On 3 Sep 2020, at 23:04, Murphy Zhou wrote: > Hi Benjamin, > > On Thu, Sep 03, 2020 at 01:54:26PM -0400, Benjamin Coddington wrote: >> >> On 11 Oct 2019, at 10:14, Trond Myklebust wrote: >>> On Fri, 2019-10-11 at 16:49 +0800, Murphy Zhou wrote: >>>> On Thu, Oct 10, 2019 at 02:46:40PM +0000, Trond Myklebust wrote: >>>>> On Thu, 2019-10-10 at 15:40 +0800, Murphy Zhou wrote: >> ... >>>>>> @@ -3367,14 +3368,16 @@ static bool >>>>>> nfs4_refresh_open_old_stateid(nfs4_stateid *dst, >>>>>> break; >>>>>> } >>>>>> seqid_open = state->open_stateid.seqid; >>>>>> - if (read_seqretry(&state->seqlock, seq)) >>>>>> - continue; >>>>>> >>>>>> dst_seqid = be32_to_cpu(dst->seqid); >>>>>> - if ((s32)(dst_seqid - be32_to_cpu(seqid_open)) >= 0) >>>>>> + if ((s32)(dst_seqid - be32_to_cpu(seqid_open)) > 0) >>>>>> dst->seqid = cpu_to_be32(dst_seqid + 1); >>>>> >>>>> This negates the whole intention of the patch you reference in the >>>>> 'Fixes:', which was to allow us to CLOSE files even if seqid bumps >>>>> have >>>>> been lost due to interrupted RPC calls e.g. when using 'soft' or >>>>> 'softerr' mounts. >>>>> With the above change, the check could just be tossed out >>>>> altogether, >>>>> because dst_seqid will never become larger than seqid_open. >>>> >>>> Hmm.. I got it wrong. Thanks for the explanation. >>> >>> So to be clear: I'm not saying that what you describe is not a problem. >>> I'm just saying that the fix you propose is really no better than >>> reverting the entire patch. I'd prefer not to do that, and would rather >>> see us look for ways to fix both problems, but if we can't find such as >>> fix then that would be the better solution. >> >> Hi Trond and Murphy Zhou, >> >> Sorry to resurrect this old thread, but I'm wondering if any progress was >> made on this front. > > This failure stoped showing up since v5.6-rc1 release cycle > in my records. Can you reproduce this on latest upstream kernel? I'm seeing it on generic/168 on a v5.8 client against a v5.3 knfsd server. When I test against v5.8 server, the test takes longer to complete and I have yet to reproduce the livelock. - on v5.3 server takes ~50 iterations to produce, each test completes in ~40 seconds - on v5.8 server my test has run ~750 iterations without getting into the lock, each test takes ~60 seconds. I suspect recent changes to the server have changed the timing of open replies such that the problem isn't reproduced on the client. Ben
