Re: trace: trace_kprobe.c always shows interrupts off

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

----- On Nov 18, 2015, at 1:51 AM, masami hiramatsu pt masami.hiramatsu.pt@hitachi.com wrote:

> Hi Mathieu, Steven,
> 
> From: Steven Rostedt [mailto:rostedt@goodmis.org]
>>
>>On Tue, 17 Nov 2015 16:35:35 +0000 (UTC)
>>Mathieu Desnoyers <mathieu.desnoyers@efficios.com> wrote:
>>
>>> Hi,
>>>
>>> I notice that trace_kprobe.c does local_save_flags() within
>>> __kprobe_trace_func(), which is called (at least on x86) with
>>> interrupts always disabled. This is then used as interrupt on/off
>>> state within the recorded event, which is misleading.
>>
>>Kinda. The kprobe itself has interrupts disabled, so it's only a white
>>lie.
>>
>>>
>>> I also don't understand why arch/x86/kernel/kprobes/ftrace.c
>>> pre handler disables interrupts, considering the following
>>> comment above arch/x86/kernel/kprobes/core.c: kprobe_int3_handler()
>>>
>>>  * Interrupts are disabled on entry as trap3 is an interrupt gate and they
>>>  * remain disabled throughout this function.
>>
>>I think you answered your own question. The key piece you may be
>>missing is that kprobe_ftrace_handler() is not called from a trap, but
>>from a function traced callback, which does not disable interrupts.
> 
> right, since the user can not forcibly change any kprobes to
> jump, those handlers should be ran under the same environment.
> 
>>> A struct pt_regs is received by this function, but I don't see
>>> any way to get the state of irq enable/disable from struct pt_regs
>>> across architectures.
>>>
>>> Any thoughts on how to fix this ?
>>
>>Create a cross arch: flags = regs_irq_save(regs) function.
>>
> 
> Good idea! Anyway, we can start on x86, in other arch we can
> prepare a dummy function to return current interrupt state(as
> we are doing now).

Derived from Steven's idea, I added a lttng_regs_irqs_disabled(struct pt_regs *regs)
to lttng-modules. The generic version returns "-1", which means that the architecture
don't support it yet. It returns 1 or 0 when implemented.

I prefer this approach compared to fall-back to irqs_disabled(), because
users then know that they can trust the information when it is
available.

Returning an unknown state with a regs_irq_save(regs) did not
seem obvious, because there is no guarantee that a value like
"-1" does not mean something specific on some architectures.
One possibility would be to make it return success/failure,
and populate the flags in an output parameter, but that seems
rather cumbersome.

When regs_irq_save() becomes widely available across architectures
in newer kernels, I'll use it of course.

Thanks,

Mathieu


> 
> Thank you,
> 
> ----
> Masami Hiramatsu

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