projects/linux·blob v4.13 · refs

1333 lines · 32.8 KiB

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
21struct bucket {
22 struct hlist_nulls_head head;
23 raw_spinlock_t lock;
24};
25
26struct bpf_htab {
27 struct bpf_map map;
28 struct bucket *buckets;
29 void *elems;
30 union {
31 struct pcpu_freelist freelist;
32 struct bpf_lru lru;
33 };
34 struct htab_elem *__percpu *extra_elems;
35 atomic_t count; /* number of elements in this hashtable */
36 u32 n_buckets; /* number of hash buckets */
37 u32 elem_size; /* size of each element in bytes */
38};
39
40/* each htab element is struct htab_elem + key + value */
41struct htab_elem {
42 union {
43 struct hlist_nulls_node hash_node;
44 struct {
45 void *padding;
46 union {
47 struct bpf_htab *htab;
48 struct pcpu_freelist_node fnode;
49 };
50 };
51 };
52 union {
53 struct rcu_head rcu;
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 bool htab_is_prealloc(const struct bpf_htab *htab)
75{
76 return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
77}
78
79static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
80 void __percpu *pptr)
81{
82 *(void __percpu **)(l->key + key_size) = pptr;
83}
84
85static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
86{
87 return *(void __percpu **)(l->key + key_size);
88}
89
90static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
91{
92 return *(void **)(l->key + roundup(map->key_size, 8));
93}
94
95static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
96{
97 return (struct htab_elem *) (htab->elems + i * htab->elem_size);
98}
99
100static void htab_free_elems(struct bpf_htab *htab)
101{
102 int i;
103
104 if (!htab_is_percpu(htab))
105 goto free_elems;
106
107 for (i = 0; i < htab->map.max_entries; i++) {
108 void __percpu *pptr;
109
110 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
111 htab->map.key_size);
112 free_percpu(pptr);
113 }
114free_elems:
115 bpf_map_area_free(htab->elems);
116}
117
118static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
119 u32 hash)
120{
121 struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
122 struct htab_elem *l;
123
124 if (node) {
125 l = container_of(node, struct htab_elem, lru_node);
126 memcpy(l->key, key, htab->map.key_size);
127 return l;
128 }
129
130 return NULL;
131}
132
133static int prealloc_init(struct bpf_htab *htab)
134{
135 u32 num_entries = htab->map.max_entries;
136 int err = -ENOMEM, i;
137
138 if (!htab_is_percpu(htab) && !htab_is_lru(htab))
139 num_entries += num_possible_cpus();
140
141 htab->elems = bpf_map_area_alloc(htab->elem_size * num_entries);
142 if (!htab->elems)
143 return -ENOMEM;
144
145 if (!htab_is_percpu(htab))
146 goto skip_percpu_elems;
147
148 for (i = 0; i < num_entries; i++) {
149 u32 size = round_up(htab->map.value_size, 8);
150 void __percpu *pptr;
151
152 pptr = __alloc_percpu_gfp(size, 8, GFP_USER | __GFP_NOWARN);
153 if (!pptr)
154 goto free_elems;
155 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
156 pptr);
157 }
158
159skip_percpu_elems:
160 if (htab_is_lru(htab))
161 err = bpf_lru_init(&htab->lru,
162 htab->map.map_flags & BPF_F_NO_COMMON_LRU,
163 offsetof(struct htab_elem, hash) -
164 offsetof(struct htab_elem, lru_node),
165 htab_lru_map_delete_node,
166 htab);
167 else
168 err = pcpu_freelist_init(&htab->freelist);
169
170 if (err)
171 goto free_elems;
172
173 if (htab_is_lru(htab))
174 bpf_lru_populate(&htab->lru, htab->elems,
175 offsetof(struct htab_elem, lru_node),
176 htab->elem_size, num_entries);
177 else
178 pcpu_freelist_populate(&htab->freelist,
179 htab->elems + offsetof(struct htab_elem, fnode),
180 htab->elem_size, num_entries);
181
182 return 0;
183
184free_elems:
185 htab_free_elems(htab);
186 return err;
187}
188
189static void prealloc_destroy(struct bpf_htab *htab)
190{
191 htab_free_elems(htab);
192
193 if (htab_is_lru(htab))
194 bpf_lru_destroy(&htab->lru);
195 else
196 pcpu_freelist_destroy(&htab->freelist);
197}
198
199static int alloc_extra_elems(struct bpf_htab *htab)
200{
201 struct htab_elem *__percpu *pptr, *l_new;
202 struct pcpu_freelist_node *l;
203 int cpu;
204
205 pptr = __alloc_percpu_gfp(sizeof(struct htab_elem *), 8,
206 GFP_USER | __GFP_NOWARN);
207 if (!pptr)
208 return -ENOMEM;
209
210 for_each_possible_cpu(cpu) {
211 l = pcpu_freelist_pop(&htab->freelist);
212 /* pop will succeed, since prealloc_init()
213 * preallocated extra num_possible_cpus elements
214 */
215 l_new = container_of(l, struct htab_elem, fnode);
216 *per_cpu_ptr(pptr, cpu) = l_new;
217 }
218 htab->extra_elems = pptr;
219 return 0;
220}
221
222/* Called from syscall */
223static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
224{
225 bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
226 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
227 bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
228 attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
229 /* percpu_lru means each cpu has its own LRU list.
230 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
231 * the map's value itself is percpu. percpu_lru has
232 * nothing to do with the map's value.
233 */
234 bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
235 bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
236 struct bpf_htab *htab;
237 int err, i;
238 u64 cost;
239
240 BUILD_BUG_ON(offsetof(struct htab_elem, htab) !=
241 offsetof(struct htab_elem, hash_node.pprev));
242 BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
243 offsetof(struct htab_elem, hash_node.pprev));
244
245 if (lru && !capable(CAP_SYS_ADMIN))
246 /* LRU implementation is much complicated than other
247 * maps. Hence, limit to CAP_SYS_ADMIN for now.
248 */
249 return ERR_PTR(-EPERM);
250
251 if (attr->map_flags & ~(BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU))
252 /* reserved bits should not be used */
253 return ERR_PTR(-EINVAL);
254
255 if (!lru && percpu_lru)
256 return ERR_PTR(-EINVAL);
257
258 if (lru && !prealloc)
259 return ERR_PTR(-ENOTSUPP);
260
261 htab = kzalloc(sizeof(*htab), GFP_USER);
262 if (!htab)
263 return ERR_PTR(-ENOMEM);
264
265 /* mandatory map attributes */
266 htab->map.map_type = attr->map_type;
267 htab->map.key_size = attr->key_size;
268 htab->map.value_size = attr->value_size;
269 htab->map.max_entries = attr->max_entries;
270 htab->map.map_flags = attr->map_flags;
271
272 /* check sanity of attributes.
273 * value_size == 0 may be allowed in the future to use map as a set
274 */
275 err = -EINVAL;
276 if (htab->map.max_entries == 0 || htab->map.key_size == 0 ||
277 htab->map.value_size == 0)
278 goto free_htab;
279
280 if (percpu_lru) {
281 /* ensure each CPU's lru list has >=1 elements.
282 * since we are at it, make each lru list has the same
283 * number of elements.
284 */
285 htab->map.max_entries = roundup(attr->max_entries,
286 num_possible_cpus());
287 if (htab->map.max_entries < attr->max_entries)
288 htab->map.max_entries = rounddown(attr->max_entries,
289 num_possible_cpus());
290 }
291
292 /* hash table size must be power of 2 */
293 htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
294
295 err = -E2BIG;
296 if (htab->map.key_size > MAX_BPF_STACK)
297 /* eBPF programs initialize keys on stack, so they cannot be
298 * larger than max stack size
299 */
300 goto free_htab;
301
302 if (htab->map.value_size >= KMALLOC_MAX_SIZE -
303 MAX_BPF_STACK - sizeof(struct htab_elem))
304 /* if value_size is bigger, the user space won't be able to
305 * access the elements via bpf syscall. This check also makes
306 * sure that the elem_size doesn't overflow and it's
307 * kmalloc-able later in htab_map_update_elem()
308 */
309 goto free_htab;
310
311 if (percpu && round_up(htab->map.value_size, 8) > PCPU_MIN_UNIT_SIZE)
312 /* make sure the size for pcpu_alloc() is reasonable */
313 goto free_htab;
314
315 htab->elem_size = sizeof(struct htab_elem) +
316 round_up(htab->map.key_size, 8);
317 if (percpu)
318 htab->elem_size += sizeof(void *);
319 else
320 htab->elem_size += round_up(htab->map.value_size, 8);
321
322 /* prevent zero size kmalloc and check for u32 overflow */
323 if (htab->n_buckets == 0 ||
324 htab->n_buckets > U32_MAX / sizeof(struct bucket))
325 goto free_htab;
326
327 cost = (u64) htab->n_buckets * sizeof(struct bucket) +
328 (u64) htab->elem_size * htab->map.max_entries;
329
330 if (percpu)
331 cost += (u64) round_up(htab->map.value_size, 8) *
332 num_possible_cpus() * htab->map.max_entries;
333 else
334 cost += (u64) htab->elem_size * num_possible_cpus();
335
336 if (cost >= U32_MAX - PAGE_SIZE)
337 /* make sure page count doesn't overflow */
338 goto free_htab;
339
340 htab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
341
342 /* if map size is larger than memlock limit, reject it early */
343 err = bpf_map_precharge_memlock(htab->map.pages);
344 if (err)
345 goto free_htab;
346
347 err = -ENOMEM;
348 htab->buckets = bpf_map_area_alloc(htab->n_buckets *
349 sizeof(struct bucket));
350 if (!htab->buckets)
351 goto free_htab;
352
353 for (i = 0; i < htab->n_buckets; i++) {
354 INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
355 raw_spin_lock_init(&htab->buckets[i].lock);
356 }
357
358 if (prealloc) {
359 err = prealloc_init(htab);
360 if (err)
361 goto free_buckets;
362
363 if (!percpu && !lru) {
364 /* lru itself can remove the least used element, so
365 * there is no need for an extra elem during map_update.
366 */
367 err = alloc_extra_elems(htab);
368 if (err)
369 goto free_prealloc;
370 }
371 }
372
373 return &htab->map;
374
375free_prealloc:
376 prealloc_destroy(htab);
377free_buckets:
378 bpf_map_area_free(htab->buckets);
379free_htab:
380 kfree(htab);
381 return ERR_PTR(err);
382}
383
384static inline u32 htab_map_hash(const void *key, u32 key_len)
385{
386 return jhash(key, key_len, 0);
387}
388
389static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
390{
391 return &htab->buckets[hash & (htab->n_buckets - 1)];
392}
393
394static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
395{
396 return &__select_bucket(htab, hash)->head;
397}
398
399/* this lookup function can only be called with bucket lock taken */
400static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
401 void *key, u32 key_size)
402{
403 struct hlist_nulls_node *n;
404 struct htab_elem *l;
405
406 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
407 if (l->hash == hash && !memcmp(&l->key, key, key_size))
408 return l;
409
410 return NULL;
411}
412
413/* can be called without bucket lock. it will repeat the loop in
414 * the unlikely event when elements moved from one bucket into another
415 * while link list is being walked
416 */
417static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
418 u32 hash, void *key,
419 u32 key_size, u32 n_buckets)
420{
421 struct hlist_nulls_node *n;
422 struct htab_elem *l;
423
424again:
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 if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
430 goto again;
431
432 return NULL;
433}
434
435/* Called from syscall or from eBPF program directly, so
436 * arguments have to match bpf_map_lookup_elem() exactly.
437 * The return value is adjusted by BPF instructions
438 * in htab_map_gen_lookup().
439 */
440static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
441{
442 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
443 struct hlist_nulls_head *head;
444 struct htab_elem *l;
445 u32 hash, key_size;
446
447 /* Must be called with rcu_read_lock. */
448 WARN_ON_ONCE(!rcu_read_lock_held());
449
450 key_size = map->key_size;
451
452 hash = htab_map_hash(key, key_size);
453
454 head = select_bucket(htab, hash);
455
456 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
457
458 return l;
459}
460
461static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
462{
463 struct htab_elem *l = __htab_map_lookup_elem(map, key);
464
465 if (l)
466 return l->key + round_up(map->key_size, 8);
467
468 return NULL;
469}
470
471/* inline bpf_map_lookup_elem() call.
472 * Instead of:
473 * bpf_prog
474 * bpf_map_lookup_elem
475 * map->ops->map_lookup_elem
476 * htab_map_lookup_elem
477 * __htab_map_lookup_elem
478 * do:
479 * bpf_prog
480 * __htab_map_lookup_elem
481 */
482static u32 htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
483{
484 struct bpf_insn *insn = insn_buf;
485 const int ret = BPF_REG_0;
486
487 *insn++ = BPF_EMIT_CALL((u64 (*)(u64, u64, u64, u64, u64))__htab_map_lookup_elem);
488 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
489 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
490 offsetof(struct htab_elem, key) +
491 round_up(map->key_size, 8));
492 return insn - insn_buf;
493}
494
495static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
496{
497 struct htab_elem *l = __htab_map_lookup_elem(map, key);
498
499 if (l) {
500 bpf_lru_node_set_ref(&l->lru_node);
501 return l->key + round_up(map->key_size, 8);
502 }
503
504 return NULL;
505}
506
507/* It is called from the bpf_lru_list when the LRU needs to delete
508 * older elements from the htab.
509 */
510static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
511{
512 struct bpf_htab *htab = (struct bpf_htab *)arg;
513 struct htab_elem *l = NULL, *tgt_l;
514 struct hlist_nulls_head *head;
515 struct hlist_nulls_node *n;
516 unsigned long flags;
517 struct bucket *b;
518
519 tgt_l = container_of(node, struct htab_elem, lru_node);
520 b = __select_bucket(htab, tgt_l->hash);
521 head = &b->head;
522
523 raw_spin_lock_irqsave(&b->lock, flags);
524
525 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
526 if (l == tgt_l) {
527 hlist_nulls_del_rcu(&l->hash_node);
528 break;
529 }
530
531 raw_spin_unlock_irqrestore(&b->lock, flags);
532
533 return l == tgt_l;
534}
535
536/* Called from syscall */
537static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
538{
539 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
540 struct hlist_nulls_head *head;
541 struct htab_elem *l, *next_l;
542 u32 hash, key_size;
543 int i = 0;
544
545 WARN_ON_ONCE(!rcu_read_lock_held());
546
547 key_size = map->key_size;
548
549 if (!key)
550 goto find_first_elem;
551
552 hash = htab_map_hash(key, key_size);
553
554 head = select_bucket(htab, hash);
555
556 /* lookup the key */
557 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
558
559 if (!l)
560 goto find_first_elem;
561
562 /* key was found, get next key in the same bucket */
563 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
564 struct htab_elem, hash_node);
565
566 if (next_l) {
567 /* if next elem in this hash list is non-zero, just return it */
568 memcpy(next_key, next_l->key, key_size);
569 return 0;
570 }
571
572 /* no more elements in this hash list, go to the next bucket */
573 i = hash & (htab->n_buckets - 1);
574 i++;
575
576find_first_elem:
577 /* iterate over buckets */
578 for (; i < htab->n_buckets; i++) {
579 head = select_bucket(htab, i);
580
581 /* pick first element in the bucket */
582 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
583 struct htab_elem, hash_node);
584 if (next_l) {
585 /* if it's not empty, just return it */
586 memcpy(next_key, next_l->key, key_size);
587 return 0;
588 }
589 }
590
591 /* iterated over all buckets and all elements */
592 return -ENOENT;
593}
594
595static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
596{
597 if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
598 free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
599 kfree(l);
600}
601
602static void htab_elem_free_rcu(struct rcu_head *head)
603{
604 struct htab_elem *l = container_of(head, struct htab_elem, rcu);
605 struct bpf_htab *htab = l->htab;
606
607 /* must increment bpf_prog_active to avoid kprobe+bpf triggering while
608 * we're calling kfree, otherwise deadlock is possible if kprobes
609 * are placed somewhere inside of slub
610 */
611 preempt_disable();
612 __this_cpu_inc(bpf_prog_active);
613 htab_elem_free(htab, l);
614 __this_cpu_dec(bpf_prog_active);
615 preempt_enable();
616}
617
618static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
619{
620 struct bpf_map *map = &htab->map;
621
622 if (map->ops->map_fd_put_ptr) {
623 void *ptr = fd_htab_map_get_ptr(map, l);
624
625 map->ops->map_fd_put_ptr(ptr);
626 }
627
628 if (htab_is_prealloc(htab)) {
629 pcpu_freelist_push(&htab->freelist, &l->fnode);
630 } else {
631 atomic_dec(&htab->count);
632 l->htab = htab;
633 call_rcu(&l->rcu, htab_elem_free_rcu);
634 }
635}
636
637static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
638 void *value, bool onallcpus)
639{
640 if (!onallcpus) {
641 /* copy true value_size bytes */
642 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
643 } else {
644 u32 size = round_up(htab->map.value_size, 8);
645 int off = 0, cpu;
646
647 for_each_possible_cpu(cpu) {
648 bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
649 value + off, size);
650 off += size;
651 }
652 }
653}
654
655static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
656{
657 return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
658 BITS_PER_LONG == 64;
659}
660
661static u32 htab_size_value(const struct bpf_htab *htab, bool percpu)
662{
663 u32 size = htab->map.value_size;
664
665 if (percpu || fd_htab_map_needs_adjust(htab))
666 size = round_up(size, 8);
667 return size;
668}
669
670static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
671 void *value, u32 key_size, u32 hash,
672 bool percpu, bool onallcpus,
673 struct htab_elem *old_elem)
674{
675 u32 size = htab_size_value(htab, percpu);
676 bool prealloc = htab_is_prealloc(htab);
677 struct htab_elem *l_new, **pl_new;
678 void __percpu *pptr;
679
680 if (prealloc) {
681 if (old_elem) {
682 /* if we're updating the existing element,
683 * use per-cpu extra elems to avoid freelist_pop/push
684 */
685 pl_new = this_cpu_ptr(htab->extra_elems);
686 l_new = *pl_new;
687 *pl_new = old_elem;
688 } else {
689 struct pcpu_freelist_node *l;
690
691 l = pcpu_freelist_pop(&htab->freelist);
692 if (!l)
693 return ERR_PTR(-E2BIG);
694 l_new = container_of(l, struct htab_elem, fnode);
695 }
696 } else {
697 if (atomic_inc_return(&htab->count) > htab->map.max_entries)
698 if (!old_elem) {
699 /* when map is full and update() is replacing
700 * old element, it's ok to allocate, since
701 * old element will be freed immediately.
702 * Otherwise return an error
703 */
704 atomic_dec(&htab->count);
705 return ERR_PTR(-E2BIG);
706 }
707 l_new = kmalloc(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN);
708 if (!l_new)
709 return ERR_PTR(-ENOMEM);
710 }
711
712 memcpy(l_new->key, key, key_size);
713 if (percpu) {
714 if (prealloc) {
715 pptr = htab_elem_get_ptr(l_new, key_size);
716 } else {
717 /* alloc_percpu zero-fills */
718 pptr = __alloc_percpu_gfp(size, 8,
719 GFP_ATOMIC | __GFP_NOWARN);
720 if (!pptr) {
721 kfree(l_new);
722 return ERR_PTR(-ENOMEM);
723 }
724 }
725
726 pcpu_copy_value(htab, pptr, value, onallcpus);
727
728 if (!prealloc)
729 htab_elem_set_ptr(l_new, key_size, pptr);
730 } else {
731 memcpy(l_new->key + round_up(key_size, 8), value, size);
732 }
733
734 l_new->hash = hash;
735 return l_new;
736}
737
738static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
739 u64 map_flags)
740{
741 if (l_old && map_flags == BPF_NOEXIST)
742 /* elem already exists */
743 return -EEXIST;
744
745 if (!l_old && map_flags == BPF_EXIST)
746 /* elem doesn't exist, cannot update it */
747 return -ENOENT;
748
749 return 0;
750}
751
752/* Called from syscall or from eBPF program */
753static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
754 u64 map_flags)
755{
756 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
757 struct htab_elem *l_new = NULL, *l_old;
758 struct hlist_nulls_head *head;
759 unsigned long flags;
760 struct bucket *b;
761 u32 key_size, hash;
762 int ret;
763
764 if (unlikely(map_flags > BPF_EXIST))
765 /* unknown flags */
766 return -EINVAL;
767
768 WARN_ON_ONCE(!rcu_read_lock_held());
769
770 key_size = map->key_size;
771
772 hash = htab_map_hash(key, key_size);
773
774 b = __select_bucket(htab, hash);
775 head = &b->head;
776
777 /* bpf_map_update_elem() can be called in_irq() */
778 raw_spin_lock_irqsave(&b->lock, flags);
779
780 l_old = lookup_elem_raw(head, hash, key, key_size);
781
782 ret = check_flags(htab, l_old, map_flags);
783 if (ret)
784 goto err;
785
786 l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
787 l_old);
788 if (IS_ERR(l_new)) {
789 /* all pre-allocated elements are in use or memory exhausted */
790 ret = PTR_ERR(l_new);
791 goto err;
792 }
793
794 /* add new element to the head of the list, so that
795 * concurrent search will find it before old elem
796 */
797 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
798 if (l_old) {
799 hlist_nulls_del_rcu(&l_old->hash_node);
800 if (!htab_is_prealloc(htab))
801 free_htab_elem(htab, l_old);
802 }
803 ret = 0;
804err:
805 raw_spin_unlock_irqrestore(&b->lock, flags);
806 return ret;
807}
808
809static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
810 u64 map_flags)
811{
812 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
813 struct htab_elem *l_new, *l_old = NULL;
814 struct hlist_nulls_head *head;
815 unsigned long flags;
816 struct bucket *b;
817 u32 key_size, hash;
818 int ret;
819
820 if (unlikely(map_flags > BPF_EXIST))
821 /* unknown flags */
822 return -EINVAL;
823
824 WARN_ON_ONCE(!rcu_read_lock_held());
825
826 key_size = map->key_size;
827
828 hash = htab_map_hash(key, key_size);
829
830 b = __select_bucket(htab, hash);
831 head = &b->head;
832
833 /* For LRU, we need to alloc before taking bucket's
834 * spinlock because getting free nodes from LRU may need
835 * to remove older elements from htab and this removal
836 * operation will need a bucket lock.
837 */
838 l_new = prealloc_lru_pop(htab, key, hash);
839 if (!l_new)
840 return -ENOMEM;
841 memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
842
843 /* bpf_map_update_elem() can be called in_irq() */
844 raw_spin_lock_irqsave(&b->lock, flags);
845
846 l_old = lookup_elem_raw(head, hash, key, key_size);
847
848 ret = check_flags(htab, l_old, map_flags);
849 if (ret)
850 goto err;
851
852 /* add new element to the head of the list, so that
853 * concurrent search will find it before old elem
854 */
855 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
856 if (l_old) {
857 bpf_lru_node_set_ref(&l_new->lru_node);
858 hlist_nulls_del_rcu(&l_old->hash_node);
859 }
860 ret = 0;
861
862err:
863 raw_spin_unlock_irqrestore(&b->lock, flags);
864
865 if (ret)
866 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
867 else if (l_old)
868 bpf_lru_push_free(&htab->lru, &l_old->lru_node);
869
870 return ret;
871}
872
873static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
874 void *value, u64 map_flags,
875 bool onallcpus)
876{
877 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
878 struct htab_elem *l_new = NULL, *l_old;
879 struct hlist_nulls_head *head;
880 unsigned long flags;
881 struct bucket *b;
882 u32 key_size, hash;
883 int ret;
884
885 if (unlikely(map_flags > BPF_EXIST))
886 /* unknown flags */
887 return -EINVAL;
888
889 WARN_ON_ONCE(!rcu_read_lock_held());
890
891 key_size = map->key_size;
892
893 hash = htab_map_hash(key, key_size);
894
895 b = __select_bucket(htab, hash);
896 head = &b->head;
897
898 /* bpf_map_update_elem() can be called in_irq() */
899 raw_spin_lock_irqsave(&b->lock, flags);
900
901 l_old = lookup_elem_raw(head, hash, key, key_size);
902
903 ret = check_flags(htab, l_old, map_flags);
904 if (ret)
905 goto err;
906
907 if (l_old) {
908 /* per-cpu hash map can update value in-place */
909 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
910 value, onallcpus);
911 } else {
912 l_new = alloc_htab_elem(htab, key, value, key_size,
913 hash, true, onallcpus, NULL);
914 if (IS_ERR(l_new)) {
915 ret = PTR_ERR(l_new);
916 goto err;
917 }
918 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
919 }
920 ret = 0;
921err:
922 raw_spin_unlock_irqrestore(&b->lock, flags);
923 return ret;
924}
925
926static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
927 void *value, u64 map_flags,
928 bool onallcpus)
929{
930 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
931 struct htab_elem *l_new = NULL, *l_old;
932 struct hlist_nulls_head *head;
933 unsigned long flags;
934 struct bucket *b;
935 u32 key_size, hash;
936 int ret;
937
938 if (unlikely(map_flags > BPF_EXIST))
939 /* unknown flags */
940 return -EINVAL;
941
942 WARN_ON_ONCE(!rcu_read_lock_held());
943
944 key_size = map->key_size;
945
946 hash = htab_map_hash(key, key_size);
947
948 b = __select_bucket(htab, hash);
949 head = &b->head;
950
951 /* For LRU, we need to alloc before taking bucket's
952 * spinlock because LRU's elem alloc may need
953 * to remove older elem from htab and this removal
954 * operation will need a bucket lock.
955 */
956 if (map_flags != BPF_EXIST) {
957 l_new = prealloc_lru_pop(htab, key, hash);
958 if (!l_new)
959 return -ENOMEM;
960 }
961
962 /* bpf_map_update_elem() can be called in_irq() */
963 raw_spin_lock_irqsave(&b->lock, flags);
964
965 l_old = lookup_elem_raw(head, hash, key, key_size);
966
967 ret = check_flags(htab, l_old, map_flags);
968 if (ret)
969 goto err;
970
971 if (l_old) {
972 bpf_lru_node_set_ref(&l_old->lru_node);
973
974 /* per-cpu hash map can update value in-place */
975 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
976 value, onallcpus);
977 } else {
978 pcpu_copy_value(htab, htab_elem_get_ptr(l_new, key_size),
979 value, onallcpus);
980 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
981 l_new = NULL;
982 }
983 ret = 0;
984err:
985 raw_spin_unlock_irqrestore(&b->lock, flags);
986 if (l_new)
987 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
988 return ret;
989}
990
991static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
992 void *value, u64 map_flags)
993{
994 return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
995}
996
997static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
998 void *value, u64 map_flags)
999{
1000 return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1001 false);
1002}
1003
1004/* Called from syscall or from eBPF program */
1005static int htab_map_delete_elem(struct bpf_map *map, void *key)
1006{
1007 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1008 struct hlist_nulls_head *head;
1009 struct bucket *b;
1010 struct htab_elem *l;
1011 unsigned long flags;
1012 u32 hash, key_size;
1013 int ret = -ENOENT;
1014
1015 WARN_ON_ONCE(!rcu_read_lock_held());
1016
1017 key_size = map->key_size;
1018
1019 hash = htab_map_hash(key, key_size);
1020 b = __select_bucket(htab, hash);
1021 head = &b->head;
1022
1023 raw_spin_lock_irqsave(&b->lock, flags);
1024
1025 l = lookup_elem_raw(head, hash, key, key_size);
1026
1027 if (l) {
1028 hlist_nulls_del_rcu(&l->hash_node);
1029 free_htab_elem(htab, l);
1030 ret = 0;
1031 }
1032
1033 raw_spin_unlock_irqrestore(&b->lock, flags);
1034 return ret;
1035}
1036
1037static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1038{
1039 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1040 struct hlist_nulls_head *head;
1041 struct bucket *b;
1042 struct htab_elem *l;
1043 unsigned long flags;
1044 u32 hash, key_size;
1045 int ret = -ENOENT;
1046
1047 WARN_ON_ONCE(!rcu_read_lock_held());
1048
1049 key_size = map->key_size;
1050
1051 hash = htab_map_hash(key, key_size);
1052 b = __select_bucket(htab, hash);
1053 head = &b->head;
1054
1055 raw_spin_lock_irqsave(&b->lock, flags);
1056
1057 l = lookup_elem_raw(head, hash, key, key_size);
1058
1059 if (l) {
1060 hlist_nulls_del_rcu(&l->hash_node);
1061 ret = 0;
1062 }
1063
1064 raw_spin_unlock_irqrestore(&b->lock, flags);
1065 if (l)
1066 bpf_lru_push_free(&htab->lru, &l->lru_node);
1067 return ret;
1068}
1069
1070static void delete_all_elements(struct bpf_htab *htab)
1071{
1072 int i;
1073
1074 for (i = 0; i < htab->n_buckets; i++) {
1075 struct hlist_nulls_head *head = select_bucket(htab, i);
1076 struct hlist_nulls_node *n;
1077 struct htab_elem *l;
1078
1079 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1080 hlist_nulls_del_rcu(&l->hash_node);
1081 htab_elem_free(htab, l);
1082 }
1083 }
1084}
1085
1086/* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1087static void htab_map_free(struct bpf_map *map)
1088{
1089 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1090
1091 /* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
1092 * so the programs (can be more than one that used this map) were
1093 * disconnected from events. Wait for outstanding critical sections in
1094 * these programs to complete
1095 */
1096 synchronize_rcu();
1097
1098 /* some of free_htab_elem() callbacks for elements of this map may
1099 * not have executed. Wait for them.
1100 */
1101 rcu_barrier();
1102 if (!htab_is_prealloc(htab))
1103 delete_all_elements(htab);
1104 else
1105 prealloc_destroy(htab);
1106
1107 free_percpu(htab->extra_elems);
1108 bpf_map_area_free(htab->buckets);
1109 kfree(htab);
1110}
1111
1112const struct bpf_map_ops htab_map_ops = {
1113 .map_alloc = htab_map_alloc,
1114 .map_free = htab_map_free,
1115 .map_get_next_key = htab_map_get_next_key,
1116 .map_lookup_elem = htab_map_lookup_elem,
1117 .map_update_elem = htab_map_update_elem,
1118 .map_delete_elem = htab_map_delete_elem,
1119 .map_gen_lookup = htab_map_gen_lookup,
1120};
1121
1122const struct bpf_map_ops htab_lru_map_ops = {
1123 .map_alloc = htab_map_alloc,
1124 .map_free = htab_map_free,
1125 .map_get_next_key = htab_map_get_next_key,
1126 .map_lookup_elem = htab_lru_map_lookup_elem,
1127 .map_update_elem = htab_lru_map_update_elem,
1128 .map_delete_elem = htab_lru_map_delete_elem,
1129};
1130
1131/* Called from eBPF program */
1132static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1133{
1134 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1135
1136 if (l)
1137 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1138 else
1139 return NULL;
1140}
1141
1142static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1143{
1144 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1145
1146 if (l) {
1147 bpf_lru_node_set_ref(&l->lru_node);
1148 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1149 }
1150
1151 return NULL;
1152}
1153
1154int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1155{
1156 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1157 struct htab_elem *l;
1158 void __percpu *pptr;
1159 int ret = -ENOENT;
1160 int cpu, off = 0;
1161 u32 size;
1162
1163 /* per_cpu areas are zero-filled and bpf programs can only
1164 * access 'value_size' of them, so copying rounded areas
1165 * will not leak any kernel data
1166 */
1167 size = round_up(map->value_size, 8);
1168 rcu_read_lock();
1169 l = __htab_map_lookup_elem(map, key);
1170 if (!l)
1171 goto out;
1172 if (htab_is_lru(htab))
1173 bpf_lru_node_set_ref(&l->lru_node);
1174 pptr = htab_elem_get_ptr(l, map->key_size);
1175 for_each_possible_cpu(cpu) {
1176 bpf_long_memcpy(value + off,
1177 per_cpu_ptr(pptr, cpu), size);
1178 off += size;
1179 }
1180 ret = 0;
1181out:
1182 rcu_read_unlock();
1183 return ret;
1184}
1185
1186int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1187 u64 map_flags)
1188{
1189 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1190 int ret;
1191
1192 rcu_read_lock();
1193 if (htab_is_lru(htab))
1194 ret = __htab_lru_percpu_map_update_elem(map, key, value,
1195 map_flags, true);
1196 else
1197 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1198 true);
1199 rcu_read_unlock();
1200
1201 return ret;
1202}
1203
1204const struct bpf_map_ops htab_percpu_map_ops = {
1205 .map_alloc = htab_map_alloc,
1206 .map_free = htab_map_free,
1207 .map_get_next_key = htab_map_get_next_key,
1208 .map_lookup_elem = htab_percpu_map_lookup_elem,
1209 .map_update_elem = htab_percpu_map_update_elem,
1210 .map_delete_elem = htab_map_delete_elem,
1211};
1212
1213const struct bpf_map_ops htab_lru_percpu_map_ops = {
1214 .map_alloc = htab_map_alloc,
1215 .map_free = htab_map_free,
1216 .map_get_next_key = htab_map_get_next_key,
1217 .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
1218 .map_update_elem = htab_lru_percpu_map_update_elem,
1219 .map_delete_elem = htab_lru_map_delete_elem,
1220};
1221
1222static struct bpf_map *fd_htab_map_alloc(union bpf_attr *attr)
1223{
1224 if (attr->value_size != sizeof(u32))
1225 return ERR_PTR(-EINVAL);
1226 return htab_map_alloc(attr);
1227}
1228
1229static void fd_htab_map_free(struct bpf_map *map)
1230{
1231 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1232 struct hlist_nulls_node *n;
1233 struct hlist_nulls_head *head;
1234 struct htab_elem *l;
1235 int i;
1236
1237 for (i = 0; i < htab->n_buckets; i++) {
1238 head = select_bucket(htab, i);
1239
1240 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1241 void *ptr = fd_htab_map_get_ptr(map, l);
1242
1243 map->ops->map_fd_put_ptr(ptr);
1244 }
1245 }
1246
1247 htab_map_free(map);
1248}
1249
1250/* only called from syscall */
1251int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
1252{
1253 void **ptr;
1254 int ret = 0;
1255
1256 if (!map->ops->map_fd_sys_lookup_elem)
1257 return -ENOTSUPP;
1258
1259 rcu_read_lock();
1260 ptr = htab_map_lookup_elem(map, key);
1261 if (ptr)
1262 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
1263 else
1264 ret = -ENOENT;
1265 rcu_read_unlock();
1266
1267 return ret;
1268}
1269
1270/* only called from syscall */
1271int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
1272 void *key, void *value, u64 map_flags)
1273{
1274 void *ptr;
1275 int ret;
1276 u32 ufd = *(u32 *)value;
1277
1278 ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
1279 if (IS_ERR(ptr))
1280 return PTR_ERR(ptr);
1281
1282 ret = htab_map_update_elem(map, key, &ptr, map_flags);
1283 if (ret)
1284 map->ops->map_fd_put_ptr(ptr);
1285
1286 return ret;
1287}
1288
1289static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
1290{
1291 struct bpf_map *map, *inner_map_meta;
1292
1293 inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
1294 if (IS_ERR(inner_map_meta))
1295 return inner_map_meta;
1296
1297 map = fd_htab_map_alloc(attr);
1298 if (IS_ERR(map)) {
1299 bpf_map_meta_free(inner_map_meta);
1300 return map;
1301 }
1302
1303 map->inner_map_meta = inner_map_meta;
1304
1305 return map;
1306}
1307
1308static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
1309{
1310 struct bpf_map **inner_map = htab_map_lookup_elem(map, key);
1311
1312 if (!inner_map)
1313 return NULL;
1314
1315 return READ_ONCE(*inner_map);
1316}
1317
1318static void htab_of_map_free(struct bpf_map *map)
1319{
1320 bpf_map_meta_free(map->inner_map_meta);
1321 fd_htab_map_free(map);
1322}
1323
1324const struct bpf_map_ops htab_of_maps_map_ops = {
1325 .map_alloc = htab_of_map_alloc,
1326 .map_free = htab_of_map_free,
1327 .map_get_next_key = htab_map_get_next_key,
1328 .map_lookup_elem = htab_of_map_lookup_elem,
1329 .map_delete_elem = htab_map_delete_elem,
1330 .map_fd_get_ptr = bpf_map_fd_get_ptr,
1331 .map_fd_put_ptr = bpf_map_fd_put_ptr,
1332 .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
1333};