Loading...
1/* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
2 * Copyright (c) 2016 Facebook
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of version 2 of the GNU General Public
6 * License as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful, but
9 * WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
12 */
13#include <linux/bpf.h>
14#include <linux/jhash.h>
15#include <linux/filter.h>
16#include "percpu_freelist.h"
17#include "bpf_lru_list.h"
18
19struct bucket {
20 struct hlist_head head;
21 raw_spinlock_t lock;
22};
23
24struct bpf_htab {
25 struct bpf_map map;
26 struct bucket *buckets;
27 void *elems;
28 union {
29 struct pcpu_freelist freelist;
30 struct bpf_lru lru;
31 };
32 void __percpu *extra_elems;
33 atomic_t count; /* number of elements in this hashtable */
34 u32 n_buckets; /* number of hash buckets */
35 u32 elem_size; /* size of each element in bytes */
36};
37
38enum extra_elem_state {
39 HTAB_NOT_AN_EXTRA_ELEM = 0,
40 HTAB_EXTRA_ELEM_FREE,
41 HTAB_EXTRA_ELEM_USED
42};
43
44/* each htab element is struct htab_elem + key + value */
45struct htab_elem {
46 union {
47 struct hlist_node hash_node;
48 struct bpf_htab *htab;
49 struct pcpu_freelist_node fnode;
50 };
51 union {
52 struct rcu_head rcu;
53 enum extra_elem_state state;
54 struct bpf_lru_node lru_node;
55 };
56 u32 hash;
57 char key[0] __aligned(8);
58};
59
60static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
61
62static bool htab_is_lru(const struct bpf_htab *htab)
63{
64 return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
65 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
66}
67
68static bool htab_is_percpu(const struct bpf_htab *htab)
69{
70 return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
71 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
72}
73
74static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
75 void __percpu *pptr)
76{
77 *(void __percpu **)(l->key + key_size) = pptr;
78}
79
80static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
81{
82 return *(void __percpu **)(l->key + key_size);
83}
84
85static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
86{
87 return (struct htab_elem *) (htab->elems + i * htab->elem_size);
88}
89
90static void htab_free_elems(struct bpf_htab *htab)
91{
92 int i;
93
94 if (!htab_is_percpu(htab))
95 goto free_elems;
96
97 for (i = 0; i < htab->map.max_entries; i++) {
98 void __percpu *pptr;
99
100 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
101 htab->map.key_size);
102 free_percpu(pptr);
103 }
104free_elems:
105 bpf_map_area_free(htab->elems);
106}
107
108static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
109 u32 hash)
110{
111 struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
112 struct htab_elem *l;
113
114 if (node) {
115 l = container_of(node, struct htab_elem, lru_node);
116 memcpy(l->key, key, htab->map.key_size);
117 return l;
118 }
119
120 return NULL;
121}
122
123static int prealloc_init(struct bpf_htab *htab)
124{
125 int err = -ENOMEM, i;
126
127 htab->elems = bpf_map_area_alloc(htab->elem_size *
128 htab->map.max_entries);
129 if (!htab->elems)
130 return -ENOMEM;
131
132 if (!htab_is_percpu(htab))
133 goto skip_percpu_elems;
134
135 for (i = 0; i < htab->map.max_entries; i++) {
136 u32 size = round_up(htab->map.value_size, 8);
137 void __percpu *pptr;
138
139 pptr = __alloc_percpu_gfp(size, 8, GFP_USER | __GFP_NOWARN);
140 if (!pptr)
141 goto free_elems;
142 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
143 pptr);
144 }
145
146skip_percpu_elems:
147 if (htab_is_lru(htab))
148 err = bpf_lru_init(&htab->lru,
149 htab->map.map_flags & BPF_F_NO_COMMON_LRU,
150 offsetof(struct htab_elem, hash) -
151 offsetof(struct htab_elem, lru_node),
152 htab_lru_map_delete_node,
153 htab);
154 else
155 err = pcpu_freelist_init(&htab->freelist);
156
157 if (err)
158 goto free_elems;
159
160 if (htab_is_lru(htab))
161 bpf_lru_populate(&htab->lru, htab->elems,
162 offsetof(struct htab_elem, lru_node),
163 htab->elem_size, htab->map.max_entries);
164 else
165 pcpu_freelist_populate(&htab->freelist, htab->elems,
166 htab->elem_size, htab->map.max_entries);
167
168 return 0;
169
170free_elems:
171 htab_free_elems(htab);
172 return err;
173}
174
175static void prealloc_destroy(struct bpf_htab *htab)
176{
177 htab_free_elems(htab);
178
179 if (htab_is_lru(htab))
180 bpf_lru_destroy(&htab->lru);
181 else
182 pcpu_freelist_destroy(&htab->freelist);
183}
184
185static int alloc_extra_elems(struct bpf_htab *htab)
186{
187 void __percpu *pptr;
188 int cpu;
189
190 pptr = __alloc_percpu_gfp(htab->elem_size, 8, GFP_USER | __GFP_NOWARN);
191 if (!pptr)
192 return -ENOMEM;
193
194 for_each_possible_cpu(cpu) {
195 ((struct htab_elem *)per_cpu_ptr(pptr, cpu))->state =
196 HTAB_EXTRA_ELEM_FREE;
197 }
198 htab->extra_elems = pptr;
199 return 0;
200}
201
202/* Called from syscall */
203static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
204{
205 bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
206 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
207 bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
208 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
209 /* percpu_lru means each cpu has its own LRU list.
210 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
211 * the map's value itself is percpu. percpu_lru has
212 * nothing to do with the map's value.
213 */
214 bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
215 bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
216 struct bpf_htab *htab;
217 int err, i;
218 u64 cost;
219
220 if (lru && !capable(CAP_SYS_ADMIN))
221 /* LRU implementation is much complicated than other
222 * maps. Hence, limit to CAP_SYS_ADMIN for now.
223 */
224 return ERR_PTR(-EPERM);
225
226 if (attr->map_flags & ~(BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU))
227 /* reserved bits should not be used */
228 return ERR_PTR(-EINVAL);
229
230 if (!lru && percpu_lru)
231 return ERR_PTR(-EINVAL);
232
233 if (lru && !prealloc)
234 return ERR_PTR(-ENOTSUPP);
235
236 htab = kzalloc(sizeof(*htab), GFP_USER);
237 if (!htab)
238 return ERR_PTR(-ENOMEM);
239
240 /* mandatory map attributes */
241 htab->map.map_type = attr->map_type;
242 htab->map.key_size = attr->key_size;
243 htab->map.value_size = attr->value_size;
244 htab->map.max_entries = attr->max_entries;
245 htab->map.map_flags = attr->map_flags;
246
247 /* check sanity of attributes.
248 * value_size == 0 may be allowed in the future to use map as a set
249 */
250 err = -EINVAL;
251 if (htab->map.max_entries == 0 || htab->map.key_size == 0 ||
252 htab->map.value_size == 0)
253 goto free_htab;
254
255 if (percpu_lru) {
256 /* ensure each CPU's lru list has >=1 elements.
257 * since we are at it, make each lru list has the same
258 * number of elements.
259 */
260 htab->map.max_entries = roundup(attr->max_entries,
261 num_possible_cpus());
262 if (htab->map.max_entries < attr->max_entries)
263 htab->map.max_entries = rounddown(attr->max_entries,
264 num_possible_cpus());
265 }
266
267 /* hash table size must be power of 2 */
268 htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
269
270 err = -E2BIG;
271 if (htab->map.key_size > MAX_BPF_STACK)
272 /* eBPF programs initialize keys on stack, so they cannot be
273 * larger than max stack size
274 */
275 goto free_htab;
276
277 if (htab->map.value_size >= KMALLOC_MAX_SIZE -
278 MAX_BPF_STACK - sizeof(struct htab_elem))
279 /* if value_size is bigger, the user space won't be able to
280 * access the elements via bpf syscall. This check also makes
281 * sure that the elem_size doesn't overflow and it's
282 * kmalloc-able later in htab_map_update_elem()
283 */
284 goto free_htab;
285
286 if (percpu && round_up(htab->map.value_size, 8) > PCPU_MIN_UNIT_SIZE)
287 /* make sure the size for pcpu_alloc() is reasonable */
288 goto free_htab;
289
290 htab->elem_size = sizeof(struct htab_elem) +
291 round_up(htab->map.key_size, 8);
292 if (percpu)
293 htab->elem_size += sizeof(void *);
294 else
295 htab->elem_size += round_up(htab->map.value_size, 8);
296
297 /* prevent zero size kmalloc and check for u32 overflow */
298 if (htab->n_buckets == 0 ||
299 htab->n_buckets > U32_MAX / sizeof(struct bucket))
300 goto free_htab;
301
302 cost = (u64) htab->n_buckets * sizeof(struct bucket) +
303 (u64) htab->elem_size * htab->map.max_entries;
304
305 if (percpu)
306 cost += (u64) round_up(htab->map.value_size, 8) *
307 num_possible_cpus() * htab->map.max_entries;
308 else
309 cost += (u64) htab->elem_size * num_possible_cpus();
310
311 if (cost >= U32_MAX - PAGE_SIZE)
312 /* make sure page count doesn't overflow */
313 goto free_htab;
314
315 htab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
316
317 /* if map size is larger than memlock limit, reject it early */
318 err = bpf_map_precharge_memlock(htab->map.pages);
319 if (err)
320 goto free_htab;
321
322 err = -ENOMEM;
323 htab->buckets = bpf_map_area_alloc(htab->n_buckets *
324 sizeof(struct bucket));
325 if (!htab->buckets)
326 goto free_htab;
327
328 for (i = 0; i < htab->n_buckets; i++) {
329 INIT_HLIST_HEAD(&htab->buckets[i].head);
330 raw_spin_lock_init(&htab->buckets[i].lock);
331 }
332
333 if (!percpu && !lru) {
334 /* lru itself can remove the least used element, so
335 * there is no need for an extra elem during map_update.
336 */
337 err = alloc_extra_elems(htab);
338 if (err)
339 goto free_buckets;
340 }
341
342 if (prealloc) {
343 err = prealloc_init(htab);
344 if (err)
345 goto free_extra_elems;
346 }
347
348 return &htab->map;
349
350free_extra_elems:
351 free_percpu(htab->extra_elems);
352free_buckets:
353 bpf_map_area_free(htab->buckets);
354free_htab:
355 kfree(htab);
356 return ERR_PTR(err);
357}
358
359static inline u32 htab_map_hash(const void *key, u32 key_len)
360{
361 return jhash(key, key_len, 0);
362}
363
364static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
365{
366 return &htab->buckets[hash & (htab->n_buckets - 1)];
367}
368
369static inline struct hlist_head *select_bucket(struct bpf_htab *htab, u32 hash)
370{
371 return &__select_bucket(htab, hash)->head;
372}
373
374static struct htab_elem *lookup_elem_raw(struct hlist_head *head, u32 hash,
375 void *key, u32 key_size)
376{
377 struct htab_elem *l;
378
379 hlist_for_each_entry_rcu(l, head, hash_node)
380 if (l->hash == hash && !memcmp(&l->key, key, key_size))
381 return l;
382
383 return NULL;
384}
385
386/* Called from syscall or from eBPF program */
387static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
388{
389 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
390 struct hlist_head *head;
391 struct htab_elem *l;
392 u32 hash, key_size;
393
394 /* Must be called with rcu_read_lock. */
395 WARN_ON_ONCE(!rcu_read_lock_held());
396
397 key_size = map->key_size;
398
399 hash = htab_map_hash(key, key_size);
400
401 head = select_bucket(htab, hash);
402
403 l = lookup_elem_raw(head, hash, key, key_size);
404
405 return l;
406}
407
408static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
409{
410 struct htab_elem *l = __htab_map_lookup_elem(map, key);
411
412 if (l)
413 return l->key + round_up(map->key_size, 8);
414
415 return NULL;
416}
417
418static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
419{
420 struct htab_elem *l = __htab_map_lookup_elem(map, key);
421
422 if (l) {
423 bpf_lru_node_set_ref(&l->lru_node);
424 return l->key + round_up(map->key_size, 8);
425 }
426
427 return NULL;
428}
429
430/* It is called from the bpf_lru_list when the LRU needs to delete
431 * older elements from the htab.
432 */
433static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
434{
435 struct bpf_htab *htab = (struct bpf_htab *)arg;
436 struct htab_elem *l, *tgt_l;
437 struct hlist_head *head;
438 unsigned long flags;
439 struct bucket *b;
440
441 tgt_l = container_of(node, struct htab_elem, lru_node);
442 b = __select_bucket(htab, tgt_l->hash);
443 head = &b->head;
444
445 raw_spin_lock_irqsave(&b->lock, flags);
446
447 hlist_for_each_entry_rcu(l, head, hash_node)
448 if (l == tgt_l) {
449 hlist_del_rcu(&l->hash_node);
450 break;
451 }
452
453 raw_spin_unlock_irqrestore(&b->lock, flags);
454
455 return l == tgt_l;
456}
457
458/* Called from syscall */
459static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
460{
461 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
462 struct hlist_head *head;
463 struct htab_elem *l, *next_l;
464 u32 hash, key_size;
465 int i;
466
467 WARN_ON_ONCE(!rcu_read_lock_held());
468
469 key_size = map->key_size;
470
471 hash = htab_map_hash(key, key_size);
472
473 head = select_bucket(htab, hash);
474
475 /* lookup the key */
476 l = lookup_elem_raw(head, hash, key, key_size);
477
478 if (!l) {
479 i = 0;
480 goto find_first_elem;
481 }
482
483 /* key was found, get next key in the same bucket */
484 next_l = hlist_entry_safe(rcu_dereference_raw(hlist_next_rcu(&l->hash_node)),
485 struct htab_elem, hash_node);
486
487 if (next_l) {
488 /* if next elem in this hash list is non-zero, just return it */
489 memcpy(next_key, next_l->key, key_size);
490 return 0;
491 }
492
493 /* no more elements in this hash list, go to the next bucket */
494 i = hash & (htab->n_buckets - 1);
495 i++;
496
497find_first_elem:
498 /* iterate over buckets */
499 for (; i < htab->n_buckets; i++) {
500 head = select_bucket(htab, i);
501
502 /* pick first element in the bucket */
503 next_l = hlist_entry_safe(rcu_dereference_raw(hlist_first_rcu(head)),
504 struct htab_elem, hash_node);
505 if (next_l) {
506 /* if it's not empty, just return it */
507 memcpy(next_key, next_l->key, key_size);
508 return 0;
509 }
510 }
511
512 /* iterated over all buckets and all elements */
513 return -ENOENT;
514}
515
516static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
517{
518 if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
519 free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
520 kfree(l);
521}
522
523static void htab_elem_free_rcu(struct rcu_head *head)
524{
525 struct htab_elem *l = container_of(head, struct htab_elem, rcu);
526 struct bpf_htab *htab = l->htab;
527
528 /* must increment bpf_prog_active to avoid kprobe+bpf triggering while
529 * we're calling kfree, otherwise deadlock is possible if kprobes
530 * are placed somewhere inside of slub
531 */
532 preempt_disable();
533 __this_cpu_inc(bpf_prog_active);
534 htab_elem_free(htab, l);
535 __this_cpu_dec(bpf_prog_active);
536 preempt_enable();
537}
538
539static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
540{
541 if (l->state == HTAB_EXTRA_ELEM_USED) {
542 l->state = HTAB_EXTRA_ELEM_FREE;
543 return;
544 }
545
546 if (!(htab->map.map_flags & BPF_F_NO_PREALLOC)) {
547 pcpu_freelist_push(&htab->freelist, &l->fnode);
548 } else {
549 atomic_dec(&htab->count);
550 l->htab = htab;
551 call_rcu(&l->rcu, htab_elem_free_rcu);
552 }
553}
554
555static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
556 void *value, bool onallcpus)
557{
558 if (!onallcpus) {
559 /* copy true value_size bytes */
560 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
561 } else {
562 u32 size = round_up(htab->map.value_size, 8);
563 int off = 0, cpu;
564
565 for_each_possible_cpu(cpu) {
566 bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
567 value + off, size);
568 off += size;
569 }
570 }
571}
572
573static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
574 void *value, u32 key_size, u32 hash,
575 bool percpu, bool onallcpus,
576 bool old_elem_exists)
577{
578 u32 size = htab->map.value_size;
579 bool prealloc = !(htab->map.map_flags & BPF_F_NO_PREALLOC);
580 struct htab_elem *l_new;
581 void __percpu *pptr;
582 int err = 0;
583
584 if (prealloc) {
585 l_new = (struct htab_elem *)pcpu_freelist_pop(&htab->freelist);
586 if (!l_new)
587 err = -E2BIG;
588 } else {
589 if (atomic_inc_return(&htab->count) > htab->map.max_entries) {
590 atomic_dec(&htab->count);
591 err = -E2BIG;
592 } else {
593 l_new = kmalloc(htab->elem_size,
594 GFP_ATOMIC | __GFP_NOWARN);
595 if (!l_new)
596 return ERR_PTR(-ENOMEM);
597 }
598 }
599
600 if (err) {
601 if (!old_elem_exists)
602 return ERR_PTR(err);
603
604 /* if we're updating the existing element and the hash table
605 * is full, use per-cpu extra elems
606 */
607 l_new = this_cpu_ptr(htab->extra_elems);
608 if (l_new->state != HTAB_EXTRA_ELEM_FREE)
609 return ERR_PTR(-E2BIG);
610 l_new->state = HTAB_EXTRA_ELEM_USED;
611 } else {
612 l_new->state = HTAB_NOT_AN_EXTRA_ELEM;
613 }
614
615 memcpy(l_new->key, key, key_size);
616 if (percpu) {
617 /* round up value_size to 8 bytes */
618 size = round_up(size, 8);
619
620 if (prealloc) {
621 pptr = htab_elem_get_ptr(l_new, key_size);
622 } else {
623 /* alloc_percpu zero-fills */
624 pptr = __alloc_percpu_gfp(size, 8,
625 GFP_ATOMIC | __GFP_NOWARN);
626 if (!pptr) {
627 kfree(l_new);
628 return ERR_PTR(-ENOMEM);
629 }
630 }
631
632 pcpu_copy_value(htab, pptr, value, onallcpus);
633
634 if (!prealloc)
635 htab_elem_set_ptr(l_new, key_size, pptr);
636 } else {
637 memcpy(l_new->key + round_up(key_size, 8), value, size);
638 }
639
640 l_new->hash = hash;
641 return l_new;
642}
643
644static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
645 u64 map_flags)
646{
647 if (l_old && map_flags == BPF_NOEXIST)
648 /* elem already exists */
649 return -EEXIST;
650
651 if (!l_old && map_flags == BPF_EXIST)
652 /* elem doesn't exist, cannot update it */
653 return -ENOENT;
654
655 return 0;
656}
657
658/* Called from syscall or from eBPF program */
659static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
660 u64 map_flags)
661{
662 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
663 struct htab_elem *l_new = NULL, *l_old;
664 struct hlist_head *head;
665 unsigned long flags;
666 struct bucket *b;
667 u32 key_size, hash;
668 int ret;
669
670 if (unlikely(map_flags > BPF_EXIST))
671 /* unknown flags */
672 return -EINVAL;
673
674 WARN_ON_ONCE(!rcu_read_lock_held());
675
676 key_size = map->key_size;
677
678 hash = htab_map_hash(key, key_size);
679
680 b = __select_bucket(htab, hash);
681 head = &b->head;
682
683 /* bpf_map_update_elem() can be called in_irq() */
684 raw_spin_lock_irqsave(&b->lock, flags);
685
686 l_old = lookup_elem_raw(head, hash, key, key_size);
687
688 ret = check_flags(htab, l_old, map_flags);
689 if (ret)
690 goto err;
691
692 l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
693 !!l_old);
694 if (IS_ERR(l_new)) {
695 /* all pre-allocated elements are in use or memory exhausted */
696 ret = PTR_ERR(l_new);
697 goto err;
698 }
699
700 /* add new element to the head of the list, so that
701 * concurrent search will find it before old elem
702 */
703 hlist_add_head_rcu(&l_new->hash_node, head);
704 if (l_old) {
705 hlist_del_rcu(&l_old->hash_node);
706 free_htab_elem(htab, l_old);
707 }
708 ret = 0;
709err:
710 raw_spin_unlock_irqrestore(&b->lock, flags);
711 return ret;
712}
713
714static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
715 u64 map_flags)
716{
717 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
718 struct htab_elem *l_new, *l_old = NULL;
719 struct hlist_head *head;
720 unsigned long flags;
721 struct bucket *b;
722 u32 key_size, hash;
723 int ret;
724
725 if (unlikely(map_flags > BPF_EXIST))
726 /* unknown flags */
727 return -EINVAL;
728
729 WARN_ON_ONCE(!rcu_read_lock_held());
730
731 key_size = map->key_size;
732
733 hash = htab_map_hash(key, key_size);
734
735 b = __select_bucket(htab, hash);
736 head = &b->head;
737
738 /* For LRU, we need to alloc before taking bucket's
739 * spinlock because getting free nodes from LRU may need
740 * to remove older elements from htab and this removal
741 * operation will need a bucket lock.
742 */
743 l_new = prealloc_lru_pop(htab, key, hash);
744 if (!l_new)
745 return -ENOMEM;
746 memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
747
748 /* bpf_map_update_elem() can be called in_irq() */
749 raw_spin_lock_irqsave(&b->lock, flags);
750
751 l_old = lookup_elem_raw(head, hash, key, key_size);
752
753 ret = check_flags(htab, l_old, map_flags);
754 if (ret)
755 goto err;
756
757 /* add new element to the head of the list, so that
758 * concurrent search will find it before old elem
759 */
760 hlist_add_head_rcu(&l_new->hash_node, head);
761 if (l_old) {
762 bpf_lru_node_set_ref(&l_new->lru_node);
763 hlist_del_rcu(&l_old->hash_node);
764 }
765 ret = 0;
766
767err:
768 raw_spin_unlock_irqrestore(&b->lock, flags);
769
770 if (ret)
771 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
772 else if (l_old)
773 bpf_lru_push_free(&htab->lru, &l_old->lru_node);
774
775 return ret;
776}
777
778static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
779 void *value, u64 map_flags,
780 bool onallcpus)
781{
782 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
783 struct htab_elem *l_new = NULL, *l_old;
784 struct hlist_head *head;
785 unsigned long flags;
786 struct bucket *b;
787 u32 key_size, hash;
788 int ret;
789
790 if (unlikely(map_flags > BPF_EXIST))
791 /* unknown flags */
792 return -EINVAL;
793
794 WARN_ON_ONCE(!rcu_read_lock_held());
795
796 key_size = map->key_size;
797
798 hash = htab_map_hash(key, key_size);
799
800 b = __select_bucket(htab, hash);
801 head = &b->head;
802
803 /* bpf_map_update_elem() can be called in_irq() */
804 raw_spin_lock_irqsave(&b->lock, flags);
805
806 l_old = lookup_elem_raw(head, hash, key, key_size);
807
808 ret = check_flags(htab, l_old, map_flags);
809 if (ret)
810 goto err;
811
812 if (l_old) {
813 /* per-cpu hash map can update value in-place */
814 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
815 value, onallcpus);
816 } else {
817 l_new = alloc_htab_elem(htab, key, value, key_size,
818 hash, true, onallcpus, false);
819 if (IS_ERR(l_new)) {
820 ret = PTR_ERR(l_new);
821 goto err;
822 }
823 hlist_add_head_rcu(&l_new->hash_node, head);
824 }
825 ret = 0;
826err:
827 raw_spin_unlock_irqrestore(&b->lock, flags);
828 return ret;
829}
830
831static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
832 void *value, u64 map_flags,
833 bool onallcpus)
834{
835 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
836 struct htab_elem *l_new = NULL, *l_old;
837 struct hlist_head *head;
838 unsigned long flags;
839 struct bucket *b;
840 u32 key_size, hash;
841 int ret;
842
843 if (unlikely(map_flags > BPF_EXIST))
844 /* unknown flags */
845 return -EINVAL;
846
847 WARN_ON_ONCE(!rcu_read_lock_held());
848
849 key_size = map->key_size;
850
851 hash = htab_map_hash(key, key_size);
852
853 b = __select_bucket(htab, hash);
854 head = &b->head;
855
856 /* For LRU, we need to alloc before taking bucket's
857 * spinlock because LRU's elem alloc may need
858 * to remove older elem from htab and this removal
859 * operation will need a bucket lock.
860 */
861 if (map_flags != BPF_EXIST) {
862 l_new = prealloc_lru_pop(htab, key, hash);
863 if (!l_new)
864 return -ENOMEM;
865 }
866
867 /* bpf_map_update_elem() can be called in_irq() */
868 raw_spin_lock_irqsave(&b->lock, flags);
869
870 l_old = lookup_elem_raw(head, hash, key, key_size);
871
872 ret = check_flags(htab, l_old, map_flags);
873 if (ret)
874 goto err;
875
876 if (l_old) {
877 bpf_lru_node_set_ref(&l_old->lru_node);
878
879 /* per-cpu hash map can update value in-place */
880 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
881 value, onallcpus);
882 } else {
883 pcpu_copy_value(htab, htab_elem_get_ptr(l_new, key_size),
884 value, onallcpus);
885 hlist_add_head_rcu(&l_new->hash_node, head);
886 l_new = NULL;
887 }
888 ret = 0;
889err:
890 raw_spin_unlock_irqrestore(&b->lock, flags);
891 if (l_new)
892 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
893 return ret;
894}
895
896static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
897 void *value, u64 map_flags)
898{
899 return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
900}
901
902static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
903 void *value, u64 map_flags)
904{
905 return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
906 false);
907}
908
909/* Called from syscall or from eBPF program */
910static int htab_map_delete_elem(struct bpf_map *map, void *key)
911{
912 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
913 struct hlist_head *head;
914 struct bucket *b;
915 struct htab_elem *l;
916 unsigned long flags;
917 u32 hash, key_size;
918 int ret = -ENOENT;
919
920 WARN_ON_ONCE(!rcu_read_lock_held());
921
922 key_size = map->key_size;
923
924 hash = htab_map_hash(key, key_size);
925 b = __select_bucket(htab, hash);
926 head = &b->head;
927
928 raw_spin_lock_irqsave(&b->lock, flags);
929
930 l = lookup_elem_raw(head, hash, key, key_size);
931
932 if (l) {
933 hlist_del_rcu(&l->hash_node);
934 free_htab_elem(htab, l);
935 ret = 0;
936 }
937
938 raw_spin_unlock_irqrestore(&b->lock, flags);
939 return ret;
940}
941
942static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
943{
944 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
945 struct hlist_head *head;
946 struct bucket *b;
947 struct htab_elem *l;
948 unsigned long flags;
949 u32 hash, key_size;
950 int ret = -ENOENT;
951
952 WARN_ON_ONCE(!rcu_read_lock_held());
953
954 key_size = map->key_size;
955
956 hash = htab_map_hash(key, key_size);
957 b = __select_bucket(htab, hash);
958 head = &b->head;
959
960 raw_spin_lock_irqsave(&b->lock, flags);
961
962 l = lookup_elem_raw(head, hash, key, key_size);
963
964 if (l) {
965 hlist_del_rcu(&l->hash_node);
966 ret = 0;
967 }
968
969 raw_spin_unlock_irqrestore(&b->lock, flags);
970 if (l)
971 bpf_lru_push_free(&htab->lru, &l->lru_node);
972 return ret;
973}
974
975static void delete_all_elements(struct bpf_htab *htab)
976{
977 int i;
978
979 for (i = 0; i < htab->n_buckets; i++) {
980 struct hlist_head *head = select_bucket(htab, i);
981 struct hlist_node *n;
982 struct htab_elem *l;
983
984 hlist_for_each_entry_safe(l, n, head, hash_node) {
985 hlist_del_rcu(&l->hash_node);
986 if (l->state != HTAB_EXTRA_ELEM_USED)
987 htab_elem_free(htab, l);
988 }
989 }
990}
991/* Called when map->refcnt goes to zero, either from workqueue or from syscall */
992static void htab_map_free(struct bpf_map *map)
993{
994 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
995
996 /* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
997 * so the programs (can be more than one that used this map) were
998 * disconnected from events. Wait for outstanding critical sections in
999 * these programs to complete
1000 */
1001 synchronize_rcu();
1002
1003 /* some of free_htab_elem() callbacks for elements of this map may
1004 * not have executed. Wait for them.
1005 */
1006 rcu_barrier();
1007 if (htab->map.map_flags & BPF_F_NO_PREALLOC)
1008 delete_all_elements(htab);
1009 else
1010 prealloc_destroy(htab);
1011
1012 free_percpu(htab->extra_elems);
1013 bpf_map_area_free(htab->buckets);
1014 kfree(htab);
1015}
1016
1017static const struct bpf_map_ops htab_ops = {
1018 .map_alloc = htab_map_alloc,
1019 .map_free = htab_map_free,
1020 .map_get_next_key = htab_map_get_next_key,
1021 .map_lookup_elem = htab_map_lookup_elem,
1022 .map_update_elem = htab_map_update_elem,
1023 .map_delete_elem = htab_map_delete_elem,
1024};
1025
1026static struct bpf_map_type_list htab_type __read_mostly = {
1027 .ops = &htab_ops,
1028 .type = BPF_MAP_TYPE_HASH,
1029};
1030
1031static const struct bpf_map_ops htab_lru_ops = {
1032 .map_alloc = htab_map_alloc,
1033 .map_free = htab_map_free,
1034 .map_get_next_key = htab_map_get_next_key,
1035 .map_lookup_elem = htab_lru_map_lookup_elem,
1036 .map_update_elem = htab_lru_map_update_elem,
1037 .map_delete_elem = htab_lru_map_delete_elem,
1038};
1039
1040static struct bpf_map_type_list htab_lru_type __read_mostly = {
1041 .ops = &htab_lru_ops,
1042 .type = BPF_MAP_TYPE_LRU_HASH,
1043};
1044
1045/* Called from eBPF program */
1046static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1047{
1048 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1049
1050 if (l)
1051 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1052 else
1053 return NULL;
1054}
1055
1056static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1057{
1058 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1059
1060 if (l) {
1061 bpf_lru_node_set_ref(&l->lru_node);
1062 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1063 }
1064
1065 return NULL;
1066}
1067
1068int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1069{
1070 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1071 struct htab_elem *l;
1072 void __percpu *pptr;
1073 int ret = -ENOENT;
1074 int cpu, off = 0;
1075 u32 size;
1076
1077 /* per_cpu areas are zero-filled and bpf programs can only
1078 * access 'value_size' of them, so copying rounded areas
1079 * will not leak any kernel data
1080 */
1081 size = round_up(map->value_size, 8);
1082 rcu_read_lock();
1083 l = __htab_map_lookup_elem(map, key);
1084 if (!l)
1085 goto out;
1086 if (htab_is_lru(htab))
1087 bpf_lru_node_set_ref(&l->lru_node);
1088 pptr = htab_elem_get_ptr(l, map->key_size);
1089 for_each_possible_cpu(cpu) {
1090 bpf_long_memcpy(value + off,
1091 per_cpu_ptr(pptr, cpu), size);
1092 off += size;
1093 }
1094 ret = 0;
1095out:
1096 rcu_read_unlock();
1097 return ret;
1098}
1099
1100int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1101 u64 map_flags)
1102{
1103 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1104 int ret;
1105
1106 rcu_read_lock();
1107 if (htab_is_lru(htab))
1108 ret = __htab_lru_percpu_map_update_elem(map, key, value,
1109 map_flags, true);
1110 else
1111 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1112 true);
1113 rcu_read_unlock();
1114
1115 return ret;
1116}
1117
1118static const struct bpf_map_ops htab_percpu_ops = {
1119 .map_alloc = htab_map_alloc,
1120 .map_free = htab_map_free,
1121 .map_get_next_key = htab_map_get_next_key,
1122 .map_lookup_elem = htab_percpu_map_lookup_elem,
1123 .map_update_elem = htab_percpu_map_update_elem,
1124 .map_delete_elem = htab_map_delete_elem,
1125};
1126
1127static struct bpf_map_type_list htab_percpu_type __read_mostly = {
1128 .ops = &htab_percpu_ops,
1129 .type = BPF_MAP_TYPE_PERCPU_HASH,
1130};
1131
1132static const struct bpf_map_ops htab_lru_percpu_ops = {
1133 .map_alloc = htab_map_alloc,
1134 .map_free = htab_map_free,
1135 .map_get_next_key = htab_map_get_next_key,
1136 .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
1137 .map_update_elem = htab_lru_percpu_map_update_elem,
1138 .map_delete_elem = htab_lru_map_delete_elem,
1139};
1140
1141static struct bpf_map_type_list htab_lru_percpu_type __read_mostly = {
1142 .ops = &htab_lru_percpu_ops,
1143 .type = BPF_MAP_TYPE_LRU_PERCPU_HASH,
1144};
1145
1146static int __init register_htab_map(void)
1147{
1148 bpf_register_map_type(&htab_type);
1149 bpf_register_map_type(&htab_percpu_type);
1150 bpf_register_map_type(&htab_lru_type);
1151 bpf_register_map_type(&htab_lru_percpu_type);
1152 return 0;
1153}
1154late_initcall(register_htab_map);
1// SPDX-License-Identifier: GPL-2.0-only
2/* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
3 * Copyright (c) 2016 Facebook
4 */
5#include <linux/bpf.h>
6#include <linux/btf.h>
7#include <linux/jhash.h>
8#include <linux/filter.h>
9#include <linux/rculist_nulls.h>
10#include <linux/rcupdate_wait.h>
11#include <linux/random.h>
12#include <uapi/linux/btf.h>
13#include <linux/rcupdate_trace.h>
14#include <linux/btf_ids.h>
15#include "percpu_freelist.h"
16#include "bpf_lru_list.h"
17#include "map_in_map.h"
18#include <linux/bpf_mem_alloc.h>
19
20#define HTAB_CREATE_FLAG_MASK \
21 (BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU | BPF_F_NUMA_NODE | \
22 BPF_F_ACCESS_MASK | BPF_F_ZERO_SEED)
23
24#define BATCH_OPS(_name) \
25 .map_lookup_batch = \
26 _name##_map_lookup_batch, \
27 .map_lookup_and_delete_batch = \
28 _name##_map_lookup_and_delete_batch, \
29 .map_update_batch = \
30 generic_map_update_batch, \
31 .map_delete_batch = \
32 generic_map_delete_batch
33
34/*
35 * The bucket lock has two protection scopes:
36 *
37 * 1) Serializing concurrent operations from BPF programs on different
38 * CPUs
39 *
40 * 2) Serializing concurrent operations from BPF programs and sys_bpf()
41 *
42 * BPF programs can execute in any context including perf, kprobes and
43 * tracing. As there are almost no limits where perf, kprobes and tracing
44 * can be invoked from the lock operations need to be protected against
45 * deadlocks. Deadlocks can be caused by recursion and by an invocation in
46 * the lock held section when functions which acquire this lock are invoked
47 * from sys_bpf(). BPF recursion is prevented by incrementing the per CPU
48 * variable bpf_prog_active, which prevents BPF programs attached to perf
49 * events, kprobes and tracing to be invoked before the prior invocation
50 * from one of these contexts completed. sys_bpf() uses the same mechanism
51 * by pinning the task to the current CPU and incrementing the recursion
52 * protection across the map operation.
53 *
54 * This has subtle implications on PREEMPT_RT. PREEMPT_RT forbids certain
55 * operations like memory allocations (even with GFP_ATOMIC) from atomic
56 * contexts. This is required because even with GFP_ATOMIC the memory
57 * allocator calls into code paths which acquire locks with long held lock
58 * sections. To ensure the deterministic behaviour these locks are regular
59 * spinlocks, which are converted to 'sleepable' spinlocks on RT. The only
60 * true atomic contexts on an RT kernel are the low level hardware
61 * handling, scheduling, low level interrupt handling, NMIs etc. None of
62 * these contexts should ever do memory allocations.
63 *
64 * As regular device interrupt handlers and soft interrupts are forced into
65 * thread context, the existing code which does
66 * spin_lock*(); alloc(GFP_ATOMIC); spin_unlock*();
67 * just works.
68 *
69 * In theory the BPF locks could be converted to regular spinlocks as well,
70 * but the bucket locks and percpu_freelist locks can be taken from
71 * arbitrary contexts (perf, kprobes, tracepoints) which are required to be
72 * atomic contexts even on RT. Before the introduction of bpf_mem_alloc,
73 * it is only safe to use raw spinlock for preallocated hash map on a RT kernel,
74 * because there is no memory allocation within the lock held sections. However
75 * after hash map was fully converted to use bpf_mem_alloc, there will be
76 * non-synchronous memory allocation for non-preallocated hash map, so it is
77 * safe to always use raw spinlock for bucket lock.
78 */
79struct bucket {
80 struct hlist_nulls_head head;
81 raw_spinlock_t raw_lock;
82};
83
84#define HASHTAB_MAP_LOCK_COUNT 8
85#define HASHTAB_MAP_LOCK_MASK (HASHTAB_MAP_LOCK_COUNT - 1)
86
87struct bpf_htab {
88 struct bpf_map map;
89 struct bpf_mem_alloc ma;
90 struct bpf_mem_alloc pcpu_ma;
91 struct bucket *buckets;
92 void *elems;
93 union {
94 struct pcpu_freelist freelist;
95 struct bpf_lru lru;
96 };
97 struct htab_elem *__percpu *extra_elems;
98 /* number of elements in non-preallocated hashtable are kept
99 * in either pcount or count
100 */
101 struct percpu_counter pcount;
102 atomic_t count;
103 bool use_percpu_counter;
104 u32 n_buckets; /* number of hash buckets */
105 u32 elem_size; /* size of each element in bytes */
106 u32 hashrnd;
107 struct lock_class_key lockdep_key;
108 int __percpu *map_locked[HASHTAB_MAP_LOCK_COUNT];
109};
110
111/* each htab element is struct htab_elem + key + value */
112struct htab_elem {
113 union {
114 struct hlist_nulls_node hash_node;
115 struct {
116 void *padding;
117 union {
118 struct pcpu_freelist_node fnode;
119 struct htab_elem *batch_flink;
120 };
121 };
122 };
123 union {
124 /* pointer to per-cpu pointer */
125 void *ptr_to_pptr;
126 struct bpf_lru_node lru_node;
127 };
128 u32 hash;
129 char key[] __aligned(8);
130};
131
132static inline bool htab_is_prealloc(const struct bpf_htab *htab)
133{
134 return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
135}
136
137static void htab_init_buckets(struct bpf_htab *htab)
138{
139 unsigned int i;
140
141 for (i = 0; i < htab->n_buckets; i++) {
142 INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
143 raw_spin_lock_init(&htab->buckets[i].raw_lock);
144 lockdep_set_class(&htab->buckets[i].raw_lock,
145 &htab->lockdep_key);
146 cond_resched();
147 }
148}
149
150static inline int htab_lock_bucket(const struct bpf_htab *htab,
151 struct bucket *b, u32 hash,
152 unsigned long *pflags)
153{
154 unsigned long flags;
155
156 hash = hash & min_t(u32, HASHTAB_MAP_LOCK_MASK, htab->n_buckets - 1);
157
158 preempt_disable();
159 local_irq_save(flags);
160 if (unlikely(__this_cpu_inc_return(*(htab->map_locked[hash])) != 1)) {
161 __this_cpu_dec(*(htab->map_locked[hash]));
162 local_irq_restore(flags);
163 preempt_enable();
164 return -EBUSY;
165 }
166
167 raw_spin_lock(&b->raw_lock);
168 *pflags = flags;
169
170 return 0;
171}
172
173static inline void htab_unlock_bucket(const struct bpf_htab *htab,
174 struct bucket *b, u32 hash,
175 unsigned long flags)
176{
177 hash = hash & min_t(u32, HASHTAB_MAP_LOCK_MASK, htab->n_buckets - 1);
178 raw_spin_unlock(&b->raw_lock);
179 __this_cpu_dec(*(htab->map_locked[hash]));
180 local_irq_restore(flags);
181 preempt_enable();
182}
183
184static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
185
186static bool htab_is_lru(const struct bpf_htab *htab)
187{
188 return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
189 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
190}
191
192static bool htab_is_percpu(const struct bpf_htab *htab)
193{
194 return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
195 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
196}
197
198static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
199 void __percpu *pptr)
200{
201 *(void __percpu **)(l->key + key_size) = pptr;
202}
203
204static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
205{
206 return *(void __percpu **)(l->key + key_size);
207}
208
209static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
210{
211 return *(void **)(l->key + roundup(map->key_size, 8));
212}
213
214static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
215{
216 return (struct htab_elem *) (htab->elems + i * (u64)htab->elem_size);
217}
218
219static bool htab_has_extra_elems(struct bpf_htab *htab)
220{
221 return !htab_is_percpu(htab) && !htab_is_lru(htab);
222}
223
224static void htab_free_prealloced_timers(struct bpf_htab *htab)
225{
226 u32 num_entries = htab->map.max_entries;
227 int i;
228
229 if (!btf_record_has_field(htab->map.record, BPF_TIMER))
230 return;
231 if (htab_has_extra_elems(htab))
232 num_entries += num_possible_cpus();
233
234 for (i = 0; i < num_entries; i++) {
235 struct htab_elem *elem;
236
237 elem = get_htab_elem(htab, i);
238 bpf_obj_free_timer(htab->map.record, elem->key + round_up(htab->map.key_size, 8));
239 cond_resched();
240 }
241}
242
243static void htab_free_prealloced_fields(struct bpf_htab *htab)
244{
245 u32 num_entries = htab->map.max_entries;
246 int i;
247
248 if (IS_ERR_OR_NULL(htab->map.record))
249 return;
250 if (htab_has_extra_elems(htab))
251 num_entries += num_possible_cpus();
252 for (i = 0; i < num_entries; i++) {
253 struct htab_elem *elem;
254
255 elem = get_htab_elem(htab, i);
256 if (htab_is_percpu(htab)) {
257 void __percpu *pptr = htab_elem_get_ptr(elem, htab->map.key_size);
258 int cpu;
259
260 for_each_possible_cpu(cpu) {
261 bpf_obj_free_fields(htab->map.record, per_cpu_ptr(pptr, cpu));
262 cond_resched();
263 }
264 } else {
265 bpf_obj_free_fields(htab->map.record, elem->key + round_up(htab->map.key_size, 8));
266 cond_resched();
267 }
268 cond_resched();
269 }
270}
271
272static void htab_free_elems(struct bpf_htab *htab)
273{
274 int i;
275
276 if (!htab_is_percpu(htab))
277 goto free_elems;
278
279 for (i = 0; i < htab->map.max_entries; i++) {
280 void __percpu *pptr;
281
282 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
283 htab->map.key_size);
284 free_percpu(pptr);
285 cond_resched();
286 }
287free_elems:
288 bpf_map_area_free(htab->elems);
289}
290
291/* The LRU list has a lock (lru_lock). Each htab bucket has a lock
292 * (bucket_lock). If both locks need to be acquired together, the lock
293 * order is always lru_lock -> bucket_lock and this only happens in
294 * bpf_lru_list.c logic. For example, certain code path of
295 * bpf_lru_pop_free(), which is called by function prealloc_lru_pop(),
296 * will acquire lru_lock first followed by acquiring bucket_lock.
297 *
298 * In hashtab.c, to avoid deadlock, lock acquisition of
299 * bucket_lock followed by lru_lock is not allowed. In such cases,
300 * bucket_lock needs to be released first before acquiring lru_lock.
301 */
302static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
303 u32 hash)
304{
305 struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
306 struct htab_elem *l;
307
308 if (node) {
309 bpf_map_inc_elem_count(&htab->map);
310 l = container_of(node, struct htab_elem, lru_node);
311 memcpy(l->key, key, htab->map.key_size);
312 return l;
313 }
314
315 return NULL;
316}
317
318static int prealloc_init(struct bpf_htab *htab)
319{
320 u32 num_entries = htab->map.max_entries;
321 int err = -ENOMEM, i;
322
323 if (htab_has_extra_elems(htab))
324 num_entries += num_possible_cpus();
325
326 htab->elems = bpf_map_area_alloc((u64)htab->elem_size * num_entries,
327 htab->map.numa_node);
328 if (!htab->elems)
329 return -ENOMEM;
330
331 if (!htab_is_percpu(htab))
332 goto skip_percpu_elems;
333
334 for (i = 0; i < num_entries; i++) {
335 u32 size = round_up(htab->map.value_size, 8);
336 void __percpu *pptr;
337
338 pptr = bpf_map_alloc_percpu(&htab->map, size, 8,
339 GFP_USER | __GFP_NOWARN);
340 if (!pptr)
341 goto free_elems;
342 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
343 pptr);
344 cond_resched();
345 }
346
347skip_percpu_elems:
348 if (htab_is_lru(htab))
349 err = bpf_lru_init(&htab->lru,
350 htab->map.map_flags & BPF_F_NO_COMMON_LRU,
351 offsetof(struct htab_elem, hash) -
352 offsetof(struct htab_elem, lru_node),
353 htab_lru_map_delete_node,
354 htab);
355 else
356 err = pcpu_freelist_init(&htab->freelist);
357
358 if (err)
359 goto free_elems;
360
361 if (htab_is_lru(htab))
362 bpf_lru_populate(&htab->lru, htab->elems,
363 offsetof(struct htab_elem, lru_node),
364 htab->elem_size, num_entries);
365 else
366 pcpu_freelist_populate(&htab->freelist,
367 htab->elems + offsetof(struct htab_elem, fnode),
368 htab->elem_size, num_entries);
369
370 return 0;
371
372free_elems:
373 htab_free_elems(htab);
374 return err;
375}
376
377static void prealloc_destroy(struct bpf_htab *htab)
378{
379 htab_free_elems(htab);
380
381 if (htab_is_lru(htab))
382 bpf_lru_destroy(&htab->lru);
383 else
384 pcpu_freelist_destroy(&htab->freelist);
385}
386
387static int alloc_extra_elems(struct bpf_htab *htab)
388{
389 struct htab_elem *__percpu *pptr, *l_new;
390 struct pcpu_freelist_node *l;
391 int cpu;
392
393 pptr = bpf_map_alloc_percpu(&htab->map, sizeof(struct htab_elem *), 8,
394 GFP_USER | __GFP_NOWARN);
395 if (!pptr)
396 return -ENOMEM;
397
398 for_each_possible_cpu(cpu) {
399 l = pcpu_freelist_pop(&htab->freelist);
400 /* pop will succeed, since prealloc_init()
401 * preallocated extra num_possible_cpus elements
402 */
403 l_new = container_of(l, struct htab_elem, fnode);
404 *per_cpu_ptr(pptr, cpu) = l_new;
405 }
406 htab->extra_elems = pptr;
407 return 0;
408}
409
410/* Called from syscall */
411static int htab_map_alloc_check(union bpf_attr *attr)
412{
413 bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
414 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
415 bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
416 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
417 /* percpu_lru means each cpu has its own LRU list.
418 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
419 * the map's value itself is percpu. percpu_lru has
420 * nothing to do with the map's value.
421 */
422 bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
423 bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
424 bool zero_seed = (attr->map_flags & BPF_F_ZERO_SEED);
425 int numa_node = bpf_map_attr_numa_node(attr);
426
427 BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
428 offsetof(struct htab_elem, hash_node.pprev));
429
430 if (zero_seed && !capable(CAP_SYS_ADMIN))
431 /* Guard against local DoS, and discourage production use. */
432 return -EPERM;
433
434 if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK ||
435 !bpf_map_flags_access_ok(attr->map_flags))
436 return -EINVAL;
437
438 if (!lru && percpu_lru)
439 return -EINVAL;
440
441 if (lru && !prealloc)
442 return -ENOTSUPP;
443
444 if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
445 return -EINVAL;
446
447 /* check sanity of attributes.
448 * value_size == 0 may be allowed in the future to use map as a set
449 */
450 if (attr->max_entries == 0 || attr->key_size == 0 ||
451 attr->value_size == 0)
452 return -EINVAL;
453
454 if ((u64)attr->key_size + attr->value_size >= KMALLOC_MAX_SIZE -
455 sizeof(struct htab_elem))
456 /* if key_size + value_size is bigger, the user space won't be
457 * able to access the elements via bpf syscall. This check
458 * also makes sure that the elem_size doesn't overflow and it's
459 * kmalloc-able later in htab_map_update_elem()
460 */
461 return -E2BIG;
462
463 return 0;
464}
465
466static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
467{
468 bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
469 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
470 bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
471 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
472 /* percpu_lru means each cpu has its own LRU list.
473 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
474 * the map's value itself is percpu. percpu_lru has
475 * nothing to do with the map's value.
476 */
477 bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
478 bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
479 struct bpf_htab *htab;
480 int err, i;
481
482 htab = bpf_map_area_alloc(sizeof(*htab), NUMA_NO_NODE);
483 if (!htab)
484 return ERR_PTR(-ENOMEM);
485
486 lockdep_register_key(&htab->lockdep_key);
487
488 bpf_map_init_from_attr(&htab->map, attr);
489
490 if (percpu_lru) {
491 /* ensure each CPU's lru list has >=1 elements.
492 * since we are at it, make each lru list has the same
493 * number of elements.
494 */
495 htab->map.max_entries = roundup(attr->max_entries,
496 num_possible_cpus());
497 if (htab->map.max_entries < attr->max_entries)
498 htab->map.max_entries = rounddown(attr->max_entries,
499 num_possible_cpus());
500 }
501
502 /* hash table size must be power of 2; roundup_pow_of_two() can overflow
503 * into UB on 32-bit arches, so check that first
504 */
505 err = -E2BIG;
506 if (htab->map.max_entries > 1UL << 31)
507 goto free_htab;
508
509 htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
510
511 htab->elem_size = sizeof(struct htab_elem) +
512 round_up(htab->map.key_size, 8);
513 if (percpu)
514 htab->elem_size += sizeof(void *);
515 else
516 htab->elem_size += round_up(htab->map.value_size, 8);
517
518 /* check for u32 overflow */
519 if (htab->n_buckets > U32_MAX / sizeof(struct bucket))
520 goto free_htab;
521
522 err = bpf_map_init_elem_count(&htab->map);
523 if (err)
524 goto free_htab;
525
526 err = -ENOMEM;
527 htab->buckets = bpf_map_area_alloc(htab->n_buckets *
528 sizeof(struct bucket),
529 htab->map.numa_node);
530 if (!htab->buckets)
531 goto free_elem_count;
532
533 for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++) {
534 htab->map_locked[i] = bpf_map_alloc_percpu(&htab->map,
535 sizeof(int),
536 sizeof(int),
537 GFP_USER);
538 if (!htab->map_locked[i])
539 goto free_map_locked;
540 }
541
542 if (htab->map.map_flags & BPF_F_ZERO_SEED)
543 htab->hashrnd = 0;
544 else
545 htab->hashrnd = get_random_u32();
546
547 htab_init_buckets(htab);
548
549/* compute_batch_value() computes batch value as num_online_cpus() * 2
550 * and __percpu_counter_compare() needs
551 * htab->max_entries - cur_number_of_elems to be more than batch * num_online_cpus()
552 * for percpu_counter to be faster than atomic_t. In practice the average bpf
553 * hash map size is 10k, which means that a system with 64 cpus will fill
554 * hashmap to 20% of 10k before percpu_counter becomes ineffective. Therefore
555 * define our own batch count as 32 then 10k hash map can be filled up to 80%:
556 * 10k - 8k > 32 _batch_ * 64 _cpus_
557 * and __percpu_counter_compare() will still be fast. At that point hash map
558 * collisions will dominate its performance anyway. Assume that hash map filled
559 * to 50+% isn't going to be O(1) and use the following formula to choose
560 * between percpu_counter and atomic_t.
561 */
562#define PERCPU_COUNTER_BATCH 32
563 if (attr->max_entries / 2 > num_online_cpus() * PERCPU_COUNTER_BATCH)
564 htab->use_percpu_counter = true;
565
566 if (htab->use_percpu_counter) {
567 err = percpu_counter_init(&htab->pcount, 0, GFP_KERNEL);
568 if (err)
569 goto free_map_locked;
570 }
571
572 if (prealloc) {
573 err = prealloc_init(htab);
574 if (err)
575 goto free_map_locked;
576
577 if (!percpu && !lru) {
578 /* lru itself can remove the least used element, so
579 * there is no need for an extra elem during map_update.
580 */
581 err = alloc_extra_elems(htab);
582 if (err)
583 goto free_prealloc;
584 }
585 } else {
586 err = bpf_mem_alloc_init(&htab->ma, htab->elem_size, false);
587 if (err)
588 goto free_map_locked;
589 if (percpu) {
590 err = bpf_mem_alloc_init(&htab->pcpu_ma,
591 round_up(htab->map.value_size, 8), true);
592 if (err)
593 goto free_map_locked;
594 }
595 }
596
597 return &htab->map;
598
599free_prealloc:
600 prealloc_destroy(htab);
601free_map_locked:
602 if (htab->use_percpu_counter)
603 percpu_counter_destroy(&htab->pcount);
604 for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
605 free_percpu(htab->map_locked[i]);
606 bpf_map_area_free(htab->buckets);
607 bpf_mem_alloc_destroy(&htab->pcpu_ma);
608 bpf_mem_alloc_destroy(&htab->ma);
609free_elem_count:
610 bpf_map_free_elem_count(&htab->map);
611free_htab:
612 lockdep_unregister_key(&htab->lockdep_key);
613 bpf_map_area_free(htab);
614 return ERR_PTR(err);
615}
616
617static inline u32 htab_map_hash(const void *key, u32 key_len, u32 hashrnd)
618{
619 if (likely(key_len % 4 == 0))
620 return jhash2(key, key_len / 4, hashrnd);
621 return jhash(key, key_len, hashrnd);
622}
623
624static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
625{
626 return &htab->buckets[hash & (htab->n_buckets - 1)];
627}
628
629static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
630{
631 return &__select_bucket(htab, hash)->head;
632}
633
634/* this lookup function can only be called with bucket lock taken */
635static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
636 void *key, u32 key_size)
637{
638 struct hlist_nulls_node *n;
639 struct htab_elem *l;
640
641 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
642 if (l->hash == hash && !memcmp(&l->key, key, key_size))
643 return l;
644
645 return NULL;
646}
647
648/* can be called without bucket lock. it will repeat the loop in
649 * the unlikely event when elements moved from one bucket into another
650 * while link list is being walked
651 */
652static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
653 u32 hash, void *key,
654 u32 key_size, u32 n_buckets)
655{
656 struct hlist_nulls_node *n;
657 struct htab_elem *l;
658
659again:
660 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
661 if (l->hash == hash && !memcmp(&l->key, key, key_size))
662 return l;
663
664 if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
665 goto again;
666
667 return NULL;
668}
669
670/* Called from syscall or from eBPF program directly, so
671 * arguments have to match bpf_map_lookup_elem() exactly.
672 * The return value is adjusted by BPF instructions
673 * in htab_map_gen_lookup().
674 */
675static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
676{
677 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
678 struct hlist_nulls_head *head;
679 struct htab_elem *l;
680 u32 hash, key_size;
681
682 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
683 !rcu_read_lock_bh_held());
684
685 key_size = map->key_size;
686
687 hash = htab_map_hash(key, key_size, htab->hashrnd);
688
689 head = select_bucket(htab, hash);
690
691 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
692
693 return l;
694}
695
696static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
697{
698 struct htab_elem *l = __htab_map_lookup_elem(map, key);
699
700 if (l)
701 return l->key + round_up(map->key_size, 8);
702
703 return NULL;
704}
705
706/* inline bpf_map_lookup_elem() call.
707 * Instead of:
708 * bpf_prog
709 * bpf_map_lookup_elem
710 * map->ops->map_lookup_elem
711 * htab_map_lookup_elem
712 * __htab_map_lookup_elem
713 * do:
714 * bpf_prog
715 * __htab_map_lookup_elem
716 */
717static int htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
718{
719 struct bpf_insn *insn = insn_buf;
720 const int ret = BPF_REG_0;
721
722 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
723 (void *(*)(struct bpf_map *map, void *key))NULL));
724 *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
725 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
726 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
727 offsetof(struct htab_elem, key) +
728 round_up(map->key_size, 8));
729 return insn - insn_buf;
730}
731
732static __always_inline void *__htab_lru_map_lookup_elem(struct bpf_map *map,
733 void *key, const bool mark)
734{
735 struct htab_elem *l = __htab_map_lookup_elem(map, key);
736
737 if (l) {
738 if (mark)
739 bpf_lru_node_set_ref(&l->lru_node);
740 return l->key + round_up(map->key_size, 8);
741 }
742
743 return NULL;
744}
745
746static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
747{
748 return __htab_lru_map_lookup_elem(map, key, true);
749}
750
751static void *htab_lru_map_lookup_elem_sys(struct bpf_map *map, void *key)
752{
753 return __htab_lru_map_lookup_elem(map, key, false);
754}
755
756static int htab_lru_map_gen_lookup(struct bpf_map *map,
757 struct bpf_insn *insn_buf)
758{
759 struct bpf_insn *insn = insn_buf;
760 const int ret = BPF_REG_0;
761 const int ref_reg = BPF_REG_1;
762
763 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
764 (void *(*)(struct bpf_map *map, void *key))NULL));
765 *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
766 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
767 *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
768 offsetof(struct htab_elem, lru_node) +
769 offsetof(struct bpf_lru_node, ref));
770 *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
771 *insn++ = BPF_ST_MEM(BPF_B, ret,
772 offsetof(struct htab_elem, lru_node) +
773 offsetof(struct bpf_lru_node, ref),
774 1);
775 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
776 offsetof(struct htab_elem, key) +
777 round_up(map->key_size, 8));
778 return insn - insn_buf;
779}
780
781static void check_and_free_fields(struct bpf_htab *htab,
782 struct htab_elem *elem)
783{
784 if (htab_is_percpu(htab)) {
785 void __percpu *pptr = htab_elem_get_ptr(elem, htab->map.key_size);
786 int cpu;
787
788 for_each_possible_cpu(cpu)
789 bpf_obj_free_fields(htab->map.record, per_cpu_ptr(pptr, cpu));
790 } else {
791 void *map_value = elem->key + round_up(htab->map.key_size, 8);
792
793 bpf_obj_free_fields(htab->map.record, map_value);
794 }
795}
796
797/* It is called from the bpf_lru_list when the LRU needs to delete
798 * older elements from the htab.
799 */
800static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
801{
802 struct bpf_htab *htab = arg;
803 struct htab_elem *l = NULL, *tgt_l;
804 struct hlist_nulls_head *head;
805 struct hlist_nulls_node *n;
806 unsigned long flags;
807 struct bucket *b;
808 int ret;
809
810 tgt_l = container_of(node, struct htab_elem, lru_node);
811 b = __select_bucket(htab, tgt_l->hash);
812 head = &b->head;
813
814 ret = htab_lock_bucket(htab, b, tgt_l->hash, &flags);
815 if (ret)
816 return false;
817
818 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
819 if (l == tgt_l) {
820 hlist_nulls_del_rcu(&l->hash_node);
821 check_and_free_fields(htab, l);
822 bpf_map_dec_elem_count(&htab->map);
823 break;
824 }
825
826 htab_unlock_bucket(htab, b, tgt_l->hash, flags);
827
828 return l == tgt_l;
829}
830
831/* Called from syscall */
832static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
833{
834 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
835 struct hlist_nulls_head *head;
836 struct htab_elem *l, *next_l;
837 u32 hash, key_size;
838 int i = 0;
839
840 WARN_ON_ONCE(!rcu_read_lock_held());
841
842 key_size = map->key_size;
843
844 if (!key)
845 goto find_first_elem;
846
847 hash = htab_map_hash(key, key_size, htab->hashrnd);
848
849 head = select_bucket(htab, hash);
850
851 /* lookup the key */
852 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
853
854 if (!l)
855 goto find_first_elem;
856
857 /* key was found, get next key in the same bucket */
858 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
859 struct htab_elem, hash_node);
860
861 if (next_l) {
862 /* if next elem in this hash list is non-zero, just return it */
863 memcpy(next_key, next_l->key, key_size);
864 return 0;
865 }
866
867 /* no more elements in this hash list, go to the next bucket */
868 i = hash & (htab->n_buckets - 1);
869 i++;
870
871find_first_elem:
872 /* iterate over buckets */
873 for (; i < htab->n_buckets; i++) {
874 head = select_bucket(htab, i);
875
876 /* pick first element in the bucket */
877 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
878 struct htab_elem, hash_node);
879 if (next_l) {
880 /* if it's not empty, just return it */
881 memcpy(next_key, next_l->key, key_size);
882 return 0;
883 }
884 }
885
886 /* iterated over all buckets and all elements */
887 return -ENOENT;
888}
889
890static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
891{
892 check_and_free_fields(htab, l);
893 if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
894 bpf_mem_cache_free(&htab->pcpu_ma, l->ptr_to_pptr);
895 bpf_mem_cache_free(&htab->ma, l);
896}
897
898static void htab_put_fd_value(struct bpf_htab *htab, struct htab_elem *l)
899{
900 struct bpf_map *map = &htab->map;
901 void *ptr;
902
903 if (map->ops->map_fd_put_ptr) {
904 ptr = fd_htab_map_get_ptr(map, l);
905 map->ops->map_fd_put_ptr(map, ptr, true);
906 }
907}
908
909static bool is_map_full(struct bpf_htab *htab)
910{
911 if (htab->use_percpu_counter)
912 return __percpu_counter_compare(&htab->pcount, htab->map.max_entries,
913 PERCPU_COUNTER_BATCH) >= 0;
914 return atomic_read(&htab->count) >= htab->map.max_entries;
915}
916
917static void inc_elem_count(struct bpf_htab *htab)
918{
919 bpf_map_inc_elem_count(&htab->map);
920
921 if (htab->use_percpu_counter)
922 percpu_counter_add_batch(&htab->pcount, 1, PERCPU_COUNTER_BATCH);
923 else
924 atomic_inc(&htab->count);
925}
926
927static void dec_elem_count(struct bpf_htab *htab)
928{
929 bpf_map_dec_elem_count(&htab->map);
930
931 if (htab->use_percpu_counter)
932 percpu_counter_add_batch(&htab->pcount, -1, PERCPU_COUNTER_BATCH);
933 else
934 atomic_dec(&htab->count);
935}
936
937
938static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
939{
940 htab_put_fd_value(htab, l);
941
942 if (htab_is_prealloc(htab)) {
943 bpf_map_dec_elem_count(&htab->map);
944 check_and_free_fields(htab, l);
945 __pcpu_freelist_push(&htab->freelist, &l->fnode);
946 } else {
947 dec_elem_count(htab);
948 htab_elem_free(htab, l);
949 }
950}
951
952static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
953 void *value, bool onallcpus)
954{
955 if (!onallcpus) {
956 /* copy true value_size bytes */
957 copy_map_value(&htab->map, this_cpu_ptr(pptr), value);
958 } else {
959 u32 size = round_up(htab->map.value_size, 8);
960 int off = 0, cpu;
961
962 for_each_possible_cpu(cpu) {
963 copy_map_value_long(&htab->map, per_cpu_ptr(pptr, cpu), value + off);
964 off += size;
965 }
966 }
967}
968
969static void pcpu_init_value(struct bpf_htab *htab, void __percpu *pptr,
970 void *value, bool onallcpus)
971{
972 /* When not setting the initial value on all cpus, zero-fill element
973 * values for other cpus. Otherwise, bpf program has no way to ensure
974 * known initial values for cpus other than current one
975 * (onallcpus=false always when coming from bpf prog).
976 */
977 if (!onallcpus) {
978 int current_cpu = raw_smp_processor_id();
979 int cpu;
980
981 for_each_possible_cpu(cpu) {
982 if (cpu == current_cpu)
983 copy_map_value_long(&htab->map, per_cpu_ptr(pptr, cpu), value);
984 else /* Since elem is preallocated, we cannot touch special fields */
985 zero_map_value(&htab->map, per_cpu_ptr(pptr, cpu));
986 }
987 } else {
988 pcpu_copy_value(htab, pptr, value, onallcpus);
989 }
990}
991
992static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
993{
994 return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
995 BITS_PER_LONG == 64;
996}
997
998static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
999 void *value, u32 key_size, u32 hash,
1000 bool percpu, bool onallcpus,
1001 struct htab_elem *old_elem)
1002{
1003 u32 size = htab->map.value_size;
1004 bool prealloc = htab_is_prealloc(htab);
1005 struct htab_elem *l_new, **pl_new;
1006 void __percpu *pptr;
1007
1008 if (prealloc) {
1009 if (old_elem) {
1010 /* if we're updating the existing element,
1011 * use per-cpu extra elems to avoid freelist_pop/push
1012 */
1013 pl_new = this_cpu_ptr(htab->extra_elems);
1014 l_new = *pl_new;
1015 htab_put_fd_value(htab, old_elem);
1016 *pl_new = old_elem;
1017 } else {
1018 struct pcpu_freelist_node *l;
1019
1020 l = __pcpu_freelist_pop(&htab->freelist);
1021 if (!l)
1022 return ERR_PTR(-E2BIG);
1023 l_new = container_of(l, struct htab_elem, fnode);
1024 bpf_map_inc_elem_count(&htab->map);
1025 }
1026 } else {
1027 if (is_map_full(htab))
1028 if (!old_elem)
1029 /* when map is full and update() is replacing
1030 * old element, it's ok to allocate, since
1031 * old element will be freed immediately.
1032 * Otherwise return an error
1033 */
1034 return ERR_PTR(-E2BIG);
1035 inc_elem_count(htab);
1036 l_new = bpf_mem_cache_alloc(&htab->ma);
1037 if (!l_new) {
1038 l_new = ERR_PTR(-ENOMEM);
1039 goto dec_count;
1040 }
1041 }
1042
1043 memcpy(l_new->key, key, key_size);
1044 if (percpu) {
1045 if (prealloc) {
1046 pptr = htab_elem_get_ptr(l_new, key_size);
1047 } else {
1048 /* alloc_percpu zero-fills */
1049 pptr = bpf_mem_cache_alloc(&htab->pcpu_ma);
1050 if (!pptr) {
1051 bpf_mem_cache_free(&htab->ma, l_new);
1052 l_new = ERR_PTR(-ENOMEM);
1053 goto dec_count;
1054 }
1055 l_new->ptr_to_pptr = pptr;
1056 pptr = *(void **)pptr;
1057 }
1058
1059 pcpu_init_value(htab, pptr, value, onallcpus);
1060
1061 if (!prealloc)
1062 htab_elem_set_ptr(l_new, key_size, pptr);
1063 } else if (fd_htab_map_needs_adjust(htab)) {
1064 size = round_up(size, 8);
1065 memcpy(l_new->key + round_up(key_size, 8), value, size);
1066 } else {
1067 copy_map_value(&htab->map,
1068 l_new->key + round_up(key_size, 8),
1069 value);
1070 }
1071
1072 l_new->hash = hash;
1073 return l_new;
1074dec_count:
1075 dec_elem_count(htab);
1076 return l_new;
1077}
1078
1079static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
1080 u64 map_flags)
1081{
1082 if (l_old && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
1083 /* elem already exists */
1084 return -EEXIST;
1085
1086 if (!l_old && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
1087 /* elem doesn't exist, cannot update it */
1088 return -ENOENT;
1089
1090 return 0;
1091}
1092
1093/* Called from syscall or from eBPF program */
1094static long htab_map_update_elem(struct bpf_map *map, void *key, void *value,
1095 u64 map_flags)
1096{
1097 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1098 struct htab_elem *l_new = NULL, *l_old;
1099 struct hlist_nulls_head *head;
1100 unsigned long flags;
1101 struct bucket *b;
1102 u32 key_size, hash;
1103 int ret;
1104
1105 if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
1106 /* unknown flags */
1107 return -EINVAL;
1108
1109 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1110 !rcu_read_lock_bh_held());
1111
1112 key_size = map->key_size;
1113
1114 hash = htab_map_hash(key, key_size, htab->hashrnd);
1115
1116 b = __select_bucket(htab, hash);
1117 head = &b->head;
1118
1119 if (unlikely(map_flags & BPF_F_LOCK)) {
1120 if (unlikely(!btf_record_has_field(map->record, BPF_SPIN_LOCK)))
1121 return -EINVAL;
1122 /* find an element without taking the bucket lock */
1123 l_old = lookup_nulls_elem_raw(head, hash, key, key_size,
1124 htab->n_buckets);
1125 ret = check_flags(htab, l_old, map_flags);
1126 if (ret)
1127 return ret;
1128 if (l_old) {
1129 /* grab the element lock and update value in place */
1130 copy_map_value_locked(map,
1131 l_old->key + round_up(key_size, 8),
1132 value, false);
1133 return 0;
1134 }
1135 /* fall through, grab the bucket lock and lookup again.
1136 * 99.9% chance that the element won't be found,
1137 * but second lookup under lock has to be done.
1138 */
1139 }
1140
1141 ret = htab_lock_bucket(htab, b, hash, &flags);
1142 if (ret)
1143 return ret;
1144
1145 l_old = lookup_elem_raw(head, hash, key, key_size);
1146
1147 ret = check_flags(htab, l_old, map_flags);
1148 if (ret)
1149 goto err;
1150
1151 if (unlikely(l_old && (map_flags & BPF_F_LOCK))) {
1152 /* first lookup without the bucket lock didn't find the element,
1153 * but second lookup with the bucket lock found it.
1154 * This case is highly unlikely, but has to be dealt with:
1155 * grab the element lock in addition to the bucket lock
1156 * and update element in place
1157 */
1158 copy_map_value_locked(map,
1159 l_old->key + round_up(key_size, 8),
1160 value, false);
1161 ret = 0;
1162 goto err;
1163 }
1164
1165 l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
1166 l_old);
1167 if (IS_ERR(l_new)) {
1168 /* all pre-allocated elements are in use or memory exhausted */
1169 ret = PTR_ERR(l_new);
1170 goto err;
1171 }
1172
1173 /* add new element to the head of the list, so that
1174 * concurrent search will find it before old elem
1175 */
1176 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1177 if (l_old) {
1178 hlist_nulls_del_rcu(&l_old->hash_node);
1179 if (!htab_is_prealloc(htab))
1180 free_htab_elem(htab, l_old);
1181 else
1182 check_and_free_fields(htab, l_old);
1183 }
1184 ret = 0;
1185err:
1186 htab_unlock_bucket(htab, b, hash, flags);
1187 return ret;
1188}
1189
1190static void htab_lru_push_free(struct bpf_htab *htab, struct htab_elem *elem)
1191{
1192 check_and_free_fields(htab, elem);
1193 bpf_map_dec_elem_count(&htab->map);
1194 bpf_lru_push_free(&htab->lru, &elem->lru_node);
1195}
1196
1197static long htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
1198 u64 map_flags)
1199{
1200 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1201 struct htab_elem *l_new, *l_old = NULL;
1202 struct hlist_nulls_head *head;
1203 unsigned long flags;
1204 struct bucket *b;
1205 u32 key_size, hash;
1206 int ret;
1207
1208 if (unlikely(map_flags > BPF_EXIST))
1209 /* unknown flags */
1210 return -EINVAL;
1211
1212 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1213 !rcu_read_lock_bh_held());
1214
1215 key_size = map->key_size;
1216
1217 hash = htab_map_hash(key, key_size, htab->hashrnd);
1218
1219 b = __select_bucket(htab, hash);
1220 head = &b->head;
1221
1222 /* For LRU, we need to alloc before taking bucket's
1223 * spinlock because getting free nodes from LRU may need
1224 * to remove older elements from htab and this removal
1225 * operation will need a bucket lock.
1226 */
1227 l_new = prealloc_lru_pop(htab, key, hash);
1228 if (!l_new)
1229 return -ENOMEM;
1230 copy_map_value(&htab->map,
1231 l_new->key + round_up(map->key_size, 8), value);
1232
1233 ret = htab_lock_bucket(htab, b, hash, &flags);
1234 if (ret)
1235 goto err_lock_bucket;
1236
1237 l_old = lookup_elem_raw(head, hash, key, key_size);
1238
1239 ret = check_flags(htab, l_old, map_flags);
1240 if (ret)
1241 goto err;
1242
1243 /* add new element to the head of the list, so that
1244 * concurrent search will find it before old elem
1245 */
1246 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1247 if (l_old) {
1248 bpf_lru_node_set_ref(&l_new->lru_node);
1249 hlist_nulls_del_rcu(&l_old->hash_node);
1250 }
1251 ret = 0;
1252
1253err:
1254 htab_unlock_bucket(htab, b, hash, flags);
1255
1256err_lock_bucket:
1257 if (ret)
1258 htab_lru_push_free(htab, l_new);
1259 else if (l_old)
1260 htab_lru_push_free(htab, l_old);
1261
1262 return ret;
1263}
1264
1265static long __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1266 void *value, u64 map_flags,
1267 bool onallcpus)
1268{
1269 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1270 struct htab_elem *l_new = NULL, *l_old;
1271 struct hlist_nulls_head *head;
1272 unsigned long flags;
1273 struct bucket *b;
1274 u32 key_size, hash;
1275 int ret;
1276
1277 if (unlikely(map_flags > BPF_EXIST))
1278 /* unknown flags */
1279 return -EINVAL;
1280
1281 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1282 !rcu_read_lock_bh_held());
1283
1284 key_size = map->key_size;
1285
1286 hash = htab_map_hash(key, key_size, htab->hashrnd);
1287
1288 b = __select_bucket(htab, hash);
1289 head = &b->head;
1290
1291 ret = htab_lock_bucket(htab, b, hash, &flags);
1292 if (ret)
1293 return ret;
1294
1295 l_old = lookup_elem_raw(head, hash, key, key_size);
1296
1297 ret = check_flags(htab, l_old, map_flags);
1298 if (ret)
1299 goto err;
1300
1301 if (l_old) {
1302 /* per-cpu hash map can update value in-place */
1303 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1304 value, onallcpus);
1305 } else {
1306 l_new = alloc_htab_elem(htab, key, value, key_size,
1307 hash, true, onallcpus, NULL);
1308 if (IS_ERR(l_new)) {
1309 ret = PTR_ERR(l_new);
1310 goto err;
1311 }
1312 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1313 }
1314 ret = 0;
1315err:
1316 htab_unlock_bucket(htab, b, hash, flags);
1317 return ret;
1318}
1319
1320static long __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1321 void *value, u64 map_flags,
1322 bool onallcpus)
1323{
1324 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1325 struct htab_elem *l_new = NULL, *l_old;
1326 struct hlist_nulls_head *head;
1327 unsigned long flags;
1328 struct bucket *b;
1329 u32 key_size, hash;
1330 int ret;
1331
1332 if (unlikely(map_flags > BPF_EXIST))
1333 /* unknown flags */
1334 return -EINVAL;
1335
1336 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1337 !rcu_read_lock_bh_held());
1338
1339 key_size = map->key_size;
1340
1341 hash = htab_map_hash(key, key_size, htab->hashrnd);
1342
1343 b = __select_bucket(htab, hash);
1344 head = &b->head;
1345
1346 /* For LRU, we need to alloc before taking bucket's
1347 * spinlock because LRU's elem alloc may need
1348 * to remove older elem from htab and this removal
1349 * operation will need a bucket lock.
1350 */
1351 if (map_flags != BPF_EXIST) {
1352 l_new = prealloc_lru_pop(htab, key, hash);
1353 if (!l_new)
1354 return -ENOMEM;
1355 }
1356
1357 ret = htab_lock_bucket(htab, b, hash, &flags);
1358 if (ret)
1359 goto err_lock_bucket;
1360
1361 l_old = lookup_elem_raw(head, hash, key, key_size);
1362
1363 ret = check_flags(htab, l_old, map_flags);
1364 if (ret)
1365 goto err;
1366
1367 if (l_old) {
1368 bpf_lru_node_set_ref(&l_old->lru_node);
1369
1370 /* per-cpu hash map can update value in-place */
1371 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1372 value, onallcpus);
1373 } else {
1374 pcpu_init_value(htab, htab_elem_get_ptr(l_new, key_size),
1375 value, onallcpus);
1376 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1377 l_new = NULL;
1378 }
1379 ret = 0;
1380err:
1381 htab_unlock_bucket(htab, b, hash, flags);
1382err_lock_bucket:
1383 if (l_new) {
1384 bpf_map_dec_elem_count(&htab->map);
1385 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1386 }
1387 return ret;
1388}
1389
1390static long htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1391 void *value, u64 map_flags)
1392{
1393 return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1394}
1395
1396static long htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1397 void *value, u64 map_flags)
1398{
1399 return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1400 false);
1401}
1402
1403/* Called from syscall or from eBPF program */
1404static long htab_map_delete_elem(struct bpf_map *map, void *key)
1405{
1406 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1407 struct hlist_nulls_head *head;
1408 struct bucket *b;
1409 struct htab_elem *l;
1410 unsigned long flags;
1411 u32 hash, key_size;
1412 int ret;
1413
1414 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1415 !rcu_read_lock_bh_held());
1416
1417 key_size = map->key_size;
1418
1419 hash = htab_map_hash(key, key_size, htab->hashrnd);
1420 b = __select_bucket(htab, hash);
1421 head = &b->head;
1422
1423 ret = htab_lock_bucket(htab, b, hash, &flags);
1424 if (ret)
1425 return ret;
1426
1427 l = lookup_elem_raw(head, hash, key, key_size);
1428
1429 if (l) {
1430 hlist_nulls_del_rcu(&l->hash_node);
1431 free_htab_elem(htab, l);
1432 } else {
1433 ret = -ENOENT;
1434 }
1435
1436 htab_unlock_bucket(htab, b, hash, flags);
1437 return ret;
1438}
1439
1440static long htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1441{
1442 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1443 struct hlist_nulls_head *head;
1444 struct bucket *b;
1445 struct htab_elem *l;
1446 unsigned long flags;
1447 u32 hash, key_size;
1448 int ret;
1449
1450 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1451 !rcu_read_lock_bh_held());
1452
1453 key_size = map->key_size;
1454
1455 hash = htab_map_hash(key, key_size, htab->hashrnd);
1456 b = __select_bucket(htab, hash);
1457 head = &b->head;
1458
1459 ret = htab_lock_bucket(htab, b, hash, &flags);
1460 if (ret)
1461 return ret;
1462
1463 l = lookup_elem_raw(head, hash, key, key_size);
1464
1465 if (l)
1466 hlist_nulls_del_rcu(&l->hash_node);
1467 else
1468 ret = -ENOENT;
1469
1470 htab_unlock_bucket(htab, b, hash, flags);
1471 if (l)
1472 htab_lru_push_free(htab, l);
1473 return ret;
1474}
1475
1476static void delete_all_elements(struct bpf_htab *htab)
1477{
1478 int i;
1479
1480 /* It's called from a worker thread, so disable migration here,
1481 * since bpf_mem_cache_free() relies on that.
1482 */
1483 migrate_disable();
1484 for (i = 0; i < htab->n_buckets; i++) {
1485 struct hlist_nulls_head *head = select_bucket(htab, i);
1486 struct hlist_nulls_node *n;
1487 struct htab_elem *l;
1488
1489 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1490 hlist_nulls_del_rcu(&l->hash_node);
1491 htab_elem_free(htab, l);
1492 }
1493 }
1494 migrate_enable();
1495}
1496
1497static void htab_free_malloced_timers(struct bpf_htab *htab)
1498{
1499 int i;
1500
1501 rcu_read_lock();
1502 for (i = 0; i < htab->n_buckets; i++) {
1503 struct hlist_nulls_head *head = select_bucket(htab, i);
1504 struct hlist_nulls_node *n;
1505 struct htab_elem *l;
1506
1507 hlist_nulls_for_each_entry(l, n, head, hash_node) {
1508 /* We only free timer on uref dropping to zero */
1509 bpf_obj_free_timer(htab->map.record, l->key + round_up(htab->map.key_size, 8));
1510 }
1511 cond_resched_rcu();
1512 }
1513 rcu_read_unlock();
1514}
1515
1516static void htab_map_free_timers(struct bpf_map *map)
1517{
1518 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1519
1520 /* We only free timer on uref dropping to zero */
1521 if (!btf_record_has_field(htab->map.record, BPF_TIMER))
1522 return;
1523 if (!htab_is_prealloc(htab))
1524 htab_free_malloced_timers(htab);
1525 else
1526 htab_free_prealloced_timers(htab);
1527}
1528
1529/* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1530static void htab_map_free(struct bpf_map *map)
1531{
1532 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1533 int i;
1534
1535 /* bpf_free_used_maps() or close(map_fd) will trigger this map_free callback.
1536 * bpf_free_used_maps() is called after bpf prog is no longer executing.
1537 * There is no need to synchronize_rcu() here to protect map elements.
1538 */
1539
1540 /* htab no longer uses call_rcu() directly. bpf_mem_alloc does it
1541 * underneath and is reponsible for waiting for callbacks to finish
1542 * during bpf_mem_alloc_destroy().
1543 */
1544 if (!htab_is_prealloc(htab)) {
1545 delete_all_elements(htab);
1546 } else {
1547 htab_free_prealloced_fields(htab);
1548 prealloc_destroy(htab);
1549 }
1550
1551 bpf_map_free_elem_count(map);
1552 free_percpu(htab->extra_elems);
1553 bpf_map_area_free(htab->buckets);
1554 bpf_mem_alloc_destroy(&htab->pcpu_ma);
1555 bpf_mem_alloc_destroy(&htab->ma);
1556 if (htab->use_percpu_counter)
1557 percpu_counter_destroy(&htab->pcount);
1558 for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
1559 free_percpu(htab->map_locked[i]);
1560 lockdep_unregister_key(&htab->lockdep_key);
1561 bpf_map_area_free(htab);
1562}
1563
1564static void htab_map_seq_show_elem(struct bpf_map *map, void *key,
1565 struct seq_file *m)
1566{
1567 void *value;
1568
1569 rcu_read_lock();
1570
1571 value = htab_map_lookup_elem(map, key);
1572 if (!value) {
1573 rcu_read_unlock();
1574 return;
1575 }
1576
1577 btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1578 seq_puts(m, ": ");
1579 btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
1580 seq_puts(m, "\n");
1581
1582 rcu_read_unlock();
1583}
1584
1585static int __htab_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1586 void *value, bool is_lru_map,
1587 bool is_percpu, u64 flags)
1588{
1589 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1590 struct hlist_nulls_head *head;
1591 unsigned long bflags;
1592 struct htab_elem *l;
1593 u32 hash, key_size;
1594 struct bucket *b;
1595 int ret;
1596
1597 key_size = map->key_size;
1598
1599 hash = htab_map_hash(key, key_size, htab->hashrnd);
1600 b = __select_bucket(htab, hash);
1601 head = &b->head;
1602
1603 ret = htab_lock_bucket(htab, b, hash, &bflags);
1604 if (ret)
1605 return ret;
1606
1607 l = lookup_elem_raw(head, hash, key, key_size);
1608 if (!l) {
1609 ret = -ENOENT;
1610 } else {
1611 if (is_percpu) {
1612 u32 roundup_value_size = round_up(map->value_size, 8);
1613 void __percpu *pptr;
1614 int off = 0, cpu;
1615
1616 pptr = htab_elem_get_ptr(l, key_size);
1617 for_each_possible_cpu(cpu) {
1618 copy_map_value_long(&htab->map, value + off, per_cpu_ptr(pptr, cpu));
1619 check_and_init_map_value(&htab->map, value + off);
1620 off += roundup_value_size;
1621 }
1622 } else {
1623 u32 roundup_key_size = round_up(map->key_size, 8);
1624
1625 if (flags & BPF_F_LOCK)
1626 copy_map_value_locked(map, value, l->key +
1627 roundup_key_size,
1628 true);
1629 else
1630 copy_map_value(map, value, l->key +
1631 roundup_key_size);
1632 /* Zeroing special fields in the temp buffer */
1633 check_and_init_map_value(map, value);
1634 }
1635
1636 hlist_nulls_del_rcu(&l->hash_node);
1637 if (!is_lru_map)
1638 free_htab_elem(htab, l);
1639 }
1640
1641 htab_unlock_bucket(htab, b, hash, bflags);
1642
1643 if (is_lru_map && l)
1644 htab_lru_push_free(htab, l);
1645
1646 return ret;
1647}
1648
1649static int htab_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1650 void *value, u64 flags)
1651{
1652 return __htab_map_lookup_and_delete_elem(map, key, value, false, false,
1653 flags);
1654}
1655
1656static int htab_percpu_map_lookup_and_delete_elem(struct bpf_map *map,
1657 void *key, void *value,
1658 u64 flags)
1659{
1660 return __htab_map_lookup_and_delete_elem(map, key, value, false, true,
1661 flags);
1662}
1663
1664static int htab_lru_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1665 void *value, u64 flags)
1666{
1667 return __htab_map_lookup_and_delete_elem(map, key, value, true, false,
1668 flags);
1669}
1670
1671static int htab_lru_percpu_map_lookup_and_delete_elem(struct bpf_map *map,
1672 void *key, void *value,
1673 u64 flags)
1674{
1675 return __htab_map_lookup_and_delete_elem(map, key, value, true, true,
1676 flags);
1677}
1678
1679static int
1680__htab_map_lookup_and_delete_batch(struct bpf_map *map,
1681 const union bpf_attr *attr,
1682 union bpf_attr __user *uattr,
1683 bool do_delete, bool is_lru_map,
1684 bool is_percpu)
1685{
1686 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1687 u32 bucket_cnt, total, key_size, value_size, roundup_key_size;
1688 void *keys = NULL, *values = NULL, *value, *dst_key, *dst_val;
1689 void __user *uvalues = u64_to_user_ptr(attr->batch.values);
1690 void __user *ukeys = u64_to_user_ptr(attr->batch.keys);
1691 void __user *ubatch = u64_to_user_ptr(attr->batch.in_batch);
1692 u32 batch, max_count, size, bucket_size, map_id;
1693 struct htab_elem *node_to_free = NULL;
1694 u64 elem_map_flags, map_flags;
1695 struct hlist_nulls_head *head;
1696 struct hlist_nulls_node *n;
1697 unsigned long flags = 0;
1698 bool locked = false;
1699 struct htab_elem *l;
1700 struct bucket *b;
1701 int ret = 0;
1702
1703 elem_map_flags = attr->batch.elem_flags;
1704 if ((elem_map_flags & ~BPF_F_LOCK) ||
1705 ((elem_map_flags & BPF_F_LOCK) && !btf_record_has_field(map->record, BPF_SPIN_LOCK)))
1706 return -EINVAL;
1707
1708 map_flags = attr->batch.flags;
1709 if (map_flags)
1710 return -EINVAL;
1711
1712 max_count = attr->batch.count;
1713 if (!max_count)
1714 return 0;
1715
1716 if (put_user(0, &uattr->batch.count))
1717 return -EFAULT;
1718
1719 batch = 0;
1720 if (ubatch && copy_from_user(&batch, ubatch, sizeof(batch)))
1721 return -EFAULT;
1722
1723 if (batch >= htab->n_buckets)
1724 return -ENOENT;
1725
1726 key_size = htab->map.key_size;
1727 roundup_key_size = round_up(htab->map.key_size, 8);
1728 value_size = htab->map.value_size;
1729 size = round_up(value_size, 8);
1730 if (is_percpu)
1731 value_size = size * num_possible_cpus();
1732 total = 0;
1733 /* while experimenting with hash tables with sizes ranging from 10 to
1734 * 1000, it was observed that a bucket can have up to 5 entries.
1735 */
1736 bucket_size = 5;
1737
1738alloc:
1739 /* We cannot do copy_from_user or copy_to_user inside
1740 * the rcu_read_lock. Allocate enough space here.
1741 */
1742 keys = kvmalloc_array(key_size, bucket_size, GFP_USER | __GFP_NOWARN);
1743 values = kvmalloc_array(value_size, bucket_size, GFP_USER | __GFP_NOWARN);
1744 if (!keys || !values) {
1745 ret = -ENOMEM;
1746 goto after_loop;
1747 }
1748
1749again:
1750 bpf_disable_instrumentation();
1751 rcu_read_lock();
1752again_nocopy:
1753 dst_key = keys;
1754 dst_val = values;
1755 b = &htab->buckets[batch];
1756 head = &b->head;
1757 /* do not grab the lock unless need it (bucket_cnt > 0). */
1758 if (locked) {
1759 ret = htab_lock_bucket(htab, b, batch, &flags);
1760 if (ret) {
1761 rcu_read_unlock();
1762 bpf_enable_instrumentation();
1763 goto after_loop;
1764 }
1765 }
1766
1767 bucket_cnt = 0;
1768 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
1769 bucket_cnt++;
1770
1771 if (bucket_cnt && !locked) {
1772 locked = true;
1773 goto again_nocopy;
1774 }
1775
1776 if (bucket_cnt > (max_count - total)) {
1777 if (total == 0)
1778 ret = -ENOSPC;
1779 /* Note that since bucket_cnt > 0 here, it is implicit
1780 * that the locked was grabbed, so release it.
1781 */
1782 htab_unlock_bucket(htab, b, batch, flags);
1783 rcu_read_unlock();
1784 bpf_enable_instrumentation();
1785 goto after_loop;
1786 }
1787
1788 if (bucket_cnt > bucket_size) {
1789 bucket_size = bucket_cnt;
1790 /* Note that since bucket_cnt > 0 here, it is implicit
1791 * that the locked was grabbed, so release it.
1792 */
1793 htab_unlock_bucket(htab, b, batch, flags);
1794 rcu_read_unlock();
1795 bpf_enable_instrumentation();
1796 kvfree(keys);
1797 kvfree(values);
1798 goto alloc;
1799 }
1800
1801 /* Next block is only safe to run if you have grabbed the lock */
1802 if (!locked)
1803 goto next_batch;
1804
1805 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1806 memcpy(dst_key, l->key, key_size);
1807
1808 if (is_percpu) {
1809 int off = 0, cpu;
1810 void __percpu *pptr;
1811
1812 pptr = htab_elem_get_ptr(l, map->key_size);
1813 for_each_possible_cpu(cpu) {
1814 copy_map_value_long(&htab->map, dst_val + off, per_cpu_ptr(pptr, cpu));
1815 check_and_init_map_value(&htab->map, dst_val + off);
1816 off += size;
1817 }
1818 } else {
1819 value = l->key + roundup_key_size;
1820 if (map->map_type == BPF_MAP_TYPE_HASH_OF_MAPS) {
1821 struct bpf_map **inner_map = value;
1822
1823 /* Actual value is the id of the inner map */
1824 map_id = map->ops->map_fd_sys_lookup_elem(*inner_map);
1825 value = &map_id;
1826 }
1827
1828 if (elem_map_flags & BPF_F_LOCK)
1829 copy_map_value_locked(map, dst_val, value,
1830 true);
1831 else
1832 copy_map_value(map, dst_val, value);
1833 /* Zeroing special fields in the temp buffer */
1834 check_and_init_map_value(map, dst_val);
1835 }
1836 if (do_delete) {
1837 hlist_nulls_del_rcu(&l->hash_node);
1838
1839 /* bpf_lru_push_free() will acquire lru_lock, which
1840 * may cause deadlock. See comments in function
1841 * prealloc_lru_pop(). Let us do bpf_lru_push_free()
1842 * after releasing the bucket lock.
1843 */
1844 if (is_lru_map) {
1845 l->batch_flink = node_to_free;
1846 node_to_free = l;
1847 } else {
1848 free_htab_elem(htab, l);
1849 }
1850 }
1851 dst_key += key_size;
1852 dst_val += value_size;
1853 }
1854
1855 htab_unlock_bucket(htab, b, batch, flags);
1856 locked = false;
1857
1858 while (node_to_free) {
1859 l = node_to_free;
1860 node_to_free = node_to_free->batch_flink;
1861 htab_lru_push_free(htab, l);
1862 }
1863
1864next_batch:
1865 /* If we are not copying data, we can go to next bucket and avoid
1866 * unlocking the rcu.
1867 */
1868 if (!bucket_cnt && (batch + 1 < htab->n_buckets)) {
1869 batch++;
1870 goto again_nocopy;
1871 }
1872
1873 rcu_read_unlock();
1874 bpf_enable_instrumentation();
1875 if (bucket_cnt && (copy_to_user(ukeys + total * key_size, keys,
1876 key_size * bucket_cnt) ||
1877 copy_to_user(uvalues + total * value_size, values,
1878 value_size * bucket_cnt))) {
1879 ret = -EFAULT;
1880 goto after_loop;
1881 }
1882
1883 total += bucket_cnt;
1884 batch++;
1885 if (batch >= htab->n_buckets) {
1886 ret = -ENOENT;
1887 goto after_loop;
1888 }
1889 goto again;
1890
1891after_loop:
1892 if (ret == -EFAULT)
1893 goto out;
1894
1895 /* copy # of entries and next batch */
1896 ubatch = u64_to_user_ptr(attr->batch.out_batch);
1897 if (copy_to_user(ubatch, &batch, sizeof(batch)) ||
1898 put_user(total, &uattr->batch.count))
1899 ret = -EFAULT;
1900
1901out:
1902 kvfree(keys);
1903 kvfree(values);
1904 return ret;
1905}
1906
1907static int
1908htab_percpu_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1909 union bpf_attr __user *uattr)
1910{
1911 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1912 false, true);
1913}
1914
1915static int
1916htab_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1917 const union bpf_attr *attr,
1918 union bpf_attr __user *uattr)
1919{
1920 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1921 false, true);
1922}
1923
1924static int
1925htab_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1926 union bpf_attr __user *uattr)
1927{
1928 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1929 false, false);
1930}
1931
1932static int
1933htab_map_lookup_and_delete_batch(struct bpf_map *map,
1934 const union bpf_attr *attr,
1935 union bpf_attr __user *uattr)
1936{
1937 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1938 false, false);
1939}
1940
1941static int
1942htab_lru_percpu_map_lookup_batch(struct bpf_map *map,
1943 const union bpf_attr *attr,
1944 union bpf_attr __user *uattr)
1945{
1946 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1947 true, true);
1948}
1949
1950static int
1951htab_lru_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1952 const union bpf_attr *attr,
1953 union bpf_attr __user *uattr)
1954{
1955 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1956 true, true);
1957}
1958
1959static int
1960htab_lru_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1961 union bpf_attr __user *uattr)
1962{
1963 return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1964 true, false);
1965}
1966
1967static int
1968htab_lru_map_lookup_and_delete_batch(struct bpf_map *map,
1969 const union bpf_attr *attr,
1970 union bpf_attr __user *uattr)
1971{
1972 return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1973 true, false);
1974}
1975
1976struct bpf_iter_seq_hash_map_info {
1977 struct bpf_map *map;
1978 struct bpf_htab *htab;
1979 void *percpu_value_buf; // non-zero means percpu hash
1980 u32 bucket_id;
1981 u32 skip_elems;
1982};
1983
1984static struct htab_elem *
1985bpf_hash_map_seq_find_next(struct bpf_iter_seq_hash_map_info *info,
1986 struct htab_elem *prev_elem)
1987{
1988 const struct bpf_htab *htab = info->htab;
1989 u32 skip_elems = info->skip_elems;
1990 u32 bucket_id = info->bucket_id;
1991 struct hlist_nulls_head *head;
1992 struct hlist_nulls_node *n;
1993 struct htab_elem *elem;
1994 struct bucket *b;
1995 u32 i, count;
1996
1997 if (bucket_id >= htab->n_buckets)
1998 return NULL;
1999
2000 /* try to find next elem in the same bucket */
2001 if (prev_elem) {
2002 /* no update/deletion on this bucket, prev_elem should be still valid
2003 * and we won't skip elements.
2004 */
2005 n = rcu_dereference_raw(hlist_nulls_next_rcu(&prev_elem->hash_node));
2006 elem = hlist_nulls_entry_safe(n, struct htab_elem, hash_node);
2007 if (elem)
2008 return elem;
2009
2010 /* not found, unlock and go to the next bucket */
2011 b = &htab->buckets[bucket_id++];
2012 rcu_read_unlock();
2013 skip_elems = 0;
2014 }
2015
2016 for (i = bucket_id; i < htab->n_buckets; i++) {
2017 b = &htab->buckets[i];
2018 rcu_read_lock();
2019
2020 count = 0;
2021 head = &b->head;
2022 hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
2023 if (count >= skip_elems) {
2024 info->bucket_id = i;
2025 info->skip_elems = count;
2026 return elem;
2027 }
2028 count++;
2029 }
2030
2031 rcu_read_unlock();
2032 skip_elems = 0;
2033 }
2034
2035 info->bucket_id = i;
2036 info->skip_elems = 0;
2037 return NULL;
2038}
2039
2040static void *bpf_hash_map_seq_start(struct seq_file *seq, loff_t *pos)
2041{
2042 struct bpf_iter_seq_hash_map_info *info = seq->private;
2043 struct htab_elem *elem;
2044
2045 elem = bpf_hash_map_seq_find_next(info, NULL);
2046 if (!elem)
2047 return NULL;
2048
2049 if (*pos == 0)
2050 ++*pos;
2051 return elem;
2052}
2053
2054static void *bpf_hash_map_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2055{
2056 struct bpf_iter_seq_hash_map_info *info = seq->private;
2057
2058 ++*pos;
2059 ++info->skip_elems;
2060 return bpf_hash_map_seq_find_next(info, v);
2061}
2062
2063static int __bpf_hash_map_seq_show(struct seq_file *seq, struct htab_elem *elem)
2064{
2065 struct bpf_iter_seq_hash_map_info *info = seq->private;
2066 u32 roundup_key_size, roundup_value_size;
2067 struct bpf_iter__bpf_map_elem ctx = {};
2068 struct bpf_map *map = info->map;
2069 struct bpf_iter_meta meta;
2070 int ret = 0, off = 0, cpu;
2071 struct bpf_prog *prog;
2072 void __percpu *pptr;
2073
2074 meta.seq = seq;
2075 prog = bpf_iter_get_info(&meta, elem == NULL);
2076 if (prog) {
2077 ctx.meta = &meta;
2078 ctx.map = info->map;
2079 if (elem) {
2080 roundup_key_size = round_up(map->key_size, 8);
2081 ctx.key = elem->key;
2082 if (!info->percpu_value_buf) {
2083 ctx.value = elem->key + roundup_key_size;
2084 } else {
2085 roundup_value_size = round_up(map->value_size, 8);
2086 pptr = htab_elem_get_ptr(elem, map->key_size);
2087 for_each_possible_cpu(cpu) {
2088 copy_map_value_long(map, info->percpu_value_buf + off,
2089 per_cpu_ptr(pptr, cpu));
2090 check_and_init_map_value(map, info->percpu_value_buf + off);
2091 off += roundup_value_size;
2092 }
2093 ctx.value = info->percpu_value_buf;
2094 }
2095 }
2096 ret = bpf_iter_run_prog(prog, &ctx);
2097 }
2098
2099 return ret;
2100}
2101
2102static int bpf_hash_map_seq_show(struct seq_file *seq, void *v)
2103{
2104 return __bpf_hash_map_seq_show(seq, v);
2105}
2106
2107static void bpf_hash_map_seq_stop(struct seq_file *seq, void *v)
2108{
2109 if (!v)
2110 (void)__bpf_hash_map_seq_show(seq, NULL);
2111 else
2112 rcu_read_unlock();
2113}
2114
2115static int bpf_iter_init_hash_map(void *priv_data,
2116 struct bpf_iter_aux_info *aux)
2117{
2118 struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
2119 struct bpf_map *map = aux->map;
2120 void *value_buf;
2121 u32 buf_size;
2122
2123 if (map->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
2124 map->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH) {
2125 buf_size = round_up(map->value_size, 8) * num_possible_cpus();
2126 value_buf = kmalloc(buf_size, GFP_USER | __GFP_NOWARN);
2127 if (!value_buf)
2128 return -ENOMEM;
2129
2130 seq_info->percpu_value_buf = value_buf;
2131 }
2132
2133 bpf_map_inc_with_uref(map);
2134 seq_info->map = map;
2135 seq_info->htab = container_of(map, struct bpf_htab, map);
2136 return 0;
2137}
2138
2139static void bpf_iter_fini_hash_map(void *priv_data)
2140{
2141 struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
2142
2143 bpf_map_put_with_uref(seq_info->map);
2144 kfree(seq_info->percpu_value_buf);
2145}
2146
2147static const struct seq_operations bpf_hash_map_seq_ops = {
2148 .start = bpf_hash_map_seq_start,
2149 .next = bpf_hash_map_seq_next,
2150 .stop = bpf_hash_map_seq_stop,
2151 .show = bpf_hash_map_seq_show,
2152};
2153
2154static const struct bpf_iter_seq_info iter_seq_info = {
2155 .seq_ops = &bpf_hash_map_seq_ops,
2156 .init_seq_private = bpf_iter_init_hash_map,
2157 .fini_seq_private = bpf_iter_fini_hash_map,
2158 .seq_priv_size = sizeof(struct bpf_iter_seq_hash_map_info),
2159};
2160
2161static long bpf_for_each_hash_elem(struct bpf_map *map, bpf_callback_t callback_fn,
2162 void *callback_ctx, u64 flags)
2163{
2164 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2165 struct hlist_nulls_head *head;
2166 struct hlist_nulls_node *n;
2167 struct htab_elem *elem;
2168 u32 roundup_key_size;
2169 int i, num_elems = 0;
2170 void __percpu *pptr;
2171 struct bucket *b;
2172 void *key, *val;
2173 bool is_percpu;
2174 u64 ret = 0;
2175
2176 if (flags != 0)
2177 return -EINVAL;
2178
2179 is_percpu = htab_is_percpu(htab);
2180
2181 roundup_key_size = round_up(map->key_size, 8);
2182 /* disable migration so percpu value prepared here will be the
2183 * same as the one seen by the bpf program with bpf_map_lookup_elem().
2184 */
2185 if (is_percpu)
2186 migrate_disable();
2187 for (i = 0; i < htab->n_buckets; i++) {
2188 b = &htab->buckets[i];
2189 rcu_read_lock();
2190 head = &b->head;
2191 hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
2192 key = elem->key;
2193 if (is_percpu) {
2194 /* current cpu value for percpu map */
2195 pptr = htab_elem_get_ptr(elem, map->key_size);
2196 val = this_cpu_ptr(pptr);
2197 } else {
2198 val = elem->key + roundup_key_size;
2199 }
2200 num_elems++;
2201 ret = callback_fn((u64)(long)map, (u64)(long)key,
2202 (u64)(long)val, (u64)(long)callback_ctx, 0);
2203 /* return value: 0 - continue, 1 - stop and return */
2204 if (ret) {
2205 rcu_read_unlock();
2206 goto out;
2207 }
2208 }
2209 rcu_read_unlock();
2210 }
2211out:
2212 if (is_percpu)
2213 migrate_enable();
2214 return num_elems;
2215}
2216
2217static u64 htab_map_mem_usage(const struct bpf_map *map)
2218{
2219 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2220 u32 value_size = round_up(htab->map.value_size, 8);
2221 bool prealloc = htab_is_prealloc(htab);
2222 bool percpu = htab_is_percpu(htab);
2223 bool lru = htab_is_lru(htab);
2224 u64 num_entries;
2225 u64 usage = sizeof(struct bpf_htab);
2226
2227 usage += sizeof(struct bucket) * htab->n_buckets;
2228 usage += sizeof(int) * num_possible_cpus() * HASHTAB_MAP_LOCK_COUNT;
2229 if (prealloc) {
2230 num_entries = map->max_entries;
2231 if (htab_has_extra_elems(htab))
2232 num_entries += num_possible_cpus();
2233
2234 usage += htab->elem_size * num_entries;
2235
2236 if (percpu)
2237 usage += value_size * num_possible_cpus() * num_entries;
2238 else if (!lru)
2239 usage += sizeof(struct htab_elem *) * num_possible_cpus();
2240 } else {
2241#define LLIST_NODE_SZ sizeof(struct llist_node)
2242
2243 num_entries = htab->use_percpu_counter ?
2244 percpu_counter_sum(&htab->pcount) :
2245 atomic_read(&htab->count);
2246 usage += (htab->elem_size + LLIST_NODE_SZ) * num_entries;
2247 if (percpu) {
2248 usage += (LLIST_NODE_SZ + sizeof(void *)) * num_entries;
2249 usage += value_size * num_possible_cpus() * num_entries;
2250 }
2251 }
2252 return usage;
2253}
2254
2255BTF_ID_LIST_SINGLE(htab_map_btf_ids, struct, bpf_htab)
2256const struct bpf_map_ops htab_map_ops = {
2257 .map_meta_equal = bpf_map_meta_equal,
2258 .map_alloc_check = htab_map_alloc_check,
2259 .map_alloc = htab_map_alloc,
2260 .map_free = htab_map_free,
2261 .map_get_next_key = htab_map_get_next_key,
2262 .map_release_uref = htab_map_free_timers,
2263 .map_lookup_elem = htab_map_lookup_elem,
2264 .map_lookup_and_delete_elem = htab_map_lookup_and_delete_elem,
2265 .map_update_elem = htab_map_update_elem,
2266 .map_delete_elem = htab_map_delete_elem,
2267 .map_gen_lookup = htab_map_gen_lookup,
2268 .map_seq_show_elem = htab_map_seq_show_elem,
2269 .map_set_for_each_callback_args = map_set_for_each_callback_args,
2270 .map_for_each_callback = bpf_for_each_hash_elem,
2271 .map_mem_usage = htab_map_mem_usage,
2272 BATCH_OPS(htab),
2273 .map_btf_id = &htab_map_btf_ids[0],
2274 .iter_seq_info = &iter_seq_info,
2275};
2276
2277const struct bpf_map_ops htab_lru_map_ops = {
2278 .map_meta_equal = bpf_map_meta_equal,
2279 .map_alloc_check = htab_map_alloc_check,
2280 .map_alloc = htab_map_alloc,
2281 .map_free = htab_map_free,
2282 .map_get_next_key = htab_map_get_next_key,
2283 .map_release_uref = htab_map_free_timers,
2284 .map_lookup_elem = htab_lru_map_lookup_elem,
2285 .map_lookup_and_delete_elem = htab_lru_map_lookup_and_delete_elem,
2286 .map_lookup_elem_sys_only = htab_lru_map_lookup_elem_sys,
2287 .map_update_elem = htab_lru_map_update_elem,
2288 .map_delete_elem = htab_lru_map_delete_elem,
2289 .map_gen_lookup = htab_lru_map_gen_lookup,
2290 .map_seq_show_elem = htab_map_seq_show_elem,
2291 .map_set_for_each_callback_args = map_set_for_each_callback_args,
2292 .map_for_each_callback = bpf_for_each_hash_elem,
2293 .map_mem_usage = htab_map_mem_usage,
2294 BATCH_OPS(htab_lru),
2295 .map_btf_id = &htab_map_btf_ids[0],
2296 .iter_seq_info = &iter_seq_info,
2297};
2298
2299/* Called from eBPF program */
2300static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
2301{
2302 struct htab_elem *l = __htab_map_lookup_elem(map, key);
2303
2304 if (l)
2305 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
2306 else
2307 return NULL;
2308}
2309
2310static void *htab_percpu_map_lookup_percpu_elem(struct bpf_map *map, void *key, u32 cpu)
2311{
2312 struct htab_elem *l;
2313
2314 if (cpu >= nr_cpu_ids)
2315 return NULL;
2316
2317 l = __htab_map_lookup_elem(map, key);
2318 if (l)
2319 return per_cpu_ptr(htab_elem_get_ptr(l, map->key_size), cpu);
2320 else
2321 return NULL;
2322}
2323
2324static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
2325{
2326 struct htab_elem *l = __htab_map_lookup_elem(map, key);
2327
2328 if (l) {
2329 bpf_lru_node_set_ref(&l->lru_node);
2330 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
2331 }
2332
2333 return NULL;
2334}
2335
2336static void *htab_lru_percpu_map_lookup_percpu_elem(struct bpf_map *map, void *key, u32 cpu)
2337{
2338 struct htab_elem *l;
2339
2340 if (cpu >= nr_cpu_ids)
2341 return NULL;
2342
2343 l = __htab_map_lookup_elem(map, key);
2344 if (l) {
2345 bpf_lru_node_set_ref(&l->lru_node);
2346 return per_cpu_ptr(htab_elem_get_ptr(l, map->key_size), cpu);
2347 }
2348
2349 return NULL;
2350}
2351
2352int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
2353{
2354 struct htab_elem *l;
2355 void __percpu *pptr;
2356 int ret = -ENOENT;
2357 int cpu, off = 0;
2358 u32 size;
2359
2360 /* per_cpu areas are zero-filled and bpf programs can only
2361 * access 'value_size' of them, so copying rounded areas
2362 * will not leak any kernel data
2363 */
2364 size = round_up(map->value_size, 8);
2365 rcu_read_lock();
2366 l = __htab_map_lookup_elem(map, key);
2367 if (!l)
2368 goto out;
2369 /* We do not mark LRU map element here in order to not mess up
2370 * eviction heuristics when user space does a map walk.
2371 */
2372 pptr = htab_elem_get_ptr(l, map->key_size);
2373 for_each_possible_cpu(cpu) {
2374 copy_map_value_long(map, value + off, per_cpu_ptr(pptr, cpu));
2375 check_and_init_map_value(map, value + off);
2376 off += size;
2377 }
2378 ret = 0;
2379out:
2380 rcu_read_unlock();
2381 return ret;
2382}
2383
2384int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
2385 u64 map_flags)
2386{
2387 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2388 int ret;
2389
2390 rcu_read_lock();
2391 if (htab_is_lru(htab))
2392 ret = __htab_lru_percpu_map_update_elem(map, key, value,
2393 map_flags, true);
2394 else
2395 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
2396 true);
2397 rcu_read_unlock();
2398
2399 return ret;
2400}
2401
2402static void htab_percpu_map_seq_show_elem(struct bpf_map *map, void *key,
2403 struct seq_file *m)
2404{
2405 struct htab_elem *l;
2406 void __percpu *pptr;
2407 int cpu;
2408
2409 rcu_read_lock();
2410
2411 l = __htab_map_lookup_elem(map, key);
2412 if (!l) {
2413 rcu_read_unlock();
2414 return;
2415 }
2416
2417 btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
2418 seq_puts(m, ": {\n");
2419 pptr = htab_elem_get_ptr(l, map->key_size);
2420 for_each_possible_cpu(cpu) {
2421 seq_printf(m, "\tcpu%d: ", cpu);
2422 btf_type_seq_show(map->btf, map->btf_value_type_id,
2423 per_cpu_ptr(pptr, cpu), m);
2424 seq_puts(m, "\n");
2425 }
2426 seq_puts(m, "}\n");
2427
2428 rcu_read_unlock();
2429}
2430
2431const struct bpf_map_ops htab_percpu_map_ops = {
2432 .map_meta_equal = bpf_map_meta_equal,
2433 .map_alloc_check = htab_map_alloc_check,
2434 .map_alloc = htab_map_alloc,
2435 .map_free = htab_map_free,
2436 .map_get_next_key = htab_map_get_next_key,
2437 .map_lookup_elem = htab_percpu_map_lookup_elem,
2438 .map_lookup_and_delete_elem = htab_percpu_map_lookup_and_delete_elem,
2439 .map_update_elem = htab_percpu_map_update_elem,
2440 .map_delete_elem = htab_map_delete_elem,
2441 .map_lookup_percpu_elem = htab_percpu_map_lookup_percpu_elem,
2442 .map_seq_show_elem = htab_percpu_map_seq_show_elem,
2443 .map_set_for_each_callback_args = map_set_for_each_callback_args,
2444 .map_for_each_callback = bpf_for_each_hash_elem,
2445 .map_mem_usage = htab_map_mem_usage,
2446 BATCH_OPS(htab_percpu),
2447 .map_btf_id = &htab_map_btf_ids[0],
2448 .iter_seq_info = &iter_seq_info,
2449};
2450
2451const struct bpf_map_ops htab_lru_percpu_map_ops = {
2452 .map_meta_equal = bpf_map_meta_equal,
2453 .map_alloc_check = htab_map_alloc_check,
2454 .map_alloc = htab_map_alloc,
2455 .map_free = htab_map_free,
2456 .map_get_next_key = htab_map_get_next_key,
2457 .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
2458 .map_lookup_and_delete_elem = htab_lru_percpu_map_lookup_and_delete_elem,
2459 .map_update_elem = htab_lru_percpu_map_update_elem,
2460 .map_delete_elem = htab_lru_map_delete_elem,
2461 .map_lookup_percpu_elem = htab_lru_percpu_map_lookup_percpu_elem,
2462 .map_seq_show_elem = htab_percpu_map_seq_show_elem,
2463 .map_set_for_each_callback_args = map_set_for_each_callback_args,
2464 .map_for_each_callback = bpf_for_each_hash_elem,
2465 .map_mem_usage = htab_map_mem_usage,
2466 BATCH_OPS(htab_lru_percpu),
2467 .map_btf_id = &htab_map_btf_ids[0],
2468 .iter_seq_info = &iter_seq_info,
2469};
2470
2471static int fd_htab_map_alloc_check(union bpf_attr *attr)
2472{
2473 if (attr->value_size != sizeof(u32))
2474 return -EINVAL;
2475 return htab_map_alloc_check(attr);
2476}
2477
2478static void fd_htab_map_free(struct bpf_map *map)
2479{
2480 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2481 struct hlist_nulls_node *n;
2482 struct hlist_nulls_head *head;
2483 struct htab_elem *l;
2484 int i;
2485
2486 for (i = 0; i < htab->n_buckets; i++) {
2487 head = select_bucket(htab, i);
2488
2489 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
2490 void *ptr = fd_htab_map_get_ptr(map, l);
2491
2492 map->ops->map_fd_put_ptr(map, ptr, false);
2493 }
2494 }
2495
2496 htab_map_free(map);
2497}
2498
2499/* only called from syscall */
2500int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
2501{
2502 void **ptr;
2503 int ret = 0;
2504
2505 if (!map->ops->map_fd_sys_lookup_elem)
2506 return -ENOTSUPP;
2507
2508 rcu_read_lock();
2509 ptr = htab_map_lookup_elem(map, key);
2510 if (ptr)
2511 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
2512 else
2513 ret = -ENOENT;
2514 rcu_read_unlock();
2515
2516 return ret;
2517}
2518
2519/* only called from syscall */
2520int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
2521 void *key, void *value, u64 map_flags)
2522{
2523 void *ptr;
2524 int ret;
2525 u32 ufd = *(u32 *)value;
2526
2527 ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
2528 if (IS_ERR(ptr))
2529 return PTR_ERR(ptr);
2530
2531 /* The htab bucket lock is always held during update operations in fd
2532 * htab map, and the following rcu_read_lock() is only used to avoid
2533 * the WARN_ON_ONCE in htab_map_update_elem().
2534 */
2535 rcu_read_lock();
2536 ret = htab_map_update_elem(map, key, &ptr, map_flags);
2537 rcu_read_unlock();
2538 if (ret)
2539 map->ops->map_fd_put_ptr(map, ptr, false);
2540
2541 return ret;
2542}
2543
2544static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
2545{
2546 struct bpf_map *map, *inner_map_meta;
2547
2548 inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
2549 if (IS_ERR(inner_map_meta))
2550 return inner_map_meta;
2551
2552 map = htab_map_alloc(attr);
2553 if (IS_ERR(map)) {
2554 bpf_map_meta_free(inner_map_meta);
2555 return map;
2556 }
2557
2558 map->inner_map_meta = inner_map_meta;
2559
2560 return map;
2561}
2562
2563static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
2564{
2565 struct bpf_map **inner_map = htab_map_lookup_elem(map, key);
2566
2567 if (!inner_map)
2568 return NULL;
2569
2570 return READ_ONCE(*inner_map);
2571}
2572
2573static int htab_of_map_gen_lookup(struct bpf_map *map,
2574 struct bpf_insn *insn_buf)
2575{
2576 struct bpf_insn *insn = insn_buf;
2577 const int ret = BPF_REG_0;
2578
2579 BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
2580 (void *(*)(struct bpf_map *map, void *key))NULL));
2581 *insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
2582 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
2583 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
2584 offsetof(struct htab_elem, key) +
2585 round_up(map->key_size, 8));
2586 *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
2587
2588 return insn - insn_buf;
2589}
2590
2591static void htab_of_map_free(struct bpf_map *map)
2592{
2593 bpf_map_meta_free(map->inner_map_meta);
2594 fd_htab_map_free(map);
2595}
2596
2597const struct bpf_map_ops htab_of_maps_map_ops = {
2598 .map_alloc_check = fd_htab_map_alloc_check,
2599 .map_alloc = htab_of_map_alloc,
2600 .map_free = htab_of_map_free,
2601 .map_get_next_key = htab_map_get_next_key,
2602 .map_lookup_elem = htab_of_map_lookup_elem,
2603 .map_delete_elem = htab_map_delete_elem,
2604 .map_fd_get_ptr = bpf_map_fd_get_ptr,
2605 .map_fd_put_ptr = bpf_map_fd_put_ptr,
2606 .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
2607 .map_gen_lookup = htab_of_map_gen_lookup,
2608 .map_check_btf = map_check_no_btf,
2609 .map_mem_usage = htab_map_mem_usage,
2610 BATCH_OPS(htab),
2611 .map_btf_id = &htab_map_btf_ids[0],
2612};