Re: New data=ordered code pushed out to btrfs-unstable

[Chris Mason <chris.mason@oracle.com>]

On Mon, 2008-07-21 at 15:23 -0400, Ric Wheeler wrote:

> >>> [ lock timeouts and stalls ]
> >>>
> >>>
> >>> Ok, I've made a few changes that should lower overall contenion on the
> >>> allocation mutex.  I'm getting better performance on a 3 million file
> >>> run, please give it a shot.
> >>
> >> After an update, clean rebuild & reboot, the test is running along and 
> >> has hit about 10 million files. I still see some messages like:
> >>
> >> INFO: task pdflush:4051 blocked for more than 120 seconds.

The latest code in btrfs-unstable has everything I can safely do right
now :)  

Basically the stalls come from someone doing IO with the allocation
mutex held.  It is surprising that we should be stalling for such a long
time, it is probably a  mixture of elevator starvation and btrfs fun.

But, btrfs-unstable also has code to replace the page lock with a
per-tree block mutex, which will allow me to get rid of the big
allocation mutex over the long term.  I was able to break up most of the
long operations and have them drop/reacquire the allocation mutex to
prevent this starvation most of the time.

-chris