1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Copyright (c) 2007-2014 Nicira, Inc.
4 */
5
6 #include "flow.h"
7 #include "datapath.h"
8 #include "flow_netlink.h"
9 #include <linux/uaccess.h>
10 #include <linux/netdevice.h>
11 #include <linux/etherdevice.h>
12 #include <linux/if_ether.h>
13 #include <linux/if_vlan.h>
14 #include <net/llc_pdu.h>
15 #include <linux/kernel.h>
16 #include <linux/jhash.h>
17 #include <linux/jiffies.h>
18 #include <linux/llc.h>
19 #include <linux/module.h>
20 #include <linux/in.h>
21 #include <linux/rcupdate.h>
22 #include <linux/cpumask.h>
23 #include <linux/if_arp.h>
24 #include <linux/ip.h>
25 #include <linux/ipv6.h>
26 #include <linux/sctp.h>
27 #include <linux/tcp.h>
28 #include <linux/udp.h>
29 #include <linux/icmp.h>
30 #include <linux/icmpv6.h>
31 #include <linux/rculist.h>
32 #include <linux/sort.h>
33 #include <net/ip.h>
34 #include <net/ipv6.h>
35 #include <net/ndisc.h>
36
37 #define TBL_MIN_BUCKETS 1024
38 #define MASK_ARRAY_SIZE_MIN 16
39 #define REHASH_INTERVAL (10 * 60 * HZ)
40
41 #define MC_DEFAULT_HASH_ENTRIES 256
42 #define MC_HASH_SHIFT 8
43 #define MC_HASH_SEGS ((sizeof(uint32_t) * 8) / MC_HASH_SHIFT)
44
45 static struct kmem_cache *flow_cache;
46 struct kmem_cache *flow_stats_cache __read_mostly;
47
48 static u16 range_n_bytes(const struct sw_flow_key_range *range)
49 {
50 return range->end - range->start;
51 }
52
53 void ovs_flow_mask_key(struct sw_flow_key *dst, const struct sw_flow_key *src,
54 bool full, const struct sw_flow_mask *mask)
55 {
56 int start = full ? 0 : mask->range.start;
57 int len = full ? sizeof *dst : range_n_bytes(&mask->range);
58 const long *m = (const long *)((const u8 *)&mask->key + start);
59 const long *s = (const long *)((const u8 *)src + start);
60 long *d = (long *)((u8 *)dst + start);
61 int i;
62
63 /* If 'full' is true then all of 'dst' is fully initialized. Otherwise,
64 * if 'full' is false the memory outside of the 'mask->range' is left
65 * uninitialized. This can be used as an optimization when further
66 * operations on 'dst' only use contents within 'mask->range'.
67 */
68 for (i = 0; i < len; i += sizeof(long))
69 *d++ = *s++ & *m++;
70 }
71
72 struct sw_flow *ovs_flow_alloc(void)
73 {
74 struct sw_flow *flow;
75 struct sw_flow_stats *stats;
76
77 flow = kmem_cache_zalloc(flow_cache, GFP_KERNEL);
78 if (!flow)
79 return ERR_PTR(-ENOMEM);
80
81 flow->stats_last_writer = -1;
82 flow->cpu_used_mask = (struct cpumask *)&flow->stats[nr_cpu_ids];
83
84 /* Initialize the default stat node. */
85 stats = kmem_cache_alloc_node(flow_stats_cache,
86 GFP_KERNEL | __GFP_ZERO,
87 node_online(0) ? 0 : NUMA_NO_NODE);
88 if (!stats)
89 goto err;
90
91 spin_lock_init(&stats->lock);
92
93 RCU_INIT_POINTER(flow->stats[0], stats);
94
95 cpumask_set_cpu(0, flow->cpu_used_mask);
96
97 return flow;
98 err:
99 kmem_cache_free(flow_cache, flow);
100 return ERR_PTR(-ENOMEM);
101 }
102
103 int ovs_flow_tbl_count(const struct flow_table *table)
104 {
105 return table->count;
106 }
107
108 static void flow_free(struct sw_flow *flow)
109 {
110 int cpu;
111
112 if (ovs_identifier_is_key(&flow->id))
113 kfree(flow->id.unmasked_key);
114 if (flow->sf_acts)
115 ovs_nla_free_flow_actions((struct sw_flow_actions __force *)
116 flow->sf_acts);
117 /* We open code this to make sure cpu 0 is always considered */
118 for (cpu = 0; cpu < nr_cpu_ids;
119 cpu = cpumask_next(cpu, flow->cpu_used_mask)) {
120 if (flow->stats[cpu])
121 kmem_cache_free(flow_stats_cache,
122 (struct sw_flow_stats __force *)flow->stats[cpu]);
123 }
124
125 kmem_cache_free(flow_cache, flow);
126 }
127
128 static void rcu_free_flow_callback(struct rcu_head *rcu)
129 {
130 struct sw_flow *flow = container_of(rcu, struct sw_flow, rcu);
131
132 flow_free(flow);
133 }
134
135 void ovs_flow_free(struct sw_flow *flow, bool deferred)
136 {
137 if (!flow)
138 return;
139
140 if (deferred)
141 call_rcu(&flow->rcu, rcu_free_flow_callback);
142 else
143 flow_free(flow);
144 }
145
146 static void __table_instance_destroy(struct table_instance *ti)
147 {
148 kvfree(ti->buckets);
149 kfree(ti);
150 }
151
152 static struct table_instance *table_instance_alloc(int new_size)
153 {
154 struct table_instance *ti = kmalloc(sizeof(*ti), GFP_KERNEL);
155 int i;
156
157 if (!ti)
158 return NULL;
159
160 ti->buckets = kvmalloc_array(new_size, sizeof(struct hlist_head),
161 GFP_KERNEL);
162 if (!ti->buckets) {
163 kfree(ti);
164 return NULL;
165 }
166
167 for (i = 0; i < new_size; i++)
168 INIT_HLIST_HEAD(&ti->buckets[i]);
169
170 ti->n_buckets = new_size;
171 ti->node_ver = 0;
172 get_random_bytes(&ti->hash_seed, sizeof(u32));
173
174 return ti;
175 }
176
177 static void __mask_array_destroy(struct mask_array *ma)
178 {
179 free_percpu(ma->masks_usage_stats);
180 kfree(ma);
181 }
182
183 static void mask_array_rcu_cb(struct rcu_head *rcu)
184 {
185 struct mask_array *ma = container_of(rcu, struct mask_array, rcu);
186
187 __mask_array_destroy(ma);
188 }
189
190 static void tbl_mask_array_reset_counters(struct mask_array *ma)
191 {
192 int i, cpu;
193
194 /* As the per CPU counters are not atomic we can not go ahead and
195 * reset them from another CPU. To be able to still have an approximate
196 * zero based counter we store the value at reset, and subtract it
197 * later when processing.
198 */
199 for (i = 0; i < ma->max; i++) {
200 ma->masks_usage_zero_cntr[i] = 0;
201
202 for_each_possible_cpu(cpu) {
203 struct mask_array_stats *stats;
204 unsigned int start;
205 u64 counter;
206
207 stats = per_cpu_ptr(ma->masks_usage_stats, cpu);
208 do {
209 start = u64_stats_fetch_begin(&stats->syncp);
210 counter = stats->usage_cntrs[i];
211 } while (u64_stats_fetch_retry(&stats->syncp, start));
212
213 ma->masks_usage_zero_cntr[i] += counter;
214 }
215 }
216 }
217
218 static struct mask_array *tbl_mask_array_alloc(int size)
219 {
220 struct mask_array *new;
221
222 size = max(MASK_ARRAY_SIZE_MIN, size);
223 new = kzalloc(struct_size(new, masks, size) +
224 sizeof(u64) * size, GFP_KERNEL);
225 if (!new)
226 return NULL;
227
228 new->masks_usage_zero_cntr = (u64 *)((u8 *)new +
229 struct_size(new, masks, size));
230
231 new->masks_usage_stats = __alloc_percpu(sizeof(struct mask_array_stats) +
232 sizeof(u64) * size,
233 __alignof__(u64));
234 if (!new->masks_usage_stats) {
235 kfree(new);
236 return NULL;
237 }
238
239 new->count = 0;
240 new->max = size;
241
242 return new;
243 }
244
245 static int tbl_mask_array_realloc(struct flow_table *tbl, int size)
246 {
247 struct mask_array *old;
248 struct mask_array *new;
249
250 new = tbl_mask_array_alloc(size);
251 if (!new)
252 return -ENOMEM;
253
254 old = ovsl_dereference(tbl->mask_array);
255 if (old) {
256 int i;
257
258 for (i = 0; i < old->max; i++) {
259 if (ovsl_dereference(old->masks[i]))
260 new->masks[new->count++] = old->masks[i];
261 }
262 call_rcu(&old->rcu, mask_array_rcu_cb);
263 }
264
265 rcu_assign_pointer(tbl->mask_array, new);
266
267 return 0;
268 }
269
270 static int tbl_mask_array_add_mask(struct flow_table *tbl,
271 struct sw_flow_mask *new)
272 {
273 struct mask_array *ma = ovsl_dereference(tbl->mask_array);
274 int err, ma_count = READ_ONCE(ma->count);
275
276 if (ma_count >= ma->max) {
277 err = tbl_mask_array_realloc(tbl, ma->max +
278 MASK_ARRAY_SIZE_MIN);
279 if (err)
280 return err;
281
282 ma = ovsl_dereference(tbl->mask_array);
283 } else {
284 /* On every add or delete we need to reset the counters so
285 * every new mask gets a fair chance of being prioritized.
286 */
287 tbl_mask_array_reset_counters(ma);
288 }
289
290 BUG_ON(ovsl_dereference(ma->masks[ma_count]));
291
292 rcu_assign_pointer(ma->masks[ma_count], new);
293 WRITE_ONCE(ma->count, ma_count + 1);
294
295 return 0;
296 }
297
298 static void tbl_mask_array_del_mask(struct flow_table *tbl,
299 struct sw_flow_mask *mask)
300 {
301 struct mask_array *ma = ovsl_dereference(tbl->mask_array);
302 int i, ma_count = READ_ONCE(ma->count);
303
304 /* Remove the deleted mask pointers from the array */
305 for (i = 0; i < ma_count; i++) {
306 if (mask == ovsl_dereference(ma->masks[i]))
307 goto found;
308 }
309
310 BUG();
311 return;
312
313 found:
314 WRITE_ONCE(ma->count, ma_count - 1);
315
316 rcu_assign_pointer(ma->masks[i], ma->masks[ma_count - 1]);
317 RCU_INIT_POINTER(ma->masks[ma_count - 1], NULL);
318
319 kfree_rcu(mask, rcu);
320
321 /* Shrink the mask array if necessary. */
322 if (ma->max >= (MASK_ARRAY_SIZE_MIN * 2) &&
323 ma_count <= (ma->max / 3))
324 tbl_mask_array_realloc(tbl, ma->max / 2);
325 else
326 tbl_mask_array_reset_counters(ma);
327
328 }
329
330 /* Remove 'mask' from the mask list, if it is not needed any more. */
331 static void flow_mask_remove(struct flow_table *tbl, struct sw_flow_mask *mask)
332 {
333 if (mask) {
334 /* ovs-lock is required to protect mask-refcount and
335 * mask list.
336 */
337 ASSERT_OVSL();
338 BUG_ON(!mask->ref_count);
339 mask->ref_count--;
340
341 if (!mask->ref_count)
342 tbl_mask_array_del_mask(tbl, mask);
343 }
344 }
345
346 static void __mask_cache_destroy(struct mask_cache *mc)
347 {
348 free_percpu(mc->mask_cache);
349 kfree(mc);
350 }
351
352 static void mask_cache_rcu_cb(struct rcu_head *rcu)
353 {
354 struct mask_cache *mc = container_of(rcu, struct mask_cache, rcu);
355
356 __mask_cache_destroy(mc);
357 }
358
359 static struct mask_cache *tbl_mask_cache_alloc(u32 size)
360 {
361 struct mask_cache_entry __percpu *cache = NULL;
362 struct mask_cache *new;
363
364 /* Only allow size to be 0, or a power of 2, and does not exceed
365 * percpu allocation size.
366 */
367 if ((!is_power_of_2(size) && size != 0) ||
368 (size * sizeof(struct mask_cache_entry)) > PCPU_MIN_UNIT_SIZE)
369 return NULL;
370
371 new = kzalloc(sizeof(*new), GFP_KERNEL);
372 if (!new)
373 return NULL;
374
375 new->cache_size = size;
376 if (new->cache_size > 0) {
377 cache = __alloc_percpu(array_size(sizeof(struct mask_cache_entry),
378 new->cache_size),
379 __alignof__(struct mask_cache_entry));
380 if (!cache) {
381 kfree(new);
382 return NULL;
383 }
384 }
385
386 new->mask_cache = cache;
387 return new;
388 }
389 int ovs_flow_tbl_masks_cache_resize(struct flow_table *table, u32 size)
390 {
391 struct mask_cache *mc = rcu_dereference_ovsl(table->mask_cache);
392 struct mask_cache *new;
393
394 if (size == mc->cache_size)
395 return 0;
396
397 if ((!is_power_of_2(size) && size != 0) ||
398 (size * sizeof(struct mask_cache_entry)) > PCPU_MIN_UNIT_SIZE)
399 return -EINVAL;
400
401 new = tbl_mask_cache_alloc(size);
402 if (!new)
403 return -ENOMEM;
404
405 rcu_assign_pointer(table->mask_cache, new);
406 call_rcu(&mc->rcu, mask_cache_rcu_cb);
407
408 return 0;
409 }
410
411 int ovs_flow_tbl_init(struct flow_table *table)
412 {
413 struct table_instance *ti, *ufid_ti;
414 struct mask_cache *mc;
415 struct mask_array *ma;
416
417 mc = tbl_mask_cache_alloc(MC_DEFAULT_HASH_ENTRIES);
418 if (!mc)
419 return -ENOMEM;
420
421 ma = tbl_mask_array_alloc(MASK_ARRAY_SIZE_MIN);
422 if (!ma)
423 goto free_mask_cache;
424
425 ti = table_instance_alloc(TBL_MIN_BUCKETS);
426 if (!ti)
427 goto free_mask_array;
428
429 ufid_ti = table_instance_alloc(TBL_MIN_BUCKETS);
430 if (!ufid_ti)
431 goto free_ti;
432
433 rcu_assign_pointer(table->ti, ti);
434 rcu_assign_pointer(table->ufid_ti, ufid_ti);
435 rcu_assign_pointer(table->mask_array, ma);
436 rcu_assign_pointer(table->mask_cache, mc);
437 table->last_rehash = jiffies;
438 table->count = 0;
439 table->ufid_count = 0;
440 return 0;
441
442 free_ti:
443 __table_instance_destroy(ti);
444 free_mask_array:
445 __mask_array_destroy(ma);
446 free_mask_cache:
447 __mask_cache_destroy(mc);
448 return -ENOMEM;
449 }
450
451 static void flow_tbl_destroy_rcu_cb(struct rcu_head *rcu)
452 {
453 struct table_instance *ti;
454
455 ti = container_of(rcu, struct table_instance, rcu);
456 __table_instance_destroy(ti);
457 }
458
459 static void table_instance_flow_free(struct flow_table *table,
460 struct table_instance *ti,
461 struct table_instance *ufid_ti,
462 struct sw_flow *flow)
463 {
464 hlist_del_rcu(&flow->flow_table.node[ti->node_ver]);
465 table->count--;
466
467 if (ovs_identifier_is_ufid(&flow->id)) {
468 hlist_del_rcu(&flow->ufid_table.node[ufid_ti->node_ver]);
469 table->ufid_count--;
470 }
471
472 flow_mask_remove(table, flow->mask);
473 }
474
475 /* Must be called with OVS mutex held. */
476 void table_instance_flow_flush(struct flow_table *table,
477 struct table_instance *ti,
478 struct table_instance *ufid_ti)
479 {
480 int i;
481
482 for (i = 0; i < ti->n_buckets; i++) {
483 struct hlist_head *head = &ti->buckets[i];
484 struct hlist_node *n;
485 struct sw_flow *flow;
486
487 hlist_for_each_entry_safe(flow, n, head,
488 flow_table.node[ti->node_ver]) {
489
490 table_instance_flow_free(table, ti, ufid_ti,
491 flow);
492 ovs_flow_free(flow, true);
493 }
494 }
495
496 if (WARN_ON(table->count != 0 ||
497 table->ufid_count != 0)) {
498 table->count = 0;
499 table->ufid_count = 0;
500 }
501 }
502
503 static void table_instance_destroy(struct table_instance *ti,
504 struct table_instance *ufid_ti)
505 {
506 call_rcu(&ti->rcu, flow_tbl_destroy_rcu_cb);
507 call_rcu(&ufid_ti->rcu, flow_tbl_destroy_rcu_cb);
508 }
509
510 /* No need for locking this function is called from RCU callback or
511 * error path.
512 */
513 void ovs_flow_tbl_destroy(struct flow_table *table)
514 {
515 struct table_instance *ti = rcu_dereference_raw(table->ti);
516 struct table_instance *ufid_ti = rcu_dereference_raw(table->ufid_ti);
517 struct mask_cache *mc = rcu_dereference_raw(table->mask_cache);
518 struct mask_array *ma = rcu_dereference_raw(table->mask_array);
519
520 call_rcu(&mc->rcu, mask_cache_rcu_cb);
521 call_rcu(&ma->rcu, mask_array_rcu_cb);
522 table_instance_destroy(ti, ufid_ti);
523 }
524
525 struct sw_flow *ovs_flow_tbl_dump_next(struct table_instance *ti,
526 u32 *bucket, u32 *last)
527 {
528 struct sw_flow *flow;
529 struct hlist_head *head;
530 int ver;
531 int i;
532
533 ver = ti->node_ver;
534 while (*bucket < ti->n_buckets) {
535 i = 0;
536 head = &ti->buckets[*bucket];
537 hlist_for_each_entry_rcu(flow, head, flow_table.node[ver]) {
538 if (i < *last) {
539 i++;
540 continue;
541 }
542 *last = i + 1;
543 return flow;
544 }
545 (*bucket)++;
546 *last = 0;
547 }
548
549 return NULL;
550 }
551
552 static struct hlist_head *find_bucket(struct table_instance *ti, u32 hash)
553 {
554 hash = jhash_1word(hash, ti->hash_seed);
555 return &ti->buckets[hash & (ti->n_buckets - 1)];
556 }
557
558 static void table_instance_insert(struct table_instance *ti,
559 struct sw_flow *flow)
560 {
561 struct hlist_head *head;
562
563 head = find_bucket(ti, flow->flow_table.hash);
564 hlist_add_head_rcu(&flow->flow_table.node[ti->node_ver], head);
565 }
566
567 static void ufid_table_instance_insert(struct table_instance *ti,
568 struct sw_flow *flow)
569 {
570 struct hlist_head *head;
571
572 head = find_bucket(ti, flow->ufid_table.hash);
573 hlist_add_head_rcu(&flow->ufid_table.node[ti->node_ver], head);
574 }
575
576 static void flow_table_copy_flows(struct table_instance *old,
577 struct table_instance *new, bool ufid)
578 {
579 int old_ver;
580 int i;
581
582 old_ver = old->node_ver;
583 new->node_ver = !old_ver;
584
585 /* Insert in new table. */
586 for (i = 0; i < old->n_buckets; i++) {
587 struct sw_flow *flow;
588 struct hlist_head *head = &old->buckets[i];
589
590 if (ufid)
591 hlist_for_each_entry_rcu(flow, head,
592 ufid_table.node[old_ver],
593 lockdep_ovsl_is_held())
594 ufid_table_instance_insert(new, flow);
595 else
596 hlist_for_each_entry_rcu(flow, head,
597 flow_table.node[old_ver],
598 lockdep_ovsl_is_held())
599 table_instance_insert(new, flow);
600 }
601 }
602
603 static struct table_instance *table_instance_rehash(struct table_instance *ti,
604 int n_buckets, bool ufid)
605 {
606 struct table_instance *new_ti;
607
608 new_ti = table_instance_alloc(n_buckets);
609 if (!new_ti)
610 return NULL;
611
612 flow_table_copy_flows(ti, new_ti, ufid);
613
614 return new_ti;
615 }
616
617 int ovs_flow_tbl_flush(struct flow_table *flow_table)
618 {
619 struct table_instance *old_ti, *new_ti;
620 struct table_instance *old_ufid_ti, *new_ufid_ti;
621
622 new_ti = table_instance_alloc(TBL_MIN_BUCKETS);
623 if (!new_ti)
624 return -ENOMEM;
625 new_ufid_ti = table_instance_alloc(TBL_MIN_BUCKETS);
626 if (!new_ufid_ti)
627 goto err_free_ti;
628
629 old_ti = ovsl_dereference(flow_table->ti);
630 old_ufid_ti = ovsl_dereference(flow_table->ufid_ti);
631
632 rcu_assign_pointer(flow_table->ti, new_ti);
633 rcu_assign_pointer(flow_table->ufid_ti, new_ufid_ti);
634 flow_table->last_rehash = jiffies;
635
636 table_instance_flow_flush(flow_table, old_ti, old_ufid_ti);
637 table_instance_destroy(old_ti, old_ufid_ti);
638 return 0;
639
640 err_free_ti:
641 __table_instance_destroy(new_ti);
642 return -ENOMEM;
643 }
644
645 static u32 flow_hash(const struct sw_flow_key *key,
646 const struct sw_flow_key_range *range)
647 {
648 const u32 *hash_key = (const u32 *)((const u8 *)key + range->start);
649
650 /* Make sure number of hash bytes are multiple of u32. */
651 int hash_u32s = range_n_bytes(range) >> 2;
652
653 return jhash2(hash_key, hash_u32s, 0);
654 }
655
656 static int flow_key_start(const struct sw_flow_key *key)
657 {
658 if (key->tun_proto)
659 return 0;
660 else
661 return rounddown(offsetof(struct sw_flow_key, phy),
662 sizeof(long));
663 }
664
665 static bool cmp_key(const struct sw_flow_key *key1,
666 const struct sw_flow_key *key2,
667 int key_start, int key_end)
668 {
669 const long *cp1 = (const long *)((const u8 *)key1 + key_start);
670 const long *cp2 = (const long *)((const u8 *)key2 + key_start);
671 int i;
672
673 for (i = key_start; i < key_end; i += sizeof(long))
674 if (*cp1++ ^ *cp2++)
675 return false;
676
677 return true;
678 }
679
680 static bool flow_cmp_masked_key(const struct sw_flow *flow,
681 const struct sw_flow_key *key,
682 const struct sw_flow_key_range *range)
683 {
684 return cmp_key(&flow->key, key, range->start, range->end);
685 }
686
687 static bool ovs_flow_cmp_unmasked_key(const struct sw_flow *flow,
688 const struct sw_flow_match *match)
689 {
690 struct sw_flow_key *key = match->key;
691 int key_start = flow_key_start(key);
692 int key_end = match->range.end;
693
694 BUG_ON(ovs_identifier_is_ufid(&flow->id));
695 return cmp_key(flow->id.unmasked_key, key, key_start, key_end);
696 }
697
698 static struct sw_flow *masked_flow_lookup(struct table_instance *ti,
699 const struct sw_flow_key *unmasked,
700 const struct sw_flow_mask *mask,
701 u32 *n_mask_hit)
702 {
703 struct sw_flow *flow;
704 struct hlist_head *head;
705 u32 hash;
706 struct sw_flow_key masked_key;
707
708 ovs_flow_mask_key(&masked_key, unmasked, false, mask);
709 hash = flow_hash(&masked_key, &mask->range);
710 head = find_bucket(ti, hash);
711 (*n_mask_hit)++;
712
713 hlist_for_each_entry_rcu(flow, head, flow_table.node[ti->node_ver],
714 lockdep_ovsl_is_held()) {
715 if (flow->mask == mask && flow->flow_table.hash == hash &&
716 flow_cmp_masked_key(flow, &masked_key, &mask->range))
717 return flow;
718 }
719 return NULL;
720 }
721
722 /* Flow lookup does full lookup on flow table. It starts with
723 * mask from index passed in *index.
724 * This function MUST be called with BH disabled due to the use
725 * of CPU specific variables.
726 */
727 static struct sw_flow *flow_lookup(struct flow_table *tbl,
728 struct table_instance *ti,
729 struct mask_array *ma,
730 const struct sw_flow_key *key,
731 u32 *n_mask_hit,
732 u32 *n_cache_hit,
733 u32 *index)
734 {
735 struct mask_array_stats *stats = this_cpu_ptr(ma->masks_usage_stats);
736 struct sw_flow *flow;
737 struct sw_flow_mask *mask;
738 int i;
739
740 if (likely(*index < ma->max)) {
741 mask = rcu_dereference_ovsl(ma->masks[*index]);
742 if (mask) {
743 flow = masked_flow_lookup(ti, key, mask, n_mask_hit);
744 if (flow) {
745 u64_stats_update_begin(&stats->syncp);
746 stats->usage_cntrs[*index]++;
747 u64_stats_update_end(&stats->syncp);
748 (*n_cache_hit)++;
749 return flow;
750 }
751 }
752 }
753
754 for (i = 0; i < ma->max; i++) {
755
756 if (i == *index)
757 continue;
758
759 mask = rcu_dereference_ovsl(ma->masks[i]);
760 if (unlikely(!mask))
761 break;
762
763 flow = masked_flow_lookup(ti, key, mask, n_mask_hit);
764 if (flow) { /* Found */
765 *index = i;
766 u64_stats_update_begin(&stats->syncp);
767 stats->usage_cntrs[*index]++;
768 u64_stats_update_end(&stats->syncp);
769 return flow;
770 }
771 }
772
773 return NULL;
774 }
775
776 /*
777 * mask_cache maps flow to probable mask. This cache is not tightly
778 * coupled cache, It means updates to mask list can result in inconsistent
779 * cache entry in mask cache.
780 * This is per cpu cache and is divided in MC_HASH_SEGS segments.
781 * In case of a hash collision the entry is hashed in next segment.
782 * */
783 struct sw_flow *ovs_flow_tbl_lookup_stats(struct flow_table *tbl,
784 const struct sw_flow_key *key,
785 u32 skb_hash,
786 u32 *n_mask_hit,
787 u32 *n_cache_hit)
788 {
789 struct mask_cache *mc = rcu_dereference(tbl->mask_cache);
790 struct mask_array *ma = rcu_dereference(tbl->mask_array);
791 struct table_instance *ti = rcu_dereference(tbl->ti);
792 struct mask_cache_entry *entries, *ce;
793 struct sw_flow *flow;
794 u32 hash;
795 int seg;
796
797 *n_mask_hit = 0;
798 *n_cache_hit = 0;
799 if (unlikely(!skb_hash || mc->cache_size == 0)) {
800 u32 mask_index = 0;
801 u32 cache = 0;
802
803 return flow_lookup(tbl, ti, ma, key, n_mask_hit, &cache,
804 &mask_index);
805 }
806
807 /* Pre and post recirulation flows usually have the same skb_hash
808 * value. To avoid hash collisions, rehash the 'skb_hash' with
809 * 'recirc_id'. */
810 if (key->recirc_id)
811 skb_hash = jhash_1word(skb_hash, key->recirc_id);
812
813 ce = NULL;
814 hash = skb_hash;
815 entries = this_cpu_ptr(mc->mask_cache);
816
817 /* Find the cache entry 'ce' to operate on. */
818 for (seg = 0; seg < MC_HASH_SEGS; seg++) {
819 int index = hash & (mc->cache_size - 1);
820 struct mask_cache_entry *e;
821
822 e = &entries[index];
823 if (e->skb_hash == skb_hash) {
824 flow = flow_lookup(tbl, ti, ma, key, n_mask_hit,
825 n_cache_hit, &e->mask_index);
826 if (!flow)
827 e->skb_hash = 0;
828 return flow;
829 }
830
831 if (!ce || e->skb_hash < ce->skb_hash)
832 ce = e; /* A better replacement cache candidate. */
833
834 hash >>= MC_HASH_SHIFT;
835 }
836
837 /* Cache miss, do full lookup. */
838 flow = flow_lookup(tbl, ti, ma, key, n_mask_hit, n_cache_hit,
839 &ce->mask_index);
840 if (flow)
841 ce->skb_hash = skb_hash;
842
843 *n_cache_hit = 0;
844 return flow;
845 }
846
847 struct sw_flow *ovs_flow_tbl_lookup(struct flow_table *tbl,
848 const struct sw_flow_key *key)
849 {
850 struct table_instance *ti = rcu_dereference_ovsl(tbl->ti);
851 struct mask_array *ma = rcu_dereference_ovsl(tbl->mask_array);
852 u32 __always_unused n_mask_hit;
853 u32 __always_unused n_cache_hit;
854 struct sw_flow *flow;
855 u32 index = 0;
856
857 /* This function gets called trough the netlink interface and therefore
858 * is preemptible. However, flow_lookup() function needs to be called
859 * with BH disabled due to CPU specific variables.
860 */
861 local_bh_disable();
862 flow = flow_lookup(tbl, ti, ma, key, &n_mask_hit, &n_cache_hit, &index);
863 local_bh_enable();
864 return flow;
865 }
866
867 struct sw_flow *ovs_flow_tbl_lookup_exact(struct flow_table *tbl,
868 const struct sw_flow_match *match)
869 {
870 struct mask_array *ma = ovsl_dereference(tbl->mask_array);
871 int i;
872
873 /* Always called under ovs-mutex. */
874 for (i = 0; i < ma->max; i++) {
875 struct table_instance *ti = rcu_dereference_ovsl(tbl->ti);
876 u32 __always_unused n_mask_hit;
877 struct sw_flow_mask *mask;
878 struct sw_flow *flow;
879
880 mask = ovsl_dereference(ma->masks[i]);
881 if (!mask)
882 continue;
883
884 flow = masked_flow_lookup(ti, match->key, mask, &n_mask_hit);
885 if (flow && ovs_identifier_is_key(&flow->id) &&
886 ovs_flow_cmp_unmasked_key(flow, match)) {
887 return flow;
888 }
889 }
890
891 return NULL;
892 }
893
894 static u32 ufid_hash(const struct sw_flow_id *sfid)
895 {
896 return jhash(sfid->ufid, sfid->ufid_len, 0);
897 }
898
899 static bool ovs_flow_cmp_ufid(const struct sw_flow *flow,
900 const struct sw_flow_id *sfid)
901 {
902 if (flow->id.ufid_len != sfid->ufid_len)
903 return false;
904
905 return !memcmp(flow->id.ufid, sfid->ufid, sfid->ufid_len);
906 }
907
908 bool ovs_flow_cmp(const struct sw_flow *flow,
909 const struct sw_flow_match *match)
910 {
911 if (ovs_identifier_is_ufid(&flow->id))
912 return flow_cmp_masked_key(flow, match->key, &match->range);
913
914 return ovs_flow_cmp_unmasked_key(flow, match);
915 }
916
917 struct sw_flow *ovs_flow_tbl_lookup_ufid(struct flow_table *tbl,
918 const struct sw_flow_id *ufid)
919 {
920 struct table_instance *ti = rcu_dereference_ovsl(tbl->ufid_ti);
921 struct sw_flow *flow;
922 struct hlist_head *head;
923 u32 hash;
924
925 hash = ufid_hash(ufid);
926 head = find_bucket(ti, hash);
927 hlist_for_each_entry_rcu(flow, head, ufid_table.node[ti->node_ver],
928 lockdep_ovsl_is_held()) {
929 if (flow->ufid_table.hash == hash &&
930 ovs_flow_cmp_ufid(flow, ufid))
931 return flow;
932 }
933 return NULL;
934 }
935
936 int ovs_flow_tbl_num_masks(const struct flow_table *table)
937 {
938 struct mask_array *ma = rcu_dereference_ovsl(table->mask_array);
939 return READ_ONCE(ma->count);
940 }
941
942 u32 ovs_flow_tbl_masks_cache_size(const struct flow_table *table)
943 {
944 struct mask_cache *mc = rcu_dereference_ovsl(table->mask_cache);
945
946 return READ_ONCE(mc->cache_size);
947 }
948
949 static struct table_instance *table_instance_expand(struct table_instance *ti,
950 bool ufid)
951 {
952 return table_instance_rehash(ti, ti->n_buckets * 2, ufid);
953 }
954
955 /* Must be called with OVS mutex held. */
956 void ovs_flow_tbl_remove(struct flow_table *table, struct sw_flow *flow)
957 {
958 struct table_instance *ti = ovsl_dereference(table->ti);
959 struct table_instance *ufid_ti = ovsl_dereference(table->ufid_ti);
960
961 BUG_ON(table->count == 0);
962 table_instance_flow_free(table, ti, ufid_ti, flow);
963 }
964
965 static struct sw_flow_mask *mask_alloc(void)
966 {
967 struct sw_flow_mask *mask;
968
969 mask = kmalloc(sizeof(*mask), GFP_KERNEL);
970 if (mask)
971 mask->ref_count = 1;
972
973 return mask;
974 }
975
976 static bool mask_equal(const struct sw_flow_mask *a,
977 const struct sw_flow_mask *b)
978 {
979 const u8 *a_ = (const u8 *)&a->key + a->range.start;
980 const u8 *b_ = (const u8 *)&b->key + b->range.start;
981
982 return (a->range.end == b->range.end)
983 && (a->range.start == b->range.start)
984 && (memcmp(a_, b_, range_n_bytes(&a->range)) == 0);
985 }
986
987 static struct sw_flow_mask *flow_mask_find(const struct flow_table *tbl,
988 const struct sw_flow_mask *mask)
989 {
990 struct mask_array *ma;
991 int i;
992
993 ma = ovsl_dereference(tbl->mask_array);
994 for (i = 0; i < ma->max; i++) {
995 struct sw_flow_mask *t;
996 t = ovsl_dereference(ma->masks[i]);
997
998 if (t && mask_equal(mask, t))
999 return t;
1000 }
1001
1002 return NULL;
1003 }
1004
1005 /* Add 'mask' into the mask list, if it is not already there. */
1006 static int flow_mask_insert(struct flow_table *tbl, struct sw_flow *flow,
1007 const struct sw_flow_mask *new)
1008 {
1009 struct sw_flow_mask *mask;
1010
1011 mask = flow_mask_find(tbl, new);
1012 if (!mask) {
1013 /* Allocate a new mask if none exists. */
1014 mask = mask_alloc();
1015 if (!mask)
1016 return -ENOMEM;
1017 mask->key = new->key;
1018 mask->range = new->range;
1019
1020 /* Add mask to mask-list. */
1021 if (tbl_mask_array_add_mask(tbl, mask)) {
1022 kfree(mask);
1023 return -ENOMEM;
1024 }
1025 } else {
1026 BUG_ON(!mask->ref_count);
1027 mask->ref_count++;
1028 }
1029
1030 flow->mask = mask;
1031 return 0;
1032 }
1033
1034 /* Must be called with OVS mutex held. */
1035 static void flow_key_insert(struct flow_table *table, struct sw_flow *flow)
1036 {
1037 struct table_instance *new_ti = NULL;
1038 struct table_instance *ti;
1039
1040 flow->flow_table.hash = flow_hash(&flow->key, &flow->mask->range);
1041 ti = ovsl_dereference(table->ti);
1042 table_instance_insert(ti, flow);
1043 table->count++;
1044
1045 /* Expand table, if necessary, to make room. */
1046 if (table->count > ti->n_buckets)
1047 new_ti = table_instance_expand(ti, false);
1048 else if (time_after(jiffies, table->last_rehash + REHASH_INTERVAL))
1049 new_ti = table_instance_rehash(ti, ti->n_buckets, false);
1050
1051 if (new_ti) {
1052 rcu_assign_pointer(table->ti, new_ti);
1053 call_rcu(&ti->rcu, flow_tbl_destroy_rcu_cb);
1054 table->last_rehash = jiffies;
1055 }
1056 }
1057
1058 /* Must be called with OVS mutex held. */
1059 static void flow_ufid_insert(struct flow_table *table, struct sw_flow *flow)
1060 {
1061 struct table_instance *ti;
1062
1063 flow->ufid_table.hash = ufid_hash(&flow->id);
1064 ti = ovsl_dereference(table->ufid_ti);
1065 ufid_table_instance_insert(ti, flow);
1066 table->ufid_count++;
1067
1068 /* Expand table, if necessary, to make room. */
1069 if (table->ufid_count > ti->n_buckets) {
1070 struct table_instance *new_ti;
1071
1072 new_ti = table_instance_expand(ti, true);
1073 if (new_ti) {
1074 rcu_assign_pointer(table->ufid_ti, new_ti);
1075 call_rcu(&ti->rcu, flow_tbl_destroy_rcu_cb);
1076 }
1077 }
1078 }
1079
1080 /* Must be called with OVS mutex held. */
1081 int ovs_flow_tbl_insert(struct flow_table *table, struct sw_flow *flow,
1082 const struct sw_flow_mask *mask)
1083 {
1084 int err;
1085
1086 err = flow_mask_insert(table, flow, mask);
1087 if (err)
1088 return err;
1089 flow_key_insert(table, flow);
1090 if (ovs_identifier_is_ufid(&flow->id))
1091 flow_ufid_insert(table, flow);
1092
1093 return 0;
1094 }
1095
1096 static int compare_mask_and_count(const void *a, const void *b)
1097 {
1098 const struct mask_count *mc_a = a;
1099 const struct mask_count *mc_b = b;
1100
1101 return (s64)mc_b->counter - (s64)mc_a->counter;
1102 }
1103
1104 /* Must be called with OVS mutex held. */
1105 void ovs_flow_masks_rebalance(struct flow_table *table)
1106 {
1107 struct mask_array *ma = rcu_dereference_ovsl(table->mask_array);
1108 struct mask_count *masks_and_count;
1109 struct mask_array *new;
1110 int masks_entries = 0;
1111 int i;
1112
1113 /* Build array of all current entries with use counters. */
1114 masks_and_count = kmalloc_array(ma->max, sizeof(*masks_and_count),
1115 GFP_KERNEL);
1116 if (!masks_and_count)
1117 return;
1118
1119 for (i = 0; i < ma->max; i++) {
1120 struct sw_flow_mask *mask;
1121 int cpu;
1122
1123 mask = rcu_dereference_ovsl(ma->masks[i]);
1124 if (unlikely(!mask))
1125 break;
1126
1127 masks_and_count[i].index = i;
1128 masks_and_count[i].counter = 0;
1129
1130 for_each_possible_cpu(cpu) {
1131 struct mask_array_stats *stats;
1132 unsigned int start;
1133 u64 counter;
1134
1135 stats = per_cpu_ptr(ma->masks_usage_stats, cpu);
1136 do {
1137 start = u64_stats_fetch_begin(&stats->syncp);
1138 counter = stats->usage_cntrs[i];
1139 } while (u64_stats_fetch_retry(&stats->syncp, start));
1140
1141 masks_and_count[i].counter += counter;
1142 }
1143
1144 /* Subtract the zero count value. */
1145 masks_and_count[i].counter -= ma->masks_usage_zero_cntr[i];
1146
1147 /* Rather than calling tbl_mask_array_reset_counters()
1148 * below when no change is needed, do it inline here.
1149 */
1150 ma->masks_usage_zero_cntr[i] += masks_and_count[i].counter;
1151 }
1152
1153 if (i == 0)
1154 goto free_mask_entries;
1155
1156 /* Sort the entries */
1157 masks_entries = i;
1158 sort(masks_and_count, masks_entries, sizeof(*masks_and_count),
1159 compare_mask_and_count, NULL);
1160
1161 /* If the order is the same, nothing to do... */
1162 for (i = 0; i < masks_entries; i++) {
1163 if (i != masks_and_count[i].index)
1164 break;
1165 }
1166 if (i == masks_entries)
1167 goto free_mask_entries;
1168
1169 /* Rebuilt the new list in order of usage. */
1170 new = tbl_mask_array_alloc(ma->max);
1171 if (!new)
1172 goto free_mask_entries;
1173
1174 for (i = 0; i < masks_entries; i++) {
1175 int index = masks_and_count[i].index;
1176
1177 if (ovsl_dereference(ma->masks[index]))
1178 new->masks[new->count++] = ma->masks[index];
1179 }
1180
1181 rcu_assign_pointer(table->mask_array, new);
1182 call_rcu(&ma->rcu, mask_array_rcu_cb);
1183
1184 free_mask_entries:
1185 kfree(masks_and_count);
1186 }
1187
1188 /* Initializes the flow module.
1189 * Returns zero if successful or a negative error code. */
1190 int ovs_flow_init(void)
1191 {
1192 BUILD_BUG_ON(__alignof__(struct sw_flow_key) % __alignof__(long));
1193 BUILD_BUG_ON(sizeof(struct sw_flow_key) % sizeof(long));
1194
1195 flow_cache = kmem_cache_create("sw_flow", sizeof(struct sw_flow)
1196 + (nr_cpu_ids
1197 * sizeof(struct sw_flow_stats *))
1198 + cpumask_size(),
1199 0, 0, NULL);
1200 if (flow_cache == NULL)
1201 return -ENOMEM;
1202
1203 flow_stats_cache
1204 = kmem_cache_create("sw_flow_stats", sizeof(struct sw_flow_stats),
1205 0, SLAB_HWCACHE_ALIGN, NULL);
1206 if (flow_stats_cache == NULL) {
1207 kmem_cache_destroy(flow_cache);
1208 flow_cache = NULL;
1209 return -ENOMEM;
1210 }
1211
1212 return 0;
1213 }
1214
1215 /* Uninitializes the flow module. */
1216 void ovs_flow_exit(void)
1217 {
1218 kmem_cache_destroy(flow_stats_cache);
1219 kmem_cache_destroy(flow_cache);
1220 }
1221
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.