lkml.org 
[lkml]   [2007]   [Mar]   [25]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [patch 1/3] fix illogical behavior in balance_dirty_pages()
On Sat, 24 Mar 2007 22:55:29 +0100 Miklos Szeredi <miklos@szeredi.hu> wrote:

> This is a slightly different take on the fix for the deadlock in fuse
> with dirty balancing. David Chinner convinced me, that per-bdi
> counters are too expensive, and that it's not worth trying to account
> the number of pages under writeback, as they will be limited by the
> queue anyway.
>
> ----
> From: Miklos Szeredi <mszeredi@suse.cz>
>
> Current behavior of balance_dirty_pages() is to try to start writeout
> into the specified queue for at least "write_chunk" number of pages.
> If "write_chunk" pages have been submitted, then return.
>
> However if there are less than "write_chunk" dirty pages for this
> queue, then it doesn't return, waiting for the global dirty+writeback
> counters to subside, but without doing any actual work.
>
> This is illogical behavior: it allows more dirtyings while there are
> dirty pages, but stops further dirtying completely if there are no
> more dirty pages.

That behaviour is perfectly logical. It prevents the number of
dirty+writeback pages from exceeding dirty_ratio.

> It also makes a deadlock possible when one filesystem is writing data
> through another, and the balance_dirty_pages() for the lower
> filesystem is stalling the writeback for the upper filesystem's
> data (*).

I still don't understand this one. I got lost when belatedly told that
i_mutex had something to do with it.

> So the exit condition should instead be:
>
> submitted at least "write_chunk" number of pages
> OR
> submitted ALL the dirty pages destined for this backing dev
> AND
> backing dev is not congested
>
> To do this, introduce a new counter in writeback_control, which counts
> the number of dirty pages encountered during writeback. This includes
> all dirty pages, even those which are already under writeback but have
> been dirtied again, and those which have been skipped due to having
> locked buffers.
>
> If this counter is zero after trying to submit some pages for
> writeback, and the backing dev is uncongested, then don't wait any
> more. After this, newly dirtied pages can quickly be written back to
> this backing dev.
>
> If there are globally no more pages to submit for writeback
> (nr_reclaimable == 0), then also don't wait for ever, only while this
> backing dev is congested.
>
> (*) For more info on this deadlock, see the following discussions:
>
> http://lkml.org/lkml/2007/3/1/9
> http://lkml.org/lkml/2007/3/12/16
>
> Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
> ---
>
> Index: linux/include/linux/writeback.h
> ===================================================================
> --- linux.orig/include/linux/writeback.h 2007-03-24 22:06:56.000000000 +0100
> +++ linux/include/linux/writeback.h 2007-03-24 22:29:02.000000000 +0100
> @@ -44,6 +44,7 @@ struct writeback_control {
> long nr_to_write; /* Write this many pages, and decrement
> this for each page written */
> long pages_skipped; /* Pages which were not written */
> + long nr_dirty; /* Number of dirty pages encountered */
>
> /*
> * For a_ops->writepages(): is start or end are non-zero then this is
> Index: linux/mm/page-writeback.c
> ===================================================================
> --- linux.orig/mm/page-writeback.c 2007-03-24 22:06:56.000000000 +0100
> +++ linux/mm/page-writeback.c 2007-03-24 22:29:02.000000000 +0100
> @@ -207,7 +207,15 @@ static void balance_dirty_pages(struct a
> * written to the server's write cache, but has not yet
> * been flushed to permanent storage.
> */
> - if (nr_reclaimable) {
> + if (!nr_reclaimable) {
> + /*
> + * If there's nothing more to write back and this queue
> + * is uncongested, then it is possible to quickly
> + * write out some more data, so let's not wait
> + */
> + if (!bdi_write_congested(bdi))
> + break;
> + } else {

This says "if there are no dirty pages in the machine at all, then go back
and dirty some more pages, regardless of the present number of
under-writeback pages".

> writeback_inodes(&wbc);
> get_dirty_limits(&background_thresh,
> &dirty_thresh, mapping);
> @@ -220,6 +228,14 @@ static void balance_dirty_pages(struct a
> pages_written += write_chunk - wbc.nr_to_write;
> if (pages_written >= write_chunk)
> break; /* We've done our duty */
> +
> + /*
> + * If there are no more dirty pages for this backing
> + * backing dev, and the queue is not congested, then
> + * it is possible to quickly write out some more data
> + */
> + if (!wbc.nr_dirty && !bdi_write_congested(bdi))
> + break;

This says "if there are no pages dirty againt this device then go back and
dirty some more pages, regardless of the present number of under-writeback
pages".



IOW: this change will allow us to 100% fill all memory with under-writeback
pages.

The VM _should_ cope with that - it kinda used to. But it is an untested
region of operation and the chances of bogus oom-killings are excellent.

-
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: 2007-03-26 01:37    [W:0.084 / U:0.780 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site