Do we need to disable preemption in flush_tlb_range()?

Do we need to disable preemption in flush_tlb_range()?

[Alexey Brodkin]

Hi Vineet,

Just noticed that in comments for smp_call_function_many() it is said that
preemption must be disabled during its execution. And that function gets executed
among other ways like that:
-------------------------->8-----------------------
  flush_tlb_range()
    -> on_each_cpu_mask()
         -> smp_call_function_many()
-------------------------->8-----------------------

I'm not seeing right now any real problem with current implementation but
some architectures do that thus the question.

-Alexey

Re: Do we need to disable preemption in flush_tlb_range()?

[Alexey Brodkin]

Ping!

On Thu, 2018-03-01 at 18:13 +0300, Alexey Brodkin wrote:
> Hi Vineet,
> 
> Just noticed that in comments for smp_call_function_many() it is said that
> preemption must be disabled during its execution. And that function gets executed
> among other ways like that:
> -------------------------->8-----------------------
>   flush_tlb_range()
>     -> on_each_cpu_mask()
>          -> smp_call_function_many()
> -------------------------->8-----------------------
> 
> I'm not seeing right now any real problem with current implementation but
> some architectures do that thus the question.
> 
> -Alexey

Re: Do we need to disable preemption in flush_tlb_range()?

[Vineet Gupta]

+CC Peter since we have his attention ;-)

On 03/01/2018 07:13 AM, Alexey Brodkin wrote:
> Hi Vineet,
> 
> Just noticed that in comments for smp_call_function_many() it is said that
> preemption must be disabled during its execution. And that function gets executed
> among other ways like that:
> -------------------------->8-----------------------
>    flush_tlb_range()
>      -> on_each_cpu_mask()
>           -> smp_call_function_many()
> -------------------------->8-----------------------

In general I prefer not to - Peter what say you ?

> 
> I'm not seeing right now any real problem with current implementation but
> some architectures do that thus the question.
> 
> -Alexey
> 

Re: Do we need to disable preemption in flush_tlb_range()?

[Peter Zijlstra]

On Wed, Mar 14, 2018 at 01:19:01PM -0700, Vineet Gupta wrote:
> +CC Peter since we have his attention ;-)

Yeah, timezone collision there, I typically sleep at 1am ;-)

> On 03/01/2018 07:13 AM, Alexey Brodkin wrote:
> > Hi Vineet,
> > 
> > Just noticed that in comments for smp_call_function_many() it is said that
> > preemption must be disabled during its execution. And that function gets executed
> > among other ways like that:
> > -------------------------->8-----------------------
> >    flush_tlb_range()
> >      -> on_each_cpu_mask()
> >           -> smp_call_function_many()
> > -------------------------->8-----------------------
> 
> In general I prefer not to - Peter what say you ?

The comment with smp_call_function_many() is correct, it relies on
preemption being disabled in a number of ways. I would expect
this_cpu_ptr() for example to complain when used with preemption
enabled (CONFIG_DEBUG_PREEMPT).

Re: Do we need to disable preemption in flush_tlb_range()?

[Alexey Brodkin]

Hi Peter,

On Thu, 2018-03-15 at 09:27 +0100, Peter Zijlstra wrote:
> On Wed, Mar 14, 2018 at 01:19:01PM -0700, Vineet Gupta wrote:
> > +CC Peter since we have his attention ;-)
> 
> Yeah, timezone collision there, I typically sleep at 1am ;-)
> 
> > On 03/01/2018 07:13 AM, Alexey Brodkin wrote:
> > > Hi Vineet,
> > > 
> > > Just noticed that in comments for smp_call_function_many() it is said that
> > > preemption must be disabled during its execution. And that function gets executed
> > > among other ways like that:
> > > -------------------------->8-----------------------
> > >    flush_tlb_range()
> > >      -> on_each_cpu_mask()
> > >           -> smp_call_function_many()
> > > -------------------------->8-----------------------
> > 
> > In general I prefer not to - Peter what say you ?
> 
> The comment with smp_call_function_many() is correct, it relies on
> preemption being disabled in a number of ways. I would expect
> this_cpu_ptr() for example to complain when used with preemption
> enabled (CONFIG_DEBUG_PREEMPT).

I just tried CONFIG_DEBUG_PREEMPT and the only thing I got was that:
-------------------------->8-----------------------
ARC perf        : 8 counters (32 bits), 32 conditions, [overflow IRQ support]
BUG: using smp_processor_id() in preemptible [00000000] code: swapper/0/1
caller is arc_pmu_device_probe+0x24e/0x29c
CPU: 0 PID: 1 Comm: swapper/0 Not tainted 4.14.14+ #67

Stack Trace:
  arc_unwind_core.constprop.1+0xd0/0xf4
  dump_stack+0x64/0x7c
  debug_smp_processor_id+0xb8/0xbc
  arc_pmu_device_probe+0x24e/0x29c
  platform_drv_probe+0x26/0x5c
  really_probe+0x288/0x338
  __driver_attach+0xc4/0xc8
  bus_for_each_dev+0x38/0x70
  bus_add_driver+0x12a/0x18c
  driver_register+0x50/0xec
  do_one_initcall+0x32/0x108
  kernel_init_freeable+0xfe/0x188
-------------------------->8-----------------------

That happens because in PMU probe routine we want to
configure IRQ handlers on all other cores:
-------------------------->8-----------------------
  arc_pmu_device_probe() ->
    on_each_cpu(arc_cpu_pmu_irq_init, &irq, 1): preempt_disable() ->
      enable_percpu_irq(irq, IRQ_TYPE_NONE) ->
        smp_processor_id() with disabled preemption.
-------------------------->8-----------------------

Which poses another preemption related question - how do IRQ setup on
all cores properly? :)

-Alexey

Re: Do we need to disable preemption in flush_tlb_range()?

[Vineet Gupta]

+CC some more folks for intc/irq insights - please see question at the bottom !

On 03/15/2018 02:39 AM, Alexey Brodkin wrote:
> Hi Peter,
> 
> On Thu, 2018-03-15 at 09:27 +0100, Peter Zijlstra wrote:
>> On Wed, Mar 14, 2018 at 01:19:01PM -0700, Vineet Gupta wrote:
>>> +CC Peter since we have his attention ;-)
>>
>> Yeah, timezone collision there, I typically sleep at 1am ;-)
>>
>>> On 03/01/2018 07:13 AM, Alexey Brodkin wrote:
>>>> Hi Vineet,
>>>>
>>>> Just noticed that in comments for smp_call_function_many() it is said that
>>>> preemption must be disabled during its execution. And that function gets executed
>>>> among other ways like that:
>>>> -------------------------->8-----------------------
>>>>     flush_tlb_range()
>>>>       -> on_each_cpu_mask()
>>>>            -> smp_call_function_many()
>>>> -------------------------->8-----------------------
>>>
>>> In general I prefer not to - Peter what say you ?
>>
>> The comment with smp_call_function_many() is correct, it relies on
>> preemption being disabled in a number of ways. I would expect
>> this_cpu_ptr() for example to complain when used with preemption
>> enabled (CONFIG_DEBUG_PREEMPT).
> 
> I just tried CONFIG_DEBUG_PREEMPT and the only thing I got was that:
> -------------------------->8-----------------------
> ARC perf        : 8 counters (32 bits), 32 conditions, [overflow IRQ support]
> BUG: using smp_processor_id() in preemptible [00000000] code: swapper/0/1
> caller is arc_pmu_device_probe+0x24e/0x29c
> CPU: 0 PID: 1 Comm: swapper/0 Not tainted 4.14.14+ #67
> 
> Stack Trace:
>    arc_unwind_core.constprop.1+0xd0/0xf4
>    dump_stack+0x64/0x7c
>    debug_smp_processor_id+0xb8/0xbc
>    arc_pmu_device_probe+0x24e/0x29c
>    platform_drv_probe+0x26/0x5c
>    really_probe+0x288/0x338
>    __driver_attach+0xc4/0xc8
>    bus_for_each_dev+0x38/0x70
>    bus_add_driver+0x12a/0x18c
>    driver_register+0x50/0xec
>    do_one_initcall+0x32/0x108
>    kernel_init_freeable+0xfe/0x188
> -------------------------->8-----------------------
> 
> That happens because in PMU probe routine we want to
> configure IRQ handlers on all other cores:
> -------------------------->8-----------------------
>    arc_pmu_device_probe() ->
>      on_each_cpu(arc_cpu_pmu_irq_init, &irq, 1): preempt_disable() ->
>        enable_percpu_irq(irq, IRQ_TYPE_NONE) ->
>          smp_processor_id() with disabled preemption.
> -------------------------->8-----------------------
> 
> Which poses another preemption related question - how do IRQ setup on
> all cores properly? :)
> 
> -Alexey
> 

Re: Do we need to disable preemption in flush_tlb_range()?

[Peter Zijlstra]

On Thu, Mar 15, 2018 at 09:39:31AM +0000, Alexey Brodkin wrote:
> Hi Peter,
> 
> On Thu, 2018-03-15 at 09:27 +0100, Peter Zijlstra wrote:
> > On Wed, Mar 14, 2018 at 01:19:01PM -0700, Vineet Gupta wrote:
> > > +CC Peter since we have his attention ;-)
> > 
> > Yeah, timezone collision there, I typically sleep at 1am ;-)
> > 
> > > On 03/01/2018 07:13 AM, Alexey Brodkin wrote:
> > > > Hi Vineet,
> > > > 
> > > > Just noticed that in comments for smp_call_function_many() it is said that
> > > > preemption must be disabled during its execution. And that function gets executed
> > > > among other ways like that:
> > > > -------------------------->8-----------------------
> > > >    flush_tlb_range()
> > > >      -> on_each_cpu_mask()
> > > >           -> smp_call_function_many()
> > > > -------------------------->8-----------------------
> > > 
> > > In general I prefer not to - Peter what say you ?
> > 
> > The comment with smp_call_function_many() is correct, it relies on
> > preemption being disabled in a number of ways. I would expect
> > this_cpu_ptr() for example to complain when used with preemption
> > enabled (CONFIG_DEBUG_PREEMPT).

So on_each_cpu_mask() already disables preemption around calling
smp_call_function_many().

Re: Do we need to disable preemption in flush_tlb_range()?

[Alexey Brodkin]

Hi Peter,

On Fri, 2018-03-16 at 11:11 +0100, Peter Zijlstra wrote:
> On Thu, Mar 15, 2018 at 09:39:31AM +0000, Alexey Brodkin wrote:
> > Hi Peter,
> > 
> > On Thu, 2018-03-15 at 09:27 +0100, Peter Zijlstra wrote:
> > > On Wed, Mar 14, 2018 at 01:19:01PM -0700, Vineet Gupta wrote:
> > > > +CC Peter since we have his attention ;-)
> > > 
> > > Yeah, timezone collision there, I typically sleep at 1am ;-)
> > > 
> > > > On 03/01/2018 07:13 AM, Alexey Brodkin wrote:
> > > > > Hi Vineet,
> > > > > 
> > > > > Just noticed that in comments for smp_call_function_many() it is said that
> > > > > preemption must be disabled during its execution. And that function gets executed
> > > > > among other ways like that:
> > > > > -------------------------->8-----------------------
> > > > >    flush_tlb_range()
> > > > >      -> on_each_cpu_mask()
> > > > >           -> smp_call_function_many()
> > > > > -------------------------->8-----------------------
> > > > 
> > > > In general I prefer not to - Peter what say you ?
> > > 
> > > The comment with smp_call_function_many() is correct, it relies on
> > > preemption being disabled in a number of ways. I would expect
> > > this_cpu_ptr() for example to complain when used with preemption
> > > enabled (CONFIG_DEBUG_PREEMPT).
> 
> So on_each_cpu_mask() already disables preemption around calling
> smp_call_function_many().

Right that happens in get_cpu() so then we're golden here.
Thanks for pointing out - was not clear immediately from the code :)

-Alexey