On Wed, Apr 28, 2021 at 07:03:16AM +0800, Qu Wenruo wrote:

+		if (entry->bytes_left == 0) {
+			set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
+			/* set_bit implies a barrier */
+			cond_wake_up_nomb(&entry->wait);

Well, no, set_bit does not imply a barrier. In general, RMW operations
do, and set_bit lacks the 'R' part. It's also in the memory-barriers.txt
document, look up set_bit. I inserted a patch that does
cond_wake_up that does the barrier, calling smp_mb_after_atomic +
cond_wake_up_nomb would also work.