On Mon, 2023-06-12 at 19:53 +0000, Trond Myklebust wrote:
> On Mon, 2023-06-12 at 15:17 -0400, Jeff Layton wrote:
> > On Mon, 2023-06-12 at 13:49 -0400, Chris Perl wrote:
> > > On Mon, Jun 12, 2023 at 1:30 PM Jeff Layton <jlayton@xxxxxxxxxx>
> > > wrote:
> > > >
> > > > On Mon, 2023-06-12 at 11:58 -0400, Jeff Layton wrote:
> > > >
> > > > >
> > > > > Got it: I think I see what's happening. filemap_sample_wb_err
> > > > > just calls
> > > > > errseq_sample, which does this:
> > > > >
> > > > > errseq_t errseq_sample(errseq_t *eseq)
> > > > > {
> > > > > errseq_t old = READ_ONCE(*eseq);
> > > > >
> > > > > /* If nobody has seen this error yet, then we can be
> > > > > the first. */
> > > > > if (!(old & ERRSEQ_SEEN))
> > > > > old = 0;
> > > > > return old;
> > > > > }
> > > > >
> > > > > Because no one has seen that error yet (ERRSEQ_SEEN is clear),
> > > > > the write
> > > > > ends up being the first to see it and it gets back a 0, even
> > > > > though the
> > > > > error happened before the sample.
> > > > >
> > > > > The above behavior is what we want for the sample that we do at
> > > > > open()
> > > > > time, but not what's needed for this use-case. We need a new
> > > > > helper that
> > > > > samples the value regardless of whether it has already been
> > > > > seen:
> > > > >
> > > > > errseq_t errseq_peek(errseq_t *eseq)
> > > > > {
> > > > > return READ_ONCE(*eseq);
> > > > > }
> > > > >
> > > > > ...but we'll also need to fix up errseq_check to handle
> > > > > differences
> > > > > between the SEEN bit.
> > > > >
> > > > > I'll see if I can spin up a patch for that. Stay tuned.
> > > >
> > > > This may not be fixable with the way that NFS is trying to use
> > > > errseq_t.
> > > >
> > > > The fundamental problem is that we need to mark the errseq_t in
> > > > the
> > > > mapping as SEEN when we sample it, to ensure that a later error
> > > > is
> > > > recorded and not ignored.
> > > >
> > > > But...if the error hasn't been reported yet and we mark it SEEN
> > > > here,
> > > > and then a later error doesn't occur, then a later open won't
> > > > have its
> > > > errseq_t set to 0, and that unseen error could be lost.
> > > >
> > > > It's a bit of a pity: as originally envisioned, the errseq_t
> > > > mechanism
> > > > would provide for this sort of use case, but we added this patch
> > > > not
> > > > long after the original code went in, and it changed those
> > > > semantics:
> > > >
> > > > b4678df184b3 errseq: Always report a writeback error once
> > > >
> > > > I don't see a good way to do this using the current errseq_t
> > > > mechanism,
> > > > given these competing needs. I'll keep thinking about it though.
> > > > Maybe
> > > > we could add some sort of store and forward mechanism for fsync
> > > > on NFS?
> > > > That could get rather complex though.
> > >
> > > Can/should it be marked SEEN when the initial close(2) from tee(1)
> > > reports the error?
> > >
> >
> > No. Most software doesn't check for errors on close(), and for good
> > reason: there's no requirement that any data be written back before
> > close() returns. A successful return is meaningless.
> >
> > It turns out that NFSv4 (usually) writes back the data before a close
> > returns, but you don't want to rely on that.
> >
> > > Part of the reason I had originally asked about `nfs_file_flush'
> > > (i.e.
> > > what close(2) calls) using `file_check_and_advance_wb_err' instead
> > > of
> > > `filemap_check_wb_err' was because I was drawn to comparing
> > > `nfs_file_flush' against `nfs_file_fsync' as it seems like in the
> > > 3.10
> > > based EL7 kernels, the former used to delegate to the latter (by
> > > way
> > > of `vfs_fsync') and so they had consistent behavior, whereas now
> > > they
> > > do not.
> >
> > I think the problem is in some of the changes to write that have come
> > into play since then. They tried to use errseq_t to track errors over
> > a
> > small window, but the underlying infrastructure is not quite suited
> > for
> > that at the moment.
> >
> > I think we can get there though by carving another flag bit out of
> > the
> > counter in the errseq_t. I'm working on a patch for that now.
> >
>
> The current NFS client code tries to do its best to match the
> description in the manpages for how errors are reported: we try to
> report them exactly once, either in write() or fsync().
> We do still return errors on close(), but that kind of opportunistic
> error return makes sure to use filemap_check_wb_err() so that we don't
> break the write() + fsync() documented semantics.
>
> The issue of picking up errors using errseq_sample() before even any
> I/O has been attempted has been raised before, but AFAIK, the current
> behaviour does actually match the promises made in the manpages, and it
> matches what can happen with other filesystems.
> I don't want to special case the NFS client, because that just leads to
> people getting confused as to whether or not it will work correctly
> with applications such as postgresql.
>
The point here would be to bring NFS more into line with how other
filesystems behave. As Chris pointed out, other filesystems don't report
an error on a new write() just because there was an earlier, unseen
writeback error on the same inode.
I think we can achieve this by carving out another flag bit from the
errseq_t counter. I'm building and testing a patch now, and I'll post it
once I'm convinced it's sane.
Cheers,
--
Jeff Layton <jlayton@xxxxxxxxxx>
