xref: /openbmc/linux/net/openvswitch/flow_table.c (revision 9dbbc3b9)
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 
83 	/* Initialize the default stat node. */
84 	stats = kmem_cache_alloc_node(flow_stats_cache,
85 				      GFP_KERNEL | __GFP_ZERO,
86 				      node_online(0) ? 0 : NUMA_NO_NODE);
87 	if (!stats)
88 		goto err;
89 
90 	spin_lock_init(&stats->lock);
91 
92 	RCU_INIT_POINTER(flow->stats[0], stats);
93 
94 	cpumask_set_cpu(0, &flow->cpu_used_mask);
95 
96 	return flow;
97 err:
98 	kmem_cache_free(flow_cache, flow);
99 	return ERR_PTR(-ENOMEM);
100 }
101 
102 int ovs_flow_tbl_count(const struct flow_table *table)
103 {
104 	return table->count;
105 }
106 
107 static void flow_free(struct sw_flow *flow)
108 {
109 	int cpu;
110 
111 	if (ovs_identifier_is_key(&flow->id))
112 		kfree(flow->id.unmasked_key);
113 	if (flow->sf_acts)
114 		ovs_nla_free_flow_actions((struct sw_flow_actions __force *)
115 					  flow->sf_acts);
116 	/* We open code this to make sure cpu 0 is always considered */
117 	for (cpu = 0; cpu < nr_cpu_ids;
118 	     cpu = cpumask_next(cpu, &flow->cpu_used_mask)) {
119 		if (flow->stats[cpu])
120 			kmem_cache_free(flow_stats_cache,
121 					(struct sw_flow_stats __force *)flow->stats[cpu]);
122 	}
123 
124 	kmem_cache_free(flow_cache, flow);
125 }
126 
127 static void rcu_free_flow_callback(struct rcu_head *rcu)
128 {
129 	struct sw_flow *flow = container_of(rcu, struct sw_flow, rcu);
130 
131 	flow_free(flow);
132 }
133 
134 void ovs_flow_free(struct sw_flow *flow, bool deferred)
135 {
136 	if (!flow)
137 		return;
138 
139 	if (deferred)
140 		call_rcu(&flow->rcu, rcu_free_flow_callback);
141 	else
142 		flow_free(flow);
143 }
144 
145 static void __table_instance_destroy(struct table_instance *ti)
146 {
147 	kvfree(ti->buckets);
148 	kfree(ti);
149 }
150 
151 static struct table_instance *table_instance_alloc(int new_size)
152 {
153 	struct table_instance *ti = kmalloc(sizeof(*ti), GFP_KERNEL);
154 	int i;
155 
156 	if (!ti)
157 		return NULL;
158 
159 	ti->buckets = kvmalloc_array(new_size, sizeof(struct hlist_head),
160 				     GFP_KERNEL);
161 	if (!ti->buckets) {
162 		kfree(ti);
163 		return NULL;
164 	}
165 
166 	for (i = 0; i < new_size; i++)
167 		INIT_HLIST_HEAD(&ti->buckets[i]);
168 
169 	ti->n_buckets = new_size;
170 	ti->node_ver = 0;
171 	get_random_bytes(&ti->hash_seed, sizeof(u32));
172 
173 	return ti;
174 }
175 
176 static void __mask_array_destroy(struct mask_array *ma)
177 {
178 	free_percpu(ma->masks_usage_stats);
179 	kfree(ma);
180 }
181 
182 static void mask_array_rcu_cb(struct rcu_head *rcu)
183 {
184 	struct mask_array *ma = container_of(rcu, struct mask_array, rcu);
185 
186 	__mask_array_destroy(ma);
187 }
188 
189 static void tbl_mask_array_reset_counters(struct mask_array *ma)
190 {
191 	int i, cpu;
192 
193 	/* As the per CPU counters are not atomic we can not go ahead and
194 	 * reset them from another CPU. To be able to still have an approximate
195 	 * zero based counter we store the value at reset, and subtract it
196 	 * later when processing.
197 	 */
198 	for (i = 0; i < ma->max; i++) {
199 		ma->masks_usage_zero_cntr[i] = 0;
200 
201 		for_each_possible_cpu(cpu) {
202 			struct mask_array_stats *stats;
203 			unsigned int start;
204 			u64 counter;
205 
206 			stats = per_cpu_ptr(ma->masks_usage_stats, cpu);
207 			do {
208 				start = u64_stats_fetch_begin_irq(&stats->syncp);
209 				counter = stats->usage_cntrs[i];
210 			} while (u64_stats_fetch_retry_irq(&stats->syncp, start));
211 
212 			ma->masks_usage_zero_cntr[i] += counter;
213 		}
214 	}
215 }
216 
217 static struct mask_array *tbl_mask_array_alloc(int size)
218 {
219 	struct mask_array *new;
220 
221 	size = max(MASK_ARRAY_SIZE_MIN, size);
222 	new = kzalloc(sizeof(struct mask_array) +
223 		      sizeof(struct sw_flow_mask *) * size +
224 		      sizeof(u64) * size, GFP_KERNEL);
225 	if (!new)
226 		return NULL;
227 
228 	new->masks_usage_zero_cntr = (u64 *)((u8 *)new +
229 					     sizeof(struct mask_array) +
230 					     sizeof(struct sw_flow_mask *) *
231 					     size);
232 
233 	new->masks_usage_stats = __alloc_percpu(sizeof(struct mask_array_stats) +
234 						sizeof(u64) * size,
235 						__alignof__(u64));
236 	if (!new->masks_usage_stats) {
237 		kfree(new);
238 		return NULL;
239 	}
240 
241 	new->count = 0;
242 	new->max = size;
243 
244 	return new;
245 }
246 
247 static int tbl_mask_array_realloc(struct flow_table *tbl, int size)
248 {
249 	struct mask_array *old;
250 	struct mask_array *new;
251 
252 	new = tbl_mask_array_alloc(size);
253 	if (!new)
254 		return -ENOMEM;
255 
256 	old = ovsl_dereference(tbl->mask_array);
257 	if (old) {
258 		int i;
259 
260 		for (i = 0; i < old->max; i++) {
261 			if (ovsl_dereference(old->masks[i]))
262 				new->masks[new->count++] = old->masks[i];
263 		}
264 		call_rcu(&old->rcu, mask_array_rcu_cb);
265 	}
266 
267 	rcu_assign_pointer(tbl->mask_array, new);
268 
269 	return 0;
270 }
271 
272 static int tbl_mask_array_add_mask(struct flow_table *tbl,
273 				   struct sw_flow_mask *new)
274 {
275 	struct mask_array *ma = ovsl_dereference(tbl->mask_array);
276 	int err, ma_count = READ_ONCE(ma->count);
277 
278 	if (ma_count >= ma->max) {
279 		err = tbl_mask_array_realloc(tbl, ma->max +
280 						  MASK_ARRAY_SIZE_MIN);
281 		if (err)
282 			return err;
283 
284 		ma = ovsl_dereference(tbl->mask_array);
285 	} else {
286 		/* On every add or delete we need to reset the counters so
287 		 * every new mask gets a fair chance of being prioritized.
288 		 */
289 		tbl_mask_array_reset_counters(ma);
290 	}
291 
292 	BUG_ON(ovsl_dereference(ma->masks[ma_count]));
293 
294 	rcu_assign_pointer(ma->masks[ma_count], new);
295 	WRITE_ONCE(ma->count, ma_count + 1);
296 
297 	return 0;
298 }
299 
300 static void tbl_mask_array_del_mask(struct flow_table *tbl,
301 				    struct sw_flow_mask *mask)
302 {
303 	struct mask_array *ma = ovsl_dereference(tbl->mask_array);
304 	int i, ma_count = READ_ONCE(ma->count);
305 
306 	/* Remove the deleted mask pointers from the array */
307 	for (i = 0; i < ma_count; i++) {
308 		if (mask == ovsl_dereference(ma->masks[i]))
309 			goto found;
310 	}
311 
312 	BUG();
313 	return;
314 
315 found:
316 	WRITE_ONCE(ma->count, ma_count - 1);
317 
318 	rcu_assign_pointer(ma->masks[i], ma->masks[ma_count - 1]);
319 	RCU_INIT_POINTER(ma->masks[ma_count - 1], NULL);
320 
321 	kfree_rcu(mask, rcu);
322 
323 	/* Shrink the mask array if necessary. */
324 	if (ma->max >= (MASK_ARRAY_SIZE_MIN * 2) &&
325 	    ma_count <= (ma->max / 3))
326 		tbl_mask_array_realloc(tbl, ma->max / 2);
327 	else
328 		tbl_mask_array_reset_counters(ma);
329 
330 }
331 
332 /* Remove 'mask' from the mask list, if it is not needed any more. */
333 static void flow_mask_remove(struct flow_table *tbl, struct sw_flow_mask *mask)
334 {
335 	if (mask) {
336 		/* ovs-lock is required to protect mask-refcount and
337 		 * mask list.
338 		 */
339 		ASSERT_OVSL();
340 		BUG_ON(!mask->ref_count);
341 		mask->ref_count--;
342 
343 		if (!mask->ref_count)
344 			tbl_mask_array_del_mask(tbl, mask);
345 	}
346 }
347 
348 static void __mask_cache_destroy(struct mask_cache *mc)
349 {
350 	free_percpu(mc->mask_cache);
351 	kfree(mc);
352 }
353 
354 static void mask_cache_rcu_cb(struct rcu_head *rcu)
355 {
356 	struct mask_cache *mc = container_of(rcu, struct mask_cache, rcu);
357 
358 	__mask_cache_destroy(mc);
359 }
360 
361 static struct mask_cache *tbl_mask_cache_alloc(u32 size)
362 {
363 	struct mask_cache_entry __percpu *cache = NULL;
364 	struct mask_cache *new;
365 
366 	/* Only allow size to be 0, or a power of 2, and does not exceed
367 	 * percpu allocation size.
368 	 */
369 	if ((!is_power_of_2(size) && size != 0) ||
370 	    (size * sizeof(struct mask_cache_entry)) > PCPU_MIN_UNIT_SIZE)
371 		return NULL;
372 
373 	new = kzalloc(sizeof(*new), GFP_KERNEL);
374 	if (!new)
375 		return NULL;
376 
377 	new->cache_size = size;
378 	if (new->cache_size > 0) {
379 		cache = __alloc_percpu(array_size(sizeof(struct mask_cache_entry),
380 						  new->cache_size),
381 				       __alignof__(struct mask_cache_entry));
382 		if (!cache) {
383 			kfree(new);
384 			return NULL;
385 		}
386 	}
387 
388 	new->mask_cache = cache;
389 	return new;
390 }
391 int ovs_flow_tbl_masks_cache_resize(struct flow_table *table, u32 size)
392 {
393 	struct mask_cache *mc = rcu_dereference_ovsl(table->mask_cache);
394 	struct mask_cache *new;
395 
396 	if (size == mc->cache_size)
397 		return 0;
398 
399 	if ((!is_power_of_2(size) && size != 0) ||
400 	    (size * sizeof(struct mask_cache_entry)) > PCPU_MIN_UNIT_SIZE)
401 		return -EINVAL;
402 
403 	new = tbl_mask_cache_alloc(size);
404 	if (!new)
405 		return -ENOMEM;
406 
407 	rcu_assign_pointer(table->mask_cache, new);
408 	call_rcu(&mc->rcu, mask_cache_rcu_cb);
409 
410 	return 0;
411 }
412 
413 int ovs_flow_tbl_init(struct flow_table *table)
414 {
415 	struct table_instance *ti, *ufid_ti;
416 	struct mask_cache *mc;
417 	struct mask_array *ma;
418 
419 	mc = tbl_mask_cache_alloc(MC_DEFAULT_HASH_ENTRIES);
420 	if (!mc)
421 		return -ENOMEM;
422 
423 	ma = tbl_mask_array_alloc(MASK_ARRAY_SIZE_MIN);
424 	if (!ma)
425 		goto free_mask_cache;
426 
427 	ti = table_instance_alloc(TBL_MIN_BUCKETS);
428 	if (!ti)
429 		goto free_mask_array;
430 
431 	ufid_ti = table_instance_alloc(TBL_MIN_BUCKETS);
432 	if (!ufid_ti)
433 		goto free_ti;
434 
435 	rcu_assign_pointer(table->ti, ti);
436 	rcu_assign_pointer(table->ufid_ti, ufid_ti);
437 	rcu_assign_pointer(table->mask_array, ma);
438 	rcu_assign_pointer(table->mask_cache, mc);
439 	table->last_rehash = jiffies;
440 	table->count = 0;
441 	table->ufid_count = 0;
442 	return 0;
443 
444 free_ti:
445 	__table_instance_destroy(ti);
446 free_mask_array:
447 	__mask_array_destroy(ma);
448 free_mask_cache:
449 	__mask_cache_destroy(mc);
450 	return -ENOMEM;
451 }
452 
453 static void flow_tbl_destroy_rcu_cb(struct rcu_head *rcu)
454 {
455 	struct table_instance *ti;
456 
457 	ti = container_of(rcu, struct table_instance, rcu);
458 	__table_instance_destroy(ti);
459 }
460 
461 static void table_instance_flow_free(struct flow_table *table,
462 				     struct table_instance *ti,
463 				     struct table_instance *ufid_ti,
464 				     struct sw_flow *flow)
465 {
466 	hlist_del_rcu(&flow->flow_table.node[ti->node_ver]);
467 	table->count--;
468 
469 	if (ovs_identifier_is_ufid(&flow->id)) {
470 		hlist_del_rcu(&flow->ufid_table.node[ufid_ti->node_ver]);
471 		table->ufid_count--;
472 	}
473 
474 	flow_mask_remove(table, flow->mask);
475 }
476 
477 /* Must be called with OVS mutex held. */
478 void table_instance_flow_flush(struct flow_table *table,
479 			       struct table_instance *ti,
480 			       struct table_instance *ufid_ti)
481 {
482 	int i;
483 
484 	for (i = 0; i < ti->n_buckets; i++) {
485 		struct hlist_head *head = &ti->buckets[i];
486 		struct hlist_node *n;
487 		struct sw_flow *flow;
488 
489 		hlist_for_each_entry_safe(flow, n, head,
490 					  flow_table.node[ti->node_ver]) {
491 
492 			table_instance_flow_free(table, ti, ufid_ti,
493 						 flow);
494 			ovs_flow_free(flow, true);
495 		}
496 	}
497 
498 	if (WARN_ON(table->count != 0 ||
499 		    table->ufid_count != 0)) {
500 		table->count = 0;
501 		table->ufid_count = 0;
502 	}
503 }
504 
505 static void table_instance_destroy(struct table_instance *ti,
506 				   struct table_instance *ufid_ti)
507 {
508 	call_rcu(&ti->rcu, flow_tbl_destroy_rcu_cb);
509 	call_rcu(&ufid_ti->rcu, flow_tbl_destroy_rcu_cb);
510 }
511 
512 /* No need for locking this function is called from RCU callback or
513  * error path.
514  */
515 void ovs_flow_tbl_destroy(struct flow_table *table)
516 {
517 	struct table_instance *ti = rcu_dereference_raw(table->ti);
518 	struct table_instance *ufid_ti = rcu_dereference_raw(table->ufid_ti);
519 	struct mask_cache *mc = rcu_dereference_raw(table->mask_cache);
520 	struct mask_array *ma = rcu_dereference_raw(table->mask_array);
521 
522 	call_rcu(&mc->rcu, mask_cache_rcu_cb);
523 	call_rcu(&ma->rcu, mask_array_rcu_cb);
524 	table_instance_destroy(ti, ufid_ti);
525 }
526 
527 struct sw_flow *ovs_flow_tbl_dump_next(struct table_instance *ti,
528 				       u32 *bucket, u32 *last)
529 {
530 	struct sw_flow *flow;
531 	struct hlist_head *head;
532 	int ver;
533 	int i;
534 
535 	ver = ti->node_ver;
536 	while (*bucket < ti->n_buckets) {
537 		i = 0;
538 		head = &ti->buckets[*bucket];
539 		hlist_for_each_entry_rcu(flow, head, flow_table.node[ver]) {
540 			if (i < *last) {
541 				i++;
542 				continue;
543 			}
544 			*last = i + 1;
545 			return flow;
546 		}
547 		(*bucket)++;
548 		*last = 0;
549 	}
550 
551 	return NULL;
552 }
553 
554 static struct hlist_head *find_bucket(struct table_instance *ti, u32 hash)
555 {
556 	hash = jhash_1word(hash, ti->hash_seed);
557 	return &ti->buckets[hash & (ti->n_buckets - 1)];
558 }
559 
560 static void table_instance_insert(struct table_instance *ti,
561 				  struct sw_flow *flow)
562 {
563 	struct hlist_head *head;
564 
565 	head = find_bucket(ti, flow->flow_table.hash);
566 	hlist_add_head_rcu(&flow->flow_table.node[ti->node_ver], head);
567 }
568 
569 static void ufid_table_instance_insert(struct table_instance *ti,
570 				       struct sw_flow *flow)
571 {
572 	struct hlist_head *head;
573 
574 	head = find_bucket(ti, flow->ufid_table.hash);
575 	hlist_add_head_rcu(&flow->ufid_table.node[ti->node_ver], head);
576 }
577 
578 static void flow_table_copy_flows(struct table_instance *old,
579 				  struct table_instance *new, bool ufid)
580 {
581 	int old_ver;
582 	int i;
583 
584 	old_ver = old->node_ver;
585 	new->node_ver = !old_ver;
586 
587 	/* Insert in new table. */
588 	for (i = 0; i < old->n_buckets; i++) {
589 		struct sw_flow *flow;
590 		struct hlist_head *head = &old->buckets[i];
591 
592 		if (ufid)
593 			hlist_for_each_entry_rcu(flow, head,
594 						 ufid_table.node[old_ver],
595 						 lockdep_ovsl_is_held())
596 				ufid_table_instance_insert(new, flow);
597 		else
598 			hlist_for_each_entry_rcu(flow, head,
599 						 flow_table.node[old_ver],
600 						 lockdep_ovsl_is_held())
601 				table_instance_insert(new, flow);
602 	}
603 }
604 
605 static struct table_instance *table_instance_rehash(struct table_instance *ti,
606 						    int n_buckets, bool ufid)
607 {
608 	struct table_instance *new_ti;
609 
610 	new_ti = table_instance_alloc(n_buckets);
611 	if (!new_ti)
612 		return NULL;
613 
614 	flow_table_copy_flows(ti, new_ti, ufid);
615 
616 	return new_ti;
617 }
618 
619 int ovs_flow_tbl_flush(struct flow_table *flow_table)
620 {
621 	struct table_instance *old_ti, *new_ti;
622 	struct table_instance *old_ufid_ti, *new_ufid_ti;
623 
624 	new_ti = table_instance_alloc(TBL_MIN_BUCKETS);
625 	if (!new_ti)
626 		return -ENOMEM;
627 	new_ufid_ti = table_instance_alloc(TBL_MIN_BUCKETS);
628 	if (!new_ufid_ti)
629 		goto err_free_ti;
630 
631 	old_ti = ovsl_dereference(flow_table->ti);
632 	old_ufid_ti = ovsl_dereference(flow_table->ufid_ti);
633 
634 	rcu_assign_pointer(flow_table->ti, new_ti);
635 	rcu_assign_pointer(flow_table->ufid_ti, new_ufid_ti);
636 	flow_table->last_rehash = jiffies;
637 
638 	table_instance_flow_flush(flow_table, old_ti, old_ufid_ti);
639 	table_instance_destroy(old_ti, old_ufid_ti);
640 	return 0;
641 
642 err_free_ti:
643 	__table_instance_destroy(new_ti);
644 	return -ENOMEM;
645 }
646 
647 static u32 flow_hash(const struct sw_flow_key *key,
648 		     const struct sw_flow_key_range *range)
649 {
650 	const u32 *hash_key = (const u32 *)((const u8 *)key + range->start);
651 
652 	/* Make sure number of hash bytes are multiple of u32. */
653 	int hash_u32s = range_n_bytes(range) >> 2;
654 
655 	return jhash2(hash_key, hash_u32s, 0);
656 }
657 
658 static int flow_key_start(const struct sw_flow_key *key)
659 {
660 	if (key->tun_proto)
661 		return 0;
662 	else
663 		return rounddown(offsetof(struct sw_flow_key, phy),
664 				 sizeof(long));
665 }
666 
667 static bool cmp_key(const struct sw_flow_key *key1,
668 		    const struct sw_flow_key *key2,
669 		    int key_start, int key_end)
670 {
671 	const long *cp1 = (const long *)((const u8 *)key1 + key_start);
672 	const long *cp2 = (const long *)((const u8 *)key2 + key_start);
673 	long diffs = 0;
674 	int i;
675 
676 	for (i = key_start; i < key_end; i += sizeof(long))
677 		diffs |= *cp1++ ^ *cp2++;
678 
679 	return diffs == 0;
680 }
681 
682 static bool flow_cmp_masked_key(const struct sw_flow *flow,
683 				const struct sw_flow_key *key,
684 				const struct sw_flow_key_range *range)
685 {
686 	return cmp_key(&flow->key, key, range->start, range->end);
687 }
688 
689 static bool ovs_flow_cmp_unmasked_key(const struct sw_flow *flow,
690 				      const struct sw_flow_match *match)
691 {
692 	struct sw_flow_key *key = match->key;
693 	int key_start = flow_key_start(key);
694 	int key_end = match->range.end;
695 
696 	BUG_ON(ovs_identifier_is_ufid(&flow->id));
697 	return cmp_key(flow->id.unmasked_key, key, key_start, key_end);
698 }
699 
700 static struct sw_flow *masked_flow_lookup(struct table_instance *ti,
701 					  const struct sw_flow_key *unmasked,
702 					  const struct sw_flow_mask *mask,
703 					  u32 *n_mask_hit)
704 {
705 	struct sw_flow *flow;
706 	struct hlist_head *head;
707 	u32 hash;
708 	struct sw_flow_key masked_key;
709 
710 	ovs_flow_mask_key(&masked_key, unmasked, false, mask);
711 	hash = flow_hash(&masked_key, &mask->range);
712 	head = find_bucket(ti, hash);
713 	(*n_mask_hit)++;
714 
715 	hlist_for_each_entry_rcu(flow, head, flow_table.node[ti->node_ver],
716 				 lockdep_ovsl_is_held()) {
717 		if (flow->mask == mask && flow->flow_table.hash == hash &&
718 		    flow_cmp_masked_key(flow, &masked_key, &mask->range))
719 			return flow;
720 	}
721 	return NULL;
722 }
723 
724 /* Flow lookup does full lookup on flow table. It starts with
725  * mask from index passed in *index.
726  * This function MUST be called with BH disabled due to the use
727  * of CPU specific variables.
728  */
729 static struct sw_flow *flow_lookup(struct flow_table *tbl,
730 				   struct table_instance *ti,
731 				   struct mask_array *ma,
732 				   const struct sw_flow_key *key,
733 				   u32 *n_mask_hit,
734 				   u32 *n_cache_hit,
735 				   u32 *index)
736 {
737 	struct mask_array_stats *stats = this_cpu_ptr(ma->masks_usage_stats);
738 	struct sw_flow *flow;
739 	struct sw_flow_mask *mask;
740 	int i;
741 
742 	if (likely(*index < ma->max)) {
743 		mask = rcu_dereference_ovsl(ma->masks[*index]);
744 		if (mask) {
745 			flow = masked_flow_lookup(ti, key, mask, n_mask_hit);
746 			if (flow) {
747 				u64_stats_update_begin(&stats->syncp);
748 				stats->usage_cntrs[*index]++;
749 				u64_stats_update_end(&stats->syncp);
750 				(*n_cache_hit)++;
751 				return flow;
752 			}
753 		}
754 	}
755 
756 	for (i = 0; i < ma->max; i++)  {
757 
758 		if (i == *index)
759 			continue;
760 
761 		mask = rcu_dereference_ovsl(ma->masks[i]);
762 		if (unlikely(!mask))
763 			break;
764 
765 		flow = masked_flow_lookup(ti, key, mask, n_mask_hit);
766 		if (flow) { /* Found */
767 			*index = i;
768 			u64_stats_update_begin(&stats->syncp);
769 			stats->usage_cntrs[*index]++;
770 			u64_stats_update_end(&stats->syncp);
771 			return flow;
772 		}
773 	}
774 
775 	return NULL;
776 }
777 
778 /*
779  * mask_cache maps flow to probable mask. This cache is not tightly
780  * coupled cache, It means updates to  mask list can result in inconsistent
781  * cache entry in mask cache.
782  * This is per cpu cache and is divided in MC_HASH_SEGS segments.
783  * In case of a hash collision the entry is hashed in next segment.
784  * */
785 struct sw_flow *ovs_flow_tbl_lookup_stats(struct flow_table *tbl,
786 					  const struct sw_flow_key *key,
787 					  u32 skb_hash,
788 					  u32 *n_mask_hit,
789 					  u32 *n_cache_hit)
790 {
791 	struct mask_cache *mc = rcu_dereference(tbl->mask_cache);
792 	struct mask_array *ma = rcu_dereference(tbl->mask_array);
793 	struct table_instance *ti = rcu_dereference(tbl->ti);
794 	struct mask_cache_entry *entries, *ce;
795 	struct sw_flow *flow;
796 	u32 hash;
797 	int seg;
798 
799 	*n_mask_hit = 0;
800 	*n_cache_hit = 0;
801 	if (unlikely(!skb_hash || mc->cache_size == 0)) {
802 		u32 mask_index = 0;
803 		u32 cache = 0;
804 
805 		return flow_lookup(tbl, ti, ma, key, n_mask_hit, &cache,
806 				   &mask_index);
807 	}
808 
809 	/* Pre and post recirulation flows usually have the same skb_hash
810 	 * value. To avoid hash collisions, rehash the 'skb_hash' with
811 	 * 'recirc_id'.  */
812 	if (key->recirc_id)
813 		skb_hash = jhash_1word(skb_hash, key->recirc_id);
814 
815 	ce = NULL;
816 	hash = skb_hash;
817 	entries = this_cpu_ptr(mc->mask_cache);
818 
819 	/* Find the cache entry 'ce' to operate on. */
820 	for (seg = 0; seg < MC_HASH_SEGS; seg++) {
821 		int index = hash & (mc->cache_size - 1);
822 		struct mask_cache_entry *e;
823 
824 		e = &entries[index];
825 		if (e->skb_hash == skb_hash) {
826 			flow = flow_lookup(tbl, ti, ma, key, n_mask_hit,
827 					   n_cache_hit, &e->mask_index);
828 			if (!flow)
829 				e->skb_hash = 0;
830 			return flow;
831 		}
832 
833 		if (!ce || e->skb_hash < ce->skb_hash)
834 			ce = e;  /* A better replacement cache candidate. */
835 
836 		hash >>= MC_HASH_SHIFT;
837 	}
838 
839 	/* Cache miss, do full lookup. */
840 	flow = flow_lookup(tbl, ti, ma, key, n_mask_hit, n_cache_hit,
841 			   &ce->mask_index);
842 	if (flow)
843 		ce->skb_hash = skb_hash;
844 
845 	*n_cache_hit = 0;
846 	return flow;
847 }
848 
849 struct sw_flow *ovs_flow_tbl_lookup(struct flow_table *tbl,
850 				    const struct sw_flow_key *key)
851 {
852 	struct table_instance *ti = rcu_dereference_ovsl(tbl->ti);
853 	struct mask_array *ma = rcu_dereference_ovsl(tbl->mask_array);
854 	u32 __always_unused n_mask_hit;
855 	u32 __always_unused n_cache_hit;
856 	struct sw_flow *flow;
857 	u32 index = 0;
858 
859 	/* This function gets called trough the netlink interface and therefore
860 	 * is preemptible. However, flow_lookup() function needs to be called
861 	 * with BH disabled due to CPU specific variables.
862 	 */
863 	local_bh_disable();
864 	flow = flow_lookup(tbl, ti, ma, key, &n_mask_hit, &n_cache_hit, &index);
865 	local_bh_enable();
866 	return flow;
867 }
868 
869 struct sw_flow *ovs_flow_tbl_lookup_exact(struct flow_table *tbl,
870 					  const struct sw_flow_match *match)
871 {
872 	struct mask_array *ma = ovsl_dereference(tbl->mask_array);
873 	int i;
874 
875 	/* Always called under ovs-mutex. */
876 	for (i = 0; i < ma->max; i++) {
877 		struct table_instance *ti = rcu_dereference_ovsl(tbl->ti);
878 		u32 __always_unused n_mask_hit;
879 		struct sw_flow_mask *mask;
880 		struct sw_flow *flow;
881 
882 		mask = ovsl_dereference(ma->masks[i]);
883 		if (!mask)
884 			continue;
885 
886 		flow = masked_flow_lookup(ti, match->key, mask, &n_mask_hit);
887 		if (flow && ovs_identifier_is_key(&flow->id) &&
888 		    ovs_flow_cmp_unmasked_key(flow, match)) {
889 			return flow;
890 		}
891 	}
892 
893 	return NULL;
894 }
895 
896 static u32 ufid_hash(const struct sw_flow_id *sfid)
897 {
898 	return jhash(sfid->ufid, sfid->ufid_len, 0);
899 }
900 
901 static bool ovs_flow_cmp_ufid(const struct sw_flow *flow,
902 			      const struct sw_flow_id *sfid)
903 {
904 	if (flow->id.ufid_len != sfid->ufid_len)
905 		return false;
906 
907 	return !memcmp(flow->id.ufid, sfid->ufid, sfid->ufid_len);
908 }
909 
910 bool ovs_flow_cmp(const struct sw_flow *flow,
911 		  const struct sw_flow_match *match)
912 {
913 	if (ovs_identifier_is_ufid(&flow->id))
914 		return flow_cmp_masked_key(flow, match->key, &match->range);
915 
916 	return ovs_flow_cmp_unmasked_key(flow, match);
917 }
918 
919 struct sw_flow *ovs_flow_tbl_lookup_ufid(struct flow_table *tbl,
920 					 const struct sw_flow_id *ufid)
921 {
922 	struct table_instance *ti = rcu_dereference_ovsl(tbl->ufid_ti);
923 	struct sw_flow *flow;
924 	struct hlist_head *head;
925 	u32 hash;
926 
927 	hash = ufid_hash(ufid);
928 	head = find_bucket(ti, hash);
929 	hlist_for_each_entry_rcu(flow, head, ufid_table.node[ti->node_ver],
930 				 lockdep_ovsl_is_held()) {
931 		if (flow->ufid_table.hash == hash &&
932 		    ovs_flow_cmp_ufid(flow, ufid))
933 			return flow;
934 	}
935 	return NULL;
936 }
937 
938 int ovs_flow_tbl_num_masks(const struct flow_table *table)
939 {
940 	struct mask_array *ma = rcu_dereference_ovsl(table->mask_array);
941 	return READ_ONCE(ma->count);
942 }
943 
944 u32 ovs_flow_tbl_masks_cache_size(const struct flow_table *table)
945 {
946 	struct mask_cache *mc = rcu_dereference_ovsl(table->mask_cache);
947 
948 	return READ_ONCE(mc->cache_size);
949 }
950 
951 static struct table_instance *table_instance_expand(struct table_instance *ti,
952 						    bool ufid)
953 {
954 	return table_instance_rehash(ti, ti->n_buckets * 2, ufid);
955 }
956 
957 /* Must be called with OVS mutex held. */
958 void ovs_flow_tbl_remove(struct flow_table *table, struct sw_flow *flow)
959 {
960 	struct table_instance *ti = ovsl_dereference(table->ti);
961 	struct table_instance *ufid_ti = ovsl_dereference(table->ufid_ti);
962 
963 	BUG_ON(table->count == 0);
964 	table_instance_flow_free(table, ti, ufid_ti, flow);
965 }
966 
967 static struct sw_flow_mask *mask_alloc(void)
968 {
969 	struct sw_flow_mask *mask;
970 
971 	mask = kmalloc(sizeof(*mask), GFP_KERNEL);
972 	if (mask)
973 		mask->ref_count = 1;
974 
975 	return mask;
976 }
977 
978 static bool mask_equal(const struct sw_flow_mask *a,
979 		       const struct sw_flow_mask *b)
980 {
981 	const u8 *a_ = (const u8 *)&a->key + a->range.start;
982 	const u8 *b_ = (const u8 *)&b->key + b->range.start;
983 
984 	return  (a->range.end == b->range.end)
985 		&& (a->range.start == b->range.start)
986 		&& (memcmp(a_, b_, range_n_bytes(&a->range)) == 0);
987 }
988 
989 static struct sw_flow_mask *flow_mask_find(const struct flow_table *tbl,
990 					   const struct sw_flow_mask *mask)
991 {
992 	struct mask_array *ma;
993 	int i;
994 
995 	ma = ovsl_dereference(tbl->mask_array);
996 	for (i = 0; i < ma->max; i++) {
997 		struct sw_flow_mask *t;
998 		t = ovsl_dereference(ma->masks[i]);
999 
1000 		if (t && mask_equal(mask, t))
1001 			return t;
1002 	}
1003 
1004 	return NULL;
1005 }
1006 
1007 /* Add 'mask' into the mask list, if it is not already there. */
1008 static int flow_mask_insert(struct flow_table *tbl, struct sw_flow *flow,
1009 			    const struct sw_flow_mask *new)
1010 {
1011 	struct sw_flow_mask *mask;
1012 
1013 	mask = flow_mask_find(tbl, new);
1014 	if (!mask) {
1015 		/* Allocate a new mask if none exsits. */
1016 		mask = mask_alloc();
1017 		if (!mask)
1018 			return -ENOMEM;
1019 		mask->key = new->key;
1020 		mask->range = new->range;
1021 
1022 		/* Add mask to mask-list. */
1023 		if (tbl_mask_array_add_mask(tbl, mask)) {
1024 			kfree(mask);
1025 			return -ENOMEM;
1026 		}
1027 	} else {
1028 		BUG_ON(!mask->ref_count);
1029 		mask->ref_count++;
1030 	}
1031 
1032 	flow->mask = mask;
1033 	return 0;
1034 }
1035 
1036 /* Must be called with OVS mutex held. */
1037 static void flow_key_insert(struct flow_table *table, struct sw_flow *flow)
1038 {
1039 	struct table_instance *new_ti = NULL;
1040 	struct table_instance *ti;
1041 
1042 	flow->flow_table.hash = flow_hash(&flow->key, &flow->mask->range);
1043 	ti = ovsl_dereference(table->ti);
1044 	table_instance_insert(ti, flow);
1045 	table->count++;
1046 
1047 	/* Expand table, if necessary, to make room. */
1048 	if (table->count > ti->n_buckets)
1049 		new_ti = table_instance_expand(ti, false);
1050 	else if (time_after(jiffies, table->last_rehash + REHASH_INTERVAL))
1051 		new_ti = table_instance_rehash(ti, ti->n_buckets, false);
1052 
1053 	if (new_ti) {
1054 		rcu_assign_pointer(table->ti, new_ti);
1055 		call_rcu(&ti->rcu, flow_tbl_destroy_rcu_cb);
1056 		table->last_rehash = jiffies;
1057 	}
1058 }
1059 
1060 /* Must be called with OVS mutex held. */
1061 static void flow_ufid_insert(struct flow_table *table, struct sw_flow *flow)
1062 {
1063 	struct table_instance *ti;
1064 
1065 	flow->ufid_table.hash = ufid_hash(&flow->id);
1066 	ti = ovsl_dereference(table->ufid_ti);
1067 	ufid_table_instance_insert(ti, flow);
1068 	table->ufid_count++;
1069 
1070 	/* Expand table, if necessary, to make room. */
1071 	if (table->ufid_count > ti->n_buckets) {
1072 		struct table_instance *new_ti;
1073 
1074 		new_ti = table_instance_expand(ti, true);
1075 		if (new_ti) {
1076 			rcu_assign_pointer(table->ufid_ti, new_ti);
1077 			call_rcu(&ti->rcu, flow_tbl_destroy_rcu_cb);
1078 		}
1079 	}
1080 }
1081 
1082 /* Must be called with OVS mutex held. */
1083 int ovs_flow_tbl_insert(struct flow_table *table, struct sw_flow *flow,
1084 			const struct sw_flow_mask *mask)
1085 {
1086 	int err;
1087 
1088 	err = flow_mask_insert(table, flow, mask);
1089 	if (err)
1090 		return err;
1091 	flow_key_insert(table, flow);
1092 	if (ovs_identifier_is_ufid(&flow->id))
1093 		flow_ufid_insert(table, flow);
1094 
1095 	return 0;
1096 }
1097 
1098 static int compare_mask_and_count(const void *a, const void *b)
1099 {
1100 	const struct mask_count *mc_a = a;
1101 	const struct mask_count *mc_b = b;
1102 
1103 	return (s64)mc_b->counter - (s64)mc_a->counter;
1104 }
1105 
1106 /* Must be called with OVS mutex held. */
1107 void ovs_flow_masks_rebalance(struct flow_table *table)
1108 {
1109 	struct mask_array *ma = rcu_dereference_ovsl(table->mask_array);
1110 	struct mask_count *masks_and_count;
1111 	struct mask_array *new;
1112 	int masks_entries = 0;
1113 	int i;
1114 
1115 	/* Build array of all current entries with use counters. */
1116 	masks_and_count = kmalloc_array(ma->max, sizeof(*masks_and_count),
1117 					GFP_KERNEL);
1118 	if (!masks_and_count)
1119 		return;
1120 
1121 	for (i = 0; i < ma->max; i++) {
1122 		struct sw_flow_mask *mask;
1123 		int cpu;
1124 
1125 		mask = rcu_dereference_ovsl(ma->masks[i]);
1126 		if (unlikely(!mask))
1127 			break;
1128 
1129 		masks_and_count[i].index = i;
1130 		masks_and_count[i].counter = 0;
1131 
1132 		for_each_possible_cpu(cpu) {
1133 			struct mask_array_stats *stats;
1134 			unsigned int start;
1135 			u64 counter;
1136 
1137 			stats = per_cpu_ptr(ma->masks_usage_stats, cpu);
1138 			do {
1139 				start = u64_stats_fetch_begin_irq(&stats->syncp);
1140 				counter = stats->usage_cntrs[i];
1141 			} while (u64_stats_fetch_retry_irq(&stats->syncp,
1142 							   start));
1143 
1144 			masks_and_count[i].counter += counter;
1145 		}
1146 
1147 		/* Subtract the zero count value. */
1148 		masks_and_count[i].counter -= ma->masks_usage_zero_cntr[i];
1149 
1150 		/* Rather than calling tbl_mask_array_reset_counters()
1151 		 * below when no change is needed, do it inline here.
1152 		 */
1153 		ma->masks_usage_zero_cntr[i] += masks_and_count[i].counter;
1154 	}
1155 
1156 	if (i == 0)
1157 		goto free_mask_entries;
1158 
1159 	/* Sort the entries */
1160 	masks_entries = i;
1161 	sort(masks_and_count, masks_entries, sizeof(*masks_and_count),
1162 	     compare_mask_and_count, NULL);
1163 
1164 	/* If the order is the same, nothing to do... */
1165 	for (i = 0; i < masks_entries; i++) {
1166 		if (i != masks_and_count[i].index)
1167 			break;
1168 	}
1169 	if (i == masks_entries)
1170 		goto free_mask_entries;
1171 
1172 	/* Rebuilt the new list in order of usage. */
1173 	new = tbl_mask_array_alloc(ma->max);
1174 	if (!new)
1175 		goto free_mask_entries;
1176 
1177 	for (i = 0; i < masks_entries; i++) {
1178 		int index = masks_and_count[i].index;
1179 
1180 		if (ovsl_dereference(ma->masks[index]))
1181 			new->masks[new->count++] = ma->masks[index];
1182 	}
1183 
1184 	rcu_assign_pointer(table->mask_array, new);
1185 	call_rcu(&ma->rcu, mask_array_rcu_cb);
1186 
1187 free_mask_entries:
1188 	kfree(masks_and_count);
1189 }
1190 
1191 /* Initializes the flow module.
1192  * Returns zero if successful or a negative error code. */
1193 int ovs_flow_init(void)
1194 {
1195 	BUILD_BUG_ON(__alignof__(struct sw_flow_key) % __alignof__(long));
1196 	BUILD_BUG_ON(sizeof(struct sw_flow_key) % sizeof(long));
1197 
1198 	flow_cache = kmem_cache_create("sw_flow", sizeof(struct sw_flow)
1199 				       + (nr_cpu_ids
1200 					  * sizeof(struct sw_flow_stats *)),
1201 				       0, 0, NULL);
1202 	if (flow_cache == NULL)
1203 		return -ENOMEM;
1204 
1205 	flow_stats_cache
1206 		= kmem_cache_create("sw_flow_stats", sizeof(struct sw_flow_stats),
1207 				    0, SLAB_HWCACHE_ALIGN, NULL);
1208 	if (flow_stats_cache == NULL) {
1209 		kmem_cache_destroy(flow_cache);
1210 		flow_cache = NULL;
1211 		return -ENOMEM;
1212 	}
1213 
1214 	return 0;
1215 }
1216 
1217 /* Uninitializes the flow module. */
1218 void ovs_flow_exit(void)
1219 {
1220 	kmem_cache_destroy(flow_stats_cache);
1221 	kmem_cache_destroy(flow_cache);
1222 }
1223