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
  74         cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap);
  75         if (cost >= U32_MAX - PAGE_SIZE)
  76                 return ERR_PTR(-E2BIG);
  77
  78         smap = bpf_map_area_alloc(cost);
  79         if (!smap)
  80                 return ERR_PTR(-ENOMEM);
  81
  82         err = -E2BIG;
  83         cost += n_buckets * (value_size + sizeof(struct stack_map_bucket));
  84         if (cost >= U32_MAX - PAGE_SIZE)
  85                 goto free_smap;
  86
  87         smap->map.map_type = attr->map_type;
  88         smap->map.key_size = attr->key_size;
  89         smap->map.value_size = value_size;
  90         smap->map.max_entries = attr->max_entries;
  91         smap->map.map_flags = attr->map_flags;
  92         smap->n_buckets = n_buckets;
  93         smap->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
  94
  95         err = bpf_map_precharge_memlock(smap->map.pages);
  96         if (err)
  97                 goto free_smap;
  98
  99         err = get_callchain_buffers(sysctl_perf_event_max_stack);
 100         if (err)
 101                 goto free_smap;
 102
 103         err = prealloc_elems_and_freelist(smap);
 104         if (err)
 105                 goto put_buffers;
 106
 107         return &smap->map;
 108
 109 put_buffers:
 110         put_callchain_buffers();
 111 free_smap:
 112         bpf_map_area_free(smap);
 113         return ERR_PTR(err);
 114 }
 115
 116 BPF_CALL_3(bpf_get_stackid, struct pt_regs *, regs, struct bpf_map *, map,
 117            u64, flags)
 118 {
 119         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 120         struct perf_callchain_entry *trace;
 121         struct stack_map_bucket *bucket, *new_bucket, *old_bucket;
 122         u32 max_depth = map->value_size / 8;
 123         /* stack_map_alloc() checks that max_depth <= sysctl_perf_event_max_stack */
 124         u32 init_nr = sysctl_perf_event_max_stack - max_depth;
 125         u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
 126         u32 hash, id, trace_nr, trace_len;
 127         bool user = flags & BPF_F_USER_STACK;
 128         bool kernel = !user;
 129         u64 *ips;
 130
 131         if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
 132                                BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
 133                 return -EINVAL;
 134
 135         trace = get_perf_callchain(regs, init_nr, kernel, user,
 136                                    sysctl_perf_event_max_stack, false, false);
 137
 138         if (unlikely(!trace))
 139                 /* couldn't fetch the stack trace */
 140                 return -EFAULT;
 141
 142         /* get_perf_callchain() guarantees that trace->nr >= init_nr
 143          * and trace-nr <= sysctl_perf_event_max_stack, so trace_nr <= max_depth
 144          */
 145         trace_nr = trace->nr - init_nr;
 146
 147         if (trace_nr <= skip)
 148                 /* skipping more than usable stack trace */
 149                 return -EFAULT;
 150
 151         trace_nr -= skip;
 152         trace_len = trace_nr * sizeof(u64);
 153         ips = trace->ip + skip + init_nr;
 154         hash = jhash2((u32 *)ips, trace_len / sizeof(u32), 0);
 155         id = hash & (smap->n_buckets - 1);
 156         bucket = READ_ONCE(smap->buckets[id]);
 157
 158         if (bucket && bucket->hash == hash) {
 159                 if (flags & BPF_F_FAST_STACK_CMP)
 160                         return id;
 161                 if (bucket->nr == trace_nr &&
 162                     memcmp(bucket->ip, ips, trace_len) == 0)
 163                         return id;
 164         }
 165
 166         /* this call stack is not in the map, try to add it */
 167         if (bucket && !(flags & BPF_F_REUSE_STACKID))
 168                 return -EEXIST;
 169
 170         new_bucket = (struct stack_map_bucket *)
 171                 pcpu_freelist_pop(&smap->freelist);
 172         if (unlikely(!new_bucket))
 173                 return -ENOMEM;
 174
 175         memcpy(new_bucket->ip, ips, trace_len);
 176         new_bucket->hash = hash;
 177         new_bucket->nr = trace_nr;
 178
 179         old_bucket = xchg(&smap->buckets[id], new_bucket);
 180         if (old_bucket)
 181                 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
 182         return id;
 183 }
 184
 185 const struct bpf_func_proto bpf_get_stackid_proto = {
 186         .func           = bpf_get_stackid,
 187         .gpl_only       = true,
 188         .ret_type       = RET_INTEGER,
 189         .arg1_type      = ARG_PTR_TO_CTX,
 190         .arg2_type      = ARG_CONST_MAP_PTR,
 191         .arg3_type      = ARG_ANYTHING,
 192 };
 193
 194 /* Called from eBPF program */
 195 static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
 196 {
 197         return NULL;
 198 }
 199
 200 /* Called from syscall */
 201 int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value)
 202 {
 203         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 204         struct stack_map_bucket *bucket, *old_bucket;
 205         u32 id = *(u32 *)key, trace_len;
 206
 207         if (unlikely(id >= smap->n_buckets))
 208                 return -ENOENT;
 209
 210         bucket = xchg(&smap->buckets[id], NULL);
 211         if (!bucket)
 212                 return -ENOENT;
 213
 214         trace_len = bucket->nr * sizeof(u64);
 215         memcpy(value, bucket->ip, trace_len);
 216         memset(value + trace_len, 0, map->value_size - trace_len);
 217
 218         old_bucket = xchg(&smap->buckets[id], bucket);
 219         if (old_bucket)
 220                 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
 221         return 0;
 222 }
 223
 224 static int stack_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 225 {
 226         return -EINVAL;
 227 }
 228
 229 static int stack_map_update_elem(struct bpf_map *map, void *key, void *value,
 230                                  u64 map_flags)
 231 {
 232         return -EINVAL;
 233 }
 234
 235 /* Called from syscall or from eBPF program */
 236 static int stack_map_delete_elem(struct bpf_map *map, void *key)
 237 {
 238         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 239         struct stack_map_bucket *old_bucket;
 240         u32 id = *(u32 *)key;
 241
 242         if (unlikely(id >= smap->n_buckets))
 243                 return -E2BIG;
 244
 245         old_bucket = xchg(&smap->buckets[id], NULL);
 246         if (old_bucket) {
 247                 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
 248                 return 0;
 249         } else {
 250                 return -ENOENT;
 251         }
 252 }
 253
 254 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
 255 static void stack_map_free(struct bpf_map *map)
 256 {
 257         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 258
 259         /* wait for bpf programs to complete before freeing stack map */
 260         synchronize_rcu();
 261
 262         bpf_map_area_free(smap->elems);
 263         pcpu_freelist_destroy(&smap->freelist);
 264         bpf_map_area_free(smap);
 265         put_callchain_buffers();
 266 }
 267
 268 static const struct bpf_map_ops stack_map_ops = {
 269         .map_alloc = stack_map_alloc,
 270         .map_free = stack_map_free,
 271         .map_get_next_key = stack_map_get_next_key,
 272         .map_lookup_elem = stack_map_lookup_elem,
 273         .map_update_elem = stack_map_update_elem,
 274         .map_delete_elem = stack_map_delete_elem,
 275 };
 276
 277 static struct bpf_map_type_list stack_map_type __read_mostly = {
 278         .ops = &stack_map_ops,
 279         .type = BPF_MAP_TYPE_STACK_TRACE,
 280 };
 281
 282 static int __init register_stack_map(void)
 283 {
 284         bpf_register_map_type(&stack_map_type);
 285         return 0;
 286 }
 287 late_initcall(register_stack_map);