kernel/bpf/queue_stack_maps.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * queue_stack_maps.c: BPF queue and stack maps
   4  *
   5  * Copyright (c) 2018 Politecnico di Torino
   6  */
   7 #include <linux/bpf.h>
   8 #include <linux/list.h>
   9 #include <linux/slab.h>
  10 #include <linux/capability.h>
  11 #include "percpu_freelist.h"
  12
  13 #define QUEUE_STACK_CREATE_FLAG_MASK \
  14         (BPF_F_NUMA_NODE | BPF_F_ACCESS_MASK)
  15
  16 struct bpf_queue_stack {
  17         struct bpf_map map;
  18         raw_spinlock_t lock;
  19         u32 head, tail;
  20         u32 size; /* max_entries + 1 */
  21
  22         char elements[0] __aligned(8);
  23 };
  24
  25 static struct bpf_queue_stack *bpf_queue_stack(struct bpf_map *map)
  26 {
  27         return container_of(map, struct bpf_queue_stack, map);
  28 }
  29
  30 static bool queue_stack_map_is_empty(struct bpf_queue_stack *qs)
  31 {
  32         return qs->head == qs->tail;
  33 }
  34
  35 static bool queue_stack_map_is_full(struct bpf_queue_stack *qs)
  36 {
  37         u32 head = qs->head + 1;
  38
  39         if (unlikely(head >= qs->size))
  40                 head = 0;
  41
  42         return head == qs->tail;
  43 }
  44
  45 /* Called from syscall */
  46 static int queue_stack_map_alloc_check(union bpf_attr *attr)
  47 {
  48         if (!capable(CAP_SYS_ADMIN))
  49                 return -EPERM;
  50
  51         /* check sanity of attributes */
  52         if (attr->max_entries == 0 || attr->key_size != 0 ||
  53             attr->value_size == 0 ||
  54             attr->map_flags & ~QUEUE_STACK_CREATE_FLAG_MASK ||
  55             !bpf_map_flags_access_ok(attr->map_flags))
  56                 return -EINVAL;
  57
  58         if (attr->value_size > KMALLOC_MAX_SIZE)
  59                 /* if value_size is bigger, the user space won't be able to
  60                  * access the elements.
  61                  */
  62                 return -E2BIG;
  63
  64         return 0;
  65 }
  66
  67 static struct bpf_map *queue_stack_map_alloc(union bpf_attr *attr)
  68 {
  69         int ret, numa_node = bpf_map_attr_numa_node(attr);
  70         struct bpf_queue_stack *qs;
  71         u64 size, queue_size, cost;
  72
  73         size = (u64) attr->max_entries + 1;
  74         cost = queue_size = sizeof(*qs) + size * attr->value_size;
  75         if (cost >= U32_MAX - PAGE_SIZE)
  76                 return ERR_PTR(-E2BIG);
  77
  78         cost = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
  79
  80         ret = bpf_map_precharge_memlock(cost);
  81         if (ret < 0)
  82                 return ERR_PTR(ret);
  83
  84         qs = bpf_map_area_alloc(queue_size, numa_node);
  85         if (!qs)
  86                 return ERR_PTR(-ENOMEM);
  87
  88         memset(qs, 0, sizeof(*qs));
  89
  90         bpf_map_init_from_attr(&qs->map, attr);
  91
  92         qs->map.pages = cost;
  93         qs->size = size;
  94
  95         raw_spin_lock_init(&qs->lock);
  96
  97         return &qs->map;
  98 }
  99
 100 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
 101 static void queue_stack_map_free(struct bpf_map *map)
 102 {
 103         struct bpf_queue_stack *qs = bpf_queue_stack(map);
 104
 105         /* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
 106          * so the programs (can be more than one that used this map) were
 107          * disconnected from events. Wait for outstanding critical sections in
 108          * these programs to complete
 109          */
 110         synchronize_rcu();
 111
 112         bpf_map_area_free(qs);
 113 }
 114
 115 static int __queue_map_get(struct bpf_map *map, void *value, bool delete)
 116 {
 117         struct bpf_queue_stack *qs = bpf_queue_stack(map);
 118         unsigned long flags;
 119         int err = 0;
 120         void *ptr;
 121
 122         raw_spin_lock_irqsave(&qs->lock, flags);
 123
 124         if (queue_stack_map_is_empty(qs)) {
 125                 memset(value, 0, qs->map.value_size);
 126                 err = -ENOENT;
 127                 goto out;
 128         }
 129
 130         ptr = &qs->elements[qs->tail * qs->map.value_size];
 131         memcpy(value, ptr, qs->map.value_size);
 132
 133         if (delete) {
 134                 if (unlikely(++qs->tail >= qs->size))
 135                         qs->tail = 0;
 136         }
 137
 138 out:
 139         raw_spin_unlock_irqrestore(&qs->lock, flags);
 140         return err;
 141 }
 142
 143
 144 static int __stack_map_get(struct bpf_map *map, void *value, bool delete)
 145 {
 146         struct bpf_queue_stack *qs = bpf_queue_stack(map);
 147         unsigned long flags;
 148         int err = 0;
 149         void *ptr;
 150         u32 index;
 151
 152         raw_spin_lock_irqsave(&qs->lock, flags);
 153
 154         if (queue_stack_map_is_empty(qs)) {
 155                 memset(value, 0, qs->map.value_size);
 156                 err = -ENOENT;
 157                 goto out;
 158         }
 159
 160         index = qs->head - 1;
 161         if (unlikely(index >= qs->size))
 162                 index = qs->size - 1;
 163
 164         ptr = &qs->elements[index * qs->map.value_size];
 165         memcpy(value, ptr, qs->map.value_size);
 166
 167         if (delete)
 168                 qs->head = index;
 169
 170 out:
 171         raw_spin_unlock_irqrestore(&qs->lock, flags);
 172         return err;
 173 }
 174
 175 /* Called from syscall or from eBPF program */
 176 static int queue_map_peek_elem(struct bpf_map *map, void *value)
 177 {
 178         return __queue_map_get(map, value, false);
 179 }
 180
 181 /* Called from syscall or from eBPF program */
 182 static int stack_map_peek_elem(struct bpf_map *map, void *value)
 183 {
 184         return __stack_map_get(map, value, false);
 185 }
 186
 187 /* Called from syscall or from eBPF program */
 188 static int queue_map_pop_elem(struct bpf_map *map, void *value)
 189 {
 190         return __queue_map_get(map, value, true);
 191 }
 192
 193 /* Called from syscall or from eBPF program */
 194 static int stack_map_pop_elem(struct bpf_map *map, void *value)
 195 {
 196         return __stack_map_get(map, value, true);
 197 }
 198
 199 /* Called from syscall or from eBPF program */
 200 static int queue_stack_map_push_elem(struct bpf_map *map, void *value,
 201                                      u64 flags)
 202 {
 203         struct bpf_queue_stack *qs = bpf_queue_stack(map);
 204         unsigned long irq_flags;
 205         int err = 0;
 206         void *dst;
 207
 208         /* BPF_EXIST is used to force making room for a new element in case the
 209          * map is full
 210          */
 211         bool replace = (flags & BPF_EXIST);
 212
 213         /* Check supported flags for queue and stack maps */
 214         if (flags & BPF_NOEXIST || flags > BPF_EXIST)
 215                 return -EINVAL;
 216
 217         raw_spin_lock_irqsave(&qs->lock, irq_flags);
 218
 219         if (queue_stack_map_is_full(qs)) {
 220                 if (!replace) {
 221                         err = -E2BIG;
 222                         goto out;
 223                 }
 224                 /* advance tail pointer to overwrite oldest element */
 225                 if (unlikely(++qs->tail >= qs->size))
 226                         qs->tail = 0;
 227         }
 228
 229         dst = &qs->elements[qs->head * qs->map.value_size];
 230         memcpy(dst, value, qs->map.value_size);
 231
 232         if (unlikely(++qs->head >= qs->size))
 233                 qs->head = 0;
 234
 235 out:
 236         raw_spin_unlock_irqrestore(&qs->lock, irq_flags);
 237         return err;
 238 }
 239
 240 /* Called from syscall or from eBPF program */
 241 static void *queue_stack_map_lookup_elem(struct bpf_map *map, void *key)
 242 {
 243         return NULL;
 244 }
 245
 246 /* Called from syscall or from eBPF program */
 247 static int queue_stack_map_update_elem(struct bpf_map *map, void *key,
 248                                        void *value, u64 flags)
 249 {
 250         return -EINVAL;
 251 }
 252
 253 /* Called from syscall or from eBPF program */
 254 static int queue_stack_map_delete_elem(struct bpf_map *map, void *key)
 255 {
 256         return -EINVAL;
 257 }
 258
 259 /* Called from syscall */
 260 static int queue_stack_map_get_next_key(struct bpf_map *map, void *key,
 261                                         void *next_key)
 262 {
 263         return -EINVAL;
 264 }
 265
 266 const struct bpf_map_ops queue_map_ops = {
 267         .map_alloc_check = queue_stack_map_alloc_check,
 268         .map_alloc = queue_stack_map_alloc,
 269         .map_free = queue_stack_map_free,
 270         .map_lookup_elem = queue_stack_map_lookup_elem,
 271         .map_update_elem = queue_stack_map_update_elem,
 272         .map_delete_elem = queue_stack_map_delete_elem,
 273         .map_push_elem = queue_stack_map_push_elem,
 274         .map_pop_elem = queue_map_pop_elem,
 275         .map_peek_elem = queue_map_peek_elem,
 276         .map_get_next_key = queue_stack_map_get_next_key,
 277 };
 278
 279 const struct bpf_map_ops stack_map_ops = {
 280         .map_alloc_check = queue_stack_map_alloc_check,
 281         .map_alloc = queue_stack_map_alloc,
 282         .map_free = queue_stack_map_free,
 283         .map_lookup_elem = queue_stack_map_lookup_elem,
 284         .map_update_elem = queue_stack_map_update_elem,
 285         .map_delete_elem = queue_stack_map_delete_elem,
 286         .map_push_elem = queue_stack_map_push_elem,
 287         .map_pop_elem = stack_map_pop_elem,
 288         .map_peek_elem = stack_map_peek_elem,
 289         .map_get_next_key = queue_stack_map_get_next_key,
 290 };