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/* 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 <linux/rculist_nulls.h>
17#include "percpu_freelist.h"
18#include "bpf_lru_list.h"
19#include "map_in_map.h"
20
21#define HTAB_CREATE_FLAG_MASK \
22 (BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU | BPF_F_NUMA_NODE | \
23 BPF_F_RDONLY | BPF_F_WRONLY)
24
25struct bucket {
26 struct hlist_nulls_head head;
27 raw_spinlock_t lock;
28};
29
30struct bpf_htab {
31 struct bpf_map map;
32 struct bucket *buckets;
33 void *elems;
34 union {
35 struct pcpu_freelist freelist;
36 struct bpf_lru lru;
37 };
38 struct htab_elem *__percpu *extra_elems;
39 atomic_t count; /* number of elements in this hashtable */
40 u32 n_buckets; /* number of hash buckets */
41 u32 elem_size; /* size of each element in bytes */
42};
43
44/* each htab element is struct htab_elem + key + value */
45struct htab_elem {
46 union {
47 struct hlist_nulls_node hash_node;
48 struct {
49 void *padding;
50 union {
51 struct bpf_htab *htab;
52 struct pcpu_freelist_node fnode;
53 };
54 };
55 };
56 union {
57 struct rcu_head rcu;
58 struct bpf_lru_node lru_node;
59 };
60 u32 hash;
61 char key[0] __aligned(8);
62};
63
64static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
65
66static bool htab_is_lru(const struct bpf_htab *htab)
67{
68 return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
69 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
70}
71
72static bool htab_is_percpu(const struct bpf_htab *htab)
73{
74 return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
75 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
76}
77
78static bool htab_is_prealloc(const struct bpf_htab *htab)
79{
80 return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
81}
82
83static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
84 void __percpu *pptr)
85{
86 *(void __percpu **)(l->key + key_size) = pptr;
87}
88
89static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
90{
91 return *(void __percpu **)(l->key + key_size);
92}
93
94static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
95{
96 return *(void **)(l->key + roundup(map->key_size, 8));
97}
98
99static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
100{
101 return (struct htab_elem *) (htab->elems + i * htab->elem_size);
102}
103
104static void htab_free_elems(struct bpf_htab *htab)
105{
106 int i;
107
108 if (!htab_is_percpu(htab))
109 goto free_elems;
110
111 for (i = 0; i < htab->map.max_entries; i++) {
112 void __percpu *pptr;
113
114 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
115 htab->map.key_size);
116 free_percpu(pptr);
117 cond_resched();
118 }
119free_elems:
120 bpf_map_area_free(htab->elems);
121}
122
123static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
124 u32 hash)
125{
126 struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
127 struct htab_elem *l;
128
129 if (node) {
130 l = container_of(node, struct htab_elem, lru_node);
131 memcpy(l->key, key, htab->map.key_size);
132 return l;
133 }
134
135 return NULL;
136}
137
138static int prealloc_init(struct bpf_htab *htab)
139{
140 u32 num_entries = htab->map.max_entries;
141 int err = -ENOMEM, i;
142
143 if (!htab_is_percpu(htab) && !htab_is_lru(htab))
144 num_entries += num_possible_cpus();
145
146 htab->elems = bpf_map_area_alloc(htab->elem_size * num_entries,
147 htab->map.numa_node);
148 if (!htab->elems)
149 return -ENOMEM;
150
151 if (!htab_is_percpu(htab))
152 goto skip_percpu_elems;
153
154 for (i = 0; i < num_entries; i++) {
155 u32 size = round_up(htab->map.value_size, 8);
156 void __percpu *pptr;
157
158 pptr = __alloc_percpu_gfp(size, 8, GFP_USER | __GFP_NOWARN);
159 if (!pptr)
160 goto free_elems;
161 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
162 pptr);
163 cond_resched();
164 }
165
166skip_percpu_elems:
167 if (htab_is_lru(htab))
168 err = bpf_lru_init(&htab->lru,
169 htab->map.map_flags & BPF_F_NO_COMMON_LRU,
170 offsetof(struct htab_elem, hash) -
171 offsetof(struct htab_elem, lru_node),
172 htab_lru_map_delete_node,
173 htab);
174 else
175 err = pcpu_freelist_init(&htab->freelist);
176
177 if (err)
178 goto free_elems;
179
180 if (htab_is_lru(htab))
181 bpf_lru_populate(&htab->lru, htab->elems,
182 offsetof(struct htab_elem, lru_node),
183 htab->elem_size, num_entries);
184 else
185 pcpu_freelist_populate(&htab->freelist,
186 htab->elems + offsetof(struct htab_elem, fnode),
187 htab->elem_size, num_entries);
188
189 return 0;
190
191free_elems:
192 htab_free_elems(htab);
193 return err;
194}
195
196static void prealloc_destroy(struct bpf_htab *htab)
197{
198 htab_free_elems(htab);
199
200 if (htab_is_lru(htab))
201 bpf_lru_destroy(&htab->lru);
202 else
203 pcpu_freelist_destroy(&htab->freelist);
204}
205
206static int alloc_extra_elems(struct bpf_htab *htab)
207{
208 struct htab_elem *__percpu *pptr, *l_new;
209 struct pcpu_freelist_node *l;
210 int cpu;
211
212 pptr = __alloc_percpu_gfp(sizeof(struct htab_elem *), 8,
213 GFP_USER | __GFP_NOWARN);
214 if (!pptr)
215 return -ENOMEM;
216
217 for_each_possible_cpu(cpu) {
218 l = pcpu_freelist_pop(&htab->freelist);
219 /* pop will succeed, since prealloc_init()
220 * preallocated extra num_possible_cpus elements
221 */
222 l_new = container_of(l, struct htab_elem, fnode);
223 *per_cpu_ptr(pptr, cpu) = l_new;
224 }
225 htab->extra_elems = pptr;
226 return 0;
227}
228
229/* Called from syscall */
230static int htab_map_alloc_check(union bpf_attr *attr)
231{
232 bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
233 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
234 bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
235 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
236 /* percpu_lru means each cpu has its own LRU list.
237 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
238 * the map's value itself is percpu. percpu_lru has
239 * nothing to do with the map's value.
240 */
241 bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
242 bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
243 int numa_node = bpf_map_attr_numa_node(attr);
244
245 BUILD_BUG_ON(offsetof(struct htab_elem, htab) !=
246 offsetof(struct htab_elem, hash_node.pprev));
247 BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
248 offsetof(struct htab_elem, hash_node.pprev));
249
250 if (lru && !capable(CAP_SYS_ADMIN))
251 /* LRU implementation is much complicated than other
252 * maps. Hence, limit to CAP_SYS_ADMIN for now.
253 */
254 return -EPERM;
255
256 if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK)
257 /* reserved bits should not be used */
258 return -EINVAL;
259
260 if (!lru && percpu_lru)
261 return -EINVAL;
262
263 if (lru && !prealloc)
264 return -ENOTSUPP;
265
266 if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
267 return -EINVAL;
268
269 /* check sanity of attributes.
270 * value_size == 0 may be allowed in the future to use map as a set
271 */
272 if (attr->max_entries == 0 || attr->key_size == 0 ||
273 attr->value_size == 0)
274 return -EINVAL;
275
276 if (attr->key_size > MAX_BPF_STACK)
277 /* eBPF programs initialize keys on stack, so they cannot be
278 * larger than max stack size
279 */
280 return -E2BIG;
281
282 if (attr->value_size >= KMALLOC_MAX_SIZE -
283 MAX_BPF_STACK - sizeof(struct htab_elem))
284 /* if value_size is bigger, the user space won't be able to
285 * access the elements via bpf syscall. This check also makes
286 * sure that the elem_size doesn't overflow and it's
287 * kmalloc-able later in htab_map_update_elem()
288 */
289 return -E2BIG;
290
291 return 0;
292}
293
294static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
295{
296 bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
297 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
298 bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
299 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
300 /* percpu_lru means each cpu has its own LRU list.
301 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
302 * the map's value itself is percpu. percpu_lru has
303 * nothing to do with the map's value.
304 */
305 bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
306 bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
307 struct bpf_htab *htab;
308 int err, i;
309 u64 cost;
310
311 htab = kzalloc(sizeof(*htab), GFP_USER);
312 if (!htab)
313 return ERR_PTR(-ENOMEM);
314
315 bpf_map_init_from_attr(&htab->map, attr);
316
317 if (percpu_lru) {
318 /* ensure each CPU's lru list has >=1 elements.
319 * since we are at it, make each lru list has the same
320 * number of elements.
321 */
322 htab->map.max_entries = roundup(attr->max_entries,
323 num_possible_cpus());
324 if (htab->map.max_entries < attr->max_entries)
325 htab->map.max_entries = rounddown(attr->max_entries,
326 num_possible_cpus());
327 }
328
329 /* hash table size must be power of 2 */
330 htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
331
332 htab->elem_size = sizeof(struct htab_elem) +
333 round_up(htab->map.key_size, 8);
334 if (percpu)
335 htab->elem_size += sizeof(void *);
336 else
337 htab->elem_size += round_up(htab->map.value_size, 8);
338
339 err = -E2BIG;
340 /* prevent zero size kmalloc and check for u32 overflow */
341 if (htab->n_buckets == 0 ||
342 htab->n_buckets > U32_MAX / sizeof(struct bucket))
343 goto free_htab;
344
345 cost = (u64) htab->n_buckets * sizeof(struct bucket) +
346 (u64) htab->elem_size * htab->map.max_entries;
347
348 if (percpu)
349 cost += (u64) round_up(htab->map.value_size, 8) *
350 num_possible_cpus() * htab->map.max_entries;
351 else
352 cost += (u64) htab->elem_size * num_possible_cpus();
353
354 if (cost >= U32_MAX - PAGE_SIZE)
355 /* make sure page count doesn't overflow */
356 goto free_htab;
357
358 htab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
359
360 /* if map size is larger than memlock limit, reject it early */
361 err = bpf_map_precharge_memlock(htab->map.pages);
362 if (err)
363 goto free_htab;
364
365 err = -ENOMEM;
366 htab->buckets = bpf_map_area_alloc(htab->n_buckets *
367 sizeof(struct bucket),
368 htab->map.numa_node);
369 if (!htab->buckets)
370 goto free_htab;
371
372 for (i = 0; i < htab->n_buckets; i++) {
373 INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
374 raw_spin_lock_init(&htab->buckets[i].lock);
375 }
376
377 if (prealloc) {
378 err = prealloc_init(htab);
379 if (err)
380 goto free_buckets;
381
382 if (!percpu && !lru) {
383 /* lru itself can remove the least used element, so
384 * there is no need for an extra elem during map_update.
385 */
386 err = alloc_extra_elems(htab);
387 if (err)
388 goto free_prealloc;
389 }
390 }
391
392 return &htab->map;
393
394free_prealloc:
395 prealloc_destroy(htab);
396free_buckets:
397 bpf_map_area_free(htab->buckets);
398free_htab:
399 kfree(htab);
400 return ERR_PTR(err);
401}
402
403static inline u32 htab_map_hash(const void *key, u32 key_len)
404{
405 return jhash(key, key_len, 0);
406}
407
408static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
409{
410 return &htab->buckets[hash & (htab->n_buckets - 1)];
411}
412
413static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
414{
415 return &__select_bucket(htab, hash)->head;
416}
417
418/* this lookup function can only be called with bucket lock taken */
419static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
420 void *key, u32 key_size)
421{
422 struct hlist_nulls_node *n;
423 struct htab_elem *l;
424
425 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
426 if (l->hash == hash && !memcmp(&l->key, key, key_size))
427 return l;
428
429 return NULL;
430}
431
432/* can be called without bucket lock. it will repeat the loop in
433 * the unlikely event when elements moved from one bucket into another
434 * while link list is being walked
435 */
436static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
437 u32 hash, void *key,
438 u32 key_size, u32 n_buckets)
439{
440 struct hlist_nulls_node *n;
441 struct htab_elem *l;
442
443again:
444 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
445 if (l->hash == hash && !memcmp(&l->key, key, key_size))
446 return l;
447
448 if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
449 goto again;
450
451 return NULL;
452}
453
454/* Called from syscall or from eBPF program directly, so
455 * arguments have to match bpf_map_lookup_elem() exactly.
456 * The return value is adjusted by BPF instructions
457 * in htab_map_gen_lookup().
458 */
459static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
460{
461 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
462 struct hlist_nulls_head *head;
463 struct htab_elem *l;
464 u32 hash, key_size;
465
466 /* Must be called with rcu_read_lock. */
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 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
476
477 return l;
478}
479
480static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
481{
482 struct htab_elem *l = __htab_map_lookup_elem(map, key);
483
484 if (l)
485 return l->key + round_up(map->key_size, 8);
486
487 return NULL;
488}
489
490/* inline bpf_map_lookup_elem() call.
491 * Instead of:
492 * bpf_prog
493 * bpf_map_lookup_elem
494 * map->ops->map_lookup_elem
495 * htab_map_lookup_elem
496 * __htab_map_lookup_elem
497 * do:
498 * bpf_prog
499 * __htab_map_lookup_elem
500 */
501static u32 htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
502{
503 struct bpf_insn *insn = insn_buf;
504 const int ret = BPF_REG_0;
505
506 *insn++ = BPF_EMIT_CALL((u64 (*)(u64, u64, u64, u64, u64))__htab_map_lookup_elem);
507 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
508 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
509 offsetof(struct htab_elem, key) +
510 round_up(map->key_size, 8));
511 return insn - insn_buf;
512}
513
514static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
515{
516 struct htab_elem *l = __htab_map_lookup_elem(map, key);
517
518 if (l) {
519 bpf_lru_node_set_ref(&l->lru_node);
520 return l->key + round_up(map->key_size, 8);
521 }
522
523 return NULL;
524}
525
526static u32 htab_lru_map_gen_lookup(struct bpf_map *map,
527 struct bpf_insn *insn_buf)
528{
529 struct bpf_insn *insn = insn_buf;
530 const int ret = BPF_REG_0;
531 const int ref_reg = BPF_REG_1;
532
533 *insn++ = BPF_EMIT_CALL((u64 (*)(u64, u64, u64, u64, u64))__htab_map_lookup_elem);
534 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
535 *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
536 offsetof(struct htab_elem, lru_node) +
537 offsetof(struct bpf_lru_node, ref));
538 *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
539 *insn++ = BPF_ST_MEM(BPF_B, ret,
540 offsetof(struct htab_elem, lru_node) +
541 offsetof(struct bpf_lru_node, ref),
542 1);
543 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
544 offsetof(struct htab_elem, key) +
545 round_up(map->key_size, 8));
546 return insn - insn_buf;
547}
548
549/* It is called from the bpf_lru_list when the LRU needs to delete
550 * older elements from the htab.
551 */
552static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
553{
554 struct bpf_htab *htab = (struct bpf_htab *)arg;
555 struct htab_elem *l = NULL, *tgt_l;
556 struct hlist_nulls_head *head;
557 struct hlist_nulls_node *n;
558 unsigned long flags;
559 struct bucket *b;
560
561 tgt_l = container_of(node, struct htab_elem, lru_node);
562 b = __select_bucket(htab, tgt_l->hash);
563 head = &b->head;
564
565 raw_spin_lock_irqsave(&b->lock, flags);
566
567 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
568 if (l == tgt_l) {
569 hlist_nulls_del_rcu(&l->hash_node);
570 break;
571 }
572
573 raw_spin_unlock_irqrestore(&b->lock, flags);
574
575 return l == tgt_l;
576}
577
578/* Called from syscall */
579static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
580{
581 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
582 struct hlist_nulls_head *head;
583 struct htab_elem *l, *next_l;
584 u32 hash, key_size;
585 int i = 0;
586
587 WARN_ON_ONCE(!rcu_read_lock_held());
588
589 key_size = map->key_size;
590
591 if (!key)
592 goto find_first_elem;
593
594 hash = htab_map_hash(key, key_size);
595
596 head = select_bucket(htab, hash);
597
598 /* lookup the key */
599 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
600
601 if (!l)
602 goto find_first_elem;
603
604 /* key was found, get next key in the same bucket */
605 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
606 struct htab_elem, hash_node);
607
608 if (next_l) {
609 /* if next elem in this hash list is non-zero, just return it */
610 memcpy(next_key, next_l->key, key_size);
611 return 0;
612 }
613
614 /* no more elements in this hash list, go to the next bucket */
615 i = hash & (htab->n_buckets - 1);
616 i++;
617
618find_first_elem:
619 /* iterate over buckets */
620 for (; i < htab->n_buckets; i++) {
621 head = select_bucket(htab, i);
622
623 /* pick first element in the bucket */
624 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
625 struct htab_elem, hash_node);
626 if (next_l) {
627 /* if it's not empty, just return it */
628 memcpy(next_key, next_l->key, key_size);
629 return 0;
630 }
631 }
632
633 /* iterated over all buckets and all elements */
634 return -ENOENT;
635}
636
637static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
638{
639 if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
640 free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
641 kfree(l);
642}
643
644static void htab_elem_free_rcu(struct rcu_head *head)
645{
646 struct htab_elem *l = container_of(head, struct htab_elem, rcu);
647 struct bpf_htab *htab = l->htab;
648
649 /* must increment bpf_prog_active to avoid kprobe+bpf triggering while
650 * we're calling kfree, otherwise deadlock is possible if kprobes
651 * are placed somewhere inside of slub
652 */
653 preempt_disable();
654 __this_cpu_inc(bpf_prog_active);
655 htab_elem_free(htab, l);
656 __this_cpu_dec(bpf_prog_active);
657 preempt_enable();
658}
659
660static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
661{
662 struct bpf_map *map = &htab->map;
663
664 if (map->ops->map_fd_put_ptr) {
665 void *ptr = fd_htab_map_get_ptr(map, l);
666
667 map->ops->map_fd_put_ptr(ptr);
668 }
669
670 if (htab_is_prealloc(htab)) {
671 pcpu_freelist_push(&htab->freelist, &l->fnode);
672 } else {
673 atomic_dec(&htab->count);
674 l->htab = htab;
675 call_rcu(&l->rcu, htab_elem_free_rcu);
676 }
677}
678
679static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
680 void *value, bool onallcpus)
681{
682 if (!onallcpus) {
683 /* copy true value_size bytes */
684 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
685 } else {
686 u32 size = round_up(htab->map.value_size, 8);
687 int off = 0, cpu;
688
689 for_each_possible_cpu(cpu) {
690 bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
691 value + off, size);
692 off += size;
693 }
694 }
695}
696
697static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
698{
699 return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
700 BITS_PER_LONG == 64;
701}
702
703static u32 htab_size_value(const struct bpf_htab *htab, bool percpu)
704{
705 u32 size = htab->map.value_size;
706
707 if (percpu || fd_htab_map_needs_adjust(htab))
708 size = round_up(size, 8);
709 return size;
710}
711
712static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
713 void *value, u32 key_size, u32 hash,
714 bool percpu, bool onallcpus,
715 struct htab_elem *old_elem)
716{
717 u32 size = htab_size_value(htab, percpu);
718 bool prealloc = htab_is_prealloc(htab);
719 struct htab_elem *l_new, **pl_new;
720 void __percpu *pptr;
721
722 if (prealloc) {
723 if (old_elem) {
724 /* if we're updating the existing element,
725 * use per-cpu extra elems to avoid freelist_pop/push
726 */
727 pl_new = this_cpu_ptr(htab->extra_elems);
728 l_new = *pl_new;
729 *pl_new = old_elem;
730 } else {
731 struct pcpu_freelist_node *l;
732
733 l = pcpu_freelist_pop(&htab->freelist);
734 if (!l)
735 return ERR_PTR(-E2BIG);
736 l_new = container_of(l, struct htab_elem, fnode);
737 }
738 } else {
739 if (atomic_inc_return(&htab->count) > htab->map.max_entries)
740 if (!old_elem) {
741 /* when map is full and update() is replacing
742 * old element, it's ok to allocate, since
743 * old element will be freed immediately.
744 * Otherwise return an error
745 */
746 atomic_dec(&htab->count);
747 return ERR_PTR(-E2BIG);
748 }
749 l_new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
750 htab->map.numa_node);
751 if (!l_new)
752 return ERR_PTR(-ENOMEM);
753 }
754
755 memcpy(l_new->key, key, key_size);
756 if (percpu) {
757 if (prealloc) {
758 pptr = htab_elem_get_ptr(l_new, key_size);
759 } else {
760 /* alloc_percpu zero-fills */
761 pptr = __alloc_percpu_gfp(size, 8,
762 GFP_ATOMIC | __GFP_NOWARN);
763 if (!pptr) {
764 kfree(l_new);
765 return ERR_PTR(-ENOMEM);
766 }
767 }
768
769 pcpu_copy_value(htab, pptr, value, onallcpus);
770
771 if (!prealloc)
772 htab_elem_set_ptr(l_new, key_size, pptr);
773 } else {
774 memcpy(l_new->key + round_up(key_size, 8), value, size);
775 }
776
777 l_new->hash = hash;
778 return l_new;
779}
780
781static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
782 u64 map_flags)
783{
784 if (l_old && map_flags == BPF_NOEXIST)
785 /* elem already exists */
786 return -EEXIST;
787
788 if (!l_old && map_flags == BPF_EXIST)
789 /* elem doesn't exist, cannot update it */
790 return -ENOENT;
791
792 return 0;
793}
794
795/* Called from syscall or from eBPF program */
796static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
797 u64 map_flags)
798{
799 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
800 struct htab_elem *l_new = NULL, *l_old;
801 struct hlist_nulls_head *head;
802 unsigned long flags;
803 struct bucket *b;
804 u32 key_size, hash;
805 int ret;
806
807 if (unlikely(map_flags > BPF_EXIST))
808 /* unknown flags */
809 return -EINVAL;
810
811 WARN_ON_ONCE(!rcu_read_lock_held());
812
813 key_size = map->key_size;
814
815 hash = htab_map_hash(key, key_size);
816
817 b = __select_bucket(htab, hash);
818 head = &b->head;
819
820 /* bpf_map_update_elem() can be called in_irq() */
821 raw_spin_lock_irqsave(&b->lock, flags);
822
823 l_old = lookup_elem_raw(head, hash, key, key_size);
824
825 ret = check_flags(htab, l_old, map_flags);
826 if (ret)
827 goto err;
828
829 l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
830 l_old);
831 if (IS_ERR(l_new)) {
832 /* all pre-allocated elements are in use or memory exhausted */
833 ret = PTR_ERR(l_new);
834 goto err;
835 }
836
837 /* add new element to the head of the list, so that
838 * concurrent search will find it before old elem
839 */
840 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
841 if (l_old) {
842 hlist_nulls_del_rcu(&l_old->hash_node);
843 if (!htab_is_prealloc(htab))
844 free_htab_elem(htab, l_old);
845 }
846 ret = 0;
847err:
848 raw_spin_unlock_irqrestore(&b->lock, flags);
849 return ret;
850}
851
852static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
853 u64 map_flags)
854{
855 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
856 struct htab_elem *l_new, *l_old = NULL;
857 struct hlist_nulls_head *head;
858 unsigned long flags;
859 struct bucket *b;
860 u32 key_size, hash;
861 int ret;
862
863 if (unlikely(map_flags > BPF_EXIST))
864 /* unknown flags */
865 return -EINVAL;
866
867 WARN_ON_ONCE(!rcu_read_lock_held());
868
869 key_size = map->key_size;
870
871 hash = htab_map_hash(key, key_size);
872
873 b = __select_bucket(htab, hash);
874 head = &b->head;
875
876 /* For LRU, we need to alloc before taking bucket's
877 * spinlock because getting free nodes from LRU may need
878 * to remove older elements from htab and this removal
879 * operation will need a bucket lock.
880 */
881 l_new = prealloc_lru_pop(htab, key, hash);
882 if (!l_new)
883 return -ENOMEM;
884 memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
885
886 /* bpf_map_update_elem() can be called in_irq() */
887 raw_spin_lock_irqsave(&b->lock, flags);
888
889 l_old = lookup_elem_raw(head, hash, key, key_size);
890
891 ret = check_flags(htab, l_old, map_flags);
892 if (ret)
893 goto err;
894
895 /* add new element to the head of the list, so that
896 * concurrent search will find it before old elem
897 */
898 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
899 if (l_old) {
900 bpf_lru_node_set_ref(&l_new->lru_node);
901 hlist_nulls_del_rcu(&l_old->hash_node);
902 }
903 ret = 0;
904
905err:
906 raw_spin_unlock_irqrestore(&b->lock, flags);
907
908 if (ret)
909 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
910 else if (l_old)
911 bpf_lru_push_free(&htab->lru, &l_old->lru_node);
912
913 return ret;
914}
915
916static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
917 void *value, u64 map_flags,
918 bool onallcpus)
919{
920 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
921 struct htab_elem *l_new = NULL, *l_old;
922 struct hlist_nulls_head *head;
923 unsigned long flags;
924 struct bucket *b;
925 u32 key_size, hash;
926 int ret;
927
928 if (unlikely(map_flags > BPF_EXIST))
929 /* unknown flags */
930 return -EINVAL;
931
932 WARN_ON_ONCE(!rcu_read_lock_held());
933
934 key_size = map->key_size;
935
936 hash = htab_map_hash(key, key_size);
937
938 b = __select_bucket(htab, hash);
939 head = &b->head;
940
941 /* bpf_map_update_elem() can be called in_irq() */
942 raw_spin_lock_irqsave(&b->lock, flags);
943
944 l_old = lookup_elem_raw(head, hash, key, key_size);
945
946 ret = check_flags(htab, l_old, map_flags);
947 if (ret)
948 goto err;
949
950 if (l_old) {
951 /* per-cpu hash map can update value in-place */
952 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
953 value, onallcpus);
954 } else {
955 l_new = alloc_htab_elem(htab, key, value, key_size,
956 hash, true, onallcpus, NULL);
957 if (IS_ERR(l_new)) {
958 ret = PTR_ERR(l_new);
959 goto err;
960 }
961 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
962 }
963 ret = 0;
964err:
965 raw_spin_unlock_irqrestore(&b->lock, flags);
966 return ret;
967}
968
969static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
970 void *value, u64 map_flags,
971 bool onallcpus)
972{
973 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
974 struct htab_elem *l_new = NULL, *l_old;
975 struct hlist_nulls_head *head;
976 unsigned long flags;
977 struct bucket *b;
978 u32 key_size, hash;
979 int ret;
980
981 if (unlikely(map_flags > BPF_EXIST))
982 /* unknown flags */
983 return -EINVAL;
984
985 WARN_ON_ONCE(!rcu_read_lock_held());
986
987 key_size = map->key_size;
988
989 hash = htab_map_hash(key, key_size);
990
991 b = __select_bucket(htab, hash);
992 head = &b->head;
993
994 /* For LRU, we need to alloc before taking bucket's
995 * spinlock because LRU's elem alloc may need
996 * to remove older elem from htab and this removal
997 * operation will need a bucket lock.
998 */
999 if (map_flags != BPF_EXIST) {
1000 l_new = prealloc_lru_pop(htab, key, hash);
1001 if (!l_new)
1002 return -ENOMEM;
1003 }
1004
1005 /* bpf_map_update_elem() can be called in_irq() */
1006 raw_spin_lock_irqsave(&b->lock, flags);
1007
1008 l_old = lookup_elem_raw(head, hash, key, key_size);
1009
1010 ret = check_flags(htab, l_old, map_flags);
1011 if (ret)
1012 goto err;
1013
1014 if (l_old) {
1015 bpf_lru_node_set_ref(&l_old->lru_node);
1016
1017 /* per-cpu hash map can update value in-place */
1018 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1019 value, onallcpus);
1020 } else {
1021 pcpu_copy_value(htab, htab_elem_get_ptr(l_new, key_size),
1022 value, onallcpus);
1023 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1024 l_new = NULL;
1025 }
1026 ret = 0;
1027err:
1028 raw_spin_unlock_irqrestore(&b->lock, flags);
1029 if (l_new)
1030 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1031 return ret;
1032}
1033
1034static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1035 void *value, u64 map_flags)
1036{
1037 return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1038}
1039
1040static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1041 void *value, u64 map_flags)
1042{
1043 return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1044 false);
1045}
1046
1047/* Called from syscall or from eBPF program */
1048static int htab_map_delete_elem(struct bpf_map *map, void *key)
1049{
1050 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1051 struct hlist_nulls_head *head;
1052 struct bucket *b;
1053 struct htab_elem *l;
1054 unsigned long flags;
1055 u32 hash, key_size;
1056 int ret = -ENOENT;
1057
1058 WARN_ON_ONCE(!rcu_read_lock_held());
1059
1060 key_size = map->key_size;
1061
1062 hash = htab_map_hash(key, key_size);
1063 b = __select_bucket(htab, hash);
1064 head = &b->head;
1065
1066 raw_spin_lock_irqsave(&b->lock, flags);
1067
1068 l = lookup_elem_raw(head, hash, key, key_size);
1069
1070 if (l) {
1071 hlist_nulls_del_rcu(&l->hash_node);
1072 free_htab_elem(htab, l);
1073 ret = 0;
1074 }
1075
1076 raw_spin_unlock_irqrestore(&b->lock, flags);
1077 return ret;
1078}
1079
1080static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1081{
1082 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1083 struct hlist_nulls_head *head;
1084 struct bucket *b;
1085 struct htab_elem *l;
1086 unsigned long flags;
1087 u32 hash, key_size;
1088 int ret = -ENOENT;
1089
1090 WARN_ON_ONCE(!rcu_read_lock_held());
1091
1092 key_size = map->key_size;
1093
1094 hash = htab_map_hash(key, key_size);
1095 b = __select_bucket(htab, hash);
1096 head = &b->head;
1097
1098 raw_spin_lock_irqsave(&b->lock, flags);
1099
1100 l = lookup_elem_raw(head, hash, key, key_size);
1101
1102 if (l) {
1103 hlist_nulls_del_rcu(&l->hash_node);
1104 ret = 0;
1105 }
1106
1107 raw_spin_unlock_irqrestore(&b->lock, flags);
1108 if (l)
1109 bpf_lru_push_free(&htab->lru, &l->lru_node);
1110 return ret;
1111}
1112
1113static void delete_all_elements(struct bpf_htab *htab)
1114{
1115 int i;
1116
1117 for (i = 0; i < htab->n_buckets; i++) {
1118 struct hlist_nulls_head *head = select_bucket(htab, i);
1119 struct hlist_nulls_node *n;
1120 struct htab_elem *l;
1121
1122 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1123 hlist_nulls_del_rcu(&l->hash_node);
1124 htab_elem_free(htab, l);
1125 }
1126 }
1127}
1128
1129/* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1130static void htab_map_free(struct bpf_map *map)
1131{
1132 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1133
1134 /* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
1135 * so the programs (can be more than one that used this map) were
1136 * disconnected from events. Wait for outstanding critical sections in
1137 * these programs to complete
1138 */
1139 synchronize_rcu();
1140
1141 /* some of free_htab_elem() callbacks for elements of this map may
1142 * not have executed. Wait for them.
1143 */
1144 rcu_barrier();
1145 if (!htab_is_prealloc(htab))
1146 delete_all_elements(htab);
1147 else
1148 prealloc_destroy(htab);
1149
1150 free_percpu(htab->extra_elems);
1151 bpf_map_area_free(htab->buckets);
1152 kfree(htab);
1153}
1154
1155const struct bpf_map_ops htab_map_ops = {
1156 .map_alloc_check = htab_map_alloc_check,
1157 .map_alloc = htab_map_alloc,
1158 .map_free = htab_map_free,
1159 .map_get_next_key = htab_map_get_next_key,
1160 .map_lookup_elem = htab_map_lookup_elem,
1161 .map_update_elem = htab_map_update_elem,
1162 .map_delete_elem = htab_map_delete_elem,
1163 .map_gen_lookup = htab_map_gen_lookup,
1164};
1165
1166const struct bpf_map_ops htab_lru_map_ops = {
1167 .map_alloc_check = htab_map_alloc_check,
1168 .map_alloc = htab_map_alloc,
1169 .map_free = htab_map_free,
1170 .map_get_next_key = htab_map_get_next_key,
1171 .map_lookup_elem = htab_lru_map_lookup_elem,
1172 .map_update_elem = htab_lru_map_update_elem,
1173 .map_delete_elem = htab_lru_map_delete_elem,
1174 .map_gen_lookup = htab_lru_map_gen_lookup,
1175};
1176
1177/* Called from eBPF program */
1178static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1179{
1180 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1181
1182 if (l)
1183 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1184 else
1185 return NULL;
1186}
1187
1188static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1189{
1190 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1191
1192 if (l) {
1193 bpf_lru_node_set_ref(&l->lru_node);
1194 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1195 }
1196
1197 return NULL;
1198}
1199
1200int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1201{
1202 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1203 struct htab_elem *l;
1204 void __percpu *pptr;
1205 int ret = -ENOENT;
1206 int cpu, off = 0;
1207 u32 size;
1208
1209 /* per_cpu areas are zero-filled and bpf programs can only
1210 * access 'value_size' of them, so copying rounded areas
1211 * will not leak any kernel data
1212 */
1213 size = round_up(map->value_size, 8);
1214 rcu_read_lock();
1215 l = __htab_map_lookup_elem(map, key);
1216 if (!l)
1217 goto out;
1218 if (htab_is_lru(htab))
1219 bpf_lru_node_set_ref(&l->lru_node);
1220 pptr = htab_elem_get_ptr(l, map->key_size);
1221 for_each_possible_cpu(cpu) {
1222 bpf_long_memcpy(value + off,
1223 per_cpu_ptr(pptr, cpu), size);
1224 off += size;
1225 }
1226 ret = 0;
1227out:
1228 rcu_read_unlock();
1229 return ret;
1230}
1231
1232int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1233 u64 map_flags)
1234{
1235 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1236 int ret;
1237
1238 rcu_read_lock();
1239 if (htab_is_lru(htab))
1240 ret = __htab_lru_percpu_map_update_elem(map, key, value,
1241 map_flags, true);
1242 else
1243 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1244 true);
1245 rcu_read_unlock();
1246
1247 return ret;
1248}
1249
1250const struct bpf_map_ops htab_percpu_map_ops = {
1251 .map_alloc_check = htab_map_alloc_check,
1252 .map_alloc = htab_map_alloc,
1253 .map_free = htab_map_free,
1254 .map_get_next_key = htab_map_get_next_key,
1255 .map_lookup_elem = htab_percpu_map_lookup_elem,
1256 .map_update_elem = htab_percpu_map_update_elem,
1257 .map_delete_elem = htab_map_delete_elem,
1258};
1259
1260const struct bpf_map_ops htab_lru_percpu_map_ops = {
1261 .map_alloc_check = htab_map_alloc_check,
1262 .map_alloc = htab_map_alloc,
1263 .map_free = htab_map_free,
1264 .map_get_next_key = htab_map_get_next_key,
1265 .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
1266 .map_update_elem = htab_lru_percpu_map_update_elem,
1267 .map_delete_elem = htab_lru_map_delete_elem,
1268};
1269
1270static int fd_htab_map_alloc_check(union bpf_attr *attr)
1271{
1272 if (attr->value_size != sizeof(u32))
1273 return -EINVAL;
1274 return htab_map_alloc_check(attr);
1275}
1276
1277static void fd_htab_map_free(struct bpf_map *map)
1278{
1279 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1280 struct hlist_nulls_node *n;
1281 struct hlist_nulls_head *head;
1282 struct htab_elem *l;
1283 int i;
1284
1285 for (i = 0; i < htab->n_buckets; i++) {
1286 head = select_bucket(htab, i);
1287
1288 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1289 void *ptr = fd_htab_map_get_ptr(map, l);
1290
1291 map->ops->map_fd_put_ptr(ptr);
1292 }
1293 }
1294
1295 htab_map_free(map);
1296}
1297
1298/* only called from syscall */
1299int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
1300{
1301 void **ptr;
1302 int ret = 0;
1303
1304 if (!map->ops->map_fd_sys_lookup_elem)
1305 return -ENOTSUPP;
1306
1307 rcu_read_lock();
1308 ptr = htab_map_lookup_elem(map, key);
1309 if (ptr)
1310 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
1311 else
1312 ret = -ENOENT;
1313 rcu_read_unlock();
1314
1315 return ret;
1316}
1317
1318/* only called from syscall */
1319int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
1320 void *key, void *value, u64 map_flags)
1321{
1322 void *ptr;
1323 int ret;
1324 u32 ufd = *(u32 *)value;
1325
1326 ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
1327 if (IS_ERR(ptr))
1328 return PTR_ERR(ptr);
1329
1330 ret = htab_map_update_elem(map, key, &ptr, map_flags);
1331 if (ret)
1332 map->ops->map_fd_put_ptr(ptr);
1333
1334 return ret;
1335}
1336
1337static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
1338{
1339 struct bpf_map *map, *inner_map_meta;
1340
1341 inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
1342 if (IS_ERR(inner_map_meta))
1343 return inner_map_meta;
1344
1345 map = htab_map_alloc(attr);
1346 if (IS_ERR(map)) {
1347 bpf_map_meta_free(inner_map_meta);
1348 return map;
1349 }
1350
1351 map->inner_map_meta = inner_map_meta;
1352
1353 return map;
1354}
1355
1356static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
1357{
1358 struct bpf_map **inner_map = htab_map_lookup_elem(map, key);
1359
1360 if (!inner_map)
1361 return NULL;
1362
1363 return READ_ONCE(*inner_map);
1364}
1365
1366static u32 htab_of_map_gen_lookup(struct bpf_map *map,
1367 struct bpf_insn *insn_buf)
1368{
1369 struct bpf_insn *insn = insn_buf;
1370 const int ret = BPF_REG_0;
1371
1372 *insn++ = BPF_EMIT_CALL((u64 (*)(u64, u64, u64, u64, u64))__htab_map_lookup_elem);
1373 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
1374 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
1375 offsetof(struct htab_elem, key) +
1376 round_up(map->key_size, 8));
1377 *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
1378
1379 return insn - insn_buf;
1380}
1381
1382static void htab_of_map_free(struct bpf_map *map)
1383{
1384 bpf_map_meta_free(map->inner_map_meta);
1385 fd_htab_map_free(map);
1386}
1387
1388const struct bpf_map_ops htab_of_maps_map_ops = {
1389 .map_alloc_check = fd_htab_map_alloc_check,
1390 .map_alloc = htab_of_map_alloc,
1391 .map_free = htab_of_map_free,
1392 .map_get_next_key = htab_map_get_next_key,
1393 .map_lookup_elem = htab_of_map_lookup_elem,
1394 .map_delete_elem = htab_map_delete_elem,
1395 .map_fd_get_ptr = bpf_map_fd_get_ptr,
1396 .map_fd_put_ptr = bpf_map_fd_put_ptr,
1397 .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
1398 .map_gen_lookup = htab_of_map_gen_lookup,
1399};