?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Linux 4.15.6
[linux/fpc-iii.git] / kernel / bpf / hashtab.c
blob3905d4bc5b80d74f0b8f9e2e8f8526a0115ce239
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
25 struct bucket {
26 struct hlist_nulls_head head;
27 raw_spinlock_t lock;
28 };
29
30 struct 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 */
45 struct 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
64 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
65
66 static 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
72 static 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
78 static bool htab_is_prealloc(const struct bpf_htab *htab)
79 {
80 return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
81 }
82
83 static 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
89 static 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
94 static 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
99 static 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
104 static 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 }
119 free_elems:
120 bpf_map_area_free(htab->elems);
121 }
122
123 static 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
138 static 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
166 skip_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
191 free_elems:
192 htab_free_elems(htab);
193 return err;
194 }
195
196 static 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
206 static 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 */
230 static struct bpf_map *htab_map_alloc(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 struct bpf_htab *htab;
245 int err, i;
246 u64 cost;
247
248 BUILD_BUG_ON(offsetof(struct htab_elem, htab) !=
249 offsetof(struct htab_elem, hash_node.pprev));
250 BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
251 offsetof(struct htab_elem, hash_node.pprev));
252
253 if (lru && !capable(CAP_SYS_ADMIN))
254 /* LRU implementation is much complicated than other
255 * maps. Hence, limit to CAP_SYS_ADMIN for now.
256 */
257 return ERR_PTR(-EPERM);
258
259 if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK)
260 /* reserved bits should not be used */
261 return ERR_PTR(-EINVAL);
262
263 if (!lru && percpu_lru)
264 return ERR_PTR(-EINVAL);
265
266 if (lru && !prealloc)
267 return ERR_PTR(-ENOTSUPP);
268
269 if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
270 return ERR_PTR(-EINVAL);
271
272 htab = kzalloc(sizeof(*htab), GFP_USER);
273 if (!htab)
274 return ERR_PTR(-ENOMEM);
275
276 /* mandatory map attributes */
277 htab->map.map_type = attr->map_type;
278 htab->map.key_size = attr->key_size;
279 htab->map.value_size = attr->value_size;
280 htab->map.max_entries = attr->max_entries;
281 htab->map.map_flags = attr->map_flags;
282 htab->map.numa_node = numa_node;
283
284 /* check sanity of attributes.
285 * value_size == 0 may be allowed in the future to use map as a set
286 */
287 err = -EINVAL;
288 if (htab->map.max_entries == 0 || htab->map.key_size == 0 ||
289 htab->map.value_size == 0)
290 goto free_htab;
291
292 if (percpu_lru) {
293 /* ensure each CPU's lru list has >=1 elements.
294 * since we are at it, make each lru list has the same
295 * number of elements.
296 */
297 htab->map.max_entries = roundup(attr->max_entries,
298 num_possible_cpus());
299 if (htab->map.max_entries < attr->max_entries)
300 htab->map.max_entries = rounddown(attr->max_entries,
301 num_possible_cpus());
302 }
303
304 /* hash table size must be power of 2 */
305 htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
306
307 err = -E2BIG;
308 if (htab->map.key_size > MAX_BPF_STACK)
309 /* eBPF programs initialize keys on stack, so they cannot be
310 * larger than max stack size
311 */
312 goto free_htab;
313
314 if (htab->map.value_size >= KMALLOC_MAX_SIZE -
315 MAX_BPF_STACK - sizeof(struct htab_elem))
316 /* if value_size is bigger, the user space won't be able to
317 * access the elements via bpf syscall. This check also makes
318 * sure that the elem_size doesn't overflow and it's
319 * kmalloc-able later in htab_map_update_elem()
320 */
321 goto free_htab;
322
323 htab->elem_size = sizeof(struct htab_elem) +
324 round_up(htab->map.key_size, 8);
325 if (percpu)
326 htab->elem_size += sizeof(void *);
327 else
328 htab->elem_size += round_up(htab->map.value_size, 8);
329
330 /* prevent zero size kmalloc and check for u32 overflow */
331 if (htab->n_buckets == 0 ||
332 htab->n_buckets > U32_MAX / sizeof(struct bucket))
333 goto free_htab;
334
335 cost = (u64) htab->n_buckets * sizeof(struct bucket) +
336 (u64) htab->elem_size * htab->map.max_entries;
337
338 if (percpu)
339 cost += (u64) round_up(htab->map.value_size, 8) *
340 num_possible_cpus() * htab->map.max_entries;
341 else
342 cost += (u64) htab->elem_size * num_possible_cpus();
343
344 if (cost >= U32_MAX - PAGE_SIZE)
345 /* make sure page count doesn't overflow */
346 goto free_htab;
347
348 htab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
349
350 /* if map size is larger than memlock limit, reject it early */
351 err = bpf_map_precharge_memlock(htab->map.pages);
352 if (err)
353 goto free_htab;
354
355 err = -ENOMEM;
356 htab->buckets = bpf_map_area_alloc(htab->n_buckets *
357 sizeof(struct bucket),
358 htab->map.numa_node);
359 if (!htab->buckets)
360 goto free_htab;
361
362 for (i = 0; i < htab->n_buckets; i++) {
363 INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
364 raw_spin_lock_init(&htab->buckets[i].lock);
365 }
366
367 if (prealloc) {
368 err = prealloc_init(htab);
369 if (err)
370 goto free_buckets;
371
372 if (!percpu && !lru) {
373 /* lru itself can remove the least used element, so
374 * there is no need for an extra elem during map_update.
375 */
376 err = alloc_extra_elems(htab);
377 if (err)
378 goto free_prealloc;
379 }
380 }
381
382 return &htab->map;
383
384 free_prealloc:
385 prealloc_destroy(htab);
386 free_buckets:
387 bpf_map_area_free(htab->buckets);
388 free_htab:
389 kfree(htab);
390 return ERR_PTR(err);
391 }
392
393 static inline u32 htab_map_hash(const void *key, u32 key_len)
394 {
395 return jhash(key, key_len, 0);
396 }
397
398 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
399 {
400 return &htab->buckets[hash & (htab->n_buckets - 1)];
401 }
402
403 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
404 {
405 return &__select_bucket(htab, hash)->head;
406 }
407
408 /* this lookup function can only be called with bucket lock taken */
409 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
410 void *key, u32 key_size)
411 {
412 struct hlist_nulls_node *n;
413 struct htab_elem *l;
414
415 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
416 if (l->hash == hash && !memcmp(&l->key, key, key_size))
417 return l;
418
419 return NULL;
420 }
421
422 /* can be called without bucket lock. it will repeat the loop in
423 * the unlikely event when elements moved from one bucket into another
424 * while link list is being walked
425 */
426 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
427 u32 hash, void *key,
428 u32 key_size, u32 n_buckets)
429 {
430 struct hlist_nulls_node *n;
431 struct htab_elem *l;
432
433 again:
434 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
435 if (l->hash == hash && !memcmp(&l->key, key, key_size))
436 return l;
437
438 if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
439 goto again;
440
441 return NULL;
442 }
443
444 /* Called from syscall or from eBPF program directly, so
445 * arguments have to match bpf_map_lookup_elem() exactly.
446 * The return value is adjusted by BPF instructions
447 * in htab_map_gen_lookup().
448 */
449 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
450 {
451 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
452 struct hlist_nulls_head *head;
453 struct htab_elem *l;
454 u32 hash, key_size;
455
456 /* Must be called with rcu_read_lock. */
457 WARN_ON_ONCE(!rcu_read_lock_held());
458
459 key_size = map->key_size;
460
461 hash = htab_map_hash(key, key_size);
462
463 head = select_bucket(htab, hash);
464
465 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
466
467 return l;
468 }
469
470 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
471 {
472 struct htab_elem *l = __htab_map_lookup_elem(map, key);
473
474 if (l)
475 return l->key + round_up(map->key_size, 8);
476
477 return NULL;
478 }
479
480 /* inline bpf_map_lookup_elem() call.
481 * Instead of:
482 * bpf_prog
483 * bpf_map_lookup_elem
484 * map->ops->map_lookup_elem
485 * htab_map_lookup_elem
486 * __htab_map_lookup_elem
487 * do:
488 * bpf_prog
489 * __htab_map_lookup_elem
490 */
491 static u32 htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
492 {
493 struct bpf_insn *insn = insn_buf;
494 const int ret = BPF_REG_0;
495
496 *insn++ = BPF_EMIT_CALL((u64 (*)(u64, u64, u64, u64, u64))__htab_map_lookup_elem);
497 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
498 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
499 offsetof(struct htab_elem, key) +
500 round_up(map->key_size, 8));
501 return insn - insn_buf;
502 }
503
504 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
505 {
506 struct htab_elem *l = __htab_map_lookup_elem(map, key);
507
508 if (l) {
509 bpf_lru_node_set_ref(&l->lru_node);
510 return l->key + round_up(map->key_size, 8);
511 }
512
513 return NULL;
514 }
515
516 static u32 htab_lru_map_gen_lookup(struct bpf_map *map,
517 struct bpf_insn *insn_buf)
518 {
519 struct bpf_insn *insn = insn_buf;
520 const int ret = BPF_REG_0;
521 const int ref_reg = BPF_REG_1;
522
523 *insn++ = BPF_EMIT_CALL((u64 (*)(u64, u64, u64, u64, u64))__htab_map_lookup_elem);
524 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
525 *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
526 offsetof(struct htab_elem, lru_node) +
527 offsetof(struct bpf_lru_node, ref));
528 *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
529 *insn++ = BPF_ST_MEM(BPF_B, ret,
530 offsetof(struct htab_elem, lru_node) +
531 offsetof(struct bpf_lru_node, ref),
532 1);
533 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
534 offsetof(struct htab_elem, key) +
535 round_up(map->key_size, 8));
536 return insn - insn_buf;
537 }
538
539 /* It is called from the bpf_lru_list when the LRU needs to delete
540 * older elements from the htab.
541 */
542 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
543 {
544 struct bpf_htab *htab = (struct bpf_htab *)arg;
545 struct htab_elem *l = NULL, *tgt_l;
546 struct hlist_nulls_head *head;
547 struct hlist_nulls_node *n;
548 unsigned long flags;
549 struct bucket *b;
550
551 tgt_l = container_of(node, struct htab_elem, lru_node);
552 b = __select_bucket(htab, tgt_l->hash);
553 head = &b->head;
554
555 raw_spin_lock_irqsave(&b->lock, flags);
556
557 hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
558 if (l == tgt_l) {
559 hlist_nulls_del_rcu(&l->hash_node);
560 break;
561 }
562
563 raw_spin_unlock_irqrestore(&b->lock, flags);
564
565 return l == tgt_l;
566 }
567
568 /* Called from syscall */
569 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
570 {
571 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
572 struct hlist_nulls_head *head;
573 struct htab_elem *l, *next_l;
574 u32 hash, key_size;
575 int i = 0;
576
577 WARN_ON_ONCE(!rcu_read_lock_held());
578
579 key_size = map->key_size;
580
581 if (!key)
582 goto find_first_elem;
583
584 hash = htab_map_hash(key, key_size);
585
586 head = select_bucket(htab, hash);
587
588 /* lookup the key */
589 l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
590
591 if (!l)
592 goto find_first_elem;
593
594 /* key was found, get next key in the same bucket */
595 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
596 struct htab_elem, hash_node);
597
598 if (next_l) {
599 /* if next elem in this hash list is non-zero, just return it */
600 memcpy(next_key, next_l->key, key_size);
601 return 0;
602 }
603
604 /* no more elements in this hash list, go to the next bucket */
605 i = hash & (htab->n_buckets - 1);
606 i++;
607
608 find_first_elem:
609 /* iterate over buckets */
610 for (; i < htab->n_buckets; i++) {
611 head = select_bucket(htab, i);
612
613 /* pick first element in the bucket */
614 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
615 struct htab_elem, hash_node);
616 if (next_l) {
617 /* if it's not empty, just return it */
618 memcpy(next_key, next_l->key, key_size);
619 return 0;
620 }
621 }
622
623 /* iterated over all buckets and all elements */
624 return -ENOENT;
625 }
626
627 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
628 {
629 if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
630 free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
631 kfree(l);
632 }
633
634 static void htab_elem_free_rcu(struct rcu_head *head)
635 {
636 struct htab_elem *l = container_of(head, struct htab_elem, rcu);
637 struct bpf_htab *htab = l->htab;
638
639 /* must increment bpf_prog_active to avoid kprobe+bpf triggering while
640 * we're calling kfree, otherwise deadlock is possible if kprobes
641 * are placed somewhere inside of slub
642 */
643 preempt_disable();
644 __this_cpu_inc(bpf_prog_active);
645 htab_elem_free(htab, l);
646 __this_cpu_dec(bpf_prog_active);
647 preempt_enable();
648 }
649
650 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
651 {
652 struct bpf_map *map = &htab->map;
653
654 if (map->ops->map_fd_put_ptr) {
655 void *ptr = fd_htab_map_get_ptr(map, l);
656
657 map->ops->map_fd_put_ptr(ptr);
658 }
659
660 if (htab_is_prealloc(htab)) {
661 pcpu_freelist_push(&htab->freelist, &l->fnode);
662 } else {
663 atomic_dec(&htab->count);
664 l->htab = htab;
665 call_rcu(&l->rcu, htab_elem_free_rcu);
666 }
667 }
668
669 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
670 void *value, bool onallcpus)
671 {
672 if (!onallcpus) {
673 /* copy true value_size bytes */
674 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
675 } else {
676 u32 size = round_up(htab->map.value_size, 8);
677 int off = 0, cpu;
678
679 for_each_possible_cpu(cpu) {
680 bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
681 value + off, size);
682 off += size;
683 }
684 }
685 }
686
687 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
688 {
689 return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
690 BITS_PER_LONG == 64;
691 }
692
693 static u32 htab_size_value(const struct bpf_htab *htab, bool percpu)
694 {
695 u32 size = htab->map.value_size;
696
697 if (percpu || fd_htab_map_needs_adjust(htab))
698 size = round_up(size, 8);
699 return size;
700 }
701
702 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
703 void *value, u32 key_size, u32 hash,
704 bool percpu, bool onallcpus,
705 struct htab_elem *old_elem)
706 {
707 u32 size = htab_size_value(htab, percpu);
708 bool prealloc = htab_is_prealloc(htab);
709 struct htab_elem *l_new, **pl_new;
710 void __percpu *pptr;
711
712 if (prealloc) {
713 if (old_elem) {
714 /* if we're updating the existing element,
715 * use per-cpu extra elems to avoid freelist_pop/push
716 */
717 pl_new = this_cpu_ptr(htab->extra_elems);
718 l_new = *pl_new;
719 *pl_new = old_elem;
720 } else {
721 struct pcpu_freelist_node *l;
722
723 l = pcpu_freelist_pop(&htab->freelist);
724 if (!l)
725 return ERR_PTR(-E2BIG);
726 l_new = container_of(l, struct htab_elem, fnode);
727 }
728 } else {
729 if (atomic_inc_return(&htab->count) > htab->map.max_entries)
730 if (!old_elem) {
731 /* when map is full and update() is replacing
732 * old element, it's ok to allocate, since
733 * old element will be freed immediately.
734 * Otherwise return an error
735 */
736 atomic_dec(&htab->count);
737 return ERR_PTR(-E2BIG);
738 }
739 l_new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
740 htab->map.numa_node);
741 if (!l_new)
742 return ERR_PTR(-ENOMEM);
743 }
744
745 memcpy(l_new->key, key, key_size);
746 if (percpu) {
747 if (prealloc) {
748 pptr = htab_elem_get_ptr(l_new, key_size);
749 } else {
750 /* alloc_percpu zero-fills */
751 pptr = __alloc_percpu_gfp(size, 8,
752 GFP_ATOMIC | __GFP_NOWARN);
753 if (!pptr) {
754 kfree(l_new);
755 return ERR_PTR(-ENOMEM);
756 }
757 }
758
759 pcpu_copy_value(htab, pptr, value, onallcpus);
760
761 if (!prealloc)
762 htab_elem_set_ptr(l_new, key_size, pptr);
763 } else {
764 memcpy(l_new->key + round_up(key_size, 8), value, size);
765 }
766
767 l_new->hash = hash;
768 return l_new;
769 }
770
771 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
772 u64 map_flags)
773 {
774 if (l_old && map_flags == BPF_NOEXIST)
775 /* elem already exists */
776 return -EEXIST;
777
778 if (!l_old && map_flags == BPF_EXIST)
779 /* elem doesn't exist, cannot update it */
780 return -ENOENT;
781
782 return 0;
783 }
784
785 /* Called from syscall or from eBPF program */
786 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
787 u64 map_flags)
788 {
789 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
790 struct htab_elem *l_new = NULL, *l_old;
791 struct hlist_nulls_head *head;
792 unsigned long flags;
793 struct bucket *b;
794 u32 key_size, hash;
795 int ret;
796
797 if (unlikely(map_flags > BPF_EXIST))
798 /* unknown flags */
799 return -EINVAL;
800
801 WARN_ON_ONCE(!rcu_read_lock_held());
802
803 key_size = map->key_size;
804
805 hash = htab_map_hash(key, key_size);
806
807 b = __select_bucket(htab, hash);
808 head = &b->head;
809
810 /* bpf_map_update_elem() can be called in_irq() */
811 raw_spin_lock_irqsave(&b->lock, flags);
812
813 l_old = lookup_elem_raw(head, hash, key, key_size);
814
815 ret = check_flags(htab, l_old, map_flags);
816 if (ret)
817 goto err;
818
819 l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
820 l_old);
821 if (IS_ERR(l_new)) {
822 /* all pre-allocated elements are in use or memory exhausted */
823 ret = PTR_ERR(l_new);
824 goto err;
825 }
826
827 /* add new element to the head of the list, so that
828 * concurrent search will find it before old elem
829 */
830 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
831 if (l_old) {
832 hlist_nulls_del_rcu(&l_old->hash_node);
833 if (!htab_is_prealloc(htab))
834 free_htab_elem(htab, l_old);
835 }
836 ret = 0;
837 err:
838 raw_spin_unlock_irqrestore(&b->lock, flags);
839 return ret;
840 }
841
842 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
843 u64 map_flags)
844 {
845 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
846 struct htab_elem *l_new, *l_old = NULL;
847 struct hlist_nulls_head *head;
848 unsigned long flags;
849 struct bucket *b;
850 u32 key_size, hash;
851 int ret;
852
853 if (unlikely(map_flags > BPF_EXIST))
854 /* unknown flags */
855 return -EINVAL;
856
857 WARN_ON_ONCE(!rcu_read_lock_held());
858
859 key_size = map->key_size;
860
861 hash = htab_map_hash(key, key_size);
862
863 b = __select_bucket(htab, hash);
864 head = &b->head;
865
866 /* For LRU, we need to alloc before taking bucket's
867 * spinlock because getting free nodes from LRU may need
868 * to remove older elements from htab and this removal
869 * operation will need a bucket lock.
870 */
871 l_new = prealloc_lru_pop(htab, key, hash);
872 if (!l_new)
873 return -ENOMEM;
874 memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
875
876 /* bpf_map_update_elem() can be called in_irq() */
877 raw_spin_lock_irqsave(&b->lock, flags);
878
879 l_old = lookup_elem_raw(head, hash, key, key_size);
880
881 ret = check_flags(htab, l_old, map_flags);
882 if (ret)
883 goto err;
884
885 /* add new element to the head of the list, so that
886 * concurrent search will find it before old elem
887 */
888 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
889 if (l_old) {
890 bpf_lru_node_set_ref(&l_new->lru_node);
891 hlist_nulls_del_rcu(&l_old->hash_node);
892 }
893 ret = 0;
894
895 err:
896 raw_spin_unlock_irqrestore(&b->lock, flags);
897
898 if (ret)
899 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
900 else if (l_old)
901 bpf_lru_push_free(&htab->lru, &l_old->lru_node);
902
903 return ret;
904 }
905
906 static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
907 void *value, u64 map_flags,
908 bool onallcpus)
909 {
910 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
911 struct htab_elem *l_new = NULL, *l_old;
912 struct hlist_nulls_head *head;
913 unsigned long flags;
914 struct bucket *b;
915 u32 key_size, hash;
916 int ret;
917
918 if (unlikely(map_flags > BPF_EXIST))
919 /* unknown flags */
920 return -EINVAL;
921
922 WARN_ON_ONCE(!rcu_read_lock_held());
923
924 key_size = map->key_size;
925
926 hash = htab_map_hash(key, key_size);
927
928 b = __select_bucket(htab, hash);
929 head = &b->head;
930
931 /* bpf_map_update_elem() can be called in_irq() */
932 raw_spin_lock_irqsave(&b->lock, flags);
933
934 l_old = lookup_elem_raw(head, hash, key, key_size);
935
936 ret = check_flags(htab, l_old, map_flags);
937 if (ret)
938 goto err;
939
940 if (l_old) {
941 /* per-cpu hash map can update value in-place */
942 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
943 value, onallcpus);
944 } else {
945 l_new = alloc_htab_elem(htab, key, value, key_size,
946 hash, true, onallcpus, NULL);
947 if (IS_ERR(l_new)) {
948 ret = PTR_ERR(l_new);
949 goto err;
950 }
951 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
952 }
953 ret = 0;
954 err:
955 raw_spin_unlock_irqrestore(&b->lock, flags);
956 return ret;
957 }
958
959 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
960 void *value, u64 map_flags,
961 bool onallcpus)
962 {
963 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
964 struct htab_elem *l_new = NULL, *l_old;
965 struct hlist_nulls_head *head;
966 unsigned long flags;
967 struct bucket *b;
968 u32 key_size, hash;
969 int ret;
970
971 if (unlikely(map_flags > BPF_EXIST))
972 /* unknown flags */
973 return -EINVAL;
974
975 WARN_ON_ONCE(!rcu_read_lock_held());
976
977 key_size = map->key_size;
978
979 hash = htab_map_hash(key, key_size);
980
981 b = __select_bucket(htab, hash);
982 head = &b->head;
983
984 /* For LRU, we need to alloc before taking bucket's
985 * spinlock because LRU's elem alloc may need
986 * to remove older elem from htab and this removal
987 * operation will need a bucket lock.
988 */
989 if (map_flags != BPF_EXIST) {
990 l_new = prealloc_lru_pop(htab, key, hash);
991 if (!l_new)
992 return -ENOMEM;
993 }
994
995 /* bpf_map_update_elem() can be called in_irq() */
996 raw_spin_lock_irqsave(&b->lock, flags);
997
998 l_old = lookup_elem_raw(head, hash, key, key_size);
999
1000 ret = check_flags(htab, l_old, map_flags);
1001 if (ret)
1002 goto err;
1003
1004 if (l_old) {
1005 bpf_lru_node_set_ref(&l_old->lru_node);
1006
1007 /* per-cpu hash map can update value in-place */
1008 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1009 value, onallcpus);
1010 } else {
1011 pcpu_copy_value(htab, htab_elem_get_ptr(l_new, key_size),
1012 value, onallcpus);
1013 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1014 l_new = NULL;
1015 }
1016 ret = 0;
1017 err:
1018 raw_spin_unlock_irqrestore(&b->lock, flags);
1019 if (l_new)
1020 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1021 return ret;
1022 }
1023
1024 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1025 void *value, u64 map_flags)
1026 {
1027 return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1028 }
1029
1030 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1031 void *value, u64 map_flags)
1032 {
1033 return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1034 false);
1035 }
1036
1037 /* Called from syscall or from eBPF program */
1038 static int htab_map_delete_elem(struct bpf_map *map, void *key)
1039 {
1040 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1041 struct hlist_nulls_head *head;
1042 struct bucket *b;
1043 struct htab_elem *l;
1044 unsigned long flags;
1045 u32 hash, key_size;
1046 int ret = -ENOENT;
1047
1048 WARN_ON_ONCE(!rcu_read_lock_held());
1049
1050 key_size = map->key_size;
1051
1052 hash = htab_map_hash(key, key_size);
1053 b = __select_bucket(htab, hash);
1054 head = &b->head;
1055
1056 raw_spin_lock_irqsave(&b->lock, flags);
1057
1058 l = lookup_elem_raw(head, hash, key, key_size);
1059
1060 if (l) {
1061 hlist_nulls_del_rcu(&l->hash_node);
1062 free_htab_elem(htab, l);
1063 ret = 0;
1064 }
1065
1066 raw_spin_unlock_irqrestore(&b->lock, flags);
1067 return ret;
1068 }
1069
1070 static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1071 {
1072 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1073 struct hlist_nulls_head *head;
1074 struct bucket *b;
1075 struct htab_elem *l;
1076 unsigned long flags;
1077 u32 hash, key_size;
1078 int ret = -ENOENT;
1079
1080 WARN_ON_ONCE(!rcu_read_lock_held());
1081
1082 key_size = map->key_size;
1083
1084 hash = htab_map_hash(key, key_size);
1085 b = __select_bucket(htab, hash);
1086 head = &b->head;
1087
1088 raw_spin_lock_irqsave(&b->lock, flags);
1089
1090 l = lookup_elem_raw(head, hash, key, key_size);
1091
1092 if (l) {
1093 hlist_nulls_del_rcu(&l->hash_node);
1094 ret = 0;
1095 }
1096
1097 raw_spin_unlock_irqrestore(&b->lock, flags);
1098 if (l)
1099 bpf_lru_push_free(&htab->lru, &l->lru_node);
1100 return ret;
1101 }
1102
1103 static void delete_all_elements(struct bpf_htab *htab)
1104 {
1105 int i;
1106
1107 for (i = 0; i < htab->n_buckets; i++) {
1108 struct hlist_nulls_head *head = select_bucket(htab, i);
1109 struct hlist_nulls_node *n;
1110 struct htab_elem *l;
1111
1112 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1113 hlist_nulls_del_rcu(&l->hash_node);
1114 htab_elem_free(htab, l);
1115 }
1116 }
1117 }
1118
1119 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1120 static void htab_map_free(struct bpf_map *map)
1121 {
1122 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1123
1124 /* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
1125 * so the programs (can be more than one that used this map) were
1126 * disconnected from events. Wait for outstanding critical sections in
1127 * these programs to complete
1128 */
1129 synchronize_rcu();
1130
1131 /* some of free_htab_elem() callbacks for elements of this map may
1132 * not have executed. Wait for them.
1133 */
1134 rcu_barrier();
1135 if (!htab_is_prealloc(htab))
1136 delete_all_elements(htab);
1137 else
1138 prealloc_destroy(htab);
1139
1140 free_percpu(htab->extra_elems);
1141 bpf_map_area_free(htab->buckets);
1142 kfree(htab);
1143 }
1144
1145 const struct bpf_map_ops htab_map_ops = {
1146 .map_alloc = htab_map_alloc,
1147 .map_free = htab_map_free,
1148 .map_get_next_key = htab_map_get_next_key,
1149 .map_lookup_elem = htab_map_lookup_elem,
1150 .map_update_elem = htab_map_update_elem,
1151 .map_delete_elem = htab_map_delete_elem,
1152 .map_gen_lookup = htab_map_gen_lookup,
1153 };
1154
1155 const struct bpf_map_ops htab_lru_map_ops = {
1156 .map_alloc = htab_map_alloc,
1157 .map_free = htab_map_free,
1158 .map_get_next_key = htab_map_get_next_key,
1159 .map_lookup_elem = htab_lru_map_lookup_elem,
1160 .map_update_elem = htab_lru_map_update_elem,
1161 .map_delete_elem = htab_lru_map_delete_elem,
1162 .map_gen_lookup = htab_lru_map_gen_lookup,
1163 };
1164
1165 /* Called from eBPF program */
1166 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1167 {
1168 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1169
1170 if (l)
1171 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1172 else
1173 return NULL;
1174 }
1175
1176 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1177 {
1178 struct htab_elem *l = __htab_map_lookup_elem(map, key);
1179
1180 if (l) {
1181 bpf_lru_node_set_ref(&l->lru_node);
1182 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1183 }
1184
1185 return NULL;
1186 }
1187
1188 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1189 {
1190 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1191 struct htab_elem *l;
1192 void __percpu *pptr;
1193 int ret = -ENOENT;
1194 int cpu, off = 0;
1195 u32 size;
1196
1197 /* per_cpu areas are zero-filled and bpf programs can only
1198 * access 'value_size' of them, so copying rounded areas
1199 * will not leak any kernel data
1200 */
1201 size = round_up(map->value_size, 8);
1202 rcu_read_lock();
1203 l = __htab_map_lookup_elem(map, key);
1204 if (!l)
1205 goto out;
1206 if (htab_is_lru(htab))
1207 bpf_lru_node_set_ref(&l->lru_node);
1208 pptr = htab_elem_get_ptr(l, map->key_size);
1209 for_each_possible_cpu(cpu) {
1210 bpf_long_memcpy(value + off,
1211 per_cpu_ptr(pptr, cpu), size);
1212 off += size;
1213 }
1214 ret = 0;
1215 out:
1216 rcu_read_unlock();
1217 return ret;
1218 }
1219
1220 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1221 u64 map_flags)
1222 {
1223 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1224 int ret;
1225
1226 rcu_read_lock();
1227 if (htab_is_lru(htab))
1228 ret = __htab_lru_percpu_map_update_elem(map, key, value,
1229 map_flags, true);
1230 else
1231 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1232 true);
1233 rcu_read_unlock();
1234
1235 return ret;
1236 }
1237
1238 const struct bpf_map_ops htab_percpu_map_ops = {
1239 .map_alloc = htab_map_alloc,
1240 .map_free = htab_map_free,
1241 .map_get_next_key = htab_map_get_next_key,
1242 .map_lookup_elem = htab_percpu_map_lookup_elem,
1243 .map_update_elem = htab_percpu_map_update_elem,
1244 .map_delete_elem = htab_map_delete_elem,
1245 };
1246
1247 const struct bpf_map_ops htab_lru_percpu_map_ops = {
1248 .map_alloc = htab_map_alloc,
1249 .map_free = htab_map_free,
1250 .map_get_next_key = htab_map_get_next_key,
1251 .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
1252 .map_update_elem = htab_lru_percpu_map_update_elem,
1253 .map_delete_elem = htab_lru_map_delete_elem,
1254 };
1255
1256 static struct bpf_map *fd_htab_map_alloc(union bpf_attr *attr)
1257 {
1258 if (attr->value_size != sizeof(u32))
1259 return ERR_PTR(-EINVAL);
1260 return htab_map_alloc(attr);
1261 }
1262
1263 static void fd_htab_map_free(struct bpf_map *map)
1264 {
1265 struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1266 struct hlist_nulls_node *n;
1267 struct hlist_nulls_head *head;
1268 struct htab_elem *l;
1269 int i;
1270
1271 for (i = 0; i < htab->n_buckets; i++) {
1272 head = select_bucket(htab, i);
1273
1274 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1275 void *ptr = fd_htab_map_get_ptr(map, l);
1276
1277 map->ops->map_fd_put_ptr(ptr);
1278 }
1279 }
1280
1281 htab_map_free(map);
1282 }
1283
1284 /* only called from syscall */
1285 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
1286 {
1287 void **ptr;
1288 int ret = 0;
1289
1290 if (!map->ops->map_fd_sys_lookup_elem)
1291 return -ENOTSUPP;
1292
1293 rcu_read_lock();
1294 ptr = htab_map_lookup_elem(map, key);
1295 if (ptr)
1296 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
1297 else
1298 ret = -ENOENT;
1299 rcu_read_unlock();
1300
1301 return ret;
1302 }
1303
1304 /* only called from syscall */
1305 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
1306 void *key, void *value, u64 map_flags)
1307 {
1308 void *ptr;
1309 int ret;
1310 u32 ufd = *(u32 *)value;
1311
1312 ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
1313 if (IS_ERR(ptr))
1314 return PTR_ERR(ptr);
1315
1316 ret = htab_map_update_elem(map, key, &ptr, map_flags);
1317 if (ret)
1318 map->ops->map_fd_put_ptr(ptr);
1319
1320 return ret;
1321 }
1322
1323 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
1324 {
1325 struct bpf_map *map, *inner_map_meta;
1326
1327 inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
1328 if (IS_ERR(inner_map_meta))
1329 return inner_map_meta;
1330
1331 map = fd_htab_map_alloc(attr);
1332 if (IS_ERR(map)) {
1333 bpf_map_meta_free(inner_map_meta);
1334 return map;
1335 }
1336
1337 map->inner_map_meta = inner_map_meta;
1338
1339 return map;
1340 }
1341
1342 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
1343 {
1344 struct bpf_map **inner_map = htab_map_lookup_elem(map, key);
1345
1346 if (!inner_map)
1347 return NULL;
1348
1349 return READ_ONCE(*inner_map);
1350 }
1351
1352 static u32 htab_of_map_gen_lookup(struct bpf_map *map,
1353 struct bpf_insn *insn_buf)
1354 {
1355 struct bpf_insn *insn = insn_buf;
1356 const int ret = BPF_REG_0;
1357
1358 *insn++ = BPF_EMIT_CALL((u64 (*)(u64, u64, u64, u64, u64))__htab_map_lookup_elem);
1359 *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
1360 *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
1361 offsetof(struct htab_elem, key) +
1362 round_up(map->key_size, 8));
1363 *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
1364
1365 return insn - insn_buf;
1366 }
1367
1368 static void htab_of_map_free(struct bpf_map *map)
1369 {
1370 bpf_map_meta_free(map->inner_map_meta);
1371 fd_htab_map_free(map);
1372 }
1373
1374 const struct bpf_map_ops htab_of_maps_map_ops = {
1375 .map_alloc = htab_of_map_alloc,
1376 .map_free = htab_of_map_free,
1377 .map_get_next_key = htab_map_get_next_key,
1378 .map_lookup_elem = htab_of_map_lookup_elem,
1379 .map_delete_elem = htab_map_delete_elem,
1380 .map_fd_get_ptr = bpf_map_fd_get_ptr,
1381 .map_fd_put_ptr = bpf_map_fd_put_ptr,
1382 .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
1383 .map_gen_lookup = htab_of_map_gen_lookup,
1384 };
Linux with added FPC-III support