Re: [RRC PATCH 2/2] vfs: Use per-cpu list for superblock's inode list

From: Dave Chinner <david@fromorbit.com>
To: Waiman Long <Waiman.Long@hpe.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>,
	Jan Kara <jack@suse.com>, Jeff Layton <jlayton@poochiereds.net>,
	"J. Bruce Fields" <bfields@fieldses.org>,
	Tejun Heo <tj@kernel.org>,
	Christoph Lameter <cl@linux-foundation.org>,
	linux-fsdevel@vger.kernel.org, linux-kernel@vger.kernel.org,
	Ingo Molnar <mingo@redhat.com>,
	Peter Zijlstra <peterz@infradead.org>,
	Andi Kleen <andi@firstfloor.org>,
	Dave Chinner <dchinner@redhat.com>,
	Scott J Norton <scott.norton@hp.com>,
	Douglas Hatch <doug.hatch@hp.com>
Subject: Re: [RRC PATCH 2/2] vfs: Use per-cpu list for superblock's inode list
Date: Wed, 17 Feb 2016 21:37:39 +1100	[thread overview]
Message-ID: <20160217103739.GP14668@dastard> (raw)
In-Reply-To: <1455672680-7153-3-git-send-email-Waiman.Long@hpe.com>

On Tue, Feb 16, 2016 at 08:31:20PM -0500, Waiman Long wrote:
> When many threads are trying to add or delete inode to or from
> a superblock's s_inodes list, spinlock contention on the list can
> become a performance bottleneck.
> 
> This patch changes the s_inodes field to become a per-cpu list with
> per-cpu spinlocks.
> 
> With an exit microbenchmark that creates a large number of threads,
> attachs many inodes to them and then exits. The runtimes of that
> microbenchmark with 1000 threads before and after the patch on a
> 4-socket Intel E7-4820 v3 system (40 cores, 80 threads) were as
> follows:
> 
>   Kernel            Elapsed Time    System Time
>   ------            ------------    -----------
>   Vanilla 4.5-rc4      65.29s         82m14s
>   Patched 4.5-rc4      22.81s         23m03s

Pretty good :)

My fsmark tests usually show up a fair bit of contention - moving
250k inodes through the cache every second over 16p does generate a
bit of load on the list. The patch makes the inode list add/del
operations disappear completely from the perf profiles, and there's
a marginal decrease in runtime (~4m40s vs 4m30s). I think the global
lock is right on the edge of breakdown under this load, though, so
if I was testing on a larger system I think the difference would be
much bigger.

I'll run some more testing on it, see if anything breaks.

A few comments on the code follow.

> @@ -1866,8 +1866,8 @@ void iterate_bdevs(void (*func)(struct block_device *, void *), void *arg)
>  {
>  	struct inode *inode, *old_inode = NULL;
>  
> -	spin_lock(&blockdev_superblock->s_inode_list_lock);
> -	list_for_each_entry(inode, &blockdev_superblock->s_inodes, i_sb_list) {
> +	for_all_percpu_list_entries_simple(inode, percpu_lock,
> +			blockdev_superblock->s_inodes_cpu, i_sb_list) {

This is kind what I meant about names getting way too long. How
about something like:

#define walk_sb_inodes(inode, sb, pcpu_lock)	\
	for_all_percpu_list_entries_simple(inode, pcpu_lock,	\
					   sb->s_inodes_list, i_sb_list)

#define walk_sb_inodes_end(pcpu_lock) end_all_percpu_list_entries(pcpu_lock)

for brevity?

> @@ -189,7 +190,7 @@ void fsnotify_unmount_inodes(struct super_block *sb)
>  		spin_unlock(&inode->i_lock);
>  
>  		/* In case the dropping of a reference would nuke next_i. */
> -		while (&next_i->i_sb_list != &sb->s_inodes) {
> +		while (&next_i->i_sb_list.list != percpu_head) {
>  			spin_lock(&next_i->i_lock);
>  			if (!(next_i->i_state & (I_FREEING | I_WILL_FREE)) &&
>  						atomic_read(&next_i->i_count)) {
> @@ -199,16 +200,16 @@ void fsnotify_unmount_inodes(struct super_block *sb)
>  				break;
>  			}
>  			spin_unlock(&next_i->i_lock);
> -			next_i = list_next_entry(next_i, i_sb_list);
> +			next_i = list_next_entry(next_i, i_sb_list.list);

pcpu_list_next_entry(next_i, i_sb_list)?

> @@ -1397,9 +1398,8 @@ struct super_block {
>  	 */
>  	int s_stack_depth;
>  
> -	/* s_inode_list_lock protects s_inodes */
> -	spinlock_t		s_inode_list_lock ____cacheline_aligned_in_smp;
> -	struct list_head	s_inodes;	/* all inodes */
> +	/* The percpu locks protect s_inodes_cpu */
> +	PERCPU_LIST_HEAD(s_inodes_cpu);	/* all inodes */

There is no need to encode the type of list into the name.
i.e. drop the "_cpu" suffix - we can see it's a percpu list from the
declaration.

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com