Re: v4.2-rc dcache regression, probably 75a6f82a0d10

[Dominique Martinet <dominique.martinet@cea.fr>]

Hi,

Hugh Dickins wrote on Fri, Jul 31, 2015:
> I think there's something not quite right with the fs/dcache.c
> commit 75a6f82a0d10 ("freeing unlinked file indefinitely delayed").
> 
> When running my old tmpfs swapping load (two repetitive make -j20
> kernel builds, one on tmpfs, one on ext4 over loop over tmpfs file,
> in limited memory with plenty of swapping; rm -rf of both trees
> in between the builds): one of the builds, always the ext4 so far,
> fails after several hours, one or another header file "No such file
> or directory", but the file's there when I check the tree afterwards.
> 
> Sounds like a dcache problem, and 75a6f82a0d10 seemed the only
> likely candidate, so I experimented with reverting it yesterday,
> and ran successfully for 24 hours.  That's a little too soon to
> be sure (I've set another run going this morning), but I'd say
> 90% certain that is to blame, and thought I'd better alert you
> sooner than later - you'll probably guess what's the matter
> long before I get back to check today's run.

Actually sounds like my thread "Race condition introduced in 4bf46a27
VFS: Impose ordering on accesses of d_inode and d_flags" (in fsdevel
only)

It's WAY easier to reproduce over a 9P/virtio mount and with multiple
level of caches (memory barrier), I was able to track it down to
4bf46a272647d ("VFS: Impose ordering on accesses of d_inode and
d_flags") and debug a bit, but this isn't code I'm familiar with so
would appreciate if you could tell me what you think.

-- 
Dominique Martinet