Re: [PATCH 1/2] tracing/ftrace: don't trace on early stage of secondary cpu boot

[Frédéric Weisbecker]

2008/12/24 Steven Rostedt <rostedt@xxxxxxxxxxx>:
>
> On Wed, 24 Dec 2008, Fr?d?ric Weisbecker wrote:
>> >>       /*
>> >> @@ -1233,7 +1233,7 @@ int trace_graph_entry(struct ftrace_graph_ent *trace)
>> >>       int cpu;
>> >>       int pc;
>> >>
>> >> -     if (!ftrace_trace_task(current))
>> >> +     if (!ftrace_trace_task(current) || ftrace_in_early_cpuinit())
>> >>               return 0;
>> >
>> > Actually, didn't you say current is not available either? We are testing
>> > current first.
>>
>> Oops. Strange, my tests didn't give me any problem.
>> I will fix it.
>
> Are you sure 'current' is dangerous? Perhaps we do not need to worry here,
> since we are testing current against something else, and if the CPU
> is just coming on line, it will not be the task we want.


Yes, a simple local_irq_save might call trace_hardirqs_off(), and then
increment something in current....
And current is found toward reading pda, hence a crash...

>>
>>
>> > I still wonder if there's a better way to find out if it is safe to run.
>> > Perhaps the arch code can test the %gs register to see if it is actually
>> > something valid?
>> >
>> > -- Steve
>>
>> I could test if MSR_GS_BASE references cpu_pda(cpu) but that remains
>> as much heavy...
>
> Perhaps just reading the %gs register and see if it faults would be good
> enough?
>
>        long dummy;
>        int faulted=0;
>
>            asm(
>                "1:     movq %%gs:0, %0\n"
>                "2:
>                ".section .fixup \"ax\"\n"
>                "3:     movl $1, %1\n"
>                "       jmp 2b\n"
>                ".previous"
>                _ASM_EXTABLE(1b, 3b)
>                : "=r"(dummy), "=r"(faulted)
>                : "1"(faulted));
>
>
> Would this work?  On the good cases, all we do is a read of the %gs
> section, which is the same as doing a smp_process_id(). On the bad case,
> we fault, but we also catch that fact. We are more interested in the good
> case being fast. but I'm not sure how well the start up code could take a
> fault, even when it is expected.


But on the worst case, that could cause a fault recursion since the first line
of do_page_fault in x86 is:

tsk = current;

> -- Steve
>
>
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/