Re: NFS/TCP crashes on MIPS/RBTX4927 in v4.20-rcX (bisected)

[Trond Myklebust <trondmy@hammerspace.com>]

Hi Geert,

On Mon, 2018-12-17 at 15:03 +0100, Geert Uytterhoeven wrote:
> Hi Trond,
> 
> On Wed, Dec 5, 2018 at 3:47 PM Geert Uytterhoeven <
> geert@linux-m68k.org> wrote:
> > On Wed, Dec 5, 2018 at 2:45 PM Trond Myklebust <
> > trondmy@hammerspace.com> wrote:
> > > On Wed, 2018-12-05 at 14:41 +0100, Geert Uytterhoeven wrote:
> > > > On Wed, Dec 5, 2018 at 2:11 PM Atsushi Nemoto <
> > > > anemo@mba.ocn.ne.jp>
> > > > wrote:
> > > > > On Tue, 4 Dec 2018 14:53:07 +0100, Geert Uytterhoeven <
> > > > > geert@linux-m68k.org> wrote:
> > > > > > I found similar crashes in a report from 2006, but of
> > > > > > course the
> > > > > > code
> > > > > > has changed too much to apply the solution proposed there
> > > > > > (
> > > > > > https://www.linux-mips.org/archives/linux-mips/2006-09/msg00169.html
> > > > > > ).
> > > > > > 
> > > > > > Userland is Debian 8 (the last release supporting "old"
> > > > > > MIPS).
> > > > > > My kernel is based on v4.20.0-rc5, but the issue happens
> > > > > > with
> > > > > > v4.20-rc1,
> > > > > > too.
> > > > > > 
> > > > > > However, I noticed it works in v4.19! Hence I've bisected
> > > > > > this,
> > > > > > to commit
> > > > > > 277e4ab7d530bf28 ("SUNRPC: Simplify TCP receive code by
> > > > > > switching
> > > > > > to using
> > > > > > iterators").
> > > > > > 
> > > > > > Dropping the ",tcp" part from the nfsroot parameter also
> > > > > > fixes
> > > > > > the issue.
> > > > > > 
> > > > > > Given RBTX4927 is little endian, just like my arm/arm64
> > > > > > boards,
> > > > > > it's probably
> > > > > > not an endianness issue.  Sparse didn't show anything
> > > > > > suspicious
> > > > > > before/after
> > > > > > the guilty commit.
> > > > > > 
> > > > > > Do you have a clue?
> > > > > 
> > > > > If it was a cache issue, disabling i-cache or d-cache
> > > > > completely
> > > > > might
> > > > > help understanding the problem.  I added TXx9 specific
> > > > > "icdisable"
> > > > > and
> > > > > "dcdisable" kernel options for debugging long ago.
> > > > > 
> > > > > I hope these options still works correctly with recent kernel
> > > > > but
> > > > > not
> > > > > sure.
> > > > > 
> > > > > Also, disabling i-cache makes your board VERY slow, of
> > > > > course.
> > > > 
> > > > Thanks!
> > > > 
> > > > When using these options, I do see a slowdown in early boot,
> > > > but the
> > > > issue
> > > > is still there.
> > > > 
> > > > My next guess is an unaligned access not using
> > > > {get,put}_unaligned(),
> > > > which
> > > > doesn't seem to work on tx4927, but doesn't cause an exception
> > > > neither.
> > > 
> > > Can you try my linux-next branch on git.linux-nfs.org? It
> > > contains a
> > > fixes for a hang that results from the above commit.
> > > 
> > > git pull git://git.linux-nfs.org/projects/trondmy/linux-nfs.git
> > > linux-next
> > 
> > Thanks for the suggestion, but unfortunately it doesn't help.
> 
> In the mean time, I tried your newer linux-next, no change.
> I tried several other things:
>   - remove the packed attribute (why did you add that?),

The packed attribute allows us to avoid a series of copy operations
when decoding the first three elements of a RPC over TCP header (which
is why they are all declared as big endian). The alternative would be
to have a 12 byte buffer there for temporary storage, and then a
duplicate set of 3 32-bit words into which we copy the buffer contents
after extracting them from the (non-blocking) socket.

>   - verify (at runtime) that all accesses to fraghdr, xid, and
> calldir
> are aligned,
>   - enable RPC_DEBUG_DATA, nothing fishy seen at first sight.
> 
> Is anyone else seeing this on MIPS, or any other platform?
> Does mounting NFS with -o nfsvers=3,tcp work on other MIPS platforms?

I have no access to any MIPS hardware for the purposes of testing so
that would be a question for the community.

One thing that I have noticed is that unlike the old code, the bvec
'generic' code does appear to fail to call flush_dcache_page(). Could
that be causing the problem here? If so, why would that not be a
problem in the context of regular block I/O?

Cheers
  Trond
-- 
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trond.myklebust@hammerspace.com