RE: [BUG] Variable stopmachine_state should be volatile

[Zhang, Yanmin]

>>-----Original Message-----
>>From: Arjan van de Ven [mailto:arjan@xxxxxxxxxxxxx]
>>Sent: 2005年12月2日 16:04
>>To: Zhang, Yanmin
>>Cc: linux-kernel@xxxxxxxxxxxxxxx; Pallipadi, Venkatesh; Shah, Rajesh
>>Subject: Re: [BUG] Variable stopmachine_state should be volatile
>>
>>On Wed, 2005-11-30 at 10:04 +0800, Zhang, Yanmin wrote:
>>> The model to access variable stopmachine_state is that a main thread
>>> writes it and other threads read it. Its declaration has no sign
>>> volatile. In the while loop in function stopmachine, this variable is
>>> read, and compiler might optimize it by reading it once before the loop
>>> and not reading it again in the loop, so the thread might enter dead
>>> loop.
>>
>>cpu_relax() includes a compiler barier..... so... what's wrong with the
>>compiler that it ignores such barriers?

You are right. I hit the problem when I compiled kernel 2.6.9 on IA64 by intel compiler.
cpu_relax has the compiler barrier if we use gcc, but cpu_relax becomes just ia64_hint which is null when I use intel compiler to compile kernel on ia64. file include/asm-ia64/intel_intrin.h defines ia64_hint as null.

Function stopmachine in kernel/stop_machine.c uses cpu_relax to prevent compiler from moving the reading of stopmachine_state out of the while loop. But when we use intel compiler, cpu_relax doesn't work because it is just null.

The right approach is to define ia64_hint to ia64_barrier in file include/asm-ia64/intel_intrin.h. I tested the new approach and it does work.

Thank Arjan, Pavel, and Venki.

-
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/