Re: [2.6.35-rc1, bug] mm: minute-long livelocks in memory reclaim

[David Rientjes]

On Mon, 23 Aug 2010, Wu Fengguang wrote:

> > I've been testing parallel create workloads over the weekend, and
> > I've seen this a couple of times now under 8 thread parallel creates
> > with XFS. I'm running on an 8p VM with 4GB RAM and a fast disk
> > subsystem. Basically I am seeing the create rate drop to zero
> > with all 8 CPUs stuck spinning for up to 2 minutes. 'echo t >
> > /proc/sysrq-trigger' while this is occurring gives the following
> > trace for all the fs-mark processes:
> > 
> > [49506.624018] fs_mark       R  running task        0  8376   7917 0x00000008
> > [49506.624018]  0000000000000000 ffffffff81b94590 00000000000008fc 0000000000000002
> > [49506.624018]  0000000000000000 0000000000000286 0000000000000297 ffffffffffffff10
> > [49506.624018]  ffffffff810b3d02 0000000000000010 0000000000000202 ffff88011df777a8
> > [49506.624018] Call Trace:
> > [49506.624018]  [<ffffffff810b3d02>] ? smp_call_function_many+0x1a2/0x210
> > [49506.624018]  [<ffffffff810b3ce5>] ? smp_call_function_many+0x185/0x210
> > [49506.624018]  [<ffffffff81109170>] ? drain_local_pages+0x0/0x20
> > [49506.624018]  [<ffffffff810b3d92>] ? smp_call_function+0x22/0x30
> > [49506.624018]  [<ffffffff810849a4>] ? on_each_cpu+0x24/0x50
> > [49506.624018]  [<ffffffff81107bec>] ? drain_all_pages+0x1c/0x20
> > [49506.624018]  [<ffffffff8110825a>] ? __alloc_pages_nodemask+0x57a/0x730
> > [49506.624018]  [<ffffffff8113c6d2>] ? kmem_getpages+0x62/0x160
> > [49506.624018]  [<ffffffff8113d2b2>] ? fallback_alloc+0x192/0x240
> > [49506.624018]  [<ffffffff8113cce1>] ? cache_grow+0x2d1/0x300
> > [49506.624018]  [<ffffffff8113d04a>] ? ____cache_alloc_node+0x9a/0x170
> > [49506.624018]  [<ffffffff8113cf6c>] ? cache_alloc_refill+0x25c/0x2a0
> > [49506.624018]  [<ffffffff8113ddb3>] ? __kmalloc+0x193/0x230
> > [49506.624018]  [<ffffffff812f59af>] ? kmem_alloc+0x8f/0xe0
> > [49506.624018]  [<ffffffff812f59af>] ? kmem_alloc+0x8f/0xe0
> > [49506.624018]  [<ffffffff812f5a9e>] ? kmem_zalloc+0x1e/0x50
> > [49506.624018]  [<ffffffff812e2f4d>] ? xfs_log_commit_cil+0x9d/0x440
> > [49506.624018]  [<ffffffff812eeec6>] ? _xfs_trans_commit+0x1e6/0x2b0
> > [49506.624018]  [<ffffffff812f2b6f>] ? xfs_create+0x51f/0x690
> > [49506.624018]  [<ffffffff812ffdb7>] ? xfs_vn_mknod+0xa7/0x1c0
> > [49506.624018]  [<ffffffff812fff00>] ? xfs_vn_create+0x10/0x20
> > [49506.624018]  [<ffffffff811510b8>] ? vfs_create+0xb8/0xf0
> > [49506.624018]  [<ffffffff81151d2c>] ? do_last+0x4dc/0x5d0
> > [49506.624018]  [<ffffffff81153bd7>] ? do_filp_open+0x207/0x5e0
> > [49506.624018]  [<ffffffff8105fc58>] ? pvclock_clocksource_read+0x58/0xd0
> > [49506.624018]  [<ffffffff8115eaca>] ? alloc_fd+0x10a/0x150
> > [49506.624018]  [<ffffffff81144005>] ? do_sys_open+0x65/0x130
> > [49506.624018]  [<ffffffff81144110>] ? sys_open+0x20/0x30
> > [49506.624018]  [<ffffffff81036072>] ? system_call_fastpath+0x16/0x1b
> > 
> > Eventually the problem goes away, and the system goes back to
> > performing at the normal rate. Any ideas on how to avoid this
> > problem? I'm using CONFIG_SLAB=y is that is relevant....
> 
> zone->lock contention? Try rip the following two lines. The change
> might be a bit aggressive though :)
> 
> Thanks,
> Fengguang
> 
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index 1bb327a..c08b8d3 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -1864,9 +1864,6 @@ __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
>  
>  	cond_resched();
>  
> -	if (order != 0)
> -		drain_all_pages();
> -
>  	if (likely(*did_some_progress))
>  		page = get_page_from_freelist(gfp_mask, nodemask, order,
>  					zonelist, high_zoneidx,

You may be interested in Mel's patchset that he just proposed for -mm 
which identifies watermark variations on machines with high cpu counts 
(perhaps even eight, as in this report).  The last patch actually reworks 
this hunk of the code as well.

	http://marc.info/?l=linux-mm&m=128255044912938
	http://marc.info/?l=linux-mm&m=128255045312950
	http://marc.info/?l=linux-mm&m=128255045012942
	http://marc.info/?l=linux-mm&m=128255045612954

Dave, it would be interesting to see if this fixes your problem.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/