lkml.org 
[lkml]   [2019]   [Feb]   [2]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] Drivers: hv: vmbus: Add mutex lock to channel show functions
On Fri, Feb 01, 2019 at 06:24:24PM +0000, Dexuan Cui wrote:
> > From: Kimberly Brown <kimbrownkd@gmail.com>
> > Sent: Thursday, January 31, 2019 9:47 AM
> > ...
> > 2) Prevent a deadlock that can occur between the proposed mutex_lock()
> > call in the vmbus_chan_attr_show() function and the sysfs/kernfs functions.
> Hi Kim,
> Can you please share more details about the deadlock?
> It's unclear to me how the deadlock happens.
>

Hi Dexuan,

I encountered the deadlock by:
1) Adding a call to msleep() before acquiring the mutex in
vmbus_chan_attr_show()
2) Opening a hv_netvsc subchannel's sysfs file
3) Removing hv_netvsc while the sysfs file is opening

Here's the lockdep output:

[ 164.167699] hv_vmbus: unregistering driver hv_netvsc

[ 164.171660] ======================================================
[ 164.171661] WARNING: possible circular locking dependency detected
[ 164.171663] 5.0.0-rc1+ #58 Not tainted
[ 164.171664] ------------------------------------------------------
[ 164.171666] kworker/0:1/12 is trying to acquire lock:
[ 164.171668] 00000000664f9893 (kn->count#43){++++}, at: kernfs_remove+0x23/0x40
[ 164.171676]
but task is already holding lock:
[ 164.171677] 000000007b9e8443 (&vmbus_connection.channel_mutex){+.+.}, at: vmbus_onoffer_rescind+0x1ae/0x210 [hv_vmbus]
[ 164.171686]
which lock already depends on the new lock.

[ 164.171687]
the existing dependency chain (in reverse order) is:
[ 164.171688]
-> #1 (&vmbus_connection.channel_mutex){+.+.}:
[ 164.171694] __mutex_lock+0x65/0x9b0
[ 164.171696] mutex_lock_nested+0x1b/0x20
[ 164.171700] vmbus_chan_attr_show+0x3f/0x90 [hv_vmbus]
[ 164.171703] sysfs_kf_seq_show+0xa4/0x130
[ 164.171705] kernfs_seq_show+0x2d/0x30
[ 164.171708] seq_read+0xe2/0x410
[ 164.171711] kernfs_fop_read+0x14e/0x1a0
[ 164.171714] __vfs_read+0x3a/0x1a0
[ 164.171716] vfs_read+0x91/0x140
[ 164.171718] ksys_read+0x58/0xc0
[ 164.171721] __x64_sys_read+0x1a/0x20
[ 164.171724] do_syscall_64+0x65/0x1b0
[ 164.171727] entry_SYSCALL_64_after_hwframe+0x49/0xbe
[ 164.171728]
-> #0 (kn->count#43){++++}:
[ 164.171732] lock_acquire+0xa3/0x180
[ 164.171734] __kernfs_remove+0x278/0x300
[ 164.171737] kernfs_remove+0x23/0x40
[ 164.171739] sysfs_remove_dir+0x51/0x60
[ 164.171741] kobject_del.part.7+0x13/0x40
[ 164.171743] kobject_put+0x6a/0x1a0
[ 164.171748] hv_process_channel_removal+0xfe/0x180 [hv_vmbus]
[ 164.171752] vmbus_onoffer_rescind+0x20a/0x210 [hv_vmbus]
[ 164.171756] vmbus_onmessage+0x5f/0x150 [hv_vmbus]
[ 164.171760] vmbus_onmessage_work+0x22/0x30 [hv_vmbus]
[ 164.171763] process_one_work+0x291/0x5c0
[ 164.171765] worker_thread+0x34/0x400
[ 164.171767] kthread+0x124/0x140
[ 164.171770] ret_from_fork+0x3a/0x50
[ 164.171771]
other info that might help us debug this:

[ 164.171772] Possible unsafe locking scenario:

[ 164.171773] CPU0 CPU1
[ 164.171775] ---- ----
[ 164.171776] lock(&vmbus_connection.channel_mutex);
[ 164.171777] lock(kn->count#43);
[ 164.171779] lock(&vmbus_connection.channel_mutex);
[ 164.171781] lock(kn->count#43);
[ 164.171783]
*** DEADLOCK ***

[ 164.171785] 3 locks held by kworker/0:1/12:
[ 164.171786] #0: 000000002128b29f ((wq_completion)"events"){+.+.}, at: process_one_work+0x20f/0x5c0
[ 164.171790] #1: 0000000041d2602c ((work_completion)(&ctx->work)){+.+.}, at: process_one_work+0x20f/0x5c0
[ 164.171794] #2: 000000007b9e8443 (&vmbus_connection.channel_mutex){+.+.}, at: vmbus_onoffer_rescind+0x1ae/0x210 [hv_vmbus]
[ 164.171799]
stack backtrace:
[ 164.171802] CPU: 0 PID: 12 Comm: kworker/0:1 Not tainted 5.0.0-rc1+ #58
[ 164.171804] Hardware name: Microsoft Corporation Virtual Machine/Virtual Machine, BIOS Hyper-V UEFI Release v3.0 03/02/2018
[ 164.171809] Workqueue: events vmbus_onmessage_work [hv_vmbus]
[ 164.171811] Call Trace:
[ 164.171816] dump_stack+0x8e/0xd5
[ 164.171819] print_circular_bug.isra.37+0x1e7/0x1f5
[ 164.171822] __lock_acquire+0x1427/0x1490
[ 164.171826] ? sched_clock+0x9/0x10
[ 164.171830] lock_acquire+0xa3/0x180
[ 164.171832] ? lock_acquire+0xa3/0x180
[ 164.171835] ? kernfs_remove+0x23/0x40
[ 164.171838] __kernfs_remove+0x278/0x300
[ 164.171840] ? kernfs_remove+0x23/0x40
[ 164.171843] kernfs_remove+0x23/0x40
[ 164.171846] sysfs_remove_dir+0x51/0x60
[ 164.171848] kobject_del.part.7+0x13/0x40
[ 164.171850] kobject_put+0x6a/0x1a0
[ 164.171855] hv_process_channel_removal+0xfe/0x180 [hv_vmbus]
[ 164.171859] vmbus_onoffer_rescind+0x20a/0x210 [hv_vmbus]
[ 164.171863] vmbus_onmessage+0x5f/0x150 [hv_vmbus]
[ 164.171868] vmbus_onmessage_work+0x22/0x30 [hv_vmbus]
[ 164.171870] process_one_work+0x291/0x5c0
[ 164.171874] worker_thread+0x34/0x400
[ 164.171877] kthread+0x124/0x140
[ 164.171879] ? process_one_work+0x5c0/0x5c0
[ 164.171881] ? kthread_park+0x90/0x90
[ 164.171884] ret_from_fork+0x3a/0x50

==========================================================================


> > I've identified two possible solutions to the deadlock:
> >
> > 1) Add a new mutex to the vmbus_channel struct which protects changes to
> > "channel->state". Use this new mutex in vmbus_chan_attr_show() instead of
> > "vmbus_connection.channel_mutex".
> >
> > 2) Use "vmbus_connection.channel_mutex" in vmbus_chan_attr_show() as
> > originally proposed, and acquire it with mutex_trylock(). If the mutex
> > cannot be acquired, return -EINVAL.
> It looks more like a workaround. IMO we should try to find a real fix. :-)
>

Good point. I think that the root cause of the deadlock problem is that
the calls to free_channel() in hv_process_channel_removal() are within a
section that's locked with vmbus_connection.channel_mutex. So, first, I
need to determine whether free_channel() needs to be called within the
locked section.

If free_channel() does need to be called within the locked section, then
there's no way to prevent the possibility of a deadlock when
vmbus_connection.channel_mutex is used in vmbus_chan_attr_show(). A
different solution, such as a new mutex to protect changes to
"channel->state", is necessary.

If free_channel() does not need to be in the locked section, then we may
be able to restructure the functions to prevent the deadlock.

I see that there was a race condition problem when
hv_process_channel_removal() was called on an already freed channel
(fixed in commit 192b2d78722f, "Fix bugs in rescind handling"), so I
suspect that free_channel() needs to stay in the locked section.


> > I'm leaning towards (2), using mutex_trylock().
> > "vmbus_connection.channel_mutex"
> > appears to be used only when a channel is being opened or closed, so
> > vmbus_chan_attr_show() should be able to acquire the mutex when the
> > channel is in use.
> >
> > If anyone has suggestions, please let me know.
> >
> > Thanks,
> > Kim
>
> Thanks,
> -- Dexuan

\
 
 \ /
  Last update: 2019-02-02 21:08    [W:0.094 / U:7.800 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site