> On Sep 27, 2023, at 5:41 AM, Mantas Mikulėnas <grawity@xxxxxxxxx> wrote: > > On 27/09/2023 11.45, Mantas Mikulėnas wrote: >> On 26/09/2023 16.57, Chuck Lever III wrote: >>> I'm wondering if I can get you to bisect the server kernel using >>> your v5.10 client to test? good = v6.4, bad = v6.5 should do it. >>>> Yeah, I will try to bisect but it'll probably take a day or two. >> >> I'm *nearly* done with bisect (most of the builds with distro config took over an hour on this aging Xeon), and I'm currently in the middle of: > > Now it's done with: > > 703d7521555504b3a316b105b4806d641b7ebc76 is the first bad commit > commit 703d7521555504b3a316b105b4806d641b7ebc76 > Author: Chuck Lever <chuck.lever@xxxxxxxxxx> > Date: Thu May 18 13:46:03 2023 -0400 > > NFSD: Hoist rq_vec preparation into nfsd_read() [step two] That's even a plausible bisect result! The difference between the v5.10 client capture and the v6.4 capture you sent us is that the v5.10 client asks for only 4092 bytes in its NFS READ request. The v6.4 client asks for 4096, so the server bug is avoided. Olga's reproducer tickles the bug by using O_DIRECT to force the client to request exactly 4092 bytes. Let me take a closer look at this. -- Chuck Lever
