projects/linux·blob v6.4 · refs

719 lines · 19.7 KiB

1// SPDX-License-Identifier: GPL-2.0-only
2/* Copyright (c) 2022 Meta Platforms, Inc. and affiliates. */
3#include <linux/mm.h>
4#include <linux/llist.h>
5#include <linux/bpf.h>
6#include <linux/irq_work.h>
7#include <linux/bpf_mem_alloc.h>
8#include <linux/memcontrol.h>
9#include <asm/local.h>
10
11/* Any context (including NMI) BPF specific memory allocator.
12 *
13 * Tracing BPF programs can attach to kprobe and fentry. Hence they
14 * run in unknown context where calling plain kmalloc() might not be safe.
15 *
16 * Front-end kmalloc() with per-cpu per-bucket cache of free elements.
17 * Refill this cache asynchronously from irq_work.
18 *
19 * CPU_0 buckets
20 * 16 32 64 96 128 196 256 512 1024 2048 4096
21 * ...
22 * CPU_N buckets
23 * 16 32 64 96 128 196 256 512 1024 2048 4096
24 *
25 * The buckets are prefilled at the start.
26 * BPF programs always run with migration disabled.
27 * It's safe to allocate from cache of the current cpu with irqs disabled.
28 * Free-ing is always done into bucket of the current cpu as well.
29 * irq_work trims extra free elements from buckets with kfree
30 * and refills them with kmalloc, so global kmalloc logic takes care
31 * of freeing objects allocated by one cpu and freed on another.
32 *
33 * Every allocated objected is padded with extra 8 bytes that contains
34 * struct llist_node.
35 */
36#define LLIST_NODE_SZ sizeof(struct llist_node)
37
38/* similar to kmalloc, but sizeof == 8 bucket is gone */
39static u8 size_index[24] __ro_after_init = {
40 3, /* 8 */
41 3, /* 16 */
42 4, /* 24 */
43 4, /* 32 */
44 5, /* 40 */
45 5, /* 48 */
46 5, /* 56 */
47 5, /* 64 */
48 1, /* 72 */
49 1, /* 80 */
50 1, /* 88 */
51 1, /* 96 */
52 6, /* 104 */
53 6, /* 112 */
54 6, /* 120 */
55 6, /* 128 */
56 2, /* 136 */
57 2, /* 144 */
58 2, /* 152 */
59 2, /* 160 */
60 2, /* 168 */
61 2, /* 176 */
62 2, /* 184 */
63 2 /* 192 */
64};
65
66static int bpf_mem_cache_idx(size_t size)
67{
68 if (!size || size > 4096)
69 return -1;
70
71 if (size <= 192)
72 return size_index[(size - 1) / 8] - 1;
73
74 return fls(size - 1) - 2;
75}
76
77#define NUM_CACHES 11
78
79struct bpf_mem_cache {
80 /* per-cpu list of free objects of size 'unit_size'.
81 * All accesses are done with interrupts disabled and 'active' counter
82 * protection with __llist_add() and __llist_del_first().
83 */
84 struct llist_head free_llist;
85 local_t active;
86
87 /* Operations on the free_list from unit_alloc/unit_free/bpf_mem_refill
88 * are sequenced by per-cpu 'active' counter. But unit_free() cannot
89 * fail. When 'active' is busy the unit_free() will add an object to
90 * free_llist_extra.
91 */
92 struct llist_head free_llist_extra;
93
94 struct irq_work refill_work;
95 struct obj_cgroup *objcg;
96 int unit_size;
97 /* count of objects in free_llist */
98 int free_cnt;
99 int low_watermark, high_watermark, batch;
100 int percpu_size;
101
102 struct rcu_head rcu;
103 struct llist_head free_by_rcu;
104 struct llist_head waiting_for_gp;
105 atomic_t call_rcu_in_progress;
106};
107
108struct bpf_mem_caches {
109 struct bpf_mem_cache cache[NUM_CACHES];
110};
111
112static struct llist_node notrace *__llist_del_first(struct llist_head *head)
113{
114 struct llist_node *entry, *next;
115
116 entry = head->first;
117 if (!entry)
118 return NULL;
119 next = entry->next;
120 head->first = next;
121 return entry;
122}
123
124static void *__alloc(struct bpf_mem_cache *c, int node, gfp_t flags)
125{
126 if (c->percpu_size) {
127 void **obj = kmalloc_node(c->percpu_size, flags, node);
128 void *pptr = __alloc_percpu_gfp(c->unit_size, 8, flags);
129
130 if (!obj || !pptr) {
131 free_percpu(pptr);
132 kfree(obj);
133 return NULL;
134 }
135 obj[1] = pptr;
136 return obj;
137 }
138
139 return kmalloc_node(c->unit_size, flags | __GFP_ZERO, node);
140}
141
142static struct mem_cgroup *get_memcg(const struct bpf_mem_cache *c)
143{
144#ifdef CONFIG_MEMCG_KMEM
145 if (c->objcg)
146 return get_mem_cgroup_from_objcg(c->objcg);
147#endif
148
149#ifdef CONFIG_MEMCG
150 return root_mem_cgroup;
151#else
152 return NULL;
153#endif
154}
155
156/* Mostly runs from irq_work except __init phase. */
157static void alloc_bulk(struct bpf_mem_cache *c, int cnt, int node)
158{
159 struct mem_cgroup *memcg = NULL, *old_memcg;
160 unsigned long flags;
161 void *obj;
162 int i;
163
164 memcg = get_memcg(c);
165 old_memcg = set_active_memcg(memcg);
166 for (i = 0; i < cnt; i++) {
167 /*
168 * free_by_rcu is only manipulated by irq work refill_work().
169 * IRQ works on the same CPU are called sequentially, so it is
170 * safe to use __llist_del_first() here. If alloc_bulk() is
171 * invoked by the initial prefill, there will be no running
172 * refill_work(), so __llist_del_first() is fine as well.
173 *
174 * In most cases, objects on free_by_rcu are from the same CPU.
175 * If some objects come from other CPUs, it doesn't incur any
176 * harm because NUMA_NO_NODE means the preference for current
177 * numa node and it is not a guarantee.
178 */
179 obj = __llist_del_first(&c->free_by_rcu);
180 if (!obj) {
181 /* Allocate, but don't deplete atomic reserves that typical
182 * GFP_ATOMIC would do. irq_work runs on this cpu and kmalloc
183 * will allocate from the current numa node which is what we
184 * want here.
185 */
186 obj = __alloc(c, node, GFP_NOWAIT | __GFP_NOWARN | __GFP_ACCOUNT);
187 if (!obj)
188 break;
189 }
190 if (IS_ENABLED(CONFIG_PREEMPT_RT))
191 /* In RT irq_work runs in per-cpu kthread, so disable
192 * interrupts to avoid preemption and interrupts and
193 * reduce the chance of bpf prog executing on this cpu
194 * when active counter is busy.
195 */
196 local_irq_save(flags);
197 /* alloc_bulk runs from irq_work which will not preempt a bpf
198 * program that does unit_alloc/unit_free since IRQs are
199 * disabled there. There is no race to increment 'active'
200 * counter. It protects free_llist from corruption in case NMI
201 * bpf prog preempted this loop.
202 */
203 WARN_ON_ONCE(local_inc_return(&c->active) != 1);
204 __llist_add(obj, &c->free_llist);
205 c->free_cnt++;
206 local_dec(&c->active);
207 if (IS_ENABLED(CONFIG_PREEMPT_RT))
208 local_irq_restore(flags);
209 }
210 set_active_memcg(old_memcg);
211 mem_cgroup_put(memcg);
212}
213
214static void free_one(struct bpf_mem_cache *c, void *obj)
215{
216 if (c->percpu_size) {
217 free_percpu(((void **)obj)[1]);
218 kfree(obj);
219 return;
220 }
221
222 kfree(obj);
223}
224
225static void __free_rcu(struct rcu_head *head)
226{
227 struct bpf_mem_cache *c = container_of(head, struct bpf_mem_cache, rcu);
228 struct llist_node *llnode = llist_del_all(&c->waiting_for_gp);
229 struct llist_node *pos, *t;
230
231 llist_for_each_safe(pos, t, llnode)
232 free_one(c, pos);
233 atomic_set(&c->call_rcu_in_progress, 0);
234}
235
236static void __free_rcu_tasks_trace(struct rcu_head *head)
237{
238 /* If RCU Tasks Trace grace period implies RCU grace period,
239 * there is no need to invoke call_rcu().
240 */
241 if (rcu_trace_implies_rcu_gp())
242 __free_rcu(head);
243 else
244 call_rcu(head, __free_rcu);
245}
246
247static void enque_to_free(struct bpf_mem_cache *c, void *obj)
248{
249 struct llist_node *llnode = obj;
250
251 /* bpf_mem_cache is a per-cpu object. Freeing happens in irq_work.
252 * Nothing races to add to free_by_rcu list.
253 */
254 __llist_add(llnode, &c->free_by_rcu);
255}
256
257static void do_call_rcu(struct bpf_mem_cache *c)
258{
259 struct llist_node *llnode, *t;
260
261 if (atomic_xchg(&c->call_rcu_in_progress, 1))
262 return;
263
264 WARN_ON_ONCE(!llist_empty(&c->waiting_for_gp));
265 llist_for_each_safe(llnode, t, __llist_del_all(&c->free_by_rcu))
266 /* There is no concurrent __llist_add(waiting_for_gp) access.
267 * It doesn't race with llist_del_all either.
268 * But there could be two concurrent llist_del_all(waiting_for_gp):
269 * from __free_rcu() and from drain_mem_cache().
270 */
271 __llist_add(llnode, &c->waiting_for_gp);
272 /* Use call_rcu_tasks_trace() to wait for sleepable progs to finish.
273 * If RCU Tasks Trace grace period implies RCU grace period, free
274 * these elements directly, else use call_rcu() to wait for normal
275 * progs to finish and finally do free_one() on each element.
276 */
277 call_rcu_tasks_trace(&c->rcu, __free_rcu_tasks_trace);
278}
279
280static void free_bulk(struct bpf_mem_cache *c)
281{
282 struct llist_node *llnode, *t;
283 unsigned long flags;
284 int cnt;
285
286 do {
287 if (IS_ENABLED(CONFIG_PREEMPT_RT))
288 local_irq_save(flags);
289 WARN_ON_ONCE(local_inc_return(&c->active) != 1);
290 llnode = __llist_del_first(&c->free_llist);
291 if (llnode)
292 cnt = --c->free_cnt;
293 else
294 cnt = 0;
295 local_dec(&c->active);
296 if (IS_ENABLED(CONFIG_PREEMPT_RT))
297 local_irq_restore(flags);
298 if (llnode)
299 enque_to_free(c, llnode);
300 } while (cnt > (c->high_watermark + c->low_watermark) / 2);
301
302 /* and drain free_llist_extra */
303 llist_for_each_safe(llnode, t, llist_del_all(&c->free_llist_extra))
304 enque_to_free(c, llnode);
305 do_call_rcu(c);
306}
307
308static void bpf_mem_refill(struct irq_work *work)
309{
310 struct bpf_mem_cache *c = container_of(work, struct bpf_mem_cache, refill_work);
311 int cnt;
312
313 /* Racy access to free_cnt. It doesn't need to be 100% accurate */
314 cnt = c->free_cnt;
315 if (cnt < c->low_watermark)
316 /* irq_work runs on this cpu and kmalloc will allocate
317 * from the current numa node which is what we want here.
318 */
319 alloc_bulk(c, c->batch, NUMA_NO_NODE);
320 else if (cnt > c->high_watermark)
321 free_bulk(c);
322}
323
324static void notrace irq_work_raise(struct bpf_mem_cache *c)
325{
326 irq_work_queue(&c->refill_work);
327}
328
329/* For typical bpf map case that uses bpf_mem_cache_alloc and single bucket
330 * the freelist cache will be elem_size * 64 (or less) on each cpu.
331 *
332 * For bpf programs that don't have statically known allocation sizes and
333 * assuming (low_mark + high_mark) / 2 as an average number of elements per
334 * bucket and all buckets are used the total amount of memory in freelists
335 * on each cpu will be:
336 * 64*16 + 64*32 + 64*64 + 64*96 + 64*128 + 64*196 + 64*256 + 32*512 + 16*1024 + 8*2048 + 4*4096
337 * == ~ 116 Kbyte using below heuristic.
338 * Initialized, but unused bpf allocator (not bpf map specific one) will
339 * consume ~ 11 Kbyte per cpu.
340 * Typical case will be between 11K and 116K closer to 11K.
341 * bpf progs can and should share bpf_mem_cache when possible.
342 */
343
344static void prefill_mem_cache(struct bpf_mem_cache *c, int cpu)
345{
346 init_irq_work(&c->refill_work, bpf_mem_refill);
347 if (c->unit_size <= 256) {
348 c->low_watermark = 32;
349 c->high_watermark = 96;
350 } else {
351 /* When page_size == 4k, order-0 cache will have low_mark == 2
352 * and high_mark == 6 with batch alloc of 3 individual pages at
353 * a time.
354 * 8k allocs and above low == 1, high == 3, batch == 1.
355 */
356 c->low_watermark = max(32 * 256 / c->unit_size, 1);
357 c->high_watermark = max(96 * 256 / c->unit_size, 3);
358 }
359 c->batch = max((c->high_watermark - c->low_watermark) / 4 * 3, 1);
360
361 /* To avoid consuming memory assume that 1st run of bpf
362 * prog won't be doing more than 4 map_update_elem from
363 * irq disabled region
364 */
365 alloc_bulk(c, c->unit_size <= 256 ? 4 : 1, cpu_to_node(cpu));
366}
367
368/* When size != 0 bpf_mem_cache for each cpu.
369 * This is typical bpf hash map use case when all elements have equal size.
370 *
371 * When size == 0 allocate 11 bpf_mem_cache-s for each cpu, then rely on
372 * kmalloc/kfree. Max allocation size is 4096 in this case.
373 * This is bpf_dynptr and bpf_kptr use case.
374 */
375int bpf_mem_alloc_init(struct bpf_mem_alloc *ma, int size, bool percpu)
376{
377 static u16 sizes[NUM_CACHES] = {96, 192, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096};
378 struct bpf_mem_caches *cc, __percpu *pcc;
379 struct bpf_mem_cache *c, __percpu *pc;
380 struct obj_cgroup *objcg = NULL;
381 int cpu, i, unit_size, percpu_size = 0;
382
383 if (size) {
384 pc = __alloc_percpu_gfp(sizeof(*pc), 8, GFP_KERNEL);
385 if (!pc)
386 return -ENOMEM;
387
388 if (percpu)
389 /* room for llist_node and per-cpu pointer */
390 percpu_size = LLIST_NODE_SZ + sizeof(void *);
391 else
392 size += LLIST_NODE_SZ; /* room for llist_node */
393 unit_size = size;
394
395#ifdef CONFIG_MEMCG_KMEM
396 if (memcg_bpf_enabled())
397 objcg = get_obj_cgroup_from_current();
398#endif
399 for_each_possible_cpu(cpu) {
400 c = per_cpu_ptr(pc, cpu);
401 c->unit_size = unit_size;
402 c->objcg = objcg;
403 c->percpu_size = percpu_size;
404 prefill_mem_cache(c, cpu);
405 }
406 ma->cache = pc;
407 return 0;
408 }
409
410 /* size == 0 && percpu is an invalid combination */
411 if (WARN_ON_ONCE(percpu))
412 return -EINVAL;
413
414 pcc = __alloc_percpu_gfp(sizeof(*cc), 8, GFP_KERNEL);
415 if (!pcc)
416 return -ENOMEM;
417#ifdef CONFIG_MEMCG_KMEM
418 objcg = get_obj_cgroup_from_current();
419#endif
420 for_each_possible_cpu(cpu) {
421 cc = per_cpu_ptr(pcc, cpu);
422 for (i = 0; i < NUM_CACHES; i++) {
423 c = &cc->cache[i];
424 c->unit_size = sizes[i];
425 c->objcg = objcg;
426 prefill_mem_cache(c, cpu);
427 }
428 }
429 ma->caches = pcc;
430 return 0;
431}
432
433static void drain_mem_cache(struct bpf_mem_cache *c)
434{
435 struct llist_node *llnode, *t;
436
437 /* No progs are using this bpf_mem_cache, but htab_map_free() called
438 * bpf_mem_cache_free() for all remaining elements and they can be in
439 * free_by_rcu or in waiting_for_gp lists, so drain those lists now.
440 *
441 * Except for waiting_for_gp list, there are no concurrent operations
442 * on these lists, so it is safe to use __llist_del_all().
443 */
444 llist_for_each_safe(llnode, t, __llist_del_all(&c->free_by_rcu))
445 free_one(c, llnode);
446 llist_for_each_safe(llnode, t, llist_del_all(&c->waiting_for_gp))
447 free_one(c, llnode);
448 llist_for_each_safe(llnode, t, __llist_del_all(&c->free_llist))
449 free_one(c, llnode);
450 llist_for_each_safe(llnode, t, __llist_del_all(&c->free_llist_extra))
451 free_one(c, llnode);
452}
453
454static void free_mem_alloc_no_barrier(struct bpf_mem_alloc *ma)
455{
456 free_percpu(ma->cache);
457 free_percpu(ma->caches);
458 ma->cache = NULL;
459 ma->caches = NULL;
460}
461
462static void free_mem_alloc(struct bpf_mem_alloc *ma)
463{
464 /* waiting_for_gp lists was drained, but __free_rcu might
465 * still execute. Wait for it now before we freeing percpu caches.
466 *
467 * rcu_barrier_tasks_trace() doesn't imply synchronize_rcu_tasks_trace(),
468 * but rcu_barrier_tasks_trace() and rcu_barrier() below are only used
469 * to wait for the pending __free_rcu_tasks_trace() and __free_rcu(),
470 * so if call_rcu(head, __free_rcu) is skipped due to
471 * rcu_trace_implies_rcu_gp(), it will be OK to skip rcu_barrier() by
472 * using rcu_trace_implies_rcu_gp() as well.
473 */
474 rcu_barrier_tasks_trace();
475 if (!rcu_trace_implies_rcu_gp())
476 rcu_barrier();
477 free_mem_alloc_no_barrier(ma);
478}
479
480static void free_mem_alloc_deferred(struct work_struct *work)
481{
482 struct bpf_mem_alloc *ma = container_of(work, struct bpf_mem_alloc, work);
483
484 free_mem_alloc(ma);
485 kfree(ma);
486}
487
488static void destroy_mem_alloc(struct bpf_mem_alloc *ma, int rcu_in_progress)
489{
490 struct bpf_mem_alloc *copy;
491
492 if (!rcu_in_progress) {
493 /* Fast path. No callbacks are pending, hence no need to do
494 * rcu_barrier-s.
495 */
496 free_mem_alloc_no_barrier(ma);
497 return;
498 }
499
500 copy = kmalloc(sizeof(*ma), GFP_KERNEL);
501 if (!copy) {
502 /* Slow path with inline barrier-s */
503 free_mem_alloc(ma);
504 return;
505 }
506
507 /* Defer barriers into worker to let the rest of map memory to be freed */
508 copy->cache = ma->cache;
509 ma->cache = NULL;
510 copy->caches = ma->caches;
511 ma->caches = NULL;
512 INIT_WORK(&copy->work, free_mem_alloc_deferred);
513 queue_work(system_unbound_wq, &copy->work);
514}
515
516void bpf_mem_alloc_destroy(struct bpf_mem_alloc *ma)
517{
518 struct bpf_mem_caches *cc;
519 struct bpf_mem_cache *c;
520 int cpu, i, rcu_in_progress;
521
522 if (ma->cache) {
523 rcu_in_progress = 0;
524 for_each_possible_cpu(cpu) {
525 c = per_cpu_ptr(ma->cache, cpu);
526 /*
527 * refill_work may be unfinished for PREEMPT_RT kernel
528 * in which irq work is invoked in a per-CPU RT thread.
529 * It is also possible for kernel with
530 * arch_irq_work_has_interrupt() being false and irq
531 * work is invoked in timer interrupt. So waiting for
532 * the completion of irq work to ease the handling of
533 * concurrency.
534 */
535 irq_work_sync(&c->refill_work);
536 drain_mem_cache(c);
537 rcu_in_progress += atomic_read(&c->call_rcu_in_progress);
538 }
539 /* objcg is the same across cpus */
540 if (c->objcg)
541 obj_cgroup_put(c->objcg);
542 destroy_mem_alloc(ma, rcu_in_progress);
543 }
544 if (ma->caches) {
545 rcu_in_progress = 0;
546 for_each_possible_cpu(cpu) {
547 cc = per_cpu_ptr(ma->caches, cpu);
548 for (i = 0; i < NUM_CACHES; i++) {
549 c = &cc->cache[i];
550 irq_work_sync(&c->refill_work);
551 drain_mem_cache(c);
552 rcu_in_progress += atomic_read(&c->call_rcu_in_progress);
553 }
554 }
555 if (c->objcg)
556 obj_cgroup_put(c->objcg);
557 destroy_mem_alloc(ma, rcu_in_progress);
558 }
559}
560
561/* notrace is necessary here and in other functions to make sure
562 * bpf programs cannot attach to them and cause llist corruptions.
563 */
564static void notrace *unit_alloc(struct bpf_mem_cache *c)
565{
566 struct llist_node *llnode = NULL;
567 unsigned long flags;
568 int cnt = 0;
569
570 /* Disable irqs to prevent the following race for majority of prog types:
571 * prog_A
572 * bpf_mem_alloc
573 * preemption or irq -> prog_B
574 * bpf_mem_alloc
575 *
576 * but prog_B could be a perf_event NMI prog.
577 * Use per-cpu 'active' counter to order free_list access between
578 * unit_alloc/unit_free/bpf_mem_refill.
579 */
580 local_irq_save(flags);
581 if (local_inc_return(&c->active) == 1) {
582 llnode = __llist_del_first(&c->free_llist);
583 if (llnode)
584 cnt = --c->free_cnt;
585 }
586 local_dec(&c->active);
587 local_irq_restore(flags);
588
589 WARN_ON(cnt < 0);
590
591 if (cnt < c->low_watermark)
592 irq_work_raise(c);
593 return llnode;
594}
595
596/* Though 'ptr' object could have been allocated on a different cpu
597 * add it to the free_llist of the current cpu.
598 * Let kfree() logic deal with it when it's later called from irq_work.
599 */
600static void notrace unit_free(struct bpf_mem_cache *c, void *ptr)
601{
602 struct llist_node *llnode = ptr - LLIST_NODE_SZ;
603 unsigned long flags;
604 int cnt = 0;
605
606 BUILD_BUG_ON(LLIST_NODE_SZ > 8);
607
608 local_irq_save(flags);
609 if (local_inc_return(&c->active) == 1) {
610 __llist_add(llnode, &c->free_llist);
611 cnt = ++c->free_cnt;
612 } else {
613 /* unit_free() cannot fail. Therefore add an object to atomic
614 * llist. free_bulk() will drain it. Though free_llist_extra is
615 * a per-cpu list we have to use atomic llist_add here, since
616 * it also can be interrupted by bpf nmi prog that does another
617 * unit_free() into the same free_llist_extra.
618 */
619 llist_add(llnode, &c->free_llist_extra);
620 }
621 local_dec(&c->active);
622 local_irq_restore(flags);
623
624 if (cnt > c->high_watermark)
625 /* free few objects from current cpu into global kmalloc pool */
626 irq_work_raise(c);
627}
628
629/* Called from BPF program or from sys_bpf syscall.
630 * In both cases migration is disabled.
631 */
632void notrace *bpf_mem_alloc(struct bpf_mem_alloc *ma, size_t size)
633{
634 int idx;
635 void *ret;
636
637 if (!size)
638 return ZERO_SIZE_PTR;
639
640 idx = bpf_mem_cache_idx(size + LLIST_NODE_SZ);
641 if (idx < 0)
642 return NULL;
643
644 ret = unit_alloc(this_cpu_ptr(ma->caches)->cache + idx);
645 return !ret ? NULL : ret + LLIST_NODE_SZ;
646}
647
648void notrace bpf_mem_free(struct bpf_mem_alloc *ma, void *ptr)
649{
650 int idx;
651
652 if (!ptr)
653 return;
654
655 idx = bpf_mem_cache_idx(ksize(ptr - LLIST_NODE_SZ));
656 if (idx < 0)
657 return;
658
659 unit_free(this_cpu_ptr(ma->caches)->cache + idx, ptr);
660}
661
662void notrace *bpf_mem_cache_alloc(struct bpf_mem_alloc *ma)
663{
664 void *ret;
665
666 ret = unit_alloc(this_cpu_ptr(ma->cache));
667 return !ret ? NULL : ret + LLIST_NODE_SZ;
668}
669
670void notrace bpf_mem_cache_free(struct bpf_mem_alloc *ma, void *ptr)
671{
672 if (!ptr)
673 return;
674
675 unit_free(this_cpu_ptr(ma->cache), ptr);
676}
677
678/* Directly does a kfree() without putting 'ptr' back to the free_llist
679 * for reuse and without waiting for a rcu_tasks_trace gp.
680 * The caller must first go through the rcu_tasks_trace gp for 'ptr'
681 * before calling bpf_mem_cache_raw_free().
682 * It could be used when the rcu_tasks_trace callback does not have
683 * a hold on the original bpf_mem_alloc object that allocated the
684 * 'ptr'. This should only be used in the uncommon code path.
685 * Otherwise, the bpf_mem_alloc's free_llist cannot be refilled
686 * and may affect performance.
687 */
688void bpf_mem_cache_raw_free(void *ptr)
689{
690 if (!ptr)
691 return;
692
693 kfree(ptr - LLIST_NODE_SZ);
694}
695
696/* When flags == GFP_KERNEL, it signals that the caller will not cause
697 * deadlock when using kmalloc. bpf_mem_cache_alloc_flags() will use
698 * kmalloc if the free_llist is empty.
699 */
700void notrace *bpf_mem_cache_alloc_flags(struct bpf_mem_alloc *ma, gfp_t flags)
701{
702 struct bpf_mem_cache *c;
703 void *ret;
704
705 c = this_cpu_ptr(ma->cache);
706
707 ret = unit_alloc(c);
708 if (!ret && flags == GFP_KERNEL) {
709 struct mem_cgroup *memcg, *old_memcg;
710
711 memcg = get_memcg(c);
712 old_memcg = set_active_memcg(memcg);
713 ret = __alloc(c, NUMA_NO_NODE, GFP_KERNEL | __GFP_NOWARN | __GFP_ACCOUNT);
714 set_active_memcg(old_memcg);
715 mem_cgroup_put(memcg);
716 }
717
718 return !ret ? NULL : ret + LLIST_NODE_SZ;
719}