Re: 2.5.60-mm2

[Andrew Morton]

Dave Jones <[e-mail]> wrote:
>
> On Fri, Feb 14, 2003 at 01:31:44AM -0800, Andrew Morton wrote:
> 
>  > . Considerable poking at the NFS MAP_SHARED OOM lockup.  It is limping
>  >   along now, but writeout bandwidth is poor and it is still struggling. 
>  >   Needs work.
>  > 
>  > . There's a one-liner which removes an O(n^2) search in the NFS writeback
>  >   path.  It increases writeout bandwidth by 4x and decreases CPU load from
>  >   100% to 3%.  Needs work.
> 
> I'm puzzled that you've had NFS stable enough to test these.

This was just writing out a single 400 megabyte file with `dd'.  I didn't try
anything fancier.

> How much testing has this stuff had? Here 2.5.60+bk clients fall over under
> moderate NFS load. (And go splat quickly under high load).
> 
> Trying to run things like dbench causes lockups, fsx/fstress made it
> reboot, plus the odd 'cheating' errors reported yesterday.

I have not tried pushing NFS with complex access patterns recently.

BTW, there's a little patch in there from Trond which I forgot to mention: it
implements sendfile for NFS, so loop-on-NFS works again.

But we have a refcounting bug somewhere:

# mount server:/dir /mnt/point
# losetup /dev/loop0 /mnt/point/file
# mount /dev/loop0 /mnt/loop0
# umount /mnt/loop0
# losetup -d /dev/loop0 
# umount /mnt/point
umount: /mnt/point: device is busy