On Fri, Oct 11, 2019 at 2:50 PM Olga Kornievskaia <aglo@xxxxxxxxx> wrote: > > On Fri, Oct 11, 2019 at 10:18 AM Trond Myklebust > <trondmy@xxxxxxxxxxxxxxx> wrote: > > > > On Thu, 2019-10-10 at 13:32 -0400, Olga Kornievskaia wrote: > > > On Thu, Oct 10, 2019 at 3:42 AM Murphy Zhou <jencce.kernel@xxxxxxxxx> > > > wrote: > > > > Since commit: > > > > [0e0cb35] NFSv4: Handle NFS4ERR_OLD_STATEID in > > > > CLOSE/OPEN_DOWNGRADE > > > > > > > > xfstests generic/168 on v4.2 starts to fail because reflink call > > > > gets: > > > > +XFS_IOC_CLONE_RANGE: Resource temporarily unavailable > > > > > > I don't believe this failure has to do with getting ERR_OLD_STATEID > > > on > > > the CLOSE. What you see on the network trace is expected as the > > > client > > > in parallel sends OPEN/CLOSE thus server will fail the CLOSE with the > > > ERR_OLD_STATEID since it already updated its stateid for the OPEN. > > > > > > > In tshark output, NFS4ERR_OLD_STATEID stands out when comparing > > > > with > > > > good ones: > > > > > > > > 5210 NFS 406 V4 Reply (Call In 5209) OPEN StateID: 0xadb5 > > > > 5211 NFS 314 V4 Call GETATTR FH: 0x8d44a6b1 > > > > 5212 NFS 250 V4 Reply (Call In 5211) GETATTR > > > > 5213 NFS 314 V4 Call GETATTR FH: 0x8d44a6b1 > > > > 5214 NFS 250 V4 Reply (Call In 5213) GETATTR > > > > 5216 NFS 422 V4 Call WRITE StateID: 0xa818 Offset: 851968 Len: > > > > 65536 > > > > 5218 NFS 266 V4 Reply (Call In 5216) WRITE > > > > 5219 NFS 382 V4 Call OPEN DH: 0x8d44a6b1/ > > > > 5220 NFS 338 V4 Call CLOSE StateID: 0xadb5 > > > > 5222 NFS 406 V4 Reply (Call In 5219) OPEN StateID: 0xa342 > > > > 5223 NFS 250 V4 Reply (Call In 5220) CLOSE Status: > > > > NFS4ERR_OLD_STATEID > > > > 5225 NFS 338 V4 Call CLOSE StateID: 0xa342 > > > > 5226 NFS 314 V4 Call GETATTR FH: 0x8d44a6b1 > > > > 5227 NFS 266 V4 Reply (Call In 5225) CLOSE > > > > 5228 NFS 250 V4 Reply (Call In 5226) GETATTR > > > > > > "resource temporarily unavailable" is more likely to do with ulimit > > > limits. > > > > > > I also saw the same error. After I increased the ulimit for the stack > > > size, the problem went away. There might still be a problem somewhere > > > in the kernel. > > > > > > Trond, is it possible that we have too many CLOSE recovery on the > > > stack that's eating up stack space? > > > > That shouldn't normally happen. CLOSE runs as an asynchronous RPC call, > > so its stack usage should be pretty minimal (limited to whatever each > > callback function uses). > > Yeah, that wasn't it. I've straced generic/168 to catch > ioctl(clone_file_range) returning EAGAIN. > > I've instrumented the kernel to see where we are returning an EAGAIN > in nfs42_proc_clone(). nfs42_proc_clone is failing on > nfs42_select_rw_context() because nfs4_copy_open_stateid() is failing > to get the open state. Basically it looks like we are trying to do a > clone on a file that's not opened. Still trying to understand > things... Trond, Generic/168 fails in 2 ways (though only 1 leads to the failure in xfs_io). Another way is having a file closed then client using the stateid for the write and getting a bad_stateid which the client recovers from (but the fact that client shouldn't have done that is a problem). Another is the clone where again happens that file is "closed" and clone is trying to use a stateid. The problem comes from the following fact. We have a racing CLOSE and OPEN. Where client did the CLOSE followed by the OPEN but the server processed OPEN and then the CLOSE. Server returns OLD_STATEID to the CLOSE. What the code does it bumps the sequence id and resends the CLOSE which inadvertently is closing a file that was opened before. While IO errors from this are recoverable, the clone error is visible to the application (I think another case would be a copy). I don't have a code solution yet. But it'll have to be something where we need to ignore a CLOSE with OLD_STATEID when another OPEN happened.
