kernel/bpf/stackmap.c

   1 /* Copyright (c) 2016 Facebook
   2  *
   3  * This program is free software; you can redistribute it and/or
   4  * modify it under the terms of version 2 of the GNU General Public
   5  * License as published by the Free Software Foundation.
   6  */
   7 #include <linux/bpf.h>
   8 #include <linux/jhash.h>
   9 #include <linux/filter.h>
  10 #include <linux/stacktrace.h>
  11 #include <linux/perf_event.h>
  12 #include "percpu_freelist.h"
  13
  14 struct stack_map_bucket {
  15         struct pcpu_freelist_node fnode;
  16         u32 hash;
  17         u32 nr;
  18         u64 ip[];
  19 };
  20
  21 struct bpf_stack_map {
  22         struct bpf_map map;
  23         void *elems;
  24         struct pcpu_freelist freelist;
  25         u32 n_buckets;
  26         struct stack_map_bucket *buckets[];
  27 };
  28
  29 static int prealloc_elems_and_freelist(struct bpf_stack_map *smap)
  30 {
  31         u32 elem_size = sizeof(struct stack_map_bucket) + smap->map.value_size;
  32         int err;
  33
  34         smap->elems = bpf_map_area_alloc(elem_size * smap->map.max_entries);
  35         if (!smap->elems)
  36                 return -ENOMEM;
  37
  38         err = pcpu_freelist_init(&smap->freelist);
  39         if (err)
  40                 goto free_elems;
  41
  42         pcpu_freelist_populate(&smap->freelist, smap->elems, elem_size,
  43                                smap->map.max_entries);
  44         return 0;
  45
  46 free_elems:
  47         bpf_map_area_free(smap->elems);
  48         return err;
  49 }
  50
  51 /* Called from syscall */
  52 static struct bpf_map *stack_map_alloc(union bpf_attr *attr)
  53 {
  54         u32 value_size = attr->value_size;
  55         struct bpf_stack_map *smap;
  56         u64 cost, n_buckets;
  57         int err;
  58
  59         if (!capable(CAP_SYS_ADMIN))
  60                 return ERR_PTR(-EPERM);
  61
  62         if (attr->map_flags)
  63                 return ERR_PTR(-EINVAL);
  64
  65         /* check sanity of attributes */
  66         if (attr->max_entries == 0 || attr->key_size != 4 ||
  67             value_size < 8 || value_size % 8 ||
  68             value_size / 8 > sysctl_perf_event_max_stack)
  69                 return ERR_PTR(-EINVAL);
  70
  71         /* hash table size must be power of 2 */
  72         n_buckets = roundup_pow_of_two(attr->max_entries);
  73         if (!n_buckets)
  74                 return ERR_PTR(-E2BIG);
  75
  76         cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap);
  77         if (cost >= U32_MAX - PAGE_SIZE)
  78                 return ERR_PTR(-E2BIG);
  79
  80         smap = bpf_map_area_alloc(cost);
  81         if (!smap)
  82                 return ERR_PTR(-ENOMEM);
  83
  84         err = -E2BIG;
  85         cost += n_buckets * (value_size + sizeof(struct stack_map_bucket));
  86         if (cost >= U32_MAX - PAGE_SIZE)
  87                 goto free_smap;
  88
  89         smap->map.map_type = attr->map_type;
  90         smap->map.key_size = attr->key_size;
  91         smap->map.value_size = value_size;
  92         smap->map.max_entries = attr->max_entries;
  93         smap->map.map_flags = attr->map_flags;
  94         smap->n_buckets = n_buckets;
  95         smap->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
  96
  97         err = bpf_map_precharge_memlock(smap->map.pages);
  98         if (err)
  99                 goto free_smap;
 100
 101         err = get_callchain_buffers(sysctl_perf_event_max_stack);
 102         if (err)
 103                 goto free_smap;
 104
 105         err = prealloc_elems_and_freelist(smap);
 106         if (err)
 107                 goto put_buffers;
 108
 109         return &smap->map;
 110
 111 put_buffers:
 112         put_callchain_buffers();
 113 free_smap:
 114         bpf_map_area_free(smap);
 115         return ERR_PTR(err);
 116 }
 117
 118 BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
 119            u64, flags)
 120 {
 121         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 122         struct perf_callchain_entry *trace;
 123         struct stack_map_bucket *bucket, *new_bucket, *old_bucket;
 124         u32 max_depth = map->value_size / 8;
 125         /* stack_map_alloc() checks that max_depth <= sysctl_perf_event_max_stack */
 126         u32 init_nr = sysctl_perf_event_max_stack - max_depth;
 127         u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
 128         u32 hash, id, trace_nr, trace_len;
 129         bool user = flags & BPF_F_USER_STACK;
 130         bool kernel = !user;
 131         u64 *ips;
 132
 133         if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
 134                                BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
 135                 return -EINVAL;
 136
 137         trace = get_perf_callchain(regs, init_nr, kernel, user,
 138                                    sysctl_perf_event_max_stack, false, false);
 139
 140         if (unlikely(!trace))
 141                 /* couldn't fetch the stack trace */
 142                 return -EFAULT;
 143
 144         /* get_perf_callchain() guarantees that trace->nr >= init_nr
 145          * and trace-nr <= sysctl_perf_event_max_stack, so trace_nr <= max_depth
 146          */
 147         trace_nr = trace->nr - init_nr;
 148
 149         if (trace_nr <= skip)
 150                 /* skipping more than usable stack trace */
 151                 return -EFAULT;
 152
 153         trace_nr -= skip;
 154         trace_len = trace_nr * sizeof(u64);
 155         ips = trace->ip + skip + init_nr;
 156         hash = jhash2((u32 *)ips, trace_len / sizeof(u32), 0);
 157         id = hash & (smap->n_buckets - 1);
 158         bucket = READ_ONCE(smap->buckets[id]);
 159
 160         if (bucket && bucket->hash == hash) {
 161                 if (flags & BPF_F_FAST_STACK_CMP)
 162                         return id;
 163                 if (bucket->nr == trace_nr &&
 164                     memcmp(bucket->ip, ips, trace_len) == 0)
 165                         return id;
 166         }
 167
 168         /* this call stack is not in the map, try to add it */
 169         if (bucket && !(flags & BPF_F_REUSE_STACKID))
 170                 return -EEXIST;
 171
 172         new_bucket = (struct stack_map_bucket *)
 173                 pcpu_freelist_pop(&smap->freelist);
 174         if (unlikely(!new_bucket))
 175                 return -ENOMEM;
 176
 177         memcpy(new_bucket->ip, ips, trace_len);
 178         new_bucket->hash = hash;
 179         new_bucket->nr = trace_nr;
 180
 181         old_bucket = xchg(&smap->buckets[id], new_bucket);
 182         if (old_bucket)
 183                 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
 184         return id;
 185 }
 186
 187 const struct bpf_func_proto bpf_get_stackid_proto = {
 188         .func           = bpf_get_stackid,
 189         .gpl_only       = true,
 190         .ret_type       = RET_INTEGER,
 191         .arg1_type      = ARG_PTR_TO_CTX,
 192         .arg2_type      = ARG_CONST_MAP_PTR,
 193         .arg3_type      = ARG_ANYTHING,
 194 };
 195
 196 /* Called from eBPF program */
 197 static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
 198 {
 199         return NULL;
 200 }
 201
 202 /* Called from syscall */
 203 int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value)
 204 {
 205         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 206         struct stack_map_bucket *bucket, *old_bucket;
 207         u32 id = *(u32 *)key, trace_len;
 208
 209         if (unlikely(id >= smap->n_buckets))
 210                 return -ENOENT;
 211
 212         bucket = xchg(&smap->buckets[id], NULL);
 213         if (!bucket)
 214                 return -ENOENT;
 215
 216         trace_len = bucket->nr * sizeof(u64);
 217         memcpy(value, bucket->ip, trace_len);
 218         memset(value + trace_len, 0, map->value_size - trace_len);
 219
 220         old_bucket = xchg(&smap->buckets[id], bucket);
 221         if (old_bucket)
 222                 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
 223         return 0;
 224 }
 225
 226 static int stack_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 227 {
 228         return -EINVAL;
 229 }
 230
 231 static int stack_map_update_elem(struct bpf_map *map, void *key, void *value,
 232                                  u64 map_flags)
 233 {
 234         return -EINVAL;
 235 }
 236
 237 /* Called from syscall or from eBPF program */
 238 static int stack_map_delete_elem(struct bpf_map *map, void *key)
 239 {
 240         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 241         struct stack_map_bucket *old_bucket;
 242         u32 id = *(u32 *)key;
 243
 244         if (unlikely(id >= smap->n_buckets))
 245                 return -E2BIG;
 246
 247         old_bucket = xchg(&smap->buckets[id], NULL);
 248         if (old_bucket) {
 249                 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
 250                 return 0;
 251         } else {
 252                 return -ENOENT;
 253         }
 254 }
 255
 256 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
 257 static void stack_map_free(struct bpf_map *map)
 258 {
 259         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 260
 261         /* wait for bpf programs to complete before freeing stack map */
 262         synchronize_rcu();
 263
 264         bpf_map_area_free(smap->elems);
 265         pcpu_freelist_destroy(&smap->freelist);
 266         bpf_map_area_free(smap);
 267         put_callchain_buffers();
 268 }
 269
 270 static const struct bpf_map_ops stack_map_ops = {
 271         .map_alloc = stack_map_alloc,
 272         .map_free = stack_map_free,
 273         .map_get_next_key = stack_map_get_next_key,
 274         .map_lookup_elem = stack_map_lookup_elem,
 275         .map_update_elem = stack_map_update_elem,
 276         .map_delete_elem = stack_map_delete_elem,
 277 };
 278
 279 static struct bpf_map_type_list stack_map_type __read_mostly = {
 280         .ops = &stack_map_ops,
 281         .type = BPF_MAP_TYPE_STACK_TRACE,
 282 };
 283
 284 static int __init register_stack_map(void)
 285 {
 286         bpf_register_map_type(&stack_map_type);
 287         return 0;
 288 }
 289 late_initcall(register_stack_map);