xref: /linux/kernel/bpf/memalloc.c (revision 3fd6c59042dbba50391e30862beac979491145fe)
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 */
41 static 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 
bpf_mem_cache_idx(size_t size)68 static 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 
81 struct 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 
122 struct bpf_mem_caches {
123 	struct bpf_mem_cache cache[NUM_CACHES];
124 };
125 
126 static const u16 sizes[NUM_CACHES] = {96, 192, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096};
127 
__llist_del_first(struct llist_head * head)128 static 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 
__alloc(struct bpf_mem_cache * c,int node,gfp_t flags)140 static 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 
get_memcg(const struct bpf_mem_cache * c)158 static 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 
inc_active(struct bpf_mem_cache * c,unsigned long * flags)169 static 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 
dec_active(struct bpf_mem_cache * c,unsigned long * flags)187 static 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 
add_obj_to_free_list(struct bpf_mem_cache * c,void * obj)194 static 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. */
alloc_bulk(struct bpf_mem_cache * c,int cnt,int node,bool atomic)205 static 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 
free_one(void * obj,bool percpu)255 static void free_one(void *obj, bool percpu)
256 {
257 	if (percpu)
258 		free_percpu(((void __percpu **)obj)[1]);
259 
260 	kfree(obj);
261 }
262 
free_all(struct llist_node * llnode,bool percpu)263 static 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 
__free_rcu(struct rcu_head * head)275 static 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 
__free_rcu_tasks_trace(struct rcu_head * head)283 static 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 
enque_to_free(struct bpf_mem_cache * c,void * obj)294 static 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 
do_call_rcu_ttrace(struct bpf_mem_cache * c)304 static 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 
free_bulk(struct bpf_mem_cache * c)333 static 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 
__free_by_rcu(struct rcu_head * head)361 static 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);
378 out:
379 	atomic_set(&c->call_rcu_in_progress, 0);
380 }
381 
check_free_by_rcu(struct bpf_mem_cache * c)382 static 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 
bpf_mem_refill(struct irq_work * work)427 static 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 
irq_work_raise(struct bpf_mem_cache * c)445 static 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  */
init_refill_work(struct bpf_mem_cache * c)467 static 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 
prefill_mem_cache(struct bpf_mem_cache * c,int cpu)488 static 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  */
bpf_mem_alloc_init(struct bpf_mem_alloc * ma,int size,bool percpu)509 int 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 
bpf_mem_alloc_percpu_init(struct bpf_mem_alloc * ma,struct obj_cgroup * objcg)577 int 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 
bpf_mem_alloc_percpu_unit_init(struct bpf_mem_alloc * ma,int size)591 int 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 
drain_mem_cache(struct bpf_mem_cache * c)627 static 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 
check_mem_cache(struct bpf_mem_cache * c)647 static 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 
check_leaked_objs(struct bpf_mem_alloc * ma)658 static 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 
free_mem_alloc_no_barrier(struct bpf_mem_alloc * ma)681 static 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 
free_mem_alloc(struct bpf_mem_alloc * ma)690 static 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 
free_mem_alloc_deferred(struct work_struct * work)709 static 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 
destroy_mem_alloc(struct bpf_mem_alloc * ma,int rcu_in_progress)717 static 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(&copy->work, free_mem_alloc_deferred);
739 	queue_work(system_unbound_wq, &copy->work);
740 }
741 
bpf_mem_alloc_destroy(struct bpf_mem_alloc * ma)742 void 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  */
unit_alloc(struct bpf_mem_cache * c)782 static 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  */
unit_free(struct bpf_mem_cache * c,void * ptr)825 static 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 
unit_free_rcu(struct bpf_mem_cache * c,void * ptr)865 static 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  */
bpf_mem_alloc(struct bpf_mem_alloc * ma,size_t size)889 void 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 
bpf_mem_free(struct bpf_mem_alloc * ma,void * ptr)907 void 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 
bpf_mem_free_rcu(struct bpf_mem_alloc * ma,void * ptr)923 void 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 
bpf_mem_cache_alloc(struct bpf_mem_alloc * ma)939 void 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 
bpf_mem_cache_free(struct bpf_mem_alloc * ma,void * ptr)947 void 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 
bpf_mem_cache_free_rcu(struct bpf_mem_alloc * ma,void * ptr)955 void 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  */
bpf_mem_cache_raw_free(void * ptr)973 void 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  */
bpf_mem_cache_alloc_flags(struct bpf_mem_alloc * ma,gfp_t flags)985 void 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 
bpf_mem_alloc_check_size(bool percpu,size_t size)1008 int 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 }
1017