On Fri, Feb 05, 2021 at 11:49:05AM +0000, Filipe Manana wrote:
> On Fri, Feb 5, 2021 at 9:26 AM Naohiro Aota <naohiro.aota@xxxxxxx> wrote:
> >
> > Since the zoned filesystem requires sequential write out of metadata, we
> > cannot proceed with a hole in tree-log pages. When such a hole exists,
> > btree_write_cache_pages() will return -EAGAIN. We cannot wait for the range
> > to be written, because it will cause a deadlock. So, let's bail out to a
> > full commit in this case.
> >
> > Cc: Filipe Manana <fdmanana@xxxxxxxxx>
> > Signed-off-by: Naohiro Aota <naohiro.aota@xxxxxxx>
> > ---
> > fs/btrfs/tree-log.c | 19 ++++++++++++++++++-
> > 1 file changed, 18 insertions(+), 1 deletion(-)
> >
> > This patch solves a regression introduced by fixing patch 40. I'm
> > sorry for the confusing patch numbering.
>
> Hum, how does patch 40 can cause this?
> And is it before the fixup or after?
With pre-5.10 code base + zoned series at that time, it passed
xfstests without this patch.
With current code base + zoned series without the fixup for patch 40,
it also passed the tests, because we are mostly bailing out to a full
commit.
The fixup now stressed the new fsync code on zoned mode and revealed
an issue to have -EAGAIN from btrfs_write_marked_extents(). This error
happens when a concurrent transaction commit is writing a dirty extent
in this tree-log commit. This issue didn't occur previously because of
a longer critical section, I guess.
>
> >
> > diff --git a/fs/btrfs/tree-log.c b/fs/btrfs/tree-log.c
> > index 4e72794342c0..629e605cd62d 100644
> > --- a/fs/btrfs/tree-log.c
> > +++ b/fs/btrfs/tree-log.c
> > @@ -3120,6 +3120,14 @@ int btrfs_sync_log(struct btrfs_trans_handle *trans,
> > */
> > blk_start_plug(&plug);
> > ret = btrfs_write_marked_extents(fs_info, &log->dirty_log_pages, mark);
> > + /*
> > + * There is a hole writing out the extents and cannot proceed it on
> > + * zoned filesystem, which require sequential writing. We can
>
> require -> requires
>
> > + * ignore the error for now, since we don't wait for completion for
> > + * now.
>
> So why can we ignore the error for now?
> Why not just bail out here and mark the log for full commit? (without
> a transaction abort)
As described above, -EAGAIN happens when a concurrent process writes
out an extent buffer of this tree-log commit. This concurrent write
out will fill a hole for us, so the next write out might
succeed. Indeed we can bail out here, but I opted to try the next
write.
> > + */
> > + if (ret == -EAGAIN)
> > + ret = 0;
> > if (ret) {
> > blk_finish_plug(&plug);
> > btrfs_abort_transaction(trans, ret);
> > @@ -3229,7 +3237,16 @@ int btrfs_sync_log(struct btrfs_trans_handle *trans,
> > &log_root_tree->dirty_log_pages,
> > EXTENT_DIRTY | EXTENT_NEW);
> > blk_finish_plug(&plug);
> > - if (ret) {
> > + /*
> > + * There is a hole in the extents, and failed to sequential write
> > + * on zoned filesystem. We cannot wait for this write outs, sinc it
>
> this -> these
>
> > + * cause a deadlock. Bail out to the full commit, instead.
> > + */
> > + if (ret == -EAGAIN) {
> > + btrfs_wait_tree_log_extents(log, mark);
> > + mutex_unlock(&log_root_tree->log_mutex);
> > + goto out_wake_log_root;
>
> Must also call btrfs_set_log_full_commit(trans);
Oops, I missed this one.
> Thanks.
>
> > + } else if (ret) {
> > btrfs_set_log_full_commit(trans);
> > btrfs_abort_transaction(trans, ret);
> > mutex_unlock(&log_root_tree->log_mutex);
> > --
> > 2.30.0
> >
>
>
> --
> Filipe David Manana,
>
> “Whether you think you can, or you think you can't — you're right.”
