xref: /openbmc/linux/kernel/bpf/stackmap.c (revision ebd09753)
1 /* Copyright (c) 2016 Facebook
2  *
3  * This program is free software; you can redistribute it and/or
4  * modify it under the terms of version 2 of the GNU General Public
5  * License as published by the Free Software Foundation.
6  */
7 #include <linux/bpf.h>
8 #include <linux/jhash.h>
9 #include <linux/filter.h>
10 #include <linux/stacktrace.h>
11 #include <linux/perf_event.h>
12 #include <linux/elf.h>
13 #include <linux/pagemap.h>
14 #include <linux/irq_work.h>
15 #include "percpu_freelist.h"
16 
17 #define STACK_CREATE_FLAG_MASK					\
18 	(BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY |	\
19 	 BPF_F_STACK_BUILD_ID)
20 
21 struct stack_map_bucket {
22 	struct pcpu_freelist_node fnode;
23 	u32 hash;
24 	u32 nr;
25 	u64 data[];
26 };
27 
28 struct bpf_stack_map {
29 	struct bpf_map map;
30 	void *elems;
31 	struct pcpu_freelist freelist;
32 	u32 n_buckets;
33 	struct stack_map_bucket *buckets[];
34 };
35 
36 /* irq_work to run up_read() for build_id lookup in nmi context */
37 struct stack_map_irq_work {
38 	struct irq_work irq_work;
39 	struct rw_semaphore *sem;
40 };
41 
42 static void do_up_read(struct irq_work *entry)
43 {
44 	struct stack_map_irq_work *work;
45 
46 	work = container_of(entry, struct stack_map_irq_work, irq_work);
47 	up_read(work->sem);
48 	work->sem = NULL;
49 }
50 
51 static DEFINE_PER_CPU(struct stack_map_irq_work, up_read_work);
52 
53 static inline bool stack_map_use_build_id(struct bpf_map *map)
54 {
55 	return (map->map_flags & BPF_F_STACK_BUILD_ID);
56 }
57 
58 static inline int stack_map_data_size(struct bpf_map *map)
59 {
60 	return stack_map_use_build_id(map) ?
61 		sizeof(struct bpf_stack_build_id) : sizeof(u64);
62 }
63 
64 static int prealloc_elems_and_freelist(struct bpf_stack_map *smap)
65 {
66 	u32 elem_size = sizeof(struct stack_map_bucket) + smap->map.value_size;
67 	int err;
68 
69 	smap->elems = bpf_map_area_alloc(elem_size * smap->map.max_entries,
70 					 smap->map.numa_node);
71 	if (!smap->elems)
72 		return -ENOMEM;
73 
74 	err = pcpu_freelist_init(&smap->freelist);
75 	if (err)
76 		goto free_elems;
77 
78 	pcpu_freelist_populate(&smap->freelist, smap->elems, elem_size,
79 			       smap->map.max_entries);
80 	return 0;
81 
82 free_elems:
83 	bpf_map_area_free(smap->elems);
84 	return err;
85 }
86 
87 /* Called from syscall */
88 static struct bpf_map *stack_map_alloc(union bpf_attr *attr)
89 {
90 	u32 value_size = attr->value_size;
91 	struct bpf_stack_map *smap;
92 	u64 cost, n_buckets;
93 	int err;
94 
95 	if (!capable(CAP_SYS_ADMIN))
96 		return ERR_PTR(-EPERM);
97 
98 	if (attr->map_flags & ~STACK_CREATE_FLAG_MASK)
99 		return ERR_PTR(-EINVAL);
100 
101 	/* check sanity of attributes */
102 	if (attr->max_entries == 0 || attr->key_size != 4 ||
103 	    value_size < 8 || value_size % 8)
104 		return ERR_PTR(-EINVAL);
105 
106 	BUILD_BUG_ON(sizeof(struct bpf_stack_build_id) % sizeof(u64));
107 	if (attr->map_flags & BPF_F_STACK_BUILD_ID) {
108 		if (value_size % sizeof(struct bpf_stack_build_id) ||
109 		    value_size / sizeof(struct bpf_stack_build_id)
110 		    > sysctl_perf_event_max_stack)
111 			return ERR_PTR(-EINVAL);
112 	} else if (value_size / 8 > sysctl_perf_event_max_stack)
113 		return ERR_PTR(-EINVAL);
114 
115 	/* hash table size must be power of 2 */
116 	n_buckets = roundup_pow_of_two(attr->max_entries);
117 
118 	cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap);
119 	if (cost >= U32_MAX - PAGE_SIZE)
120 		return ERR_PTR(-E2BIG);
121 
122 	smap = bpf_map_area_alloc(cost, bpf_map_attr_numa_node(attr));
123 	if (!smap)
124 		return ERR_PTR(-ENOMEM);
125 
126 	err = -E2BIG;
127 	cost += n_buckets * (value_size + sizeof(struct stack_map_bucket));
128 	if (cost >= U32_MAX - PAGE_SIZE)
129 		goto free_smap;
130 
131 	bpf_map_init_from_attr(&smap->map, attr);
132 	smap->map.value_size = value_size;
133 	smap->n_buckets = n_buckets;
134 	smap->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
135 
136 	err = bpf_map_precharge_memlock(smap->map.pages);
137 	if (err)
138 		goto free_smap;
139 
140 	err = get_callchain_buffers(sysctl_perf_event_max_stack);
141 	if (err)
142 		goto free_smap;
143 
144 	err = prealloc_elems_and_freelist(smap);
145 	if (err)
146 		goto put_buffers;
147 
148 	return &smap->map;
149 
150 put_buffers:
151 	put_callchain_buffers();
152 free_smap:
153 	bpf_map_area_free(smap);
154 	return ERR_PTR(err);
155 }
156 
157 #define BPF_BUILD_ID 3
158 /*
159  * Parse build id from the note segment. This logic can be shared between
160  * 32-bit and 64-bit system, because Elf32_Nhdr and Elf64_Nhdr are
161  * identical.
162  */
163 static inline int stack_map_parse_build_id(void *page_addr,
164 					   unsigned char *build_id,
165 					   void *note_start,
166 					   Elf32_Word note_size)
167 {
168 	Elf32_Word note_offs = 0, new_offs;
169 
170 	/* check for overflow */
171 	if (note_start < page_addr || note_start + note_size < note_start)
172 		return -EINVAL;
173 
174 	/* only supports note that fits in the first page */
175 	if (note_start + note_size > page_addr + PAGE_SIZE)
176 		return -EINVAL;
177 
178 	while (note_offs + sizeof(Elf32_Nhdr) < note_size) {
179 		Elf32_Nhdr *nhdr = (Elf32_Nhdr *)(note_start + note_offs);
180 
181 		if (nhdr->n_type == BPF_BUILD_ID &&
182 		    nhdr->n_namesz == sizeof("GNU") &&
183 		    nhdr->n_descsz == BPF_BUILD_ID_SIZE) {
184 			memcpy(build_id,
185 			       note_start + note_offs +
186 			       ALIGN(sizeof("GNU"), 4) + sizeof(Elf32_Nhdr),
187 			       BPF_BUILD_ID_SIZE);
188 			return 0;
189 		}
190 		new_offs = note_offs + sizeof(Elf32_Nhdr) +
191 			ALIGN(nhdr->n_namesz, 4) + ALIGN(nhdr->n_descsz, 4);
192 		if (new_offs <= note_offs)  /* overflow */
193 			break;
194 		note_offs = new_offs;
195 	}
196 	return -EINVAL;
197 }
198 
199 /* Parse build ID from 32-bit ELF */
200 static int stack_map_get_build_id_32(void *page_addr,
201 				     unsigned char *build_id)
202 {
203 	Elf32_Ehdr *ehdr = (Elf32_Ehdr *)page_addr;
204 	Elf32_Phdr *phdr;
205 	int i;
206 
207 	/* only supports phdr that fits in one page */
208 	if (ehdr->e_phnum >
209 	    (PAGE_SIZE - sizeof(Elf32_Ehdr)) / sizeof(Elf32_Phdr))
210 		return -EINVAL;
211 
212 	phdr = (Elf32_Phdr *)(page_addr + sizeof(Elf32_Ehdr));
213 
214 	for (i = 0; i < ehdr->e_phnum; ++i)
215 		if (phdr[i].p_type == PT_NOTE)
216 			return stack_map_parse_build_id(page_addr, build_id,
217 					page_addr + phdr[i].p_offset,
218 					phdr[i].p_filesz);
219 	return -EINVAL;
220 }
221 
222 /* Parse build ID from 64-bit ELF */
223 static int stack_map_get_build_id_64(void *page_addr,
224 				     unsigned char *build_id)
225 {
226 	Elf64_Ehdr *ehdr = (Elf64_Ehdr *)page_addr;
227 	Elf64_Phdr *phdr;
228 	int i;
229 
230 	/* only supports phdr that fits in one page */
231 	if (ehdr->e_phnum >
232 	    (PAGE_SIZE - sizeof(Elf64_Ehdr)) / sizeof(Elf64_Phdr))
233 		return -EINVAL;
234 
235 	phdr = (Elf64_Phdr *)(page_addr + sizeof(Elf64_Ehdr));
236 
237 	for (i = 0; i < ehdr->e_phnum; ++i)
238 		if (phdr[i].p_type == PT_NOTE)
239 			return stack_map_parse_build_id(page_addr, build_id,
240 					page_addr + phdr[i].p_offset,
241 					phdr[i].p_filesz);
242 	return -EINVAL;
243 }
244 
245 /* Parse build ID of ELF file mapped to vma */
246 static int stack_map_get_build_id(struct vm_area_struct *vma,
247 				  unsigned char *build_id)
248 {
249 	Elf32_Ehdr *ehdr;
250 	struct page *page;
251 	void *page_addr;
252 	int ret;
253 
254 	/* only works for page backed storage  */
255 	if (!vma->vm_file)
256 		return -EINVAL;
257 
258 	page = find_get_page(vma->vm_file->f_mapping, 0);
259 	if (!page)
260 		return -EFAULT;	/* page not mapped */
261 
262 	ret = -EINVAL;
263 	page_addr = page_address(page);
264 	ehdr = (Elf32_Ehdr *)page_addr;
265 
266 	/* compare magic x7f "ELF" */
267 	if (memcmp(ehdr->e_ident, ELFMAG, SELFMAG) != 0)
268 		goto out;
269 
270 	/* only support executable file and shared object file */
271 	if (ehdr->e_type != ET_EXEC && ehdr->e_type != ET_DYN)
272 		goto out;
273 
274 	if (ehdr->e_ident[EI_CLASS] == ELFCLASS32)
275 		ret = stack_map_get_build_id_32(page_addr, build_id);
276 	else if (ehdr->e_ident[EI_CLASS] == ELFCLASS64)
277 		ret = stack_map_get_build_id_64(page_addr, build_id);
278 out:
279 	put_page(page);
280 	return ret;
281 }
282 
283 static void stack_map_get_build_id_offset(struct bpf_stack_build_id *id_offs,
284 					  u64 *ips, u32 trace_nr, bool user)
285 {
286 	int i;
287 	struct vm_area_struct *vma;
288 	bool irq_work_busy = false;
289 	struct stack_map_irq_work *work = NULL;
290 
291 	if (in_nmi()) {
292 		work = this_cpu_ptr(&up_read_work);
293 		if (work->irq_work.flags & IRQ_WORK_BUSY)
294 			/* cannot queue more up_read, fallback */
295 			irq_work_busy = true;
296 	}
297 
298 	/*
299 	 * We cannot do up_read() in nmi context. To do build_id lookup
300 	 * in nmi context, we need to run up_read() in irq_work. We use
301 	 * a percpu variable to do the irq_work. If the irq_work is
302 	 * already used by another lookup, we fall back to report ips.
303 	 *
304 	 * Same fallback is used for kernel stack (!user) on a stackmap
305 	 * with build_id.
306 	 */
307 	if (!user || !current || !current->mm || irq_work_busy ||
308 	    down_read_trylock(&current->mm->mmap_sem) == 0) {
309 		/* cannot access current->mm, fall back to ips */
310 		for (i = 0; i < trace_nr; i++) {
311 			id_offs[i].status = BPF_STACK_BUILD_ID_IP;
312 			id_offs[i].ip = ips[i];
313 		}
314 		return;
315 	}
316 
317 	for (i = 0; i < trace_nr; i++) {
318 		vma = find_vma(current->mm, ips[i]);
319 		if (!vma || stack_map_get_build_id(vma, id_offs[i].build_id)) {
320 			/* per entry fall back to ips */
321 			id_offs[i].status = BPF_STACK_BUILD_ID_IP;
322 			id_offs[i].ip = ips[i];
323 			continue;
324 		}
325 		id_offs[i].offset = (vma->vm_pgoff << PAGE_SHIFT) + ips[i]
326 			- vma->vm_start;
327 		id_offs[i].status = BPF_STACK_BUILD_ID_VALID;
328 	}
329 
330 	if (!work) {
331 		up_read(&current->mm->mmap_sem);
332 	} else {
333 		work->sem = &current->mm->mmap_sem;
334 		irq_work_queue(&work->irq_work);
335 	}
336 }
337 
338 BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
339 	   u64, flags)
340 {
341 	struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
342 	struct perf_callchain_entry *trace;
343 	struct stack_map_bucket *bucket, *new_bucket, *old_bucket;
344 	u32 max_depth = map->value_size / stack_map_data_size(map);
345 	/* stack_map_alloc() checks that max_depth <= sysctl_perf_event_max_stack */
346 	u32 init_nr = sysctl_perf_event_max_stack - max_depth;
347 	u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
348 	u32 hash, id, trace_nr, trace_len;
349 	bool user = flags & BPF_F_USER_STACK;
350 	bool kernel = !user;
351 	u64 *ips;
352 	bool hash_matches;
353 
354 	if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
355 			       BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
356 		return -EINVAL;
357 
358 	trace = get_perf_callchain(regs, init_nr, kernel, user,
359 				   sysctl_perf_event_max_stack, false, false);
360 
361 	if (unlikely(!trace))
362 		/* couldn't fetch the stack trace */
363 		return -EFAULT;
364 
365 	/* get_perf_callchain() guarantees that trace->nr >= init_nr
366 	 * and trace-nr <= sysctl_perf_event_max_stack, so trace_nr <= max_depth
367 	 */
368 	trace_nr = trace->nr - init_nr;
369 
370 	if (trace_nr <= skip)
371 		/* skipping more than usable stack trace */
372 		return -EFAULT;
373 
374 	trace_nr -= skip;
375 	trace_len = trace_nr * sizeof(u64);
376 	ips = trace->ip + skip + init_nr;
377 	hash = jhash2((u32 *)ips, trace_len / sizeof(u32), 0);
378 	id = hash & (smap->n_buckets - 1);
379 	bucket = READ_ONCE(smap->buckets[id]);
380 
381 	hash_matches = bucket && bucket->hash == hash;
382 	/* fast cmp */
383 	if (hash_matches && flags & BPF_F_FAST_STACK_CMP)
384 		return id;
385 
386 	if (stack_map_use_build_id(map)) {
387 		/* for build_id+offset, pop a bucket before slow cmp */
388 		new_bucket = (struct stack_map_bucket *)
389 			pcpu_freelist_pop(&smap->freelist);
390 		if (unlikely(!new_bucket))
391 			return -ENOMEM;
392 		new_bucket->nr = trace_nr;
393 		stack_map_get_build_id_offset(
394 			(struct bpf_stack_build_id *)new_bucket->data,
395 			ips, trace_nr, user);
396 		trace_len = trace_nr * sizeof(struct bpf_stack_build_id);
397 		if (hash_matches && bucket->nr == trace_nr &&
398 		    memcmp(bucket->data, new_bucket->data, trace_len) == 0) {
399 			pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
400 			return id;
401 		}
402 		if (bucket && !(flags & BPF_F_REUSE_STACKID)) {
403 			pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
404 			return -EEXIST;
405 		}
406 	} else {
407 		if (hash_matches && bucket->nr == trace_nr &&
408 		    memcmp(bucket->data, ips, trace_len) == 0)
409 			return id;
410 		if (bucket && !(flags & BPF_F_REUSE_STACKID))
411 			return -EEXIST;
412 
413 		new_bucket = (struct stack_map_bucket *)
414 			pcpu_freelist_pop(&smap->freelist);
415 		if (unlikely(!new_bucket))
416 			return -ENOMEM;
417 		memcpy(new_bucket->data, ips, trace_len);
418 	}
419 
420 	new_bucket->hash = hash;
421 	new_bucket->nr = trace_nr;
422 
423 	old_bucket = xchg(&smap->buckets[id], new_bucket);
424 	if (old_bucket)
425 		pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
426 	return id;
427 }
428 
429 const struct bpf_func_proto bpf_get_stackid_proto = {
430 	.func		= bpf_get_stackid,
431 	.gpl_only	= true,
432 	.ret_type	= RET_INTEGER,
433 	.arg1_type	= ARG_PTR_TO_CTX,
434 	.arg2_type	= ARG_CONST_MAP_PTR,
435 	.arg3_type	= ARG_ANYTHING,
436 };
437 
438 BPF_CALL_4(bpf_get_stack, struct pt_regs *, regs, void *, buf, u32, size,
439 	   u64, flags)
440 {
441 	u32 init_nr, trace_nr, copy_len, elem_size, num_elem;
442 	bool user_build_id = flags & BPF_F_USER_BUILD_ID;
443 	u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
444 	bool user = flags & BPF_F_USER_STACK;
445 	struct perf_callchain_entry *trace;
446 	bool kernel = !user;
447 	int err = -EINVAL;
448 	u64 *ips;
449 
450 	if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
451 			       BPF_F_USER_BUILD_ID)))
452 		goto clear;
453 	if (kernel && user_build_id)
454 		goto clear;
455 
456 	elem_size = (user && user_build_id) ? sizeof(struct bpf_stack_build_id)
457 					    : sizeof(u64);
458 	if (unlikely(size % elem_size))
459 		goto clear;
460 
461 	num_elem = size / elem_size;
462 	if (sysctl_perf_event_max_stack < num_elem)
463 		init_nr = 0;
464 	else
465 		init_nr = sysctl_perf_event_max_stack - num_elem;
466 	trace = get_perf_callchain(regs, init_nr, kernel, user,
467 				   sysctl_perf_event_max_stack, false, false);
468 	if (unlikely(!trace))
469 		goto err_fault;
470 
471 	trace_nr = trace->nr - init_nr;
472 	if (trace_nr < skip)
473 		goto err_fault;
474 
475 	trace_nr -= skip;
476 	trace_nr = (trace_nr <= num_elem) ? trace_nr : num_elem;
477 	copy_len = trace_nr * elem_size;
478 	ips = trace->ip + skip + init_nr;
479 	if (user && user_build_id)
480 		stack_map_get_build_id_offset(buf, ips, trace_nr, user);
481 	else
482 		memcpy(buf, ips, copy_len);
483 
484 	if (size > copy_len)
485 		memset(buf + copy_len, 0, size - copy_len);
486 	return copy_len;
487 
488 err_fault:
489 	err = -EFAULT;
490 clear:
491 	memset(buf, 0, size);
492 	return err;
493 }
494 
495 const struct bpf_func_proto bpf_get_stack_proto = {
496 	.func		= bpf_get_stack,
497 	.gpl_only	= true,
498 	.ret_type	= RET_INTEGER,
499 	.arg1_type	= ARG_PTR_TO_CTX,
500 	.arg2_type	= ARG_PTR_TO_UNINIT_MEM,
501 	.arg3_type	= ARG_CONST_SIZE_OR_ZERO,
502 	.arg4_type	= ARG_ANYTHING,
503 };
504 
505 /* Called from eBPF program */
506 static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
507 {
508 	return ERR_PTR(-EOPNOTSUPP);
509 }
510 
511 /* Called from syscall */
512 int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value)
513 {
514 	struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
515 	struct stack_map_bucket *bucket, *old_bucket;
516 	u32 id = *(u32 *)key, trace_len;
517 
518 	if (unlikely(id >= smap->n_buckets))
519 		return -ENOENT;
520 
521 	bucket = xchg(&smap->buckets[id], NULL);
522 	if (!bucket)
523 		return -ENOENT;
524 
525 	trace_len = bucket->nr * stack_map_data_size(map);
526 	memcpy(value, bucket->data, trace_len);
527 	memset(value + trace_len, 0, map->value_size - trace_len);
528 
529 	old_bucket = xchg(&smap->buckets[id], bucket);
530 	if (old_bucket)
531 		pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
532 	return 0;
533 }
534 
535 static int stack_map_get_next_key(struct bpf_map *map, void *key,
536 				  void *next_key)
537 {
538 	struct bpf_stack_map *smap = container_of(map,
539 						  struct bpf_stack_map, map);
540 	u32 id;
541 
542 	WARN_ON_ONCE(!rcu_read_lock_held());
543 
544 	if (!key) {
545 		id = 0;
546 	} else {
547 		id = *(u32 *)key;
548 		if (id >= smap->n_buckets || !smap->buckets[id])
549 			id = 0;
550 		else
551 			id++;
552 	}
553 
554 	while (id < smap->n_buckets && !smap->buckets[id])
555 		id++;
556 
557 	if (id >= smap->n_buckets)
558 		return -ENOENT;
559 
560 	*(u32 *)next_key = id;
561 	return 0;
562 }
563 
564 static int stack_map_update_elem(struct bpf_map *map, void *key, void *value,
565 				 u64 map_flags)
566 {
567 	return -EINVAL;
568 }
569 
570 /* Called from syscall or from eBPF program */
571 static int stack_map_delete_elem(struct bpf_map *map, void *key)
572 {
573 	struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
574 	struct stack_map_bucket *old_bucket;
575 	u32 id = *(u32 *)key;
576 
577 	if (unlikely(id >= smap->n_buckets))
578 		return -E2BIG;
579 
580 	old_bucket = xchg(&smap->buckets[id], NULL);
581 	if (old_bucket) {
582 		pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
583 		return 0;
584 	} else {
585 		return -ENOENT;
586 	}
587 }
588 
589 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
590 static void stack_map_free(struct bpf_map *map)
591 {
592 	struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
593 
594 	/* wait for bpf programs to complete before freeing stack map */
595 	synchronize_rcu();
596 
597 	bpf_map_area_free(smap->elems);
598 	pcpu_freelist_destroy(&smap->freelist);
599 	bpf_map_area_free(smap);
600 	put_callchain_buffers();
601 }
602 
603 const struct bpf_map_ops stack_trace_map_ops = {
604 	.map_alloc = stack_map_alloc,
605 	.map_free = stack_map_free,
606 	.map_get_next_key = stack_map_get_next_key,
607 	.map_lookup_elem = stack_map_lookup_elem,
608 	.map_update_elem = stack_map_update_elem,
609 	.map_delete_elem = stack_map_delete_elem,
610 	.map_check_btf = map_check_no_btf,
611 };
612 
613 static int __init stack_map_init(void)
614 {
615 	int cpu;
616 	struct stack_map_irq_work *work;
617 
618 	for_each_possible_cpu(cpu) {
619 		work = per_cpu_ptr(&up_read_work, cpu);
620 		init_irq_work(&work->irq_work, do_up_read);
621 	}
622 	return 0;
623 }
624 subsys_initcall(stack_map_init);
625