On 05/30/2014 10:07 AM, Jan Kara wrote:
> On Fri 30-05-14 09:58:14, Peter Hurley wrote:
>> > On 05/30/2014 09:11 AM, Sasha Levin wrote:
>>> > >Hi all,
>>> > >
>>> > >I sometime see lockups when booting my KVM guest with the latest -next kernel,
>>> > >it basically hangs right when it should start 'init', and after a while I get
>>> > >the following spew:
>>> > >
>>> > >[ 30.790833] BUG: spinlock lockup suspected on CPU#1, swapper/1/0
>> >
>> > Maybe related to this report: https://lkml.org/lkml/2014/5/30/26
>> > from Jet Chen which was bisected to
>> >
>> > commit bafe980f5afc7ccc693fd8c81c8aa5a02fbb5ae0
>> > Author: Jan Kara <[email protected]>
>> > AuthorDate: Thu May 22 10:43:35 2014 +1000
>> > Commit: Stephen Rothwell <[email protected]>
>> > CommitDate: Thu May 22 10:43:35 2014 +1000
>> >
>> > printk: enable interrupts before calling console_trylock_for_printk()
>> > We need interrupts disabled when calling console_trylock_for_printk() only
>> > so that cpu id we pass to can_use_console() remains valid (for other
>> > things console_sem provides all the exclusion we need and deadlocks on
>> > console_sem due to interrupts are impossible because we use
>> > down_trylock()). However if we are rescheduled, we are guaranteed to run
>> > on an online cpu so we can easily just get the cpu id in
>> > can_use_console().
>> > We can lose a bit of performance when we enable interrupts in
>> > vprintk_emit() and then disable them again in console_unlock() but OTOH it
>> > can somewhat reduce interrupt latency caused by console_unlock()
>> > especially since later in the patch series we will want to spin on
>> > console_sem in console_trylock_for_printk().
>> > Signed-off-by: Jan Kara <[email protected]>
>> > Signed-off-by: Andrew Morton <[email protected]>
>> >
>> > ?
> Yeah, very likely. I think I see the problem, I'll send the fix shortly.
Hi Jan,
It seems that the issue I'm seeing is different from the "[prink] BUG: spinlock
lockup suspected on CPU#0, swapper/1".
Is there anything else I could try here? The issue is very common during testing.
Thanks,
Sasha
On 06/06/2014 03:05 PM, Sasha Levin wrote:
> On 05/30/2014 10:07 AM, Jan Kara wrote:
>> On Fri 30-05-14 09:58:14, Peter Hurley wrote:
>>>> On 05/30/2014 09:11 AM, Sasha Levin wrote:
>>>>>> Hi all,
>>>>>>
>>>>>> I sometime see lockups when booting my KVM guest with the latest -next kernel,
>>>>>> it basically hangs right when it should start 'init', and after a while I get
>>>>>> the following spew:
>>>>>>
>>>>>> [ 30.790833] BUG: spinlock lockup suspected on CPU#1, swapper/1/0
>>>>
>>>> Maybe related to this report: https://lkml.org/lkml/2014/5/30/26
>>>> from Jet Chen which was bisected to
>>>>
>>>> commit bafe980f5afc7ccc693fd8c81c8aa5a02fbb5ae0
>>>> Author: Jan Kara <[email protected]>
>>>> AuthorDate: Thu May 22 10:43:35 2014 +1000
>>>> Commit: Stephen Rothwell <[email protected]>
>>>> CommitDate: Thu May 22 10:43:35 2014 +1000
>>>>
>>>> printk: enable interrupts before calling console_trylock_for_printk()
>>>> We need interrupts disabled when calling console_trylock_for_printk() only
>>>> so that cpu id we pass to can_use_console() remains valid (for other
>>>> things console_sem provides all the exclusion we need and deadlocks on
>>>> console_sem due to interrupts are impossible because we use
>>>> down_trylock()). However if we are rescheduled, we are guaranteed to run
>>>> on an online cpu so we can easily just get the cpu id in
>>>> can_use_console().
>>>> We can lose a bit of performance when we enable interrupts in
>>>> vprintk_emit() and then disable them again in console_unlock() but OTOH it
>>>> can somewhat reduce interrupt latency caused by console_unlock()
>>>> especially since later in the patch series we will want to spin on
>>>> console_sem in console_trylock_for_printk().
>>>> Signed-off-by: Jan Kara <[email protected]>
>>>> Signed-off-by: Andrew Morton <[email protected]>
>>>>
>>>> ?
>> Yeah, very likely. I think I see the problem, I'll send the fix shortly.
>
> Hi Jan,
>
> It seems that the issue I'm seeing is different from the "[prink] BUG: spinlock
> lockup suspected on CPU#0, swapper/1".
>
> Is there anything else I could try here? The issue is very common during testing.
Sasha,
Is this bisectable? Maybe that's the best way forward here.
Regards,
Peter Hurley
On 06/10/2014 11:59 AM, Peter Hurley wrote:
> On 06/06/2014 03:05 PM, Sasha Levin wrote:
>> On 05/30/2014 10:07 AM, Jan Kara wrote:
>>> On Fri 30-05-14 09:58:14, Peter Hurley wrote:
>>>>> On 05/30/2014 09:11 AM, Sasha Levin wrote:
>>>>>>> Hi all,
>>>>>>>
>>>>>>> I sometime see lockups when booting my KVM guest with the latest -next kernel,
>>>>>>> it basically hangs right when it should start 'init', and after a while I get
>>>>>>> the following spew:
>>>>>>>
>>>>>>> [ 30.790833] BUG: spinlock lockup suspected on CPU#1, swapper/1/0
>>>>>
>>>>> Maybe related to this report: https://lkml.org/lkml/2014/5/30/26
>>>>> from Jet Chen which was bisected to
>>>>>
>>>>> commit bafe980f5afc7ccc693fd8c81c8aa5a02fbb5ae0
>>>>> Author: Jan Kara <[email protected]>
>>>>> AuthorDate: Thu May 22 10:43:35 2014 +1000
>>>>> Commit: Stephen Rothwell <[email protected]>
>>>>> CommitDate: Thu May 22 10:43:35 2014 +1000
>>>>>
>>>>> printk: enable interrupts before calling console_trylock_for_printk()
>>>>> We need interrupts disabled when calling console_trylock_for_printk() only
>>>>> so that cpu id we pass to can_use_console() remains valid (for other
>>>>> things console_sem provides all the exclusion we need and deadlocks on
>>>>> console_sem due to interrupts are impossible because we use
>>>>> down_trylock()). However if we are rescheduled, we are guaranteed to run
>>>>> on an online cpu so we can easily just get the cpu id in
>>>>> can_use_console().
>>>>> We can lose a bit of performance when we enable interrupts in
>>>>> vprintk_emit() and then disable them again in console_unlock() but OTOH it
>>>>> can somewhat reduce interrupt latency caused by console_unlock()
>>>>> especially since later in the patch series we will want to spin on
>>>>> console_sem in console_trylock_for_printk().
>>>>> Signed-off-by: Jan Kara <[email protected]>
>>>>> Signed-off-by: Andrew Morton <[email protected]>
>>>>>
>>>>> ?
>>> Yeah, very likely. I think I see the problem, I'll send the fix shortly.
>>
>> Hi Jan,
>>
>> It seems that the issue I'm seeing is different from the "[prink] BUG: spinlock
>> lockup suspected on CPU#0, swapper/1".
>>
>> Is there anything else I could try here? The issue is very common during testing.
>
> Sasha,
>
> Is this bisectable? Maybe that's the best way forward here.
I've ran a bisection again and ended up at the same commit as Jet Chen (the commit
unfortunately already made it to Linus's tree).
Note that I did try Jan's proposed fix and that didn't solve the issue for me, I
believe we're seeing different issues caused by the same commit.
939f04bec1a4ef6ba4370b0f34b01decc844b1b1 is the first bad commit
commit 939f04bec1a4ef6ba4370b0f34b01decc844b1b1
Author: Jan Kara <[email protected]>
Date: Wed Jun 4 16:11:37 2014 -0700
printk: enable interrupts before calling console_trylock_for_printk()
We need interrupts disabled when calling console_trylock_for_printk()
only so that cpu id we pass to can_use_console() remains valid (for
other things console_sem provides all the exclusion we need and
deadlocks on console_sem due to interrupts are impossible because we use
down_trylock()). However if we are rescheduled, we are guaranteed to
run on an online cpu so we can easily just get the cpu id in
can_use_console().
We can lose a bit of performance when we enable interrupts in
vprintk_emit() and then disable them again in console_unlock() but OTOH
it can somewhat reduce interrupt latency caused by console_unlock()
especially since later in the patch series we will want to spin on
console_sem in console_trylock_for_printk().
Signed-off-by: Jan Kara <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Thanks,
Sasha
On 06/11/2014 10:55 AM, Sasha Levin wrote:
> On 06/10/2014 11:59 AM, Peter Hurley wrote:
>> On 06/06/2014 03:05 PM, Sasha Levin wrote:
>>> On 05/30/2014 10:07 AM, Jan Kara wrote:
>>>> On Fri 30-05-14 09:58:14, Peter Hurley wrote:
>>>>>> On 05/30/2014 09:11 AM, Sasha Levin wrote:
>>>>>>>> Hi all,
>>>>>>>>
>>>>>>>> I sometime see lockups when booting my KVM guest with the latest -next kernel,
>>>>>>>> it basically hangs right when it should start 'init', and after a while I get
>>>>>>>> the following spew:
>>>>>>>>
>>>>>>>> [ 30.790833] BUG: spinlock lockup suspected on CPU#1, swapper/1/0
>>>>>>
>>>>>> Maybe related to this report: https://lkml.org/lkml/2014/5/30/26
>>>>>> from Jet Chen which was bisected to
>>>>>>
>>>>>> commit bafe980f5afc7ccc693fd8c81c8aa5a02fbb5ae0
>>>>>> Author: Jan Kara <[email protected]>
>>>>>> AuthorDate: Thu May 22 10:43:35 2014 +1000
>>>>>> Commit: Stephen Rothwell <[email protected]>
>>>>>> CommitDate: Thu May 22 10:43:35 2014 +1000
>>>>>>
>>>>>> printk: enable interrupts before calling console_trylock_for_printk()
>>>>>> We need interrupts disabled when calling console_trylock_for_printk() only
>>>>>> so that cpu id we pass to can_use_console() remains valid (for other
>>>>>> things console_sem provides all the exclusion we need and deadlocks on
>>>>>> console_sem due to interrupts are impossible because we use
>>>>>> down_trylock()). However if we are rescheduled, we are guaranteed to run
>>>>>> on an online cpu so we can easily just get the cpu id in
>>>>>> can_use_console().
>>>>>> We can lose a bit of performance when we enable interrupts in
>>>>>> vprintk_emit() and then disable them again in console_unlock() but OTOH it
>>>>>> can somewhat reduce interrupt latency caused by console_unlock()
>>>>>> especially since later in the patch series we will want to spin on
>>>>>> console_sem in console_trylock_for_printk().
>>>>>> Signed-off-by: Jan Kara <[email protected]>
>>>>>> Signed-off-by: Andrew Morton <[email protected]>
>>>>>>
>>>>>> ?
>>>> Yeah, very likely. I think I see the problem, I'll send the fix shortly.
>>>
>>> Hi Jan,
>>>
>>> It seems that the issue I'm seeing is different from the "[prink] BUG: spinlock
>>> lockup suspected on CPU#0, swapper/1".
>>>
>>> Is there anything else I could try here? The issue is very common during testing.
>>
>> Sasha,
>>
>> Is this bisectable? Maybe that's the best way forward here.
>
> I've ran a bisection again and ended up at the same commit as Jet Chen (the commit
> unfortunately already made it to Linus's tree).
>
> Note that I did try Jan's proposed fix and that didn't solve the issue for me, I
> believe we're seeing different issues caused by the same commit.
>
>
> 939f04bec1a4ef6ba4370b0f34b01decc844b1b1 is the first bad commit
> commit 939f04bec1a4ef6ba4370b0f34b01decc844b1b1
> Author: Jan Kara <[email protected]>
> Date: Wed Jun 4 16:11:37 2014 -0700
>
> printk: enable interrupts before calling console_trylock_for_printk()
>
> We need interrupts disabled when calling console_trylock_for_printk()
> only so that cpu id we pass to can_use_console() remains valid (for
> other things console_sem provides all the exclusion we need and
> deadlocks on console_sem due to interrupts are impossible because we use
> down_trylock()). However if we are rescheduled, we are guaranteed to
> run on an online cpu so we can easily just get the cpu id in
> can_use_console().
>
> We can lose a bit of performance when we enable interrupts in
> vprintk_emit() and then disable them again in console_unlock() but OTOH
> it can somewhat reduce interrupt latency caused by console_unlock()
> especially since later in the patch series we will want to spin on
> console_sem in console_trylock_for_printk().
>
> Signed-off-by: Jan Kara <[email protected]>
> Signed-off-by: Andrew Morton <[email protected]>
> Signed-off-by: Linus Torvalds <[email protected]>
I apologize; I didn't look at the patch very closely, but now that I do, this
sticks out:
@@ -1597,17 +1599,22 @@ asmlinkage int vprintk_emit(int facility, int level,
logbuf_cpu = UINT_MAX;
raw_spin_unlock(&logbuf_lock);
+ lockdep_on();
+ local_irq_restore(flags);
+
What prevents cpu migration right here?
If nothing, then logbuf_cpu is now stale and the recursion test at
the top of vprintk_emit is doing nothing to prevent recursion.
+ /*
+ * Disable preemption to avoid being preempted while holding
+ * console_sem which would prevent anyone from printing to console
+ */
+ preempt_disable();
/*
* Try to acquire and then immediately release the console semaphore.
* The release will print out buffers and wake up /dev/kmsg and syslog()
* users.
*/
- if (console_trylock_for_printk(this_cpu))
+ if (console_trylock_for_printk())
console_unlock();
Regards,
Peter Hurley
On Wed, Jun 11, 2014 at 7:55 AM, Sasha Levin <[email protected]> wrote:
>
> I've ran a bisection again and ended up at the same commit as Jet Chen (the commit
> unfortunately already made it to Linus's tree).
Ok, I'm going to revert that commit. The games it plays with cpu
numbers and preemption make me worried, and it looks silly.
If the only reason for that patch is interrupt latency, then if the
"preempt_disable()" had been done *before* restoring irq's, none of
those "let's reload the CPU number" games would have been necessary,
and the patch could have been smaller. So even if that is the solution
to the problems people report, it's actually *better* to just revert
the patch entirely, and try again with a smaller and simpler version,
rather than try to change the patch after-the-fact.
Linus
On Wed, Jun 11, 2014 at 10:38 AM, Linus Torvalds
<[email protected]> wrote:
>
> Ok, I'm going to revert that commit. The games it plays with cpu
> numbers and preemption make me worried, and it looks silly.
>
> If the only reason for that patch is interrupt latency, then if the
> "preempt_disable()" had been done *before* restoring irq's, none of
> those "let's reload the CPU number" games would have been necessary,
> and the patch could have been smaller.
Ugh. It doesn't revert cleanly, so I guess I'm not just reverting it
after all. Jan?
Linus
On Wed 11-06-14 11:34:28, Peter Hurley wrote:
> On 06/11/2014 10:55 AM, Sasha Levin wrote:
> >On 06/10/2014 11:59 AM, Peter Hurley wrote:
> >>On 06/06/2014 03:05 PM, Sasha Levin wrote:
> >>>On 05/30/2014 10:07 AM, Jan Kara wrote:
> >>>>On Fri 30-05-14 09:58:14, Peter Hurley wrote:
> >>>>>>On 05/30/2014 09:11 AM, Sasha Levin wrote:
> >>>>>>>>Hi all,
> >>>>>>>>
> >>>>>>>>I sometime see lockups when booting my KVM guest with the latest -next kernel,
> >>>>>>>>it basically hangs right when it should start 'init', and after a while I get
> >>>>>>>>the following spew:
> >>>>>>>>
> >>>>>>>>[ 30.790833] BUG: spinlock lockup suspected on CPU#1, swapper/1/0
> >>>>>>
> >>>>>>Maybe related to this report: https://lkml.org/lkml/2014/5/30/26
> >>>>>>from Jet Chen which was bisected to
> >>>>>>
> >>>>>>commit bafe980f5afc7ccc693fd8c81c8aa5a02fbb5ae0
> >>>>>>Author: Jan Kara <[email protected]>
> >>>>>>AuthorDate: Thu May 22 10:43:35 2014 +1000
> >>>>>>Commit: Stephen Rothwell <[email protected]>
> >>>>>>CommitDate: Thu May 22 10:43:35 2014 +1000
> >>>>>>
> >>>>>> printk: enable interrupts before calling console_trylock_for_printk()
> >>>>>> We need interrupts disabled when calling console_trylock_for_printk() only
> >>>>>> so that cpu id we pass to can_use_console() remains valid (for other
> >>>>>> things console_sem provides all the exclusion we need and deadlocks on
> >>>>>> console_sem due to interrupts are impossible because we use
> >>>>>> down_trylock()). However if we are rescheduled, we are guaranteed to run
> >>>>>> on an online cpu so we can easily just get the cpu id in
> >>>>>> can_use_console().
> >>>>>> We can lose a bit of performance when we enable interrupts in
> >>>>>> vprintk_emit() and then disable them again in console_unlock() but OTOH it
> >>>>>> can somewhat reduce interrupt latency caused by console_unlock()
> >>>>>> especially since later in the patch series we will want to spin on
> >>>>>> console_sem in console_trylock_for_printk().
> >>>>>> Signed-off-by: Jan Kara <[email protected]>
> >>>>>> Signed-off-by: Andrew Morton <[email protected]>
> >>>>>>
> >>>>>>?
> >>>> Yeah, very likely. I think I see the problem, I'll send the fix shortly.
> >>>
> >>>Hi Jan,
> >>>
> >>>It seems that the issue I'm seeing is different from the "[prink] BUG: spinlock
> >>>lockup suspected on CPU#0, swapper/1".
> >>>
> >>>Is there anything else I could try here? The issue is very common during testing.
> >>
> >>Sasha,
> >>
> >>Is this bisectable? Maybe that's the best way forward here.
> >
> >I've ran a bisection again and ended up at the same commit as Jet Chen (the commit
> >unfortunately already made it to Linus's tree).
> >
> >Note that I did try Jan's proposed fix and that didn't solve the issue for me, I
> >believe we're seeing different issues caused by the same commit.
> >
> >
> >939f04bec1a4ef6ba4370b0f34b01decc844b1b1 is the first bad commit
> >commit 939f04bec1a4ef6ba4370b0f34b01decc844b1b1
> >Author: Jan Kara <[email protected]>
> >Date: Wed Jun 4 16:11:37 2014 -0700
> >
> > printk: enable interrupts before calling console_trylock_for_printk()
> >
> > We need interrupts disabled when calling console_trylock_for_printk()
> > only so that cpu id we pass to can_use_console() remains valid (for
> > other things console_sem provides all the exclusion we need and
> > deadlocks on console_sem due to interrupts are impossible because we use
> > down_trylock()). However if we are rescheduled, we are guaranteed to
> > run on an online cpu so we can easily just get the cpu id in
> > can_use_console().
> >
> > We can lose a bit of performance when we enable interrupts in
> > vprintk_emit() and then disable them again in console_unlock() but OTOH
> > it can somewhat reduce interrupt latency caused by console_unlock()
> > especially since later in the patch series we will want to spin on
> > console_sem in console_trylock_for_printk().
> >
> > Signed-off-by: Jan Kara <[email protected]>
> > Signed-off-by: Andrew Morton <[email protected]>
> > Signed-off-by: Linus Torvalds <[email protected]>
>
> I apologize; I didn't look at the patch very closely, but now that I do,
> this sticks out:
>
> @@ -1597,17 +1599,22 @@ asmlinkage int vprintk_emit(int facility, int level,
>
> logbuf_cpu = UINT_MAX;
> raw_spin_unlock(&logbuf_lock);
> + lockdep_on();
> + local_irq_restore(flags);
> +
>
> What prevents cpu migration right here?
Nothing.
> If nothing, then logbuf_cpu is now stale and the recursion test at
> the top of vprintk_emit is doing nothing to prevent recursion.
Well, note that logbuf_cpu has just been set to UINT_MAX (i.e. undefined)
two lines above. So my patch changes nothing wrt. how printk recursion
detection works (at least AFAICT).
> + /*
> + * Disable preemption to avoid being preempted while holding
> + * console_sem which would prevent anyone from printing to console
> + */
> + preempt_disable();
> /*
> * Try to acquire and then immediately release the console semaphore.
> * The release will print out buffers and wake up /dev/kmsg and syslog()
> * users.
> */
> - if (console_trylock_for_printk(this_cpu))
> + if (console_trylock_for_printk())
> console_unlock();
Honza
--
Jan Kara <[email protected]>
SUSE Labs, CR
On Wed 11-06-14 10:55:55, Sasha Levin wrote:
> On 06/10/2014 11:59 AM, Peter Hurley wrote:
> > On 06/06/2014 03:05 PM, Sasha Levin wrote:
> >> On 05/30/2014 10:07 AM, Jan Kara wrote:
> >>> On Fri 30-05-14 09:58:14, Peter Hurley wrote:
> >>>>> On 05/30/2014 09:11 AM, Sasha Levin wrote:
> >>>>>>> Hi all,
> >>>>>>>
> >>>>>>> I sometime see lockups when booting my KVM guest with the latest -next kernel,
> >>>>>>> it basically hangs right when it should start 'init', and after a while I get
> >>>>>>> the following spew:
> >>>>>>>
> >>>>>>> [ 30.790833] BUG: spinlock lockup suspected on CPU#1, swapper/1/0
> >>>>>
> >>>>> Maybe related to this report: https://lkml.org/lkml/2014/5/30/26
> >>>>> from Jet Chen which was bisected to
> >>>>>
> >>>>> commit bafe980f5afc7ccc693fd8c81c8aa5a02fbb5ae0
> >>>>> Author: Jan Kara <[email protected]>
> >>>>> AuthorDate: Thu May 22 10:43:35 2014 +1000
> >>>>> Commit: Stephen Rothwell <[email protected]>
> >>>>> CommitDate: Thu May 22 10:43:35 2014 +1000
> >>>>>
> >>>>> printk: enable interrupts before calling console_trylock_for_printk()
> >>>>> We need interrupts disabled when calling console_trylock_for_printk() only
> >>>>> so that cpu id we pass to can_use_console() remains valid (for other
> >>>>> things console_sem provides all the exclusion we need and deadlocks on
> >>>>> console_sem due to interrupts are impossible because we use
> >>>>> down_trylock()). However if we are rescheduled, we are guaranteed to run
> >>>>> on an online cpu so we can easily just get the cpu id in
> >>>>> can_use_console().
> >>>>> We can lose a bit of performance when we enable interrupts in
> >>>>> vprintk_emit() and then disable them again in console_unlock() but OTOH it
> >>>>> can somewhat reduce interrupt latency caused by console_unlock()
> >>>>> especially since later in the patch series we will want to spin on
> >>>>> console_sem in console_trylock_for_printk().
> >>>>> Signed-off-by: Jan Kara <[email protected]>
> >>>>> Signed-off-by: Andrew Morton <[email protected]>
> >>>>>
> >>>>> ?
> >>> Yeah, very likely. I think I see the problem, I'll send the fix shortly.
> >>
> >> Hi Jan,
> >>
> >> It seems that the issue I'm seeing is different from the "[prink] BUG: spinlock
> >> lockup suspected on CPU#0, swapper/1".
> >>
> >> Is there anything else I could try here? The issue is very common during testing.
> >
> > Sasha,
> >
> > Is this bisectable? Maybe that's the best way forward here.
>
> I've ran a bisection again and ended up at the same commit as Jet Chen
> (the commit unfortunately already made it to Linus's tree).
>
> Note that I did try Jan's proposed fix and that didn't solve the issue
> for me, I believe we're seeing different issues caused by the same
> commit.
Sorry it has been busy time lately and I didn't have as much time to look
into this as would be needed.
>
>
> 939f04bec1a4ef6ba4370b0f34b01decc844b1b1 is the first bad commit
> commit 939f04bec1a4ef6ba4370b0f34b01decc844b1b1
> Author: Jan Kara <[email protected]>
> Date: Wed Jun 4 16:11:37 2014 -0700
>
> printk: enable interrupts before calling console_trylock_for_printk()
>
> We need interrupts disabled when calling console_trylock_for_printk()
> only so that cpu id we pass to can_use_console() remains valid (for
> other things console_sem provides all the exclusion we need and
> deadlocks on console_sem due to interrupts are impossible because we use
> down_trylock()). However if we are rescheduled, we are guaranteed to
> run on an online cpu so we can easily just get the cpu id in
> can_use_console().
>
> We can lose a bit of performance when we enable interrupts in
> vprintk_emit() and then disable them again in console_unlock() but OTOH
> it can somewhat reduce interrupt latency caused by console_unlock()
> especially since later in the patch series we will want to spin on
> console_sem in console_trylock_for_printk().
>
> Signed-off-by: Jan Kara <[email protected]>
> Signed-off-by: Andrew Morton <[email protected]>
> Signed-off-by: Linus Torvalds <[email protected]>
>
>
> Thanks,
> Sasha
--
Jan Kara <[email protected]>
SUSE Labs, CR
On Wed 11-06-14 22:34:36, Jan Kara wrote:
> On Wed 11-06-14 10:55:55, Sasha Levin wrote:
> > On 06/10/2014 11:59 AM, Peter Hurley wrote:
> > > On 06/06/2014 03:05 PM, Sasha Levin wrote:
> > >> On 05/30/2014 10:07 AM, Jan Kara wrote:
> > >>> On Fri 30-05-14 09:58:14, Peter Hurley wrote:
> > >>>>> On 05/30/2014 09:11 AM, Sasha Levin wrote:
> > >>>>>>> Hi all,
> > >>>>>>>
> > >>>>>>> I sometime see lockups when booting my KVM guest with the latest -next kernel,
> > >>>>>>> it basically hangs right when it should start 'init', and after a while I get
> > >>>>>>> the following spew:
> > >>>>>>>
> > >>>>>>> [ 30.790833] BUG: spinlock lockup suspected on CPU#1, swapper/1/0
> > >>>>>
> > >>>>> Maybe related to this report: https://lkml.org/lkml/2014/5/30/26
> > >>>>> from Jet Chen which was bisected to
> > >>>>>
> > >>>>> commit bafe980f5afc7ccc693fd8c81c8aa5a02fbb5ae0
> > >>>>> Author: Jan Kara <[email protected]>
> > >>>>> AuthorDate: Thu May 22 10:43:35 2014 +1000
> > >>>>> Commit: Stephen Rothwell <[email protected]>
> > >>>>> CommitDate: Thu May 22 10:43:35 2014 +1000
> > >>>>>
> > >>>>> printk: enable interrupts before calling console_trylock_for_printk()
> > >>>>> We need interrupts disabled when calling console_trylock_for_printk() only
> > >>>>> so that cpu id we pass to can_use_console() remains valid (for other
> > >>>>> things console_sem provides all the exclusion we need and deadlocks on
> > >>>>> console_sem due to interrupts are impossible because we use
> > >>>>> down_trylock()). However if we are rescheduled, we are guaranteed to run
> > >>>>> on an online cpu so we can easily just get the cpu id in
> > >>>>> can_use_console().
> > >>>>> We can lose a bit of performance when we enable interrupts in
> > >>>>> vprintk_emit() and then disable them again in console_unlock() but OTOH it
> > >>>>> can somewhat reduce interrupt latency caused by console_unlock()
> > >>>>> especially since later in the patch series we will want to spin on
> > >>>>> console_sem in console_trylock_for_printk().
> > >>>>> Signed-off-by: Jan Kara <[email protected]>
> > >>>>> Signed-off-by: Andrew Morton <[email protected]>
> > >>>>>
> > >>>>> ?
> > >>> Yeah, very likely. I think I see the problem, I'll send the fix shortly.
> > >>
> > >> Hi Jan,
> > >>
> > >> It seems that the issue I'm seeing is different from the "[prink] BUG: spinlock
> > >> lockup suspected on CPU#0, swapper/1".
> > >>
> > >> Is there anything else I could try here? The issue is very common during testing.
> > >
> > > Sasha,
> > >
> > > Is this bisectable? Maybe that's the best way forward here.
> >
> > I've ran a bisection again and ended up at the same commit as Jet Chen
> > (the commit unfortunately already made it to Linus's tree).
> >
> > Note that I did try Jan's proposed fix and that didn't solve the issue
> > for me, I believe we're seeing different issues caused by the same
> > commit.
> Sorry it has been busy time lately and I didn't have as much time to look
> into this as would be needed.
Oops, pressed send too early... So I have two debug patches for you. Can
you try whether the problem reproduces with the first one or with both of
them applied?
The stacktrace from spinlock lockup is somewhat weird - the spinlock lockup
message comes from NMI context where we tried to do print stack dump.
Indeed calling printk from NMI context *can* lead to this sort of lockups
(and the current process is a worker thread which apparently runs
fb_flashcursor() function which calls console_unlock() which can hold
logbuf_lock so all these things fit together). What I really fail to see
is how my patch makes the problem happen to you pretty reliably.
Another somewhat strange thing to me is that although lock->owner in your
dump looks fine and points to currently running process, lock->owner_cpu is
-1 which is a combination that shouldn't ever happen as I'm looking into
spinlock debug code.
So for now I'm still puzzled.
Honza
--
Jan Kara <[email protected]>
SUSE Labs, CR
On 06/11/2014 05:31 PM, Jan Kara wrote:
> On Wed 11-06-14 22:34:36, Jan Kara wrote:
>> > On Wed 11-06-14 10:55:55, Sasha Levin wrote:
>>> > > On 06/10/2014 11:59 AM, Peter Hurley wrote:
>>>> > > > On 06/06/2014 03:05 PM, Sasha Levin wrote:
>>>>> > > >> On 05/30/2014 10:07 AM, Jan Kara wrote:
>>>>>> > > >>> On Fri 30-05-14 09:58:14, Peter Hurley wrote:
>>>>>>>> > > >>>>> On 05/30/2014 09:11 AM, Sasha Levin wrote:
>>>>>>>>>> > > >>>>>>> Hi all,
>>>>>>>>>> > > >>>>>>>
>>>>>>>>>> > > >>>>>>> I sometime see lockups when booting my KVM guest with the latest -next kernel,
>>>>>>>>>> > > >>>>>>> it basically hangs right when it should start 'init', and after a while I get
>>>>>>>>>> > > >>>>>>> the following spew:
>>>>>>>>>> > > >>>>>>>
>>>>>>>>>> > > >>>>>>> [ 30.790833] BUG: spinlock lockup suspected on CPU#1, swapper/1/0
>>>>>>>> > > >>>>>
>>>>>>>> > > >>>>> Maybe related to this report: https://lkml.org/lkml/2014/5/30/26
>>>>>>>> > > >>>>> from Jet Chen which was bisected to
>>>>>>>> > > >>>>>
>>>>>>>> > > >>>>> commit bafe980f5afc7ccc693fd8c81c8aa5a02fbb5ae0
>>>>>>>> > > >>>>> Author: Jan Kara <[email protected]>
>>>>>>>> > > >>>>> AuthorDate: Thu May 22 10:43:35 2014 +1000
>>>>>>>> > > >>>>> Commit: Stephen Rothwell <[email protected]>
>>>>>>>> > > >>>>> CommitDate: Thu May 22 10:43:35 2014 +1000
>>>>>>>> > > >>>>>
>>>>>>>> > > >>>>> printk: enable interrupts before calling console_trylock_for_printk()
>>>>>>>> > > >>>>> We need interrupts disabled when calling console_trylock_for_printk() only
>>>>>>>> > > >>>>> so that cpu id we pass to can_use_console() remains valid (for other
>>>>>>>> > > >>>>> things console_sem provides all the exclusion we need and deadlocks on
>>>>>>>> > > >>>>> console_sem due to interrupts are impossible because we use
>>>>>>>> > > >>>>> down_trylock()). However if we are rescheduled, we are guaranteed to run
>>>>>>>> > > >>>>> on an online cpu so we can easily just get the cpu id in
>>>>>>>> > > >>>>> can_use_console().
>>>>>>>> > > >>>>> We can lose a bit of performance when we enable interrupts in
>>>>>>>> > > >>>>> vprintk_emit() and then disable them again in console_unlock() but OTOH it
>>>>>>>> > > >>>>> can somewhat reduce interrupt latency caused by console_unlock()
>>>>>>>> > > >>>>> especially since later in the patch series we will want to spin on
>>>>>>>> > > >>>>> console_sem in console_trylock_for_printk().
>>>>>>>> > > >>>>> Signed-off-by: Jan Kara <[email protected]>
>>>>>>>> > > >>>>> Signed-off-by: Andrew Morton <[email protected]>
>>>>>>>> > > >>>>>
>>>>>>>> > > >>>>> ?
>>>>>> > > >>> Yeah, very likely. I think I see the problem, I'll send the fix shortly.
>>>>> > > >>
>>>>> > > >> Hi Jan,
>>>>> > > >>
>>>>> > > >> It seems that the issue I'm seeing is different from the "[prink] BUG: spinlock
>>>>> > > >> lockup suspected on CPU#0, swapper/1".
>>>>> > > >>
>>>>> > > >> Is there anything else I could try here? The issue is very common during testing.
>>>> > > >
>>>> > > > Sasha,
>>>> > > >
>>>> > > > Is this bisectable? Maybe that's the best way forward here.
>>> > >
>>> > > I've ran a bisection again and ended up at the same commit as Jet Chen
>>> > > (the commit unfortunately already made it to Linus's tree).
>>> > >
>>> > > Note that I did try Jan's proposed fix and that didn't solve the issue
>>> > > for me, I believe we're seeing different issues caused by the same
>>> > > commit.
>> > Sorry it has been busy time lately and I didn't have as much time to look
>> > into this as would be needed.
> Oops, pressed send too early... So I have two debug patches for you. Can
> you try whether the problem reproduces with the first one or with both of
> them applied?
The first patch fixed it (I assumed that there's no need to try the second).
Thanks,
Sasha
On Wed 11-06-14 23:07:04, Sasha Levin wrote:
> On 06/11/2014 05:31 PM, Jan Kara wrote:
> > On Wed 11-06-14 22:34:36, Jan Kara wrote:
> >> > On Wed 11-06-14 10:55:55, Sasha Levin wrote:
> >>> > > On 06/10/2014 11:59 AM, Peter Hurley wrote:
> >>>> > > > On 06/06/2014 03:05 PM, Sasha Levin wrote:
> >>>>> > > >> On 05/30/2014 10:07 AM, Jan Kara wrote:
> >>>>>> > > >>> On Fri 30-05-14 09:58:14, Peter Hurley wrote:
> >>>>>>>> > > >>>>> On 05/30/2014 09:11 AM, Sasha Levin wrote:
> >>>>>>>>>> > > >>>>>>> Hi all,
> >>>>>>>>>> > > >>>>>>>
> >>>>>>>>>> > > >>>>>>> I sometime see lockups when booting my KVM guest with the latest -next kernel,
> >>>>>>>>>> > > >>>>>>> it basically hangs right when it should start 'init', and after a while I get
> >>>>>>>>>> > > >>>>>>> the following spew:
> >>>>>>>>>> > > >>>>>>>
> >>>>>>>>>> > > >>>>>>> [ 30.790833] BUG: spinlock lockup suspected on CPU#1, swapper/1/0
> >>>>>>>> > > >>>>>
> >>>>>>>> > > >>>>> Maybe related to this report: https://lkml.org/lkml/2014/5/30/26
> >>>>>>>> > > >>>>> from Jet Chen which was bisected to
> >>>>>>>> > > >>>>>
> >>>>>>>> > > >>>>> commit bafe980f5afc7ccc693fd8c81c8aa5a02fbb5ae0
> >>>>>>>> > > >>>>> Author: Jan Kara <[email protected]>
> >>>>>>>> > > >>>>> AuthorDate: Thu May 22 10:43:35 2014 +1000
> >>>>>>>> > > >>>>> Commit: Stephen Rothwell <[email protected]>
> >>>>>>>> > > >>>>> CommitDate: Thu May 22 10:43:35 2014 +1000
> >>>>>>>> > > >>>>>
> >>>>>>>> > > >>>>> printk: enable interrupts before calling console_trylock_for_printk()
> >>>>>>>> > > >>>>> We need interrupts disabled when calling console_trylock_for_printk() only
> >>>>>>>> > > >>>>> so that cpu id we pass to can_use_console() remains valid (for other
> >>>>>>>> > > >>>>> things console_sem provides all the exclusion we need and deadlocks on
> >>>>>>>> > > >>>>> console_sem due to interrupts are impossible because we use
> >>>>>>>> > > >>>>> down_trylock()). However if we are rescheduled, we are guaranteed to run
> >>>>>>>> > > >>>>> on an online cpu so we can easily just get the cpu id in
> >>>>>>>> > > >>>>> can_use_console().
> >>>>>>>> > > >>>>> We can lose a bit of performance when we enable interrupts in
> >>>>>>>> > > >>>>> vprintk_emit() and then disable them again in console_unlock() but OTOH it
> >>>>>>>> > > >>>>> can somewhat reduce interrupt latency caused by console_unlock()
> >>>>>>>> > > >>>>> especially since later in the patch series we will want to spin on
> >>>>>>>> > > >>>>> console_sem in console_trylock_for_printk().
> >>>>>>>> > > >>>>> Signed-off-by: Jan Kara <[email protected]>
> >>>>>>>> > > >>>>> Signed-off-by: Andrew Morton <[email protected]>
> >>>>>>>> > > >>>>>
> >>>>>>>> > > >>>>> ?
> >>>>>> > > >>> Yeah, very likely. I think I see the problem, I'll send the fix shortly.
> >>>>> > > >>
> >>>>> > > >> Hi Jan,
> >>>>> > > >>
> >>>>> > > >> It seems that the issue I'm seeing is different from the "[prink] BUG: spinlock
> >>>>> > > >> lockup suspected on CPU#0, swapper/1".
> >>>>> > > >>
> >>>>> > > >> Is there anything else I could try here? The issue is very common during testing.
> >>>> > > >
> >>>> > > > Sasha,
> >>>> > > >
> >>>> > > > Is this bisectable? Maybe that's the best way forward here.
> >>> > >
> >>> > > I've ran a bisection again and ended up at the same commit as Jet Chen
> >>> > > (the commit unfortunately already made it to Linus's tree).
> >>> > >
> >>> > > Note that I did try Jan's proposed fix and that didn't solve the issue
> >>> > > for me, I believe we're seeing different issues caused by the same
> >>> > > commit.
> >> > Sorry it has been busy time lately and I didn't have as much time to look
> >> > into this as would be needed.
> > Oops, pressed send too early... So I have two debug patches for you. Can
> > you try whether the problem reproduces with the first one or with both of
> > them applied?
>
> The first patch fixed it (I assumed that there's no need to try the second).
Good. So that shows that it is the increased lockdep coverage which is
causing problems - with my patch, lockdep is able to identify some problem
because console drivers are now called with lockdep enabled. But because
the problem was found in some difficult context, lockdep just hung the
machine when trying to report it... Sadly the stacktraces you posted don't
tell us what lockdep found.
Adding Peter Zijlstra to CC. Peter, any idea how lockdep could report
problems when holding logbuf_lock? One possibility would be to extend
logbuf_cpu recursion logic to every holder of logbuf_lock. That will at
least avoid the spinlock recursion killing the machine but we won't be able
to see what lockdep found...
Honza
--
Jan Kara <[email protected]>
SUSE Labs, CR
On Thu, 2014-06-12 at 10:26 +0200, Jan Kara wrote:
> On Wed 11-06-14 23:07:04, Sasha Levin wrote:
> > The first patch fixed it (I assumed that there's no need to try the second).
> Good. So that shows that it is the increased lockdep coverage which is
> causing problems - with my patch, lockdep is able to identify some problem
> because console drivers are now called with lockdep enabled. But because
> the problem was found in some difficult context, lockdep just hung the
> machine when trying to report it... Sadly the stacktraces you posted don't
> tell us what lockdep found.
>
> Adding Peter Zijlstra to CC. Peter, any idea how lockdep could report
> problems when holding logbuf_lock? One possibility would be to extend
> logbuf_cpu recursion logic to every holder of logbuf_lock. That will at
> least avoid the spinlock recursion killing the machine but we won't be able
> to see what lockdep found...
Could tell lockdep to use trace_printk().
On Wed 11-06-14 23:07:04, Sasha Levin wrote:
> On 06/11/2014 05:31 PM, Jan Kara wrote:
> > On Wed 11-06-14 22:34:36, Jan Kara wrote:
> >> > On Wed 11-06-14 10:55:55, Sasha Levin wrote:
> >>> > > On 06/10/2014 11:59 AM, Peter Hurley wrote:
> >>>> > > > On 06/06/2014 03:05 PM, Sasha Levin wrote:
> >>>>> > > >> On 05/30/2014 10:07 AM, Jan Kara wrote:
> >>>>>> > > >>> On Fri 30-05-14 09:58:14, Peter Hurley wrote:
> >>>>>>>> > > >>>>> On 05/30/2014 09:11 AM, Sasha Levin wrote:
> >>>>>>>>>> > > >>>>>>> Hi all,
> >>>>>>>>>> > > >>>>>>>
> >>>>>>>>>> > > >>>>>>> I sometime see lockups when booting my KVM guest with the latest -next kernel,
> >>>>>>>>>> > > >>>>>>> it basically hangs right when it should start 'init', and after a while I get
> >>>>>>>>>> > > >>>>>>> the following spew:
> >>>>>>>>>> > > >>>>>>>
> >>>>>>>>>> > > >>>>>>> [ 30.790833] BUG: spinlock lockup suspected on CPU#1, swapper/1/0
> >>>>>>>> > > >>>>>
> >>>>>>>> > > >>>>> Maybe related to this report: https://lkml.org/lkml/2014/5/30/26
> >>>>>>>> > > >>>>> from Jet Chen which was bisected to
> >>>>>>>> > > >>>>>
> >>>>>>>> > > >>>>> commit bafe980f5afc7ccc693fd8c81c8aa5a02fbb5ae0
> >>>>>>>> > > >>>>> Author: Jan Kara <[email protected]>
> >>>>>>>> > > >>>>> AuthorDate: Thu May 22 10:43:35 2014 +1000
> >>>>>>>> > > >>>>> Commit: Stephen Rothwell <[email protected]>
> >>>>>>>> > > >>>>> CommitDate: Thu May 22 10:43:35 2014 +1000
> >>>>>>>> > > >>>>>
> >>>>>>>> > > >>>>> printk: enable interrupts before calling console_trylock_for_printk()
> >>>>>>>> > > >>>>> We need interrupts disabled when calling console_trylock_for_printk() only
> >>>>>>>> > > >>>>> so that cpu id we pass to can_use_console() remains valid (for other
> >>>>>>>> > > >>>>> things console_sem provides all the exclusion we need and deadlocks on
> >>>>>>>> > > >>>>> console_sem due to interrupts are impossible because we use
> >>>>>>>> > > >>>>> down_trylock()). However if we are rescheduled, we are guaranteed to run
> >>>>>>>> > > >>>>> on an online cpu so we can easily just get the cpu id in
> >>>>>>>> > > >>>>> can_use_console().
> >>>>>>>> > > >>>>> We can lose a bit of performance when we enable interrupts in
> >>>>>>>> > > >>>>> vprintk_emit() and then disable them again in console_unlock() but OTOH it
> >>>>>>>> > > >>>>> can somewhat reduce interrupt latency caused by console_unlock()
> >>>>>>>> > > >>>>> especially since later in the patch series we will want to spin on
> >>>>>>>> > > >>>>> console_sem in console_trylock_for_printk().
> >>>>>>>> > > >>>>> Signed-off-by: Jan Kara <[email protected]>
> >>>>>>>> > > >>>>> Signed-off-by: Andrew Morton <[email protected]>
> >>>>>>>> > > >>>>>
> >>>>>>>> > > >>>>> ?
> >>>>>> > > >>> Yeah, very likely. I think I see the problem, I'll send the fix shortly.
> >>>>> > > >>
> >>>>> > > >> Hi Jan,
> >>>>> > > >>
> >>>>> > > >> It seems that the issue I'm seeing is different from the "[prink] BUG: spinlock
> >>>>> > > >> lockup suspected on CPU#0, swapper/1".
> >>>>> > > >>
> >>>>> > > >> Is there anything else I could try here? The issue is very common during testing.
> >>>> > > >
> >>>> > > > Sasha,
> >>>> > > >
> >>>> > > > Is this bisectable? Maybe that's the best way forward here.
> >>> > >
> >>> > > I've ran a bisection again and ended up at the same commit as Jet Chen
> >>> > > (the commit unfortunately already made it to Linus's tree).
> >>> > >
> >>> > > Note that I did try Jan's proposed fix and that didn't solve the issue
> >>> > > for me, I believe we're seeing different issues caused by the same
> >>> > > commit.
> >> > Sorry it has been busy time lately and I didn't have as much time to look
> >> > into this as would be needed.
> > Oops, pressed send too early... So I have two debug patches for you. Can
> > you try whether the problem reproduces with the first one or with both of
> > them applied?
>
> The first patch fixed it (I assumed that there's no need to try the second).
OK, so I got back to it. Can you try with attached patch and without my
"fix"? lockdep should dump complaints using trace buffer (so you should
check it from time to time) instead of using printk and thus we should be
able to see what it is complaining about. Thanks!
Honza
--
Jan Kara <[email protected]>
SUSE Labs, CR
On Thu, Jun 12, 2014 at 10:54:22AM +0200, Mike Galbraith wrote:
> On Thu, 2014-06-12 at 10:26 +0200, Jan Kara wrote:
> > On Wed 11-06-14 23:07:04, Sasha Levin wrote:
>
> > > The first patch fixed it (I assumed that there's no need to try the second).
> > Good. So that shows that it is the increased lockdep coverage which is
> > causing problems - with my patch, lockdep is able to identify some problem
> > because console drivers are now called with lockdep enabled. But because
> > the problem was found in some difficult context, lockdep just hung the
> > machine when trying to report it... Sadly the stacktraces you posted don't
> > tell us what lockdep found.
> >
> > Adding Peter Zijlstra to CC. Peter, any idea how lockdep could report
> > problems when holding logbuf_lock? One possibility would be to extend
> > logbuf_cpu recursion logic to every holder of logbuf_lock. That will at
> > least avoid the spinlock recursion killing the machine but we won't be able
> > to see what lockdep found...
>
> Could tell lockdep to use trace_printk().
lkml.kernel.org/r/[email protected]