reiserfs-devel.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
* [patch 0/7] per-bdi flushing model improvements. reiser4
@ 2010-02-02  1:39 Edward Shishkin
  0 siblings, 0 replies; 9+ messages in thread
From: Edward Shishkin @ 2010-02-02  1:39 UTC (permalink / raw)
  To: Andrew Morton, ReiserFS Development List
  Cc: linux-fsdevel, linux-kernel, xfs, jens.axboe

Hello.

Andrew Morton wrote:
> reiser4 is currently disabled in -mm (via reiser4-disable.patch)
> because recent changes to fs/fs/writeback.c wrecked the build.  I fixed
> it about ten times as the underlying code was churning, then gave up.  It
> would be nice if you take a look at that sometime please.
>
>

I have taken a look at fs/fs-writeback.c and found that per-superblock
flushing interface is eliminated. However migrating to per-bdi flushing
model doesn't necessarily means that such interface doesn't exist or is
not needed anymore. Flushing in accordance with the scheme "data-inode-
data-inode-..." would be very suboptimal for reiser4. Also xfs people
were unhappy with such flushing model:
http://article.gmane.org/gmane.linux.file-systems/30153

Moreover, the current stuff looks rather ugly. Why do we pin/unpin
superblock for every inode? It would be more reasonable to pin it for the
whole group of inodes and call a flushing handler for them. The patch 4
introduces such handler writeback_sb_inodes (which resembles dropped
sync_sb_inodes, the difference is that the newer version doesn't flush
necessarily all inodes of the superblock). Please, consider pushing this
patch to mainline.

The patch 5 adds super operation .writeback_inodes (former .sync_inodes)
which allows a file system to make optimizations. It can happen that
reiser4 will flush a bit more inodes then generic implementation suggests.
"a bit more" doesn't mean "all dirty inodes of the superblock" (see a
comment about atoms in the header of patch 6).

Finally, some file systems have its own means for periodical writeout
of dirty data. Since b_io contains inodes of many superblocks we need
to evict our inodes back to dirty list when flushing is going on with
for_kupdate flag installed. The new library function
writeback_skip_sb_inodes() provides such possibility.

Patch 7 fixes a race in checkin-checkout jnodes for entd task (reiser4).
Please, apply.

Thanks,
Edward.

^ permalink raw reply	[flat|nested] 9+ messages in thread
* [patch 0/7] per-bdi flushing model improvements. reiser4
@ 2010-02-02  1:55 Edward Shishkin
  2010-02-02  8:17 ` Christoph Hellwig
  0 siblings, 1 reply; 9+ messages in thread
From: Edward Shishkin @ 2010-02-02  1:55 UTC (permalink / raw)
  To: Andrew Morton, ReiserFS Development List
  Cc: linux-fsdevel, linux-kernel, xfs, jens.axboe

Hello.

Andrew Morton wrote:
> reiser4 is currently disabled in -mm (via reiser4-disable.patch)
> because recent changes to fs/fs/writeback.c wrecked the build.  I fixed
> it about ten times as the underlying code was churning, then gave up.  It
> would be nice if you take a look at that sometime please.
>
>

I have taken a look at fs/fs-writeback.c and found that per-superblock
flushing interface is eliminated. However migrating to per-bdi flushing
model doesn't necessarily means that such interface doesn't exist or is
not needed anymore. Flushing in accordance with the scheme "data-inode-
data-inode-..." would be very suboptimal for reiser4. Also xfs people
were unhappy with such flushing model:
http://article.gmane.org/gmane.linux.file-systems/30153

Moreover, current stuff doesn't look fine. Why do we pin/unpin
superblock for every inode? It would be more reasonable to pin it for the
whole group of inodes and call a flushing handler for them. The patch 4
introduces such handler writeback_sb_inodes (which resembles dropped
sync_sb_inodes, the difference is that the newer version doesn't flush
necessarily all inodes of the superblock). Please, consider pushing this
patch to mainline.

The patch 5 adds a super operation .writeback_inodes (former .sync_inodes)
which allows a file system to make optimizations. It can happen that
reiser4 will flush a bit more inodes then generic implementation suggests.
"a bit more" doesn't mean "all dirty inodes of the superblock" (see a
comment about atoms in the header of patch 6).

Finally, some file systems have its own means for periodical writeout
of dirty data. Since b_io contains inodes of many superblocks we need
to evict our inodes back to dirty list when flushing is going on with
for_kupdate flag installed. The new library function
writeback_skip_sb_inodes() provides such possibility.

Please, apply.

Thanks,
Edward.

^ permalink raw reply	[flat|nested] 9+ messages in thread
end of thread, other threads:[~2010-02-24  7:12 UTC | newest]

Thread overview: 9+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2010-02-02  1:39 [patch 0/7] per-bdi flushing model improvements. reiser4 Edward Shishkin
  -- strict thread matches above, loose matches on Subject: below --
2010-02-02  1:55 Edward Shishkin
2010-02-02  8:17 ` Christoph Hellwig
2010-02-02 15:25   ` Edward Shishkin
2010-02-02 19:42     ` Jens Axboe
2010-02-02 21:45       ` Ronni Holm-Nielsen
2010-02-02 22:26         ` Edward Shishkin
     [not found]           ` <20100222084753.0a5d613c.buchner.johannes@gmx.at>
2010-02-24  7:12             ` Christian Kujau
2010-02-02 21:47       ` Ronni Holm-Nielsen

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).