]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014, 2016, 2017 Nicira, Inc. | |
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> | |
18 | #include "dpif-netdev.h" | |
19 | ||
20 | #include <ctype.h> | |
21 | #include <errno.h> | |
22 | #include <fcntl.h> | |
23 | #include <inttypes.h> | |
24 | #include <net/if.h> | |
25 | #include <netinet/in.h> | |
26 | #include <stdint.h> | |
27 | #include <stdlib.h> | |
28 | #include <string.h> | |
29 | #include <sys/ioctl.h> | |
30 | #include <sys/socket.h> | |
31 | #include <sys/stat.h> | |
32 | #include <unistd.h> | |
33 | ||
34 | #ifdef DPDK_NETDEV | |
35 | #include <rte_cycles.h> | |
36 | #endif | |
37 | ||
38 | #include "bitmap.h" | |
39 | #include "cmap.h" | |
40 | #include "conntrack.h" | |
41 | #include "coverage.h" | |
42 | #include "ct-dpif.h" | |
43 | #include "csum.h" | |
44 | #include "dp-packet.h" | |
45 | #include "dpif.h" | |
46 | #include "dpif-provider.h" | |
47 | #include "dummy.h" | |
48 | #include "fat-rwlock.h" | |
49 | #include "flow.h" | |
50 | #include "hmapx.h" | |
51 | #include "latch.h" | |
52 | #include "netdev.h" | |
53 | #include "netdev-vport.h" | |
54 | #include "netlink.h" | |
55 | #include "odp-execute.h" | |
56 | #include "odp-util.h" | |
57 | #include "openvswitch/dynamic-string.h" | |
58 | #include "openvswitch/list.h" | |
59 | #include "openvswitch/match.h" | |
60 | #include "openvswitch/ofp-print.h" | |
61 | #include "openvswitch/ofp-util.h" | |
62 | #include "openvswitch/ofpbuf.h" | |
63 | #include "openvswitch/shash.h" | |
64 | #include "openvswitch/vlog.h" | |
65 | #include "ovs-numa.h" | |
66 | #include "ovs-rcu.h" | |
67 | #include "packets.h" | |
68 | #include "poll-loop.h" | |
69 | #include "pvector.h" | |
70 | #include "random.h" | |
71 | #include "seq.h" | |
72 | #include "smap.h" | |
73 | #include "sset.h" | |
74 | #include "timeval.h" | |
75 | #include "tnl-neigh-cache.h" | |
76 | #include "tnl-ports.h" | |
77 | #include "unixctl.h" | |
78 | #include "util.h" | |
79 | ||
80 | VLOG_DEFINE_THIS_MODULE(dpif_netdev); | |
81 | ||
82 | #define FLOW_DUMP_MAX_BATCH 50 | |
83 | /* Use per thread recirc_depth to prevent recirculation loop. */ | |
84 | #define MAX_RECIRC_DEPTH 5 | |
85 | DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth, 0) | |
86 | ||
87 | /* Configuration parameters. */ | |
88 | enum { MAX_FLOWS = 65536 }; /* Maximum number of flows in flow table. */ | |
89 | enum { MAX_METERS = 65536 }; /* Maximum number of meters. */ | |
90 | enum { MAX_BANDS = 8 }; /* Maximum number of bands / meter. */ | |
91 | enum { N_METER_LOCKS = 64 }; /* Maximum number of meters. */ | |
92 | ||
93 | /* Protects against changes to 'dp_netdevs'. */ | |
94 | static struct ovs_mutex dp_netdev_mutex = OVS_MUTEX_INITIALIZER; | |
95 | ||
96 | /* Contains all 'struct dp_netdev's. */ | |
97 | static struct shash dp_netdevs OVS_GUARDED_BY(dp_netdev_mutex) | |
98 | = SHASH_INITIALIZER(&dp_netdevs); | |
99 | ||
100 | static struct vlog_rate_limit upcall_rl = VLOG_RATE_LIMIT_INIT(600, 600); | |
101 | ||
102 | #define DP_NETDEV_CS_SUPPORTED_MASK (CS_NEW | CS_ESTABLISHED | CS_RELATED \ | |
103 | | CS_INVALID | CS_REPLY_DIR | CS_TRACKED) | |
104 | #define DP_NETDEV_CS_UNSUPPORTED_MASK (~(uint32_t)DP_NETDEV_CS_SUPPORTED_MASK) | |
105 | ||
106 | static struct odp_support dp_netdev_support = { | |
107 | .max_vlan_headers = SIZE_MAX, | |
108 | .max_mpls_depth = SIZE_MAX, | |
109 | .recirc = true, | |
110 | .ct_state = true, | |
111 | .ct_zone = true, | |
112 | .ct_mark = true, | |
113 | .ct_label = true, | |
114 | }; | |
115 | ||
116 | /* Stores a miniflow with inline values */ | |
117 | ||
118 | struct netdev_flow_key { | |
119 | uint32_t hash; /* Hash function differs for different users. */ | |
120 | uint32_t len; /* Length of the following miniflow (incl. map). */ | |
121 | struct miniflow mf; | |
122 | uint64_t buf[FLOW_MAX_PACKET_U64S]; | |
123 | }; | |
124 | ||
125 | /* Exact match cache for frequently used flows | |
126 | * | |
127 | * The cache uses a 32-bit hash of the packet (which can be the RSS hash) to | |
128 | * search its entries for a miniflow that matches exactly the miniflow of the | |
129 | * packet. It stores the 'dpcls_rule' (rule) that matches the miniflow. | |
130 | * | |
131 | * A cache entry holds a reference to its 'dp_netdev_flow'. | |
132 | * | |
133 | * A miniflow with a given hash can be in one of EM_FLOW_HASH_SEGS different | |
134 | * entries. The 32-bit hash is split into EM_FLOW_HASH_SEGS values (each of | |
135 | * them is EM_FLOW_HASH_SHIFT bits wide and the remainder is thrown away). Each | |
136 | * value is the index of a cache entry where the miniflow could be. | |
137 | * | |
138 | * | |
139 | * Thread-safety | |
140 | * ============= | |
141 | * | |
142 | * Each pmd_thread has its own private exact match cache. | |
143 | * If dp_netdev_input is not called from a pmd thread, a mutex is used. | |
144 | */ | |
145 | ||
146 | #define EM_FLOW_HASH_SHIFT 13 | |
147 | #define EM_FLOW_HASH_ENTRIES (1u << EM_FLOW_HASH_SHIFT) | |
148 | #define EM_FLOW_HASH_MASK (EM_FLOW_HASH_ENTRIES - 1) | |
149 | #define EM_FLOW_HASH_SEGS 2 | |
150 | ||
151 | /* Default EMC insert probability is 1 / DEFAULT_EM_FLOW_INSERT_INV_PROB */ | |
152 | #define DEFAULT_EM_FLOW_INSERT_INV_PROB 100 | |
153 | #define DEFAULT_EM_FLOW_INSERT_MIN (UINT32_MAX / \ | |
154 | DEFAULT_EM_FLOW_INSERT_INV_PROB) | |
155 | ||
156 | struct emc_entry { | |
157 | struct dp_netdev_flow *flow; | |
158 | struct netdev_flow_key key; /* key.hash used for emc hash value. */ | |
159 | }; | |
160 | ||
161 | struct emc_cache { | |
162 | struct emc_entry entries[EM_FLOW_HASH_ENTRIES]; | |
163 | int sweep_idx; /* For emc_cache_slow_sweep(). */ | |
164 | }; | |
165 | ||
166 | /* Iterate in the exact match cache through every entry that might contain a | |
167 | * miniflow with hash 'HASH'. */ | |
168 | #define EMC_FOR_EACH_POS_WITH_HASH(EMC, CURRENT_ENTRY, HASH) \ | |
169 | for (uint32_t i__ = 0, srch_hash__ = (HASH); \ | |
170 | (CURRENT_ENTRY) = &(EMC)->entries[srch_hash__ & EM_FLOW_HASH_MASK], \ | |
171 | i__ < EM_FLOW_HASH_SEGS; \ | |
172 | i__++, srch_hash__ >>= EM_FLOW_HASH_SHIFT) | |
173 | \f | |
174 | /* Simple non-wildcarding single-priority classifier. */ | |
175 | ||
176 | /* Time in ms between successive optimizations of the dpcls subtable vector */ | |
177 | #define DPCLS_OPTIMIZATION_INTERVAL 1000 | |
178 | ||
179 | struct dpcls { | |
180 | struct cmap_node node; /* Within dp_netdev_pmd_thread.classifiers */ | |
181 | odp_port_t in_port; | |
182 | struct cmap subtables_map; | |
183 | struct pvector subtables; | |
184 | }; | |
185 | ||
186 | /* A rule to be inserted to the classifier. */ | |
187 | struct dpcls_rule { | |
188 | struct cmap_node cmap_node; /* Within struct dpcls_subtable 'rules'. */ | |
189 | struct netdev_flow_key *mask; /* Subtable's mask. */ | |
190 | struct netdev_flow_key flow; /* Matching key. */ | |
191 | /* 'flow' must be the last field, additional space is allocated here. */ | |
192 | }; | |
193 | ||
194 | static void dpcls_init(struct dpcls *); | |
195 | static void dpcls_destroy(struct dpcls *); | |
196 | static void dpcls_sort_subtable_vector(struct dpcls *); | |
197 | static void dpcls_insert(struct dpcls *, struct dpcls_rule *, | |
198 | const struct netdev_flow_key *mask); | |
199 | static void dpcls_remove(struct dpcls *, struct dpcls_rule *); | |
200 | static bool dpcls_lookup(struct dpcls *cls, | |
201 | const struct netdev_flow_key keys[], | |
202 | struct dpcls_rule **rules, size_t cnt, | |
203 | int *num_lookups_p); | |
204 | \f | |
205 | /* Set of supported meter flags */ | |
206 | #define DP_SUPPORTED_METER_FLAGS_MASK \ | |
207 | (OFPMF13_STATS | OFPMF13_PKTPS | OFPMF13_KBPS | OFPMF13_BURST) | |
208 | ||
209 | /* Set of supported meter band types */ | |
210 | #define DP_SUPPORTED_METER_BAND_TYPES \ | |
211 | ( 1 << OFPMBT13_DROP ) | |
212 | ||
213 | struct dp_meter_band { | |
214 | struct ofputil_meter_band up; /* type, prec_level, pad, rate, burst_size */ | |
215 | uint32_t bucket; /* In 1/1000 packets (for PKTPS), or in bits (for KBPS) */ | |
216 | uint64_t packet_count; | |
217 | uint64_t byte_count; | |
218 | }; | |
219 | ||
220 | struct dp_meter { | |
221 | uint16_t flags; | |
222 | uint16_t n_bands; | |
223 | uint32_t max_delta_t; | |
224 | uint64_t used; | |
225 | uint64_t packet_count; | |
226 | uint64_t byte_count; | |
227 | struct dp_meter_band bands[]; | |
228 | }; | |
229 | ||
230 | /* Datapath based on the network device interface from netdev.h. | |
231 | * | |
232 | * | |
233 | * Thread-safety | |
234 | * ============= | |
235 | * | |
236 | * Some members, marked 'const', are immutable. Accessing other members | |
237 | * requires synchronization, as noted in more detail below. | |
238 | * | |
239 | * Acquisition order is, from outermost to innermost: | |
240 | * | |
241 | * dp_netdev_mutex (global) | |
242 | * port_mutex | |
243 | * non_pmd_mutex | |
244 | */ | |
245 | struct dp_netdev { | |
246 | const struct dpif_class *const class; | |
247 | const char *const name; | |
248 | struct dpif *dpif; | |
249 | struct ovs_refcount ref_cnt; | |
250 | atomic_flag destroyed; | |
251 | ||
252 | /* Ports. | |
253 | * | |
254 | * Any lookup into 'ports' or any access to the dp_netdev_ports found | |
255 | * through 'ports' requires taking 'port_mutex'. */ | |
256 | struct ovs_mutex port_mutex; | |
257 | struct hmap ports; | |
258 | struct seq *port_seq; /* Incremented whenever a port changes. */ | |
259 | ||
260 | /* Meters. */ | |
261 | struct ovs_mutex meter_locks[N_METER_LOCKS]; | |
262 | struct dp_meter *meters[MAX_METERS]; /* Meter bands. */ | |
263 | ||
264 | /* Probability of EMC insertions is a factor of 'emc_insert_min'.*/ | |
265 | OVS_ALIGNED_VAR(CACHE_LINE_SIZE) atomic_uint32_t emc_insert_min; | |
266 | ||
267 | /* Protects access to ofproto-dpif-upcall interface during revalidator | |
268 | * thread synchronization. */ | |
269 | struct fat_rwlock upcall_rwlock; | |
270 | upcall_callback *upcall_cb; /* Callback function for executing upcalls. */ | |
271 | void *upcall_aux; | |
272 | ||
273 | /* Callback function for notifying the purging of dp flows (during | |
274 | * reseting pmd deletion). */ | |
275 | dp_purge_callback *dp_purge_cb; | |
276 | void *dp_purge_aux; | |
277 | ||
278 | /* Stores all 'struct dp_netdev_pmd_thread's. */ | |
279 | struct cmap poll_threads; | |
280 | ||
281 | /* Protects the access of the 'struct dp_netdev_pmd_thread' | |
282 | * instance for non-pmd thread. */ | |
283 | struct ovs_mutex non_pmd_mutex; | |
284 | ||
285 | /* Each pmd thread will store its pointer to | |
286 | * 'struct dp_netdev_pmd_thread' in 'per_pmd_key'. */ | |
287 | ovsthread_key_t per_pmd_key; | |
288 | ||
289 | struct seq *reconfigure_seq; | |
290 | uint64_t last_reconfigure_seq; | |
291 | ||
292 | /* Cpu mask for pin of pmd threads. */ | |
293 | char *pmd_cmask; | |
294 | ||
295 | uint64_t last_tnl_conf_seq; | |
296 | ||
297 | struct conntrack conntrack; | |
298 | }; | |
299 | ||
300 | static void meter_lock(const struct dp_netdev *dp, uint32_t meter_id) | |
301 | OVS_ACQUIRES(dp->meter_locks[meter_id % N_METER_LOCKS]) | |
302 | { | |
303 | ovs_mutex_lock(&dp->meter_locks[meter_id % N_METER_LOCKS]); | |
304 | } | |
305 | ||
306 | static void meter_unlock(const struct dp_netdev *dp, uint32_t meter_id) | |
307 | OVS_RELEASES(dp->meter_locks[meter_id % N_METER_LOCKS]) | |
308 | { | |
309 | ovs_mutex_unlock(&dp->meter_locks[meter_id % N_METER_LOCKS]); | |
310 | } | |
311 | ||
312 | ||
313 | static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp, | |
314 | odp_port_t) | |
315 | OVS_REQUIRES(dp->port_mutex); | |
316 | ||
317 | enum dp_stat_type { | |
318 | DP_STAT_EXACT_HIT, /* Packets that had an exact match (emc). */ | |
319 | DP_STAT_MASKED_HIT, /* Packets that matched in the flow table. */ | |
320 | DP_STAT_MISS, /* Packets that did not match. */ | |
321 | DP_STAT_LOST, /* Packets not passed up to the client. */ | |
322 | DP_STAT_LOOKUP_HIT, /* Number of subtable lookups for flow table | |
323 | hits */ | |
324 | DP_N_STATS | |
325 | }; | |
326 | ||
327 | enum pmd_cycles_counter_type { | |
328 | PMD_CYCLES_POLLING, /* Cycles spent polling NICs. */ | |
329 | PMD_CYCLES_PROCESSING, /* Cycles spent processing packets */ | |
330 | PMD_N_CYCLES | |
331 | }; | |
332 | ||
333 | #define XPS_TIMEOUT_MS 500LL | |
334 | ||
335 | /* Contained by struct dp_netdev_port's 'rxqs' member. */ | |
336 | struct dp_netdev_rxq { | |
337 | struct dp_netdev_port *port; | |
338 | struct netdev_rxq *rx; | |
339 | unsigned core_id; /* Core to which this queue should be | |
340 | pinned. OVS_CORE_UNSPEC if the | |
341 | queue doesn't need to be pinned to a | |
342 | particular core. */ | |
343 | struct dp_netdev_pmd_thread *pmd; /* pmd thread that polls this queue. */ | |
344 | }; | |
345 | ||
346 | /* A port in a netdev-based datapath. */ | |
347 | struct dp_netdev_port { | |
348 | odp_port_t port_no; | |
349 | struct netdev *netdev; | |
350 | struct hmap_node node; /* Node in dp_netdev's 'ports'. */ | |
351 | struct netdev_saved_flags *sf; | |
352 | struct dp_netdev_rxq *rxqs; | |
353 | unsigned n_rxq; /* Number of elements in 'rxq' */ | |
354 | bool dynamic_txqs; /* If true XPS will be used. */ | |
355 | unsigned *txq_used; /* Number of threads that use each tx queue. */ | |
356 | struct ovs_mutex txq_used_mutex; | |
357 | char *type; /* Port type as requested by user. */ | |
358 | char *rxq_affinity_list; /* Requested affinity of rx queues. */ | |
359 | bool need_reconfigure; /* True if we should reconfigure netdev. */ | |
360 | }; | |
361 | ||
362 | /* Contained by struct dp_netdev_flow's 'stats' member. */ | |
363 | struct dp_netdev_flow_stats { | |
364 | atomic_llong used; /* Last used time, in monotonic msecs. */ | |
365 | atomic_ullong packet_count; /* Number of packets matched. */ | |
366 | atomic_ullong byte_count; /* Number of bytes matched. */ | |
367 | atomic_uint16_t tcp_flags; /* Bitwise-OR of seen tcp_flags values. */ | |
368 | }; | |
369 | ||
370 | /* A flow in 'dp_netdev_pmd_thread's 'flow_table'. | |
371 | * | |
372 | * | |
373 | * Thread-safety | |
374 | * ============= | |
375 | * | |
376 | * Except near the beginning or ending of its lifespan, rule 'rule' belongs to | |
377 | * its pmd thread's classifier. The text below calls this classifier 'cls'. | |
378 | * | |
379 | * Motivation | |
380 | * ---------- | |
381 | * | |
382 | * The thread safety rules described here for "struct dp_netdev_flow" are | |
383 | * motivated by two goals: | |
384 | * | |
385 | * - Prevent threads that read members of "struct dp_netdev_flow" from | |
386 | * reading bad data due to changes by some thread concurrently modifying | |
387 | * those members. | |
388 | * | |
389 | * - Prevent two threads making changes to members of a given "struct | |
390 | * dp_netdev_flow" from interfering with each other. | |
391 | * | |
392 | * | |
393 | * Rules | |
394 | * ----- | |
395 | * | |
396 | * A flow 'flow' may be accessed without a risk of being freed during an RCU | |
397 | * grace period. Code that needs to hold onto a flow for a while | |
398 | * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref(). | |
399 | * | |
400 | * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the | |
401 | * flow from being deleted from 'cls' and it doesn't protect members of 'flow' | |
402 | * from modification. | |
403 | * | |
404 | * Some members, marked 'const', are immutable. Accessing other members | |
405 | * requires synchronization, as noted in more detail below. | |
406 | */ | |
407 | struct dp_netdev_flow { | |
408 | const struct flow flow; /* Unmasked flow that created this entry. */ | |
409 | /* Hash table index by unmasked flow. */ | |
410 | const struct cmap_node node; /* In owning dp_netdev_pmd_thread's */ | |
411 | /* 'flow_table'. */ | |
412 | const ovs_u128 ufid; /* Unique flow identifier. */ | |
413 | const unsigned pmd_id; /* The 'core_id' of pmd thread owning this */ | |
414 | /* flow. */ | |
415 | ||
416 | /* Number of references. | |
417 | * The classifier owns one reference. | |
418 | * Any thread trying to keep a rule from being freed should hold its own | |
419 | * reference. */ | |
420 | struct ovs_refcount ref_cnt; | |
421 | ||
422 | bool dead; | |
423 | ||
424 | /* Statistics. */ | |
425 | struct dp_netdev_flow_stats stats; | |
426 | ||
427 | /* Actions. */ | |
428 | OVSRCU_TYPE(struct dp_netdev_actions *) actions; | |
429 | ||
430 | /* While processing a group of input packets, the datapath uses the next | |
431 | * member to store a pointer to the output batch for the flow. It is | |
432 | * reset after the batch has been sent out (See dp_netdev_queue_batches(), | |
433 | * packet_batch_per_flow_init() and packet_batch_per_flow_execute()). */ | |
434 | struct packet_batch_per_flow *batch; | |
435 | ||
436 | /* Packet classification. */ | |
437 | struct dpcls_rule cr; /* In owning dp_netdev's 'cls'. */ | |
438 | /* 'cr' must be the last member. */ | |
439 | }; | |
440 | ||
441 | static void dp_netdev_flow_unref(struct dp_netdev_flow *); | |
442 | static bool dp_netdev_flow_ref(struct dp_netdev_flow *); | |
443 | static int dpif_netdev_flow_from_nlattrs(const struct nlattr *, uint32_t, | |
444 | struct flow *, bool); | |
445 | ||
446 | /* A set of datapath actions within a "struct dp_netdev_flow". | |
447 | * | |
448 | * | |
449 | * Thread-safety | |
450 | * ============= | |
451 | * | |
452 | * A struct dp_netdev_actions 'actions' is protected with RCU. */ | |
453 | struct dp_netdev_actions { | |
454 | /* These members are immutable: they do not change during the struct's | |
455 | * lifetime. */ | |
456 | unsigned int size; /* Size of 'actions', in bytes. */ | |
457 | struct nlattr actions[]; /* Sequence of OVS_ACTION_ATTR_* attributes. */ | |
458 | }; | |
459 | ||
460 | struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *, | |
461 | size_t); | |
462 | struct dp_netdev_actions *dp_netdev_flow_get_actions( | |
463 | const struct dp_netdev_flow *); | |
464 | static void dp_netdev_actions_free(struct dp_netdev_actions *); | |
465 | ||
466 | /* Contained by struct dp_netdev_pmd_thread's 'stats' member. */ | |
467 | struct dp_netdev_pmd_stats { | |
468 | /* Indexed by DP_STAT_*. */ | |
469 | atomic_ullong n[DP_N_STATS]; | |
470 | }; | |
471 | ||
472 | /* Contained by struct dp_netdev_pmd_thread's 'cycle' member. */ | |
473 | struct dp_netdev_pmd_cycles { | |
474 | /* Indexed by PMD_CYCLES_*. */ | |
475 | atomic_ullong n[PMD_N_CYCLES]; | |
476 | }; | |
477 | ||
478 | struct polled_queue { | |
479 | struct netdev_rxq *rx; | |
480 | odp_port_t port_no; | |
481 | }; | |
482 | ||
483 | /* Contained by struct dp_netdev_pmd_thread's 'poll_list' member. */ | |
484 | struct rxq_poll { | |
485 | struct dp_netdev_rxq *rxq; | |
486 | struct hmap_node node; | |
487 | }; | |
488 | ||
489 | /* Contained by struct dp_netdev_pmd_thread's 'send_port_cache', | |
490 | * 'tnl_port_cache' or 'tx_ports'. */ | |
491 | struct tx_port { | |
492 | struct dp_netdev_port *port; | |
493 | int qid; | |
494 | long long last_used; | |
495 | struct hmap_node node; | |
496 | }; | |
497 | ||
498 | /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate | |
499 | * the performance overhead of interrupt processing. Therefore netdev can | |
500 | * not implement rx-wait for these devices. dpif-netdev needs to poll | |
501 | * these device to check for recv buffer. pmd-thread does polling for | |
502 | * devices assigned to itself. | |
503 | * | |
504 | * DPDK used PMD for accessing NIC. | |
505 | * | |
506 | * Note, instance with cpu core id NON_PMD_CORE_ID will be reserved for | |
507 | * I/O of all non-pmd threads. There will be no actual thread created | |
508 | * for the instance. | |
509 | * | |
510 | * Each struct has its own flow cache and classifier per managed ingress port. | |
511 | * For packets received on ingress port, a look up is done on corresponding PMD | |
512 | * thread's flow cache and in case of a miss, lookup is performed in the | |
513 | * corresponding classifier of port. Packets are executed with the found | |
514 | * actions in either case. | |
515 | * */ | |
516 | struct dp_netdev_pmd_thread { | |
517 | struct dp_netdev *dp; | |
518 | struct ovs_refcount ref_cnt; /* Every reference must be refcount'ed. */ | |
519 | struct cmap_node node; /* In 'dp->poll_threads'. */ | |
520 | ||
521 | pthread_cond_t cond; /* For synchronizing pmd thread reload. */ | |
522 | struct ovs_mutex cond_mutex; /* Mutex for condition variable. */ | |
523 | ||
524 | /* Per thread exact-match cache. Note, the instance for cpu core | |
525 | * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly | |
526 | * need to be protected by 'non_pmd_mutex'. Every other instance | |
527 | * will only be accessed by its own pmd thread. */ | |
528 | struct emc_cache flow_cache; | |
529 | ||
530 | /* Flow-Table and classifiers | |
531 | * | |
532 | * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding | |
533 | * changes to 'classifiers' must be made while still holding the | |
534 | * 'flow_mutex'. | |
535 | */ | |
536 | struct ovs_mutex flow_mutex; | |
537 | struct cmap flow_table OVS_GUARDED; /* Flow table. */ | |
538 | ||
539 | /* One classifier per in_port polled by the pmd */ | |
540 | struct cmap classifiers; | |
541 | /* Periodically sort subtable vectors according to hit frequencies */ | |
542 | long long int next_optimization; | |
543 | ||
544 | /* Statistics. */ | |
545 | struct dp_netdev_pmd_stats stats; | |
546 | ||
547 | /* Cycles counters */ | |
548 | struct dp_netdev_pmd_cycles cycles; | |
549 | ||
550 | /* Used to count cicles. See 'cycles_counter_end()' */ | |
551 | unsigned long long last_cycles; | |
552 | ||
553 | struct latch exit_latch; /* For terminating the pmd thread. */ | |
554 | struct seq *reload_seq; | |
555 | uint64_t last_reload_seq; | |
556 | atomic_bool reload; /* Do we need to reload ports? */ | |
557 | pthread_t thread; | |
558 | unsigned core_id; /* CPU core id of this pmd thread. */ | |
559 | int numa_id; /* numa node id of this pmd thread. */ | |
560 | bool isolated; | |
561 | ||
562 | /* Queue id used by this pmd thread to send packets on all netdevs if | |
563 | * XPS disabled for this netdev. All static_tx_qid's are unique and less | |
564 | * than 'cmap_count(dp->poll_threads)'. */ | |
565 | const int static_tx_qid; | |
566 | ||
567 | struct ovs_mutex port_mutex; /* Mutex for 'poll_list' and 'tx_ports'. */ | |
568 | /* List of rx queues to poll. */ | |
569 | struct hmap poll_list OVS_GUARDED; | |
570 | /* Map of 'tx_port's used for transmission. Written by the main thread, | |
571 | * read by the pmd thread. */ | |
572 | struct hmap tx_ports OVS_GUARDED; | |
573 | ||
574 | /* These are thread-local copies of 'tx_ports'. One contains only tunnel | |
575 | * ports (that support push_tunnel/pop_tunnel), the other contains ports | |
576 | * with at least one txq (that support send). A port can be in both. | |
577 | * | |
578 | * There are two separate maps to make sure that we don't try to execute | |
579 | * OUTPUT on a device which has 0 txqs or PUSH/POP on a non-tunnel device. | |
580 | * | |
581 | * The instances for cpu core NON_PMD_CORE_ID can be accessed by multiple | |
582 | * threads, and thusly need to be protected by 'non_pmd_mutex'. Every | |
583 | * other instance will only be accessed by its own pmd thread. */ | |
584 | struct hmap tnl_port_cache; | |
585 | struct hmap send_port_cache; | |
586 | ||
587 | /* Only a pmd thread can write on its own 'cycles' and 'stats'. | |
588 | * The main thread keeps 'stats_zero' and 'cycles_zero' as base | |
589 | * values and subtracts them from 'stats' and 'cycles' before | |
590 | * reporting to the user */ | |
591 | unsigned long long stats_zero[DP_N_STATS]; | |
592 | uint64_t cycles_zero[PMD_N_CYCLES]; | |
593 | ||
594 | /* Set to true if the pmd thread needs to be reloaded. */ | |
595 | bool need_reload; | |
596 | }; | |
597 | ||
598 | /* Interface to netdev-based datapath. */ | |
599 | struct dpif_netdev { | |
600 | struct dpif dpif; | |
601 | struct dp_netdev *dp; | |
602 | uint64_t last_port_seq; | |
603 | }; | |
604 | ||
605 | static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no, | |
606 | struct dp_netdev_port **portp) | |
607 | OVS_REQUIRES(dp->port_mutex); | |
608 | static int get_port_by_name(struct dp_netdev *dp, const char *devname, | |
609 | struct dp_netdev_port **portp) | |
610 | OVS_REQUIRES(dp->port_mutex); | |
611 | static void dp_netdev_free(struct dp_netdev *) | |
612 | OVS_REQUIRES(dp_netdev_mutex); | |
613 | static int do_add_port(struct dp_netdev *dp, const char *devname, | |
614 | const char *type, odp_port_t port_no) | |
615 | OVS_REQUIRES(dp->port_mutex); | |
616 | static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *) | |
617 | OVS_REQUIRES(dp->port_mutex); | |
618 | static int dpif_netdev_open(const struct dpif_class *, const char *name, | |
619 | bool create, struct dpif **); | |
620 | static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd, | |
621 | struct dp_packet_batch *, | |
622 | bool may_steal, const struct flow *flow, | |
623 | const struct nlattr *actions, | |
624 | size_t actions_len, | |
625 | long long now); | |
626 | static void dp_netdev_input(struct dp_netdev_pmd_thread *, | |
627 | struct dp_packet_batch *, odp_port_t port_no); | |
628 | static void dp_netdev_recirculate(struct dp_netdev_pmd_thread *, | |
629 | struct dp_packet_batch *); | |
630 | ||
631 | static void dp_netdev_disable_upcall(struct dp_netdev *); | |
632 | static void dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd); | |
633 | static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, | |
634 | struct dp_netdev *dp, unsigned core_id, | |
635 | int numa_id); | |
636 | static void dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd); | |
637 | static void dp_netdev_set_nonpmd(struct dp_netdev *dp) | |
638 | OVS_REQUIRES(dp->port_mutex); | |
639 | ||
640 | static void *pmd_thread_main(void *); | |
641 | static struct dp_netdev_pmd_thread *dp_netdev_get_pmd(struct dp_netdev *dp, | |
642 | unsigned core_id); | |
643 | static struct dp_netdev_pmd_thread * | |
644 | dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos); | |
645 | static void dp_netdev_destroy_all_pmds(struct dp_netdev *dp, bool non_pmd); | |
646 | static void dp_netdev_pmd_clear_ports(struct dp_netdev_pmd_thread *pmd); | |
647 | static void dp_netdev_add_port_tx_to_pmd(struct dp_netdev_pmd_thread *pmd, | |
648 | struct dp_netdev_port *port) | |
649 | OVS_REQUIRES(pmd->port_mutex); | |
650 | static void dp_netdev_del_port_tx_from_pmd(struct dp_netdev_pmd_thread *pmd, | |
651 | struct tx_port *tx) | |
652 | OVS_REQUIRES(pmd->port_mutex); | |
653 | static void dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread *pmd, | |
654 | struct dp_netdev_rxq *rxq) | |
655 | OVS_REQUIRES(pmd->port_mutex); | |
656 | static void dp_netdev_del_rxq_from_pmd(struct dp_netdev_pmd_thread *pmd, | |
657 | struct rxq_poll *poll) | |
658 | OVS_REQUIRES(pmd->port_mutex); | |
659 | static void reconfigure_datapath(struct dp_netdev *dp) | |
660 | OVS_REQUIRES(dp->port_mutex); | |
661 | static bool dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd); | |
662 | static void dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd); | |
663 | static void dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd); | |
664 | static void pmd_load_cached_ports(struct dp_netdev_pmd_thread *pmd) | |
665 | OVS_REQUIRES(pmd->port_mutex); | |
666 | static inline void | |
667 | dp_netdev_pmd_try_optimize(struct dp_netdev_pmd_thread *pmd); | |
668 | ||
669 | static void | |
670 | dpif_netdev_xps_revalidate_pmd(const struct dp_netdev_pmd_thread *pmd, | |
671 | long long now, bool purge); | |
672 | static int dpif_netdev_xps_get_tx_qid(const struct dp_netdev_pmd_thread *pmd, | |
673 | struct tx_port *tx, long long now); | |
674 | ||
675 | static inline bool emc_entry_alive(struct emc_entry *ce); | |
676 | static void emc_clear_entry(struct emc_entry *ce); | |
677 | ||
678 | static void | |
679 | emc_cache_init(struct emc_cache *flow_cache) | |
680 | { | |
681 | int i; | |
682 | ||
683 | flow_cache->sweep_idx = 0; | |
684 | for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) { | |
685 | flow_cache->entries[i].flow = NULL; | |
686 | flow_cache->entries[i].key.hash = 0; | |
687 | flow_cache->entries[i].key.len = sizeof(struct miniflow); | |
688 | flowmap_init(&flow_cache->entries[i].key.mf.map); | |
689 | } | |
690 | } | |
691 | ||
692 | static void | |
693 | emc_cache_uninit(struct emc_cache *flow_cache) | |
694 | { | |
695 | int i; | |
696 | ||
697 | for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) { | |
698 | emc_clear_entry(&flow_cache->entries[i]); | |
699 | } | |
700 | } | |
701 | ||
702 | /* Check and clear dead flow references slowly (one entry at each | |
703 | * invocation). */ | |
704 | static void | |
705 | emc_cache_slow_sweep(struct emc_cache *flow_cache) | |
706 | { | |
707 | struct emc_entry *entry = &flow_cache->entries[flow_cache->sweep_idx]; | |
708 | ||
709 | if (!emc_entry_alive(entry)) { | |
710 | emc_clear_entry(entry); | |
711 | } | |
712 | flow_cache->sweep_idx = (flow_cache->sweep_idx + 1) & EM_FLOW_HASH_MASK; | |
713 | } | |
714 | ||
715 | /* Returns true if 'dpif' is a netdev or dummy dpif, false otherwise. */ | |
716 | bool | |
717 | dpif_is_netdev(const struct dpif *dpif) | |
718 | { | |
719 | return dpif->dpif_class->open == dpif_netdev_open; | |
720 | } | |
721 | ||
722 | static struct dpif_netdev * | |
723 | dpif_netdev_cast(const struct dpif *dpif) | |
724 | { | |
725 | ovs_assert(dpif_is_netdev(dpif)); | |
726 | return CONTAINER_OF(dpif, struct dpif_netdev, dpif); | |
727 | } | |
728 | ||
729 | static struct dp_netdev * | |
730 | get_dp_netdev(const struct dpif *dpif) | |
731 | { | |
732 | return dpif_netdev_cast(dpif)->dp; | |
733 | } | |
734 | \f | |
735 | enum pmd_info_type { | |
736 | PMD_INFO_SHOW_STATS, /* Show how cpu cycles are spent. */ | |
737 | PMD_INFO_CLEAR_STATS, /* Set the cycles count to 0. */ | |
738 | PMD_INFO_SHOW_RXQ /* Show poll-lists of pmd threads. */ | |
739 | }; | |
740 | ||
741 | static void | |
742 | pmd_info_show_stats(struct ds *reply, | |
743 | struct dp_netdev_pmd_thread *pmd, | |
744 | unsigned long long stats[DP_N_STATS], | |
745 | uint64_t cycles[PMD_N_CYCLES]) | |
746 | { | |
747 | unsigned long long total_packets = 0; | |
748 | uint64_t total_cycles = 0; | |
749 | int i; | |
750 | ||
751 | /* These loops subtracts reference values ('*_zero') from the counters. | |
752 | * Since loads and stores are relaxed, it might be possible for a '*_zero' | |
753 | * value to be more recent than the current value we're reading from the | |
754 | * counter. This is not a big problem, since these numbers are not | |
755 | * supposed to be too accurate, but we should at least make sure that | |
756 | * the result is not negative. */ | |
757 | for (i = 0; i < DP_N_STATS; i++) { | |
758 | if (stats[i] > pmd->stats_zero[i]) { | |
759 | stats[i] -= pmd->stats_zero[i]; | |
760 | } else { | |
761 | stats[i] = 0; | |
762 | } | |
763 | ||
764 | if (i != DP_STAT_LOST) { | |
765 | /* Lost packets are already included in DP_STAT_MISS */ | |
766 | total_packets += stats[i]; | |
767 | } | |
768 | } | |
769 | ||
770 | for (i = 0; i < PMD_N_CYCLES; i++) { | |
771 | if (cycles[i] > pmd->cycles_zero[i]) { | |
772 | cycles[i] -= pmd->cycles_zero[i]; | |
773 | } else { | |
774 | cycles[i] = 0; | |
775 | } | |
776 | ||
777 | total_cycles += cycles[i]; | |
778 | } | |
779 | ||
780 | ds_put_cstr(reply, (pmd->core_id == NON_PMD_CORE_ID) | |
781 | ? "main thread" : "pmd thread"); | |
782 | ||
783 | if (pmd->numa_id != OVS_NUMA_UNSPEC) { | |
784 | ds_put_format(reply, " numa_id %d", pmd->numa_id); | |
785 | } | |
786 | if (pmd->core_id != OVS_CORE_UNSPEC && pmd->core_id != NON_PMD_CORE_ID) { | |
787 | ds_put_format(reply, " core_id %u", pmd->core_id); | |
788 | } | |
789 | ds_put_cstr(reply, ":\n"); | |
790 | ||
791 | ds_put_format(reply, | |
792 | "\temc hits:%llu\n\tmegaflow hits:%llu\n" | |
793 | "\tavg. subtable lookups per hit:%.2f\n" | |
794 | "\tmiss:%llu\n\tlost:%llu\n", | |
795 | stats[DP_STAT_EXACT_HIT], stats[DP_STAT_MASKED_HIT], | |
796 | stats[DP_STAT_MASKED_HIT] > 0 | |
797 | ? (1.0*stats[DP_STAT_LOOKUP_HIT])/stats[DP_STAT_MASKED_HIT] | |
798 | : 0, | |
799 | stats[DP_STAT_MISS], stats[DP_STAT_LOST]); | |
800 | ||
801 | if (total_cycles == 0) { | |
802 | return; | |
803 | } | |
804 | ||
805 | ds_put_format(reply, | |
806 | "\tpolling cycles:%"PRIu64" (%.02f%%)\n" | |
807 | "\tprocessing cycles:%"PRIu64" (%.02f%%)\n", | |
808 | cycles[PMD_CYCLES_POLLING], | |
809 | cycles[PMD_CYCLES_POLLING] / (double)total_cycles * 100, | |
810 | cycles[PMD_CYCLES_PROCESSING], | |
811 | cycles[PMD_CYCLES_PROCESSING] / (double)total_cycles * 100); | |
812 | ||
813 | if (total_packets == 0) { | |
814 | return; | |
815 | } | |
816 | ||
817 | ds_put_format(reply, | |
818 | "\tavg cycles per packet: %.02f (%"PRIu64"/%llu)\n", | |
819 | total_cycles / (double)total_packets, | |
820 | total_cycles, total_packets); | |
821 | ||
822 | ds_put_format(reply, | |
823 | "\tavg processing cycles per packet: " | |
824 | "%.02f (%"PRIu64"/%llu)\n", | |
825 | cycles[PMD_CYCLES_PROCESSING] / (double)total_packets, | |
826 | cycles[PMD_CYCLES_PROCESSING], total_packets); | |
827 | } | |
828 | ||
829 | static void | |
830 | pmd_info_clear_stats(struct ds *reply OVS_UNUSED, | |
831 | struct dp_netdev_pmd_thread *pmd, | |
832 | unsigned long long stats[DP_N_STATS], | |
833 | uint64_t cycles[PMD_N_CYCLES]) | |
834 | { | |
835 | int i; | |
836 | ||
837 | /* We cannot write 'stats' and 'cycles' (because they're written by other | |
838 | * threads) and we shouldn't change 'stats' (because they're used to count | |
839 | * datapath stats, which must not be cleared here). Instead, we save the | |
840 | * current values and subtract them from the values to be displayed in the | |
841 | * future */ | |
842 | for (i = 0; i < DP_N_STATS; i++) { | |
843 | pmd->stats_zero[i] = stats[i]; | |
844 | } | |
845 | for (i = 0; i < PMD_N_CYCLES; i++) { | |
846 | pmd->cycles_zero[i] = cycles[i]; | |
847 | } | |
848 | } | |
849 | ||
850 | static int | |
851 | compare_poll_list(const void *a_, const void *b_) | |
852 | { | |
853 | const struct rxq_poll *a = a_; | |
854 | const struct rxq_poll *b = b_; | |
855 | ||
856 | const char *namea = netdev_rxq_get_name(a->rxq->rx); | |
857 | const char *nameb = netdev_rxq_get_name(b->rxq->rx); | |
858 | ||
859 | int cmp = strcmp(namea, nameb); | |
860 | if (!cmp) { | |
861 | return netdev_rxq_get_queue_id(a->rxq->rx) | |
862 | - netdev_rxq_get_queue_id(b->rxq->rx); | |
863 | } else { | |
864 | return cmp; | |
865 | } | |
866 | } | |
867 | ||
868 | static void | |
869 | sorted_poll_list(struct dp_netdev_pmd_thread *pmd, struct rxq_poll **list, | |
870 | size_t *n) | |
871 | { | |
872 | struct rxq_poll *ret, *poll; | |
873 | size_t i; | |
874 | ||
875 | *n = hmap_count(&pmd->poll_list); | |
876 | if (!*n) { | |
877 | ret = NULL; | |
878 | } else { | |
879 | ret = xcalloc(*n, sizeof *ret); | |
880 | i = 0; | |
881 | HMAP_FOR_EACH (poll, node, &pmd->poll_list) { | |
882 | ret[i] = *poll; | |
883 | i++; | |
884 | } | |
885 | ovs_assert(i == *n); | |
886 | } | |
887 | ||
888 | qsort(ret, *n, sizeof *ret, compare_poll_list); | |
889 | ||
890 | *list = ret; | |
891 | } | |
892 | ||
893 | static void | |
894 | pmd_info_show_rxq(struct ds *reply, struct dp_netdev_pmd_thread *pmd) | |
895 | { | |
896 | if (pmd->core_id != NON_PMD_CORE_ID) { | |
897 | const char *prev_name = NULL; | |
898 | struct rxq_poll *list; | |
899 | size_t i, n; | |
900 | ||
901 | ds_put_format(reply, | |
902 | "pmd thread numa_id %d core_id %u:\n\tisolated : %s\n", | |
903 | pmd->numa_id, pmd->core_id, (pmd->isolated) | |
904 | ? "true" : "false"); | |
905 | ||
906 | ovs_mutex_lock(&pmd->port_mutex); | |
907 | sorted_poll_list(pmd, &list, &n); | |
908 | for (i = 0; i < n; i++) { | |
909 | const char *name = netdev_rxq_get_name(list[i].rxq->rx); | |
910 | ||
911 | if (!prev_name || strcmp(name, prev_name)) { | |
912 | if (prev_name) { | |
913 | ds_put_cstr(reply, "\n"); | |
914 | } | |
915 | ds_put_format(reply, "\tport: %s\tqueue-id:", name); | |
916 | } | |
917 | ds_put_format(reply, " %d", | |
918 | netdev_rxq_get_queue_id(list[i].rxq->rx)); | |
919 | prev_name = name; | |
920 | } | |
921 | ovs_mutex_unlock(&pmd->port_mutex); | |
922 | ds_put_cstr(reply, "\n"); | |
923 | free(list); | |
924 | } | |
925 | } | |
926 | ||
927 | static int | |
928 | compare_poll_thread_list(const void *a_, const void *b_) | |
929 | { | |
930 | const struct dp_netdev_pmd_thread *a, *b; | |
931 | ||
932 | a = *(struct dp_netdev_pmd_thread **)a_; | |
933 | b = *(struct dp_netdev_pmd_thread **)b_; | |
934 | ||
935 | if (a->core_id < b->core_id) { | |
936 | return -1; | |
937 | } | |
938 | if (a->core_id > b->core_id) { | |
939 | return 1; | |
940 | } | |
941 | return 0; | |
942 | } | |
943 | ||
944 | /* Create a sorted list of pmd's from the dp->poll_threads cmap. We can use | |
945 | * this list, as long as we do not go to quiescent state. */ | |
946 | static void | |
947 | sorted_poll_thread_list(struct dp_netdev *dp, | |
948 | struct dp_netdev_pmd_thread ***list, | |
949 | size_t *n) | |
950 | { | |
951 | struct dp_netdev_pmd_thread *pmd; | |
952 | struct dp_netdev_pmd_thread **pmd_list; | |
953 | size_t k = 0, n_pmds; | |
954 | ||
955 | n_pmds = cmap_count(&dp->poll_threads); | |
956 | pmd_list = xcalloc(n_pmds, sizeof *pmd_list); | |
957 | ||
958 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
959 | if (k >= n_pmds) { | |
960 | break; | |
961 | } | |
962 | pmd_list[k++] = pmd; | |
963 | } | |
964 | ||
965 | qsort(pmd_list, k, sizeof *pmd_list, compare_poll_thread_list); | |
966 | ||
967 | *list = pmd_list; | |
968 | *n = k; | |
969 | } | |
970 | ||
971 | static void | |
972 | dpif_netdev_pmd_info(struct unixctl_conn *conn, int argc, const char *argv[], | |
973 | void *aux) | |
974 | { | |
975 | struct ds reply = DS_EMPTY_INITIALIZER; | |
976 | struct dp_netdev_pmd_thread **pmd_list; | |
977 | struct dp_netdev *dp = NULL; | |
978 | size_t n; | |
979 | enum pmd_info_type type = *(enum pmd_info_type *) aux; | |
980 | ||
981 | ovs_mutex_lock(&dp_netdev_mutex); | |
982 | ||
983 | if (argc == 2) { | |
984 | dp = shash_find_data(&dp_netdevs, argv[1]); | |
985 | } else if (shash_count(&dp_netdevs) == 1) { | |
986 | /* There's only one datapath */ | |
987 | dp = shash_first(&dp_netdevs)->data; | |
988 | } | |
989 | ||
990 | if (!dp) { | |
991 | ovs_mutex_unlock(&dp_netdev_mutex); | |
992 | unixctl_command_reply_error(conn, | |
993 | "please specify an existing datapath"); | |
994 | return; | |
995 | } | |
996 | ||
997 | sorted_poll_thread_list(dp, &pmd_list, &n); | |
998 | for (size_t i = 0; i < n; i++) { | |
999 | struct dp_netdev_pmd_thread *pmd = pmd_list[i]; | |
1000 | if (!pmd) { | |
1001 | break; | |
1002 | } | |
1003 | ||
1004 | if (type == PMD_INFO_SHOW_RXQ) { | |
1005 | pmd_info_show_rxq(&reply, pmd); | |
1006 | } else { | |
1007 | unsigned long long stats[DP_N_STATS]; | |
1008 | uint64_t cycles[PMD_N_CYCLES]; | |
1009 | int i; | |
1010 | ||
1011 | /* Read current stats and cycle counters */ | |
1012 | for (i = 0; i < ARRAY_SIZE(stats); i++) { | |
1013 | atomic_read_relaxed(&pmd->stats.n[i], &stats[i]); | |
1014 | } | |
1015 | for (i = 0; i < ARRAY_SIZE(cycles); i++) { | |
1016 | atomic_read_relaxed(&pmd->cycles.n[i], &cycles[i]); | |
1017 | } | |
1018 | ||
1019 | if (type == PMD_INFO_CLEAR_STATS) { | |
1020 | pmd_info_clear_stats(&reply, pmd, stats, cycles); | |
1021 | } else if (type == PMD_INFO_SHOW_STATS) { | |
1022 | pmd_info_show_stats(&reply, pmd, stats, cycles); | |
1023 | } | |
1024 | } | |
1025 | } | |
1026 | free(pmd_list); | |
1027 | ||
1028 | ovs_mutex_unlock(&dp_netdev_mutex); | |
1029 | ||
1030 | unixctl_command_reply(conn, ds_cstr(&reply)); | |
1031 | ds_destroy(&reply); | |
1032 | } | |
1033 | \f | |
1034 | static int | |
1035 | dpif_netdev_init(void) | |
1036 | { | |
1037 | static enum pmd_info_type show_aux = PMD_INFO_SHOW_STATS, | |
1038 | clear_aux = PMD_INFO_CLEAR_STATS, | |
1039 | poll_aux = PMD_INFO_SHOW_RXQ; | |
1040 | ||
1041 | unixctl_command_register("dpif-netdev/pmd-stats-show", "[dp]", | |
1042 | 0, 1, dpif_netdev_pmd_info, | |
1043 | (void *)&show_aux); | |
1044 | unixctl_command_register("dpif-netdev/pmd-stats-clear", "[dp]", | |
1045 | 0, 1, dpif_netdev_pmd_info, | |
1046 | (void *)&clear_aux); | |
1047 | unixctl_command_register("dpif-netdev/pmd-rxq-show", "[dp]", | |
1048 | 0, 1, dpif_netdev_pmd_info, | |
1049 | (void *)&poll_aux); | |
1050 | return 0; | |
1051 | } | |
1052 | ||
1053 | static int | |
1054 | dpif_netdev_enumerate(struct sset *all_dps, | |
1055 | const struct dpif_class *dpif_class) | |
1056 | { | |
1057 | struct shash_node *node; | |
1058 | ||
1059 | ovs_mutex_lock(&dp_netdev_mutex); | |
1060 | SHASH_FOR_EACH(node, &dp_netdevs) { | |
1061 | struct dp_netdev *dp = node->data; | |
1062 | if (dpif_class != dp->class) { | |
1063 | /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs. | |
1064 | * If the class doesn't match, skip this dpif. */ | |
1065 | continue; | |
1066 | } | |
1067 | sset_add(all_dps, node->name); | |
1068 | } | |
1069 | ovs_mutex_unlock(&dp_netdev_mutex); | |
1070 | ||
1071 | return 0; | |
1072 | } | |
1073 | ||
1074 | static bool | |
1075 | dpif_netdev_class_is_dummy(const struct dpif_class *class) | |
1076 | { | |
1077 | return class != &dpif_netdev_class; | |
1078 | } | |
1079 | ||
1080 | static const char * | |
1081 | dpif_netdev_port_open_type(const struct dpif_class *class, const char *type) | |
1082 | { | |
1083 | return strcmp(type, "internal") ? type | |
1084 | : dpif_netdev_class_is_dummy(class) ? "dummy-internal" | |
1085 | : "tap"; | |
1086 | } | |
1087 | ||
1088 | static struct dpif * | |
1089 | create_dpif_netdev(struct dp_netdev *dp) | |
1090 | { | |
1091 | uint16_t netflow_id = hash_string(dp->name, 0); | |
1092 | struct dpif_netdev *dpif; | |
1093 | ||
1094 | ovs_refcount_ref(&dp->ref_cnt); | |
1095 | ||
1096 | dpif = xmalloc(sizeof *dpif); | |
1097 | dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id); | |
1098 | dpif->dp = dp; | |
1099 | dpif->last_port_seq = seq_read(dp->port_seq); | |
1100 | ||
1101 | return &dpif->dpif; | |
1102 | } | |
1103 | ||
1104 | /* Choose an unused, non-zero port number and return it on success. | |
1105 | * Return ODPP_NONE on failure. */ | |
1106 | static odp_port_t | |
1107 | choose_port(struct dp_netdev *dp, const char *name) | |
1108 | OVS_REQUIRES(dp->port_mutex) | |
1109 | { | |
1110 | uint32_t port_no; | |
1111 | ||
1112 | if (dp->class != &dpif_netdev_class) { | |
1113 | const char *p; | |
1114 | int start_no = 0; | |
1115 | ||
1116 | /* If the port name begins with "br", start the number search at | |
1117 | * 100 to make writing tests easier. */ | |
1118 | if (!strncmp(name, "br", 2)) { | |
1119 | start_no = 100; | |
1120 | } | |
1121 | ||
1122 | /* If the port name contains a number, try to assign that port number. | |
1123 | * This can make writing unit tests easier because port numbers are | |
1124 | * predictable. */ | |
1125 | for (p = name; *p != '\0'; p++) { | |
1126 | if (isdigit((unsigned char) *p)) { | |
1127 | port_no = start_no + strtol(p, NULL, 10); | |
1128 | if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE) | |
1129 | && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) { | |
1130 | return u32_to_odp(port_no); | |
1131 | } | |
1132 | break; | |
1133 | } | |
1134 | } | |
1135 | } | |
1136 | ||
1137 | for (port_no = 1; port_no <= UINT16_MAX; port_no++) { | |
1138 | if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) { | |
1139 | return u32_to_odp(port_no); | |
1140 | } | |
1141 | } | |
1142 | ||
1143 | return ODPP_NONE; | |
1144 | } | |
1145 | ||
1146 | static int | |
1147 | create_dp_netdev(const char *name, const struct dpif_class *class, | |
1148 | struct dp_netdev **dpp) | |
1149 | OVS_REQUIRES(dp_netdev_mutex) | |
1150 | { | |
1151 | struct dp_netdev *dp; | |
1152 | int error; | |
1153 | ||
1154 | dp = xzalloc(sizeof *dp); | |
1155 | shash_add(&dp_netdevs, name, dp); | |
1156 | ||
1157 | *CONST_CAST(const struct dpif_class **, &dp->class) = class; | |
1158 | *CONST_CAST(const char **, &dp->name) = xstrdup(name); | |
1159 | ovs_refcount_init(&dp->ref_cnt); | |
1160 | atomic_flag_clear(&dp->destroyed); | |
1161 | ||
1162 | ovs_mutex_init(&dp->port_mutex); | |
1163 | hmap_init(&dp->ports); | |
1164 | dp->port_seq = seq_create(); | |
1165 | fat_rwlock_init(&dp->upcall_rwlock); | |
1166 | ||
1167 | dp->reconfigure_seq = seq_create(); | |
1168 | dp->last_reconfigure_seq = seq_read(dp->reconfigure_seq); | |
1169 | ||
1170 | for (int i = 0; i < N_METER_LOCKS; ++i) { | |
1171 | ovs_mutex_init_adaptive(&dp->meter_locks[i]); | |
1172 | } | |
1173 | ||
1174 | /* Disable upcalls by default. */ | |
1175 | dp_netdev_disable_upcall(dp); | |
1176 | dp->upcall_aux = NULL; | |
1177 | dp->upcall_cb = NULL; | |
1178 | ||
1179 | conntrack_init(&dp->conntrack); | |
1180 | ||
1181 | atomic_init(&dp->emc_insert_min, DEFAULT_EM_FLOW_INSERT_MIN); | |
1182 | ||
1183 | cmap_init(&dp->poll_threads); | |
1184 | ovs_mutex_init_recursive(&dp->non_pmd_mutex); | |
1185 | ovsthread_key_create(&dp->per_pmd_key, NULL); | |
1186 | ||
1187 | ovs_mutex_lock(&dp->port_mutex); | |
1188 | dp_netdev_set_nonpmd(dp); | |
1189 | ||
1190 | error = do_add_port(dp, name, dpif_netdev_port_open_type(dp->class, | |
1191 | "internal"), | |
1192 | ODPP_LOCAL); | |
1193 | ovs_mutex_unlock(&dp->port_mutex); | |
1194 | if (error) { | |
1195 | dp_netdev_free(dp); | |
1196 | return error; | |
1197 | } | |
1198 | ||
1199 | dp->last_tnl_conf_seq = seq_read(tnl_conf_seq); | |
1200 | *dpp = dp; | |
1201 | return 0; | |
1202 | } | |
1203 | ||
1204 | static void | |
1205 | dp_netdev_request_reconfigure(struct dp_netdev *dp) | |
1206 | { | |
1207 | seq_change(dp->reconfigure_seq); | |
1208 | } | |
1209 | ||
1210 | static bool | |
1211 | dp_netdev_is_reconf_required(struct dp_netdev *dp) | |
1212 | { | |
1213 | return seq_read(dp->reconfigure_seq) != dp->last_reconfigure_seq; | |
1214 | } | |
1215 | ||
1216 | static int | |
1217 | dpif_netdev_open(const struct dpif_class *class, const char *name, | |
1218 | bool create, struct dpif **dpifp) | |
1219 | { | |
1220 | struct dp_netdev *dp; | |
1221 | int error; | |
1222 | ||
1223 | ovs_mutex_lock(&dp_netdev_mutex); | |
1224 | dp = shash_find_data(&dp_netdevs, name); | |
1225 | if (!dp) { | |
1226 | error = create ? create_dp_netdev(name, class, &dp) : ENODEV; | |
1227 | } else { | |
1228 | error = (dp->class != class ? EINVAL | |
1229 | : create ? EEXIST | |
1230 | : 0); | |
1231 | } | |
1232 | if (!error) { | |
1233 | *dpifp = create_dpif_netdev(dp); | |
1234 | dp->dpif = *dpifp; | |
1235 | } | |
1236 | ovs_mutex_unlock(&dp_netdev_mutex); | |
1237 | ||
1238 | return error; | |
1239 | } | |
1240 | ||
1241 | static void | |
1242 | dp_netdev_destroy_upcall_lock(struct dp_netdev *dp) | |
1243 | OVS_NO_THREAD_SAFETY_ANALYSIS | |
1244 | { | |
1245 | /* Check that upcalls are disabled, i.e. that the rwlock is taken */ | |
1246 | ovs_assert(fat_rwlock_tryrdlock(&dp->upcall_rwlock)); | |
1247 | ||
1248 | /* Before freeing a lock we should release it */ | |
1249 | fat_rwlock_unlock(&dp->upcall_rwlock); | |
1250 | fat_rwlock_destroy(&dp->upcall_rwlock); | |
1251 | } | |
1252 | ||
1253 | static void | |
1254 | dp_delete_meter(struct dp_netdev *dp, uint32_t meter_id) | |
1255 | OVS_REQUIRES(dp->meter_locks[meter_id % N_METER_LOCKS]) | |
1256 | { | |
1257 | if (dp->meters[meter_id]) { | |
1258 | free(dp->meters[meter_id]); | |
1259 | dp->meters[meter_id] = NULL; | |
1260 | } | |
1261 | } | |
1262 | ||
1263 | /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp' | |
1264 | * through the 'dp_netdevs' shash while freeing 'dp'. */ | |
1265 | static void | |
1266 | dp_netdev_free(struct dp_netdev *dp) | |
1267 | OVS_REQUIRES(dp_netdev_mutex) | |
1268 | { | |
1269 | struct dp_netdev_port *port, *next; | |
1270 | ||
1271 | shash_find_and_delete(&dp_netdevs, dp->name); | |
1272 | ||
1273 | ovs_mutex_lock(&dp->port_mutex); | |
1274 | HMAP_FOR_EACH_SAFE (port, next, node, &dp->ports) { | |
1275 | do_del_port(dp, port); | |
1276 | } | |
1277 | ovs_mutex_unlock(&dp->port_mutex); | |
1278 | ||
1279 | dp_netdev_destroy_all_pmds(dp, true); | |
1280 | cmap_destroy(&dp->poll_threads); | |
1281 | ||
1282 | ovs_mutex_destroy(&dp->non_pmd_mutex); | |
1283 | ovsthread_key_delete(dp->per_pmd_key); | |
1284 | ||
1285 | conntrack_destroy(&dp->conntrack); | |
1286 | ||
1287 | ||
1288 | seq_destroy(dp->reconfigure_seq); | |
1289 | ||
1290 | seq_destroy(dp->port_seq); | |
1291 | hmap_destroy(&dp->ports); | |
1292 | ovs_mutex_destroy(&dp->port_mutex); | |
1293 | ||
1294 | /* Upcalls must be disabled at this point */ | |
1295 | dp_netdev_destroy_upcall_lock(dp); | |
1296 | ||
1297 | int i; | |
1298 | ||
1299 | for (i = 0; i < MAX_METERS; ++i) { | |
1300 | meter_lock(dp, i); | |
1301 | dp_delete_meter(dp, i); | |
1302 | meter_unlock(dp, i); | |
1303 | } | |
1304 | for (i = 0; i < N_METER_LOCKS; ++i) { | |
1305 | ovs_mutex_destroy(&dp->meter_locks[i]); | |
1306 | } | |
1307 | ||
1308 | free(dp->pmd_cmask); | |
1309 | free(CONST_CAST(char *, dp->name)); | |
1310 | free(dp); | |
1311 | } | |
1312 | ||
1313 | static void | |
1314 | dp_netdev_unref(struct dp_netdev *dp) | |
1315 | { | |
1316 | if (dp) { | |
1317 | /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't | |
1318 | * get a new reference to 'dp' through the 'dp_netdevs' shash. */ | |
1319 | ovs_mutex_lock(&dp_netdev_mutex); | |
1320 | if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) { | |
1321 | dp_netdev_free(dp); | |
1322 | } | |
1323 | ovs_mutex_unlock(&dp_netdev_mutex); | |
1324 | } | |
1325 | } | |
1326 | ||
1327 | static void | |
1328 | dpif_netdev_close(struct dpif *dpif) | |
1329 | { | |
1330 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
1331 | ||
1332 | dp_netdev_unref(dp); | |
1333 | free(dpif); | |
1334 | } | |
1335 | ||
1336 | static int | |
1337 | dpif_netdev_destroy(struct dpif *dpif) | |
1338 | { | |
1339 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
1340 | ||
1341 | if (!atomic_flag_test_and_set(&dp->destroyed)) { | |
1342 | if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) { | |
1343 | /* Can't happen: 'dpif' still owns a reference to 'dp'. */ | |
1344 | OVS_NOT_REACHED(); | |
1345 | } | |
1346 | } | |
1347 | ||
1348 | return 0; | |
1349 | } | |
1350 | ||
1351 | /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed | |
1352 | * load/store semantics. While the increment is not atomic, the load and | |
1353 | * store operations are, making it impossible to read inconsistent values. | |
1354 | * | |
1355 | * This is used to update thread local stats counters. */ | |
1356 | static void | |
1357 | non_atomic_ullong_add(atomic_ullong *var, unsigned long long n) | |
1358 | { | |
1359 | unsigned long long tmp; | |
1360 | ||
1361 | atomic_read_relaxed(var, &tmp); | |
1362 | tmp += n; | |
1363 | atomic_store_relaxed(var, tmp); | |
1364 | } | |
1365 | ||
1366 | static int | |
1367 | dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats) | |
1368 | { | |
1369 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
1370 | struct dp_netdev_pmd_thread *pmd; | |
1371 | ||
1372 | stats->n_flows = stats->n_hit = stats->n_missed = stats->n_lost = 0; | |
1373 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
1374 | unsigned long long n; | |
1375 | stats->n_flows += cmap_count(&pmd->flow_table); | |
1376 | ||
1377 | atomic_read_relaxed(&pmd->stats.n[DP_STAT_MASKED_HIT], &n); | |
1378 | stats->n_hit += n; | |
1379 | atomic_read_relaxed(&pmd->stats.n[DP_STAT_EXACT_HIT], &n); | |
1380 | stats->n_hit += n; | |
1381 | atomic_read_relaxed(&pmd->stats.n[DP_STAT_MISS], &n); | |
1382 | stats->n_missed += n; | |
1383 | atomic_read_relaxed(&pmd->stats.n[DP_STAT_LOST], &n); | |
1384 | stats->n_lost += n; | |
1385 | } | |
1386 | stats->n_masks = UINT32_MAX; | |
1387 | stats->n_mask_hit = UINT64_MAX; | |
1388 | ||
1389 | return 0; | |
1390 | } | |
1391 | ||
1392 | static void | |
1393 | dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread *pmd) | |
1394 | { | |
1395 | if (pmd->core_id == NON_PMD_CORE_ID) { | |
1396 | ovs_mutex_lock(&pmd->dp->non_pmd_mutex); | |
1397 | ovs_mutex_lock(&pmd->port_mutex); | |
1398 | pmd_load_cached_ports(pmd); | |
1399 | ovs_mutex_unlock(&pmd->port_mutex); | |
1400 | ovs_mutex_unlock(&pmd->dp->non_pmd_mutex); | |
1401 | return; | |
1402 | } | |
1403 | ||
1404 | ovs_mutex_lock(&pmd->cond_mutex); | |
1405 | seq_change(pmd->reload_seq); | |
1406 | atomic_store_relaxed(&pmd->reload, true); | |
1407 | ovs_mutex_cond_wait(&pmd->cond, &pmd->cond_mutex); | |
1408 | ovs_mutex_unlock(&pmd->cond_mutex); | |
1409 | } | |
1410 | ||
1411 | static uint32_t | |
1412 | hash_port_no(odp_port_t port_no) | |
1413 | { | |
1414 | return hash_int(odp_to_u32(port_no), 0); | |
1415 | } | |
1416 | ||
1417 | static int | |
1418 | port_create(const char *devname, const char *type, | |
1419 | odp_port_t port_no, struct dp_netdev_port **portp) | |
1420 | { | |
1421 | struct netdev_saved_flags *sf; | |
1422 | struct dp_netdev_port *port; | |
1423 | enum netdev_flags flags; | |
1424 | struct netdev *netdev; | |
1425 | int error; | |
1426 | ||
1427 | *portp = NULL; | |
1428 | ||
1429 | /* Open and validate network device. */ | |
1430 | error = netdev_open(devname, type, &netdev); | |
1431 | if (error) { | |
1432 | return error; | |
1433 | } | |
1434 | /* XXX reject non-Ethernet devices */ | |
1435 | ||
1436 | netdev_get_flags(netdev, &flags); | |
1437 | if (flags & NETDEV_LOOPBACK) { | |
1438 | VLOG_ERR("%s: cannot add a loopback device", devname); | |
1439 | error = EINVAL; | |
1440 | goto out; | |
1441 | } | |
1442 | ||
1443 | error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf); | |
1444 | if (error) { | |
1445 | VLOG_ERR("%s: cannot set promisc flag", devname); | |
1446 | goto out; | |
1447 | } | |
1448 | ||
1449 | port = xzalloc(sizeof *port); | |
1450 | port->port_no = port_no; | |
1451 | port->netdev = netdev; | |
1452 | port->type = xstrdup(type); | |
1453 | port->sf = sf; | |
1454 | port->need_reconfigure = true; | |
1455 | ovs_mutex_init(&port->txq_used_mutex); | |
1456 | ||
1457 | *portp = port; | |
1458 | ||
1459 | return 0; | |
1460 | ||
1461 | out: | |
1462 | netdev_close(netdev); | |
1463 | return error; | |
1464 | } | |
1465 | ||
1466 | static int | |
1467 | do_add_port(struct dp_netdev *dp, const char *devname, const char *type, | |
1468 | odp_port_t port_no) | |
1469 | OVS_REQUIRES(dp->port_mutex) | |
1470 | { | |
1471 | struct dp_netdev_port *port; | |
1472 | int error; | |
1473 | ||
1474 | /* Reject devices already in 'dp'. */ | |
1475 | if (!get_port_by_name(dp, devname, &port)) { | |
1476 | return EEXIST; | |
1477 | } | |
1478 | ||
1479 | error = port_create(devname, type, port_no, &port); | |
1480 | if (error) { | |
1481 | return error; | |
1482 | } | |
1483 | ||
1484 | hmap_insert(&dp->ports, &port->node, hash_port_no(port_no)); | |
1485 | seq_change(dp->port_seq); | |
1486 | ||
1487 | reconfigure_datapath(dp); | |
1488 | ||
1489 | return 0; | |
1490 | } | |
1491 | ||
1492 | static int | |
1493 | dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev, | |
1494 | odp_port_t *port_nop) | |
1495 | { | |
1496 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
1497 | char namebuf[NETDEV_VPORT_NAME_BUFSIZE]; | |
1498 | const char *dpif_port; | |
1499 | odp_port_t port_no; | |
1500 | int error; | |
1501 | ||
1502 | ovs_mutex_lock(&dp->port_mutex); | |
1503 | dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf); | |
1504 | if (*port_nop != ODPP_NONE) { | |
1505 | port_no = *port_nop; | |
1506 | error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0; | |
1507 | } else { | |
1508 | port_no = choose_port(dp, dpif_port); | |
1509 | error = port_no == ODPP_NONE ? EFBIG : 0; | |
1510 | } | |
1511 | if (!error) { | |
1512 | *port_nop = port_no; | |
1513 | error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no); | |
1514 | } | |
1515 | ovs_mutex_unlock(&dp->port_mutex); | |
1516 | ||
1517 | return error; | |
1518 | } | |
1519 | ||
1520 | static int | |
1521 | dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no) | |
1522 | { | |
1523 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
1524 | int error; | |
1525 | ||
1526 | ovs_mutex_lock(&dp->port_mutex); | |
1527 | if (port_no == ODPP_LOCAL) { | |
1528 | error = EINVAL; | |
1529 | } else { | |
1530 | struct dp_netdev_port *port; | |
1531 | ||
1532 | error = get_port_by_number(dp, port_no, &port); | |
1533 | if (!error) { | |
1534 | do_del_port(dp, port); | |
1535 | } | |
1536 | } | |
1537 | ovs_mutex_unlock(&dp->port_mutex); | |
1538 | ||
1539 | return error; | |
1540 | } | |
1541 | ||
1542 | static bool | |
1543 | is_valid_port_number(odp_port_t port_no) | |
1544 | { | |
1545 | return port_no != ODPP_NONE; | |
1546 | } | |
1547 | ||
1548 | static struct dp_netdev_port * | |
1549 | dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no) | |
1550 | OVS_REQUIRES(dp->port_mutex) | |
1551 | { | |
1552 | struct dp_netdev_port *port; | |
1553 | ||
1554 | HMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) { | |
1555 | if (port->port_no == port_no) { | |
1556 | return port; | |
1557 | } | |
1558 | } | |
1559 | return NULL; | |
1560 | } | |
1561 | ||
1562 | static int | |
1563 | get_port_by_number(struct dp_netdev *dp, | |
1564 | odp_port_t port_no, struct dp_netdev_port **portp) | |
1565 | OVS_REQUIRES(dp->port_mutex) | |
1566 | { | |
1567 | if (!is_valid_port_number(port_no)) { | |
1568 | *portp = NULL; | |
1569 | return EINVAL; | |
1570 | } else { | |
1571 | *portp = dp_netdev_lookup_port(dp, port_no); | |
1572 | return *portp ? 0 : ENODEV; | |
1573 | } | |
1574 | } | |
1575 | ||
1576 | static void | |
1577 | port_destroy(struct dp_netdev_port *port) | |
1578 | { | |
1579 | if (!port) { | |
1580 | return; | |
1581 | } | |
1582 | ||
1583 | netdev_close(port->netdev); | |
1584 | netdev_restore_flags(port->sf); | |
1585 | ||
1586 | for (unsigned i = 0; i < port->n_rxq; i++) { | |
1587 | netdev_rxq_close(port->rxqs[i].rx); | |
1588 | } | |
1589 | ovs_mutex_destroy(&port->txq_used_mutex); | |
1590 | free(port->rxq_affinity_list); | |
1591 | free(port->txq_used); | |
1592 | free(port->rxqs); | |
1593 | free(port->type); | |
1594 | free(port); | |
1595 | } | |
1596 | ||
1597 | static int | |
1598 | get_port_by_name(struct dp_netdev *dp, | |
1599 | const char *devname, struct dp_netdev_port **portp) | |
1600 | OVS_REQUIRES(dp->port_mutex) | |
1601 | { | |
1602 | struct dp_netdev_port *port; | |
1603 | ||
1604 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
1605 | if (!strcmp(netdev_get_name(port->netdev), devname)) { | |
1606 | *portp = port; | |
1607 | return 0; | |
1608 | } | |
1609 | } | |
1610 | ||
1611 | /* Callers of dpif_netdev_port_query_by_name() expect ENODEV for a non | |
1612 | * existing port. */ | |
1613 | return ENODEV; | |
1614 | } | |
1615 | ||
1616 | /* Returns 'true' if there is a port with pmd netdev. */ | |
1617 | static bool | |
1618 | has_pmd_port(struct dp_netdev *dp) | |
1619 | OVS_REQUIRES(dp->port_mutex) | |
1620 | { | |
1621 | struct dp_netdev_port *port; | |
1622 | ||
1623 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
1624 | if (netdev_is_pmd(port->netdev)) { | |
1625 | return true; | |
1626 | } | |
1627 | } | |
1628 | ||
1629 | return false; | |
1630 | } | |
1631 | ||
1632 | static void | |
1633 | do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port) | |
1634 | OVS_REQUIRES(dp->port_mutex) | |
1635 | { | |
1636 | hmap_remove(&dp->ports, &port->node); | |
1637 | seq_change(dp->port_seq); | |
1638 | ||
1639 | reconfigure_datapath(dp); | |
1640 | ||
1641 | port_destroy(port); | |
1642 | } | |
1643 | ||
1644 | static void | |
1645 | answer_port_query(const struct dp_netdev_port *port, | |
1646 | struct dpif_port *dpif_port) | |
1647 | { | |
1648 | dpif_port->name = xstrdup(netdev_get_name(port->netdev)); | |
1649 | dpif_port->type = xstrdup(port->type); | |
1650 | dpif_port->port_no = port->port_no; | |
1651 | } | |
1652 | ||
1653 | static int | |
1654 | dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no, | |
1655 | struct dpif_port *dpif_port) | |
1656 | { | |
1657 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
1658 | struct dp_netdev_port *port; | |
1659 | int error; | |
1660 | ||
1661 | ovs_mutex_lock(&dp->port_mutex); | |
1662 | error = get_port_by_number(dp, port_no, &port); | |
1663 | if (!error && dpif_port) { | |
1664 | answer_port_query(port, dpif_port); | |
1665 | } | |
1666 | ovs_mutex_unlock(&dp->port_mutex); | |
1667 | ||
1668 | return error; | |
1669 | } | |
1670 | ||
1671 | static int | |
1672 | dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname, | |
1673 | struct dpif_port *dpif_port) | |
1674 | { | |
1675 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
1676 | struct dp_netdev_port *port; | |
1677 | int error; | |
1678 | ||
1679 | ovs_mutex_lock(&dp->port_mutex); | |
1680 | error = get_port_by_name(dp, devname, &port); | |
1681 | if (!error && dpif_port) { | |
1682 | answer_port_query(port, dpif_port); | |
1683 | } | |
1684 | ovs_mutex_unlock(&dp->port_mutex); | |
1685 | ||
1686 | return error; | |
1687 | } | |
1688 | ||
1689 | static void | |
1690 | dp_netdev_flow_free(struct dp_netdev_flow *flow) | |
1691 | { | |
1692 | dp_netdev_actions_free(dp_netdev_flow_get_actions(flow)); | |
1693 | free(flow); | |
1694 | } | |
1695 | ||
1696 | static void dp_netdev_flow_unref(struct dp_netdev_flow *flow) | |
1697 | { | |
1698 | if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) { | |
1699 | ovsrcu_postpone(dp_netdev_flow_free, flow); | |
1700 | } | |
1701 | } | |
1702 | ||
1703 | static uint32_t | |
1704 | dp_netdev_flow_hash(const ovs_u128 *ufid) | |
1705 | { | |
1706 | return ufid->u32[0]; | |
1707 | } | |
1708 | ||
1709 | static inline struct dpcls * | |
1710 | dp_netdev_pmd_lookup_dpcls(struct dp_netdev_pmd_thread *pmd, | |
1711 | odp_port_t in_port) | |
1712 | { | |
1713 | struct dpcls *cls; | |
1714 | uint32_t hash = hash_port_no(in_port); | |
1715 | CMAP_FOR_EACH_WITH_HASH (cls, node, hash, &pmd->classifiers) { | |
1716 | if (cls->in_port == in_port) { | |
1717 | /* Port classifier exists already */ | |
1718 | return cls; | |
1719 | } | |
1720 | } | |
1721 | return NULL; | |
1722 | } | |
1723 | ||
1724 | static inline struct dpcls * | |
1725 | dp_netdev_pmd_find_dpcls(struct dp_netdev_pmd_thread *pmd, | |
1726 | odp_port_t in_port) | |
1727 | OVS_REQUIRES(pmd->flow_mutex) | |
1728 | { | |
1729 | struct dpcls *cls = dp_netdev_pmd_lookup_dpcls(pmd, in_port); | |
1730 | uint32_t hash = hash_port_no(in_port); | |
1731 | ||
1732 | if (!cls) { | |
1733 | /* Create new classifier for in_port */ | |
1734 | cls = xmalloc(sizeof(*cls)); | |
1735 | dpcls_init(cls); | |
1736 | cls->in_port = in_port; | |
1737 | cmap_insert(&pmd->classifiers, &cls->node, hash); | |
1738 | VLOG_DBG("Creating dpcls %p for in_port %d", cls, in_port); | |
1739 | } | |
1740 | return cls; | |
1741 | } | |
1742 | ||
1743 | static void | |
1744 | dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread *pmd, | |
1745 | struct dp_netdev_flow *flow) | |
1746 | OVS_REQUIRES(pmd->flow_mutex) | |
1747 | { | |
1748 | struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node); | |
1749 | struct dpcls *cls; | |
1750 | odp_port_t in_port = flow->flow.in_port.odp_port; | |
1751 | ||
1752 | cls = dp_netdev_pmd_lookup_dpcls(pmd, in_port); | |
1753 | ovs_assert(cls != NULL); | |
1754 | dpcls_remove(cls, &flow->cr); | |
1755 | cmap_remove(&pmd->flow_table, node, dp_netdev_flow_hash(&flow->ufid)); | |
1756 | flow->dead = true; | |
1757 | ||
1758 | dp_netdev_flow_unref(flow); | |
1759 | } | |
1760 | ||
1761 | static void | |
1762 | dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd) | |
1763 | { | |
1764 | struct dp_netdev_flow *netdev_flow; | |
1765 | ||
1766 | ovs_mutex_lock(&pmd->flow_mutex); | |
1767 | CMAP_FOR_EACH (netdev_flow, node, &pmd->flow_table) { | |
1768 | dp_netdev_pmd_remove_flow(pmd, netdev_flow); | |
1769 | } | |
1770 | ovs_mutex_unlock(&pmd->flow_mutex); | |
1771 | } | |
1772 | ||
1773 | static int | |
1774 | dpif_netdev_flow_flush(struct dpif *dpif) | |
1775 | { | |
1776 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
1777 | struct dp_netdev_pmd_thread *pmd; | |
1778 | ||
1779 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
1780 | dp_netdev_pmd_flow_flush(pmd); | |
1781 | } | |
1782 | ||
1783 | return 0; | |
1784 | } | |
1785 | ||
1786 | struct dp_netdev_port_state { | |
1787 | struct hmap_position position; | |
1788 | char *name; | |
1789 | }; | |
1790 | ||
1791 | static int | |
1792 | dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep) | |
1793 | { | |
1794 | *statep = xzalloc(sizeof(struct dp_netdev_port_state)); | |
1795 | return 0; | |
1796 | } | |
1797 | ||
1798 | static int | |
1799 | dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_, | |
1800 | struct dpif_port *dpif_port) | |
1801 | { | |
1802 | struct dp_netdev_port_state *state = state_; | |
1803 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
1804 | struct hmap_node *node; | |
1805 | int retval; | |
1806 | ||
1807 | ovs_mutex_lock(&dp->port_mutex); | |
1808 | node = hmap_at_position(&dp->ports, &state->position); | |
1809 | if (node) { | |
1810 | struct dp_netdev_port *port; | |
1811 | ||
1812 | port = CONTAINER_OF(node, struct dp_netdev_port, node); | |
1813 | ||
1814 | free(state->name); | |
1815 | state->name = xstrdup(netdev_get_name(port->netdev)); | |
1816 | dpif_port->name = state->name; | |
1817 | dpif_port->type = port->type; | |
1818 | dpif_port->port_no = port->port_no; | |
1819 | ||
1820 | retval = 0; | |
1821 | } else { | |
1822 | retval = EOF; | |
1823 | } | |
1824 | ovs_mutex_unlock(&dp->port_mutex); | |
1825 | ||
1826 | return retval; | |
1827 | } | |
1828 | ||
1829 | static int | |
1830 | dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_) | |
1831 | { | |
1832 | struct dp_netdev_port_state *state = state_; | |
1833 | free(state->name); | |
1834 | free(state); | |
1835 | return 0; | |
1836 | } | |
1837 | ||
1838 | static int | |
1839 | dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED) | |
1840 | { | |
1841 | struct dpif_netdev *dpif = dpif_netdev_cast(dpif_); | |
1842 | uint64_t new_port_seq; | |
1843 | int error; | |
1844 | ||
1845 | new_port_seq = seq_read(dpif->dp->port_seq); | |
1846 | if (dpif->last_port_seq != new_port_seq) { | |
1847 | dpif->last_port_seq = new_port_seq; | |
1848 | error = ENOBUFS; | |
1849 | } else { | |
1850 | error = EAGAIN; | |
1851 | } | |
1852 | ||
1853 | return error; | |
1854 | } | |
1855 | ||
1856 | static void | |
1857 | dpif_netdev_port_poll_wait(const struct dpif *dpif_) | |
1858 | { | |
1859 | struct dpif_netdev *dpif = dpif_netdev_cast(dpif_); | |
1860 | ||
1861 | seq_wait(dpif->dp->port_seq, dpif->last_port_seq); | |
1862 | } | |
1863 | ||
1864 | static struct dp_netdev_flow * | |
1865 | dp_netdev_flow_cast(const struct dpcls_rule *cr) | |
1866 | { | |
1867 | return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL; | |
1868 | } | |
1869 | ||
1870 | static bool dp_netdev_flow_ref(struct dp_netdev_flow *flow) | |
1871 | { | |
1872 | return ovs_refcount_try_ref_rcu(&flow->ref_cnt); | |
1873 | } | |
1874 | ||
1875 | /* netdev_flow_key utilities. | |
1876 | * | |
1877 | * netdev_flow_key is basically a miniflow. We use these functions | |
1878 | * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow | |
1879 | * functions (miniflow_clone_inline, miniflow_equal, ...), because: | |
1880 | * | |
1881 | * - Since we are dealing exclusively with miniflows created by | |
1882 | * miniflow_extract(), if the map is different the miniflow is different. | |
1883 | * Therefore we can be faster by comparing the map and the miniflow in a | |
1884 | * single memcmp(). | |
1885 | * - These functions can be inlined by the compiler. */ | |
1886 | ||
1887 | /* Given the number of bits set in miniflow's maps, returns the size of the | |
1888 | * 'netdev_flow_key.mf' */ | |
1889 | static inline size_t | |
1890 | netdev_flow_key_size(size_t flow_u64s) | |
1891 | { | |
1892 | return sizeof(struct miniflow) + MINIFLOW_VALUES_SIZE(flow_u64s); | |
1893 | } | |
1894 | ||
1895 | static inline bool | |
1896 | netdev_flow_key_equal(const struct netdev_flow_key *a, | |
1897 | const struct netdev_flow_key *b) | |
1898 | { | |
1899 | /* 'b->len' may be not set yet. */ | |
1900 | return a->hash == b->hash && !memcmp(&a->mf, &b->mf, a->len); | |
1901 | } | |
1902 | ||
1903 | /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow. | |
1904 | * The maps are compared bitwise, so both 'key->mf' and 'mf' must have been | |
1905 | * generated by miniflow_extract. */ | |
1906 | static inline bool | |
1907 | netdev_flow_key_equal_mf(const struct netdev_flow_key *key, | |
1908 | const struct miniflow *mf) | |
1909 | { | |
1910 | return !memcmp(&key->mf, mf, key->len); | |
1911 | } | |
1912 | ||
1913 | static inline void | |
1914 | netdev_flow_key_clone(struct netdev_flow_key *dst, | |
1915 | const struct netdev_flow_key *src) | |
1916 | { | |
1917 | memcpy(dst, src, | |
1918 | offsetof(struct netdev_flow_key, mf) + src->len); | |
1919 | } | |
1920 | ||
1921 | /* Initialize a netdev_flow_key 'mask' from 'match'. */ | |
1922 | static inline void | |
1923 | netdev_flow_mask_init(struct netdev_flow_key *mask, | |
1924 | const struct match *match) | |
1925 | { | |
1926 | uint64_t *dst = miniflow_values(&mask->mf); | |
1927 | struct flowmap fmap; | |
1928 | uint32_t hash = 0; | |
1929 | size_t idx; | |
1930 | ||
1931 | /* Only check masks that make sense for the flow. */ | |
1932 | flow_wc_map(&match->flow, &fmap); | |
1933 | flowmap_init(&mask->mf.map); | |
1934 | ||
1935 | FLOWMAP_FOR_EACH_INDEX(idx, fmap) { | |
1936 | uint64_t mask_u64 = flow_u64_value(&match->wc.masks, idx); | |
1937 | ||
1938 | if (mask_u64) { | |
1939 | flowmap_set(&mask->mf.map, idx, 1); | |
1940 | *dst++ = mask_u64; | |
1941 | hash = hash_add64(hash, mask_u64); | |
1942 | } | |
1943 | } | |
1944 | ||
1945 | map_t map; | |
1946 | ||
1947 | FLOWMAP_FOR_EACH_MAP (map, mask->mf.map) { | |
1948 | hash = hash_add64(hash, map); | |
1949 | } | |
1950 | ||
1951 | size_t n = dst - miniflow_get_values(&mask->mf); | |
1952 | ||
1953 | mask->hash = hash_finish(hash, n * 8); | |
1954 | mask->len = netdev_flow_key_size(n); | |
1955 | } | |
1956 | ||
1957 | /* Initializes 'dst' as a copy of 'flow' masked with 'mask'. */ | |
1958 | static inline void | |
1959 | netdev_flow_key_init_masked(struct netdev_flow_key *dst, | |
1960 | const struct flow *flow, | |
1961 | const struct netdev_flow_key *mask) | |
1962 | { | |
1963 | uint64_t *dst_u64 = miniflow_values(&dst->mf); | |
1964 | const uint64_t *mask_u64 = miniflow_get_values(&mask->mf); | |
1965 | uint32_t hash = 0; | |
1966 | uint64_t value; | |
1967 | ||
1968 | dst->len = mask->len; | |
1969 | dst->mf = mask->mf; /* Copy maps. */ | |
1970 | ||
1971 | FLOW_FOR_EACH_IN_MAPS(value, flow, mask->mf.map) { | |
1972 | *dst_u64 = value & *mask_u64++; | |
1973 | hash = hash_add64(hash, *dst_u64++); | |
1974 | } | |
1975 | dst->hash = hash_finish(hash, | |
1976 | (dst_u64 - miniflow_get_values(&dst->mf)) * 8); | |
1977 | } | |
1978 | ||
1979 | /* Iterate through netdev_flow_key TNL u64 values specified by 'FLOWMAP'. */ | |
1980 | #define NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(VALUE, KEY, FLOWMAP) \ | |
1981 | MINIFLOW_FOR_EACH_IN_FLOWMAP(VALUE, &(KEY)->mf, FLOWMAP) | |
1982 | ||
1983 | /* Returns a hash value for the bits of 'key' where there are 1-bits in | |
1984 | * 'mask'. */ | |
1985 | static inline uint32_t | |
1986 | netdev_flow_key_hash_in_mask(const struct netdev_flow_key *key, | |
1987 | const struct netdev_flow_key *mask) | |
1988 | { | |
1989 | const uint64_t *p = miniflow_get_values(&mask->mf); | |
1990 | uint32_t hash = 0; | |
1991 | uint64_t value; | |
1992 | ||
1993 | NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value, key, mask->mf.map) { | |
1994 | hash = hash_add64(hash, value & *p++); | |
1995 | } | |
1996 | ||
1997 | return hash_finish(hash, (p - miniflow_get_values(&mask->mf)) * 8); | |
1998 | } | |
1999 | ||
2000 | static inline bool | |
2001 | emc_entry_alive(struct emc_entry *ce) | |
2002 | { | |
2003 | return ce->flow && !ce->flow->dead; | |
2004 | } | |
2005 | ||
2006 | static void | |
2007 | emc_clear_entry(struct emc_entry *ce) | |
2008 | { | |
2009 | if (ce->flow) { | |
2010 | dp_netdev_flow_unref(ce->flow); | |
2011 | ce->flow = NULL; | |
2012 | } | |
2013 | } | |
2014 | ||
2015 | static inline void | |
2016 | emc_change_entry(struct emc_entry *ce, struct dp_netdev_flow *flow, | |
2017 | const struct netdev_flow_key *key) | |
2018 | { | |
2019 | if (ce->flow != flow) { | |
2020 | if (ce->flow) { | |
2021 | dp_netdev_flow_unref(ce->flow); | |
2022 | } | |
2023 | ||
2024 | if (dp_netdev_flow_ref(flow)) { | |
2025 | ce->flow = flow; | |
2026 | } else { | |
2027 | ce->flow = NULL; | |
2028 | } | |
2029 | } | |
2030 | if (key) { | |
2031 | netdev_flow_key_clone(&ce->key, key); | |
2032 | } | |
2033 | } | |
2034 | ||
2035 | static inline void | |
2036 | emc_insert(struct emc_cache *cache, const struct netdev_flow_key *key, | |
2037 | struct dp_netdev_flow *flow) | |
2038 | { | |
2039 | struct emc_entry *to_be_replaced = NULL; | |
2040 | struct emc_entry *current_entry; | |
2041 | ||
2042 | EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) { | |
2043 | if (netdev_flow_key_equal(¤t_entry->key, key)) { | |
2044 | /* We found the entry with the 'mf' miniflow */ | |
2045 | emc_change_entry(current_entry, flow, NULL); | |
2046 | return; | |
2047 | } | |
2048 | ||
2049 | /* Replacement policy: put the flow in an empty (not alive) entry, or | |
2050 | * in the first entry where it can be */ | |
2051 | if (!to_be_replaced | |
2052 | || (emc_entry_alive(to_be_replaced) | |
2053 | && !emc_entry_alive(current_entry)) | |
2054 | || current_entry->key.hash < to_be_replaced->key.hash) { | |
2055 | to_be_replaced = current_entry; | |
2056 | } | |
2057 | } | |
2058 | /* We didn't find the miniflow in the cache. | |
2059 | * The 'to_be_replaced' entry is where the new flow will be stored */ | |
2060 | ||
2061 | emc_change_entry(to_be_replaced, flow, key); | |
2062 | } | |
2063 | ||
2064 | static inline void | |
2065 | emc_probabilistic_insert(struct dp_netdev_pmd_thread *pmd, | |
2066 | const struct netdev_flow_key *key, | |
2067 | struct dp_netdev_flow *flow) | |
2068 | { | |
2069 | /* Insert an entry into the EMC based on probability value 'min'. By | |
2070 | * default the value is UINT32_MAX / 100 which yields an insertion | |
2071 | * probability of 1/100 ie. 1% */ | |
2072 | ||
2073 | uint32_t min; | |
2074 | atomic_read_relaxed(&pmd->dp->emc_insert_min, &min); | |
2075 | ||
2076 | #ifdef DPDK_NETDEV | |
2077 | if (min && (key->hash ^ (uint32_t) pmd->last_cycles) <= min) { | |
2078 | #else | |
2079 | if (min && (key->hash ^ random_uint32()) <= min) { | |
2080 | #endif | |
2081 | emc_insert(&pmd->flow_cache, key, flow); | |
2082 | } | |
2083 | } | |
2084 | ||
2085 | static inline struct dp_netdev_flow * | |
2086 | emc_lookup(struct emc_cache *cache, const struct netdev_flow_key *key) | |
2087 | { | |
2088 | struct emc_entry *current_entry; | |
2089 | ||
2090 | EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) { | |
2091 | if (current_entry->key.hash == key->hash | |
2092 | && emc_entry_alive(current_entry) | |
2093 | && netdev_flow_key_equal_mf(¤t_entry->key, &key->mf)) { | |
2094 | ||
2095 | /* We found the entry with the 'key->mf' miniflow */ | |
2096 | return current_entry->flow; | |
2097 | } | |
2098 | } | |
2099 | ||
2100 | return NULL; | |
2101 | } | |
2102 | ||
2103 | static struct dp_netdev_flow * | |
2104 | dp_netdev_pmd_lookup_flow(struct dp_netdev_pmd_thread *pmd, | |
2105 | const struct netdev_flow_key *key, | |
2106 | int *lookup_num_p) | |
2107 | { | |
2108 | struct dpcls *cls; | |
2109 | struct dpcls_rule *rule; | |
2110 | odp_port_t in_port = u32_to_odp(MINIFLOW_GET_U32(&key->mf, in_port)); | |
2111 | struct dp_netdev_flow *netdev_flow = NULL; | |
2112 | ||
2113 | cls = dp_netdev_pmd_lookup_dpcls(pmd, in_port); | |
2114 | if (OVS_LIKELY(cls)) { | |
2115 | dpcls_lookup(cls, key, &rule, 1, lookup_num_p); | |
2116 | netdev_flow = dp_netdev_flow_cast(rule); | |
2117 | } | |
2118 | return netdev_flow; | |
2119 | } | |
2120 | ||
2121 | static struct dp_netdev_flow * | |
2122 | dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread *pmd, | |
2123 | const ovs_u128 *ufidp, const struct nlattr *key, | |
2124 | size_t key_len) | |
2125 | { | |
2126 | struct dp_netdev_flow *netdev_flow; | |
2127 | struct flow flow; | |
2128 | ovs_u128 ufid; | |
2129 | ||
2130 | /* If a UFID is not provided, determine one based on the key. */ | |
2131 | if (!ufidp && key && key_len | |
2132 | && !dpif_netdev_flow_from_nlattrs(key, key_len, &flow, false)) { | |
2133 | dpif_flow_hash(pmd->dp->dpif, &flow, sizeof flow, &ufid); | |
2134 | ufidp = &ufid; | |
2135 | } | |
2136 | ||
2137 | if (ufidp) { | |
2138 | CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, dp_netdev_flow_hash(ufidp), | |
2139 | &pmd->flow_table) { | |
2140 | if (ovs_u128_equals(netdev_flow->ufid, *ufidp)) { | |
2141 | return netdev_flow; | |
2142 | } | |
2143 | } | |
2144 | } | |
2145 | ||
2146 | return NULL; | |
2147 | } | |
2148 | ||
2149 | static void | |
2150 | get_dpif_flow_stats(const struct dp_netdev_flow *netdev_flow_, | |
2151 | struct dpif_flow_stats *stats) | |
2152 | { | |
2153 | struct dp_netdev_flow *netdev_flow; | |
2154 | unsigned long long n; | |
2155 | long long used; | |
2156 | uint16_t flags; | |
2157 | ||
2158 | netdev_flow = CONST_CAST(struct dp_netdev_flow *, netdev_flow_); | |
2159 | ||
2160 | atomic_read_relaxed(&netdev_flow->stats.packet_count, &n); | |
2161 | stats->n_packets = n; | |
2162 | atomic_read_relaxed(&netdev_flow->stats.byte_count, &n); | |
2163 | stats->n_bytes = n; | |
2164 | atomic_read_relaxed(&netdev_flow->stats.used, &used); | |
2165 | stats->used = used; | |
2166 | atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags); | |
2167 | stats->tcp_flags = flags; | |
2168 | } | |
2169 | ||
2170 | /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for | |
2171 | * storing the netlink-formatted key/mask. 'key_buf' may be the same as | |
2172 | * 'mask_buf'. Actions will be returned without copying, by relying on RCU to | |
2173 | * protect them. */ | |
2174 | static void | |
2175 | dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow *netdev_flow, | |
2176 | struct ofpbuf *key_buf, struct ofpbuf *mask_buf, | |
2177 | struct dpif_flow *flow, bool terse) | |
2178 | { | |
2179 | if (terse) { | |
2180 | memset(flow, 0, sizeof *flow); | |
2181 | } else { | |
2182 | struct flow_wildcards wc; | |
2183 | struct dp_netdev_actions *actions; | |
2184 | size_t offset; | |
2185 | struct odp_flow_key_parms odp_parms = { | |
2186 | .flow = &netdev_flow->flow, | |
2187 | .mask = &wc.masks, | |
2188 | .support = dp_netdev_support, | |
2189 | }; | |
2190 | ||
2191 | miniflow_expand(&netdev_flow->cr.mask->mf, &wc.masks); | |
2192 | /* in_port is exact matched, but we have left it out from the mask for | |
2193 | * optimnization reasons. Add in_port back to the mask. */ | |
2194 | wc.masks.in_port.odp_port = ODPP_NONE; | |
2195 | ||
2196 | /* Key */ | |
2197 | offset = key_buf->size; | |
2198 | flow->key = ofpbuf_tail(key_buf); | |
2199 | odp_flow_key_from_flow(&odp_parms, key_buf); | |
2200 | flow->key_len = key_buf->size - offset; | |
2201 | ||
2202 | /* Mask */ | |
2203 | offset = mask_buf->size; | |
2204 | flow->mask = ofpbuf_tail(mask_buf); | |
2205 | odp_parms.key_buf = key_buf; | |
2206 | odp_flow_key_from_mask(&odp_parms, mask_buf); | |
2207 | flow->mask_len = mask_buf->size - offset; | |
2208 | ||
2209 | /* Actions */ | |
2210 | actions = dp_netdev_flow_get_actions(netdev_flow); | |
2211 | flow->actions = actions->actions; | |
2212 | flow->actions_len = actions->size; | |
2213 | } | |
2214 | ||
2215 | flow->ufid = netdev_flow->ufid; | |
2216 | flow->ufid_present = true; | |
2217 | flow->pmd_id = netdev_flow->pmd_id; | |
2218 | get_dpif_flow_stats(netdev_flow, &flow->stats); | |
2219 | } | |
2220 | ||
2221 | static int | |
2222 | dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len, | |
2223 | const struct nlattr *mask_key, | |
2224 | uint32_t mask_key_len, const struct flow *flow, | |
2225 | struct flow_wildcards *wc, bool probe) | |
2226 | { | |
2227 | enum odp_key_fitness fitness; | |
2228 | ||
2229 | fitness = odp_flow_key_to_mask(mask_key, mask_key_len, wc, flow); | |
2230 | if (fitness) { | |
2231 | if (!probe) { | |
2232 | /* This should not happen: it indicates that | |
2233 | * odp_flow_key_from_mask() and odp_flow_key_to_mask() | |
2234 | * disagree on the acceptable form of a mask. Log the problem | |
2235 | * as an error, with enough details to enable debugging. */ | |
2236 | static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5); | |
2237 | ||
2238 | if (!VLOG_DROP_ERR(&rl)) { | |
2239 | struct ds s; | |
2240 | ||
2241 | ds_init(&s); | |
2242 | odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s, | |
2243 | true); | |
2244 | VLOG_ERR("internal error parsing flow mask %s (%s)", | |
2245 | ds_cstr(&s), odp_key_fitness_to_string(fitness)); | |
2246 | ds_destroy(&s); | |
2247 | } | |
2248 | } | |
2249 | ||
2250 | return EINVAL; | |
2251 | } | |
2252 | ||
2253 | return 0; | |
2254 | } | |
2255 | ||
2256 | static int | |
2257 | dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len, | |
2258 | struct flow *flow, bool probe) | |
2259 | { | |
2260 | odp_port_t in_port; | |
2261 | ||
2262 | if (odp_flow_key_to_flow(key, key_len, flow)) { | |
2263 | if (!probe) { | |
2264 | /* This should not happen: it indicates that | |
2265 | * odp_flow_key_from_flow() and odp_flow_key_to_flow() disagree on | |
2266 | * the acceptable form of a flow. Log the problem as an error, | |
2267 | * with enough details to enable debugging. */ | |
2268 | static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5); | |
2269 | ||
2270 | if (!VLOG_DROP_ERR(&rl)) { | |
2271 | struct ds s; | |
2272 | ||
2273 | ds_init(&s); | |
2274 | odp_flow_format(key, key_len, NULL, 0, NULL, &s, true); | |
2275 | VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s)); | |
2276 | ds_destroy(&s); | |
2277 | } | |
2278 | } | |
2279 | ||
2280 | return EINVAL; | |
2281 | } | |
2282 | ||
2283 | in_port = flow->in_port.odp_port; | |
2284 | if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) { | |
2285 | return EINVAL; | |
2286 | } | |
2287 | ||
2288 | if (flow->ct_state & DP_NETDEV_CS_UNSUPPORTED_MASK) { | |
2289 | return EINVAL; | |
2290 | } | |
2291 | ||
2292 | return 0; | |
2293 | } | |
2294 | ||
2295 | static int | |
2296 | dpif_netdev_flow_get(const struct dpif *dpif, const struct dpif_flow_get *get) | |
2297 | { | |
2298 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
2299 | struct dp_netdev_flow *netdev_flow; | |
2300 | struct dp_netdev_pmd_thread *pmd; | |
2301 | struct hmapx to_find = HMAPX_INITIALIZER(&to_find); | |
2302 | struct hmapx_node *node; | |
2303 | int error = EINVAL; | |
2304 | ||
2305 | if (get->pmd_id == PMD_ID_NULL) { | |
2306 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
2307 | if (dp_netdev_pmd_try_ref(pmd) && !hmapx_add(&to_find, pmd)) { | |
2308 | dp_netdev_pmd_unref(pmd); | |
2309 | } | |
2310 | } | |
2311 | } else { | |
2312 | pmd = dp_netdev_get_pmd(dp, get->pmd_id); | |
2313 | if (!pmd) { | |
2314 | goto out; | |
2315 | } | |
2316 | hmapx_add(&to_find, pmd); | |
2317 | } | |
2318 | ||
2319 | if (!hmapx_count(&to_find)) { | |
2320 | goto out; | |
2321 | } | |
2322 | ||
2323 | HMAPX_FOR_EACH (node, &to_find) { | |
2324 | pmd = (struct dp_netdev_pmd_thread *) node->data; | |
2325 | netdev_flow = dp_netdev_pmd_find_flow(pmd, get->ufid, get->key, | |
2326 | get->key_len); | |
2327 | if (netdev_flow) { | |
2328 | dp_netdev_flow_to_dpif_flow(netdev_flow, get->buffer, get->buffer, | |
2329 | get->flow, false); | |
2330 | error = 0; | |
2331 | break; | |
2332 | } else { | |
2333 | error = ENOENT; | |
2334 | } | |
2335 | } | |
2336 | ||
2337 | HMAPX_FOR_EACH (node, &to_find) { | |
2338 | pmd = (struct dp_netdev_pmd_thread *) node->data; | |
2339 | dp_netdev_pmd_unref(pmd); | |
2340 | } | |
2341 | out: | |
2342 | hmapx_destroy(&to_find); | |
2343 | return error; | |
2344 | } | |
2345 | ||
2346 | static struct dp_netdev_flow * | |
2347 | dp_netdev_flow_add(struct dp_netdev_pmd_thread *pmd, | |
2348 | struct match *match, const ovs_u128 *ufid, | |
2349 | const struct nlattr *actions, size_t actions_len) | |
2350 | OVS_REQUIRES(pmd->flow_mutex) | |
2351 | { | |
2352 | struct dp_netdev_flow *flow; | |
2353 | struct netdev_flow_key mask; | |
2354 | struct dpcls *cls; | |
2355 | ||
2356 | /* Make sure in_port is exact matched before we read it. */ | |
2357 | ovs_assert(match->wc.masks.in_port.odp_port == ODPP_NONE); | |
2358 | odp_port_t in_port = match->flow.in_port.odp_port; | |
2359 | ||
2360 | /* As we select the dpcls based on the port number, each netdev flow | |
2361 | * belonging to the same dpcls will have the same odp_port value. | |
2362 | * For performance reasons we wildcard odp_port here in the mask. In the | |
2363 | * typical case dp_hash is also wildcarded, and the resulting 8-byte | |
2364 | * chunk {dp_hash, in_port} will be ignored by netdev_flow_mask_init() and | |
2365 | * will not be part of the subtable mask. | |
2366 | * This will speed up the hash computation during dpcls_lookup() because | |
2367 | * there is one less call to hash_add64() in this case. */ | |
2368 | match->wc.masks.in_port.odp_port = 0; | |
2369 | netdev_flow_mask_init(&mask, match); | |
2370 | match->wc.masks.in_port.odp_port = ODPP_NONE; | |
2371 | ||
2372 | /* Make sure wc does not have metadata. */ | |
2373 | ovs_assert(!FLOWMAP_HAS_FIELD(&mask.mf.map, metadata) | |
2374 | && !FLOWMAP_HAS_FIELD(&mask.mf.map, regs)); | |
2375 | ||
2376 | /* Do not allocate extra space. */ | |
2377 | flow = xmalloc(sizeof *flow - sizeof flow->cr.flow.mf + mask.len); | |
2378 | memset(&flow->stats, 0, sizeof flow->stats); | |
2379 | flow->dead = false; | |
2380 | flow->batch = NULL; | |
2381 | *CONST_CAST(unsigned *, &flow->pmd_id) = pmd->core_id; | |
2382 | *CONST_CAST(struct flow *, &flow->flow) = match->flow; | |
2383 | *CONST_CAST(ovs_u128 *, &flow->ufid) = *ufid; | |
2384 | ovs_refcount_init(&flow->ref_cnt); | |
2385 | ovsrcu_set(&flow->actions, dp_netdev_actions_create(actions, actions_len)); | |
2386 | ||
2387 | netdev_flow_key_init_masked(&flow->cr.flow, &match->flow, &mask); | |
2388 | ||
2389 | /* Select dpcls for in_port. Relies on in_port to be exact match. */ | |
2390 | cls = dp_netdev_pmd_find_dpcls(pmd, in_port); | |
2391 | dpcls_insert(cls, &flow->cr, &mask); | |
2392 | ||
2393 | cmap_insert(&pmd->flow_table, CONST_CAST(struct cmap_node *, &flow->node), | |
2394 | dp_netdev_flow_hash(&flow->ufid)); | |
2395 | ||
2396 | if (OVS_UNLIKELY(!VLOG_DROP_DBG((&upcall_rl)))) { | |
2397 | struct ds ds = DS_EMPTY_INITIALIZER; | |
2398 | struct ofpbuf key_buf, mask_buf; | |
2399 | struct odp_flow_key_parms odp_parms = { | |
2400 | .flow = &match->flow, | |
2401 | .mask = &match->wc.masks, | |
2402 | .support = dp_netdev_support, | |
2403 | }; | |
2404 | ||
2405 | ofpbuf_init(&key_buf, 0); | |
2406 | ofpbuf_init(&mask_buf, 0); | |
2407 | ||
2408 | odp_flow_key_from_flow(&odp_parms, &key_buf); | |
2409 | odp_parms.key_buf = &key_buf; | |
2410 | odp_flow_key_from_mask(&odp_parms, &mask_buf); | |
2411 | ||
2412 | ds_put_cstr(&ds, "flow_add: "); | |
2413 | odp_format_ufid(ufid, &ds); | |
2414 | ds_put_cstr(&ds, " "); | |
2415 | odp_flow_format(key_buf.data, key_buf.size, | |
2416 | mask_buf.data, mask_buf.size, | |
2417 | NULL, &ds, false); | |
2418 | ds_put_cstr(&ds, ", actions:"); | |
2419 | format_odp_actions(&ds, actions, actions_len); | |
2420 | ||
2421 | VLOG_DBG("%s", ds_cstr(&ds)); | |
2422 | ||
2423 | ofpbuf_uninit(&key_buf); | |
2424 | ofpbuf_uninit(&mask_buf); | |
2425 | ||
2426 | /* Add a printout of the actual match installed. */ | |
2427 | struct match m; | |
2428 | ds_clear(&ds); | |
2429 | ds_put_cstr(&ds, "flow match: "); | |
2430 | miniflow_expand(&flow->cr.flow.mf, &m.flow); | |
2431 | miniflow_expand(&flow->cr.mask->mf, &m.wc.masks); | |
2432 | match_format(&m, NULL, &ds, OFP_DEFAULT_PRIORITY); | |
2433 | ||
2434 | VLOG_DBG("%s", ds_cstr(&ds)); | |
2435 | ||
2436 | ds_destroy(&ds); | |
2437 | } | |
2438 | ||
2439 | return flow; | |
2440 | } | |
2441 | ||
2442 | static int | |
2443 | flow_put_on_pmd(struct dp_netdev_pmd_thread *pmd, | |
2444 | struct netdev_flow_key *key, | |
2445 | struct match *match, | |
2446 | ovs_u128 *ufid, | |
2447 | const struct dpif_flow_put *put, | |
2448 | struct dpif_flow_stats *stats) | |
2449 | { | |
2450 | struct dp_netdev_flow *netdev_flow; | |
2451 | int error = 0; | |
2452 | ||
2453 | if (stats) { | |
2454 | memset(stats, 0, sizeof *stats); | |
2455 | } | |
2456 | ||
2457 | ovs_mutex_lock(&pmd->flow_mutex); | |
2458 | netdev_flow = dp_netdev_pmd_lookup_flow(pmd, key, NULL); | |
2459 | if (!netdev_flow) { | |
2460 | if (put->flags & DPIF_FP_CREATE) { | |
2461 | if (cmap_count(&pmd->flow_table) < MAX_FLOWS) { | |
2462 | dp_netdev_flow_add(pmd, match, ufid, put->actions, | |
2463 | put->actions_len); | |
2464 | error = 0; | |
2465 | } else { | |
2466 | error = EFBIG; | |
2467 | } | |
2468 | } else { | |
2469 | error = ENOENT; | |
2470 | } | |
2471 | } else { | |
2472 | if (put->flags & DPIF_FP_MODIFY) { | |
2473 | struct dp_netdev_actions *new_actions; | |
2474 | struct dp_netdev_actions *old_actions; | |
2475 | ||
2476 | new_actions = dp_netdev_actions_create(put->actions, | |
2477 | put->actions_len); | |
2478 | ||
2479 | old_actions = dp_netdev_flow_get_actions(netdev_flow); | |
2480 | ovsrcu_set(&netdev_flow->actions, new_actions); | |
2481 | ||
2482 | if (stats) { | |
2483 | get_dpif_flow_stats(netdev_flow, stats); | |
2484 | } | |
2485 | if (put->flags & DPIF_FP_ZERO_STATS) { | |
2486 | /* XXX: The userspace datapath uses thread local statistics | |
2487 | * (for flows), which should be updated only by the owning | |
2488 | * thread. Since we cannot write on stats memory here, | |
2489 | * we choose not to support this flag. Please note: | |
2490 | * - This feature is currently used only by dpctl commands with | |
2491 | * option --clear. | |
2492 | * - Should the need arise, this operation can be implemented | |
2493 | * by keeping a base value (to be update here) for each | |
2494 | * counter, and subtracting it before outputting the stats */ | |
2495 | error = EOPNOTSUPP; | |
2496 | } | |
2497 | ||
2498 | ovsrcu_postpone(dp_netdev_actions_free, old_actions); | |
2499 | } else if (put->flags & DPIF_FP_CREATE) { | |
2500 | error = EEXIST; | |
2501 | } else { | |
2502 | /* Overlapping flow. */ | |
2503 | error = EINVAL; | |
2504 | } | |
2505 | } | |
2506 | ovs_mutex_unlock(&pmd->flow_mutex); | |
2507 | return error; | |
2508 | } | |
2509 | ||
2510 | static int | |
2511 | dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put) | |
2512 | { | |
2513 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
2514 | struct netdev_flow_key key, mask; | |
2515 | struct dp_netdev_pmd_thread *pmd; | |
2516 | struct match match; | |
2517 | ovs_u128 ufid; | |
2518 | int error; | |
2519 | bool probe = put->flags & DPIF_FP_PROBE; | |
2520 | ||
2521 | if (put->stats) { | |
2522 | memset(put->stats, 0, sizeof *put->stats); | |
2523 | } | |
2524 | error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow, | |
2525 | probe); | |
2526 | if (error) { | |
2527 | return error; | |
2528 | } | |
2529 | error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len, | |
2530 | put->mask, put->mask_len, | |
2531 | &match.flow, &match.wc, probe); | |
2532 | if (error) { | |
2533 | return error; | |
2534 | } | |
2535 | ||
2536 | if (put->ufid) { | |
2537 | ufid = *put->ufid; | |
2538 | } else { | |
2539 | dpif_flow_hash(dpif, &match.flow, sizeof match.flow, &ufid); | |
2540 | } | |
2541 | ||
2542 | /* Must produce a netdev_flow_key for lookup. | |
2543 | * Use the same method as employed to create the key when adding | |
2544 | * the flow to the dplcs to make sure they match. */ | |
2545 | netdev_flow_mask_init(&mask, &match); | |
2546 | netdev_flow_key_init_masked(&key, &match.flow, &mask); | |
2547 | ||
2548 | if (put->pmd_id == PMD_ID_NULL) { | |
2549 | if (cmap_count(&dp->poll_threads) == 0) { | |
2550 | return EINVAL; | |
2551 | } | |
2552 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
2553 | struct dpif_flow_stats pmd_stats; | |
2554 | int pmd_error; | |
2555 | ||
2556 | pmd_error = flow_put_on_pmd(pmd, &key, &match, &ufid, put, | |
2557 | &pmd_stats); | |
2558 | if (pmd_error) { | |
2559 | error = pmd_error; | |
2560 | } else if (put->stats) { | |
2561 | put->stats->n_packets += pmd_stats.n_packets; | |
2562 | put->stats->n_bytes += pmd_stats.n_bytes; | |
2563 | put->stats->used = MAX(put->stats->used, pmd_stats.used); | |
2564 | put->stats->tcp_flags |= pmd_stats.tcp_flags; | |
2565 | } | |
2566 | } | |
2567 | } else { | |
2568 | pmd = dp_netdev_get_pmd(dp, put->pmd_id); | |
2569 | if (!pmd) { | |
2570 | return EINVAL; | |
2571 | } | |
2572 | error = flow_put_on_pmd(pmd, &key, &match, &ufid, put, put->stats); | |
2573 | dp_netdev_pmd_unref(pmd); | |
2574 | } | |
2575 | ||
2576 | return error; | |
2577 | } | |
2578 | ||
2579 | static int | |
2580 | flow_del_on_pmd(struct dp_netdev_pmd_thread *pmd, | |
2581 | struct dpif_flow_stats *stats, | |
2582 | const struct dpif_flow_del *del) | |
2583 | { | |
2584 | struct dp_netdev_flow *netdev_flow; | |
2585 | int error = 0; | |
2586 | ||
2587 | ovs_mutex_lock(&pmd->flow_mutex); | |
2588 | netdev_flow = dp_netdev_pmd_find_flow(pmd, del->ufid, del->key, | |
2589 | del->key_len); | |
2590 | if (netdev_flow) { | |
2591 | if (stats) { | |
2592 | get_dpif_flow_stats(netdev_flow, stats); | |
2593 | } | |
2594 | dp_netdev_pmd_remove_flow(pmd, netdev_flow); | |
2595 | } else { | |
2596 | error = ENOENT; | |
2597 | } | |
2598 | ovs_mutex_unlock(&pmd->flow_mutex); | |
2599 | ||
2600 | return error; | |
2601 | } | |
2602 | ||
2603 | static int | |
2604 | dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del) | |
2605 | { | |
2606 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
2607 | struct dp_netdev_pmd_thread *pmd; | |
2608 | int error = 0; | |
2609 | ||
2610 | if (del->stats) { | |
2611 | memset(del->stats, 0, sizeof *del->stats); | |
2612 | } | |
2613 | ||
2614 | if (del->pmd_id == PMD_ID_NULL) { | |
2615 | if (cmap_count(&dp->poll_threads) == 0) { | |
2616 | return EINVAL; | |
2617 | } | |
2618 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
2619 | struct dpif_flow_stats pmd_stats; | |
2620 | int pmd_error; | |
2621 | ||
2622 | pmd_error = flow_del_on_pmd(pmd, &pmd_stats, del); | |
2623 | if (pmd_error) { | |
2624 | error = pmd_error; | |
2625 | } else if (del->stats) { | |
2626 | del->stats->n_packets += pmd_stats.n_packets; | |
2627 | del->stats->n_bytes += pmd_stats.n_bytes; | |
2628 | del->stats->used = MAX(del->stats->used, pmd_stats.used); | |
2629 | del->stats->tcp_flags |= pmd_stats.tcp_flags; | |
2630 | } | |
2631 | } | |
2632 | } else { | |
2633 | pmd = dp_netdev_get_pmd(dp, del->pmd_id); | |
2634 | if (!pmd) { | |
2635 | return EINVAL; | |
2636 | } | |
2637 | error = flow_del_on_pmd(pmd, del->stats, del); | |
2638 | dp_netdev_pmd_unref(pmd); | |
2639 | } | |
2640 | ||
2641 | ||
2642 | return error; | |
2643 | } | |
2644 | ||
2645 | struct dpif_netdev_flow_dump { | |
2646 | struct dpif_flow_dump up; | |
2647 | struct cmap_position poll_thread_pos; | |
2648 | struct cmap_position flow_pos; | |
2649 | struct dp_netdev_pmd_thread *cur_pmd; | |
2650 | int status; | |
2651 | struct ovs_mutex mutex; | |
2652 | }; | |
2653 | ||
2654 | static struct dpif_netdev_flow_dump * | |
2655 | dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump) | |
2656 | { | |
2657 | return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up); | |
2658 | } | |
2659 | ||
2660 | static struct dpif_flow_dump * | |
2661 | dpif_netdev_flow_dump_create(const struct dpif *dpif_, bool terse) | |
2662 | { | |
2663 | struct dpif_netdev_flow_dump *dump; | |
2664 | ||
2665 | dump = xzalloc(sizeof *dump); | |
2666 | dpif_flow_dump_init(&dump->up, dpif_); | |
2667 | dump->up.terse = terse; | |
2668 | ovs_mutex_init(&dump->mutex); | |
2669 | ||
2670 | return &dump->up; | |
2671 | } | |
2672 | ||
2673 | static int | |
2674 | dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_) | |
2675 | { | |
2676 | struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_); | |
2677 | ||
2678 | ovs_mutex_destroy(&dump->mutex); | |
2679 | free(dump); | |
2680 | return 0; | |
2681 | } | |
2682 | ||
2683 | struct dpif_netdev_flow_dump_thread { | |
2684 | struct dpif_flow_dump_thread up; | |
2685 | struct dpif_netdev_flow_dump *dump; | |
2686 | struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH]; | |
2687 | struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH]; | |
2688 | }; | |
2689 | ||
2690 | static struct dpif_netdev_flow_dump_thread * | |
2691 | dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread) | |
2692 | { | |
2693 | return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up); | |
2694 | } | |
2695 | ||
2696 | static struct dpif_flow_dump_thread * | |
2697 | dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_) | |
2698 | { | |
2699 | struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_); | |
2700 | struct dpif_netdev_flow_dump_thread *thread; | |
2701 | ||
2702 | thread = xmalloc(sizeof *thread); | |
2703 | dpif_flow_dump_thread_init(&thread->up, &dump->up); | |
2704 | thread->dump = dump; | |
2705 | return &thread->up; | |
2706 | } | |
2707 | ||
2708 | static void | |
2709 | dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_) | |
2710 | { | |
2711 | struct dpif_netdev_flow_dump_thread *thread | |
2712 | = dpif_netdev_flow_dump_thread_cast(thread_); | |
2713 | ||
2714 | free(thread); | |
2715 | } | |
2716 | ||
2717 | static int | |
2718 | dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_, | |
2719 | struct dpif_flow *flows, int max_flows) | |
2720 | { | |
2721 | struct dpif_netdev_flow_dump_thread *thread | |
2722 | = dpif_netdev_flow_dump_thread_cast(thread_); | |
2723 | struct dpif_netdev_flow_dump *dump = thread->dump; | |
2724 | struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH]; | |
2725 | int n_flows = 0; | |
2726 | int i; | |
2727 | ||
2728 | ovs_mutex_lock(&dump->mutex); | |
2729 | if (!dump->status) { | |
2730 | struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif); | |
2731 | struct dp_netdev *dp = get_dp_netdev(&dpif->dpif); | |
2732 | struct dp_netdev_pmd_thread *pmd = dump->cur_pmd; | |
2733 | int flow_limit = MIN(max_flows, FLOW_DUMP_MAX_BATCH); | |
2734 | ||
2735 | /* First call to dump_next(), extracts the first pmd thread. | |
2736 | * If there is no pmd thread, returns immediately. */ | |
2737 | if (!pmd) { | |
2738 | pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos); | |
2739 | if (!pmd) { | |
2740 | ovs_mutex_unlock(&dump->mutex); | |
2741 | return n_flows; | |
2742 | ||
2743 | } | |
2744 | } | |
2745 | ||
2746 | do { | |
2747 | for (n_flows = 0; n_flows < flow_limit; n_flows++) { | |
2748 | struct cmap_node *node; | |
2749 | ||
2750 | node = cmap_next_position(&pmd->flow_table, &dump->flow_pos); | |
2751 | if (!node) { | |
2752 | break; | |
2753 | } | |
2754 | netdev_flows[n_flows] = CONTAINER_OF(node, | |
2755 | struct dp_netdev_flow, | |
2756 | node); | |
2757 | } | |
2758 | /* When finishing dumping the current pmd thread, moves to | |
2759 | * the next. */ | |
2760 | if (n_flows < flow_limit) { | |
2761 | memset(&dump->flow_pos, 0, sizeof dump->flow_pos); | |
2762 | dp_netdev_pmd_unref(pmd); | |
2763 | pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos); | |
2764 | if (!pmd) { | |
2765 | dump->status = EOF; | |
2766 | break; | |
2767 | } | |
2768 | } | |
2769 | /* Keeps the reference to next caller. */ | |
2770 | dump->cur_pmd = pmd; | |
2771 | ||
2772 | /* If the current dump is empty, do not exit the loop, since the | |
2773 | * remaining pmds could have flows to be dumped. Just dumps again | |
2774 | * on the new 'pmd'. */ | |
2775 | } while (!n_flows); | |
2776 | } | |
2777 | ovs_mutex_unlock(&dump->mutex); | |
2778 | ||
2779 | for (i = 0; i < n_flows; i++) { | |
2780 | struct odputil_keybuf *maskbuf = &thread->maskbuf[i]; | |
2781 | struct odputil_keybuf *keybuf = &thread->keybuf[i]; | |
2782 | struct dp_netdev_flow *netdev_flow = netdev_flows[i]; | |
2783 | struct dpif_flow *f = &flows[i]; | |
2784 | struct ofpbuf key, mask; | |
2785 | ||
2786 | ofpbuf_use_stack(&key, keybuf, sizeof *keybuf); | |
2787 | ofpbuf_use_stack(&mask, maskbuf, sizeof *maskbuf); | |
2788 | dp_netdev_flow_to_dpif_flow(netdev_flow, &key, &mask, f, | |
2789 | dump->up.terse); | |
2790 | } | |
2791 | ||
2792 | return n_flows; | |
2793 | } | |
2794 | ||
2795 | static int | |
2796 | dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute) | |
2797 | OVS_NO_THREAD_SAFETY_ANALYSIS | |
2798 | { | |
2799 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
2800 | struct dp_netdev_pmd_thread *pmd; | |
2801 | struct dp_packet_batch pp; | |
2802 | ||
2803 | if (dp_packet_size(execute->packet) < ETH_HEADER_LEN || | |
2804 | dp_packet_size(execute->packet) > UINT16_MAX) { | |
2805 | return EINVAL; | |
2806 | } | |
2807 | ||
2808 | /* Tries finding the 'pmd'. If NULL is returned, that means | |
2809 | * the current thread is a non-pmd thread and should use | |
2810 | * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */ | |
2811 | pmd = ovsthread_getspecific(dp->per_pmd_key); | |
2812 | if (!pmd) { | |
2813 | pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID); | |
2814 | if (!pmd) { | |
2815 | return EBUSY; | |
2816 | } | |
2817 | } | |
2818 | ||
2819 | if (execute->probe) { | |
2820 | /* If this is part of a probe, Drop the packet, since executing | |
2821 | * the action may actually cause spurious packets be sent into | |
2822 | * the network. */ | |
2823 | return 0; | |
2824 | } | |
2825 | ||
2826 | /* If the current thread is non-pmd thread, acquires | |
2827 | * the 'non_pmd_mutex'. */ | |
2828 | if (pmd->core_id == NON_PMD_CORE_ID) { | |
2829 | ovs_mutex_lock(&dp->non_pmd_mutex); | |
2830 | } | |
2831 | ||
2832 | /* The action processing expects the RSS hash to be valid, because | |
2833 | * it's always initialized at the beginning of datapath processing. | |
2834 | * In this case, though, 'execute->packet' may not have gone through | |
2835 | * the datapath at all, it may have been generated by the upper layer | |
2836 | * (OpenFlow packet-out, BFD frame, ...). */ | |
2837 | if (!dp_packet_rss_valid(execute->packet)) { | |
2838 | dp_packet_set_rss_hash(execute->packet, | |
2839 | flow_hash_5tuple(execute->flow, 0)); | |
2840 | } | |
2841 | ||
2842 | dp_packet_batch_init_packet(&pp, execute->packet); | |
2843 | dp_netdev_execute_actions(pmd, &pp, false, execute->flow, | |
2844 | execute->actions, execute->actions_len, | |
2845 | time_msec()); | |
2846 | ||
2847 | if (pmd->core_id == NON_PMD_CORE_ID) { | |
2848 | ovs_mutex_unlock(&dp->non_pmd_mutex); | |
2849 | dp_netdev_pmd_unref(pmd); | |
2850 | } | |
2851 | ||
2852 | return 0; | |
2853 | } | |
2854 | ||
2855 | static void | |
2856 | dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops) | |
2857 | { | |
2858 | size_t i; | |
2859 | ||
2860 | for (i = 0; i < n_ops; i++) { | |
2861 | struct dpif_op *op = ops[i]; | |
2862 | ||
2863 | switch (op->type) { | |
2864 | case DPIF_OP_FLOW_PUT: | |
2865 | op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put); | |
2866 | break; | |
2867 | ||
2868 | case DPIF_OP_FLOW_DEL: | |
2869 | op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del); | |
2870 | break; | |
2871 | ||
2872 | case DPIF_OP_EXECUTE: | |
2873 | op->error = dpif_netdev_execute(dpif, &op->u.execute); | |
2874 | break; | |
2875 | ||
2876 | case DPIF_OP_FLOW_GET: | |
2877 | op->error = dpif_netdev_flow_get(dpif, &op->u.flow_get); | |
2878 | break; | |
2879 | } | |
2880 | } | |
2881 | } | |
2882 | ||
2883 | /* Applies datapath configuration from the database. Some of the changes are | |
2884 | * actually applied in dpif_netdev_run(). */ | |
2885 | static int | |
2886 | dpif_netdev_set_config(struct dpif *dpif, const struct smap *other_config) | |
2887 | { | |
2888 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
2889 | const char *cmask = smap_get(other_config, "pmd-cpu-mask"); | |
2890 | unsigned long long insert_prob = | |
2891 | smap_get_ullong(other_config, "emc-insert-inv-prob", | |
2892 | DEFAULT_EM_FLOW_INSERT_INV_PROB); | |
2893 | uint32_t insert_min, cur_min; | |
2894 | ||
2895 | if (!nullable_string_is_equal(dp->pmd_cmask, cmask)) { | |
2896 | free(dp->pmd_cmask); | |
2897 | dp->pmd_cmask = nullable_xstrdup(cmask); | |
2898 | dp_netdev_request_reconfigure(dp); | |
2899 | } | |
2900 | ||
2901 | atomic_read_relaxed(&dp->emc_insert_min, &cur_min); | |
2902 | if (insert_prob <= UINT32_MAX) { | |
2903 | insert_min = insert_prob == 0 ? 0 : UINT32_MAX / insert_prob; | |
2904 | } else { | |
2905 | insert_min = DEFAULT_EM_FLOW_INSERT_MIN; | |
2906 | insert_prob = DEFAULT_EM_FLOW_INSERT_INV_PROB; | |
2907 | } | |
2908 | ||
2909 | if (insert_min != cur_min) { | |
2910 | atomic_store_relaxed(&dp->emc_insert_min, insert_min); | |
2911 | if (insert_min == 0) { | |
2912 | VLOG_INFO("EMC has been disabled"); | |
2913 | } else { | |
2914 | VLOG_INFO("EMC insertion probability changed to 1/%llu (~%.2f%%)", | |
2915 | insert_prob, (100 / (float)insert_prob)); | |
2916 | } | |
2917 | } | |
2918 | ||
2919 | return 0; | |
2920 | } | |
2921 | ||
2922 | /* Parses affinity list and returns result in 'core_ids'. */ | |
2923 | static int | |
2924 | parse_affinity_list(const char *affinity_list, unsigned *core_ids, int n_rxq) | |
2925 | { | |
2926 | unsigned i; | |
2927 | char *list, *copy, *key, *value; | |
2928 | int error = 0; | |
2929 | ||
2930 | for (i = 0; i < n_rxq; i++) { | |
2931 | core_ids[i] = OVS_CORE_UNSPEC; | |
2932 | } | |
2933 | ||
2934 | if (!affinity_list) { | |
2935 | return 0; | |
2936 | } | |
2937 | ||
2938 | list = copy = xstrdup(affinity_list); | |
2939 | ||
2940 | while (ofputil_parse_key_value(&list, &key, &value)) { | |
2941 | int rxq_id, core_id; | |
2942 | ||
2943 | if (!str_to_int(key, 0, &rxq_id) || rxq_id < 0 | |
2944 | || !str_to_int(value, 0, &core_id) || core_id < 0) { | |
2945 | error = EINVAL; | |
2946 | break; | |
2947 | } | |
2948 | ||
2949 | if (rxq_id < n_rxq) { | |
2950 | core_ids[rxq_id] = core_id; | |
2951 | } | |
2952 | } | |
2953 | ||
2954 | free(copy); | |
2955 | return error; | |
2956 | } | |
2957 | ||
2958 | /* Parses 'affinity_list' and applies configuration if it is valid. */ | |
2959 | static int | |
2960 | dpif_netdev_port_set_rxq_affinity(struct dp_netdev_port *port, | |
2961 | const char *affinity_list) | |
2962 | { | |
2963 | unsigned *core_ids, i; | |
2964 | int error = 0; | |
2965 | ||
2966 | core_ids = xmalloc(port->n_rxq * sizeof *core_ids); | |
2967 | if (parse_affinity_list(affinity_list, core_ids, port->n_rxq)) { | |
2968 | error = EINVAL; | |
2969 | goto exit; | |
2970 | } | |
2971 | ||
2972 | for (i = 0; i < port->n_rxq; i++) { | |
2973 | port->rxqs[i].core_id = core_ids[i]; | |
2974 | } | |
2975 | ||
2976 | exit: | |
2977 | free(core_ids); | |
2978 | return error; | |
2979 | } | |
2980 | ||
2981 | /* Changes the affinity of port's rx queues. The changes are actually applied | |
2982 | * in dpif_netdev_run(). */ | |
2983 | static int | |
2984 | dpif_netdev_port_set_config(struct dpif *dpif, odp_port_t port_no, | |
2985 | const struct smap *cfg) | |
2986 | { | |
2987 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
2988 | struct dp_netdev_port *port; | |
2989 | int error = 0; | |
2990 | const char *affinity_list = smap_get(cfg, "pmd-rxq-affinity"); | |
2991 | ||
2992 | ovs_mutex_lock(&dp->port_mutex); | |
2993 | error = get_port_by_number(dp, port_no, &port); | |
2994 | if (error || !netdev_is_pmd(port->netdev) | |
2995 | || nullable_string_is_equal(affinity_list, port->rxq_affinity_list)) { | |
2996 | goto unlock; | |
2997 | } | |
2998 | ||
2999 | error = dpif_netdev_port_set_rxq_affinity(port, affinity_list); | |
3000 | if (error) { | |
3001 | goto unlock; | |
3002 | } | |
3003 | free(port->rxq_affinity_list); | |
3004 | port->rxq_affinity_list = nullable_xstrdup(affinity_list); | |
3005 | ||
3006 | dp_netdev_request_reconfigure(dp); | |
3007 | unlock: | |
3008 | ovs_mutex_unlock(&dp->port_mutex); | |
3009 | return error; | |
3010 | } | |
3011 | ||
3012 | static int | |
3013 | dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED, | |
3014 | uint32_t queue_id, uint32_t *priority) | |
3015 | { | |
3016 | *priority = queue_id; | |
3017 | return 0; | |
3018 | } | |
3019 | ||
3020 | \f | |
3021 | /* Creates and returns a new 'struct dp_netdev_actions', whose actions are | |
3022 | * a copy of the 'ofpacts_len' bytes of 'ofpacts'. */ | |
3023 | struct dp_netdev_actions * | |
3024 | dp_netdev_actions_create(const struct nlattr *actions, size_t size) | |
3025 | { | |
3026 | struct dp_netdev_actions *netdev_actions; | |
3027 | ||
3028 | netdev_actions = xmalloc(sizeof *netdev_actions + size); | |
3029 | memcpy(netdev_actions->actions, actions, size); | |
3030 | netdev_actions->size = size; | |
3031 | ||
3032 | return netdev_actions; | |
3033 | } | |
3034 | ||
3035 | struct dp_netdev_actions * | |
3036 | dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow) | |
3037 | { | |
3038 | return ovsrcu_get(struct dp_netdev_actions *, &flow->actions); | |
3039 | } | |
3040 | ||
3041 | static void | |
3042 | dp_netdev_actions_free(struct dp_netdev_actions *actions) | |
3043 | { | |
3044 | free(actions); | |
3045 | } | |
3046 | \f | |
3047 | static inline unsigned long long | |
3048 | cycles_counter(void) | |
3049 | { | |
3050 | #ifdef DPDK_NETDEV | |
3051 | return rte_get_tsc_cycles(); | |
3052 | #else | |
3053 | return 0; | |
3054 | #endif | |
3055 | } | |
3056 | ||
3057 | /* Fake mutex to make sure that the calls to cycles_count_* are balanced */ | |
3058 | extern struct ovs_mutex cycles_counter_fake_mutex; | |
3059 | ||
3060 | /* Start counting cycles. Must be followed by 'cycles_count_end()' */ | |
3061 | static inline void | |
3062 | cycles_count_start(struct dp_netdev_pmd_thread *pmd) | |
3063 | OVS_ACQUIRES(&cycles_counter_fake_mutex) | |
3064 | OVS_NO_THREAD_SAFETY_ANALYSIS | |
3065 | { | |
3066 | pmd->last_cycles = cycles_counter(); | |
3067 | } | |
3068 | ||
3069 | /* Stop counting cycles and add them to the counter 'type' */ | |
3070 | static inline void | |
3071 | cycles_count_end(struct dp_netdev_pmd_thread *pmd, | |
3072 | enum pmd_cycles_counter_type type) | |
3073 | OVS_RELEASES(&cycles_counter_fake_mutex) | |
3074 | OVS_NO_THREAD_SAFETY_ANALYSIS | |
3075 | { | |
3076 | unsigned long long interval = cycles_counter() - pmd->last_cycles; | |
3077 | ||
3078 | non_atomic_ullong_add(&pmd->cycles.n[type], interval); | |
3079 | } | |
3080 | ||
3081 | static void | |
3082 | dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd, | |
3083 | struct netdev_rxq *rx, | |
3084 | odp_port_t port_no) | |
3085 | { | |
3086 | struct dp_packet_batch batch; | |
3087 | int error; | |
3088 | ||
3089 | dp_packet_batch_init(&batch); | |
3090 | cycles_count_start(pmd); | |
3091 | error = netdev_rxq_recv(rx, &batch); | |
3092 | cycles_count_end(pmd, PMD_CYCLES_POLLING); | |
3093 | if (!error) { | |
3094 | *recirc_depth_get() = 0; | |
3095 | ||
3096 | cycles_count_start(pmd); | |
3097 | dp_netdev_input(pmd, &batch, port_no); | |
3098 | cycles_count_end(pmd, PMD_CYCLES_PROCESSING); | |
3099 | } else if (error != EAGAIN && error != EOPNOTSUPP) { | |
3100 | static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5); | |
3101 | ||
3102 | VLOG_ERR_RL(&rl, "error receiving data from %s: %s", | |
3103 | netdev_rxq_get_name(rx), ovs_strerror(error)); | |
3104 | } | |
3105 | } | |
3106 | ||
3107 | static struct tx_port * | |
3108 | tx_port_lookup(const struct hmap *hmap, odp_port_t port_no) | |
3109 | { | |
3110 | struct tx_port *tx; | |
3111 | ||
3112 | HMAP_FOR_EACH_IN_BUCKET (tx, node, hash_port_no(port_no), hmap) { | |
3113 | if (tx->port->port_no == port_no) { | |
3114 | return tx; | |
3115 | } | |
3116 | } | |
3117 | ||
3118 | return NULL; | |
3119 | } | |
3120 | ||
3121 | static int | |
3122 | port_reconfigure(struct dp_netdev_port *port) | |
3123 | { | |
3124 | struct netdev *netdev = port->netdev; | |
3125 | int i, err; | |
3126 | ||
3127 | port->need_reconfigure = false; | |
3128 | ||
3129 | /* Closes the existing 'rxq's. */ | |
3130 | for (i = 0; i < port->n_rxq; i++) { | |
3131 | netdev_rxq_close(port->rxqs[i].rx); | |
3132 | port->rxqs[i].rx = NULL; | |
3133 | } | |
3134 | port->n_rxq = 0; | |
3135 | ||
3136 | /* Allows 'netdev' to apply the pending configuration changes. */ | |
3137 | if (netdev_is_reconf_required(netdev)) { | |
3138 | err = netdev_reconfigure(netdev); | |
3139 | if (err && (err != EOPNOTSUPP)) { | |
3140 | VLOG_ERR("Failed to set interface %s new configuration", | |
3141 | netdev_get_name(netdev)); | |
3142 | return err; | |
3143 | } | |
3144 | } | |
3145 | /* If the netdev_reconfigure() above succeeds, reopens the 'rxq's. */ | |
3146 | port->rxqs = xrealloc(port->rxqs, | |
3147 | sizeof *port->rxqs * netdev_n_rxq(netdev)); | |
3148 | /* Realloc 'used' counters for tx queues. */ | |
3149 | free(port->txq_used); | |
3150 | port->txq_used = xcalloc(netdev_n_txq(netdev), sizeof *port->txq_used); | |
3151 | ||
3152 | for (i = 0; i < netdev_n_rxq(netdev); i++) { | |
3153 | port->rxqs[i].port = port; | |
3154 | err = netdev_rxq_open(netdev, &port->rxqs[i].rx, i); | |
3155 | if (err) { | |
3156 | return err; | |
3157 | } | |
3158 | port->n_rxq++; | |
3159 | } | |
3160 | ||
3161 | /* Parse affinity list to apply configuration for new queues. */ | |
3162 | dpif_netdev_port_set_rxq_affinity(port, port->rxq_affinity_list); | |
3163 | ||
3164 | return 0; | |
3165 | } | |
3166 | ||
3167 | struct rr_numa_list { | |
3168 | struct hmap numas; /* Contains 'struct rr_numa' */ | |
3169 | }; | |
3170 | ||
3171 | struct rr_numa { | |
3172 | struct hmap_node node; | |
3173 | ||
3174 | int numa_id; | |
3175 | ||
3176 | /* Non isolated pmds on numa node 'numa_id' */ | |
3177 | struct dp_netdev_pmd_thread **pmds; | |
3178 | int n_pmds; | |
3179 | ||
3180 | int cur_index; | |
3181 | }; | |
3182 | ||
3183 | static struct rr_numa * | |
3184 | rr_numa_list_lookup(struct rr_numa_list *rr, int numa_id) | |
3185 | { | |
3186 | struct rr_numa *numa; | |
3187 | ||
3188 | HMAP_FOR_EACH_WITH_HASH (numa, node, hash_int(numa_id, 0), &rr->numas) { | |
3189 | if (numa->numa_id == numa_id) { | |
3190 | return numa; | |
3191 | } | |
3192 | } | |
3193 | ||
3194 | return NULL; | |
3195 | } | |
3196 | ||
3197 | static void | |
3198 | rr_numa_list_populate(struct dp_netdev *dp, struct rr_numa_list *rr) | |
3199 | { | |
3200 | struct dp_netdev_pmd_thread *pmd; | |
3201 | struct rr_numa *numa; | |
3202 | ||
3203 | hmap_init(&rr->numas); | |
3204 | ||
3205 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
3206 | if (pmd->core_id == NON_PMD_CORE_ID || pmd->isolated) { | |
3207 | continue; | |
3208 | } | |
3209 | ||
3210 | numa = rr_numa_list_lookup(rr, pmd->numa_id); | |
3211 | if (!numa) { | |
3212 | numa = xzalloc(sizeof *numa); | |
3213 | numa->numa_id = pmd->numa_id; | |
3214 | hmap_insert(&rr->numas, &numa->node, hash_int(pmd->numa_id, 0)); | |
3215 | } | |
3216 | numa->n_pmds++; | |
3217 | numa->pmds = xrealloc(numa->pmds, numa->n_pmds * sizeof *numa->pmds); | |
3218 | numa->pmds[numa->n_pmds - 1] = pmd; | |
3219 | } | |
3220 | } | |
3221 | ||
3222 | static struct dp_netdev_pmd_thread * | |
3223 | rr_numa_get_pmd(struct rr_numa *numa) | |
3224 | { | |
3225 | return numa->pmds[numa->cur_index++ % numa->n_pmds]; | |
3226 | } | |
3227 | ||
3228 | static void | |
3229 | rr_numa_list_destroy(struct rr_numa_list *rr) | |
3230 | { | |
3231 | struct rr_numa *numa; | |
3232 | ||
3233 | HMAP_FOR_EACH_POP (numa, node, &rr->numas) { | |
3234 | free(numa->pmds); | |
3235 | free(numa); | |
3236 | } | |
3237 | hmap_destroy(&rr->numas); | |
3238 | } | |
3239 | ||
3240 | /* Assign pmds to queues. If 'pinned' is true, assign pmds to pinned | |
3241 | * queues and marks the pmds as isolated. Otherwise, assign non isolated | |
3242 | * pmds to unpinned queues. | |
3243 | * | |
3244 | * The function doesn't touch the pmd threads, it just stores the assignment | |
3245 | * in the 'pmd' member of each rxq. */ | |
3246 | static void | |
3247 | rxq_scheduling(struct dp_netdev *dp, bool pinned) OVS_REQUIRES(dp->port_mutex) | |
3248 | { | |
3249 | struct dp_netdev_port *port; | |
3250 | struct rr_numa_list rr; | |
3251 | ||
3252 | rr_numa_list_populate(dp, &rr); | |
3253 | ||
3254 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
3255 | struct rr_numa *numa; | |
3256 | int numa_id; | |
3257 | ||
3258 | if (!netdev_is_pmd(port->netdev)) { | |
3259 | continue; | |
3260 | } | |
3261 | ||
3262 | numa_id = netdev_get_numa_id(port->netdev); | |
3263 | numa = rr_numa_list_lookup(&rr, numa_id); | |
3264 | ||
3265 | for (int qid = 0; qid < port->n_rxq; qid++) { | |
3266 | struct dp_netdev_rxq *q = &port->rxqs[qid]; | |
3267 | ||
3268 | if (pinned && q->core_id != OVS_CORE_UNSPEC) { | |
3269 | struct dp_netdev_pmd_thread *pmd; | |
3270 | ||
3271 | pmd = dp_netdev_get_pmd(dp, q->core_id); | |
3272 | if (!pmd) { | |
3273 | VLOG_WARN("There is no PMD thread on core %d. Queue " | |
3274 | "%d on port \'%s\' will not be polled.", | |
3275 | q->core_id, qid, netdev_get_name(port->netdev)); | |
3276 | } else { | |
3277 | q->pmd = pmd; | |
3278 | pmd->isolated = true; | |
3279 | dp_netdev_pmd_unref(pmd); | |
3280 | } | |
3281 | } else if (!pinned && q->core_id == OVS_CORE_UNSPEC) { | |
3282 | if (!numa) { | |
3283 | VLOG_WARN("There's no available (non isolated) pmd thread " | |
3284 | "on numa node %d. Queue %d on port \'%s\' will " | |
3285 | "not be polled.", | |
3286 | numa_id, qid, netdev_get_name(port->netdev)); | |
3287 | } else { | |
3288 | q->pmd = rr_numa_get_pmd(numa); | |
3289 | } | |
3290 | } | |
3291 | } | |
3292 | } | |
3293 | ||
3294 | rr_numa_list_destroy(&rr); | |
3295 | } | |
3296 | ||
3297 | static void | |
3298 | reconfigure_pmd_threads(struct dp_netdev *dp) | |
3299 | OVS_REQUIRES(dp->port_mutex) | |
3300 | { | |
3301 | struct dp_netdev_pmd_thread *pmd; | |
3302 | struct ovs_numa_dump *pmd_cores; | |
3303 | bool changed = false; | |
3304 | ||
3305 | /* The pmd threads should be started only if there's a pmd port in the | |
3306 | * datapath. If the user didn't provide any "pmd-cpu-mask", we start | |
3307 | * NR_PMD_THREADS per numa node. */ | |
3308 | if (!has_pmd_port(dp)) { | |
3309 | pmd_cores = ovs_numa_dump_n_cores_per_numa(0); | |
3310 | } else if (dp->pmd_cmask && dp->pmd_cmask[0]) { | |
3311 | pmd_cores = ovs_numa_dump_cores_with_cmask(dp->pmd_cmask); | |
3312 | } else { | |
3313 | pmd_cores = ovs_numa_dump_n_cores_per_numa(NR_PMD_THREADS); | |
3314 | } | |
3315 | ||
3316 | /* Check for changed configuration */ | |
3317 | if (ovs_numa_dump_count(pmd_cores) != cmap_count(&dp->poll_threads) - 1) { | |
3318 | changed = true; | |
3319 | } else { | |
3320 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
3321 | if (pmd->core_id != NON_PMD_CORE_ID | |
3322 | && !ovs_numa_dump_contains_core(pmd_cores, | |
3323 | pmd->numa_id, | |
3324 | pmd->core_id)) { | |
3325 | changed = true; | |
3326 | break; | |
3327 | } | |
3328 | } | |
3329 | } | |
3330 | ||
3331 | /* Destroy the old and recreate the new pmd threads. We don't perform an | |
3332 | * incremental update because we would have to adjust 'static_tx_qid'. */ | |
3333 | if (changed) { | |
3334 | struct ovs_numa_info_core *core; | |
3335 | struct ovs_numa_info_numa *numa; | |
3336 | ||
3337 | /* Do not destroy the non pmd thread. */ | |
3338 | dp_netdev_destroy_all_pmds(dp, false); | |
3339 | FOR_EACH_CORE_ON_DUMP (core, pmd_cores) { | |
3340 | struct dp_netdev_pmd_thread *pmd = xzalloc(sizeof *pmd); | |
3341 | ||
3342 | dp_netdev_configure_pmd(pmd, dp, core->core_id, core->numa_id); | |
3343 | ||
3344 | pmd->thread = ovs_thread_create("pmd", pmd_thread_main, pmd); | |
3345 | } | |
3346 | ||
3347 | /* Log the number of pmd threads per numa node. */ | |
3348 | FOR_EACH_NUMA_ON_DUMP (numa, pmd_cores) { | |
3349 | VLOG_INFO("Created %"PRIuSIZE" pmd threads on numa node %d", | |
3350 | numa->n_cores, numa->numa_id); | |
3351 | } | |
3352 | } | |
3353 | ||
3354 | ovs_numa_dump_destroy(pmd_cores); | |
3355 | } | |
3356 | ||
3357 | static void | |
3358 | reload_affected_pmds(struct dp_netdev *dp) | |
3359 | { | |
3360 | struct dp_netdev_pmd_thread *pmd; | |
3361 | ||
3362 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
3363 | if (pmd->need_reload) { | |
3364 | dp_netdev_reload_pmd__(pmd); | |
3365 | pmd->need_reload = false; | |
3366 | } | |
3367 | } | |
3368 | } | |
3369 | ||
3370 | static void | |
3371 | pmd_remove_stale_ports(struct dp_netdev *dp, | |
3372 | struct dp_netdev_pmd_thread *pmd) | |
3373 | OVS_EXCLUDED(pmd->port_mutex) | |
3374 | OVS_REQUIRES(dp->port_mutex) | |
3375 | { | |
3376 | struct rxq_poll *poll, *poll_next; | |
3377 | struct tx_port *tx, *tx_next; | |
3378 | ||
3379 | ovs_mutex_lock(&pmd->port_mutex); | |
3380 | HMAP_FOR_EACH_SAFE (poll, poll_next, node, &pmd->poll_list) { | |
3381 | struct dp_netdev_port *port = poll->rxq->port; | |
3382 | ||
3383 | if (port->need_reconfigure | |
3384 | || !hmap_contains(&dp->ports, &port->node)) { | |
3385 | dp_netdev_del_rxq_from_pmd(pmd, poll); | |
3386 | } | |
3387 | } | |
3388 | HMAP_FOR_EACH_SAFE (tx, tx_next, node, &pmd->tx_ports) { | |
3389 | struct dp_netdev_port *port = tx->port; | |
3390 | ||
3391 | if (port->need_reconfigure | |
3392 | || !hmap_contains(&dp->ports, &port->node)) { | |
3393 | dp_netdev_del_port_tx_from_pmd(pmd, tx); | |
3394 | } | |
3395 | } | |
3396 | ovs_mutex_unlock(&pmd->port_mutex); | |
3397 | } | |
3398 | ||
3399 | /* Must be called each time a port is added/removed or the cmask changes. | |
3400 | * This creates and destroys pmd threads, reconfigures ports, opens their | |
3401 | * rxqs and assigns all rxqs/txqs to pmd threads. */ | |
3402 | static void | |
3403 | reconfigure_datapath(struct dp_netdev *dp) | |
3404 | OVS_REQUIRES(dp->port_mutex) | |
3405 | { | |
3406 | struct dp_netdev_pmd_thread *pmd; | |
3407 | struct dp_netdev_port *port; | |
3408 | int wanted_txqs; | |
3409 | ||
3410 | dp->last_reconfigure_seq = seq_read(dp->reconfigure_seq); | |
3411 | ||
3412 | /* Step 1: Adjust the pmd threads based on the datapath ports, the cores | |
3413 | * on the system and the user configuration. */ | |
3414 | reconfigure_pmd_threads(dp); | |
3415 | ||
3416 | wanted_txqs = cmap_count(&dp->poll_threads); | |
3417 | ||
3418 | /* The number of pmd threads might have changed, or a port can be new: | |
3419 | * adjust the txqs. */ | |
3420 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
3421 | netdev_set_tx_multiq(port->netdev, wanted_txqs); | |
3422 | } | |
3423 | ||
3424 | /* Step 2: Remove from the pmd threads ports that have been removed or | |
3425 | * need reconfiguration. */ | |
3426 | ||
3427 | /* Check for all the ports that need reconfiguration. We cache this in | |
3428 | * 'port->reconfigure', because netdev_is_reconf_required() can change at | |
3429 | * any time. */ | |
3430 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
3431 | if (netdev_is_reconf_required(port->netdev)) { | |
3432 | port->need_reconfigure = true; | |
3433 | } | |
3434 | } | |
3435 | ||
3436 | /* Remove from the pmd threads all the ports that have been deleted or | |
3437 | * need reconfiguration. */ | |
3438 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
3439 | pmd_remove_stale_ports(dp, pmd); | |
3440 | } | |
3441 | ||
3442 | /* Reload affected pmd threads. We must wait for the pmd threads before | |
3443 | * reconfiguring the ports, because a port cannot be reconfigured while | |
3444 | * it's being used. */ | |
3445 | reload_affected_pmds(dp); | |
3446 | ||
3447 | /* Step 3: Reconfigure ports. */ | |
3448 | ||
3449 | /* We only reconfigure the ports that we determined above, because they're | |
3450 | * not being used by any pmd thread at the moment. If a port fails to | |
3451 | * reconfigure we remove it from the datapath. */ | |
3452 | struct dp_netdev_port *next_port; | |
3453 | HMAP_FOR_EACH_SAFE (port, next_port, node, &dp->ports) { | |
3454 | int err; | |
3455 | ||
3456 | if (!port->need_reconfigure) { | |
3457 | continue; | |
3458 | } | |
3459 | ||
3460 | err = port_reconfigure(port); | |
3461 | if (err) { | |
3462 | hmap_remove(&dp->ports, &port->node); | |
3463 | seq_change(dp->port_seq); | |
3464 | port_destroy(port); | |
3465 | } else { | |
3466 | port->dynamic_txqs = netdev_n_txq(port->netdev) < wanted_txqs; | |
3467 | } | |
3468 | } | |
3469 | ||
3470 | /* Step 4: Compute new rxq scheduling. We don't touch the pmd threads | |
3471 | * for now, we just update the 'pmd' pointer in each rxq to point to the | |
3472 | * wanted thread according to the scheduling policy. */ | |
3473 | ||
3474 | /* Reset all the pmd threads to non isolated. */ | |
3475 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
3476 | pmd->isolated = false; | |
3477 | } | |
3478 | ||
3479 | /* Reset all the queues to unassigned */ | |
3480 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
3481 | for (int i = 0; i < port->n_rxq; i++) { | |
3482 | port->rxqs[i].pmd = NULL; | |
3483 | } | |
3484 | } | |
3485 | ||
3486 | /* Add pinned queues and mark pmd threads isolated. */ | |
3487 | rxq_scheduling(dp, true); | |
3488 | ||
3489 | /* Add non-pinned queues. */ | |
3490 | rxq_scheduling(dp, false); | |
3491 | ||
3492 | /* Step 5: Remove queues not compliant with new scheduling. */ | |
3493 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
3494 | struct rxq_poll *poll, *poll_next; | |
3495 | ||
3496 | ovs_mutex_lock(&pmd->port_mutex); | |
3497 | HMAP_FOR_EACH_SAFE (poll, poll_next, node, &pmd->poll_list) { | |
3498 | if (poll->rxq->pmd != pmd) { | |
3499 | dp_netdev_del_rxq_from_pmd(pmd, poll); | |
3500 | } | |
3501 | } | |
3502 | ovs_mutex_unlock(&pmd->port_mutex); | |
3503 | } | |
3504 | ||
3505 | /* Reload affected pmd threads. We must wait for the pmd threads to remove | |
3506 | * the old queues before readding them, otherwise a queue can be polled by | |
3507 | * two threads at the same time. */ | |
3508 | reload_affected_pmds(dp); | |
3509 | ||
3510 | /* Step 6: Add queues from scheduling, if they're not there already. */ | |
3511 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
3512 | if (!netdev_is_pmd(port->netdev)) { | |
3513 | continue; | |
3514 | } | |
3515 | ||
3516 | for (int qid = 0; qid < port->n_rxq; qid++) { | |
3517 | struct dp_netdev_rxq *q = &port->rxqs[qid]; | |
3518 | ||
3519 | if (q->pmd) { | |
3520 | ovs_mutex_lock(&q->pmd->port_mutex); | |
3521 | dp_netdev_add_rxq_to_pmd(q->pmd, q); | |
3522 | ovs_mutex_unlock(&q->pmd->port_mutex); | |
3523 | } | |
3524 | } | |
3525 | } | |
3526 | ||
3527 | /* Add every port to the tx cache of every pmd thread, if it's not | |
3528 | * there already and if this pmd has at least one rxq to poll. */ | |
3529 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
3530 | ovs_mutex_lock(&pmd->port_mutex); | |
3531 | if (hmap_count(&pmd->poll_list) || pmd->core_id == NON_PMD_CORE_ID) { | |
3532 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
3533 | dp_netdev_add_port_tx_to_pmd(pmd, port); | |
3534 | } | |
3535 | } | |
3536 | ovs_mutex_unlock(&pmd->port_mutex); | |
3537 | } | |
3538 | ||
3539 | /* Reload affected pmd threads. */ | |
3540 | reload_affected_pmds(dp); | |
3541 | } | |
3542 | ||
3543 | /* Returns true if one of the netdevs in 'dp' requires a reconfiguration */ | |
3544 | static bool | |
3545 | ports_require_restart(const struct dp_netdev *dp) | |
3546 | OVS_REQUIRES(dp->port_mutex) | |
3547 | { | |
3548 | struct dp_netdev_port *port; | |
3549 | ||
3550 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
3551 | if (netdev_is_reconf_required(port->netdev)) { | |
3552 | return true; | |
3553 | } | |
3554 | } | |
3555 | ||
3556 | return false; | |
3557 | } | |
3558 | ||
3559 | /* Return true if needs to revalidate datapath flows. */ | |
3560 | static bool | |
3561 | dpif_netdev_run(struct dpif *dpif) | |
3562 | { | |
3563 | struct dp_netdev_port *port; | |
3564 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
3565 | struct dp_netdev_pmd_thread *non_pmd; | |
3566 | uint64_t new_tnl_seq; | |
3567 | ||
3568 | ovs_mutex_lock(&dp->port_mutex); | |
3569 | non_pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID); | |
3570 | if (non_pmd) { | |
3571 | ovs_mutex_lock(&dp->non_pmd_mutex); | |
3572 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
3573 | if (!netdev_is_pmd(port->netdev)) { | |
3574 | int i; | |
3575 | ||
3576 | for (i = 0; i < port->n_rxq; i++) { | |
3577 | dp_netdev_process_rxq_port(non_pmd, port->rxqs[i].rx, | |
3578 | port->port_no); | |
3579 | } | |
3580 | } | |
3581 | } | |
3582 | dpif_netdev_xps_revalidate_pmd(non_pmd, time_msec(), false); | |
3583 | ovs_mutex_unlock(&dp->non_pmd_mutex); | |
3584 | ||
3585 | dp_netdev_pmd_unref(non_pmd); | |
3586 | } | |
3587 | ||
3588 | if (dp_netdev_is_reconf_required(dp) || ports_require_restart(dp)) { | |
3589 | reconfigure_datapath(dp); | |
3590 | } | |
3591 | ovs_mutex_unlock(&dp->port_mutex); | |
3592 | ||
3593 | tnl_neigh_cache_run(); | |
3594 | tnl_port_map_run(); | |
3595 | new_tnl_seq = seq_read(tnl_conf_seq); | |
3596 | ||
3597 | if (dp->last_tnl_conf_seq != new_tnl_seq) { | |
3598 | dp->last_tnl_conf_seq = new_tnl_seq; | |
3599 | return true; | |
3600 | } | |
3601 | return false; | |
3602 | } | |
3603 | ||
3604 | static void | |
3605 | dpif_netdev_wait(struct dpif *dpif) | |
3606 | { | |
3607 | struct dp_netdev_port *port; | |
3608 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
3609 | ||
3610 | ovs_mutex_lock(&dp_netdev_mutex); | |
3611 | ovs_mutex_lock(&dp->port_mutex); | |
3612 | HMAP_FOR_EACH (port, node, &dp->ports) { | |
3613 | netdev_wait_reconf_required(port->netdev); | |
3614 | if (!netdev_is_pmd(port->netdev)) { | |
3615 | int i; | |
3616 | ||
3617 | for (i = 0; i < port->n_rxq; i++) { | |
3618 | netdev_rxq_wait(port->rxqs[i].rx); | |
3619 | } | |
3620 | } | |
3621 | } | |
3622 | ovs_mutex_unlock(&dp->port_mutex); | |
3623 | ovs_mutex_unlock(&dp_netdev_mutex); | |
3624 | seq_wait(tnl_conf_seq, dp->last_tnl_conf_seq); | |
3625 | } | |
3626 | ||
3627 | static void | |
3628 | pmd_free_cached_ports(struct dp_netdev_pmd_thread *pmd) | |
3629 | { | |
3630 | struct tx_port *tx_port_cached; | |
3631 | ||
3632 | /* Free all used tx queue ids. */ | |
3633 | dpif_netdev_xps_revalidate_pmd(pmd, 0, true); | |
3634 | ||
3635 | HMAP_FOR_EACH_POP (tx_port_cached, node, &pmd->tnl_port_cache) { | |
3636 | free(tx_port_cached); | |
3637 | } | |
3638 | HMAP_FOR_EACH_POP (tx_port_cached, node, &pmd->send_port_cache) { | |
3639 | free(tx_port_cached); | |
3640 | } | |
3641 | } | |
3642 | ||
3643 | /* Copies ports from 'pmd->tx_ports' (shared with the main thread) to | |
3644 | * 'pmd->port_cache' (thread local) */ | |
3645 | static void | |
3646 | pmd_load_cached_ports(struct dp_netdev_pmd_thread *pmd) | |
3647 | OVS_REQUIRES(pmd->port_mutex) | |
3648 | { | |
3649 | struct tx_port *tx_port, *tx_port_cached; | |
3650 | ||
3651 | pmd_free_cached_ports(pmd); | |
3652 | hmap_shrink(&pmd->send_port_cache); | |
3653 | hmap_shrink(&pmd->tnl_port_cache); | |
3654 | ||
3655 | HMAP_FOR_EACH (tx_port, node, &pmd->tx_ports) { | |
3656 | if (netdev_has_tunnel_push_pop(tx_port->port->netdev)) { | |
3657 | tx_port_cached = xmemdup(tx_port, sizeof *tx_port_cached); | |
3658 | hmap_insert(&pmd->tnl_port_cache, &tx_port_cached->node, | |
3659 | hash_port_no(tx_port_cached->port->port_no)); | |
3660 | } | |
3661 | ||
3662 | if (netdev_n_txq(tx_port->port->netdev)) { | |
3663 | tx_port_cached = xmemdup(tx_port, sizeof *tx_port_cached); | |
3664 | hmap_insert(&pmd->send_port_cache, &tx_port_cached->node, | |
3665 | hash_port_no(tx_port_cached->port->port_no)); | |
3666 | } | |
3667 | } | |
3668 | } | |
3669 | ||
3670 | static int | |
3671 | pmd_load_queues_and_ports(struct dp_netdev_pmd_thread *pmd, | |
3672 | struct polled_queue **ppoll_list) | |
3673 | { | |
3674 | struct polled_queue *poll_list = *ppoll_list; | |
3675 | struct rxq_poll *poll; | |
3676 | int i; | |
3677 | ||
3678 | ovs_mutex_lock(&pmd->port_mutex); | |
3679 | poll_list = xrealloc(poll_list, hmap_count(&pmd->poll_list) | |
3680 | * sizeof *poll_list); | |
3681 | ||
3682 | i = 0; | |
3683 | HMAP_FOR_EACH (poll, node, &pmd->poll_list) { | |
3684 | poll_list[i].rx = poll->rxq->rx; | |
3685 | poll_list[i].port_no = poll->rxq->port->port_no; | |
3686 | i++; | |
3687 | } | |
3688 | ||
3689 | pmd_load_cached_ports(pmd); | |
3690 | ||
3691 | ovs_mutex_unlock(&pmd->port_mutex); | |
3692 | ||
3693 | *ppoll_list = poll_list; | |
3694 | return i; | |
3695 | } | |
3696 | ||
3697 | static void * | |
3698 | pmd_thread_main(void *f_) | |
3699 | { | |
3700 | struct dp_netdev_pmd_thread *pmd = f_; | |
3701 | unsigned int lc = 0; | |
3702 | struct polled_queue *poll_list; | |
3703 | bool exiting; | |
3704 | int poll_cnt; | |
3705 | int i; | |
3706 | ||
3707 | poll_list = NULL; | |
3708 | ||
3709 | /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */ | |
3710 | ovsthread_setspecific(pmd->dp->per_pmd_key, pmd); | |
3711 | ovs_numa_thread_setaffinity_core(pmd->core_id); | |
3712 | dpdk_set_lcore_id(pmd->core_id); | |
3713 | poll_cnt = pmd_load_queues_and_ports(pmd, &poll_list); | |
3714 | reload: | |
3715 | emc_cache_init(&pmd->flow_cache); | |
3716 | ||
3717 | /* List port/core affinity */ | |
3718 | for (i = 0; i < poll_cnt; i++) { | |
3719 | VLOG_DBG("Core %d processing port \'%s\' with queue-id %d\n", | |
3720 | pmd->core_id, netdev_rxq_get_name(poll_list[i].rx), | |
3721 | netdev_rxq_get_queue_id(poll_list[i].rx)); | |
3722 | } | |
3723 | ||
3724 | if (!poll_cnt) { | |
3725 | while (seq_read(pmd->reload_seq) == pmd->last_reload_seq) { | |
3726 | seq_wait(pmd->reload_seq, pmd->last_reload_seq); | |
3727 | poll_block(); | |
3728 | } | |
3729 | lc = UINT_MAX; | |
3730 | } | |
3731 | ||
3732 | for (;;) { | |
3733 | for (i = 0; i < poll_cnt; i++) { | |
3734 | dp_netdev_process_rxq_port(pmd, poll_list[i].rx, | |
3735 | poll_list[i].port_no); | |
3736 | } | |
3737 | ||
3738 | if (lc++ > 1024) { | |
3739 | bool reload; | |
3740 | ||
3741 | lc = 0; | |
3742 | ||
3743 | coverage_try_clear(); | |
3744 | dp_netdev_pmd_try_optimize(pmd); | |
3745 | if (!ovsrcu_try_quiesce()) { | |
3746 | emc_cache_slow_sweep(&pmd->flow_cache); | |
3747 | } | |
3748 | ||
3749 | atomic_read_relaxed(&pmd->reload, &reload); | |
3750 | if (reload) { | |
3751 | break; | |
3752 | } | |
3753 | } | |
3754 | } | |
3755 | ||
3756 | poll_cnt = pmd_load_queues_and_ports(pmd, &poll_list); | |
3757 | exiting = latch_is_set(&pmd->exit_latch); | |
3758 | /* Signal here to make sure the pmd finishes | |
3759 | * reloading the updated configuration. */ | |
3760 | dp_netdev_pmd_reload_done(pmd); | |
3761 | ||
3762 | emc_cache_uninit(&pmd->flow_cache); | |
3763 | ||
3764 | if (!exiting) { | |
3765 | goto reload; | |
3766 | } | |
3767 | ||
3768 | free(poll_list); | |
3769 | pmd_free_cached_ports(pmd); | |
3770 | return NULL; | |
3771 | } | |
3772 | ||
3773 | static void | |
3774 | dp_netdev_disable_upcall(struct dp_netdev *dp) | |
3775 | OVS_ACQUIRES(dp->upcall_rwlock) | |
3776 | { | |
3777 | fat_rwlock_wrlock(&dp->upcall_rwlock); | |
3778 | } | |
3779 | ||
3780 | \f | |
3781 | /* Meters */ | |
3782 | static void | |
3783 | dpif_netdev_meter_get_features(const struct dpif * dpif OVS_UNUSED, | |
3784 | struct ofputil_meter_features *features) | |
3785 | { | |
3786 | features->max_meters = MAX_METERS; | |
3787 | features->band_types = DP_SUPPORTED_METER_BAND_TYPES; | |
3788 | features->capabilities = DP_SUPPORTED_METER_FLAGS_MASK; | |
3789 | features->max_bands = MAX_BANDS; | |
3790 | features->max_color = 0; | |
3791 | } | |
3792 | ||
3793 | /* Returns false when packet needs to be dropped. */ | |
3794 | static void | |
3795 | dp_netdev_run_meter(struct dp_netdev *dp, struct dp_packet_batch *packets_, | |
3796 | uint32_t meter_id, long long int now) | |
3797 | { | |
3798 | struct dp_meter *meter; | |
3799 | struct dp_meter_band *band; | |
3800 | long long int long_delta_t; /* msec */ | |
3801 | uint32_t delta_t; /* msec */ | |
3802 | int i; | |
3803 | int cnt = packets_->count; | |
3804 | uint32_t bytes, volume; | |
3805 | int exceeded_band[NETDEV_MAX_BURST]; | |
3806 | uint32_t exceeded_rate[NETDEV_MAX_BURST]; | |
3807 | int exceeded_pkt = cnt; /* First packet that exceeded a band rate. */ | |
3808 | ||
3809 | if (meter_id >= MAX_METERS) { | |
3810 | return; | |
3811 | } | |
3812 | ||
3813 | meter_lock(dp, meter_id); | |
3814 | meter = dp->meters[meter_id]; | |
3815 | if (!meter) { | |
3816 | goto out; | |
3817 | } | |
3818 | ||
3819 | /* Initialize as negative values. */ | |
3820 | memset(exceeded_band, 0xff, cnt * sizeof *exceeded_band); | |
3821 | /* Initialize as zeroes. */ | |
3822 | memset(exceeded_rate, 0, cnt * sizeof *exceeded_rate); | |
3823 | ||
3824 | /* All packets will hit the meter at the same time. */ | |
3825 | long_delta_t = (now - meter->used); /* msec */ | |
3826 | ||
3827 | /* Make sure delta_t will not be too large, so that bucket will not | |
3828 | * wrap around below. */ | |
3829 | delta_t = (long_delta_t > (long long int)meter->max_delta_t) | |
3830 | ? meter->max_delta_t : (uint32_t)long_delta_t; | |
3831 | ||
3832 | /* Update meter stats. */ | |
3833 | meter->used = now; | |
3834 | meter->packet_count += cnt; | |
3835 | bytes = 0; | |
3836 | for (i = 0; i < cnt; i++) { | |
3837 | bytes += dp_packet_size(packets_->packets[i]); | |
3838 | } | |
3839 | meter->byte_count += bytes; | |
3840 | ||
3841 | /* Meters can operate in terms of packets per second or kilobits per | |
3842 | * second. */ | |
3843 | if (meter->flags & OFPMF13_PKTPS) { | |
3844 | /* Rate in packets/second, bucket 1/1000 packets. */ | |
3845 | /* msec * packets/sec = 1/1000 packets. */ | |
3846 | volume = cnt * 1000; /* Take 'cnt' packets from the bucket. */ | |
3847 | } else { | |
3848 | /* Rate in kbps, bucket in bits. */ | |
3849 | /* msec * kbps = bits */ | |
3850 | volume = bytes * 8; | |
3851 | } | |
3852 | ||
3853 | /* Update all bands and find the one hit with the highest rate for each | |
3854 | * packet (if any). */ | |
3855 | for (int m = 0; m < meter->n_bands; ++m) { | |
3856 | band = &meter->bands[m]; | |
3857 | ||
3858 | /* Update band's bucket. */ | |
3859 | band->bucket += delta_t * band->up.rate; | |
3860 | if (band->bucket > band->up.burst_size) { | |
3861 | band->bucket = band->up.burst_size; | |
3862 | } | |
3863 | ||
3864 | /* Drain the bucket for all the packets, if possible. */ | |
3865 | if (band->bucket >= volume) { | |
3866 | band->bucket -= volume; | |
3867 | } else { | |
3868 | int band_exceeded_pkt; | |
3869 | ||
3870 | /* Band limit hit, must process packet-by-packet. */ | |
3871 | if (meter->flags & OFPMF13_PKTPS) { | |
3872 | band_exceeded_pkt = band->bucket / 1000; | |
3873 | band->bucket %= 1000; /* Remainder stays in bucket. */ | |
3874 | ||
3875 | /* Update the exceeding band for each exceeding packet. | |
3876 | * (Only one band will be fired by a packet, and that | |
3877 | * can be different for each packet.) */ | |
3878 | for (i = band_exceeded_pkt; i < cnt; i++) { | |
3879 | if (band->up.rate > exceeded_rate[i]) { | |
3880 | exceeded_rate[i] = band->up.rate; | |
3881 | exceeded_band[i] = m; | |
3882 | } | |
3883 | } | |
3884 | } else { | |
3885 | /* Packet sizes differ, must process one-by-one. */ | |
3886 | band_exceeded_pkt = cnt; | |
3887 | for (i = 0; i < cnt; i++) { | |
3888 | uint32_t bits = dp_packet_size(packets_->packets[i]) * 8; | |
3889 | ||
3890 | if (band->bucket >= bits) { | |
3891 | band->bucket -= bits; | |
3892 | } else { | |
3893 | if (i < band_exceeded_pkt) { | |
3894 | band_exceeded_pkt = i; | |
3895 | } | |
3896 | /* Update the exceeding band for the exceeding packet. | |
3897 | * (Only one band will be fired by a packet, and that | |
3898 | * can be different for each packet.) */ | |
3899 | if (band->up.rate > exceeded_rate[i]) { | |
3900 | exceeded_rate[i] = band->up.rate; | |
3901 | exceeded_band[i] = m; | |
3902 | } | |
3903 | } | |
3904 | } | |
3905 | } | |
3906 | /* Remember the first exceeding packet. */ | |
3907 | if (exceeded_pkt > band_exceeded_pkt) { | |
3908 | exceeded_pkt = band_exceeded_pkt; | |
3909 | } | |
3910 | } | |
3911 | } | |
3912 | ||
3913 | /* Fire the highest rate band exceeded by each packet. | |
3914 | * Drop packets if needed, by swapping packet to the end that will be | |
3915 | * ignored. */ | |
3916 | const size_t size = dp_packet_batch_size(packets_); | |
3917 | struct dp_packet *packet; | |
3918 | size_t j; | |
3919 | DP_PACKET_BATCH_REFILL_FOR_EACH (j, size, packet, packets_) { | |
3920 | if (exceeded_band[j] >= 0) { | |
3921 | /* Meter drop packet. */ | |
3922 | band = &meter->bands[exceeded_band[j]]; | |
3923 | band->packet_count += 1; | |
3924 | band->byte_count += dp_packet_size(packet); | |
3925 | ||
3926 | dp_packet_delete(packet); | |
3927 | } else { | |
3928 | /* Meter accepts packet. */ | |
3929 | dp_packet_batch_refill(packets_, packet, j); | |
3930 | } | |
3931 | } | |
3932 | out: | |
3933 | meter_unlock(dp, meter_id); | |
3934 | } | |
3935 | ||
3936 | /* Meter set/get/del processing is still single-threaded. */ | |
3937 | static int | |
3938 | dpif_netdev_meter_set(struct dpif *dpif, ofproto_meter_id *meter_id, | |
3939 | struct ofputil_meter_config *config) | |
3940 | { | |
3941 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
3942 | uint32_t mid = meter_id->uint32; | |
3943 | struct dp_meter *meter; | |
3944 | int i; | |
3945 | ||
3946 | if (mid >= MAX_METERS) { | |
3947 | return EFBIG; /* Meter_id out of range. */ | |
3948 | } | |
3949 | ||
3950 | if (config->flags & ~DP_SUPPORTED_METER_FLAGS_MASK || | |
3951 | !(config->flags & (OFPMF13_KBPS | OFPMF13_PKTPS))) { | |
3952 | return EBADF; /* Unsupported flags set */ | |
3953 | } | |
3954 | /* Validate bands */ | |
3955 | if (config->n_bands == 0 || config->n_bands > MAX_BANDS) { | |
3956 | return EINVAL; /* Too many bands */ | |
3957 | } | |
3958 | for (i = 0; i < config->n_bands; ++i) { | |
3959 | switch (config->bands[i].type) { | |
3960 | case OFPMBT13_DROP: | |
3961 | break; | |
3962 | default: | |
3963 | return ENODEV; /* Unsupported band type */ | |
3964 | } | |
3965 | } | |
3966 | ||
3967 | /* Allocate meter */ | |
3968 | meter = xzalloc(sizeof *meter | |
3969 | + config->n_bands * sizeof(struct dp_meter_band)); | |
3970 | if (meter) { | |
3971 | meter->flags = config->flags; | |
3972 | meter->n_bands = config->n_bands; | |
3973 | meter->max_delta_t = 0; | |
3974 | meter->used = time_msec(); | |
3975 | ||
3976 | /* set up bands */ | |
3977 | for (i = 0; i < config->n_bands; ++i) { | |
3978 | uint32_t band_max_delta_t; | |
3979 | ||
3980 | /* Set burst size to a workable value if none specified. */ | |
3981 | if (config->bands[i].burst_size == 0) { | |
3982 | config->bands[i].burst_size = config->bands[i].rate; | |
3983 | } | |
3984 | ||
3985 | meter->bands[i].up = config->bands[i]; | |
3986 | /* Convert burst size to the bucket units: */ | |
3987 | /* pkts => 1/1000 packets, kilobits => bits. */ | |
3988 | meter->bands[i].up.burst_size *= 1000; | |
3989 | /* Initialize bucket to empty. */ | |
3990 | meter->bands[i].bucket = 0; | |
3991 | ||
3992 | /* Figure out max delta_t that is enough to fill any bucket. */ | |
3993 | band_max_delta_t | |
3994 | = meter->bands[i].up.burst_size / meter->bands[i].up.rate; | |
3995 | if (band_max_delta_t > meter->max_delta_t) { | |
3996 | meter->max_delta_t = band_max_delta_t; | |
3997 | } | |
3998 | } | |
3999 | ||
4000 | meter_lock(dp, mid); | |
4001 | dp_delete_meter(dp, mid); /* Free existing meter, if any */ | |
4002 | dp->meters[mid] = meter; | |
4003 | meter_unlock(dp, mid); | |
4004 | ||
4005 | return 0; | |
4006 | } | |
4007 | return ENOMEM; | |
4008 | } | |
4009 | ||
4010 | static int | |
4011 | dpif_netdev_meter_get(const struct dpif *dpif, | |
4012 | ofproto_meter_id meter_id_, | |
4013 | struct ofputil_meter_stats *stats, uint16_t n_bands) | |
4014 | { | |
4015 | const struct dp_netdev *dp = get_dp_netdev(dpif); | |
4016 | const struct dp_meter *meter; | |
4017 | uint32_t meter_id = meter_id_.uint32; | |
4018 | ||
4019 | if (meter_id >= MAX_METERS) { | |
4020 | return EFBIG; | |
4021 | } | |
4022 | meter = dp->meters[meter_id]; | |
4023 | if (!meter) { | |
4024 | return ENOENT; | |
4025 | } | |
4026 | if (stats) { | |
4027 | int i = 0; | |
4028 | ||
4029 | meter_lock(dp, meter_id); | |
4030 | stats->packet_in_count = meter->packet_count; | |
4031 | stats->byte_in_count = meter->byte_count; | |
4032 | ||
4033 | for (i = 0; i < n_bands && i < meter->n_bands; ++i) { | |
4034 | stats->bands[i].packet_count = meter->bands[i].packet_count; | |
4035 | stats->bands[i].byte_count = meter->bands[i].byte_count; | |
4036 | } | |
4037 | meter_unlock(dp, meter_id); | |
4038 | ||
4039 | stats->n_bands = i; | |
4040 | } | |
4041 | return 0; | |
4042 | } | |
4043 | ||
4044 | static int | |
4045 | dpif_netdev_meter_del(struct dpif *dpif, | |
4046 | ofproto_meter_id meter_id_, | |
4047 | struct ofputil_meter_stats *stats, uint16_t n_bands) | |
4048 | { | |
4049 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
4050 | int error; | |
4051 | ||
4052 | error = dpif_netdev_meter_get(dpif, meter_id_, stats, n_bands); | |
4053 | if (!error) { | |
4054 | uint32_t meter_id = meter_id_.uint32; | |
4055 | ||
4056 | meter_lock(dp, meter_id); | |
4057 | dp_delete_meter(dp, meter_id); | |
4058 | meter_unlock(dp, meter_id); | |
4059 | } | |
4060 | return error; | |
4061 | } | |
4062 | ||
4063 | \f | |
4064 | static void | |
4065 | dpif_netdev_disable_upcall(struct dpif *dpif) | |
4066 | OVS_NO_THREAD_SAFETY_ANALYSIS | |
4067 | { | |
4068 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
4069 | dp_netdev_disable_upcall(dp); | |
4070 | } | |
4071 | ||
4072 | static void | |
4073 | dp_netdev_enable_upcall(struct dp_netdev *dp) | |
4074 | OVS_RELEASES(dp->upcall_rwlock) | |
4075 | { | |
4076 | fat_rwlock_unlock(&dp->upcall_rwlock); | |
4077 | } | |
4078 | ||
4079 | static void | |
4080 | dpif_netdev_enable_upcall(struct dpif *dpif) | |
4081 | OVS_NO_THREAD_SAFETY_ANALYSIS | |
4082 | { | |
4083 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
4084 | dp_netdev_enable_upcall(dp); | |
4085 | } | |
4086 | ||
4087 | static void | |
4088 | dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd) | |
4089 | { | |
4090 | ovs_mutex_lock(&pmd->cond_mutex); | |
4091 | atomic_store_relaxed(&pmd->reload, false); | |
4092 | pmd->last_reload_seq = seq_read(pmd->reload_seq); | |
4093 | xpthread_cond_signal(&pmd->cond); | |
4094 | ovs_mutex_unlock(&pmd->cond_mutex); | |
4095 | } | |
4096 | ||
4097 | /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns | |
4098 | * the pointer if succeeds, otherwise, NULL (it can return NULL even if | |
4099 | * 'core_id' is NON_PMD_CORE_ID). | |
4100 | * | |
4101 | * Caller must unrefs the returned reference. */ | |
4102 | static struct dp_netdev_pmd_thread * | |
4103 | dp_netdev_get_pmd(struct dp_netdev *dp, unsigned core_id) | |
4104 | { | |
4105 | struct dp_netdev_pmd_thread *pmd; | |
4106 | const struct cmap_node *pnode; | |
4107 | ||
4108 | pnode = cmap_find(&dp->poll_threads, hash_int(core_id, 0)); | |
4109 | if (!pnode) { | |
4110 | return NULL; | |
4111 | } | |
4112 | pmd = CONTAINER_OF(pnode, struct dp_netdev_pmd_thread, node); | |
4113 | ||
4114 | return dp_netdev_pmd_try_ref(pmd) ? pmd : NULL; | |
4115 | } | |
4116 | ||
4117 | /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */ | |
4118 | static void | |
4119 | dp_netdev_set_nonpmd(struct dp_netdev *dp) | |
4120 | OVS_REQUIRES(dp->port_mutex) | |
4121 | { | |
4122 | struct dp_netdev_pmd_thread *non_pmd; | |
4123 | ||
4124 | non_pmd = xzalloc(sizeof *non_pmd); | |
4125 | dp_netdev_configure_pmd(non_pmd, dp, NON_PMD_CORE_ID, OVS_NUMA_UNSPEC); | |
4126 | } | |
4127 | ||
4128 | /* Caller must have valid pointer to 'pmd'. */ | |
4129 | static bool | |
4130 | dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd) | |
4131 | { | |
4132 | return ovs_refcount_try_ref_rcu(&pmd->ref_cnt); | |
4133 | } | |
4134 | ||
4135 | static void | |
4136 | dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd) | |
4137 | { | |
4138 | if (pmd && ovs_refcount_unref(&pmd->ref_cnt) == 1) { | |
4139 | ovsrcu_postpone(dp_netdev_destroy_pmd, pmd); | |
4140 | } | |
4141 | } | |
4142 | ||
4143 | /* Given cmap position 'pos', tries to ref the next node. If try_ref() | |
4144 | * fails, keeps checking for next node until reaching the end of cmap. | |
4145 | * | |
4146 | * Caller must unrefs the returned reference. */ | |
4147 | static struct dp_netdev_pmd_thread * | |
4148 | dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos) | |
4149 | { | |
4150 | struct dp_netdev_pmd_thread *next; | |
4151 | ||
4152 | do { | |
4153 | struct cmap_node *node; | |
4154 | ||
4155 | node = cmap_next_position(&dp->poll_threads, pos); | |
4156 | next = node ? CONTAINER_OF(node, struct dp_netdev_pmd_thread, node) | |
4157 | : NULL; | |
4158 | } while (next && !dp_netdev_pmd_try_ref(next)); | |
4159 | ||
4160 | return next; | |
4161 | } | |
4162 | ||
4163 | /* Configures the 'pmd' based on the input argument. */ | |
4164 | static void | |
4165 | dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp, | |
4166 | unsigned core_id, int numa_id) | |
4167 | { | |
4168 | pmd->dp = dp; | |
4169 | pmd->core_id = core_id; | |
4170 | pmd->numa_id = numa_id; | |
4171 | pmd->need_reload = false; | |
4172 | ||
4173 | *CONST_CAST(int *, &pmd->static_tx_qid) = cmap_count(&dp->poll_threads); | |
4174 | ||
4175 | ovs_refcount_init(&pmd->ref_cnt); | |
4176 | latch_init(&pmd->exit_latch); | |
4177 | pmd->reload_seq = seq_create(); | |
4178 | pmd->last_reload_seq = seq_read(pmd->reload_seq); | |
4179 | atomic_init(&pmd->reload, false); | |
4180 | xpthread_cond_init(&pmd->cond, NULL); | |
4181 | ovs_mutex_init(&pmd->cond_mutex); | |
4182 | ovs_mutex_init(&pmd->flow_mutex); | |
4183 | ovs_mutex_init(&pmd->port_mutex); | |
4184 | cmap_init(&pmd->flow_table); | |
4185 | cmap_init(&pmd->classifiers); | |
4186 | pmd->next_optimization = time_msec() + DPCLS_OPTIMIZATION_INTERVAL; | |
4187 | hmap_init(&pmd->poll_list); | |
4188 | hmap_init(&pmd->tx_ports); | |
4189 | hmap_init(&pmd->tnl_port_cache); | |
4190 | hmap_init(&pmd->send_port_cache); | |
4191 | /* init the 'flow_cache' since there is no | |
4192 | * actual thread created for NON_PMD_CORE_ID. */ | |
4193 | if (core_id == NON_PMD_CORE_ID) { | |
4194 | emc_cache_init(&pmd->flow_cache); | |
4195 | } | |
4196 | cmap_insert(&dp->poll_threads, CONST_CAST(struct cmap_node *, &pmd->node), | |
4197 | hash_int(core_id, 0)); | |
4198 | } | |
4199 | ||
4200 | static void | |
4201 | dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd) | |
4202 | { | |
4203 | struct dpcls *cls; | |
4204 | ||
4205 | dp_netdev_pmd_flow_flush(pmd); | |
4206 | hmap_destroy(&pmd->send_port_cache); | |
4207 | hmap_destroy(&pmd->tnl_port_cache); | |
4208 | hmap_destroy(&pmd->tx_ports); | |
4209 | hmap_destroy(&pmd->poll_list); | |
4210 | /* All flows (including their dpcls_rules) have been deleted already */ | |
4211 | CMAP_FOR_EACH (cls, node, &pmd->classifiers) { | |
4212 | dpcls_destroy(cls); | |
4213 | ovsrcu_postpone(free, cls); | |
4214 | } | |
4215 | cmap_destroy(&pmd->classifiers); | |
4216 | cmap_destroy(&pmd->flow_table); | |
4217 | ovs_mutex_destroy(&pmd->flow_mutex); | |
4218 | latch_destroy(&pmd->exit_latch); | |
4219 | seq_destroy(pmd->reload_seq); | |
4220 | xpthread_cond_destroy(&pmd->cond); | |
4221 | ovs_mutex_destroy(&pmd->cond_mutex); | |
4222 | ovs_mutex_destroy(&pmd->port_mutex); | |
4223 | free(pmd); | |
4224 | } | |
4225 | ||
4226 | /* Stops the pmd thread, removes it from the 'dp->poll_threads', | |
4227 | * and unrefs the struct. */ | |
4228 | static void | |
4229 | dp_netdev_del_pmd(struct dp_netdev *dp, struct dp_netdev_pmd_thread *pmd) | |
4230 | { | |
4231 | /* NON_PMD_CORE_ID doesn't have a thread, so we don't have to synchronize, | |
4232 | * but extra cleanup is necessary */ | |
4233 | if (pmd->core_id == NON_PMD_CORE_ID) { | |
4234 | ovs_mutex_lock(&dp->non_pmd_mutex); | |
4235 | emc_cache_uninit(&pmd->flow_cache); | |
4236 | pmd_free_cached_ports(pmd); | |
4237 | ovs_mutex_unlock(&dp->non_pmd_mutex); | |
4238 | } else { | |
4239 | latch_set(&pmd->exit_latch); | |
4240 | dp_netdev_reload_pmd__(pmd); | |
4241 | xpthread_join(pmd->thread, NULL); | |
4242 | } | |
4243 | ||
4244 | dp_netdev_pmd_clear_ports(pmd); | |
4245 | ||
4246 | /* Purges the 'pmd''s flows after stopping the thread, but before | |
4247 | * destroying the flows, so that the flow stats can be collected. */ | |
4248 | if (dp->dp_purge_cb) { | |
4249 | dp->dp_purge_cb(dp->dp_purge_aux, pmd->core_id); | |
4250 | } | |
4251 | cmap_remove(&pmd->dp->poll_threads, &pmd->node, hash_int(pmd->core_id, 0)); | |
4252 | dp_netdev_pmd_unref(pmd); | |
4253 | } | |
4254 | ||
4255 | /* Destroys all pmd threads. If 'non_pmd' is true it also destroys the non pmd | |
4256 | * thread. */ | |
4257 | static void | |
4258 | dp_netdev_destroy_all_pmds(struct dp_netdev *dp, bool non_pmd) | |
4259 | { | |
4260 | struct dp_netdev_pmd_thread *pmd; | |
4261 | struct dp_netdev_pmd_thread **pmd_list; | |
4262 | size_t k = 0, n_pmds; | |
4263 | ||
4264 | n_pmds = cmap_count(&dp->poll_threads); | |
4265 | pmd_list = xcalloc(n_pmds, sizeof *pmd_list); | |
4266 | ||
4267 | CMAP_FOR_EACH (pmd, node, &dp->poll_threads) { | |
4268 | if (!non_pmd && pmd->core_id == NON_PMD_CORE_ID) { | |
4269 | continue; | |
4270 | } | |
4271 | /* We cannot call dp_netdev_del_pmd(), since it alters | |
4272 | * 'dp->poll_threads' (while we're iterating it) and it | |
4273 | * might quiesce. */ | |
4274 | ovs_assert(k < n_pmds); | |
4275 | pmd_list[k++] = pmd; | |
4276 | } | |
4277 | ||
4278 | for (size_t i = 0; i < k; i++) { | |
4279 | dp_netdev_del_pmd(dp, pmd_list[i]); | |
4280 | } | |
4281 | free(pmd_list); | |
4282 | } | |
4283 | ||
4284 | /* Deletes all rx queues from pmd->poll_list and all the ports from | |
4285 | * pmd->tx_ports. */ | |
4286 | static void | |
4287 | dp_netdev_pmd_clear_ports(struct dp_netdev_pmd_thread *pmd) | |
4288 | { | |
4289 | struct rxq_poll *poll; | |
4290 | struct tx_port *port; | |
4291 | ||
4292 | ovs_mutex_lock(&pmd->port_mutex); | |
4293 | HMAP_FOR_EACH_POP (poll, node, &pmd->poll_list) { | |
4294 | free(poll); | |
4295 | } | |
4296 | HMAP_FOR_EACH_POP (port, node, &pmd->tx_ports) { | |
4297 | free(port); | |
4298 | } | |
4299 | ovs_mutex_unlock(&pmd->port_mutex); | |
4300 | } | |
4301 | ||
4302 | /* Adds rx queue to poll_list of PMD thread, if it's not there already. */ | |
4303 | static void | |
4304 | dp_netdev_add_rxq_to_pmd(struct dp_netdev_pmd_thread *pmd, | |
4305 | struct dp_netdev_rxq *rxq) | |
4306 | OVS_REQUIRES(pmd->port_mutex) | |
4307 | { | |
4308 | int qid = netdev_rxq_get_queue_id(rxq->rx); | |
4309 | uint32_t hash = hash_2words(odp_to_u32(rxq->port->port_no), qid); | |
4310 | struct rxq_poll *poll; | |
4311 | ||
4312 | HMAP_FOR_EACH_WITH_HASH (poll, node, hash, &pmd->poll_list) { | |
4313 | if (poll->rxq == rxq) { | |
4314 | /* 'rxq' is already polled by this thread. Do nothing. */ | |
4315 | return; | |
4316 | } | |
4317 | } | |
4318 | ||
4319 | poll = xmalloc(sizeof *poll); | |
4320 | poll->rxq = rxq; | |
4321 | hmap_insert(&pmd->poll_list, &poll->node, hash); | |
4322 | ||
4323 | pmd->need_reload = true; | |
4324 | } | |
4325 | ||
4326 | /* Delete 'poll' from poll_list of PMD thread. */ | |
4327 | static void | |
4328 | dp_netdev_del_rxq_from_pmd(struct dp_netdev_pmd_thread *pmd, | |
4329 | struct rxq_poll *poll) | |
4330 | OVS_REQUIRES(pmd->port_mutex) | |
4331 | { | |
4332 | hmap_remove(&pmd->poll_list, &poll->node); | |
4333 | free(poll); | |
4334 | ||
4335 | pmd->need_reload = true; | |
4336 | } | |
4337 | ||
4338 | /* Add 'port' to the tx port cache of 'pmd', which must be reloaded for the | |
4339 | * changes to take effect. */ | |
4340 | static void | |
4341 | dp_netdev_add_port_tx_to_pmd(struct dp_netdev_pmd_thread *pmd, | |
4342 | struct dp_netdev_port *port) | |
4343 | OVS_REQUIRES(pmd->port_mutex) | |
4344 | { | |
4345 | struct tx_port *tx; | |
4346 | ||
4347 | tx = tx_port_lookup(&pmd->tx_ports, port->port_no); | |
4348 | if (tx) { | |
4349 | /* 'port' is already on this thread tx cache. Do nothing. */ | |
4350 | return; | |
4351 | } | |
4352 | ||
4353 | tx = xzalloc(sizeof *tx); | |
4354 | ||
4355 | tx->port = port; | |
4356 | tx->qid = -1; | |
4357 | ||
4358 | hmap_insert(&pmd->tx_ports, &tx->node, hash_port_no(tx->port->port_no)); | |
4359 | pmd->need_reload = true; | |
4360 | } | |
4361 | ||
4362 | /* Del 'tx' from the tx port cache of 'pmd', which must be reloaded for the | |
4363 | * changes to take effect. */ | |
4364 | static void | |
4365 | dp_netdev_del_port_tx_from_pmd(struct dp_netdev_pmd_thread *pmd, | |
4366 | struct tx_port *tx) | |
4367 | OVS_REQUIRES(pmd->port_mutex) | |
4368 | { | |
4369 | hmap_remove(&pmd->tx_ports, &tx->node); | |
4370 | free(tx); | |
4371 | pmd->need_reload = true; | |
4372 | } | |
4373 | \f | |
4374 | static char * | |
4375 | dpif_netdev_get_datapath_version(void) | |
4376 | { | |
4377 | return xstrdup("<built-in>"); | |
4378 | } | |
4379 | ||
4380 | static void | |
4381 | dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow, int cnt, int size, | |
4382 | uint16_t tcp_flags, long long now) | |
4383 | { | |
4384 | uint16_t flags; | |
4385 | ||
4386 | atomic_store_relaxed(&netdev_flow->stats.used, now); | |
4387 | non_atomic_ullong_add(&netdev_flow->stats.packet_count, cnt); | |
4388 | non_atomic_ullong_add(&netdev_flow->stats.byte_count, size); | |
4389 | atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags); | |
4390 | flags |= tcp_flags; | |
4391 | atomic_store_relaxed(&netdev_flow->stats.tcp_flags, flags); | |
4392 | } | |
4393 | ||
4394 | static void | |
4395 | dp_netdev_count_packet(struct dp_netdev_pmd_thread *pmd, | |
4396 | enum dp_stat_type type, int cnt) | |
4397 | { | |
4398 | non_atomic_ullong_add(&pmd->stats.n[type], cnt); | |
4399 | } | |
4400 | ||
4401 | static int | |
4402 | dp_netdev_upcall(struct dp_netdev_pmd_thread *pmd, struct dp_packet *packet_, | |
4403 | struct flow *flow, struct flow_wildcards *wc, ovs_u128 *ufid, | |
4404 | enum dpif_upcall_type type, const struct nlattr *userdata, | |
4405 | struct ofpbuf *actions, struct ofpbuf *put_actions) | |
4406 | { | |
4407 | struct dp_netdev *dp = pmd->dp; | |
4408 | ||
4409 | if (OVS_UNLIKELY(!dp->upcall_cb)) { | |
4410 | return ENODEV; | |
4411 | } | |
4412 | ||
4413 | if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl))) { | |
4414 | struct ds ds = DS_EMPTY_INITIALIZER; | |
4415 | char *packet_str; | |
4416 | struct ofpbuf key; | |
4417 | struct odp_flow_key_parms odp_parms = { | |
4418 | .flow = flow, | |
4419 | .mask = wc ? &wc->masks : NULL, | |
4420 | .support = dp_netdev_support, | |
4421 | }; | |
4422 | ||
4423 | ofpbuf_init(&key, 0); | |
4424 | odp_flow_key_from_flow(&odp_parms, &key); | |
4425 | packet_str = ofp_dp_packet_to_string(packet_); | |
4426 | ||
4427 | odp_flow_key_format(key.data, key.size, &ds); | |
4428 | ||
4429 | VLOG_DBG("%s: %s upcall:\n%s\n%s", dp->name, | |
4430 | dpif_upcall_type_to_string(type), ds_cstr(&ds), packet_str); | |
4431 | ||
4432 | ofpbuf_uninit(&key); | |
4433 | free(packet_str); | |
4434 | ||
4435 | ds_destroy(&ds); | |
4436 | } | |
4437 | ||
4438 | return dp->upcall_cb(packet_, flow, ufid, pmd->core_id, type, userdata, | |
4439 | actions, wc, put_actions, dp->upcall_aux); | |
4440 | } | |
4441 | ||
4442 | static inline uint32_t | |
4443 | dpif_netdev_packet_get_rss_hash(struct dp_packet *packet, | |
4444 | const struct miniflow *mf) | |
4445 | { | |
4446 | uint32_t hash, recirc_depth; | |
4447 | ||
4448 | if (OVS_LIKELY(dp_packet_rss_valid(packet))) { | |
4449 | hash = dp_packet_get_rss_hash(packet); | |
4450 | } else { | |
4451 | hash = miniflow_hash_5tuple(mf, 0); | |
4452 | dp_packet_set_rss_hash(packet, hash); | |
4453 | } | |
4454 | ||
4455 | /* The RSS hash must account for the recirculation depth to avoid | |
4456 | * collisions in the exact match cache */ | |
4457 | recirc_depth = *recirc_depth_get_unsafe(); | |
4458 | if (OVS_UNLIKELY(recirc_depth)) { | |
4459 | hash = hash_finish(hash, recirc_depth); | |
4460 | dp_packet_set_rss_hash(packet, hash); | |
4461 | } | |
4462 | return hash; | |
4463 | } | |
4464 | ||
4465 | struct packet_batch_per_flow { | |
4466 | unsigned int byte_count; | |
4467 | uint16_t tcp_flags; | |
4468 | struct dp_netdev_flow *flow; | |
4469 | ||
4470 | struct dp_packet_batch array; | |
4471 | }; | |
4472 | ||
4473 | static inline void | |
4474 | packet_batch_per_flow_update(struct packet_batch_per_flow *batch, | |
4475 | struct dp_packet *packet, | |
4476 | const struct miniflow *mf) | |
4477 | { | |
4478 | batch->byte_count += dp_packet_size(packet); | |
4479 | batch->tcp_flags |= miniflow_get_tcp_flags(mf); | |
4480 | batch->array.packets[batch->array.count++] = packet; | |
4481 | } | |
4482 | ||
4483 | static inline void | |
4484 | packet_batch_per_flow_init(struct packet_batch_per_flow *batch, | |
4485 | struct dp_netdev_flow *flow) | |
4486 | { | |
4487 | flow->batch = batch; | |
4488 | ||
4489 | batch->flow = flow; | |
4490 | dp_packet_batch_init(&batch->array); | |
4491 | batch->byte_count = 0; | |
4492 | batch->tcp_flags = 0; | |
4493 | } | |
4494 | ||
4495 | static inline void | |
4496 | packet_batch_per_flow_execute(struct packet_batch_per_flow *batch, | |
4497 | struct dp_netdev_pmd_thread *pmd, | |
4498 | long long now) | |
4499 | { | |
4500 | struct dp_netdev_actions *actions; | |
4501 | struct dp_netdev_flow *flow = batch->flow; | |
4502 | ||
4503 | dp_netdev_flow_used(flow, batch->array.count, batch->byte_count, | |
4504 | batch->tcp_flags, now); | |
4505 | ||
4506 | actions = dp_netdev_flow_get_actions(flow); | |
4507 | ||
4508 | dp_netdev_execute_actions(pmd, &batch->array, true, &flow->flow, | |
4509 | actions->actions, actions->size, now); | |
4510 | } | |
4511 | ||
4512 | static inline void | |
4513 | dp_netdev_queue_batches(struct dp_packet *pkt, | |
4514 | struct dp_netdev_flow *flow, const struct miniflow *mf, | |
4515 | struct packet_batch_per_flow *batches, | |
4516 | size_t *n_batches) | |
4517 | { | |
4518 | struct packet_batch_per_flow *batch = flow->batch; | |
4519 | ||
4520 | if (OVS_UNLIKELY(!batch)) { | |
4521 | batch = &batches[(*n_batches)++]; | |
4522 | packet_batch_per_flow_init(batch, flow); | |
4523 | } | |
4524 | ||
4525 | packet_batch_per_flow_update(batch, pkt, mf); | |
4526 | } | |
4527 | ||
4528 | /* Try to process all ('cnt') the 'packets' using only the exact match cache | |
4529 | * 'pmd->flow_cache'. If a flow is not found for a packet 'packets[i]', the | |
4530 | * miniflow is copied into 'keys' and the packet pointer is moved at the | |
4531 | * beginning of the 'packets' array. | |
4532 | * | |
4533 | * The function returns the number of packets that needs to be processed in the | |
4534 | * 'packets' array (they have been moved to the beginning of the vector). | |
4535 | * | |
4536 | * If 'md_is_valid' is false, the metadata in 'packets' is not valid and must | |
4537 | * be initialized by this function using 'port_no'. | |
4538 | */ | |
4539 | static inline size_t | |
4540 | emc_processing(struct dp_netdev_pmd_thread *pmd, | |
4541 | struct dp_packet_batch *packets_, | |
4542 | struct netdev_flow_key *keys, | |
4543 | struct packet_batch_per_flow batches[], size_t *n_batches, | |
4544 | bool md_is_valid, odp_port_t port_no) | |
4545 | { | |
4546 | struct emc_cache *flow_cache = &pmd->flow_cache; | |
4547 | struct netdev_flow_key *key = &keys[0]; | |
4548 | size_t n_missed = 0, n_dropped = 0; | |
4549 | struct dp_packet *packet; | |
4550 | const size_t size = dp_packet_batch_size(packets_); | |
4551 | uint32_t cur_min; | |
4552 | int i; | |
4553 | ||
4554 | atomic_read_relaxed(&pmd->dp->emc_insert_min, &cur_min); | |
4555 | ||
4556 | DP_PACKET_BATCH_REFILL_FOR_EACH (i, size, packet, packets_) { | |
4557 | struct dp_netdev_flow *flow; | |
4558 | ||
4559 | if (OVS_UNLIKELY(dp_packet_size(packet) < ETH_HEADER_LEN)) { | |
4560 | dp_packet_delete(packet); | |
4561 | n_dropped++; | |
4562 | continue; | |
4563 | } | |
4564 | ||
4565 | if (i != size - 1) { | |
4566 | struct dp_packet **packets = packets_->packets; | |
4567 | /* Prefetch next packet data and metadata. */ | |
4568 | OVS_PREFETCH(dp_packet_data(packets[i+1])); | |
4569 | pkt_metadata_prefetch_init(&packets[i+1]->md); | |
4570 | } | |
4571 | ||
4572 | if (!md_is_valid) { | |
4573 | pkt_metadata_init(&packet->md, port_no); | |
4574 | } | |
4575 | miniflow_extract(packet, &key->mf); | |
4576 | key->len = 0; /* Not computed yet. */ | |
4577 | key->hash = dpif_netdev_packet_get_rss_hash(packet, &key->mf); | |
4578 | ||
4579 | /* If EMC is disabled skip emc_lookup */ | |
4580 | flow = (cur_min == 0) ? NULL: emc_lookup(flow_cache, key); | |
4581 | if (OVS_LIKELY(flow)) { | |
4582 | dp_netdev_queue_batches(packet, flow, &key->mf, batches, | |
4583 | n_batches); | |
4584 | } else { | |
4585 | /* Exact match cache missed. Group missed packets together at | |
4586 | * the beginning of the 'packets' array. */ | |
4587 | dp_packet_batch_refill(packets_, packet, i); | |
4588 | /* 'key[n_missed]' contains the key of the current packet and it | |
4589 | * must be returned to the caller. The next key should be extracted | |
4590 | * to 'keys[n_missed + 1]'. */ | |
4591 | key = &keys[++n_missed]; | |
4592 | } | |
4593 | } | |
4594 | ||
4595 | dp_netdev_count_packet(pmd, DP_STAT_EXACT_HIT, | |
4596 | size - n_dropped - n_missed); | |
4597 | ||
4598 | return dp_packet_batch_size(packets_); | |
4599 | } | |
4600 | ||
4601 | static inline void | |
4602 | handle_packet_upcall(struct dp_netdev_pmd_thread *pmd, | |
4603 | struct dp_packet *packet, | |
4604 | const struct netdev_flow_key *key, | |
4605 | struct ofpbuf *actions, struct ofpbuf *put_actions, | |
4606 | int *lost_cnt, long long now) | |
4607 | { | |
4608 | struct ofpbuf *add_actions; | |
4609 | struct dp_packet_batch b; | |
4610 | struct match match; | |
4611 | ovs_u128 ufid; | |
4612 | int error; | |
4613 | ||
4614 | match.tun_md.valid = false; | |
4615 | miniflow_expand(&key->mf, &match.flow); | |
4616 | ||
4617 | ofpbuf_clear(actions); | |
4618 | ofpbuf_clear(put_actions); | |
4619 | ||
4620 | dpif_flow_hash(pmd->dp->dpif, &match.flow, sizeof match.flow, &ufid); | |
4621 | error = dp_netdev_upcall(pmd, packet, &match.flow, &match.wc, | |
4622 | &ufid, DPIF_UC_MISS, NULL, actions, | |
4623 | put_actions); | |
4624 | if (OVS_UNLIKELY(error && error != ENOSPC)) { | |
4625 | dp_packet_delete(packet); | |
4626 | (*lost_cnt)++; | |
4627 | return; | |
4628 | } | |
4629 | ||
4630 | /* The Netlink encoding of datapath flow keys cannot express | |
4631 | * wildcarding the presence of a VLAN tag. Instead, a missing VLAN | |
4632 | * tag is interpreted as exact match on the fact that there is no | |
4633 | * VLAN. Unless we refactor a lot of code that translates between | |
4634 | * Netlink and struct flow representations, we have to do the same | |
4635 | * here. */ | |
4636 | if (!match.wc.masks.vlans[0].tci) { | |
4637 | match.wc.masks.vlans[0].tci = htons(0xffff); | |
4638 | } | |
4639 | ||
4640 | /* We can't allow the packet batching in the next loop to execute | |
4641 | * the actions. Otherwise, if there are any slow path actions, | |
4642 | * we'll send the packet up twice. */ | |
4643 | dp_packet_batch_init_packet(&b, packet); | |
4644 | dp_netdev_execute_actions(pmd, &b, true, &match.flow, | |
4645 | actions->data, actions->size, now); | |
4646 | ||
4647 | add_actions = put_actions->size ? put_actions : actions; | |
4648 | if (OVS_LIKELY(error != ENOSPC)) { | |
4649 | struct dp_netdev_flow *netdev_flow; | |
4650 | ||
4651 | /* XXX: There's a race window where a flow covering this packet | |
4652 | * could have already been installed since we last did the flow | |
4653 | * lookup before upcall. This could be solved by moving the | |
4654 | * mutex lock outside the loop, but that's an awful long time | |
4655 | * to be locking everyone out of making flow installs. If we | |
4656 | * move to a per-core classifier, it would be reasonable. */ | |
4657 | ovs_mutex_lock(&pmd->flow_mutex); | |
4658 | netdev_flow = dp_netdev_pmd_lookup_flow(pmd, key, NULL); | |
4659 | if (OVS_LIKELY(!netdev_flow)) { | |
4660 | netdev_flow = dp_netdev_flow_add(pmd, &match, &ufid, | |
4661 | add_actions->data, | |
4662 | add_actions->size); | |
4663 | } | |
4664 | ovs_mutex_unlock(&pmd->flow_mutex); | |
4665 | emc_probabilistic_insert(pmd, key, netdev_flow); | |
4666 | } | |
4667 | } | |
4668 | ||
4669 | static inline void | |
4670 | fast_path_processing(struct dp_netdev_pmd_thread *pmd, | |
4671 | struct dp_packet_batch *packets_, | |
4672 | struct netdev_flow_key *keys, | |
4673 | struct packet_batch_per_flow batches[], size_t *n_batches, | |
4674 | odp_port_t in_port, | |
4675 | long long now) | |
4676 | { | |
4677 | int cnt = packets_->count; | |
4678 | #if !defined(__CHECKER__) && !defined(_WIN32) | |
4679 | const size_t PKT_ARRAY_SIZE = cnt; | |
4680 | #else | |
4681 | /* Sparse or MSVC doesn't like variable length array. */ | |
4682 | enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST }; | |
4683 | #endif | |
4684 | struct dp_packet **packets = packets_->packets; | |
4685 | struct dpcls *cls; | |
4686 | struct dpcls_rule *rules[PKT_ARRAY_SIZE]; | |
4687 | struct dp_netdev *dp = pmd->dp; | |
4688 | int miss_cnt = 0, lost_cnt = 0; | |
4689 | int lookup_cnt = 0, add_lookup_cnt; | |
4690 | bool any_miss; | |
4691 | size_t i; | |
4692 | ||
4693 | for (i = 0; i < cnt; i++) { | |
4694 | /* Key length is needed in all the cases, hash computed on demand. */ | |
4695 | keys[i].len = netdev_flow_key_size(miniflow_n_values(&keys[i].mf)); | |
4696 | } | |
4697 | /* Get the classifier for the in_port */ | |
4698 | cls = dp_netdev_pmd_lookup_dpcls(pmd, in_port); | |
4699 | if (OVS_LIKELY(cls)) { | |
4700 | any_miss = !dpcls_lookup(cls, keys, rules, cnt, &lookup_cnt); | |
4701 | } else { | |
4702 | any_miss = true; | |
4703 | memset(rules, 0, sizeof(rules)); | |
4704 | } | |
4705 | if (OVS_UNLIKELY(any_miss) && !fat_rwlock_tryrdlock(&dp->upcall_rwlock)) { | |
4706 | uint64_t actions_stub[512 / 8], slow_stub[512 / 8]; | |
4707 | struct ofpbuf actions, put_actions; | |
4708 | ||
4709 | ofpbuf_use_stub(&actions, actions_stub, sizeof actions_stub); | |
4710 | ofpbuf_use_stub(&put_actions, slow_stub, sizeof slow_stub); | |
4711 | ||
4712 | for (i = 0; i < cnt; i++) { | |
4713 | struct dp_netdev_flow *netdev_flow; | |
4714 | ||
4715 | if (OVS_LIKELY(rules[i])) { | |
4716 | continue; | |
4717 | } | |
4718 | ||
4719 | /* It's possible that an earlier slow path execution installed | |
4720 | * a rule covering this flow. In this case, it's a lot cheaper | |
4721 | * to catch it here than execute a miss. */ | |
4722 | netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i], | |
4723 | &add_lookup_cnt); | |
4724 | if (netdev_flow) { | |
4725 | lookup_cnt += add_lookup_cnt; | |
4726 | rules[i] = &netdev_flow->cr; | |
4727 | continue; | |
4728 | } | |
4729 | ||
4730 | miss_cnt++; | |
4731 | handle_packet_upcall(pmd, packets[i], &keys[i], &actions, | |
4732 | &put_actions, &lost_cnt, now); | |
4733 | } | |
4734 | ||
4735 | ofpbuf_uninit(&actions); | |
4736 | ofpbuf_uninit(&put_actions); | |
4737 | fat_rwlock_unlock(&dp->upcall_rwlock); | |
4738 | } else if (OVS_UNLIKELY(any_miss)) { | |
4739 | for (i = 0; i < cnt; i++) { | |
4740 | if (OVS_UNLIKELY(!rules[i])) { | |
4741 | dp_packet_delete(packets[i]); | |
4742 | lost_cnt++; | |
4743 | miss_cnt++; | |
4744 | } | |
4745 | } | |
4746 | } | |
4747 | ||
4748 | for (i = 0; i < cnt; i++) { | |
4749 | struct dp_packet *packet = packets[i]; | |
4750 | struct dp_netdev_flow *flow; | |
4751 | ||
4752 | if (OVS_UNLIKELY(!rules[i])) { | |
4753 | continue; | |
4754 | } | |
4755 | ||
4756 | flow = dp_netdev_flow_cast(rules[i]); | |
4757 | ||
4758 | emc_probabilistic_insert(pmd, &keys[i], flow); | |
4759 | dp_netdev_queue_batches(packet, flow, &keys[i].mf, batches, n_batches); | |
4760 | } | |
4761 | ||
4762 | dp_netdev_count_packet(pmd, DP_STAT_MASKED_HIT, cnt - miss_cnt); | |
4763 | dp_netdev_count_packet(pmd, DP_STAT_LOOKUP_HIT, lookup_cnt); | |
4764 | dp_netdev_count_packet(pmd, DP_STAT_MISS, miss_cnt); | |
4765 | dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt); | |
4766 | } | |
4767 | ||
4768 | /* Packets enter the datapath from a port (or from recirculation) here. | |
4769 | * | |
4770 | * For performance reasons a caller may choose not to initialize the metadata | |
4771 | * in 'packets': in this case 'mdinit' is false and this function needs to | |
4772 | * initialize it using 'port_no'. If the metadata in 'packets' is already | |
4773 | * valid, 'md_is_valid' must be true and 'port_no' will be ignored. */ | |
4774 | static void | |
4775 | dp_netdev_input__(struct dp_netdev_pmd_thread *pmd, | |
4776 | struct dp_packet_batch *packets, | |
4777 | bool md_is_valid, odp_port_t port_no) | |
4778 | { | |
4779 | int cnt = packets->count; | |
4780 | #if !defined(__CHECKER__) && !defined(_WIN32) | |
4781 | const size_t PKT_ARRAY_SIZE = cnt; | |
4782 | #else | |
4783 | /* Sparse or MSVC doesn't like variable length array. */ | |
4784 | enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST }; | |
4785 | #endif | |
4786 | OVS_ALIGNED_VAR(CACHE_LINE_SIZE) | |
4787 | struct netdev_flow_key keys[PKT_ARRAY_SIZE]; | |
4788 | struct packet_batch_per_flow batches[PKT_ARRAY_SIZE]; | |
4789 | long long now = time_msec(); | |
4790 | size_t n_batches; | |
4791 | odp_port_t in_port; | |
4792 | ||
4793 | n_batches = 0; | |
4794 | emc_processing(pmd, packets, keys, batches, &n_batches, | |
4795 | md_is_valid, port_no); | |
4796 | if (!dp_packet_batch_is_empty(packets)) { | |
4797 | /* Get ingress port from first packet's metadata. */ | |
4798 | in_port = packets->packets[0]->md.in_port.odp_port; | |
4799 | fast_path_processing(pmd, packets, keys, batches, &n_batches, | |
4800 | in_port, now); | |
4801 | } | |
4802 | ||
4803 | /* All the flow batches need to be reset before any call to | |
4804 | * packet_batch_per_flow_execute() as it could potentially trigger | |
4805 | * recirculation. When a packet matching flow ‘j’ happens to be | |
4806 | * recirculated, the nested call to dp_netdev_input__() could potentially | |
4807 | * classify the packet as matching another flow - say 'k'. It could happen | |
4808 | * that in the previous call to dp_netdev_input__() that same flow 'k' had | |
4809 | * already its own batches[k] still waiting to be served. So if its | |
4810 | * ‘batch’ member is not reset, the recirculated packet would be wrongly | |
4811 | * appended to batches[k] of the 1st call to dp_netdev_input__(). */ | |
4812 | size_t i; | |
4813 | for (i = 0; i < n_batches; i++) { | |
4814 | batches[i].flow->batch = NULL; | |
4815 | } | |
4816 | ||
4817 | for (i = 0; i < n_batches; i++) { | |
4818 | packet_batch_per_flow_execute(&batches[i], pmd, now); | |
4819 | } | |
4820 | } | |
4821 | ||
4822 | static void | |
4823 | dp_netdev_input(struct dp_netdev_pmd_thread *pmd, | |
4824 | struct dp_packet_batch *packets, | |
4825 | odp_port_t port_no) | |
4826 | { | |
4827 | dp_netdev_input__(pmd, packets, false, port_no); | |
4828 | } | |
4829 | ||
4830 | static void | |
4831 | dp_netdev_recirculate(struct dp_netdev_pmd_thread *pmd, | |
4832 | struct dp_packet_batch *packets) | |
4833 | { | |
4834 | dp_netdev_input__(pmd, packets, true, 0); | |
4835 | } | |
4836 | ||
4837 | struct dp_netdev_execute_aux { | |
4838 | struct dp_netdev_pmd_thread *pmd; | |
4839 | long long now; | |
4840 | const struct flow *flow; | |
4841 | }; | |
4842 | ||
4843 | static void | |
4844 | dpif_netdev_register_dp_purge_cb(struct dpif *dpif, dp_purge_callback *cb, | |
4845 | void *aux) | |
4846 | { | |
4847 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
4848 | dp->dp_purge_aux = aux; | |
4849 | dp->dp_purge_cb = cb; | |
4850 | } | |
4851 | ||
4852 | static void | |
4853 | dpif_netdev_register_upcall_cb(struct dpif *dpif, upcall_callback *cb, | |
4854 | void *aux) | |
4855 | { | |
4856 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
4857 | dp->upcall_aux = aux; | |
4858 | dp->upcall_cb = cb; | |
4859 | } | |
4860 | ||
4861 | static void | |
4862 | dpif_netdev_xps_revalidate_pmd(const struct dp_netdev_pmd_thread *pmd, | |
4863 | long long now, bool purge) | |
4864 | { | |
4865 | struct tx_port *tx; | |
4866 | struct dp_netdev_port *port; | |
4867 | long long interval; | |
4868 | ||
4869 | HMAP_FOR_EACH (tx, node, &pmd->send_port_cache) { | |
4870 | if (!tx->port->dynamic_txqs) { | |
4871 | continue; | |
4872 | } | |
4873 | interval = now - tx->last_used; | |
4874 | if (tx->qid >= 0 && (purge || interval >= XPS_TIMEOUT_MS)) { | |
4875 | port = tx->port; | |
4876 | ovs_mutex_lock(&port->txq_used_mutex); | |
4877 | port->txq_used[tx->qid]--; | |
4878 | ovs_mutex_unlock(&port->txq_used_mutex); | |
4879 | tx->qid = -1; | |
4880 | } | |
4881 | } | |
4882 | } | |
4883 | ||
4884 | static int | |
4885 | dpif_netdev_xps_get_tx_qid(const struct dp_netdev_pmd_thread *pmd, | |
4886 | struct tx_port *tx, long long now) | |
4887 | { | |
4888 | struct dp_netdev_port *port; | |
4889 | long long interval; | |
4890 | int i, min_cnt, min_qid; | |
4891 | ||
4892 | if (OVS_UNLIKELY(!now)) { | |
4893 | now = time_msec(); | |
4894 | } | |
4895 | ||
4896 | interval = now - tx->last_used; | |
4897 | tx->last_used = now; | |
4898 | ||
4899 | if (OVS_LIKELY(tx->qid >= 0 && interval < XPS_TIMEOUT_MS)) { | |
4900 | return tx->qid; | |
4901 | } | |
4902 | ||
4903 | port = tx->port; | |
4904 | ||
4905 | ovs_mutex_lock(&port->txq_used_mutex); | |
4906 | if (tx->qid >= 0) { | |
4907 | port->txq_used[tx->qid]--; | |
4908 | tx->qid = -1; | |
4909 | } | |
4910 | ||
4911 | min_cnt = -1; | |
4912 | min_qid = 0; | |
4913 | for (i = 0; i < netdev_n_txq(port->netdev); i++) { | |
4914 | if (port->txq_used[i] < min_cnt || min_cnt == -1) { | |
4915 | min_cnt = port->txq_used[i]; | |
4916 | min_qid = i; | |
4917 | } | |
4918 | } | |
4919 | ||
4920 | port->txq_used[min_qid]++; | |
4921 | tx->qid = min_qid; | |
4922 | ||
4923 | ovs_mutex_unlock(&port->txq_used_mutex); | |
4924 | ||
4925 | dpif_netdev_xps_revalidate_pmd(pmd, now, false); | |
4926 | ||
4927 | VLOG_DBG("Core %d: New TX queue ID %d for port \'%s\'.", | |
4928 | pmd->core_id, tx->qid, netdev_get_name(tx->port->netdev)); | |
4929 | return min_qid; | |
4930 | } | |
4931 | ||
4932 | static struct tx_port * | |
4933 | pmd_tnl_port_cache_lookup(const struct dp_netdev_pmd_thread *pmd, | |
4934 | odp_port_t port_no) | |
4935 | { | |
4936 | return tx_port_lookup(&pmd->tnl_port_cache, port_no); | |
4937 | } | |
4938 | ||
4939 | static struct tx_port * | |
4940 | pmd_send_port_cache_lookup(const struct dp_netdev_pmd_thread *pmd, | |
4941 | odp_port_t port_no) | |
4942 | { | |
4943 | return tx_port_lookup(&pmd->send_port_cache, port_no); | |
4944 | } | |
4945 | ||
4946 | static int | |
4947 | push_tnl_action(const struct dp_netdev_pmd_thread *pmd, | |
4948 | const struct nlattr *attr, | |
4949 | struct dp_packet_batch *batch) | |
4950 | { | |
4951 | struct tx_port *tun_port; | |
4952 | const struct ovs_action_push_tnl *data; | |
4953 | int err; | |
4954 | ||
4955 | data = nl_attr_get(attr); | |
4956 | ||
4957 | tun_port = pmd_tnl_port_cache_lookup(pmd, u32_to_odp(data->tnl_port)); | |
4958 | if (!tun_port) { | |
4959 | err = -EINVAL; | |
4960 | goto error; | |
4961 | } | |
4962 | err = netdev_push_header(tun_port->port->netdev, batch, data); | |
4963 | if (!err) { | |
4964 | return 0; | |
4965 | } | |
4966 | error: | |
4967 | dp_packet_delete_batch(batch, true); | |
4968 | return err; | |
4969 | } | |
4970 | ||
4971 | static void | |
4972 | dp_execute_userspace_action(struct dp_netdev_pmd_thread *pmd, | |
4973 | struct dp_packet *packet, bool may_steal, | |
4974 | struct flow *flow, ovs_u128 *ufid, | |
4975 | struct ofpbuf *actions, | |
4976 | const struct nlattr *userdata, long long now) | |
4977 | { | |
4978 | struct dp_packet_batch b; | |
4979 | int error; | |
4980 | ||
4981 | ofpbuf_clear(actions); | |
4982 | ||
4983 | error = dp_netdev_upcall(pmd, packet, flow, NULL, ufid, | |
4984 | DPIF_UC_ACTION, userdata, actions, | |
4985 | NULL); | |
4986 | if (!error || error == ENOSPC) { | |
4987 | dp_packet_batch_init_packet(&b, packet); | |
4988 | dp_netdev_execute_actions(pmd, &b, may_steal, flow, | |
4989 | actions->data, actions->size, now); | |
4990 | } else if (may_steal) { | |
4991 | dp_packet_delete(packet); | |
4992 | } | |
4993 | } | |
4994 | ||
4995 | static void | |
4996 | dp_execute_cb(void *aux_, struct dp_packet_batch *packets_, | |
4997 | const struct nlattr *a, bool may_steal) | |
4998 | OVS_NO_THREAD_SAFETY_ANALYSIS | |
4999 | { | |
5000 | struct dp_netdev_execute_aux *aux = aux_; | |
5001 | uint32_t *depth = recirc_depth_get(); | |
5002 | struct dp_netdev_pmd_thread *pmd = aux->pmd; | |
5003 | struct dp_netdev *dp = pmd->dp; | |
5004 | int type = nl_attr_type(a); | |
5005 | long long now = aux->now; | |
5006 | struct tx_port *p; | |
5007 | ||
5008 | switch ((enum ovs_action_attr)type) { | |
5009 | case OVS_ACTION_ATTR_OUTPUT: | |
5010 | p = pmd_send_port_cache_lookup(pmd, nl_attr_get_odp_port(a)); | |
5011 | if (OVS_LIKELY(p)) { | |
5012 | int tx_qid; | |
5013 | bool dynamic_txqs; | |
5014 | ||
5015 | dynamic_txqs = p->port->dynamic_txqs; | |
5016 | if (dynamic_txqs) { | |
5017 | tx_qid = dpif_netdev_xps_get_tx_qid(pmd, p, now); | |
5018 | } else { | |
5019 | tx_qid = pmd->static_tx_qid; | |
5020 | } | |
5021 | ||
5022 | netdev_send(p->port->netdev, tx_qid, packets_, may_steal, | |
5023 | dynamic_txqs); | |
5024 | return; | |
5025 | } | |
5026 | break; | |
5027 | ||
5028 | case OVS_ACTION_ATTR_TUNNEL_PUSH: | |
5029 | if (*depth < MAX_RECIRC_DEPTH) { | |
5030 | struct dp_packet_batch tnl_pkt; | |
5031 | struct dp_packet_batch *orig_packets_ = packets_; | |
5032 | int err; | |
5033 | ||
5034 | if (!may_steal) { | |
5035 | dp_packet_batch_clone(&tnl_pkt, packets_); | |
5036 | packets_ = &tnl_pkt; | |
5037 | dp_packet_batch_reset_cutlen(orig_packets_); | |
5038 | } | |
5039 | ||
5040 | dp_packet_batch_apply_cutlen(packets_); | |
5041 | ||
5042 | err = push_tnl_action(pmd, a, packets_); | |
5043 | if (!err) { | |
5044 | (*depth)++; | |
5045 | dp_netdev_recirculate(pmd, packets_); | |
5046 | (*depth)--; | |
5047 | } | |
5048 | return; | |
5049 | } | |
5050 | break; | |
5051 | ||
5052 | case OVS_ACTION_ATTR_TUNNEL_POP: | |
5053 | if (*depth < MAX_RECIRC_DEPTH) { | |
5054 | struct dp_packet_batch *orig_packets_ = packets_; | |
5055 | odp_port_t portno = nl_attr_get_odp_port(a); | |
5056 | ||
5057 | p = pmd_tnl_port_cache_lookup(pmd, portno); | |
5058 | if (p) { | |
5059 | struct dp_packet_batch tnl_pkt; | |
5060 | ||
5061 | if (!may_steal) { | |
5062 | dp_packet_batch_clone(&tnl_pkt, packets_); | |
5063 | packets_ = &tnl_pkt; | |
5064 | dp_packet_batch_reset_cutlen(orig_packets_); | |
5065 | } | |
5066 | ||
5067 | dp_packet_batch_apply_cutlen(packets_); | |
5068 | ||
5069 | netdev_pop_header(p->port->netdev, packets_); | |
5070 | if (dp_packet_batch_is_empty(packets_)) { | |
5071 | return; | |
5072 | } | |
5073 | ||
5074 | struct dp_packet *packet; | |
5075 | DP_PACKET_BATCH_FOR_EACH (packet, packets_) { | |
5076 | packet->md.in_port.odp_port = portno; | |
5077 | } | |
5078 | ||
5079 | (*depth)++; | |
5080 | dp_netdev_recirculate(pmd, packets_); | |
5081 | (*depth)--; | |
5082 | return; | |
5083 | } | |
5084 | } | |
5085 | break; | |
5086 | ||
5087 | case OVS_ACTION_ATTR_USERSPACE: | |
5088 | if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) { | |
5089 | struct dp_packet_batch *orig_packets_ = packets_; | |
5090 | const struct nlattr *userdata; | |
5091 | struct dp_packet_batch usr_pkt; | |
5092 | struct ofpbuf actions; | |
5093 | struct flow flow; | |
5094 | ovs_u128 ufid; | |
5095 | bool clone = false; | |
5096 | ||
5097 | userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA); | |
5098 | ofpbuf_init(&actions, 0); | |
5099 | ||
5100 | if (packets_->trunc) { | |
5101 | if (!may_steal) { | |
5102 | dp_packet_batch_clone(&usr_pkt, packets_); | |
5103 | packets_ = &usr_pkt; | |
5104 | clone = true; | |
5105 | dp_packet_batch_reset_cutlen(orig_packets_); | |
5106 | } | |
5107 | ||
5108 | dp_packet_batch_apply_cutlen(packets_); | |
5109 | } | |
5110 | ||
5111 | struct dp_packet *packet; | |
5112 | DP_PACKET_BATCH_FOR_EACH (packet, packets_) { | |
5113 | flow_extract(packet, &flow); | |
5114 | dpif_flow_hash(dp->dpif, &flow, sizeof flow, &ufid); | |
5115 | dp_execute_userspace_action(pmd, packet, may_steal, &flow, | |
5116 | &ufid, &actions, userdata, now); | |
5117 | } | |
5118 | ||
5119 | if (clone) { | |
5120 | dp_packet_delete_batch(packets_, true); | |
5121 | } | |
5122 | ||
5123 | ofpbuf_uninit(&actions); | |
5124 | fat_rwlock_unlock(&dp->upcall_rwlock); | |
5125 | ||
5126 | return; | |
5127 | } | |
5128 | break; | |
5129 | ||
5130 | case OVS_ACTION_ATTR_RECIRC: | |
5131 | if (*depth < MAX_RECIRC_DEPTH) { | |
5132 | struct dp_packet_batch recirc_pkts; | |
5133 | ||
5134 | if (!may_steal) { | |
5135 | dp_packet_batch_clone(&recirc_pkts, packets_); | |
5136 | packets_ = &recirc_pkts; | |
5137 | } | |
5138 | ||
5139 | struct dp_packet *packet; | |
5140 | DP_PACKET_BATCH_FOR_EACH (packet, packets_) { | |
5141 | packet->md.recirc_id = nl_attr_get_u32(a); | |
5142 | } | |
5143 | ||
5144 | (*depth)++; | |
5145 | dp_netdev_recirculate(pmd, packets_); | |
5146 | (*depth)--; | |
5147 | ||
5148 | return; | |
5149 | } | |
5150 | ||
5151 | VLOG_WARN("Packet dropped. Max recirculation depth exceeded."); | |
5152 | break; | |
5153 | ||
5154 | case OVS_ACTION_ATTR_CT: { | |
5155 | const struct nlattr *b; | |
5156 | bool force = false; | |
5157 | bool commit = false; | |
5158 | unsigned int left; | |
5159 | uint16_t zone = 0; | |
5160 | const char *helper = NULL; | |
5161 | const uint32_t *setmark = NULL; | |
5162 | const struct ovs_key_ct_labels *setlabel = NULL; | |
5163 | ||
5164 | NL_ATTR_FOR_EACH_UNSAFE (b, left, nl_attr_get(a), | |
5165 | nl_attr_get_size(a)) { | |
5166 | enum ovs_ct_attr sub_type = nl_attr_type(b); | |
5167 | ||
5168 | switch(sub_type) { | |
5169 | case OVS_CT_ATTR_FORCE_COMMIT: | |
5170 | force = true; | |
5171 | /* fall through. */ | |
5172 | case OVS_CT_ATTR_COMMIT: | |
5173 | commit = true; | |
5174 | break; | |
5175 | case OVS_CT_ATTR_ZONE: | |
5176 | zone = nl_attr_get_u16(b); | |
5177 | break; | |
5178 | case OVS_CT_ATTR_HELPER: | |
5179 | helper = nl_attr_get_string(b); | |
5180 | break; | |
5181 | case OVS_CT_ATTR_MARK: | |
5182 | setmark = nl_attr_get(b); | |
5183 | break; | |
5184 | case OVS_CT_ATTR_LABELS: | |
5185 | setlabel = nl_attr_get(b); | |
5186 | break; | |
5187 | case OVS_CT_ATTR_EVENTMASK: | |
5188 | /* Silently ignored, as userspace datapath does not generate | |
5189 | * netlink events. */ | |
5190 | break; | |
5191 | case OVS_CT_ATTR_NAT: | |
5192 | case OVS_CT_ATTR_UNSPEC: | |
5193 | case __OVS_CT_ATTR_MAX: | |
5194 | OVS_NOT_REACHED(); | |
5195 | } | |
5196 | } | |
5197 | ||
5198 | conntrack_execute(&dp->conntrack, packets_, aux->flow->dl_type, force, | |
5199 | commit, zone, setmark, setlabel, helper); | |
5200 | break; | |
5201 | } | |
5202 | ||
5203 | case OVS_ACTION_ATTR_METER: | |
5204 | dp_netdev_run_meter(pmd->dp, packets_, nl_attr_get_u32(a), | |
5205 | time_msec()); | |
5206 | break; | |
5207 | ||
5208 | case OVS_ACTION_ATTR_PUSH_VLAN: | |
5209 | case OVS_ACTION_ATTR_POP_VLAN: | |
5210 | case OVS_ACTION_ATTR_PUSH_MPLS: | |
5211 | case OVS_ACTION_ATTR_POP_MPLS: | |
5212 | case OVS_ACTION_ATTR_SET: | |
5213 | case OVS_ACTION_ATTR_SET_MASKED: | |
5214 | case OVS_ACTION_ATTR_SAMPLE: | |
5215 | case OVS_ACTION_ATTR_HASH: | |
5216 | case OVS_ACTION_ATTR_UNSPEC: | |
5217 | case OVS_ACTION_ATTR_TRUNC: | |
5218 | case OVS_ACTION_ATTR_PUSH_ETH: | |
5219 | case OVS_ACTION_ATTR_POP_ETH: | |
5220 | case OVS_ACTION_ATTR_CLONE: | |
5221 | case __OVS_ACTION_ATTR_MAX: | |
5222 | OVS_NOT_REACHED(); | |
5223 | } | |
5224 | ||
5225 | dp_packet_delete_batch(packets_, may_steal); | |
5226 | } | |
5227 | ||
5228 | static void | |
5229 | dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd, | |
5230 | struct dp_packet_batch *packets, | |
5231 | bool may_steal, const struct flow *flow, | |
5232 | const struct nlattr *actions, size_t actions_len, | |
5233 | long long now) | |
5234 | { | |
5235 | struct dp_netdev_execute_aux aux = { pmd, now, flow }; | |
5236 | ||
5237 | odp_execute_actions(&aux, packets, may_steal, actions, | |
5238 | actions_len, dp_execute_cb); | |
5239 | } | |
5240 | ||
5241 | struct dp_netdev_ct_dump { | |
5242 | struct ct_dpif_dump_state up; | |
5243 | struct conntrack_dump dump; | |
5244 | struct conntrack *ct; | |
5245 | struct dp_netdev *dp; | |
5246 | }; | |
5247 | ||
5248 | static int | |
5249 | dpif_netdev_ct_dump_start(struct dpif *dpif, struct ct_dpif_dump_state **dump_, | |
5250 | const uint16_t *pzone) | |
5251 | { | |
5252 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
5253 | struct dp_netdev_ct_dump *dump; | |
5254 | ||
5255 | dump = xzalloc(sizeof *dump); | |
5256 | dump->dp = dp; | |
5257 | dump->ct = &dp->conntrack; | |
5258 | ||
5259 | conntrack_dump_start(&dp->conntrack, &dump->dump, pzone); | |
5260 | ||
5261 | *dump_ = &dump->up; | |
5262 | ||
5263 | return 0; | |
5264 | } | |
5265 | ||
5266 | static int | |
5267 | dpif_netdev_ct_dump_next(struct dpif *dpif OVS_UNUSED, | |
5268 | struct ct_dpif_dump_state *dump_, | |
5269 | struct ct_dpif_entry *entry) | |
5270 | { | |
5271 | struct dp_netdev_ct_dump *dump; | |
5272 | ||
5273 | INIT_CONTAINER(dump, dump_, up); | |
5274 | ||
5275 | return conntrack_dump_next(&dump->dump, entry); | |
5276 | } | |
5277 | ||
5278 | static int | |
5279 | dpif_netdev_ct_dump_done(struct dpif *dpif OVS_UNUSED, | |
5280 | struct ct_dpif_dump_state *dump_) | |
5281 | { | |
5282 | struct dp_netdev_ct_dump *dump; | |
5283 | int err; | |
5284 | ||
5285 | INIT_CONTAINER(dump, dump_, up); | |
5286 | ||
5287 | err = conntrack_dump_done(&dump->dump); | |
5288 | ||
5289 | free(dump); | |
5290 | ||
5291 | return err; | |
5292 | } | |
5293 | ||
5294 | static int | |
5295 | dpif_netdev_ct_flush(struct dpif *dpif, const uint16_t *zone) | |
5296 | { | |
5297 | struct dp_netdev *dp = get_dp_netdev(dpif); | |
5298 | ||
5299 | return conntrack_flush(&dp->conntrack, zone); | |
5300 | } | |
5301 | ||
5302 | const struct dpif_class dpif_netdev_class = { | |
5303 | "netdev", | |
5304 | dpif_netdev_init, | |
5305 | dpif_netdev_enumerate, | |
5306 | dpif_netdev_port_open_type, | |
5307 | dpif_netdev_open, | |
5308 | dpif_netdev_close, | |
5309 | dpif_netdev_destroy, | |
5310 | dpif_netdev_run, | |
5311 | dpif_netdev_wait, | |
5312 | dpif_netdev_get_stats, | |
5313 | dpif_netdev_port_add, | |
5314 | dpif_netdev_port_del, | |
5315 | dpif_netdev_port_set_config, | |
5316 | dpif_netdev_port_query_by_number, | |
5317 | dpif_netdev_port_query_by_name, | |
5318 | NULL, /* port_get_pid */ | |
5319 | dpif_netdev_port_dump_start, | |
5320 | dpif_netdev_port_dump_next, | |
5321 | dpif_netdev_port_dump_done, | |
5322 | dpif_netdev_port_poll, | |
5323 | dpif_netdev_port_poll_wait, | |
5324 | dpif_netdev_flow_flush, | |
5325 | dpif_netdev_flow_dump_create, | |
5326 | dpif_netdev_flow_dump_destroy, | |
5327 | dpif_netdev_flow_dump_thread_create, | |
5328 | dpif_netdev_flow_dump_thread_destroy, | |
5329 | dpif_netdev_flow_dump_next, | |
5330 | dpif_netdev_operate, | |
5331 | NULL, /* recv_set */ | |
5332 | NULL, /* handlers_set */ | |
5333 | dpif_netdev_set_config, | |
5334 | dpif_netdev_queue_to_priority, | |
5335 | NULL, /* recv */ | |
5336 | NULL, /* recv_wait */ | |
5337 | NULL, /* recv_purge */ | |
5338 | dpif_netdev_register_dp_purge_cb, | |
5339 | dpif_netdev_register_upcall_cb, | |
5340 | dpif_netdev_enable_upcall, | |
5341 | dpif_netdev_disable_upcall, | |
5342 | dpif_netdev_get_datapath_version, | |
5343 | dpif_netdev_ct_dump_start, | |
5344 | dpif_netdev_ct_dump_next, | |
5345 | dpif_netdev_ct_dump_done, | |
5346 | dpif_netdev_ct_flush, | |
5347 | dpif_netdev_meter_get_features, | |
5348 | dpif_netdev_meter_set, | |
5349 | dpif_netdev_meter_get, | |
5350 | dpif_netdev_meter_del, | |
5351 | }; | |
5352 | ||
5353 | static void | |
5354 | dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED, | |
5355 | const char *argv[], void *aux OVS_UNUSED) | |
5356 | { | |
5357 | struct dp_netdev_port *port; | |
5358 | struct dp_netdev *dp; | |
5359 | odp_port_t port_no; | |
5360 | ||
5361 | ovs_mutex_lock(&dp_netdev_mutex); | |
5362 | dp = shash_find_data(&dp_netdevs, argv[1]); | |
5363 | if (!dp || !dpif_netdev_class_is_dummy(dp->class)) { | |
5364 | ovs_mutex_unlock(&dp_netdev_mutex); | |
5365 | unixctl_command_reply_error(conn, "unknown datapath or not a dummy"); | |
5366 | return; | |
5367 | } | |
5368 | ovs_refcount_ref(&dp->ref_cnt); | |
5369 | ovs_mutex_unlock(&dp_netdev_mutex); | |
5370 | ||
5371 | ovs_mutex_lock(&dp->port_mutex); | |
5372 | if (get_port_by_name(dp, argv[2], &port)) { | |
5373 | unixctl_command_reply_error(conn, "unknown port"); | |
5374 | goto exit; | |
5375 | } | |
5376 | ||
5377 | port_no = u32_to_odp(atoi(argv[3])); | |
5378 | if (!port_no || port_no == ODPP_NONE) { | |
5379 | unixctl_command_reply_error(conn, "bad port number"); | |
5380 | goto exit; | |
5381 | } | |
5382 | if (dp_netdev_lookup_port(dp, port_no)) { | |
5383 | unixctl_command_reply_error(conn, "port number already in use"); | |
5384 | goto exit; | |
5385 | } | |
5386 | ||
5387 | /* Remove port. */ | |
5388 | hmap_remove(&dp->ports, &port->node); | |
5389 | reconfigure_datapath(dp); | |
5390 | ||
5391 | /* Reinsert with new port number. */ | |
5392 | port->port_no = port_no; | |
5393 | hmap_insert(&dp->ports, &port->node, hash_port_no(port_no)); | |
5394 | reconfigure_datapath(dp); | |
5395 | ||
5396 | seq_change(dp->port_seq); | |
5397 | unixctl_command_reply(conn, NULL); | |
5398 | ||
5399 | exit: | |
5400 | ovs_mutex_unlock(&dp->port_mutex); | |
5401 | dp_netdev_unref(dp); | |
5402 | } | |
5403 | ||
5404 | static void | |
5405 | dpif_dummy_register__(const char *type) | |
5406 | { | |
5407 | struct dpif_class *class; | |
5408 | ||
5409 | class = xmalloc(sizeof *class); | |
5410 | *class = dpif_netdev_class; | |
5411 | class->type = xstrdup(type); | |
5412 | dp_register_provider(class); | |
5413 | } | |
5414 | ||
5415 | static void | |
5416 | dpif_dummy_override(const char *type) | |
5417 | { | |
5418 | int error; | |
5419 | ||
5420 | /* | |
5421 | * Ignore EAFNOSUPPORT to allow --enable-dummy=system with | |
5422 | * a userland-only build. It's useful for testsuite. | |
5423 | */ | |
5424 | error = dp_unregister_provider(type); | |
5425 | if (error == 0 || error == EAFNOSUPPORT) { | |
5426 | dpif_dummy_register__(type); | |
5427 | } | |
5428 | } | |
5429 | ||
5430 | void | |
5431 | dpif_dummy_register(enum dummy_level level) | |
5432 | { | |
5433 | if (level == DUMMY_OVERRIDE_ALL) { | |
5434 | struct sset types; | |
5435 | const char *type; | |
5436 | ||
5437 | sset_init(&types); | |
5438 | dp_enumerate_types(&types); | |
5439 | SSET_FOR_EACH (type, &types) { | |
5440 | dpif_dummy_override(type); | |
5441 | } | |
5442 | sset_destroy(&types); | |
5443 | } else if (level == DUMMY_OVERRIDE_SYSTEM) { | |
5444 | dpif_dummy_override("system"); | |
5445 | } | |
5446 | ||
5447 | dpif_dummy_register__("dummy"); | |
5448 | ||
5449 | unixctl_command_register("dpif-dummy/change-port-number", | |
5450 | "dp port new-number", | |
5451 | 3, 3, dpif_dummy_change_port_number, NULL); | |
5452 | } | |
5453 | \f | |
5454 | /* Datapath Classifier. */ | |
5455 | ||
5456 | /* A set of rules that all have the same fields wildcarded. */ | |
5457 | struct dpcls_subtable { | |
5458 | /* The fields are only used by writers. */ | |
5459 | struct cmap_node cmap_node OVS_GUARDED; /* Within dpcls 'subtables_map'. */ | |
5460 | ||
5461 | /* These fields are accessed by readers. */ | |
5462 | struct cmap rules; /* Contains "struct dpcls_rule"s. */ | |
5463 | uint32_t hit_cnt; /* Number of match hits in subtable in current | |
5464 | optimization interval. */ | |
5465 | struct netdev_flow_key mask; /* Wildcards for fields (const). */ | |
5466 | /* 'mask' must be the last field, additional space is allocated here. */ | |
5467 | }; | |
5468 | ||
5469 | /* Initializes 'cls' as a classifier that initially contains no classification | |
5470 | * rules. */ | |
5471 | static void | |
5472 | dpcls_init(struct dpcls *cls) | |
5473 | { | |
5474 | cmap_init(&cls->subtables_map); | |
5475 | pvector_init(&cls->subtables); | |
5476 | } | |
5477 | ||
5478 | static void | |
5479 | dpcls_destroy_subtable(struct dpcls *cls, struct dpcls_subtable *subtable) | |
5480 | { | |
5481 | VLOG_DBG("Destroying subtable %p for in_port %d", subtable, cls->in_port); | |
5482 | pvector_remove(&cls->subtables, subtable); | |
5483 | cmap_remove(&cls->subtables_map, &subtable->cmap_node, | |
5484 | subtable->mask.hash); | |
5485 | cmap_destroy(&subtable->rules); | |
5486 | ovsrcu_postpone(free, subtable); | |
5487 | } | |
5488 | ||
5489 | /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the | |
5490 | * caller's responsibility. | |
5491 | * May only be called after all the readers have been terminated. */ | |
5492 | static void | |
5493 | dpcls_destroy(struct dpcls *cls) | |
5494 | { | |
5495 | if (cls) { | |
5496 | struct dpcls_subtable *subtable; | |
5497 | ||
5498 | CMAP_FOR_EACH (subtable, cmap_node, &cls->subtables_map) { | |
5499 | ovs_assert(cmap_count(&subtable->rules) == 0); | |
5500 | dpcls_destroy_subtable(cls, subtable); | |
5501 | } | |
5502 | cmap_destroy(&cls->subtables_map); | |
5503 | pvector_destroy(&cls->subtables); | |
5504 | } | |
5505 | } | |
5506 | ||
5507 | static struct dpcls_subtable * | |
5508 | dpcls_create_subtable(struct dpcls *cls, const struct netdev_flow_key *mask) | |
5509 | { | |
5510 | struct dpcls_subtable *subtable; | |
5511 | ||
5512 | /* Need to add one. */ | |
5513 | subtable = xmalloc(sizeof *subtable | |
5514 | - sizeof subtable->mask.mf + mask->len); | |
5515 | cmap_init(&subtable->rules); | |
5516 | subtable->hit_cnt = 0; | |
5517 | netdev_flow_key_clone(&subtable->mask, mask); | |
5518 | cmap_insert(&cls->subtables_map, &subtable->cmap_node, mask->hash); | |
5519 | /* Add the new subtable at the end of the pvector (with no hits yet) */ | |
5520 | pvector_insert(&cls->subtables, subtable, 0); | |
5521 | VLOG_DBG("Creating %"PRIuSIZE". subtable %p for in_port %d", | |
5522 | cmap_count(&cls->subtables_map), subtable, cls->in_port); | |
5523 | pvector_publish(&cls->subtables); | |
5524 | ||
5525 | return subtable; | |
5526 | } | |
5527 | ||
5528 | static inline struct dpcls_subtable * | |
5529 | dpcls_find_subtable(struct dpcls *cls, const struct netdev_flow_key *mask) | |
5530 | { | |
5531 | struct dpcls_subtable *subtable; | |
5532 | ||
5533 | CMAP_FOR_EACH_WITH_HASH (subtable, cmap_node, mask->hash, | |
5534 | &cls->subtables_map) { | |
5535 | if (netdev_flow_key_equal(&subtable->mask, mask)) { | |
5536 | return subtable; | |
5537 | } | |
5538 | } | |
5539 | return dpcls_create_subtable(cls, mask); | |
5540 | } | |
5541 | ||
5542 | ||
5543 | /* Periodically sort the dpcls subtable vectors according to hit counts */ | |
5544 | static void | |
5545 | dpcls_sort_subtable_vector(struct dpcls *cls) | |
5546 | { | |
5547 | struct pvector *pvec = &cls->subtables; | |
5548 | struct dpcls_subtable *subtable; | |
5549 | ||
5550 | PVECTOR_FOR_EACH (subtable, pvec) { | |
5551 | pvector_change_priority(pvec, subtable, subtable->hit_cnt); | |
5552 | subtable->hit_cnt = 0; | |
5553 | } | |
5554 | pvector_publish(pvec); | |
5555 | } | |
5556 | ||
5557 | static inline void | |
5558 | dp_netdev_pmd_try_optimize(struct dp_netdev_pmd_thread *pmd) | |
5559 | { | |
5560 | struct dpcls *cls; | |
5561 | long long int now = time_msec(); | |
5562 | ||
5563 | if (now > pmd->next_optimization) { | |
5564 | /* Try to obtain the flow lock to block out revalidator threads. | |
5565 | * If not possible, just try next time. */ | |
5566 | if (!ovs_mutex_trylock(&pmd->flow_mutex)) { | |
5567 | /* Optimize each classifier */ | |
5568 | CMAP_FOR_EACH (cls, node, &pmd->classifiers) { | |
5569 | dpcls_sort_subtable_vector(cls); | |
5570 | } | |
5571 | ovs_mutex_unlock(&pmd->flow_mutex); | |
5572 | /* Start new measuring interval */ | |
5573 | pmd->next_optimization = now + DPCLS_OPTIMIZATION_INTERVAL; | |
5574 | } | |
5575 | } | |
5576 | } | |
5577 | ||
5578 | /* Insert 'rule' into 'cls'. */ | |
5579 | static void | |
5580 | dpcls_insert(struct dpcls *cls, struct dpcls_rule *rule, | |
5581 | const struct netdev_flow_key *mask) | |
5582 | { | |
5583 | struct dpcls_subtable *subtable = dpcls_find_subtable(cls, mask); | |
5584 | ||
5585 | /* Refer to subtable's mask, also for later removal. */ | |
5586 | rule->mask = &subtable->mask; | |
5587 | cmap_insert(&subtable->rules, &rule->cmap_node, rule->flow.hash); | |
5588 | } | |
5589 | ||
5590 | /* Removes 'rule' from 'cls', also destructing the 'rule'. */ | |
5591 | static void | |
5592 | dpcls_remove(struct dpcls *cls, struct dpcls_rule *rule) | |
5593 | { | |
5594 | struct dpcls_subtable *subtable; | |
5595 | ||
5596 | ovs_assert(rule->mask); | |
5597 | ||
5598 | /* Get subtable from reference in rule->mask. */ | |
5599 | INIT_CONTAINER(subtable, rule->mask, mask); | |
5600 | if (cmap_remove(&subtable->rules, &rule->cmap_node, rule->flow.hash) | |
5601 | == 0) { | |
5602 | /* Delete empty subtable. */ | |
5603 | dpcls_destroy_subtable(cls, subtable); | |
5604 | pvector_publish(&cls->subtables); | |
5605 | } | |
5606 | } | |
5607 | ||
5608 | /* Returns true if 'target' satisfies 'key' in 'mask', that is, if each 1-bit | |
5609 | * in 'mask' the values in 'key' and 'target' are the same. */ | |
5610 | static inline bool | |
5611 | dpcls_rule_matches_key(const struct dpcls_rule *rule, | |
5612 | const struct netdev_flow_key *target) | |
5613 | { | |
5614 | const uint64_t *keyp = miniflow_get_values(&rule->flow.mf); | |
5615 | const uint64_t *maskp = miniflow_get_values(&rule->mask->mf); | |
5616 | uint64_t value; | |
5617 | ||
5618 | NETDEV_FLOW_KEY_FOR_EACH_IN_FLOWMAP(value, target, rule->flow.mf.map) { | |
5619 | if (OVS_UNLIKELY((value & *maskp++) != *keyp++)) { | |
5620 | return false; | |
5621 | } | |
5622 | } | |
5623 | return true; | |
5624 | } | |
5625 | ||
5626 | /* For each miniflow in 'keys' performs a classifier lookup writing the result | |
5627 | * into the corresponding slot in 'rules'. If a particular entry in 'keys' is | |
5628 | * NULL it is skipped. | |
5629 | * | |
5630 | * This function is optimized for use in the userspace datapath and therefore | |
5631 | * does not implement a lot of features available in the standard | |
5632 | * classifier_lookup() function. Specifically, it does not implement | |
5633 | * priorities, instead returning any rule which matches the flow. | |
5634 | * | |
5635 | * Returns true if all miniflows found a corresponding rule. */ | |
5636 | static bool | |
5637 | dpcls_lookup(struct dpcls *cls, const struct netdev_flow_key keys[], | |
5638 | struct dpcls_rule **rules, const size_t cnt, | |
5639 | int *num_lookups_p) | |
5640 | { | |
5641 | /* The received 'cnt' miniflows are the search-keys that will be processed | |
5642 | * to find a matching entry into the available subtables. | |
5643 | * The number of bits in map_type is equal to NETDEV_MAX_BURST. */ | |
5644 | typedef uint32_t map_type; | |
5645 | #define MAP_BITS (sizeof(map_type) * CHAR_BIT) | |
5646 | BUILD_ASSERT_DECL(MAP_BITS >= NETDEV_MAX_BURST); | |
5647 | ||
5648 | struct dpcls_subtable *subtable; | |
5649 | ||
5650 | map_type keys_map = TYPE_MAXIMUM(map_type); /* Set all bits. */ | |
5651 | map_type found_map; | |
5652 | uint32_t hashes[MAP_BITS]; | |
5653 | const struct cmap_node *nodes[MAP_BITS]; | |
5654 | ||
5655 | if (cnt != MAP_BITS) { | |
5656 | keys_map >>= MAP_BITS - cnt; /* Clear extra bits. */ | |
5657 | } | |
5658 | memset(rules, 0, cnt * sizeof *rules); | |
5659 | ||
5660 | int lookups_match = 0, subtable_pos = 1; | |
5661 | ||
5662 | /* The Datapath classifier - aka dpcls - is composed of subtables. | |
5663 | * Subtables are dynamically created as needed when new rules are inserted. | |
5664 | * Each subtable collects rules with matches on a specific subset of packet | |
5665 | * fields as defined by the subtable's mask. We proceed to process every | |
5666 | * search-key against each subtable, but when a match is found for a | |
5667 | * search-key, the search for that key can stop because the rules are | |
5668 | * non-overlapping. */ | |
5669 | PVECTOR_FOR_EACH (subtable, &cls->subtables) { | |
5670 | int i; | |
5671 | ||
5672 | /* Compute hashes for the remaining keys. Each search-key is | |
5673 | * masked with the subtable's mask to avoid hashing the wildcarded | |
5674 | * bits. */ | |
5675 | ULLONG_FOR_EACH_1(i, keys_map) { | |
5676 | hashes[i] = netdev_flow_key_hash_in_mask(&keys[i], | |
5677 | &subtable->mask); | |
5678 | } | |
5679 | /* Lookup. */ | |
5680 | found_map = cmap_find_batch(&subtable->rules, keys_map, hashes, nodes); | |
5681 | /* Check results. When the i-th bit of found_map is set, it means | |
5682 | * that a set of nodes with a matching hash value was found for the | |
5683 | * i-th search-key. Due to possible hash collisions we need to check | |
5684 | * which of the found rules, if any, really matches our masked | |
5685 | * search-key. */ | |
5686 | ULLONG_FOR_EACH_1(i, found_map) { | |
5687 | struct dpcls_rule *rule; | |
5688 | ||
5689 | CMAP_NODE_FOR_EACH (rule, cmap_node, nodes[i]) { | |
5690 | if (OVS_LIKELY(dpcls_rule_matches_key(rule, &keys[i]))) { | |
5691 | rules[i] = rule; | |
5692 | /* Even at 20 Mpps the 32-bit hit_cnt cannot wrap | |
5693 | * within one second optimization interval. */ | |
5694 | subtable->hit_cnt++; | |
5695 | lookups_match += subtable_pos; | |
5696 | goto next; | |
5697 | } | |
5698 | } | |
5699 | /* None of the found rules was a match. Reset the i-th bit to | |
5700 | * keep searching this key in the next subtable. */ | |
5701 | ULLONG_SET0(found_map, i); /* Did not match. */ | |
5702 | next: | |
5703 | ; /* Keep Sparse happy. */ | |
5704 | } | |
5705 | keys_map &= ~found_map; /* Clear the found rules. */ | |
5706 | if (!keys_map) { | |
5707 | if (num_lookups_p) { | |
5708 | *num_lookups_p = lookups_match; | |
5709 | } | |
5710 | return true; /* All found. */ | |
5711 | } | |
5712 | subtable_pos++; | |
5713 | } | |
5714 | if (num_lookups_p) { | |
5715 | *num_lookups_p = lookups_match; | |
5716 | } | |
5717 | return false; /* Some misses. */ | |
5718 | } |