1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2022 Meta Platforms, Inc. and affiliates. */
4 #include <linux/llist.h>
6 #include <linux/irq_work.h>
7 #include <linux/bpf_mem_alloc.h>
8 #include <linux/memcontrol.h>
11 /* Any context (including NMI) BPF specific memory allocator.
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.
16 * Front-end kmalloc() with per-cpu per-bucket cache of free elements.
17 * Refill this cache asynchronously from irq_work.
20 * 16 32 64 96 128 196 256 512 1024 2048 4096
23 * 16 32 64 96 128 196 256 512 1024 2048 4096
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.
33 * Every allocated objected is padded with extra 8 bytes that contains
36 #define LLIST_NODE_SZ sizeof(struct llist_node)
38 /* similar to kmalloc, but sizeof == 8 bucket is gone */
39 static u8 size_index[24] __ro_after_init = {
66 static int bpf_mem_cache_idx(size_t size)
68 if (!size || size > 4096)
72 return size_index[(size - 1) / 8] - 1;
74 return fls(size - 1) - 2;
79 struct 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().
84 struct llist_head free_llist;
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
92 struct llist_head free_llist_extra;
94 struct irq_work refill_work;
95 struct obj_cgroup *objcg;
97 /* count of objects in free_llist */
99 int low_watermark, high_watermark, batch;
102 struct bpf_mem_cache *tgt;
104 /* list of objects to be freed after RCU GP */
105 struct llist_head free_by_rcu;
106 struct llist_node *free_by_rcu_tail;
107 struct llist_head waiting_for_gp;
108 struct llist_node *waiting_for_gp_tail;
110 atomic_t call_rcu_in_progress;
111 struct llist_head free_llist_extra_rcu;
113 /* list of objects to be freed after RCU tasks trace GP */
114 struct llist_head free_by_rcu_ttrace;
115 struct llist_head waiting_for_gp_ttrace;
116 struct rcu_head rcu_ttrace;
117 atomic_t call_rcu_ttrace_in_progress;
120 struct bpf_mem_caches {
121 struct bpf_mem_cache cache[NUM_CACHES];
124 static struct llist_node notrace *__llist_del_first(struct llist_head *head)
126 struct llist_node *entry, *next;
136 static void *__alloc(struct bpf_mem_cache *c, int node, gfp_t flags)
138 if (c->percpu_size) {
139 void **obj = kmalloc_node(c->percpu_size, flags, node);
140 void *pptr = __alloc_percpu_gfp(c->unit_size, 8, flags);
151 return kmalloc_node(c->unit_size, flags | __GFP_ZERO, node);
154 static struct mem_cgroup *get_memcg(const struct bpf_mem_cache *c)
156 #ifdef CONFIG_MEMCG_KMEM
158 return get_mem_cgroup_from_objcg(c->objcg);
162 return root_mem_cgroup;
168 static void inc_active(struct bpf_mem_cache *c, unsigned long *flags)
170 if (IS_ENABLED(CONFIG_PREEMPT_RT))
171 /* In RT irq_work runs in per-cpu kthread, so disable
172 * interrupts to avoid preemption and interrupts and
173 * reduce the chance of bpf prog executing on this cpu
174 * when active counter is busy.
176 local_irq_save(*flags);
177 /* alloc_bulk runs from irq_work which will not preempt a bpf
178 * program that does unit_alloc/unit_free since IRQs are
179 * disabled there. There is no race to increment 'active'
180 * counter. It protects free_llist from corruption in case NMI
181 * bpf prog preempted this loop.
183 WARN_ON_ONCE(local_inc_return(&c->active) != 1);
186 static void dec_active(struct bpf_mem_cache *c, unsigned long *flags)
188 local_dec(&c->active);
189 if (IS_ENABLED(CONFIG_PREEMPT_RT))
190 local_irq_restore(*flags);
193 static void add_obj_to_free_list(struct bpf_mem_cache *c, void *obj)
197 inc_active(c, &flags);
198 __llist_add(obj, &c->free_llist);
200 dec_active(c, &flags);
203 /* Mostly runs from irq_work except __init phase. */
204 static void alloc_bulk(struct bpf_mem_cache *c, int cnt, int node, bool atomic)
206 struct mem_cgroup *memcg = NULL, *old_memcg;
211 gfp = __GFP_NOWARN | __GFP_ACCOUNT;
212 gfp |= atomic ? GFP_NOWAIT : GFP_KERNEL;
214 for (i = 0; i < cnt; i++) {
216 * For every 'c' llist_del_first(&c->free_by_rcu_ttrace); is
217 * done only by one CPU == current CPU. Other CPUs might
218 * llist_add() and llist_del_all() in parallel.
220 obj = llist_del_first(&c->free_by_rcu_ttrace);
223 add_obj_to_free_list(c, obj);
228 for (; i < cnt; i++) {
229 obj = llist_del_first(&c->waiting_for_gp_ttrace);
232 add_obj_to_free_list(c, obj);
237 memcg = get_memcg(c);
238 old_memcg = set_active_memcg(memcg);
239 for (; i < cnt; i++) {
240 /* Allocate, but don't deplete atomic reserves that typical
241 * GFP_ATOMIC would do. irq_work runs on this cpu and kmalloc
242 * will allocate from the current numa node which is what we
245 obj = __alloc(c, node, gfp);
248 add_obj_to_free_list(c, obj);
250 set_active_memcg(old_memcg);
251 mem_cgroup_put(memcg);
254 static void free_one(void *obj, bool percpu)
257 free_percpu(((void **)obj)[1]);
265 static int free_all(struct llist_node *llnode, bool percpu)
267 struct llist_node *pos, *t;
270 llist_for_each_safe(pos, t, llnode) {
271 free_one(pos, percpu);
277 static void __free_rcu(struct rcu_head *head)
279 struct bpf_mem_cache *c = container_of(head, struct bpf_mem_cache, rcu_ttrace);
281 free_all(llist_del_all(&c->waiting_for_gp_ttrace), !!c->percpu_size);
282 atomic_set(&c->call_rcu_ttrace_in_progress, 0);
285 static void __free_rcu_tasks_trace(struct rcu_head *head)
287 /* If RCU Tasks Trace grace period implies RCU grace period,
288 * there is no need to invoke call_rcu().
290 if (rcu_trace_implies_rcu_gp())
293 call_rcu(head, __free_rcu);
296 static void enque_to_free(struct bpf_mem_cache *c, void *obj)
298 struct llist_node *llnode = obj;
300 /* bpf_mem_cache is a per-cpu object. Freeing happens in irq_work.
301 * Nothing races to add to free_by_rcu_ttrace list.
303 llist_add(llnode, &c->free_by_rcu_ttrace);
306 static void do_call_rcu_ttrace(struct bpf_mem_cache *c)
308 struct llist_node *llnode, *t;
310 if (atomic_xchg(&c->call_rcu_ttrace_in_progress, 1)) {
311 if (unlikely(READ_ONCE(c->draining))) {
312 llnode = llist_del_all(&c->free_by_rcu_ttrace);
313 free_all(llnode, !!c->percpu_size);
318 WARN_ON_ONCE(!llist_empty(&c->waiting_for_gp_ttrace));
319 llist_for_each_safe(llnode, t, llist_del_all(&c->free_by_rcu_ttrace))
320 llist_add(llnode, &c->waiting_for_gp_ttrace);
322 if (unlikely(READ_ONCE(c->draining))) {
323 __free_rcu(&c->rcu_ttrace);
327 /* Use call_rcu_tasks_trace() to wait for sleepable progs to finish.
328 * If RCU Tasks Trace grace period implies RCU grace period, free
329 * these elements directly, else use call_rcu() to wait for normal
330 * progs to finish and finally do free_one() on each element.
332 call_rcu_tasks_trace(&c->rcu_ttrace, __free_rcu_tasks_trace);
335 static void free_bulk(struct bpf_mem_cache *c)
337 struct bpf_mem_cache *tgt = c->tgt;
338 struct llist_node *llnode, *t;
342 WARN_ON_ONCE(tgt->unit_size != c->unit_size);
345 inc_active(c, &flags);
346 llnode = __llist_del_first(&c->free_llist);
351 dec_active(c, &flags);
353 enque_to_free(tgt, llnode);
354 } while (cnt > (c->high_watermark + c->low_watermark) / 2);
356 /* and drain free_llist_extra */
357 llist_for_each_safe(llnode, t, llist_del_all(&c->free_llist_extra))
358 enque_to_free(tgt, llnode);
359 do_call_rcu_ttrace(tgt);
362 static void __free_by_rcu(struct rcu_head *head)
364 struct bpf_mem_cache *c = container_of(head, struct bpf_mem_cache, rcu);
365 struct bpf_mem_cache *tgt = c->tgt;
366 struct llist_node *llnode;
368 llnode = llist_del_all(&c->waiting_for_gp);
372 llist_add_batch(llnode, c->waiting_for_gp_tail, &tgt->free_by_rcu_ttrace);
374 /* Objects went through regular RCU GP. Send them to RCU tasks trace */
375 do_call_rcu_ttrace(tgt);
377 atomic_set(&c->call_rcu_in_progress, 0);
380 static void check_free_by_rcu(struct bpf_mem_cache *c)
382 struct llist_node *llnode, *t;
385 /* drain free_llist_extra_rcu */
386 if (unlikely(!llist_empty(&c->free_llist_extra_rcu))) {
387 inc_active(c, &flags);
388 llist_for_each_safe(llnode, t, llist_del_all(&c->free_llist_extra_rcu))
389 if (__llist_add(llnode, &c->free_by_rcu))
390 c->free_by_rcu_tail = llnode;
391 dec_active(c, &flags);
394 if (llist_empty(&c->free_by_rcu))
397 if (atomic_xchg(&c->call_rcu_in_progress, 1)) {
399 * Instead of kmalloc-ing new rcu_head and triggering 10k
400 * call_rcu() to hit rcutree.qhimark and force RCU to notice
401 * the overload just ask RCU to hurry up. There could be many
402 * objects in free_by_rcu list.
403 * This hint reduces memory consumption for an artificial
404 * benchmark from 2 Gbyte to 150 Mbyte.
406 rcu_request_urgent_qs_task(current);
410 WARN_ON_ONCE(!llist_empty(&c->waiting_for_gp));
412 inc_active(c, &flags);
413 WRITE_ONCE(c->waiting_for_gp.first, __llist_del_all(&c->free_by_rcu));
414 c->waiting_for_gp_tail = c->free_by_rcu_tail;
415 dec_active(c, &flags);
417 if (unlikely(READ_ONCE(c->draining))) {
418 free_all(llist_del_all(&c->waiting_for_gp), !!c->percpu_size);
419 atomic_set(&c->call_rcu_in_progress, 0);
421 call_rcu_hurry(&c->rcu, __free_by_rcu);
425 static void bpf_mem_refill(struct irq_work *work)
427 struct bpf_mem_cache *c = container_of(work, struct bpf_mem_cache, refill_work);
430 /* Racy access to free_cnt. It doesn't need to be 100% accurate */
432 if (cnt < c->low_watermark)
433 /* irq_work runs on this cpu and kmalloc will allocate
434 * from the current numa node which is what we want here.
436 alloc_bulk(c, c->batch, NUMA_NO_NODE, true);
437 else if (cnt > c->high_watermark)
440 check_free_by_rcu(c);
443 static void notrace irq_work_raise(struct bpf_mem_cache *c)
445 irq_work_queue(&c->refill_work);
448 /* For typical bpf map case that uses bpf_mem_cache_alloc and single bucket
449 * the freelist cache will be elem_size * 64 (or less) on each cpu.
451 * For bpf programs that don't have statically known allocation sizes and
452 * assuming (low_mark + high_mark) / 2 as an average number of elements per
453 * bucket and all buckets are used the total amount of memory in freelists
454 * on each cpu will be:
455 * 64*16 + 64*32 + 64*64 + 64*96 + 64*128 + 64*196 + 64*256 + 32*512 + 16*1024 + 8*2048 + 4*4096
456 * == ~ 116 Kbyte using below heuristic.
457 * Initialized, but unused bpf allocator (not bpf map specific one) will
458 * consume ~ 11 Kbyte per cpu.
459 * Typical case will be between 11K and 116K closer to 11K.
460 * bpf progs can and should share bpf_mem_cache when possible.
463 static void prefill_mem_cache(struct bpf_mem_cache *c, int cpu)
465 init_irq_work(&c->refill_work, bpf_mem_refill);
466 if (c->unit_size <= 256) {
467 c->low_watermark = 32;
468 c->high_watermark = 96;
470 /* When page_size == 4k, order-0 cache will have low_mark == 2
471 * and high_mark == 6 with batch alloc of 3 individual pages at
473 * 8k allocs and above low == 1, high == 3, batch == 1.
475 c->low_watermark = max(32 * 256 / c->unit_size, 1);
476 c->high_watermark = max(96 * 256 / c->unit_size, 3);
478 c->batch = max((c->high_watermark - c->low_watermark) / 4 * 3, 1);
480 /* To avoid consuming memory assume that 1st run of bpf
481 * prog won't be doing more than 4 map_update_elem from
482 * irq disabled region
484 alloc_bulk(c, c->unit_size <= 256 ? 4 : 1, cpu_to_node(cpu), false);
487 /* When size != 0 bpf_mem_cache for each cpu.
488 * This is typical bpf hash map use case when all elements have equal size.
490 * When size == 0 allocate 11 bpf_mem_cache-s for each cpu, then rely on
491 * kmalloc/kfree. Max allocation size is 4096 in this case.
492 * This is bpf_dynptr and bpf_kptr use case.
494 int bpf_mem_alloc_init(struct bpf_mem_alloc *ma, int size, bool percpu)
496 static u16 sizes[NUM_CACHES] = {96, 192, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096};
497 struct bpf_mem_caches *cc, __percpu *pcc;
498 struct bpf_mem_cache *c, __percpu *pc;
499 struct obj_cgroup *objcg = NULL;
500 int cpu, i, unit_size, percpu_size = 0;
503 pc = __alloc_percpu_gfp(sizeof(*pc), 8, GFP_KERNEL);
508 /* room for llist_node and per-cpu pointer */
509 percpu_size = LLIST_NODE_SZ + sizeof(void *);
511 size += LLIST_NODE_SZ; /* room for llist_node */
514 #ifdef CONFIG_MEMCG_KMEM
515 if (memcg_bpf_enabled())
516 objcg = get_obj_cgroup_from_current();
518 for_each_possible_cpu(cpu) {
519 c = per_cpu_ptr(pc, cpu);
520 c->unit_size = unit_size;
522 c->percpu_size = percpu_size;
524 prefill_mem_cache(c, cpu);
530 /* size == 0 && percpu is an invalid combination */
531 if (WARN_ON_ONCE(percpu))
534 pcc = __alloc_percpu_gfp(sizeof(*cc), 8, GFP_KERNEL);
537 #ifdef CONFIG_MEMCG_KMEM
538 objcg = get_obj_cgroup_from_current();
540 for_each_possible_cpu(cpu) {
541 cc = per_cpu_ptr(pcc, cpu);
542 for (i = 0; i < NUM_CACHES; i++) {
544 c->unit_size = sizes[i];
547 prefill_mem_cache(c, cpu);
554 static void drain_mem_cache(struct bpf_mem_cache *c)
556 bool percpu = !!c->percpu_size;
558 /* No progs are using this bpf_mem_cache, but htab_map_free() called
559 * bpf_mem_cache_free() for all remaining elements and they can be in
560 * free_by_rcu_ttrace or in waiting_for_gp_ttrace lists, so drain those lists now.
562 * Except for waiting_for_gp_ttrace list, there are no concurrent operations
563 * on these lists, so it is safe to use __llist_del_all().
565 free_all(llist_del_all(&c->free_by_rcu_ttrace), percpu);
566 free_all(llist_del_all(&c->waiting_for_gp_ttrace), percpu);
567 free_all(__llist_del_all(&c->free_llist), percpu);
568 free_all(__llist_del_all(&c->free_llist_extra), percpu);
569 free_all(__llist_del_all(&c->free_by_rcu), percpu);
570 free_all(__llist_del_all(&c->free_llist_extra_rcu), percpu);
571 free_all(llist_del_all(&c->waiting_for_gp), percpu);
574 static void check_mem_cache(struct bpf_mem_cache *c)
576 WARN_ON_ONCE(!llist_empty(&c->free_by_rcu_ttrace));
577 WARN_ON_ONCE(!llist_empty(&c->waiting_for_gp_ttrace));
578 WARN_ON_ONCE(!llist_empty(&c->free_llist));
579 WARN_ON_ONCE(!llist_empty(&c->free_llist_extra));
580 WARN_ON_ONCE(!llist_empty(&c->free_by_rcu));
581 WARN_ON_ONCE(!llist_empty(&c->free_llist_extra_rcu));
582 WARN_ON_ONCE(!llist_empty(&c->waiting_for_gp));
585 static void check_leaked_objs(struct bpf_mem_alloc *ma)
587 struct bpf_mem_caches *cc;
588 struct bpf_mem_cache *c;
592 for_each_possible_cpu(cpu) {
593 c = per_cpu_ptr(ma->cache, cpu);
598 for_each_possible_cpu(cpu) {
599 cc = per_cpu_ptr(ma->caches, cpu);
600 for (i = 0; i < NUM_CACHES; i++) {
608 static void free_mem_alloc_no_barrier(struct bpf_mem_alloc *ma)
610 check_leaked_objs(ma);
611 free_percpu(ma->cache);
612 free_percpu(ma->caches);
617 static void free_mem_alloc(struct bpf_mem_alloc *ma)
619 /* waiting_for_gp[_ttrace] lists were drained, but RCU callbacks
620 * might still execute. Wait for them.
622 * rcu_barrier_tasks_trace() doesn't imply synchronize_rcu_tasks_trace(),
623 * but rcu_barrier_tasks_trace() and rcu_barrier() below are only used
624 * to wait for the pending __free_rcu_tasks_trace() and __free_rcu(),
625 * so if call_rcu(head, __free_rcu) is skipped due to
626 * rcu_trace_implies_rcu_gp(), it will be OK to skip rcu_barrier() by
627 * using rcu_trace_implies_rcu_gp() as well.
629 rcu_barrier(); /* wait for __free_by_rcu */
630 rcu_barrier_tasks_trace(); /* wait for __free_rcu */
631 if (!rcu_trace_implies_rcu_gp())
633 free_mem_alloc_no_barrier(ma);
636 static void free_mem_alloc_deferred(struct work_struct *work)
638 struct bpf_mem_alloc *ma = container_of(work, struct bpf_mem_alloc, work);
644 static void destroy_mem_alloc(struct bpf_mem_alloc *ma, int rcu_in_progress)
646 struct bpf_mem_alloc *copy;
648 if (!rcu_in_progress) {
649 /* Fast path. No callbacks are pending, hence no need to do
652 free_mem_alloc_no_barrier(ma);
656 copy = kmemdup(ma, sizeof(*ma), GFP_KERNEL);
658 /* Slow path with inline barrier-s */
663 /* Defer barriers into worker to let the rest of map memory to be freed */
664 memset(ma, 0, sizeof(*ma));
665 INIT_WORK(©->work, free_mem_alloc_deferred);
666 queue_work(system_unbound_wq, ©->work);
669 void bpf_mem_alloc_destroy(struct bpf_mem_alloc *ma)
671 struct bpf_mem_caches *cc;
672 struct bpf_mem_cache *c;
673 int cpu, i, rcu_in_progress;
677 for_each_possible_cpu(cpu) {
678 c = per_cpu_ptr(ma->cache, cpu);
679 WRITE_ONCE(c->draining, true);
680 irq_work_sync(&c->refill_work);
682 rcu_in_progress += atomic_read(&c->call_rcu_ttrace_in_progress);
683 rcu_in_progress += atomic_read(&c->call_rcu_in_progress);
685 /* objcg is the same across cpus */
687 obj_cgroup_put(c->objcg);
688 destroy_mem_alloc(ma, rcu_in_progress);
692 for_each_possible_cpu(cpu) {
693 cc = per_cpu_ptr(ma->caches, cpu);
694 for (i = 0; i < NUM_CACHES; i++) {
696 WRITE_ONCE(c->draining, true);
697 irq_work_sync(&c->refill_work);
699 rcu_in_progress += atomic_read(&c->call_rcu_ttrace_in_progress);
700 rcu_in_progress += atomic_read(&c->call_rcu_in_progress);
704 obj_cgroup_put(c->objcg);
705 destroy_mem_alloc(ma, rcu_in_progress);
709 /* notrace is necessary here and in other functions to make sure
710 * bpf programs cannot attach to them and cause llist corruptions.
712 static void notrace *unit_alloc(struct bpf_mem_cache *c)
714 struct llist_node *llnode = NULL;
718 /* Disable irqs to prevent the following race for majority of prog types:
721 * preemption or irq -> prog_B
724 * but prog_B could be a perf_event NMI prog.
725 * Use per-cpu 'active' counter to order free_list access between
726 * unit_alloc/unit_free/bpf_mem_refill.
728 local_irq_save(flags);
729 if (local_inc_return(&c->active) == 1) {
730 llnode = __llist_del_first(&c->free_llist);
733 *(struct bpf_mem_cache **)llnode = c;
736 local_dec(&c->active);
737 local_irq_restore(flags);
741 if (cnt < c->low_watermark)
746 /* Though 'ptr' object could have been allocated on a different cpu
747 * add it to the free_llist of the current cpu.
748 * Let kfree() logic deal with it when it's later called from irq_work.
750 static void notrace unit_free(struct bpf_mem_cache *c, void *ptr)
752 struct llist_node *llnode = ptr - LLIST_NODE_SZ;
756 BUILD_BUG_ON(LLIST_NODE_SZ > 8);
759 * Remember bpf_mem_cache that allocated this object.
760 * The hint is not accurate.
762 c->tgt = *(struct bpf_mem_cache **)llnode;
764 local_irq_save(flags);
765 if (local_inc_return(&c->active) == 1) {
766 __llist_add(llnode, &c->free_llist);
769 /* unit_free() cannot fail. Therefore add an object to atomic
770 * llist. free_bulk() will drain it. Though free_llist_extra is
771 * a per-cpu list we have to use atomic llist_add here, since
772 * it also can be interrupted by bpf nmi prog that does another
773 * unit_free() into the same free_llist_extra.
775 llist_add(llnode, &c->free_llist_extra);
777 local_dec(&c->active);
778 local_irq_restore(flags);
780 if (cnt > c->high_watermark)
781 /* free few objects from current cpu into global kmalloc pool */
785 static void notrace unit_free_rcu(struct bpf_mem_cache *c, void *ptr)
787 struct llist_node *llnode = ptr - LLIST_NODE_SZ;
790 c->tgt = *(struct bpf_mem_cache **)llnode;
792 local_irq_save(flags);
793 if (local_inc_return(&c->active) == 1) {
794 if (__llist_add(llnode, &c->free_by_rcu))
795 c->free_by_rcu_tail = llnode;
797 llist_add(llnode, &c->free_llist_extra_rcu);
799 local_dec(&c->active);
800 local_irq_restore(flags);
802 if (!atomic_read(&c->call_rcu_in_progress))
806 /* Called from BPF program or from sys_bpf syscall.
807 * In both cases migration is disabled.
809 void notrace *bpf_mem_alloc(struct bpf_mem_alloc *ma, size_t size)
815 return ZERO_SIZE_PTR;
817 idx = bpf_mem_cache_idx(size + LLIST_NODE_SZ);
821 ret = unit_alloc(this_cpu_ptr(ma->caches)->cache + idx);
822 return !ret ? NULL : ret + LLIST_NODE_SZ;
825 void notrace bpf_mem_free(struct bpf_mem_alloc *ma, void *ptr)
832 idx = bpf_mem_cache_idx(ksize(ptr - LLIST_NODE_SZ));
836 unit_free(this_cpu_ptr(ma->caches)->cache + idx, ptr);
839 void notrace bpf_mem_free_rcu(struct bpf_mem_alloc *ma, void *ptr)
846 idx = bpf_mem_cache_idx(ksize(ptr - LLIST_NODE_SZ));
850 unit_free_rcu(this_cpu_ptr(ma->caches)->cache + idx, ptr);
853 void notrace *bpf_mem_cache_alloc(struct bpf_mem_alloc *ma)
857 ret = unit_alloc(this_cpu_ptr(ma->cache));
858 return !ret ? NULL : ret + LLIST_NODE_SZ;
861 void notrace bpf_mem_cache_free(struct bpf_mem_alloc *ma, void *ptr)
866 unit_free(this_cpu_ptr(ma->cache), ptr);
869 void notrace bpf_mem_cache_free_rcu(struct bpf_mem_alloc *ma, void *ptr)
874 unit_free_rcu(this_cpu_ptr(ma->cache), ptr);
877 /* Directly does a kfree() without putting 'ptr' back to the free_llist
878 * for reuse and without waiting for a rcu_tasks_trace gp.
879 * The caller must first go through the rcu_tasks_trace gp for 'ptr'
880 * before calling bpf_mem_cache_raw_free().
881 * It could be used when the rcu_tasks_trace callback does not have
882 * a hold on the original bpf_mem_alloc object that allocated the
883 * 'ptr'. This should only be used in the uncommon code path.
884 * Otherwise, the bpf_mem_alloc's free_llist cannot be refilled
885 * and may affect performance.
887 void bpf_mem_cache_raw_free(void *ptr)
892 kfree(ptr - LLIST_NODE_SZ);
895 /* When flags == GFP_KERNEL, it signals that the caller will not cause
896 * deadlock when using kmalloc. bpf_mem_cache_alloc_flags() will use
897 * kmalloc if the free_llist is empty.
899 void notrace *bpf_mem_cache_alloc_flags(struct bpf_mem_alloc *ma, gfp_t flags)
901 struct bpf_mem_cache *c;
904 c = this_cpu_ptr(ma->cache);
907 if (!ret && flags == GFP_KERNEL) {
908 struct mem_cgroup *memcg, *old_memcg;
910 memcg = get_memcg(c);
911 old_memcg = set_active_memcg(memcg);
912 ret = __alloc(c, NUMA_NO_NODE, GFP_KERNEL | __GFP_NOWARN | __GFP_ACCOUNT);
913 set_active_memcg(old_memcg);
914 mem_cgroup_put(memcg);
917 return !ret ? NULL : ret + LLIST_NODE_SZ;