Re: [PATCH] nfs: lookupcache coherence bugs in WCC update path (revised)

["J. Bruce Fields" <bfields@fieldses.org>]

On Thu, Aug 12, 2010 at 10:16:57PM -0700, Patrick J. LoPresti wrote:
> OK, I found the problem.  Whether it is a bug depends on your point of
> view, I suppose.
> 
> Although I am using XFS on my file server, and XFS has
> nanosecond-granularity timestamps, the true granularity of ctime/mtime
> is ultimately determined by the resolution of current_fs_time() which
> calls current_kernel_time(); i.e. jiffies; i.e. 1/HZ.
> 
> On my system (SLES 11 SP1), HZ is 250.  In my failing application, 4
> ms is long enough for many filesystem operations, even over NFS. (My
> network is 10GigE with a 300 microsecond round trip time, and my
> systems are very new.)
> 
> Anyway, I instrumented the VFS code on the NFS server to catch it in
> the act; specifically, I saw the following sequence:
> 
> file A is created on server, updating directory mtime
> NFS client does LOOKUP on file B, gets nfserr_noent
> file B created on server, does not update directory's mtime
> 
> ...all within 4 milliseconds (which is why the creation of file B did
> not update the directory's mtime).
> 
> The result is that the lookup cache on the client is stale and stays
> stale until some other client (or the server) updates the directory.
> Even making changes from the client does not invalidate the cache,
> thanks to the clever WCC logic that Trond had to explain to me
> earlier.
> 
> This is not exactly an NFS specific question, but I will ask anyway...
>  If I were to propose modifying current_fs_time() to call
> getnstimeofday() instead of current_kernel_time(), would the VFS folks
> laugh me out the door?

Good question.... Cc'ing linux-fsdevel.

If you have an easy reproducer it might also be worth experimenting with
NFSv4 exports of an ext4 system mounted with the i_version option.

Actually: should the NFSv4 server always be using i_version as the
change attribute for directories?  (Does every exportable filesystem
update it on every directory modification?)

(And if so, maybe on directories we should factor the i_version into the
low bits of the mtime reported to NFSv3 client?)

--b.

> 
> 1/HZ granularity for file timestamps just seems so...  90s or
> something.  4ms really is a lot of time these days.
> 
>  - Pat
> --
> To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html