Re: [PATCH] EXT2: Remove superblock lock contention in ext2_statfs

[Valerie Henson <val_henson@linux.intel.com>]

On Mon, Sep 18, 2006 at 06:38:05PM -0500, Dave Kleikamp wrote:
> On Mon, 2006-09-18 at 15:36 -0500, Dave Kleikamp wrote:
> > EXT2: Remove superblock lock contention in ext2_statfs
> > 
> > Fix a performance degradation introduced in 2.6.17.  (30% degradation running
> > dbench with 16 threads)
> > 
> > Patch 21730eed11de42f22afcbd43f450a1872a0b5ea1, which claims to make
> > EXT2_DEBUG work again, moves the taking of the kernel lock out of debug-only
> > code in ext2_count_free_inodes and ext2_count_free_blocks and into
> > ext2_statfs.  This patch reverses that part of the patch.
> > 
> > Signed-off-by: Dave Kleikamp <shaggy@austin.ibm.com>
> 
> Eric Sandeen pointed out to me that taking the superblock lock in
> ext2_count_free_* will cause a deadlock when EXT2FS_DEBUG is enabled,
> since the superblock is locked in write_super().
> 
> We found that the same problem was fixed in ext3 with this patch
> (forgive the long link):
> http://git.kernel.org/git/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h=5b11687924e40790deb0d5f959247ade82196665;hp=2384f55f8aa520172c995965bd2f8a9740d53095
> 
> The patch below just removes the use of the superblock lock in the debug
> code.

(Sorry for the delay; been on vacation.)

Heh, I ran into the same lock nesting issues as you when I first tried
to fix this; the lock debugging code found it for me.  I asked for
feedback on the locking issue when I submitted the patch, but no one
had any opinions then, so I chose consistency over possible
contention.  Al Viro snorted at the idea of consistency in the results
of statfs (I paraphrase his IRC remarks), and thinking about it
further, I realized the debug code should not be doing these checks in
statfs anyway; only on mount and unmount.  This is because it appears
that the block group accounting and the overall fs accounting are done
non-atomically - see group_reserve_blocks() for example - and part of
what the code does is reconcile these two numbers.  It is legal for a
valid fs to have the block group summaries and the fs-wide summaries
out of sync, so the debug code could erroneously report an error,
leading some poor soul on a wild goose chase.  Removing this code from
statfs also happens to fix the locking issues nicely.

Rewriting this has been on my todo list for about 6 months now -
anyone interested in grabbing it?  I'm on #linuxfs on irc.oftc.net if
anyone wants to chat about it.

-VAL