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