Re: [PATCH] block/mq: Cure cpu hotplug lock inversion

[Jens Axboe]

On Thu, May 04 2017, Peter Zijlstra wrote:
> On Thu, May 04, 2017 at 07:56:57AM -0600, Jens Axboe wrote:
> > On 05/04/2017 07:05 AM, Peter Zijlstra wrote:
> > > 
> > > By poking at /debug/sched_features I triggered the following splat:
> > > 
> > >  [] ======================================================
> > >  [] WARNING: possible circular locking dependency detected
> > >  [] 4.11.0-00873-g964c8b7-dirty #694 Not tainted
> > >  [] ------------------------------------------------------
> > >  [] bash/2109 is trying to acquire lock:
> > >  []  (cpu_hotplug_lock.rw_sem){++++++}, at: [<ffffffff8120cb8b>] static_key_slow_dec+0x1b/0x50
> > >  [] 
> > >  [] but task is already holding lock:
> > >  []  (&sb->s_type->i_mutex_key#4){+++++.}, at: [<ffffffff81140216>] sched_feat_write+0x86/0x170
> > >  [] 
> > >  [] which lock already depends on the new lock.
> > >  [] 
> > >  [] 
> > >  [] the existing dependency chain (in reverse order) is:
> > >  [] 
> > >  [] -> #2 (&sb->s_type->i_mutex_key#4){+++++.}:
> > >  []        lock_acquire+0x100/0x210
> > >  []        down_write+0x28/0x60
> > >  []        start_creating+0x5e/0xf0
> > >  []        debugfs_create_dir+0x13/0x110
> > >  []        blk_mq_debugfs_register+0x21/0x70
> > >  []        blk_mq_register_dev+0x64/0xd0
> > >  []        blk_register_queue+0x6a/0x170
> > >  []        device_add_disk+0x22d/0x440
> > >  []        loop_add+0x1f3/0x280
> > >  []        loop_init+0x104/0x142
> > >  []        do_one_initcall+0x43/0x180
> > >  []        kernel_init_freeable+0x1de/0x266
> > >  []        kernel_init+0xe/0x100
> > >  []        ret_from_fork+0x31/0x40
> > >  [] 
> > >  [] -> #1 (all_q_mutex){+.+.+.}:
> > >  []        lock_acquire+0x100/0x210
> > >  []        __mutex_lock+0x6c/0x960
> > >  []        mutex_lock_nested+0x1b/0x20
> > >  []        blk_mq_init_allocated_queue+0x37c/0x4e0
> > >  []        blk_mq_init_queue+0x3a/0x60
> > >  []        loop_add+0xe5/0x280
> > >  []        loop_init+0x104/0x142
> > >  []        do_one_initcall+0x43/0x180
> > >  []        kernel_init_freeable+0x1de/0x266
> > >  []        kernel_init+0xe/0x100
> > >  []        ret_from_fork+0x31/0x40
> > > 
> > >  []  *** DEADLOCK ***
> > >  [] 
> > >  [] 3 locks held by bash/2109:
> > >  []  #0:  (sb_writers#11){.+.+.+}, at: [<ffffffff81292bcd>] vfs_write+0x17d/0x1a0
> > >  []  #1:  (debugfs_srcu){......}, at: [<ffffffff8155a90d>] full_proxy_write+0x5d/0xd0
> > >  []  #2:  (&sb->s_type->i_mutex_key#4){+++++.}, at: [<ffffffff81140216>] sched_feat_write+0x86/0x170
> > >  [] 
> > >  [] stack backtrace:
> > >  [] CPU: 9 PID: 2109 Comm: bash Not tainted 4.11.0-00873-g964c8b7-dirty #694
> > >  [] Hardware name: Intel Corporation S2600GZ/S2600GZ, BIOS SE5C600.86B.02.02.0002.122320131210 12/23/2013
> > >  [] Call Trace:
> > > 
> > >  []  lock_acquire+0x100/0x210
> > >  []  get_online_cpus+0x2a/0x90
> > >  []  static_key_slow_dec+0x1b/0x50
> > >  []  static_key_disable+0x20/0x30
> > >  []  sched_feat_write+0x131/0x170
> > >  []  full_proxy_write+0x97/0xd0
> > >  []  __vfs_write+0x28/0x120
> > >  []  vfs_write+0xb5/0x1a0
> > >  []  SyS_write+0x49/0xa0
> > >  []  entry_SYSCALL_64_fastpath+0x23/0xc2
> > > 
> > > This is because of the cpu hotplug lock rework. Break the chain at #1
> > > by reversing the lock acquisition order. This way i_mutex_key#4 no
> > > longer depends on cpu_hotplug_lock and things are good.
> > 
> > Thanks Peter, applied.
> 
> Note that the hotplug rework is still work-in-progress and lives in a
> -tip branch.
> 
> That said, the patch is harmless outside of that, so yes it can travel
> upstream independently. But note that mainline cannot yet trigger that
> splat.

Yes, may as well just get it in and avoid any dependencies.

-- 
Jens Axboe