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/stacktrace.h>
   8 #include <linux/perf_event.h>
   9 #include <linux/elf.h>
  10 #include <linux/pagemap.h>
  11 #include <linux/irq_work.h>
  12 #include <linux/btf_ids.h>
  13 #include "percpu_freelist.h"
  14
  15 #define STACK_CREATE_FLAG_MASK                                  \
  16         (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY |        \
  17          BPF_F_STACK_BUILD_ID)
  18
  19 struct stack_map_bucket {
  20         struct pcpu_freelist_node fnode;
  21         u32 hash;
  22         u32 nr;
  23         u64 data[];
  24 };
  25
  26 struct bpf_stack_map {
  27         struct bpf_map map;
  28         void *elems;
  29         struct pcpu_freelist freelist;
  30         u32 n_buckets;
  31         struct stack_map_bucket *buckets[];
  32 };
  33
  34 /* irq_work to run up_read() for build_id lookup in nmi context */
  35 struct stack_map_irq_work {
  36         struct irq_work irq_work;
  37         struct mm_struct *mm;
  38 };
  39
  40 static void do_up_read(struct irq_work *entry)
  41 {
  42         struct stack_map_irq_work *work;
  43
  44         if (WARN_ON_ONCE(IS_ENABLED(CONFIG_PREEMPT_RT)))
  45                 return;
  46
  47         work = container_of(entry, struct stack_map_irq_work, irq_work);
  48         mmap_read_unlock_non_owner(work->mm);
  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         struct bpf_map_memory mem;
  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
 119         cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap);
 120         cost += n_buckets * (value_size + sizeof(struct stack_map_bucket));
 121         err = bpf_map_charge_init(&mem, cost);
 122         if (err)
 123                 return ERR_PTR(err);
 124
 125         smap = bpf_map_area_alloc(cost, bpf_map_attr_numa_node(attr));
 126         if (!smap) {
 127                 bpf_map_charge_finish(&mem);
 128                 return ERR_PTR(-ENOMEM);
 129         }
 130
 131         bpf_map_init_from_attr(&smap->map, attr);
 132         smap->map.value_size = value_size;
 133         smap->n_buckets = n_buckets;
 134
 135         err = get_callchain_buffers(sysctl_perf_event_max_stack);
 136         if (err)
 137                 goto free_charge;
 138
 139         err = prealloc_elems_and_freelist(smap);
 140         if (err)
 141                 goto put_buffers;
 142
 143         bpf_map_charge_move(&smap->map.memory, &mem);
 144
 145         return &smap->map;
 146
 147 put_buffers:
 148         put_callchain_buffers();
 149 free_charge:
 150         bpf_map_charge_finish(&mem);
 151         bpf_map_area_free(smap);
 152         return ERR_PTR(err);
 153 }
 154
 155 #define BPF_BUILD_ID 3
 156 /*
 157  * Parse build id from the note segment. This logic can be shared between
 158  * 32-bit and 64-bit system, because Elf32_Nhdr and Elf64_Nhdr are
 159  * identical.
 160  */
 161 static inline int stack_map_parse_build_id(void *page_addr,
 162                                            unsigned char *build_id,
 163                                            void *note_start,
 164                                            Elf32_Word note_size)
 165 {
 166         Elf32_Word note_offs = 0, new_offs;
 167
 168         /* check for overflow */
 169         if (note_start < page_addr || note_start + note_size < note_start)
 170                 return -EINVAL;
 171
 172         /* only supports note that fits in the first page */
 173         if (note_start + note_size > page_addr + PAGE_SIZE)
 174                 return -EINVAL;
 175
 176         while (note_offs + sizeof(Elf32_Nhdr) < note_size) {
 177                 Elf32_Nhdr *nhdr = (Elf32_Nhdr *)(note_start + note_offs);
 178
 179                 if (nhdr->n_type == BPF_BUILD_ID &&
 180                     nhdr->n_namesz == sizeof("GNU") &&
 181                     nhdr->n_descsz > 0 &&
 182                     nhdr->n_descsz <= BPF_BUILD_ID_SIZE) {
 183                         memcpy(build_id,
 184                                note_start + note_offs +
 185                                ALIGN(sizeof("GNU"), 4) + sizeof(Elf32_Nhdr),
 186                                nhdr->n_descsz);
 187                         memset(build_id + nhdr->n_descsz, 0,
 188                                BPF_BUILD_ID_SIZE - nhdr->n_descsz);
 189                         return 0;
 190                 }
 191                 new_offs = note_offs + sizeof(Elf32_Nhdr) +
 192                         ALIGN(nhdr->n_namesz, 4) + ALIGN(nhdr->n_descsz, 4);
 193                 if (new_offs <= note_offs)  /* overflow */
 194                         break;
 195                 note_offs = new_offs;
 196         }
 197         return -EINVAL;
 198 }
 199
 200 /* Parse build ID from 32-bit ELF */
 201 static int stack_map_get_build_id_32(void *page_addr,
 202                                      unsigned char *build_id)
 203 {
 204         Elf32_Ehdr *ehdr = (Elf32_Ehdr *)page_addr;
 205         Elf32_Phdr *phdr;
 206         int i;
 207
 208         /* only supports phdr that fits in one page */
 209         if (ehdr->e_phnum >
 210             (PAGE_SIZE - sizeof(Elf32_Ehdr)) / sizeof(Elf32_Phdr))
 211                 return -EINVAL;
 212
 213         phdr = (Elf32_Phdr *)(page_addr + sizeof(Elf32_Ehdr));
 214
 215         for (i = 0; i < ehdr->e_phnum; ++i)
 216                 if (phdr[i].p_type == PT_NOTE)
 217                         return stack_map_parse_build_id(page_addr, build_id,
 218                                         page_addr + phdr[i].p_offset,
 219                                         phdr[i].p_filesz);
 220         return -EINVAL;
 221 }
 222
 223 /* Parse build ID from 64-bit ELF */
 224 static int stack_map_get_build_id_64(void *page_addr,
 225                                      unsigned char *build_id)
 226 {
 227         Elf64_Ehdr *ehdr = (Elf64_Ehdr *)page_addr;
 228         Elf64_Phdr *phdr;
 229         int i;
 230
 231         /* only supports phdr that fits in one page */
 232         if (ehdr->e_phnum >
 233             (PAGE_SIZE - sizeof(Elf64_Ehdr)) / sizeof(Elf64_Phdr))
 234                 return -EINVAL;
 235
 236         phdr = (Elf64_Phdr *)(page_addr + sizeof(Elf64_Ehdr));
 237
 238         for (i = 0; i < ehdr->e_phnum; ++i)
 239                 if (phdr[i].p_type == PT_NOTE)
 240                         return stack_map_parse_build_id(page_addr, build_id,
 241                                         page_addr + phdr[i].p_offset,
 242                                         phdr[i].p_filesz);
 243         return -EINVAL;
 244 }
 245
 246 /* Parse build ID of ELF file mapped to vma */
 247 static int stack_map_get_build_id(struct vm_area_struct *vma,
 248                                   unsigned char *build_id)
 249 {
 250         Elf32_Ehdr *ehdr;
 251         struct page *page;
 252         void *page_addr;
 253         int ret;
 254
 255         /* only works for page backed storage  */
 256         if (!vma->vm_file)
 257                 return -EINVAL;
 258
 259         page = find_get_page(vma->vm_file->f_mapping, 0);
 260         if (!page)
 261                 return -EFAULT; /* page not mapped */
 262
 263         ret = -EINVAL;
 264         page_addr = kmap_atomic(page);
 265         ehdr = (Elf32_Ehdr *)page_addr;
 266
 267         /* compare magic x7f "ELF" */
 268         if (memcmp(ehdr->e_ident, ELFMAG, SELFMAG) != 0)
 269                 goto out;
 270
 271         /* only support executable file and shared object file */
 272         if (ehdr->e_type != ET_EXEC && ehdr->e_type != ET_DYN)
 273                 goto out;
 274
 275         if (ehdr->e_ident[EI_CLASS] == ELFCLASS32)
 276                 ret = stack_map_get_build_id_32(page_addr, build_id);
 277         else if (ehdr->e_ident[EI_CLASS] == ELFCLASS64)
 278                 ret = stack_map_get_build_id_64(page_addr, build_id);
 279 out:
 280         kunmap_atomic(page_addr);
 281         put_page(page);
 282         return ret;
 283 }
 284
 285 static void stack_map_get_build_id_offset(struct bpf_stack_build_id *id_offs,
 286                                           u64 *ips, u32 trace_nr, bool user)
 287 {
 288         int i;
 289         struct vm_area_struct *vma;
 290         bool irq_work_busy = false;
 291         struct stack_map_irq_work *work = NULL;
 292
 293         if (irqs_disabled()) {
 294                 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) {
 295                         work = this_cpu_ptr(&up_read_work);
 296                         if (atomic_read(&work->irq_work.flags) & IRQ_WORK_BUSY) {
 297                                 /* cannot queue more up_read, fallback */
 298                                 irq_work_busy = true;
 299                         }
 300                 } else {
 301                         /*
 302                          * PREEMPT_RT does not allow to trylock mmap sem in
 303                          * interrupt disabled context. Force the fallback code.
 304                          */
 305                         irq_work_busy = true;
 306                 }
 307         }
 308
 309         /*
 310          * We cannot do up_read() when the irq is disabled, because of
 311          * risk to deadlock with rq_lock. To do build_id lookup when the
 312          * irqs are disabled, we need to run up_read() in irq_work. We use
 313          * a percpu variable to do the irq_work. If the irq_work is
 314          * already used by another lookup, we fall back to report ips.
 315          *
 316          * Same fallback is used for kernel stack (!user) on a stackmap
 317          * with build_id.
 318          */
 319         if (!user || !current || !current->mm || irq_work_busy ||
 320             !mmap_read_trylock_non_owner(current->mm)) {
 321                 /* cannot access current->mm, fall back to ips */
 322                 for (i = 0; i < trace_nr; i++) {
 323                         id_offs[i].status = BPF_STACK_BUILD_ID_IP;
 324                         id_offs[i].ip = ips[i];
 325                         memset(id_offs[i].build_id, 0, BPF_BUILD_ID_SIZE);
 326                 }
 327                 return;
 328         }
 329
 330         for (i = 0; i < trace_nr; i++) {
 331                 vma = find_vma(current->mm, ips[i]);
 332                 if (!vma || stack_map_get_build_id(vma, id_offs[i].build_id)) {
 333                         /* per entry fall back to ips */
 334                         id_offs[i].status = BPF_STACK_BUILD_ID_IP;
 335                         id_offs[i].ip = ips[i];
 336                         memset(id_offs[i].build_id, 0, BPF_BUILD_ID_SIZE);
 337                         continue;
 338                 }
 339                 id_offs[i].offset = (vma->vm_pgoff << PAGE_SHIFT) + ips[i]
 340                         - vma->vm_start;
 341                 id_offs[i].status = BPF_STACK_BUILD_ID_VALID;
 342         }
 343
 344         if (!work) {
 345                 mmap_read_unlock_non_owner(current->mm);
 346         } else {
 347                 work->mm = current->mm;
 348                 irq_work_queue(&work->irq_work);
 349         }
 350 }
 351
 352 static struct perf_callchain_entry *
 353 get_callchain_entry_for_task(struct task_struct *task, u32 init_nr)
 354 {
 355 #ifdef CONFIG_STACKTRACE
 356         struct perf_callchain_entry *entry;
 357         int rctx;
 358
 359         entry = get_callchain_entry(&rctx);
 360
 361         if (!entry)
 362                 return NULL;
 363
 364         entry->nr = init_nr +
 365                 stack_trace_save_tsk(task, (unsigned long *)(entry->ip + init_nr),
 366                                      sysctl_perf_event_max_stack - init_nr, 0);
 367
 368         /* stack_trace_save_tsk() works on unsigned long array, while
 369          * perf_callchain_entry uses u64 array. For 32-bit systems, it is
 370          * necessary to fix this mismatch.
 371          */
 372         if (__BITS_PER_LONG != 64) {
 373                 unsigned long *from = (unsigned long *) entry->ip;
 374                 u64 *to = entry->ip;
 375                 int i;
 376
 377                 /* copy data from the end to avoid using extra buffer */
 378                 for (i = entry->nr - 1; i >= (int)init_nr; i--)
 379                         to[i] = (u64)(from[i]);
 380         }
 381
 382         put_callchain_entry(rctx);
 383
 384         return entry;
 385 #else /* CONFIG_STACKTRACE */
 386         return NULL;
 387 #endif
 388 }
 389
 390 BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
 391            u64, flags)
 392 {
 393         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 394         struct perf_callchain_entry *trace;
 395         struct stack_map_bucket *bucket, *new_bucket, *old_bucket;
 396         u32 max_depth = map->value_size / stack_map_data_size(map);
 397         /* stack_map_alloc() checks that max_depth <= sysctl_perf_event_max_stack */
 398         u32 init_nr = sysctl_perf_event_max_stack - max_depth;
 399         u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
 400         u32 hash, id, trace_nr, trace_len;
 401         bool user = flags & BPF_F_USER_STACK;
 402         bool kernel = !user;
 403         u64 *ips;
 404         bool hash_matches;
 405
 406         if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
 407                                BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
 408                 return -EINVAL;
 409
 410         trace = get_perf_callchain(regs, init_nr, kernel, user,
 411                                    sysctl_perf_event_max_stack, false, false);
 412
 413         if (unlikely(!trace))
 414                 /* couldn't fetch the stack trace */
 415                 return -EFAULT;
 416
 417         /* get_perf_callchain() guarantees that trace->nr >= init_nr
 418          * and trace-nr <= sysctl_perf_event_max_stack, so trace_nr <= max_depth
 419          */
 420         trace_nr = trace->nr - init_nr;
 421
 422         if (trace_nr <= skip)
 423                 /* skipping more than usable stack trace */
 424                 return -EFAULT;
 425
 426         trace_nr -= skip;
 427         trace_len = trace_nr * sizeof(u64);
 428         ips = trace->ip + skip + init_nr;
 429         hash = jhash2((u32 *)ips, trace_len / sizeof(u32), 0);
 430         id = hash & (smap->n_buckets - 1);
 431         bucket = READ_ONCE(smap->buckets[id]);
 432
 433         hash_matches = bucket && bucket->hash == hash;
 434         /* fast cmp */
 435         if (hash_matches && flags & BPF_F_FAST_STACK_CMP)
 436                 return id;
 437
 438         if (stack_map_use_build_id(map)) {
 439                 /* for build_id+offset, pop a bucket before slow cmp */
 440                 new_bucket = (struct stack_map_bucket *)
 441                         pcpu_freelist_pop(&smap->freelist);
 442                 if (unlikely(!new_bucket))
 443                         return -ENOMEM;
 444                 new_bucket->nr = trace_nr;
 445                 stack_map_get_build_id_offset(
 446                         (struct bpf_stack_build_id *)new_bucket->data,
 447                         ips, trace_nr, user);
 448                 trace_len = trace_nr * sizeof(struct bpf_stack_build_id);
 449                 if (hash_matches && bucket->nr == trace_nr &&
 450                     memcmp(bucket->data, new_bucket->data, trace_len) == 0) {
 451                         pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
 452                         return id;
 453                 }
 454                 if (bucket && !(flags & BPF_F_REUSE_STACKID)) {
 455                         pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
 456                         return -EEXIST;
 457                 }
 458         } else {
 459                 if (hash_matches && bucket->nr == trace_nr &&
 460                     memcmp(bucket->data, ips, trace_len) == 0)
 461                         return id;
 462                 if (bucket && !(flags & BPF_F_REUSE_STACKID))
 463                         return -EEXIST;
 464
 465                 new_bucket = (struct stack_map_bucket *)
 466                         pcpu_freelist_pop(&smap->freelist);
 467                 if (unlikely(!new_bucket))
 468                         return -ENOMEM;
 469                 memcpy(new_bucket->data, ips, trace_len);
 470         }
 471
 472         new_bucket->hash = hash;
 473         new_bucket->nr = trace_nr;
 474
 475         old_bucket = xchg(&smap->buckets[id], new_bucket);
 476         if (old_bucket)
 477                 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
 478         return id;
 479 }
 480
 481 const struct bpf_func_proto bpf_get_stackid_proto = {
 482         .func           = bpf_get_stackid,
 483         .gpl_only       = true,
 484         .ret_type       = RET_INTEGER,
 485         .arg1_type      = ARG_PTR_TO_CTX,
 486         .arg2_type      = ARG_CONST_MAP_PTR,
 487         .arg3_type      = ARG_ANYTHING,
 488 };
 489
 490 static long __bpf_get_stack(struct pt_regs *regs, struct task_struct *task,
 491                             void *buf, u32 size, u64 flags)
 492 {
 493         u32 init_nr, trace_nr, copy_len, elem_size, num_elem;
 494         bool user_build_id = flags & BPF_F_USER_BUILD_ID;
 495         u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
 496         bool user = flags & BPF_F_USER_STACK;
 497         struct perf_callchain_entry *trace;
 498         bool kernel = !user;
 499         int err = -EINVAL;
 500         u64 *ips;
 501
 502         if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
 503                                BPF_F_USER_BUILD_ID)))
 504                 goto clear;
 505         if (kernel && user_build_id)
 506                 goto clear;
 507
 508         elem_size = (user && user_build_id) ? sizeof(struct bpf_stack_build_id)
 509                                             : sizeof(u64);
 510         if (unlikely(size % elem_size))
 511                 goto clear;
 512
 513         /* cannot get valid user stack for task without user_mode regs */
 514         if (task && user && !user_mode(regs))
 515                 goto err_fault;
 516
 517         num_elem = size / elem_size;
 518         if (sysctl_perf_event_max_stack < num_elem)
 519                 init_nr = 0;
 520         else
 521                 init_nr = sysctl_perf_event_max_stack - num_elem;
 522
 523         if (kernel && task)
 524                 trace = get_callchain_entry_for_task(task, init_nr);
 525         else
 526                 trace = get_perf_callchain(regs, init_nr, kernel, user,
 527                                            sysctl_perf_event_max_stack,
 528                                            false, false);
 529         if (unlikely(!trace))
 530                 goto err_fault;
 531
 532         trace_nr = trace->nr - init_nr;
 533         if (trace_nr < skip)
 534                 goto err_fault;
 535
 536         trace_nr -= skip;
 537         trace_nr = (trace_nr <= num_elem) ? trace_nr : num_elem;
 538         copy_len = trace_nr * elem_size;
 539         ips = trace->ip + skip + init_nr;
 540         if (user && user_build_id)
 541                 stack_map_get_build_id_offset(buf, ips, trace_nr, user);
 542         else
 543                 memcpy(buf, ips, copy_len);
 544
 545         if (size > copy_len)
 546                 memset(buf + copy_len, 0, size - copy_len);
 547         return copy_len;
 548
 549 err_fault:
 550         err = -EFAULT;
 551 clear:
 552         memset(buf, 0, size);
 553         return err;
 554 }
 555
 556 BPF_CALL_4(bpf_get_stack, struct pt_regs *, regs, void *, buf, u32, size,
 557            u64, flags)
 558 {
 559         return __bpf_get_stack(regs, NULL, buf, size, flags);
 560 }
 561
 562 const struct bpf_func_proto bpf_get_stack_proto = {
 563         .func           = bpf_get_stack,
 564         .gpl_only       = true,
 565         .ret_type       = RET_INTEGER,
 566         .arg1_type      = ARG_PTR_TO_CTX,
 567         .arg2_type      = ARG_PTR_TO_UNINIT_MEM,
 568         .arg3_type      = ARG_CONST_SIZE_OR_ZERO,
 569         .arg4_type      = ARG_ANYTHING,
 570 };
 571
 572 BPF_CALL_4(bpf_get_task_stack, struct task_struct *, task, void *, buf,
 573            u32, size, u64, flags)
 574 {
 575         struct pt_regs *regs = task_pt_regs(task);
 576
 577         return __bpf_get_stack(regs, task, buf, size, flags);
 578 }
 579
 580 BTF_ID_LIST(bpf_get_task_stack_btf_ids)
 581 BTF_ID(struct, task_struct)
 582
 583 const struct bpf_func_proto bpf_get_task_stack_proto = {
 584         .func           = bpf_get_task_stack,
 585         .gpl_only       = false,
 586         .ret_type       = RET_INTEGER,
 587         .arg1_type      = ARG_PTR_TO_BTF_ID,
 588         .arg2_type      = ARG_PTR_TO_UNINIT_MEM,
 589         .arg3_type      = ARG_CONST_SIZE_OR_ZERO,
 590         .arg4_type      = ARG_ANYTHING,
 591         .btf_id         = bpf_get_task_stack_btf_ids,
 592 };
 593
 594 /* Called from eBPF program */
 595 static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
 596 {
 597         return ERR_PTR(-EOPNOTSUPP);
 598 }
 599
 600 /* Called from syscall */
 601 int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value)
 602 {
 603         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 604         struct stack_map_bucket *bucket, *old_bucket;
 605         u32 id = *(u32 *)key, trace_len;
 606
 607         if (unlikely(id >= smap->n_buckets))
 608                 return -ENOENT;
 609
 610         bucket = xchg(&smap->buckets[id], NULL);
 611         if (!bucket)
 612                 return -ENOENT;
 613
 614         trace_len = bucket->nr * stack_map_data_size(map);
 615         memcpy(value, bucket->data, trace_len);
 616         memset(value + trace_len, 0, map->value_size - trace_len);
 617
 618         old_bucket = xchg(&smap->buckets[id], bucket);
 619         if (old_bucket)
 620                 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
 621         return 0;
 622 }
 623
 624 static int stack_map_get_next_key(struct bpf_map *map, void *key,
 625                                   void *next_key)
 626 {
 627         struct bpf_stack_map *smap = container_of(map,
 628                                                   struct bpf_stack_map, map);
 629         u32 id;
 630
 631         WARN_ON_ONCE(!rcu_read_lock_held());
 632
 633         if (!key) {
 634                 id = 0;
 635         } else {
 636                 id = *(u32 *)key;
 637                 if (id >= smap->n_buckets || !smap->buckets[id])
 638                         id = 0;
 639                 else
 640                         id++;
 641         }
 642
 643         while (id < smap->n_buckets && !smap->buckets[id])
 644                 id++;
 645
 646         if (id >= smap->n_buckets)
 647                 return -ENOENT;
 648
 649         *(u32 *)next_key = id;
 650         return 0;
 651 }
 652
 653 static int stack_map_update_elem(struct bpf_map *map, void *key, void *value,
 654                                  u64 map_flags)
 655 {
 656         return -EINVAL;
 657 }
 658
 659 /* Called from syscall or from eBPF program */
 660 static int stack_map_delete_elem(struct bpf_map *map, void *key)
 661 {
 662         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 663         struct stack_map_bucket *old_bucket;
 664         u32 id = *(u32 *)key;
 665
 666         if (unlikely(id >= smap->n_buckets))
 667                 return -E2BIG;
 668
 669         old_bucket = xchg(&smap->buckets[id], NULL);
 670         if (old_bucket) {
 671                 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
 672                 return 0;
 673         } else {
 674                 return -ENOENT;
 675         }
 676 }
 677
 678 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
 679 static void stack_map_free(struct bpf_map *map)
 680 {
 681         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 682
 683         bpf_map_area_free(smap->elems);
 684         pcpu_freelist_destroy(&smap->freelist);
 685         bpf_map_area_free(smap);
 686         put_callchain_buffers();
 687 }
 688
 689 static int stack_trace_map_btf_id;
 690 const struct bpf_map_ops stack_trace_map_ops = {
 691         .map_alloc = stack_map_alloc,
 692         .map_free = stack_map_free,
 693         .map_get_next_key = stack_map_get_next_key,
 694         .map_lookup_elem = stack_map_lookup_elem,
 695         .map_update_elem = stack_map_update_elem,
 696         .map_delete_elem = stack_map_delete_elem,
 697         .map_check_btf = map_check_no_btf,
 698         .map_btf_name = "bpf_stack_map",
 699         .map_btf_id = &stack_trace_map_btf_id,
 700 };
 701
 702 static int __init stack_map_init(void)
 703 {
 704         int cpu;
 705         struct stack_map_irq_work *work;
 706
 707         for_each_possible_cpu(cpu) {
 708                 work = per_cpu_ptr(&up_read_work, cpu);
 709                 init_irq_work(&work->irq_work, do_up_read);
 710         }
 711         return 0;
 712 }
 713 subsys_initcall(stack_map_init);