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/vmalloc.h>
  11 #include <linux/stacktrace.h>
  12 #include <linux/perf_event.h>
  13 #include "percpu_freelist.h"
  14
  15 struct stack_map_bucket {
  16         struct pcpu_freelist_node fnode;
  17         u32 hash;
  18         u32 nr;
  19         u64 ip[];
  20 };
  21
  22 struct bpf_stack_map {
  23         struct bpf_map map;
  24         void *elems;
  25         struct pcpu_freelist freelist;
  26         u32 n_buckets;
  27         struct stack_map_bucket *buckets[];
  28 };
  29
  30 static int prealloc_elems_and_freelist(struct bpf_stack_map *smap)
  31 {
  32         u32 elem_size = sizeof(struct stack_map_bucket) + smap->map.value_size;
  33         int err;
  34
  35         smap->elems = vzalloc(elem_size * smap->map.max_entries);
  36         if (!smap->elems)
  37                 return -ENOMEM;
  38
  39         err = pcpu_freelist_init(&smap->freelist);
  40         if (err)
  41                 goto free_elems;
  42
  43         pcpu_freelist_populate(&smap->freelist, smap->elems, elem_size,
  44                                smap->map.max_entries);
  45         return 0;
  46
  47 free_elems:
  48         vfree(smap->elems);
  49         return err;
  50 }
  51
  52 /* Called from syscall */
  53 static struct bpf_map *stack_map_alloc(union bpf_attr *attr)
  54 {
  55         u32 value_size = attr->value_size;
  56         struct bpf_stack_map *smap;
  57         u64 cost, n_buckets;
  58         int err;
  59
  60         if (!capable(CAP_SYS_ADMIN))
  61                 return ERR_PTR(-EPERM);
  62
  63         if (attr->map_flags)
  64                 return ERR_PTR(-EINVAL);
  65
  66         /* check sanity of attributes */
  67         if (attr->max_entries == 0 || attr->key_size != 4 ||
  68             value_size < 8 || value_size % 8 ||
  69             value_size / 8 > sysctl_perf_event_max_stack)
  70                 return ERR_PTR(-EINVAL);
  71
  72         /* hash table size must be power of 2 */
  73         n_buckets = roundup_pow_of_two(attr->max_entries);
  74
  75         cost = n_buckets * sizeof(struct stack_map_bucket *) + sizeof(*smap);
  76         if (cost >= U32_MAX - PAGE_SIZE)
  77                 return ERR_PTR(-E2BIG);
  78
  79         smap = kzalloc(cost, GFP_USER | __GFP_NOWARN);
  80         if (!smap) {
  81                 smap = vzalloc(cost);
  82                 if (!smap)
  83                         return ERR_PTR(-ENOMEM);
  84         }
  85
  86         err = -E2BIG;
  87         cost += n_buckets * (value_size + sizeof(struct stack_map_bucket));
  88         if (cost >= U32_MAX - PAGE_SIZE)
  89                 goto free_smap;
  90
  91         smap->map.map_type = attr->map_type;
  92         smap->map.key_size = attr->key_size;
  93         smap->map.value_size = value_size;
  94         smap->map.max_entries = attr->max_entries;
  95         smap->n_buckets = n_buckets;
  96         smap->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
  97
  98         err = bpf_map_precharge_memlock(smap->map.pages);
  99         if (err)
 100                 goto free_smap;
 101
 102         err = get_callchain_buffers();
 103         if (err)
 104                 goto free_smap;
 105
 106         err = prealloc_elems_and_freelist(smap);
 107         if (err)
 108                 goto put_buffers;
 109
 110         return &smap->map;
 111
 112 put_buffers:
 113         put_callchain_buffers();
 114 free_smap:
 115         kvfree(smap);
 116         return ERR_PTR(err);
 117 }
 118
 119 u64 bpf_get_stackid(u64 r1, u64 r2, u64 flags, u64 r4, u64 r5)
 120 {
 121         struct pt_regs *regs = (struct pt_regs *) (long) r1;
 122         struct bpf_map *map = (struct bpf_map *) (long) r2;
 123         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 124         struct perf_callchain_entry *trace;
 125         struct stack_map_bucket *bucket, *new_bucket, *old_bucket;
 126         u32 max_depth = map->value_size / 8;
 127         /* stack_map_alloc() checks that max_depth <= sysctl_perf_event_max_stack */
 128         u32 init_nr = sysctl_perf_event_max_stack - max_depth;
 129         u32 skip = flags & BPF_F_SKIP_FIELD_MASK;
 130         u32 hash, id, trace_nr, trace_len;
 131         bool user = flags & BPF_F_USER_STACK;
 132         bool kernel = !user;
 133         u64 *ips;
 134
 135         if (unlikely(flags & ~(BPF_F_SKIP_FIELD_MASK | BPF_F_USER_STACK |
 136                                BPF_F_FAST_STACK_CMP | BPF_F_REUSE_STACKID)))
 137                 return -EINVAL;
 138
 139         trace = get_perf_callchain(regs, init_nr, kernel, user,
 140                                    sysctl_perf_event_max_stack, false, false);
 141
 142         if (unlikely(!trace))
 143                 /* couldn't fetch the stack trace */
 144                 return -EFAULT;
 145
 146         /* get_perf_callchain() guarantees that trace->nr >= init_nr
 147          * and trace-nr <= sysctl_perf_event_max_stack, so trace_nr <= max_depth
 148          */
 149         trace_nr = trace->nr - init_nr;
 150
 151         if (trace_nr <= skip)
 152                 /* skipping more than usable stack trace */
 153                 return -EFAULT;
 154
 155         trace_nr -= skip;
 156         trace_len = trace_nr * sizeof(u64);
 157         ips = trace->ip + skip + init_nr;
 158         hash = jhash2((u32 *)ips, trace_len / sizeof(u32), 0);
 159         id = hash & (smap->n_buckets - 1);
 160         bucket = READ_ONCE(smap->buckets[id]);
 161
 162         if (bucket && bucket->hash == hash) {
 163                 if (flags & BPF_F_FAST_STACK_CMP)
 164                         return id;
 165                 if (bucket->nr == trace_nr &&
 166                     memcmp(bucket->ip, ips, trace_len) == 0)
 167                         return id;
 168         }
 169
 170         /* this call stack is not in the map, try to add it */
 171         if (bucket && !(flags & BPF_F_REUSE_STACKID))
 172                 return -EEXIST;
 173
 174         new_bucket = (struct stack_map_bucket *)
 175                 pcpu_freelist_pop(&smap->freelist);
 176         if (unlikely(!new_bucket))
 177                 return -ENOMEM;
 178
 179         memcpy(new_bucket->ip, ips, trace_len);
 180         new_bucket->hash = hash;
 181         new_bucket->nr = trace_nr;
 182
 183         old_bucket = xchg(&smap->buckets[id], new_bucket);
 184         if (old_bucket)
 185                 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
 186         return id;
 187 }
 188
 189 const struct bpf_func_proto bpf_get_stackid_proto = {
 190         .func           = bpf_get_stackid,
 191         .gpl_only       = true,
 192         .ret_type       = RET_INTEGER,
 193         .arg1_type      = ARG_PTR_TO_CTX,
 194         .arg2_type      = ARG_CONST_MAP_PTR,
 195         .arg3_type      = ARG_ANYTHING,
 196 };
 197
 198 /* Called from eBPF program */
 199 static void *stack_map_lookup_elem(struct bpf_map *map, void *key)
 200 {
 201         return NULL;
 202 }
 203
 204 /* Called from syscall */
 205 int bpf_stackmap_copy(struct bpf_map *map, void *key, void *value)
 206 {
 207         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 208         struct stack_map_bucket *bucket, *old_bucket;
 209         u32 id = *(u32 *)key, trace_len;
 210
 211         if (unlikely(id >= smap->n_buckets))
 212                 return -ENOENT;
 213
 214         bucket = xchg(&smap->buckets[id], NULL);
 215         if (!bucket)
 216                 return -ENOENT;
 217
 218         trace_len = bucket->nr * sizeof(u64);
 219         memcpy(value, bucket->ip, trace_len);
 220         memset(value + trace_len, 0, map->value_size - trace_len);
 221
 222         old_bucket = xchg(&smap->buckets[id], bucket);
 223         if (old_bucket)
 224                 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
 225         return 0;
 226 }
 227
 228 static int stack_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 229 {
 230         return -EINVAL;
 231 }
 232
 233 static int stack_map_update_elem(struct bpf_map *map, void *key, void *value,
 234                                  u64 map_flags)
 235 {
 236         return -EINVAL;
 237 }
 238
 239 /* Called from syscall or from eBPF program */
 240 static int stack_map_delete_elem(struct bpf_map *map, void *key)
 241 {
 242         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 243         struct stack_map_bucket *old_bucket;
 244         u32 id = *(u32 *)key;
 245
 246         if (unlikely(id >= smap->n_buckets))
 247                 return -E2BIG;
 248
 249         old_bucket = xchg(&smap->buckets[id], NULL);
 250         if (old_bucket) {
 251                 pcpu_freelist_push(&smap->freelist, &old_bucket->fnode);
 252                 return 0;
 253         } else {
 254                 return -ENOENT;
 255         }
 256 }
 257
 258 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
 259 static void stack_map_free(struct bpf_map *map)
 260 {
 261         struct bpf_stack_map *smap = container_of(map, struct bpf_stack_map, map);
 262
 263         /* wait for bpf programs to complete before freeing stack map */
 264         synchronize_rcu();
 265
 266         vfree(smap->elems);
 267         pcpu_freelist_destroy(&smap->freelist);
 268         kvfree(smap);
 269         put_callchain_buffers();
 270 }
 271
 272 static const struct bpf_map_ops stack_map_ops = {
 273         .map_alloc = stack_map_alloc,
 274         .map_free = stack_map_free,
 275         .map_get_next_key = stack_map_get_next_key,
 276         .map_lookup_elem = stack_map_lookup_elem,
 277         .map_update_elem = stack_map_update_elem,
 278         .map_delete_elem = stack_map_delete_elem,
 279 };
 280
 281 static struct bpf_map_type_list stack_map_type __read_mostly = {
 282         .ops = &stack_map_ops,
 283         .type = BPF_MAP_TYPE_STACK_TRACE,
 284 };
 285
 286 static int __init register_stack_map(void)
 287 {
 288         bpf_register_map_type(&stack_map_type);
 289         return 0;
 290 }
 291 late_initcall(register_stack_map);