kernel/bpf/stackmap.c

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