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);
343 WARN_ON_ONCE(tgt->percpu_size != c->percpu_size);
346 inc_active(c, &flags);
347 llnode = __llist_del_first(&c->free_llist);
352 dec_active(c, &flags);
354 enque_to_free(tgt, llnode);
355 } while (cnt > (c->high_watermark + c->low_watermark) / 2);
357 /* and drain free_llist_extra */
358 llist_for_each_safe(llnode, t, llist_del_all(&c->free_llist_extra))
359 enque_to_free(tgt, llnode);
360 do_call_rcu_ttrace(tgt);
363 static void __free_by_rcu(struct rcu_head *head)
365 struct bpf_mem_cache *c = container_of(head, struct bpf_mem_cache, rcu);
366 struct bpf_mem_cache *tgt = c->tgt;
367 struct llist_node *llnode;
369 WARN_ON_ONCE(tgt->unit_size != c->unit_size);
370 WARN_ON_ONCE(tgt->percpu_size != c->percpu_size);
372 llnode = llist_del_all(&c->waiting_for_gp);
376 llist_add_batch(llnode, c->waiting_for_gp_tail, &tgt->free_by_rcu_ttrace);
378 /* Objects went through regular RCU GP. Send them to RCU tasks trace */
379 do_call_rcu_ttrace(tgt);
381 atomic_set(&c->call_rcu_in_progress, 0);
384 static void check_free_by_rcu(struct bpf_mem_cache *c)
386 struct llist_node *llnode, *t;
389 /* drain free_llist_extra_rcu */
390 if (unlikely(!llist_empty(&c->free_llist_extra_rcu))) {
391 inc_active(c, &flags);
392 llist_for_each_safe(llnode, t, llist_del_all(&c->free_llist_extra_rcu))
393 if (__llist_add(llnode, &c->free_by_rcu))
394 c->free_by_rcu_tail = llnode;
395 dec_active(c, &flags);
398 if (llist_empty(&c->free_by_rcu))
401 if (atomic_xchg(&c->call_rcu_in_progress, 1)) {
403 * Instead of kmalloc-ing new rcu_head and triggering 10k
404 * call_rcu() to hit rcutree.qhimark and force RCU to notice
405 * the overload just ask RCU to hurry up. There could be many
406 * objects in free_by_rcu list.
407 * This hint reduces memory consumption for an artificial
408 * benchmark from 2 Gbyte to 150 Mbyte.
410 rcu_request_urgent_qs_task(current);
414 WARN_ON_ONCE(!llist_empty(&c->waiting_for_gp));
416 inc_active(c, &flags);
417 WRITE_ONCE(c->waiting_for_gp.first, __llist_del_all(&c->free_by_rcu));
418 c->waiting_for_gp_tail = c->free_by_rcu_tail;
419 dec_active(c, &flags);
421 if (unlikely(READ_ONCE(c->draining))) {
422 free_all(llist_del_all(&c->waiting_for_gp), !!c->percpu_size);
423 atomic_set(&c->call_rcu_in_progress, 0);
425 call_rcu_hurry(&c->rcu, __free_by_rcu);
429 static void bpf_mem_refill(struct irq_work *work)
431 struct bpf_mem_cache *c = container_of(work, struct bpf_mem_cache, refill_work);
434 /* Racy access to free_cnt. It doesn't need to be 100% accurate */
436 if (cnt < c->low_watermark)
437 /* irq_work runs on this cpu and kmalloc will allocate
438 * from the current numa node which is what we want here.
440 alloc_bulk(c, c->batch, NUMA_NO_NODE, true);
441 else if (cnt > c->high_watermark)
444 check_free_by_rcu(c);
447 static void notrace irq_work_raise(struct bpf_mem_cache *c)
449 irq_work_queue(&c->refill_work);
452 /* For typical bpf map case that uses bpf_mem_cache_alloc and single bucket
453 * the freelist cache will be elem_size * 64 (or less) on each cpu.
455 * For bpf programs that don't have statically known allocation sizes and
456 * assuming (low_mark + high_mark) / 2 as an average number of elements per
457 * bucket and all buckets are used the total amount of memory in freelists
458 * on each cpu will be:
459 * 64*16 + 64*32 + 64*64 + 64*96 + 64*128 + 64*196 + 64*256 + 32*512 + 16*1024 + 8*2048 + 4*4096
460 * == ~ 116 Kbyte using below heuristic.
461 * Initialized, but unused bpf allocator (not bpf map specific one) will
462 * consume ~ 11 Kbyte per cpu.
463 * Typical case will be between 11K and 116K closer to 11K.
464 * bpf progs can and should share bpf_mem_cache when possible.
466 static void init_refill_work(struct bpf_mem_cache *c)
468 init_irq_work(&c->refill_work, bpf_mem_refill);
469 if (c->unit_size <= 256) {
470 c->low_watermark = 32;
471 c->high_watermark = 96;
473 /* When page_size == 4k, order-0 cache will have low_mark == 2
474 * and high_mark == 6 with batch alloc of 3 individual pages at
476 * 8k allocs and above low == 1, high == 3, batch == 1.
478 c->low_watermark = max(32 * 256 / c->unit_size, 1);
479 c->high_watermark = max(96 * 256 / c->unit_size, 3);
481 c->batch = max((c->high_watermark - c->low_watermark) / 4 * 3, 1);
484 static void prefill_mem_cache(struct bpf_mem_cache *c, int cpu)
486 /* To avoid consuming memory assume that 1st run of bpf
487 * prog won't be doing more than 4 map_update_elem from
488 * irq disabled region
490 alloc_bulk(c, c->unit_size <= 256 ? 4 : 1, cpu_to_node(cpu), false);
493 static int check_obj_size(struct bpf_mem_cache *c, unsigned int idx)
495 struct llist_node *first;
496 unsigned int obj_size;
498 first = c->free_llist.first;
503 obj_size = pcpu_alloc_size(((void **)first)[1]);
505 obj_size = ksize(first);
506 if (obj_size != c->unit_size) {
507 WARN_ONCE(1, "bpf_mem_cache[%u]: percpu %d, unexpected object size %u, expect %u\n",
508 idx, c->percpu_size, obj_size, c->unit_size);
514 /* When size != 0 bpf_mem_cache for each cpu.
515 * This is typical bpf hash map use case when all elements have equal size.
517 * When size == 0 allocate 11 bpf_mem_cache-s for each cpu, then rely on
518 * kmalloc/kfree. Max allocation size is 4096 in this case.
519 * This is bpf_dynptr and bpf_kptr use case.
521 int bpf_mem_alloc_init(struct bpf_mem_alloc *ma, int size, bool percpu)
523 static u16 sizes[NUM_CACHES] = {96, 192, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096};
524 int cpu, i, err, unit_size, percpu_size = 0;
525 struct bpf_mem_caches *cc, __percpu *pcc;
526 struct bpf_mem_cache *c, __percpu *pc;
527 struct obj_cgroup *objcg = NULL;
529 /* room for llist_node and per-cpu pointer */
531 percpu_size = LLIST_NODE_SZ + sizeof(void *);
535 pc = __alloc_percpu_gfp(sizeof(*pc), 8, GFP_KERNEL);
540 size += LLIST_NODE_SZ; /* room for llist_node */
543 #ifdef CONFIG_MEMCG_KMEM
544 if (memcg_bpf_enabled())
545 objcg = get_obj_cgroup_from_current();
547 for_each_possible_cpu(cpu) {
548 c = per_cpu_ptr(pc, cpu);
549 c->unit_size = unit_size;
551 c->percpu_size = percpu_size;
554 prefill_mem_cache(c, cpu);
560 pcc = __alloc_percpu_gfp(sizeof(*cc), 8, GFP_KERNEL);
564 #ifdef CONFIG_MEMCG_KMEM
565 objcg = get_obj_cgroup_from_current();
567 for_each_possible_cpu(cpu) {
568 cc = per_cpu_ptr(pcc, cpu);
569 for (i = 0; i < NUM_CACHES; i++) {
571 c->unit_size = sizes[i];
573 c->percpu_size = percpu_size;
577 /* Another bpf_mem_cache will be used when allocating
578 * c->unit_size in bpf_mem_alloc(), so doesn't prefill
579 * for the bpf_mem_cache because these free objects will
582 if (i != bpf_mem_cache_idx(c->unit_size))
584 prefill_mem_cache(c, cpu);
585 err = check_obj_size(c, i);
593 /* refill_work is either zeroed or initialized, so it is safe to
594 * call irq_work_sync().
597 bpf_mem_alloc_destroy(ma);
601 static void drain_mem_cache(struct bpf_mem_cache *c)
603 bool percpu = !!c->percpu_size;
605 /* No progs are using this bpf_mem_cache, but htab_map_free() called
606 * bpf_mem_cache_free() for all remaining elements and they can be in
607 * free_by_rcu_ttrace or in waiting_for_gp_ttrace lists, so drain those lists now.
609 * Except for waiting_for_gp_ttrace list, there are no concurrent operations
610 * on these lists, so it is safe to use __llist_del_all().
612 free_all(llist_del_all(&c->free_by_rcu_ttrace), percpu);
613 free_all(llist_del_all(&c->waiting_for_gp_ttrace), percpu);
614 free_all(__llist_del_all(&c->free_llist), percpu);
615 free_all(__llist_del_all(&c->free_llist_extra), percpu);
616 free_all(__llist_del_all(&c->free_by_rcu), percpu);
617 free_all(__llist_del_all(&c->free_llist_extra_rcu), percpu);
618 free_all(llist_del_all(&c->waiting_for_gp), percpu);
621 static void check_mem_cache(struct bpf_mem_cache *c)
623 WARN_ON_ONCE(!llist_empty(&c->free_by_rcu_ttrace));
624 WARN_ON_ONCE(!llist_empty(&c->waiting_for_gp_ttrace));
625 WARN_ON_ONCE(!llist_empty(&c->free_llist));
626 WARN_ON_ONCE(!llist_empty(&c->free_llist_extra));
627 WARN_ON_ONCE(!llist_empty(&c->free_by_rcu));
628 WARN_ON_ONCE(!llist_empty(&c->free_llist_extra_rcu));
629 WARN_ON_ONCE(!llist_empty(&c->waiting_for_gp));
632 static void check_leaked_objs(struct bpf_mem_alloc *ma)
634 struct bpf_mem_caches *cc;
635 struct bpf_mem_cache *c;
639 for_each_possible_cpu(cpu) {
640 c = per_cpu_ptr(ma->cache, cpu);
645 for_each_possible_cpu(cpu) {
646 cc = per_cpu_ptr(ma->caches, cpu);
647 for (i = 0; i < NUM_CACHES; i++) {
655 static void free_mem_alloc_no_barrier(struct bpf_mem_alloc *ma)
657 check_leaked_objs(ma);
658 free_percpu(ma->cache);
659 free_percpu(ma->caches);
664 static void free_mem_alloc(struct bpf_mem_alloc *ma)
666 /* waiting_for_gp[_ttrace] lists were drained, but RCU callbacks
667 * might still execute. Wait for them.
669 * rcu_barrier_tasks_trace() doesn't imply synchronize_rcu_tasks_trace(),
670 * but rcu_barrier_tasks_trace() and rcu_barrier() below are only used
671 * to wait for the pending __free_rcu_tasks_trace() and __free_rcu(),
672 * so if call_rcu(head, __free_rcu) is skipped due to
673 * rcu_trace_implies_rcu_gp(), it will be OK to skip rcu_barrier() by
674 * using rcu_trace_implies_rcu_gp() as well.
676 rcu_barrier(); /* wait for __free_by_rcu */
677 rcu_barrier_tasks_trace(); /* wait for __free_rcu */
678 if (!rcu_trace_implies_rcu_gp())
680 free_mem_alloc_no_barrier(ma);
683 static void free_mem_alloc_deferred(struct work_struct *work)
685 struct bpf_mem_alloc *ma = container_of(work, struct bpf_mem_alloc, work);
691 static void destroy_mem_alloc(struct bpf_mem_alloc *ma, int rcu_in_progress)
693 struct bpf_mem_alloc *copy;
695 if (!rcu_in_progress) {
696 /* Fast path. No callbacks are pending, hence no need to do
699 free_mem_alloc_no_barrier(ma);
703 copy = kmemdup(ma, sizeof(*ma), GFP_KERNEL);
705 /* Slow path with inline barrier-s */
710 /* Defer barriers into worker to let the rest of map memory to be freed */
711 memset(ma, 0, sizeof(*ma));
712 INIT_WORK(©->work, free_mem_alloc_deferred);
713 queue_work(system_unbound_wq, ©->work);
716 void bpf_mem_alloc_destroy(struct bpf_mem_alloc *ma)
718 struct bpf_mem_caches *cc;
719 struct bpf_mem_cache *c;
720 int cpu, i, rcu_in_progress;
724 for_each_possible_cpu(cpu) {
725 c = per_cpu_ptr(ma->cache, cpu);
726 WRITE_ONCE(c->draining, true);
727 irq_work_sync(&c->refill_work);
729 rcu_in_progress += atomic_read(&c->call_rcu_ttrace_in_progress);
730 rcu_in_progress += atomic_read(&c->call_rcu_in_progress);
732 /* objcg is the same across cpus */
734 obj_cgroup_put(c->objcg);
735 destroy_mem_alloc(ma, rcu_in_progress);
739 for_each_possible_cpu(cpu) {
740 cc = per_cpu_ptr(ma->caches, cpu);
741 for (i = 0; i < NUM_CACHES; i++) {
743 WRITE_ONCE(c->draining, true);
744 irq_work_sync(&c->refill_work);
746 rcu_in_progress += atomic_read(&c->call_rcu_ttrace_in_progress);
747 rcu_in_progress += atomic_read(&c->call_rcu_in_progress);
751 obj_cgroup_put(c->objcg);
752 destroy_mem_alloc(ma, rcu_in_progress);
756 /* notrace is necessary here and in other functions to make sure
757 * bpf programs cannot attach to them and cause llist corruptions.
759 static void notrace *unit_alloc(struct bpf_mem_cache *c)
761 struct llist_node *llnode = NULL;
765 /* Disable irqs to prevent the following race for majority of prog types:
768 * preemption or irq -> prog_B
771 * but prog_B could be a perf_event NMI prog.
772 * Use per-cpu 'active' counter to order free_list access between
773 * unit_alloc/unit_free/bpf_mem_refill.
775 local_irq_save(flags);
776 if (local_inc_return(&c->active) == 1) {
777 llnode = __llist_del_first(&c->free_llist);
780 *(struct bpf_mem_cache **)llnode = c;
783 local_dec(&c->active);
787 if (cnt < c->low_watermark)
789 /* Enable IRQ after the enqueue of irq work completes, so irq work
790 * will run after IRQ is enabled and free_llist may be refilled by
791 * irq work before other task preempts current task.
793 local_irq_restore(flags);
798 /* Though 'ptr' object could have been allocated on a different cpu
799 * add it to the free_llist of the current cpu.
800 * Let kfree() logic deal with it when it's later called from irq_work.
802 static void notrace unit_free(struct bpf_mem_cache *c, void *ptr)
804 struct llist_node *llnode = ptr - LLIST_NODE_SZ;
808 BUILD_BUG_ON(LLIST_NODE_SZ > 8);
811 * Remember bpf_mem_cache that allocated this object.
812 * The hint is not accurate.
814 c->tgt = *(struct bpf_mem_cache **)llnode;
816 local_irq_save(flags);
817 if (local_inc_return(&c->active) == 1) {
818 __llist_add(llnode, &c->free_llist);
821 /* unit_free() cannot fail. Therefore add an object to atomic
822 * llist. free_bulk() will drain it. Though free_llist_extra is
823 * a per-cpu list we have to use atomic llist_add here, since
824 * it also can be interrupted by bpf nmi prog that does another
825 * unit_free() into the same free_llist_extra.
827 llist_add(llnode, &c->free_llist_extra);
829 local_dec(&c->active);
831 if (cnt > c->high_watermark)
832 /* free few objects from current cpu into global kmalloc pool */
834 /* Enable IRQ after irq_work_raise() completes, otherwise when current
835 * task is preempted by task which does unit_alloc(), unit_alloc() may
836 * return NULL unexpectedly because irq work is already pending but can
837 * not been triggered and free_llist can not be refilled timely.
839 local_irq_restore(flags);
842 static void notrace unit_free_rcu(struct bpf_mem_cache *c, void *ptr)
844 struct llist_node *llnode = ptr - LLIST_NODE_SZ;
847 c->tgt = *(struct bpf_mem_cache **)llnode;
849 local_irq_save(flags);
850 if (local_inc_return(&c->active) == 1) {
851 if (__llist_add(llnode, &c->free_by_rcu))
852 c->free_by_rcu_tail = llnode;
854 llist_add(llnode, &c->free_llist_extra_rcu);
856 local_dec(&c->active);
858 if (!atomic_read(&c->call_rcu_in_progress))
860 local_irq_restore(flags);
863 /* Called from BPF program or from sys_bpf syscall.
864 * In both cases migration is disabled.
866 void notrace *bpf_mem_alloc(struct bpf_mem_alloc *ma, size_t size)
872 return ZERO_SIZE_PTR;
874 idx = bpf_mem_cache_idx(size + LLIST_NODE_SZ);
878 ret = unit_alloc(this_cpu_ptr(ma->caches)->cache + idx);
879 return !ret ? NULL : ret + LLIST_NODE_SZ;
882 static notrace int bpf_mem_free_idx(void *ptr, bool percpu)
887 size = pcpu_alloc_size(*((void **)ptr));
889 size = ksize(ptr - LLIST_NODE_SZ);
890 return bpf_mem_cache_idx(size);
893 void notrace bpf_mem_free(struct bpf_mem_alloc *ma, void *ptr)
900 idx = bpf_mem_free_idx(ptr, ma->percpu);
904 unit_free(this_cpu_ptr(ma->caches)->cache + idx, ptr);
907 void notrace bpf_mem_free_rcu(struct bpf_mem_alloc *ma, void *ptr)
914 idx = bpf_mem_free_idx(ptr, ma->percpu);
918 unit_free_rcu(this_cpu_ptr(ma->caches)->cache + idx, ptr);
921 void notrace *bpf_mem_cache_alloc(struct bpf_mem_alloc *ma)
925 ret = unit_alloc(this_cpu_ptr(ma->cache));
926 return !ret ? NULL : ret + LLIST_NODE_SZ;
929 void notrace bpf_mem_cache_free(struct bpf_mem_alloc *ma, void *ptr)
934 unit_free(this_cpu_ptr(ma->cache), ptr);
937 void notrace bpf_mem_cache_free_rcu(struct bpf_mem_alloc *ma, void *ptr)
942 unit_free_rcu(this_cpu_ptr(ma->cache), ptr);
945 /* Directly does a kfree() without putting 'ptr' back to the free_llist
946 * for reuse and without waiting for a rcu_tasks_trace gp.
947 * The caller must first go through the rcu_tasks_trace gp for 'ptr'
948 * before calling bpf_mem_cache_raw_free().
949 * It could be used when the rcu_tasks_trace callback does not have
950 * a hold on the original bpf_mem_alloc object that allocated the
951 * 'ptr'. This should only be used in the uncommon code path.
952 * Otherwise, the bpf_mem_alloc's free_llist cannot be refilled
953 * and may affect performance.
955 void bpf_mem_cache_raw_free(void *ptr)
960 kfree(ptr - LLIST_NODE_SZ);
963 /* When flags == GFP_KERNEL, it signals that the caller will not cause
964 * deadlock when using kmalloc. bpf_mem_cache_alloc_flags() will use
965 * kmalloc if the free_llist is empty.
967 void notrace *bpf_mem_cache_alloc_flags(struct bpf_mem_alloc *ma, gfp_t flags)
969 struct bpf_mem_cache *c;
972 c = this_cpu_ptr(ma->cache);
975 if (!ret && flags == GFP_KERNEL) {
976 struct mem_cgroup *memcg, *old_memcg;
978 memcg = get_memcg(c);
979 old_memcg = set_active_memcg(memcg);
980 ret = __alloc(c, NUMA_NO_NODE, GFP_KERNEL | __GFP_NOWARN | __GFP_ACCOUNT);
982 *(struct bpf_mem_cache **)ret = c;
983 set_active_memcg(old_memcg);
984 mem_cgroup_put(memcg);
987 return !ret ? NULL : ret + LLIST_NODE_SZ;
990 /* The alignment of dynamic per-cpu area is 8, so c->unit_size and the
991 * actual size of dynamic per-cpu area will always be matched and there is
992 * no need to adjust size_index for per-cpu allocation. However for the
993 * simplicity of the implementation, use an unified size_index for both
994 * kmalloc and per-cpu allocation.
996 static __init int bpf_mem_cache_adjust_size(void)
1000 /* Adjusting the indexes in size_index() according to the object_size
1001 * of underlying slab cache, so bpf_mem_alloc() will select a
1002 * bpf_mem_cache with unit_size equal to the object_size of
1003 * the underlying slab cache.
1005 * The maximal value of KMALLOC_MIN_SIZE and __kmalloc_minalign() is
1006 * 256-bytes, so only do adjustment for [8-bytes, 192-bytes].
1008 for (size = 192; size >= 8; size -= 8) {
1009 unsigned int kmalloc_size, index;
1011 kmalloc_size = kmalloc_size_roundup(size);
1012 if (kmalloc_size == size)
1015 if (kmalloc_size <= 192)
1016 index = size_index[(kmalloc_size - 1) / 8];
1018 index = fls(kmalloc_size - 1) - 1;
1019 /* Only overwrite if necessary */
1020 if (size_index[(size - 1) / 8] != index)
1021 size_index[(size - 1) / 8] = index;
1026 subsys_initcall(bpf_mem_cache_adjust_size);