kernel/bpf/devmap.c

   1 /* Copyright (c) 2017 Covalent IO, Inc. http://covalent.io
   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  * This program is distributed in the hope that it will be useful, but
   8  * WITHOUT ANY WARRANTY; without even the implied warranty of
   9  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  10  * General Public License for more details.
  11  */
  12
  13 /* Devmaps primary use is as a backend map for XDP BPF helper call
  14  * bpf_redirect_map(). Because XDP is mostly concerned with performance we
  15  * spent some effort to ensure the datapath with redirect maps does not use
  16  * any locking. This is a quick note on the details.
  17  *
  18  * We have three possible paths to get into the devmap control plane bpf
  19  * syscalls, bpf programs, and driver side xmit/flush operations. A bpf syscall
  20  * will invoke an update, delete, or lookup operation. To ensure updates and
  21  * deletes appear atomic from the datapath side xchg() is used to modify the
  22  * netdev_map array. Then because the datapath does a lookup into the netdev_map
  23  * array (read-only) from an RCU critical section we use call_rcu() to wait for
  24  * an rcu grace period before free'ing the old data structures. This ensures the
  25  * datapath always has a valid copy. However, the datapath does a "flush"
  26  * operation that pushes any pending packets in the driver outside the RCU
  27  * critical section. Each bpf_dtab_netdev tracks these pending operations using
  28  * an atomic per-cpu bitmap. The bpf_dtab_netdev object will not be destroyed
  29  * until all bits are cleared indicating outstanding flush operations have
  30  * completed.
  31  *
  32  * BPF syscalls may race with BPF program calls on any of the update, delete
  33  * or lookup operations. As noted above the xchg() operation also keep the
  34  * netdev_map consistent in this case. From the devmap side BPF programs
  35  * calling into these operations are the same as multiple user space threads
  36  * making system calls.
  37  *
  38  * Finally, any of the above may race with a netdev_unregister notifier. The
  39  * unregister notifier must search for net devices in the map structure that
  40  * contain a reference to the net device and remove them. This is a two step
  41  * process (a) dereference the bpf_dtab_netdev object in netdev_map and (b)
  42  * check to see if the ifindex is the same as the net_device being removed.
  43  * When removing the dev a cmpxchg() is used to ensure the correct dev is
  44  * removed, in the case of a concurrent update or delete operation it is
  45  * possible that the initially referenced dev is no longer in the map. As the
  46  * notifier hook walks the map we know that new dev references can not be
  47  * added by the user because core infrastructure ensures dev_get_by_index()
  48  * calls will fail at this point.
  49  */
  50 #include <linux/bpf.h>
  51 #include <linux/filter.h>
  52
  53 struct bpf_dtab_netdev {
  54         struct net_device *dev;
  55         struct bpf_dtab *dtab;
  56         unsigned int bit;
  57         struct rcu_head rcu;
  58 };
  59
  60 struct bpf_dtab {
  61         struct bpf_map map;
  62         struct bpf_dtab_netdev **netdev_map;
  63         unsigned long __percpu *flush_needed;
  64         struct list_head list;
  65 };
  66
  67 static DEFINE_SPINLOCK(dev_map_lock);
  68 static LIST_HEAD(dev_map_list);
  69
  70 static u64 dev_map_bitmap_size(const union bpf_attr *attr)
  71 {
  72         return BITS_TO_LONGS(attr->max_entries) * sizeof(unsigned long);
  73 }
  74
  75 static struct bpf_map *dev_map_alloc(union bpf_attr *attr)
  76 {
  77         struct bpf_dtab *dtab;
  78         u64 cost;
  79         int err;
  80
  81         /* check sanity of attributes */
  82         if (attr->max_entries == 0 || attr->key_size != 4 ||
  83             attr->value_size != 4 || attr->map_flags & ~BPF_F_NUMA_NODE)
  84                 return ERR_PTR(-EINVAL);
  85
  86         dtab = kzalloc(sizeof(*dtab), GFP_USER);
  87         if (!dtab)
  88                 return ERR_PTR(-ENOMEM);
  89
  90         /* mandatory map attributes */
  91         dtab->map.map_type = attr->map_type;
  92         dtab->map.key_size = attr->key_size;
  93         dtab->map.value_size = attr->value_size;
  94         dtab->map.max_entries = attr->max_entries;
  95         dtab->map.map_flags = attr->map_flags;
  96         dtab->map.numa_node = bpf_map_attr_numa_node(attr);
  97
  98         /* make sure page count doesn't overflow */
  99         cost = (u64) dtab->map.max_entries * sizeof(struct bpf_dtab_netdev *);
 100         cost += dev_map_bitmap_size(attr) * num_possible_cpus();
 101         if (cost >= U32_MAX - PAGE_SIZE)
 102                 goto free_dtab;
 103
 104         dtab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
 105
 106         /* if map size is larger than memlock limit, reject it early */
 107         err = bpf_map_precharge_memlock(dtab->map.pages);
 108         if (err)
 109                 goto free_dtab;
 110
 111         /* A per cpu bitfield with a bit per possible net device */
 112         dtab->flush_needed = __alloc_percpu(dev_map_bitmap_size(attr),
 113                                             __alignof__(unsigned long));
 114         if (!dtab->flush_needed)
 115                 goto free_dtab;
 116
 117         dtab->netdev_map = bpf_map_area_alloc(dtab->map.max_entries *
 118                                               sizeof(struct bpf_dtab_netdev *),
 119                                               dtab->map.numa_node);
 120         if (!dtab->netdev_map)
 121                 goto free_dtab;
 122
 123         spin_lock(&dev_map_lock);
 124         list_add_tail_rcu(&dtab->list, &dev_map_list);
 125         spin_unlock(&dev_map_lock);
 126
 127         return &dtab->map;
 128 free_dtab:
 129         free_percpu(dtab->flush_needed);
 130         kfree(dtab);
 131         return ERR_PTR(-ENOMEM);
 132 }
 133
 134 static void dev_map_free(struct bpf_map *map)
 135 {
 136         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 137         int i, cpu;
 138
 139         /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
 140          * so the programs (can be more than one that used this map) were
 141          * disconnected from events. Wait for outstanding critical sections in
 142          * these programs to complete. The rcu critical section only guarantees
 143          * no further reads against netdev_map. It does __not__ ensure pending
 144          * flush operations (if any) are complete.
 145          */
 146
 147         spin_lock(&dev_map_lock);
 148         list_del_rcu(&dtab->list);
 149         spin_unlock(&dev_map_lock);
 150
 151         synchronize_rcu();
 152
 153         /* To ensure all pending flush operations have completed wait for flush
 154          * bitmap to indicate all flush_needed bits to be zero on _all_ cpus.
 155          * Because the above synchronize_rcu() ensures the map is disconnected
 156          * from the program we can assume no new bits will be set.
 157          */
 158         for_each_online_cpu(cpu) {
 159                 unsigned long *bitmap = per_cpu_ptr(dtab->flush_needed, cpu);
 160
 161                 while (!bitmap_empty(bitmap, dtab->map.max_entries))
 162                         cpu_relax();
 163         }
 164
 165         for (i = 0; i < dtab->map.max_entries; i++) {
 166                 struct bpf_dtab_netdev *dev;
 167
 168                 dev = dtab->netdev_map[i];
 169                 if (!dev)
 170                         continue;
 171
 172                 dev_put(dev->dev);
 173                 kfree(dev);
 174         }
 175
 176         free_percpu(dtab->flush_needed);
 177         bpf_map_area_free(dtab->netdev_map);
 178         kfree(dtab);
 179 }
 180
 181 static int dev_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 182 {
 183         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 184         u32 index = key ? *(u32 *)key : U32_MAX;
 185         u32 *next = next_key;
 186
 187         if (index >= dtab->map.max_entries) {
 188                 *next = 0;
 189                 return 0;
 190         }
 191
 192         if (index == dtab->map.max_entries - 1)
 193                 return -ENOENT;
 194         *next = index + 1;
 195         return 0;
 196 }
 197
 198 void __dev_map_insert_ctx(struct bpf_map *map, u32 bit)
 199 {
 200         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 201         unsigned long *bitmap = this_cpu_ptr(dtab->flush_needed);
 202
 203         __set_bit(bit, bitmap);
 204 }
 205
 206 /* __dev_map_flush is called from xdp_do_flush_map() which _must_ be signaled
 207  * from the driver before returning from its napi->poll() routine. The poll()
 208  * routine is called either from busy_poll context or net_rx_action signaled
 209  * from NET_RX_SOFTIRQ. Either way the poll routine must complete before the
 210  * net device can be torn down. On devmap tear down we ensure the ctx bitmap
 211  * is zeroed before completing to ensure all flush operations have completed.
 212  */
 213 void __dev_map_flush(struct bpf_map *map)
 214 {
 215         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 216         unsigned long *bitmap = this_cpu_ptr(dtab->flush_needed);
 217         u32 bit;
 218
 219         for_each_set_bit(bit, bitmap, map->max_entries) {
 220                 struct bpf_dtab_netdev *dev = READ_ONCE(dtab->netdev_map[bit]);
 221                 struct net_device *netdev;
 222
 223                 /* This is possible if the dev entry is removed by user space
 224                  * between xdp redirect and flush op.
 225                  */
 226                 if (unlikely(!dev))
 227                         continue;
 228
 229                 __clear_bit(bit, bitmap);
 230                 netdev = dev->dev;
 231                 if (likely(netdev->netdev_ops->ndo_xdp_flush))
 232                         netdev->netdev_ops->ndo_xdp_flush(netdev);
 233         }
 234 }
 235
 236 /* rcu_read_lock (from syscall and BPF contexts) ensures that if a delete and/or
 237  * update happens in parallel here a dev_put wont happen until after reading the
 238  * ifindex.
 239  */
 240 struct net_device  *__dev_map_lookup_elem(struct bpf_map *map, u32 key)
 241 {
 242         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 243         struct bpf_dtab_netdev *dev;
 244
 245         if (key >= map->max_entries)
 246                 return NULL;
 247
 248         dev = READ_ONCE(dtab->netdev_map[key]);
 249         return dev ? dev->dev : NULL;
 250 }
 251
 252 static void *dev_map_lookup_elem(struct bpf_map *map, void *key)
 253 {
 254         struct net_device *dev = __dev_map_lookup_elem(map, *(u32 *)key);
 255
 256         return dev ? &dev->ifindex : NULL;
 257 }
 258
 259 static void dev_map_flush_old(struct bpf_dtab_netdev *dev)
 260 {
 261         if (dev->dev->netdev_ops->ndo_xdp_flush) {
 262                 struct net_device *fl = dev->dev;
 263                 unsigned long *bitmap;
 264                 int cpu;
 265
 266                 for_each_online_cpu(cpu) {
 267                         bitmap = per_cpu_ptr(dev->dtab->flush_needed, cpu);
 268                         __clear_bit(dev->bit, bitmap);
 269
 270                         fl->netdev_ops->ndo_xdp_flush(dev->dev);
 271                 }
 272         }
 273 }
 274
 275 static void __dev_map_entry_free(struct rcu_head *rcu)
 276 {
 277         struct bpf_dtab_netdev *dev;
 278
 279         dev = container_of(rcu, struct bpf_dtab_netdev, rcu);
 280         dev_map_flush_old(dev);
 281         dev_put(dev->dev);
 282         kfree(dev);
 283 }
 284
 285 static int dev_map_delete_elem(struct bpf_map *map, void *key)
 286 {
 287         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 288         struct bpf_dtab_netdev *old_dev;
 289         int k = *(u32 *)key;
 290
 291         if (k >= map->max_entries)
 292                 return -EINVAL;
 293
 294         /* Use call_rcu() here to ensure any rcu critical sections have
 295          * completed, but this does not guarantee a flush has happened
 296          * yet. Because driver side rcu_read_lock/unlock only protects the
 297          * running XDP program. However, for pending flush operations the
 298          * dev and ctx are stored in another per cpu map. And additionally,
 299          * the driver tear down ensures all soft irqs are complete before
 300          * removing the net device in the case of dev_put equals zero.
 301          */
 302         old_dev = xchg(&dtab->netdev_map[k], NULL);
 303         if (old_dev)
 304                 call_rcu(&old_dev->rcu, __dev_map_entry_free);
 305         return 0;
 306 }
 307
 308 static int dev_map_update_elem(struct bpf_map *map, void *key, void *value,
 309                                 u64 map_flags)
 310 {
 311         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 312         struct net *net = current->nsproxy->net_ns;
 313         struct bpf_dtab_netdev *dev, *old_dev;
 314         u32 i = *(u32 *)key;
 315         u32 ifindex = *(u32 *)value;
 316
 317         if (unlikely(map_flags > BPF_EXIST))
 318                 return -EINVAL;
 319         if (unlikely(i >= dtab->map.max_entries))
 320                 return -E2BIG;
 321         if (unlikely(map_flags == BPF_NOEXIST))
 322                 return -EEXIST;
 323
 324         if (!ifindex) {
 325                 dev = NULL;
 326         } else {
 327                 dev = kmalloc_node(sizeof(*dev), GFP_ATOMIC | __GFP_NOWARN,
 328                                    map->numa_node);
 329                 if (!dev)
 330                         return -ENOMEM;
 331
 332                 dev->dev = dev_get_by_index(net, ifindex);
 333                 if (!dev->dev) {
 334                         kfree(dev);
 335                         return -EINVAL;
 336                 }
 337
 338                 dev->bit = i;
 339                 dev->dtab = dtab;
 340         }
 341
 342         /* Use call_rcu() here to ensure rcu critical sections have completed
 343          * Remembering the driver side flush operation will happen before the
 344          * net device is removed.
 345          */
 346         old_dev = xchg(&dtab->netdev_map[i], dev);
 347         if (old_dev)
 348                 call_rcu(&old_dev->rcu, __dev_map_entry_free);
 349
 350         return 0;
 351 }
 352
 353 const struct bpf_map_ops dev_map_ops = {
 354         .map_alloc = dev_map_alloc,
 355         .map_free = dev_map_free,
 356         .map_get_next_key = dev_map_get_next_key,
 357         .map_lookup_elem = dev_map_lookup_elem,
 358         .map_update_elem = dev_map_update_elem,
 359         .map_delete_elem = dev_map_delete_elem,
 360 };
 361
 362 static int dev_map_notification(struct notifier_block *notifier,
 363                                 ulong event, void *ptr)
 364 {
 365         struct net_device *netdev = netdev_notifier_info_to_dev(ptr);
 366         struct bpf_dtab *dtab;
 367         int i;
 368
 369         switch (event) {
 370         case NETDEV_UNREGISTER:
 371                 /* This rcu_read_lock/unlock pair is needed because
 372                  * dev_map_list is an RCU list AND to ensure a delete
 373                  * operation does not free a netdev_map entry while we
 374                  * are comparing it against the netdev being unregistered.
 375                  */
 376                 rcu_read_lock();
 377                 list_for_each_entry_rcu(dtab, &dev_map_list, list) {
 378                         for (i = 0; i < dtab->map.max_entries; i++) {
 379                                 struct bpf_dtab_netdev *dev, *odev;
 380
 381                                 dev = READ_ONCE(dtab->netdev_map[i]);
 382                                 if (!dev ||
 383                                     dev->dev->ifindex != netdev->ifindex)
 384                                         continue;
 385                                 odev = cmpxchg(&dtab->netdev_map[i], dev, NULL);
 386                                 if (dev == odev)
 387                                         call_rcu(&dev->rcu,
 388                                                  __dev_map_entry_free);
 389                         }
 390                 }
 391                 rcu_read_unlock();
 392                 break;
 393         default:
 394                 break;
 395         }
 396         return NOTIFY_OK;
 397 }
 398
 399 static struct notifier_block dev_map_notifier = {
 400         .notifier_call = dev_map_notification,
 401 };
 402
 403 static int __init dev_map_init(void)
 404 {
 405         register_netdevice_notifier(&dev_map_notifier);
 406         return 0;
 407 }
 408
 409 subsys_initcall(dev_map_init);