Skip site navigation (1)Skip section navigation (2)
Date:      Mon, 1 Feb 2010 09:46:57 -0500
From:      John Baldwin <jhb@freebsd.org>
To:        Marius Strobl <marius@alchemy.franken.de>
Cc:        rmacklem@freebsd.org, dfr@freebsd.org, freebsd-stable@freebsd.org, Peter Jeremy <peterjeremy@acm.org>
Subject:   Re: uma_zalloc_arg complaining about non-sleepable locks
Message-ID:  <201002010946.57253.jhb@freebsd.org>
In-Reply-To: <20100131162854.GC77522@alchemy.franken.de>
References:  <20100126073336.GA1955@server.vk2pj.dyndns.org> <20100131010618.GA1864@server.vk2pj.dyndns.org> <20100131162854.GC77522@alchemy.franken.de>

next in thread | previous in thread | raw e-mail | index | archive | help
On Sunday 31 January 2010 11:28:54 am Marius Strobl wrote:
> On Sun, Jan 31, 2010 at 12:06:18PM +1100, Peter Jeremy wrote:
> > Sorry for the delay, I was trying to avoid rebooting my server.
> > I've setup a similar environment in VirtualBox to test it.
> > 
> > On 2010-Jan-27 12:52:29 +0100, Marius Strobl <marius@alchemy.franken.de> 
wrote:
> > >Ah, I forgot that using nfsm_aligned() causes nfs_realign() to
> > >be a NOP on architectures without strict alignment requirements
> > >for performance reasons. That's generally fine but unfortunately
> > >that way you don't actually exercise the code which caused the
> > >problem before (unfortunately I still don't manage to hit the
> > >unaligned case myself).
> > 
> > >Could you please test with #ifdef __NO_STRICT_ALIGNMENT replaced
> > >with #if 0 in sys/nfs/nfs_common.h? The vfs.nfs.realign_count
> > >counter should also increase then.
> > 
> > I'm not sure what triggers the unaligned case either - I tried
> > roughly "tar -cf - -C /mnt/usr src | tar -xf - -C /mnt/tmp" and
> > that caused some unaligned accesses (but also completely wedged
> > the VBox host).  I also tried copying a pile of files off my
> > NFS client (FreeBSD-8.x/i386) and that also triggered some
> > unaligned accesses without any errors being reported.
> > 
> > Currently, I have:
> > vfs.nfs.realign_count: 12
> > vfs.nfs.realign_test: 188817
> > 
> > I'd say that your patch works.
> 
> John, are you okay with that patch?
> http://people.freebsd.org/~marius/fha_extract_info_realign2.diff
> 
> It's intention is to:
> - Move nfs_realign() from the NFS client to the shared NFS code and
>   remove the NFS server version in order to reduce code duplication.
>   The shared version now uses a second parameter how, which is passed
>   on to m_get(9) and m_getcl(9) as the server used M_WAIT while the
>   client requires M_DONTWAIT, and replaces the the previously unused
>   parameter hsiz.
> - Change nfs_realign() to use nfsm_aligned() so as with other NFS code
>   the alignment check isn't actually performed on platforms without
>   strict alignment requirements for performance reasons because as the
>   comment suggests only occasionally occurs with TCP.
> - Change fha_extract_info() to use nfs_realign() with M_NOWAIT rather
>   than M_DONTWAIT because it's called with the RPC sp_lock held.
> 
> The only downside of the shared nfs_realign() are the combined
> SYSCTL counters but the fact we incremented them non-atomically
> so far I think already indicates that their intention only is a
> rough indication rather than exact values for client and server.

This all sounds good to me, but isn't M_NOWAIT == M_DONTWAIT?  Hmm, reading 
the code more closely, it looks like fha_extract_info() was using M_WAIT 
rather than M_DONTWAIT previously.  Also, I think you should be careful to use 
M_DONTWAIT instead of M_NOWAIT for mbufs, so I would fix that in 
fha_extract_info().

-- 
John Baldwin



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?201002010946.57253.jhb>