lkml.org 
[lkml]   [2007]   [Apr]   [24]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH 10/10] mm: per device dirty threshold
From
Date
On Tue, 2007-04-24 at 10:19 +0200, Miklos Szeredi wrote:
> > > This is probably a
> > > reasonable thing to do but it doesn't feel like the right place. I
> > > think get_dirty_limits should return the raw threshold, and
> > > balance_dirty_pages should do both tests - the bdi-local test and the
> > > system-wide test.
> >
> > Ok, that makes sense I guess.
>
> Well, my narrow minded world view says it's not such a good idea,
> because it would again introduce the deadlock scenario, we're trying
> to avoid.

I was only referring to the placement of the clipping; and exactly where
that happens does not affect the deadlock.

> In a sense allowing a queue to go over the global limit just a little
> bit is a good thing. Actually the very original code does that: if
> writeback was started for "write_chunk" number of pages, then we allow
> "ratelimit" (8) _new_ pages to be dirtied, effectively ignoring the
> global limit.

It might be time to get rid of that rate-limiting.
balance_dirty_pages()'s fast path is not nearly as heavy as it used to
be. All these fancy counter systems have removed quite a bit of
iteration from there.

> That's why I've been saying, that the current code is so unfair: if
> there are lots of dirty pages to be written back to a particular
> device, then balance_dirty_pages() allows the dirty producer to make
> even more pages dirty, but if there are _no_ dirty pages for a device,
> and we are over the limit, then that dirty producer is allowed
> absolutely no new dirty pages until the global counts subside.

Well, that got fixed on a per device basis with this patch, it is still
true for multiple tasks writing to the same device.

> I'm still not quite sure what purpose the above "soft" limiting
> serves. It seems to just give advantage to writers, which managed to
> accumulate lots of dirty pages, and then can convert that into even
> more dirtyings.

The queues only limit the actual in-flight writeback pages,
balance_dirty_pages() considers all pages that might become writeback as
well as those that are.

> Would it make sense to remove this behavior, and ensure that
> balance_dirty_pages() doesn't return until the per-queue limits have
> been complied with?

I don't think that will help, balance_dirty_pages drives the queues.
That is, it converts pages from mere dirty to writeback.

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2007-04-24 10:35    [from the cache]
©2003-2014 Jasper Spaans. Advertise on this site