]> git.proxmox.com Git - ovs.git/blob - lib/dpif-netdev.c
dpif-netdev: Clear flow batches before execute.
[ovs.git] / lib / dpif-netdev.c
1 /*
2 * Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014 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 <netinet/in.h>
25 #include <sys/socket.h>
26 #include <net/if.h>
27 #include <stdint.h>
28 #include <stdlib.h>
29 #include <string.h>
30 #include <sys/ioctl.h>
31 #include <sys/stat.h>
32 #include <unistd.h>
33
34 #include "cmap.h"
35 #include "csum.h"
36 #include "dp-packet.h"
37 #include "dpif.h"
38 #include "dpif-provider.h"
39 #include "dummy.h"
40 #include "dynamic-string.h"
41 #include "fat-rwlock.h"
42 #include "flow.h"
43 #include "cmap.h"
44 #include "latch.h"
45 #include "list.h"
46 #include "match.h"
47 #include "meta-flow.h"
48 #include "netdev.h"
49 #include "netdev-dpdk.h"
50 #include "netdev-vport.h"
51 #include "netlink.h"
52 #include "odp-execute.h"
53 #include "odp-util.h"
54 #include "ofp-print.h"
55 #include "ofpbuf.h"
56 #include "ovs-numa.h"
57 #include "ovs-rcu.h"
58 #include "packets.h"
59 #include "poll-loop.h"
60 #include "pvector.h"
61 #include "random.h"
62 #include "seq.h"
63 #include "shash.h"
64 #include "sset.h"
65 #include "timeval.h"
66 #include "tnl-arp-cache.h"
67 #include "unixctl.h"
68 #include "util.h"
69 #include "openvswitch/vlog.h"
70
71 VLOG_DEFINE_THIS_MODULE(dpif_netdev);
72
73 #define FLOW_DUMP_MAX_BATCH 50
74 /* Use per thread recirc_depth to prevent recirculation loop. */
75 #define MAX_RECIRC_DEPTH 5
76 DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth, 0)
77
78 /* Configuration parameters. */
79 enum { MAX_FLOWS = 65536 }; /* Maximum number of flows in flow table. */
80
81 /* Protects against changes to 'dp_netdevs'. */
82 static struct ovs_mutex dp_netdev_mutex = OVS_MUTEX_INITIALIZER;
83
84 /* Contains all 'struct dp_netdev's. */
85 static struct shash dp_netdevs OVS_GUARDED_BY(dp_netdev_mutex)
86 = SHASH_INITIALIZER(&dp_netdevs);
87
88 static struct vlog_rate_limit upcall_rl = VLOG_RATE_LIMIT_INIT(600, 600);
89
90 /* Stores a miniflow with inline values */
91
92 struct netdev_flow_key {
93 uint32_t hash; /* Hash function differs for different users. */
94 uint32_t len; /* Length of the following miniflow (incl. map). */
95 struct miniflow mf;
96 uint64_t buf[FLOW_MAX_PACKET_U64S - MINI_N_INLINE];
97 };
98
99 /* Exact match cache for frequently used flows
100 *
101 * The cache uses a 32-bit hash of the packet (which can be the RSS hash) to
102 * search its entries for a miniflow that matches exactly the miniflow of the
103 * packet. It stores the 'dpcls_rule' (rule) that matches the miniflow.
104 *
105 * A cache entry holds a reference to its 'dp_netdev_flow'.
106 *
107 * A miniflow with a given hash can be in one of EM_FLOW_HASH_SEGS different
108 * entries. The 32-bit hash is split into EM_FLOW_HASH_SEGS values (each of
109 * them is EM_FLOW_HASH_SHIFT bits wide and the remainder is thrown away). Each
110 * value is the index of a cache entry where the miniflow could be.
111 *
112 *
113 * Thread-safety
114 * =============
115 *
116 * Each pmd_thread has its own private exact match cache.
117 * If dp_netdev_input is not called from a pmd thread, a mutex is used.
118 */
119
120 #define EM_FLOW_HASH_SHIFT 13
121 #define EM_FLOW_HASH_ENTRIES (1u << EM_FLOW_HASH_SHIFT)
122 #define EM_FLOW_HASH_MASK (EM_FLOW_HASH_ENTRIES - 1)
123 #define EM_FLOW_HASH_SEGS 2
124
125 struct emc_entry {
126 struct dp_netdev_flow *flow;
127 struct netdev_flow_key key; /* key.hash used for emc hash value. */
128 };
129
130 struct emc_cache {
131 struct emc_entry entries[EM_FLOW_HASH_ENTRIES];
132 int sweep_idx; /* For emc_cache_slow_sweep(). */
133 };
134
135 /* Iterate in the exact match cache through every entry that might contain a
136 * miniflow with hash 'HASH'. */
137 #define EMC_FOR_EACH_POS_WITH_HASH(EMC, CURRENT_ENTRY, HASH) \
138 for (uint32_t i__ = 0, srch_hash__ = (HASH); \
139 (CURRENT_ENTRY) = &(EMC)->entries[srch_hash__ & EM_FLOW_HASH_MASK], \
140 i__ < EM_FLOW_HASH_SEGS; \
141 i__++, srch_hash__ >>= EM_FLOW_HASH_SHIFT)
142 \f
143 /* Simple non-wildcarding single-priority classifier. */
144
145 struct dpcls {
146 struct cmap subtables_map;
147 struct pvector subtables;
148 };
149
150 /* A rule to be inserted to the classifier. */
151 struct dpcls_rule {
152 struct cmap_node cmap_node; /* Within struct dpcls_subtable 'rules'. */
153 struct netdev_flow_key *mask; /* Subtable's mask. */
154 struct netdev_flow_key flow; /* Matching key. */
155 /* 'flow' must be the last field, additional space is allocated here. */
156 };
157
158 static void dpcls_init(struct dpcls *);
159 static void dpcls_destroy(struct dpcls *);
160 static void dpcls_insert(struct dpcls *, struct dpcls_rule *,
161 const struct netdev_flow_key *mask);
162 static void dpcls_remove(struct dpcls *, struct dpcls_rule *);
163 static bool dpcls_lookup(const struct dpcls *cls,
164 const struct netdev_flow_key keys[],
165 struct dpcls_rule **rules, size_t cnt);
166 \f
167 /* Datapath based on the network device interface from netdev.h.
168 *
169 *
170 * Thread-safety
171 * =============
172 *
173 * Some members, marked 'const', are immutable. Accessing other members
174 * requires synchronization, as noted in more detail below.
175 *
176 * Acquisition order is, from outermost to innermost:
177 *
178 * dp_netdev_mutex (global)
179 * port_mutex
180 */
181 struct dp_netdev {
182 const struct dpif_class *const class;
183 const char *const name;
184 struct dpif *dpif;
185 struct ovs_refcount ref_cnt;
186 atomic_flag destroyed;
187
188 /* Ports.
189 *
190 * Protected by RCU. Take the mutex to add or remove ports. */
191 struct ovs_mutex port_mutex;
192 struct cmap ports;
193 struct seq *port_seq; /* Incremented whenever a port changes. */
194
195 /* Protects access to ofproto-dpif-upcall interface during revalidator
196 * thread synchronization. */
197 struct fat_rwlock upcall_rwlock;
198 upcall_callback *upcall_cb; /* Callback function for executing upcalls. */
199 void *upcall_aux;
200
201 /* Stores all 'struct dp_netdev_pmd_thread's. */
202 struct cmap poll_threads;
203
204 /* Protects the access of the 'struct dp_netdev_pmd_thread'
205 * instance for non-pmd thread. */
206 struct ovs_mutex non_pmd_mutex;
207
208 /* Each pmd thread will store its pointer to
209 * 'struct dp_netdev_pmd_thread' in 'per_pmd_key'. */
210 ovsthread_key_t per_pmd_key;
211
212 /* Number of rx queues for each dpdk interface and the cpu mask
213 * for pin of pmd threads. */
214 size_t n_dpdk_rxqs;
215 char *pmd_cmask;
216 uint64_t last_tnl_conf_seq;
217 };
218
219 static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp,
220 odp_port_t);
221
222 enum dp_stat_type {
223 DP_STAT_EXACT_HIT, /* Packets that had an exact match (emc). */
224 DP_STAT_MASKED_HIT, /* Packets that matched in the flow table. */
225 DP_STAT_MISS, /* Packets that did not match. */
226 DP_STAT_LOST, /* Packets not passed up to the client. */
227 DP_N_STATS
228 };
229
230 enum pmd_cycles_counter_type {
231 PMD_CYCLES_POLLING, /* Cycles spent polling NICs. */
232 PMD_CYCLES_PROCESSING, /* Cycles spent processing packets */
233 PMD_N_CYCLES
234 };
235
236 /* A port in a netdev-based datapath. */
237 struct dp_netdev_port {
238 struct pkt_metadata md;
239 struct netdev *netdev;
240 struct cmap_node node; /* Node in dp_netdev's 'ports'. */
241 struct netdev_saved_flags *sf;
242 struct netdev_rxq **rxq;
243 struct ovs_refcount ref_cnt;
244 char *type; /* Port type as requested by user. */
245 };
246
247 /* Contained by struct dp_netdev_flow's 'stats' member. */
248 struct dp_netdev_flow_stats {
249 atomic_llong used; /* Last used time, in monotonic msecs. */
250 atomic_ullong packet_count; /* Number of packets matched. */
251 atomic_ullong byte_count; /* Number of bytes matched. */
252 atomic_uint16_t tcp_flags; /* Bitwise-OR of seen tcp_flags values. */
253 };
254
255 /* A flow in 'dp_netdev_pmd_thread's 'flow_table'.
256 *
257 *
258 * Thread-safety
259 * =============
260 *
261 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
262 * its pmd thread's classifier. The text below calls this classifier 'cls'.
263 *
264 * Motivation
265 * ----------
266 *
267 * The thread safety rules described here for "struct dp_netdev_flow" are
268 * motivated by two goals:
269 *
270 * - Prevent threads that read members of "struct dp_netdev_flow" from
271 * reading bad data due to changes by some thread concurrently modifying
272 * those members.
273 *
274 * - Prevent two threads making changes to members of a given "struct
275 * dp_netdev_flow" from interfering with each other.
276 *
277 *
278 * Rules
279 * -----
280 *
281 * A flow 'flow' may be accessed without a risk of being freed during an RCU
282 * grace period. Code that needs to hold onto a flow for a while
283 * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref().
284 *
285 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
286 * flow from being deleted from 'cls' and it doesn't protect members of 'flow'
287 * from modification.
288 *
289 * Some members, marked 'const', are immutable. Accessing other members
290 * requires synchronization, as noted in more detail below.
291 */
292 struct dp_netdev_flow {
293 const struct flow flow; /* Unmasked flow that created this entry. */
294 /* Hash table index by unmasked flow. */
295 const struct cmap_node node; /* In owning dp_netdev_pmd_thread's */
296 /* 'flow_table'. */
297 const ovs_u128 ufid; /* Unique flow identifier. */
298 const int pmd_id; /* The 'core_id' of pmd thread owning this */
299 /* flow. */
300
301 /* Number of references.
302 * The classifier owns one reference.
303 * Any thread trying to keep a rule from being freed should hold its own
304 * reference. */
305 struct ovs_refcount ref_cnt;
306
307 bool dead;
308
309 /* Statistics. */
310 struct dp_netdev_flow_stats stats;
311
312 /* Actions. */
313 OVSRCU_TYPE(struct dp_netdev_actions *) actions;
314
315 /* While processing a group of input packets, the datapath uses the next
316 * member to store a pointer to the output batch for the flow. It is
317 * reset after the batch has been sent out (See dp_netdev_queue_batches(),
318 * packet_batch_init() and packet_batch_execute()). */
319 struct packet_batch *batch;
320
321 /* Packet classification. */
322 struct dpcls_rule cr; /* In owning dp_netdev's 'cls'. */
323 /* 'cr' must be the last member. */
324 };
325
326 static void dp_netdev_flow_unref(struct dp_netdev_flow *);
327 static bool dp_netdev_flow_ref(struct dp_netdev_flow *);
328 static int dpif_netdev_flow_from_nlattrs(const struct nlattr *, uint32_t,
329 struct flow *);
330
331 /* A set of datapath actions within a "struct dp_netdev_flow".
332 *
333 *
334 * Thread-safety
335 * =============
336 *
337 * A struct dp_netdev_actions 'actions' is protected with RCU. */
338 struct dp_netdev_actions {
339 /* These members are immutable: they do not change during the struct's
340 * lifetime. */
341 unsigned int size; /* Size of 'actions', in bytes. */
342 struct nlattr actions[]; /* Sequence of OVS_ACTION_ATTR_* attributes. */
343 };
344
345 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
346 size_t);
347 struct dp_netdev_actions *dp_netdev_flow_get_actions(
348 const struct dp_netdev_flow *);
349 static void dp_netdev_actions_free(struct dp_netdev_actions *);
350
351 /* Contained by struct dp_netdev_pmd_thread's 'stats' member. */
352 struct dp_netdev_pmd_stats {
353 /* Indexed by DP_STAT_*. */
354 atomic_ullong n[DP_N_STATS];
355 };
356
357 /* Contained by struct dp_netdev_pmd_thread's 'cycle' member. */
358 struct dp_netdev_pmd_cycles {
359 /* Indexed by PMD_CYCLES_*. */
360 atomic_ullong n[PMD_N_CYCLES];
361 };
362
363 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
364 * the performance overhead of interrupt processing. Therefore netdev can
365 * not implement rx-wait for these devices. dpif-netdev needs to poll
366 * these device to check for recv buffer. pmd-thread does polling for
367 * devices assigned to itself.
368 *
369 * DPDK used PMD for accessing NIC.
370 *
371 * Note, instance with cpu core id NON_PMD_CORE_ID will be reserved for
372 * I/O of all non-pmd threads. There will be no actual thread created
373 * for the instance.
374 *
375 * Each struct has its own flow table and classifier. Packets received
376 * from managed ports are looked up in the corresponding pmd thread's
377 * flow table, and are executed with the found actions.
378 * */
379 struct dp_netdev_pmd_thread {
380 struct dp_netdev *dp;
381 struct ovs_refcount ref_cnt; /* Every reference must be refcount'ed. */
382 struct cmap_node node; /* In 'dp->poll_threads'. */
383
384 pthread_cond_t cond; /* For synchronizing pmd thread reload. */
385 struct ovs_mutex cond_mutex; /* Mutex for condition variable. */
386
387 /* Per thread exact-match cache. Note, the instance for cpu core
388 * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly
389 * need to be protected (e.g. by 'dp_netdev_mutex'). All other
390 * instances will only be accessed by its own pmd thread. */
391 struct emc_cache flow_cache;
392
393 /* Classifier and Flow-Table.
394 *
395 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
396 * changes to 'cls' must be made while still holding the 'flow_mutex'.
397 */
398 struct ovs_mutex flow_mutex;
399 struct dpcls cls;
400 struct cmap flow_table OVS_GUARDED; /* Flow table. */
401
402 /* Statistics. */
403 struct dp_netdev_pmd_stats stats;
404
405 /* Cycles counters */
406 struct dp_netdev_pmd_cycles cycles;
407
408 /* Used to count cicles. See 'cycles_counter_end()' */
409 unsigned long long last_cycles;
410
411 struct latch exit_latch; /* For terminating the pmd thread. */
412 atomic_uint change_seq; /* For reloading pmd ports. */
413 pthread_t thread;
414 int index; /* Idx of this pmd thread among pmd*/
415 /* threads on same numa node. */
416 int core_id; /* CPU core id of this pmd thread. */
417 int numa_id; /* numa node id of this pmd thread. */
418
419 /* Only a pmd thread can write on its own 'cycles' and 'stats'.
420 * The main thread keeps 'stats_zero' and 'cycles_zero' as base
421 * values and subtracts them from 'stats' and 'cycles' before
422 * reporting to the user */
423 unsigned long long stats_zero[DP_N_STATS];
424 uint64_t cycles_zero[PMD_N_CYCLES];
425 };
426
427 #define PMD_INITIAL_SEQ 1
428
429 /* Interface to netdev-based datapath. */
430 struct dpif_netdev {
431 struct dpif dpif;
432 struct dp_netdev *dp;
433 uint64_t last_port_seq;
434 };
435
436 static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no,
437 struct dp_netdev_port **portp);
438 static int get_port_by_name(struct dp_netdev *dp, const char *devname,
439 struct dp_netdev_port **portp);
440 static void dp_netdev_free(struct dp_netdev *)
441 OVS_REQUIRES(dp_netdev_mutex);
442 static int do_add_port(struct dp_netdev *dp, const char *devname,
443 const char *type, odp_port_t port_no)
444 OVS_REQUIRES(dp->port_mutex);
445 static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *)
446 OVS_REQUIRES(dp->port_mutex);
447 static int dpif_netdev_open(const struct dpif_class *, const char *name,
448 bool create, struct dpif **);
449 static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
450 struct dp_packet **, int c,
451 bool may_steal,
452 const struct nlattr *actions,
453 size_t actions_len);
454 static void dp_netdev_input(struct dp_netdev_pmd_thread *,
455 struct dp_packet **, int cnt);
456
457 static void dp_netdev_disable_upcall(struct dp_netdev *);
458 void dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd);
459 static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd,
460 struct dp_netdev *dp, int index,
461 int core_id, int numa_id);
462 static void dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd);
463 static void dp_netdev_set_nonpmd(struct dp_netdev *dp);
464 static struct dp_netdev_pmd_thread *dp_netdev_get_pmd(struct dp_netdev *dp,
465 int core_id);
466 static struct dp_netdev_pmd_thread *
467 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos);
468 static void dp_netdev_destroy_all_pmds(struct dp_netdev *dp);
469 static void dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id);
470 static void dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id);
471 static void dp_netdev_reset_pmd_threads(struct dp_netdev *dp);
472 static bool dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd);
473 static void dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd);
474 static void dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd);
475
476 static inline bool emc_entry_alive(struct emc_entry *ce);
477 static void emc_clear_entry(struct emc_entry *ce);
478
479 static void
480 emc_cache_init(struct emc_cache *flow_cache)
481 {
482 int i;
483
484 BUILD_ASSERT(offsetof(struct miniflow, inline_values) == sizeof(uint64_t));
485
486 flow_cache->sweep_idx = 0;
487 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
488 flow_cache->entries[i].flow = NULL;
489 flow_cache->entries[i].key.hash = 0;
490 flow_cache->entries[i].key.len
491 = offsetof(struct miniflow, inline_values);
492 miniflow_initialize(&flow_cache->entries[i].key.mf,
493 flow_cache->entries[i].key.buf);
494 }
495 }
496
497 static void
498 emc_cache_uninit(struct emc_cache *flow_cache)
499 {
500 int i;
501
502 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
503 emc_clear_entry(&flow_cache->entries[i]);
504 }
505 }
506
507 /* Check and clear dead flow references slowly (one entry at each
508 * invocation). */
509 static void
510 emc_cache_slow_sweep(struct emc_cache *flow_cache)
511 {
512 struct emc_entry *entry = &flow_cache->entries[flow_cache->sweep_idx];
513
514 if (!emc_entry_alive(entry)) {
515 emc_clear_entry(entry);
516 }
517 flow_cache->sweep_idx = (flow_cache->sweep_idx + 1) & EM_FLOW_HASH_MASK;
518 }
519
520 static struct dpif_netdev *
521 dpif_netdev_cast(const struct dpif *dpif)
522 {
523 ovs_assert(dpif->dpif_class->open == dpif_netdev_open);
524 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
525 }
526
527 static struct dp_netdev *
528 get_dp_netdev(const struct dpif *dpif)
529 {
530 return dpif_netdev_cast(dpif)->dp;
531 }
532 \f
533 enum pmd_info_type {
534 PMD_INFO_SHOW_STATS, /* show how cpu cycles are spent */
535 PMD_INFO_CLEAR_STATS /* set the cycles count to 0 */
536 };
537
538 static void
539 pmd_info_show_stats(struct ds *reply,
540 struct dp_netdev_pmd_thread *pmd,
541 unsigned long long stats[DP_N_STATS],
542 uint64_t cycles[PMD_N_CYCLES])
543 {
544 unsigned long long total_packets = 0;
545 uint64_t total_cycles = 0;
546 int i;
547
548 /* These loops subtracts reference values ('*_zero') from the counters.
549 * Since loads and stores are relaxed, it might be possible for a '*_zero'
550 * value to be more recent than the current value we're reading from the
551 * counter. This is not a big problem, since these numbers are not
552 * supposed to be too accurate, but we should at least make sure that
553 * the result is not negative. */
554 for (i = 0; i < DP_N_STATS; i++) {
555 if (stats[i] > pmd->stats_zero[i]) {
556 stats[i] -= pmd->stats_zero[i];
557 } else {
558 stats[i] = 0;
559 }
560
561 if (i != DP_STAT_LOST) {
562 /* Lost packets are already included in DP_STAT_MISS */
563 total_packets += stats[i];
564 }
565 }
566
567 for (i = 0; i < PMD_N_CYCLES; i++) {
568 if (cycles[i] > pmd->cycles_zero[i]) {
569 cycles[i] -= pmd->cycles_zero[i];
570 } else {
571 cycles[i] = 0;
572 }
573
574 total_cycles += cycles[i];
575 }
576
577 ds_put_cstr(reply, (pmd->core_id == NON_PMD_CORE_ID)
578 ? "main thread" : "pmd thread");
579
580 if (pmd->numa_id != OVS_NUMA_UNSPEC) {
581 ds_put_format(reply, " numa_id %d", pmd->numa_id);
582 }
583 if (pmd->core_id != OVS_CORE_UNSPEC) {
584 ds_put_format(reply, " core_id %d", pmd->core_id);
585 }
586 ds_put_cstr(reply, ":\n");
587
588 ds_put_format(reply,
589 "\temc hits:%llu\n\tmegaflow hits:%llu\n"
590 "\tmiss:%llu\n\tlost:%llu\n",
591 stats[DP_STAT_EXACT_HIT], stats[DP_STAT_MASKED_HIT],
592 stats[DP_STAT_MISS], stats[DP_STAT_LOST]);
593
594 if (total_cycles == 0) {
595 return;
596 }
597
598 ds_put_format(reply,
599 "\tpolling cycles:%"PRIu64" (%.02f%%)\n"
600 "\tprocessing cycles:%"PRIu64" (%.02f%%)\n",
601 cycles[PMD_CYCLES_POLLING],
602 cycles[PMD_CYCLES_POLLING] / (double)total_cycles * 100,
603 cycles[PMD_CYCLES_PROCESSING],
604 cycles[PMD_CYCLES_PROCESSING] / (double)total_cycles * 100);
605
606 if (total_packets == 0) {
607 return;
608 }
609
610 ds_put_format(reply,
611 "\tavg cycles per packet: %.02f (%"PRIu64"/%llu)\n",
612 total_cycles / (double)total_packets,
613 total_cycles, total_packets);
614
615 ds_put_format(reply,
616 "\tavg processing cycles per packet: "
617 "%.02f (%"PRIu64"/%llu)\n",
618 cycles[PMD_CYCLES_PROCESSING] / (double)total_packets,
619 cycles[PMD_CYCLES_PROCESSING], total_packets);
620 }
621
622 static void
623 pmd_info_clear_stats(struct ds *reply OVS_UNUSED,
624 struct dp_netdev_pmd_thread *pmd,
625 unsigned long long stats[DP_N_STATS],
626 uint64_t cycles[PMD_N_CYCLES])
627 {
628 int i;
629
630 /* We cannot write 'stats' and 'cycles' (because they're written by other
631 * threads) and we shouldn't change 'stats' (because they're used to count
632 * datapath stats, which must not be cleared here). Instead, we save the
633 * current values and subtract them from the values to be displayed in the
634 * future */
635 for (i = 0; i < DP_N_STATS; i++) {
636 pmd->stats_zero[i] = stats[i];
637 }
638 for (i = 0; i < PMD_N_CYCLES; i++) {
639 pmd->cycles_zero[i] = cycles[i];
640 }
641 }
642
643 static void
644 dpif_netdev_pmd_info(struct unixctl_conn *conn, int argc, const char *argv[],
645 void *aux)
646 {
647 struct ds reply = DS_EMPTY_INITIALIZER;
648 struct dp_netdev_pmd_thread *pmd;
649 struct dp_netdev *dp = NULL;
650 enum pmd_info_type type = *(enum pmd_info_type *) aux;
651
652 ovs_mutex_lock(&dp_netdev_mutex);
653
654 if (argc == 2) {
655 dp = shash_find_data(&dp_netdevs, argv[1]);
656 } else if (shash_count(&dp_netdevs) == 1) {
657 /* There's only one datapath */
658 dp = shash_first(&dp_netdevs)->data;
659 }
660
661 if (!dp) {
662 ovs_mutex_unlock(&dp_netdev_mutex);
663 unixctl_command_reply_error(conn,
664 "please specify an existing datapath");
665 return;
666 }
667
668 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
669 unsigned long long stats[DP_N_STATS];
670 uint64_t cycles[PMD_N_CYCLES];
671 int i;
672
673 /* Read current stats and cycle counters */
674 for (i = 0; i < ARRAY_SIZE(stats); i++) {
675 atomic_read_relaxed(&pmd->stats.n[i], &stats[i]);
676 }
677 for (i = 0; i < ARRAY_SIZE(cycles); i++) {
678 atomic_read_relaxed(&pmd->cycles.n[i], &cycles[i]);
679 }
680
681 if (type == PMD_INFO_CLEAR_STATS) {
682 pmd_info_clear_stats(&reply, pmd, stats, cycles);
683 } else if (type == PMD_INFO_SHOW_STATS) {
684 pmd_info_show_stats(&reply, pmd, stats, cycles);
685 }
686 }
687
688 ovs_mutex_unlock(&dp_netdev_mutex);
689
690 unixctl_command_reply(conn, ds_cstr(&reply));
691 ds_destroy(&reply);
692 }
693 \f
694 static int
695 dpif_netdev_init(void)
696 {
697 static enum pmd_info_type show_aux = PMD_INFO_SHOW_STATS,
698 clear_aux = PMD_INFO_CLEAR_STATS;
699
700 unixctl_command_register("dpif-netdev/pmd-stats-show", "[dp]",
701 0, 1, dpif_netdev_pmd_info,
702 (void *)&show_aux);
703 unixctl_command_register("dpif-netdev/pmd-stats-clear", "[dp]",
704 0, 1, dpif_netdev_pmd_info,
705 (void *)&clear_aux);
706 return 0;
707 }
708
709 static int
710 dpif_netdev_enumerate(struct sset *all_dps,
711 const struct dpif_class *dpif_class)
712 {
713 struct shash_node *node;
714
715 ovs_mutex_lock(&dp_netdev_mutex);
716 SHASH_FOR_EACH(node, &dp_netdevs) {
717 struct dp_netdev *dp = node->data;
718 if (dpif_class != dp->class) {
719 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
720 * If the class doesn't match, skip this dpif. */
721 continue;
722 }
723 sset_add(all_dps, node->name);
724 }
725 ovs_mutex_unlock(&dp_netdev_mutex);
726
727 return 0;
728 }
729
730 static bool
731 dpif_netdev_class_is_dummy(const struct dpif_class *class)
732 {
733 return class != &dpif_netdev_class;
734 }
735
736 static const char *
737 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
738 {
739 return strcmp(type, "internal") ? type
740 : dpif_netdev_class_is_dummy(class) ? "dummy"
741 : "tap";
742 }
743
744 static struct dpif *
745 create_dpif_netdev(struct dp_netdev *dp)
746 {
747 uint16_t netflow_id = hash_string(dp->name, 0);
748 struct dpif_netdev *dpif;
749
750 ovs_refcount_ref(&dp->ref_cnt);
751
752 dpif = xmalloc(sizeof *dpif);
753 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
754 dpif->dp = dp;
755 dpif->last_port_seq = seq_read(dp->port_seq);
756
757 return &dpif->dpif;
758 }
759
760 /* Choose an unused, non-zero port number and return it on success.
761 * Return ODPP_NONE on failure. */
762 static odp_port_t
763 choose_port(struct dp_netdev *dp, const char *name)
764 OVS_REQUIRES(dp->port_mutex)
765 {
766 uint32_t port_no;
767
768 if (dp->class != &dpif_netdev_class) {
769 const char *p;
770 int start_no = 0;
771
772 /* If the port name begins with "br", start the number search at
773 * 100 to make writing tests easier. */
774 if (!strncmp(name, "br", 2)) {
775 start_no = 100;
776 }
777
778 /* If the port name contains a number, try to assign that port number.
779 * This can make writing unit tests easier because port numbers are
780 * predictable. */
781 for (p = name; *p != '\0'; p++) {
782 if (isdigit((unsigned char) *p)) {
783 port_no = start_no + strtol(p, NULL, 10);
784 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
785 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
786 return u32_to_odp(port_no);
787 }
788 break;
789 }
790 }
791 }
792
793 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
794 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
795 return u32_to_odp(port_no);
796 }
797 }
798
799 return ODPP_NONE;
800 }
801
802 static int
803 create_dp_netdev(const char *name, const struct dpif_class *class,
804 struct dp_netdev **dpp)
805 OVS_REQUIRES(dp_netdev_mutex)
806 {
807 struct dp_netdev *dp;
808 int error;
809
810 dp = xzalloc(sizeof *dp);
811 shash_add(&dp_netdevs, name, dp);
812
813 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
814 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
815 ovs_refcount_init(&dp->ref_cnt);
816 atomic_flag_clear(&dp->destroyed);
817
818 ovs_mutex_init(&dp->port_mutex);
819 cmap_init(&dp->ports);
820 dp->port_seq = seq_create();
821 fat_rwlock_init(&dp->upcall_rwlock);
822
823 /* Disable upcalls by default. */
824 dp_netdev_disable_upcall(dp);
825 dp->upcall_aux = NULL;
826 dp->upcall_cb = NULL;
827
828 cmap_init(&dp->poll_threads);
829 ovs_mutex_init_recursive(&dp->non_pmd_mutex);
830 ovsthread_key_create(&dp->per_pmd_key, NULL);
831
832 /* Reserves the core NON_PMD_CORE_ID for all non-pmd threads. */
833 ovs_numa_try_pin_core_specific(NON_PMD_CORE_ID);
834 dp_netdev_set_nonpmd(dp);
835 dp->n_dpdk_rxqs = NR_QUEUE;
836
837 ovs_mutex_lock(&dp->port_mutex);
838 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
839 ovs_mutex_unlock(&dp->port_mutex);
840 if (error) {
841 dp_netdev_free(dp);
842 return error;
843 }
844
845 dp->last_tnl_conf_seq = seq_read(tnl_conf_seq);
846 *dpp = dp;
847 return 0;
848 }
849
850 static int
851 dpif_netdev_open(const struct dpif_class *class, const char *name,
852 bool create, struct dpif **dpifp)
853 {
854 struct dp_netdev *dp;
855 int error;
856
857 ovs_mutex_lock(&dp_netdev_mutex);
858 dp = shash_find_data(&dp_netdevs, name);
859 if (!dp) {
860 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
861 } else {
862 error = (dp->class != class ? EINVAL
863 : create ? EEXIST
864 : 0);
865 }
866 if (!error) {
867 *dpifp = create_dpif_netdev(dp);
868 dp->dpif = *dpifp;
869 }
870 ovs_mutex_unlock(&dp_netdev_mutex);
871
872 return error;
873 }
874
875 static void
876 dp_netdev_destroy_upcall_lock(struct dp_netdev *dp)
877 OVS_NO_THREAD_SAFETY_ANALYSIS
878 {
879 /* Check that upcalls are disabled, i.e. that the rwlock is taken */
880 ovs_assert(fat_rwlock_tryrdlock(&dp->upcall_rwlock));
881
882 /* Before freeing a lock we should release it */
883 fat_rwlock_unlock(&dp->upcall_rwlock);
884 fat_rwlock_destroy(&dp->upcall_rwlock);
885 }
886
887 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
888 * through the 'dp_netdevs' shash while freeing 'dp'. */
889 static void
890 dp_netdev_free(struct dp_netdev *dp)
891 OVS_REQUIRES(dp_netdev_mutex)
892 {
893 struct dp_netdev_port *port;
894
895 shash_find_and_delete(&dp_netdevs, dp->name);
896
897 dp_netdev_destroy_all_pmds(dp);
898 cmap_destroy(&dp->poll_threads);
899 ovs_mutex_destroy(&dp->non_pmd_mutex);
900 ovsthread_key_delete(dp->per_pmd_key);
901
902 ovs_mutex_lock(&dp->port_mutex);
903 CMAP_FOR_EACH (port, node, &dp->ports) {
904 do_del_port(dp, port);
905 }
906 ovs_mutex_unlock(&dp->port_mutex);
907
908 seq_destroy(dp->port_seq);
909 cmap_destroy(&dp->ports);
910
911 /* Upcalls must be disabled at this point */
912 dp_netdev_destroy_upcall_lock(dp);
913
914 free(dp->pmd_cmask);
915 free(CONST_CAST(char *, dp->name));
916 free(dp);
917 }
918
919 static void
920 dp_netdev_unref(struct dp_netdev *dp)
921 {
922 if (dp) {
923 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
924 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
925 ovs_mutex_lock(&dp_netdev_mutex);
926 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
927 dp_netdev_free(dp);
928 }
929 ovs_mutex_unlock(&dp_netdev_mutex);
930 }
931 }
932
933 static void
934 dpif_netdev_close(struct dpif *dpif)
935 {
936 struct dp_netdev *dp = get_dp_netdev(dpif);
937
938 dp_netdev_unref(dp);
939 free(dpif);
940 }
941
942 static int
943 dpif_netdev_destroy(struct dpif *dpif)
944 {
945 struct dp_netdev *dp = get_dp_netdev(dpif);
946
947 if (!atomic_flag_test_and_set(&dp->destroyed)) {
948 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
949 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
950 OVS_NOT_REACHED();
951 }
952 }
953
954 return 0;
955 }
956
957 /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed
958 * load/store semantics. While the increment is not atomic, the load and
959 * store operations are, making it impossible to read inconsistent values.
960 *
961 * This is used to update thread local stats counters. */
962 static void
963 non_atomic_ullong_add(atomic_ullong *var, unsigned long long n)
964 {
965 unsigned long long tmp;
966
967 atomic_read_relaxed(var, &tmp);
968 tmp += n;
969 atomic_store_relaxed(var, tmp);
970 }
971
972 static int
973 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
974 {
975 struct dp_netdev *dp = get_dp_netdev(dpif);
976 struct dp_netdev_pmd_thread *pmd;
977
978 stats->n_flows = stats->n_hit = stats->n_missed = stats->n_lost = 0;
979 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
980 unsigned long long n;
981 stats->n_flows += cmap_count(&pmd->flow_table);
982
983 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MASKED_HIT], &n);
984 stats->n_hit += n;
985 atomic_read_relaxed(&pmd->stats.n[DP_STAT_EXACT_HIT], &n);
986 stats->n_hit += n;
987 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MISS], &n);
988 stats->n_missed += n;
989 atomic_read_relaxed(&pmd->stats.n[DP_STAT_LOST], &n);
990 stats->n_lost += n;
991 }
992 stats->n_masks = UINT32_MAX;
993 stats->n_mask_hit = UINT64_MAX;
994
995 return 0;
996 }
997
998 static void
999 dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread *pmd)
1000 {
1001 int old_seq;
1002
1003 if (pmd->core_id == NON_PMD_CORE_ID) {
1004 return;
1005 }
1006
1007 ovs_mutex_lock(&pmd->cond_mutex);
1008 atomic_add_relaxed(&pmd->change_seq, 1, &old_seq);
1009 ovs_mutex_cond_wait(&pmd->cond, &pmd->cond_mutex);
1010 ovs_mutex_unlock(&pmd->cond_mutex);
1011 }
1012
1013 /* Causes all pmd threads to reload its tx/rx devices.
1014 * Must be called after adding/removing ports. */
1015 static void
1016 dp_netdev_reload_pmds(struct dp_netdev *dp)
1017 {
1018 struct dp_netdev_pmd_thread *pmd;
1019
1020 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1021 dp_netdev_reload_pmd__(pmd);
1022 }
1023 }
1024
1025 static uint32_t
1026 hash_port_no(odp_port_t port_no)
1027 {
1028 return hash_int(odp_to_u32(port_no), 0);
1029 }
1030
1031 static int
1032 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
1033 odp_port_t port_no)
1034 OVS_REQUIRES(dp->port_mutex)
1035 {
1036 struct netdev_saved_flags *sf;
1037 struct dp_netdev_port *port;
1038 struct netdev *netdev;
1039 enum netdev_flags flags;
1040 const char *open_type;
1041 int error;
1042 int i;
1043
1044 /* Reject devices already in 'dp'. */
1045 if (!get_port_by_name(dp, devname, &port)) {
1046 return EEXIST;
1047 }
1048
1049 /* Open and validate network device. */
1050 open_type = dpif_netdev_port_open_type(dp->class, type);
1051 error = netdev_open(devname, open_type, &netdev);
1052 if (error) {
1053 return error;
1054 }
1055 /* XXX reject non-Ethernet devices */
1056
1057 netdev_get_flags(netdev, &flags);
1058 if (flags & NETDEV_LOOPBACK) {
1059 VLOG_ERR("%s: cannot add a loopback device", devname);
1060 netdev_close(netdev);
1061 return EINVAL;
1062 }
1063
1064 if (netdev_is_pmd(netdev)) {
1065 int n_cores = ovs_numa_get_n_cores();
1066
1067 if (n_cores == OVS_CORE_UNSPEC) {
1068 VLOG_ERR("%s, cannot get cpu core info", devname);
1069 return ENOENT;
1070 }
1071 /* There can only be ovs_numa_get_n_cores() pmd threads,
1072 * so creates a txq for each. */
1073 error = netdev_set_multiq(netdev, n_cores, dp->n_dpdk_rxqs);
1074 if (error && (error != EOPNOTSUPP)) {
1075 VLOG_ERR("%s, cannot set multiq", devname);
1076 return errno;
1077 }
1078 }
1079 port = xzalloc(sizeof *port);
1080 port->md = PKT_METADATA_INITIALIZER(port_no);
1081 port->netdev = netdev;
1082 port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev));
1083 port->type = xstrdup(type);
1084 for (i = 0; i < netdev_n_rxq(netdev); i++) {
1085 error = netdev_rxq_open(netdev, &port->rxq[i], i);
1086 if (error
1087 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
1088 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
1089 devname, ovs_strerror(errno));
1090 netdev_close(netdev);
1091 free(port->type);
1092 free(port->rxq);
1093 free(port);
1094 return error;
1095 }
1096 }
1097
1098 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
1099 if (error) {
1100 for (i = 0; i < netdev_n_rxq(netdev); i++) {
1101 netdev_rxq_close(port->rxq[i]);
1102 }
1103 netdev_close(netdev);
1104 free(port->type);
1105 free(port->rxq);
1106 free(port);
1107 return error;
1108 }
1109 port->sf = sf;
1110
1111 ovs_refcount_init(&port->ref_cnt);
1112 cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
1113
1114 if (netdev_is_pmd(netdev)) {
1115 dp_netdev_set_pmds_on_numa(dp, netdev_get_numa_id(netdev));
1116 dp_netdev_reload_pmds(dp);
1117 }
1118 seq_change(dp->port_seq);
1119
1120 return 0;
1121 }
1122
1123 static int
1124 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
1125 odp_port_t *port_nop)
1126 {
1127 struct dp_netdev *dp = get_dp_netdev(dpif);
1128 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
1129 const char *dpif_port;
1130 odp_port_t port_no;
1131 int error;
1132
1133 ovs_mutex_lock(&dp->port_mutex);
1134 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
1135 if (*port_nop != ODPP_NONE) {
1136 port_no = *port_nop;
1137 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
1138 } else {
1139 port_no = choose_port(dp, dpif_port);
1140 error = port_no == ODPP_NONE ? EFBIG : 0;
1141 }
1142 if (!error) {
1143 *port_nop = port_no;
1144 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
1145 }
1146 ovs_mutex_unlock(&dp->port_mutex);
1147
1148 return error;
1149 }
1150
1151 static int
1152 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
1153 {
1154 struct dp_netdev *dp = get_dp_netdev(dpif);
1155 int error;
1156
1157 ovs_mutex_lock(&dp->port_mutex);
1158 if (port_no == ODPP_LOCAL) {
1159 error = EINVAL;
1160 } else {
1161 struct dp_netdev_port *port;
1162
1163 error = get_port_by_number(dp, port_no, &port);
1164 if (!error) {
1165 do_del_port(dp, port);
1166 }
1167 }
1168 ovs_mutex_unlock(&dp->port_mutex);
1169
1170 return error;
1171 }
1172
1173 static bool
1174 is_valid_port_number(odp_port_t port_no)
1175 {
1176 return port_no != ODPP_NONE;
1177 }
1178
1179 static struct dp_netdev_port *
1180 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
1181 {
1182 struct dp_netdev_port *port;
1183
1184 CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
1185 if (port->md.in_port.odp_port == port_no) {
1186 return port;
1187 }
1188 }
1189 return NULL;
1190 }
1191
1192 static int
1193 get_port_by_number(struct dp_netdev *dp,
1194 odp_port_t port_no, struct dp_netdev_port **portp)
1195 {
1196 if (!is_valid_port_number(port_no)) {
1197 *portp = NULL;
1198 return EINVAL;
1199 } else {
1200 *portp = dp_netdev_lookup_port(dp, port_no);
1201 return *portp ? 0 : ENOENT;
1202 }
1203 }
1204
1205 static void
1206 port_ref(struct dp_netdev_port *port)
1207 {
1208 if (port) {
1209 ovs_refcount_ref(&port->ref_cnt);
1210 }
1211 }
1212
1213 static bool
1214 port_try_ref(struct dp_netdev_port *port)
1215 {
1216 if (port) {
1217 return ovs_refcount_try_ref_rcu(&port->ref_cnt);
1218 }
1219
1220 return false;
1221 }
1222
1223 static void
1224 port_unref(struct dp_netdev_port *port)
1225 {
1226 if (port && ovs_refcount_unref_relaxed(&port->ref_cnt) == 1) {
1227 int n_rxq = netdev_n_rxq(port->netdev);
1228 int i;
1229
1230 netdev_close(port->netdev);
1231 netdev_restore_flags(port->sf);
1232
1233 for (i = 0; i < n_rxq; i++) {
1234 netdev_rxq_close(port->rxq[i]);
1235 }
1236 free(port->rxq);
1237 free(port->type);
1238 free(port);
1239 }
1240 }
1241
1242 static int
1243 get_port_by_name(struct dp_netdev *dp,
1244 const char *devname, struct dp_netdev_port **portp)
1245 OVS_REQUIRES(dp->port_mutex)
1246 {
1247 struct dp_netdev_port *port;
1248
1249 CMAP_FOR_EACH (port, node, &dp->ports) {
1250 if (!strcmp(netdev_get_name(port->netdev), devname)) {
1251 *portp = port;
1252 return 0;
1253 }
1254 }
1255 return ENOENT;
1256 }
1257
1258 static int
1259 get_n_pmd_threads_on_numa(struct dp_netdev *dp, int numa_id)
1260 {
1261 struct dp_netdev_pmd_thread *pmd;
1262 int n_pmds = 0;
1263
1264 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1265 if (pmd->numa_id == numa_id) {
1266 n_pmds++;
1267 }
1268 }
1269
1270 return n_pmds;
1271 }
1272
1273 /* Returns 'true' if there is a port with pmd netdev and the netdev
1274 * is on numa node 'numa_id'. */
1275 static bool
1276 has_pmd_port_for_numa(struct dp_netdev *dp, int numa_id)
1277 {
1278 struct dp_netdev_port *port;
1279
1280 CMAP_FOR_EACH (port, node, &dp->ports) {
1281 if (netdev_is_pmd(port->netdev)
1282 && netdev_get_numa_id(port->netdev) == numa_id) {
1283 return true;
1284 }
1285 }
1286
1287 return false;
1288 }
1289
1290
1291 static void
1292 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
1293 OVS_REQUIRES(dp->port_mutex)
1294 {
1295 cmap_remove(&dp->ports, &port->node,
1296 hash_odp_port(port->md.in_port.odp_port));
1297 seq_change(dp->port_seq);
1298 if (netdev_is_pmd(port->netdev)) {
1299 int numa_id = netdev_get_numa_id(port->netdev);
1300
1301 /* If there is no netdev on the numa node, deletes the pmd threads
1302 * for that numa. Else, just reloads the queues. */
1303 if (!has_pmd_port_for_numa(dp, numa_id)) {
1304 dp_netdev_del_pmds_on_numa(dp, numa_id);
1305 }
1306 dp_netdev_reload_pmds(dp);
1307 }
1308
1309 port_unref(port);
1310 }
1311
1312 static void
1313 answer_port_query(const struct dp_netdev_port *port,
1314 struct dpif_port *dpif_port)
1315 {
1316 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
1317 dpif_port->type = xstrdup(port->type);
1318 dpif_port->port_no = port->md.in_port.odp_port;
1319 }
1320
1321 static int
1322 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
1323 struct dpif_port *dpif_port)
1324 {
1325 struct dp_netdev *dp = get_dp_netdev(dpif);
1326 struct dp_netdev_port *port;
1327 int error;
1328
1329 error = get_port_by_number(dp, port_no, &port);
1330 if (!error && dpif_port) {
1331 answer_port_query(port, dpif_port);
1332 }
1333
1334 return error;
1335 }
1336
1337 static int
1338 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
1339 struct dpif_port *dpif_port)
1340 {
1341 struct dp_netdev *dp = get_dp_netdev(dpif);
1342 struct dp_netdev_port *port;
1343 int error;
1344
1345 ovs_mutex_lock(&dp->port_mutex);
1346 error = get_port_by_name(dp, devname, &port);
1347 if (!error && dpif_port) {
1348 answer_port_query(port, dpif_port);
1349 }
1350 ovs_mutex_unlock(&dp->port_mutex);
1351
1352 return error;
1353 }
1354
1355 static void
1356 dp_netdev_flow_free(struct dp_netdev_flow *flow)
1357 {
1358 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
1359 free(flow);
1360 }
1361
1362 static void dp_netdev_flow_unref(struct dp_netdev_flow *flow)
1363 {
1364 if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) {
1365 ovsrcu_postpone(dp_netdev_flow_free, flow);
1366 }
1367 }
1368
1369 static uint32_t
1370 dp_netdev_flow_hash(const ovs_u128 *ufid)
1371 {
1372 return ufid->u32[0];
1373 }
1374
1375 static void
1376 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread *pmd,
1377 struct dp_netdev_flow *flow)
1378 OVS_REQUIRES(pmd->flow_mutex)
1379 {
1380 struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
1381
1382 dpcls_remove(&pmd->cls, &flow->cr);
1383 cmap_remove(&pmd->flow_table, node, dp_netdev_flow_hash(&flow->ufid));
1384 flow->dead = true;
1385
1386 dp_netdev_flow_unref(flow);
1387 }
1388
1389 static void
1390 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd)
1391 {
1392 struct dp_netdev_flow *netdev_flow;
1393
1394 ovs_mutex_lock(&pmd->flow_mutex);
1395 CMAP_FOR_EACH (netdev_flow, node, &pmd->flow_table) {
1396 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
1397 }
1398 ovs_mutex_unlock(&pmd->flow_mutex);
1399 }
1400
1401 static int
1402 dpif_netdev_flow_flush(struct dpif *dpif)
1403 {
1404 struct dp_netdev *dp = get_dp_netdev(dpif);
1405 struct dp_netdev_pmd_thread *pmd;
1406
1407 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1408 dp_netdev_pmd_flow_flush(pmd);
1409 }
1410
1411 return 0;
1412 }
1413
1414 struct dp_netdev_port_state {
1415 struct cmap_position position;
1416 char *name;
1417 };
1418
1419 static int
1420 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
1421 {
1422 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
1423 return 0;
1424 }
1425
1426 static int
1427 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
1428 struct dpif_port *dpif_port)
1429 {
1430 struct dp_netdev_port_state *state = state_;
1431 struct dp_netdev *dp = get_dp_netdev(dpif);
1432 struct cmap_node *node;
1433 int retval;
1434
1435 node = cmap_next_position(&dp->ports, &state->position);
1436 if (node) {
1437 struct dp_netdev_port *port;
1438
1439 port = CONTAINER_OF(node, struct dp_netdev_port, node);
1440
1441 free(state->name);
1442 state->name = xstrdup(netdev_get_name(port->netdev));
1443 dpif_port->name = state->name;
1444 dpif_port->type = port->type;
1445 dpif_port->port_no = port->md.in_port.odp_port;
1446
1447 retval = 0;
1448 } else {
1449 retval = EOF;
1450 }
1451
1452 return retval;
1453 }
1454
1455 static int
1456 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1457 {
1458 struct dp_netdev_port_state *state = state_;
1459 free(state->name);
1460 free(state);
1461 return 0;
1462 }
1463
1464 static int
1465 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1466 {
1467 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1468 uint64_t new_port_seq;
1469 int error;
1470
1471 new_port_seq = seq_read(dpif->dp->port_seq);
1472 if (dpif->last_port_seq != new_port_seq) {
1473 dpif->last_port_seq = new_port_seq;
1474 error = ENOBUFS;
1475 } else {
1476 error = EAGAIN;
1477 }
1478
1479 return error;
1480 }
1481
1482 static void
1483 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1484 {
1485 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1486
1487 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1488 }
1489
1490 static struct dp_netdev_flow *
1491 dp_netdev_flow_cast(const struct dpcls_rule *cr)
1492 {
1493 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1494 }
1495
1496 static bool dp_netdev_flow_ref(struct dp_netdev_flow *flow)
1497 {
1498 return ovs_refcount_try_ref_rcu(&flow->ref_cnt);
1499 }
1500
1501 /* netdev_flow_key utilities.
1502 *
1503 * netdev_flow_key is basically a miniflow. We use these functions
1504 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
1505 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
1506 *
1507 * - Since we are dealing exclusively with miniflows created by
1508 * miniflow_extract(), if the map is different the miniflow is different.
1509 * Therefore we can be faster by comparing the map and the miniflow in a
1510 * single memcmp().
1511 * _ netdev_flow_key's miniflow has always inline values.
1512 * - These functions can be inlined by the compiler.
1513 *
1514 * The following assertions make sure that what we're doing with miniflow is
1515 * safe
1516 */
1517 BUILD_ASSERT_DECL(offsetof(struct miniflow, inline_values)
1518 == sizeof(uint64_t));
1519
1520 /* Given the number of bits set in the miniflow map, returns the size of the
1521 * 'netdev_flow_key.mf' */
1522 static inline uint32_t
1523 netdev_flow_key_size(uint32_t flow_u32s)
1524 {
1525 return offsetof(struct miniflow, inline_values) +
1526 MINIFLOW_VALUES_SIZE(flow_u32s);
1527 }
1528
1529 static inline bool
1530 netdev_flow_key_equal(const struct netdev_flow_key *a,
1531 const struct netdev_flow_key *b)
1532 {
1533 /* 'b->len' may be not set yet. */
1534 return a->hash == b->hash && !memcmp(&a->mf, &b->mf, a->len);
1535 }
1536
1537 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
1538 * The maps are compared bitwise, so both 'key->mf' 'mf' must have been
1539 * generated by miniflow_extract. */
1540 static inline bool
1541 netdev_flow_key_equal_mf(const struct netdev_flow_key *key,
1542 const struct miniflow *mf)
1543 {
1544 return !memcmp(&key->mf, mf, key->len);
1545 }
1546
1547 static inline void
1548 netdev_flow_key_clone(struct netdev_flow_key *dst,
1549 const struct netdev_flow_key *src)
1550 {
1551 memcpy(dst, src,
1552 offsetof(struct netdev_flow_key, mf) + src->len);
1553 }
1554
1555 /* Slow. */
1556 static void
1557 netdev_flow_key_from_flow(struct netdev_flow_key *dst,
1558 const struct flow *src)
1559 {
1560 struct dp_packet packet;
1561 uint64_t buf_stub[512 / 8];
1562
1563 miniflow_initialize(&dst->mf, dst->buf);
1564
1565 dp_packet_use_stub(&packet, buf_stub, sizeof buf_stub);
1566 pkt_metadata_from_flow(&packet.md, src);
1567 flow_compose(&packet, src);
1568 miniflow_extract(&packet, &dst->mf);
1569 dp_packet_uninit(&packet);
1570
1571 dst->len = netdev_flow_key_size(count_1bits(dst->mf.map));
1572 dst->hash = 0; /* Not computed yet. */
1573 }
1574
1575 /* Initialize a netdev_flow_key 'mask' from 'match'. */
1576 static inline void
1577 netdev_flow_mask_init(struct netdev_flow_key *mask,
1578 const struct match *match)
1579 {
1580 const uint64_t *mask_u64 = (const uint64_t *) &match->wc.masks;
1581 uint64_t *dst = mask->mf.inline_values;
1582 uint64_t map, mask_map = 0;
1583 uint32_t hash = 0;
1584 int n;
1585
1586 /* Only check masks that make sense for the flow. */
1587 map = flow_wc_map(&match->flow);
1588
1589 while (map) {
1590 uint64_t rm1bit = rightmost_1bit(map);
1591 int i = raw_ctz(map);
1592
1593 if (mask_u64[i]) {
1594 mask_map |= rm1bit;
1595 *dst++ = mask_u64[i];
1596 hash = hash_add64(hash, mask_u64[i]);
1597 }
1598 map -= rm1bit;
1599 }
1600
1601 mask->mf.values_inline = true;
1602 mask->mf.map = mask_map;
1603
1604 hash = hash_add64(hash, mask_map);
1605
1606 n = dst - mask->mf.inline_values;
1607
1608 mask->hash = hash_finish(hash, n * 8);
1609 mask->len = netdev_flow_key_size(n);
1610 }
1611
1612 /* Initializes 'dst' as a copy of 'src' masked with 'mask'. */
1613 static inline void
1614 netdev_flow_key_init_masked(struct netdev_flow_key *dst,
1615 const struct flow *flow,
1616 const struct netdev_flow_key *mask)
1617 {
1618 uint64_t *dst_u64 = dst->mf.inline_values;
1619 const uint64_t *mask_u64 = mask->mf.inline_values;
1620 uint32_t hash = 0;
1621 uint64_t value;
1622
1623 dst->len = mask->len;
1624 dst->mf.values_inline = true;
1625 dst->mf.map = mask->mf.map;
1626
1627 FLOW_FOR_EACH_IN_MAP(value, flow, mask->mf.map) {
1628 *dst_u64 = value & *mask_u64++;
1629 hash = hash_add64(hash, *dst_u64++);
1630 }
1631 dst->hash = hash_finish(hash, (dst_u64 - dst->mf.inline_values) * 8);
1632 }
1633
1634 /* Iterate through all netdev_flow_key u64 values specified by 'MAP' */
1635 #define NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(VALUE, KEY, MAP) \
1636 for (struct mf_for_each_in_map_aux aux__ \
1637 = { (KEY)->mf.inline_values, (KEY)->mf.map, MAP }; \
1638 mf_get_next_in_map(&aux__, &(VALUE)); \
1639 )
1640
1641 /* Returns a hash value for the bits of 'key' where there are 1-bits in
1642 * 'mask'. */
1643 static inline uint32_t
1644 netdev_flow_key_hash_in_mask(const struct netdev_flow_key *key,
1645 const struct netdev_flow_key *mask)
1646 {
1647 const uint64_t *p = mask->mf.inline_values;
1648 uint32_t hash = 0;
1649 uint64_t key_u64;
1650
1651 NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(key_u64, key, mask->mf.map) {
1652 hash = hash_add64(hash, key_u64 & *p++);
1653 }
1654
1655 return hash_finish(hash, (p - mask->mf.inline_values) * 8);
1656 }
1657
1658 static inline bool
1659 emc_entry_alive(struct emc_entry *ce)
1660 {
1661 return ce->flow && !ce->flow->dead;
1662 }
1663
1664 static void
1665 emc_clear_entry(struct emc_entry *ce)
1666 {
1667 if (ce->flow) {
1668 dp_netdev_flow_unref(ce->flow);
1669 ce->flow = NULL;
1670 }
1671 }
1672
1673 static inline void
1674 emc_change_entry(struct emc_entry *ce, struct dp_netdev_flow *flow,
1675 const struct netdev_flow_key *key)
1676 {
1677 if (ce->flow != flow) {
1678 if (ce->flow) {
1679 dp_netdev_flow_unref(ce->flow);
1680 }
1681
1682 if (dp_netdev_flow_ref(flow)) {
1683 ce->flow = flow;
1684 } else {
1685 ce->flow = NULL;
1686 }
1687 }
1688 if (key) {
1689 netdev_flow_key_clone(&ce->key, key);
1690 }
1691 }
1692
1693 static inline void
1694 emc_insert(struct emc_cache *cache, const struct netdev_flow_key *key,
1695 struct dp_netdev_flow *flow)
1696 {
1697 struct emc_entry *to_be_replaced = NULL;
1698 struct emc_entry *current_entry;
1699
1700 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1701 if (netdev_flow_key_equal(&current_entry->key, key)) {
1702 /* We found the entry with the 'mf' miniflow */
1703 emc_change_entry(current_entry, flow, NULL);
1704 return;
1705 }
1706
1707 /* Replacement policy: put the flow in an empty (not alive) entry, or
1708 * in the first entry where it can be */
1709 if (!to_be_replaced
1710 || (emc_entry_alive(to_be_replaced)
1711 && !emc_entry_alive(current_entry))
1712 || current_entry->key.hash < to_be_replaced->key.hash) {
1713 to_be_replaced = current_entry;
1714 }
1715 }
1716 /* We didn't find the miniflow in the cache.
1717 * The 'to_be_replaced' entry is where the new flow will be stored */
1718
1719 emc_change_entry(to_be_replaced, flow, key);
1720 }
1721
1722 static inline struct dp_netdev_flow *
1723 emc_lookup(struct emc_cache *cache, const struct netdev_flow_key *key)
1724 {
1725 struct emc_entry *current_entry;
1726
1727 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1728 if (current_entry->key.hash == key->hash
1729 && emc_entry_alive(current_entry)
1730 && netdev_flow_key_equal_mf(&current_entry->key, &key->mf)) {
1731
1732 /* We found the entry with the 'key->mf' miniflow */
1733 return current_entry->flow;
1734 }
1735 }
1736
1737 return NULL;
1738 }
1739
1740 static struct dp_netdev_flow *
1741 dp_netdev_pmd_lookup_flow(const struct dp_netdev_pmd_thread *pmd,
1742 const struct netdev_flow_key *key)
1743 {
1744 struct dp_netdev_flow *netdev_flow;
1745 struct dpcls_rule *rule;
1746
1747 dpcls_lookup(&pmd->cls, key, &rule, 1);
1748 netdev_flow = dp_netdev_flow_cast(rule);
1749
1750 return netdev_flow;
1751 }
1752
1753 static struct dp_netdev_flow *
1754 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread *pmd,
1755 const ovs_u128 *ufidp, const struct nlattr *key,
1756 size_t key_len)
1757 {
1758 struct dp_netdev_flow *netdev_flow;
1759 struct flow flow;
1760 ovs_u128 ufid;
1761
1762 /* If a UFID is not provided, determine one based on the key. */
1763 if (!ufidp && key && key_len
1764 && !dpif_netdev_flow_from_nlattrs(key, key_len, &flow)) {
1765 dpif_flow_hash(pmd->dp->dpif, &flow, sizeof flow, &ufid);
1766 ufidp = &ufid;
1767 }
1768
1769 if (ufidp) {
1770 CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, dp_netdev_flow_hash(ufidp),
1771 &pmd->flow_table) {
1772 if (ovs_u128_equal(&netdev_flow->ufid, ufidp)) {
1773 return netdev_flow;
1774 }
1775 }
1776 }
1777
1778 return NULL;
1779 }
1780
1781 static void
1782 get_dpif_flow_stats(const struct dp_netdev_flow *netdev_flow_,
1783 struct dpif_flow_stats *stats)
1784 {
1785 struct dp_netdev_flow *netdev_flow;
1786 unsigned long long n;
1787 long long used;
1788 uint16_t flags;
1789
1790 netdev_flow = CONST_CAST(struct dp_netdev_flow *, netdev_flow_);
1791
1792 atomic_read_relaxed(&netdev_flow->stats.packet_count, &n);
1793 stats->n_packets = n;
1794 atomic_read_relaxed(&netdev_flow->stats.byte_count, &n);
1795 stats->n_bytes = n;
1796 atomic_read_relaxed(&netdev_flow->stats.used, &used);
1797 stats->used = used;
1798 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
1799 stats->tcp_flags = flags;
1800 }
1801
1802 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
1803 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
1804 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
1805 * protect them. */
1806 static void
1807 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow *netdev_flow,
1808 struct ofpbuf *key_buf, struct ofpbuf *mask_buf,
1809 struct dpif_flow *flow, bool terse)
1810 {
1811 if (terse) {
1812 memset(flow, 0, sizeof *flow);
1813 } else {
1814 struct flow_wildcards wc;
1815 struct dp_netdev_actions *actions;
1816 size_t offset;
1817
1818 miniflow_expand(&netdev_flow->cr.mask->mf, &wc.masks);
1819
1820 /* Key */
1821 offset = key_buf->size;
1822 flow->key = ofpbuf_tail(key_buf);
1823 odp_flow_key_from_flow(key_buf, &netdev_flow->flow, &wc.masks,
1824 netdev_flow->flow.in_port.odp_port, true);
1825 flow->key_len = key_buf->size - offset;
1826
1827 /* Mask */
1828 offset = mask_buf->size;
1829 flow->mask = ofpbuf_tail(mask_buf);
1830 odp_flow_key_from_mask(mask_buf, &wc.masks, &netdev_flow->flow,
1831 odp_to_u32(wc.masks.in_port.odp_port),
1832 SIZE_MAX, true);
1833 flow->mask_len = mask_buf->size - offset;
1834
1835 /* Actions */
1836 actions = dp_netdev_flow_get_actions(netdev_flow);
1837 flow->actions = actions->actions;
1838 flow->actions_len = actions->size;
1839 }
1840
1841 flow->ufid = netdev_flow->ufid;
1842 flow->ufid_present = true;
1843 flow->pmd_id = netdev_flow->pmd_id;
1844 get_dpif_flow_stats(netdev_flow, &flow->stats);
1845 }
1846
1847 static int
1848 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1849 const struct nlattr *mask_key,
1850 uint32_t mask_key_len, const struct flow *flow,
1851 struct flow *mask)
1852 {
1853 if (mask_key_len) {
1854 enum odp_key_fitness fitness;
1855
1856 fitness = odp_flow_key_to_mask(mask_key, mask_key_len, mask, flow);
1857 if (fitness) {
1858 /* This should not happen: it indicates that
1859 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1860 * disagree on the acceptable form of a mask. Log the problem
1861 * as an error, with enough details to enable debugging. */
1862 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1863
1864 if (!VLOG_DROP_ERR(&rl)) {
1865 struct ds s;
1866
1867 ds_init(&s);
1868 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1869 true);
1870 VLOG_ERR("internal error parsing flow mask %s (%s)",
1871 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1872 ds_destroy(&s);
1873 }
1874
1875 return EINVAL;
1876 }
1877 } else {
1878 enum mf_field_id id;
1879 /* No mask key, unwildcard everything except fields whose
1880 * prerequisities are not met. */
1881 memset(mask, 0x0, sizeof *mask);
1882
1883 for (id = 0; id < MFF_N_IDS; ++id) {
1884 /* Skip registers and metadata. */
1885 if (!(id >= MFF_REG0 && id < MFF_REG0 + FLOW_N_REGS)
1886 && id != MFF_METADATA) {
1887 const struct mf_field *mf = mf_from_id(id);
1888 if (mf_are_prereqs_ok(mf, flow)) {
1889 mf_mask_field(mf, mask);
1890 }
1891 }
1892 }
1893 }
1894
1895 /* Force unwildcard the in_port.
1896 *
1897 * We need to do this even in the case where we unwildcard "everything"
1898 * above because "everything" only includes the 16-bit OpenFlow port number
1899 * mask->in_port.ofp_port, which only covers half of the 32-bit datapath
1900 * port number mask->in_port.odp_port. */
1901 mask->in_port.odp_port = u32_to_odp(UINT32_MAX);
1902
1903 return 0;
1904 }
1905
1906 static int
1907 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1908 struct flow *flow)
1909 {
1910 odp_port_t in_port;
1911
1912 if (odp_flow_key_to_flow(key, key_len, flow)) {
1913 /* This should not happen: it indicates that odp_flow_key_from_flow()
1914 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1915 * flow. Log the problem as an error, with enough details to enable
1916 * debugging. */
1917 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1918
1919 if (!VLOG_DROP_ERR(&rl)) {
1920 struct ds s;
1921
1922 ds_init(&s);
1923 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1924 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1925 ds_destroy(&s);
1926 }
1927
1928 return EINVAL;
1929 }
1930
1931 in_port = flow->in_port.odp_port;
1932 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1933 return EINVAL;
1934 }
1935
1936 return 0;
1937 }
1938
1939 static int
1940 dpif_netdev_flow_get(const struct dpif *dpif, const struct dpif_flow_get *get)
1941 {
1942 struct dp_netdev *dp = get_dp_netdev(dpif);
1943 struct dp_netdev_flow *netdev_flow;
1944 struct dp_netdev_pmd_thread *pmd;
1945 int pmd_id = get->pmd_id == PMD_ID_NULL ? NON_PMD_CORE_ID : get->pmd_id;
1946 int error = 0;
1947
1948 pmd = dp_netdev_get_pmd(dp, pmd_id);
1949 if (!pmd) {
1950 return EINVAL;
1951 }
1952
1953 netdev_flow = dp_netdev_pmd_find_flow(pmd, get->ufid, get->key,
1954 get->key_len);
1955 if (netdev_flow) {
1956 dp_netdev_flow_to_dpif_flow(netdev_flow, get->buffer, get->buffer,
1957 get->flow, false);
1958 } else {
1959 error = ENOENT;
1960 }
1961 dp_netdev_pmd_unref(pmd);
1962
1963
1964 return error;
1965 }
1966
1967 static struct dp_netdev_flow *
1968 dp_netdev_flow_add(struct dp_netdev_pmd_thread *pmd,
1969 struct match *match, const ovs_u128 *ufid,
1970 const struct nlattr *actions, size_t actions_len)
1971 OVS_REQUIRES(pmd->flow_mutex)
1972 {
1973 struct dp_netdev_flow *flow;
1974 struct netdev_flow_key mask;
1975
1976 netdev_flow_mask_init(&mask, match);
1977 /* Make sure wc does not have metadata. */
1978 ovs_assert(!(mask.mf.map & (MINIFLOW_MAP(metadata) | MINIFLOW_MAP(regs))));
1979
1980 /* Do not allocate extra space. */
1981 flow = xmalloc(sizeof *flow - sizeof flow->cr.flow.mf + mask.len);
1982 memset(&flow->stats, 0, sizeof flow->stats);
1983 flow->dead = false;
1984 flow->batch = NULL;
1985 *CONST_CAST(int *, &flow->pmd_id) = pmd->core_id;
1986 *CONST_CAST(struct flow *, &flow->flow) = match->flow;
1987 *CONST_CAST(ovs_u128 *, &flow->ufid) = *ufid;
1988 ovs_refcount_init(&flow->ref_cnt);
1989 ovsrcu_set(&flow->actions, dp_netdev_actions_create(actions, actions_len));
1990
1991 netdev_flow_key_init_masked(&flow->cr.flow, &match->flow, &mask);
1992 dpcls_insert(&pmd->cls, &flow->cr, &mask);
1993
1994 cmap_insert(&pmd->flow_table, CONST_CAST(struct cmap_node *, &flow->node),
1995 dp_netdev_flow_hash(&flow->ufid));
1996
1997 if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
1998 struct match match;
1999 struct ds ds = DS_EMPTY_INITIALIZER;
2000
2001 match.flow = flow->flow;
2002 miniflow_expand(&flow->cr.mask->mf, &match.wc.masks);
2003
2004 ds_put_cstr(&ds, "flow_add: ");
2005 odp_format_ufid(ufid, &ds);
2006 ds_put_cstr(&ds, " ");
2007 match_format(&match, &ds, OFP_DEFAULT_PRIORITY);
2008 ds_put_cstr(&ds, ", actions:");
2009 format_odp_actions(&ds, actions, actions_len);
2010
2011 VLOG_DBG_RL(&upcall_rl, "%s", ds_cstr(&ds));
2012
2013 ds_destroy(&ds);
2014 }
2015
2016 return flow;
2017 }
2018
2019 static int
2020 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
2021 {
2022 struct dp_netdev *dp = get_dp_netdev(dpif);
2023 struct dp_netdev_flow *netdev_flow;
2024 struct netdev_flow_key key;
2025 struct dp_netdev_pmd_thread *pmd;
2026 struct match match;
2027 ovs_u128 ufid;
2028 int pmd_id = put->pmd_id == PMD_ID_NULL ? NON_PMD_CORE_ID : put->pmd_id;
2029 int error;
2030
2031 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow);
2032 if (error) {
2033 return error;
2034 }
2035 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
2036 put->mask, put->mask_len,
2037 &match.flow, &match.wc.masks);
2038 if (error) {
2039 return error;
2040 }
2041
2042 pmd = dp_netdev_get_pmd(dp, pmd_id);
2043 if (!pmd) {
2044 return EINVAL;
2045 }
2046
2047 /* Must produce a netdev_flow_key for lookup.
2048 * This interface is no longer performance critical, since it is not used
2049 * for upcall processing any more. */
2050 netdev_flow_key_from_flow(&key, &match.flow);
2051
2052 if (put->ufid) {
2053 ufid = *put->ufid;
2054 } else {
2055 dpif_flow_hash(dpif, &match.flow, sizeof match.flow, &ufid);
2056 }
2057
2058 ovs_mutex_lock(&pmd->flow_mutex);
2059 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &key);
2060 if (!netdev_flow) {
2061 if (put->flags & DPIF_FP_CREATE) {
2062 if (cmap_count(&pmd->flow_table) < MAX_FLOWS) {
2063 if (put->stats) {
2064 memset(put->stats, 0, sizeof *put->stats);
2065 }
2066 dp_netdev_flow_add(pmd, &match, &ufid, put->actions,
2067 put->actions_len);
2068 error = 0;
2069 } else {
2070 error = EFBIG;
2071 }
2072 } else {
2073 error = ENOENT;
2074 }
2075 } else {
2076 if (put->flags & DPIF_FP_MODIFY
2077 && flow_equal(&match.flow, &netdev_flow->flow)) {
2078 struct dp_netdev_actions *new_actions;
2079 struct dp_netdev_actions *old_actions;
2080
2081 new_actions = dp_netdev_actions_create(put->actions,
2082 put->actions_len);
2083
2084 old_actions = dp_netdev_flow_get_actions(netdev_flow);
2085 ovsrcu_set(&netdev_flow->actions, new_actions);
2086
2087 if (put->stats) {
2088 get_dpif_flow_stats(netdev_flow, put->stats);
2089 }
2090 if (put->flags & DPIF_FP_ZERO_STATS) {
2091 /* XXX: The userspace datapath uses thread local statistics
2092 * (for flows), which should be updated only by the owning
2093 * thread. Since we cannot write on stats memory here,
2094 * we choose not to support this flag. Please note:
2095 * - This feature is currently used only by dpctl commands with
2096 * option --clear.
2097 * - Should the need arise, this operation can be implemented
2098 * by keeping a base value (to be update here) for each
2099 * counter, and subtracting it before outputting the stats */
2100 error = EOPNOTSUPP;
2101 }
2102
2103 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
2104 } else if (put->flags & DPIF_FP_CREATE) {
2105 error = EEXIST;
2106 } else {
2107 /* Overlapping flow. */
2108 error = EINVAL;
2109 }
2110 }
2111 ovs_mutex_unlock(&pmd->flow_mutex);
2112 dp_netdev_pmd_unref(pmd);
2113
2114 return error;
2115 }
2116
2117 static int
2118 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
2119 {
2120 struct dp_netdev *dp = get_dp_netdev(dpif);
2121 struct dp_netdev_flow *netdev_flow;
2122 struct dp_netdev_pmd_thread *pmd;
2123 int pmd_id = del->pmd_id == PMD_ID_NULL ? NON_PMD_CORE_ID : del->pmd_id;
2124 int error = 0;
2125
2126 pmd = dp_netdev_get_pmd(dp, pmd_id);
2127 if (!pmd) {
2128 return EINVAL;
2129 }
2130
2131 ovs_mutex_lock(&pmd->flow_mutex);
2132 netdev_flow = dp_netdev_pmd_find_flow(pmd, del->ufid, del->key,
2133 del->key_len);
2134 if (netdev_flow) {
2135 if (del->stats) {
2136 get_dpif_flow_stats(netdev_flow, del->stats);
2137 }
2138 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
2139 } else {
2140 error = ENOENT;
2141 }
2142 ovs_mutex_unlock(&pmd->flow_mutex);
2143 dp_netdev_pmd_unref(pmd);
2144
2145 return error;
2146 }
2147
2148 struct dpif_netdev_flow_dump {
2149 struct dpif_flow_dump up;
2150 struct cmap_position poll_thread_pos;
2151 struct cmap_position flow_pos;
2152 struct dp_netdev_pmd_thread *cur_pmd;
2153 int status;
2154 struct ovs_mutex mutex;
2155 };
2156
2157 static struct dpif_netdev_flow_dump *
2158 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
2159 {
2160 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
2161 }
2162
2163 static struct dpif_flow_dump *
2164 dpif_netdev_flow_dump_create(const struct dpif *dpif_, bool terse)
2165 {
2166 struct dpif_netdev_flow_dump *dump;
2167
2168 dump = xzalloc(sizeof *dump);
2169 dpif_flow_dump_init(&dump->up, dpif_);
2170 dump->up.terse = terse;
2171 ovs_mutex_init(&dump->mutex);
2172
2173 return &dump->up;
2174 }
2175
2176 static int
2177 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
2178 {
2179 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2180
2181 ovs_mutex_destroy(&dump->mutex);
2182 free(dump);
2183 return 0;
2184 }
2185
2186 struct dpif_netdev_flow_dump_thread {
2187 struct dpif_flow_dump_thread up;
2188 struct dpif_netdev_flow_dump *dump;
2189 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
2190 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
2191 };
2192
2193 static struct dpif_netdev_flow_dump_thread *
2194 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
2195 {
2196 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
2197 }
2198
2199 static struct dpif_flow_dump_thread *
2200 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
2201 {
2202 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2203 struct dpif_netdev_flow_dump_thread *thread;
2204
2205 thread = xmalloc(sizeof *thread);
2206 dpif_flow_dump_thread_init(&thread->up, &dump->up);
2207 thread->dump = dump;
2208 return &thread->up;
2209 }
2210
2211 static void
2212 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
2213 {
2214 struct dpif_netdev_flow_dump_thread *thread
2215 = dpif_netdev_flow_dump_thread_cast(thread_);
2216
2217 free(thread);
2218 }
2219
2220 static int
2221 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
2222 struct dpif_flow *flows, int max_flows)
2223 {
2224 struct dpif_netdev_flow_dump_thread *thread
2225 = dpif_netdev_flow_dump_thread_cast(thread_);
2226 struct dpif_netdev_flow_dump *dump = thread->dump;
2227 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
2228 int n_flows = 0;
2229 int i;
2230
2231 ovs_mutex_lock(&dump->mutex);
2232 if (!dump->status) {
2233 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
2234 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
2235 struct dp_netdev_pmd_thread *pmd = dump->cur_pmd;
2236 int flow_limit = MIN(max_flows, FLOW_DUMP_MAX_BATCH);
2237
2238 /* First call to dump_next(), extracts the first pmd thread.
2239 * If there is no pmd thread, returns immediately. */
2240 if (!pmd) {
2241 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2242 if (!pmd) {
2243 ovs_mutex_unlock(&dump->mutex);
2244 return n_flows;
2245
2246 }
2247 }
2248
2249 do {
2250 for (n_flows = 0; n_flows < flow_limit; n_flows++) {
2251 struct cmap_node *node;
2252
2253 node = cmap_next_position(&pmd->flow_table, &dump->flow_pos);
2254 if (!node) {
2255 break;
2256 }
2257 netdev_flows[n_flows] = CONTAINER_OF(node,
2258 struct dp_netdev_flow,
2259 node);
2260 }
2261 /* When finishing dumping the current pmd thread, moves to
2262 * the next. */
2263 if (n_flows < flow_limit) {
2264 memset(&dump->flow_pos, 0, sizeof dump->flow_pos);
2265 dp_netdev_pmd_unref(pmd);
2266 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2267 if (!pmd) {
2268 dump->status = EOF;
2269 break;
2270 }
2271 }
2272 /* Keeps the reference to next caller. */
2273 dump->cur_pmd = pmd;
2274
2275 /* If the current dump is empty, do not exit the loop, since the
2276 * remaining pmds could have flows to be dumped. Just dumps again
2277 * on the new 'pmd'. */
2278 } while (!n_flows);
2279 }
2280 ovs_mutex_unlock(&dump->mutex);
2281
2282 for (i = 0; i < n_flows; i++) {
2283 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
2284 struct odputil_keybuf *keybuf = &thread->keybuf[i];
2285 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
2286 struct dpif_flow *f = &flows[i];
2287 struct ofpbuf key, mask;
2288
2289 ofpbuf_use_stack(&key, keybuf, sizeof *keybuf);
2290 ofpbuf_use_stack(&mask, maskbuf, sizeof *maskbuf);
2291 dp_netdev_flow_to_dpif_flow(netdev_flow, &key, &mask, f,
2292 dump->up.terse);
2293 }
2294
2295 return n_flows;
2296 }
2297
2298 static int
2299 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
2300 OVS_NO_THREAD_SAFETY_ANALYSIS
2301 {
2302 struct dp_netdev *dp = get_dp_netdev(dpif);
2303 struct dp_netdev_pmd_thread *pmd;
2304 struct dp_packet *pp;
2305
2306 if (dp_packet_size(execute->packet) < ETH_HEADER_LEN ||
2307 dp_packet_size(execute->packet) > UINT16_MAX) {
2308 return EINVAL;
2309 }
2310
2311 /* Tries finding the 'pmd'. If NULL is returned, that means
2312 * the current thread is a non-pmd thread and should use
2313 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
2314 pmd = ovsthread_getspecific(dp->per_pmd_key);
2315 if (!pmd) {
2316 pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID);
2317 }
2318
2319 /* If the current thread is non-pmd thread, acquires
2320 * the 'non_pmd_mutex'. */
2321 if (pmd->core_id == NON_PMD_CORE_ID) {
2322 ovs_mutex_lock(&dp->non_pmd_mutex);
2323 ovs_mutex_lock(&dp->port_mutex);
2324 }
2325
2326 pp = execute->packet;
2327 dp_netdev_execute_actions(pmd, &pp, 1, false, execute->actions,
2328 execute->actions_len);
2329 if (pmd->core_id == NON_PMD_CORE_ID) {
2330 dp_netdev_pmd_unref(pmd);
2331 ovs_mutex_unlock(&dp->port_mutex);
2332 ovs_mutex_unlock(&dp->non_pmd_mutex);
2333 }
2334
2335 return 0;
2336 }
2337
2338 static void
2339 dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
2340 {
2341 size_t i;
2342
2343 for (i = 0; i < n_ops; i++) {
2344 struct dpif_op *op = ops[i];
2345
2346 switch (op->type) {
2347 case DPIF_OP_FLOW_PUT:
2348 op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put);
2349 break;
2350
2351 case DPIF_OP_FLOW_DEL:
2352 op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del);
2353 break;
2354
2355 case DPIF_OP_EXECUTE:
2356 op->error = dpif_netdev_execute(dpif, &op->u.execute);
2357 break;
2358
2359 case DPIF_OP_FLOW_GET:
2360 op->error = dpif_netdev_flow_get(dpif, &op->u.flow_get);
2361 break;
2362 }
2363 }
2364 }
2365
2366 /* Returns true if the configuration for rx queues or cpu mask
2367 * is changed. */
2368 static bool
2369 pmd_config_changed(const struct dp_netdev *dp, size_t rxqs, const char *cmask)
2370 {
2371 if (dp->n_dpdk_rxqs != rxqs) {
2372 return true;
2373 } else {
2374 if (dp->pmd_cmask != NULL && cmask != NULL) {
2375 return strcmp(dp->pmd_cmask, cmask);
2376 } else {
2377 return (dp->pmd_cmask != NULL || cmask != NULL);
2378 }
2379 }
2380 }
2381
2382 /* Resets pmd threads if the configuration for 'rxq's or cpu mask changes. */
2383 static int
2384 dpif_netdev_pmd_set(struct dpif *dpif, unsigned int n_rxqs, const char *cmask)
2385 {
2386 struct dp_netdev *dp = get_dp_netdev(dpif);
2387
2388 if (pmd_config_changed(dp, n_rxqs, cmask)) {
2389 struct dp_netdev_port *port;
2390
2391 dp_netdev_destroy_all_pmds(dp);
2392
2393 CMAP_FOR_EACH (port, node, &dp->ports) {
2394 if (netdev_is_pmd(port->netdev)) {
2395 int i, err;
2396
2397 /* Closes the existing 'rxq's. */
2398 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2399 netdev_rxq_close(port->rxq[i]);
2400 port->rxq[i] = NULL;
2401 }
2402
2403 /* Sets the new rx queue config. */
2404 err = netdev_set_multiq(port->netdev, ovs_numa_get_n_cores(),
2405 n_rxqs);
2406 if (err && (err != EOPNOTSUPP)) {
2407 VLOG_ERR("Failed to set dpdk interface %s rx_queue to:"
2408 " %u", netdev_get_name(port->netdev),
2409 n_rxqs);
2410 return err;
2411 }
2412
2413 /* If the set_multiq() above succeeds, reopens the 'rxq's. */
2414 port->rxq = xrealloc(port->rxq, sizeof *port->rxq
2415 * netdev_n_rxq(port->netdev));
2416 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2417 netdev_rxq_open(port->netdev, &port->rxq[i], i);
2418 }
2419 }
2420 }
2421 dp->n_dpdk_rxqs = n_rxqs;
2422
2423 /* Reconfigures the cpu mask. */
2424 ovs_numa_set_cpu_mask(cmask);
2425 free(dp->pmd_cmask);
2426 dp->pmd_cmask = cmask ? xstrdup(cmask) : NULL;
2427
2428 /* Restores the non-pmd. */
2429 dp_netdev_set_nonpmd(dp);
2430 /* Restores all pmd threads. */
2431 dp_netdev_reset_pmd_threads(dp);
2432 }
2433
2434 return 0;
2435 }
2436
2437 static int
2438 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
2439 uint32_t queue_id, uint32_t *priority)
2440 {
2441 *priority = queue_id;
2442 return 0;
2443 }
2444
2445 \f
2446 /* Creates and returns a new 'struct dp_netdev_actions', whose actions are
2447 * a copy of the 'ofpacts_len' bytes of 'ofpacts'. */
2448 struct dp_netdev_actions *
2449 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
2450 {
2451 struct dp_netdev_actions *netdev_actions;
2452
2453 netdev_actions = xmalloc(sizeof *netdev_actions + size);
2454 memcpy(netdev_actions->actions, actions, size);
2455 netdev_actions->size = size;
2456
2457 return netdev_actions;
2458 }
2459
2460 struct dp_netdev_actions *
2461 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
2462 {
2463 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
2464 }
2465
2466 static void
2467 dp_netdev_actions_free(struct dp_netdev_actions *actions)
2468 {
2469 free(actions);
2470 }
2471 \f
2472 static inline unsigned long long
2473 cycles_counter(void)
2474 {
2475 #ifdef DPDK_NETDEV
2476 return rte_get_tsc_cycles();
2477 #else
2478 return 0;
2479 #endif
2480 }
2481
2482 /* Fake mutex to make sure that the calls to cycles_count_* are balanced */
2483 extern struct ovs_mutex cycles_counter_fake_mutex;
2484
2485 /* Start counting cycles. Must be followed by 'cycles_count_end()' */
2486 static inline void
2487 cycles_count_start(struct dp_netdev_pmd_thread *pmd)
2488 OVS_ACQUIRES(&cycles_counter_fake_mutex)
2489 OVS_NO_THREAD_SAFETY_ANALYSIS
2490 {
2491 pmd->last_cycles = cycles_counter();
2492 }
2493
2494 /* Stop counting cycles and add them to the counter 'type' */
2495 static inline void
2496 cycles_count_end(struct dp_netdev_pmd_thread *pmd,
2497 enum pmd_cycles_counter_type type)
2498 OVS_RELEASES(&cycles_counter_fake_mutex)
2499 OVS_NO_THREAD_SAFETY_ANALYSIS
2500 {
2501 unsigned long long interval = cycles_counter() - pmd->last_cycles;
2502
2503 non_atomic_ullong_add(&pmd->cycles.n[type], interval);
2504 }
2505
2506 static void
2507 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd,
2508 struct dp_netdev_port *port,
2509 struct netdev_rxq *rxq)
2510 {
2511 struct dp_packet *packets[NETDEV_MAX_BURST];
2512 int error, cnt;
2513
2514 cycles_count_start(pmd);
2515 error = netdev_rxq_recv(rxq, packets, &cnt);
2516 cycles_count_end(pmd, PMD_CYCLES_POLLING);
2517 if (!error) {
2518 int i;
2519
2520 *recirc_depth_get() = 0;
2521
2522 /* XXX: initialize md in netdev implementation. */
2523 for (i = 0; i < cnt; i++) {
2524 packets[i]->md = port->md;
2525 }
2526 cycles_count_start(pmd);
2527 dp_netdev_input(pmd, packets, cnt);
2528 cycles_count_end(pmd, PMD_CYCLES_PROCESSING);
2529 } else if (error != EAGAIN && error != EOPNOTSUPP) {
2530 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
2531
2532 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
2533 netdev_get_name(port->netdev), ovs_strerror(error));
2534 }
2535 }
2536
2537 /* Return true if needs to revalidate datapath flows. */
2538 static bool
2539 dpif_netdev_run(struct dpif *dpif)
2540 {
2541 struct dp_netdev_port *port;
2542 struct dp_netdev *dp = get_dp_netdev(dpif);
2543 struct dp_netdev_pmd_thread *non_pmd = dp_netdev_get_pmd(dp,
2544 NON_PMD_CORE_ID);
2545 uint64_t new_tnl_seq;
2546
2547 ovs_mutex_lock(&dp->non_pmd_mutex);
2548 CMAP_FOR_EACH (port, node, &dp->ports) {
2549 if (!netdev_is_pmd(port->netdev)) {
2550 int i;
2551
2552 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2553 dp_netdev_process_rxq_port(non_pmd, port, port->rxq[i]);
2554 }
2555 }
2556 }
2557 ovs_mutex_unlock(&dp->non_pmd_mutex);
2558 dp_netdev_pmd_unref(non_pmd);
2559
2560 tnl_arp_cache_run();
2561 new_tnl_seq = seq_read(tnl_conf_seq);
2562
2563 if (dp->last_tnl_conf_seq != new_tnl_seq) {
2564 dp->last_tnl_conf_seq = new_tnl_seq;
2565 return true;
2566 }
2567 return false;
2568 }
2569
2570 static void
2571 dpif_netdev_wait(struct dpif *dpif)
2572 {
2573 struct dp_netdev_port *port;
2574 struct dp_netdev *dp = get_dp_netdev(dpif);
2575
2576 ovs_mutex_lock(&dp_netdev_mutex);
2577 CMAP_FOR_EACH (port, node, &dp->ports) {
2578 if (!netdev_is_pmd(port->netdev)) {
2579 int i;
2580
2581 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2582 netdev_rxq_wait(port->rxq[i]);
2583 }
2584 }
2585 }
2586 ovs_mutex_unlock(&dp_netdev_mutex);
2587 seq_wait(tnl_conf_seq, dp->last_tnl_conf_seq);
2588 }
2589
2590 struct rxq_poll {
2591 struct dp_netdev_port *port;
2592 struct netdev_rxq *rx;
2593 };
2594
2595 static int
2596 pmd_load_queues(struct dp_netdev_pmd_thread *pmd,
2597 struct rxq_poll **ppoll_list, int poll_cnt)
2598 {
2599 struct rxq_poll *poll_list = *ppoll_list;
2600 struct dp_netdev_port *port;
2601 int n_pmds_on_numa, index, i;
2602
2603 /* Simple scheduler for netdev rx polling. */
2604 for (i = 0; i < poll_cnt; i++) {
2605 port_unref(poll_list[i].port);
2606 }
2607
2608 poll_cnt = 0;
2609 n_pmds_on_numa = get_n_pmd_threads_on_numa(pmd->dp, pmd->numa_id);
2610 index = 0;
2611
2612 CMAP_FOR_EACH (port, node, &pmd->dp->ports) {
2613 /* Calls port_try_ref() to prevent the main thread
2614 * from deleting the port. */
2615 if (port_try_ref(port)) {
2616 if (netdev_is_pmd(port->netdev)
2617 && netdev_get_numa_id(port->netdev) == pmd->numa_id) {
2618 int i;
2619
2620 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2621 if ((index % n_pmds_on_numa) == pmd->index) {
2622 poll_list = xrealloc(poll_list,
2623 sizeof *poll_list * (poll_cnt + 1));
2624
2625 port_ref(port);
2626 poll_list[poll_cnt].port = port;
2627 poll_list[poll_cnt].rx = port->rxq[i];
2628 poll_cnt++;
2629 }
2630 index++;
2631 }
2632 }
2633 /* Unrefs the port_try_ref(). */
2634 port_unref(port);
2635 }
2636 }
2637
2638 *ppoll_list = poll_list;
2639 return poll_cnt;
2640 }
2641
2642 static void *
2643 pmd_thread_main(void *f_)
2644 {
2645 struct dp_netdev_pmd_thread *pmd = f_;
2646 unsigned int lc = 0;
2647 struct rxq_poll *poll_list;
2648 unsigned int port_seq = PMD_INITIAL_SEQ;
2649 int poll_cnt;
2650 int i;
2651
2652 poll_cnt = 0;
2653 poll_list = NULL;
2654
2655 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
2656 ovsthread_setspecific(pmd->dp->per_pmd_key, pmd);
2657 pmd_thread_setaffinity_cpu(pmd->core_id);
2658 reload:
2659 emc_cache_init(&pmd->flow_cache);
2660 poll_cnt = pmd_load_queues(pmd, &poll_list, poll_cnt);
2661
2662 /* Signal here to make sure the pmd finishes
2663 * reloading the updated configuration. */
2664 dp_netdev_pmd_reload_done(pmd);
2665
2666 for (;;) {
2667 int i;
2668
2669 for (i = 0; i < poll_cnt; i++) {
2670 dp_netdev_process_rxq_port(pmd, poll_list[i].port, poll_list[i].rx);
2671 }
2672
2673 if (lc++ > 1024) {
2674 unsigned int seq;
2675
2676 lc = 0;
2677
2678 emc_cache_slow_sweep(&pmd->flow_cache);
2679 ovsrcu_quiesce();
2680
2681 atomic_read_relaxed(&pmd->change_seq, &seq);
2682 if (seq != port_seq) {
2683 port_seq = seq;
2684 break;
2685 }
2686 }
2687 }
2688
2689 emc_cache_uninit(&pmd->flow_cache);
2690
2691 if (!latch_is_set(&pmd->exit_latch)){
2692 goto reload;
2693 }
2694
2695 for (i = 0; i < poll_cnt; i++) {
2696 port_unref(poll_list[i].port);
2697 }
2698
2699 dp_netdev_pmd_reload_done(pmd);
2700
2701 free(poll_list);
2702 return NULL;
2703 }
2704
2705 static void
2706 dp_netdev_disable_upcall(struct dp_netdev *dp)
2707 OVS_ACQUIRES(dp->upcall_rwlock)
2708 {
2709 fat_rwlock_wrlock(&dp->upcall_rwlock);
2710 }
2711
2712 static void
2713 dpif_netdev_disable_upcall(struct dpif *dpif)
2714 OVS_NO_THREAD_SAFETY_ANALYSIS
2715 {
2716 struct dp_netdev *dp = get_dp_netdev(dpif);
2717 dp_netdev_disable_upcall(dp);
2718 }
2719
2720 static void
2721 dp_netdev_enable_upcall(struct dp_netdev *dp)
2722 OVS_RELEASES(dp->upcall_rwlock)
2723 {
2724 fat_rwlock_unlock(&dp->upcall_rwlock);
2725 }
2726
2727 static void
2728 dpif_netdev_enable_upcall(struct dpif *dpif)
2729 OVS_NO_THREAD_SAFETY_ANALYSIS
2730 {
2731 struct dp_netdev *dp = get_dp_netdev(dpif);
2732 dp_netdev_enable_upcall(dp);
2733 }
2734
2735 void
2736 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd)
2737 {
2738 ovs_mutex_lock(&pmd->cond_mutex);
2739 xpthread_cond_signal(&pmd->cond);
2740 ovs_mutex_unlock(&pmd->cond_mutex);
2741 }
2742
2743 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
2744 * the pointer if succeeds, otherwise, NULL.
2745 *
2746 * Caller must unrefs the returned reference. */
2747 static struct dp_netdev_pmd_thread *
2748 dp_netdev_get_pmd(struct dp_netdev *dp, int core_id)
2749 {
2750 struct dp_netdev_pmd_thread *pmd;
2751 const struct cmap_node *pnode;
2752
2753 pnode = cmap_find(&dp->poll_threads, hash_int(core_id, 0));
2754 if (!pnode) {
2755 return NULL;
2756 }
2757 pmd = CONTAINER_OF(pnode, struct dp_netdev_pmd_thread, node);
2758
2759 return dp_netdev_pmd_try_ref(pmd) ? pmd : NULL;
2760 }
2761
2762 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
2763 static void
2764 dp_netdev_set_nonpmd(struct dp_netdev *dp)
2765 {
2766 struct dp_netdev_pmd_thread *non_pmd;
2767
2768 non_pmd = xzalloc(sizeof *non_pmd);
2769 dp_netdev_configure_pmd(non_pmd, dp, 0, NON_PMD_CORE_ID,
2770 OVS_NUMA_UNSPEC);
2771 }
2772
2773 /* Caller must have valid pointer to 'pmd'. */
2774 static bool
2775 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd)
2776 {
2777 return ovs_refcount_try_ref_rcu(&pmd->ref_cnt);
2778 }
2779
2780 static void
2781 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd)
2782 {
2783 if (pmd && ovs_refcount_unref(&pmd->ref_cnt) == 1) {
2784 ovsrcu_postpone(dp_netdev_destroy_pmd, pmd);
2785 }
2786 }
2787
2788 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
2789 * fails, keeps checking for next node until reaching the end of cmap.
2790 *
2791 * Caller must unrefs the returned reference. */
2792 static struct dp_netdev_pmd_thread *
2793 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos)
2794 {
2795 struct dp_netdev_pmd_thread *next;
2796
2797 do {
2798 struct cmap_node *node;
2799
2800 node = cmap_next_position(&dp->poll_threads, pos);
2801 next = node ? CONTAINER_OF(node, struct dp_netdev_pmd_thread, node)
2802 : NULL;
2803 } while (next && !dp_netdev_pmd_try_ref(next));
2804
2805 return next;
2806 }
2807
2808 /* Configures the 'pmd' based on the input argument. */
2809 static void
2810 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp,
2811 int index, int core_id, int numa_id)
2812 {
2813 pmd->dp = dp;
2814 pmd->index = index;
2815 pmd->core_id = core_id;
2816 pmd->numa_id = numa_id;
2817
2818 ovs_refcount_init(&pmd->ref_cnt);
2819 latch_init(&pmd->exit_latch);
2820 atomic_init(&pmd->change_seq, PMD_INITIAL_SEQ);
2821 xpthread_cond_init(&pmd->cond, NULL);
2822 ovs_mutex_init(&pmd->cond_mutex);
2823 ovs_mutex_init(&pmd->flow_mutex);
2824 dpcls_init(&pmd->cls);
2825 cmap_init(&pmd->flow_table);
2826 /* init the 'flow_cache' since there is no
2827 * actual thread created for NON_PMD_CORE_ID. */
2828 if (core_id == NON_PMD_CORE_ID) {
2829 emc_cache_init(&pmd->flow_cache);
2830 }
2831 cmap_insert(&dp->poll_threads, CONST_CAST(struct cmap_node *, &pmd->node),
2832 hash_int(core_id, 0));
2833 }
2834
2835 static void
2836 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd)
2837 {
2838 dp_netdev_pmd_flow_flush(pmd);
2839 dpcls_destroy(&pmd->cls);
2840 cmap_destroy(&pmd->flow_table);
2841 ovs_mutex_destroy(&pmd->flow_mutex);
2842 latch_destroy(&pmd->exit_latch);
2843 xpthread_cond_destroy(&pmd->cond);
2844 ovs_mutex_destroy(&pmd->cond_mutex);
2845 free(pmd);
2846 }
2847
2848 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
2849 * and unrefs the struct. */
2850 static void
2851 dp_netdev_del_pmd(struct dp_netdev_pmd_thread *pmd)
2852 {
2853 /* Uninit the 'flow_cache' since there is
2854 * no actual thread uninit it for NON_PMD_CORE_ID. */
2855 if (pmd->core_id == NON_PMD_CORE_ID) {
2856 emc_cache_uninit(&pmd->flow_cache);
2857 } else {
2858 latch_set(&pmd->exit_latch);
2859 dp_netdev_reload_pmd__(pmd);
2860 ovs_numa_unpin_core(pmd->core_id);
2861 xpthread_join(pmd->thread, NULL);
2862 }
2863 cmap_remove(&pmd->dp->poll_threads, &pmd->node, hash_int(pmd->core_id, 0));
2864 dp_netdev_pmd_unref(pmd);
2865 }
2866
2867 /* Destroys all pmd threads. */
2868 static void
2869 dp_netdev_destroy_all_pmds(struct dp_netdev *dp)
2870 {
2871 struct dp_netdev_pmd_thread *pmd;
2872
2873 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2874 dp_netdev_del_pmd(pmd);
2875 }
2876 }
2877
2878 /* Deletes all pmd threads on numa node 'numa_id'. */
2879 static void
2880 dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2881 {
2882 struct dp_netdev_pmd_thread *pmd;
2883
2884 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2885 if (pmd->numa_id == numa_id) {
2886 dp_netdev_del_pmd(pmd);
2887 }
2888 }
2889 }
2890
2891 /* Checks the numa node id of 'netdev' and starts pmd threads for
2892 * the numa node. */
2893 static void
2894 dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2895 {
2896 int n_pmds;
2897
2898 if (!ovs_numa_numa_id_is_valid(numa_id)) {
2899 VLOG_ERR("Cannot create pmd threads due to numa id (%d)"
2900 "invalid", numa_id);
2901 return ;
2902 }
2903
2904 n_pmds = get_n_pmd_threads_on_numa(dp, numa_id);
2905
2906 /* If there are already pmd threads created for the numa node
2907 * in which 'netdev' is on, do nothing. Else, creates the
2908 * pmd threads for the numa node. */
2909 if (!n_pmds) {
2910 int can_have, n_unpinned, i;
2911
2912 n_unpinned = ovs_numa_get_n_unpinned_cores_on_numa(numa_id);
2913 if (!n_unpinned) {
2914 VLOG_ERR("Cannot create pmd threads due to out of unpinned "
2915 "cores on numa node");
2916 return;
2917 }
2918
2919 /* If cpu mask is specified, uses all unpinned cores, otherwise
2920 * tries creating NR_PMD_THREADS pmd threads. */
2921 can_have = dp->pmd_cmask ? n_unpinned : MIN(n_unpinned, NR_PMD_THREADS);
2922 for (i = 0; i < can_have; i++) {
2923 struct dp_netdev_pmd_thread *pmd = xzalloc(sizeof *pmd);
2924 int core_id = ovs_numa_get_unpinned_core_on_numa(numa_id);
2925
2926 dp_netdev_configure_pmd(pmd, dp, i, core_id, numa_id);
2927 /* Each thread will distribute all devices rx-queues among
2928 * themselves. */
2929 pmd->thread = ovs_thread_create("pmd", pmd_thread_main, pmd);
2930 }
2931 VLOG_INFO("Created %d pmd threads on numa node %d", can_have, numa_id);
2932 }
2933 }
2934
2935 \f
2936 /* Called after pmd threads config change. Restarts pmd threads with
2937 * new configuration. */
2938 static void
2939 dp_netdev_reset_pmd_threads(struct dp_netdev *dp)
2940 {
2941 struct dp_netdev_port *port;
2942
2943 CMAP_FOR_EACH (port, node, &dp->ports) {
2944 if (netdev_is_pmd(port->netdev)) {
2945 int numa_id = netdev_get_numa_id(port->netdev);
2946
2947 dp_netdev_set_pmds_on_numa(dp, numa_id);
2948 }
2949 }
2950 }
2951
2952 static char *
2953 dpif_netdev_get_datapath_version(void)
2954 {
2955 return xstrdup("<built-in>");
2956 }
2957
2958 static void
2959 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow, int cnt, int size,
2960 uint16_t tcp_flags, long long now)
2961 {
2962 uint16_t flags;
2963
2964 atomic_store_relaxed(&netdev_flow->stats.used, now);
2965 non_atomic_ullong_add(&netdev_flow->stats.packet_count, cnt);
2966 non_atomic_ullong_add(&netdev_flow->stats.byte_count, size);
2967 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
2968 flags |= tcp_flags;
2969 atomic_store_relaxed(&netdev_flow->stats.tcp_flags, flags);
2970 }
2971
2972 static void
2973 dp_netdev_count_packet(struct dp_netdev_pmd_thread *pmd,
2974 enum dp_stat_type type, int cnt)
2975 {
2976 non_atomic_ullong_add(&pmd->stats.n[type], cnt);
2977 }
2978
2979 static int
2980 dp_netdev_upcall(struct dp_netdev_pmd_thread *pmd, struct dp_packet *packet_,
2981 struct flow *flow, struct flow_wildcards *wc, ovs_u128 *ufid,
2982 enum dpif_upcall_type type, const struct nlattr *userdata,
2983 struct ofpbuf *actions, struct ofpbuf *put_actions)
2984 {
2985 struct dp_netdev *dp = pmd->dp;
2986
2987 if (OVS_UNLIKELY(!dp->upcall_cb)) {
2988 return ENODEV;
2989 }
2990
2991 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl))) {
2992 struct ds ds = DS_EMPTY_INITIALIZER;
2993 char *packet_str;
2994 struct ofpbuf key;
2995
2996 ofpbuf_init(&key, 0);
2997 odp_flow_key_from_flow(&key, flow, &wc->masks, flow->in_port.odp_port,
2998 true);
2999 packet_str = ofp_packet_to_string(dp_packet_data(packet_),
3000 dp_packet_size(packet_));
3001
3002 odp_flow_key_format(key.data, key.size, &ds);
3003
3004 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp->name,
3005 dpif_upcall_type_to_string(type), ds_cstr(&ds), packet_str);
3006
3007 ofpbuf_uninit(&key);
3008 free(packet_str);
3009
3010 ds_destroy(&ds);
3011 }
3012
3013 return dp->upcall_cb(packet_, flow, ufid, pmd->core_id, type, userdata,
3014 actions, wc, put_actions, dp->upcall_aux);
3015 }
3016
3017 static inline uint32_t
3018 dpif_netdev_packet_get_dp_hash(struct dp_packet *packet,
3019 const struct miniflow *mf)
3020 {
3021 uint32_t hash;
3022
3023 hash = dp_packet_get_rss_hash(packet);
3024 if (OVS_UNLIKELY(!hash)) {
3025 hash = miniflow_hash_5tuple(mf, 0);
3026 dp_packet_set_rss_hash(packet, hash);
3027 }
3028 return hash;
3029 }
3030
3031 struct packet_batch {
3032 unsigned int packet_count;
3033 unsigned int byte_count;
3034 uint16_t tcp_flags;
3035
3036 struct dp_netdev_flow *flow;
3037
3038 struct dp_packet *packets[NETDEV_MAX_BURST];
3039 };
3040
3041 static inline void
3042 packet_batch_update(struct packet_batch *batch, struct dp_packet *packet,
3043 const struct miniflow *mf)
3044 {
3045 batch->tcp_flags |= miniflow_get_tcp_flags(mf);
3046 batch->packets[batch->packet_count++] = packet;
3047 batch->byte_count += dp_packet_size(packet);
3048 }
3049
3050 static inline void
3051 packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow)
3052 {
3053 flow->batch = batch;
3054
3055 batch->flow = flow;
3056 batch->packet_count = 0;
3057 batch->byte_count = 0;
3058 batch->tcp_flags = 0;
3059 }
3060
3061 static inline void
3062 packet_batch_execute(struct packet_batch *batch,
3063 struct dp_netdev_pmd_thread *pmd,
3064 long long now)
3065 {
3066 struct dp_netdev_actions *actions;
3067 struct dp_netdev_flow *flow = batch->flow;
3068
3069 dp_netdev_flow_used(flow, batch->packet_count, batch->byte_count,
3070 batch->tcp_flags, now);
3071
3072 actions = dp_netdev_flow_get_actions(flow);
3073
3074 dp_netdev_execute_actions(pmd, batch->packets, batch->packet_count, true,
3075 actions->actions, actions->size);
3076 }
3077
3078 static inline void
3079 dp_netdev_queue_batches(struct dp_packet *pkt,
3080 struct dp_netdev_flow *flow, const struct miniflow *mf,
3081 struct packet_batch *batches, size_t *n_batches)
3082 {
3083 struct packet_batch *batch = flow->batch;
3084
3085 if (OVS_LIKELY(batch)) {
3086 packet_batch_update(batch, pkt, mf);
3087 return;
3088 }
3089
3090 batch = &batches[(*n_batches)++];
3091 packet_batch_init(batch, flow);
3092 packet_batch_update(batch, pkt, mf);
3093 }
3094
3095 static inline void
3096 dp_packet_swap(struct dp_packet **a, struct dp_packet **b)
3097 {
3098 struct dp_packet *tmp = *a;
3099 *a = *b;
3100 *b = tmp;
3101 }
3102
3103 /* Try to process all ('cnt') the 'packets' using only the exact match cache
3104 * 'flow_cache'. If a flow is not found for a packet 'packets[i]', the
3105 * miniflow is copied into 'keys' and the packet pointer is moved at the
3106 * beginning of the 'packets' array.
3107 *
3108 * The function returns the number of packets that needs to be processed in the
3109 * 'packets' array (they have been moved to the beginning of the vector).
3110 */
3111 static inline size_t
3112 emc_processing(struct dp_netdev_pmd_thread *pmd, struct dp_packet **packets,
3113 size_t cnt, struct netdev_flow_key *keys,
3114 struct packet_batch batches[], size_t *n_batches)
3115 {
3116 struct emc_cache *flow_cache = &pmd->flow_cache;
3117 struct netdev_flow_key key;
3118 size_t i, notfound_cnt = 0;
3119
3120 miniflow_initialize(&key.mf, key.buf);
3121 for (i = 0; i < cnt; i++) {
3122 struct dp_netdev_flow *flow;
3123
3124 if (OVS_UNLIKELY(dp_packet_size(packets[i]) < ETH_HEADER_LEN)) {
3125 dp_packet_delete(packets[i]);
3126 continue;
3127 }
3128
3129 miniflow_extract(packets[i], &key.mf);
3130 key.len = 0; /* Not computed yet. */
3131 key.hash = dpif_netdev_packet_get_dp_hash(packets[i], &key.mf);
3132
3133 flow = emc_lookup(flow_cache, &key);
3134 if (OVS_LIKELY(flow)) {
3135 dp_netdev_queue_batches(packets[i], flow, &key.mf, batches,
3136 n_batches);
3137 } else {
3138 if (i != notfound_cnt) {
3139 dp_packet_swap(&packets[i], &packets[notfound_cnt]);
3140 }
3141
3142 keys[notfound_cnt++] = key;
3143 }
3144 }
3145
3146 dp_netdev_count_packet(pmd, DP_STAT_EXACT_HIT, cnt - notfound_cnt);
3147
3148 return notfound_cnt;
3149 }
3150
3151 static inline void
3152 fast_path_processing(struct dp_netdev_pmd_thread *pmd,
3153 struct dp_packet **packets, size_t cnt,
3154 struct netdev_flow_key *keys,
3155 struct packet_batch batches[], size_t *n_batches)
3156 {
3157 #if !defined(__CHECKER__) && !defined(_WIN32)
3158 const size_t PKT_ARRAY_SIZE = cnt;
3159 #else
3160 /* Sparse or MSVC doesn't like variable length array. */
3161 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3162 #endif
3163 struct dpcls_rule *rules[PKT_ARRAY_SIZE];
3164 struct dp_netdev *dp = pmd->dp;
3165 struct emc_cache *flow_cache = &pmd->flow_cache;
3166 int miss_cnt = 0, lost_cnt = 0;
3167 bool any_miss;
3168 size_t i;
3169
3170 for (i = 0; i < cnt; i++) {
3171 /* Key length is needed in all the cases, hash computed on demand. */
3172 keys[i].len = netdev_flow_key_size(count_1bits(keys[i].mf.map));
3173 }
3174 any_miss = !dpcls_lookup(&pmd->cls, keys, rules, cnt);
3175 if (OVS_UNLIKELY(any_miss) && !fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3176 uint64_t actions_stub[512 / 8], slow_stub[512 / 8];
3177 struct ofpbuf actions, put_actions;
3178 ovs_u128 ufid;
3179
3180 ofpbuf_use_stub(&actions, actions_stub, sizeof actions_stub);
3181 ofpbuf_use_stub(&put_actions, slow_stub, sizeof slow_stub);
3182
3183 for (i = 0; i < cnt; i++) {
3184 struct dp_netdev_flow *netdev_flow;
3185 struct ofpbuf *add_actions;
3186 struct match match;
3187 int error;
3188
3189 if (OVS_LIKELY(rules[i])) {
3190 continue;
3191 }
3192
3193 /* It's possible that an earlier slow path execution installed
3194 * a rule covering this flow. In this case, it's a lot cheaper
3195 * to catch it here than execute a miss. */
3196 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3197 if (netdev_flow) {
3198 rules[i] = &netdev_flow->cr;
3199 continue;
3200 }
3201
3202 miss_cnt++;
3203
3204 miniflow_expand(&keys[i].mf, &match.flow);
3205
3206 ofpbuf_clear(&actions);
3207 ofpbuf_clear(&put_actions);
3208
3209 dpif_flow_hash(dp->dpif, &match.flow, sizeof match.flow, &ufid);
3210 error = dp_netdev_upcall(pmd, packets[i], &match.flow, &match.wc,
3211 &ufid, DPIF_UC_MISS, NULL, &actions,
3212 &put_actions);
3213 if (OVS_UNLIKELY(error && error != ENOSPC)) {
3214 dp_packet_delete(packets[i]);
3215 lost_cnt++;
3216 continue;
3217 }
3218
3219 /* We can't allow the packet batching in the next loop to execute
3220 * the actions. Otherwise, if there are any slow path actions,
3221 * we'll send the packet up twice. */
3222 dp_netdev_execute_actions(pmd, &packets[i], 1, true,
3223 actions.data, actions.size);
3224
3225 add_actions = put_actions.size ? &put_actions : &actions;
3226 if (OVS_LIKELY(error != ENOSPC)) {
3227 /* XXX: There's a race window where a flow covering this packet
3228 * could have already been installed since we last did the flow
3229 * lookup before upcall. This could be solved by moving the
3230 * mutex lock outside the loop, but that's an awful long time
3231 * to be locking everyone out of making flow installs. If we
3232 * move to a per-core classifier, it would be reasonable. */
3233 ovs_mutex_lock(&pmd->flow_mutex);
3234 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3235 if (OVS_LIKELY(!netdev_flow)) {
3236 netdev_flow = dp_netdev_flow_add(pmd, &match, &ufid,
3237 add_actions->data,
3238 add_actions->size);
3239 }
3240 ovs_mutex_unlock(&pmd->flow_mutex);
3241
3242 emc_insert(flow_cache, &keys[i], netdev_flow);
3243 }
3244 }
3245
3246 ofpbuf_uninit(&actions);
3247 ofpbuf_uninit(&put_actions);
3248 fat_rwlock_unlock(&dp->upcall_rwlock);
3249 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3250 } else if (OVS_UNLIKELY(any_miss)) {
3251 for (i = 0; i < cnt; i++) {
3252 if (OVS_UNLIKELY(!rules[i])) {
3253 dp_packet_delete(packets[i]);
3254 lost_cnt++;
3255 miss_cnt++;
3256 }
3257 }
3258 }
3259
3260 for (i = 0; i < cnt; i++) {
3261 struct dp_packet *packet = packets[i];
3262 struct dp_netdev_flow *flow;
3263
3264 if (OVS_UNLIKELY(!rules[i])) {
3265 continue;
3266 }
3267
3268 flow = dp_netdev_flow_cast(rules[i]);
3269
3270 emc_insert(flow_cache, &keys[i], flow);
3271 dp_netdev_queue_batches(packet, flow, &keys[i].mf, batches, n_batches);
3272 }
3273
3274 dp_netdev_count_packet(pmd, DP_STAT_MASKED_HIT, cnt - miss_cnt);
3275 dp_netdev_count_packet(pmd, DP_STAT_MISS, miss_cnt);
3276 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3277 }
3278
3279 static void
3280 dp_netdev_input(struct dp_netdev_pmd_thread *pmd,
3281 struct dp_packet **packets, int cnt)
3282 {
3283 #if !defined(__CHECKER__) && !defined(_WIN32)
3284 const size_t PKT_ARRAY_SIZE = cnt;
3285 #else
3286 /* Sparse or MSVC doesn't like variable length array. */
3287 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3288 #endif
3289 struct netdev_flow_key keys[PKT_ARRAY_SIZE];
3290 struct packet_batch batches[PKT_ARRAY_SIZE];
3291 long long now = time_msec();
3292 size_t newcnt, n_batches, i;
3293
3294 n_batches = 0;
3295 newcnt = emc_processing(pmd, packets, cnt, keys, batches, &n_batches);
3296 if (OVS_UNLIKELY(newcnt)) {
3297 fast_path_processing(pmd, packets, newcnt, keys, batches, &n_batches);
3298 }
3299
3300 for (i = 0; i < n_batches; i++) {
3301 batches[i].flow->batch = NULL;
3302 }
3303
3304 for (i = 0; i < n_batches; i++) {
3305 packet_batch_execute(&batches[i], pmd, now);
3306 }
3307 }
3308
3309 struct dp_netdev_execute_aux {
3310 struct dp_netdev_pmd_thread *pmd;
3311 };
3312
3313 static void
3314 dpif_netdev_register_upcall_cb(struct dpif *dpif, upcall_callback *cb,
3315 void *aux)
3316 {
3317 struct dp_netdev *dp = get_dp_netdev(dpif);
3318 dp->upcall_aux = aux;
3319 dp->upcall_cb = cb;
3320 }
3321
3322 static void
3323 dp_netdev_drop_packets(struct dp_packet ** packets, int cnt, bool may_steal)
3324 {
3325 if (may_steal) {
3326 int i;
3327
3328 for (i = 0; i < cnt; i++) {
3329 dp_packet_delete(packets[i]);
3330 }
3331 }
3332 }
3333
3334 static int
3335 push_tnl_action(const struct dp_netdev *dp,
3336 const struct nlattr *attr,
3337 struct dp_packet **packets, int cnt)
3338 {
3339 struct dp_netdev_port *tun_port;
3340 const struct ovs_action_push_tnl *data;
3341
3342 data = nl_attr_get(attr);
3343
3344 tun_port = dp_netdev_lookup_port(dp, u32_to_odp(data->tnl_port));
3345 if (!tun_port) {
3346 return -EINVAL;
3347 }
3348 netdev_push_header(tun_port->netdev, packets, cnt, data);
3349
3350 return 0;
3351 }
3352
3353 static void
3354 dp_netdev_clone_pkt_batch(struct dp_packet **dst_pkts,
3355 struct dp_packet **src_pkts, int cnt)
3356 {
3357 int i;
3358
3359 for (i = 0; i < cnt; i++) {
3360 dst_pkts[i] = dp_packet_clone(src_pkts[i]);
3361 }
3362 }
3363
3364 static void
3365 dp_execute_cb(void *aux_, struct dp_packet **packets, int cnt,
3366 const struct nlattr *a, bool may_steal)
3367 OVS_NO_THREAD_SAFETY_ANALYSIS
3368 {
3369 struct dp_netdev_execute_aux *aux = aux_;
3370 uint32_t *depth = recirc_depth_get();
3371 struct dp_netdev_pmd_thread *pmd = aux->pmd;
3372 struct dp_netdev *dp = pmd->dp;
3373 int type = nl_attr_type(a);
3374 struct dp_netdev_port *p;
3375 int i;
3376
3377 switch ((enum ovs_action_attr)type) {
3378 case OVS_ACTION_ATTR_OUTPUT:
3379 p = dp_netdev_lookup_port(dp, u32_to_odp(nl_attr_get_u32(a)));
3380 if (OVS_LIKELY(p)) {
3381 netdev_send(p->netdev, pmd->core_id, packets, cnt, may_steal);
3382 return;
3383 }
3384 break;
3385
3386 case OVS_ACTION_ATTR_TUNNEL_PUSH:
3387 if (*depth < MAX_RECIRC_DEPTH) {
3388 struct dp_packet *tnl_pkt[NETDEV_MAX_BURST];
3389 int err;
3390
3391 if (!may_steal) {
3392 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3393 packets = tnl_pkt;
3394 }
3395
3396 err = push_tnl_action(dp, a, packets, cnt);
3397 if (!err) {
3398 (*depth)++;
3399 dp_netdev_input(pmd, packets, cnt);
3400 (*depth)--;
3401 } else {
3402 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3403 }
3404 return;
3405 }
3406 break;
3407
3408 case OVS_ACTION_ATTR_TUNNEL_POP:
3409 if (*depth < MAX_RECIRC_DEPTH) {
3410 odp_port_t portno = u32_to_odp(nl_attr_get_u32(a));
3411
3412 p = dp_netdev_lookup_port(dp, portno);
3413 if (p) {
3414 struct dp_packet *tnl_pkt[NETDEV_MAX_BURST];
3415 int err;
3416
3417 if (!may_steal) {
3418 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3419 packets = tnl_pkt;
3420 }
3421
3422 err = netdev_pop_header(p->netdev, packets, cnt);
3423 if (!err) {
3424
3425 for (i = 0; i < cnt; i++) {
3426 packets[i]->md.in_port.odp_port = portno;
3427 }
3428
3429 (*depth)++;
3430 dp_netdev_input(pmd, packets, cnt);
3431 (*depth)--;
3432 } else {
3433 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3434 }
3435 return;
3436 }
3437 }
3438 break;
3439
3440 case OVS_ACTION_ATTR_USERSPACE:
3441 if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3442 const struct nlattr *userdata;
3443 struct ofpbuf actions;
3444 struct flow flow;
3445 ovs_u128 ufid;
3446
3447 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
3448 ofpbuf_init(&actions, 0);
3449
3450 for (i = 0; i < cnt; i++) {
3451 int error;
3452
3453 ofpbuf_clear(&actions);
3454
3455 flow_extract(packets[i], &flow);
3456 dpif_flow_hash(dp->dpif, &flow, sizeof flow, &ufid);
3457 error = dp_netdev_upcall(pmd, packets[i], &flow, NULL, &ufid,
3458 DPIF_UC_ACTION, userdata,&actions,
3459 NULL);
3460 if (!error || error == ENOSPC) {
3461 dp_netdev_execute_actions(pmd, &packets[i], 1, may_steal,
3462 actions.data, actions.size);
3463 } else if (may_steal) {
3464 dp_packet_delete(packets[i]);
3465 }
3466 }
3467 ofpbuf_uninit(&actions);
3468 fat_rwlock_unlock(&dp->upcall_rwlock);
3469
3470 return;
3471 }
3472 break;
3473
3474 case OVS_ACTION_ATTR_RECIRC:
3475 if (*depth < MAX_RECIRC_DEPTH) {
3476 struct dp_packet *recirc_pkts[NETDEV_MAX_BURST];
3477
3478 if (!may_steal) {
3479 dp_netdev_clone_pkt_batch(recirc_pkts, packets, cnt);
3480 packets = recirc_pkts;
3481 }
3482
3483 for (i = 0; i < cnt; i++) {
3484 packets[i]->md.recirc_id = nl_attr_get_u32(a);
3485 }
3486
3487 (*depth)++;
3488 dp_netdev_input(pmd, packets, cnt);
3489 (*depth)--;
3490
3491 return;
3492 }
3493
3494 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
3495 break;
3496
3497 case OVS_ACTION_ATTR_PUSH_VLAN:
3498 case OVS_ACTION_ATTR_POP_VLAN:
3499 case OVS_ACTION_ATTR_PUSH_MPLS:
3500 case OVS_ACTION_ATTR_POP_MPLS:
3501 case OVS_ACTION_ATTR_SET:
3502 case OVS_ACTION_ATTR_SET_MASKED:
3503 case OVS_ACTION_ATTR_SAMPLE:
3504 case OVS_ACTION_ATTR_HASH:
3505 case OVS_ACTION_ATTR_UNSPEC:
3506 case __OVS_ACTION_ATTR_MAX:
3507 OVS_NOT_REACHED();
3508 }
3509
3510 dp_netdev_drop_packets(packets, cnt, may_steal);
3511 }
3512
3513 static void
3514 dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
3515 struct dp_packet **packets, int cnt,
3516 bool may_steal,
3517 const struct nlattr *actions, size_t actions_len)
3518 {
3519 struct dp_netdev_execute_aux aux = { pmd };
3520
3521 odp_execute_actions(&aux, packets, cnt, may_steal, actions,
3522 actions_len, dp_execute_cb);
3523 }
3524
3525 const struct dpif_class dpif_netdev_class = {
3526 "netdev",
3527 dpif_netdev_init,
3528 dpif_netdev_enumerate,
3529 dpif_netdev_port_open_type,
3530 dpif_netdev_open,
3531 dpif_netdev_close,
3532 dpif_netdev_destroy,
3533 dpif_netdev_run,
3534 dpif_netdev_wait,
3535 dpif_netdev_get_stats,
3536 dpif_netdev_port_add,
3537 dpif_netdev_port_del,
3538 dpif_netdev_port_query_by_number,
3539 dpif_netdev_port_query_by_name,
3540 NULL, /* port_get_pid */
3541 dpif_netdev_port_dump_start,
3542 dpif_netdev_port_dump_next,
3543 dpif_netdev_port_dump_done,
3544 dpif_netdev_port_poll,
3545 dpif_netdev_port_poll_wait,
3546 dpif_netdev_flow_flush,
3547 dpif_netdev_flow_dump_create,
3548 dpif_netdev_flow_dump_destroy,
3549 dpif_netdev_flow_dump_thread_create,
3550 dpif_netdev_flow_dump_thread_destroy,
3551 dpif_netdev_flow_dump_next,
3552 dpif_netdev_operate,
3553 NULL, /* recv_set */
3554 NULL, /* handlers_set */
3555 dpif_netdev_pmd_set,
3556 dpif_netdev_queue_to_priority,
3557 NULL, /* recv */
3558 NULL, /* recv_wait */
3559 NULL, /* recv_purge */
3560 dpif_netdev_register_upcall_cb,
3561 dpif_netdev_enable_upcall,
3562 dpif_netdev_disable_upcall,
3563 dpif_netdev_get_datapath_version,
3564 };
3565
3566 static void
3567 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
3568 const char *argv[], void *aux OVS_UNUSED)
3569 {
3570 struct dp_netdev_port *old_port;
3571 struct dp_netdev_port *new_port;
3572 struct dp_netdev *dp;
3573 odp_port_t port_no;
3574
3575 ovs_mutex_lock(&dp_netdev_mutex);
3576 dp = shash_find_data(&dp_netdevs, argv[1]);
3577 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3578 ovs_mutex_unlock(&dp_netdev_mutex);
3579 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3580 return;
3581 }
3582 ovs_refcount_ref(&dp->ref_cnt);
3583 ovs_mutex_unlock(&dp_netdev_mutex);
3584
3585 ovs_mutex_lock(&dp->port_mutex);
3586 if (get_port_by_name(dp, argv[2], &old_port)) {
3587 unixctl_command_reply_error(conn, "unknown port");
3588 goto exit;
3589 }
3590
3591 port_no = u32_to_odp(atoi(argv[3]));
3592 if (!port_no || port_no == ODPP_NONE) {
3593 unixctl_command_reply_error(conn, "bad port number");
3594 goto exit;
3595 }
3596 if (dp_netdev_lookup_port(dp, port_no)) {
3597 unixctl_command_reply_error(conn, "port number already in use");
3598 goto exit;
3599 }
3600
3601 /* Remove old port. */
3602 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->md.in_port.odp_port));
3603 ovsrcu_postpone(free, old_port);
3604
3605 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
3606 new_port = xmemdup(old_port, sizeof *old_port);
3607 new_port->md.in_port.odp_port = port_no;
3608 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
3609
3610 seq_change(dp->port_seq);
3611 unixctl_command_reply(conn, NULL);
3612
3613 exit:
3614 ovs_mutex_unlock(&dp->port_mutex);
3615 dp_netdev_unref(dp);
3616 }
3617
3618 static void
3619 dpif_dummy_delete_port(struct unixctl_conn *conn, int argc OVS_UNUSED,
3620 const char *argv[], void *aux OVS_UNUSED)
3621 {
3622 struct dp_netdev_port *port;
3623 struct dp_netdev *dp;
3624
3625 ovs_mutex_lock(&dp_netdev_mutex);
3626 dp = shash_find_data(&dp_netdevs, argv[1]);
3627 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3628 ovs_mutex_unlock(&dp_netdev_mutex);
3629 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3630 return;
3631 }
3632 ovs_refcount_ref(&dp->ref_cnt);
3633 ovs_mutex_unlock(&dp_netdev_mutex);
3634
3635 ovs_mutex_lock(&dp->port_mutex);
3636 if (get_port_by_name(dp, argv[2], &port)) {
3637 unixctl_command_reply_error(conn, "unknown port");
3638 } else if (port->md.in_port.odp_port == ODPP_LOCAL) {
3639 unixctl_command_reply_error(conn, "can't delete local port");
3640 } else {
3641 do_del_port(dp, port);
3642 unixctl_command_reply(conn, NULL);
3643 }
3644 ovs_mutex_unlock(&dp->port_mutex);
3645
3646 dp_netdev_unref(dp);
3647 }
3648
3649 static void
3650 dpif_dummy_register__(const char *type)
3651 {
3652 struct dpif_class *class;
3653
3654 class = xmalloc(sizeof *class);
3655 *class = dpif_netdev_class;
3656 class->type = xstrdup(type);
3657 dp_register_provider(class);
3658 }
3659
3660 void
3661 dpif_dummy_register(bool override)
3662 {
3663 if (override) {
3664 struct sset types;
3665 const char *type;
3666
3667 sset_init(&types);
3668 dp_enumerate_types(&types);
3669 SSET_FOR_EACH (type, &types) {
3670 if (!dp_unregister_provider(type)) {
3671 dpif_dummy_register__(type);
3672 }
3673 }
3674 sset_destroy(&types);
3675 }
3676
3677 dpif_dummy_register__("dummy");
3678
3679 unixctl_command_register("dpif-dummy/change-port-number",
3680 "dp port new-number",
3681 3, 3, dpif_dummy_change_port_number, NULL);
3682 unixctl_command_register("dpif-dummy/delete-port", "dp port",
3683 2, 2, dpif_dummy_delete_port, NULL);
3684 }
3685 \f
3686 /* Datapath Classifier. */
3687
3688 /* A set of rules that all have the same fields wildcarded. */
3689 struct dpcls_subtable {
3690 /* The fields are only used by writers. */
3691 struct cmap_node cmap_node OVS_GUARDED; /* Within dpcls 'subtables_map'. */
3692
3693 /* These fields are accessed by readers. */
3694 struct cmap rules; /* Contains "struct dpcls_rule"s. */
3695 struct netdev_flow_key mask; /* Wildcards for fields (const). */
3696 /* 'mask' must be the last field, additional space is allocated here. */
3697 };
3698
3699 /* Initializes 'cls' as a classifier that initially contains no classification
3700 * rules. */
3701 static void
3702 dpcls_init(struct dpcls *cls)
3703 {
3704 cmap_init(&cls->subtables_map);
3705 pvector_init(&cls->subtables);
3706 }
3707
3708 static void
3709 dpcls_destroy_subtable(struct dpcls *cls, struct dpcls_subtable *subtable)
3710 {
3711 pvector_remove(&cls->subtables, subtable);
3712 cmap_remove(&cls->subtables_map, &subtable->cmap_node,
3713 subtable->mask.hash);
3714 cmap_destroy(&subtable->rules);
3715 ovsrcu_postpone(free, subtable);
3716 }
3717
3718 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
3719 * caller's responsibility.
3720 * May only be called after all the readers have been terminated. */
3721 static void
3722 dpcls_destroy(struct dpcls *cls)
3723 {
3724 if (cls) {
3725 struct dpcls_subtable *subtable;
3726
3727 CMAP_FOR_EACH (subtable, cmap_node, &cls->subtables_map) {
3728 dpcls_destroy_subtable(cls, subtable);
3729 }
3730 cmap_destroy(&cls->subtables_map);
3731 pvector_destroy(&cls->subtables);
3732 }
3733 }
3734
3735 static struct dpcls_subtable *
3736 dpcls_create_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
3737 {
3738 struct dpcls_subtable *subtable;
3739
3740 /* Need to add one. */
3741 subtable = xmalloc(sizeof *subtable
3742 - sizeof subtable->mask.mf + mask->len);
3743 cmap_init(&subtable->rules);
3744 netdev_flow_key_clone(&subtable->mask, mask);
3745 cmap_insert(&cls->subtables_map, &subtable->cmap_node, mask->hash);
3746 pvector_insert(&cls->subtables, subtable, 0);
3747 pvector_publish(&cls->subtables);
3748
3749 return subtable;
3750 }
3751
3752 static inline struct dpcls_subtable *
3753 dpcls_find_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
3754 {
3755 struct dpcls_subtable *subtable;
3756
3757 CMAP_FOR_EACH_WITH_HASH (subtable, cmap_node, mask->hash,
3758 &cls->subtables_map) {
3759 if (netdev_flow_key_equal(&subtable->mask, mask)) {
3760 return subtable;
3761 }
3762 }
3763 return dpcls_create_subtable(cls, mask);
3764 }
3765
3766 /* Insert 'rule' into 'cls'. */
3767 static void
3768 dpcls_insert(struct dpcls *cls, struct dpcls_rule *rule,
3769 const struct netdev_flow_key *mask)
3770 {
3771 struct dpcls_subtable *subtable = dpcls_find_subtable(cls, mask);
3772
3773 rule->mask = &subtable->mask;
3774 cmap_insert(&subtable->rules, &rule->cmap_node, rule->flow.hash);
3775 }
3776
3777 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
3778 static void
3779 dpcls_remove(struct dpcls *cls, struct dpcls_rule *rule)
3780 {
3781 struct dpcls_subtable *subtable;
3782
3783 ovs_assert(rule->mask);
3784
3785 INIT_CONTAINER(subtable, rule->mask, mask);
3786
3787 if (cmap_remove(&subtable->rules, &rule->cmap_node, rule->flow.hash)
3788 == 0) {
3789 dpcls_destroy_subtable(cls, subtable);
3790 pvector_publish(&cls->subtables);
3791 }
3792 }
3793
3794 /* Returns true if 'target' satisifies 'key' in 'mask', that is, if each 1-bit
3795 * in 'mask' the values in 'key' and 'target' are the same.
3796 *
3797 * Note: 'key' and 'mask' have the same mask, and 'key' is already masked. */
3798 static inline bool
3799 dpcls_rule_matches_key(const struct dpcls_rule *rule,
3800 const struct netdev_flow_key *target)
3801 {
3802 const uint64_t *keyp = rule->flow.mf.inline_values;
3803 const uint64_t *maskp = rule->mask->mf.inline_values;
3804 uint64_t target_u64;
3805
3806 NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(target_u64, target, rule->flow.mf.map) {
3807 if (OVS_UNLIKELY((target_u64 & *maskp++) != *keyp++)) {
3808 return false;
3809 }
3810 }
3811 return true;
3812 }
3813
3814 /* For each miniflow in 'flows' performs a classifier lookup writing the result
3815 * into the corresponding slot in 'rules'. If a particular entry in 'flows' is
3816 * NULL it is skipped.
3817 *
3818 * This function is optimized for use in the userspace datapath and therefore
3819 * does not implement a lot of features available in the standard
3820 * classifier_lookup() function. Specifically, it does not implement
3821 * priorities, instead returning any rule which matches the flow.
3822 *
3823 * Returns true if all flows found a corresponding rule. */
3824 static bool
3825 dpcls_lookup(const struct dpcls *cls, const struct netdev_flow_key keys[],
3826 struct dpcls_rule **rules, const size_t cnt)
3827 {
3828 /* The batch size 16 was experimentally found faster than 8 or 32. */
3829 typedef uint16_t map_type;
3830 #define MAP_BITS (sizeof(map_type) * CHAR_BIT)
3831
3832 #if !defined(__CHECKER__) && !defined(_WIN32)
3833 const int N_MAPS = DIV_ROUND_UP(cnt, MAP_BITS);
3834 #else
3835 enum { N_MAPS = DIV_ROUND_UP(NETDEV_MAX_BURST, MAP_BITS) };
3836 #endif
3837 map_type maps[N_MAPS];
3838 struct dpcls_subtable *subtable;
3839
3840 memset(maps, 0xff, sizeof maps);
3841 if (cnt % MAP_BITS) {
3842 maps[N_MAPS - 1] >>= MAP_BITS - cnt % MAP_BITS; /* Clear extra bits. */
3843 }
3844 memset(rules, 0, cnt * sizeof *rules);
3845
3846 PVECTOR_FOR_EACH (subtable, &cls->subtables) {
3847 const struct netdev_flow_key *mkeys = keys;
3848 struct dpcls_rule **mrules = rules;
3849 map_type remains = 0;
3850 int m;
3851
3852 BUILD_ASSERT_DECL(sizeof remains == sizeof *maps);
3853
3854 for (m = 0; m < N_MAPS; m++, mkeys += MAP_BITS, mrules += MAP_BITS) {
3855 uint32_t hashes[MAP_BITS];
3856 const struct cmap_node *nodes[MAP_BITS];
3857 unsigned long map = maps[m];
3858 int i;
3859
3860 if (!map) {
3861 continue; /* Skip empty maps. */
3862 }
3863
3864 /* Compute hashes for the remaining keys. */
3865 ULONG_FOR_EACH_1(i, map) {
3866 hashes[i] = netdev_flow_key_hash_in_mask(&mkeys[i],
3867 &subtable->mask);
3868 }
3869 /* Lookup. */
3870 map = cmap_find_batch(&subtable->rules, map, hashes, nodes);
3871 /* Check results. */
3872 ULONG_FOR_EACH_1(i, map) {
3873 struct dpcls_rule *rule;
3874
3875 CMAP_NODE_FOR_EACH (rule, cmap_node, nodes[i]) {
3876 if (OVS_LIKELY(dpcls_rule_matches_key(rule, &mkeys[i]))) {
3877 mrules[i] = rule;
3878 goto next;
3879 }
3880 }
3881 ULONG_SET0(map, i); /* Did not match. */
3882 next:
3883 ; /* Keep Sparse happy. */
3884 }
3885 maps[m] &= ~map; /* Clear the found rules. */
3886 remains |= maps[m];
3887 }
3888 if (!remains) {
3889 return true; /* All found. */
3890 }
3891 }
3892 return false; /* Some misses. */
3893 }