Re: [PATCH 00/46] rcu-walk and dcache scaling

From: Paul E. McKenney
Date: Sun Dec 12 2010 - 09:54:26 EST

Next message: Arnaldo Carvalho de Melo: "[PATCH 1/2] perf session: Fallback to unordered processing if no sample_id_all"
Previous message: Russell King - ARM Linux: "Re: [PATCH] dw_mmc: Add Synopsys DesignWare mmc host driver."
In reply to: Paul E. McKenney: "Re: [PATCH 00/46] rcu-walk and dcache scaling"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

On Fri, Dec 10, 2010 at 12:32:57PM -0800, Paul E. McKenney wrote:
> On Wed, Dec 08, 2010 at 06:09:09PM +1100, Nick Piggin wrote:
> > On Wed, Dec 08, 2010 at 03:28:16PM +1100, Dave Chinner wrote:
> > > On Wed, Dec 08, 2010 at 02:32:12PM +1100, Dave Chinner wrote:
> > > > On Wed, Dec 08, 2010 at 12:47:42PM +1100, Nick Piggin wrote:
> > > > > On Wed, Dec 8, 2010 at 8:56 AM, Dave Chinner <david@xxxxxxxxxxxxx> wrote:
> > > > > > On Sat, Nov 27, 2010 at 09:15:58PM +1100, Nick Piggin wrote:
> > > > > >>
> > > > > >> git://git.kernel.org/pub/scm/linux/kernel/git/npiggin/linux-npiggin.git vfs-scale-working
> > > > > >>
> > > > > >> Here is an new set of vfs patches for review, not that there was much interest
> > > > > >> last time they were posted. It is structured like:
> > > > > >>
> > > > > >> * preparation patches
> > > > > >> * introduce new locks to take over dcache_lock, then remove it
> > > > > >> * cleaning up and reworking things for new locks
> > > > > >> * rcu-walk path walking
> > > > > >> * start on some fine grained locking steps
> > > > > >
> > > > > > Stress test doing:
> > > > > >
> > > > > > single thread 50M inode create
> > > > > > single thread rm -rf
> > > > > > 2-way 50M inode create
> > > > > > 2-way rm -rf
> > > > > > 4-way 50M inode create
> > > > > > 4-way rm -rf
> > > > > > 8-way 50M inode create
> > > > > > 8-way rm -rf
> > > > > > 8-way 250M inode create
> > > > > > 8-way rm -rf
> > > > > >
> > > > > > Failed about 5 minutes into the "4-way rm -rf" (~3 hours into the test)
> > > > > > with a CPU stuck spinning on here:
> > > > > >
> > > > > > [37372.084012] NMI backtrace for cpu 5
> > > > > > [37372.084012] CPU 5
> > > > > > [37372.084012] Modules linked in:
> > > > > > [37372.084012]
> > > > > > [37372.084012] Pid: 15214, comm: rm Not tainted 2.6.37-rc4-dgc+ #797 /Bochs
> > > > > > [37372.084012] RIP: 0010:[<ffffffff810643c4>] [<ffffffff810643c4>] __ticket_spin_lock+0x14/0x20
> > > > > > [37372.084012] RSP: 0018:ffff880114643c98 EFLAGS: 00000213
> > > > > > [37372.084012] RAX: 0000000000008801 RBX: ffff8800687be6c0 RCX: ffff8800c4eb2688
> > > > > > [37372.084012] RDX: ffff880114643d38 RSI: ffff8800dfd4ea80 RDI: ffff880114643d14
> > > > > > [37372.084012] RBP: ffff880114643c98 R08: 0000000000000003 R09: 0000000000000000
> > > > > > [37372.084012] R10: 0000000000000000 R11: dead000000200200 R12: ffff880114643d14
> > > > > > [37372.084012] R13: ffff880114643cb8 R14: ffff880114643d38 R15: ffff8800687be71c
> > > > > > [37372.084012] FS: 00007fd6d7c93700(0000) GS:ffff8800dfd40000(0000) knlGS:0000000000000000
> > > > > > [37372.084012] CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b
> > > > > > [37372.084012] CR2: 0000000000bbd108 CR3: 0000000107146000 CR4: 00000000000006e0
> > > > > > [37372.084012] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
> > > > > > [37372.084012] DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400
> > > > > > [37372.084012] Process rm (pid: 15214, threadinfo ffff880114642000, task ffff88011b16f890)
> > > > > > [37372.084012] Stack:
> > > > > > [37372.084012] ffff880114643ca8 ffffffff81ad044e ffff880114643cf8 ffffffff81167ae7
> > > > > > [37372.084012] 0000000000000000 ffff880114643d38 000000000000000e ffff88011901d800
> > > > > > [37372.084012] ffff8800cdb7cf5c ffff88011901d8e0 0000000000000000 0000000000000000
> > > > > > [37372.084012] Call Trace:
> > > > > > [37372.084012] [<ffffffff81ad044e>] _raw_spin_lock+0xe/0x20
> > > > > > [37372.084012] [<ffffffff81167ae7>] shrink_dentry_list+0x47/0x370
> > > > > > [37372.084012] [<ffffffff81167f5e>] __shrink_dcache_sb+0x14e/0x1e0
> > > > > > [37372.084012] [<ffffffff81168456>] shrink_dcache_parent+0x276/0x2d0
> > > > > > [37372.084012] [<ffffffff81ad044e>] ? _raw_spin_lock+0xe/0x20
> > > > > > [37372.084012] [<ffffffff8115daa2>] dentry_unhash+0x42/0x80
> > > > > > [37372.084012] [<ffffffff8115db48>] vfs_rmdir+0x68/0x100
> > > > > > [37372.084012] [<ffffffff8115fd93>] do_rmdir+0x113/0x130
> > > > > > [37372.084012] [<ffffffff8114f5ad>] ? filp_close+0x5d/0x90
> > > > > > [37372.084012] [<ffffffff8115fde5>] sys_unlinkat+0x35/0x40
> > > > > > [37372.084012] [<ffffffff8103a002>] system_call_fastpath+0x16/0x1b
> > > > >
> > > > > OK good, with any luck, that's the same bug.
> > > > >
> > > > > Is this XFS?
> > > >
> > > > Yes.
> > > >
> > > > > Is there any concurrent activity happening on the same dentries?
> > > >
> > > > Not from an application perspective.
> > > >
> > > > > Ie. are the rm -rf threads running on the same directories,
> > > >
> > > > No, each thread operating on a different directory.
> >
> > This is probably fixed by the same patch as the lockdep splat trace.
> >
> >
> > > > > or is there any reclaim happening in the background?
> > > >
> > > > IIRC, kswapd was consuming about 5-10% of a CPU during parallel
> > > > unlink tests. Mainly reclaiming XFS inodes, I think, but there may
> > > > be dentry cache reclaim going as well.
> > >
> > > Turns out that the kswapd peaks are upwards of 50% of a CPU for a
> > > few seconds, then idle for 10-15s. Typical perf top output of kswapd
> > > while it is active during unlinks is:
> > >
> > > samples pcnt function DSO
> > > _______ _____ ___________________________ _________________
> > >
> > > 17168.00 10.2% __call_rcu [kernel.kallsyms]
> > > 13223.00 7.8% kmem_cache_free [kernel.kallsyms]
> > > 12917.00 7.6% down_write [kernel.kallsyms]
> > > 12665.00 7.5% xfs_iunlock [kernel.kallsyms]
> > > 10493.00 6.2% xfs_reclaim_inode_grab [kernel.kallsyms]
> > > 9314.00 5.5% __lookup_tag [kernel.kallsyms]
> > > 9040.00 5.4% radix_tree_delete [kernel.kallsyms]
> > > 8694.00 5.1% is_bad_inode [kernel.kallsyms]
> > > 7639.00 4.5% __ticket_spin_lock [kernel.kallsyms]
> > > 6821.00 4.0% _raw_spin_unlock_irqrestore [kernel.kallsyms]
> > > 5484.00 3.2% __d_drop [kernel.kallsyms]
> > > 5114.00 3.0% xfs_reclaim_inode [kernel.kallsyms]
> > > 4626.00 2.7% __rcu_process_callbacks [kernel.kallsyms]
> > > 3556.00 2.1% up_write [kernel.kallsyms]
> > > 3206.00 1.9% _cond_resched [kernel.kallsyms]
> > > 3129.00 1.9% xfs_qm_dqdetach [kernel.kallsyms]
> > > 2327.00 1.4% radix_tree_tag_clear [kernel.kallsyms]
> > > 2327.00 1.4% call_rcu_sched [kernel.kallsyms]
> > > 2262.00 1.3% __ticket_spin_unlock [kernel.kallsyms]
> > > 2215.00 1.3% xfs_ilock [kernel.kallsyms]
> > > 2200.00 1.3% radix_tree_gang_lookup_tag [kernel.kallsyms]
> > > 1982.00 1.2% xfs_reclaim_inodes_ag [kernel.kallsyms]
> > > 1736.00 1.0% xfs_trans_unlocked_item [kernel.kallsyms]
> > > 1707.00 1.0% __ticket_spin_trylock [kernel.kallsyms]
> > > 1688.00 1.0% xfs_perag_get_tag [kernel.kallsyms]
> > > 1660.00 1.0% flat_send_IPI_mask [kernel.kallsyms]
> > > 1538.00 0.9% xfs_inode_item_destroy [kernel.kallsyms]
> > > 1312.00 0.8% __shrink_dcache_sb [kernel.kallsyms]
> > > 940.00 0.6% xfs_perag_put [kernel.kallsyms]
> > >
> > > So there is some dentry cache reclaim going on.
> > >
> > > FWIW, it appears there is quite a lot of RCU freeing overhead (~15%
> > > more CPU time) in the work kswapd is doing during these unlinks, too.
> > > I just had a look at kswapd when a 8-way create is running - it's running at
> > > 50-60% of a cpu for seconds at a time. I caught this while it was doing pure
> > > XFS inode cache reclaim (~10s sample, kswapd reclaimed ~1M inodes):
> > >
> > > samples pcnt function DSO
> > > _______ _____ ___________________________ _________________
> > >
> > > 27171.00 9.0% __call_rcu [kernel.kallsyms]
> > > 21491.00 7.1% down_write [kernel.kallsyms]
> > > 20916.00 6.9% xfs_reclaim_inode [kernel.kallsyms]
> > > 20313.00 6.7% radix_tree_delete [kernel.kallsyms]
> > > 15828.00 5.3% kmem_cache_free [kernel.kallsyms]
> > > 15819.00 5.2% xfs_idestroy_fork [kernel.kallsyms]
> > > 14893.00 4.9% is_bad_inode [kernel.kallsyms]
> > > 14666.00 4.9% _raw_spin_unlock_irqrestore [kernel.kallsyms]
> > > 14191.00 4.7% xfs_reclaim_inode_grab [kernel.kallsyms]
> > > 14105.00 4.7% xfs_iunlock [kernel.kallsyms]
> > > 10916.00 3.6% __ticket_spin_lock [kernel.kallsyms]
> > > 10125.00 3.4% xfs_iflush_cluster [kernel.kallsyms]
> > > 8221.00 2.7% xfs_qm_dqdetach [kernel.kallsyms]
> > > 7639.00 2.5% xfs_trans_unlocked_item [kernel.kallsyms]
> > > 7028.00 2.3% xfs_synchronize_times [kernel.kallsyms]
> > > 6974.00 2.3% up_write [kernel.kallsyms]
> > > 5870.00 1.9% call_rcu_sched [kernel.kallsyms]
> > > 5634.00 1.9% _cond_resched [kernel.kallsyms]
> > >
> > > Which is showing a similar amount of RCU overhead as the unlink as above.
> > > And this while it was doing dentry cache reclaim (~10s sample):
> > >
> > > 35921.00 15.7% __d_drop [kernel.kallsyms]
> > > 30056.00 13.1% __ticket_spin_trylock [kernel.kallsyms]
> > > 29066.00 12.7% __ticket_spin_lock [kernel.kallsyms]
> > > 19043.00 8.3% __call_rcu [kernel.kallsyms]
> > > 10098.00 4.4% iput [kernel.kallsyms]
> > > 7013.00 3.1% __shrink_dcache_sb [kernel.kallsyms]
> > > 6774.00 3.0% __percpu_counter_add [kernel.kallsyms]
> > > 6708.00 2.9% radix_tree_tag_set [kernel.kallsyms]
> > > 5362.00 2.3% xfs_inactive [kernel.kallsyms]
> > > 5130.00 2.2% __ticket_spin_unlock [kernel.kallsyms]
> > > 4884.00 2.1% call_rcu_sched [kernel.kallsyms]
> > > 4621.00 2.0% dentry_lru_del [kernel.kallsyms]
> > > 3735.00 1.6% bit_waitqueue [kernel.kallsyms]
> > > 3727.00 1.6% dentry_iput [kernel.kallsyms]
> > > 3473.00 1.5% shrink_icache_memory [kernel.kallsyms]
> > > 3279.00 1.4% kfree [kernel.kallsyms]
> > > 3101.00 1.4% xfs_perag_get [kernel.kallsyms]
> > > 2516.00 1.1% kmem_cache_free [kernel.kallsyms]
> > > 2272.00 1.0% shrink_dentry_list [kernel.kallsyms]
> > >
> > > I've never really seen any signficant dentry cache reclaim overhead
> > > in profiles of these workloads before, so this was a bit of a
> > > surprise....
> >
> > call_rcu shouldn't be doing much, except for disabling irqs and linking
> > the object into the list. I have a patch somewhere to reduce the irq
> > disable overhead a bit, but it really shouldn't be doing a lot of work.
>
> Could you please enable CONFIG_RCU_TRACE, mount debugfs somewhere, and
> look at rcu/rcudata? There will be a "ql=" number printed for each
> CPU, and if that number is too large, __call_rcu() does take what it
> considers to be corrective action, which can incur some overhead.
>
> If this is the problem, then increasing the value of the qhimark module
> parameter might help.
>
> If this is not the problem, I could make a patch that disables some of
> __call_rcu()'s grace-period acceleration code if you are willing to try
> it out.

Another thing that might help is to reduce the value of CONFIG_RCU_FANOUT
to something like 16. If this does help, then there is a reasonably
straightforward change I can make to RCU.

Thanx, Paul

> > Sometimes you find that touching the rcu head field needs to get a
> > cacheline exclusive, so a bit of work gets transferred there....
> >
> > But it may also be something going a bit wrong in RCU. I blew it up
> > once already, after the files_lock splitup that enabled all CPUs to
> > create and destroy files :)
>
> I would certainly like the opportunity to fix any bugs that might be
> in RCU... ;-)
>
> Thanx, Paul
--
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/

Next message: Arnaldo Carvalho de Melo: "[PATCH 1/2] perf session: Fallback to unordered processing if no sample_id_all"
Previous message: Russell King - ARM Linux: "Re: [PATCH] dw_mmc: Add Synopsys DesignWare mmc host driver."
In reply to: Paul E. McKenney: "Re: [PATCH 00/46] rcu-walk and dcache scaling"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]