Re: [PATCH][5/?] count writeback pages in nr_scanned

[Andrea Arcangeli <andrea@suse.de>]

On Wed, Jan 05, 2005 at 09:37:04PM -0800, Andrew Morton wrote:
> Andrea Arcangeli <andrea@suse.de> wrote:
> >
> > 2) we won't need unreliable anti-deadlock timeouts anymore
> 
> The timeouts are for:
> 
> a) A fallback for backing stores which don't wake up waiters in
>    blk_congestion_wait() (eg: NFS).

that anti-deadlock will be unnecessary too with the new logic.

> b) handling the race case where the request queue suddenly goes empty
>    before the sleeper gets onto the waitqueue.

as I mentioned with proper locking setting task in uninterruptible and
then registering into the new per classzone waitqueue, the timeout will
be unnecessary even for this.

> It can probably be removed with some work, and additional locking.

The additional locking will then remove the current locking in
blk_congestion_wait so it's new locking but it will replace the current
locking. But I agree registering in the waitqueue inside the
blk_congestion_wait was simpler. It's just I've an hard time to like the
timeout.  Timeout is always wrong when it triggers: if it triggers it
always triggers either too late (wasted resources) or too early (early
oom kills). So unless it messes everything up, it'd be nice to the
locking strict. anyway point 1 and 2 can be implemented separately, at
first we can leave the timeout if the race is too hard to handle. 

Ideally if we keep the total number of oustanding writebacks
per-classzone (not sure if we account for it already somewhere, I guess
if something we've the global number and not the per-classzone one), we
could remove the timeout without having to expose the locking outside
blk_congestion_wait. With the number of oustanding writebacks
per-classzone we could truly fix the race and obsolete the timeout in a
self contained manner. Though it will require a proper amount of memory
barriers around the account increment/decrement/read.