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 <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};
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/vmalloc.h>
17#include "percpu_freelist.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 struct pcpu_freelist freelist;
29 atomic_t count; /* number of elements in this hashtable */
30 u32 n_buckets; /* number of hash buckets */
31 u32 elem_size; /* size of each element in bytes */
32};
33
34/* each htab element is struct htab_elem + key + value */
35struct htab_elem {
36 union {
37 struct hlist_node hash_node;
38 struct bpf_htab *htab;
39 struct pcpu_freelist_node fnode;
40 };
41 struct rcu_head rcu;
42 u32 hash;
43 char key[0] __aligned(8);
44};
45
46static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
47 void __percpu *pptr)
48{
49 *(void __percpu **)(l->key + key_size) = pptr;
50}
51
52static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
53{
54 return *(void __percpu **)(l->key + key_size);
55}
56
57static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
58{
59 return (struct htab_elem *) (htab->elems + i * htab->elem_size);
60}
61
62static void htab_free_elems(struct bpf_htab *htab)
63{
64 int i;
65
66 if (htab->map.map_type != BPF_MAP_TYPE_PERCPU_HASH)
67 goto free_elems;
68
69 for (i = 0; i < htab->map.max_entries; i++) {
70 void __percpu *pptr;
71
72 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
73 htab->map.key_size);
74 free_percpu(pptr);
75 }
76free_elems:
77 vfree(htab->elems);
78}
79
80static int prealloc_elems_and_freelist(struct bpf_htab *htab)
81{
82 int err = -ENOMEM, i;
83
84 htab->elems = vzalloc(htab->elem_size * htab->map.max_entries);
85 if (!htab->elems)
86 return -ENOMEM;
87
88 if (htab->map.map_type != BPF_MAP_TYPE_PERCPU_HASH)
89 goto skip_percpu_elems;
90
91 for (i = 0; i < htab->map.max_entries; i++) {
92 u32 size = round_up(htab->map.value_size, 8);
93 void __percpu *pptr;
94
95 pptr = __alloc_percpu_gfp(size, 8, GFP_USER | __GFP_NOWARN);
96 if (!pptr)
97 goto free_elems;
98 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
99 pptr);
100 }
101
102skip_percpu_elems:
103 err = pcpu_freelist_init(&htab->freelist);
104 if (err)
105 goto free_elems;
106
107 pcpu_freelist_populate(&htab->freelist, htab->elems, htab->elem_size,
108 htab->map.max_entries);
109 return 0;
110
111free_elems:
112 htab_free_elems(htab);
113 return err;
114}
115
116/* Called from syscall */
117static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
118{
119 bool percpu = attr->map_type == BPF_MAP_TYPE_PERCPU_HASH;
120 struct bpf_htab *htab;
121 int err, i;
122 u64 cost;
123
124 if (attr->map_flags & ~BPF_F_NO_PREALLOC)
125 /* reserved bits should not be used */
126 return ERR_PTR(-EINVAL);
127
128 htab = kzalloc(sizeof(*htab), GFP_USER);
129 if (!htab)
130 return ERR_PTR(-ENOMEM);
131
132 /* mandatory map attributes */
133 htab->map.map_type = attr->map_type;
134 htab->map.key_size = attr->key_size;
135 htab->map.value_size = attr->value_size;
136 htab->map.max_entries = attr->max_entries;
137 htab->map.map_flags = attr->map_flags;
138
139 /* check sanity of attributes.
140 * value_size == 0 may be allowed in the future to use map as a set
141 */
142 err = -EINVAL;
143 if (htab->map.max_entries == 0 || htab->map.key_size == 0 ||
144 htab->map.value_size == 0)
145 goto free_htab;
146
147 /* hash table size must be power of 2 */
148 htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
149
150 err = -E2BIG;
151 if (htab->map.key_size > MAX_BPF_STACK)
152 /* eBPF programs initialize keys on stack, so they cannot be
153 * larger than max stack size
154 */
155 goto free_htab;
156
157 if (htab->map.value_size >= (1 << (KMALLOC_SHIFT_MAX - 1)) -
158 MAX_BPF_STACK - sizeof(struct htab_elem))
159 /* if value_size is bigger, the user space won't be able to
160 * access the elements via bpf syscall. This check also makes
161 * sure that the elem_size doesn't overflow and it's
162 * kmalloc-able later in htab_map_update_elem()
163 */
164 goto free_htab;
165
166 if (percpu && round_up(htab->map.value_size, 8) > PCPU_MIN_UNIT_SIZE)
167 /* make sure the size for pcpu_alloc() is reasonable */
168 goto free_htab;
169
170 htab->elem_size = sizeof(struct htab_elem) +
171 round_up(htab->map.key_size, 8);
172 if (percpu)
173 htab->elem_size += sizeof(void *);
174 else
175 htab->elem_size += round_up(htab->map.value_size, 8);
176
177 /* prevent zero size kmalloc and check for u32 overflow */
178 if (htab->n_buckets == 0 ||
179 htab->n_buckets > U32_MAX / sizeof(struct bucket))
180 goto free_htab;
181
182 cost = (u64) htab->n_buckets * sizeof(struct bucket) +
183 (u64) htab->elem_size * htab->map.max_entries;
184
185 if (percpu)
186 cost += (u64) round_up(htab->map.value_size, 8) *
187 num_possible_cpus() * htab->map.max_entries;
188
189 if (cost >= U32_MAX - PAGE_SIZE)
190 /* make sure page count doesn't overflow */
191 goto free_htab;
192
193 htab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
194
195 /* if map size is larger than memlock limit, reject it early */
196 err = bpf_map_precharge_memlock(htab->map.pages);
197 if (err)
198 goto free_htab;
199
200 err = -ENOMEM;
201 htab->buckets = kmalloc_array(htab->n_buckets, sizeof(struct bucket),
202 GFP_USER | __GFP_NOWARN);
203
204 if (!htab->buckets) {
205 htab->buckets = vmalloc(htab->n_buckets * sizeof(struct bucket));
206 if (!htab->buckets)
207 goto free_htab;
208 }
209
210 for (i = 0; i < htab->n_buckets; i++) {
211 INIT_HLIST_HEAD(&htab->buckets[i].head);
212 raw_spin_lock_init(&htab->buckets[i].lock);
213 }
214
215 if (!(attr->map_flags & BPF_F_NO_PREALLOC)) {
216 err = prealloc_elems_and_freelist(htab);
217 if (err)
218 goto free_buckets;
219 }
220
221 return &htab->map;
222
223free_buckets:
224 kvfree(htab->buckets);
225free_htab:
226 kfree(htab);
227 return ERR_PTR(err);
228}
229
230static inline u32 htab_map_hash(const void *key, u32 key_len)
231{
232 return jhash(key, key_len, 0);
233}
234
235static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
236{
237 return &htab->buckets[hash & (htab->n_buckets - 1)];
238}
239
240static inline struct hlist_head *select_bucket(struct bpf_htab *htab, u32 hash)
241{
242 return &__select_bucket(htab, hash)->head;
243}
244
245static struct htab_elem *lookup_elem_raw(struct hlist_head *head, u32 hash,
246 void *key, u32 key_size)
247{
248 struct htab_elem *l;
249
250 hlist_for_each_entry_rcu(l, head, hash_node)
251 if (l->hash == hash && !memcmp(&l->key, key, key_size))
252 return l;
253
254 return NULL;
255}
256
257/* Called from syscall or from eBPF program */
258static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
259{
260 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
261 struct hlist_head *head;
262 struct htab_elem *l;
263 u32 hash, key_size;
264
265 /* Must be called with rcu_read_lock. */
266 WARN_ON_ONCE(!rcu_read_lock_held());
267
268 key_size = map->key_size;
269
270 hash = htab_map_hash(key, key_size);
271
272 head = select_bucket(htab, hash);
273
274 l = lookup_elem_raw(head, hash, key, key_size);
275
276 return l;
277}
278
279static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
280{
281 struct htab_elem *l = __htab_map_lookup_elem(map, key);
282
283 if (l)
284 return l->key + round_up(map->key_size, 8);
285
286 return NULL;
287}
288
289/* Called from syscall */
290static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
291{
292 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
293 struct hlist_head *head;
294 struct htab_elem *l, *next_l;
295 u32 hash, key_size;
296 int i;
297
298 WARN_ON_ONCE(!rcu_read_lock_held());
299
300 key_size = map->key_size;
301
302 hash = htab_map_hash(key, key_size);
303
304 head = select_bucket(htab, hash);
305
306 /* lookup the key */
307 l = lookup_elem_raw(head, hash, key, key_size);
308
309 if (!l) {
310 i = 0;
311 goto find_first_elem;
312 }
313
314 /* key was found, get next key in the same bucket */
315 next_l = hlist_entry_safe(rcu_dereference_raw(hlist_next_rcu(&l->hash_node)),
316 struct htab_elem, hash_node);
317
318 if (next_l) {
319 /* if next elem in this hash list is non-zero, just return it */
320 memcpy(next_key, next_l->key, key_size);
321 return 0;
322 }
323
324 /* no more elements in this hash list, go to the next bucket */
325 i = hash & (htab->n_buckets - 1);
326 i++;
327
328find_first_elem:
329 /* iterate over buckets */
330 for (; i < htab->n_buckets; i++) {
331 head = select_bucket(htab, i);
332
333 /* pick first element in the bucket */
334 next_l = hlist_entry_safe(rcu_dereference_raw(hlist_first_rcu(head)),
335 struct htab_elem, hash_node);
336 if (next_l) {
337 /* if it's not empty, just return it */
338 memcpy(next_key, next_l->key, key_size);
339 return 0;
340 }
341 }
342
343 /* iterated over all buckets and all elements */
344 return -ENOENT;
345}
346
347static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
348{
349 if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
350 free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
351 kfree(l);
352
353}
354
355static void htab_elem_free_rcu(struct rcu_head *head)
356{
357 struct htab_elem *l = container_of(head, struct htab_elem, rcu);
358 struct bpf_htab *htab = l->htab;
359
360 /* must increment bpf_prog_active to avoid kprobe+bpf triggering while
361 * we're calling kfree, otherwise deadlock is possible if kprobes
362 * are placed somewhere inside of slub
363 */
364 preempt_disable();
365 __this_cpu_inc(bpf_prog_active);
366 htab_elem_free(htab, l);
367 __this_cpu_dec(bpf_prog_active);
368 preempt_enable();
369}
370
371static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
372{
373 if (!(htab->map.map_flags & BPF_F_NO_PREALLOC)) {
374 pcpu_freelist_push(&htab->freelist, &l->fnode);
375 } else {
376 atomic_dec(&htab->count);
377 l->htab = htab;
378 call_rcu(&l->rcu, htab_elem_free_rcu);
379 }
380}
381
382static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
383 void *value, u32 key_size, u32 hash,
384 bool percpu, bool onallcpus)
385{
386 u32 size = htab->map.value_size;
387 bool prealloc = !(htab->map.map_flags & BPF_F_NO_PREALLOC);
388 struct htab_elem *l_new;
389 void __percpu *pptr;
390
391 if (prealloc) {
392 l_new = (struct htab_elem *)pcpu_freelist_pop(&htab->freelist);
393 if (!l_new)
394 return ERR_PTR(-E2BIG);
395 } else {
396 if (atomic_inc_return(&htab->count) > htab->map.max_entries) {
397 atomic_dec(&htab->count);
398 return ERR_PTR(-E2BIG);
399 }
400 l_new = kmalloc(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN);
401 if (!l_new)
402 return ERR_PTR(-ENOMEM);
403 }
404
405 memcpy(l_new->key, key, key_size);
406 if (percpu) {
407 /* round up value_size to 8 bytes */
408 size = round_up(size, 8);
409
410 if (prealloc) {
411 pptr = htab_elem_get_ptr(l_new, key_size);
412 } else {
413 /* alloc_percpu zero-fills */
414 pptr = __alloc_percpu_gfp(size, 8,
415 GFP_ATOMIC | __GFP_NOWARN);
416 if (!pptr) {
417 kfree(l_new);
418 return ERR_PTR(-ENOMEM);
419 }
420 }
421
422 if (!onallcpus) {
423 /* copy true value_size bytes */
424 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
425 } else {
426 int off = 0, cpu;
427
428 for_each_possible_cpu(cpu) {
429 bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
430 value + off, size);
431 off += size;
432 }
433 }
434 if (!prealloc)
435 htab_elem_set_ptr(l_new, key_size, pptr);
436 } else {
437 memcpy(l_new->key + round_up(key_size, 8), value, size);
438 }
439
440 l_new->hash = hash;
441 return l_new;
442}
443
444static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
445 u64 map_flags)
446{
447 if (l_old && map_flags == BPF_NOEXIST)
448 /* elem already exists */
449 return -EEXIST;
450
451 if (!l_old && map_flags == BPF_EXIST)
452 /* elem doesn't exist, cannot update it */
453 return -ENOENT;
454
455 return 0;
456}
457
458/* Called from syscall or from eBPF program */
459static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
460 u64 map_flags)
461{
462 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
463 struct htab_elem *l_new = NULL, *l_old;
464 struct hlist_head *head;
465 unsigned long flags;
466 struct bucket *b;
467 u32 key_size, hash;
468 int ret;
469
470 if (unlikely(map_flags > BPF_EXIST))
471 /* unknown flags */
472 return -EINVAL;
473
474 WARN_ON_ONCE(!rcu_read_lock_held());
475
476 key_size = map->key_size;
477
478 hash = htab_map_hash(key, key_size);
479
480 b = __select_bucket(htab, hash);
481 head = &b->head;
482
483 /* bpf_map_update_elem() can be called in_irq() */
484 raw_spin_lock_irqsave(&b->lock, flags);
485
486 l_old = lookup_elem_raw(head, hash, key, key_size);
487
488 ret = check_flags(htab, l_old, map_flags);
489 if (ret)
490 goto err;
491
492 l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false);
493 if (IS_ERR(l_new)) {
494 /* all pre-allocated elements are in use or memory exhausted */
495 ret = PTR_ERR(l_new);
496 goto err;
497 }
498
499 /* add new element to the head of the list, so that
500 * concurrent search will find it before old elem
501 */
502 hlist_add_head_rcu(&l_new->hash_node, head);
503 if (l_old) {
504 hlist_del_rcu(&l_old->hash_node);
505 free_htab_elem(htab, l_old);
506 }
507 ret = 0;
508err:
509 raw_spin_unlock_irqrestore(&b->lock, flags);
510 return ret;
511}
512
513static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
514 void *value, u64 map_flags,
515 bool onallcpus)
516{
517 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
518 struct htab_elem *l_new = NULL, *l_old;
519 struct hlist_head *head;
520 unsigned long flags;
521 struct bucket *b;
522 u32 key_size, hash;
523 int ret;
524
525 if (unlikely(map_flags > BPF_EXIST))
526 /* unknown flags */
527 return -EINVAL;
528
529 WARN_ON_ONCE(!rcu_read_lock_held());
530
531 key_size = map->key_size;
532
533 hash = htab_map_hash(key, key_size);
534
535 b = __select_bucket(htab, hash);
536 head = &b->head;
537
538 /* bpf_map_update_elem() can be called in_irq() */
539 raw_spin_lock_irqsave(&b->lock, flags);
540
541 l_old = lookup_elem_raw(head, hash, key, key_size);
542
543 ret = check_flags(htab, l_old, map_flags);
544 if (ret)
545 goto err;
546
547 if (l_old) {
548 void __percpu *pptr = htab_elem_get_ptr(l_old, key_size);
549 u32 size = htab->map.value_size;
550
551 /* per-cpu hash map can update value in-place */
552 if (!onallcpus) {
553 memcpy(this_cpu_ptr(pptr), value, size);
554 } else {
555 int off = 0, cpu;
556
557 size = round_up(size, 8);
558 for_each_possible_cpu(cpu) {
559 bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
560 value + off, size);
561 off += size;
562 }
563 }
564 } else {
565 l_new = alloc_htab_elem(htab, key, value, key_size,
566 hash, true, onallcpus);
567 if (IS_ERR(l_new)) {
568 ret = PTR_ERR(l_new);
569 goto err;
570 }
571 hlist_add_head_rcu(&l_new->hash_node, head);
572 }
573 ret = 0;
574err:
575 raw_spin_unlock_irqrestore(&b->lock, flags);
576 return ret;
577}
578
579static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
580 void *value, u64 map_flags)
581{
582 return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
583}
584
585/* Called from syscall or from eBPF program */
586static int htab_map_delete_elem(struct bpf_map *map, void *key)
587{
588 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
589 struct hlist_head *head;
590 struct bucket *b;
591 struct htab_elem *l;
592 unsigned long flags;
593 u32 hash, key_size;
594 int ret = -ENOENT;
595
596 WARN_ON_ONCE(!rcu_read_lock_held());
597
598 key_size = map->key_size;
599
600 hash = htab_map_hash(key, key_size);
601 b = __select_bucket(htab, hash);
602 head = &b->head;
603
604 raw_spin_lock_irqsave(&b->lock, flags);
605
606 l = lookup_elem_raw(head, hash, key, key_size);
607
608 if (l) {
609 hlist_del_rcu(&l->hash_node);
610 free_htab_elem(htab, l);
611 ret = 0;
612 }
613
614 raw_spin_unlock_irqrestore(&b->lock, flags);
615 return ret;
616}
617
618static void delete_all_elements(struct bpf_htab *htab)
619{
620 int i;
621
622 for (i = 0; i < htab->n_buckets; i++) {
623 struct hlist_head *head = select_bucket(htab, i);
624 struct hlist_node *n;
625 struct htab_elem *l;
626
627 hlist_for_each_entry_safe(l, n, head, hash_node) {
628 hlist_del_rcu(&l->hash_node);
629 htab_elem_free(htab, l);
630 }
631 }
632}
633/* Called when map->refcnt goes to zero, either from workqueue or from syscall */
634static void htab_map_free(struct bpf_map *map)
635{
636 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
637
638 /* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
639 * so the programs (can be more than one that used this map) were
640 * disconnected from events. Wait for outstanding critical sections in
641 * these programs to complete
642 */
643 synchronize_rcu();
644
645 /* some of free_htab_elem() callbacks for elements of this map may
646 * not have executed. Wait for them.
647 */
648 rcu_barrier();
649 if (htab->map.map_flags & BPF_F_NO_PREALLOC) {
650 delete_all_elements(htab);
651 } else {
652 htab_free_elems(htab);
653 pcpu_freelist_destroy(&htab->freelist);
654 }
655 kvfree(htab->buckets);
656 kfree(htab);
657}
658
659static const struct bpf_map_ops htab_ops = {
660 .map_alloc = htab_map_alloc,
661 .map_free = htab_map_free,
662 .map_get_next_key = htab_map_get_next_key,
663 .map_lookup_elem = htab_map_lookup_elem,
664 .map_update_elem = htab_map_update_elem,
665 .map_delete_elem = htab_map_delete_elem,
666};
667
668static struct bpf_map_type_list htab_type __read_mostly = {
669 .ops = &htab_ops,
670 .type = BPF_MAP_TYPE_HASH,
671};
672
673/* Called from eBPF program */
674static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
675{
676 struct htab_elem *l = __htab_map_lookup_elem(map, key);
677
678 if (l)
679 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
680 else
681 return NULL;
682}
683
684int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
685{
686 struct htab_elem *l;
687 void __percpu *pptr;
688 int ret = -ENOENT;
689 int cpu, off = 0;
690 u32 size;
691
692 /* per_cpu areas are zero-filled and bpf programs can only
693 * access 'value_size' of them, so copying rounded areas
694 * will not leak any kernel data
695 */
696 size = round_up(map->value_size, 8);
697 rcu_read_lock();
698 l = __htab_map_lookup_elem(map, key);
699 if (!l)
700 goto out;
701 pptr = htab_elem_get_ptr(l, map->key_size);
702 for_each_possible_cpu(cpu) {
703 bpf_long_memcpy(value + off,
704 per_cpu_ptr(pptr, cpu), size);
705 off += size;
706 }
707 ret = 0;
708out:
709 rcu_read_unlock();
710 return ret;
711}
712
713int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
714 u64 map_flags)
715{
716 int ret;
717
718 rcu_read_lock();
719 ret = __htab_percpu_map_update_elem(map, key, value, map_flags, true);
720 rcu_read_unlock();
721
722 return ret;
723}
724
725static const struct bpf_map_ops htab_percpu_ops = {
726 .map_alloc = htab_map_alloc,
727 .map_free = htab_map_free,
728 .map_get_next_key = htab_map_get_next_key,
729 .map_lookup_elem = htab_percpu_map_lookup_elem,
730 .map_update_elem = htab_percpu_map_update_elem,
731 .map_delete_elem = htab_map_delete_elem,
732};
733
734static struct bpf_map_type_list htab_percpu_type __read_mostly = {
735 .ops = &htab_percpu_ops,
736 .type = BPF_MAP_TYPE_PERCPU_HASH,
737};
738
739static int __init register_htab_map(void)
740{
741 bpf_register_map_type(&htab_type);
742 bpf_register_map_type(&htab_percpu_type);
743 return 0;
744}
745late_initcall(register_htab_map);