On Sat, Feb 13, 2021 at 1:07 AM Wang Yugui [off-list ref] wrote:

Hi,

quoted

This bug only affects 5.10 kernels, and the regression was introduced in
5.10-rc1 by commit 0eb79294dbe328 ("btrfs: dio iomap DSYNC workaround").
The bug does not exist in 5.11 kernels due to commit ecfdc08b8cc65d
("btrfs: remove dio iomap DSYNC workaround"), which depends on other
changes that went into the merge window for 5.11. So this is a fix only
for 5.10.x stable kernels, as there are people hitting this.

It is OK too to backport commit ecfdc08b8cc65d
 ("btrfs: remove dio iomap DSYNC workaround") to 5.10 for this problem?

the iomap issue for commit 0eb79294dbe328 ("btrfs: dio iomap DSYNC workaround")
is already fixed in 5.10?

Quoting the changelog:

"commit ecfdc08b8cc65d
("btrfs: remove dio iomap DSYNC workaround"), which depends on other
changes that went into the merge window for 5.11."

All the changes, are (at least):

commit ecfdc08b8cc65d737eebc26a1ee1875a097fd6a0   --> 5.11-rc1
Author: Goldwyn Rodrigues [off-list ref]
Date:   Thu Sep 24 11:39:21 2020 -0500

    btrfs: remove dio iomap DSYNC workaround

commit a42fa643169d2325602572633fcaa16862990e28
Author: Goldwyn Rodrigues [off-list ref]
Date:   Thu Sep 24 11:39:20 2020 -0500

    btrfs: call iomap_dio_complete() without inode_lock

commit 502756b380938022c848761837f8fa3976906aa1
Author: Goldwyn Rodrigues [off-list ref]
Date:   Thu Sep 24 11:39:19 2020 -0500

    btrfs: remove btrfs_inode::dio_sem

commit e9adabb9712ef9424cbbeeaa027d962ab5262e19
Author: Goldwyn Rodrigues [off-list ref]
Date:   Thu Sep 24 11:39:18 2020 -0500

    btrfs: use shared lock for direct writes within EOF

commit c352370633400d13765cc88080c969799ea51108
Author: Goldwyn Rodrigues [off-list ref]
Date:   Thu Sep 24 11:39:17 2020 -0500

    btrfs: push inode locking and unlocking into buffered/direct write

commit a14b78ad06aba0fa7e76d2bc13c5ba581a7f331a
Author: Goldwyn Rodrigues [off-list ref]
Date:   Thu Sep 24 11:39:16 2020 -0500

    btrfs: introduce btrfs_inode_lock()/unlock()

commit b8d8e1fd570a194904f545b135efc880d96a41a4
Author: Goldwyn Rodrigues [off-list ref]
Date:   Thu Sep 24 11:39:15 2020 -0500

    btrfs: introduce btrfs_write_check()

That's probably too much to add to stable at once, plus I'm assuming
all required iomap dependencies are in 5.10 already (it seems so,
unless I missed something).

Usually we don't add patches to stable that didn't go through Linus'
tree either (there were 1 or 2 very rare exceptions in the past I
think), but when a backport depends on so many patches, and not all
from the same patchset, the risk of getting something wrong is
significant. That's why I opted to send this patch, which is much more
simple.

David has more experience on that and it's up to him to decide.

Best Regards
Wang Yugui (wangyugui@e16-tech.com)
2021/02/13

quoted

Fixes: 0eb79294dbe328 ("btrfs: dio iomap DSYNC workaround")
CC: stable@vger.kernel.org # 5.10 (and only 5.10)
Bugzilla: https://bugzilla.suse.com/show_bug.cgi?id=1181605
Signed-off-by: Filipe Manana <redacted>
---
 fs/btrfs/inode.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c
index acc47e2ffb46..b536d21541a9 100644
--- a/fs/btrfs/inode.c
+++ b/fs/btrfs/inode.c

@@ -8026,8 +8026,12 @@ ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
      bool relock = false;
      ssize_t ret;

-     if (check_direct_IO(fs_info, iter, offset))
+     if (check_direct_IO(fs_info, iter, offset)) {
+             ASSERT(current->journal_info == NULL ||
+                    current->journal_info == BTRFS_DIO_SYNC_STUB);
+             current->journal_info = NULL;
              return 0;
+     }

      count = iov_iter_count(iter);
      if (iov_iter_rw(iter) == WRITE) {
--

2.28.0