There is a race between ext4 buffer write and direct_IO read with dioread_nolock mount option enabled. The problem is that we clear PageWriteback flag during end_io time but will do uninitialized-to-initialized extent conversion later with dioread_nolock. If an O_direct read request comes in during this period, ext4 will return zero instead of the recently written data. This patch checks whether there are any pending uninitialized-to-initialized extent conversion requests before doing O_direct read to close the race. Note that this is just a bandaid fix. The fundamental issue is that we clear PageWriteback flag before we really complete an IO, which is problem-prone. To fix the fundamental issue, we may need to implement an extent tree cache that we can use to look up pending to-be-converted extents.
Unified · Split
| @@ -800,12 +800,17 @@ ssize_t ext4_ind_direct_IO(int rw, struct kiocb *iocb, · 799 unmodified lines | ||
| 800 | 800 | } |
| 801 | 801 | |
| 802 | 802 | retry: |
| 803 | − if (rw == READ && ext4_should_dioread_nolock(inode)) | |
| 803 | + if (rw == READ && ext4_should_dioread_nolock(inode)) { | |
| 804 | + if (unlikely(!list_empty(&ei->i_completed_io_list))) { | |
| 805 | + mutex_lock(&inode->i_mutex); | |
| 806 | + ext4_flush_completed_IO(inode); | |
| 807 | + mutex_unlock(&inode->i_mutex); | |
| 808 | + } | |
| 804 | 809 | ret = __blockdev_direct_IO(rw, iocb, inode, |
| 805 | 810 | inode->i_sb->s_bdev, iov, |
| 806 | 811 | offset, nr_segs, |
| 807 | 812 | ext4_get_block, NULL, NULL, 0); |
| 808 | − else { | |
| 813 | + } else { | |
| 809 | 814 | ret = blockdev_direct_IO(rw, iocb, inode, iov, |
| 810 | 815 | offset, nr_segs, ext4_get_block); |
| 811 | 816 | |