[PATCH] xen/manage: Always freeze/thaw processes when suspend/resuming

[PATCH] xen/manage: Always freeze/thaw processes when suspend/resuming

[Ross Lagerwall]

Always freeze processes when suspending and thaw processes when resuming
to prevent a race noticeable with HVM guests.

This prevents a deadlock where the khubd kthread (which is designed to
be freezable) acquires a usb device lock and then tries to allocate
memory which requires the disk which hasn't been resumed yet.
Meanwhile, the xenwatch thread deadlocks waiting for the usb device
lock.

Freezing processes fixes this because the khubd thread is only thawed
after the xenwatch thread finishes resuming all the devices.

Signed-off-by: Ross Lagerwall <ross.lagerwall@citrix.com>
---
 drivers/xen/manage.c | 7 -------
 1 file changed, 7 deletions(-)

diff --git a/drivers/xen/manage.c b/drivers/xen/manage.c
index 5f1e1f3..f8bb36f 100644
--- a/drivers/xen/manage.c
+++ b/drivers/xen/manage.c
@@ -103,16 +103,11 @@ static void do_suspend(void)
 
 	shutting_down = SHUTDOWN_SUSPEND;
 
-#ifdef CONFIG_PREEMPT
-	/* If the kernel is preemptible, we need to freeze all the processes
-	   to prevent them from being in the middle of a pagetable update
-	   during suspend. */
 	err = freeze_processes();
 	if (err) {
 		pr_err("%s: freeze failed %d\n", __func__, err);
 		goto out;
 	}
-#endif
 
 	err = dpm_suspend_start(PMSG_FREEZE);
 	if (err) {
@@ -157,10 +152,8 @@ out_resume:
 	dpm_resume_end(si.cancelled ? PMSG_THAW : PMSG_RESTORE);
 
 out_thaw:
-#ifdef CONFIG_PREEMPT
 	thaw_processes();
 out:
-#endif
 	shutting_down = SHUTDOWN_INVALID;
 }
 #endif	/* CONFIG_HIBERNATE_CALLBACKS */
-- 
1.9.3

Re: [PATCH] xen/manage: Always freeze/thaw processes when suspend/resuming

[David Vrabel]

On 18/08/14 10:41, Ross Lagerwall wrote:
> Always freeze processes when suspending and thaw processes when resuming
> to prevent a race noticeable with HVM guests.
> 
> This prevents a deadlock where the khubd kthread (which is designed to
> be freezable) acquires a usb device lock and then tries to allocate
> memory which requires the disk which hasn't been resumed yet.
> Meanwhile, the xenwatch thread deadlocks waiting for the usb device
> lock.
> 
> Freezing processes fixes this because the khubd thread is only thawed
> after the xenwatch thread finishes resuming all the devices.

I think this is the safest fix for this issue but I'm a bit concerned
about how much extra time the freeze_processes() call will add to the
suspend.

Can you measure how long it takes for a typical VM?

David

Re: [PATCH] xen/manage: Always freeze/thaw processes when suspend/resuming

[Ross Lagerwall]

On 08/22/2014 12:00 PM, David Vrabel wrote:
> On 18/08/14 10:41, Ross Lagerwall wrote:
>> Always freeze processes when suspending and thaw processes when resuming
>> to prevent a race noticeable with HVM guests.
>>
>> This prevents a deadlock where the khubd kthread (which is designed to
>> be freezable) acquires a usb device lock and then tries to allocate
>> memory which requires the disk which hasn't been resumed yet.
>> Meanwhile, the xenwatch thread deadlocks waiting for the usb device
>> lock.
>>
>> Freezing processes fixes this because the khubd thread is only thawed
>> after the xenwatch thread finishes resuming all the devices.
>
> I think this is the safest fix for this issue but I'm a bit concerned
> about how much extra time the freeze_processes() call will add to the
> suspend.
>
> Can you measure how long it takes for a typical VM?
>

With 2 vCPUs and 1024MB RAM, 8ms for freeze_processes() and 6ms for 
thaw_processes().

With 4 vCPUs and 4096MB RAM, 8ms for freeze_processes() and 5ms for 
thaw_processes().

With 4 vCPUs and 4096MB RAM running stress -c 4, 12ms for 
freeze_processes() and 66ms for thaw_processes().

It doesn't seem to take too long...

Cheers
-- 
Ross Lagerwall