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 
bpf_mem_cache_idx(size_t size)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) - 1;
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 
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 
108 struct bpf_mem_caches {
109 	struct bpf_mem_cache cache[NUM_CACHES];
110 };
111 
__llist_del_first(struct llist_head * head)112 static 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 
__alloc(struct bpf_mem_cache * c,int node)124 static 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 
get_memcg(const struct bpf_mem_cache * c)149 static 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. */
alloc_bulk(struct bpf_mem_cache * c,int cnt,int node)164 static 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 		obj = __alloc(c, node);
175 		if (!obj)
176 			break;
177 		if (IS_ENABLED(CONFIG_PREEMPT_RT))
178 			/* In RT irq_work runs in per-cpu kthread, so disable
179 			 * interrupts to avoid preemption and interrupts and
180 			 * reduce the chance of bpf prog executing on this cpu
181 			 * when active counter is busy.
182 			 */
183 			local_irq_save(flags);
184 		/* alloc_bulk runs from irq_work which will not preempt a bpf
185 		 * program that does unit_alloc/unit_free since IRQs are
186 		 * disabled there. There is no race to increment 'active'
187 		 * counter. It protects free_llist from corruption in case NMI
188 		 * bpf prog preempted this loop.
189 		 */
190 		WARN_ON_ONCE(local_inc_return(&c->active) != 1);
191 		__llist_add(obj, &c->free_llist);
192 		c->free_cnt++;
193 		local_dec(&c->active);
194 		if (IS_ENABLED(CONFIG_PREEMPT_RT))
195 			local_irq_restore(flags);
196 	}
197 	set_active_memcg(old_memcg);
198 	mem_cgroup_put(memcg);
199 }
200 
free_one(struct bpf_mem_cache * c,void * obj)201 static void free_one(struct bpf_mem_cache *c, void *obj)
202 {
203 	if (c->percpu_size) {
204 		free_percpu(((void **)obj)[1]);
205 		kfree(obj);
206 		return;
207 	}
208 
209 	kfree(obj);
210 }
211 
__free_rcu(struct rcu_head * head)212 static void __free_rcu(struct rcu_head *head)
213 {
214 	struct bpf_mem_cache *c = container_of(head, struct bpf_mem_cache, rcu);
215 	struct llist_node *llnode = llist_del_all(&c->waiting_for_gp);
216 	struct llist_node *pos, *t;
217 
218 	llist_for_each_safe(pos, t, llnode)
219 		free_one(c, pos);
220 	atomic_set(&c->call_rcu_in_progress, 0);
221 }
222 
__free_rcu_tasks_trace(struct rcu_head * head)223 static void __free_rcu_tasks_trace(struct rcu_head *head)
224 {
225 	struct bpf_mem_cache *c = container_of(head, struct bpf_mem_cache, rcu);
226 
227 	call_rcu(&c->rcu, __free_rcu);
228 }
229 
enque_to_free(struct bpf_mem_cache * c,void * obj)230 static void enque_to_free(struct bpf_mem_cache *c, void *obj)
231 {
232 	struct llist_node *llnode = obj;
233 
234 	/* bpf_mem_cache is a per-cpu object. Freeing happens in irq_work.
235 	 * Nothing races to add to free_by_rcu list.
236 	 */
237 	__llist_add(llnode, &c->free_by_rcu);
238 }
239 
do_call_rcu(struct bpf_mem_cache * c)240 static void do_call_rcu(struct bpf_mem_cache *c)
241 {
242 	struct llist_node *llnode, *t;
243 
244 	if (atomic_xchg(&c->call_rcu_in_progress, 1))
245 		return;
246 
247 	WARN_ON_ONCE(!llist_empty(&c->waiting_for_gp));
248 	llist_for_each_safe(llnode, t, __llist_del_all(&c->free_by_rcu))
249 		/* There is no concurrent __llist_add(waiting_for_gp) access.
250 		 * It doesn't race with llist_del_all either.
251 		 * But there could be two concurrent llist_del_all(waiting_for_gp):
252 		 * from __free_rcu() and from drain_mem_cache().
253 		 */
254 		__llist_add(llnode, &c->waiting_for_gp);
255 	/* Use call_rcu_tasks_trace() to wait for sleepable progs to finish.
256 	 * Then use call_rcu() to wait for normal progs to finish
257 	 * and finally do free_one() on each element.
258 	 */
259 	call_rcu_tasks_trace(&c->rcu, __free_rcu_tasks_trace);
260 }
261 
free_bulk(struct bpf_mem_cache * c)262 static void free_bulk(struct bpf_mem_cache *c)
263 {
264 	struct llist_node *llnode, *t;
265 	unsigned long flags;
266 	int cnt;
267 
268 	do {
269 		if (IS_ENABLED(CONFIG_PREEMPT_RT))
270 			local_irq_save(flags);
271 		WARN_ON_ONCE(local_inc_return(&c->active) != 1);
272 		llnode = __llist_del_first(&c->free_llist);
273 		if (llnode)
274 			cnt = --c->free_cnt;
275 		else
276 			cnt = 0;
277 		local_dec(&c->active);
278 		if (IS_ENABLED(CONFIG_PREEMPT_RT))
279 			local_irq_restore(flags);
280 		if (llnode)
281 			enque_to_free(c, llnode);
282 	} while (cnt > (c->high_watermark + c->low_watermark) / 2);
283 
284 	/* and drain free_llist_extra */
285 	llist_for_each_safe(llnode, t, llist_del_all(&c->free_llist_extra))
286 		enque_to_free(c, llnode);
287 	do_call_rcu(c);
288 }
289 
bpf_mem_refill(struct irq_work * work)290 static void bpf_mem_refill(struct irq_work *work)
291 {
292 	struct bpf_mem_cache *c = container_of(work, struct bpf_mem_cache, refill_work);
293 	int cnt;
294 
295 	/* Racy access to free_cnt. It doesn't need to be 100% accurate */
296 	cnt = c->free_cnt;
297 	if (cnt < c->low_watermark)
298 		/* irq_work runs on this cpu and kmalloc will allocate
299 		 * from the current numa node which is what we want here.
300 		 */
301 		alloc_bulk(c, c->batch, NUMA_NO_NODE);
302 	else if (cnt > c->high_watermark)
303 		free_bulk(c);
304 }
305 
irq_work_raise(struct bpf_mem_cache * c)306 static void notrace irq_work_raise(struct bpf_mem_cache *c)
307 {
308 	irq_work_queue(&c->refill_work);
309 }
310 
311 /* For typical bpf map case that uses bpf_mem_cache_alloc and single bucket
312  * the freelist cache will be elem_size * 64 (or less) on each cpu.
313  *
314  * For bpf programs that don't have statically known allocation sizes and
315  * assuming (low_mark + high_mark) / 2 as an average number of elements per
316  * bucket and all buckets are used the total amount of memory in freelists
317  * on each cpu will be:
318  * 64*16 + 64*32 + 64*64 + 64*96 + 64*128 + 64*196 + 64*256 + 32*512 + 16*1024 + 8*2048 + 4*4096
319  * == ~ 116 Kbyte using below heuristic.
320  * Initialized, but unused bpf allocator (not bpf map specific one) will
321  * consume ~ 11 Kbyte per cpu.
322  * Typical case will be between 11K and 116K closer to 11K.
323  * bpf progs can and should share bpf_mem_cache when possible.
324  */
325 
prefill_mem_cache(struct bpf_mem_cache * c,int cpu)326 static void prefill_mem_cache(struct bpf_mem_cache *c, int cpu)
327 {
328 	init_irq_work(&c->refill_work, bpf_mem_refill);
329 	if (c->unit_size <= 256) {
330 		c->low_watermark = 32;
331 		c->high_watermark = 96;
332 	} else {
333 		/* When page_size == 4k, order-0 cache will have low_mark == 2
334 		 * and high_mark == 6 with batch alloc of 3 individual pages at
335 		 * a time.
336 		 * 8k allocs and above low == 1, high == 3, batch == 1.
337 		 */
338 		c->low_watermark = max(32 * 256 / c->unit_size, 1);
339 		c->high_watermark = max(96 * 256 / c->unit_size, 3);
340 	}
341 	c->batch = max((c->high_watermark - c->low_watermark) / 4 * 3, 1);
342 
343 	/* To avoid consuming memory assume that 1st run of bpf
344 	 * prog won't be doing more than 4 map_update_elem from
345 	 * irq disabled region
346 	 */
347 	alloc_bulk(c, c->unit_size <= 256 ? 4 : 1, cpu_to_node(cpu));
348 }
349 
350 /* When size != 0 bpf_mem_cache for each cpu.
351  * This is typical bpf hash map use case when all elements have equal size.
352  *
353  * When size == 0 allocate 11 bpf_mem_cache-s for each cpu, then rely on
354  * kmalloc/kfree. Max allocation size is 4096 in this case.
355  * This is bpf_dynptr and bpf_kptr use case.
356  */
bpf_mem_alloc_init(struct bpf_mem_alloc * ma,int size,bool percpu)357 int bpf_mem_alloc_init(struct bpf_mem_alloc *ma, int size, bool percpu)
358 {
359 	static u16 sizes[NUM_CACHES] = {96, 192, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096};
360 	struct bpf_mem_caches *cc, __percpu *pcc;
361 	struct bpf_mem_cache *c, __percpu *pc;
362 	struct obj_cgroup *objcg = NULL;
363 	int cpu, i, unit_size, percpu_size = 0;
364 
365 	if (size) {
366 		pc = __alloc_percpu_gfp(sizeof(*pc), 8, GFP_KERNEL);
367 		if (!pc)
368 			return -ENOMEM;
369 
370 		if (percpu)
371 			/* room for llist_node and per-cpu pointer */
372 			percpu_size = LLIST_NODE_SZ + sizeof(void *);
373 		else
374 			size += LLIST_NODE_SZ; /* room for llist_node */
375 		unit_size = size;
376 
377 #ifdef CONFIG_MEMCG_KMEM
378 		objcg = get_obj_cgroup_from_current();
379 #endif
380 		for_each_possible_cpu(cpu) {
381 			c = per_cpu_ptr(pc, cpu);
382 			c->unit_size = unit_size;
383 			c->objcg = objcg;
384 			c->percpu_size = percpu_size;
385 			prefill_mem_cache(c, cpu);
386 		}
387 		ma->cache = pc;
388 		return 0;
389 	}
390 
391 	/* size == 0 && percpu is an invalid combination */
392 	if (WARN_ON_ONCE(percpu))
393 		return -EINVAL;
394 
395 	pcc = __alloc_percpu_gfp(sizeof(*cc), 8, GFP_KERNEL);
396 	if (!pcc)
397 		return -ENOMEM;
398 #ifdef CONFIG_MEMCG_KMEM
399 	objcg = get_obj_cgroup_from_current();
400 #endif
401 	for_each_possible_cpu(cpu) {
402 		cc = per_cpu_ptr(pcc, cpu);
403 		for (i = 0; i < NUM_CACHES; i++) {
404 			c = &cc->cache[i];
405 			c->unit_size = sizes[i];
406 			c->objcg = objcg;
407 			prefill_mem_cache(c, cpu);
408 		}
409 	}
410 	ma->caches = pcc;
411 	return 0;
412 }
413 
drain_mem_cache(struct bpf_mem_cache * c)414 static void drain_mem_cache(struct bpf_mem_cache *c)
415 {
416 	struct llist_node *llnode, *t;
417 
418 	/* No progs are using this bpf_mem_cache, but htab_map_free() called
419 	 * bpf_mem_cache_free() for all remaining elements and they can be in
420 	 * free_by_rcu or in waiting_for_gp lists, so drain those lists now.
421 	 *
422 	 * Except for waiting_for_gp list, there are no concurrent operations
423 	 * on these lists, so it is safe to use __llist_del_all().
424 	 */
425 	llist_for_each_safe(llnode, t, __llist_del_all(&c->free_by_rcu))
426 		free_one(c, llnode);
427 	llist_for_each_safe(llnode, t, llist_del_all(&c->waiting_for_gp))
428 		free_one(c, llnode);
429 	llist_for_each_safe(llnode, t, __llist_del_all(&c->free_llist))
430 		free_one(c, llnode);
431 	llist_for_each_safe(llnode, t, __llist_del_all(&c->free_llist_extra))
432 		free_one(c, llnode);
433 }
434 
free_mem_alloc_no_barrier(struct bpf_mem_alloc * ma)435 static void free_mem_alloc_no_barrier(struct bpf_mem_alloc *ma)
436 {
437 	free_percpu(ma->cache);
438 	free_percpu(ma->caches);
439 	ma->cache = NULL;
440 	ma->caches = NULL;
441 }
442 
free_mem_alloc(struct bpf_mem_alloc * ma)443 static void free_mem_alloc(struct bpf_mem_alloc *ma)
444 {
445 	/* waiting_for_gp lists was drained, but __free_rcu might
446 	 * still execute. Wait for it now before we freeing percpu caches.
447 	 */
448 	rcu_barrier_tasks_trace();
449 	rcu_barrier();
450 	free_mem_alloc_no_barrier(ma);
451 }
452 
free_mem_alloc_deferred(struct work_struct * work)453 static void free_mem_alloc_deferred(struct work_struct *work)
454 {
455 	struct bpf_mem_alloc *ma = container_of(work, struct bpf_mem_alloc, work);
456 
457 	free_mem_alloc(ma);
458 	kfree(ma);
459 }
460 
destroy_mem_alloc(struct bpf_mem_alloc * ma,int rcu_in_progress)461 static void destroy_mem_alloc(struct bpf_mem_alloc *ma, int rcu_in_progress)
462 {
463 	struct bpf_mem_alloc *copy;
464 
465 	if (!rcu_in_progress) {
466 		/* Fast path. No callbacks are pending, hence no need to do
467 		 * rcu_barrier-s.
468 		 */
469 		free_mem_alloc_no_barrier(ma);
470 		return;
471 	}
472 
473 	copy = kmalloc(sizeof(*ma), GFP_KERNEL);
474 	if (!copy) {
475 		/* Slow path with inline barrier-s */
476 		free_mem_alloc(ma);
477 		return;
478 	}
479 
480 	/* Defer barriers into worker to let the rest of map memory to be freed */
481 	copy->cache = ma->cache;
482 	ma->cache = NULL;
483 	copy->caches = ma->caches;
484 	ma->caches = NULL;
485 	INIT_WORK(&copy->work, free_mem_alloc_deferred);
486 	queue_work(system_unbound_wq, &copy->work);
487 }
488 
bpf_mem_alloc_destroy(struct bpf_mem_alloc * ma)489 void bpf_mem_alloc_destroy(struct bpf_mem_alloc *ma)
490 {
491 	struct bpf_mem_caches *cc;
492 	struct bpf_mem_cache *c;
493 	int cpu, i, rcu_in_progress;
494 
495 	if (ma->cache) {
496 		rcu_in_progress = 0;
497 		for_each_possible_cpu(cpu) {
498 			c = per_cpu_ptr(ma->cache, cpu);
499 			/*
500 			 * refill_work may be unfinished for PREEMPT_RT kernel
501 			 * in which irq work is invoked in a per-CPU RT thread.
502 			 * It is also possible for kernel with
503 			 * arch_irq_work_has_interrupt() being false and irq
504 			 * work is invoked in timer interrupt. So waiting for
505 			 * the completion of irq work to ease the handling of
506 			 * concurrency.
507 			 */
508 			irq_work_sync(&c->refill_work);
509 			drain_mem_cache(c);
510 			rcu_in_progress += atomic_read(&c->call_rcu_in_progress);
511 		}
512 		/* objcg is the same across cpus */
513 		if (c->objcg)
514 			obj_cgroup_put(c->objcg);
515 		destroy_mem_alloc(ma, rcu_in_progress);
516 	}
517 	if (ma->caches) {
518 		rcu_in_progress = 0;
519 		for_each_possible_cpu(cpu) {
520 			cc = per_cpu_ptr(ma->caches, cpu);
521 			for (i = 0; i < NUM_CACHES; i++) {
522 				c = &cc->cache[i];
523 				irq_work_sync(&c->refill_work);
524 				drain_mem_cache(c);
525 				rcu_in_progress += atomic_read(&c->call_rcu_in_progress);
526 			}
527 		}
528 		if (c->objcg)
529 			obj_cgroup_put(c->objcg);
530 		destroy_mem_alloc(ma, rcu_in_progress);
531 	}
532 }
533 
534 /* notrace is necessary here and in other functions to make sure
535  * bpf programs cannot attach to them and cause llist corruptions.
536  */
unit_alloc(struct bpf_mem_cache * c)537 static void notrace *unit_alloc(struct bpf_mem_cache *c)
538 {
539 	struct llist_node *llnode = NULL;
540 	unsigned long flags;
541 	int cnt = 0;
542 
543 	/* Disable irqs to prevent the following race for majority of prog types:
544 	 * prog_A
545 	 *   bpf_mem_alloc
546 	 *      preemption or irq -> prog_B
547 	 *        bpf_mem_alloc
548 	 *
549 	 * but prog_B could be a perf_event NMI prog.
550 	 * Use per-cpu 'active' counter to order free_list access between
551 	 * unit_alloc/unit_free/bpf_mem_refill.
552 	 */
553 	local_irq_save(flags);
554 	if (local_inc_return(&c->active) == 1) {
555 		llnode = __llist_del_first(&c->free_llist);
556 		if (llnode)
557 			cnt = --c->free_cnt;
558 	}
559 	local_dec(&c->active);
560 	local_irq_restore(flags);
561 
562 	WARN_ON(cnt < 0);
563 
564 	if (cnt < c->low_watermark)
565 		irq_work_raise(c);
566 	return llnode;
567 }
568 
569 /* Though 'ptr' object could have been allocated on a different cpu
570  * add it to the free_llist of the current cpu.
571  * Let kfree() logic deal with it when it's later called from irq_work.
572  */
unit_free(struct bpf_mem_cache * c,void * ptr)573 static void notrace unit_free(struct bpf_mem_cache *c, void *ptr)
574 {
575 	struct llist_node *llnode = ptr - LLIST_NODE_SZ;
576 	unsigned long flags;
577 	int cnt = 0;
578 
579 	BUILD_BUG_ON(LLIST_NODE_SZ > 8);
580 
581 	local_irq_save(flags);
582 	if (local_inc_return(&c->active) == 1) {
583 		__llist_add(llnode, &c->free_llist);
584 		cnt = ++c->free_cnt;
585 	} else {
586 		/* unit_free() cannot fail. Therefore add an object to atomic
587 		 * llist. free_bulk() will drain it. Though free_llist_extra is
588 		 * a per-cpu list we have to use atomic llist_add here, since
589 		 * it also can be interrupted by bpf nmi prog that does another
590 		 * unit_free() into the same free_llist_extra.
591 		 */
592 		llist_add(llnode, &c->free_llist_extra);
593 	}
594 	local_dec(&c->active);
595 	local_irq_restore(flags);
596 
597 	if (cnt > c->high_watermark)
598 		/* free few objects from current cpu into global kmalloc pool */
599 		irq_work_raise(c);
600 }
601 
602 /* Called from BPF program or from sys_bpf syscall.
603  * In both cases migration is disabled.
604  */
bpf_mem_alloc(struct bpf_mem_alloc * ma,size_t size)605 void notrace *bpf_mem_alloc(struct bpf_mem_alloc *ma, size_t size)
606 {
607 	int idx;
608 	void *ret;
609 
610 	if (!size)
611 		return ZERO_SIZE_PTR;
612 
613 	idx = bpf_mem_cache_idx(size + LLIST_NODE_SZ);
614 	if (idx < 0)
615 		return NULL;
616 
617 	ret = unit_alloc(this_cpu_ptr(ma->caches)->cache + idx);
618 	return !ret ? NULL : ret + LLIST_NODE_SZ;
619 }
620 
bpf_mem_free(struct bpf_mem_alloc * ma,void * ptr)621 void notrace bpf_mem_free(struct bpf_mem_alloc *ma, void *ptr)
622 {
623 	int idx;
624 
625 	if (!ptr)
626 		return;
627 
628 	idx = bpf_mem_cache_idx(ksize(ptr - LLIST_NODE_SZ));
629 	if (idx < 0)
630 		return;
631 
632 	unit_free(this_cpu_ptr(ma->caches)->cache + idx, ptr);
633 }
634 
bpf_mem_cache_alloc(struct bpf_mem_alloc * ma)635 void notrace *bpf_mem_cache_alloc(struct bpf_mem_alloc *ma)
636 {
637 	void *ret;
638 
639 	ret = unit_alloc(this_cpu_ptr(ma->cache));
640 	return !ret ? NULL : ret + LLIST_NODE_SZ;
641 }
642 
bpf_mem_cache_free(struct bpf_mem_alloc * ma,void * ptr)643 void notrace bpf_mem_cache_free(struct bpf_mem_alloc *ma, void *ptr)
644 {
645 	if (!ptr)
646 		return;
647 
648 	unit_free(this_cpu_ptr(ma->cache), ptr);
649 }
650