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