Interesting csd deadlock on ARC

DISCUSSION Vineet Gupta 5 participants 12 messages Feb 19, 2016 – Feb 25, 2016

Subsystems (2)

Generic include/asm header files Synopsys ARC architecture

lore

Share & tools

lore

Vineet Gupta

author

2016-02-19 06:47 UTC #0 lore raw

Hi Peter,

I've been debugging a csd_lock_wait() deadlock on SMP+PREEMPT ARC HS38x2 and it turned out to be lot more interesting than I'd hoped for. This is stock v4.4

Trouble starts with an IPI to self which doesn't get delivered as the inter-core interrupt providing h/w is not capable of IPI to self (which I found as part of debugging this). Subsequent IPIs from other cores to this core get elided as well due to the IPI coalescing optimization in arch/arc/kernel/smp.c: ipi_send_msg_one()

There are ways to use a different h/w mechanism to solve the trigger issue and I'd hoped to just implement arch_irq_work_raise(). But the trouble is the call stack for this issue: IPI to self is triggered from

sys_sched_setscheduler
__balance_callback

       pull_rt_task
         irq_work_queue_on  <-- called with @cpu == self

Looking into irq_work.c, irq_work_queue() is what is semantically needed, specifically arch_irq_work_raise() will not be called, which means I need arch_send_call_function_single_ipi() to be able to IPI to self cpu also. Is that expected from arch code....

Just wanted to understand before writing patches...

Test case triggering is harmless looking LTP: trace_sched -c 1
It is kind of scheduler fizzer as it triggers a whole bunch of sched activity.

Thx,
-Vineet

Vineet Gupta

author

2016-02-23 05:21 UTC #1 lore raw

Quoted 16 lines · Vineet Gupta

Hi Peter,
I've been debugging a csd_lock_wait() deadlock on SMP+PREEMPT ARC HS38x2 and it turned out to be lot more interesting than I'd hoped for. This is stock v4.4
Trouble starts with an IPI to self which doesn't get delivered as the inter-core interrupt providing h/w is not capable of IPI to self (which I found as part of debugging this). Subsequent IPIs from other cores to this core get elided as well due to the IPI coalescing optimization in arch/arc/kernel/smp.c: ipi_send_msg_one()
There are ways to use a different h/w mechanism to solve the trigger issue and I'd hoped to just implement arch_irq_work_raise(). But the trouble is the call stack for this issue: IPI to self is triggered from
sys_sched_setscheduler
__balance_callback

Quoted 7 lines

       pull_rt_task
         irq_work_queue_on  <-- called with @cpu == self
Looking into irq_work.c, irq_work_queue() is what is semantically needed, specifically arch_irq_work_raise() will not be called, which means I need arch_send_call_function_single_ipi() to be able to IPI to self cpu also. Is that expected from arch code....

What I actually meant was is it OK for irq_work_queue_on() to be called locally (is this a sched bug/optimization(. Further if it is OK to be called, does it need to do behave more like irq_work_queue() i.e. call arch_irq_work_raise() or arch_send_call_function_single_ipi() is expected to handle sending IPI to self !

Quoted 9 lines

Just wanted to understand before writing patches...
Test case triggering is harmless looking LTP: trace_sched -c 1
It is kind of scheduler fizzer as it triggers a whole bunch of sched activity.
Thx,
-Vineet

Peter Zijlstra

2016-02-23 09:58 UTC #2 lore raw

Quoted 6 lines · Vineet Gupta

Quoted 5 lines · Vineet Gupta
Hi Peter,
I've been debugging a csd_lock_wait() deadlock on SMP+PREEMPT ARC HS38x2 and it turned out to be lot more interesting than I'd hoped for. This is stock v4.4

Quoted 7 lines

Quoted 7 lines
Trouble starts with an IPI to self which doesn't get delivered as the inter-core interrupt providing h/w is not capable of IPI to self (which I found as part of debugging this). Subsequent IPIs from other cores to this core get elided as well due to the IPI coalescing optimization in arch/arc/kernel/smp.c: ipi_send_msg_one()
There are ways to use a different h/w mechanism to solve the trigger issue and I'd hoped to just implement arch_irq_work_raise().

Yes, there are other architectures that use other means for self-IPI, IIRC PowerPC has to program their timer in the past to generate a local interrupt.

Quoted 13 lines

Quoted 5 lines
But the trouble is the call stack
for this issue: IPI to self is triggered from
sys_sched_setscheduler
__balance_callback
Quoted 7 lines
       pull_rt_task
         irq_work_queue_on  <-- called with @cpu == self
Looking into irq_work.c, irq_work_queue() is what is semantically needed, specifically arch_irq_work_raise() will not be called, which means I need arch_send_call_function_single_ipi() to be able to IPI to self cpu also. Is that expected from arch code....

Quoted 4 lines

What I actually meant was is it OK for irq_work_queue_on() to be called locally (is this a sched bug/optimization(. Further if it is OK to be called, does it need to do behave more like irq_work_queue() i.e. call arch_irq_work_raise() or arch_send_call_function_single_ipi() is expected to handle sending IPI to self !

Right, so I'm not actually sure we started out with this requirement. But you're not the first to run into this, see:

lkml.kernel.org/r/CAJZ5v0gLankSuziQq25qTCyNqeOX43yD9jnJu_XXwbdyajfmKg@mail.gmail.com

Initially I think irq_work_queue_on() was only used remotely, but I think it makes sense to allow the current cpu, esp. since people seem to be using it like that.

Now the distinct difference between arch_irq_work_raise() and arch_send_call_function_single_ipi() is that arch_irq_work_raise() should be NMI-safe.

So on x86 it has to be extra careful about the lapic state, whereas the regular IPI code doesn't.

I seem to have forgotten the status of NMIs on ARC, but this is something to make a note of.

Vineet Gupta

author

2016-02-23 10:21 UTC #3 lore raw

Quoted 9 lines

Quoted 4 lines
What I actually meant was is it OK for irq_work_queue_on() to be called locally (is this a sched bug/optimization(. Further if it is OK to be called, does it need to do behave more like irq_work_queue() i.e. call arch_irq_work_raise() or arch_send_call_function_single_ipi() is expected to handle sending IPI to self !
Right, so I'm not actually sure we started out with this requirement. But you're not the first to run into this, see:
lkml.kernel.org/r/CAJZ5v0gLankSuziQq25qTCyNqeOX43yD9jnJu_XXwbdyajfmKg@mail.gmail.com

Thx for the link, very helpful. I've posted fix for ARC which uses software interrupt and is thus UP/SMP safe.

Quoted 7 lines

Initially I think irq_work_queue_on() was only used remotely, but I think it makes sense to allow the current cpu, esp. since people seem to be using it like that.
Now the distinct difference between arch_irq_work_raise() and arch_send_call_function_single_ipi() is that arch_irq_work_raise() should be NMI-safe.

Ok - so when I implement interrupt priorities (aka NMI for ARC), this needs to be highest.

Quoted 6 lines

So on x86 it has to be extra careful about the lapic state, whereas the regular IPI code doesn't.
I seem to have forgotten the status of NMIs on ARC, but this is something to make a note of.

Not had a chance to go back to it since we last discussed. I've just been swamped with bug fixing like this one :-(

Thx,
-Vineet

Peter Zijlstra

2016-02-23 10:39 UTC #4 lore raw

Quoted 7 lines · Vineet Gupta

Quoted 4 lines
Now the distinct difference between arch_irq_work_raise() and arch_send_call_function_single_ipi() is that arch_irq_work_raise() should be NMI-safe.
Ok - so when I implement interrupt priorities (aka NMI for ARC), this needs to be highest.

So on x86 the issue is that the NMI can interrupt someone else writing to the lapic. So there's a bit of extra care to be taken.

If your platform doesn't suffer such issues, then that should be fine.

The only requirement for irq_work is that it runs after the NMI completes and runs from regular IRQ context. There are no strict interrupt priority requirements, only that it happens.

Quoted 5 lines

Quoted 2 lines
I seem to have forgotten the status of NMIs on ARC, but this is something to make a note of.
Not had a chance to go back to it since we last discussed. I've just been swamped with bug fixing like this one :-(

Yeah, I'm familiar with the problem ...

Noam Camus

2016-02-23 10:58 UTC #5 lore raw

Quoted 4 lines

From: Peter Zijlstra [mailto:[email protected]]
Sent: Tuesday, February 23, 2016 12:40 PM
The only requirement for irq_work is that it runs after the NMI completes and runs from regular IRQ context. >There are no strict interrupt priority requirements, only that it happens.

We here already use self IPI and irq_work on ARC (with no NMI concerns). Please see patch at:
https://github.com/EZchip/linux/commit/e42556738e610295f917c79dae166373cd11de88

Noam

Vineet Gupta

author

2016-02-24 04:45 UTC #6 lore raw

Quoted 8 lines · Noam Camus

Quoted 4 lines
From: Peter Zijlstra [mailto:[email protected]]
Sent: Tuesday, February 23, 2016 12:40 PM
The only requirement for irq_work is that it runs after the NMI completes and runs from regular IRQ context. >There are no strict interrupt priority requirements, only that it happens.
We here already use self IPI and irq_work on ARC (with no NMI concerns). Please see patch at:
https://github.com/EZchip/linux/commit/e42556738e610295f917c79dae166373cd11de88

So this relies on 2nd level intc providing the IPI.
For UP we still need the software triggered interrupt provided by core intc.

Vineet Gupta

author

2016-02-24 04:51 UTC #7 lore raw

In reply to Peter Zijlstra

Quoted 37 lines · Peter Zijlstra

Quoted 6 lines · Vineet Gupta
Quoted 5 lines · Vineet Gupta
Hi Peter,
I've been debugging a csd_lock_wait() deadlock on SMP+PREEMPT ARC HS38x2 and it turned out to be lot more interesting than I'd hoped for. This is stock v4.4
Quoted 7 lines
Quoted 7 lines
Trouble starts with an IPI to self which doesn't get delivered as the inter-core interrupt providing h/w is not capable of IPI to self (which I found as part of debugging this). Subsequent IPIs from other cores to this core get elided as well due to the IPI coalescing optimization in arch/arc/kernel/smp.c: ipi_send_msg_one()
There are ways to use a different h/w mechanism to solve the trigger issue and I'd hoped to just implement arch_irq_work_raise().
Yes, there are other architectures that use other means for self-IPI, IIRC PowerPC has to program their timer in the past to generate a local interrupt.
Quoted 13 lines
Quoted 5 lines
But the trouble is the call stack
for this issue: IPI to self is triggered from
sys_sched_setscheduler
__balance_callback
Quoted 7 lines
       pull_rt_task
         irq_work_queue_on  <-- called with @cpu == self
Looking into irq_work.c, irq_work_queue() is what is semantically needed, specifically arch_irq_work_raise() will not be called, which means I need arch_send_call_function_single_ipi() to be able to IPI to self cpu also. Is that expected from arch code....
Quoted 4 lines
What I actually meant was is it OK for irq_work_queue_on() to be called locally (is this a sched bug/optimization(. Further if it is OK to be called, does it need to do behave more like irq_work_queue() i.e. call arch_irq_work_raise() or arch_send_call_function_single_ipi() is expected to handle sending IPI to self !

Quoted 8 lines

Right, so I'm not actually sure we started out with this requirement. But you're not the first to run into this, see:
lkml.kernel.org/r/CAJZ5v0gLankSuziQq25qTCyNqeOX43yD9jnJu_XXwbdyajfmKg@mail.gmail.com
Initially I think irq_work_queue_on() was only used remotely, but I think it makes sense to allow the current cpu, esp. since people seem to be using it like that.

So it seems Russell's questions in the thread above stands still. IMO we need to massage irq_work_queue_on() to handle the case of called for local cpu. This will automatically take care of CONFIG_SMP kernel running on UP hardware.

Quoted 11 lines

Now the distinct difference between arch_irq_work_raise() and arch_send_call_function_single_ipi() is that arch_irq_work_raise() should be NMI-safe.
So on x86 it has to be extra careful about the lapic state, whereas the regular IPI code doesn't.
I seem to have forgotten the status of NMIs on ARC, but this is something to make a note of.

Peter Zijlstra

2016-02-25 14:06 UTC #8 lore raw

Quoted 10 lines · Vineet Gupta

Quoted 10 lines
Quoted 4 lines
What I actually meant was is it OK for irq_work_queue_on() to be called locally (is this a sched bug/optimization(. Further if it is OK to be called, does it need to do behave more like irq_work_queue() i.e. call arch_irq_work_raise() or arch_send_call_function_single_ipi() is expected to handle sending IPI to self !
Right, so I'm not actually sure we started out with this requirement. But you're not the first to run into this, see:
lkml.kernel.org/r/CAJZ5v0gLankSuziQq25qTCyNqeOX43yD9jnJu_XXwbdyajfmKg@mail.gmail.com

Quoted 7 lines

Quoted 3 lines
Initially I think irq_work_queue_on() was only used remotely, but I think it makes sense to allow the current cpu, esp. since people seem to be using it like that.
So it seems Russell's questions in the thread above stands still. IMO we need to massage irq_work_queue_on() to handle the case of called for local cpu. This will automatically take care of CONFIG_SMP kernel running on UP hardware.

Hmm, I missed that there was still an open question.

Afaict the only thing that needs doing to the generic code is drop the CONFIG_SMP guard, no?

Vineet Gupta

author

2016-02-25 14:23 UTC #9 lore raw

Quoted 15 lines · Peter Zijlstra

Quoted 14 lines · Vineet Gupta
Quoted 13 lines
Quoted 4 lines
What I actually meant was is it OK for irq_work_queue_on() to be called locally (is this a sched bug/optimization(. Further if it is OK to be called, does it need to do behave more like irq_work_queue() i.e. call arch_irq_work_raise() or arch_send_call_function_single_ipi() is expected to handle sending IPI to self !
Right, so I'm not actually sure we started out with this requirement. But you're not the first to run into this, see:
lkml.kernel.org/r/CAJZ5v0gLankSuziQq25qTCyNqeOX43yD9jnJu_XXwbdyajfmKg at mail.gmail.com
Initially I think irq_work_queue_on() was only used remotely, but I think it makes sense to allow the current cpu, esp. since people seem to be using it like that.

Quoted 8 lines

Quoted 3 lines
So it seems Russell's questions in the thread above stands still. IMO we need to massage irq_work_queue_on() to handle the case of called for local cpu. This will automatically take care of CONFIG_SMP kernel running on UP hardware.
Hmm, I missed that there was still an open question.
Afaict the only thing that needs doing to the generic code is drop the CONFIG_SMP guard, no?

But then ARM CONFIG_SMP on UP hardware will still crap out because there is no way to send IPI to self. Same as the bug in above discussion. I'm surprised they way ARM guys worked around it.

But yeah what you propose needs to be done and additionally make irq_work_queue_on() behave like irq_work_queue() if @cpu == smp_processor_id(). So the IPI will not be involved at all. But that means arch has to implement arch_irq_work_raise() or maybe not - per comment there "Lame architectures will get the timer tick callback" !

This would be a cleaner solution IMHO !

Russell King - ARM Linux

2016-02-25 14:30 UTC #10 lore raw

Quoted 3 lines · Vineet Gupta

But then ARM CONFIG_SMP on UP hardware will still crap out because there is no way to send IPI to self. Same as the bug in above discussion. I'm surprised they way ARM guys worked around it.

We haven't worked around it - the code which provoked the oops that was seen (in the cpufreq code) was changed not to call it, which has the effect of making the problem "go away", at least for now.

We still have the problem that if it does get called on UP, it'll blow up - and I don't see any point in complicating the code for something that never happens right now.

Signature · 3 lines

RMK's Patch system: http://www.arm.linux.org.uk/developer/patches/
FTTC broadband for 0.8mile line: currently at 9.6Mbps down 400kbps up
according to speedtest.net.

Vineet Gupta

2016-02-25 15:58 UTC #11 lore raw

Quoted 5 lines · Russell King - ARM Linux

Quoted 3 lines · Vineet Gupta
But then ARM CONFIG_SMP on UP hardware will still crap out because there is no way to send IPI to self. Same as the bug in above discussion. I'm surprised they way ARM guys worked around it.

Quoted 7 lines

We haven't worked around it - the code which provoked the oops that was seen (in the cpufreq code) was changed not to call it, which has the effect of making the problem "go away", at least for now.
We still have the problem that if it does get called on UP, it'll blow up - and I don't see any point in complicating the code for something that never happens right now.

Right so my statement "workaround" was technically incorrect. But like you say, it's a ticking bomb which will certainly go off on your SMP on UP systems the moment someone adds irq_work_queue_on() in some obscure corner of generic code.

And I think this merits fixing in generic code !

-Vineet