update meter stats. */ | |
4418 | meter->used = now; | |
4419 | meter->packet_count += cnt; | |
4420 | bytes = 0; | |
e883448e | 4421 | DP_PACKET_BATCH_FOR_EACH (i, packet, packets_) { |
79c81260 | 4422 | bytes += dp_packet_size(packet); |
4b27db64 JR |
4423 | } |
4424 | meter->byte_count += bytes; | |
4425 | ||
4426 | /* Meters can operate in terms of packets per second or kilobits per | |
4427 | * second. */ | |
4428 | if (meter->flags & OFPMF13_PKTPS) { | |
4429 | /* Rate in packets/second, bucket 1/1000 packets. */ | |
4430 | /* msec * packets/sec = 1/1000 packets. */ | |
4431 | volume = cnt * 1000; /* Take 'cnt' packets from the bucket. */ | |
4432 | } else { | |
4433 | /* Rate in kbps, bucket in bits. */ | |
4434 | /* msec * kbps = bits */ | |
4435 | volume = bytes * 8; | |
4436 | } | |
4437 | ||
4438 | /* Update all bands and find the one hit with the highest rate for each | |
4439 | * packet (if any). */ | |
4440 | for (int m = 0; m < meter->n_bands; ++m) { | |
4441 | band = &meter->bands[m]; | |
4442 | ||
4443 | /* Update band's bucket. */ | |
4444 | band->bucket += delta_t * band->up.rate; | |
4445 | if (band->bucket > band->up.burst_size) { | |
4446 | band->bucket = band->up.burst_size; | |
4447 | } | |
4448 | ||
4449 | /* Drain the bucket for all the packets, if possible. */ | |
4450 | if (band->bucket >= volume) { | |
4451 | band->bucket -= volume; | |
4452 | } else { | |
4453 | int band_exceeded_pkt; | |
4454 | ||
4455 | /* Band limit hit, must process packet-by-packet. */ | |
4456 | if (meter->flags & OFPMF13_PKTPS) { | |
4457 | band_exceeded_pkt = band->bucket / 1000; | |
4458 | band->bucket %= 1000; /* Remainder stays in bucket. */ | |
4459 | ||
4460 | /* Update the exceeding band for each exceeding packet. | |
4461 | * (Only one band will be fired by a packet, and that | |
4462 | * can be different for each packet.) */ | |
e883448e | 4463 | for (int i = band_exceeded_pkt; i < cnt; i++) { |
4b27db64 JR |
4464 | if (band->up.rate > exceeded_rate[i]) { |
4465 | exceeded_rate[i] = band->up.rate; | |
4466 | exceeded_band[i] = m; | |
4467 | } | |
4468 | } | |
4469 | } else { | |
4470 | /* Packet sizes differ, must process one-by-one. */ | |
4471 | band_exceeded_pkt = cnt; | |
e883448e | 4472 | DP_PACKET_BATCH_FOR_EACH (i, packet, packets_) { |
79c81260 | 4473 | uint32_t bits = dp_packet_size(packet) * 8; |
4b27db64 JR |
4474 | |
4475 | if (band->bucket >= bits) { | |
4476 | band->bucket -= bits; | |
4477 | } else { | |
4478 | if (i < band_exceeded_pkt) { | |
4479 | band_exceeded_pkt = i; | |
4480 | } | |
4481 | /* Update the exceeding band for the exceeding packet. | |
4482 | * (Only one band will be fired by a packet, and that | |
4483 | * can be different for each packet.) */ | |
4484 | if (band->up.rate > exceeded_rate[i]) { | |
4485 | exceeded_rate[i] = band->up.rate; | |
4486 | exceeded_band[i] = m; | |
4487 | } | |
4488 | } | |
4489 | } | |
4490 | } | |
4491 | /* Remember the first exceeding packet. */ | |
4492 | if (exceeded_pkt > band_exceeded_pkt) { | |
4493 | exceeded_pkt = band_exceeded_pkt; | |
4494 | } | |
4495 | } | |
4496 | } | |
4497 | ||
4498 | /* Fire the highest rate band exceeded by each packet. | |
4499 | * Drop packets if needed, by swapping packet to the end that will be | |
4500 | * ignored. */ | |
4b27db64 | 4501 | size_t j; |
79c81260 | 4502 | DP_PACKET_BATCH_REFILL_FOR_EACH (j, cnt, packet, packets_) { |
4b27db64 JR |
4503 | if (exceeded_band[j] >= 0) { |
4504 | /* Meter drop packet. */ | |
4505 | band = &meter->bands[exceeded_band[j]]; | |
4506 | band->packet_count += 1; | |
4507 | band->byte_count += dp_packet_size(packet); | |
4508 | ||
4509 | dp_packet_delete(packet); | |
4510 | } else { | |
4511 | /* Meter accepts packet. */ | |
4512 | dp_packet_batch_refill(packets_, packet, j); | |
4513 | } | |
4514 | } | |
4515 | out: | |
4516 | meter_unlock(dp, meter_id); | |
4517 | } | |
4518 | ||
4519 | /* Meter set/get/del processing is still single-threaded. */ | |
5dddf960 | 4520 | static int |
4b27db64 JR |
4521 | dpif_netdev_meter_set(struct dpif *dpif, ofproto_meter_id *meter_id, |
4522 | struct ofputil_meter_config *config) | |
5dddf960 | 4523 | { |
4b27db64 JR |
4524 | struct dp_netdev *dp = get_dp_netdev(dpif); |
4525 | uint32_t mid = meter_id->uint32; | |
4526 | struct dp_meter *meter; | |
4527 | int i; | |
4528 | ||
4b27db64 JR |
4529 | if (mid >= MAX_METERS) { |
4530 | return EFBIG; /* Meter_id out of range. */ | |
4531 | } | |
4532 | ||
4533 | if (config->flags & ~DP_SUPPORTED_METER_FLAGS_MASK || | |
4534 | !(config->flags & (OFPMF13_KBPS | OFPMF13_PKTPS))) { | |
4535 | return EBADF; /* Unsupported flags set */ | |
4536 | } | |
2029ce9a | 4537 | |
4b27db64 JR |
4538 | /* Validate bands */ |
4539 | if (config->n_bands == 0 || config->n_bands > MAX_BANDS) { | |
4540 | return EINVAL; /* Too many bands */ | |
4541 | } | |
2029ce9a AVA |
4542 | |
4543 | /* Validate rates */ | |
4544 | for (i = 0; i < config->n_bands; i++) { | |
4545 | if (config->bands[i].rate == 0) { | |
66a396d4 | 4546 | return EDOM; /* rate must be non-zero */ |
2029ce9a AVA |
4547 | } |
4548 | } | |
4549 | ||
4b27db64 JR |
4550 | for (i = 0; i < config->n_bands; ++i) { |
4551 | switch (config->bands[i].type) { | |
4552 | case OFPMBT13_DROP: | |
4553 | break; | |
4554 | default: | |
4555 | return ENODEV; /* Unsupported band type */ | |
4556 | } | |
4557 | } | |
4558 | ||
4559 | /* Allocate meter */ | |
4560 | meter = xzalloc(sizeof *meter | |
4561 | + config->n_bands * sizeof(struct dp_meter_band)); | |
4562 | if (meter) { | |
4563 | meter->flags = config->flags; | |
4564 | meter->n_bands = config->n_bands; | |
4565 | meter->max_delta_t = 0; | |
05f9e707 | 4566 | meter->used = time_usec(); |
4b27db64 JR |
4567 | |
4568 | /* set up bands */ | |
4569 | for (i = 0; i < config->n_bands; ++i) { | |
4570 | uint32_t band_max_delta_t; | |
4571 | ||
4572 | /* Set burst size to a workable value if none specified. */ | |
4573 | if (config->bands[i].burst_size == 0) { | |
4574 | config->bands[i].burst_size = config->bands[i].rate; | |
4575 | } | |
4576 | ||
4577 | meter->bands[i].up = config->bands[i]; | |
4578 | /* Convert burst size to the bucket units: */ | |
4579 | /* pkts => 1/1000 packets, kilobits => bits. */ | |
4580 | meter->bands[i].up.burst_size *= 1000; | |
4581 | /* Initialize bucket to empty. */ | |
4582 | meter->bands[i].bucket = 0; | |
4583 | ||
4584 | /* Figure out max delta_t that is enough to fill any bucket. */ | |
4585 | band_max_delta_t | |
4586 | = meter->bands[i].up.burst_size / meter->bands[i].up.rate; | |
4587 | if (band_max_delta_t > meter->max_delta_t) { | |
4588 | meter->max_delta_t = band_max_delta_t; | |
4589 | } | |
4590 | } | |
4591 | ||
4592 | meter_lock(dp, mid); | |
4593 | dp_delete_meter(dp, mid); /* Free existing meter, if any */ | |
4594 | dp->meters[mid] = meter; | |
4595 | meter_unlock(dp, mid); | |
4596 | ||
4b27db64 JR |
4597 | return 0; |
4598 | } | |
4599 | return ENOMEM; | |
5dddf960 JR |
4600 | } |
4601 | ||
4602 | static int | |
4b27db64 JR |
4603 | dpif_netdev_meter_get(const struct dpif *dpif, |
4604 | ofproto_meter_id meter_id_, | |
4605 | struct ofputil_meter_stats *stats, uint16_t n_bands) | |
5dddf960 | 4606 | { |
4b27db64 JR |
4607 | const struct dp_netdev *dp = get_dp_netdev(dpif); |
4608 | const struct dp_meter *meter; | |
4609 | uint32_t meter_id = meter_id_.uint32; | |
4610 | ||
4611 | if (meter_id >= MAX_METERS) { | |
4612 | return EFBIG; | |
4613 | } | |
4614 | meter = dp->meters[meter_id]; | |
4615 | if (!meter) { | |
4616 | return ENOENT; | |
4617 | } | |
4618 | if (stats) { | |
4619 | int i = 0; | |
4620 | ||
4621 | meter_lock(dp, meter_id); | |
4622 | stats->packet_in_count = meter->packet_count; | |
4623 | stats->byte_in_count = meter->byte_count; | |
4624 | ||
4625 | for (i = 0; i < n_bands && i < meter->n_bands; ++i) { | |
4626 | stats->bands[i].packet_count = meter->bands[i].packet_count; | |
4627 | stats->bands[i].byte_count = meter->bands[i].byte_count; | |
4628 | } | |
4629 | meter_unlock(dp, meter_id); | |
4630 | ||
4631 | stats->n_bands = i; | |
4632 | } | |
4633 | return 0; | |
5dddf960 JR |
4634 | } |
4635 | ||
4636 | static int | |
4b27db64 JR |
4637 | dpif_netdev_meter_del(struct dpif *dpif, |
4638 | ofproto_meter_id meter_id_, | |
4639 | struct ofputil_meter_stats *stats, uint16_t n_bands) | |
5dddf960 | 4640 | { |
4b27db64 JR |
4641 | struct dp_netdev *dp = get_dp_netdev(dpif); |
4642 | int error; | |
4643 | ||
4644 | error = dpif_netdev_meter_get(dpif, meter_id_, stats, n_bands); | |
4645 | if (!error) { | |
4646 | uint32_t meter_id = meter_id_.uint32; | |
4647 | ||
4648 | meter_lock(dp, meter_id); | |
4649 | dp_delete_meter(dp, meter_id); | |
4650 | meter_unlock(dp, meter_id); | |
4b27db64 JR |
4651 | } |
4652 | return error; | |
5dddf960 JR |
4653 | } |
4654 | ||
4655 | \f | |
6b31e073 RW |
4656 | static void |
4657 | dpif_netdev_disable_upcall(struct dpif *dpif) | |
4658 | OVS_NO_THREAD_SAFETY_ANALYSIS | |
4659 | { | |
4660 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
4661 | dp_netdev_disable_upcall(dp); | |
4662 | } | |
4663 | ||
4664 | static void | |
4665 | dp_netdev_enable_upcall(struct dp_netdev *dp) | |
4666 | OVS_RELEASES(dp->upcall_rwlock) | |
4667 | { | |
4668 | fat_rwlock_unlock(&dp->upcall_rwlock); | |
4669 | } | |
4670 | ||
4671 | static void | |
4672 | dpif_netdev_enable_upcall(struct dpif *dpif) | |
4673 | OVS_NO_THREAD_SAFETY_ANALYSIS | |
4674 | { | |
4675 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
4676 | dp_netdev_enable_upcall(dp); | |
4677 | } | |
4678 | ||
ae7ad0a1 | 4679 | static void |
accf8626 AW |
4680 | dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd) |
4681 | { | |
4682 | ovs_mutex_lock(&pmd->cond_mutex); | |
14e3e12a | 4683 | atomic_store_relaxed(&pmd->reload, false); |
2788a1b1 | 4684 | pmd->last_reload_seq = seq_read(pmd->reload_seq); |
accf8626 AW |
4685 | xpthread_cond_signal(&pmd->cond); |
4686 | ovs_mutex_unlock(&pmd->cond_mutex); | |
4687 | } | |
4688 | ||
1c1e46ed | 4689 | /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns |
546e57d4 DDP |
4690 | * the pointer if succeeds, otherwise, NULL (it can return NULL even if |
4691 | * 'core_id' is NON_PMD_CORE_ID). | |
1c1e46ed AW |
4692 | * |
4693 | * Caller must unrefs the returned reference. */ | |
65f13b50 | 4694 | static struct dp_netdev_pmd_thread * |
bd5131ba | 4695 | dp_netdev_get_pmd(struct dp_netdev *dp, unsigned core_id) |
65f13b50 AW |
4696 | { |
4697 | struct dp_netdev_pmd_thread *pmd; | |
55847abe | 4698 | const struct cmap_node *pnode; |
65f13b50 | 4699 | |
b19befae | 4700 | pnode = cmap_find(&dp->poll_threads, hash_int(core_id, 0)); |
1c1e46ed AW |
4701 | if (!pnode) { |
4702 | return NULL; | |
4703 | } | |
65f13b50 AW |
4704 | pmd = CONTAINER_OF(pnode, struct dp_netdev_pmd_thread, node); |
4705 | ||
1c1e46ed | 4706 | return dp_netdev_pmd_try_ref(pmd) ? pmd : NULL; |
65f13b50 AW |
4707 | } |
4708 | ||
f2eee189 AW |
4709 | /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */ |
4710 | static void | |
4711 | dp_netdev_set_nonpmd(struct dp_netdev *dp) | |
e9985d6a | 4712 | OVS_REQUIRES(dp->port_mutex) |
f2eee189 AW |
4713 | { |
4714 | struct dp_netdev_pmd_thread *non_pmd; | |
4715 | ||
4716 | non_pmd = xzalloc(sizeof *non_pmd); | |
00873463 | 4717 | dp_netdev_configure_pmd(non_pmd, dp, NON_PMD_CORE_ID, OVS_NUMA_UNSPEC); |
f2eee189 AW |
4718 | } |
4719 | ||
1c1e46ed AW |
4720 | /* Caller must have valid pointer to 'pmd'. */ |
4721 | static bool | |
4722 | dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd) | |
4723 | { | |
4724 | return ovs_refcount_try_ref_rcu(&pmd->ref_cnt); | |
4725 | } | |
4726 | ||
4727 | static void | |
4728 | dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd) | |
4729 | { | |
4730 | if (pmd && ovs_refcount_unref(&pmd->ref_cnt) == 1) { | |
4731 | ovsrcu_postpone(dp_netdev_destroy_pmd, pmd); | |
4732 | } | |
4733 | } | |
4734 | ||
4735 | /* Given cmap position 'pos', tries to ref the next node. If try_ref() | |
4736 | * fails, keeps checking for next node until reaching the end of cmap. | |
4737 | * | |
4738 | * Caller must unrefs the returned reference. */ | |
4739 | static struct dp_netdev_pmd_thread * | |
4740 | dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos) | |
4741 | { | |
4742 | struct dp_netdev_pmd_thread *next; | |
4743 | ||
4744 | do { | |
4745 | struct cmap_node *node; | |
4746 | ||
4747 | node = cmap_next_position(&dp->poll_threads, pos); | |
4748 | next = node ? CONTAINER_OF(node, struct dp_netdev_pmd_thread, node) | |
4749 | : NULL; | |
4750 | } while (next && !dp_netdev_pmd_try_ref(next)); | |
4751 | ||
4752 | return next; | |
4753 | } | |
4754 | ||
65f13b50 | 4755 | /* Configures the 'pmd' based on the input argument. */ |
6c3eee82 | 4756 | static void |
65f13b50 | 4757 | dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp, |
00873463 | 4758 | unsigned core_id, int numa_id) |
65f13b50 AW |
4759 | { |
4760 | pmd->dp = dp; | |
65f13b50 AW |
4761 | pmd->core_id = core_id; |
4762 | pmd->numa_id = numa_id; | |
e32971b8 | 4763 | pmd->need_reload = false; |
c71ea3c4 | 4764 | pmd->n_output_batches = 0; |
1c1e46ed AW |
4765 | |
4766 | ovs_refcount_init(&pmd->ref_cnt); | |
65f13b50 | 4767 | latch_init(&pmd->exit_latch); |
2788a1b1 DDP |
4768 | pmd->reload_seq = seq_create(); |
4769 | pmd->last_reload_seq = seq_read(pmd->reload_seq); | |
14e3e12a | 4770 | atomic_init(&pmd->reload, false); |
accf8626 AW |
4771 | xpthread_cond_init(&pmd->cond, NULL); |
4772 | ovs_mutex_init(&pmd->cond_mutex); | |
1c1e46ed | 4773 | ovs_mutex_init(&pmd->flow_mutex); |
d0cca6c3 | 4774 | ovs_mutex_init(&pmd->port_mutex); |
1c1e46ed | 4775 | cmap_init(&pmd->flow_table); |
3453b4d6 | 4776 | cmap_init(&pmd->classifiers); |
58ed6df0 | 4777 | pmd->ctx.last_rxq = NULL; |
b010be17 IM |
4778 | pmd_thread_ctx_time_update(pmd); |
4779 | pmd->next_optimization = pmd->ctx.now + DPCLS_OPTIMIZATION_INTERVAL; | |
4780 | pmd->rxq_next_cycle_store = pmd->ctx.now + PMD_RXQ_INTERVAL_LEN; | |
947dc567 | 4781 | hmap_init(&pmd->poll_list); |
d0cca6c3 | 4782 | hmap_init(&pmd->tx_ports); |
57eebbb4 DDP |
4783 | hmap_init(&pmd->tnl_port_cache); |
4784 | hmap_init(&pmd->send_port_cache); | |
65f13b50 AW |
4785 | /* init the 'flow_cache' since there is no |
4786 | * actual thread created for NON_PMD_CORE_ID. */ | |
4787 | if (core_id == NON_PMD_CORE_ID) { | |
4788 | emc_cache_init(&pmd->flow_cache); | |
140dd699 | 4789 | pmd_alloc_static_tx_qid(pmd); |
65f13b50 | 4790 | } |
82a48ead | 4791 | pmd_perf_stats_init(&pmd->perf_stats); |
65f13b50 AW |
4792 | cmap_insert(&dp->poll_threads, CONST_CAST(struct cmap_node *, &pmd->node), |
4793 | hash_int(core_id, 0)); | |
4794 | } | |
4795 | ||
1c1e46ed AW |
4796 | static void |
4797 | dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd) | |
4798 | { | |
3453b4d6 JS |
4799 | struct dpcls *cls; |
4800 | ||
1c1e46ed | 4801 | dp_netdev_pmd_flow_flush(pmd); |
57eebbb4 DDP |
4802 | hmap_destroy(&pmd->send_port_cache); |
4803 | hmap_destroy(&pmd->tnl_port_cache); | |
d0cca6c3 | 4804 | hmap_destroy(&pmd->tx_ports); |
947dc567 | 4805 | hmap_destroy(&pmd->poll_list); |
3453b4d6 JS |
4806 | /* All flows (including their dpcls_rules) have been deleted already */ |
4807 | CMAP_FOR_EACH (cls, node, &pmd->classifiers) { | |
4808 | dpcls_destroy(cls); | |
7c269972 | 4809 | ovsrcu_postpone(free, cls); |
3453b4d6 JS |
4810 | } |
4811 | cmap_destroy(&pmd->classifiers); | |
1c1e46ed AW |
4812 | cmap_destroy(&pmd->flow_table); |
4813 | ovs_mutex_destroy(&pmd->flow_mutex); | |
4814 | latch_destroy(&pmd->exit_latch); | |
2788a1b1 | 4815 | seq_destroy(pmd->reload_seq); |
1c1e46ed AW |
4816 | xpthread_cond_destroy(&pmd->cond); |
4817 | ovs_mutex_destroy(&pmd->cond_mutex); | |
d0cca6c3 | 4818 | ovs_mutex_destroy(&pmd->port_mutex); |
1c1e46ed AW |
4819 | free(pmd); |
4820 | } | |
4821 | ||
4822 | /* Stops the pmd thread, removes it from the 'dp->poll_threads', | |
4823 | * and unrefs the struct. */ | |
65f13b50 | 4824 | static void |
e4e74c3a | 4825 | dp_netdev_del_pmd(struct dp_netdev *dp, struct dp_netdev_pmd_thread *pmd) |
6c3eee82 | 4826 | { |
d0cca6c3 DDP |
4827 | /* NON_PMD_CORE_ID doesn't have a thread, so we don't have to synchronize, |
4828 | * but extra cleanup is necessary */ | |
65f13b50 | 4829 | if (pmd->core_id == NON_PMD_CORE_ID) { |
febf4a7a | 4830 | ovs_mutex_lock(&dp->non_pmd_mutex); |
65f13b50 | 4831 | emc_cache_uninit(&pmd->flow_cache); |
d0cca6c3 | 4832 | pmd_free_cached_ports(pmd); |
140dd699 | 4833 | pmd_free_static_tx_qid(pmd); |
febf4a7a | 4834 | ovs_mutex_unlock(&dp->non_pmd_mutex); |
65f13b50 AW |
4835 | } else { |
4836 | latch_set(&pmd->exit_latch); | |
4837 | dp_netdev_reload_pmd__(pmd); | |
65f13b50 AW |
4838 | xpthread_join(pmd->thread, NULL); |
4839 | } | |
ae7ad0a1 | 4840 | |
d0cca6c3 | 4841 | dp_netdev_pmd_clear_ports(pmd); |
ae7ad0a1 | 4842 | |
e4e74c3a AW |
4843 | /* Purges the 'pmd''s flows after stopping the thread, but before |
4844 | * destroying the flows, so that the flow stats can be collected. */ | |
4845 | if (dp->dp_purge_cb) { | |
4846 | dp->dp_purge_cb(dp->dp_purge_aux, pmd->core_id); | |
4847 | } | |
65f13b50 | 4848 | cmap_remove(&pmd->dp->poll_threads, &pmd->node, hash_int(pmd->core_id, 0)); |
1c1e46ed | 4849 | dp_netdev_pmd_unref(pmd); |
65f13b50 | 4850 | } |
6c3eee82 | 4851 | |
e32971b8 DDP |
4852 | /* Destroys all pmd threads. If 'non_pmd' is true it also destroys the non pmd |
4853 | * thread. */ | |
65f13b50 | 4854 | static void |
e32971b8 | 4855 | dp_netdev_destroy_all_pmds(struct dp_netdev *dp, bool non_pmd) |
65f13b50 AW |
4856 | { |
4857 | struct dp_netdev_pmd_thread *pmd; | |
d916785c DDP |
4858 | struct dp_netdev_pmd_thread **pmd_list; |
4859 | size_t k = 0, n_pmds; | |
4860 | ||
e32971b8 | 4861 | n_pmds = cmap_count(&dp->poll_threads); |
d916785c | 4862 | pmd_list = xcalloc(n_pmds, sizeof *pmd_list); |
65f13b50 AW |
4863 | |
4864 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
e32971b8 | 4865 | if (!non_pmd && pmd->core_id == NON_PMD_CORE_ID) { |
b9584f21 DDP |
4866 | continue; |
4867 | } | |
d916785c DDP |
4868 | /* We cannot call dp_netdev_del_pmd(), since it alters |
4869 | * 'dp->poll_threads' (while we're iterating it) and it | |
4870 | * might quiesce. */ | |
4871 | ovs_assert(k < n_pmds); | |
4872 | pmd_list[k++] = pmd; | |
6c3eee82 | 4873 | } |
d916785c DDP |
4874 | |
4875 | for (size_t i = 0; i < k; i++) { | |
4876 | dp_netdev_del_pmd(dp, pmd_list[i]); | |
4877 | } | |
4878 | free(pmd_list); | |
65f13b50 | 4879 | } |
6c3eee82 | 4880 | |
d0cca6c3 DDP |
4881 | /* Deletes all rx queues from pmd->poll_list and all the ports from |
4882 | * pmd->tx_ports. */ | |
cc245ce8 | 4883 | static void |
d0cca6c3 | 4884 | dp_netdev_pmd_clear_ports(struct dp_netdev_pmd_thread *pmd) |
cc245ce8 IM |
4885 | { |
4886 | struct rxq_poll *poll; | |
d0cca6c3 | 4887 | struct tx_port *port; |
cc245ce8 | 4888 | |
d0cca6c3 | 4889 | ovs_mutex_lock(&pmd->port_mutex); |
947dc567 | 4890 | HMAP_FOR_EACH_POP (poll, node, &pmd->poll_list) { |
cc245ce8 IM |
4891 | free(poll); |
4892 | } | |
d0cca6c3 DDP |
4893 | HMAP_FOR_EACH_POP (port, node, &pmd->tx_ports) { |
4894 | free(port); | |
4895 | } | |
4896 | ovs_mutex_unlock(&pmd->port_mutex); | |
cc245ce8 IM |
4897 | } |
4898 | ||
e32971b8 | 4899 | /* Adds rx queue to poll_list of PMD thread, if it's not there already. */ |
b68872d8 | 4900 | static void |
e32971b8 DDP |
4901 | dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread *pmd, |
4902 | struct dp_netdev_rxq *rxq) | |
4903 | OVS_REQUIRES(pmd->port_mutex) | |
b68872d8 | 4904 | { |
e32971b8 DDP |
4905 | int qid = netdev_rxq_get_queue_id(rxq->rx); |
4906 | uint32_t hash = hash_2words(odp_to_u32(rxq->port->port_no), qid); | |
4907 | struct rxq_poll *poll; | |
b68872d8 | 4908 | |
e32971b8 DDP |
4909 | HMAP_FOR_EACH_WITH_HASH (poll, node, hash, &pmd->poll_list) { |
4910 | if (poll->rxq == rxq) { | |
4911 | /* 'rxq' is already polled by this thread. Do nothing. */ | |
4912 | return; | |
d0cca6c3 | 4913 | } |
cc245ce8 | 4914 | } |
cc245ce8 | 4915 | |
e32971b8 DDP |
4916 | poll = xmalloc(sizeof *poll); |
4917 | poll->rxq = rxq; | |
4918 | hmap_insert(&pmd->poll_list, &poll->node, hash); | |
b68872d8 | 4919 | |
e32971b8 | 4920 | pmd->need_reload = true; |
ae7ad0a1 IM |
4921 | } |
4922 | ||
e32971b8 | 4923 | /* Delete 'poll' from poll_list of PMD thread. */ |
ae7ad0a1 | 4924 | static void |
e32971b8 DDP |
4925 | dp_netdev_del_rxq_from_pmd(struct dp_netdev_pmd_thread *pmd, |
4926 | struct rxq_poll *poll) | |
d0cca6c3 | 4927 | OVS_REQUIRES(pmd->port_mutex) |
ae7ad0a1 | 4928 | { |
e32971b8 DDP |
4929 | hmap_remove(&pmd->poll_list, &poll->node); |
4930 | free(poll); | |
ae7ad0a1 | 4931 | |
e32971b8 | 4932 | pmd->need_reload = true; |
ae7ad0a1 IM |
4933 | } |
4934 | ||
d0cca6c3 DDP |
4935 | /* Add 'port' to the tx port cache of 'pmd', which must be reloaded for the |
4936 | * changes to take effect. */ | |
cc245ce8 | 4937 | static void |
d0cca6c3 DDP |
4938 | dp_netdev_add_port_tx_to_pmd(struct dp_netdev_pmd_thread *pmd, |
4939 | struct dp_netdev_port *port) | |
e32971b8 | 4940 | OVS_REQUIRES(pmd->port_mutex) |
d0cca6c3 | 4941 | { |
57eebbb4 DDP |
4942 | struct tx_port *tx; |
4943 | ||
e32971b8 DDP |
4944 | tx = tx_port_lookup(&pmd->tx_ports, port->port_no); |
4945 | if (tx) { | |
4946 | /* 'port' is already on this thread tx cache. Do nothing. */ | |
4947 | return; | |
4948 | } | |
4949 | ||
57eebbb4 | 4950 | tx = xzalloc(sizeof *tx); |
d0cca6c3 | 4951 | |
324c8374 IM |
4952 | tx->port = port; |
4953 | tx->qid = -1; | |
c71ea3c4 | 4954 | tx->flush_time = 0LL; |
009e0033 | 4955 | dp_packet_batch_init(&tx->output_pkts); |
d0cca6c3 | 4956 | |
324c8374 | 4957 | hmap_insert(&pmd->tx_ports, &tx->node, hash_port_no(tx->port->port_no)); |
e32971b8 | 4958 | pmd->need_reload = true; |
d0cca6c3 DDP |
4959 | } |
4960 | ||
e32971b8 DDP |
4961 | /* Del 'tx' from the tx port cache of 'pmd', which must be reloaded for the |
4962 | * changes to take effect. */ | |
b9584f21 | 4963 | static void |
e32971b8 DDP |
4964 | dp_netdev_del_port_tx_from_pmd(struct dp_netdev_pmd_thread *pmd, |
4965 | struct tx_port *tx) | |
4966 | OVS_REQUIRES(pmd->port_mutex) | |
b9584f21 | 4967 | { |
e32971b8 DDP |
4968 | hmap_remove(&pmd->tx_ports, &tx->node); |
4969 | free(tx); | |
4970 | pmd->need_reload = true; | |
6c3eee82 BP |
4971 | } |
4972 | \f | |
b5cbbcf6 AZ |
4973 | static char * |
4974 | dpif_netdev_get_datapath_version(void) | |
4975 | { | |
4976 | return xstrdup("<built-in>"); | |
4977 | } | |
4978 | ||
72865317 | 4979 | static void |
1c1e46ed | 4980 | dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow, int cnt, int size, |
11bfdadd | 4981 | uint16_t tcp_flags, long long now) |
72865317 | 4982 | { |
eb94da30 | 4983 | uint16_t flags; |
72865317 | 4984 | |
eb94da30 DDP |
4985 | atomic_store_relaxed(&netdev_flow->stats.used, now); |
4986 | non_atomic_ullong_add(&netdev_flow->stats.packet_count, cnt); | |
4987 | non_atomic_ullong_add(&netdev_flow->stats.byte_count, size); | |
4988 | atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags); | |
4989 | flags |= tcp_flags; | |
4990 | atomic_store_relaxed(&netdev_flow->stats.tcp_flags, flags); | |
51852a57 BP |
4991 | } |
4992 | ||
623540e4 | 4993 | static int |
e14deea0 | 4994 | dp_netdev_upcall(struct dp_netdev_pmd_thread *pmd, struct dp_packet *packet_, |
7af12bd7 | 4995 | struct flow *flow, struct flow_wildcards *wc, ovs_u128 *ufid, |
623540e4 EJ |
4996 | enum dpif_upcall_type type, const struct nlattr *userdata, |
4997 | struct ofpbuf *actions, struct ofpbuf *put_actions) | |
4998 | { | |
1c1e46ed | 4999 | struct dp_netdev *dp = pmd->dp; |
623540e4 | 5000 | |
623540e4 EJ |
5001 | if (OVS_UNLIKELY(!dp->upcall_cb)) { |
5002 | return ENODEV; | |
5003 | } | |
5004 | ||
5005 | if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl))) { | |
5006 | struct ds ds = DS_EMPTY_INITIALIZER; | |
623540e4 | 5007 | char *packet_str; |
cf62fa4c | 5008 | struct ofpbuf key; |
5262eea1 JG |
5009 | struct odp_flow_key_parms odp_parms = { |
5010 | .flow = flow, | |
1dea1435 | 5011 | .mask = wc ? &wc->masks : NULL, |
2494ccd7 | 5012 | .support = dp_netdev_support, |
5262eea1 | 5013 | }; |
623540e4 EJ |
5014 | |
5015 | ofpbuf_init(&key, 0); | |
5262eea1 | 5016 | odp_flow_key_from_flow(&odp_parms, &key); |
2482b0b0 | 5017 | packet_str = ofp_dp_packet_to_string(packet_); |
623540e4 | 5018 | |
6fd6ed71 | 5019 | odp_flow_key_format(key.data, key.size, &ds); |
623540e4 EJ |
5020 | |
5021 | VLOG_DBG("%s: %s upcall:\n%s\n%s", dp->name, | |
5022 | dpif_upcall_type_to_string(type), ds_cstr(&ds), packet_str); | |
5023 | ||
5024 | ofpbuf_uninit(&key); | |
5025 | free(packet_str); | |
6fd6ed71 | 5026 | |
623540e4 EJ |
5027 | ds_destroy(&ds); |
5028 | } | |
5029 | ||
8d8ab6c2 JG |
5030 | return dp->upcall_cb(packet_, flow, ufid, pmd->core_id, type, userdata, |
5031 | actions, wc, put_actions, dp->upcall_aux); | |
623540e4 EJ |
5032 | } |
5033 | ||
bde94613 FA |
5034 | static inline uint32_t |
5035 | dpif_netdev_packet_get_rss_hash_orig_pkt(struct dp_packet *packet, | |
5036 | const struct miniflow *mf) | |
5037 | { | |
5038 | uint32_t hash; | |
5039 | ||
5040 | if (OVS_LIKELY(dp_packet_rss_valid(packet))) { | |
5041 | hash = dp_packet_get_rss_hash(packet); | |
5042 | } else { | |
5043 | hash = miniflow_hash_5tuple(mf, 0); | |
5044 | dp_packet_set_rss_hash(packet, hash); | |
5045 | } | |
5046 | ||
5047 | return hash; | |
5048 | } | |
5049 | ||
9bbf1c3d | 5050 | static inline uint32_t |
048963aa DDP |
5051 | dpif_netdev_packet_get_rss_hash(struct dp_packet *packet, |
5052 | const struct miniflow *mf) | |
9bbf1c3d | 5053 | { |
048963aa | 5054 | uint32_t hash, recirc_depth; |
9bbf1c3d | 5055 | |
f2f44f5d DDP |
5056 | if (OVS_LIKELY(dp_packet_rss_valid(packet))) { |
5057 | hash = dp_packet_get_rss_hash(packet); | |
5058 | } else { | |
9bbf1c3d | 5059 | hash = miniflow_hash_5tuple(mf, 0); |
2bc1bbd2 | 5060 | dp_packet_set_rss_hash(packet, hash); |
9bbf1c3d | 5061 | } |
048963aa DDP |
5062 | |
5063 | /* The RSS hash must account for the recirculation depth to avoid | |
5064 | * collisions in the exact match cache */ | |
5065 | recirc_depth = *recirc_depth_get_unsafe(); | |
5066 | if (OVS_UNLIKELY(recirc_depth)) { | |
5067 | hash = hash_finish(hash, recirc_depth); | |
5068 | dp_packet_set_rss_hash(packet, hash); | |
5069 | } | |
9bbf1c3d DDP |
5070 | return hash; |
5071 | } | |
5072 | ||
f7ce4811 | 5073 | struct packet_batch_per_flow { |
8cbf4f47 DDP |
5074 | unsigned int byte_count; |
5075 | uint16_t tcp_flags; | |
8cbf4f47 DDP |
5076 | struct dp_netdev_flow *flow; |
5077 | ||
1895cc8d | 5078 | struct dp_packet_batch array; |
8cbf4f47 DDP |
5079 | }; |
5080 | ||
5081 | static inline void | |
f7ce4811 PS |
5082 | packet_batch_per_flow_update(struct packet_batch_per_flow *batch, |
5083 | struct dp_packet *packet, | |
5084 | const struct miniflow *mf) | |
8cbf4f47 | 5085 | { |
cf62fa4c | 5086 | batch->byte_count += dp_packet_size(packet); |
1895cc8d PS |
5087 | batch->tcp_flags |= miniflow_get_tcp_flags(mf); |
5088 | batch->array.packets[batch->array.count++] = packet; | |
8cbf4f47 DDP |
5089 | } |
5090 | ||
5091 | static inline void | |
f7ce4811 PS |
5092 | packet_batch_per_flow_init(struct packet_batch_per_flow *batch, |
5093 | struct dp_netdev_flow *flow) | |
8cbf4f47 | 5094 | { |
11e5cf1f | 5095 | flow->batch = batch; |
8cbf4f47 | 5096 | |
11e5cf1f | 5097 | batch->flow = flow; |
1895cc8d | 5098 | dp_packet_batch_init(&batch->array); |
8cbf4f47 DDP |
5099 | batch->byte_count = 0; |
5100 | batch->tcp_flags = 0; | |
8cbf4f47 DDP |
5101 | } |
5102 | ||
5103 | static inline void | |
f7ce4811 | 5104 | packet_batch_per_flow_execute(struct packet_batch_per_flow *batch, |
b010be17 | 5105 | struct dp_netdev_pmd_thread *pmd) |
8cbf4f47 DDP |
5106 | { |
5107 | struct dp_netdev_actions *actions; | |
5108 | struct dp_netdev_flow *flow = batch->flow; | |
5109 | ||
1895cc8d | 5110 | dp_netdev_flow_used(flow, batch->array.count, batch->byte_count, |
05f9e707 | 5111 | batch->tcp_flags, pmd->ctx.now / 1000); |
8cbf4f47 DDP |
5112 | |
5113 | actions = dp_netdev_flow_get_actions(flow); | |
5114 | ||
66e4ad8a | 5115 | dp_netdev_execute_actions(pmd, &batch->array, true, &flow->flow, |
b010be17 | 5116 | actions->actions, actions->size); |
8cbf4f47 DDP |
5117 | } |
5118 | ||
8aaa125d | 5119 | static inline void |
e14deea0 | 5120 | dp_netdev_queue_batches(struct dp_packet *pkt, |
9bbf1c3d | 5121 | struct dp_netdev_flow *flow, const struct miniflow *mf, |
47a45d86 KT |
5122 | struct packet_batch_per_flow *batches, |
5123 | size_t *n_batches) | |
9bbf1c3d | 5124 | { |
f7ce4811 | 5125 | struct packet_batch_per_flow *batch = flow->batch; |
11e5cf1f | 5126 | |
f9fe365b AZ |
5127 | if (OVS_UNLIKELY(!batch)) { |
5128 | batch = &batches[(*n_batches)++]; | |
f7ce4811 | 5129 | packet_batch_per_flow_init(batch, flow); |
9bbf1c3d DDP |
5130 | } |
5131 | ||
f7ce4811 | 5132 | packet_batch_per_flow_update(batch, pkt, mf); |
9bbf1c3d DDP |
5133 | } |
5134 | ||
9bbf1c3d | 5135 | /* Try to process all ('cnt') the 'packets' using only the exact match cache |
a90ed026 | 5136 | * 'pmd->flow_cache'. If a flow is not found for a packet 'packets[i]', the |
8aaa125d DDP |
5137 | * miniflow is copied into 'keys' and the packet pointer is moved at the |
5138 | * beginning of the 'packets' array. | |
9bbf1c3d DDP |
5139 | * |
5140 | * The function returns the number of packets that needs to be processed in the | |
5141 | * 'packets' array (they have been moved to the beginning of the vector). | |
a90ed026 | 5142 | * |
02305520 FA |
5143 | * For performance reasons a caller may choose not to initialize the metadata |
5144 | * in 'packets_'. If 'md_is_valid' is false, the metadata in 'packets' | |
5145 | * is not valid and must be initialized by this function using 'port_no'. | |
5146 | * If 'md_is_valid' is true, the metadata is already valid and 'port_no' | |
5147 | * will be ignored. | |
9bbf1c3d DDP |
5148 | */ |
5149 | static inline size_t | |
72c84bc2 AZ |
5150 | emc_processing(struct dp_netdev_pmd_thread *pmd, |
5151 | struct dp_packet_batch *packets_, | |
1895cc8d | 5152 | struct netdev_flow_key *keys, |
f7ce4811 | 5153 | struct packet_batch_per_flow batches[], size_t *n_batches, |
a90ed026 | 5154 | bool md_is_valid, odp_port_t port_no) |
72865317 | 5155 | { |
65f13b50 | 5156 | struct emc_cache *flow_cache = &pmd->flow_cache; |
b89c678b | 5157 | struct netdev_flow_key *key = &keys[0]; |
72c84bc2 AZ |
5158 | size_t n_missed = 0, n_dropped = 0; |
5159 | struct dp_packet *packet; | |
45df9fef | 5160 | const size_t cnt = dp_packet_batch_size(packets_); |
f79b1ddb | 5161 | uint32_t cur_min; |
72c84bc2 | 5162 | int i; |
8cbf4f47 | 5163 | |
f79b1ddb | 5164 | atomic_read_relaxed(&pmd->dp->emc_insert_min, &cur_min); |
82a48ead JS |
5165 | pmd_perf_update_counter(&pmd->perf_stats, |
5166 | md_is_valid ? PMD_STAT_RECIRC : PMD_STAT_RECV, | |
5167 | cnt); | |
f79b1ddb | 5168 | |
45df9fef | 5169 | DP_PACKET_BATCH_REFILL_FOR_EACH (i, cnt, packet, packets_) { |
9bbf1c3d | 5170 | struct dp_netdev_flow *flow; |
9bbf1c3d | 5171 | |
5a2fed48 AZ |
5172 | if (OVS_UNLIKELY(dp_packet_size(packet) < ETH_HEADER_LEN)) { |
5173 | dp_packet_delete(packet); | |
3d88a620 | 5174 | n_dropped++; |
84d6d5eb EJ |
5175 | continue; |
5176 | } | |
8cbf4f47 | 5177 | |
45df9fef | 5178 | if (i != cnt - 1) { |
72c84bc2 | 5179 | struct dp_packet **packets = packets_->packets; |
a90ed026 | 5180 | /* Prefetch next packet data and metadata. */ |
72a5e2b8 | 5181 | OVS_PREFETCH(dp_packet_data(packets[i+1])); |
a90ed026 | 5182 | pkt_metadata_prefetch_init(&packets[i+1]->md); |
72a5e2b8 DDP |
5183 | } |
5184 | ||
a90ed026 DDP |
5185 | if (!md_is_valid) { |
5186 | pkt_metadata_init(&packet->md, port_no); | |
5187 | } | |
5a2fed48 | 5188 | miniflow_extract(packet, &key->mf); |
d262ac2c | 5189 | key->len = 0; /* Not computed yet. */ |
bde94613 FA |
5190 | /* If EMC is disabled skip hash computation and emc_lookup */ |
5191 | if (cur_min) { | |
5192 | if (!md_is_valid) { | |
5193 | key->hash = dpif_netdev_packet_get_rss_hash_orig_pkt(packet, | |
5194 | &key->mf); | |
5195 | } else { | |
5196 | key->hash = dpif_netdev_packet_get_rss_hash(packet, &key->mf); | |
5197 | } | |
5198 | flow = emc_lookup(flow_cache, key); | |
5199 | } else { | |
5200 | flow = NULL; | |
5201 | } | |
8aaa125d | 5202 | if (OVS_LIKELY(flow)) { |
5a2fed48 | 5203 | dp_netdev_queue_batches(packet, flow, &key->mf, batches, |
8aaa125d DDP |
5204 | n_batches); |
5205 | } else { | |
d1aa0b94 | 5206 | /* Exact match cache missed. Group missed packets together at |
72c84bc2 AZ |
5207 | * the beginning of the 'packets' array. */ |
5208 | dp_packet_batch_refill(packets_, packet, i); | |
400486f7 DDP |
5209 | /* 'key[n_missed]' contains the key of the current packet and it |
5210 | * must be returned to the caller. The next key should be extracted | |
5211 | * to 'keys[n_missed + 1]'. */ | |
5212 | key = &keys[++n_missed]; | |
9bbf1c3d DDP |
5213 | } |
5214 | } | |
5215 | ||
82a48ead JS |
5216 | pmd_perf_update_counter(&pmd->perf_stats, PMD_STAT_EXACT_HIT, |
5217 | cnt - n_dropped - n_missed); | |
4f150744 | 5218 | |
72c84bc2 | 5219 | return dp_packet_batch_size(packets_); |
9bbf1c3d DDP |
5220 | } |
5221 | ||
82a48ead | 5222 | static inline int |
47a45d86 KT |
5223 | handle_packet_upcall(struct dp_netdev_pmd_thread *pmd, |
5224 | struct dp_packet *packet, | |
a260d966 | 5225 | const struct netdev_flow_key *key, |
82a48ead | 5226 | struct ofpbuf *actions, struct ofpbuf *put_actions) |
a260d966 PS |
5227 | { |
5228 | struct ofpbuf *add_actions; | |
5229 | struct dp_packet_batch b; | |
5230 | struct match match; | |
5231 | ovs_u128 ufid; | |
5232 | int error; | |
79f36875 | 5233 | uint64_t cycles = cycles_counter_update(&pmd->perf_stats); |
a260d966 PS |
5234 | |
5235 | match.tun_md.valid = false; | |
5236 | miniflow_expand(&key->mf, &match.flow); | |
5237 | ||
5238 | ofpbuf_clear(actions); | |
5239 | ofpbuf_clear(put_actions); | |
5240 | ||
5241 | dpif_flow_hash(pmd->dp->dpif, &match.flow, sizeof match.flow, &ufid); | |
5242 | error = dp_netdev_upcall(pmd, packet, &match.flow, &match.wc, | |
5243 | &ufid, DPIF_UC_MISS, NULL, actions, | |
5244 | put_actions); | |
5245 | if (OVS_UNLIKELY(error && error != ENOSPC)) { | |
5246 | dp_packet_delete(packet); | |
82a48ead | 5247 | return error; |
a260d966 PS |
5248 | } |
5249 | ||
5250 | /* The Netlink encoding of datapath flow keys cannot express | |
5251 | * wildcarding the presence of a VLAN tag. Instead, a missing VLAN | |
5252 | * tag is interpreted as exact match on the fact that there is no | |
5253 | * VLAN. Unless we refactor a lot of code that translates between | |
5254 | * Netlink and struct flow representations, we have to do the same | |
5255 | * here. */ | |
f0fb825a EG |
5256 | if (!match.wc.masks.vlans[0].tci) { |
5257 | match.wc.masks.vlans[0].tci = htons(0xffff); | |
a260d966 PS |
5258 | } |
5259 | ||
5260 | /* We can't allow the packet batching in the next loop to execute | |
5261 | * the actions. Otherwise, if there are any slow path actions, | |
5262 | * we'll send the packet up twice. */ | |
72c84bc2 | 5263 | dp_packet_batch_init_packet(&b, packet); |
66e4ad8a | 5264 | dp_netdev_execute_actions(pmd, &b, true, &match.flow, |
b010be17 | 5265 | actions->data, actions->size); |
a260d966 PS |
5266 | |
5267 | add_actions = put_actions->size ? put_actions : actions; | |
5268 | if (OVS_LIKELY(error != ENOSPC)) { | |
5269 | struct dp_netdev_flow *netdev_flow; | |
5270 | ||
5271 | /* XXX: There's a race window where a flow covering this packet | |
5272 | * could have already been installed since we last did the flow | |
5273 | * lookup before upcall. This could be solved by moving the | |
5274 | * mutex lock outside the loop, but that's an awful long time | |
5275 | * to be locking everyone out of making flow installs. If we | |
5276 | * move to a per-core classifier, it would be reasonable. */ | |
5277 | ovs_mutex_lock(&pmd->flow_mutex); | |
3453b4d6 | 5278 | netdev_flow = dp_netdev_pmd_lookup_flow(pmd, key, NULL); |
a260d966 PS |
5279 | if (OVS_LIKELY(!netdev_flow)) { |
5280 | netdev_flow = dp_netdev_flow_add(pmd, &match, &ufid, | |
5281 | add_actions->data, | |
5282 | add_actions->size); | |
5283 | } | |
5284 | ovs_mutex_unlock(&pmd->flow_mutex); | |
4c30b246 | 5285 | emc_probabilistic_insert(pmd, key, netdev_flow); |
a260d966 | 5286 | } |
79f36875 JS |
5287 | if (pmd_perf_metrics_enabled(pmd)) { |
5288 | /* Update upcall stats. */ | |
5289 | cycles = cycles_counter_update(&pmd->perf_stats) - cycles; | |
5290 | struct pmd_perf_stats *s = &pmd->perf_stats; | |
5291 | s->current.upcalls++; | |
5292 | s->current.upcall_cycles += cycles; | |
5293 | histogram_add_sample(&s->cycles_per_upcall, cycles); | |
5294 | } | |
82a48ead | 5295 | return error; |
a260d966 PS |
5296 | } |
5297 | ||
9bbf1c3d | 5298 | static inline void |
65f13b50 | 5299 | fast_path_processing(struct dp_netdev_pmd_thread *pmd, |
1895cc8d | 5300 | struct dp_packet_batch *packets_, |
8aaa125d | 5301 | struct netdev_flow_key *keys, |
b010be17 IM |
5302 | struct packet_batch_per_flow batches[], |
5303 | size_t *n_batches, | |
5304 | odp_port_t in_port) | |
9bbf1c3d | 5305 | { |
31c82130 | 5306 | const size_t cnt = dp_packet_batch_size(packets_); |
1a0d5831 | 5307 | #if !defined(__CHECKER__) && !defined(_WIN32) |
9bbf1c3d DDP |
5308 | const size_t PKT_ARRAY_SIZE = cnt; |
5309 | #else | |
1a0d5831 | 5310 | /* Sparse or MSVC doesn't like variable length array. */ |
cd159f1a | 5311 | enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST }; |
9bbf1c3d | 5312 | #endif |
31c82130 | 5313 | struct dp_packet *packet; |
3453b4d6 | 5314 | struct dpcls *cls; |
0de8783a | 5315 | struct dpcls_rule *rules[PKT_ARRAY_SIZE]; |
65f13b50 | 5316 | struct dp_netdev *dp = pmd->dp; |
82a48ead | 5317 | int upcall_ok_cnt = 0, upcall_fail_cnt = 0; |
3453b4d6 | 5318 | int lookup_cnt = 0, add_lookup_cnt; |
9bbf1c3d DDP |
5319 | bool any_miss; |
5320 | ||
e883448e | 5321 | for (size_t i = 0; i < cnt; i++) { |
0de8783a | 5322 | /* Key length is needed in all the cases, hash computed on demand. */ |
361d808d | 5323 | keys[i].len = netdev_flow_key_size(miniflow_n_values(&keys[i].mf)); |
9bbf1c3d | 5324 | } |
3453b4d6 JS |
5325 | /* Get the classifier for the in_port */ |
5326 | cls = dp_netdev_pmd_lookup_dpcls(pmd, in_port); | |
5327 | if (OVS_LIKELY(cls)) { | |
5328 | any_miss = !dpcls_lookup(cls, keys, rules, cnt, &lookup_cnt); | |
5329 | } else { | |
5330 | any_miss = true; | |
5331 | memset(rules, 0, sizeof(rules)); | |
5332 | } | |
623540e4 EJ |
5333 | if (OVS_UNLIKELY(any_miss) && !fat_rwlock_tryrdlock(&dp->upcall_rwlock)) { |
5334 | uint64_t actions_stub[512 / 8], slow_stub[512 / 8]; | |
5335 | struct ofpbuf actions, put_actions; | |
623540e4 EJ |
5336 | |
5337 | ofpbuf_use_stub(&actions, actions_stub, sizeof actions_stub); | |
5338 | ofpbuf_use_stub(&put_actions, slow_stub, sizeof slow_stub); | |
5339 | ||
e883448e | 5340 | DP_PACKET_BATCH_FOR_EACH (i, packet, packets_) { |
0de8783a | 5341 | struct dp_netdev_flow *netdev_flow; |
623540e4 | 5342 | |
0de8783a | 5343 | if (OVS_LIKELY(rules[i])) { |
623540e4 EJ |
5344 | continue; |
5345 | } | |
5346 | ||
5347 | /* It's possible that an earlier slow path execution installed | |
0de8783a | 5348 | * a rule covering this flow. In this case, it's a lot cheaper |
623540e4 | 5349 | * to catch it here than execute a miss. */ |
3453b4d6 JS |
5350 | netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i], |
5351 | &add_lookup_cnt); | |
623540e4 | 5352 | if (netdev_flow) { |
3453b4d6 | 5353 | lookup_cnt += add_lookup_cnt; |
0de8783a | 5354 | rules[i] = &netdev_flow->cr; |
623540e4 EJ |
5355 | continue; |
5356 | } | |
5357 | ||
82a48ead JS |
5358 | int error = handle_packet_upcall(pmd, packet, &keys[i], |
5359 | &actions, &put_actions); | |
5360 | ||
5361 | if (OVS_UNLIKELY(error)) { | |
5362 | upcall_fail_cnt++; | |
5363 | } else { | |
5364 | upcall_ok_cnt++; | |
5365 | } | |
623540e4 EJ |
5366 | } |
5367 | ||
5368 | ofpbuf_uninit(&actions); | |
5369 | ofpbuf_uninit(&put_actions); | |
5370 | fat_rwlock_unlock(&dp->upcall_rwlock); | |
ac8c2081 | 5371 | } else if (OVS_UNLIKELY(any_miss)) { |
e883448e | 5372 | DP_PACKET_BATCH_FOR_EACH (i, packet, packets_) { |
0de8783a | 5373 | if (OVS_UNLIKELY(!rules[i])) { |
31c82130 | 5374 | dp_packet_delete(packet); |
82a48ead | 5375 | upcall_fail_cnt++; |
ac8c2081 DDP |
5376 | } |
5377 | } | |
623540e4 | 5378 | } |
84d6d5eb | 5379 | |
e883448e | 5380 | DP_PACKET_BATCH_FOR_EACH (i, packet, packets_) { |
84d6d5eb | 5381 | struct dp_netdev_flow *flow; |
8cbf4f47 | 5382 | |
0de8783a | 5383 | if (OVS_UNLIKELY(!rules[i])) { |
84d6d5eb EJ |
5384 | continue; |
5385 | } | |
5386 | ||
84d6d5eb | 5387 | flow = dp_netdev_flow_cast(rules[i]); |
0de8783a | 5388 | |
4c30b246 | 5389 | emc_probabilistic_insert(pmd, &keys[i], flow); |
8aaa125d | 5390 | dp_netdev_queue_batches(packet, flow, &keys[i].mf, batches, n_batches); |
8cbf4f47 DDP |
5391 | } |
5392 | ||
82a48ead JS |
5393 | pmd_perf_update_counter(&pmd->perf_stats, PMD_STAT_MASKED_HIT, |
5394 | cnt - upcall_ok_cnt - upcall_fail_cnt); | |
5395 | pmd_perf_update_counter(&pmd->perf_stats, PMD_STAT_MASKED_LOOKUP, | |
5396 | lookup_cnt); | |
5397 | pmd_perf_update_counter(&pmd->perf_stats, PMD_STAT_MISS, | |
5398 | upcall_ok_cnt); | |
5399 | pmd_perf_update_counter(&pmd->perf_stats, PMD_STAT_LOST, | |
5400 | upcall_fail_cnt); | |
72865317 BP |
5401 | } |
5402 | ||
a90ed026 DDP |
5403 | /* Packets enter the datapath from a port (or from recirculation) here. |
5404 | * | |
02305520 FA |
5405 | * When 'md_is_valid' is true the metadata in 'packets' are already valid. |
5406 | * When false the metadata in 'packets' need to be initialized. */ | |
adcf00ba | 5407 | static void |
a90ed026 | 5408 | dp_netdev_input__(struct dp_netdev_pmd_thread *pmd, |
1895cc8d | 5409 | struct dp_packet_batch *packets, |
a90ed026 | 5410 | bool md_is_valid, odp_port_t port_no) |
9bbf1c3d | 5411 | { |
1a0d5831 | 5412 | #if !defined(__CHECKER__) && !defined(_WIN32) |
37eabc70 | 5413 | const size_t PKT_ARRAY_SIZE = dp_packet_batch_size(packets); |
9bbf1c3d | 5414 | #else |
1a0d5831 | 5415 | /* Sparse or MSVC doesn't like variable length array. */ |
cd159f1a | 5416 | enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST }; |
9bbf1c3d | 5417 | #endif |
47a45d86 KT |
5418 | OVS_ALIGNED_VAR(CACHE_LINE_SIZE) |
5419 | struct netdev_flow_key keys[PKT_ARRAY_SIZE]; | |
f7ce4811 | 5420 | struct packet_batch_per_flow batches[PKT_ARRAY_SIZE]; |
72c84bc2 | 5421 | size_t n_batches; |
3453b4d6 | 5422 | odp_port_t in_port; |
9bbf1c3d | 5423 | |
8aaa125d | 5424 | n_batches = 0; |
72c84bc2 | 5425 | emc_processing(pmd, packets, keys, batches, &n_batches, |
a90ed026 | 5426 | md_is_valid, port_no); |
72c84bc2 | 5427 | if (!dp_packet_batch_is_empty(packets)) { |
3453b4d6 JS |
5428 | /* Get ingress port from first packet's metadata. */ |
5429 | in_port = packets->packets[0]->md.in_port.odp_port; | |
b010be17 IM |
5430 | fast_path_processing(pmd, packets, keys, |
5431 | batches, &n_batches, in_port); | |
8aaa125d DDP |
5432 | } |
5433 | ||
ad9f0581 BB |
5434 | /* All the flow batches need to be reset before any call to |
5435 | * packet_batch_per_flow_execute() as it could potentially trigger | |
5436 | * recirculation. When a packet matching flow ‘j’ happens to be | |
5437 | * recirculated, the nested call to dp_netdev_input__() could potentially | |
5438 | * classify the packet as matching another flow - say 'k'. It could happen | |
5439 | * that in the previous call to dp_netdev_input__() that same flow 'k' had | |
5440 | * already its own batches[k] still waiting to be served. So if its | |
5441 | * ‘batch’ member is not reset, the recirculated packet would be wrongly | |
5442 | * appended to batches[k] of the 1st call to dp_netdev_input__(). */ | |
72c84bc2 | 5443 | size_t i; |
603f2ce0 EJ |
5444 | for (i = 0; i < n_batches; i++) { |
5445 | batches[i].flow->batch = NULL; | |
5446 | } | |
5447 | ||
8aaa125d | 5448 | for (i = 0; i < n_batches; i++) { |
b010be17 | 5449 | packet_batch_per_flow_execute(&batches[i], pmd); |
9bbf1c3d DDP |
5450 | } |
5451 | } | |
5452 | ||
a90ed026 DDP |
5453 | static void |
5454 | dp_netdev_input(struct dp_netdev_pmd_thread *pmd, | |
1895cc8d | 5455 | struct dp_packet_batch *packets, |
a90ed026 DDP |
5456 | odp_port_t port_no) |
5457 | { | |
3453b4d6 | 5458 | dp_netdev_input__(pmd, packets, false, port_no); |
a90ed026 DDP |
5459 | } |
5460 | ||
5461 | static void | |
5462 | dp_netdev_recirculate(struct dp_netdev_pmd_thread *pmd, | |
1895cc8d | 5463 | struct dp_packet_batch *packets) |
a90ed026 | 5464 | { |
3453b4d6 | 5465 | dp_netdev_input__(pmd, packets, true, 0); |
a90ed026 DDP |
5466 | } |
5467 | ||
9080a111 | 5468 | struct dp_netdev_execute_aux { |
65f13b50 | 5469 | struct dp_netdev_pmd_thread *pmd; |
66e4ad8a | 5470 | const struct flow *flow; |
9080a111 JR |
5471 | }; |
5472 | ||
e4e74c3a AW |
5473 | static void |
5474 | dpif_netdev_register_dp_purge_cb(struct dpif *dpif, dp_purge_callback *cb, | |
5475 | void *aux) | |
5476 | { | |
5477 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
5478 | dp->dp_purge_aux = aux; | |
5479 | dp->dp_purge_cb = cb; | |
5480 | } | |
5481 | ||
6b31e073 | 5482 | static void |
623540e4 EJ |
5483 | dpif_netdev_register_upcall_cb(struct dpif *dpif, upcall_callback *cb, |
5484 | void *aux) | |
6b31e073 RW |
5485 | { |
5486 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
623540e4 | 5487 | dp->upcall_aux = aux; |
6b31e073 RW |
5488 | dp->upcall_cb = cb; |
5489 | } | |
5490 | ||
324c8374 IM |
5491 | static void |
5492 | dpif_netdev_xps_revalidate_pmd(const struct dp_netdev_pmd_thread *pmd, | |
b010be17 | 5493 | bool purge) |
324c8374 IM |
5494 | { |
5495 | struct tx_port *tx; | |
5496 | struct dp_netdev_port *port; | |
5497 | long long interval; | |
5498 | ||
57eebbb4 | 5499 | HMAP_FOR_EACH (tx, node, &pmd->send_port_cache) { |
9f7a3035 | 5500 | if (!tx->port->dynamic_txqs) { |
324c8374 IM |
5501 | continue; |
5502 | } | |
b010be17 | 5503 | interval = pmd->ctx.now - tx->last_used; |
05f9e707 | 5504 | if (tx->qid >= 0 && (purge || interval >= XPS_TIMEOUT)) { |
324c8374 IM |
5505 | port = tx->port; |
5506 | ovs_mutex_lock(&port->txq_used_mutex); | |
5507 | port->txq_used[tx->qid]--; | |
5508 | ovs_mutex_unlock(&port->txq_used_mutex); | |
5509 | tx->qid = -1; | |
5510 | } | |
5511 | } | |
5512 | } | |
5513 | ||
5514 | static int | |
5515 | dpif_netdev_xps_get_tx_qid(const struct dp_netdev_pmd_thread *pmd, | |
b010be17 | 5516 | struct tx_port *tx) |
324c8374 IM |
5517 | { |
5518 | struct dp_netdev_port *port; | |
5519 | long long interval; | |
5520 | int i, min_cnt, min_qid; | |
5521 | ||
b010be17 IM |
5522 | interval = pmd->ctx.now - tx->last_used; |
5523 | tx->last_used = pmd->ctx.now; | |
324c8374 | 5524 | |
05f9e707 | 5525 | if (OVS_LIKELY(tx->qid >= 0 && interval < XPS_TIMEOUT)) { |
324c8374 IM |
5526 | return tx->qid; |
5527 | } | |
5528 | ||
5529 | port = tx->port; | |
5530 | ||
5531 | ovs_mutex_lock(&port->txq_used_mutex); | |
5532 | if (tx->qid >= 0) { | |
5533 | port->txq_used[tx->qid]--; | |
5534 | tx->qid = -1; | |
5535 | } | |
5536 | ||
5537 | min_cnt = -1; | |
5538 | min_qid = 0; | |
5539 | for (i = 0; i < netdev_n_txq(port->netdev); i++) { | |
5540 | if (port->txq_used[i] < min_cnt || min_cnt == -1) { | |
5541 | min_cnt = port->txq_used[i]; | |
5542 | min_qid = i; | |
5543 | } | |
5544 | } | |
5545 | ||
5546 | port->txq_used[min_qid]++; | |
5547 | tx->qid = min_qid; | |
5548 | ||
5549 | ovs_mutex_unlock(&port->txq_used_mutex); | |
5550 | ||
b010be17 | 5551 | dpif_netdev_xps_revalidate_pmd(pmd, false); |
324c8374 IM |
5552 | |
5553 | VLOG_DBG("Core %d: New TX queue ID %d for port \'%s\'.", | |
5554 | pmd->core_id, tx->qid, netdev_get_name(tx->port->netdev)); | |
5555 | return min_qid; | |
5556 | } | |
5557 | ||
d0cca6c3 | 5558 | static struct tx_port * |
57eebbb4 DDP |
5559 | pmd_tnl_port_cache_lookup(const struct dp_netdev_pmd_thread *pmd, |
5560 | odp_port_t port_no) | |
5561 | { | |
5562 | return tx_port_lookup(&pmd->tnl_port_cache, port_no); | |
5563 | } | |
5564 | ||
5565 | static struct tx_port * | |
5566 | pmd_send_port_cache_lookup(const struct dp_netdev_pmd_thread *pmd, | |
5567 | odp_port_t port_no) | |
d0cca6c3 | 5568 | { |
57eebbb4 | 5569 | return tx_port_lookup(&pmd->send_port_cache, port_no); |
d0cca6c3 DDP |
5570 | } |
5571 | ||
a36de779 | 5572 | static int |
d0cca6c3 | 5573 | push_tnl_action(const struct dp_netdev_pmd_thread *pmd, |
1895cc8d PS |
5574 | const struct nlattr *attr, |
5575 | struct dp_packet_batch *batch) | |
a36de779 | 5576 | { |
d0cca6c3 | 5577 | struct tx_port *tun_port; |
a36de779 | 5578 | const struct ovs_action_push_tnl *data; |
4c742796 | 5579 | int err; |
a36de779 PS |
5580 | |
5581 | data = nl_attr_get(attr); | |
5582 | ||
81765c00 | 5583 | tun_port = pmd_tnl_port_cache_lookup(pmd, data->tnl_port); |
a36de779 | 5584 | if (!tun_port) { |
4c742796 PS |
5585 | err = -EINVAL; |
5586 | goto error; | |
a36de779 | 5587 | } |
324c8374 | 5588 | err = netdev_push_header(tun_port->port->netdev, batch, data); |
4c742796 PS |
5589 | if (!err) { |
5590 | return 0; | |
5591 | } | |
5592 | error: | |
5593 | dp_packet_delete_batch(batch, true); | |
5594 | return err; | |
a36de779 PS |
5595 | } |
5596 | ||
66525ef3 PS |
5597 | static void |
5598 | dp_execute_userspace_action(struct dp_netdev_pmd_thread *pmd, | |
7d7ded7a | 5599 | struct dp_packet *packet, bool should_steal, |
66525ef3 PS |
5600 | struct flow *flow, ovs_u128 *ufid, |
5601 | struct ofpbuf *actions, | |
b010be17 | 5602 | const struct nlattr *userdata) |
66525ef3 PS |
5603 | { |
5604 | struct dp_packet_batch b; | |
5605 | int error; | |
5606 | ||
5607 | ofpbuf_clear(actions); | |
5608 | ||
5609 | error = dp_netdev_upcall(pmd, packet, flow, NULL, ufid, | |
5610 | DPIF_UC_ACTION, userdata, actions, | |
5611 | NULL); | |
5612 | if (!error || error == ENOSPC) { | |
72c84bc2 | 5613 | dp_packet_batch_init_packet(&b, packet); |
7d7ded7a | 5614 | dp_netdev_execute_actions(pmd, &b, should_steal, flow, |
b010be17 | 5615 | actions->data, actions->size); |
7d7ded7a | 5616 | } else if (should_steal) { |
66525ef3 PS |
5617 | dp_packet_delete(packet); |
5618 | } | |
5619 | } | |
5620 | ||
a36de779 | 5621 | static void |
1895cc8d | 5622 | dp_execute_cb(void *aux_, struct dp_packet_batch *packets_, |
7d7ded7a | 5623 | const struct nlattr *a, bool should_steal) |
4b27db64 | 5624 | OVS_NO_THREAD_SAFETY_ANALYSIS |
9080a111 JR |
5625 | { |
5626 | struct dp_netdev_execute_aux *aux = aux_; | |
623540e4 | 5627 | uint32_t *depth = recirc_depth_get(); |
28e2fa02 DDP |
5628 | struct dp_netdev_pmd_thread *pmd = aux->pmd; |
5629 | struct dp_netdev *dp = pmd->dp; | |
09f9da0b | 5630 | int type = nl_attr_type(a); |
d0cca6c3 | 5631 | struct tx_port *p; |
9080a111 | 5632 | |
09f9da0b JR |
5633 | switch ((enum ovs_action_attr)type) { |
5634 | case OVS_ACTION_ATTR_OUTPUT: | |
57eebbb4 | 5635 | p = pmd_send_port_cache_lookup(pmd, nl_attr_get_odp_port(a)); |
26a5075b | 5636 | if (OVS_LIKELY(p)) { |
009e0033 IM |
5637 | struct dp_packet *packet; |
5638 | struct dp_packet_batch out; | |
347ba9bb | 5639 | |
7d7ded7a | 5640 | if (!should_steal) { |
009e0033 IM |
5641 | dp_packet_batch_clone(&out, packets_); |
5642 | dp_packet_batch_reset_cutlen(packets_); | |
5643 | packets_ = &out; | |
324c8374 | 5644 | } |
009e0033 | 5645 | dp_packet_batch_apply_cutlen(packets_); |
347ba9bb | 5646 | |
009e0033 IM |
5647 | #ifdef DPDK_NETDEV |
5648 | if (OVS_UNLIKELY(!dp_packet_batch_is_empty(&p->output_pkts) | |
5649 | && packets_->packets[0]->source | |
5650 | != p->output_pkts.packets[0]->source)) { | |
5651 | /* XXX: netdev-dpdk assumes that all packets in a single | |
5652 | * output batch has the same source. Flush here to | |
5653 | * avoid memory access issues. */ | |
5654 | dp_netdev_pmd_flush_output_on_port(pmd, p); | |
5655 | } | |
5656 | #endif | |
c71ea3c4 IM |
5657 | if (dp_packet_batch_size(&p->output_pkts) |
5658 | + dp_packet_batch_size(packets_) > NETDEV_MAX_BURST) { | |
5659 | /* Flush here to avoid overflow. */ | |
009e0033 IM |
5660 | dp_netdev_pmd_flush_output_on_port(pmd, p); |
5661 | } | |
c71ea3c4 IM |
5662 | |
5663 | if (dp_packet_batch_is_empty(&p->output_pkts)) { | |
5664 | pmd->n_output_batches++; | |
5665 | } | |
5666 | ||
e883448e | 5667 | DP_PACKET_BATCH_FOR_EACH (i, packet, packets_) { |
58ed6df0 IM |
5668 | p->output_pkts_rxqs[dp_packet_batch_size(&p->output_pkts)] = |
5669 | pmd->ctx.last_rxq; | |
009e0033 IM |
5670 | dp_packet_batch_add(&p->output_pkts, packet); |
5671 | } | |
ac8c2081 | 5672 | return; |
8a4e3a85 | 5673 | } |
09f9da0b JR |
5674 | break; |
5675 | ||
a36de779 PS |
5676 | case OVS_ACTION_ATTR_TUNNEL_PUSH: |
5677 | if (*depth < MAX_RECIRC_DEPTH) { | |
aaca4fe0 | 5678 | dp_packet_batch_apply_cutlen(packets_); |
7c12dfc5 | 5679 | push_tnl_action(pmd, a, packets_); |
a36de779 PS |
5680 | return; |
5681 | } | |
5682 | break; | |
5683 | ||
5684 | case OVS_ACTION_ATTR_TUNNEL_POP: | |
5685 | if (*depth < MAX_RECIRC_DEPTH) { | |
aaca4fe0 | 5686 | struct dp_packet_batch *orig_packets_ = packets_; |
8611f9a4 | 5687 | odp_port_t portno = nl_attr_get_odp_port(a); |
a36de779 | 5688 | |
57eebbb4 | 5689 | p = pmd_tnl_port_cache_lookup(pmd, portno); |
a36de779 | 5690 | if (p) { |
1895cc8d | 5691 | struct dp_packet_batch tnl_pkt; |
a36de779 | 5692 | |
7d7ded7a | 5693 | if (!should_steal) { |
aaca4fe0 WT |
5694 | dp_packet_batch_clone(&tnl_pkt, packets_); |
5695 | packets_ = &tnl_pkt; | |
5696 | dp_packet_batch_reset_cutlen(orig_packets_); | |
a36de779 PS |
5697 | } |
5698 | ||
aaca4fe0 WT |
5699 | dp_packet_batch_apply_cutlen(packets_); |
5700 | ||
324c8374 | 5701 | netdev_pop_header(p->port->netdev, packets_); |
72c84bc2 | 5702 | if (dp_packet_batch_is_empty(packets_)) { |
1c8f98d9 PS |
5703 | return; |
5704 | } | |
9235b479 | 5705 | |
72c84bc2 | 5706 | struct dp_packet *packet; |
e883448e | 5707 | DP_PACKET_BATCH_FOR_EACH (i, packet, packets_) { |
72c84bc2 | 5708 | packet->md.in_port.odp_port = portno; |
a36de779 | 5709 | } |
9235b479 PS |
5710 | |
5711 | (*depth)++; | |
5712 | dp_netdev_recirculate(pmd, packets_); | |
5713 | (*depth)--; | |
a36de779 PS |
5714 | return; |
5715 | } | |
5716 | } | |
5717 | break; | |
5718 | ||
623540e4 EJ |
5719 | case OVS_ACTION_ATTR_USERSPACE: |
5720 | if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) { | |
aaca4fe0 | 5721 | struct dp_packet_batch *orig_packets_ = packets_; |
623540e4 | 5722 | const struct nlattr *userdata; |
aaca4fe0 | 5723 | struct dp_packet_batch usr_pkt; |
623540e4 EJ |
5724 | struct ofpbuf actions; |
5725 | struct flow flow; | |
7af12bd7 | 5726 | ovs_u128 ufid; |
aaca4fe0 | 5727 | bool clone = false; |
4fc65926 | 5728 | |
623540e4 EJ |
5729 | userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA); |
5730 | ofpbuf_init(&actions, 0); | |
8cbf4f47 | 5731 | |
aaca4fe0 | 5732 | if (packets_->trunc) { |
7d7ded7a | 5733 | if (!should_steal) { |
aaca4fe0 WT |
5734 | dp_packet_batch_clone(&usr_pkt, packets_); |
5735 | packets_ = &usr_pkt; | |
aaca4fe0 WT |
5736 | clone = true; |
5737 | dp_packet_batch_reset_cutlen(orig_packets_); | |
5738 | } | |
5739 | ||
5740 | dp_packet_batch_apply_cutlen(packets_); | |
5741 | } | |
5742 | ||
72c84bc2 | 5743 | struct dp_packet *packet; |
e883448e | 5744 | DP_PACKET_BATCH_FOR_EACH (i, packet, packets_) { |
72c84bc2 | 5745 | flow_extract(packet, &flow); |
7af12bd7 | 5746 | dpif_flow_hash(dp->dpif, &flow, sizeof flow, &ufid); |
7d7ded7a | 5747 | dp_execute_userspace_action(pmd, packet, should_steal, &flow, |
b010be17 | 5748 | &ufid, &actions, userdata); |
db73f716 | 5749 | } |
aaca4fe0 WT |
5750 | |
5751 | if (clone) { | |
5752 | dp_packet_delete_batch(packets_, true); | |
5753 | } | |
5754 | ||
623540e4 EJ |
5755 | ofpbuf_uninit(&actions); |
5756 | fat_rwlock_unlock(&dp->upcall_rwlock); | |
6b31e073 | 5757 | |
ac8c2081 DDP |
5758 | return; |
5759 | } | |
09f9da0b | 5760 | break; |
572f732a | 5761 | |
adcf00ba AZ |
5762 | case OVS_ACTION_ATTR_RECIRC: |
5763 | if (*depth < MAX_RECIRC_DEPTH) { | |
1895cc8d | 5764 | struct dp_packet_batch recirc_pkts; |
572f732a | 5765 | |
7d7ded7a | 5766 | if (!should_steal) { |
1895cc8d PS |
5767 | dp_packet_batch_clone(&recirc_pkts, packets_); |
5768 | packets_ = &recirc_pkts; | |
28e2fa02 | 5769 | } |
8cbf4f47 | 5770 | |
72c84bc2 | 5771 | struct dp_packet *packet; |
e883448e | 5772 | DP_PACKET_BATCH_FOR_EACH (i, packet, packets_) { |
72c84bc2 | 5773 | packet->md.recirc_id = nl_attr_get_u32(a); |
8cbf4f47 | 5774 | } |
28e2fa02 DDP |
5775 | |
5776 | (*depth)++; | |
1895cc8d | 5777 | dp_netdev_recirculate(pmd, packets_); |
adcf00ba AZ |
5778 | (*depth)--; |
5779 | ||
ac8c2081 | 5780 | return; |
adcf00ba | 5781 | } |
ac8c2081 DDP |
5782 | |
5783 | VLOG_WARN("Packet dropped. Max recirculation depth exceeded."); | |
572f732a | 5784 | break; |
572f732a | 5785 | |
5cf3edb3 DDP |
5786 | case OVS_ACTION_ATTR_CT: { |
5787 | const struct nlattr *b; | |
a76a37ef | 5788 | bool force = false; |
5cf3edb3 DDP |
5789 | bool commit = false; |
5790 | unsigned int left; | |
5791 | uint16_t zone = 0; | |
5792 | const char *helper = NULL; | |
5793 | const uint32_t *setmark = NULL; | |
5794 | const struct ovs_key_ct_labels *setlabel = NULL; | |
4cddb1f0 DB |
5795 | struct nat_action_info_t nat_action_info; |
5796 | struct nat_action_info_t *nat_action_info_ref = NULL; | |
5797 | bool nat_config = false; | |
5cf3edb3 DDP |
5798 | |
5799 | NL_ATTR_FOR_EACH_UNSAFE (b, left, nl_attr_get(a), | |
5800 | nl_attr_get_size(a)) { | |
5801 | enum ovs_ct_attr sub_type = nl_attr_type(b); | |
5802 | ||
5803 | switch(sub_type) { | |
b80e259f | 5804 | case OVS_CT_ATTR_FORCE_COMMIT: |
a76a37ef JR |
5805 | force = true; |
5806 | /* fall through. */ | |
5cf3edb3 DDP |
5807 | case OVS_CT_ATTR_COMMIT: |
5808 | commit = true; | |
5809 | break; | |
5810 | case OVS_CT_ATTR_ZONE: | |
5811 | zone = nl_attr_get_u16(b); | |
5812 | break; | |
5813 | case OVS_CT_ATTR_HELPER: | |
5814 | helper = nl_attr_get_string(b); | |
5815 | break; | |
5816 | case OVS_CT_ATTR_MARK: | |
5817 | setmark = nl_attr_get(b); | |
5818 | break; | |
5819 | case OVS_CT_ATTR_LABELS: | |
5820 | setlabel = nl_attr_get(b); | |
5821 | break; | |
8e83854c JR |
5822 | case OVS_CT_ATTR_EVENTMASK: |
5823 | /* Silently ignored, as userspace datapath does not generate | |
5824 | * netlink events. */ | |
5825 | break; | |
4cddb1f0 DB |
5826 | case OVS_CT_ATTR_NAT: { |
5827 | const struct nlattr *b_nest; | |
5828 | unsigned int left_nest; | |
5829 | bool ip_min_specified = false; | |
5830 | bool proto_num_min_specified = false; | |
5831 | bool ip_max_specified = false; | |
5832 | bool proto_num_max_specified = false; | |
5833 | memset(&nat_action_info, 0, sizeof nat_action_info); | |
5834 | nat_action_info_ref = &nat_action_info; | |
5835 | ||
5836 | NL_NESTED_FOR_EACH_UNSAFE (b_nest, left_nest, b) { | |
5837 | enum ovs_nat_attr sub_type_nest = nl_attr_type(b_nest); | |
5838 | ||
5839 | switch (sub_type_nest) { | |
5840 | case OVS_NAT_ATTR_SRC: | |
5841 | case OVS_NAT_ATTR_DST: | |
5842 | nat_config = true; | |
5843 | nat_action_info.nat_action |= | |
5844 | ((sub_type_nest == OVS_NAT_ATTR_SRC) | |
5845 | ? NAT_ACTION_SRC : NAT_ACTION_DST); | |
5846 | break; | |
5847 | case OVS_NAT_ATTR_IP_MIN: | |
5848 | memcpy(&nat_action_info.min_addr, | |
5849 | nl_attr_get(b_nest), | |
5850 | nl_attr_get_size(b_nest)); | |
5851 | ip_min_specified = true; | |
5852 | break; | |
5853 | case OVS_NAT_ATTR_IP_MAX: | |
5854 | memcpy(&nat_action_info.max_addr, | |
5855 | nl_attr_get(b_nest), | |
5856 | nl_attr_get_size(b_nest)); | |
5857 | ip_max_specified = true; | |
5858 | break; | |
5859 | case OVS_NAT_ATTR_PROTO_MIN: | |
5860 | nat_action_info.min_port = | |
5861 | nl_attr_get_u16(b_nest); | |
5862 | proto_num_min_specified = true; | |
5863 | break; | |
5864 | case OVS_NAT_ATTR_PROTO_MAX: | |
5865 | nat_action_info.max_port = | |
5866 | nl_attr_get_u16(b_nest); | |
5867 | proto_num_max_specified = true; | |
5868 | break; | |
5869 | case OVS_NAT_ATTR_PERSISTENT: | |
5870 | case OVS_NAT_ATTR_PROTO_HASH: | |
5871 | case OVS_NAT_ATTR_PROTO_RANDOM: | |
5872 | break; | |
5873 | case OVS_NAT_ATTR_UNSPEC: | |
5874 | case __OVS_NAT_ATTR_MAX: | |
5875 | OVS_NOT_REACHED(); | |
5876 | } | |
5877 | } | |
5878 | ||
5879 | if (ip_min_specified && !ip_max_specified) { | |
5880 | nat_action_info.max_addr = nat_action_info.min_addr; | |
5881 | } | |
5882 | if (proto_num_min_specified && !proto_num_max_specified) { | |
5883 | nat_action_info.max_port = nat_action_info.min_port; | |
5884 | } | |
5885 | if (proto_num_min_specified || proto_num_max_specified) { | |
5886 | if (nat_action_info.nat_action & NAT_ACTION_SRC) { | |
5887 | nat_action_info.nat_action |= NAT_ACTION_SRC_PORT; | |
5888 | } else if (nat_action_info.nat_action & NAT_ACTION_DST) { | |
5889 | nat_action_info.nat_action |= NAT_ACTION_DST_PORT; | |
5890 | } | |
5891 | } | |
5892 | break; | |
5893 | } | |
5cf3edb3 DDP |
5894 | case OVS_CT_ATTR_UNSPEC: |
5895 | case __OVS_CT_ATTR_MAX: | |
5896 | OVS_NOT_REACHED(); | |
5897 | } | |
5898 | } | |
5899 | ||
4cddb1f0 DB |
5900 | /* We won't be able to function properly in this case, hence |
5901 | * complain loudly. */ | |
5902 | if (nat_config && !commit) { | |
5903 | static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 5); | |
5904 | VLOG_WARN_RL(&rl, "NAT specified without commit."); | |
5905 | } | |
5906 | ||
a76a37ef | 5907 | conntrack_execute(&dp->conntrack, packets_, aux->flow->dl_type, force, |
bd7d93f8 | 5908 | commit, zone, setmark, setlabel, aux->flow->tp_src, |
b010be17 | 5909 | aux->flow->tp_dst, helper, nat_action_info_ref, |
05f9e707 | 5910 | pmd->ctx.now / 1000); |
07659514 | 5911 | break; |
5cf3edb3 | 5912 | } |
07659514 | 5913 | |
5dddf960 | 5914 | case OVS_ACTION_ATTR_METER: |
4b27db64 | 5915 | dp_netdev_run_meter(pmd->dp, packets_, nl_attr_get_u32(a), |
b010be17 | 5916 | pmd->ctx.now); |
4b27db64 JR |
5917 | break; |
5918 | ||
09f9da0b JR |
5919 | case OVS_ACTION_ATTR_PUSH_VLAN: |
5920 | case OVS_ACTION_ATTR_POP_VLAN: | |
5921 | case OVS_ACTION_ATTR_PUSH_MPLS: | |
5922 | case OVS_ACTION_ATTR_POP_MPLS: | |
5923 | case OVS_ACTION_ATTR_SET: | |
6d670e7f | 5924 | case OVS_ACTION_ATTR_SET_MASKED: |
09f9da0b | 5925 | case OVS_ACTION_ATTR_SAMPLE: |
53e1d6f1 | 5926 | case OVS_ACTION_ATTR_HASH: |
09f9da0b | 5927 | case OVS_ACTION_ATTR_UNSPEC: |
aaca4fe0 | 5928 | case OVS_ACTION_ATTR_TRUNC: |
6fcecb85 YY |
5929 | case OVS_ACTION_ATTR_PUSH_ETH: |
5930 | case OVS_ACTION_ATTR_POP_ETH: | |
535e3acf | 5931 | case OVS_ACTION_ATTR_CLONE: |
f59cb331 YY |
5932 | case OVS_ACTION_ATTR_PUSH_NSH: |
5933 | case OVS_ACTION_ATTR_POP_NSH: | |
1fe178d2 | 5934 | case OVS_ACTION_ATTR_CT_CLEAR: |
09f9da0b JR |
5935 | case __OVS_ACTION_ATTR_MAX: |
5936 | OVS_NOT_REACHED(); | |
da546e07 | 5937 | } |
ac8c2081 | 5938 | |
7d7ded7a | 5939 | dp_packet_delete_batch(packets_, should_steal); |
98403001 BP |
5940 | } |
5941 | ||
4edb9ae9 | 5942 | static void |
65f13b50 | 5943 | dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd, |
1895cc8d | 5944 | struct dp_packet_batch *packets, |
7d7ded7a | 5945 | bool should_steal, const struct flow *flow, |
b010be17 | 5946 | const struct nlattr *actions, size_t actions_len) |
72865317 | 5947 | { |
b010be17 | 5948 | struct dp_netdev_execute_aux aux = { pmd, flow }; |
9080a111 | 5949 | |
7d7ded7a | 5950 | odp_execute_actions(&aux, packets, should_steal, actions, |
8cbf4f47 | 5951 | actions_len, dp_execute_cb); |
72865317 BP |
5952 | } |
5953 | ||
4d4e68ed DDP |
5954 | struct dp_netdev_ct_dump { |
5955 | struct ct_dpif_dump_state up; | |
5956 | struct conntrack_dump dump; | |
5957 | struct conntrack *ct; | |
5958 | struct dp_netdev *dp; | |
5959 | }; | |
5960 | ||
5961 | static int | |
5962 | dpif_netdev_ct_dump_start(struct dpif *dpif, struct ct_dpif_dump_state **dump_, | |
ded30c74 | 5963 | const uint16_t *pzone, int *ptot_bkts) |
4d4e68ed DDP |
5964 | { |
5965 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
5966 | struct dp_netdev_ct_dump *dump; | |
5967 | ||
5968 | dump = xzalloc(sizeof *dump); | |
5969 | dump->dp = dp; | |
5970 | dump->ct = &dp->conntrack; | |
5971 | ||
ded30c74 | 5972 | conntrack_dump_start(&dp->conntrack, &dump->dump, pzone, ptot_bkts); |
4d4e68ed DDP |
5973 | |
5974 | *dump_ = &dump->up; | |
5975 | ||
5976 | return 0; | |
5977 | } | |
5978 | ||
5979 | static int | |
5980 | dpif_netdev_ct_dump_next(struct dpif *dpif OVS_UNUSED, | |
5981 | struct ct_dpif_dump_state *dump_, | |
5982 | struct ct_dpif_entry *entry) | |
5983 | { | |
5984 | struct dp_netdev_ct_dump *dump; | |
5985 | ||
5986 | INIT_CONTAINER(dump, dump_, up); | |
5987 | ||
5988 | return conntrack_dump_next(&dump->dump, entry); | |
5989 | } | |
5990 | ||
5991 | static int | |
5992 | dpif_netdev_ct_dump_done(struct dpif *dpif OVS_UNUSED, | |
5993 | struct ct_dpif_dump_state *dump_) | |
5994 | { | |
5995 | struct dp_netdev_ct_dump *dump; | |
5996 | int err; | |
5997 | ||
5998 | INIT_CONTAINER(dump, dump_, up); | |
5999 | ||
6000 | err = conntrack_dump_done(&dump->dump); | |
6001 | ||
6002 | free(dump); | |
6003 | ||
6004 | return err; | |
6005 | } | |
6006 | ||
5d9cbb4c | 6007 | static int |
817a7657 YHW |
6008 | dpif_netdev_ct_flush(struct dpif *dpif, const uint16_t *zone, |
6009 | const struct ct_dpif_tuple *tuple) | |
5d9cbb4c DDP |
6010 | { |
6011 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
6012 | ||
817a7657 | 6013 | if (tuple) { |
271e48a0 | 6014 | return conntrack_flush_tuple(&dp->conntrack, tuple, zone ? *zone : 0); |
817a7657 | 6015 | } |
5d9cbb4c DDP |
6016 | return conntrack_flush(&dp->conntrack, zone); |
6017 | } | |
6018 | ||
c92339ad DB |
6019 | static int |
6020 | dpif_netdev_ct_set_maxconns(struct dpif *dpif, uint32_t maxconns) | |
6021 | { | |
6022 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
6023 | ||
6024 | return conntrack_set_maxconns(&dp->conntrack, maxconns); | |
6025 | } | |
6026 | ||
6027 | static int | |
6028 | dpif_netdev_ct_get_maxconns(struct dpif *dpif, uint32_t *maxconns) | |
6029 | { | |
6030 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
6031 | ||
6032 | return conntrack_get_maxconns(&dp->conntrack, maxconns); | |
6033 | } | |
6034 | ||
875075b3 DB |
6035 | static int |
6036 | dpif_netdev_ct_get_nconns(struct dpif *dpif, uint32_t *nconns) | |
6037 | { | |
6038 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
6039 | ||
6040 | return conntrack_get_nconns(&dp->conntrack, nconns); | |
6041 | } | |
6042 | ||
72865317 | 6043 | const struct dpif_class dpif_netdev_class = { |
72865317 | 6044 | "netdev", |
6553d06b | 6045 | dpif_netdev_init, |
2197d7ab | 6046 | dpif_netdev_enumerate, |
0aeaabc8 | 6047 | dpif_netdev_port_open_type, |
72865317 BP |
6048 | dpif_netdev_open, |
6049 | dpif_netdev_close, | |
7dab847a | 6050 | dpif_netdev_destroy, |
e4cfed38 PS |
6051 | dpif_netdev_run, |
6052 | dpif_netdev_wait, | |
72865317 | 6053 | dpif_netdev_get_stats, |
72865317 BP |
6054 | dpif_netdev_port_add, |
6055 | dpif_netdev_port_del, | |
3eb67853 | 6056 | dpif_netdev_port_set_config, |
72865317 BP |
6057 | dpif_netdev_port_query_by_number, |
6058 | dpif_netdev_port_query_by_name, | |
98403001 | 6059 | NULL, /* port_get_pid */ |
b0ec0f27 BP |
6060 | dpif_netdev_port_dump_start, |
6061 | dpif_netdev_port_dump_next, | |
6062 | dpif_netdev_port_dump_done, | |
72865317 BP |
6063 | dpif_netdev_port_poll, |
6064 | dpif_netdev_port_poll_wait, | |
72865317 | 6065 | dpif_netdev_flow_flush, |
ac64794a BP |
6066 | dpif_netdev_flow_dump_create, |
6067 | dpif_netdev_flow_dump_destroy, | |
6068 | dpif_netdev_flow_dump_thread_create, | |
6069 | dpif_netdev_flow_dump_thread_destroy, | |
704a1e09 | 6070 | dpif_netdev_flow_dump_next, |
1a0c894a | 6071 | dpif_netdev_operate, |
6b31e073 RW |
6072 | NULL, /* recv_set */ |
6073 | NULL, /* handlers_set */ | |
d4f6865c | 6074 | dpif_netdev_set_config, |
5bf93d67 | 6075 | dpif_netdev_queue_to_priority, |
6b31e073 RW |
6076 | NULL, /* recv */ |
6077 | NULL, /* recv_wait */ | |
6078 | NULL, /* recv_purge */ | |
e4e74c3a | 6079 | dpif_netdev_register_dp_purge_cb, |
6b31e073 RW |
6080 | dpif_netdev_register_upcall_cb, |
6081 | dpif_netdev_enable_upcall, | |
6082 | dpif_netdev_disable_upcall, | |
b5cbbcf6 | 6083 | dpif_netdev_get_datapath_version, |
4d4e68ed DDP |
6084 | dpif_netdev_ct_dump_start, |
6085 | dpif_netdev_ct_dump_next, | |
6086 | dpif_netdev_ct_dump_done, | |
5d9cbb4c | 6087 | dpif_netdev_ct_flush, |
c92339ad DB |
6088 | dpif_netdev_ct_set_maxconns, |
6089 | dpif_netdev_ct_get_maxconns, | |
875075b3 | 6090 | dpif_netdev_ct_get_nconns, |
5dddf960 JR |
6091 | dpif_netdev_meter_get_features, |
6092 | dpif_netdev_meter_set, | |
6093 | dpif_netdev_meter_get, | |
6094 | dpif_netdev_meter_del, | |
72865317 | 6095 | }; |
614c4892 | 6096 | |
74cc3969 BP |
6097 | static void |
6098 | dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED, | |
6099 | const char *argv[], void *aux OVS_UNUSED) | |
6100 | { | |
e9985d6a | 6101 | struct dp_netdev_port *port; |
74cc3969 | 6102 | struct dp_netdev *dp; |
ff073a71 | 6103 | odp_port_t port_no; |
74cc3969 | 6104 | |
8a4e3a85 | 6105 | ovs_mutex_lock(&dp_netdev_mutex); |
74cc3969 BP |
6106 | dp = shash_find_data(&dp_netdevs, argv[1]); |
6107 | if (!dp || !dpif_netdev_class_is_dummy(dp->class)) { | |
8a4e3a85 | 6108 | ovs_mutex_unlock(&dp_netdev_mutex); |
74cc3969 BP |
6109 | unixctl_command_reply_error(conn, "unknown datapath or not a dummy"); |
6110 | return; | |
6111 | } | |
8a4e3a85 BP |
6112 | ovs_refcount_ref(&dp->ref_cnt); |
6113 | ovs_mutex_unlock(&dp_netdev_mutex); | |
74cc3969 | 6114 | |
59e6d833 | 6115 | ovs_mutex_lock(&dp->port_mutex); |
e9985d6a | 6116 | if (get_port_by_name(dp, argv[2], &port)) { |
74cc3969 | 6117 | unixctl_command_reply_error(conn, "unknown port"); |
8a4e3a85 | 6118 | goto exit; |
74cc3969 BP |
6119 | } |
6120 | ||
ff073a71 BP |
6121 | port_no = u32_to_odp(atoi(argv[3])); |
6122 | if (!port_no || port_no == ODPP_NONE) { | |
74cc3969 | 6123 | unixctl_command_reply_error(conn, "bad port number"); |
8a4e3a85 | 6124 | goto exit; |
74cc3969 | 6125 | } |
ff073a71 | 6126 | if (dp_netdev_lookup_port(dp, port_no)) { |
74cc3969 | 6127 | unixctl_command_reply_error(conn, "port number already in use"); |
8a4e3a85 | 6128 | goto exit; |
74cc3969 | 6129 | } |
59e6d833 | 6130 | |
e9985d6a DDP |
6131 | /* Remove port. */ |
6132 | hmap_remove(&dp->ports, &port->node); | |
e32971b8 | 6133 | reconfigure_datapath(dp); |
59e6d833 | 6134 | |
e9985d6a DDP |
6135 | /* Reinsert with new port number. */ |
6136 | port->port_no = port_no; | |
6137 | hmap_insert(&dp->ports, &port->node, hash_port_no(port_no)); | |
e32971b8 | 6138 | reconfigure_datapath(dp); |
59e6d833 | 6139 | |
d33ed218 | 6140 | seq_change(dp->port_seq); |
74cc3969 | 6141 | unixctl_command_reply(conn, NULL); |
8a4e3a85 BP |
6142 | |
6143 | exit: | |
59e6d833 | 6144 | ovs_mutex_unlock(&dp->port_mutex); |
8a4e3a85 | 6145 | dp_netdev_unref(dp); |
74cc3969 BP |
6146 | } |
6147 | ||
0cbfe35d BP |
6148 | static void |
6149 | dpif_dummy_register__(const char *type) | |
6150 | { | |
6151 | struct dpif_class *class; | |
6152 | ||
6153 | class = xmalloc(sizeof *class); | |
6154 | *class = dpif_netdev_class; | |
6155 | class->type = xstrdup(type); | |
6156 | dp_register_provider(class); | |
6157 | } | |
6158 | ||
8420c7ad BP |
6159 | static void |
6160 | dpif_dummy_override(const char *type) | |
6161 | { | |
65d43fdc YT |
6162 | int error; |
6163 | ||
6164 | /* | |
6165 | * Ignore EAFNOSUPPORT to allow --enable-dummy=system with | |
6166 | * a userland-only build. It's useful for testsuite. | |
6167 | */ | |
6168 | error = dp_unregister_provider(type); | |
6169 | if (error == 0 || error == EAFNOSUPPORT) { | |
8420c7ad BP |
6170 | dpif_dummy_register__(type); |
6171 | } | |
6172 | } | |
6173 | ||
614c4892 | 6174 | void |
8420c7ad | 6175 | dpif_dummy_register(enum dummy_level level) |
614c4892 | 6176 | { |
8420c7ad | 6177 | if (level == DUMMY_OVERRIDE_ALL) { |
0cbfe35d BP |
6178 | struct sset types; |
6179 | const char *type; | |
6180 | ||
6181 | sset_init(&types); | |
6182 | dp_enumerate_types(&types); | |
6183 | SSET_FOR_EACH (type, &types) { | |
8420c7ad | 6184 | dpif_dummy_override(type); |
0cbfe35d BP |
6185 | } |
6186 | sset_destroy(&types); | |
8420c7ad BP |
6187 | } else if (level == DUMMY_OVERRIDE_SYSTEM) { |
6188 | dpif_dummy_override("system"); | |
614c4892 | 6189 | } |
0cbfe35d BP |
6190 | |
6191 | dpif_dummy_register__("dummy"); | |
74cc3969 BP |
6192 | |
6193 | unixctl_command_register("dpif-dummy/change-port-number", | |
74467d5c | 6194 | "dp port new-number", |
74cc3969 | 6195 | 3, 3, dpif_dummy_change_port_number, NULL); |
614c4892 | 6196 | } |
0de8783a JR |
6197 | \f |
6198 | /* Datapath Classifier. */ | |
6199 | ||
6200 | /* A set of rules that all have the same fields wildcarded. */ | |
6201 | struct dpcls_subtable { | |
6202 | /* The fields are only used by writers. */ | |
6203 | struct cmap_node cmap_node OVS_GUARDED; /* Within dpcls 'subtables_map'. */ | |
6204 | ||
6205 | /* These fields are accessed by readers. */ | |
6206 | struct cmap rules; /* Contains "struct dpcls_rule"s. */ | |
3453b4d6 JS |
6207 | uint32_t hit_cnt; /* Number of match hits in subtable in current |
6208 | optimization interval. */ | |
0de8783a JR |
6209 | struct netdev_flow_key mask; /* Wildcards for fields (const). */ |
6210 | /* 'mask' must be the last field, additional space is allocated here. */ | |
6211 | }; | |
6212 | ||
6213 | /* Initializes 'cls' as a classifier that initially contains no classification | |
6214 | * rules. */ | |
6215 | static void | |
6216 | dpcls_init(struct dpcls *cls) | |
6217 | { | |
6218 | cmap_init(&cls->subtables_map); | |
da9cfca6 | 6219 | pvector_init(&cls->subtables); |
0de8783a JR |
6220 | } |
6221 | ||
6222 | static void | |
6223 | dpcls_destroy_subtable(struct dpcls *cls, struct dpcls_subtable *subtable) | |
6224 | { | |
3453b4d6 | 6225 | VLOG_DBG("Destroying subtable %p for in_port %d", subtable, cls->in_port); |
da9cfca6 | 6226 | pvector_remove(&cls->subtables, subtable); |
0de8783a JR |
6227 | cmap_remove(&cls->subtables_map, &subtable->cmap_node, |
6228 | subtable->mask.hash); | |
6229 | cmap_destroy(&subtable->rules); | |
6230 | ovsrcu_postpone(free, subtable); | |
6231 | } | |
6232 | ||
6233 | /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the | |
6234 | * caller's responsibility. | |
6235 | * May only be called after all the readers have been terminated. */ | |
6236 | static void | |
6237 | dpcls_destroy(struct dpcls *cls) | |
6238 | { | |
6239 | if (cls) { | |
6240 | struct dpcls_subtable *subtable; | |
6241 | ||
6242 | CMAP_FOR_EACH (subtable, cmap_node, &cls->subtables_map) { | |
361d808d | 6243 | ovs_assert(cmap_count(&subtable->rules) == 0); |
0de8783a JR |
6244 | dpcls_destroy_subtable(cls, subtable); |
6245 | } | |
6246 | cmap_destroy(&cls->subtables_map); | |
da9cfca6 | 6247 | pvector_destroy(&cls->subtables); |
0de8783a JR |
6248 | } |
6249 | } | |
6250 | ||
6251 | static struct dpcls_subtable * | |
6252 | dpcls_create_subtable(struct dpcls *cls, const struct netdev_flow_key *mask) | |
6253 | { | |
6254 | struct dpcls_subtable *subtable; | |
6255 | ||
6256 | /* Need to add one. */ | |
caeb4906 JR |
6257 | subtable = xmalloc(sizeof *subtable |
6258 | - sizeof subtable->mask.mf + mask->len); | |
0de8783a | 6259 | cmap_init(&subtable->rules); |
3453b4d6 | 6260 | subtable->hit_cnt = 0; |
0de8783a JR |
6261 | netdev_flow_key_clone(&subtable->mask, mask); |
6262 | cmap_insert(&cls->subtables_map, &subtable->cmap_node, mask->hash); | |
3453b4d6 | 6263 | /* Add the new subtable at the end of the pvector (with no hits yet) */ |
da9cfca6 | 6264 | pvector_insert(&cls->subtables, subtable, 0); |
84dbfb2b | 6265 | VLOG_DBG("Creating %"PRIuSIZE". subtable %p for in_port %d", |
3453b4d6 | 6266 | cmap_count(&cls->subtables_map), subtable, cls->in_port); |
da9cfca6 | 6267 | pvector_publish(&cls->subtables); |
0de8783a JR |
6268 | |
6269 | return subtable; | |
6270 | } | |
6271 | ||
6272 | static inline struct dpcls_subtable * | |
6273 | dpcls_find_subtable(struct dpcls *cls, const struct netdev_flow_key *mask) | |
6274 | { | |
6275 | struct dpcls_subtable *subtable; | |
6276 | ||
6277 | CMAP_FOR_EACH_WITH_HASH (subtable, cmap_node, mask->hash, | |
6278 | &cls->subtables_map) { | |
6279 | if (netdev_flow_key_equal(&subtable->mask, mask)) { | |
6280 | return subtable; | |
6281 | } | |
6282 | } | |
6283 | return dpcls_create_subtable(cls, mask); | |
6284 | } | |
6285 | ||
3453b4d6 JS |
6286 | |
6287 | /* Periodically sort the dpcls subtable vectors according to hit counts */ | |
6288 | static void | |
6289 | dpcls_sort_subtable_vector(struct dpcls *cls) | |
6290 | { | |
6291 | struct pvector *pvec = &cls->subtables; | |
6292 | struct dpcls_subtable *subtable; | |
6293 | ||
6294 | PVECTOR_FOR_EACH (subtable, pvec) { | |
6295 | pvector_change_priority(pvec, subtable, subtable->hit_cnt); | |
6296 | subtable->hit_cnt = 0; | |
6297 | } | |
6298 | pvector_publish(pvec); | |
6299 | } | |
6300 | ||
6301 | static inline void | |
4809891b KT |
6302 | dp_netdev_pmd_try_optimize(struct dp_netdev_pmd_thread *pmd, |
6303 | struct polled_queue *poll_list, int poll_cnt) | |
3453b4d6 JS |
6304 | { |
6305 | struct dpcls *cls; | |
3453b4d6 | 6306 | |
b010be17 | 6307 | if (pmd->ctx.now > pmd->rxq_next_cycle_store) { |
2a2c67b4 | 6308 | uint64_t curr_tsc; |
4809891b KT |
6309 | /* Get the cycles that were used to process each queue and store. */ |
6310 | for (unsigned i = 0; i < poll_cnt; i++) { | |
6311 | uint64_t rxq_cyc_curr = dp_netdev_rxq_get_cycles(poll_list[i].rxq, | |
6312 | RXQ_CYCLES_PROC_CURR); | |
6313 | dp_netdev_rxq_set_intrvl_cycles(poll_list[i].rxq, rxq_cyc_curr); | |
6314 | dp_netdev_rxq_set_cycles(poll_list[i].rxq, RXQ_CYCLES_PROC_CURR, | |
6315 | 0); | |
6316 | } | |
2a2c67b4 KT |
6317 | curr_tsc = cycles_counter_update(&pmd->perf_stats); |
6318 | if (pmd->intrvl_tsc_prev) { | |
6319 | /* There is a prev timestamp, store a new intrvl cycle count. */ | |
6320 | atomic_store_relaxed(&pmd->intrvl_cycles, | |
6321 | curr_tsc - pmd->intrvl_tsc_prev); | |
6322 | } | |
6323 | pmd->intrvl_tsc_prev = curr_tsc; | |
4809891b | 6324 | /* Start new measuring interval */ |
b010be17 | 6325 | pmd->rxq_next_cycle_store = pmd->ctx.now + PMD_RXQ_INTERVAL_LEN; |
4809891b KT |
6326 | } |
6327 | ||
b010be17 | 6328 | if (pmd->ctx.now > pmd->next_optimization) { |
3453b4d6 JS |
6329 | /* Try to obtain the flow lock to block out revalidator threads. |
6330 | * If not possible, just try next time. */ | |
6331 | if (!ovs_mutex_trylock(&pmd->flow_mutex)) { | |
6332 | /* Optimize each classifier */ | |
6333 | CMAP_FOR_EACH (cls, node, &pmd->classifiers) { | |
6334 | dpcls_sort_subtable_vector(cls); | |
6335 | } | |
6336 | ovs_mutex_unlock(&pmd->flow_mutex); | |
6337 | /* Start new measuring interval */ | |
b010be17 IM |
6338 | pmd->next_optimization = pmd->ctx.now |
6339 | + DPCLS_OPTIMIZATION_INTERVAL; | |
3453b4d6 JS |
6340 | } |
6341 | } | |
6342 | } | |
6343 | ||
0de8783a JR |
6344 | /* Insert 'rule' into 'cls'. */ |
6345 | static void | |
6346 | dpcls_insert(struct dpcls *cls, struct dpcls_rule *rule, | |
6347 | const struct netdev_flow_key *mask) | |
6348 | { | |
6349 | struct dpcls_subtable *subtable = dpcls_find_subtable(cls, mask); | |
6350 | ||
3453b4d6 | 6351 | /* Refer to subtable's mask, also for later removal. */ |
0de8783a JR |
6352 | rule->mask = &subtable->mask; |
6353 | cmap_insert(&subtable->rules, &rule->cmap_node, rule->flow.hash); | |
6354 | } | |
6355 | ||
6356 | /* Removes 'rule' from 'cls', also destructing the 'rule'. */ | |
6357 | static void | |
6358 | dpcls_remove(struct dpcls *cls, struct dpcls_rule *rule) | |
6359 | { | |
6360 | struct dpcls_subtable *subtable; | |
6361 | ||
6362 | ovs_assert(rule->mask); | |
6363 | ||
3453b4d6 | 6364 | /* Get subtable from reference in rule->mask. */ |
0de8783a | 6365 | INIT_CONTAINER(subtable, rule->mask, mask); |
0de8783a JR |
6366 | if (cmap_remove(&subtable->rules, &rule->cmap_node, rule->flow.hash) |
6367 | == 0) { | |
3453b4d6 | 6368 | /* Delete empty subtable. */ |
0de8783a | 6369 | dpcls_destroy_subtable(cls, subtable); |
da9cfca6 | 6370 | pvector_publish(&cls->subtables); |
0de8783a JR |
6371 | } |
6372 | } | |
6373 | ||
361d808d JR |
6374 | /* Returns true if 'target' satisfies 'key' in 'mask', that is, if each 1-bit |
6375 | * in 'mask' the values in 'key' and 'target' are the same. */ | |
0de8783a JR |
6376 | static inline bool |
6377 | dpcls_rule_matches_key(const struct dpcls_rule *rule, | |
6378 | const struct netdev_flow_key *target) | |
6379 | { | |
09b0fa9c JR |
6380 | const uint64_t *keyp = miniflow_get_values(&rule->flow.mf); |
6381 | const uint64_t *maskp = miniflow_get_values(&rule->mask->mf); | |
5fcff47b | 6382 | uint64_t value; |
0de8783a | 6383 | |
5fcff47b JR |
6384 | NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value, target, rule->flow.mf.map) { |
6385 | if (OVS_UNLIKELY((value & *maskp++) != *keyp++)) { | |
0de8783a JR |
6386 | return false; |
6387 | } | |
6388 | } | |
6389 | return true; | |
6390 | } | |
6391 | ||
5b1c9c78 FA |
6392 | /* For each miniflow in 'keys' performs a classifier lookup writing the result |
6393 | * into the corresponding slot in 'rules'. If a particular entry in 'keys' is | |
0de8783a JR |
6394 | * NULL it is skipped. |
6395 | * | |
6396 | * This function is optimized for use in the userspace datapath and therefore | |
6397 | * does not implement a lot of features available in the standard | |
6398 | * classifier_lookup() function. Specifically, it does not implement | |
6399 | * priorities, instead returning any rule which matches the flow. | |
6400 | * | |
5b1c9c78 | 6401 | * Returns true if all miniflows found a corresponding rule. */ |
0de8783a | 6402 | static bool |
3453b4d6 JS |
6403 | dpcls_lookup(struct dpcls *cls, const struct netdev_flow_key keys[], |
6404 | struct dpcls_rule **rules, const size_t cnt, | |
6405 | int *num_lookups_p) | |
0de8783a | 6406 | { |
5b1c9c78 | 6407 | /* The received 'cnt' miniflows are the search-keys that will be processed |
63906f18 BB |
6408 | * to find a matching entry into the available subtables. |
6409 | * The number of bits in map_type is equal to NETDEV_MAX_BURST. */ | |
6410 | typedef uint32_t map_type; | |
0de8783a | 6411 | #define MAP_BITS (sizeof(map_type) * CHAR_BIT) |
63906f18 | 6412 | BUILD_ASSERT_DECL(MAP_BITS >= NETDEV_MAX_BURST); |
0de8783a | 6413 | |
0de8783a JR |
6414 | struct dpcls_subtable *subtable; |
6415 | ||
63906f18 BB |
6416 | map_type keys_map = TYPE_MAXIMUM(map_type); /* Set all bits. */ |
6417 | map_type found_map; | |
6418 | uint32_t hashes[MAP_BITS]; | |
6419 | const struct cmap_node *nodes[MAP_BITS]; | |
6420 | ||
6421 | if (cnt != MAP_BITS) { | |
6422 | keys_map >>= MAP_BITS - cnt; /* Clear extra bits. */ | |
0de8783a JR |
6423 | } |
6424 | memset(rules, 0, cnt * sizeof *rules); | |
6425 | ||
3453b4d6 JS |
6426 | int lookups_match = 0, subtable_pos = 1; |
6427 | ||
5b1c9c78 FA |
6428 | /* The Datapath classifier - aka dpcls - is composed of subtables. |
6429 | * Subtables are dynamically created as needed when new rules are inserted. | |
6430 | * Each subtable collects rules with matches on a specific subset of packet | |
6431 | * fields as defined by the subtable's mask. We proceed to process every | |
6432 | * search-key against each subtable, but when a match is found for a | |
6433 | * search-key, the search for that key can stop because the rules are | |
6434 | * non-overlapping. */ | |
da9cfca6 | 6435 | PVECTOR_FOR_EACH (subtable, &cls->subtables) { |
63906f18 BB |
6436 | int i; |
6437 | ||
6438 | /* Compute hashes for the remaining keys. Each search-key is | |
6439 | * masked with the subtable's mask to avoid hashing the wildcarded | |
6440 | * bits. */ | |
6441 | ULLONG_FOR_EACH_1(i, keys_map) { | |
6442 | hashes[i] = netdev_flow_key_hash_in_mask(&keys[i], | |
6443 | &subtable->mask); | |
6444 | } | |
6445 | /* Lookup. */ | |
6446 | found_map = cmap_find_batch(&subtable->rules, keys_map, hashes, nodes); | |
6447 | /* Check results. When the i-th bit of found_map is set, it means | |
6448 | * that a set of nodes with a matching hash value was found for the | |
6449 | * i-th search-key. Due to possible hash collisions we need to check | |
6450 | * which of the found rules, if any, really matches our masked | |
6451 | * search-key. */ | |
6452 | ULLONG_FOR_EACH_1(i, found_map) { | |
6453 | struct dpcls_rule *rule; | |
6454 | ||
6455 | CMAP_NODE_FOR_EACH (rule, cmap_node, nodes[i]) { | |
6456 | if (OVS_LIKELY(dpcls_rule_matches_key(rule, &keys[i]))) { | |
6457 | rules[i] = rule; | |
6458 | /* Even at 20 Mpps the 32-bit hit_cnt cannot wrap | |
6459 | * within one second optimization interval. */ | |
6460 | subtable->hit_cnt++; | |
6461 | lookups_match += subtable_pos; | |
6462 | goto next; | |
0de8783a | 6463 | } |
0de8783a | 6464 | } |
63906f18 BB |
6465 | /* None of the found rules was a match. Reset the i-th bit to |
6466 | * keep searching this key in the next subtable. */ | |
6467 | ULLONG_SET0(found_map, i); /* Did not match. */ | |
6468 | next: | |
6469 | ; /* Keep Sparse happy. */ | |
0de8783a | 6470 | } |
63906f18 BB |
6471 | keys_map &= ~found_map; /* Clear the found rules. */ |
6472 | if (!keys_map) { | |
3453b4d6 JS |
6473 | if (num_lookups_p) { |
6474 | *num_lookups_p = lookups_match; | |
6475 | } | |
0de8783a JR |
6476 | return true; /* All found. */ |
6477 | } | |
3453b4d6 JS |
6478 | subtable_pos++; |
6479 | } | |
6480 | if (num_lookups_p) { | |
6481 | *num_lookups_p = lookups_match; | |
0de8783a JR |
6482 | } |
6483 | return false; /* Some misses. */ | |
6484 | } |