[RFC,PATCHv2,1/2] drivers: power: Add watchdog timer to catch drivers which lockup during suspend/resume.

Commit Message

Zoran Markovic May 10, 2013, 9:28 p.m. UTC

From: Benoit Goby <[e-mail]>

Below is a patch from android kernel that detects a driver suspend/resume
lockup and captures dump in the kernel log. Please review and provide
comments.

Rather than hard-lock the kernel, dump the suspend/resume thread stack and
BUG() when a driver takes too long to suspend/resume.  The timeout is set to
12 seconds to be longer than the usbhid 10 second timeout.

Exclude from the watchdog the time spent waiting for children that
are resumed asynchronously and time every device, whether or not they
resumed synchronously.

This patch is targeted for mobile devices where a suspend/resume lockup
could cause a system reboot and catch user's attention. Information
about failing device can later be retrieved from captured log in
subsequent boot session.

The hardware watchdog timer is likely suspended during this time and
couldn't be relied upon. The soft-lockup detector would eventually tell
that tasks are not scheduled, but would provide little context as to why.
The patch hence uses system timer and assumes it is still active while the
devices are suspended/resumed.

Cc: Android Kernel Team <[e-mail]>
Cc: Colin Cross <[e-mail]>
Cc: Todd Poynor <[e-mail]>
Cc: San Mehat <[e-mail]>
Cc: Benoit Goby <[e-mail]>
Cc: John Stultz <[e-mail]>
Cc: Pavel Machek <[e-mail]>
Cc: Rafael J. Wysocki <[e-mail]>
Cc: Len Brown <[e-mail]>
Cc: Greg Kroah-Hartman <[e-mail]>
Original-author: San Mehat <[e-mail]>
Signed-off-by: Benoit Goby <[e-mail]>
[[e-mail]: Changed printk(KERN_EMERG,...) to pr_emerg(...),
tweaked commit message. Minor changes to merge code into kernel tip.]
Signed-off-by: Zoran Markovic <[e-mail]>
---
 drivers/base/power/main.c |   66 +++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 66 insertions(+)

Comments

Colin Cross May 11, 2013, 6:13 a.m. UTC | #1

On Fri, May 10, 2013 at 2:28 PM, Zoran Markovic
<[e-mail]> wrote:
> From: Benoit Goby <[e-mail]>
>
> Below is a patch from android kernel that detects a driver suspend/resume
> lockup and captures dump in the kernel log. Please review and provide
> comments.

This paragraph should go below the --- line so it doesn't end up in
the final commit message.

> Rather than hard-lock the kernel, dump the suspend/resume thread stack and
> BUG() when a driver takes too long to suspend/resume.  The timeout is set to
> 12 seconds to be longer than the usbhid 10 second timeout.
>
> Exclude from the watchdog the time spent waiting for children that
> are resumed asynchronously and time every device, whether or not they
> resumed synchronously.
>
> This patch is targeted for mobile devices where a suspend/resume lockup
> could cause a system reboot and catch user's attention. Information
> about failing device can later be retrieved from captured log in
> subsequent boot session.

I would take out the phrase "catch user's attention", the intention of
this patch is actually the opposite - get the system back to working
normally as fast as possible, while still putting enough information
to debug the problem into the log.

> The hardware watchdog timer is likely suspended during this time and
> couldn't be relied upon. The soft-lockup detector would eventually tell
> that tasks are not scheduled, but would provide little context as to why.
> The patch hence uses system timer and assumes it is still active while the
> devices are suspended/resumed.
>
> Cc: Android Kernel Team <[e-mail]>
> Cc: Colin Cross <[e-mail]>
> Cc: Todd Poynor <[e-mail]>
> Cc: San Mehat <[e-mail]>
> Cc: Benoit Goby <[e-mail]>
> Cc: John Stultz <[e-mail]>
> Cc: Pavel Machek <[e-mail]>
> Cc: Rafael J. Wysocki <[e-mail]>
> Cc: Len Brown <[e-mail]>
> Cc: Greg Kroah-Hartman <[e-mail]>
> Original-author: San Mehat <[e-mail]>
> Signed-off-by: Benoit Goby <[e-mail]>
> [[e-mail]: Changed printk(KERN_EMERG,...) to pr_emerg(...),
> tweaked commit message. Minor changes to merge code into kernel tip.]
> Signed-off-by: Zoran Markovic <[e-mail]>
> ---
>  drivers/base/power/main.c |   66 +++++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 66 insertions(+)
>
> diff --git a/drivers/base/power/main.c b/drivers/base/power/main.c
> index 5a9b656..a6a02c0 100644
> --- a/drivers/base/power/main.c
> +++ b/drivers/base/power/main.c
> @@ -29,6 +29,8 @@
>  #include <linux/async.h>
>  #include <linux/suspend.h>
>  #include <linux/cpuidle.h>
> +#include <linux/timer.h>
> +
>  #include "../base.h"
>  #include "power.h"
>
> @@ -54,6 +56,12 @@ struct suspend_stats suspend_stats;
>  static DEFINE_MUTEX(dpm_list_mtx);
>  static pm_message_t pm_transition;
>
> +struct dpm_watchdog {
> +       struct device           *dev;
> +       struct task_struct      *tsk;
> +       struct timer_list       timer;
> +};
> +
>  static int async_error;
>
>  /**
> @@ -384,6 +392,56 @@ static int dpm_run_callback(pm_callback_t cb, struct device *dev,
>         return error;
>  }
>
> +/**
> + * dpm_wd_handler - Driver suspend / resume watchdog handler.
> + *
> + * Called when a driver has timed out suspending or resuming.
> + * There's not much we can do here to recover so BUG() out for
> + * a crash-dump
> + */
> +static void dpm_wd_handler(unsigned long data)
> +{
> +       struct dpm_watchdog *wd = (void *)data;
> +       struct device *dev      = wd->dev;
> +       struct task_struct *tsk = wd->tsk;
> +
> +       dev_emerg(dev, "**** DPM device timeout ****\n");
> +       show_stack(tsk, NULL);
> +
> +       BUG();
> +}
> +
> +/**
> + * dpm_wd_set - Enable pm watchdog for given device.
> + * @wd: Watchdog. Must be allocated on the stack.
> + * @dev: Device to handle.
> + */
> +static void dpm_wd_set(struct dpm_watchdog *wd, struct device *dev)
> +{
> +       struct timer_list *timer = &wd->timer;
> +
> +       wd->dev = dev;
> +       wd->tsk = get_current();
> +
> +       init_timer_on_stack(timer);
> +       timer->expires = jiffies + HZ * 12;
> +       timer->function = dpm_wd_handler;
> +       timer->data = (unsigned long)wd;
> +       add_timer(timer);
> +}
> +
> +/**
> + * dpm_wd_clear - Disable pm watchdog.
> + * @wd: Watchdog to disable.
> + */
> +static void dpm_wd_clear(struct dpm_watchdog *wd)
> +{
> +       struct timer_list *timer = &wd->timer;
> +
> +       del_timer_sync(timer);
> +       destroy_timer_on_stack(timer);
> +}
> +
>  /*------------------------- Resume routines -------------------------*/
>
>  /**
> @@ -570,6 +628,7 @@ static int device_resume(struct device *dev, pm_message_t state, bool async)
>         pm_callback_t callback = NULL;
>         char *info = NULL;
>         int error = 0;
> +       struct dpm_watchdog wd;
>
>         TRACE_DEVICE(dev);
>         TRACE_RESUME(0);
> @@ -585,6 +644,7 @@ static int device_resume(struct device *dev, pm_message_t state, bool async)
>          * a resumed device, even if the device hasn't been completed yet.
>          */
>         dev->power.is_prepared = false;
> +       dpm_wd_set(&wd, dev);
>
>         if (!dev->power.is_suspended)
>                 goto Unlock;
> @@ -636,6 +696,7 @@ static int device_resume(struct device *dev, pm_message_t state, bool async)
>
>   Unlock:
>         device_unlock(dev);
> +       dpm_wd_clear(&wd);
>
>   Complete:
>         complete_all(&dev->power.completion);
> @@ -1053,6 +1114,7 @@ static int __device_suspend(struct device *dev, pm_message_t state, bool async)
>         pm_callback_t callback = NULL;
>         char *info = NULL;
>         int error = 0;
> +       struct dpm_watchdog wd;
>
>         dpm_wait_for_children(dev, async);
>
> @@ -1076,6 +1138,8 @@ static int __device_suspend(struct device *dev, pm_message_t state, bool async)
>         if (dev->power.syscore)
>                 goto Complete;
>
> +       dpm_wd_set(&wd, dev);
> +
>         device_lock(dev);
>
>         if (dev->pm_domain) {
> @@ -1131,6 +1195,8 @@ static int __device_suspend(struct device *dev, pm_message_t state, bool async)
>
>         device_unlock(dev);
>
> +       dpm_wd_clear(&wd);
> +
>   Complete:
>         complete_all(&dev->power.completion);
>         if (error)
> --
> 1.7.9.5
>
--
To unsubscribe from this list: send the line "unsubscribe linux-pm" in
the body of a message to [e-mail]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Rafael Wysocki May 12, 2013, 12:39 a.m. UTC | #2

On Friday, May 10, 2013 11:13:27 PM Colin Cross wrote:
> On Fri, May 10, 2013 at 2:28 PM, Zoran Markovic
> <[e-mail]> wrote:
> > From: Benoit Goby <[e-mail]>
> >
> > Below is a patch from android kernel that detects a driver suspend/resume
> > lockup and captures dump in the kernel log. Please review and provide
> > comments.
> 
> This paragraph should go below the --- line so it doesn't end up in
> the final commit message.
> 
> > Rather than hard-lock the kernel, dump the suspend/resume thread stack and
> > BUG() when a driver takes too long to suspend/resume.

And how exactly is that different from hanging the kernel?

> > The timeout is set to 12 seconds to be longer than the usbhid 10
> > second timeout.

Which is kind of arbitrary.

> > Exclude from the watchdog the time spent waiting for children that
> > are resumed asynchronously and time every device, whether or not they
> > resumed synchronously.

What about changing the code to use wait_for_completion_timeout() instead
of wait_for_completion() and actually trying to recover if one of them times
out?  [It could try to unregister the device in question and if *that* hangs
indefinitely, *then* commit a panic() or something similar, but if it succeeds,
abort the ongoing suspend or complete the resume without that device.]

Of course, that would involve changing things not to depend on async, but
might be better than slapping a timer on top.

> > This patch is targeted for mobile devices where a suspend/resume lockup
> > could cause a system reboot and catch user's attention. Information
> > about failing device can later be retrieved from captured log in
> > subsequent boot session.
> 
> I would take out the phrase "catch user's attention", the intention of
> this patch is actually the opposite - get the system back to working
> normally as fast as possible, while still putting enough information
> to debug the problem into the log.
> 
> > The hardware watchdog timer is likely suspended during this time and
> > couldn't be relied upon. The soft-lockup detector would eventually tell
> > that tasks are not scheduled, but would provide little context as to why.
> > The patch hence uses system timer and assumes it is still active while the
> > devices are suspended/resumed.

I must say I'm not particularly impressed by this patch series.  I understand
the motivation, but I'm wondering if that's the best we can do.

Thanks,
Rafael

Colin Cross May 12, 2013, 7:15 p.m. UTC | #3

On Sat, May 11, 2013 at 5:39 PM, Rafael J. Wysocki <[e-mail]> wrote:
> On Friday, May 10, 2013 11:13:27 PM Colin Cross wrote:
>> On Fri, May 10, 2013 at 2:28 PM, Zoran Markovic
>> <[e-mail]> wrote:
>> > From: Benoit Goby <[e-mail]>
>> >
>> > Below is a patch from android kernel that detects a driver suspend/resume
>> > lockup and captures dump in the kernel log. Please review and provide
>> > comments.
>>
>> This paragraph should go below the --- line so it doesn't end up in
>> the final commit message.
>>
>> > Rather than hard-lock the kernel, dump the suspend/resume thread stack and
>> > BUG() when a driver takes too long to suspend/resume.
>
> And how exactly is that different from hanging the kernel?

This works best with reboot on panic + pstore ram console, which
results in rebooting to a kernel in a good state that can inspect the
stack trace.

>> > The timeout is set to 12 seconds to be longer than the usbhid 10
>> > second timeout.
>
> Which is kind of arbitrary.

The next patch (which should be squashed into this one) makes this
timeout configurable.

>> > Exclude from the watchdog the time spent waiting for children that
>> > are resumed asynchronously and time every device, whether or not they
>> > resumed synchronously.
>
> What about changing the code to use wait_for_completion_timeout() instead
> of wait_for_completion() and actually trying to recover if one of them times
> out?  [It could try to unregister the device in question and if *that* hangs
> indefinitely, *then* commit a panic() or something similar, but if it succeeds,
> abort the ongoing suspend or complete the resume without that device.]

I don't think this is feasible.  What do you do if userspace was using
the device?  What do you do if the suspend function later starts
running again if the blocking condition clears up, and starts
accessing hardware for an unregistered device?  Trying to unload a
driver that is known to be misbehaving sounds to me like it is going
to result in an even more unstable situation.  The purpose of this
patch is to get the end user's phone back into a usable state as fast
as possible, while still providing the information needed to debug the
issue to power users or developers.

> Of course, that would involve changing things not to depend on async, but
> might be better than slapping a timer on top.
>
>> > This patch is targeted for mobile devices where a suspend/resume lockup
>> > could cause a system reboot and catch user's attention. Information
>> > about failing device can later be retrieved from captured log in
>> > subsequent boot session.
>>
>> I would take out the phrase "catch user's attention", the intention of
>> this patch is actually the opposite - get the system back to working
>> normally as fast as possible, while still putting enough information
>> to debug the problem into the log.
>>
>> > The hardware watchdog timer is likely suspended during this time and
>> > couldn't be relied upon. The soft-lockup detector would eventually tell
>> > that tasks are not scheduled, but would provide little context as to why.
>> > The patch hence uses system timer and assumes it is still active while the
>> > devices are suspended/resumed.
>
> I must say I'm not particularly impressed by this patch series.  I understand
> the motivation, but I'm wondering if that's the best we can do.

This feature is most useful for Android devices, which is why we never
pushed it upstream ourselves.  It has been extremely valuable to us
during development, we have been using it for years and it turns a
class of bugs that is very difficult to debug ("phone is stuck with
screen off", none of the regular debugging tools are useful because
most drivers are suspended) into something that can be trivially
collected out of the logs on the next reboot.
--
To unsubscribe from this list: send the line "unsubscribe linux-pm" in
the body of a message to [e-mail]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Rafael Wysocki May 13, 2013, 11:26 a.m. UTC | #4

On Sunday, May 12, 2013 12:15:06 PM Colin Cross wrote:
> On Sat, May 11, 2013 at 5:39 PM, Rafael J. Wysocki <[e-mail]> wrote:
> > On Friday, May 10, 2013 11:13:27 PM Colin Cross wrote:
> >> On Fri, May 10, 2013 at 2:28 PM, Zoran Markovic
> >> <[e-mail]> wrote:
> >> > From: Benoit Goby <[e-mail]>
> >> >
> >> > Below is a patch from android kernel that detects a driver suspend/resume
> >> > lockup and captures dump in the kernel log. Please review and provide
> >> > comments.
> >>
> >> This paragraph should go below the --- line so it doesn't end up in
> >> the final commit message.
> >>
> >> > Rather than hard-lock the kernel, dump the suspend/resume thread stack and
> >> > BUG() when a driver takes too long to suspend/resume.
> >
> > And how exactly is that different from hanging the kernel?
> 
> This works best with reboot on panic + pstore ram console, which
> results in rebooting to a kernel in a good state that can inspect the
> stack trace.
> 
> >> > The timeout is set to 12 seconds to be longer than the usbhid 10
> >> > second timeout.
> >
> > Which is kind of arbitrary.
> 
> The next patch (which should be squashed into this one) makes this
> timeout configurable.
> 
> >> > Exclude from the watchdog the time spent waiting for children that
> >> > are resumed asynchronously and time every device, whether or not they
> >> > resumed synchronously.
> >
> > What about changing the code to use wait_for_completion_timeout() instead
> > of wait_for_completion() and actually trying to recover if one of them times
> > out?  [It could try to unregister the device in question and if *that* hangs
> > indefinitely, *then* commit a panic() or something similar, but if it succeeds,
> > abort the ongoing suspend or complete the resume without that device.]
> 
> I don't think this is feasible.  What do you do if userspace was using
> the device?  What do you do if the suspend function later starts
> running again if the blocking condition clears up, and starts
> accessing hardware for an unregistered device?  Trying to unload a
> driver that is known to be misbehaving sounds to me like it is going
> to result in an even more unstable situation.  The purpose of this
> patch is to get the end user's phone back into a usable state as fast
> as possible, while still providing the information needed to debug the
> issue to power users or developers.

And using BUG() is probably one of the worst ways to achieve that.

On the system I'm using, for example, it will just hang indefinitely and I'm
not going to see the call trace anyway because of the console suspend that
happens before suspending device drivers.

What about this:
 - Add one more list_head to struct dev_pm_info.
 - Make dpm_prepare() create a new list for the next steps instead of moving
   devices out of dpm_list.
 - Start an async work to carry out dpm_suspend() and make the main thread
   do wait_for_completion_timeout() for every device in dpm_list (in the
   reverse order).
 - If it times out, mark the device in question as unusable, possibly resume
   the already suspended devices (except for descendants of the failed one)
   and abort the suspend.  Return a specific error code to user space so that
   it knows what happened.  [You can make this step configurable to BUG()
   instead of doing all those things if you think that will be more useful for
   platforms you care about.]
 - Disable future suspends.
And analogously for resume.

That should allow people to investigate what happened on a system that
(hopefully) is not completely dead and you still can have your "reboot if
suspend hangs" feature if you like.

> > Of course, that would involve changing things not to depend on async, but
> > might be better than slapping a timer on top.
> >
> >> > This patch is targeted for mobile devices where a suspend/resume lockup
> >> > could cause a system reboot and catch user's attention. Information
> >> > about failing device can later be retrieved from captured log in
> >> > subsequent boot session.
> >>
> >> I would take out the phrase "catch user's attention", the intention of
> >> this patch is actually the opposite - get the system back to working
> >> normally as fast as possible, while still putting enough information
> >> to debug the problem into the log.
> >>
> >> > The hardware watchdog timer is likely suspended during this time and
> >> > couldn't be relied upon. The soft-lockup detector would eventually tell
> >> > that tasks are not scheduled, but would provide little context as to why.
> >> > The patch hence uses system timer and assumes it is still active while the
> >> > devices are suspended/resumed.
> >
> > I must say I'm not particularly impressed by this patch series.  I understand
> > the motivation, but I'm wondering if that's the best we can do.
> 
> This feature is most useful for Android devices, which is why we never
> pushed it upstream ourselves.

And that was OK.

> It has been extremely valuable to us during development, we have been using
> it for years and it turns a class of bugs that is very difficult to debug
> ("phone is stuck with screen off", none of the regular debugging tools are
> useful because most drivers are suspended) into something that can be
> trivially collected out of the logs on the next reboot.

That's *if* you have those logs and that's a pretty big "if".

As I said, I understand the motivation and I know why it may be useful.  Still,
my question is if we can possibly do better here.

Thanks,
Rafael

Zoran Markovic May 28, 2013, 6:26 p.m. UTC | #5

> What about this:
>  - Add one more list_head to struct dev_pm_info.
>  - Make dpm_prepare() create a new list for the next steps instead of moving
>    devices out of dpm_list.
>  - Start an async work to carry out dpm_suspend() and make the main thread
>    do wait_for_completion_timeout() for every device in dpm_list (in the
>    reverse order).
>  - If it times out, mark the device in question as unusable, possibly resume
>    the already suspended devices (except for descendants of the failed one)
>    and abort the suspend.  Return a specific error code to user space so that
>    it knows what happened.  [You can make this step configurable to BUG()
>    instead of doing all those things if you think that will be more useful for
>    platforms you care about.]
>  - Disable future suspends.
> And analogously for resume.
>
> That should allow people to investigate what happened on a system that
> (hopefully) is not completely dead and you still can have your "reboot if
> suspend hangs" feature if you like.

I looked into implementing this. The problem that I encountered is
that there is no reliable way of canceling an async task, and hence
the asynchronous __device_suspend() would be left racing with a
recovery from a suspend timeout. We could do cancel_work_sync() as a
recovery, but that call blocks until the running async task is
flushed, which might never happen. So doing a panic() is pretty much
the only option for recovering.
- Zoran
--
To unsubscribe from this list: send the line "unsubscribe linux-pm" in
the body of a message to [e-mail]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Rafael Wysocki May 28, 2013, 8:49 p.m. UTC | #6

On Tuesday, May 28, 2013 11:26:09 AM Zoran Markovic wrote:
> > What about this:
> >  - Add one more list_head to struct dev_pm_info.
> >  - Make dpm_prepare() create a new list for the next steps instead of moving
> >    devices out of dpm_list.
> >  - Start an async work to carry out dpm_suspend() and make the main thread
> >    do wait_for_completion_timeout() for every device in dpm_list (in the
> >    reverse order).
> >  - If it times out, mark the device in question as unusable, possibly resume
> >    the already suspended devices (except for descendants of the failed one)
> >    and abort the suspend.  Return a specific error code to user space so that
> >    it knows what happened.  [You can make this step configurable to BUG()
> >    instead of doing all those things if you think that will be more useful for
> >    platforms you care about.]
> >  - Disable future suspends.
> > And analogously for resume.
> >
> > That should allow people to investigate what happened on a system that
> > (hopefully) is not completely dead and you still can have your "reboot if
> > suspend hangs" feature if you like.
> 
> I looked into implementing this. The problem that I encountered is
> that there is no reliable way of canceling an async task, and hence
> the asynchronous __device_suspend() would be left racing with a
> recovery from a suspend timeout.

Why exactly would it be racing?  We wouldn't call device_resume() for
the device that timed out (and its descendants).

> We could do cancel_work_sync() as a recovery, but that call blocks until the
> running async task is flushed, which might never happen. So doing a panic()
> is pretty much the only option for recovering.

Well, its usefulness is quite limited, then.  That said I'm still not convinced
that this actually is the case.

Thanks,
Rafael

Zoran Markovic May 31, 2013, 9:13 p.m. UTC | #7

>> We could do cancel_work_sync() as a recovery, but that call blocks until the
>> running async task is flushed, which might never happen. So doing a panic()
>> is pretty much the only option for recovering.
>
> Well, its usefulness is quite limited, then.  That said I'm still not convinced
> that this actually is the case.

It does block in my environment, AFAICS. Looking a bit further in the
code, it looks like dpm_suspend() does an async_synchronize_full()
which would wait for all async tasks to complete. This is a
show-stopper because (under the circumstances) the assumption that
every async suspend routine eventually completes doesn't hold.

We could possibly select which async tasks to wait for, but this would
add unnecessary complexity to a feature targeted for debugging. It
seems that this approach - although sounding reasonable - needs to
wait until we have a mechanism to cancel an async task.

Regards, Zoran
--
To unsubscribe from this list: send the line "unsubscribe linux-pm" in
the body of a message to [e-mail]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Zoran Markovic June 5, 2013, 10:17 p.m. UTC | #8

Rafael,

>>> We could do cancel_work_sync() as a recovery, but that call blocks until the
>>> running async task is flushed, which might never happen. So doing a panic()
>>> is pretty much the only option for recovering.
>>
>> Well, its usefulness is quite limited, then.  That said I'm still not convinced
>> that this actually is the case.
>
> It does block in my environment, AFAICS. Looking a bit further in the
> code, it looks like dpm_suspend() does an async_synchronize_full()
> which would wait for all async tasks to complete. This is a
> show-stopper because (under the circumstances) the assumption that
> every async suspend routine eventually completes doesn't hold.
>
> We could possibly select which async tasks to wait for, but this would
> add unnecessary complexity to a feature targeted for debugging. It
> seems that this approach - although sounding reasonable - needs to
> wait until we have a mechanism to cancel an async task.

Looks like the implementation of proposal for an async suspend +
wait_for_completion_timeout is quite complex due to above limitations.
How do we proceed from here? We have the following options:
1. Give up on the idea of having a suspend/resume watchdog.
2. Use the timer implementation (with possible modifications).
3. Wait for the implementation of (or implement) killing of an already
running async work.

Are there any other ideas floating around?

Thanks,
Zoran
--
To unsubscribe from this list: send the line "unsubscribe linux-pm" in
the body of a message to [e-mail]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Rafael Wysocki June 5, 2013, 10:29 p.m. UTC | #9

On Wednesday, June 05, 2013 03:17:59 PM Zoran Markovic wrote:
> Rafael,

Hi,

> >>> We could do cancel_work_sync() as a recovery, but that call blocks until the
> >>> running async task is flushed, which might never happen. So doing a panic()
> >>> is pretty much the only option for recovering.
> >>
> >> Well, its usefulness is quite limited, then.  That said I'm still not convinced
> >> that this actually is the case.
> >
> > It does block in my environment, AFAICS. Looking a bit further in the
> > code, it looks like dpm_suspend() does an async_synchronize_full()
> > which would wait for all async tasks to complete. This is a
> > show-stopper because (under the circumstances) the assumption that
> > every async suspend routine eventually completes doesn't hold.
> >
> > We could possibly select which async tasks to wait for, but this would
> > add unnecessary complexity to a feature targeted for debugging. It
> > seems that this approach - although sounding reasonable - needs to
> > wait until we have a mechanism to cancel an async task.
> 
> Looks like the implementation of proposal for an async suspend +
> wait_for_completion_timeout is quite complex due to above limitations.
> How do we proceed from here? We have the following options:
> 1. Give up on the idea of having a suspend/resume watchdog.
> 2. Use the timer implementation (with possible modifications).
> 3. Wait for the implementation of (or implement) killing of an already
> running async work.
> 
> Are there any other ideas floating around?

I'm not aware of any at the moment, but I really don't think this is urgent.

I think we can revisit it during the 3.12 cycle and decide how to proceed
then.

Thanks,
Rafael

Alan Stern June 6, 2013, 2:12 p.m. UTC | #10

On Wed, 5 Jun 2013, Zoran Markovic wrote:

> > It does block in my environment, AFAICS. Looking a bit further in the
> > code, it looks like dpm_suspend() does an async_synchronize_full()
> > which would wait for all async tasks to complete. This is a
> > show-stopper because (under the circumstances) the assumption that
> > every async suspend routine eventually completes doesn't hold.
> >
> > We could possibly select which async tasks to wait for, but this would
> > add unnecessary complexity to a feature targeted for debugging. It
> > seems that this approach - although sounding reasonable - needs to
> > wait until we have a mechanism to cancel an async task.
> 
> Looks like the implementation of proposal for an async suspend +
> wait_for_completion_timeout is quite complex due to above limitations.
> How do we proceed from here? We have the following options:
> 1. Give up on the idea of having a suspend/resume watchdog.
> 2. Use the timer implementation (with possible modifications).
> 3. Wait for the implementation of (or implement) killing of an already
> running async work.
> 
> Are there any other ideas floating around?

In general, the kernel is not designed to operate when kernel threads 
get killed at random times.  It's also not designed to operate normally 
while in the middle of a system suspend.

This means there is basically no hope of recovering from a hung async 
suspend task.  (In much the same way, there is no hope of recovering 
from any hung kernel thread.)  The best you can accomplish is to store 
some useful information somewhere and either panic or force a reboot.

Given that the usual storage media may be inaccessible, it may not be 
easy to find a place to store the information.

(By the way, what do you do if a _synchronous_ suspend routine hangs?  
The two problems are fairly similar.)

Alan Stern

--
To unsubscribe from this list: send the line "unsubscribe linux-pm" in
the body of a message to [e-mail]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Colin Cross June 10, 2013, 9:25 p.m. UTC | #11

On Thu, Jun 6, 2013 at 7:12 AM, Alan Stern <[e-mail]> wrote:
> On Wed, 5 Jun 2013, Zoran Markovic wrote:
>
>> > It does block in my environment, AFAICS. Looking a bit further in the
>> > code, it looks like dpm_suspend() does an async_synchronize_full()
>> > which would wait for all async tasks to complete. This is a
>> > show-stopper because (under the circumstances) the assumption that
>> > every async suspend routine eventually completes doesn't hold.
>> >
>> > We could possibly select which async tasks to wait for, but this would
>> > add unnecessary complexity to a feature targeted for debugging. It
>> > seems that this approach - although sounding reasonable - needs to
>> > wait until we have a mechanism to cancel an async task.
>>
>> Looks like the implementation of proposal for an async suspend +
>> wait_for_completion_timeout is quite complex due to above limitations.
>> How do we proceed from here? We have the following options:
>> 1. Give up on the idea of having a suspend/resume watchdog.
>> 2. Use the timer implementation (with possible modifications).
>> 3. Wait for the implementation of (or implement) killing of an already
>> running async work.
>>
>> Are there any other ideas floating around?
>
> In general, the kernel is not designed to operate when kernel threads
> get killed at random times.  It's also not designed to operate normally
> while in the middle of a system suspend.
>
> This means there is basically no hope of recovering from a hung async
> suspend task.  (In much the same way, there is no hope of recovering
> from any hung kernel thread.)  The best you can accomplish is to store
> some useful information somewhere and either panic or force a reboot.
>
> Given that the usual storage media may be inaccessible, it may not be
> easy to find a place to store the information.
>
> (By the way, what do you do if a _synchronous_ suspend routine hangs?
> The two problems are fairly similar.)

This is why the original patch dumped a stack trace of the offending
task and panic'd.  There is nothing else useful you can do.
--
To unsubscribe from this list: send the line "unsubscribe linux-pm" in
the body of a message to [e-mail]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Patch

diff --git a/drivers/base/power/main.c b/drivers/base/power/main.c
index 5a9b656..a6a02c0 100644
--- a/drivers/base/power/main.c
+++ b/drivers/base/power/main.c
@@ -29,6 +29,8 @@ 
 #include <linux/async.h>
 #include <linux/suspend.h>
 #include <linux/cpuidle.h>
+#include <linux/timer.h>
+
 #include "../base.h"
 #include "power.h"
 
@@ -54,6 +56,12 @@  struct suspend_stats suspend_stats;
 static DEFINE_MUTEX(dpm_list_mtx);
 static pm_message_t pm_transition;
 
+struct dpm_watchdog {
+	struct device		*dev;
+	struct task_struct	*tsk;
+	struct timer_list	timer;
+};
+
 static int async_error;
 
 /**
@@ -384,6 +392,56 @@  static int dpm_run_callback(pm_callback_t cb, struct device *dev,
 	return error;
 }
 
+/**
+ * dpm_wd_handler - Driver suspend / resume watchdog handler.
+ *
+ * Called when a driver has timed out suspending or resuming.
+ * There's not much we can do here to recover so BUG() out for
+ * a crash-dump
+ */
+static void dpm_wd_handler(unsigned long data)
+{
+	struct dpm_watchdog *wd = (void *)data;
+	struct device *dev      = wd->dev;
+	struct task_struct *tsk = wd->tsk;
+
+	dev_emerg(dev, "**** DPM device timeout ****\n");
+	show_stack(tsk, NULL);
+
+	BUG();
+}
+
+/**
+ * dpm_wd_set - Enable pm watchdog for given device.
+ * @wd: Watchdog. Must be allocated on the stack.
+ * @dev: Device to handle.
+ */
+static void dpm_wd_set(struct dpm_watchdog *wd, struct device *dev)
+{
+	struct timer_list *timer = &wd->timer;
+
+	wd->dev = dev;
+	wd->tsk = get_current();
+
+	init_timer_on_stack(timer);
+	timer->expires = jiffies + HZ * 12;
+	timer->function = dpm_wd_handler;
+	timer->data = (unsigned long)wd;
+	add_timer(timer);
+}
+
+/**
+ * dpm_wd_clear - Disable pm watchdog.
+ * @wd: Watchdog to disable.
+ */
+static void dpm_wd_clear(struct dpm_watchdog *wd)
+{
+	struct timer_list *timer = &wd->timer;
+
+	del_timer_sync(timer);
+	destroy_timer_on_stack(timer);
+}
+
 /*------------------------- Resume routines -------------------------*/
 
 /**
@@ -570,6 +628,7 @@  static int device_resume(struct device *dev, pm_message_t state, bool async)
 	pm_callback_t callback = NULL;
 	char *info = NULL;
 	int error = 0;
+	struct dpm_watchdog wd;
 
 	TRACE_DEVICE(dev);
 	TRACE_RESUME(0);
@@ -585,6 +644,7 @@  static int device_resume(struct device *dev, pm_message_t state, bool async)
 	 * a resumed device, even if the device hasn't been completed yet.
 	 */
 	dev->power.is_prepared = false;
+	dpm_wd_set(&wd, dev);
 
 	if (!dev->power.is_suspended)
 		goto Unlock;
@@ -636,6 +696,7 @@  static int device_resume(struct device *dev, pm_message_t state, bool async)
 
  Unlock:
 	device_unlock(dev);
+	dpm_wd_clear(&wd);
 
  Complete:
 	complete_all(&dev->power.completion);
@@ -1053,6 +1114,7 @@  static int __device_suspend(struct device *dev, pm_message_t state, bool async)
 	pm_callback_t callback = NULL;
 	char *info = NULL;
 	int error = 0;
+	struct dpm_watchdog wd;
 
 	dpm_wait_for_children(dev, async);
 
@@ -1076,6 +1138,8 @@  static int __device_suspend(struct device *dev, pm_message_t state, bool async)
 	if (dev->power.syscore)
 		goto Complete;
 
+	dpm_wd_set(&wd, dev);
+
 	device_lock(dev);
 
 	if (dev->pm_domain) {
@@ -1131,6 +1195,8 @@  static int __device_suspend(struct device *dev, pm_message_t state, bool async)
 
 	device_unlock(dev);
 
+	dpm_wd_clear(&wd);
+
  Complete:
 	complete_all(&dev->power.completion);
 	if (error)