Re: [RFC v4 3/4] irqflags: Avoid unnecessary calls to trace_ if you can

[Mathieu Desnoyers <mathieu.desnoyers@efficios.com>]

----- On Apr 25, 2018, at 7:13 PM, Joel Fernandes joelaf@google.com wrote:

> Hi Mathieu,
> 
> On Wed, Apr 25, 2018 at 2:40 PM, Mathieu Desnoyers
> <mathieu.desnoyers@efficios.com> wrote:
>> ----- On Apr 25, 2018, at 5:27 PM, Joel Fernandes joelaf@google.com wrote:
>>
>>> On Tue, Apr 24, 2018 at 9:20 PM, Paul E. McKenney
>>> <paulmck@linux.vnet.ibm.com> wrote:
>>> [..]
>>>>> >
>>>>> > Sounds good, thanks.
>>>>> >
>>>>> > Also I found the reason for my boot issue. It was because the
>>>>> > init_srcu_struct in the prototype was being done in an initcall.
>>>>> > Instead if I do it in start_kernel before the tracepoint is used, it
>>>>> > fixes it (although I don't know if this is dangerous to do like this
>>>>> > but I can get it to boot atleast.. Let me know if this isn't the
>>>>> > right way to do it, or if something else could go wrong)
>>>>> >
>>>>> > diff --git a/init/main.c b/init/main.c
>>>>> > index 34823072ef9e..ecc88319c6da 100644
>>>>> > --- a/init/main.c
>>>>> > +++ b/init/main.c
>>>>> > @@ -631,6 +631,7 @@ asmlinkage __visible void __init start_kernel(void)
>>>>> >         WARN(!irqs_disabled(), "Interrupts were enabled early\n");
>>>>> >         early_boot_irqs_disabled = false;
>>>>> >
>>>>> > +       init_srcu_struct(&tracepoint_srcu);
>>>>> >         lockdep_init_early();
>>>>> >
>>>>> >         local_irq_enable();
>>>>> > --
>>>>> >
>>>>> > I benchmarked it and the performance also looks quite good compared
>>>>> > to the rcu tracepoint version.
>>>>> >
>>>>> > If you, Paul and other think doing the init_srcu_struct like this
>>>>> > should be Ok, then I can try to work more on your srcu prototype and
>>>>> > roll into my series and post them in the next RFC series (or let me
>>>>> > know if you wanted to work your srcu stuff in a separate series..).
>>>>>
>>>>> That is definitely not what I was expecting, but let's see if it works
>>>>> anyway...  ;-)
>>>>>
>>>>> But first, I was instead expecting something like this:
>>>>>
>>>>> DEFINE_SRCU(tracepoint_srcu);
>>>>>
>>>>> With this approach, some of the initialization happens at compile time
>>>>> and the rest happens at the first call_srcu().
>>>>>
>>>>> This will work -only- if the first call_srcu() doesn't happen until after
>>>>> workqueue_init_early() has been invoked.  Which I believe must have been
>>>>> the case in your testing, because otherwise it looks like __call_srcu()
>>>>> would have complained bitterly.
>>>>>
>>>>> On the other hand, if you need to invoke call_srcu() before the call
>>>>> to workqueue_init_early(), then you need the patch that I am beating
>>>>> into shape.  Plus you would need to use DEFINE_SRCU() and to avoid
>>>>> invoking init_srcu_struct().
>>>>
>>>> And here is the patch.  I do not intend to send it upstream unless it
>>>> actually proves necessary, and it appears that current SRCU does what
>>>> you need.
>>>>
>>>> You would only need this patch if you wanted to invoke call_srcu()
>>>> before workqueue_init_early() was called, which does not seem likely.
>>>
>>> Cool. So I was chatting with Paul and just to update everyone as well,
>>> I tried the DEFINE_SRCU instead of the late init_srcu_struct call and
>>> can make it past boot too (thanks Paul!). Also I don't see a reason we
>>> need the RCU callback to execute early and its fine if it runs later.
>>>
>>> Also, I was thinking of introducing a separate trace_*event*_srcu API
>>> as a replacement to the _rcuidle API. Then I can make use of it for my
>>> tracepoints, and then later can use it for the other tracepoints
>>> needing _rcuidle. After that we can finally get rid of the _rcuidle
>>> API if there are no other users of it. This is just a rough plan, but
>>> let me know if there's any issue with this plan that you can think
>>> off.
>>> IMO, I believe its simpler if the caller worries about whether it can
>>> tolerate if tracepoint probes can block or not, than making it a
>>> property of the tracepoint. That would also simplify the patch to
>>> introduce srcu and keep the tracepoint creation API simple and less
>>> confusing, but let me know if I'm missing something about this.
>>
>> One problem with your approach is that you can have multiple callers
>> for the same tracepoint name, where some could be non-preemptible and
>> others blocking. Also, there is then no clear way for the callback
> 
> Shouldn't it be responsibility of the caller to make sure it calls
> correct API? So if you're wanting to allow probes to block, then you'd
> call trace*blocking, if not then you don't. So the caller side can
> just always do the right thing. That's a caller side issue.

The issue there is that tracepoint.c has APIs both for instrumentation
and for registration of probe providers (callbacks). I want tracepoint.c
to provide guarantees that it won't connect incompatible probes and
callsites together.

> 
>>
>> Regarding the name, I'm OK with having something along the lines of
>> trace_*event*_blocking or such. Please don't use "srcu" or other naming
>> that is explicitly tied to the underlying mechanism used internally
>> however: what we want to convey is that this specific tracepoint probe
> 
> Problem is that _blocking isn't the right word either. In my IRQ trace
> point case, it will look something like this then:
> 
> local_irq_disable();
> // IRQs are now off.
> trace_irq_disable_blocking(..);
> 
> This wouldn't make sense. What we really want is to use the SRCU
> implementation so that its low overhead...
> 
> So it would be something like:
> 
> local_irq_disable();
> // IRQs are now off.
> trace_irq_disable_srcu(..);
> 
> I also Ok if, as Paul was saying in his last email, that just for
> _rcuidle, we use SRCU so that we don't have to do the rcu_enter_irq
> stuff. Or we kill the _rcuidle API completely and use _srcu for those
> users instead. We already have 1 implementation specific name anyway
> (rcuidle), we're just replacing it with another one. If in the future,
> if we want to change that name we can always do so (Also if you will,
> correcting the existing already bad naming is a different problem and
> we're not making it any worse tbh).

Using SRCU rather than the sched-rcu tracepoint synchronization in your
use-case it caused by a limitation of sched-rcu: it cannot be efficiently
used within idle code. So you don't care about the "can_sleep" property
of SRCU. You could event mix SRCU and sched-rcu callsites for the same
probe name, and it would be perfectly valid.

So even though both "can_sleep" and "rcuidle" caller variants would end
up using SRCU under the hood, each can have its own caller API, e.g.:

* trace_<event>() -> only non-sleeping probes can register to those.
                     Uses sched-rcu under the hood.

* trace_<event>_can_sleep() -> both sleeping and non-sleeping probes can
                               register to those. Uses SRCU under the hood.

* trace_<event>_rcuidle() -> only non-sleeping probes can register to those,
                             uses SRCU under the hood.

> 
>> can be preempted and block. The underlying implementation could move to
>> a different RCU flavor brand in the future, and it should not impact
>> users of the tracepoint APIs.
>>
>> In order to ensure that probes that may block only register themselves
>> to tracepoints that allow blocking, we should introduce new tracepoint
>> declaration/definition *and* registration APIs also contain the
>> "BLOCKING/blocking" keywords (or such), so we can ensure that a
>> tracepoint probe being registered to a "blocking" tracepoint is indeed
>> allowed to block.
> 
> I feel this problem you're describing is slightly out of the scope of
> the issues we're talking about, I think. Even right now, someone can
> write a callback that blocks and then bad things will happen. If I
> understand correctly, all callbacks right now will execute in a
> preempt disabled section because of rcu_read_lock_sched. So we already
> have a problem (without the SRCU changes) that if a callback blocks,
> then we'll have hard to diagnose sleeping while atomic issues. Sorry
> if I missed your point.

The current situation is that no callback whatsoever can sleep. If we
introduce an API allowing some callbacks to sleep, I want to make sure
we don't end up registering sleepable callbacks to non-preemptible callsites.
Considering that the callback can be provided by a kernel module whereas the
callsite is within the kernel, having this kind of correctness validation
within tracepoint.c appears important.

Thanks!

Mathieu


> 
> thanks,
> 
> - Joel

-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com