Re: Failing to send a CLOSE if file is opened WRONLY and server reboots on a 4.x mount

[Trond Myklebust <trond.myklebust@primarydata.com>]

On Mon, Sep 14, 2015 at 7:54 PM, Olga Kornievskaia <aglo@umich.edu> wrote:
> A test case is as the description says:
> open(foobar, O_WRONLY);
> sleep()  --> reboot the server
> close(foobar)
>
> The bug is because in nfs4state.c in nfs4_reclaim_open_state() a few
> line before going to restart, there is
> clear_bit(NFS4CLNT_RECLAIM_NOGRACE, &state->flags).
>
> NFS4CLNT_RECLAIM_NOGRACE is a flag for the client states not open
> owner states. Value of NFS4CLNT_RECLAIM_NOGRACE is 4 which is the
> value of NFS_O_WRONLY_STATE in nfs4_state->flags. So clearing it wipes
> out state and when we go to close it, “call_close” doesn’t get set as
> state flag is not set and CLOSE doesn’t go on the wire.
>
> That line was introduced to fix an infinite loop for OPEN recovery
> upon receiving a BAD_STATEID error: commit e8d975e73. I have tested
> injecting BAD_STATEID error using the patch below and the code
> recovers without problems. However, I'm not sure the clearing of the
> bit is needed any more. I have tested for infinite loop by reverting
> the patch and didn't hit the infinite loop.
>
> diff --git a/fs/nfs/nfs4state.c b/fs/nfs/nfs4state.c
> index da73bc4..5db3246 100644
> --- a/fs/nfs/nfs4state.c
> +++ b/fs/nfs/nfs4state.c
> @@ -1481,7 +1481,7 @@ restart:
>                                         spin_unlock(&state->state_lock);
>                                 }
>                                 nfs4_put_open_state(state);
> -                               clear_bit(NFS4CLNT_RECLAIM_NOGRACE,
> +                               clear_bit(NFS_STATE_RECLAIM_NOGRACE,
>                                         &state->flags);
>                                 spin_lock(&sp->so_lock);
>                                 goto restart;

That's an obvious typo. Thanks for spotting it!

As for whether or not the bit clear is needed at all, I think it is
for NFSv4 on older kernels. On newer kernels, we do have the NFSv4
state recovery drain the slot table (just like we've always done for
NFSv4.1) and so I agree that those kernels probably won't be
afflicted.

Cheers
  Trond