Received-SPF: pass (google.com: domain of linux-nfs+bounces-3288-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.48.161 as permitted sender) client-ip=147.75.48.161;
Content-Type: text/plain; charset="utf-8"
Content-Transfer-Encoding: 7bit
Precedence: bulk
MIME-Version: 1.0
From: "NeilBrown" <neilb@suse.de>
To: linux-nfs@vger.kernel.org
Subject: Re: When are layouts destroyed on server reboot
In-reply-to: <171617508551.18863.10945867499281791407@noble.neil.brown.name>
References: <171617508551.18863.10945867499281791407@noble.neil.brown.name>
Date: Mon, 20 May 2024 13:52:42 +1000
Message-id: <171617716265.18863.8448767801936248728@noble.neil.brown.name>

On Mon, 20 May 2024, NeilBrown wrote:
> 
> Hi,
>  I have a customer who is experiencing a problem on an older kernel.
>  After a server restart (actually a cluster fail-over) the client
>  recovers all opens and locks correctly, but then sends a (previously
>  queued) WRITE request with a filehandle/stateid for a LAYOUT that was
>  provided by the server before the restart.  Unsurprisingly this doesn't
>  work.
> 
>  I've been hunting through the code to find out where the code attempts
>  to invalidate all layouts as required by 12.7.4 in RFC-5661.  But I
>  cannot find it.
> 
>  I'm guessing that adding a call to pnfs_destroy_layout() in
>  __nfs4_reclaim_open_state() would do the trick but maybe there is a
>  better way that is already implemented in later kernels, that I cannot
>  find.
> 
>  Any pointers?

I think I found it. nfs4_establish_lease() calls pnfs_destroy_layouts().
And that is present in the old kernel that isn't working... both.  Must
be something else.
Sorry for the noise.

NeilBrown