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