Re: [PATCH 0/7] breaking the global file_list_lock

[Christoph Hellwig <hch@infradead.org>]

On Sun, Jan 28, 2007 at 02:43:25PM +0000, Christoph Hellwig wrote:
> On Sun, Jan 28, 2007 at 12:51:18PM +0100, Peter Zijlstra wrote:
> > This patch-set breaks up the global file_list_lock which was found to be a
> > severe contention point under basically any filesystem intensive workload.
> 
> Benchmarks, please.  Where exactly do you see contention for this?
> 
> 
> filesystem intensive workload apparently means namespace operation heavy
> workload, right?  The biggest bottleneck I've seen with those is dcache lock.
> 
> Even if this is becoming a real problem there must be simpler ways to fix
> this than introducing various new locking primitives and all kinds of
> complexity.

One good way to fix scalability without all this braindamage is
to get rid of sb->s_files.  Current uses are:

 - fs/dquot.c:add_dquot_ref()

	This performs it's actual operation on inodes.  We should
	be able to check inode->i_writecount to see which inodes
	need quota initialization.

 - fs/file_table.c:fs_may_remount_ro()

	This one is gone in Dave Hansens per-mountpoint r/o patchkit

 - fs/proc/generic.c:proc_kill_inodes()

	This can be done with a list inside procfs.

 - fs/super.c:mark_files_ro()

	This one is only used for do_emergency_remount(), which is
	and utter hack.  It might be much better to just deny any
	kind of write access through a superblock flag here.

 - fs/selinuxfs.c:sel_remove_bools()

	Utter madness.  I have no idea how this ever got merged.
	Maybe the selinux folks can explain what crack they were
	on when writing this.  The problem would go away with
	a generic rewoke infrastructure.

Once sb->s_files is gone we can also kill of fu_list entirely and
replace it by a list head entirely in the tty code and make the lock
for it per-tty.