]> git.proxmox.com Git - ovs.git/blob - lib/dpif-netdev.c
flow: Make compile with MSVC.
[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 static struct odp_support dp_netdev_support = {
91 .max_mpls_depth = SIZE_MAX,
92 .recirc = true,
93 };
94
95 /* Stores a miniflow with inline values */
96
97 struct netdev_flow_key {
98 uint32_t hash; /* Hash function differs for different users. */
99 uint32_t len; /* Length of the following miniflow (incl. map). */
100 struct miniflow mf;
101 uint64_t buf[FLOW_MAX_PACKET_U64S];
102 };
103
104 /* Exact match cache for frequently used flows
105 *
106 * The cache uses a 32-bit hash of the packet (which can be the RSS hash) to
107 * search its entries for a miniflow that matches exactly the miniflow of the
108 * packet. It stores the 'dpcls_rule' (rule) that matches the miniflow.
109 *
110 * A cache entry holds a reference to its 'dp_netdev_flow'.
111 *
112 * A miniflow with a given hash can be in one of EM_FLOW_HASH_SEGS different
113 * entries. The 32-bit hash is split into EM_FLOW_HASH_SEGS values (each of
114 * them is EM_FLOW_HASH_SHIFT bits wide and the remainder is thrown away). Each
115 * value is the index of a cache entry where the miniflow could be.
116 *
117 *
118 * Thread-safety
119 * =============
120 *
121 * Each pmd_thread has its own private exact match cache.
122 * If dp_netdev_input is not called from a pmd thread, a mutex is used.
123 */
124
125 #define EM_FLOW_HASH_SHIFT 13
126 #define EM_FLOW_HASH_ENTRIES (1u << EM_FLOW_HASH_SHIFT)
127 #define EM_FLOW_HASH_MASK (EM_FLOW_HASH_ENTRIES - 1)
128 #define EM_FLOW_HASH_SEGS 2
129
130 struct emc_entry {
131 struct dp_netdev_flow *flow;
132 struct netdev_flow_key key; /* key.hash used for emc hash value. */
133 };
134
135 struct emc_cache {
136 struct emc_entry entries[EM_FLOW_HASH_ENTRIES];
137 int sweep_idx; /* For emc_cache_slow_sweep(). */
138 };
139
140 /* Iterate in the exact match cache through every entry that might contain a
141 * miniflow with hash 'HASH'. */
142 #define EMC_FOR_EACH_POS_WITH_HASH(EMC, CURRENT_ENTRY, HASH) \
143 for (uint32_t i__ = 0, srch_hash__ = (HASH); \
144 (CURRENT_ENTRY) = &(EMC)->entries[srch_hash__ & EM_FLOW_HASH_MASK], \
145 i__ < EM_FLOW_HASH_SEGS; \
146 i__++, srch_hash__ >>= EM_FLOW_HASH_SHIFT)
147 \f
148 /* Simple non-wildcarding single-priority classifier. */
149
150 struct dpcls {
151 struct cmap subtables_map;
152 struct pvector subtables;
153 };
154
155 /* A rule to be inserted to the classifier. */
156 struct dpcls_rule {
157 struct cmap_node cmap_node; /* Within struct dpcls_subtable 'rules'. */
158 struct netdev_flow_key *mask; /* Subtable's mask. */
159 struct netdev_flow_key flow; /* Matching key. */
160 /* 'flow' must be the last field, additional space is allocated here. */
161 };
162
163 static void dpcls_init(struct dpcls *);
164 static void dpcls_destroy(struct dpcls *);
165 static void dpcls_insert(struct dpcls *, struct dpcls_rule *,
166 const struct netdev_flow_key *mask);
167 static void dpcls_remove(struct dpcls *, struct dpcls_rule *);
168 static bool dpcls_lookup(const struct dpcls *cls,
169 const struct netdev_flow_key keys[],
170 struct dpcls_rule **rules, size_t cnt);
171 \f
172 /* Datapath based on the network device interface from netdev.h.
173 *
174 *
175 * Thread-safety
176 * =============
177 *
178 * Some members, marked 'const', are immutable. Accessing other members
179 * requires synchronization, as noted in more detail below.
180 *
181 * Acquisition order is, from outermost to innermost:
182 *
183 * dp_netdev_mutex (global)
184 * port_mutex
185 */
186 struct dp_netdev {
187 const struct dpif_class *const class;
188 const char *const name;
189 struct dpif *dpif;
190 struct ovs_refcount ref_cnt;
191 atomic_flag destroyed;
192
193 /* Ports.
194 *
195 * Protected by RCU. Take the mutex to add or remove ports. */
196 struct ovs_mutex port_mutex;
197 struct cmap ports;
198 struct seq *port_seq; /* Incremented whenever a port changes. */
199
200 /* Protects access to ofproto-dpif-upcall interface during revalidator
201 * thread synchronization. */
202 struct fat_rwlock upcall_rwlock;
203 upcall_callback *upcall_cb; /* Callback function for executing upcalls. */
204 void *upcall_aux;
205
206 /* Stores all 'struct dp_netdev_pmd_thread's. */
207 struct cmap poll_threads;
208
209 /* Protects the access of the 'struct dp_netdev_pmd_thread'
210 * instance for non-pmd thread. */
211 struct ovs_mutex non_pmd_mutex;
212
213 /* Each pmd thread will store its pointer to
214 * 'struct dp_netdev_pmd_thread' in 'per_pmd_key'. */
215 ovsthread_key_t per_pmd_key;
216
217 /* Number of rx queues for each dpdk interface and the cpu mask
218 * for pin of pmd threads. */
219 size_t n_dpdk_rxqs;
220 char *pmd_cmask;
221 uint64_t last_tnl_conf_seq;
222 };
223
224 static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp,
225 odp_port_t);
226
227 enum dp_stat_type {
228 DP_STAT_EXACT_HIT, /* Packets that had an exact match (emc). */
229 DP_STAT_MASKED_HIT, /* Packets that matched in the flow table. */
230 DP_STAT_MISS, /* Packets that did not match. */
231 DP_STAT_LOST, /* Packets not passed up to the client. */
232 DP_N_STATS
233 };
234
235 enum pmd_cycles_counter_type {
236 PMD_CYCLES_POLLING, /* Cycles spent polling NICs. */
237 PMD_CYCLES_PROCESSING, /* Cycles spent processing packets */
238 PMD_N_CYCLES
239 };
240
241 /* A port in a netdev-based datapath. */
242 struct dp_netdev_port {
243 odp_port_t port_no;
244 struct netdev *netdev;
245 struct cmap_node node; /* Node in dp_netdev's 'ports'. */
246 struct netdev_saved_flags *sf;
247 struct netdev_rxq **rxq;
248 struct ovs_refcount ref_cnt;
249 char *type; /* Port type as requested by user. */
250 };
251
252 /* Contained by struct dp_netdev_flow's 'stats' member. */
253 struct dp_netdev_flow_stats {
254 atomic_llong used; /* Last used time, in monotonic msecs. */
255 atomic_ullong packet_count; /* Number of packets matched. */
256 atomic_ullong byte_count; /* Number of bytes matched. */
257 atomic_uint16_t tcp_flags; /* Bitwise-OR of seen tcp_flags values. */
258 };
259
260 /* A flow in 'dp_netdev_pmd_thread's 'flow_table'.
261 *
262 *
263 * Thread-safety
264 * =============
265 *
266 * Except near the beginning or ending of its lifespan, rule 'rule' belongs to
267 * its pmd thread's classifier. The text below calls this classifier 'cls'.
268 *
269 * Motivation
270 * ----------
271 *
272 * The thread safety rules described here for "struct dp_netdev_flow" are
273 * motivated by two goals:
274 *
275 * - Prevent threads that read members of "struct dp_netdev_flow" from
276 * reading bad data due to changes by some thread concurrently modifying
277 * those members.
278 *
279 * - Prevent two threads making changes to members of a given "struct
280 * dp_netdev_flow" from interfering with each other.
281 *
282 *
283 * Rules
284 * -----
285 *
286 * A flow 'flow' may be accessed without a risk of being freed during an RCU
287 * grace period. Code that needs to hold onto a flow for a while
288 * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref().
289 *
290 * 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
291 * flow from being deleted from 'cls' and it doesn't protect members of 'flow'
292 * from modification.
293 *
294 * Some members, marked 'const', are immutable. Accessing other members
295 * requires synchronization, as noted in more detail below.
296 */
297 struct dp_netdev_flow {
298 const struct flow flow; /* Unmasked flow that created this entry. */
299 /* Hash table index by unmasked flow. */
300 const struct cmap_node node; /* In owning dp_netdev_pmd_thread's */
301 /* 'flow_table'. */
302 const ovs_u128 ufid; /* Unique flow identifier. */
303 const unsigned pmd_id; /* The 'core_id' of pmd thread owning this */
304 /* flow. */
305
306 /* Number of references.
307 * The classifier owns one reference.
308 * Any thread trying to keep a rule from being freed should hold its own
309 * reference. */
310 struct ovs_refcount ref_cnt;
311
312 bool dead;
313
314 /* Statistics. */
315 struct dp_netdev_flow_stats stats;
316
317 /* Actions. */
318 OVSRCU_TYPE(struct dp_netdev_actions *) actions;
319
320 /* While processing a group of input packets, the datapath uses the next
321 * member to store a pointer to the output batch for the flow. It is
322 * reset after the batch has been sent out (See dp_netdev_queue_batches(),
323 * packet_batch_init() and packet_batch_execute()). */
324 struct packet_batch *batch;
325
326 /* Packet classification. */
327 struct dpcls_rule cr; /* In owning dp_netdev's 'cls'. */
328 /* 'cr' must be the last member. */
329 };
330
331 static void dp_netdev_flow_unref(struct dp_netdev_flow *);
332 static bool dp_netdev_flow_ref(struct dp_netdev_flow *);
333 static int dpif_netdev_flow_from_nlattrs(const struct nlattr *, uint32_t,
334 struct flow *);
335
336 /* A set of datapath actions within a "struct dp_netdev_flow".
337 *
338 *
339 * Thread-safety
340 * =============
341 *
342 * A struct dp_netdev_actions 'actions' is protected with RCU. */
343 struct dp_netdev_actions {
344 /* These members are immutable: they do not change during the struct's
345 * lifetime. */
346 unsigned int size; /* Size of 'actions', in bytes. */
347 struct nlattr actions[]; /* Sequence of OVS_ACTION_ATTR_* attributes. */
348 };
349
350 struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
351 size_t);
352 struct dp_netdev_actions *dp_netdev_flow_get_actions(
353 const struct dp_netdev_flow *);
354 static void dp_netdev_actions_free(struct dp_netdev_actions *);
355
356 /* Contained by struct dp_netdev_pmd_thread's 'stats' member. */
357 struct dp_netdev_pmd_stats {
358 /* Indexed by DP_STAT_*. */
359 atomic_ullong n[DP_N_STATS];
360 };
361
362 /* Contained by struct dp_netdev_pmd_thread's 'cycle' member. */
363 struct dp_netdev_pmd_cycles {
364 /* Indexed by PMD_CYCLES_*. */
365 atomic_ullong n[PMD_N_CYCLES];
366 };
367
368 /* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
369 * the performance overhead of interrupt processing. Therefore netdev can
370 * not implement rx-wait for these devices. dpif-netdev needs to poll
371 * these device to check for recv buffer. pmd-thread does polling for
372 * devices assigned to itself.
373 *
374 * DPDK used PMD for accessing NIC.
375 *
376 * Note, instance with cpu core id NON_PMD_CORE_ID will be reserved for
377 * I/O of all non-pmd threads. There will be no actual thread created
378 * for the instance.
379 *
380 * Each struct has its own flow table and classifier. Packets received
381 * from managed ports are looked up in the corresponding pmd thread's
382 * flow table, and are executed with the found actions.
383 * */
384 struct dp_netdev_pmd_thread {
385 struct dp_netdev *dp;
386 struct ovs_refcount ref_cnt; /* Every reference must be refcount'ed. */
387 struct cmap_node node; /* In 'dp->poll_threads'. */
388
389 pthread_cond_t cond; /* For synchronizing pmd thread reload. */
390 struct ovs_mutex cond_mutex; /* Mutex for condition variable. */
391
392 /* Per thread exact-match cache. Note, the instance for cpu core
393 * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly
394 * need to be protected (e.g. by 'dp_netdev_mutex'). All other
395 * instances will only be accessed by its own pmd thread. */
396 struct emc_cache flow_cache;
397
398 /* Classifier and Flow-Table.
399 *
400 * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
401 * changes to 'cls' must be made while still holding the 'flow_mutex'.
402 */
403 struct ovs_mutex flow_mutex;
404 struct dpcls cls;
405 struct cmap flow_table OVS_GUARDED; /* Flow table. */
406
407 /* Statistics. */
408 struct dp_netdev_pmd_stats stats;
409
410 /* Cycles counters */
411 struct dp_netdev_pmd_cycles cycles;
412
413 /* Used to count cicles. See 'cycles_counter_end()' */
414 unsigned long long last_cycles;
415
416 struct latch exit_latch; /* For terminating the pmd thread. */
417 atomic_uint change_seq; /* For reloading pmd ports. */
418 pthread_t thread;
419 int index; /* Idx of this pmd thread among pmd*/
420 /* threads on same numa node. */
421 unsigned core_id; /* CPU core id of this pmd thread. */
422 int numa_id; /* numa node id of this pmd thread. */
423 int tx_qid; /* Queue id used by this pmd thread to
424 * send packets on all netdevs */
425
426 /* Only a pmd thread can write on its own 'cycles' and 'stats'.
427 * The main thread keeps 'stats_zero' and 'cycles_zero' as base
428 * values and subtracts them from 'stats' and 'cycles' before
429 * reporting to the user */
430 unsigned long long stats_zero[DP_N_STATS];
431 uint64_t cycles_zero[PMD_N_CYCLES];
432 };
433
434 #define PMD_INITIAL_SEQ 1
435
436 /* Interface to netdev-based datapath. */
437 struct dpif_netdev {
438 struct dpif dpif;
439 struct dp_netdev *dp;
440 uint64_t last_port_seq;
441 };
442
443 static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no,
444 struct dp_netdev_port **portp);
445 static int get_port_by_name(struct dp_netdev *dp, const char *devname,
446 struct dp_netdev_port **portp);
447 static void dp_netdev_free(struct dp_netdev *)
448 OVS_REQUIRES(dp_netdev_mutex);
449 static int do_add_port(struct dp_netdev *dp, const char *devname,
450 const char *type, odp_port_t port_no)
451 OVS_REQUIRES(dp->port_mutex);
452 static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *)
453 OVS_REQUIRES(dp->port_mutex);
454 static int dpif_netdev_open(const struct dpif_class *, const char *name,
455 bool create, struct dpif **);
456 static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
457 struct dp_packet **, int c,
458 bool may_steal,
459 const struct nlattr *actions,
460 size_t actions_len);
461 static void dp_netdev_input(struct dp_netdev_pmd_thread *,
462 struct dp_packet **, int cnt);
463
464 static void dp_netdev_disable_upcall(struct dp_netdev *);
465 void dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd);
466 static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd,
467 struct dp_netdev *dp, int index,
468 unsigned core_id, int numa_id);
469 static void dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd);
470 static void dp_netdev_set_nonpmd(struct dp_netdev *dp);
471 static struct dp_netdev_pmd_thread *dp_netdev_get_pmd(struct dp_netdev *dp,
472 unsigned core_id);
473 static struct dp_netdev_pmd_thread *
474 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos);
475 static void dp_netdev_destroy_all_pmds(struct dp_netdev *dp);
476 static void dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id);
477 static void dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id);
478 static void dp_netdev_reset_pmd_threads(struct dp_netdev *dp);
479 static bool dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd);
480 static void dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd);
481 static void dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd);
482
483 static inline bool emc_entry_alive(struct emc_entry *ce);
484 static void emc_clear_entry(struct emc_entry *ce);
485
486 static void
487 emc_cache_init(struct emc_cache *flow_cache)
488 {
489 int i;
490
491 BUILD_ASSERT(sizeof(struct miniflow) == sizeof(uint64_t));
492
493 flow_cache->sweep_idx = 0;
494 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
495 flow_cache->entries[i].flow = NULL;
496 flow_cache->entries[i].key.hash = 0;
497 flow_cache->entries[i].key.len = sizeof(struct miniflow);
498 flow_cache->entries[i].key.mf.map = 0;
499 }
500 }
501
502 static void
503 emc_cache_uninit(struct emc_cache *flow_cache)
504 {
505 int i;
506
507 for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
508 emc_clear_entry(&flow_cache->entries[i]);
509 }
510 }
511
512 /* Check and clear dead flow references slowly (one entry at each
513 * invocation). */
514 static void
515 emc_cache_slow_sweep(struct emc_cache *flow_cache)
516 {
517 struct emc_entry *entry = &flow_cache->entries[flow_cache->sweep_idx];
518
519 if (!emc_entry_alive(entry)) {
520 emc_clear_entry(entry);
521 }
522 flow_cache->sweep_idx = (flow_cache->sweep_idx + 1) & EM_FLOW_HASH_MASK;
523 }
524
525 /* Returns true if 'dpif' is a netdev or dummy dpif, false otherwise. */
526 bool
527 dpif_is_netdev(const struct dpif *dpif)
528 {
529 return dpif->dpif_class->open == dpif_netdev_open;
530 }
531
532 static struct dpif_netdev *
533 dpif_netdev_cast(const struct dpif *dpif)
534 {
535 ovs_assert(dpif_is_netdev(dpif));
536 return CONTAINER_OF(dpif, struct dpif_netdev, dpif);
537 }
538
539 static struct dp_netdev *
540 get_dp_netdev(const struct dpif *dpif)
541 {
542 return dpif_netdev_cast(dpif)->dp;
543 }
544 \f
545 enum pmd_info_type {
546 PMD_INFO_SHOW_STATS, /* show how cpu cycles are spent */
547 PMD_INFO_CLEAR_STATS /* set the cycles count to 0 */
548 };
549
550 static void
551 pmd_info_show_stats(struct ds *reply,
552 struct dp_netdev_pmd_thread *pmd,
553 unsigned long long stats[DP_N_STATS],
554 uint64_t cycles[PMD_N_CYCLES])
555 {
556 unsigned long long total_packets = 0;
557 uint64_t total_cycles = 0;
558 int i;
559
560 /* These loops subtracts reference values ('*_zero') from the counters.
561 * Since loads and stores are relaxed, it might be possible for a '*_zero'
562 * value to be more recent than the current value we're reading from the
563 * counter. This is not a big problem, since these numbers are not
564 * supposed to be too accurate, but we should at least make sure that
565 * the result is not negative. */
566 for (i = 0; i < DP_N_STATS; i++) {
567 if (stats[i] > pmd->stats_zero[i]) {
568 stats[i] -= pmd->stats_zero[i];
569 } else {
570 stats[i] = 0;
571 }
572
573 if (i != DP_STAT_LOST) {
574 /* Lost packets are already included in DP_STAT_MISS */
575 total_packets += stats[i];
576 }
577 }
578
579 for (i = 0; i < PMD_N_CYCLES; i++) {
580 if (cycles[i] > pmd->cycles_zero[i]) {
581 cycles[i] -= pmd->cycles_zero[i];
582 } else {
583 cycles[i] = 0;
584 }
585
586 total_cycles += cycles[i];
587 }
588
589 ds_put_cstr(reply, (pmd->core_id == NON_PMD_CORE_ID)
590 ? "main thread" : "pmd thread");
591
592 if (pmd->numa_id != OVS_NUMA_UNSPEC) {
593 ds_put_format(reply, " numa_id %d", pmd->numa_id);
594 }
595 if (pmd->core_id != OVS_CORE_UNSPEC && pmd->core_id != NON_PMD_CORE_ID) {
596 ds_put_format(reply, " core_id %u", pmd->core_id);
597 }
598 ds_put_cstr(reply, ":\n");
599
600 ds_put_format(reply,
601 "\temc hits:%llu\n\tmegaflow hits:%llu\n"
602 "\tmiss:%llu\n\tlost:%llu\n",
603 stats[DP_STAT_EXACT_HIT], stats[DP_STAT_MASKED_HIT],
604 stats[DP_STAT_MISS], stats[DP_STAT_LOST]);
605
606 if (total_cycles == 0) {
607 return;
608 }
609
610 ds_put_format(reply,
611 "\tpolling cycles:%"PRIu64" (%.02f%%)\n"
612 "\tprocessing cycles:%"PRIu64" (%.02f%%)\n",
613 cycles[PMD_CYCLES_POLLING],
614 cycles[PMD_CYCLES_POLLING] / (double)total_cycles * 100,
615 cycles[PMD_CYCLES_PROCESSING],
616 cycles[PMD_CYCLES_PROCESSING] / (double)total_cycles * 100);
617
618 if (total_packets == 0) {
619 return;
620 }
621
622 ds_put_format(reply,
623 "\tavg cycles per packet: %.02f (%"PRIu64"/%llu)\n",
624 total_cycles / (double)total_packets,
625 total_cycles, total_packets);
626
627 ds_put_format(reply,
628 "\tavg processing cycles per packet: "
629 "%.02f (%"PRIu64"/%llu)\n",
630 cycles[PMD_CYCLES_PROCESSING] / (double)total_packets,
631 cycles[PMD_CYCLES_PROCESSING], total_packets);
632 }
633
634 static void
635 pmd_info_clear_stats(struct ds *reply OVS_UNUSED,
636 struct dp_netdev_pmd_thread *pmd,
637 unsigned long long stats[DP_N_STATS],
638 uint64_t cycles[PMD_N_CYCLES])
639 {
640 int i;
641
642 /* We cannot write 'stats' and 'cycles' (because they're written by other
643 * threads) and we shouldn't change 'stats' (because they're used to count
644 * datapath stats, which must not be cleared here). Instead, we save the
645 * current values and subtract them from the values to be displayed in the
646 * future */
647 for (i = 0; i < DP_N_STATS; i++) {
648 pmd->stats_zero[i] = stats[i];
649 }
650 for (i = 0; i < PMD_N_CYCLES; i++) {
651 pmd->cycles_zero[i] = cycles[i];
652 }
653 }
654
655 static void
656 dpif_netdev_pmd_info(struct unixctl_conn *conn, int argc, const char *argv[],
657 void *aux)
658 {
659 struct ds reply = DS_EMPTY_INITIALIZER;
660 struct dp_netdev_pmd_thread *pmd;
661 struct dp_netdev *dp = NULL;
662 enum pmd_info_type type = *(enum pmd_info_type *) aux;
663
664 ovs_mutex_lock(&dp_netdev_mutex);
665
666 if (argc == 2) {
667 dp = shash_find_data(&dp_netdevs, argv[1]);
668 } else if (shash_count(&dp_netdevs) == 1) {
669 /* There's only one datapath */
670 dp = shash_first(&dp_netdevs)->data;
671 }
672
673 if (!dp) {
674 ovs_mutex_unlock(&dp_netdev_mutex);
675 unixctl_command_reply_error(conn,
676 "please specify an existing datapath");
677 return;
678 }
679
680 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
681 unsigned long long stats[DP_N_STATS];
682 uint64_t cycles[PMD_N_CYCLES];
683 int i;
684
685 /* Read current stats and cycle counters */
686 for (i = 0; i < ARRAY_SIZE(stats); i++) {
687 atomic_read_relaxed(&pmd->stats.n[i], &stats[i]);
688 }
689 for (i = 0; i < ARRAY_SIZE(cycles); i++) {
690 atomic_read_relaxed(&pmd->cycles.n[i], &cycles[i]);
691 }
692
693 if (type == PMD_INFO_CLEAR_STATS) {
694 pmd_info_clear_stats(&reply, pmd, stats, cycles);
695 } else if (type == PMD_INFO_SHOW_STATS) {
696 pmd_info_show_stats(&reply, pmd, stats, cycles);
697 }
698 }
699
700 ovs_mutex_unlock(&dp_netdev_mutex);
701
702 unixctl_command_reply(conn, ds_cstr(&reply));
703 ds_destroy(&reply);
704 }
705 \f
706 static int
707 dpif_netdev_init(void)
708 {
709 static enum pmd_info_type show_aux = PMD_INFO_SHOW_STATS,
710 clear_aux = PMD_INFO_CLEAR_STATS;
711
712 unixctl_command_register("dpif-netdev/pmd-stats-show", "[dp]",
713 0, 1, dpif_netdev_pmd_info,
714 (void *)&show_aux);
715 unixctl_command_register("dpif-netdev/pmd-stats-clear", "[dp]",
716 0, 1, dpif_netdev_pmd_info,
717 (void *)&clear_aux);
718 return 0;
719 }
720
721 static int
722 dpif_netdev_enumerate(struct sset *all_dps,
723 const struct dpif_class *dpif_class)
724 {
725 struct shash_node *node;
726
727 ovs_mutex_lock(&dp_netdev_mutex);
728 SHASH_FOR_EACH(node, &dp_netdevs) {
729 struct dp_netdev *dp = node->data;
730 if (dpif_class != dp->class) {
731 /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
732 * If the class doesn't match, skip this dpif. */
733 continue;
734 }
735 sset_add(all_dps, node->name);
736 }
737 ovs_mutex_unlock(&dp_netdev_mutex);
738
739 return 0;
740 }
741
742 static bool
743 dpif_netdev_class_is_dummy(const struct dpif_class *class)
744 {
745 return class != &dpif_netdev_class;
746 }
747
748 static const char *
749 dpif_netdev_port_open_type(const struct dpif_class *class, const char *type)
750 {
751 return strcmp(type, "internal") ? type
752 : dpif_netdev_class_is_dummy(class) ? "dummy"
753 : "tap";
754 }
755
756 static struct dpif *
757 create_dpif_netdev(struct dp_netdev *dp)
758 {
759 uint16_t netflow_id = hash_string(dp->name, 0);
760 struct dpif_netdev *dpif;
761
762 ovs_refcount_ref(&dp->ref_cnt);
763
764 dpif = xmalloc(sizeof *dpif);
765 dpif_init(&dpif->dpif, dp->class, dp->name, netflow_id >> 8, netflow_id);
766 dpif->dp = dp;
767 dpif->last_port_seq = seq_read(dp->port_seq);
768
769 return &dpif->dpif;
770 }
771
772 /* Choose an unused, non-zero port number and return it on success.
773 * Return ODPP_NONE on failure. */
774 static odp_port_t
775 choose_port(struct dp_netdev *dp, const char *name)
776 OVS_REQUIRES(dp->port_mutex)
777 {
778 uint32_t port_no;
779
780 if (dp->class != &dpif_netdev_class) {
781 const char *p;
782 int start_no = 0;
783
784 /* If the port name begins with "br", start the number search at
785 * 100 to make writing tests easier. */
786 if (!strncmp(name, "br", 2)) {
787 start_no = 100;
788 }
789
790 /* If the port name contains a number, try to assign that port number.
791 * This can make writing unit tests easier because port numbers are
792 * predictable. */
793 for (p = name; *p != '\0'; p++) {
794 if (isdigit((unsigned char) *p)) {
795 port_no = start_no + strtol(p, NULL, 10);
796 if (port_no > 0 && port_no != odp_to_u32(ODPP_NONE)
797 && !dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
798 return u32_to_odp(port_no);
799 }
800 break;
801 }
802 }
803 }
804
805 for (port_no = 1; port_no <= UINT16_MAX; port_no++) {
806 if (!dp_netdev_lookup_port(dp, u32_to_odp(port_no))) {
807 return u32_to_odp(port_no);
808 }
809 }
810
811 return ODPP_NONE;
812 }
813
814 static int
815 create_dp_netdev(const char *name, const struct dpif_class *class,
816 struct dp_netdev **dpp)
817 OVS_REQUIRES(dp_netdev_mutex)
818 {
819 struct dp_netdev *dp;
820 int error;
821
822 dp = xzalloc(sizeof *dp);
823 shash_add(&dp_netdevs, name, dp);
824
825 *CONST_CAST(const struct dpif_class **, &dp->class) = class;
826 *CONST_CAST(const char **, &dp->name) = xstrdup(name);
827 ovs_refcount_init(&dp->ref_cnt);
828 atomic_flag_clear(&dp->destroyed);
829
830 ovs_mutex_init(&dp->port_mutex);
831 cmap_init(&dp->ports);
832 dp->port_seq = seq_create();
833 fat_rwlock_init(&dp->upcall_rwlock);
834
835 /* Disable upcalls by default. */
836 dp_netdev_disable_upcall(dp);
837 dp->upcall_aux = NULL;
838 dp->upcall_cb = NULL;
839
840 cmap_init(&dp->poll_threads);
841 ovs_mutex_init_recursive(&dp->non_pmd_mutex);
842 ovsthread_key_create(&dp->per_pmd_key, NULL);
843
844 dp_netdev_set_nonpmd(dp);
845 dp->n_dpdk_rxqs = NR_QUEUE;
846
847 ovs_mutex_lock(&dp->port_mutex);
848 error = do_add_port(dp, name, "internal", ODPP_LOCAL);
849 ovs_mutex_unlock(&dp->port_mutex);
850 if (error) {
851 dp_netdev_free(dp);
852 return error;
853 }
854
855 dp->last_tnl_conf_seq = seq_read(tnl_conf_seq);
856 *dpp = dp;
857 return 0;
858 }
859
860 static int
861 dpif_netdev_open(const struct dpif_class *class, const char *name,
862 bool create, struct dpif **dpifp)
863 {
864 struct dp_netdev *dp;
865 int error;
866
867 ovs_mutex_lock(&dp_netdev_mutex);
868 dp = shash_find_data(&dp_netdevs, name);
869 if (!dp) {
870 error = create ? create_dp_netdev(name, class, &dp) : ENODEV;
871 } else {
872 error = (dp->class != class ? EINVAL
873 : create ? EEXIST
874 : 0);
875 }
876 if (!error) {
877 *dpifp = create_dpif_netdev(dp);
878 dp->dpif = *dpifp;
879 }
880 ovs_mutex_unlock(&dp_netdev_mutex);
881
882 return error;
883 }
884
885 static void
886 dp_netdev_destroy_upcall_lock(struct dp_netdev *dp)
887 OVS_NO_THREAD_SAFETY_ANALYSIS
888 {
889 /* Check that upcalls are disabled, i.e. that the rwlock is taken */
890 ovs_assert(fat_rwlock_tryrdlock(&dp->upcall_rwlock));
891
892 /* Before freeing a lock we should release it */
893 fat_rwlock_unlock(&dp->upcall_rwlock);
894 fat_rwlock_destroy(&dp->upcall_rwlock);
895 }
896
897 /* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
898 * through the 'dp_netdevs' shash while freeing 'dp'. */
899 static void
900 dp_netdev_free(struct dp_netdev *dp)
901 OVS_REQUIRES(dp_netdev_mutex)
902 {
903 struct dp_netdev_port *port;
904
905 shash_find_and_delete(&dp_netdevs, dp->name);
906
907 dp_netdev_destroy_all_pmds(dp);
908 cmap_destroy(&dp->poll_threads);
909 ovs_mutex_destroy(&dp->non_pmd_mutex);
910 ovsthread_key_delete(dp->per_pmd_key);
911
912 ovs_mutex_lock(&dp->port_mutex);
913 CMAP_FOR_EACH (port, node, &dp->ports) {
914 do_del_port(dp, port);
915 }
916 ovs_mutex_unlock(&dp->port_mutex);
917
918 seq_destroy(dp->port_seq);
919 cmap_destroy(&dp->ports);
920
921 /* Upcalls must be disabled at this point */
922 dp_netdev_destroy_upcall_lock(dp);
923
924 free(dp->pmd_cmask);
925 free(CONST_CAST(char *, dp->name));
926 free(dp);
927 }
928
929 static void
930 dp_netdev_unref(struct dp_netdev *dp)
931 {
932 if (dp) {
933 /* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
934 * get a new reference to 'dp' through the 'dp_netdevs' shash. */
935 ovs_mutex_lock(&dp_netdev_mutex);
936 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
937 dp_netdev_free(dp);
938 }
939 ovs_mutex_unlock(&dp_netdev_mutex);
940 }
941 }
942
943 static void
944 dpif_netdev_close(struct dpif *dpif)
945 {
946 struct dp_netdev *dp = get_dp_netdev(dpif);
947
948 dp_netdev_unref(dp);
949 free(dpif);
950 }
951
952 static int
953 dpif_netdev_destroy(struct dpif *dpif)
954 {
955 struct dp_netdev *dp = get_dp_netdev(dpif);
956
957 if (!atomic_flag_test_and_set(&dp->destroyed)) {
958 if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
959 /* Can't happen: 'dpif' still owns a reference to 'dp'. */
960 OVS_NOT_REACHED();
961 }
962 }
963
964 return 0;
965 }
966
967 /* Add 'n' to the atomic variable 'var' non-atomically and using relaxed
968 * load/store semantics. While the increment is not atomic, the load and
969 * store operations are, making it impossible to read inconsistent values.
970 *
971 * This is used to update thread local stats counters. */
972 static void
973 non_atomic_ullong_add(atomic_ullong *var, unsigned long long n)
974 {
975 unsigned long long tmp;
976
977 atomic_read_relaxed(var, &tmp);
978 tmp += n;
979 atomic_store_relaxed(var, tmp);
980 }
981
982 static int
983 dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
984 {
985 struct dp_netdev *dp = get_dp_netdev(dpif);
986 struct dp_netdev_pmd_thread *pmd;
987
988 stats->n_flows = stats->n_hit = stats->n_missed = stats->n_lost = 0;
989 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
990 unsigned long long n;
991 stats->n_flows += cmap_count(&pmd->flow_table);
992
993 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MASKED_HIT], &n);
994 stats->n_hit += n;
995 atomic_read_relaxed(&pmd->stats.n[DP_STAT_EXACT_HIT], &n);
996 stats->n_hit += n;
997 atomic_read_relaxed(&pmd->stats.n[DP_STAT_MISS], &n);
998 stats->n_missed += n;
999 atomic_read_relaxed(&pmd->stats.n[DP_STAT_LOST], &n);
1000 stats->n_lost += n;
1001 }
1002 stats->n_masks = UINT32_MAX;
1003 stats->n_mask_hit = UINT64_MAX;
1004
1005 return 0;
1006 }
1007
1008 static void
1009 dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread *pmd)
1010 {
1011 int old_seq;
1012
1013 if (pmd->core_id == NON_PMD_CORE_ID) {
1014 return;
1015 }
1016
1017 ovs_mutex_lock(&pmd->cond_mutex);
1018 atomic_add_relaxed(&pmd->change_seq, 1, &old_seq);
1019 ovs_mutex_cond_wait(&pmd->cond, &pmd->cond_mutex);
1020 ovs_mutex_unlock(&pmd->cond_mutex);
1021 }
1022
1023 /* Causes all pmd threads to reload its tx/rx devices.
1024 * Must be called after adding/removing ports. */
1025 static void
1026 dp_netdev_reload_pmds(struct dp_netdev *dp)
1027 {
1028 struct dp_netdev_pmd_thread *pmd;
1029
1030 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1031 dp_netdev_reload_pmd__(pmd);
1032 }
1033 }
1034
1035 static uint32_t
1036 hash_port_no(odp_port_t port_no)
1037 {
1038 return hash_int(odp_to_u32(port_no), 0);
1039 }
1040
1041 static int
1042 do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
1043 odp_port_t port_no)
1044 OVS_REQUIRES(dp->port_mutex)
1045 {
1046 struct netdev_saved_flags *sf;
1047 struct dp_netdev_port *port;
1048 struct netdev *netdev;
1049 enum netdev_flags flags;
1050 const char *open_type;
1051 int error;
1052 int i;
1053
1054 /* Reject devices already in 'dp'. */
1055 if (!get_port_by_name(dp, devname, &port)) {
1056 return EEXIST;
1057 }
1058
1059 /* Open and validate network device. */
1060 open_type = dpif_netdev_port_open_type(dp->class, type);
1061 error = netdev_open(devname, open_type, &netdev);
1062 if (error) {
1063 return error;
1064 }
1065 /* XXX reject non-Ethernet devices */
1066
1067 netdev_get_flags(netdev, &flags);
1068 if (flags & NETDEV_LOOPBACK) {
1069 VLOG_ERR("%s: cannot add a loopback device", devname);
1070 netdev_close(netdev);
1071 return EINVAL;
1072 }
1073
1074 if (netdev_is_pmd(netdev)) {
1075 int n_cores = ovs_numa_get_n_cores();
1076
1077 if (n_cores == OVS_CORE_UNSPEC) {
1078 VLOG_ERR("%s, cannot get cpu core info", devname);
1079 return ENOENT;
1080 }
1081 /* There can only be ovs_numa_get_n_cores() pmd threads,
1082 * so creates a txq for each, and one extra for the non
1083 * pmd threads. */
1084 error = netdev_set_multiq(netdev, n_cores + 1, dp->n_dpdk_rxqs);
1085 if (error && (error != EOPNOTSUPP)) {
1086 VLOG_ERR("%s, cannot set multiq", devname);
1087 return errno;
1088 }
1089 }
1090 port = xzalloc(sizeof *port);
1091 port->port_no = port_no;
1092 port->netdev = netdev;
1093 port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev));
1094 port->type = xstrdup(type);
1095 for (i = 0; i < netdev_n_rxq(netdev); i++) {
1096 error = netdev_rxq_open(netdev, &port->rxq[i], i);
1097 if (error
1098 && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
1099 VLOG_ERR("%s: cannot receive packets on this network device (%s)",
1100 devname, ovs_strerror(errno));
1101 netdev_close(netdev);
1102 free(port->type);
1103 free(port->rxq);
1104 free(port);
1105 return error;
1106 }
1107 }
1108
1109 error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
1110 if (error) {
1111 for (i = 0; i < netdev_n_rxq(netdev); i++) {
1112 netdev_rxq_close(port->rxq[i]);
1113 }
1114 netdev_close(netdev);
1115 free(port->type);
1116 free(port->rxq);
1117 free(port);
1118 return error;
1119 }
1120 port->sf = sf;
1121
1122 ovs_refcount_init(&port->ref_cnt);
1123 cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
1124
1125 if (netdev_is_pmd(netdev)) {
1126 dp_netdev_set_pmds_on_numa(dp, netdev_get_numa_id(netdev));
1127 dp_netdev_reload_pmds(dp);
1128 }
1129 seq_change(dp->port_seq);
1130
1131 return 0;
1132 }
1133
1134 static int
1135 dpif_netdev_port_add(struct dpif *dpif, struct netdev *netdev,
1136 odp_port_t *port_nop)
1137 {
1138 struct dp_netdev *dp = get_dp_netdev(dpif);
1139 char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
1140 const char *dpif_port;
1141 odp_port_t port_no;
1142 int error;
1143
1144 ovs_mutex_lock(&dp->port_mutex);
1145 dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
1146 if (*port_nop != ODPP_NONE) {
1147 port_no = *port_nop;
1148 error = dp_netdev_lookup_port(dp, *port_nop) ? EBUSY : 0;
1149 } else {
1150 port_no = choose_port(dp, dpif_port);
1151 error = port_no == ODPP_NONE ? EFBIG : 0;
1152 }
1153 if (!error) {
1154 *port_nop = port_no;
1155 error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
1156 }
1157 ovs_mutex_unlock(&dp->port_mutex);
1158
1159 return error;
1160 }
1161
1162 static int
1163 dpif_netdev_port_del(struct dpif *dpif, odp_port_t port_no)
1164 {
1165 struct dp_netdev *dp = get_dp_netdev(dpif);
1166 int error;
1167
1168 ovs_mutex_lock(&dp->port_mutex);
1169 if (port_no == ODPP_LOCAL) {
1170 error = EINVAL;
1171 } else {
1172 struct dp_netdev_port *port;
1173
1174 error = get_port_by_number(dp, port_no, &port);
1175 if (!error) {
1176 do_del_port(dp, port);
1177 }
1178 }
1179 ovs_mutex_unlock(&dp->port_mutex);
1180
1181 return error;
1182 }
1183
1184 static bool
1185 is_valid_port_number(odp_port_t port_no)
1186 {
1187 return port_no != ODPP_NONE;
1188 }
1189
1190 static struct dp_netdev_port *
1191 dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
1192 {
1193 struct dp_netdev_port *port;
1194
1195 CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
1196 if (port->port_no == port_no) {
1197 return port;
1198 }
1199 }
1200 return NULL;
1201 }
1202
1203 static int
1204 get_port_by_number(struct dp_netdev *dp,
1205 odp_port_t port_no, struct dp_netdev_port **portp)
1206 {
1207 if (!is_valid_port_number(port_no)) {
1208 *portp = NULL;
1209 return EINVAL;
1210 } else {
1211 *portp = dp_netdev_lookup_port(dp, port_no);
1212 return *portp ? 0 : ENOENT;
1213 }
1214 }
1215
1216 static void
1217 port_ref(struct dp_netdev_port *port)
1218 {
1219 if (port) {
1220 ovs_refcount_ref(&port->ref_cnt);
1221 }
1222 }
1223
1224 static bool
1225 port_try_ref(struct dp_netdev_port *port)
1226 {
1227 if (port) {
1228 return ovs_refcount_try_ref_rcu(&port->ref_cnt);
1229 }
1230
1231 return false;
1232 }
1233
1234 static void
1235 port_unref(struct dp_netdev_port *port)
1236 {
1237 if (port && ovs_refcount_unref_relaxed(&port->ref_cnt) == 1) {
1238 int n_rxq = netdev_n_rxq(port->netdev);
1239 int i;
1240
1241 netdev_close(port->netdev);
1242 netdev_restore_flags(port->sf);
1243
1244 for (i = 0; i < n_rxq; i++) {
1245 netdev_rxq_close(port->rxq[i]);
1246 }
1247 free(port->rxq);
1248 free(port->type);
1249 free(port);
1250 }
1251 }
1252
1253 static int
1254 get_port_by_name(struct dp_netdev *dp,
1255 const char *devname, struct dp_netdev_port **portp)
1256 OVS_REQUIRES(dp->port_mutex)
1257 {
1258 struct dp_netdev_port *port;
1259
1260 CMAP_FOR_EACH (port, node, &dp->ports) {
1261 if (!strcmp(netdev_get_name(port->netdev), devname)) {
1262 *portp = port;
1263 return 0;
1264 }
1265 }
1266 return ENOENT;
1267 }
1268
1269 static int
1270 get_n_pmd_threads_on_numa(struct dp_netdev *dp, int numa_id)
1271 {
1272 struct dp_netdev_pmd_thread *pmd;
1273 int n_pmds = 0;
1274
1275 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1276 if (pmd->numa_id == numa_id) {
1277 n_pmds++;
1278 }
1279 }
1280
1281 return n_pmds;
1282 }
1283
1284 /* Returns 'true' if there is a port with pmd netdev and the netdev
1285 * is on numa node 'numa_id'. */
1286 static bool
1287 has_pmd_port_for_numa(struct dp_netdev *dp, int numa_id)
1288 {
1289 struct dp_netdev_port *port;
1290
1291 CMAP_FOR_EACH (port, node, &dp->ports) {
1292 if (netdev_is_pmd(port->netdev)
1293 && netdev_get_numa_id(port->netdev) == numa_id) {
1294 return true;
1295 }
1296 }
1297
1298 return false;
1299 }
1300
1301
1302 static void
1303 do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
1304 OVS_REQUIRES(dp->port_mutex)
1305 {
1306 cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no));
1307 seq_change(dp->port_seq);
1308 if (netdev_is_pmd(port->netdev)) {
1309 int numa_id = netdev_get_numa_id(port->netdev);
1310
1311 /* If there is no netdev on the numa node, deletes the pmd threads
1312 * for that numa. Else, just reloads the queues. */
1313 if (!has_pmd_port_for_numa(dp, numa_id)) {
1314 dp_netdev_del_pmds_on_numa(dp, numa_id);
1315 }
1316 dp_netdev_reload_pmds(dp);
1317 }
1318
1319 port_unref(port);
1320 }
1321
1322 static void
1323 answer_port_query(const struct dp_netdev_port *port,
1324 struct dpif_port *dpif_port)
1325 {
1326 dpif_port->name = xstrdup(netdev_get_name(port->netdev));
1327 dpif_port->type = xstrdup(port->type);
1328 dpif_port->port_no = port->port_no;
1329 }
1330
1331 static int
1332 dpif_netdev_port_query_by_number(const struct dpif *dpif, odp_port_t port_no,
1333 struct dpif_port *dpif_port)
1334 {
1335 struct dp_netdev *dp = get_dp_netdev(dpif);
1336 struct dp_netdev_port *port;
1337 int error;
1338
1339 error = get_port_by_number(dp, port_no, &port);
1340 if (!error && dpif_port) {
1341 answer_port_query(port, dpif_port);
1342 }
1343
1344 return error;
1345 }
1346
1347 static int
1348 dpif_netdev_port_query_by_name(const struct dpif *dpif, const char *devname,
1349 struct dpif_port *dpif_port)
1350 {
1351 struct dp_netdev *dp = get_dp_netdev(dpif);
1352 struct dp_netdev_port *port;
1353 int error;
1354
1355 ovs_mutex_lock(&dp->port_mutex);
1356 error = get_port_by_name(dp, devname, &port);
1357 if (!error && dpif_port) {
1358 answer_port_query(port, dpif_port);
1359 }
1360 ovs_mutex_unlock(&dp->port_mutex);
1361
1362 return error;
1363 }
1364
1365 static void
1366 dp_netdev_flow_free(struct dp_netdev_flow *flow)
1367 {
1368 dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
1369 free(flow);
1370 }
1371
1372 static void dp_netdev_flow_unref(struct dp_netdev_flow *flow)
1373 {
1374 if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) {
1375 ovsrcu_postpone(dp_netdev_flow_free, flow);
1376 }
1377 }
1378
1379 static uint32_t
1380 dp_netdev_flow_hash(const ovs_u128 *ufid)
1381 {
1382 return ufid->u32[0];
1383 }
1384
1385 static void
1386 dp_netdev_pmd_remove_flow(struct dp_netdev_pmd_thread *pmd,
1387 struct dp_netdev_flow *flow)
1388 OVS_REQUIRES(pmd->flow_mutex)
1389 {
1390 struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
1391
1392 dpcls_remove(&pmd->cls, &flow->cr);
1393 cmap_remove(&pmd->flow_table, node, dp_netdev_flow_hash(&flow->ufid));
1394 flow->dead = true;
1395
1396 dp_netdev_flow_unref(flow);
1397 }
1398
1399 static void
1400 dp_netdev_pmd_flow_flush(struct dp_netdev_pmd_thread *pmd)
1401 {
1402 struct dp_netdev_flow *netdev_flow;
1403
1404 ovs_mutex_lock(&pmd->flow_mutex);
1405 CMAP_FOR_EACH (netdev_flow, node, &pmd->flow_table) {
1406 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
1407 }
1408 ovs_mutex_unlock(&pmd->flow_mutex);
1409 }
1410
1411 static int
1412 dpif_netdev_flow_flush(struct dpif *dpif)
1413 {
1414 struct dp_netdev *dp = get_dp_netdev(dpif);
1415 struct dp_netdev_pmd_thread *pmd;
1416
1417 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
1418 dp_netdev_pmd_flow_flush(pmd);
1419 }
1420
1421 return 0;
1422 }
1423
1424 struct dp_netdev_port_state {
1425 struct cmap_position position;
1426 char *name;
1427 };
1428
1429 static int
1430 dpif_netdev_port_dump_start(const struct dpif *dpif OVS_UNUSED, void **statep)
1431 {
1432 *statep = xzalloc(sizeof(struct dp_netdev_port_state));
1433 return 0;
1434 }
1435
1436 static int
1437 dpif_netdev_port_dump_next(const struct dpif *dpif, void *state_,
1438 struct dpif_port *dpif_port)
1439 {
1440 struct dp_netdev_port_state *state = state_;
1441 struct dp_netdev *dp = get_dp_netdev(dpif);
1442 struct cmap_node *node;
1443 int retval;
1444
1445 node = cmap_next_position(&dp->ports, &state->position);
1446 if (node) {
1447 struct dp_netdev_port *port;
1448
1449 port = CONTAINER_OF(node, struct dp_netdev_port, node);
1450
1451 free(state->name);
1452 state->name = xstrdup(netdev_get_name(port->netdev));
1453 dpif_port->name = state->name;
1454 dpif_port->type = port->type;
1455 dpif_port->port_no = port->port_no;
1456
1457 retval = 0;
1458 } else {
1459 retval = EOF;
1460 }
1461
1462 return retval;
1463 }
1464
1465 static int
1466 dpif_netdev_port_dump_done(const struct dpif *dpif OVS_UNUSED, void *state_)
1467 {
1468 struct dp_netdev_port_state *state = state_;
1469 free(state->name);
1470 free(state);
1471 return 0;
1472 }
1473
1474 static int
1475 dpif_netdev_port_poll(const struct dpif *dpif_, char **devnamep OVS_UNUSED)
1476 {
1477 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1478 uint64_t new_port_seq;
1479 int error;
1480
1481 new_port_seq = seq_read(dpif->dp->port_seq);
1482 if (dpif->last_port_seq != new_port_seq) {
1483 dpif->last_port_seq = new_port_seq;
1484 error = ENOBUFS;
1485 } else {
1486 error = EAGAIN;
1487 }
1488
1489 return error;
1490 }
1491
1492 static void
1493 dpif_netdev_port_poll_wait(const struct dpif *dpif_)
1494 {
1495 struct dpif_netdev *dpif = dpif_netdev_cast(dpif_);
1496
1497 seq_wait(dpif->dp->port_seq, dpif->last_port_seq);
1498 }
1499
1500 static struct dp_netdev_flow *
1501 dp_netdev_flow_cast(const struct dpcls_rule *cr)
1502 {
1503 return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
1504 }
1505
1506 static bool dp_netdev_flow_ref(struct dp_netdev_flow *flow)
1507 {
1508 return ovs_refcount_try_ref_rcu(&flow->ref_cnt);
1509 }
1510
1511 /* netdev_flow_key utilities.
1512 *
1513 * netdev_flow_key is basically a miniflow. We use these functions
1514 * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
1515 * functions (miniflow_clone_inline, miniflow_equal, ...), because:
1516 *
1517 * - Since we are dealing exclusively with miniflows created by
1518 * miniflow_extract(), if the map is different the miniflow is different.
1519 * Therefore we can be faster by comparing the map and the miniflow in a
1520 * single memcmp().
1521 * _ netdev_flow_key's miniflow has always inline values.
1522 * - These functions can be inlined by the compiler.
1523 *
1524 * The following assertions make sure that what we're doing with miniflow is
1525 * safe
1526 */
1527 BUILD_ASSERT_DECL(sizeof(struct miniflow) == sizeof(uint64_t));
1528
1529 /* Given the number of bits set in the miniflow map, returns the size of the
1530 * 'netdev_flow_key.mf' */
1531 static inline uint32_t
1532 netdev_flow_key_size(uint32_t flow_u32s)
1533 {
1534 return sizeof(struct miniflow) + MINIFLOW_VALUES_SIZE(flow_u32s);
1535 }
1536
1537 static inline bool
1538 netdev_flow_key_equal(const struct netdev_flow_key *a,
1539 const struct netdev_flow_key *b)
1540 {
1541 /* 'b->len' may be not set yet. */
1542 return a->hash == b->hash && !memcmp(&a->mf, &b->mf, a->len);
1543 }
1544
1545 /* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
1546 * The maps are compared bitwise, so both 'key->mf' 'mf' must have been
1547 * generated by miniflow_extract. */
1548 static inline bool
1549 netdev_flow_key_equal_mf(const struct netdev_flow_key *key,
1550 const struct miniflow *mf)
1551 {
1552 return !memcmp(&key->mf, mf, key->len);
1553 }
1554
1555 static inline void
1556 netdev_flow_key_clone(struct netdev_flow_key *dst,
1557 const struct netdev_flow_key *src)
1558 {
1559 memcpy(dst, src,
1560 offsetof(struct netdev_flow_key, mf) + src->len);
1561 }
1562
1563 /* Slow. */
1564 static void
1565 netdev_flow_key_from_flow(struct netdev_flow_key *dst,
1566 const struct flow *src)
1567 {
1568 struct dp_packet packet;
1569 uint64_t buf_stub[512 / 8];
1570
1571 dp_packet_use_stub(&packet, buf_stub, sizeof buf_stub);
1572 pkt_metadata_from_flow(&packet.md, src);
1573 flow_compose(&packet, src);
1574 miniflow_extract(&packet, &dst->mf);
1575 dp_packet_uninit(&packet);
1576
1577 dst->len = netdev_flow_key_size(count_1bits(dst->mf.map));
1578 dst->hash = 0; /* Not computed yet. */
1579 }
1580
1581 /* Initialize a netdev_flow_key 'mask' from 'match'. */
1582 static inline void
1583 netdev_flow_mask_init(struct netdev_flow_key *mask,
1584 const struct match *match)
1585 {
1586 const uint64_t *mask_u64 = (const uint64_t *) &match->wc.masks;
1587 uint64_t *dst = miniflow_values(&mask->mf);
1588 uint64_t map, mask_map = 0;
1589 uint32_t hash = 0;
1590 int n;
1591
1592 /* Only check masks that make sense for the flow. */
1593 map = flow_wc_map(&match->flow);
1594
1595 while (map) {
1596 uint64_t rm1bit = rightmost_1bit(map);
1597 int i = raw_ctz(map);
1598
1599 if (mask_u64[i]) {
1600 mask_map |= rm1bit;
1601 *dst++ = mask_u64[i];
1602 hash = hash_add64(hash, mask_u64[i]);
1603 }
1604 map -= rm1bit;
1605 }
1606
1607 mask->mf.map = mask_map;
1608
1609 hash = hash_add64(hash, mask_map);
1610
1611 n = dst - miniflow_get_values(&mask->mf);
1612
1613 mask->hash = hash_finish(hash, n * 8);
1614 mask->len = netdev_flow_key_size(n);
1615 }
1616
1617 /* Initializes 'dst' as a copy of 'src' masked with 'mask'. */
1618 static inline void
1619 netdev_flow_key_init_masked(struct netdev_flow_key *dst,
1620 const struct flow *flow,
1621 const struct netdev_flow_key *mask)
1622 {
1623 uint64_t *dst_u64 = miniflow_values(&dst->mf);
1624 const uint64_t *mask_u64 = miniflow_get_values(&mask->mf);
1625 uint32_t hash = 0;
1626 uint64_t value;
1627
1628 dst->len = mask->len;
1629 dst->mf.map = mask->mf.map;
1630
1631 FLOW_FOR_EACH_IN_MAP(value, flow, mask->mf.map) {
1632 *dst_u64 = value & *mask_u64++;
1633 hash = hash_add64(hash, *dst_u64++);
1634 }
1635 dst->hash = hash_finish(hash,
1636 (dst_u64 - miniflow_get_values(&dst->mf)) * 8);
1637 }
1638
1639 /* Iterate through all netdev_flow_key u64 values specified by 'MAP' */
1640 #define NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(VALUE, KEY, MAP) \
1641 for (struct mf_for_each_in_map_aux aux__ \
1642 = { miniflow_get_values(&(KEY)->mf), (KEY)->mf.map, MAP }; \
1643 mf_get_next_in_map(&aux__, &(VALUE)); \
1644 )
1645
1646 /* Returns a hash value for the bits of 'key' where there are 1-bits in
1647 * 'mask'. */
1648 static inline uint32_t
1649 netdev_flow_key_hash_in_mask(const struct netdev_flow_key *key,
1650 const struct netdev_flow_key *mask)
1651 {
1652 const uint64_t *p = miniflow_get_values(&mask->mf);
1653 uint32_t hash = 0;
1654 uint64_t key_u64;
1655
1656 NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(key_u64, key, mask->mf.map) {
1657 hash = hash_add64(hash, key_u64 & *p++);
1658 }
1659
1660 return hash_finish(hash, (p - miniflow_get_values(&mask->mf)) * 8);
1661 }
1662
1663 static inline bool
1664 emc_entry_alive(struct emc_entry *ce)
1665 {
1666 return ce->flow && !ce->flow->dead;
1667 }
1668
1669 static void
1670 emc_clear_entry(struct emc_entry *ce)
1671 {
1672 if (ce->flow) {
1673 dp_netdev_flow_unref(ce->flow);
1674 ce->flow = NULL;
1675 }
1676 }
1677
1678 static inline void
1679 emc_change_entry(struct emc_entry *ce, struct dp_netdev_flow *flow,
1680 const struct netdev_flow_key *key)
1681 {
1682 if (ce->flow != flow) {
1683 if (ce->flow) {
1684 dp_netdev_flow_unref(ce->flow);
1685 }
1686
1687 if (dp_netdev_flow_ref(flow)) {
1688 ce->flow = flow;
1689 } else {
1690 ce->flow = NULL;
1691 }
1692 }
1693 if (key) {
1694 netdev_flow_key_clone(&ce->key, key);
1695 }
1696 }
1697
1698 static inline void
1699 emc_insert(struct emc_cache *cache, const struct netdev_flow_key *key,
1700 struct dp_netdev_flow *flow)
1701 {
1702 struct emc_entry *to_be_replaced = NULL;
1703 struct emc_entry *current_entry;
1704
1705 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1706 if (netdev_flow_key_equal(&current_entry->key, key)) {
1707 /* We found the entry with the 'mf' miniflow */
1708 emc_change_entry(current_entry, flow, NULL);
1709 return;
1710 }
1711
1712 /* Replacement policy: put the flow in an empty (not alive) entry, or
1713 * in the first entry where it can be */
1714 if (!to_be_replaced
1715 || (emc_entry_alive(to_be_replaced)
1716 && !emc_entry_alive(current_entry))
1717 || current_entry->key.hash < to_be_replaced->key.hash) {
1718 to_be_replaced = current_entry;
1719 }
1720 }
1721 /* We didn't find the miniflow in the cache.
1722 * The 'to_be_replaced' entry is where the new flow will be stored */
1723
1724 emc_change_entry(to_be_replaced, flow, key);
1725 }
1726
1727 static inline struct dp_netdev_flow *
1728 emc_lookup(struct emc_cache *cache, const struct netdev_flow_key *key)
1729 {
1730 struct emc_entry *current_entry;
1731
1732 EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
1733 if (current_entry->key.hash == key->hash
1734 && emc_entry_alive(current_entry)
1735 && netdev_flow_key_equal_mf(&current_entry->key, &key->mf)) {
1736
1737 /* We found the entry with the 'key->mf' miniflow */
1738 return current_entry->flow;
1739 }
1740 }
1741
1742 return NULL;
1743 }
1744
1745 static struct dp_netdev_flow *
1746 dp_netdev_pmd_lookup_flow(const struct dp_netdev_pmd_thread *pmd,
1747 const struct netdev_flow_key *key)
1748 {
1749 struct dp_netdev_flow *netdev_flow;
1750 struct dpcls_rule *rule;
1751
1752 dpcls_lookup(&pmd->cls, key, &rule, 1);
1753 netdev_flow = dp_netdev_flow_cast(rule);
1754
1755 return netdev_flow;
1756 }
1757
1758 static struct dp_netdev_flow *
1759 dp_netdev_pmd_find_flow(const struct dp_netdev_pmd_thread *pmd,
1760 const ovs_u128 *ufidp, const struct nlattr *key,
1761 size_t key_len)
1762 {
1763 struct dp_netdev_flow *netdev_flow;
1764 struct flow flow;
1765 ovs_u128 ufid;
1766
1767 /* If a UFID is not provided, determine one based on the key. */
1768 if (!ufidp && key && key_len
1769 && !dpif_netdev_flow_from_nlattrs(key, key_len, &flow)) {
1770 dpif_flow_hash(pmd->dp->dpif, &flow, sizeof flow, &ufid);
1771 ufidp = &ufid;
1772 }
1773
1774 if (ufidp) {
1775 CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, dp_netdev_flow_hash(ufidp),
1776 &pmd->flow_table) {
1777 if (ovs_u128_equals(&netdev_flow->ufid, ufidp)) {
1778 return netdev_flow;
1779 }
1780 }
1781 }
1782
1783 return NULL;
1784 }
1785
1786 static void
1787 get_dpif_flow_stats(const struct dp_netdev_flow *netdev_flow_,
1788 struct dpif_flow_stats *stats)
1789 {
1790 struct dp_netdev_flow *netdev_flow;
1791 unsigned long long n;
1792 long long used;
1793 uint16_t flags;
1794
1795 netdev_flow = CONST_CAST(struct dp_netdev_flow *, netdev_flow_);
1796
1797 atomic_read_relaxed(&netdev_flow->stats.packet_count, &n);
1798 stats->n_packets = n;
1799 atomic_read_relaxed(&netdev_flow->stats.byte_count, &n);
1800 stats->n_bytes = n;
1801 atomic_read_relaxed(&netdev_flow->stats.used, &used);
1802 stats->used = used;
1803 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
1804 stats->tcp_flags = flags;
1805 }
1806
1807 /* Converts to the dpif_flow format, using 'key_buf' and 'mask_buf' for
1808 * storing the netlink-formatted key/mask. 'key_buf' may be the same as
1809 * 'mask_buf'. Actions will be returned without copying, by relying on RCU to
1810 * protect them. */
1811 static void
1812 dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow *netdev_flow,
1813 struct ofpbuf *key_buf, struct ofpbuf *mask_buf,
1814 struct dpif_flow *flow, bool terse)
1815 {
1816 if (terse) {
1817 memset(flow, 0, sizeof *flow);
1818 } else {
1819 struct flow_wildcards wc;
1820 struct dp_netdev_actions *actions;
1821 size_t offset;
1822 struct odp_flow_key_parms odp_parms = {
1823 .flow = &netdev_flow->flow,
1824 .mask = &wc.masks,
1825 .support = dp_netdev_support,
1826 };
1827
1828 miniflow_expand(&netdev_flow->cr.mask->mf, &wc.masks);
1829
1830 /* Key */
1831 offset = key_buf->size;
1832 flow->key = ofpbuf_tail(key_buf);
1833 odp_parms.odp_in_port = netdev_flow->flow.in_port.odp_port;
1834 odp_flow_key_from_flow(&odp_parms, key_buf);
1835 flow->key_len = key_buf->size - offset;
1836
1837 /* Mask */
1838 offset = mask_buf->size;
1839 flow->mask = ofpbuf_tail(mask_buf);
1840 odp_parms.odp_in_port = wc.masks.in_port.odp_port;
1841 odp_parms.key_buf = key_buf;
1842 odp_flow_key_from_mask(&odp_parms, mask_buf);
1843 flow->mask_len = mask_buf->size - offset;
1844
1845 /* Actions */
1846 actions = dp_netdev_flow_get_actions(netdev_flow);
1847 flow->actions = actions->actions;
1848 flow->actions_len = actions->size;
1849 }
1850
1851 flow->ufid = netdev_flow->ufid;
1852 flow->ufid_present = true;
1853 flow->pmd_id = netdev_flow->pmd_id;
1854 get_dpif_flow_stats(netdev_flow, &flow->stats);
1855 }
1856
1857 static int
1858 dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1859 const struct nlattr *mask_key,
1860 uint32_t mask_key_len, const struct flow *flow,
1861 struct flow *mask)
1862 {
1863 if (mask_key_len) {
1864 enum odp_key_fitness fitness;
1865
1866 fitness = odp_flow_key_to_mask(mask_key, mask_key_len, key, key_len,
1867 mask, flow);
1868 if (fitness) {
1869 /* This should not happen: it indicates that
1870 * odp_flow_key_from_mask() and odp_flow_key_to_mask()
1871 * disagree on the acceptable form of a mask. Log the problem
1872 * as an error, with enough details to enable debugging. */
1873 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1874
1875 if (!VLOG_DROP_ERR(&rl)) {
1876 struct ds s;
1877
1878 ds_init(&s);
1879 odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
1880 true);
1881 VLOG_ERR("internal error parsing flow mask %s (%s)",
1882 ds_cstr(&s), odp_key_fitness_to_string(fitness));
1883 ds_destroy(&s);
1884 }
1885
1886 return EINVAL;
1887 }
1888 } else {
1889 enum mf_field_id id;
1890 /* No mask key, unwildcard everything except fields whose
1891 * prerequisities are not met. */
1892 memset(mask, 0x0, sizeof *mask);
1893
1894 for (id = 0; id < MFF_N_IDS; ++id) {
1895 /* Skip registers and metadata. */
1896 if (!(id >= MFF_REG0 && id < MFF_REG0 + FLOW_N_REGS)
1897 && id != MFF_METADATA) {
1898 const struct mf_field *mf = mf_from_id(id);
1899 if (mf_are_prereqs_ok(mf, flow)) {
1900 mf_mask_field(mf, mask);
1901 }
1902 }
1903 }
1904 }
1905
1906 /* Force unwildcard the in_port.
1907 *
1908 * We need to do this even in the case where we unwildcard "everything"
1909 * above because "everything" only includes the 16-bit OpenFlow port number
1910 * mask->in_port.ofp_port, which only covers half of the 32-bit datapath
1911 * port number mask->in_port.odp_port. */
1912 mask->in_port.odp_port = u32_to_odp(UINT32_MAX);
1913
1914 return 0;
1915 }
1916
1917 static int
1918 dpif_netdev_flow_from_nlattrs(const struct nlattr *key, uint32_t key_len,
1919 struct flow *flow)
1920 {
1921 odp_port_t in_port;
1922
1923 if (odp_flow_key_to_flow(key, key_len, flow)) {
1924 /* This should not happen: it indicates that odp_flow_key_from_flow()
1925 * and odp_flow_key_to_flow() disagree on the acceptable form of a
1926 * flow. Log the problem as an error, with enough details to enable
1927 * debugging. */
1928 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
1929
1930 if (!VLOG_DROP_ERR(&rl)) {
1931 struct ds s;
1932
1933 ds_init(&s);
1934 odp_flow_format(key, key_len, NULL, 0, NULL, &s, true);
1935 VLOG_ERR("internal error parsing flow key %s", ds_cstr(&s));
1936 ds_destroy(&s);
1937 }
1938
1939 return EINVAL;
1940 }
1941
1942 in_port = flow->in_port.odp_port;
1943 if (!is_valid_port_number(in_port) && in_port != ODPP_NONE) {
1944 return EINVAL;
1945 }
1946
1947 return 0;
1948 }
1949
1950 static int
1951 dpif_netdev_flow_get(const struct dpif *dpif, const struct dpif_flow_get *get)
1952 {
1953 struct dp_netdev *dp = get_dp_netdev(dpif);
1954 struct dp_netdev_flow *netdev_flow;
1955 struct dp_netdev_pmd_thread *pmd;
1956 unsigned pmd_id = get->pmd_id == PMD_ID_NULL
1957 ? NON_PMD_CORE_ID : get->pmd_id;
1958 int error = 0;
1959
1960 pmd = dp_netdev_get_pmd(dp, pmd_id);
1961 if (!pmd) {
1962 return EINVAL;
1963 }
1964
1965 netdev_flow = dp_netdev_pmd_find_flow(pmd, get->ufid, get->key,
1966 get->key_len);
1967 if (netdev_flow) {
1968 dp_netdev_flow_to_dpif_flow(netdev_flow, get->buffer, get->buffer,
1969 get->flow, false);
1970 } else {
1971 error = ENOENT;
1972 }
1973 dp_netdev_pmd_unref(pmd);
1974
1975
1976 return error;
1977 }
1978
1979 static struct dp_netdev_flow *
1980 dp_netdev_flow_add(struct dp_netdev_pmd_thread *pmd,
1981 struct match *match, const ovs_u128 *ufid,
1982 const struct nlattr *actions, size_t actions_len)
1983 OVS_REQUIRES(pmd->flow_mutex)
1984 {
1985 struct dp_netdev_flow *flow;
1986 struct netdev_flow_key mask;
1987
1988 netdev_flow_mask_init(&mask, match);
1989 /* Make sure wc does not have metadata. */
1990 ovs_assert(!(mask.mf.map & (MINIFLOW_MAP(metadata) | MINIFLOW_MAP(regs))));
1991
1992 /* Do not allocate extra space. */
1993 flow = xmalloc(sizeof *flow - sizeof flow->cr.flow.mf + mask.len);
1994 memset(&flow->stats, 0, sizeof flow->stats);
1995 flow->dead = false;
1996 flow->batch = NULL;
1997 *CONST_CAST(unsigned *, &flow->pmd_id) = pmd->core_id;
1998 *CONST_CAST(struct flow *, &flow->flow) = match->flow;
1999 *CONST_CAST(ovs_u128 *, &flow->ufid) = *ufid;
2000 ovs_refcount_init(&flow->ref_cnt);
2001 ovsrcu_set(&flow->actions, dp_netdev_actions_create(actions, actions_len));
2002
2003 netdev_flow_key_init_masked(&flow->cr.flow, &match->flow, &mask);
2004 dpcls_insert(&pmd->cls, &flow->cr, &mask);
2005
2006 cmap_insert(&pmd->flow_table, CONST_CAST(struct cmap_node *, &flow->node),
2007 dp_netdev_flow_hash(&flow->ufid));
2008
2009 if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
2010 struct match match;
2011 struct ds ds = DS_EMPTY_INITIALIZER;
2012
2013 match.flow = flow->flow;
2014 miniflow_expand(&flow->cr.mask->mf, &match.wc.masks);
2015
2016 ds_put_cstr(&ds, "flow_add: ");
2017 odp_format_ufid(ufid, &ds);
2018 ds_put_cstr(&ds, " ");
2019 match_format(&match, &ds, OFP_DEFAULT_PRIORITY);
2020 ds_put_cstr(&ds, ", actions:");
2021 format_odp_actions(&ds, actions, actions_len);
2022
2023 VLOG_DBG_RL(&upcall_rl, "%s", ds_cstr(&ds));
2024
2025 ds_destroy(&ds);
2026 }
2027
2028 return flow;
2029 }
2030
2031 static int
2032 dpif_netdev_flow_put(struct dpif *dpif, const struct dpif_flow_put *put)
2033 {
2034 struct dp_netdev *dp = get_dp_netdev(dpif);
2035 struct dp_netdev_flow *netdev_flow;
2036 struct netdev_flow_key key;
2037 struct dp_netdev_pmd_thread *pmd;
2038 struct match match;
2039 ovs_u128 ufid;
2040 unsigned pmd_id = put->pmd_id == PMD_ID_NULL
2041 ? NON_PMD_CORE_ID : put->pmd_id;
2042 int error;
2043
2044 error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow);
2045 if (error) {
2046 return error;
2047 }
2048 error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
2049 put->mask, put->mask_len,
2050 &match.flow, &match.wc.masks);
2051 if (error) {
2052 return error;
2053 }
2054
2055 pmd = dp_netdev_get_pmd(dp, pmd_id);
2056 if (!pmd) {
2057 return EINVAL;
2058 }
2059
2060 /* Must produce a netdev_flow_key for lookup.
2061 * This interface is no longer performance critical, since it is not used
2062 * for upcall processing any more. */
2063 netdev_flow_key_from_flow(&key, &match.flow);
2064
2065 if (put->ufid) {
2066 ufid = *put->ufid;
2067 } else {
2068 dpif_flow_hash(dpif, &match.flow, sizeof match.flow, &ufid);
2069 }
2070
2071 ovs_mutex_lock(&pmd->flow_mutex);
2072 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &key);
2073 if (!netdev_flow) {
2074 if (put->flags & DPIF_FP_CREATE) {
2075 if (cmap_count(&pmd->flow_table) < MAX_FLOWS) {
2076 if (put->stats) {
2077 memset(put->stats, 0, sizeof *put->stats);
2078 }
2079 dp_netdev_flow_add(pmd, &match, &ufid, put->actions,
2080 put->actions_len);
2081 error = 0;
2082 } else {
2083 error = EFBIG;
2084 }
2085 } else {
2086 error = ENOENT;
2087 }
2088 } else {
2089 if (put->flags & DPIF_FP_MODIFY
2090 && flow_equal(&match.flow, &netdev_flow->flow)) {
2091 struct dp_netdev_actions *new_actions;
2092 struct dp_netdev_actions *old_actions;
2093
2094 new_actions = dp_netdev_actions_create(put->actions,
2095 put->actions_len);
2096
2097 old_actions = dp_netdev_flow_get_actions(netdev_flow);
2098 ovsrcu_set(&netdev_flow->actions, new_actions);
2099
2100 if (put->stats) {
2101 get_dpif_flow_stats(netdev_flow, put->stats);
2102 }
2103 if (put->flags & DPIF_FP_ZERO_STATS) {
2104 /* XXX: The userspace datapath uses thread local statistics
2105 * (for flows), which should be updated only by the owning
2106 * thread. Since we cannot write on stats memory here,
2107 * we choose not to support this flag. Please note:
2108 * - This feature is currently used only by dpctl commands with
2109 * option --clear.
2110 * - Should the need arise, this operation can be implemented
2111 * by keeping a base value (to be update here) for each
2112 * counter, and subtracting it before outputting the stats */
2113 error = EOPNOTSUPP;
2114 }
2115
2116 ovsrcu_postpone(dp_netdev_actions_free, old_actions);
2117 } else if (put->flags & DPIF_FP_CREATE) {
2118 error = EEXIST;
2119 } else {
2120 /* Overlapping flow. */
2121 error = EINVAL;
2122 }
2123 }
2124 ovs_mutex_unlock(&pmd->flow_mutex);
2125 dp_netdev_pmd_unref(pmd);
2126
2127 return error;
2128 }
2129
2130 static int
2131 dpif_netdev_flow_del(struct dpif *dpif, const struct dpif_flow_del *del)
2132 {
2133 struct dp_netdev *dp = get_dp_netdev(dpif);
2134 struct dp_netdev_flow *netdev_flow;
2135 struct dp_netdev_pmd_thread *pmd;
2136 unsigned pmd_id = del->pmd_id == PMD_ID_NULL
2137 ? NON_PMD_CORE_ID : del->pmd_id;
2138 int error = 0;
2139
2140 pmd = dp_netdev_get_pmd(dp, pmd_id);
2141 if (!pmd) {
2142 return EINVAL;
2143 }
2144
2145 ovs_mutex_lock(&pmd->flow_mutex);
2146 netdev_flow = dp_netdev_pmd_find_flow(pmd, del->ufid, del->key,
2147 del->key_len);
2148 if (netdev_flow) {
2149 if (del->stats) {
2150 get_dpif_flow_stats(netdev_flow, del->stats);
2151 }
2152 dp_netdev_pmd_remove_flow(pmd, netdev_flow);
2153 } else {
2154 error = ENOENT;
2155 }
2156 ovs_mutex_unlock(&pmd->flow_mutex);
2157 dp_netdev_pmd_unref(pmd);
2158
2159 return error;
2160 }
2161
2162 struct dpif_netdev_flow_dump {
2163 struct dpif_flow_dump up;
2164 struct cmap_position poll_thread_pos;
2165 struct cmap_position flow_pos;
2166 struct dp_netdev_pmd_thread *cur_pmd;
2167 int status;
2168 struct ovs_mutex mutex;
2169 };
2170
2171 static struct dpif_netdev_flow_dump *
2172 dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
2173 {
2174 return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
2175 }
2176
2177 static struct dpif_flow_dump *
2178 dpif_netdev_flow_dump_create(const struct dpif *dpif_, bool terse)
2179 {
2180 struct dpif_netdev_flow_dump *dump;
2181
2182 dump = xzalloc(sizeof *dump);
2183 dpif_flow_dump_init(&dump->up, dpif_);
2184 dump->up.terse = terse;
2185 ovs_mutex_init(&dump->mutex);
2186
2187 return &dump->up;
2188 }
2189
2190 static int
2191 dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
2192 {
2193 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2194
2195 ovs_mutex_destroy(&dump->mutex);
2196 free(dump);
2197 return 0;
2198 }
2199
2200 struct dpif_netdev_flow_dump_thread {
2201 struct dpif_flow_dump_thread up;
2202 struct dpif_netdev_flow_dump *dump;
2203 struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
2204 struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
2205 };
2206
2207 static struct dpif_netdev_flow_dump_thread *
2208 dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
2209 {
2210 return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
2211 }
2212
2213 static struct dpif_flow_dump_thread *
2214 dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
2215 {
2216 struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
2217 struct dpif_netdev_flow_dump_thread *thread;
2218
2219 thread = xmalloc(sizeof *thread);
2220 dpif_flow_dump_thread_init(&thread->up, &dump->up);
2221 thread->dump = dump;
2222 return &thread->up;
2223 }
2224
2225 static void
2226 dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
2227 {
2228 struct dpif_netdev_flow_dump_thread *thread
2229 = dpif_netdev_flow_dump_thread_cast(thread_);
2230
2231 free(thread);
2232 }
2233
2234 static int
2235 dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
2236 struct dpif_flow *flows, int max_flows)
2237 {
2238 struct dpif_netdev_flow_dump_thread *thread
2239 = dpif_netdev_flow_dump_thread_cast(thread_);
2240 struct dpif_netdev_flow_dump *dump = thread->dump;
2241 struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
2242 int n_flows = 0;
2243 int i;
2244
2245 ovs_mutex_lock(&dump->mutex);
2246 if (!dump->status) {
2247 struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
2248 struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
2249 struct dp_netdev_pmd_thread *pmd = dump->cur_pmd;
2250 int flow_limit = MIN(max_flows, FLOW_DUMP_MAX_BATCH);
2251
2252 /* First call to dump_next(), extracts the first pmd thread.
2253 * If there is no pmd thread, returns immediately. */
2254 if (!pmd) {
2255 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2256 if (!pmd) {
2257 ovs_mutex_unlock(&dump->mutex);
2258 return n_flows;
2259
2260 }
2261 }
2262
2263 do {
2264 for (n_flows = 0; n_flows < flow_limit; n_flows++) {
2265 struct cmap_node *node;
2266
2267 node = cmap_next_position(&pmd->flow_table, &dump->flow_pos);
2268 if (!node) {
2269 break;
2270 }
2271 netdev_flows[n_flows] = CONTAINER_OF(node,
2272 struct dp_netdev_flow,
2273 node);
2274 }
2275 /* When finishing dumping the current pmd thread, moves to
2276 * the next. */
2277 if (n_flows < flow_limit) {
2278 memset(&dump->flow_pos, 0, sizeof dump->flow_pos);
2279 dp_netdev_pmd_unref(pmd);
2280 pmd = dp_netdev_pmd_get_next(dp, &dump->poll_thread_pos);
2281 if (!pmd) {
2282 dump->status = EOF;
2283 break;
2284 }
2285 }
2286 /* Keeps the reference to next caller. */
2287 dump->cur_pmd = pmd;
2288
2289 /* If the current dump is empty, do not exit the loop, since the
2290 * remaining pmds could have flows to be dumped. Just dumps again
2291 * on the new 'pmd'. */
2292 } while (!n_flows);
2293 }
2294 ovs_mutex_unlock(&dump->mutex);
2295
2296 for (i = 0; i < n_flows; i++) {
2297 struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
2298 struct odputil_keybuf *keybuf = &thread->keybuf[i];
2299 struct dp_netdev_flow *netdev_flow = netdev_flows[i];
2300 struct dpif_flow *f = &flows[i];
2301 struct ofpbuf key, mask;
2302
2303 ofpbuf_use_stack(&key, keybuf, sizeof *keybuf);
2304 ofpbuf_use_stack(&mask, maskbuf, sizeof *maskbuf);
2305 dp_netdev_flow_to_dpif_flow(netdev_flow, &key, &mask, f,
2306 dump->up.terse);
2307 }
2308
2309 return n_flows;
2310 }
2311
2312 static int
2313 dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
2314 OVS_NO_THREAD_SAFETY_ANALYSIS
2315 {
2316 struct dp_netdev *dp = get_dp_netdev(dpif);
2317 struct dp_netdev_pmd_thread *pmd;
2318 struct dp_packet *pp;
2319
2320 if (dp_packet_size(execute->packet) < ETH_HEADER_LEN ||
2321 dp_packet_size(execute->packet) > UINT16_MAX) {
2322 return EINVAL;
2323 }
2324
2325 /* Tries finding the 'pmd'. If NULL is returned, that means
2326 * the current thread is a non-pmd thread and should use
2327 * dp_netdev_get_pmd(dp, NON_PMD_CORE_ID). */
2328 pmd = ovsthread_getspecific(dp->per_pmd_key);
2329 if (!pmd) {
2330 pmd = dp_netdev_get_pmd(dp, NON_PMD_CORE_ID);
2331 }
2332
2333 /* If the current thread is non-pmd thread, acquires
2334 * the 'non_pmd_mutex'. */
2335 if (pmd->core_id == NON_PMD_CORE_ID) {
2336 ovs_mutex_lock(&dp->non_pmd_mutex);
2337 ovs_mutex_lock(&dp->port_mutex);
2338 }
2339
2340 pp = execute->packet;
2341 dp_netdev_execute_actions(pmd, &pp, 1, false, execute->actions,
2342 execute->actions_len);
2343 if (pmd->core_id == NON_PMD_CORE_ID) {
2344 dp_netdev_pmd_unref(pmd);
2345 ovs_mutex_unlock(&dp->port_mutex);
2346 ovs_mutex_unlock(&dp->non_pmd_mutex);
2347 }
2348
2349 return 0;
2350 }
2351
2352 static void
2353 dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
2354 {
2355 size_t i;
2356
2357 for (i = 0; i < n_ops; i++) {
2358 struct dpif_op *op = ops[i];
2359
2360 switch (op->type) {
2361 case DPIF_OP_FLOW_PUT:
2362 op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put);
2363 break;
2364
2365 case DPIF_OP_FLOW_DEL:
2366 op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del);
2367 break;
2368
2369 case DPIF_OP_EXECUTE:
2370 op->error = dpif_netdev_execute(dpif, &op->u.execute);
2371 break;
2372
2373 case DPIF_OP_FLOW_GET:
2374 op->error = dpif_netdev_flow_get(dpif, &op->u.flow_get);
2375 break;
2376 }
2377 }
2378 }
2379
2380 /* Returns true if the configuration for rx queues or cpu mask
2381 * is changed. */
2382 static bool
2383 pmd_config_changed(const struct dp_netdev *dp, size_t rxqs, const char *cmask)
2384 {
2385 if (dp->n_dpdk_rxqs != rxqs) {
2386 return true;
2387 } else {
2388 if (dp->pmd_cmask != NULL && cmask != NULL) {
2389 return strcmp(dp->pmd_cmask, cmask);
2390 } else {
2391 return (dp->pmd_cmask != NULL || cmask != NULL);
2392 }
2393 }
2394 }
2395
2396 /* Resets pmd threads if the configuration for 'rxq's or cpu mask changes. */
2397 static int
2398 dpif_netdev_pmd_set(struct dpif *dpif, unsigned int n_rxqs, const char *cmask)
2399 {
2400 struct dp_netdev *dp = get_dp_netdev(dpif);
2401
2402 if (pmd_config_changed(dp, n_rxqs, cmask)) {
2403 struct dp_netdev_port *port;
2404
2405 dp_netdev_destroy_all_pmds(dp);
2406
2407 CMAP_FOR_EACH (port, node, &dp->ports) {
2408 if (netdev_is_pmd(port->netdev)) {
2409 int i, err;
2410
2411 /* Closes the existing 'rxq's. */
2412 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2413 netdev_rxq_close(port->rxq[i]);
2414 port->rxq[i] = NULL;
2415 }
2416
2417 /* Sets the new rx queue config. */
2418 err = netdev_set_multiq(port->netdev,
2419 ovs_numa_get_n_cores() + 1,
2420 n_rxqs);
2421 if (err && (err != EOPNOTSUPP)) {
2422 VLOG_ERR("Failed to set dpdk interface %s rx_queue to:"
2423 " %u", netdev_get_name(port->netdev),
2424 n_rxqs);
2425 return err;
2426 }
2427
2428 /* If the set_multiq() above succeeds, reopens the 'rxq's. */
2429 port->rxq = xrealloc(port->rxq, sizeof *port->rxq
2430 * netdev_n_rxq(port->netdev));
2431 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2432 netdev_rxq_open(port->netdev, &port->rxq[i], i);
2433 }
2434 }
2435 }
2436 dp->n_dpdk_rxqs = n_rxqs;
2437
2438 /* Reconfigures the cpu mask. */
2439 ovs_numa_set_cpu_mask(cmask);
2440 free(dp->pmd_cmask);
2441 dp->pmd_cmask = cmask ? xstrdup(cmask) : NULL;
2442
2443 /* Restores the non-pmd. */
2444 dp_netdev_set_nonpmd(dp);
2445 /* Restores all pmd threads. */
2446 dp_netdev_reset_pmd_threads(dp);
2447 }
2448
2449 return 0;
2450 }
2451
2452 static int
2453 dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
2454 uint32_t queue_id, uint32_t *priority)
2455 {
2456 *priority = queue_id;
2457 return 0;
2458 }
2459
2460 \f
2461 /* Creates and returns a new 'struct dp_netdev_actions', whose actions are
2462 * a copy of the 'ofpacts_len' bytes of 'ofpacts'. */
2463 struct dp_netdev_actions *
2464 dp_netdev_actions_create(const struct nlattr *actions, size_t size)
2465 {
2466 struct dp_netdev_actions *netdev_actions;
2467
2468 netdev_actions = xmalloc(sizeof *netdev_actions + size);
2469 memcpy(netdev_actions->actions, actions, size);
2470 netdev_actions->size = size;
2471
2472 return netdev_actions;
2473 }
2474
2475 struct dp_netdev_actions *
2476 dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
2477 {
2478 return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
2479 }
2480
2481 static void
2482 dp_netdev_actions_free(struct dp_netdev_actions *actions)
2483 {
2484 free(actions);
2485 }
2486 \f
2487 static inline unsigned long long
2488 cycles_counter(void)
2489 {
2490 #ifdef DPDK_NETDEV
2491 return rte_get_tsc_cycles();
2492 #else
2493 return 0;
2494 #endif
2495 }
2496
2497 /* Fake mutex to make sure that the calls to cycles_count_* are balanced */
2498 extern struct ovs_mutex cycles_counter_fake_mutex;
2499
2500 /* Start counting cycles. Must be followed by 'cycles_count_end()' */
2501 static inline void
2502 cycles_count_start(struct dp_netdev_pmd_thread *pmd)
2503 OVS_ACQUIRES(&cycles_counter_fake_mutex)
2504 OVS_NO_THREAD_SAFETY_ANALYSIS
2505 {
2506 pmd->last_cycles = cycles_counter();
2507 }
2508
2509 /* Stop counting cycles and add them to the counter 'type' */
2510 static inline void
2511 cycles_count_end(struct dp_netdev_pmd_thread *pmd,
2512 enum pmd_cycles_counter_type type)
2513 OVS_RELEASES(&cycles_counter_fake_mutex)
2514 OVS_NO_THREAD_SAFETY_ANALYSIS
2515 {
2516 unsigned long long interval = cycles_counter() - pmd->last_cycles;
2517
2518 non_atomic_ullong_add(&pmd->cycles.n[type], interval);
2519 }
2520
2521 static void
2522 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd,
2523 struct dp_netdev_port *port,
2524 struct netdev_rxq *rxq)
2525 {
2526 struct dp_packet *packets[NETDEV_MAX_BURST];
2527 int error, cnt;
2528
2529 cycles_count_start(pmd);
2530 error = netdev_rxq_recv(rxq, packets, &cnt);
2531 cycles_count_end(pmd, PMD_CYCLES_POLLING);
2532 if (!error) {
2533 int i;
2534
2535 *recirc_depth_get() = 0;
2536
2537 /* XXX: initialize md in netdev implementation. */
2538 for (i = 0; i < cnt; i++) {
2539 pkt_metadata_init(&packets[i]->md, port->port_no);
2540 }
2541 cycles_count_start(pmd);
2542 dp_netdev_input(pmd, packets, cnt);
2543 cycles_count_end(pmd, PMD_CYCLES_PROCESSING);
2544 } else if (error != EAGAIN && error != EOPNOTSUPP) {
2545 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
2546
2547 VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
2548 netdev_get_name(port->netdev), ovs_strerror(error));
2549 }
2550 }
2551
2552 /* Return true if needs to revalidate datapath flows. */
2553 static bool
2554 dpif_netdev_run(struct dpif *dpif)
2555 {
2556 struct dp_netdev_port *port;
2557 struct dp_netdev *dp = get_dp_netdev(dpif);
2558 struct dp_netdev_pmd_thread *non_pmd = dp_netdev_get_pmd(dp,
2559 NON_PMD_CORE_ID);
2560 uint64_t new_tnl_seq;
2561
2562 ovs_mutex_lock(&dp->non_pmd_mutex);
2563 CMAP_FOR_EACH (port, node, &dp->ports) {
2564 if (!netdev_is_pmd(port->netdev)) {
2565 int i;
2566
2567 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2568 dp_netdev_process_rxq_port(non_pmd, port, port->rxq[i]);
2569 }
2570 }
2571 }
2572 ovs_mutex_unlock(&dp->non_pmd_mutex);
2573 dp_netdev_pmd_unref(non_pmd);
2574
2575 tnl_arp_cache_run();
2576 new_tnl_seq = seq_read(tnl_conf_seq);
2577
2578 if (dp->last_tnl_conf_seq != new_tnl_seq) {
2579 dp->last_tnl_conf_seq = new_tnl_seq;
2580 return true;
2581 }
2582 return false;
2583 }
2584
2585 static void
2586 dpif_netdev_wait(struct dpif *dpif)
2587 {
2588 struct dp_netdev_port *port;
2589 struct dp_netdev *dp = get_dp_netdev(dpif);
2590
2591 ovs_mutex_lock(&dp_netdev_mutex);
2592 CMAP_FOR_EACH (port, node, &dp->ports) {
2593 if (!netdev_is_pmd(port->netdev)) {
2594 int i;
2595
2596 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2597 netdev_rxq_wait(port->rxq[i]);
2598 }
2599 }
2600 }
2601 ovs_mutex_unlock(&dp_netdev_mutex);
2602 seq_wait(tnl_conf_seq, dp->last_tnl_conf_seq);
2603 }
2604
2605 struct rxq_poll {
2606 struct dp_netdev_port *port;
2607 struct netdev_rxq *rx;
2608 };
2609
2610 static int
2611 pmd_load_queues(struct dp_netdev_pmd_thread *pmd,
2612 struct rxq_poll **ppoll_list, int poll_cnt)
2613 {
2614 struct rxq_poll *poll_list = *ppoll_list;
2615 struct dp_netdev_port *port;
2616 int n_pmds_on_numa, index, i;
2617
2618 /* Simple scheduler for netdev rx polling. */
2619 for (i = 0; i < poll_cnt; i++) {
2620 port_unref(poll_list[i].port);
2621 }
2622
2623 poll_cnt = 0;
2624 n_pmds_on_numa = get_n_pmd_threads_on_numa(pmd->dp, pmd->numa_id);
2625 index = 0;
2626
2627 CMAP_FOR_EACH (port, node, &pmd->dp->ports) {
2628 /* Calls port_try_ref() to prevent the main thread
2629 * from deleting the port. */
2630 if (port_try_ref(port)) {
2631 if (netdev_is_pmd(port->netdev)
2632 && netdev_get_numa_id(port->netdev) == pmd->numa_id) {
2633 int i;
2634
2635 for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
2636 if ((index % n_pmds_on_numa) == pmd->index) {
2637 poll_list = xrealloc(poll_list,
2638 sizeof *poll_list * (poll_cnt + 1));
2639
2640 port_ref(port);
2641 poll_list[poll_cnt].port = port;
2642 poll_list[poll_cnt].rx = port->rxq[i];
2643 poll_cnt++;
2644 }
2645 index++;
2646 }
2647 }
2648 /* Unrefs the port_try_ref(). */
2649 port_unref(port);
2650 }
2651 }
2652
2653 *ppoll_list = poll_list;
2654 return poll_cnt;
2655 }
2656
2657 static void *
2658 pmd_thread_main(void *f_)
2659 {
2660 struct dp_netdev_pmd_thread *pmd = f_;
2661 unsigned int lc = 0;
2662 struct rxq_poll *poll_list;
2663 unsigned int port_seq = PMD_INITIAL_SEQ;
2664 int poll_cnt;
2665 int i;
2666
2667 poll_cnt = 0;
2668 poll_list = NULL;
2669
2670 /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
2671 ovsthread_setspecific(pmd->dp->per_pmd_key, pmd);
2672 pmd_thread_setaffinity_cpu(pmd->core_id);
2673 reload:
2674 emc_cache_init(&pmd->flow_cache);
2675 poll_cnt = pmd_load_queues(pmd, &poll_list, poll_cnt);
2676
2677 /* List port/core affinity */
2678 for (i = 0; i < poll_cnt; i++) {
2679 VLOG_INFO("Core %d processing port \'%s\'\n", pmd->core_id, netdev_get_name(poll_list[i].port->netdev));
2680 }
2681
2682 /* Signal here to make sure the pmd finishes
2683 * reloading the updated configuration. */
2684 dp_netdev_pmd_reload_done(pmd);
2685
2686 for (;;) {
2687 int i;
2688
2689 for (i = 0; i < poll_cnt; i++) {
2690 dp_netdev_process_rxq_port(pmd, poll_list[i].port, poll_list[i].rx);
2691 }
2692
2693 if (lc++ > 1024) {
2694 unsigned int seq;
2695
2696 lc = 0;
2697
2698 emc_cache_slow_sweep(&pmd->flow_cache);
2699 ovsrcu_quiesce();
2700
2701 atomic_read_relaxed(&pmd->change_seq, &seq);
2702 if (seq != port_seq) {
2703 port_seq = seq;
2704 break;
2705 }
2706 }
2707 }
2708
2709 emc_cache_uninit(&pmd->flow_cache);
2710
2711 if (!latch_is_set(&pmd->exit_latch)){
2712 goto reload;
2713 }
2714
2715 for (i = 0; i < poll_cnt; i++) {
2716 port_unref(poll_list[i].port);
2717 }
2718
2719 dp_netdev_pmd_reload_done(pmd);
2720
2721 free(poll_list);
2722 return NULL;
2723 }
2724
2725 static void
2726 dp_netdev_disable_upcall(struct dp_netdev *dp)
2727 OVS_ACQUIRES(dp->upcall_rwlock)
2728 {
2729 fat_rwlock_wrlock(&dp->upcall_rwlock);
2730 }
2731
2732 static void
2733 dpif_netdev_disable_upcall(struct dpif *dpif)
2734 OVS_NO_THREAD_SAFETY_ANALYSIS
2735 {
2736 struct dp_netdev *dp = get_dp_netdev(dpif);
2737 dp_netdev_disable_upcall(dp);
2738 }
2739
2740 static void
2741 dp_netdev_enable_upcall(struct dp_netdev *dp)
2742 OVS_RELEASES(dp->upcall_rwlock)
2743 {
2744 fat_rwlock_unlock(&dp->upcall_rwlock);
2745 }
2746
2747 static void
2748 dpif_netdev_enable_upcall(struct dpif *dpif)
2749 OVS_NO_THREAD_SAFETY_ANALYSIS
2750 {
2751 struct dp_netdev *dp = get_dp_netdev(dpif);
2752 dp_netdev_enable_upcall(dp);
2753 }
2754
2755 void
2756 dp_netdev_pmd_reload_done(struct dp_netdev_pmd_thread *pmd)
2757 {
2758 ovs_mutex_lock(&pmd->cond_mutex);
2759 xpthread_cond_signal(&pmd->cond);
2760 ovs_mutex_unlock(&pmd->cond_mutex);
2761 }
2762
2763 /* Finds and refs the dp_netdev_pmd_thread on core 'core_id'. Returns
2764 * the pointer if succeeds, otherwise, NULL.
2765 *
2766 * Caller must unrefs the returned reference. */
2767 static struct dp_netdev_pmd_thread *
2768 dp_netdev_get_pmd(struct dp_netdev *dp, unsigned core_id)
2769 {
2770 struct dp_netdev_pmd_thread *pmd;
2771 const struct cmap_node *pnode;
2772
2773 pnode = cmap_find(&dp->poll_threads, hash_int(core_id, 0));
2774 if (!pnode) {
2775 return NULL;
2776 }
2777 pmd = CONTAINER_OF(pnode, struct dp_netdev_pmd_thread, node);
2778
2779 return dp_netdev_pmd_try_ref(pmd) ? pmd : NULL;
2780 }
2781
2782 /* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
2783 static void
2784 dp_netdev_set_nonpmd(struct dp_netdev *dp)
2785 {
2786 struct dp_netdev_pmd_thread *non_pmd;
2787
2788 non_pmd = xzalloc(sizeof *non_pmd);
2789 dp_netdev_configure_pmd(non_pmd, dp, 0, NON_PMD_CORE_ID,
2790 OVS_NUMA_UNSPEC);
2791 }
2792
2793 /* Caller must have valid pointer to 'pmd'. */
2794 static bool
2795 dp_netdev_pmd_try_ref(struct dp_netdev_pmd_thread *pmd)
2796 {
2797 return ovs_refcount_try_ref_rcu(&pmd->ref_cnt);
2798 }
2799
2800 static void
2801 dp_netdev_pmd_unref(struct dp_netdev_pmd_thread *pmd)
2802 {
2803 if (pmd && ovs_refcount_unref(&pmd->ref_cnt) == 1) {
2804 ovsrcu_postpone(dp_netdev_destroy_pmd, pmd);
2805 }
2806 }
2807
2808 /* Given cmap position 'pos', tries to ref the next node. If try_ref()
2809 * fails, keeps checking for next node until reaching the end of cmap.
2810 *
2811 * Caller must unrefs the returned reference. */
2812 static struct dp_netdev_pmd_thread *
2813 dp_netdev_pmd_get_next(struct dp_netdev *dp, struct cmap_position *pos)
2814 {
2815 struct dp_netdev_pmd_thread *next;
2816
2817 do {
2818 struct cmap_node *node;
2819
2820 node = cmap_next_position(&dp->poll_threads, pos);
2821 next = node ? CONTAINER_OF(node, struct dp_netdev_pmd_thread, node)
2822 : NULL;
2823 } while (next && !dp_netdev_pmd_try_ref(next));
2824
2825 return next;
2826 }
2827
2828 static int
2829 core_id_to_qid(unsigned core_id)
2830 {
2831 if (core_id != NON_PMD_CORE_ID) {
2832 return core_id;
2833 } else {
2834 return ovs_numa_get_n_cores();
2835 }
2836 }
2837
2838 /* Configures the 'pmd' based on the input argument. */
2839 static void
2840 dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp,
2841 int index, unsigned core_id, int numa_id)
2842 {
2843 pmd->dp = dp;
2844 pmd->index = index;
2845 pmd->core_id = core_id;
2846 pmd->tx_qid = core_id_to_qid(core_id);
2847 pmd->numa_id = numa_id;
2848
2849 ovs_refcount_init(&pmd->ref_cnt);
2850 latch_init(&pmd->exit_latch);
2851 atomic_init(&pmd->change_seq, PMD_INITIAL_SEQ);
2852 xpthread_cond_init(&pmd->cond, NULL);
2853 ovs_mutex_init(&pmd->cond_mutex);
2854 ovs_mutex_init(&pmd->flow_mutex);
2855 dpcls_init(&pmd->cls);
2856 cmap_init(&pmd->flow_table);
2857 /* init the 'flow_cache' since there is no
2858 * actual thread created for NON_PMD_CORE_ID. */
2859 if (core_id == NON_PMD_CORE_ID) {
2860 emc_cache_init(&pmd->flow_cache);
2861 }
2862 cmap_insert(&dp->poll_threads, CONST_CAST(struct cmap_node *, &pmd->node),
2863 hash_int(core_id, 0));
2864 }
2865
2866 static void
2867 dp_netdev_destroy_pmd(struct dp_netdev_pmd_thread *pmd)
2868 {
2869 dp_netdev_pmd_flow_flush(pmd);
2870 dpcls_destroy(&pmd->cls);
2871 cmap_destroy(&pmd->flow_table);
2872 ovs_mutex_destroy(&pmd->flow_mutex);
2873 latch_destroy(&pmd->exit_latch);
2874 xpthread_cond_destroy(&pmd->cond);
2875 ovs_mutex_destroy(&pmd->cond_mutex);
2876 free(pmd);
2877 }
2878
2879 /* Stops the pmd thread, removes it from the 'dp->poll_threads',
2880 * and unrefs the struct. */
2881 static void
2882 dp_netdev_del_pmd(struct dp_netdev_pmd_thread *pmd)
2883 {
2884 /* Uninit the 'flow_cache' since there is
2885 * no actual thread uninit it for NON_PMD_CORE_ID. */
2886 if (pmd->core_id == NON_PMD_CORE_ID) {
2887 emc_cache_uninit(&pmd->flow_cache);
2888 } else {
2889 latch_set(&pmd->exit_latch);
2890 dp_netdev_reload_pmd__(pmd);
2891 ovs_numa_unpin_core(pmd->core_id);
2892 xpthread_join(pmd->thread, NULL);
2893 }
2894 cmap_remove(&pmd->dp->poll_threads, &pmd->node, hash_int(pmd->core_id, 0));
2895 dp_netdev_pmd_unref(pmd);
2896 }
2897
2898 /* Destroys all pmd threads. */
2899 static void
2900 dp_netdev_destroy_all_pmds(struct dp_netdev *dp)
2901 {
2902 struct dp_netdev_pmd_thread *pmd;
2903
2904 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2905 dp_netdev_del_pmd(pmd);
2906 }
2907 }
2908
2909 /* Deletes all pmd threads on numa node 'numa_id'. */
2910 static void
2911 dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2912 {
2913 struct dp_netdev_pmd_thread *pmd;
2914
2915 CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
2916 if (pmd->numa_id == numa_id) {
2917 dp_netdev_del_pmd(pmd);
2918 }
2919 }
2920 }
2921
2922 /* Checks the numa node id of 'netdev' and starts pmd threads for
2923 * the numa node. */
2924 static void
2925 dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id)
2926 {
2927 int n_pmds;
2928
2929 if (!ovs_numa_numa_id_is_valid(numa_id)) {
2930 VLOG_ERR("Cannot create pmd threads due to numa id (%d)"
2931 "invalid", numa_id);
2932 return ;
2933 }
2934
2935 n_pmds = get_n_pmd_threads_on_numa(dp, numa_id);
2936
2937 /* If there are already pmd threads created for the numa node
2938 * in which 'netdev' is on, do nothing. Else, creates the
2939 * pmd threads for the numa node. */
2940 if (!n_pmds) {
2941 int can_have, n_unpinned, i;
2942
2943 n_unpinned = ovs_numa_get_n_unpinned_cores_on_numa(numa_id);
2944 if (!n_unpinned) {
2945 VLOG_ERR("Cannot create pmd threads due to out of unpinned "
2946 "cores on numa node");
2947 return;
2948 }
2949
2950 /* If cpu mask is specified, uses all unpinned cores, otherwise
2951 * tries creating NR_PMD_THREADS pmd threads. */
2952 can_have = dp->pmd_cmask ? n_unpinned : MIN(n_unpinned, NR_PMD_THREADS);
2953 for (i = 0; i < can_have; i++) {
2954 struct dp_netdev_pmd_thread *pmd = xzalloc(sizeof *pmd);
2955 unsigned core_id = ovs_numa_get_unpinned_core_on_numa(numa_id);
2956
2957 dp_netdev_configure_pmd(pmd, dp, i, core_id, numa_id);
2958 /* Each thread will distribute all devices rx-queues among
2959 * themselves. */
2960 pmd->thread = ovs_thread_create("pmd", pmd_thread_main, pmd);
2961 }
2962 VLOG_INFO("Created %d pmd threads on numa node %d", can_have, numa_id);
2963 }
2964 }
2965
2966 \f
2967 /* Called after pmd threads config change. Restarts pmd threads with
2968 * new configuration. */
2969 static void
2970 dp_netdev_reset_pmd_threads(struct dp_netdev *dp)
2971 {
2972 struct dp_netdev_port *port;
2973
2974 CMAP_FOR_EACH (port, node, &dp->ports) {
2975 if (netdev_is_pmd(port->netdev)) {
2976 int numa_id = netdev_get_numa_id(port->netdev);
2977
2978 dp_netdev_set_pmds_on_numa(dp, numa_id);
2979 }
2980 }
2981 }
2982
2983 static char *
2984 dpif_netdev_get_datapath_version(void)
2985 {
2986 return xstrdup("<built-in>");
2987 }
2988
2989 static void
2990 dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow, int cnt, int size,
2991 uint16_t tcp_flags, long long now)
2992 {
2993 uint16_t flags;
2994
2995 atomic_store_relaxed(&netdev_flow->stats.used, now);
2996 non_atomic_ullong_add(&netdev_flow->stats.packet_count, cnt);
2997 non_atomic_ullong_add(&netdev_flow->stats.byte_count, size);
2998 atomic_read_relaxed(&netdev_flow->stats.tcp_flags, &flags);
2999 flags |= tcp_flags;
3000 atomic_store_relaxed(&netdev_flow->stats.tcp_flags, flags);
3001 }
3002
3003 static void
3004 dp_netdev_count_packet(struct dp_netdev_pmd_thread *pmd,
3005 enum dp_stat_type type, int cnt)
3006 {
3007 non_atomic_ullong_add(&pmd->stats.n[type], cnt);
3008 }
3009
3010 static int
3011 dp_netdev_upcall(struct dp_netdev_pmd_thread *pmd, struct dp_packet *packet_,
3012 struct flow *flow, struct flow_wildcards *wc, ovs_u128 *ufid,
3013 enum dpif_upcall_type type, const struct nlattr *userdata,
3014 struct ofpbuf *actions, struct ofpbuf *put_actions)
3015 {
3016 struct dp_netdev *dp = pmd->dp;
3017
3018 if (OVS_UNLIKELY(!dp->upcall_cb)) {
3019 return ENODEV;
3020 }
3021
3022 if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl))) {
3023 struct ds ds = DS_EMPTY_INITIALIZER;
3024 char *packet_str;
3025 struct ofpbuf key;
3026 struct odp_flow_key_parms odp_parms = {
3027 .flow = flow,
3028 .mask = &wc->masks,
3029 .odp_in_port = flow->in_port.odp_port,
3030 .support = dp_netdev_support,
3031 };
3032
3033 ofpbuf_init(&key, 0);
3034 odp_flow_key_from_flow(&odp_parms, &key);
3035 packet_str = ofp_packet_to_string(dp_packet_data(packet_),
3036 dp_packet_size(packet_));
3037
3038 odp_flow_key_format(key.data, key.size, &ds);
3039
3040 VLOG_DBG("%s: %s upcall:\n%s\n%s", dp->name,
3041 dpif_upcall_type_to_string(type), ds_cstr(&ds), packet_str);
3042
3043 ofpbuf_uninit(&key);
3044 free(packet_str);
3045
3046 ds_destroy(&ds);
3047 }
3048
3049 return dp->upcall_cb(packet_, flow, ufid, pmd->core_id, type, userdata,
3050 actions, wc, put_actions, dp->upcall_aux);
3051 }
3052
3053 static inline uint32_t
3054 dpif_netdev_packet_get_rss_hash(struct dp_packet *packet,
3055 const struct miniflow *mf)
3056 {
3057 uint32_t hash, recirc_depth;
3058
3059 hash = dp_packet_get_rss_hash(packet);
3060 if (OVS_UNLIKELY(!hash)) {
3061 hash = miniflow_hash_5tuple(mf, 0);
3062 dp_packet_set_rss_hash(packet, hash);
3063 }
3064
3065 /* The RSS hash must account for the recirculation depth to avoid
3066 * collisions in the exact match cache */
3067 recirc_depth = *recirc_depth_get_unsafe();
3068 if (OVS_UNLIKELY(recirc_depth)) {
3069 hash = hash_finish(hash, recirc_depth);
3070 dp_packet_set_rss_hash(packet, hash);
3071 }
3072 return hash;
3073 }
3074
3075 struct packet_batch {
3076 unsigned int packet_count;
3077 unsigned int byte_count;
3078 uint16_t tcp_flags;
3079
3080 struct dp_netdev_flow *flow;
3081
3082 struct dp_packet *packets[NETDEV_MAX_BURST];
3083 };
3084
3085 static inline void
3086 packet_batch_update(struct packet_batch *batch, struct dp_packet *packet,
3087 const struct miniflow *mf)
3088 {
3089 batch->tcp_flags |= miniflow_get_tcp_flags(mf);
3090 batch->packets[batch->packet_count++] = packet;
3091 batch->byte_count += dp_packet_size(packet);
3092 }
3093
3094 static inline void
3095 packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow)
3096 {
3097 flow->batch = batch;
3098
3099 batch->flow = flow;
3100 batch->packet_count = 0;
3101 batch->byte_count = 0;
3102 batch->tcp_flags = 0;
3103 }
3104
3105 static inline void
3106 packet_batch_execute(struct packet_batch *batch,
3107 struct dp_netdev_pmd_thread *pmd,
3108 long long now)
3109 {
3110 struct dp_netdev_actions *actions;
3111 struct dp_netdev_flow *flow = batch->flow;
3112
3113 dp_netdev_flow_used(flow, batch->packet_count, batch->byte_count,
3114 batch->tcp_flags, now);
3115
3116 actions = dp_netdev_flow_get_actions(flow);
3117
3118 dp_netdev_execute_actions(pmd, batch->packets, batch->packet_count, true,
3119 actions->actions, actions->size);
3120 }
3121
3122 static inline void
3123 dp_netdev_queue_batches(struct dp_packet *pkt,
3124 struct dp_netdev_flow *flow, const struct miniflow *mf,
3125 struct packet_batch *batches, size_t *n_batches)
3126 {
3127 struct packet_batch *batch = flow->batch;
3128
3129 if (OVS_LIKELY(batch)) {
3130 packet_batch_update(batch, pkt, mf);
3131 return;
3132 }
3133
3134 batch = &batches[(*n_batches)++];
3135 packet_batch_init(batch, flow);
3136 packet_batch_update(batch, pkt, mf);
3137 }
3138
3139 static inline void
3140 dp_packet_swap(struct dp_packet **a, struct dp_packet **b)
3141 {
3142 struct dp_packet *tmp = *a;
3143 *a = *b;
3144 *b = tmp;
3145 }
3146
3147 /* Try to process all ('cnt') the 'packets' using only the exact match cache
3148 * 'flow_cache'. If a flow is not found for a packet 'packets[i]', the
3149 * miniflow is copied into 'keys' and the packet pointer is moved at the
3150 * beginning of the 'packets' array.
3151 *
3152 * The function returns the number of packets that needs to be processed in the
3153 * 'packets' array (they have been moved to the beginning of the vector).
3154 */
3155 static inline size_t
3156 emc_processing(struct dp_netdev_pmd_thread *pmd, struct dp_packet **packets,
3157 size_t cnt, struct netdev_flow_key *keys,
3158 struct packet_batch batches[], size_t *n_batches)
3159 {
3160 struct emc_cache *flow_cache = &pmd->flow_cache;
3161 struct netdev_flow_key key;
3162 size_t i, notfound_cnt = 0;
3163
3164 for (i = 0; i < cnt; i++) {
3165 struct dp_netdev_flow *flow;
3166
3167 if (OVS_UNLIKELY(dp_packet_size(packets[i]) < ETH_HEADER_LEN)) {
3168 dp_packet_delete(packets[i]);
3169 continue;
3170 }
3171
3172 if (i != cnt - 1) {
3173 /* Prefetch next packet data */
3174 OVS_PREFETCH(dp_packet_data(packets[i+1]));
3175 }
3176
3177 miniflow_extract(packets[i], &key.mf);
3178 key.len = 0; /* Not computed yet. */
3179 key.hash = dpif_netdev_packet_get_rss_hash(packets[i], &key.mf);
3180
3181 flow = emc_lookup(flow_cache, &key);
3182 if (OVS_LIKELY(flow)) {
3183 dp_netdev_queue_batches(packets[i], flow, &key.mf, batches,
3184 n_batches);
3185 } else {
3186 if (i != notfound_cnt) {
3187 dp_packet_swap(&packets[i], &packets[notfound_cnt]);
3188 }
3189
3190 keys[notfound_cnt++] = key;
3191 }
3192 }
3193
3194 dp_netdev_count_packet(pmd, DP_STAT_EXACT_HIT, cnt - notfound_cnt);
3195
3196 return notfound_cnt;
3197 }
3198
3199 static inline void
3200 fast_path_processing(struct dp_netdev_pmd_thread *pmd,
3201 struct dp_packet **packets, size_t cnt,
3202 struct netdev_flow_key *keys,
3203 struct packet_batch batches[], size_t *n_batches)
3204 {
3205 #if !defined(__CHECKER__) && !defined(_WIN32)
3206 const size_t PKT_ARRAY_SIZE = cnt;
3207 #else
3208 /* Sparse or MSVC doesn't like variable length array. */
3209 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3210 #endif
3211 struct dpcls_rule *rules[PKT_ARRAY_SIZE];
3212 struct dp_netdev *dp = pmd->dp;
3213 struct emc_cache *flow_cache = &pmd->flow_cache;
3214 int miss_cnt = 0, lost_cnt = 0;
3215 bool any_miss;
3216 size_t i;
3217
3218 for (i = 0; i < cnt; i++) {
3219 /* Key length is needed in all the cases, hash computed on demand. */
3220 keys[i].len = netdev_flow_key_size(count_1bits(keys[i].mf.map));
3221 }
3222 any_miss = !dpcls_lookup(&pmd->cls, keys, rules, cnt);
3223 if (OVS_UNLIKELY(any_miss) && !fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3224 uint64_t actions_stub[512 / 8], slow_stub[512 / 8];
3225 struct ofpbuf actions, put_actions;
3226 ovs_u128 ufid;
3227
3228 ofpbuf_use_stub(&actions, actions_stub, sizeof actions_stub);
3229 ofpbuf_use_stub(&put_actions, slow_stub, sizeof slow_stub);
3230
3231 for (i = 0; i < cnt; i++) {
3232 struct dp_netdev_flow *netdev_flow;
3233 struct ofpbuf *add_actions;
3234 struct match match;
3235 int error;
3236
3237 if (OVS_LIKELY(rules[i])) {
3238 continue;
3239 }
3240
3241 /* It's possible that an earlier slow path execution installed
3242 * a rule covering this flow. In this case, it's a lot cheaper
3243 * to catch it here than execute a miss. */
3244 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3245 if (netdev_flow) {
3246 rules[i] = &netdev_flow->cr;
3247 continue;
3248 }
3249
3250 miss_cnt++;
3251
3252 miniflow_expand(&keys[i].mf, &match.flow);
3253
3254 ofpbuf_clear(&actions);
3255 ofpbuf_clear(&put_actions);
3256
3257 dpif_flow_hash(dp->dpif, &match.flow, sizeof match.flow, &ufid);
3258 error = dp_netdev_upcall(pmd, packets[i], &match.flow, &match.wc,
3259 &ufid, DPIF_UC_MISS, NULL, &actions,
3260 &put_actions);
3261 if (OVS_UNLIKELY(error && error != ENOSPC)) {
3262 dp_packet_delete(packets[i]);
3263 lost_cnt++;
3264 continue;
3265 }
3266
3267 /* We can't allow the packet batching in the next loop to execute
3268 * the actions. Otherwise, if there are any slow path actions,
3269 * we'll send the packet up twice. */
3270 dp_netdev_execute_actions(pmd, &packets[i], 1, true,
3271 actions.data, actions.size);
3272
3273 add_actions = put_actions.size ? &put_actions : &actions;
3274 if (OVS_LIKELY(error != ENOSPC)) {
3275 /* XXX: There's a race window where a flow covering this packet
3276 * could have already been installed since we last did the flow
3277 * lookup before upcall. This could be solved by moving the
3278 * mutex lock outside the loop, but that's an awful long time
3279 * to be locking everyone out of making flow installs. If we
3280 * move to a per-core classifier, it would be reasonable. */
3281 ovs_mutex_lock(&pmd->flow_mutex);
3282 netdev_flow = dp_netdev_pmd_lookup_flow(pmd, &keys[i]);
3283 if (OVS_LIKELY(!netdev_flow)) {
3284 netdev_flow = dp_netdev_flow_add(pmd, &match, &ufid,
3285 add_actions->data,
3286 add_actions->size);
3287 }
3288 ovs_mutex_unlock(&pmd->flow_mutex);
3289
3290 emc_insert(flow_cache, &keys[i], netdev_flow);
3291 }
3292 }
3293
3294 ofpbuf_uninit(&actions);
3295 ofpbuf_uninit(&put_actions);
3296 fat_rwlock_unlock(&dp->upcall_rwlock);
3297 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3298 } else if (OVS_UNLIKELY(any_miss)) {
3299 for (i = 0; i < cnt; i++) {
3300 if (OVS_UNLIKELY(!rules[i])) {
3301 dp_packet_delete(packets[i]);
3302 lost_cnt++;
3303 miss_cnt++;
3304 }
3305 }
3306 }
3307
3308 for (i = 0; i < cnt; i++) {
3309 struct dp_packet *packet = packets[i];
3310 struct dp_netdev_flow *flow;
3311
3312 if (OVS_UNLIKELY(!rules[i])) {
3313 continue;
3314 }
3315
3316 flow = dp_netdev_flow_cast(rules[i]);
3317
3318 emc_insert(flow_cache, &keys[i], flow);
3319 dp_netdev_queue_batches(packet, flow, &keys[i].mf, batches, n_batches);
3320 }
3321
3322 dp_netdev_count_packet(pmd, DP_STAT_MASKED_HIT, cnt - miss_cnt);
3323 dp_netdev_count_packet(pmd, DP_STAT_MISS, miss_cnt);
3324 dp_netdev_count_packet(pmd, DP_STAT_LOST, lost_cnt);
3325 }
3326
3327 static void
3328 dp_netdev_input(struct dp_netdev_pmd_thread *pmd,
3329 struct dp_packet **packets, int cnt)
3330 {
3331 #if !defined(__CHECKER__) && !defined(_WIN32)
3332 const size_t PKT_ARRAY_SIZE = cnt;
3333 #else
3334 /* Sparse or MSVC doesn't like variable length array. */
3335 enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
3336 #endif
3337 struct netdev_flow_key keys[PKT_ARRAY_SIZE];
3338 struct packet_batch batches[PKT_ARRAY_SIZE];
3339 long long now = time_msec();
3340 size_t newcnt, n_batches, i;
3341
3342 n_batches = 0;
3343 newcnt = emc_processing(pmd, packets, cnt, keys, batches, &n_batches);
3344 if (OVS_UNLIKELY(newcnt)) {
3345 fast_path_processing(pmd, packets, newcnt, keys, batches, &n_batches);
3346 }
3347
3348 for (i = 0; i < n_batches; i++) {
3349 batches[i].flow->batch = NULL;
3350 }
3351
3352 for (i = 0; i < n_batches; i++) {
3353 packet_batch_execute(&batches[i], pmd, now);
3354 }
3355 }
3356
3357 struct dp_netdev_execute_aux {
3358 struct dp_netdev_pmd_thread *pmd;
3359 };
3360
3361 static void
3362 dpif_netdev_register_upcall_cb(struct dpif *dpif, upcall_callback *cb,
3363 void *aux)
3364 {
3365 struct dp_netdev *dp = get_dp_netdev(dpif);
3366 dp->upcall_aux = aux;
3367 dp->upcall_cb = cb;
3368 }
3369
3370 static void
3371 dp_netdev_drop_packets(struct dp_packet **packets, int cnt, bool may_steal)
3372 {
3373 if (may_steal) {
3374 int i;
3375
3376 for (i = 0; i < cnt; i++) {
3377 dp_packet_delete(packets[i]);
3378 }
3379 }
3380 }
3381
3382 static int
3383 push_tnl_action(const struct dp_netdev *dp,
3384 const struct nlattr *attr,
3385 struct dp_packet **packets, int cnt)
3386 {
3387 struct dp_netdev_port *tun_port;
3388 const struct ovs_action_push_tnl *data;
3389
3390 data = nl_attr_get(attr);
3391
3392 tun_port = dp_netdev_lookup_port(dp, u32_to_odp(data->tnl_port));
3393 if (!tun_port) {
3394 return -EINVAL;
3395 }
3396 netdev_push_header(tun_port->netdev, packets, cnt, data);
3397
3398 return 0;
3399 }
3400
3401 static void
3402 dp_netdev_clone_pkt_batch(struct dp_packet **dst_pkts,
3403 struct dp_packet **src_pkts, int cnt)
3404 {
3405 int i;
3406
3407 for (i = 0; i < cnt; i++) {
3408 dst_pkts[i] = dp_packet_clone(src_pkts[i]);
3409 }
3410 }
3411
3412 static void
3413 dp_execute_cb(void *aux_, struct dp_packet **packets, int cnt,
3414 const struct nlattr *a, bool may_steal)
3415 OVS_NO_THREAD_SAFETY_ANALYSIS
3416 {
3417 struct dp_netdev_execute_aux *aux = aux_;
3418 uint32_t *depth = recirc_depth_get();
3419 struct dp_netdev_pmd_thread *pmd = aux->pmd;
3420 struct dp_netdev *dp = pmd->dp;
3421 int type = nl_attr_type(a);
3422 struct dp_netdev_port *p;
3423 int i;
3424
3425 switch ((enum ovs_action_attr)type) {
3426 case OVS_ACTION_ATTR_OUTPUT:
3427 p = dp_netdev_lookup_port(dp, u32_to_odp(nl_attr_get_u32(a)));
3428 if (OVS_LIKELY(p)) {
3429 netdev_send(p->netdev, pmd->tx_qid, packets, cnt, may_steal);
3430 return;
3431 }
3432 break;
3433
3434 case OVS_ACTION_ATTR_TUNNEL_PUSH:
3435 if (*depth < MAX_RECIRC_DEPTH) {
3436 struct dp_packet *tnl_pkt[NETDEV_MAX_BURST];
3437 int err;
3438
3439 if (!may_steal) {
3440 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3441 packets = tnl_pkt;
3442 }
3443
3444 err = push_tnl_action(dp, a, packets, cnt);
3445 if (!err) {
3446 (*depth)++;
3447 dp_netdev_input(pmd, packets, cnt);
3448 (*depth)--;
3449 } else {
3450 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3451 }
3452 return;
3453 }
3454 break;
3455
3456 case OVS_ACTION_ATTR_TUNNEL_POP:
3457 if (*depth < MAX_RECIRC_DEPTH) {
3458 odp_port_t portno = u32_to_odp(nl_attr_get_u32(a));
3459
3460 p = dp_netdev_lookup_port(dp, portno);
3461 if (p) {
3462 struct dp_packet *tnl_pkt[NETDEV_MAX_BURST];
3463 int err;
3464
3465 if (!may_steal) {
3466 dp_netdev_clone_pkt_batch(tnl_pkt, packets, cnt);
3467 packets = tnl_pkt;
3468 }
3469
3470 err = netdev_pop_header(p->netdev, packets, cnt);
3471 if (!err) {
3472
3473 for (i = 0; i < cnt; i++) {
3474 packets[i]->md.in_port.odp_port = portno;
3475 }
3476
3477 (*depth)++;
3478 dp_netdev_input(pmd, packets, cnt);
3479 (*depth)--;
3480 } else {
3481 dp_netdev_drop_packets(tnl_pkt, cnt, !may_steal);
3482 }
3483 return;
3484 }
3485 }
3486 break;
3487
3488 case OVS_ACTION_ATTR_USERSPACE:
3489 if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
3490 const struct nlattr *userdata;
3491 struct ofpbuf actions;
3492 struct flow flow;
3493 ovs_u128 ufid;
3494
3495 userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
3496 ofpbuf_init(&actions, 0);
3497
3498 for (i = 0; i < cnt; i++) {
3499 int error;
3500
3501 ofpbuf_clear(&actions);
3502
3503 flow_extract(packets[i], &flow);
3504 dpif_flow_hash(dp->dpif, &flow, sizeof flow, &ufid);
3505 error = dp_netdev_upcall(pmd, packets[i], &flow, NULL, &ufid,
3506 DPIF_UC_ACTION, userdata,&actions,
3507 NULL);
3508 if (!error || error == ENOSPC) {
3509 dp_netdev_execute_actions(pmd, &packets[i], 1, may_steal,
3510 actions.data, actions.size);
3511 } else if (may_steal) {
3512 dp_packet_delete(packets[i]);
3513 }
3514 }
3515 ofpbuf_uninit(&actions);
3516 fat_rwlock_unlock(&dp->upcall_rwlock);
3517
3518 return;
3519 }
3520 break;
3521
3522 case OVS_ACTION_ATTR_RECIRC:
3523 if (*depth < MAX_RECIRC_DEPTH) {
3524 struct dp_packet *recirc_pkts[NETDEV_MAX_BURST];
3525
3526 if (!may_steal) {
3527 dp_netdev_clone_pkt_batch(recirc_pkts, packets, cnt);
3528 packets = recirc_pkts;
3529 }
3530
3531 for (i = 0; i < cnt; i++) {
3532 packets[i]->md.recirc_id = nl_attr_get_u32(a);
3533 }
3534
3535 (*depth)++;
3536 dp_netdev_input(pmd, packets, cnt);
3537 (*depth)--;
3538
3539 return;
3540 }
3541
3542 VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
3543 break;
3544
3545 case OVS_ACTION_ATTR_PUSH_VLAN:
3546 case OVS_ACTION_ATTR_POP_VLAN:
3547 case OVS_ACTION_ATTR_PUSH_MPLS:
3548 case OVS_ACTION_ATTR_POP_MPLS:
3549 case OVS_ACTION_ATTR_SET:
3550 case OVS_ACTION_ATTR_SET_MASKED:
3551 case OVS_ACTION_ATTR_SAMPLE:
3552 case OVS_ACTION_ATTR_HASH:
3553 case OVS_ACTION_ATTR_UNSPEC:
3554 case __OVS_ACTION_ATTR_MAX:
3555 OVS_NOT_REACHED();
3556 }
3557
3558 dp_netdev_drop_packets(packets, cnt, may_steal);
3559 }
3560
3561 static void
3562 dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
3563 struct dp_packet **packets, int cnt,
3564 bool may_steal,
3565 const struct nlattr *actions, size_t actions_len)
3566 {
3567 struct dp_netdev_execute_aux aux = { pmd };
3568
3569 odp_execute_actions(&aux, packets, cnt, may_steal, actions,
3570 actions_len, dp_execute_cb);
3571 }
3572
3573 const struct dpif_class dpif_netdev_class = {
3574 "netdev",
3575 dpif_netdev_init,
3576 dpif_netdev_enumerate,
3577 dpif_netdev_port_open_type,
3578 dpif_netdev_open,
3579 dpif_netdev_close,
3580 dpif_netdev_destroy,
3581 dpif_netdev_run,
3582 dpif_netdev_wait,
3583 dpif_netdev_get_stats,
3584 dpif_netdev_port_add,
3585 dpif_netdev_port_del,
3586 dpif_netdev_port_query_by_number,
3587 dpif_netdev_port_query_by_name,
3588 NULL, /* port_get_pid */
3589 dpif_netdev_port_dump_start,
3590 dpif_netdev_port_dump_next,
3591 dpif_netdev_port_dump_done,
3592 dpif_netdev_port_poll,
3593 dpif_netdev_port_poll_wait,
3594 dpif_netdev_flow_flush,
3595 dpif_netdev_flow_dump_create,
3596 dpif_netdev_flow_dump_destroy,
3597 dpif_netdev_flow_dump_thread_create,
3598 dpif_netdev_flow_dump_thread_destroy,
3599 dpif_netdev_flow_dump_next,
3600 dpif_netdev_operate,
3601 NULL, /* recv_set */
3602 NULL, /* handlers_set */
3603 dpif_netdev_pmd_set,
3604 dpif_netdev_queue_to_priority,
3605 NULL, /* recv */
3606 NULL, /* recv_wait */
3607 NULL, /* recv_purge */
3608 dpif_netdev_register_upcall_cb,
3609 dpif_netdev_enable_upcall,
3610 dpif_netdev_disable_upcall,
3611 dpif_netdev_get_datapath_version,
3612 };
3613
3614 static void
3615 dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
3616 const char *argv[], void *aux OVS_UNUSED)
3617 {
3618 struct dp_netdev_port *old_port;
3619 struct dp_netdev_port *new_port;
3620 struct dp_netdev *dp;
3621 odp_port_t port_no;
3622
3623 ovs_mutex_lock(&dp_netdev_mutex);
3624 dp = shash_find_data(&dp_netdevs, argv[1]);
3625 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3626 ovs_mutex_unlock(&dp_netdev_mutex);
3627 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3628 return;
3629 }
3630 ovs_refcount_ref(&dp->ref_cnt);
3631 ovs_mutex_unlock(&dp_netdev_mutex);
3632
3633 ovs_mutex_lock(&dp->port_mutex);
3634 if (get_port_by_name(dp, argv[2], &old_port)) {
3635 unixctl_command_reply_error(conn, "unknown port");
3636 goto exit;
3637 }
3638
3639 port_no = u32_to_odp(atoi(argv[3]));
3640 if (!port_no || port_no == ODPP_NONE) {
3641 unixctl_command_reply_error(conn, "bad port number");
3642 goto exit;
3643 }
3644 if (dp_netdev_lookup_port(dp, port_no)) {
3645 unixctl_command_reply_error(conn, "port number already in use");
3646 goto exit;
3647 }
3648
3649 /* Remove old port. */
3650 cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->port_no));
3651 ovsrcu_postpone(free, old_port);
3652
3653 /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
3654 new_port = xmemdup(old_port, sizeof *old_port);
3655 new_port->port_no = port_no;
3656 cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
3657
3658 seq_change(dp->port_seq);
3659 unixctl_command_reply(conn, NULL);
3660
3661 exit:
3662 ovs_mutex_unlock(&dp->port_mutex);
3663 dp_netdev_unref(dp);
3664 }
3665
3666 static void
3667 dpif_dummy_delete_port(struct unixctl_conn *conn, int argc OVS_UNUSED,
3668 const char *argv[], void *aux OVS_UNUSED)
3669 {
3670 struct dp_netdev_port *port;
3671 struct dp_netdev *dp;
3672
3673 ovs_mutex_lock(&dp_netdev_mutex);
3674 dp = shash_find_data(&dp_netdevs, argv[1]);
3675 if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
3676 ovs_mutex_unlock(&dp_netdev_mutex);
3677 unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
3678 return;
3679 }
3680 ovs_refcount_ref(&dp->ref_cnt);
3681 ovs_mutex_unlock(&dp_netdev_mutex);
3682
3683 ovs_mutex_lock(&dp->port_mutex);
3684 if (get_port_by_name(dp, argv[2], &port)) {
3685 unixctl_command_reply_error(conn, "unknown port");
3686 } else if (port->port_no == ODPP_LOCAL) {
3687 unixctl_command_reply_error(conn, "can't delete local port");
3688 } else {
3689 do_del_port(dp, port);
3690 unixctl_command_reply(conn, NULL);
3691 }
3692 ovs_mutex_unlock(&dp->port_mutex);
3693
3694 dp_netdev_unref(dp);
3695 }
3696
3697 static void
3698 dpif_dummy_register__(const char *type)
3699 {
3700 struct dpif_class *class;
3701
3702 class = xmalloc(sizeof *class);
3703 *class = dpif_netdev_class;
3704 class->type = xstrdup(type);
3705 dp_register_provider(class);
3706 }
3707
3708 static void
3709 dpif_dummy_override(const char *type)
3710 {
3711 if (!dp_unregister_provider(type)) {
3712 dpif_dummy_register__(type);
3713 }
3714 }
3715
3716 void
3717 dpif_dummy_register(enum dummy_level level)
3718 {
3719 if (level == DUMMY_OVERRIDE_ALL) {
3720 struct sset types;
3721 const char *type;
3722
3723 sset_init(&types);
3724 dp_enumerate_types(&types);
3725 SSET_FOR_EACH (type, &types) {
3726 dpif_dummy_override(type);
3727 }
3728 sset_destroy(&types);
3729 } else if (level == DUMMY_OVERRIDE_SYSTEM) {
3730 dpif_dummy_override("system");
3731 }
3732
3733 dpif_dummy_register__("dummy");
3734
3735 unixctl_command_register("dpif-dummy/change-port-number",
3736 "dp port new-number",
3737 3, 3, dpif_dummy_change_port_number, NULL);
3738 unixctl_command_register("dpif-dummy/delete-port", "dp port",
3739 2, 2, dpif_dummy_delete_port, NULL);
3740 }
3741 \f
3742 /* Datapath Classifier. */
3743
3744 /* A set of rules that all have the same fields wildcarded. */
3745 struct dpcls_subtable {
3746 /* The fields are only used by writers. */
3747 struct cmap_node cmap_node OVS_GUARDED; /* Within dpcls 'subtables_map'. */
3748
3749 /* These fields are accessed by readers. */
3750 struct cmap rules; /* Contains "struct dpcls_rule"s. */
3751 struct netdev_flow_key mask; /* Wildcards for fields (const). */
3752 /* 'mask' must be the last field, additional space is allocated here. */
3753 };
3754
3755 /* Initializes 'cls' as a classifier that initially contains no classification
3756 * rules. */
3757 static void
3758 dpcls_init(struct dpcls *cls)
3759 {
3760 cmap_init(&cls->subtables_map);
3761 pvector_init(&cls->subtables);
3762 }
3763
3764 static void
3765 dpcls_destroy_subtable(struct dpcls *cls, struct dpcls_subtable *subtable)
3766 {
3767 pvector_remove(&cls->subtables, subtable);
3768 cmap_remove(&cls->subtables_map, &subtable->cmap_node,
3769 subtable->mask.hash);
3770 cmap_destroy(&subtable->rules);
3771 ovsrcu_postpone(free, subtable);
3772 }
3773
3774 /* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
3775 * caller's responsibility.
3776 * May only be called after all the readers have been terminated. */
3777 static void
3778 dpcls_destroy(struct dpcls *cls)
3779 {
3780 if (cls) {
3781 struct dpcls_subtable *subtable;
3782
3783 CMAP_FOR_EACH (subtable, cmap_node, &cls->subtables_map) {
3784 dpcls_destroy_subtable(cls, subtable);
3785 }
3786 cmap_destroy(&cls->subtables_map);
3787 pvector_destroy(&cls->subtables);
3788 }
3789 }
3790
3791 static struct dpcls_subtable *
3792 dpcls_create_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
3793 {
3794 struct dpcls_subtable *subtable;
3795
3796 /* Need to add one. */
3797 subtable = xmalloc(sizeof *subtable
3798 - sizeof subtable->mask.mf + mask->len);
3799 cmap_init(&subtable->rules);
3800 netdev_flow_key_clone(&subtable->mask, mask);
3801 cmap_insert(&cls->subtables_map, &subtable->cmap_node, mask->hash);
3802 pvector_insert(&cls->subtables, subtable, 0);
3803 pvector_publish(&cls->subtables);
3804
3805 return subtable;
3806 }
3807
3808 static inline struct dpcls_subtable *
3809 dpcls_find_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
3810 {
3811 struct dpcls_subtable *subtable;
3812
3813 CMAP_FOR_EACH_WITH_HASH (subtable, cmap_node, mask->hash,
3814 &cls->subtables_map) {
3815 if (netdev_flow_key_equal(&subtable->mask, mask)) {
3816 return subtable;
3817 }
3818 }
3819 return dpcls_create_subtable(cls, mask);
3820 }
3821
3822 /* Insert 'rule' into 'cls'. */
3823 static void
3824 dpcls_insert(struct dpcls *cls, struct dpcls_rule *rule,
3825 const struct netdev_flow_key *mask)
3826 {
3827 struct dpcls_subtable *subtable = dpcls_find_subtable(cls, mask);
3828
3829 rule->mask = &subtable->mask;
3830 cmap_insert(&subtable->rules, &rule->cmap_node, rule->flow.hash);
3831 }
3832
3833 /* Removes 'rule' from 'cls', also destructing the 'rule'. */
3834 static void
3835 dpcls_remove(struct dpcls *cls, struct dpcls_rule *rule)
3836 {
3837 struct dpcls_subtable *subtable;
3838
3839 ovs_assert(rule->mask);
3840
3841 INIT_CONTAINER(subtable, rule->mask, mask);
3842
3843 if (cmap_remove(&subtable->rules, &rule->cmap_node, rule->flow.hash)
3844 == 0) {
3845 dpcls_destroy_subtable(cls, subtable);
3846 pvector_publish(&cls->subtables);
3847 }
3848 }
3849
3850 /* Returns true if 'target' satisifies 'key' in 'mask', that is, if each 1-bit
3851 * in 'mask' the values in 'key' and 'target' are the same.
3852 *
3853 * Note: 'key' and 'mask' have the same mask, and 'key' is already masked. */
3854 static inline bool
3855 dpcls_rule_matches_key(const struct dpcls_rule *rule,
3856 const struct netdev_flow_key *target)
3857 {
3858 const uint64_t *keyp = miniflow_get_values(&rule->flow.mf);
3859 const uint64_t *maskp = miniflow_get_values(&rule->mask->mf);
3860 uint64_t target_u64;
3861
3862 NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(target_u64, target, rule->flow.mf.map) {
3863 if (OVS_UNLIKELY((target_u64 & *maskp++) != *keyp++)) {
3864 return false;
3865 }
3866 }
3867 return true;
3868 }
3869
3870 /* For each miniflow in 'flows' performs a classifier lookup writing the result
3871 * into the corresponding slot in 'rules'. If a particular entry in 'flows' is
3872 * NULL it is skipped.
3873 *
3874 * This function is optimized for use in the userspace datapath and therefore
3875 * does not implement a lot of features available in the standard
3876 * classifier_lookup() function. Specifically, it does not implement
3877 * priorities, instead returning any rule which matches the flow.
3878 *
3879 * Returns true if all flows found a corresponding rule. */
3880 static bool
3881 dpcls_lookup(const struct dpcls *cls, const struct netdev_flow_key keys[],
3882 struct dpcls_rule **rules, const size_t cnt)
3883 {
3884 /* The batch size 16 was experimentally found faster than 8 or 32. */
3885 typedef uint16_t map_type;
3886 #define MAP_BITS (sizeof(map_type) * CHAR_BIT)
3887
3888 #if !defined(__CHECKER__) && !defined(_WIN32)
3889 const int N_MAPS = DIV_ROUND_UP(cnt, MAP_BITS);
3890 #else
3891 enum { N_MAPS = DIV_ROUND_UP(NETDEV_MAX_BURST, MAP_BITS) };
3892 #endif
3893 map_type maps[N_MAPS];
3894 struct dpcls_subtable *subtable;
3895
3896 memset(maps, 0xff, sizeof maps);
3897 if (cnt % MAP_BITS) {
3898 maps[N_MAPS - 1] >>= MAP_BITS - cnt % MAP_BITS; /* Clear extra bits. */
3899 }
3900 memset(rules, 0, cnt * sizeof *rules);
3901
3902 PVECTOR_FOR_EACH (subtable, &cls->subtables) {
3903 const struct netdev_flow_key *mkeys = keys;
3904 struct dpcls_rule **mrules = rules;
3905 map_type remains = 0;
3906 int m;
3907
3908 BUILD_ASSERT_DECL(sizeof remains == sizeof *maps);
3909
3910 for (m = 0; m < N_MAPS; m++, mkeys += MAP_BITS, mrules += MAP_BITS) {
3911 uint32_t hashes[MAP_BITS];
3912 const struct cmap_node *nodes[MAP_BITS];
3913 unsigned long map = maps[m];
3914 int i;
3915
3916 if (!map) {
3917 continue; /* Skip empty maps. */
3918 }
3919
3920 /* Compute hashes for the remaining keys. */
3921 ULLONG_FOR_EACH_1(i, map) {
3922 hashes[i] = netdev_flow_key_hash_in_mask(&mkeys[i],
3923 &subtable->mask);
3924 }
3925 /* Lookup. */
3926 map = cmap_find_batch(&subtable->rules, map, hashes, nodes);
3927 /* Check results. */
3928 ULLONG_FOR_EACH_1(i, map) {
3929 struct dpcls_rule *rule;
3930
3931 CMAP_NODE_FOR_EACH (rule, cmap_node, nodes[i]) {
3932 if (OVS_LIKELY(dpcls_rule_matches_key(rule, &mkeys[i]))) {
3933 mrules[i] = rule;
3934 goto next;
3935 }
3936 }
3937 ULLONG_SET0(map, i); /* Did not match. */
3938 next:
3939 ; /* Keep Sparse happy. */
3940 }
3941 maps[m] &= ~map; /* Clear the found rules. */
3942 remains |= maps[m];
3943 }
3944 if (!remains) {
3945 return true; /* All found. */
3946 }
3947 }
3948 return false; /* Some misses. */
3949 }