bpf: Do not use bucket_lock for hashmap iterator
authorYonghong Song <yhs@fb.com>
Wed, 2 Sep 2020 23:53:40 +0000 (16:53 -0700)
committerAlexei Starovoitov <ast@kernel.org>
Fri, 4 Sep 2020 00:36:41 +0000 (17:36 -0700)
Currently, for hashmap, the bpf iterator will grab a bucket lock, a
spinlock, before traversing the elements in the bucket. This can ensure
all bpf visted elements are valid. But this mechanism may cause
deadlock if update/deletion happens to the same bucket of the
visited map in the program. For example, if we added bpf_map_update_elem()
call to the same visited element in selftests bpf_iter_bpf_hash_map.c,
we will have the following deadlock:

  ============================================
  WARNING: possible recursive locking detected
  5.9.0-rc1+ #841 Not tainted
  --------------------------------------------
  test_progs/1750 is trying to acquire lock:
  ffff9a5bb73c5e70 (&htab->buckets[i].raw_lock){....}-{2:2}, at: htab_map_update_elem+0x1cf/0x410

  but task is already holding lock:
  ffff9a5bb73c5e20 (&htab->buckets[i].raw_lock){....}-{2:2}, at: bpf_hash_map_seq_find_next+0x94/0x120

  other info that might help us debug this:
   Possible unsafe locking scenario:

         CPU0
         ----
    lock(&htab->buckets[i].raw_lock);
    lock(&htab->buckets[i].raw_lock);

   *** DEADLOCK ***
   ...
  Call Trace:
   dump_stack+0x78/0xa0
   __lock_acquire.cold.74+0x209/0x2e3
   lock_acquire+0xba/0x380
   ? htab_map_update_elem+0x1cf/0x410
   ? __lock_acquire+0x639/0x20c0
   _raw_spin_lock_irqsave+0x3b/0x80
   ? htab_map_update_elem+0x1cf/0x410
   htab_map_update_elem+0x1cf/0x410
   ? lock_acquire+0xba/0x380
   bpf_prog_ad6dab10433b135d_dump_bpf_hash_map+0x88/0xa9c
   ? find_held_lock+0x34/0xa0
   bpf_iter_run_prog+0x81/0x16e
   __bpf_hash_map_seq_show+0x145/0x180
   bpf_seq_read+0xff/0x3d0
   vfs_read+0xad/0x1c0
   ksys_read+0x5f/0xe0
   do_syscall_64+0x33/0x40
   entry_SYSCALL_64_after_hwframe+0x44/0xa9
  ...

The bucket_lock first grabbed in seq_ops->next() called by bpf_seq_read(),
and then grabbed again in htab_map_update_elem() in the bpf program, causing
deadlocks.

Actually, we do not need bucket_lock here, we can just use rcu_read_lock()
similar to netlink iterator where the rcu_read_{lock,unlock} likes below:
 seq_ops->start():
     rcu_read_lock();
 seq_ops->next():
     rcu_read_unlock();
     /* next element */
     rcu_read_lock();
 seq_ops->stop();
     rcu_read_unlock();

Compared to old bucket_lock mechanism, if concurrent updata/delete happens,
we may visit stale elements, miss some elements, or repeat some elements.
I think this is a reasonable compromise. For users wanting to avoid
stale, missing/repeated accesses, bpf_map batch access syscall interface
can be used.

Signed-off-by: Yonghong Song <yhs@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20200902235340.2001375-1-yhs@fb.com
kernel/bpf/hashtab.c

index 78dfff6..7df28a4 100644 (file)
@@ -1622,7 +1622,6 @@ struct bpf_iter_seq_hash_map_info {
        struct bpf_map *map;
        struct bpf_htab *htab;
        void *percpu_value_buf; // non-zero means percpu hash
-       unsigned long flags;
        u32 bucket_id;
        u32 skip_elems;
 };
@@ -1632,7 +1631,6 @@ bpf_hash_map_seq_find_next(struct bpf_iter_seq_hash_map_info *info,
                           struct htab_elem *prev_elem)
 {
        const struct bpf_htab *htab = info->htab;
-       unsigned long flags = info->flags;
        u32 skip_elems = info->skip_elems;
        u32 bucket_id = info->bucket_id;
        struct hlist_nulls_head *head;
@@ -1656,19 +1654,18 @@ bpf_hash_map_seq_find_next(struct bpf_iter_seq_hash_map_info *info,
 
                /* not found, unlock and go to the next bucket */
                b = &htab->buckets[bucket_id++];
-               htab_unlock_bucket(htab, b, flags);
+               rcu_read_unlock();
                skip_elems = 0;
        }
 
        for (i = bucket_id; i < htab->n_buckets; i++) {
                b = &htab->buckets[i];
-               flags = htab_lock_bucket(htab, b);
+               rcu_read_lock();
 
                count = 0;
                head = &b->head;
                hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
                        if (count >= skip_elems) {
-                               info->flags = flags;
                                info->bucket_id = i;
                                info->skip_elems = count;
                                return elem;
@@ -1676,7 +1673,7 @@ bpf_hash_map_seq_find_next(struct bpf_iter_seq_hash_map_info *info,
                        count++;
                }
 
-               htab_unlock_bucket(htab, b, flags);
+               rcu_read_unlock();
                skip_elems = 0;
        }
 
@@ -1754,14 +1751,10 @@ static int bpf_hash_map_seq_show(struct seq_file *seq, void *v)
 
 static void bpf_hash_map_seq_stop(struct seq_file *seq, void *v)
 {
-       struct bpf_iter_seq_hash_map_info *info = seq->private;
-
        if (!v)
                (void)__bpf_hash_map_seq_show(seq, NULL);
        else
-               htab_unlock_bucket(info->htab,
-                                  &info->htab->buckets[info->bucket_id],
-                                  info->flags);
+               rcu_read_unlock();
 }
 
 static int bpf_iter_init_hash_map(void *priv_data,