btrfs: Only check first key for committed tree blocks

When looping btrfs/074 with many cpus (>= 8), it's possible to trigger
kernel warning due to first key verification:

[ 4239.523446] WARNING: CPU: 5 PID: 2381 at fs/btrfs/disk-io.c:460 btree_read_extent_buffer_pages+0x1ad/0x210
[ 4239.523830] Modules linked in:
[ 4239.524630] RIP: 0010:btree_read_extent_buffer_pages+0x1ad/0x210
[ 4239.527101] Call Trace:
[ 4239.527251]  read_tree_block+0x42/0x70
[ 4239.527434]  read_node_slot+0xd2/0x110
[ 4239.527632]  push_leaf_right+0xad/0x1b0
[ 4239.527809]  split_leaf+0x4ea/0x700
[ 4239.527988]  ? leaf_space_used+0xbc/0xe0
[ 4239.528192]  ? btrfs_set_lock_blocking_rw+0x99/0xb0
[ 4239.528416]  btrfs_search_slot+0x8cc/0xa40
[ 4239.528605]  btrfs_insert_empty_items+0x71/0xc0
[ 4239.528798]  __btrfs_run_delayed_refs+0xa98/0x1680
[ 4239.529013]  btrfs_run_delayed_refs+0x10b/0x1b0
[ 4239.529205]  btrfs_commit_transaction+0x33/0xaf0
[ 4239.529445]  ? start_transaction+0xa8/0x4f0
[ 4239.529630]  btrfs_alloc_data_chunk_ondemand+0x1b0/0x4e0
[ 4239.529833]  btrfs_check_data_free_space+0x54/0xa0
[ 4239.530045]  btrfs_delalloc_reserve_space+0x25/0x70
[ 4239.531907]  btrfs_direct_IO+0x233/0x3d0
[ 4239.532098]  generic_file_direct_write+0xcb/0x170
[ 4239.532296]  btrfs_file_write_iter+0x2bb/0x5f4
[ 4239.532491]  aio_write+0xe2/0x180
[ 4239.532669]  ? lock_acquire+0xac/0x1e0
[ 4239.532839]  ? __might_fault+0x3e/0x90
[ 4239.533032]  do_io_submit+0x594/0x860
[ 4239.533223]  ? do_io_submit+0x594/0x860
[ 4239.533398]  SyS_io_submit+0x10/0x20
[ 4239.533560]  ? SyS_io_submit+0x10/0x20
[ 4239.533729]  do_syscall_64+0x75/0x1d0
[ 4239.533979]  entry_SYSCALL_64_after_hwframe+0x42/0xb7
[ 4239.534182] RIP: 0033:0x7f8519741697

The problem here is, at btree_read_extent_buffer_pages() we don't have
acquired read/write lock on that extent buffer, only basic info like
level/bytenr is reliable.

So race condition leads to such false alert.

However in current call site, it's impossible to acquire proper lock
without race window.
To fix the problem, we only verify first key for committed tree blocks
(whose generation is no larger than fs_info->last_trans_committed), so
the content of such tree blocks will not change and there is no need to
get read/write lock.

Reported-by: Nikolay Borisov <nborisov@suse.com>
Fixes: 581c1760415c ("btrfs: Validate child tree block's level and first key")
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>

diff --git a/fs/btrfs/disk-io.c b/fs/btrfs/disk-io.c
index 74cbc342b29ad35f8d6fb401c8bf6550c8cf27a3..4ac8b1d21bafd07e00a686d6c15228f6be6aef72 100644 (file)
--- a/fs/btrfs/disk-io.c
+++ b/fs/btrfs/disk-io.c
@@ -436,6 +436,14 @@ static int verify_level_key(struct btrfs_fs_info *fs_info,
        if (!first_key)
                return 0;
 
+       /*
+        * For live tree block (new tree blocks in current transaction),
+        * we need proper lock context to avoid race, which is impossible here.
+        * So we only checks tree blocks which is read from disk, whose
+        * generation <= fs_info->last_trans_committed.
+        */
+       if (btrfs_header_generation(eb) > fs_info->last_trans_committed)
+               return 0;
        if (found_level)
                btrfs_node_key_to_cpu(eb, &found_key, 0);
        else