MIME-Version: 1.0
In-Reply-To: <CAN-5tyHGp4ihz3pzcmX+xjm5aba8SA3Pfxw-nxNTKv9tFikaGw@mail.gmail.com>
References: <CAN-5tyHxP276Re_Zm4mV9F+EY1Czq5SfbPJ1Os_CQd0hfhFzdA@mail.gmail.com>
	<CAHQdGtSg5E0jrEhK-aLht6WdmMTFFBUuG8tUX7sHJDUj+67-4Q@mail.gmail.com>
	<CAN-5tyHMCMekT0gJE1zNQLO0P5bVgjr_PuQF1BYhNy7n4KBvHg@mail.gmail.com>
	<CAHQdGtQfBZ4yRCkZ0-voFekTEx8fJ1kHCzrmtjr3RHh-dqOuww@mail.gmail.com>
	<CAN-5tyGme1rm+hD0JpOc3co0=_u2hzPwR7eN3AT4_TjVA4NkDA@mail.gmail.com>
	<CAN-5tyFZRU3zCSKU7wNoGzwDF9hn2Vvq6+JL9cWQnuhwoBftyg@mail.gmail.com>
	<CAHQdGtRzyiqjEj8Z2eBjk6ugzG8maK+F2BnO69JsX4ntCOzCcw@mail.gmail.com>
	<CAN-5tyHGp4ihz3pzcmX+xjm5aba8SA3Pfxw-nxNTKv9tFikaGw@mail.gmail.com>
Date: Tue, 13 Oct 2015 13:28:08 -0400
Message-ID: <CAHQdGtQJm6rFiNgF+SVCx_3SAnzBgHu8mXToWu_uNwJHYCZzqA@mail.gmail.com>
Subject: Re: [PATCH 1/1] Adding issync field to delegreturn_exit tracepoint
From: Trond Myklebust <trond.myklebust@primarydata.com>
To: Olga Kornievskaia <aglo@umich.edu>
Cc: linux-nfs <linux-nfs@vger.kernel.org>
Content-Type: text/plain; charset=UTF-8
Sender: linux-nfs-owner@vger.kernel.org

On Tue, Oct 13, 2015 at 1:23 PM, Olga Kornievskaia <aglo@umich.edu> wrote:
>
> On Tue, Oct 13, 2015 at 12:56 PM, Trond Myklebust
> <trond.myklebust@primarydata.com> wrote:
> > On Tue, Oct 13, 2015 at 12:24 PM, Olga Kornievskaia <aglo@umich.edu> wrote:
> >> On Tue, Oct 13, 2015 at 10:13 AM, Olga Kornievskaia <aglo@umich.edu> wrote:
> >>> On Tue, Oct 13, 2015 at 9:27 AM, Trond Myklebust
> >>> <trond.myklebust@primarydata.com> wrote:
> >>>> On Tue, Oct 13, 2015 at 8:26 AM, Olga Kornievskaia <aglo@umich.edu> wrote:
> >>>>>
> >>>>> On Mon, Oct 12, 2015 at 11:47 PM, Trond Myklebust
> >>>>> <trond.myklebust@primarydata.com> wrote:
> >>>>> > On Mon, Oct 12, 2015 at 5:55 PM, Olga Kornievskaia <aglo@umich.edu> wrote:
> >>>>> >> It'll be nice to know when we return delegations synchronously or not.
> >>>>> >
> >>>>> > Why? This patch forces us to carry an otherwise completely unnecessary
> >>>>> > parameter, so at the very minimum we should have a discussion of what
> >>>>> > the real use cases are.
> >>>>>
> >>>>> I used it to diagnose the race of open and delegreturn. If it's kept
> >>>>
> >>>> How were you using it?
> >>>
> >>>  I added two more traces points in the beginning of delegreturn and in
> >>> nfs4_do_open before sending the rpc. I can see that a given file
> >>> handle:
> >>> -- delegreturn prepare tracepoint is happening,
> >>> -- then the tracepoint of before sending the open is logged,
> >>> -- then delegreturn prepare is logged again,
> >>> -- then tracepoint for nfs4_open_file which is after receiving reply
> >>> to the open from the server
> >>> -- then delegreturn_exit tracepoint
> >>>
> >>>     kworker/1:0H-14168 [001] ....   576.571636:
> >>> nfs4_delegreturn_prepare: error=0 (OK) dev=00:2a fhandle=0x84792ca9
> >>> issync=0
> >>>
> >>>           hammer-13955 [000] ....   576.942632: nfs4_open_file_begin:
> >>> flags=32768 (0x8000) fmode=READ|0x801c fileid=00:2a:0
> >>> fhandle=0x00000000 name=00:2a:904/000002CB.ham
> >>>
> >>>           hammer-13955 [001] ....   577.043084: nfs4_open_file:
> >>> error=0 (OK) flags=32768 (0x8000) fmode=READ|0x801c fileid=00:2a:7708
> >>> fhandle=0x84792ca9 name=00:2a:904/000002CB.ham
> >>>
> >>>     kworker/0:1H-431   [000] ....   577.064013:
> >>> nfs4_delegreturn_prepare: error=0 (OK) dev=00:2a fhandle=0x84792ca9
> >>> issync=0
> >>>
> >>>     kworker/0:1H-431   [000] ....   577.101076: nfs4_delegreturn_exit:
> >>> error=0 (OK) dev=00:2a fhandle=0x84792ca9
> >>>
> >>>     kworker/0:1H-431   [000] ....   577.113021: nfs4_read:
> >>> error=-10025 (BAD_STATEID) fileid=00:2a:7708 fhandle=0x84792ca9
> >>> offset=0 count=64
> >>>
> >>>
> >>>>
> >>>>> that some delegreturns are synchronous and others are not I think the
> >>>>> information is useful.
> >>>>
> >>>> The only difference between synchronous and asynchronous in this case
> >>>> is whether or not the process that launched the delegreturn actually
> >>>> waits for it to complete; a signal could easily prevent it from doing
> >>>> so without interrupting the delegreturn call itself.
> >>>> IOW: for complete information when debugging races here, you really
> >>>> need to examine the return value from the wait call.
> >>>>
> >>>>> Speaking of there is a race between state manager thread returning
> >>>>> used delegations and new open. Previously I thought it was evict
> >>>>> inode...
> >>>>
> >>>> Is this with commit 5e99b532bb95 ("nfs4: reset states to use
> >>>> open_stateid when returning delegation voluntarily") applied?
> >>>
> >>> No I have not. I will try that. Thanks.
> >>
> >> This patch does not help. The race is still present.
> >
> > OK. So what are the symptoms? I'm having trouble seeing how a race can
> > happen, given a correctly coded server.
>
> Here's what the server sees:
> open (foobar) replies back with a delegation
> various operations including a close()
> some time goes by...
> open (foobar) replies back with the same delegation

Why? Olga, we already had this discussion. That sort of server
behaviour is never going to work without races and is the root cause
of your problem. We simply won't ever support servers that do this.

> delegreturn
> read (foobar) using delegation