Loading...
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)
125{
126 /* Allocate, but don't deplete atomic reserves that typical
127 * GFP_ATOMIC would do. irq_work runs on this cpu and kmalloc
128 * will allocate from the current numa node which is what we
129 * want here.
130 */
131 gfp_t flags = GFP_NOWAIT | __GFP_NOWARN | __GFP_ACCOUNT;
132
133 if (c->percpu_size) {
134 void **obj = kmalloc_node(c->percpu_size, flags, node);
135 void *pptr = __alloc_percpu_gfp(c->unit_size, 8, flags);
136
137 if (!obj || !pptr) {
138 free_percpu(pptr);
139 kfree(obj);
140 return NULL;
141 }
142 obj[1] = pptr;
143 return obj;
144 }
145
146 return kmalloc_node(c->unit_size, flags, node);
147}
148
149static struct mem_cgroup *get_memcg(const struct bpf_mem_cache *c)
150{
151#ifdef CONFIG_MEMCG_KMEM
152 if (c->objcg)
153 return get_mem_cgroup_from_objcg(c->objcg);
154#endif
155
156#ifdef CONFIG_MEMCG
157 return root_mem_cgroup;
158#else
159 return NULL;
160#endif
161}
162
163/* Mostly runs from irq_work except __init phase. */
164static void alloc_bulk(struct bpf_mem_cache *c, int cnt, int node)
165{
166 struct mem_cgroup *memcg = NULL, *old_memcg;
167 unsigned long flags;
168 void *obj;
169 int i;
170
171 memcg = get_memcg(c);
172 old_memcg = set_active_memcg(memcg);
173 for (i = 0; i < cnt; i++) {
174 /*
175 * free_by_rcu is only manipulated by irq work refill_work().
176 * IRQ works on the same CPU are called sequentially, so it is
177 * safe to use __llist_del_first() here. If alloc_bulk() is
178 * invoked by the initial prefill, there will be no running
179 * refill_work(), so __llist_del_first() is fine as well.
180 *
181 * In most cases, objects on free_by_rcu are from the same CPU.
182 * If some objects come from other CPUs, it doesn't incur any
183 * harm because NUMA_NO_NODE means the preference for current
184 * numa node and it is not a guarantee.
185 */
186 obj = __llist_del_first(&c->free_by_rcu);
187 if (!obj) {
188 obj = __alloc(c, node);
189 if (!obj)
190 break;
191 }
192 if (IS_ENABLED(CONFIG_PREEMPT_RT))
193 /* In RT irq_work runs in per-cpu kthread, so disable
194 * interrupts to avoid preemption and interrupts and
195 * reduce the chance of bpf prog executing on this cpu
196 * when active counter is busy.
197 */
198 local_irq_save(flags);
199 /* alloc_bulk runs from irq_work which will not preempt a bpf
200 * program that does unit_alloc/unit_free since IRQs are
201 * disabled there. There is no race to increment 'active'
202 * counter. It protects free_llist from corruption in case NMI
203 * bpf prog preempted this loop.
204 */
205 WARN_ON_ONCE(local_inc_return(&c->active) != 1);
206 __llist_add(obj, &c->free_llist);
207 c->free_cnt++;
208 local_dec(&c->active);
209 if (IS_ENABLED(CONFIG_PREEMPT_RT))
210 local_irq_restore(flags);
211 }
212 set_active_memcg(old_memcg);
213 mem_cgroup_put(memcg);
214}
215
216static void free_one(struct bpf_mem_cache *c, void *obj)
217{
218 if (c->percpu_size) {
219 free_percpu(((void **)obj)[1]);
220 kfree(obj);
221 return;
222 }
223
224 kfree(obj);
225}
226
227static void __free_rcu(struct rcu_head *head)
228{
229 struct bpf_mem_cache *c = container_of(head, struct bpf_mem_cache, rcu);
230 struct llist_node *llnode = llist_del_all(&c->waiting_for_gp);
231 struct llist_node *pos, *t;
232
233 llist_for_each_safe(pos, t, llnode)
234 free_one(c, pos);
235 atomic_set(&c->call_rcu_in_progress, 0);
236}
237
238static void __free_rcu_tasks_trace(struct rcu_head *head)
239{
240 /* If RCU Tasks Trace grace period implies RCU grace period,
241 * there is no need to invoke call_rcu().
242 */
243 if (rcu_trace_implies_rcu_gp())
244 __free_rcu(head);
245 else
246 call_rcu(head, __free_rcu);
247}
248
249static void enque_to_free(struct bpf_mem_cache *c, void *obj)
250{
251 struct llist_node *llnode = obj;
252
253 /* bpf_mem_cache is a per-cpu object. Freeing happens in irq_work.
254 * Nothing races to add to free_by_rcu list.
255 */
256 __llist_add(llnode, &c->free_by_rcu);
257}
258
259static void do_call_rcu(struct bpf_mem_cache *c)
260{
261 struct llist_node *llnode, *t;
262
263 if (atomic_xchg(&c->call_rcu_in_progress, 1))
264 return;
265
266 WARN_ON_ONCE(!llist_empty(&c->waiting_for_gp));
267 llist_for_each_safe(llnode, t, __llist_del_all(&c->free_by_rcu))
268 /* There is no concurrent __llist_add(waiting_for_gp) access.
269 * It doesn't race with llist_del_all either.
270 * But there could be two concurrent llist_del_all(waiting_for_gp):
271 * from __free_rcu() and from drain_mem_cache().
272 */
273 __llist_add(llnode, &c->waiting_for_gp);
274 /* Use call_rcu_tasks_trace() to wait for sleepable progs to finish.
275 * If RCU Tasks Trace grace period implies RCU grace period, free
276 * these elements directly, else use call_rcu() to wait for normal
277 * progs to finish and finally do free_one() on each element.
278 */
279 call_rcu_tasks_trace(&c->rcu, __free_rcu_tasks_trace);
280}
281
282static void free_bulk(struct bpf_mem_cache *c)
283{
284 struct llist_node *llnode, *t;
285 unsigned long flags;
286 int cnt;
287
288 do {
289 if (IS_ENABLED(CONFIG_PREEMPT_RT))
290 local_irq_save(flags);
291 WARN_ON_ONCE(local_inc_return(&c->active) != 1);
292 llnode = __llist_del_first(&c->free_llist);
293 if (llnode)
294 cnt = --c->free_cnt;
295 else
296 cnt = 0;
297 local_dec(&c->active);
298 if (IS_ENABLED(CONFIG_PREEMPT_RT))
299 local_irq_restore(flags);
300 if (llnode)
301 enque_to_free(c, llnode);
302 } while (cnt > (c->high_watermark + c->low_watermark) / 2);
303
304 /* and drain free_llist_extra */
305 llist_for_each_safe(llnode, t, llist_del_all(&c->free_llist_extra))
306 enque_to_free(c, llnode);
307 do_call_rcu(c);
308}
309
310static void bpf_mem_refill(struct irq_work *work)
311{
312 struct bpf_mem_cache *c = container_of(work, struct bpf_mem_cache, refill_work);
313 int cnt;
314
315 /* Racy access to free_cnt. It doesn't need to be 100% accurate */
316 cnt = c->free_cnt;
317 if (cnt < c->low_watermark)
318 /* irq_work runs on this cpu and kmalloc will allocate
319 * from the current numa node which is what we want here.
320 */
321 alloc_bulk(c, c->batch, NUMA_NO_NODE);
322 else if (cnt > c->high_watermark)
323 free_bulk(c);
324}
325
326static void notrace irq_work_raise(struct bpf_mem_cache *c)
327{
328 irq_work_queue(&c->refill_work);
329}
330
331/* For typical bpf map case that uses bpf_mem_cache_alloc and single bucket
332 * the freelist cache will be elem_size * 64 (or less) on each cpu.
333 *
334 * For bpf programs that don't have statically known allocation sizes and
335 * assuming (low_mark + high_mark) / 2 as an average number of elements per
336 * bucket and all buckets are used the total amount of memory in freelists
337 * on each cpu will be:
338 * 64*16 + 64*32 + 64*64 + 64*96 + 64*128 + 64*196 + 64*256 + 32*512 + 16*1024 + 8*2048 + 4*4096
339 * == ~ 116 Kbyte using below heuristic.
340 * Initialized, but unused bpf allocator (not bpf map specific one) will
341 * consume ~ 11 Kbyte per cpu.
342 * Typical case will be between 11K and 116K closer to 11K.
343 * bpf progs can and should share bpf_mem_cache when possible.
344 */
345
346static void prefill_mem_cache(struct bpf_mem_cache *c, int cpu)
347{
348 init_irq_work(&c->refill_work, bpf_mem_refill);
349 if (c->unit_size <= 256) {
350 c->low_watermark = 32;
351 c->high_watermark = 96;
352 } else {
353 /* When page_size == 4k, order-0 cache will have low_mark == 2
354 * and high_mark == 6 with batch alloc of 3 individual pages at
355 * a time.
356 * 8k allocs and above low == 1, high == 3, batch == 1.
357 */
358 c->low_watermark = max(32 * 256 / c->unit_size, 1);
359 c->high_watermark = max(96 * 256 / c->unit_size, 3);
360 }
361 c->batch = max((c->high_watermark - c->low_watermark) / 4 * 3, 1);
362
363 /* To avoid consuming memory assume that 1st run of bpf
364 * prog won't be doing more than 4 map_update_elem from
365 * irq disabled region
366 */
367 alloc_bulk(c, c->unit_size <= 256 ? 4 : 1, cpu_to_node(cpu));
368}
369
370/* When size != 0 bpf_mem_cache for each cpu.
371 * This is typical bpf hash map use case when all elements have equal size.
372 *
373 * When size == 0 allocate 11 bpf_mem_cache-s for each cpu, then rely on
374 * kmalloc/kfree. Max allocation size is 4096 in this case.
375 * This is bpf_dynptr and bpf_kptr use case.
376 */
377int bpf_mem_alloc_init(struct bpf_mem_alloc *ma, int size, bool percpu)
378{
379 static u16 sizes[NUM_CACHES] = {96, 192, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096};
380 struct bpf_mem_caches *cc, __percpu *pcc;
381 struct bpf_mem_cache *c, __percpu *pc;
382 struct obj_cgroup *objcg = NULL;
383 int cpu, i, unit_size, percpu_size = 0;
384
385 if (size) {
386 pc = __alloc_percpu_gfp(sizeof(*pc), 8, GFP_KERNEL);
387 if (!pc)
388 return -ENOMEM;
389
390 if (percpu)
391 /* room for llist_node and per-cpu pointer */
392 percpu_size = LLIST_NODE_SZ + sizeof(void *);
393 else
394 size += LLIST_NODE_SZ; /* room for llist_node */
395 unit_size = size;
396
397#ifdef CONFIG_MEMCG_KMEM
398 objcg = get_obj_cgroup_from_current();
399#endif
400 for_each_possible_cpu(cpu) {
401 c = per_cpu_ptr(pc, cpu);
402 c->unit_size = unit_size;
403 c->objcg = objcg;
404 c->percpu_size = percpu_size;
405 prefill_mem_cache(c, cpu);
406 }
407 ma->cache = pc;
408 return 0;
409 }
410
411 /* size == 0 && percpu is an invalid combination */
412 if (WARN_ON_ONCE(percpu))
413 return -EINVAL;
414
415 pcc = __alloc_percpu_gfp(sizeof(*cc), 8, GFP_KERNEL);
416 if (!pcc)
417 return -ENOMEM;
418#ifdef CONFIG_MEMCG_KMEM
419 objcg = get_obj_cgroup_from_current();
420#endif
421 for_each_possible_cpu(cpu) {
422 cc = per_cpu_ptr(pcc, cpu);
423 for (i = 0; i < NUM_CACHES; i++) {
424 c = &cc->cache[i];
425 c->unit_size = sizes[i];
426 c->objcg = objcg;
427 prefill_mem_cache(c, cpu);
428 }
429 }
430 ma->caches = pcc;
431 return 0;
432}
433
434static void drain_mem_cache(struct bpf_mem_cache *c)
435{
436 struct llist_node *llnode, *t;
437
438 /* No progs are using this bpf_mem_cache, but htab_map_free() called
439 * bpf_mem_cache_free() for all remaining elements and they can be in
440 * free_by_rcu or in waiting_for_gp lists, so drain those lists now.
441 *
442 * Except for waiting_for_gp list, there are no concurrent operations
443 * on these lists, so it is safe to use __llist_del_all().
444 */
445 llist_for_each_safe(llnode, t, __llist_del_all(&c->free_by_rcu))
446 free_one(c, llnode);
447 llist_for_each_safe(llnode, t, llist_del_all(&c->waiting_for_gp))
448 free_one(c, llnode);
449 llist_for_each_safe(llnode, t, __llist_del_all(&c->free_llist))
450 free_one(c, llnode);
451 llist_for_each_safe(llnode, t, __llist_del_all(&c->free_llist_extra))
452 free_one(c, llnode);
453}
454
455static void free_mem_alloc_no_barrier(struct bpf_mem_alloc *ma)
456{
457 free_percpu(ma->cache);
458 free_percpu(ma->caches);
459 ma->cache = NULL;
460 ma->caches = NULL;
461}
462
463static void free_mem_alloc(struct bpf_mem_alloc *ma)
464{
465 /* waiting_for_gp lists was drained, but __free_rcu might
466 * still execute. Wait for it now before we freeing percpu caches.
467 *
468 * rcu_barrier_tasks_trace() doesn't imply synchronize_rcu_tasks_trace(),
469 * but rcu_barrier_tasks_trace() and rcu_barrier() below are only used
470 * to wait for the pending __free_rcu_tasks_trace() and __free_rcu(),
471 * so if call_rcu(head, __free_rcu) is skipped due to
472 * rcu_trace_implies_rcu_gp(), it will be OK to skip rcu_barrier() by
473 * using rcu_trace_implies_rcu_gp() as well.
474 */
475 rcu_barrier_tasks_trace();
476 if (!rcu_trace_implies_rcu_gp())
477 rcu_barrier();
478 free_mem_alloc_no_barrier(ma);
479}
480
481static void free_mem_alloc_deferred(struct work_struct *work)
482{
483 struct bpf_mem_alloc *ma = container_of(work, struct bpf_mem_alloc, work);
484
485 free_mem_alloc(ma);
486 kfree(ma);
487}
488
489static void destroy_mem_alloc(struct bpf_mem_alloc *ma, int rcu_in_progress)
490{
491 struct bpf_mem_alloc *copy;
492
493 if (!rcu_in_progress) {
494 /* Fast path. No callbacks are pending, hence no need to do
495 * rcu_barrier-s.
496 */
497 free_mem_alloc_no_barrier(ma);
498 return;
499 }
500
501 copy = kmalloc(sizeof(*ma), GFP_KERNEL);
502 if (!copy) {
503 /* Slow path with inline barrier-s */
504 free_mem_alloc(ma);
505 return;
506 }
507
508 /* Defer barriers into worker to let the rest of map memory to be freed */
509 copy->cache = ma->cache;
510 ma->cache = NULL;
511 copy->caches = ma->caches;
512 ma->caches = NULL;
513 INIT_WORK(©->work, free_mem_alloc_deferred);
514 queue_work(system_unbound_wq, ©->work);
515}
516
517void bpf_mem_alloc_destroy(struct bpf_mem_alloc *ma)
518{
519 struct bpf_mem_caches *cc;
520 struct bpf_mem_cache *c;
521 int cpu, i, rcu_in_progress;
522
523 if (ma->cache) {
524 rcu_in_progress = 0;
525 for_each_possible_cpu(cpu) {
526 c = per_cpu_ptr(ma->cache, cpu);
527 /*
528 * refill_work may be unfinished for PREEMPT_RT kernel
529 * in which irq work is invoked in a per-CPU RT thread.
530 * It is also possible for kernel with
531 * arch_irq_work_has_interrupt() being false and irq
532 * work is invoked in timer interrupt. So waiting for
533 * the completion of irq work to ease the handling of
534 * concurrency.
535 */
536 irq_work_sync(&c->refill_work);
537 drain_mem_cache(c);
538 rcu_in_progress += atomic_read(&c->call_rcu_in_progress);
539 }
540 /* objcg is the same across cpus */
541 if (c->objcg)
542 obj_cgroup_put(c->objcg);
543 destroy_mem_alloc(ma, rcu_in_progress);
544 }
545 if (ma->caches) {
546 rcu_in_progress = 0;
547 for_each_possible_cpu(cpu) {
548 cc = per_cpu_ptr(ma->caches, cpu);
549 for (i = 0; i < NUM_CACHES; i++) {
550 c = &cc->cache[i];
551 irq_work_sync(&c->refill_work);
552 drain_mem_cache(c);
553 rcu_in_progress += atomic_read(&c->call_rcu_in_progress);
554 }
555 }
556 if (c->objcg)
557 obj_cgroup_put(c->objcg);
558 destroy_mem_alloc(ma, rcu_in_progress);
559 }
560}
561
562/* notrace is necessary here and in other functions to make sure
563 * bpf programs cannot attach to them and cause llist corruptions.
564 */
565static void notrace *unit_alloc(struct bpf_mem_cache *c)
566{
567 struct llist_node *llnode = NULL;
568 unsigned long flags;
569 int cnt = 0;
570
571 /* Disable irqs to prevent the following race for majority of prog types:
572 * prog_A
573 * bpf_mem_alloc
574 * preemption or irq -> prog_B
575 * bpf_mem_alloc
576 *
577 * but prog_B could be a perf_event NMI prog.
578 * Use per-cpu 'active' counter to order free_list access between
579 * unit_alloc/unit_free/bpf_mem_refill.
580 */
581 local_irq_save(flags);
582 if (local_inc_return(&c->active) == 1) {
583 llnode = __llist_del_first(&c->free_llist);
584 if (llnode)
585 cnt = --c->free_cnt;
586 }
587 local_dec(&c->active);
588 local_irq_restore(flags);
589
590 WARN_ON(cnt < 0);
591
592 if (cnt < c->low_watermark)
593 irq_work_raise(c);
594 return llnode;
595}
596
597/* Though 'ptr' object could have been allocated on a different cpu
598 * add it to the free_llist of the current cpu.
599 * Let kfree() logic deal with it when it's later called from irq_work.
600 */
601static void notrace unit_free(struct bpf_mem_cache *c, void *ptr)
602{
603 struct llist_node *llnode = ptr - LLIST_NODE_SZ;
604 unsigned long flags;
605 int cnt = 0;
606
607 BUILD_BUG_ON(LLIST_NODE_SZ > 8);
608
609 local_irq_save(flags);
610 if (local_inc_return(&c->active) == 1) {
611 __llist_add(llnode, &c->free_llist);
612 cnt = ++c->free_cnt;
613 } else {
614 /* unit_free() cannot fail. Therefore add an object to atomic
615 * llist. free_bulk() will drain it. Though free_llist_extra is
616 * a per-cpu list we have to use atomic llist_add here, since
617 * it also can be interrupted by bpf nmi prog that does another
618 * unit_free() into the same free_llist_extra.
619 */
620 llist_add(llnode, &c->free_llist_extra);
621 }
622 local_dec(&c->active);
623 local_irq_restore(flags);
624
625 if (cnt > c->high_watermark)
626 /* free few objects from current cpu into global kmalloc pool */
627 irq_work_raise(c);
628}
629
630/* Called from BPF program or from sys_bpf syscall.
631 * In both cases migration is disabled.
632 */
633void notrace *bpf_mem_alloc(struct bpf_mem_alloc *ma, size_t size)
634{
635 int idx;
636 void *ret;
637
638 if (!size)
639 return ZERO_SIZE_PTR;
640
641 idx = bpf_mem_cache_idx(size + LLIST_NODE_SZ);
642 if (idx < 0)
643 return NULL;
644
645 ret = unit_alloc(this_cpu_ptr(ma->caches)->cache + idx);
646 return !ret ? NULL : ret + LLIST_NODE_SZ;
647}
648
649void notrace bpf_mem_free(struct bpf_mem_alloc *ma, void *ptr)
650{
651 int idx;
652
653 if (!ptr)
654 return;
655
656 idx = bpf_mem_cache_idx(ksize(ptr - LLIST_NODE_SZ));
657 if (idx < 0)
658 return;
659
660 unit_free(this_cpu_ptr(ma->caches)->cache + idx, ptr);
661}
662
663void notrace *bpf_mem_cache_alloc(struct bpf_mem_alloc *ma)
664{
665 void *ret;
666
667 ret = unit_alloc(this_cpu_ptr(ma->cache));
668 return !ret ? NULL : ret + LLIST_NODE_SZ;
669}
670
671void notrace bpf_mem_cache_free(struct bpf_mem_alloc *ma, void *ptr)
672{
673 if (!ptr)
674 return;
675
676 unit_free(this_cpu_ptr(ma->cache), ptr);
677}
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#define BPF_MEM_ALLOC_SIZE_MAX 4096
39
40/* similar to kmalloc, but sizeof == 8 bucket is gone */
41static u8 size_index[24] __ro_after_init = {
42 3, /* 8 */
43 3, /* 16 */
44 4, /* 24 */
45 4, /* 32 */
46 5, /* 40 */
47 5, /* 48 */
48 5, /* 56 */
49 5, /* 64 */
50 1, /* 72 */
51 1, /* 80 */
52 1, /* 88 */
53 1, /* 96 */
54 6, /* 104 */
55 6, /* 112 */
56 6, /* 120 */
57 6, /* 128 */
58 2, /* 136 */
59 2, /* 144 */
60 2, /* 152 */
61 2, /* 160 */
62 2, /* 168 */
63 2, /* 176 */
64 2, /* 184 */
65 2 /* 192 */
66};
67
68static int bpf_mem_cache_idx(size_t size)
69{
70 if (!size || size > BPF_MEM_ALLOC_SIZE_MAX)
71 return -1;
72
73 if (size <= 192)
74 return size_index[(size - 1) / 8] - 1;
75
76 return fls(size - 1) - 2;
77}
78
79#define NUM_CACHES 11
80
81struct bpf_mem_cache {
82 /* per-cpu list of free objects of size 'unit_size'.
83 * All accesses are done with interrupts disabled and 'active' counter
84 * protection with __llist_add() and __llist_del_first().
85 */
86 struct llist_head free_llist;
87 local_t active;
88
89 /* Operations on the free_list from unit_alloc/unit_free/bpf_mem_refill
90 * are sequenced by per-cpu 'active' counter. But unit_free() cannot
91 * fail. When 'active' is busy the unit_free() will add an object to
92 * free_llist_extra.
93 */
94 struct llist_head free_llist_extra;
95
96 struct irq_work refill_work;
97 struct obj_cgroup *objcg;
98 int unit_size;
99 /* count of objects in free_llist */
100 int free_cnt;
101 int low_watermark, high_watermark, batch;
102 int percpu_size;
103 bool draining;
104 struct bpf_mem_cache *tgt;
105
106 /* list of objects to be freed after RCU GP */
107 struct llist_head free_by_rcu;
108 struct llist_node *free_by_rcu_tail;
109 struct llist_head waiting_for_gp;
110 struct llist_node *waiting_for_gp_tail;
111 struct rcu_head rcu;
112 atomic_t call_rcu_in_progress;
113 struct llist_head free_llist_extra_rcu;
114
115 /* list of objects to be freed after RCU tasks trace GP */
116 struct llist_head free_by_rcu_ttrace;
117 struct llist_head waiting_for_gp_ttrace;
118 struct rcu_head rcu_ttrace;
119 atomic_t call_rcu_ttrace_in_progress;
120};
121
122struct bpf_mem_caches {
123 struct bpf_mem_cache cache[NUM_CACHES];
124};
125
126static const u16 sizes[NUM_CACHES] = {96, 192, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096};
127
128static struct llist_node notrace *__llist_del_first(struct llist_head *head)
129{
130 struct llist_node *entry, *next;
131
132 entry = head->first;
133 if (!entry)
134 return NULL;
135 next = entry->next;
136 head->first = next;
137 return entry;
138}
139
140static void *__alloc(struct bpf_mem_cache *c, int node, gfp_t flags)
141{
142 if (c->percpu_size) {
143 void __percpu **obj = kmalloc_node(c->percpu_size, flags, node);
144 void __percpu *pptr = __alloc_percpu_gfp(c->unit_size, 8, flags);
145
146 if (!obj || !pptr) {
147 free_percpu(pptr);
148 kfree(obj);
149 return NULL;
150 }
151 obj[1] = pptr;
152 return obj;
153 }
154
155 return kmalloc_node(c->unit_size, flags | __GFP_ZERO, node);
156}
157
158static struct mem_cgroup *get_memcg(const struct bpf_mem_cache *c)
159{
160#ifdef CONFIG_MEMCG
161 if (c->objcg)
162 return get_mem_cgroup_from_objcg(c->objcg);
163 return root_mem_cgroup;
164#else
165 return NULL;
166#endif
167}
168
169static void inc_active(struct bpf_mem_cache *c, unsigned long *flags)
170{
171 if (IS_ENABLED(CONFIG_PREEMPT_RT))
172 /* In RT irq_work runs in per-cpu kthread, so disable
173 * interrupts to avoid preemption and interrupts and
174 * reduce the chance of bpf prog executing on this cpu
175 * when active counter is busy.
176 */
177 local_irq_save(*flags);
178 /* alloc_bulk runs from irq_work which will not preempt a bpf
179 * program that does unit_alloc/unit_free since IRQs are
180 * disabled there. There is no race to increment 'active'
181 * counter. It protects free_llist from corruption in case NMI
182 * bpf prog preempted this loop.
183 */
184 WARN_ON_ONCE(local_inc_return(&c->active) != 1);
185}
186
187static void dec_active(struct bpf_mem_cache *c, unsigned long *flags)
188{
189 local_dec(&c->active);
190 if (IS_ENABLED(CONFIG_PREEMPT_RT))
191 local_irq_restore(*flags);
192}
193
194static void add_obj_to_free_list(struct bpf_mem_cache *c, void *obj)
195{
196 unsigned long flags;
197
198 inc_active(c, &flags);
199 __llist_add(obj, &c->free_llist);
200 c->free_cnt++;
201 dec_active(c, &flags);
202}
203
204/* Mostly runs from irq_work except __init phase. */
205static void alloc_bulk(struct bpf_mem_cache *c, int cnt, int node, bool atomic)
206{
207 struct mem_cgroup *memcg = NULL, *old_memcg;
208 gfp_t gfp;
209 void *obj;
210 int i;
211
212 gfp = __GFP_NOWARN | __GFP_ACCOUNT;
213 gfp |= atomic ? GFP_NOWAIT : GFP_KERNEL;
214
215 for (i = 0; i < cnt; i++) {
216 /*
217 * For every 'c' llist_del_first(&c->free_by_rcu_ttrace); is
218 * done only by one CPU == current CPU. Other CPUs might
219 * llist_add() and llist_del_all() in parallel.
220 */
221 obj = llist_del_first(&c->free_by_rcu_ttrace);
222 if (!obj)
223 break;
224 add_obj_to_free_list(c, obj);
225 }
226 if (i >= cnt)
227 return;
228
229 for (; i < cnt; i++) {
230 obj = llist_del_first(&c->waiting_for_gp_ttrace);
231 if (!obj)
232 break;
233 add_obj_to_free_list(c, obj);
234 }
235 if (i >= cnt)
236 return;
237
238 memcg = get_memcg(c);
239 old_memcg = set_active_memcg(memcg);
240 for (; i < cnt; i++) {
241 /* Allocate, but don't deplete atomic reserves that typical
242 * GFP_ATOMIC would do. irq_work runs on this cpu and kmalloc
243 * will allocate from the current numa node which is what we
244 * want here.
245 */
246 obj = __alloc(c, node, gfp);
247 if (!obj)
248 break;
249 add_obj_to_free_list(c, obj);
250 }
251 set_active_memcg(old_memcg);
252 mem_cgroup_put(memcg);
253}
254
255static void free_one(void *obj, bool percpu)
256{
257 if (percpu)
258 free_percpu(((void __percpu **)obj)[1]);
259
260 kfree(obj);
261}
262
263static int free_all(struct llist_node *llnode, bool percpu)
264{
265 struct llist_node *pos, *t;
266 int cnt = 0;
267
268 llist_for_each_safe(pos, t, llnode) {
269 free_one(pos, percpu);
270 cnt++;
271 }
272 return cnt;
273}
274
275static void __free_rcu(struct rcu_head *head)
276{
277 struct bpf_mem_cache *c = container_of(head, struct bpf_mem_cache, rcu_ttrace);
278
279 free_all(llist_del_all(&c->waiting_for_gp_ttrace), !!c->percpu_size);
280 atomic_set(&c->call_rcu_ttrace_in_progress, 0);
281}
282
283static void __free_rcu_tasks_trace(struct rcu_head *head)
284{
285 /* If RCU Tasks Trace grace period implies RCU grace period,
286 * there is no need to invoke call_rcu().
287 */
288 if (rcu_trace_implies_rcu_gp())
289 __free_rcu(head);
290 else
291 call_rcu(head, __free_rcu);
292}
293
294static void enque_to_free(struct bpf_mem_cache *c, void *obj)
295{
296 struct llist_node *llnode = obj;
297
298 /* bpf_mem_cache is a per-cpu object. Freeing happens in irq_work.
299 * Nothing races to add to free_by_rcu_ttrace list.
300 */
301 llist_add(llnode, &c->free_by_rcu_ttrace);
302}
303
304static void do_call_rcu_ttrace(struct bpf_mem_cache *c)
305{
306 struct llist_node *llnode, *t;
307
308 if (atomic_xchg(&c->call_rcu_ttrace_in_progress, 1)) {
309 if (unlikely(READ_ONCE(c->draining))) {
310 llnode = llist_del_all(&c->free_by_rcu_ttrace);
311 free_all(llnode, !!c->percpu_size);
312 }
313 return;
314 }
315
316 WARN_ON_ONCE(!llist_empty(&c->waiting_for_gp_ttrace));
317 llist_for_each_safe(llnode, t, llist_del_all(&c->free_by_rcu_ttrace))
318 llist_add(llnode, &c->waiting_for_gp_ttrace);
319
320 if (unlikely(READ_ONCE(c->draining))) {
321 __free_rcu(&c->rcu_ttrace);
322 return;
323 }
324
325 /* Use call_rcu_tasks_trace() to wait for sleepable progs to finish.
326 * If RCU Tasks Trace grace period implies RCU grace period, free
327 * these elements directly, else use call_rcu() to wait for normal
328 * progs to finish and finally do free_one() on each element.
329 */
330 call_rcu_tasks_trace(&c->rcu_ttrace, __free_rcu_tasks_trace);
331}
332
333static void free_bulk(struct bpf_mem_cache *c)
334{
335 struct bpf_mem_cache *tgt = c->tgt;
336 struct llist_node *llnode, *t;
337 unsigned long flags;
338 int cnt;
339
340 WARN_ON_ONCE(tgt->unit_size != c->unit_size);
341 WARN_ON_ONCE(tgt->percpu_size != c->percpu_size);
342
343 do {
344 inc_active(c, &flags);
345 llnode = __llist_del_first(&c->free_llist);
346 if (llnode)
347 cnt = --c->free_cnt;
348 else
349 cnt = 0;
350 dec_active(c, &flags);
351 if (llnode)
352 enque_to_free(tgt, llnode);
353 } while (cnt > (c->high_watermark + c->low_watermark) / 2);
354
355 /* and drain free_llist_extra */
356 llist_for_each_safe(llnode, t, llist_del_all(&c->free_llist_extra))
357 enque_to_free(tgt, llnode);
358 do_call_rcu_ttrace(tgt);
359}
360
361static void __free_by_rcu(struct rcu_head *head)
362{
363 struct bpf_mem_cache *c = container_of(head, struct bpf_mem_cache, rcu);
364 struct bpf_mem_cache *tgt = c->tgt;
365 struct llist_node *llnode;
366
367 WARN_ON_ONCE(tgt->unit_size != c->unit_size);
368 WARN_ON_ONCE(tgt->percpu_size != c->percpu_size);
369
370 llnode = llist_del_all(&c->waiting_for_gp);
371 if (!llnode)
372 goto out;
373
374 llist_add_batch(llnode, c->waiting_for_gp_tail, &tgt->free_by_rcu_ttrace);
375
376 /* Objects went through regular RCU GP. Send them to RCU tasks trace */
377 do_call_rcu_ttrace(tgt);
378out:
379 atomic_set(&c->call_rcu_in_progress, 0);
380}
381
382static void check_free_by_rcu(struct bpf_mem_cache *c)
383{
384 struct llist_node *llnode, *t;
385 unsigned long flags;
386
387 /* drain free_llist_extra_rcu */
388 if (unlikely(!llist_empty(&c->free_llist_extra_rcu))) {
389 inc_active(c, &flags);
390 llist_for_each_safe(llnode, t, llist_del_all(&c->free_llist_extra_rcu))
391 if (__llist_add(llnode, &c->free_by_rcu))
392 c->free_by_rcu_tail = llnode;
393 dec_active(c, &flags);
394 }
395
396 if (llist_empty(&c->free_by_rcu))
397 return;
398
399 if (atomic_xchg(&c->call_rcu_in_progress, 1)) {
400 /*
401 * Instead of kmalloc-ing new rcu_head and triggering 10k
402 * call_rcu() to hit rcutree.qhimark and force RCU to notice
403 * the overload just ask RCU to hurry up. There could be many
404 * objects in free_by_rcu list.
405 * This hint reduces memory consumption for an artificial
406 * benchmark from 2 Gbyte to 150 Mbyte.
407 */
408 rcu_request_urgent_qs_task(current);
409 return;
410 }
411
412 WARN_ON_ONCE(!llist_empty(&c->waiting_for_gp));
413
414 inc_active(c, &flags);
415 WRITE_ONCE(c->waiting_for_gp.first, __llist_del_all(&c->free_by_rcu));
416 c->waiting_for_gp_tail = c->free_by_rcu_tail;
417 dec_active(c, &flags);
418
419 if (unlikely(READ_ONCE(c->draining))) {
420 free_all(llist_del_all(&c->waiting_for_gp), !!c->percpu_size);
421 atomic_set(&c->call_rcu_in_progress, 0);
422 } else {
423 call_rcu_hurry(&c->rcu, __free_by_rcu);
424 }
425}
426
427static void bpf_mem_refill(struct irq_work *work)
428{
429 struct bpf_mem_cache *c = container_of(work, struct bpf_mem_cache, refill_work);
430 int cnt;
431
432 /* Racy access to free_cnt. It doesn't need to be 100% accurate */
433 cnt = c->free_cnt;
434 if (cnt < c->low_watermark)
435 /* irq_work runs on this cpu and kmalloc will allocate
436 * from the current numa node which is what we want here.
437 */
438 alloc_bulk(c, c->batch, NUMA_NO_NODE, true);
439 else if (cnt > c->high_watermark)
440 free_bulk(c);
441
442 check_free_by_rcu(c);
443}
444
445static void notrace irq_work_raise(struct bpf_mem_cache *c)
446{
447 irq_work_queue(&c->refill_work);
448}
449
450/* For typical bpf map case that uses bpf_mem_cache_alloc and single bucket
451 * the freelist cache will be elem_size * 64 (or less) on each cpu.
452 *
453 * For bpf programs that don't have statically known allocation sizes and
454 * assuming (low_mark + high_mark) / 2 as an average number of elements per
455 * bucket and all buckets are used the total amount of memory in freelists
456 * on each cpu will be:
457 * 64*16 + 64*32 + 64*64 + 64*96 + 64*128 + 64*196 + 64*256 + 32*512 + 16*1024 + 8*2048 + 4*4096
458 * == ~ 116 Kbyte using below heuristic.
459 * Initialized, but unused bpf allocator (not bpf map specific one) will
460 * consume ~ 11 Kbyte per cpu.
461 * Typical case will be between 11K and 116K closer to 11K.
462 * bpf progs can and should share bpf_mem_cache when possible.
463 *
464 * Percpu allocation is typically rare. To avoid potential unnecessary large
465 * memory consumption, set low_mark = 1 and high_mark = 3, resulting in c->batch = 1.
466 */
467static void init_refill_work(struct bpf_mem_cache *c)
468{
469 init_irq_work(&c->refill_work, bpf_mem_refill);
470 if (c->percpu_size) {
471 c->low_watermark = 1;
472 c->high_watermark = 3;
473 } else if (c->unit_size <= 256) {
474 c->low_watermark = 32;
475 c->high_watermark = 96;
476 } else {
477 /* When page_size == 4k, order-0 cache will have low_mark == 2
478 * and high_mark == 6 with batch alloc of 3 individual pages at
479 * a time.
480 * 8k allocs and above low == 1, high == 3, batch == 1.
481 */
482 c->low_watermark = max(32 * 256 / c->unit_size, 1);
483 c->high_watermark = max(96 * 256 / c->unit_size, 3);
484 }
485 c->batch = max((c->high_watermark - c->low_watermark) / 4 * 3, 1);
486}
487
488static void prefill_mem_cache(struct bpf_mem_cache *c, int cpu)
489{
490 int cnt = 1;
491
492 /* To avoid consuming memory, for non-percpu allocation, assume that
493 * 1st run of bpf prog won't be doing more than 4 map_update_elem from
494 * irq disabled region if unit size is less than or equal to 256.
495 * For all other cases, let us just do one allocation.
496 */
497 if (!c->percpu_size && c->unit_size <= 256)
498 cnt = 4;
499 alloc_bulk(c, cnt, cpu_to_node(cpu), false);
500}
501
502/* When size != 0 bpf_mem_cache for each cpu.
503 * This is typical bpf hash map use case when all elements have equal size.
504 *
505 * When size == 0 allocate 11 bpf_mem_cache-s for each cpu, then rely on
506 * kmalloc/kfree. Max allocation size is 4096 in this case.
507 * This is bpf_dynptr and bpf_kptr use case.
508 */
509int bpf_mem_alloc_init(struct bpf_mem_alloc *ma, int size, bool percpu)
510{
511 struct bpf_mem_caches *cc; struct bpf_mem_caches __percpu *pcc;
512 struct bpf_mem_cache *c; struct bpf_mem_cache __percpu *pc;
513 struct obj_cgroup *objcg = NULL;
514 int cpu, i, unit_size, percpu_size = 0;
515
516 if (percpu && size == 0)
517 return -EINVAL;
518
519 /* room for llist_node and per-cpu pointer */
520 if (percpu)
521 percpu_size = LLIST_NODE_SZ + sizeof(void *);
522 ma->percpu = percpu;
523
524 if (size) {
525 pc = __alloc_percpu_gfp(sizeof(*pc), 8, GFP_KERNEL);
526 if (!pc)
527 return -ENOMEM;
528
529 if (!percpu)
530 size += LLIST_NODE_SZ; /* room for llist_node */
531 unit_size = size;
532
533#ifdef CONFIG_MEMCG
534 if (memcg_bpf_enabled())
535 objcg = get_obj_cgroup_from_current();
536#endif
537 ma->objcg = objcg;
538
539 for_each_possible_cpu(cpu) {
540 c = per_cpu_ptr(pc, cpu);
541 c->unit_size = unit_size;
542 c->objcg = objcg;
543 c->percpu_size = percpu_size;
544 c->tgt = c;
545 init_refill_work(c);
546 prefill_mem_cache(c, cpu);
547 }
548 ma->cache = pc;
549 return 0;
550 }
551
552 pcc = __alloc_percpu_gfp(sizeof(*cc), 8, GFP_KERNEL);
553 if (!pcc)
554 return -ENOMEM;
555#ifdef CONFIG_MEMCG
556 objcg = get_obj_cgroup_from_current();
557#endif
558 ma->objcg = objcg;
559 for_each_possible_cpu(cpu) {
560 cc = per_cpu_ptr(pcc, cpu);
561 for (i = 0; i < NUM_CACHES; i++) {
562 c = &cc->cache[i];
563 c->unit_size = sizes[i];
564 c->objcg = objcg;
565 c->percpu_size = percpu_size;
566 c->tgt = c;
567
568 init_refill_work(c);
569 prefill_mem_cache(c, cpu);
570 }
571 }
572
573 ma->caches = pcc;
574 return 0;
575}
576
577int bpf_mem_alloc_percpu_init(struct bpf_mem_alloc *ma, struct obj_cgroup *objcg)
578{
579 struct bpf_mem_caches __percpu *pcc;
580
581 pcc = __alloc_percpu_gfp(sizeof(struct bpf_mem_caches), 8, GFP_KERNEL);
582 if (!pcc)
583 return -ENOMEM;
584
585 ma->caches = pcc;
586 ma->objcg = objcg;
587 ma->percpu = true;
588 return 0;
589}
590
591int bpf_mem_alloc_percpu_unit_init(struct bpf_mem_alloc *ma, int size)
592{
593 struct bpf_mem_caches *cc; struct bpf_mem_caches __percpu *pcc;
594 int cpu, i, unit_size, percpu_size;
595 struct obj_cgroup *objcg;
596 struct bpf_mem_cache *c;
597
598 i = bpf_mem_cache_idx(size);
599 if (i < 0)
600 return -EINVAL;
601
602 /* room for llist_node and per-cpu pointer */
603 percpu_size = LLIST_NODE_SZ + sizeof(void *);
604
605 unit_size = sizes[i];
606 objcg = ma->objcg;
607 pcc = ma->caches;
608
609 for_each_possible_cpu(cpu) {
610 cc = per_cpu_ptr(pcc, cpu);
611 c = &cc->cache[i];
612 if (c->unit_size)
613 break;
614
615 c->unit_size = unit_size;
616 c->objcg = objcg;
617 c->percpu_size = percpu_size;
618 c->tgt = c;
619
620 init_refill_work(c);
621 prefill_mem_cache(c, cpu);
622 }
623
624 return 0;
625}
626
627static void drain_mem_cache(struct bpf_mem_cache *c)
628{
629 bool percpu = !!c->percpu_size;
630
631 /* No progs are using this bpf_mem_cache, but htab_map_free() called
632 * bpf_mem_cache_free() for all remaining elements and they can be in
633 * free_by_rcu_ttrace or in waiting_for_gp_ttrace lists, so drain those lists now.
634 *
635 * Except for waiting_for_gp_ttrace list, there are no concurrent operations
636 * on these lists, so it is safe to use __llist_del_all().
637 */
638 free_all(llist_del_all(&c->free_by_rcu_ttrace), percpu);
639 free_all(llist_del_all(&c->waiting_for_gp_ttrace), percpu);
640 free_all(__llist_del_all(&c->free_llist), percpu);
641 free_all(__llist_del_all(&c->free_llist_extra), percpu);
642 free_all(__llist_del_all(&c->free_by_rcu), percpu);
643 free_all(__llist_del_all(&c->free_llist_extra_rcu), percpu);
644 free_all(llist_del_all(&c->waiting_for_gp), percpu);
645}
646
647static void check_mem_cache(struct bpf_mem_cache *c)
648{
649 WARN_ON_ONCE(!llist_empty(&c->free_by_rcu_ttrace));
650 WARN_ON_ONCE(!llist_empty(&c->waiting_for_gp_ttrace));
651 WARN_ON_ONCE(!llist_empty(&c->free_llist));
652 WARN_ON_ONCE(!llist_empty(&c->free_llist_extra));
653 WARN_ON_ONCE(!llist_empty(&c->free_by_rcu));
654 WARN_ON_ONCE(!llist_empty(&c->free_llist_extra_rcu));
655 WARN_ON_ONCE(!llist_empty(&c->waiting_for_gp));
656}
657
658static void check_leaked_objs(struct bpf_mem_alloc *ma)
659{
660 struct bpf_mem_caches *cc;
661 struct bpf_mem_cache *c;
662 int cpu, i;
663
664 if (ma->cache) {
665 for_each_possible_cpu(cpu) {
666 c = per_cpu_ptr(ma->cache, cpu);
667 check_mem_cache(c);
668 }
669 }
670 if (ma->caches) {
671 for_each_possible_cpu(cpu) {
672 cc = per_cpu_ptr(ma->caches, cpu);
673 for (i = 0; i < NUM_CACHES; i++) {
674 c = &cc->cache[i];
675 check_mem_cache(c);
676 }
677 }
678 }
679}
680
681static void free_mem_alloc_no_barrier(struct bpf_mem_alloc *ma)
682{
683 check_leaked_objs(ma);
684 free_percpu(ma->cache);
685 free_percpu(ma->caches);
686 ma->cache = NULL;
687 ma->caches = NULL;
688}
689
690static void free_mem_alloc(struct bpf_mem_alloc *ma)
691{
692 /* waiting_for_gp[_ttrace] lists were drained, but RCU callbacks
693 * might still execute. Wait for them.
694 *
695 * rcu_barrier_tasks_trace() doesn't imply synchronize_rcu_tasks_trace(),
696 * but rcu_barrier_tasks_trace() and rcu_barrier() below are only used
697 * to wait for the pending __free_rcu_tasks_trace() and __free_rcu(),
698 * so if call_rcu(head, __free_rcu) is skipped due to
699 * rcu_trace_implies_rcu_gp(), it will be OK to skip rcu_barrier() by
700 * using rcu_trace_implies_rcu_gp() as well.
701 */
702 rcu_barrier(); /* wait for __free_by_rcu */
703 rcu_barrier_tasks_trace(); /* wait for __free_rcu */
704 if (!rcu_trace_implies_rcu_gp())
705 rcu_barrier();
706 free_mem_alloc_no_barrier(ma);
707}
708
709static void free_mem_alloc_deferred(struct work_struct *work)
710{
711 struct bpf_mem_alloc *ma = container_of(work, struct bpf_mem_alloc, work);
712
713 free_mem_alloc(ma);
714 kfree(ma);
715}
716
717static void destroy_mem_alloc(struct bpf_mem_alloc *ma, int rcu_in_progress)
718{
719 struct bpf_mem_alloc *copy;
720
721 if (!rcu_in_progress) {
722 /* Fast path. No callbacks are pending, hence no need to do
723 * rcu_barrier-s.
724 */
725 free_mem_alloc_no_barrier(ma);
726 return;
727 }
728
729 copy = kmemdup(ma, sizeof(*ma), GFP_KERNEL);
730 if (!copy) {
731 /* Slow path with inline barrier-s */
732 free_mem_alloc(ma);
733 return;
734 }
735
736 /* Defer barriers into worker to let the rest of map memory to be freed */
737 memset(ma, 0, sizeof(*ma));
738 INIT_WORK(©->work, free_mem_alloc_deferred);
739 queue_work(system_unbound_wq, ©->work);
740}
741
742void bpf_mem_alloc_destroy(struct bpf_mem_alloc *ma)
743{
744 struct bpf_mem_caches *cc;
745 struct bpf_mem_cache *c;
746 int cpu, i, rcu_in_progress;
747
748 if (ma->cache) {
749 rcu_in_progress = 0;
750 for_each_possible_cpu(cpu) {
751 c = per_cpu_ptr(ma->cache, cpu);
752 WRITE_ONCE(c->draining, true);
753 irq_work_sync(&c->refill_work);
754 drain_mem_cache(c);
755 rcu_in_progress += atomic_read(&c->call_rcu_ttrace_in_progress);
756 rcu_in_progress += atomic_read(&c->call_rcu_in_progress);
757 }
758 obj_cgroup_put(ma->objcg);
759 destroy_mem_alloc(ma, rcu_in_progress);
760 }
761 if (ma->caches) {
762 rcu_in_progress = 0;
763 for_each_possible_cpu(cpu) {
764 cc = per_cpu_ptr(ma->caches, cpu);
765 for (i = 0; i < NUM_CACHES; i++) {
766 c = &cc->cache[i];
767 WRITE_ONCE(c->draining, true);
768 irq_work_sync(&c->refill_work);
769 drain_mem_cache(c);
770 rcu_in_progress += atomic_read(&c->call_rcu_ttrace_in_progress);
771 rcu_in_progress += atomic_read(&c->call_rcu_in_progress);
772 }
773 }
774 obj_cgroup_put(ma->objcg);
775 destroy_mem_alloc(ma, rcu_in_progress);
776 }
777}
778
779/* notrace is necessary here and in other functions to make sure
780 * bpf programs cannot attach to them and cause llist corruptions.
781 */
782static void notrace *unit_alloc(struct bpf_mem_cache *c)
783{
784 struct llist_node *llnode = NULL;
785 unsigned long flags;
786 int cnt = 0;
787
788 /* Disable irqs to prevent the following race for majority of prog types:
789 * prog_A
790 * bpf_mem_alloc
791 * preemption or irq -> prog_B
792 * bpf_mem_alloc
793 *
794 * but prog_B could be a perf_event NMI prog.
795 * Use per-cpu 'active' counter to order free_list access between
796 * unit_alloc/unit_free/bpf_mem_refill.
797 */
798 local_irq_save(flags);
799 if (local_inc_return(&c->active) == 1) {
800 llnode = __llist_del_first(&c->free_llist);
801 if (llnode) {
802 cnt = --c->free_cnt;
803 *(struct bpf_mem_cache **)llnode = c;
804 }
805 }
806 local_dec(&c->active);
807
808 WARN_ON(cnt < 0);
809
810 if (cnt < c->low_watermark)
811 irq_work_raise(c);
812 /* Enable IRQ after the enqueue of irq work completes, so irq work
813 * will run after IRQ is enabled and free_llist may be refilled by
814 * irq work before other task preempts current task.
815 */
816 local_irq_restore(flags);
817
818 return llnode;
819}
820
821/* Though 'ptr' object could have been allocated on a different cpu
822 * add it to the free_llist of the current cpu.
823 * Let kfree() logic deal with it when it's later called from irq_work.
824 */
825static void notrace unit_free(struct bpf_mem_cache *c, void *ptr)
826{
827 struct llist_node *llnode = ptr - LLIST_NODE_SZ;
828 unsigned long flags;
829 int cnt = 0;
830
831 BUILD_BUG_ON(LLIST_NODE_SZ > 8);
832
833 /*
834 * Remember bpf_mem_cache that allocated this object.
835 * The hint is not accurate.
836 */
837 c->tgt = *(struct bpf_mem_cache **)llnode;
838
839 local_irq_save(flags);
840 if (local_inc_return(&c->active) == 1) {
841 __llist_add(llnode, &c->free_llist);
842 cnt = ++c->free_cnt;
843 } else {
844 /* unit_free() cannot fail. Therefore add an object to atomic
845 * llist. free_bulk() will drain it. Though free_llist_extra is
846 * a per-cpu list we have to use atomic llist_add here, since
847 * it also can be interrupted by bpf nmi prog that does another
848 * unit_free() into the same free_llist_extra.
849 */
850 llist_add(llnode, &c->free_llist_extra);
851 }
852 local_dec(&c->active);
853
854 if (cnt > c->high_watermark)
855 /* free few objects from current cpu into global kmalloc pool */
856 irq_work_raise(c);
857 /* Enable IRQ after irq_work_raise() completes, otherwise when current
858 * task is preempted by task which does unit_alloc(), unit_alloc() may
859 * return NULL unexpectedly because irq work is already pending but can
860 * not been triggered and free_llist can not be refilled timely.
861 */
862 local_irq_restore(flags);
863}
864
865static void notrace unit_free_rcu(struct bpf_mem_cache *c, void *ptr)
866{
867 struct llist_node *llnode = ptr - LLIST_NODE_SZ;
868 unsigned long flags;
869
870 c->tgt = *(struct bpf_mem_cache **)llnode;
871
872 local_irq_save(flags);
873 if (local_inc_return(&c->active) == 1) {
874 if (__llist_add(llnode, &c->free_by_rcu))
875 c->free_by_rcu_tail = llnode;
876 } else {
877 llist_add(llnode, &c->free_llist_extra_rcu);
878 }
879 local_dec(&c->active);
880
881 if (!atomic_read(&c->call_rcu_in_progress))
882 irq_work_raise(c);
883 local_irq_restore(flags);
884}
885
886/* Called from BPF program or from sys_bpf syscall.
887 * In both cases migration is disabled.
888 */
889void notrace *bpf_mem_alloc(struct bpf_mem_alloc *ma, size_t size)
890{
891 int idx;
892 void *ret;
893
894 if (!size)
895 return NULL;
896
897 if (!ma->percpu)
898 size += LLIST_NODE_SZ;
899 idx = bpf_mem_cache_idx(size);
900 if (idx < 0)
901 return NULL;
902
903 ret = unit_alloc(this_cpu_ptr(ma->caches)->cache + idx);
904 return !ret ? NULL : ret + LLIST_NODE_SZ;
905}
906
907void notrace bpf_mem_free(struct bpf_mem_alloc *ma, void *ptr)
908{
909 struct bpf_mem_cache *c;
910 int idx;
911
912 if (!ptr)
913 return;
914
915 c = *(void **)(ptr - LLIST_NODE_SZ);
916 idx = bpf_mem_cache_idx(c->unit_size);
917 if (WARN_ON_ONCE(idx < 0))
918 return;
919
920 unit_free(this_cpu_ptr(ma->caches)->cache + idx, ptr);
921}
922
923void notrace bpf_mem_free_rcu(struct bpf_mem_alloc *ma, void *ptr)
924{
925 struct bpf_mem_cache *c;
926 int idx;
927
928 if (!ptr)
929 return;
930
931 c = *(void **)(ptr - LLIST_NODE_SZ);
932 idx = bpf_mem_cache_idx(c->unit_size);
933 if (WARN_ON_ONCE(idx < 0))
934 return;
935
936 unit_free_rcu(this_cpu_ptr(ma->caches)->cache + idx, ptr);
937}
938
939void notrace *bpf_mem_cache_alloc(struct bpf_mem_alloc *ma)
940{
941 void *ret;
942
943 ret = unit_alloc(this_cpu_ptr(ma->cache));
944 return !ret ? NULL : ret + LLIST_NODE_SZ;
945}
946
947void notrace bpf_mem_cache_free(struct bpf_mem_alloc *ma, void *ptr)
948{
949 if (!ptr)
950 return;
951
952 unit_free(this_cpu_ptr(ma->cache), ptr);
953}
954
955void notrace bpf_mem_cache_free_rcu(struct bpf_mem_alloc *ma, void *ptr)
956{
957 if (!ptr)
958 return;
959
960 unit_free_rcu(this_cpu_ptr(ma->cache), ptr);
961}
962
963/* Directly does a kfree() without putting 'ptr' back to the free_llist
964 * for reuse and without waiting for a rcu_tasks_trace gp.
965 * The caller must first go through the rcu_tasks_trace gp for 'ptr'
966 * before calling bpf_mem_cache_raw_free().
967 * It could be used when the rcu_tasks_trace callback does not have
968 * a hold on the original bpf_mem_alloc object that allocated the
969 * 'ptr'. This should only be used in the uncommon code path.
970 * Otherwise, the bpf_mem_alloc's free_llist cannot be refilled
971 * and may affect performance.
972 */
973void bpf_mem_cache_raw_free(void *ptr)
974{
975 if (!ptr)
976 return;
977
978 kfree(ptr - LLIST_NODE_SZ);
979}
980
981/* When flags == GFP_KERNEL, it signals that the caller will not cause
982 * deadlock when using kmalloc. bpf_mem_cache_alloc_flags() will use
983 * kmalloc if the free_llist is empty.
984 */
985void notrace *bpf_mem_cache_alloc_flags(struct bpf_mem_alloc *ma, gfp_t flags)
986{
987 struct bpf_mem_cache *c;
988 void *ret;
989
990 c = this_cpu_ptr(ma->cache);
991
992 ret = unit_alloc(c);
993 if (!ret && flags == GFP_KERNEL) {
994 struct mem_cgroup *memcg, *old_memcg;
995
996 memcg = get_memcg(c);
997 old_memcg = set_active_memcg(memcg);
998 ret = __alloc(c, NUMA_NO_NODE, GFP_KERNEL | __GFP_NOWARN | __GFP_ACCOUNT);
999 if (ret)
1000 *(struct bpf_mem_cache **)ret = c;
1001 set_active_memcg(old_memcg);
1002 mem_cgroup_put(memcg);
1003 }
1004
1005 return !ret ? NULL : ret + LLIST_NODE_SZ;
1006}
1007
1008int bpf_mem_alloc_check_size(bool percpu, size_t size)
1009{
1010 /* The size of percpu allocation doesn't have LLIST_NODE_SZ overhead */
1011 if ((percpu && size > BPF_MEM_ALLOC_SIZE_MAX) ||
1012 (!percpu && size > BPF_MEM_ALLOC_SIZE_MAX - LLIST_NODE_SZ))
1013 return -E2BIG;
1014
1015 return 0;
1016}