On Sat, May 09, 2020 at 01:34:31PM +0800, Jia-Ju Bai wrote: > The functions btrfs_block_rsv_release() and > btrfs_update_delayed_refs_rsv() are concurrently executed at runtime in > the following call contexts: > > Thread 1: > btrfs_file_write_iter() > btrfs_buffered_write() > btrfs_delalloc_release_extents() > btrfs_inode_rsv_release() > __btrfs_block_rsv_release() > > Thread 2: > finish_ordered_fn() > btrfs_finish_ordered_io() > insert_reserved_file_extent() > __btrfs_drop_extents() > btrfs_free_extent() > btrfs_add_delayed_data_ref() > btrfs_update_delayed_refs_rsv() > > In __btrfs_block_rsv_release(): > else if (... && !delayed_rsv->full) > > In btrfs_update_delayed_refs_rsv(): > spin_lock(&delayed_rsv->lock); > delayed_rsv->size += num_bytes; > delayed_rsv->full = 0; > spin_unlock(&delayed_rsv->lock); > > Thus a data race for delayed_rsv->full can occur. > This race was found and actually reproduced by our conccurency fuzzer. > > To fix this race, the spinlock delayed_rsv->lock is used to > protect the access to delayed_rsv->full in btrfs_block_rsv_release(). > > Signed-off-by: Jia-Ju Bai <baijiaju1990@xxxxxxxxx> > --- > fs/btrfs/block-rsv.c | 7 ++++++- > 1 file changed, 6 insertions(+), 1 deletion(-) > > diff --git a/fs/btrfs/block-rsv.c b/fs/btrfs/block-rsv.c > index 27efec8f7c5b..89c53a7137b4 100644 > --- a/fs/btrfs/block-rsv.c > +++ b/fs/btrfs/block-rsv.c > @@ -277,6 +277,11 @@ u64 btrfs_block_rsv_release(struct btrfs_fs_info *fs_info, > struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv; > struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_refs_rsv; > struct btrfs_block_rsv *target = NULL; > + unsigned short full = 0; > + > + spin_lock(&delayed_rsv->lock); > + full = delayed_rsv->full; > + spin_unlock(&delayed_rsv->lock); > > /* > * If we are the delayed_rsv then push to the global rsv, otherwise dump > @@ -284,7 +289,7 @@ u64 btrfs_block_rsv_release(struct btrfs_fs_info *fs_info, > */ > if (block_rsv == delayed_rsv) > target = global_rsv; > - else if (block_rsv != global_rsv && !delayed_rsv->full) > + else if (block_rsv != global_rsv && !full) This has been reported as suspicous https://lore.kernel.org/linux-btrfs/CAAwBoOJDjei5Hnem155N_cJwiEkVwJYvgN-tQrwWbZQGhFU=cA@xxxxxxxxxxxxxx/ and there's an answer that this is racy but does not cause any unexpected behaviour.
