On Fri, Nov 27, 2020 at 11:21 AM Sargun Dhillon <sargun@xxxxxxxxx> wrote: > > This documents behaviour that was discussed in a thread about the volatile > feature. Specifically, how failures can go hidden from asynchronous writes > (such as from mmap, or writes that are not immediately flushed to the > filesystem). Although we pass through calls like msync, fallocate, and > write, and will still return errors on those, it doesn't guarantee all > kinds of errors will happen at those times, and thus may hide errors. > > In the future, we can add error checking to all interactions with the > upperdir, and pass through errseq_t from the upperdir on mappings, > and other interactions with the filesystem[1]. > > [1]: https://lore.kernel.org/linux-unionfs/20201116045758.21774-1-sargun@xxxxxxxxx/T/#m7d501f375e031056efad626e471a1392dd3aad33 > > Signed-off-by: Sargun Dhillon <sargun@xxxxxxxxx> > Cc: linux-fsdevel@xxxxxxxxxxxxxxx > Cc: linux-unionfs@xxxxxxxxxxxxxxx > Cc: Miklos Szeredi <miklos@xxxxxxxxxx> > Cc: Amir Goldstein <amir73il@xxxxxxxxx> > Cc: Vivek Goyal <vgoyal@xxxxxxxxxx> > --- > Documentation/filesystems/overlayfs.rst | 6 +++++- > 1 file changed, 5 insertions(+), 1 deletion(-) > > diff --git a/Documentation/filesystems/overlayfs.rst b/Documentation/filesystems/overlayfs.rst > index 580ab9a0fe31..c6e30c1bc2f2 100644 > --- a/Documentation/filesystems/overlayfs.rst > +++ b/Documentation/filesystems/overlayfs.rst > @@ -570,7 +570,11 @@ Volatile mount > This is enabled with the "volatile" mount option. Volatile mounts are not > guaranteed to survive a crash. It is strongly recommended that volatile > mounts are only used if data written to the overlay can be recreated > -without significant effort. > +without significant effort. In addition to this, the sync family of syscalls > +are not sufficient to determine whether a write failed as sync calls are > +omitted. For this reason, it is important that the filesystem used by the > +upperdir handles failure in a fashion that's suitable for the user. For > +example, upon detecting a fault, ext4 can be configured to panic. > Reading this now, I think I may have wrongly analysed the issue. Specifically, when I wrote that the very minimum is to document the issue, it was under the assumption that a proper fix is hard. I think I was wrong and that the very minimum is to check for errseq since mount on the fsync and syncfs calls. Why? first of all because it is very very simple and goes a long way to fix the broken contract with applications, not the contract about durability obviously, but the contract about write+fsync+read expects to find the written data (during the same mount era). Second, because the sentence you added above is hard for users to understand and out of context. If we properly handle the writeback error in fsync/syncfs, then this sentence becomes irrelevant. The fact that ext4 can lose data if application did not fsync after write is true for volatile as well as non-volatile and it is therefore not relevant in the context of overlayfs documentation at all. Am I wrong saying that it is very very simple to fix? Would you mind making that fix at the bottom of the patch series, so it can be easily applied to stable kernels? Thanks, Amir.
