• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2016 Facebook
3  */
4 #include <linux/bpf.h>
5 #include <linux/jhash.h>
6 #include <linux/filter.h>
7 #include <linux/kernel.h>
8 #include <linux/stacktrace.h>
9 #include <linux/perf_event.h>
10 #include <linux/elf.h>
11 #include <linux/pagemap.h>
12 #include <linux/irq_work.h>
13 #include <linux/btf_ids.h>
14 #include "percpu_freelist.h"
15 
16 #define STACK_CREATE_FLAG_MASK					\
17 	(BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY |	\
18 	 BPF_F_STACK_BUILD_ID)
19 
20 struct stack_map_bucket {
21 	struct pcpu_freelist_node fnode;
22 	u32 hash;
23 	u32 nr;
24 	u64 data[];
25 };
26 
27 struct bpf_stack_map {
28 	struct bpf_map map;
29 	void *elems;
30 	struct pcpu_freelist freelist;
31 	u32 n_buckets;
32 	struct stack_map_bucket *buckets[];
33 };
34 
35 /* irq_work to run up_read() for build_id lookup in nmi context */
36 struct stack_map_irq_work {
37 	struct irq_work irq_work;
38 	struct mm_struct *mm;
39 };
40 
do_up_read(struct irq_work * entry)41 static void do_up_read(struct irq_work *entry)
42 {
43 	struct stack_map_irq_work *work;
44 
45 	if (WARN_ON_ONCE(IS_ENABLED(CONFIG_PREEMPT_RT)))
46 		return;
47 
48 	work = container_of(entry, struct stack_map_irq_work, irq_work);
49 	mmap_read_unlock_non_owner(work->mm);
50 }
51 
52 static DEFINE_PER_CPU(struct stack_map_irq_work, up_read_work);
53 
stack_map_use_build_id(struct bpf_map * map)54 static inline bool stack_map_use_build_id(struct bpf_map *map)
55 {
56 	return (map->map_flags & BPF_F_STACK_BUILD_ID);
57 }
58 
stack_map_data_size(struct bpf_map * map)59 static inline int stack_map_data_size(struct bpf_map *map)
60 {
61 	return stack_map_use_build_id(map) ?
62 		sizeof(struct bpf_stack_build_id) : sizeof(u64);
63 }
64 
prealloc_elems_and_freelist(struct bpf_stack_map * smap)65 static int prealloc_elems_and_freelist(struct bpf_stack_map *smap)
66 {
67 	u64 elem_size = sizeof(struct stack_map_bucket) +
68 			(u64)smap->map.value_size;
69 	int err;
70 
71 	smap->elems = bpf_map_area_alloc(elem_size * smap->map.max_entries,
72 					 smap->map.numa_node);
73 	if (!smap->elems)
74 		return -ENOMEM;
75 
76 	err = pcpu_freelist_init(&smap->freelist);
77 	if (err)
78 		goto free_elems;
79 
80 	pcpu_freelist_populate(&smap->freelist, smap->elems, elem_size,
81 			       smap->map.max_entries);
82 	return 0;
83 
84 free_elems:
85 	bpf_map_area_free(smap->elems);
86 	return err;
87 }
88 
89 /* Called from syscall */
stack_map_alloc(union bpf_attr * attr)90 static struct bpf_map *stack_map_alloc(union bpf_attr *attr)
91 {
92 	u32 value_size = attr->value_size;
93 	struct bpf_stack_map *smap;
94 	struct bpf_map_memory mem;
95 	u64 cost, n_buckets;
96 	int err;
97 
98 	if (!bpf_capable())
99 		return ERR_PTR(-EPERM);
100 
101 	if (attr->map_flags & ~STACK_CREATE_FLAG_MASK)
102 		return ERR_PTR(-EINVAL);
103 
104 	/* check sanity of attributes */
105 	if (attr->max_entries == 0 || attr->key_size != 4 ||
106 	    value_size < 8 || value_size % 8)
107 		return ERR_PTR(-EINVAL);
108 
109 	BUILD_BUG_ON(sizeof(struct bpf_stack_build_id) % sizeof(u64));
110 	if (attr->map_flags & BPF_F_STACK_BUILD_ID) {
111 		if (value_size % sizeof(struct bpf_stack_build_id) ||
112 		    value_size / sizeof(struct bpf_stack_build_id)
113 		    > sysctl_perf_event_max_stack)
114 			return ERR_PTR(-EINVAL);
115 	} else if (value_size / 8 > sysctl_perf_event_max_stack)
116 		return ERR_PTR(-EINVAL);
117 
118 	/* hash table size must be power of 2 */
119 	n_buckets = roundup_pow_of_two(attr->max_entries);
120 	if (!n_buckets)
121 		return ERR_PTR(-E2BIG);
122 
123 	cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap);
124 	cost += n_buckets * (value_size + sizeof(struct stack_map_bucket));
125 	err = bpf_map_charge_init(&mem, cost);
126 	if (err)
127 		return ERR_PTR(err);
128 
129 	smap = bpf_map_area_alloc(cost, bpf_map_attr_numa_node(attr));
130 	if (!smap) {
131 		bpf_map_charge_finish(&mem);
132 		return ERR_PTR(-ENOMEM);
133 	}
134 
135 	bpf_map_init_from_attr(&smap->map, attr);
136 	smap->map.value_size = value_size;
137 	smap->n_buckets = n_buckets;
138 
139 	err = get_callchain_buffers(sysctl_perf_event_max_stack);
140 	if (err)
141 		goto free_charge;
142 
143 	err = prealloc_elems_and_freelist(smap);
144 	if (err)
145 		goto put_buffers;
146 
147 	bpf_map_charge_move(&smap->map.memory, &mem);
148 
149 	return &smap->map;
150 
151 put_buffers:
152 	put_callchain_buffers();
153 free_charge:
154 	bpf_map_charge_finish(&mem);
155 	bpf_map_area_free(smap);
156 	return ERR_PTR(err);
157 }
158 
159 #define BPF_BUILD_ID 3
160 /*
161  * Parse build id from the note segment. This logic can be shared between
162  * 32-bit and 64-bit system, because Elf32_Nhdr and Elf64_Nhdr are
163  * identical.
164  */
stack_map_parse_build_id(void * page_addr,unsigned char * build_id,void * note_start,Elf32_Word note_size)165 static inline int stack_map_parse_build_id(void *page_addr,
166 					   unsigned char *build_id,
167 					   void *note_start,
168 					   Elf32_Word note_size)
169 {
170 	Elf32_Word note_offs = 0, new_offs;
171 
172 	/* check for overflow */
173 	if (note_start < page_addr || note_start + note_size < note_start)
174 		return -EINVAL;
175 
176 	/* only supports note that fits in the first page */
177 	if (note_start + note_size > page_addr + PAGE_SIZE)
178 		return -EINVAL;
179 
180 	while (note_offs + sizeof(Elf32_Nhdr) < note_size) {
181 		Elf32_Nhdr *nhdr = (Elf32_Nhdr *)(note_start + note_offs);
182 
183 		if (nhdr->n_type == BPF_BUILD_ID &&
184 		    nhdr->n_namesz == sizeof("GNU") &&
185 		    nhdr->n_descsz > 0 &&
186 		    nhdr->n_descsz <= BPF_BUILD_ID_SIZE) {
187 			memcpy(build_id,
188 			       note_start + note_offs +
189 			       ALIGN(sizeof("GNU"), 4) + sizeof(Elf32_Nhdr),
190 			       nhdr->n_descsz);
191 			memset(build_id + nhdr->n_descsz, 0,
192 			       BPF_BUILD_ID_SIZE - nhdr->n_descsz);
193 			return 0;
194 		}
195 		new_offs = note_offs + sizeof(Elf32_Nhdr) +
196 			ALIGN(nhdr->n_namesz, 4) + ALIGN(nhdr->n_descsz, 4);
197 		if (new_offs <= note_offs)  /* overflow */
198 			break;
199 		note_offs = new_offs;
200 	}
201 	return -EINVAL;
202 }
203 
204 /* Parse build ID from 32-bit ELF */
stack_map_get_build_id_32(void * page_addr,unsigned char * build_id)205 static int stack_map_get_build_id_32(void *page_addr,
206 				     unsigned char *build_id)
207 {
208 	Elf32_Ehdr *ehdr = (Elf32_Ehdr *)page_addr;
209 	Elf32_Phdr *phdr;
210 	int i;
211 
212 	/* only supports phdr that fits in one page */
213 	if (ehdr->e_phnum >
214 	    (PAGE_SIZE - sizeof(Elf32_Ehdr)) / sizeof(Elf32_Phdr))
215 		return -EINVAL;
216 
217 	phdr = (Elf32_Phdr *)(page_addr + sizeof(Elf32_Ehdr));
218 
219 	for (i = 0; i < ehdr->e_phnum; ++i) {
220 		if (phdr[i].p_type == PT_NOTE &&
221 		    !stack_map_parse_build_id(page_addr, build_id,
222 					      page_addr + phdr[i].p_offset,
223 					      phdr[i].p_filesz))
224 			return 0;
225 	}
226 	return -EINVAL;
227 }
228 
229 /* Parse build ID from 64-bit ELF */
stack_map_get_build_id_64(void * page_addr,unsigned char * build_id)230 static int stack_map_get_build_id_64(void *page_addr,
231 				     unsigned char *build_id)
232 {
233 	Elf64_Ehdr *ehdr = (Elf64_Ehdr *)page_addr;
234 	Elf64_Phdr *phdr;
235 	int i;
236 
237 	/* only supports phdr that fits in one page */
238 	if (ehdr->e_phnum >
239 	    (PAGE_SIZE - sizeof(Elf64_Ehdr)) / sizeof(Elf64_Phdr))
240 		return -EINVAL;
241 
242 	phdr = (Elf64_Phdr *)(page_addr + sizeof(Elf64_Ehdr));
243 
244 	for (i = 0; i < ehdr->e_phnum; ++i) {
245 		if (phdr[i].p_type == PT_NOTE &&
246 		    !stack_map_parse_build_id(page_addr, build_id,
247 					      page_addr + phdr[i].p_offset,
248 					      phdr[i].p_filesz))
249 			return 0;
250 	}
251 	return -EINVAL;
252 }
253 
254 /* Parse build ID of ELF file mapped to vma */
stack_map_get_build_id(struct vm_area_struct * vma,unsigned char * build_id)255 static int stack_map_get_build_id(struct vm_area_struct *vma,
256 				  unsigned char *build_id)
257 {
258 	Elf32_Ehdr *ehdr;
259 	struct page *page;
260 	void *page_addr;
261 	int ret;
262 
263 	/* only works for page backed storage  */
264 	if (!vma->vm_file)
265 		return -EINVAL;
266 
267 	page = find_get_page(vma->vm_file->f_mapping, 0);
268 	if (!page)
269 		return -EFAULT;	/* page not mapped */
270 
271 	ret = -EINVAL;
272 	page_addr = kmap_atomic(page);
273 	ehdr = (Elf32_Ehdr *)page_addr;
274 
275 	/* compare magic x7f "ELF" */
276 	if (memcmp(ehdr->e_ident, ELFMAG, SELFMAG) != 0)
277 		goto out;
278 
279 	/* only support executable file and shared object file */
280 	if (ehdr->e_type != ET_EXEC && ehdr->e_type != ET_DYN)
281 		goto out;
282 
283 	if (ehdr->e_ident[EI_CLASS] == ELFCLASS32)
284 		ret = stack_map_get_build_id_32(page_addr, build_id);
285 	else if (ehdr->e_ident[EI_CLASS] == ELFCLASS64)
286 		ret = stack_map_get_build_id_64(page_addr, build_id);
287 out:
288 	kunmap_atomic(page_addr);
289 	put_page(page);
290 	return ret;
291 }
292 
stack_map_get_build_id_offset(struct bpf_stack_build_id * id_offs,u64 * ips,u32 trace_nr,bool user)293 static void stack_map_get_build_id_offset(struct bpf_stack_build_id *id_offs,
294 					  u64 *ips, u32 trace_nr, bool user)
295 {
296 	int i;
297 	struct vm_area_struct *vma;
298 	bool irq_work_busy = false;
299 	struct stack_map_irq_work *work = NULL;
300 
301 	if (irqs_disabled()) {
302 		if (!IS_ENABLED(CONFIG_PREEMPT_RT)) {
303 			work = this_cpu_ptr(&up_read_work);
304 			if (atomic_read(&work->irq_work.flags) & IRQ_WORK_BUSY) {
305 				/* cannot queue more up_read, fallback */
306 				irq_work_busy = true;
307 			}
308 		} else {
309 			/*
310 			 * PREEMPT_RT does not allow to trylock mmap sem in
311 			 * interrupt disabled context. Force the fallback code.
312 			 */
313 			irq_work_busy = true;
314 		}
315 	}
316 
317 	/*
318 	 * We cannot do up_read() when the irq is disabled, because of
319 	 * risk to deadlock with rq_lock. To do build_id lookup when the
320 	 * irqs are disabled, we need to run up_read() in irq_work. We use
321 	 * a percpu variable to do the irq_work. If the irq_work is
322 	 * already used by another lookup, we fall back to report ips.
323 	 *
324 	 * Same fallback is used for kernel stack (!user) on a stackmap
325 	 * with build_id.
326 	 */
327 	if (!user || !current || !current->mm || irq_work_busy ||
328 	    !mmap_read_trylock_non_owner(current->mm)) {
329 		/* cannot access current->mm, fall back to ips */
330 		for (i = 0; i < trace_nr; i++) {
331 			id_offs[i].status = BPF_STACK_BUILD_ID_IP;
332 			id_offs[i].ip = ips[i];
333 			memset(id_offs[i].build_id, 0, BPF_BUILD_ID_SIZE);
334 		}
335 		return;
336 	}
337 
338 	for (i = 0; i < trace_nr; i++) {
339 		vma = find_vma(current->mm, ips[i]);
340 		if (!vma || stack_map_get_build_id(vma, id_offs[i].build_id)) {
341 			/* per entry fall back to ips */
342 			id_offs[i].status = BPF_STACK_BUILD_ID_IP;
343 			id_offs[i].ip = ips[i];
344 			memset(id_offs[i].build_id, 0, BPF_BUILD_ID_SIZE);
345 			continue;
346 		}
347 		id_offs[i].offset = (vma->vm_pgoff << PAGE_SHIFT) + ips[i]
348 			- vma->vm_start;
349 		id_offs[i].status = BPF_STACK_BUILD_ID_VALID;
350 	}
351 
352 	if (!work) {
353 		mmap_read_unlock_non_owner(current->mm);
354 	} else {
355 		work->mm = current->mm;
356 		irq_work_queue(&work->irq_work);
357 	}
358 }
359 
360 static struct perf_callchain_entry *
get_callchain_entry_for_task(struct task_struct * task,u32 init_nr)361 get_callchain_entry_for_task(struct task_struct *task, u32 init_nr)
362 {
363 #ifdef CONFIG_STACKTRACE
364 	struct perf_callchain_entry *entry;
365 	int rctx;
366 
367 	entry = get_callchain_entry(&rctx);
368 
369 	if (!entry)
370 		return NULL;
371 
372 	entry->nr = init_nr +
373 		stack_trace_save_tsk(task, (unsigned long *)(entry->ip + init_nr),
374 				     sysctl_perf_event_max_stack - init_nr, 0);
375 
376 	/* stack_trace_save_tsk() works on unsigned long array, while
377 	 * perf_callchain_entry uses u64 array. For 32-bit systems, it is
378 	 * necessary to fix this mismatch.
379 	 */
380 	if (__BITS_PER_LONG != 64) {
381 		unsigned long *from = (unsigned long *) entry->ip;
382 		u64 *to = entry->ip;
383 		int i;
384 
385 		/* copy data from the end to avoid using extra buffer */
386 		for (i = entry->nr - 1; i >= (int)init_nr; i--)
387 			to[i] = (u64)(from[i]);
388 	}
389 
390 	put_callchain_entry(rctx);
391 
392 	return entry;
393 #else /* CONFIG_STACKTRACE */
394 	return NULL;
395 #endif
396 }
397 
__bpf_get_stackid(struct bpf_map * map,struct perf_callchain_entry * trace,u64 flags)398 static long __bpf_get_stackid(struct bpf_map *map,
399 			      struct perf_callchain_entry *trace, u64 flags)
400 {
401 	struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
402 	struct stack_map_bucket *bucket, *new_bucket, *old_bucket;
403 	u32 max_depth = map->value_size / stack_map_data_size(map);
404 	/* stack_map_alloc() checks that max_depth <= sysctl_perf_event_max_stack */
405 	u32 init_nr = sysctl_perf_event_max_stack - max_depth;
406 	u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
407 	u32 hash, id, trace_nr, trace_len;
408 	bool user = flags & BPF_F_USER_STACK;
409 	u64 *ips;
410 	bool hash_matches;
411 
412 	/* get_perf_callchain() guarantees that trace->nr >= init_nr
413 	 * and trace-nr <= sysctl_perf_event_max_stack, so trace_nr <= max_depth
414 	 */
415 	trace_nr = trace->nr - init_nr;
416 
417 	if (trace_nr <= skip)
418 		/* skipping more than usable stack trace */
419 		return -EFAULT;
420 
421 	trace_nr -= skip;
422 	trace_len = trace_nr * sizeof(u64);
423 	ips = trace->ip + skip + init_nr;
424 	hash = jhash2((u32 *)ips, trace_len / sizeof(u32), 0);
425 	id = hash & (smap->n_buckets - 1);
426 	bucket = READ_ONCE(smap->buckets[id]);
427 
428 	hash_matches = bucket && bucket->hash == hash;
429 	/* fast cmp */
430 	if (hash_matches && flags & BPF_F_FAST_STACK_CMP)
431 		return id;
432 
433 	if (stack_map_use_build_id(map)) {
434 		/* for build_id+offset, pop a bucket before slow cmp */
435 		new_bucket = (struct stack_map_bucket *)
436 			pcpu_freelist_pop(&smap->freelist);
437 		if (unlikely(!new_bucket))
438 			return -ENOMEM;
439 		new_bucket->nr = trace_nr;
440 		stack_map_get_build_id_offset(
441 			(struct bpf_stack_build_id *)new_bucket->data,
442 			ips, trace_nr, user);
443 		trace_len = trace_nr * sizeof(struct bpf_stack_build_id);
444 		if (hash_matches && bucket->nr == trace_nr &&
445 		    memcmp(bucket->data, new_bucket->data, trace_len) == 0) {
446 			pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
447 			return id;
448 		}
449 		if (bucket && !(flags & BPF_F_REUSE_STACKID)) {
450 			pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
451 			return -EEXIST;
452 		}
453 	} else {
454 		if (hash_matches && bucket->nr == trace_nr &&
455 		    memcmp(bucket->data, ips, trace_len) == 0)
456 			return id;
457 		if (bucket && !(flags & BPF_F_REUSE_STACKID))
458 			return -EEXIST;
459 
460 		new_bucket = (struct stack_map_bucket *)
461 			pcpu_freelist_pop(&smap->freelist);
462 		if (unlikely(!new_bucket))
463 			return -ENOMEM;
464 		memcpy(new_bucket->data, ips, trace_len);
465 	}
466 
467 	new_bucket->hash = hash;
468 	new_bucket->nr = trace_nr;
469 
470 	old_bucket = xchg(&smap->buckets[id], new_bucket);
471 	if (old_bucket)
472 		pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
473 	return id;
474 }
475 
BPF_CALL_3(bpf_get_stackid,struct pt_regs *,regs,struct bpf_map *,map,u64,flags)476 BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
477 	   u64, flags)
478 {
479 	u32 max_depth = map->value_size / stack_map_data_size(map);
480 	/* stack_map_alloc() checks that max_depth <= sysctl_perf_event_max_stack */
481 	u32 init_nr = sysctl_perf_event_max_stack - max_depth;
482 	bool user = flags & BPF_F_USER_STACK;
483 	struct perf_callchain_entry *trace;
484 	bool kernel = !user;
485 
486 	if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
487 			       BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
488 		return -EINVAL;
489 
490 	trace = get_perf_callchain(regs, init_nr, kernel, user,
491 				   sysctl_perf_event_max_stack, false, false);
492 
493 	if (unlikely(!trace))
494 		/* couldn't fetch the stack trace */
495 		return -EFAULT;
496 
497 	return __bpf_get_stackid(map, trace, flags);
498 }
499 
500 const struct bpf_func_proto bpf_get_stackid_proto = {
501 	.func		= bpf_get_stackid,
502 	.gpl_only	= true,
503 	.ret_type	= RET_INTEGER,
504 	.arg1_type	= ARG_PTR_TO_CTX,
505 	.arg2_type	= ARG_CONST_MAP_PTR,
506 	.arg3_type	= ARG_ANYTHING,
507 };
508 
count_kernel_ip(struct perf_callchain_entry * trace)509 static __u64 count_kernel_ip(struct perf_callchain_entry *trace)
510 {
511 	__u64 nr_kernel = 0;
512 
513 	while (nr_kernel < trace->nr) {
514 		if (trace->ip[nr_kernel] == PERF_CONTEXT_USER)
515 			break;
516 		nr_kernel++;
517 	}
518 	return nr_kernel;
519 }
520 
BPF_CALL_3(bpf_get_stackid_pe,struct bpf_perf_event_data_kern *,ctx,struct bpf_map *,map,u64,flags)521 BPF_CALL_3(bpf_get_stackid_pe, struct bpf_perf_event_data_kern *, ctx,
522 	   struct bpf_map *, map, u64, flags)
523 {
524 	struct perf_event *event = ctx->event;
525 	struct perf_callchain_entry *trace;
526 	bool kernel, user;
527 	__u64 nr_kernel;
528 	int ret;
529 
530 	/* perf_sample_data doesn't have callchain, use bpf_get_stackid */
531 	if (!(event->attr.sample_type & __PERF_SAMPLE_CALLCHAIN_EARLY))
532 		return bpf_get_stackid((unsigned long)(ctx->regs),
533 				       (unsigned long) map, flags, 0, 0);
534 
535 	if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
536 			       BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
537 		return -EINVAL;
538 
539 	user = flags & BPF_F_USER_STACK;
540 	kernel = !user;
541 
542 	trace = ctx->data->callchain;
543 	if (unlikely(!trace))
544 		return -EFAULT;
545 
546 	nr_kernel = count_kernel_ip(trace);
547 
548 	if (kernel) {
549 		__u64 nr = trace->nr;
550 
551 		trace->nr = nr_kernel;
552 		ret = __bpf_get_stackid(map, trace, flags);
553 
554 		/* restore nr */
555 		trace->nr = nr;
556 	} else { /* user */
557 		u64 skip = flags & BPF_F_SKIP_FIELD_MASK;
558 
559 		skip += nr_kernel;
560 		if (skip > BPF_F_SKIP_FIELD_MASK)
561 			return -EFAULT;
562 
563 		flags = (flags & ~BPF_F_SKIP_FIELD_MASK) | skip;
564 		ret = __bpf_get_stackid(map, trace, flags);
565 	}
566 	return ret;
567 }
568 
569 const struct bpf_func_proto bpf_get_stackid_proto_pe = {
570 	.func		= bpf_get_stackid_pe,
571 	.gpl_only	= false,
572 	.ret_type	= RET_INTEGER,
573 	.arg1_type	= ARG_PTR_TO_CTX,
574 	.arg2_type	= ARG_CONST_MAP_PTR,
575 	.arg3_type	= ARG_ANYTHING,
576 };
577 
__bpf_get_stack(struct pt_regs * regs,struct task_struct * task,struct perf_callchain_entry * trace_in,void * buf,u32 size,u64 flags)578 static long __bpf_get_stack(struct pt_regs *regs, struct task_struct *task,
579 			    struct perf_callchain_entry *trace_in,
580 			    void *buf, u32 size, u64 flags)
581 {
582 	u32 init_nr, trace_nr, copy_len, elem_size, num_elem;
583 	bool user_build_id = flags & BPF_F_USER_BUILD_ID;
584 	u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
585 	bool user = flags & BPF_F_USER_STACK;
586 	struct perf_callchain_entry *trace;
587 	bool kernel = !user;
588 	int err = -EINVAL;
589 	u64 *ips;
590 
591 	if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
592 			       BPF_F_USER_BUILD_ID)))
593 		goto clear;
594 	if (kernel && user_build_id)
595 		goto clear;
596 
597 	elem_size = (user && user_build_id) ? sizeof(struct bpf_stack_build_id)
598 					    : sizeof(u64);
599 	if (unlikely(size % elem_size))
600 		goto clear;
601 
602 	/* cannot get valid user stack for task without user_mode regs */
603 	if (task && user && !user_mode(regs))
604 		goto err_fault;
605 
606 	num_elem = size / elem_size;
607 	if (sysctl_perf_event_max_stack < num_elem)
608 		init_nr = 0;
609 	else
610 		init_nr = sysctl_perf_event_max_stack - num_elem;
611 
612 	if (trace_in)
613 		trace = trace_in;
614 	else if (kernel && task)
615 		trace = get_callchain_entry_for_task(task, init_nr);
616 	else
617 		trace = get_perf_callchain(regs, init_nr, kernel, user,
618 					   sysctl_perf_event_max_stack,
619 					   false, false);
620 	if (unlikely(!trace))
621 		goto err_fault;
622 
623 	trace_nr = trace->nr - init_nr;
624 	if (trace_nr < skip)
625 		goto err_fault;
626 
627 	trace_nr -= skip;
628 	trace_nr = (trace_nr <= num_elem) ? trace_nr : num_elem;
629 	copy_len = trace_nr * elem_size;
630 	ips = trace->ip + skip + init_nr;
631 	if (user && user_build_id)
632 		stack_map_get_build_id_offset(buf, ips, trace_nr, user);
633 	else
634 		memcpy(buf, ips, copy_len);
635 
636 	if (size > copy_len)
637 		memset(buf + copy_len, 0, size - copy_len);
638 	return copy_len;
639 
640 err_fault:
641 	err = -EFAULT;
642 clear:
643 	memset(buf, 0, size);
644 	return err;
645 }
646 
BPF_CALL_4(bpf_get_stack,struct pt_regs *,regs,void *,buf,u32,size,u64,flags)647 BPF_CALL_4(bpf_get_stack, struct pt_regs *, regs, void *, buf, u32, size,
648 	   u64, flags)
649 {
650 	return __bpf_get_stack(regs, NULL, NULL, buf, size, flags);
651 }
652 
653 const struct bpf_func_proto bpf_get_stack_proto = {
654 	.func		= bpf_get_stack,
655 	.gpl_only	= true,
656 	.ret_type	= RET_INTEGER,
657 	.arg1_type	= ARG_PTR_TO_CTX,
658 	.arg2_type	= ARG_PTR_TO_UNINIT_MEM,
659 	.arg3_type	= ARG_CONST_SIZE_OR_ZERO,
660 	.arg4_type	= ARG_ANYTHING,
661 };
662 
BPF_CALL_4(bpf_get_task_stack,struct task_struct *,task,void *,buf,u32,size,u64,flags)663 BPF_CALL_4(bpf_get_task_stack, struct task_struct *, task, void *, buf,
664 	   u32, size, u64, flags)
665 {
666 	struct pt_regs *regs;
667 	long res = -EINVAL;
668 
669 	if (!try_get_task_stack(task))
670 		return -EFAULT;
671 
672 	regs = task_pt_regs(task);
673 	if (regs)
674 		res = __bpf_get_stack(regs, task, NULL, buf, size, flags);
675 	put_task_stack(task);
676 
677 	return res;
678 }
679 
680 BTF_ID_LIST_SINGLE(bpf_get_task_stack_btf_ids, struct, task_struct)
681 
682 const struct bpf_func_proto bpf_get_task_stack_proto = {
683 	.func		= bpf_get_task_stack,
684 	.gpl_only	= false,
685 	.ret_type	= RET_INTEGER,
686 	.arg1_type	= ARG_PTR_TO_BTF_ID,
687 	.arg1_btf_id	= &bpf_get_task_stack_btf_ids[0],
688 	.arg2_type	= ARG_PTR_TO_UNINIT_MEM,
689 	.arg3_type	= ARG_CONST_SIZE_OR_ZERO,
690 	.arg4_type	= ARG_ANYTHING,
691 };
692 
BPF_CALL_4(bpf_get_stack_pe,struct bpf_perf_event_data_kern *,ctx,void *,buf,u32,size,u64,flags)693 BPF_CALL_4(bpf_get_stack_pe, struct bpf_perf_event_data_kern *, ctx,
694 	   void *, buf, u32, size, u64, flags)
695 {
696 	struct pt_regs *regs = (struct pt_regs *)(ctx->regs);
697 	struct perf_event *event = ctx->event;
698 	struct perf_callchain_entry *trace;
699 	bool kernel, user;
700 	int err = -EINVAL;
701 	__u64 nr_kernel;
702 
703 	if (!(event->attr.sample_type & __PERF_SAMPLE_CALLCHAIN_EARLY))
704 		return __bpf_get_stack(regs, NULL, NULL, buf, size, flags);
705 
706 	if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
707 			       BPF_F_USER_BUILD_ID)))
708 		goto clear;
709 
710 	user = flags & BPF_F_USER_STACK;
711 	kernel = !user;
712 
713 	err = -EFAULT;
714 	trace = ctx->data->callchain;
715 	if (unlikely(!trace))
716 		goto clear;
717 
718 	nr_kernel = count_kernel_ip(trace);
719 
720 	if (kernel) {
721 		__u64 nr = trace->nr;
722 
723 		trace->nr = nr_kernel;
724 		err = __bpf_get_stack(regs, NULL, trace, buf, size, flags);
725 
726 		/* restore nr */
727 		trace->nr = nr;
728 	} else { /* user */
729 		u64 skip = flags & BPF_F_SKIP_FIELD_MASK;
730 
731 		skip += nr_kernel;
732 		if (skip > BPF_F_SKIP_FIELD_MASK)
733 			goto clear;
734 
735 		flags = (flags & ~BPF_F_SKIP_FIELD_MASK) | skip;
736 		err = __bpf_get_stack(regs, NULL, trace, buf, size, flags);
737 	}
738 	return err;
739 
740 clear:
741 	memset(buf, 0, size);
742 	return err;
743 
744 }
745 
746 const struct bpf_func_proto bpf_get_stack_proto_pe = {
747 	.func		= bpf_get_stack_pe,
748 	.gpl_only	= true,
749 	.ret_type	= RET_INTEGER,
750 	.arg1_type	= ARG_PTR_TO_CTX,
751 	.arg2_type	= ARG_PTR_TO_UNINIT_MEM,
752 	.arg3_type	= ARG_CONST_SIZE_OR_ZERO,
753 	.arg4_type	= ARG_ANYTHING,
754 };
755 
756 /* Called from eBPF program */
stack_map_lookup_elem(struct bpf_map * map,void * key)757 static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
758 {
759 	return ERR_PTR(-EOPNOTSUPP);
760 }
761 
762 /* Called from syscall */
bpf_stackmap_copy(struct bpf_map * map,void * key,void * value)763 int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value)
764 {
765 	struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
766 	struct stack_map_bucket *bucket, *old_bucket;
767 	u32 id = *(u32 *)key, trace_len;
768 
769 	if (unlikely(id >= smap->n_buckets))
770 		return -ENOENT;
771 
772 	bucket = xchg(&smap->buckets[id], NULL);
773 	if (!bucket)
774 		return -ENOENT;
775 
776 	trace_len = bucket->nr * stack_map_data_size(map);
777 	memcpy(value, bucket->data, trace_len);
778 	memset(value + trace_len, 0, map->value_size - trace_len);
779 
780 	old_bucket = xchg(&smap->buckets[id], bucket);
781 	if (old_bucket)
782 		pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
783 	return 0;
784 }
785 
stack_map_get_next_key(struct bpf_map * map,void * key,void * next_key)786 static int stack_map_get_next_key(struct bpf_map *map, void *key,
787 				  void *next_key)
788 {
789 	struct bpf_stack_map *smap = container_of(map,
790 						  struct bpf_stack_map, map);
791 	u32 id;
792 
793 	WARN_ON_ONCE(!rcu_read_lock_held());
794 
795 	if (!key) {
796 		id = 0;
797 	} else {
798 		id = *(u32 *)key;
799 		if (id >= smap->n_buckets || !smap->buckets[id])
800 			id = 0;
801 		else
802 			id++;
803 	}
804 
805 	while (id < smap->n_buckets && !smap->buckets[id])
806 		id++;
807 
808 	if (id >= smap->n_buckets)
809 		return -ENOENT;
810 
811 	*(u32 *)next_key = id;
812 	return 0;
813 }
814 
stack_map_update_elem(struct bpf_map * map,void * key,void * value,u64 map_flags)815 static int stack_map_update_elem(struct bpf_map *map, void *key, void *value,
816 				 u64 map_flags)
817 {
818 	return -EINVAL;
819 }
820 
821 /* Called from syscall or from eBPF program */
stack_map_delete_elem(struct bpf_map * map,void * key)822 static int stack_map_delete_elem(struct bpf_map *map, void *key)
823 {
824 	struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
825 	struct stack_map_bucket *old_bucket;
826 	u32 id = *(u32 *)key;
827 
828 	if (unlikely(id >= smap->n_buckets))
829 		return -E2BIG;
830 
831 	old_bucket = xchg(&smap->buckets[id], NULL);
832 	if (old_bucket) {
833 		pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
834 		return 0;
835 	} else {
836 		return -ENOENT;
837 	}
838 }
839 
840 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
stack_map_free(struct bpf_map * map)841 static void stack_map_free(struct bpf_map *map)
842 {
843 	struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
844 
845 	bpf_map_area_free(smap->elems);
846 	pcpu_freelist_destroy(&smap->freelist);
847 	bpf_map_area_free(smap);
848 	put_callchain_buffers();
849 }
850 
851 static int stack_trace_map_btf_id;
852 const struct bpf_map_ops stack_trace_map_ops = {
853 	.map_meta_equal = bpf_map_meta_equal,
854 	.map_alloc = stack_map_alloc,
855 	.map_free = stack_map_free,
856 	.map_get_next_key = stack_map_get_next_key,
857 	.map_lookup_elem = stack_map_lookup_elem,
858 	.map_update_elem = stack_map_update_elem,
859 	.map_delete_elem = stack_map_delete_elem,
860 	.map_check_btf = map_check_no_btf,
861 	.map_btf_name = "bpf_stack_map",
862 	.map_btf_id = &stack_trace_map_btf_id,
863 };
864 
stack_map_init(void)865 static int __init stack_map_init(void)
866 {
867 	int cpu;
868 	struct stack_map_irq_work *work;
869 
870 	for_each_possible_cpu(cpu) {
871 		work = per_cpu_ptr(&up_read_work, cpu);
872 		init_irq_work(&work->irq_work, do_up_read);
873 	}
874 	return 0;
875 }
876 subsys_initcall(stack_map_init);
877