Commit | Line | Data |
---|---|---|
450af8d0 KS |
1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* Copyright (c) 2019 Facebook */ | |
3 | #include <linux/rculist.h> | |
4 | #include <linux/list.h> | |
5 | #include <linux/hash.h> | |
6 | #include <linux/types.h> | |
7 | #include <linux/spinlock.h> | |
8 | #include <linux/bpf.h> | |
9 | #include <linux/btf_ids.h> | |
10 | #include <linux/bpf_local_storage.h> | |
11 | #include <net/sock.h> | |
12 | #include <uapi/linux/sock_diag.h> | |
13 | #include <uapi/linux/btf.h> | |
0fe4b381 KS |
14 | #include <linux/rcupdate.h> |
15 | #include <linux/rcupdate_trace.h> | |
16 | #include <linux/rcupdate_wait.h> | |
450af8d0 KS |
17 | |
18 | #define BPF_LOCAL_STORAGE_CREATE_FLAG_MASK (BPF_F_NO_PREALLOC | BPF_F_CLONE) | |
19 | ||
20 | static struct bpf_local_storage_map_bucket * | |
21 | select_bucket(struct bpf_local_storage_map *smap, | |
22 | struct bpf_local_storage_elem *selem) | |
23 | { | |
24 | return &smap->buckets[hash_ptr(selem, smap->bucket_log)]; | |
25 | } | |
26 | ||
27 | static int mem_charge(struct bpf_local_storage_map *smap, void *owner, u32 size) | |
28 | { | |
29 | struct bpf_map *map = &smap->map; | |
30 | ||
31 | if (!map->ops->map_local_storage_charge) | |
32 | return 0; | |
33 | ||
34 | return map->ops->map_local_storage_charge(smap, owner, size); | |
35 | } | |
36 | ||
37 | static void mem_uncharge(struct bpf_local_storage_map *smap, void *owner, | |
38 | u32 size) | |
39 | { | |
40 | struct bpf_map *map = &smap->map; | |
41 | ||
42 | if (map->ops->map_local_storage_uncharge) | |
43 | map->ops->map_local_storage_uncharge(smap, owner, size); | |
44 | } | |
45 | ||
46 | static struct bpf_local_storage __rcu ** | |
47 | owner_storage(struct bpf_local_storage_map *smap, void *owner) | |
48 | { | |
49 | struct bpf_map *map = &smap->map; | |
50 | ||
51 | return map->ops->map_owner_storage_ptr(owner); | |
52 | } | |
53 | ||
54 | static bool selem_linked_to_storage(const struct bpf_local_storage_elem *selem) | |
55 | { | |
56 | return !hlist_unhashed(&selem->snode); | |
57 | } | |
58 | ||
59 | static bool selem_linked_to_map(const struct bpf_local_storage_elem *selem) | |
60 | { | |
61 | return !hlist_unhashed(&selem->map_node); | |
62 | } | |
63 | ||
64 | struct bpf_local_storage_elem * | |
65 | bpf_selem_alloc(struct bpf_local_storage_map *smap, void *owner, | |
b00fa38a | 66 | void *value, bool charge_mem, gfp_t gfp_flags) |
450af8d0 KS |
67 | { |
68 | struct bpf_local_storage_elem *selem; | |
69 | ||
70 | if (charge_mem && mem_charge(smap, owner, smap->elem_size)) | |
71 | return NULL; | |
72 | ||
e9aae8be | 73 | selem = bpf_map_kzalloc(&smap->map, smap->elem_size, |
b00fa38a | 74 | gfp_flags | __GFP_NOWARN); |
450af8d0 KS |
75 | if (selem) { |
76 | if (value) | |
836e49e1 | 77 | copy_map_value(&smap->map, SDATA(selem)->data, value); |
450af8d0 KS |
78 | return selem; |
79 | } | |
80 | ||
81 | if (charge_mem) | |
82 | mem_uncharge(smap, owner, smap->elem_size); | |
83 | ||
84 | return NULL; | |
85 | } | |
86 | ||
0fe4b381 KS |
87 | void bpf_local_storage_free_rcu(struct rcu_head *rcu) |
88 | { | |
89 | struct bpf_local_storage *local_storage; | |
90 | ||
d39d1445 HT |
91 | /* If RCU Tasks Trace grace period implies RCU grace period, do |
92 | * kfree(), else do kfree_rcu(). | |
93 | */ | |
0fe4b381 | 94 | local_storage = container_of(rcu, struct bpf_local_storage, rcu); |
d39d1445 HT |
95 | if (rcu_trace_implies_rcu_gp()) |
96 | kfree(local_storage); | |
97 | else | |
98 | kfree_rcu(local_storage, rcu); | |
0fe4b381 KS |
99 | } |
100 | ||
101 | static void bpf_selem_free_rcu(struct rcu_head *rcu) | |
102 | { | |
103 | struct bpf_local_storage_elem *selem; | |
104 | ||
105 | selem = container_of(rcu, struct bpf_local_storage_elem, rcu); | |
d39d1445 HT |
106 | if (rcu_trace_implies_rcu_gp()) |
107 | kfree(selem); | |
108 | else | |
109 | kfree_rcu(selem, rcu); | |
0fe4b381 KS |
110 | } |
111 | ||
450af8d0 KS |
112 | /* local_storage->lock must be held and selem->local_storage == local_storage. |
113 | * The caller must ensure selem->smap is still valid to be | |
114 | * dereferenced for its smap->elem_size and smap->cache_idx. | |
115 | */ | |
c83597fa YS |
116 | static bool bpf_selem_unlink_storage_nolock(struct bpf_local_storage *local_storage, |
117 | struct bpf_local_storage_elem *selem, | |
118 | bool uncharge_mem, bool use_trace_rcu) | |
450af8d0 KS |
119 | { |
120 | struct bpf_local_storage_map *smap; | |
121 | bool free_local_storage; | |
122 | void *owner; | |
123 | ||
0fe4b381 | 124 | smap = rcu_dereference_check(SDATA(selem)->smap, bpf_rcu_lock_held()); |
450af8d0 KS |
125 | owner = local_storage->owner; |
126 | ||
127 | /* All uncharging on the owner must be done first. | |
128 | * The owner may be freed once the last selem is unlinked | |
129 | * from local_storage. | |
130 | */ | |
131 | if (uncharge_mem) | |
132 | mem_uncharge(smap, owner, smap->elem_size); | |
133 | ||
134 | free_local_storage = hlist_is_singular_node(&selem->snode, | |
135 | &local_storage->list); | |
136 | if (free_local_storage) { | |
137 | mem_uncharge(smap, owner, sizeof(struct bpf_local_storage)); | |
138 | local_storage->owner = NULL; | |
139 | ||
140 | /* After this RCU_INIT, owner may be freed and cannot be used */ | |
141 | RCU_INIT_POINTER(*owner_storage(smap, owner), NULL); | |
142 | ||
143 | /* local_storage is not freed now. local_storage->lock is | |
144 | * still held and raw_spin_unlock_bh(&local_storage->lock) | |
145 | * will be done by the caller. | |
146 | * | |
147 | * Although the unlock will be done under | |
c561d110 | 148 | * rcu_read_lock(), it is more intuitive to |
0fe4b381 | 149 | * read if the freeing of the storage is done |
450af8d0 KS |
150 | * after the raw_spin_unlock_bh(&local_storage->lock). |
151 | * | |
152 | * Hence, a "bool free_local_storage" is returned | |
0fe4b381 KS |
153 | * to the caller which then calls then frees the storage after |
154 | * all the RCU grace periods have expired. | |
450af8d0 KS |
155 | */ |
156 | } | |
157 | hlist_del_init_rcu(&selem->snode); | |
158 | if (rcu_access_pointer(local_storage->cache[smap->cache_idx]) == | |
159 | SDATA(selem)) | |
160 | RCU_INIT_POINTER(local_storage->cache[smap->cache_idx], NULL); | |
161 | ||
dcf456c9 KS |
162 | if (use_trace_rcu) |
163 | call_rcu_tasks_trace(&selem->rcu, bpf_selem_free_rcu); | |
164 | else | |
165 | kfree_rcu(selem, rcu); | |
166 | ||
450af8d0 KS |
167 | return free_local_storage; |
168 | } | |
169 | ||
dcf456c9 KS |
170 | static void __bpf_selem_unlink_storage(struct bpf_local_storage_elem *selem, |
171 | bool use_trace_rcu) | |
450af8d0 KS |
172 | { |
173 | struct bpf_local_storage *local_storage; | |
174 | bool free_local_storage = false; | |
a10787e6 | 175 | unsigned long flags; |
450af8d0 KS |
176 | |
177 | if (unlikely(!selem_linked_to_storage(selem))) | |
178 | /* selem has already been unlinked from sk */ | |
179 | return; | |
180 | ||
0fe4b381 KS |
181 | local_storage = rcu_dereference_check(selem->local_storage, |
182 | bpf_rcu_lock_held()); | |
a10787e6 | 183 | raw_spin_lock_irqsave(&local_storage->lock, flags); |
450af8d0 KS |
184 | if (likely(selem_linked_to_storage(selem))) |
185 | free_local_storage = bpf_selem_unlink_storage_nolock( | |
dcf456c9 | 186 | local_storage, selem, true, use_trace_rcu); |
a10787e6 | 187 | raw_spin_unlock_irqrestore(&local_storage->lock, flags); |
450af8d0 | 188 | |
dcf456c9 KS |
189 | if (free_local_storage) { |
190 | if (use_trace_rcu) | |
191 | call_rcu_tasks_trace(&local_storage->rcu, | |
0fe4b381 | 192 | bpf_local_storage_free_rcu); |
dcf456c9 KS |
193 | else |
194 | kfree_rcu(local_storage, rcu); | |
195 | } | |
450af8d0 KS |
196 | } |
197 | ||
198 | void bpf_selem_link_storage_nolock(struct bpf_local_storage *local_storage, | |
199 | struct bpf_local_storage_elem *selem) | |
200 | { | |
201 | RCU_INIT_POINTER(selem->local_storage, local_storage); | |
70b97111 | 202 | hlist_add_head_rcu(&selem->snode, &local_storage->list); |
450af8d0 KS |
203 | } |
204 | ||
205 | void bpf_selem_unlink_map(struct bpf_local_storage_elem *selem) | |
206 | { | |
207 | struct bpf_local_storage_map *smap; | |
208 | struct bpf_local_storage_map_bucket *b; | |
a10787e6 | 209 | unsigned long flags; |
450af8d0 KS |
210 | |
211 | if (unlikely(!selem_linked_to_map(selem))) | |
212 | /* selem has already be unlinked from smap */ | |
213 | return; | |
214 | ||
0fe4b381 | 215 | smap = rcu_dereference_check(SDATA(selem)->smap, bpf_rcu_lock_held()); |
450af8d0 | 216 | b = select_bucket(smap, selem); |
a10787e6 | 217 | raw_spin_lock_irqsave(&b->lock, flags); |
450af8d0 KS |
218 | if (likely(selem_linked_to_map(selem))) |
219 | hlist_del_init_rcu(&selem->map_node); | |
a10787e6 | 220 | raw_spin_unlock_irqrestore(&b->lock, flags); |
450af8d0 KS |
221 | } |
222 | ||
223 | void bpf_selem_link_map(struct bpf_local_storage_map *smap, | |
224 | struct bpf_local_storage_elem *selem) | |
225 | { | |
226 | struct bpf_local_storage_map_bucket *b = select_bucket(smap, selem); | |
a10787e6 | 227 | unsigned long flags; |
450af8d0 | 228 | |
a10787e6 | 229 | raw_spin_lock_irqsave(&b->lock, flags); |
450af8d0 KS |
230 | RCU_INIT_POINTER(SDATA(selem)->smap, smap); |
231 | hlist_add_head_rcu(&selem->map_node, &b->list); | |
a10787e6 | 232 | raw_spin_unlock_irqrestore(&b->lock, flags); |
450af8d0 KS |
233 | } |
234 | ||
dcf456c9 | 235 | void bpf_selem_unlink(struct bpf_local_storage_elem *selem, bool use_trace_rcu) |
450af8d0 KS |
236 | { |
237 | /* Always unlink from map before unlinking from local_storage | |
238 | * because selem will be freed after successfully unlinked from | |
239 | * the local_storage. | |
240 | */ | |
241 | bpf_selem_unlink_map(selem); | |
dcf456c9 | 242 | __bpf_selem_unlink_storage(selem, use_trace_rcu); |
450af8d0 KS |
243 | } |
244 | ||
e8b02296 | 245 | /* If cacheit_lockit is false, this lookup function is lockless */ |
450af8d0 KS |
246 | struct bpf_local_storage_data * |
247 | bpf_local_storage_lookup(struct bpf_local_storage *local_storage, | |
248 | struct bpf_local_storage_map *smap, | |
249 | bool cacheit_lockit) | |
250 | { | |
251 | struct bpf_local_storage_data *sdata; | |
252 | struct bpf_local_storage_elem *selem; | |
253 | ||
254 | /* Fast path (cache hit) */ | |
0fe4b381 KS |
255 | sdata = rcu_dereference_check(local_storage->cache[smap->cache_idx], |
256 | bpf_rcu_lock_held()); | |
450af8d0 KS |
257 | if (sdata && rcu_access_pointer(sdata->smap) == smap) |
258 | return sdata; | |
259 | ||
260 | /* Slow path (cache miss) */ | |
0fe4b381 KS |
261 | hlist_for_each_entry_rcu(selem, &local_storage->list, snode, |
262 | rcu_read_lock_trace_held()) | |
450af8d0 KS |
263 | if (rcu_access_pointer(SDATA(selem)->smap) == smap) |
264 | break; | |
265 | ||
266 | if (!selem) | |
267 | return NULL; | |
268 | ||
269 | sdata = SDATA(selem); | |
270 | if (cacheit_lockit) { | |
a10787e6 SL |
271 | unsigned long flags; |
272 | ||
450af8d0 KS |
273 | /* spinlock is needed to avoid racing with the |
274 | * parallel delete. Otherwise, publishing an already | |
275 | * deleted sdata to the cache will become a use-after-free | |
276 | * problem in the next bpf_local_storage_lookup(). | |
277 | */ | |
a10787e6 | 278 | raw_spin_lock_irqsave(&local_storage->lock, flags); |
450af8d0 KS |
279 | if (selem_linked_to_storage(selem)) |
280 | rcu_assign_pointer(local_storage->cache[smap->cache_idx], | |
281 | sdata); | |
a10787e6 | 282 | raw_spin_unlock_irqrestore(&local_storage->lock, flags); |
450af8d0 KS |
283 | } |
284 | ||
285 | return sdata; | |
286 | } | |
287 | ||
288 | static int check_flags(const struct bpf_local_storage_data *old_sdata, | |
289 | u64 map_flags) | |
290 | { | |
291 | if (old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST) | |
292 | /* elem already exists */ | |
293 | return -EEXIST; | |
294 | ||
295 | if (!old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_EXIST) | |
296 | /* elem doesn't exist, cannot update it */ | |
297 | return -ENOENT; | |
298 | ||
299 | return 0; | |
300 | } | |
301 | ||
302 | int bpf_local_storage_alloc(void *owner, | |
303 | struct bpf_local_storage_map *smap, | |
b00fa38a JK |
304 | struct bpf_local_storage_elem *first_selem, |
305 | gfp_t gfp_flags) | |
450af8d0 KS |
306 | { |
307 | struct bpf_local_storage *prev_storage, *storage; | |
308 | struct bpf_local_storage **owner_storage_ptr; | |
309 | int err; | |
310 | ||
311 | err = mem_charge(smap, owner, sizeof(*storage)); | |
312 | if (err) | |
313 | return err; | |
314 | ||
e9aae8be | 315 | storage = bpf_map_kzalloc(&smap->map, sizeof(*storage), |
b00fa38a | 316 | gfp_flags | __GFP_NOWARN); |
450af8d0 KS |
317 | if (!storage) { |
318 | err = -ENOMEM; | |
319 | goto uncharge; | |
320 | } | |
321 | ||
322 | INIT_HLIST_HEAD(&storage->list); | |
323 | raw_spin_lock_init(&storage->lock); | |
324 | storage->owner = owner; | |
325 | ||
326 | bpf_selem_link_storage_nolock(storage, first_selem); | |
327 | bpf_selem_link_map(smap, first_selem); | |
328 | ||
329 | owner_storage_ptr = | |
330 | (struct bpf_local_storage **)owner_storage(smap, owner); | |
331 | /* Publish storage to the owner. | |
332 | * Instead of using any lock of the kernel object (i.e. owner), | |
333 | * cmpxchg will work with any kernel object regardless what | |
334 | * the running context is, bh, irq...etc. | |
335 | * | |
336 | * From now on, the owner->storage pointer (e.g. sk->sk_bpf_storage) | |
337 | * is protected by the storage->lock. Hence, when freeing | |
338 | * the owner->storage, the storage->lock must be held before | |
339 | * setting owner->storage ptr to NULL. | |
340 | */ | |
341 | prev_storage = cmpxchg(owner_storage_ptr, NULL, storage); | |
342 | if (unlikely(prev_storage)) { | |
343 | bpf_selem_unlink_map(first_selem); | |
344 | err = -EAGAIN; | |
345 | goto uncharge; | |
346 | ||
347 | /* Note that even first_selem was linked to smap's | |
348 | * bucket->list, first_selem can be freed immediately | |
349 | * (instead of kfree_rcu) because | |
350 | * bpf_local_storage_map_free() does a | |
0fe4b381 KS |
351 | * synchronize_rcu_mult (waiting for both sleepable and |
352 | * normal programs) before walking the bucket->list. | |
450af8d0 KS |
353 | * Hence, no one is accessing selem from the |
354 | * bucket->list under rcu_read_lock(). | |
355 | */ | |
356 | } | |
357 | ||
358 | return 0; | |
359 | ||
360 | uncharge: | |
361 | kfree(storage); | |
362 | mem_uncharge(smap, owner, sizeof(*storage)); | |
363 | return err; | |
364 | } | |
365 | ||
366 | /* sk cannot be going away because it is linking new elem | |
367 | * to sk->sk_bpf_storage. (i.e. sk->sk_refcnt cannot be 0). | |
368 | * Otherwise, it will become a leak (and other memory issues | |
369 | * during map destruction). | |
370 | */ | |
371 | struct bpf_local_storage_data * | |
372 | bpf_local_storage_update(void *owner, struct bpf_local_storage_map *smap, | |
b00fa38a | 373 | void *value, u64 map_flags, gfp_t gfp_flags) |
450af8d0 KS |
374 | { |
375 | struct bpf_local_storage_data *old_sdata = NULL; | |
b00fa38a | 376 | struct bpf_local_storage_elem *selem = NULL; |
450af8d0 | 377 | struct bpf_local_storage *local_storage; |
a10787e6 | 378 | unsigned long flags; |
450af8d0 KS |
379 | int err; |
380 | ||
381 | /* BPF_EXIST and BPF_NOEXIST cannot be both set */ | |
382 | if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST) || | |
383 | /* BPF_F_LOCK can only be used in a value with spin_lock */ | |
384 | unlikely((map_flags & BPF_F_LOCK) && | |
db559117 | 385 | !btf_record_has_field(smap->map.record, BPF_SPIN_LOCK))) |
450af8d0 KS |
386 | return ERR_PTR(-EINVAL); |
387 | ||
b00fa38a JK |
388 | if (gfp_flags == GFP_KERNEL && (map_flags & ~BPF_F_LOCK) != BPF_NOEXIST) |
389 | return ERR_PTR(-EINVAL); | |
390 | ||
0fe4b381 KS |
391 | local_storage = rcu_dereference_check(*owner_storage(smap, owner), |
392 | bpf_rcu_lock_held()); | |
450af8d0 KS |
393 | if (!local_storage || hlist_empty(&local_storage->list)) { |
394 | /* Very first elem for the owner */ | |
395 | err = check_flags(NULL, map_flags); | |
396 | if (err) | |
397 | return ERR_PTR(err); | |
398 | ||
b00fa38a | 399 | selem = bpf_selem_alloc(smap, owner, value, true, gfp_flags); |
450af8d0 KS |
400 | if (!selem) |
401 | return ERR_PTR(-ENOMEM); | |
402 | ||
b00fa38a | 403 | err = bpf_local_storage_alloc(owner, smap, selem, gfp_flags); |
450af8d0 KS |
404 | if (err) { |
405 | kfree(selem); | |
406 | mem_uncharge(smap, owner, smap->elem_size); | |
407 | return ERR_PTR(err); | |
408 | } | |
409 | ||
410 | return SDATA(selem); | |
411 | } | |
412 | ||
413 | if ((map_flags & BPF_F_LOCK) && !(map_flags & BPF_NOEXIST)) { | |
414 | /* Hoping to find an old_sdata to do inline update | |
415 | * such that it can avoid taking the local_storage->lock | |
416 | * and changing the lists. | |
417 | */ | |
418 | old_sdata = | |
419 | bpf_local_storage_lookup(local_storage, smap, false); | |
420 | err = check_flags(old_sdata, map_flags); | |
421 | if (err) | |
422 | return ERR_PTR(err); | |
423 | if (old_sdata && selem_linked_to_storage(SELEM(old_sdata))) { | |
424 | copy_map_value_locked(&smap->map, old_sdata->data, | |
425 | value, false); | |
426 | return old_sdata; | |
427 | } | |
428 | } | |
429 | ||
b00fa38a JK |
430 | if (gfp_flags == GFP_KERNEL) { |
431 | selem = bpf_selem_alloc(smap, owner, value, true, gfp_flags); | |
432 | if (!selem) | |
433 | return ERR_PTR(-ENOMEM); | |
434 | } | |
435 | ||
a10787e6 | 436 | raw_spin_lock_irqsave(&local_storage->lock, flags); |
450af8d0 KS |
437 | |
438 | /* Recheck local_storage->list under local_storage->lock */ | |
439 | if (unlikely(hlist_empty(&local_storage->list))) { | |
440 | /* A parallel del is happening and local_storage is going | |
441 | * away. It has just been checked before, so very | |
442 | * unlikely. Return instead of retry to keep things | |
443 | * simple. | |
444 | */ | |
445 | err = -EAGAIN; | |
446 | goto unlock_err; | |
447 | } | |
448 | ||
449 | old_sdata = bpf_local_storage_lookup(local_storage, smap, false); | |
450 | err = check_flags(old_sdata, map_flags); | |
451 | if (err) | |
452 | goto unlock_err; | |
453 | ||
454 | if (old_sdata && (map_flags & BPF_F_LOCK)) { | |
455 | copy_map_value_locked(&smap->map, old_sdata->data, value, | |
456 | false); | |
457 | selem = SELEM(old_sdata); | |
458 | goto unlock; | |
459 | } | |
460 | ||
b00fa38a JK |
461 | if (gfp_flags != GFP_KERNEL) { |
462 | /* local_storage->lock is held. Hence, we are sure | |
463 | * we can unlink and uncharge the old_sdata successfully | |
464 | * later. Hence, instead of charging the new selem now | |
465 | * and then uncharge the old selem later (which may cause | |
466 | * a potential but unnecessary charge failure), avoid taking | |
467 | * a charge at all here (the "!old_sdata" check) and the | |
468 | * old_sdata will not be uncharged later during | |
469 | * bpf_selem_unlink_storage_nolock(). | |
470 | */ | |
471 | selem = bpf_selem_alloc(smap, owner, value, !old_sdata, gfp_flags); | |
472 | if (!selem) { | |
473 | err = -ENOMEM; | |
474 | goto unlock_err; | |
475 | } | |
450af8d0 KS |
476 | } |
477 | ||
478 | /* First, link the new selem to the map */ | |
479 | bpf_selem_link_map(smap, selem); | |
480 | ||
481 | /* Second, link (and publish) the new selem to local_storage */ | |
482 | bpf_selem_link_storage_nolock(local_storage, selem); | |
483 | ||
484 | /* Third, remove old selem, SELEM(old_sdata) */ | |
485 | if (old_sdata) { | |
486 | bpf_selem_unlink_map(SELEM(old_sdata)); | |
487 | bpf_selem_unlink_storage_nolock(local_storage, SELEM(old_sdata), | |
dcf456c9 | 488 | false, true); |
450af8d0 KS |
489 | } |
490 | ||
491 | unlock: | |
a10787e6 | 492 | raw_spin_unlock_irqrestore(&local_storage->lock, flags); |
450af8d0 KS |
493 | return SDATA(selem); |
494 | ||
495 | unlock_err: | |
a10787e6 | 496 | raw_spin_unlock_irqrestore(&local_storage->lock, flags); |
b00fa38a JK |
497 | if (selem) { |
498 | mem_uncharge(smap, owner, smap->elem_size); | |
499 | kfree(selem); | |
500 | } | |
450af8d0 KS |
501 | return ERR_PTR(err); |
502 | } | |
503 | ||
c83597fa | 504 | static u16 bpf_local_storage_cache_idx_get(struct bpf_local_storage_cache *cache) |
450af8d0 KS |
505 | { |
506 | u64 min_usage = U64_MAX; | |
507 | u16 i, res = 0; | |
508 | ||
509 | spin_lock(&cache->idx_lock); | |
510 | ||
511 | for (i = 0; i < BPF_LOCAL_STORAGE_CACHE_SIZE; i++) { | |
512 | if (cache->idx_usage_counts[i] < min_usage) { | |
513 | min_usage = cache->idx_usage_counts[i]; | |
514 | res = i; | |
515 | ||
516 | /* Found a free cache_idx */ | |
517 | if (!min_usage) | |
518 | break; | |
519 | } | |
520 | } | |
521 | cache->idx_usage_counts[res]++; | |
522 | ||
523 | spin_unlock(&cache->idx_lock); | |
524 | ||
525 | return res; | |
526 | } | |
527 | ||
c83597fa YS |
528 | static void bpf_local_storage_cache_idx_free(struct bpf_local_storage_cache *cache, |
529 | u16 idx) | |
450af8d0 KS |
530 | { |
531 | spin_lock(&cache->idx_lock); | |
532 | cache->idx_usage_counts[idx]--; | |
533 | spin_unlock(&cache->idx_lock); | |
534 | } | |
535 | ||
450af8d0 KS |
536 | int bpf_local_storage_map_alloc_check(union bpf_attr *attr) |
537 | { | |
538 | if (attr->map_flags & ~BPF_LOCAL_STORAGE_CREATE_FLAG_MASK || | |
539 | !(attr->map_flags & BPF_F_NO_PREALLOC) || | |
540 | attr->max_entries || | |
541 | attr->key_size != sizeof(int) || !attr->value_size || | |
542 | /* Enforce BTF for userspace sk dumping */ | |
543 | !attr->btf_key_type_id || !attr->btf_value_type_id) | |
544 | return -EINVAL; | |
545 | ||
546 | if (!bpf_capable()) | |
547 | return -EPERM; | |
548 | ||
549 | if (attr->value_size > BPF_LOCAL_STORAGE_MAX_VALUE_SIZE) | |
550 | return -E2BIG; | |
551 | ||
552 | return 0; | |
553 | } | |
554 | ||
c83597fa | 555 | static struct bpf_local_storage_map *__bpf_local_storage_map_alloc(union bpf_attr *attr) |
450af8d0 KS |
556 | { |
557 | struct bpf_local_storage_map *smap; | |
558 | unsigned int i; | |
559 | u32 nbuckets; | |
450af8d0 | 560 | |
73cf09a3 | 561 | smap = bpf_map_area_alloc(sizeof(*smap), NUMA_NO_NODE); |
450af8d0 KS |
562 | if (!smap) |
563 | return ERR_PTR(-ENOMEM); | |
564 | bpf_map_init_from_attr(&smap->map, attr); | |
565 | ||
566 | nbuckets = roundup_pow_of_two(num_possible_cpus()); | |
567 | /* Use at least 2 buckets, select_bucket() is undefined behavior with 1 bucket */ | |
568 | nbuckets = max_t(u32, 2, nbuckets); | |
569 | smap->bucket_log = ilog2(nbuckets); | |
450af8d0 | 570 | |
ddef81b5 YS |
571 | smap->buckets = bpf_map_kvcalloc(&smap->map, sizeof(*smap->buckets), |
572 | nbuckets, GFP_USER | __GFP_NOWARN); | |
450af8d0 | 573 | if (!smap->buckets) { |
73cf09a3 | 574 | bpf_map_area_free(smap); |
450af8d0 KS |
575 | return ERR_PTR(-ENOMEM); |
576 | } | |
577 | ||
578 | for (i = 0; i < nbuckets; i++) { | |
579 | INIT_HLIST_HEAD(&smap->buckets[i].list); | |
580 | raw_spin_lock_init(&smap->buckets[i].lock); | |
581 | } | |
582 | ||
552d42a3 MKL |
583 | smap->elem_size = offsetof(struct bpf_local_storage_elem, |
584 | sdata.data[attr->value_size]); | |
450af8d0 KS |
585 | |
586 | return smap; | |
587 | } | |
588 | ||
589 | int bpf_local_storage_map_check_btf(const struct bpf_map *map, | |
590 | const struct btf *btf, | |
591 | const struct btf_type *key_type, | |
592 | const struct btf_type *value_type) | |
593 | { | |
594 | u32 int_data; | |
595 | ||
596 | if (BTF_INFO_KIND(key_type->info) != BTF_KIND_INT) | |
597 | return -EINVAL; | |
598 | ||
599 | int_data = *(u32 *)(key_type + 1); | |
600 | if (BTF_INT_BITS(int_data) != 32 || BTF_INT_OFFSET(int_data)) | |
601 | return -EINVAL; | |
602 | ||
603 | return 0; | |
604 | } | |
c83597fa YS |
605 | |
606 | bool bpf_local_storage_unlink_nolock(struct bpf_local_storage *local_storage) | |
607 | { | |
608 | struct bpf_local_storage_elem *selem; | |
609 | bool free_storage = false; | |
610 | struct hlist_node *n; | |
611 | ||
612 | /* Neither the bpf_prog nor the bpf_map's syscall | |
613 | * could be modifying the local_storage->list now. | |
614 | * Thus, no elem can be added to or deleted from the | |
615 | * local_storage->list by the bpf_prog or by the bpf_map's syscall. | |
616 | * | |
617 | * It is racing with bpf_local_storage_map_free() alone | |
618 | * when unlinking elem from the local_storage->list and | |
619 | * the map's bucket->list. | |
620 | */ | |
621 | hlist_for_each_entry_safe(selem, n, &local_storage->list, snode) { | |
622 | /* Always unlink from map before unlinking from | |
623 | * local_storage. | |
624 | */ | |
625 | bpf_selem_unlink_map(selem); | |
626 | /* If local_storage list has only one element, the | |
627 | * bpf_selem_unlink_storage_nolock() will return true. | |
628 | * Otherwise, it will return false. The current loop iteration | |
629 | * intends to remove all local storage. So the last iteration | |
630 | * of the loop will set the free_cgroup_storage to true. | |
631 | */ | |
632 | free_storage = bpf_selem_unlink_storage_nolock( | |
633 | local_storage, selem, false, false); | |
634 | } | |
635 | ||
636 | return free_storage; | |
637 | } | |
638 | ||
639 | struct bpf_map * | |
640 | bpf_local_storage_map_alloc(union bpf_attr *attr, | |
641 | struct bpf_local_storage_cache *cache) | |
642 | { | |
643 | struct bpf_local_storage_map *smap; | |
644 | ||
645 | smap = __bpf_local_storage_map_alloc(attr); | |
646 | if (IS_ERR(smap)) | |
647 | return ERR_CAST(smap); | |
648 | ||
649 | smap->cache_idx = bpf_local_storage_cache_idx_get(cache); | |
650 | return &smap->map; | |
651 | } | |
652 | ||
653 | void bpf_local_storage_map_free(struct bpf_map *map, | |
654 | struct bpf_local_storage_cache *cache, | |
655 | int __percpu *busy_counter) | |
656 | { | |
657 | struct bpf_local_storage_map_bucket *b; | |
658 | struct bpf_local_storage_elem *selem; | |
659 | struct bpf_local_storage_map *smap; | |
660 | unsigned int i; | |
661 | ||
662 | smap = (struct bpf_local_storage_map *)map; | |
663 | bpf_local_storage_cache_idx_free(cache, smap->cache_idx); | |
664 | ||
665 | /* Note that this map might be concurrently cloned from | |
666 | * bpf_sk_storage_clone. Wait for any existing bpf_sk_storage_clone | |
667 | * RCU read section to finish before proceeding. New RCU | |
668 | * read sections should be prevented via bpf_map_inc_not_zero. | |
669 | */ | |
670 | synchronize_rcu(); | |
671 | ||
672 | /* bpf prog and the userspace can no longer access this map | |
673 | * now. No new selem (of this map) can be added | |
674 | * to the owner->storage or to the map bucket's list. | |
675 | * | |
676 | * The elem of this map can be cleaned up here | |
677 | * or when the storage is freed e.g. | |
678 | * by bpf_sk_storage_free() during __sk_destruct(). | |
679 | */ | |
680 | for (i = 0; i < (1U << smap->bucket_log); i++) { | |
681 | b = &smap->buckets[i]; | |
682 | ||
683 | rcu_read_lock(); | |
684 | /* No one is adding to b->list now */ | |
685 | while ((selem = hlist_entry_safe( | |
686 | rcu_dereference_raw(hlist_first_rcu(&b->list)), | |
687 | struct bpf_local_storage_elem, map_node))) { | |
688 | if (busy_counter) { | |
689 | migrate_disable(); | |
690 | this_cpu_inc(*busy_counter); | |
691 | } | |
692 | bpf_selem_unlink(selem, false); | |
693 | if (busy_counter) { | |
694 | this_cpu_dec(*busy_counter); | |
695 | migrate_enable(); | |
696 | } | |
697 | cond_resched_rcu(); | |
698 | } | |
699 | rcu_read_unlock(); | |
700 | } | |
701 | ||
702 | /* While freeing the storage we may still need to access the map. | |
703 | * | |
704 | * e.g. when bpf_sk_storage_free() has unlinked selem from the map | |
705 | * which then made the above while((selem = ...)) loop | |
706 | * exit immediately. | |
707 | * | |
708 | * However, while freeing the storage one still needs to access the | |
709 | * smap->elem_size to do the uncharging in | |
710 | * bpf_selem_unlink_storage_nolock(). | |
711 | * | |
712 | * Hence, wait another rcu grace period for the storage to be freed. | |
713 | */ | |
714 | synchronize_rcu(); | |
715 | ||
716 | kvfree(smap->buckets); | |
717 | bpf_map_area_free(smap); | |
718 | } |