Re: [REGRESSION] perf/core: PMU interrupts dropped if we entered the kernel in the "skid" region

From: Kyle Huey
Date: Wed Jun 28 2017 - 03:30:37 EST


On Tue, Jun 27, 2017 at 10:35 PM, Jin, Yao <yao.jin@xxxxxxxxxxxxxxx> wrote:
> Hi Kyle,
>
> I understand your requirement. Sorry I don't know the detail of rr debugger,
> but I guess if it just uses counter overflow to deliver a signal, could it
> set the counter without "exclude_kernel"?

Unfortunately we cannot. We depend on the counter value being exactly
the same between recording and replay, and dropping `exclude_kernel`
would introduce non-determinism.

> Another consideration is, I'm not sure if the kernel can accurately realize
> that if the interrupt is to trigger sampling or just deliver a signal.
> Userspace may know that but kernel may not.

After looking at this code a bit more, I think that changing the
`is_sample_allowed` check from an early return to a guard around the
invocation of `overflow_handler` would fix this. I believe, but have
not tested, that `perf_event_fasync` is what must run to deliver our
signal, while the `overflow_handler` is what copies the kernel RIP/etc
into the output buffer that you want to skip.

- Kyle

> Anyway let's listen to more comments from community.
>
> Thanks
>
> Jin Yao
>
>
>
> On 6/28/2017 12:51 PM, Kyle Huey wrote:
>>
>> On Tue, Jun 27, 2017 at 7:09 PM, Jin, Yao <yao.jin@xxxxxxxxxxxxxxx> wrote:
>>>
>>> Hi,
>>>
>>> In theory, the PMI interrupts in skid region should be dropped, right?
>>
>> No, why would they be dropped?
>>
>> My understanding of the situation is as follows:
>>
>> There is some time, call it t_0, where the hardware counter overflows.
>> The PMU triggers an interrupt, but this is not instantaneous. Call
>> the time when the interrupt is actually delivered t_1. Then t_1 - t_0
>> is the "skid".
>>
>> Note that if the counter is `exclude_kernel`, then at t_0 the CPU
>> *must* be running a userspace program. But by t_1, the CPU may be
>> doing something else. Your patch changed things so that if at t_1 the
>> CPU is in the kernel, then the interrupt is discarded. But rr has
>> programmed the counter to deliver a signal on overflow (via F_SETSIG
>> on the fd returned by perf_event_open). This change results in the
>> signal never being delivered, because the interrupt was ignored.
>> (More accurately, the signal is delivered the *next* time the counter
>> overflows, which is far past where we wanted to inject our
>> asynchronous event into our tracee.
>>
>> It seems to me that it might be reasonable to ignore the interrupt if
>> the purpose of the interrupt is to trigger sampling of the CPUs
>> register state. But if the interrupt will trigger some other
>> operation, such as a signal on an fd, then there's no reason to drop
>> it.
>>
>>> For a userspace debugger, is it the only choice that relies on the *skid*
>>> PMI interrupt?
>>
>> I don't understand this question, but hopefully the above clarified
>> things.
>>
>> - Kyle
>>
>>> Thanks
>>> Jin Yao
>>>
>>>
>>> On 6/28/2017 9:01 AM, Kyle Huey wrote:
>>>>
>>>> Sent again with LKML CCd, sorry for the noise.
>>>>
>>>> - Kyle
>>>>
>>>> On Tue, Jun 27, 2017 at 5:38 PM, Kyle Huey <me@xxxxxxxxxxxx> wrote:
>>>>>
>>>>> cc1582c231ea introduced a regression in v4.12.0-rc5, and appears to be
>>>>> a candidate for backporting to stable branches.
>>>>>
>>>>> rr, a userspace record and replay debugger[0], uses the PMU interrupt
>>>>> to stop a program during replay to inject asynchronous events such as
>>>>> signals. We are counting retired conditional branches in userspace
>>>>> only. This changeset causes the kernel to drop interrupts on the
>>>>> floor if, during the PMU interrupt's "skid" region, the CPU enters
>>>>> kernel mode for whatever reason. When replaying traces of complex
>>>>> programs such as Firefox, we intermittently fail to deliver
>>>>> asynchronous events on time, leading the replay to diverge from the
>>>>> recorded state.
>>>>>
>>>>> It seems like this change should, at a bare minimum, be limited to
>>>>> counters that actually perform sampling of register state when the
>>>>> interrupt fires. In our case, with the retired conditional branches
>>>>> counter restricted to counting userspace events only, it makes no
>>>>> difference that the PMU interrupt happened to be delivered in the
>>>>> kernel.
>>>>>
>>>>> As this makes rr unusable on complex applications and cannot be
>>>>> efficiently worked around, we would appreciate this being addressed
>>>>> before 4.12 is finalized, and the regression not being introduced to
>>>>> stable branches.
>>>>>
>>>>> Thanks,
>>>>>
>>>>> - Kyle
>>>>>
>>>>> [0] http://rr-project.org/
>>>
>>>
>