2023-09-12 21:31:33

by Chuck Lever III

[permalink] [raw]
Subject: Re: [PATCH] nfsd: Don't reset the write verifier on a commit EAGAIN



> On Sep 11, 2023, at 7:42 PM, Trond Myklebust <[email protected]> wrote:
>
> On Mon, 2023-09-11 at 22:10 +0000, Chuck Lever III wrote:
>>
>>> On Sep 11, 2023, at 4:54 PM, Trond Myklebust
>>> <[email protected]> wrote:
>>>
>>> On Mon, 2023-09-11 at 16:14 -0400, Chuck Lever wrote:
>>>> On Mon, Sep 11, 2023 at 02:43:57PM -0400,
>>>> [email protected] wrote:
>>>>> From: Trond Myklebust <[email protected]>
>>>>>
>>>>> If fsync() is returning EAGAIN, then we can assume that the
>>>>> filesystem
>>>>> being exported is something like NFS with the 'softerr' mount
>>>>> option
>>>>> enabled, and that it is just asking us to replay the fsync()
>>>>> operation
>>>>> at a later date.
>>>>> If we see an ESTALE, then ditto: the file is gone, so there is
>>>>> no
>>>>> danger
>>>>> of losing the error.
>>>>> For those cases, do not reset the write verifier.
>>>>
>>>> Out of interest, what's the hazard in a write verifier change in
>>>> these cases? There could be a slight performance penalty, I
>>>> imagine,
>>>> but how frequently does this happen?
>>>
>>> When re-exporting to NFSv4 clients, it should be less of a problem,
>>> since any REMOVE will result in a sillyrenamed file that only
>>> disappears once the file is closed. However with NFSv3 clients,
>>> that is
>>> circumvented by the fact that the filecache closes the files when
>>> they
>>> are inactive. We've seen this occur frequently with VMware vmdks:
>>> their
>>> lock files appear to generate a lot of these phantom ESTALE writes.
>>>
>>> As for EAGAIN, I just pushed out a 2 patch client series that makes
>>> it
>>> a lot more frequent when re-exporting NFSv4 with 'softerr'.
>>>
>>> Finally, it is worth noting that a write verifier change has a
>>> global
>>> effect, causing retransmission by all clients of all uncommitted
>>> unstable writes for all files, so is worth mitigating where
>>> possible.
>>
>> Good info. I've added some of this to the patch description.
>>
>>
>>>> One more below.
>>>>
>>>>
>>>>> Signed-off-by: Trond Myklebust
>>>>> <[email protected]>
>>>>> ---
>>>>> fs/nfsd/vfs.c | 29 +++++++++++++++++++----------
>>>>> 1 file changed, 19 insertions(+), 10 deletions(-)
>>>>>
>>>>> diff --git a/fs/nfsd/vfs.c b/fs/nfsd/vfs.c
>>>>> index 98fa4fd0556d..31daf9f63572 100644
>>>>> --- a/fs/nfsd/vfs.c
>>>>> +++ b/fs/nfsd/vfs.c
>>>>> @@ -337,6 +337,20 @@ nfsd_lookup(struct svc_rqst *rqstp, struct
>>>>> svc_fh *fhp, const char *name,
>>>>> return err;
>>>>> }
>>>>>
>>>>> +static void
>>>>> +commit_reset_write_verifier(struct nfsd_net *nn, struct
>>>>> svc_rqst
>>>>> *rqstp,
>>>>> + int err)
>>>>> +{
>>>>> + switch (err) {
>>>>> + case -EAGAIN:
>>>>> + case -ESTALE:
>>>>> + break;
>>>>> + default:
>>>>> + nfsd_reset_write_verifier(nn);
>>>>> + trace_nfsd_writeverf_reset(nn, rqstp, err);
>>>>> + }
>>>>> +}
>>>>> +
>>>>> /*
>>>>> * Commit metadata changes to stable storage.
>>>>> */
>>>>> @@ -647,8 +661,7 @@ __be32 nfsd4_clone_file_range(struct
>>>>> svc_rqst
>>>>> *rqstp,
>>>>>
>>>>> &nfsd4_get_cstate(rqstp)-
>>>>>> current_fh,
>>>>> dst_pos,
>>>>> count, status);
>>>>> - nfsd_reset_write_verifier(nn);
>>>>> - trace_nfsd_writeverf_reset(nn, rqstp,
>>>>> status);
>>>>> + commit_reset_write_verifier(nn, rqstp,
>>>>> status);
>>>>> ret = nfserrno(status);
>>>>> }
>>>>> }
>>>>> @@ -1170,8 +1183,7 @@ nfsd_vfs_write(struct svc_rqst *rqstp,
>>>>> struct
>>>>> svc_fh *fhp, struct nfsd_file *nf,
>>>>> host_err = vfs_iter_write(file, &iter, &pos, flags);
>>>>> file_end_write(file);
>>>>> if (host_err < 0) {
>>>>> - nfsd_reset_write_verifier(nn);
>>>>> - trace_nfsd_writeverf_reset(nn, rqstp,
>>>>> host_err);
>>>>> + commit_reset_write_verifier(nn, rqstp,
>>>>> host_err);
>>>>
>>>> Can generic_file_write_iter() or its brethren return STALE or
>>>> AGAIN
>>>> before they get to the generic_write_sync() call ?
>>>
>>> The call to nfs_revalidate_file_size(), which can occur when you
>>> are
>>> appending to the file (whether or not O_APPEND is set) could indeed
>>> return ESTALE.
>>> With the new patchset mentioned above, it could also return EAGAIN.
>>
>> Sounds like I should drop this hunk when applying this fix.
>
> I'm not understanding. Why would you not keep it?

generic_file_write_iter() and its brethren are two calls in
one, if I'm following this correctly:

1. write
2. sync

All the other places you change are "sync" only, so it's
fairly obvious that those callers get a return code that
reflects a failure of "sync".

I asked above if it's possible for the "write" part of
generic_file_write_iter() to fail with STALE/AGAIN before the
sync part is even called.

You seemed to be answering "yes, the 'write' part can fail
that way" but I may have misunderstood your response.

If the "write" step can fail, isn't that something that should
be reflected in a write verifier change? If yes, I don't see
how this particular call site can distinguish between a "write"
failure versus a "sync" failure.

Or, if the vfs_iter_write() call here is guaranteed to never
be a sync write request, then again, I think we want to reflect
all failures here with a write verifier change.

However, if STALE and AGAIN have the exact same semantics
for "write" as they do for "sync", those failures can be
thrown away too, and I can keep this hunk. Are you saying
this is the case?

(this is /only/ for the vfs_iter_write() call site. The others
look OK to me).


>>>>> goto out_nfserr;
>>>>> }
>>>>> *cnt = host_err;
>>>>> @@ -1183,10 +1195,8 @@ nfsd_vfs_write(struct svc_rqst *rqstp,
>>>>> struct svc_fh *fhp, struct nfsd_file *nf,
>>>>>
>>>>> if (stable && use_wgather) {
>>>>> host_err = wait_for_concurrent_writes(file);
>>>>> - if (host_err < 0) {
>>>>> - nfsd_reset_write_verifier(nn);
>>>>> - trace_nfsd_writeverf_reset(nn, rqstp,
>>>>> host_err);
>>>>> - }
>>>>> + if (host_err < 0)
>>>>> + commit_reset_write_verifier(nn, rqstp,
>>>>> host_err);
>>>>> }
>>>>>
>>>>> out_nfserr:
>>>>> @@ -1329,8 +1339,7 @@ nfsd_commit(struct svc_rqst *rqstp,
>>>>> struct
>>>>> svc_fh *fhp, struct nfsd_file *nf,
>>>>> err = nfserr_notsupp;
>>>>> break;
>>>>> default:
>>>>> - nfsd_reset_write_verifier(nn);
>>>>> - trace_nfsd_writeverf_reset(nn, rqstp,
>>>>> err2);
>>>>> + commit_reset_write_verifier(nn, rqstp,
>>>>> err2);
>>>>> err = nfserrno(err2);
>>>>> }
>>>>> } else
>>>>> --
>>>>> 2.41.0
>>>>>
>>>>
>>>
>>> --
>>> Trond Myklebust
>>> Linux NFS client maintainer, Hammerspace
>>> [email protected]
>>
>>
>> --
>> Chuck Lever
>>
>>
>
> --
> Trond Myklebust
> Linux NFS client maintainer, Hammerspace
> [email protected]
>
>

--
Chuck Lever