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/irq_work.h>
  11 #include <linux/btf_ids.h>
  12 #include <linux/buildid.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         u64 cost, n_buckets;
  93         int err;
  94
  95         if (!bpf_capable())
  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         if (!n_buckets)
 118                 return ERR_PTR(-E2BIG);
 119
 120         cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap);
 121         cost += n_buckets * (value_size + sizeof(struct stack_map_bucket));
 122         smap = bpf_map_area_alloc(cost, bpf_map_attr_numa_node(attr));
 123         if (!smap)
 124                 return ERR_PTR(-ENOMEM);
 125
 126         bpf_map_init_from_attr(&smap->map, attr);
 127         smap->map.value_size = value_size;
 128         smap->n_buckets = n_buckets;
 129
 130         err = get_callchain_buffers(sysctl_perf_event_max_stack);
 131         if (err)
 132                 goto free_smap;
 133
 134         err = prealloc_elems_and_freelist(smap);
 135         if (err)
 136                 goto put_buffers;
 137
 138         return &smap->map;
 139
 140 put_buffers:
 141         put_callchain_buffers();
 142 free_smap:
 143         bpf_map_area_free(smap);
 144         return ERR_PTR(err);
 145 }
 146
 147 static void stack_map_get_build_id_offset(struct bpf_stack_build_id *id_offs,
 148                                           u64 *ips, u32 trace_nr, bool user)
 149 {
 150         int i;
 151         struct vm_area_struct *vma;
 152         bool irq_work_busy = false;
 153         struct stack_map_irq_work *work = NULL;
 154
 155         if (irqs_disabled()) {
 156                 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) {
 157                         work = this_cpu_ptr(&up_read_work);
 158                         if (irq_work_is_busy(&work->irq_work)) {
 159                                 /* cannot queue more up_read, fallback */
 160                                 irq_work_busy = true;
 161                         }
 162                 } else {
 163                         /*
 164                          * PREEMPT_RT does not allow to trylock mmap sem in
 165                          * interrupt disabled context. Force the fallback code.
 166                          */
 167                         irq_work_busy = true;
 168                 }
 169         }
 170
 171         /*
 172          * We cannot do up_read() when the irq is disabled, because of
 173          * risk to deadlock with rq_lock. To do build_id lookup when the
 174          * irqs are disabled, we need to run up_read() in irq_work. We use
 175          * a percpu variable to do the irq_work. If the irq_work is
 176          * already used by another lookup, we fall back to report ips.
 177          *
 178          * Same fallback is used for kernel stack (!user) on a stackmap
 179          * with build_id.
 180          */
 181         if (!user || !current || !current->mm || irq_work_busy ||
 182             !mmap_read_trylock_non_owner(current->mm)) {
 183                 /* cannot access current->mm, fall back to ips */
 184                 for (i = 0; i < trace_nr; i++) {
 185                         id_offs[i].status = BPF_STACK_BUILD_ID_IP;
 186                         id_offs[i].ip = ips[i];
 187                         memset(id_offs[i].build_id, 0, BUILD_ID_SIZE_MAX);
 188                 }
 189                 return;
 190         }
 191
 192         for (i = 0; i < trace_nr; i++) {
 193                 vma = find_vma(current->mm, ips[i]);
 194                 if (!vma || build_id_parse(vma, id_offs[i].build_id, NULL)) {
 195                         /* per entry fall back to ips */
 196                         id_offs[i].status = BPF_STACK_BUILD_ID_IP;
 197                         id_offs[i].ip = ips[i];
 198                         memset(id_offs[i].build_id, 0, BUILD_ID_SIZE_MAX);
 199                         continue;
 200                 }
 201                 id_offs[i].offset = (vma->vm_pgoff << PAGE_SHIFT) + ips[i]
 202                         - vma->vm_start;
 203                 id_offs[i].status = BPF_STACK_BUILD_ID_VALID;
 204         }
 205
 206         if (!work) {
 207                 mmap_read_unlock_non_owner(current->mm);
 208         } else {
 209                 work->mm = current->mm;
 210                 irq_work_queue(&work->irq_work);
 211         }
 212 }
 213
 214 static struct perf_callchain_entry *
 215 get_callchain_entry_for_task(struct task_struct *task, u32 init_nr)
 216 {
 217 #ifdef CONFIG_STACKTRACE
 218         struct perf_callchain_entry *entry;
 219         int rctx;
 220
 221         entry = get_callchain_entry(&rctx);
 222
 223         if (!entry)
 224                 return NULL;
 225
 226         entry->nr = init_nr +
 227                 stack_trace_save_tsk(task, (unsigned long *)(entry->ip + init_nr),
 228                                      sysctl_perf_event_max_stack - init_nr, 0);
 229
 230         /* stack_trace_save_tsk() works on unsigned long array, while
 231          * perf_callchain_entry uses u64 array. For 32-bit systems, it is
 232          * necessary to fix this mismatch.
 233          */
 234         if (__BITS_PER_LONG != 64) {
 235                 unsigned long *from = (unsigned long *) entry->ip;
 236                 u64 *to = entry->ip;
 237                 int i;
 238
 239                 /* copy data from the end to avoid using extra buffer */
 240                 for (i = entry->nr - 1; i >= (int)init_nr; i--)
 241                         to[i] = (u64)(from[i]);
 242         }
 243
 244         put_callchain_entry(rctx);
 245
 246         return entry;
 247 #else /* CONFIG_STACKTRACE */
 248         return NULL;
 249 #endif
 250 }
 251
 252 static long __bpf_get_stackid(struct bpf_map *map,
 253                               struct perf_callchain_entry *trace, u64 flags)
 254 {
 255         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 256         struct stack_map_bucket *bucket, *new_bucket, *old_bucket;
 257         u32 max_depth = map->value_size / stack_map_data_size(map);
 258         /* stack_map_alloc() checks that max_depth <= sysctl_perf_event_max_stack */
 259         u32 init_nr = sysctl_perf_event_max_stack - max_depth;
 260         u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
 261         u32 hash, id, trace_nr, trace_len;
 262         bool user = flags & BPF_F_USER_STACK;
 263         u64 *ips;
 264         bool hash_matches;
 265
 266         /* get_perf_callchain() guarantees that trace->nr >= init_nr
 267          * and trace-nr <= sysctl_perf_event_max_stack, so trace_nr <= max_depth
 268          */
 269         trace_nr = trace->nr - init_nr;
 270
 271         if (trace_nr <= skip)
 272                 /* skipping more than usable stack trace */
 273                 return -EFAULT;
 274
 275         trace_nr -= skip;
 276         trace_len = trace_nr * sizeof(u64);
 277         ips = trace->ip + skip + init_nr;
 278         hash = jhash2((u32 *)ips, trace_len / sizeof(u32), 0);
 279         id = hash & (smap->n_buckets - 1);
 280         bucket = READ_ONCE(smap->buckets[id]);
 281
 282         hash_matches = bucket && bucket->hash == hash;
 283         /* fast cmp */
 284         if (hash_matches && flags & BPF_F_FAST_STACK_CMP)
 285                 return id;
 286
 287         if (stack_map_use_build_id(map)) {
 288                 /* for build_id+offset, pop a bucket before slow cmp */
 289                 new_bucket = (struct stack_map_bucket *)
 290                         pcpu_freelist_pop(&smap->freelist);
 291                 if (unlikely(!new_bucket))
 292                         return -ENOMEM;
 293                 new_bucket->nr = trace_nr;
 294                 stack_map_get_build_id_offset(
 295                         (struct bpf_stack_build_id *)new_bucket->data,
 296                         ips, trace_nr, user);
 297                 trace_len = trace_nr * sizeof(struct bpf_stack_build_id);
 298                 if (hash_matches && bucket->nr == trace_nr &&
 299                     memcmp(bucket->data, new_bucket->data, trace_len) == 0) {
 300                         pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
 301                         return id;
 302                 }
 303                 if (bucket && !(flags & BPF_F_REUSE_STACKID)) {
 304                         pcpu_freelist_push(&smap->freelist, &new_bucket->fnode);
 305                         return -EEXIST;
 306                 }
 307         } else {
 308                 if (hash_matches && bucket->nr == trace_nr &&
 309                     memcmp(bucket->data, ips, trace_len) == 0)
 310                         return id;
 311                 if (bucket && !(flags & BPF_F_REUSE_STACKID))
 312                         return -EEXIST;
 313
 314                 new_bucket = (struct stack_map_bucket *)
 315                         pcpu_freelist_pop(&smap->freelist);
 316                 if (unlikely(!new_bucket))
 317                         return -ENOMEM;
 318                 memcpy(new_bucket->data, ips, trace_len);
 319         }
 320
 321         new_bucket->hash = hash;
 322         new_bucket->nr = trace_nr;
 323
 324         old_bucket = xchg(&smap->buckets[id], new_bucket);
 325         if (old_bucket)
 326                 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
 327         return id;
 328 }
 329
 330 BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
 331            u64, flags)
 332 {
 333         u32 max_depth = map->value_size / stack_map_data_size(map);
 334         /* stack_map_alloc() checks that max_depth <= sysctl_perf_event_max_stack */
 335         u32 init_nr = sysctl_perf_event_max_stack - max_depth;
 336         bool user = flags & BPF_F_USER_STACK;
 337         struct perf_callchain_entry *trace;
 338         bool kernel = !user;
 339
 340         if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
 341                                BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
 342                 return -EINVAL;
 343
 344         trace = get_perf_callchain(regs, init_nr, kernel, user,
 345                                    sysctl_perf_event_max_stack, false, false);
 346
 347         if (unlikely(!trace))
 348                 /* couldn't fetch the stack trace */
 349                 return -EFAULT;
 350
 351         return __bpf_get_stackid(map, trace, flags);
 352 }
 353
 354 const struct bpf_func_proto bpf_get_stackid_proto = {
 355         .func           = bpf_get_stackid,
 356         .gpl_only       = true,
 357         .ret_type       = RET_INTEGER,
 358         .arg1_type      = ARG_PTR_TO_CTX,
 359         .arg2_type      = ARG_CONST_MAP_PTR,
 360         .arg3_type      = ARG_ANYTHING,
 361 };
 362
 363 static __u64 count_kernel_ip(struct perf_callchain_entry *trace)
 364 {
 365         __u64 nr_kernel = 0;
 366
 367         while (nr_kernel < trace->nr) {
 368                 if (trace->ip[nr_kernel] == PERF_CONTEXT_USER)
 369                         break;
 370                 nr_kernel++;
 371         }
 372         return nr_kernel;
 373 }
 374
 375 BPF_CALL_3(bpf_get_stackid_pe, struct bpf_perf_event_data_kern *, ctx,
 376            struct bpf_map *, map, u64, flags)
 377 {
 378         struct perf_event *event = ctx->event;
 379         struct perf_callchain_entry *trace;
 380         bool kernel, user;
 381         __u64 nr_kernel;
 382         int ret;
 383
 384         /* perf_sample_data doesn't have callchain, use bpf_get_stackid */
 385         if (!(event->attr.sample_type & __PERF_SAMPLE_CALLCHAIN_EARLY))
 386                 return bpf_get_stackid((unsigned long)(ctx->regs),
 387                                        (unsigned long) map, flags, 0, 0);
 388
 389         if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
 390                                BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
 391                 return -EINVAL;
 392
 393         user = flags & BPF_F_USER_STACK;
 394         kernel = !user;
 395
 396         trace = ctx->data->callchain;
 397         if (unlikely(!trace))
 398                 return -EFAULT;
 399
 400         nr_kernel = count_kernel_ip(trace);
 401
 402         if (kernel) {
 403                 __u64 nr = trace->nr;
 404
 405                 trace->nr = nr_kernel;
 406                 ret = __bpf_get_stackid(map, trace, flags);
 407
 408                 /* restore nr */
 409                 trace->nr = nr;
 410         } else { /* user */
 411                 u64 skip = flags & BPF_F_SKIP_FIELD_MASK;
 412
 413                 skip += nr_kernel;
 414                 if (skip > BPF_F_SKIP_FIELD_MASK)
 415                         return -EFAULT;
 416
 417                 flags = (flags & ~BPF_F_SKIP_FIELD_MASK) | skip;
 418                 ret = __bpf_get_stackid(map, trace, flags);
 419         }
 420         return ret;
 421 }
 422
 423 const struct bpf_func_proto bpf_get_stackid_proto_pe = {
 424         .func           = bpf_get_stackid_pe,
 425         .gpl_only       = false,
 426         .ret_type       = RET_INTEGER,
 427         .arg1_type      = ARG_PTR_TO_CTX,
 428         .arg2_type      = ARG_CONST_MAP_PTR,
 429         .arg3_type      = ARG_ANYTHING,
 430 };
 431
 432 static long __bpf_get_stack(struct pt_regs *regs, struct task_struct *task,
 433                             struct perf_callchain_entry *trace_in,
 434                             void *buf, u32 size, u64 flags)
 435 {
 436         u32 init_nr, trace_nr, copy_len, elem_size, num_elem;
 437         bool user_build_id = flags & BPF_F_USER_BUILD_ID;
 438         u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
 439         bool user = flags & BPF_F_USER_STACK;
 440         struct perf_callchain_entry *trace;
 441         bool kernel = !user;
 442         int err = -EINVAL;
 443         u64 *ips;
 444
 445         if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
 446                                BPF_F_USER_BUILD_ID)))
 447                 goto clear;
 448         if (kernel && user_build_id)
 449                 goto clear;
 450
 451         elem_size = (user && user_build_id) ? sizeof(struct bpf_stack_build_id)
 452                                             : sizeof(u64);
 453         if (unlikely(size % elem_size))
 454                 goto clear;
 455
 456         /* cannot get valid user stack for task without user_mode regs */
 457         if (task && user && !user_mode(regs))
 458                 goto err_fault;
 459
 460         num_elem = size / elem_size;
 461         if (sysctl_perf_event_max_stack < num_elem)
 462                 init_nr = 0;
 463         else
 464                 init_nr = sysctl_perf_event_max_stack - num_elem;
 465
 466         if (trace_in)
 467                 trace = trace_in;
 468         else if (kernel && task)
 469                 trace = get_callchain_entry_for_task(task, init_nr);
 470         else
 471                 trace = get_perf_callchain(regs, init_nr, kernel, user,
 472                                            sysctl_perf_event_max_stack,
 473                                            false, false);
 474         if (unlikely(!trace))
 475                 goto err_fault;
 476
 477         trace_nr = trace->nr - init_nr;
 478         if (trace_nr < skip)
 479                 goto err_fault;
 480
 481         trace_nr -= skip;
 482         trace_nr = (trace_nr <= num_elem) ? trace_nr : num_elem;
 483         copy_len = trace_nr * elem_size;
 484         ips = trace->ip + skip + init_nr;
 485         if (user && user_build_id)
 486                 stack_map_get_build_id_offset(buf, ips, trace_nr, user);
 487         else
 488                 memcpy(buf, ips, copy_len);
 489
 490         if (size > copy_len)
 491                 memset(buf + copy_len, 0, size - copy_len);
 492         return copy_len;
 493
 494 err_fault:
 495         err = -EFAULT;
 496 clear:
 497         memset(buf, 0, size);
 498         return err;
 499 }
 500
 501 BPF_CALL_4(bpf_get_stack, struct pt_regs *, regs, void *, buf, u32, size,
 502            u64, flags)
 503 {
 504         return __bpf_get_stack(regs, NULL, NULL, buf, size, flags);
 505 }
 506
 507 const struct bpf_func_proto bpf_get_stack_proto = {
 508         .func           = bpf_get_stack,
 509         .gpl_only       = true,
 510         .ret_type       = RET_INTEGER,
 511         .arg1_type      = ARG_PTR_TO_CTX,
 512         .arg2_type      = ARG_PTR_TO_UNINIT_MEM,
 513         .arg3_type      = ARG_CONST_SIZE_OR_ZERO,
 514         .arg4_type      = ARG_ANYTHING,
 515 };
 516
 517 BPF_CALL_4(bpf_get_task_stack, struct task_struct *, task, void *, buf,
 518            u32, size, u64, flags)
 519 {
 520         struct pt_regs *regs;
 521         long res;
 522
 523         if (!try_get_task_stack(task))
 524                 return -EFAULT;
 525
 526         regs = task_pt_regs(task);
 527         res = __bpf_get_stack(regs, task, NULL, buf, size, flags);
 528         put_task_stack(task);
 529
 530         return res;
 531 }
 532
 533 const struct bpf_func_proto bpf_get_task_stack_proto = {
 534         .func           = bpf_get_task_stack,
 535         .gpl_only       = false,
 536         .ret_type       = RET_INTEGER,
 537         .arg1_type      = ARG_PTR_TO_BTF_ID,
 538         .arg1_btf_id    = &btf_task_struct_ids[0],
 539         .arg2_type      = ARG_PTR_TO_UNINIT_MEM,
 540         .arg3_type      = ARG_CONST_SIZE_OR_ZERO,
 541         .arg4_type      = ARG_ANYTHING,
 542 };
 543
 544 BPF_CALL_4(bpf_get_stack_pe, struct bpf_perf_event_data_kern *, ctx,
 545            void *, buf, u32, size, u64, flags)
 546 {
 547         struct pt_regs *regs = (struct pt_regs *)(ctx->regs);
 548         struct perf_event *event = ctx->event;
 549         struct perf_callchain_entry *trace;
 550         bool kernel, user;
 551         int err = -EINVAL;
 552         __u64 nr_kernel;
 553
 554         if (!(event->attr.sample_type & __PERF_SAMPLE_CALLCHAIN_EARLY))
 555                 return __bpf_get_stack(regs, NULL, NULL, buf, size, flags);
 556
 557         if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
 558                                BPF_F_USER_BUILD_ID)))
 559                 goto clear;
 560
 561         user = flags & BPF_F_USER_STACK;
 562         kernel = !user;
 563
 564         err = -EFAULT;
 565         trace = ctx->data->callchain;
 566         if (unlikely(!trace))
 567                 goto clear;
 568
 569         nr_kernel = count_kernel_ip(trace);
 570
 571         if (kernel) {
 572                 __u64 nr = trace->nr;
 573
 574                 trace->nr = nr_kernel;
 575                 err = __bpf_get_stack(regs, NULL, trace, buf, size, flags);
 576
 577                 /* restore nr */
 578                 trace->nr = nr;
 579         } else { /* user */
 580                 u64 skip = flags & BPF_F_SKIP_FIELD_MASK;
 581
 582                 skip += nr_kernel;
 583                 if (skip > BPF_F_SKIP_FIELD_MASK)
 584                         goto clear;
 585
 586                 flags = (flags & ~BPF_F_SKIP_FIELD_MASK) | skip;
 587                 err = __bpf_get_stack(regs, NULL, trace, buf, size, flags);
 588         }
 589         return err;
 590
 591 clear:
 592         memset(buf, 0, size);
 593         return err;
 594
 595 }
 596
 597 const struct bpf_func_proto bpf_get_stack_proto_pe = {
 598         .func           = bpf_get_stack_pe,
 599         .gpl_only       = true,
 600         .ret_type       = RET_INTEGER,
 601         .arg1_type      = ARG_PTR_TO_CTX,
 602         .arg2_type      = ARG_PTR_TO_UNINIT_MEM,
 603         .arg3_type      = ARG_CONST_SIZE_OR_ZERO,
 604         .arg4_type      = ARG_ANYTHING,
 605 };
 606
 607 /* Called from eBPF program */
 608 static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
 609 {
 610         return ERR_PTR(-EOPNOTSUPP);
 611 }
 612
 613 /* Called from syscall */
 614 int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value)
 615 {
 616         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 617         struct stack_map_bucket *bucket, *old_bucket;
 618         u32 id = *(u32 *)key, trace_len;
 619
 620         if (unlikely(id >= smap->n_buckets))
 621                 return -ENOENT;
 622
 623         bucket = xchg(&smap->buckets[id], NULL);
 624         if (!bucket)
 625                 return -ENOENT;
 626
 627         trace_len = bucket->nr * stack_map_data_size(map);
 628         memcpy(value, bucket->data, trace_len);
 629         memset(value + trace_len, 0, map->value_size - trace_len);
 630
 631         old_bucket = xchg(&smap->buckets[id], bucket);
 632         if (old_bucket)
 633                 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
 634         return 0;
 635 }
 636
 637 static int stack_map_get_next_key(struct bpf_map *map, void *key,
 638                                   void *next_key)
 639 {
 640         struct bpf_stack_map *smap = container_of(map,
 641                                                   struct bpf_stack_map, map);
 642         u32 id;
 643
 644         WARN_ON_ONCE(!rcu_read_lock_held());
 645
 646         if (!key) {
 647                 id = 0;
 648         } else {
 649                 id = *(u32 *)key;
 650                 if (id >= smap->n_buckets || !smap->buckets[id])
 651                         id = 0;
 652                 else
 653                         id++;
 654         }
 655
 656         while (id < smap->n_buckets && !smap->buckets[id])
 657                 id++;
 658
 659         if (id >= smap->n_buckets)
 660                 return -ENOENT;
 661
 662         *(u32 *)next_key = id;
 663         return 0;
 664 }
 665
 666 static int stack_map_update_elem(struct bpf_map *map, void *key, void *value,
 667                                  u64 map_flags)
 668 {
 669         return -EINVAL;
 670 }
 671
 672 /* Called from syscall or from eBPF program */
 673 static int stack_map_delete_elem(struct bpf_map *map, void *key)
 674 {
 675         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 676         struct stack_map_bucket *old_bucket;
 677         u32 id = *(u32 *)key;
 678
 679         if (unlikely(id >= smap->n_buckets))
 680                 return -E2BIG;
 681
 682         old_bucket = xchg(&smap->buckets[id], NULL);
 683         if (old_bucket) {
 684                 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
 685                 return 0;
 686         } else {
 687                 return -ENOENT;
 688         }
 689 }
 690
 691 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
 692 static void stack_map_free(struct bpf_map *map)
 693 {
 694         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 695
 696         bpf_map_area_free(smap->elems);
 697         pcpu_freelist_destroy(&smap->freelist);
 698         bpf_map_area_free(smap);
 699         put_callchain_buffers();
 700 }
 701
 702 static int stack_trace_map_btf_id;
 703 const struct bpf_map_ops stack_trace_map_ops = {
 704         .map_meta_equal = bpf_map_meta_equal,
 705         .map_alloc = stack_map_alloc,
 706         .map_free = stack_map_free,
 707         .map_get_next_key = stack_map_get_next_key,
 708         .map_lookup_elem = stack_map_lookup_elem,
 709         .map_update_elem = stack_map_update_elem,
 710         .map_delete_elem = stack_map_delete_elem,
 711         .map_check_btf = map_check_no_btf,
 712         .map_btf_name = "bpf_stack_map",
 713         .map_btf_id = &stack_trace_map_btf_id,
 714 };
 715
 716 static int __init stack_map_init(void)
 717 {
 718         int cpu;
 719         struct stack_map_irq_work *work;
 720
 721         for_each_possible_cpu(cpu) {
 722                 work = per_cpu_ptr(&up_read_work, cpu);
 723                 init_irq_work(&work->irq_work, do_up_read);
 724         }
 725         return 0;
 726 }
 727 subsys_initcall(stack_map_init);