still nfs problems [Was: Linux 2.6.37-rc8]

[Trond Myklebust]

On Wed, 2011-01-05 at 16:01 +0100, Marc Kleine-Budde wrote: 
> On 01/05/2011 03:53 PM, Trond Myklebust wrote:
> > On Wed, 2011-01-05 at 14:40 +0100, Uwe Kleine-König wrote: 
> >> Hi Russell,
> >>
> >> On Wed, Jan 05, 2011 at 11:27:01AM +0000, Russell King - ARM Linux wrote:
> >>> On Wed, Jan 05, 2011 at 12:05:17PM +0100, Uwe Kleine-König wrote:
> >>>> Hello Trond,
> >>>>
> >>>> On Wed, Jan 05, 2011 at 09:40:14AM +0100, Uwe Kleine-König wrote:
> >>>>> On Mon, Jan 03, 2011 at 07:22:38PM -0500, Trond Myklebust wrote:
> >>>>>> The question is whether this is something happening on the server or the
> >>>>>> client. Does an older client kernel boot without any trouble?
> >>>>> I will set up a boot test with 2.6.37 (for statistics) and 2.6.36 to
> >>>>> compare with.  If you don't consider .36 to be old enough let me now.
> >>>>> Once the setup is done it should be easy to test .35 (say), too.
> >>>>>
> >>>> Marc (cc'd) saw similar[1] problems with .37, when using .36.2 the
> >>>> problems didn't occur.  This was more reliable to trigger and he was so
> >>>> kind to bisect the problem.
> >>>>
> >>>> When testing v2.6.36-rc3-51-gafa8ccc init hanged.
> >>>> (babddc72a9468884ce1a23db3c3d54b0afa299f0 is the first bad commit with
> >>>> this hang.)  Commit 56e4ebf877b6043c289bda32a5a7385b80c17dee makes the
> >>>> "init hangs" problem the "fileid changed on tab" problem.
> >>>>
> >>>> I could only reproduce that on armv5 machines (imx27, imx28 and at91)
> >>>> but not on armv6 (imx35).
> >>>
> >>> FYI, I've seen the "fileid changed" problem, and it looked like a 32-bit
> >>> truncation of the fileid.  It occurred several times on successive
> >>> reboots, so I tried to capture a tcpdump trace off the server (Linux
> >>> 2.6.23-rc8-ga64314e6 - its ancient because I've had issues with buggy
> >>> IDE drivers trying to move it forward.)  However, for the last couple
> >>> of weeks I've been unable to reproduce it.
> >>>
> >>> The client was based on 2.6.37-rc6.
> >>>
> >>> The "fileid changed" messages popped up after mounting an export with
> >>> 'nolock,intr,rsize=4096,soft', and then trying to use bash completion
> >>> and 'ls' in a few subdirectories - and entries were missing from the
> >>> directory lists without 'ls' reporting any errors (which I think is bad
> >>> behaviour in itself.)
> >> There was a bug in at least -rc5[1] that was considered already fixed in
> >> -rc4[2]. The later announcements didn't mention it anymore. 
> >>
> >>> I don't know why it's stopped producing the errors, although once it
> >>> went I never investigated it any further (was far too busy trying to
> >>> get AMBA DMA support working.)
> >> It seems it was fixed for most users though. Trond?
> > 
> > As I said, I can't reproduce it.
> > 
> > I'm seeing a lot of mention of ARM above. Is anyone seeing this bug on
> > x86, or does it appear to be architecture-specific?
> 
> It _seems_ to be ARMv5 specific[1]. Uwe did some tests and figured out
> that disabling dcache on ARMv5 "fixes" the problem, but
> CONFIG_CPU_DCACHE_WRITETHROUGH isn't enough.
> 
> [1] Uwe fails to reproduce it on ARMv6. The ARMv6 has a L2 cache and
> uses IIRC different instructions to flush the L1 caches. (please correct
> me, if I'm wrong, ARM guys :)
> 
> cheers, Marc

OK. So,the new behaviour in 2.6.37 is that we're writing to a series of
pages via the usual kmap_atomic()/kunmap_atomic() and kmap()/kunmap()
interfaces, but we can end up reading them via a virtual address range
that gets set up via vm_map_ram() (that range gets set up before the
write occurs).

Do we perhaps need an invalidate_kernel_vmap_range() before we can read
the data on ARM in this kind of scenario?

-- 
Trond Myklebust
Linux NFS client maintainer

NetApp
Trond.Myklebust at netapp.com
www.netapp.com