Re: [PATCH 1/2] Make the TSC safe to be used by gettimeofday().

[dean gaudet]

On Mon, 13 Nov 2006, dean gaudet wrote:

> next an implementation which relies on the kernel restarting the computation when
> necessary.  this would be achieved by testing to see when the task to be restarted
> is on the vsyscall page and backtracking the task to the vsyscall entry point.
> 
> this is challenging when the vsyscall is implemented in C -- because of potential
> stack usage.  there are ways to get this to work though, even without resorting to
> assembly.  i'm presenting this only as a best case scenario should such an effort
> be undertaken.  (i have a crazy idea involving the direction flag which i need to
> mock up.)

nevermind the crazy idea using DF... i was hoping to use DF as a generic 
"restart a vsyscall" indicator -- switch_to() would note the task is on 
the vsyscall page and unilaterally clear DF before restoring eflags.

then a vsyscall critical section could be surrounded like so:

	unsigned long tmp;
	do {
		asm volatile("std");

		critical section

		asm volatile(
			"\n	pushf"
			"\n	pop %0"
			"\n	cld"
			: "=r" (tmp));
	} while ((tmp & 0x400) == 0);

it works great on k8 ... but DF manipulation hurts way too much on core2 and p4.

i even tried reading DF using a string instruction:

	long tmp;
	do {
		asm volatile("std");

		critical section

		asm volatile(
			"\n	mov %%rsp,%%rsi"
			"\n	lodsl"
			"\n	sub %%rsp,%%rsi"
			"\n	cld"
			: "=S" (tmp));
	} while (tmp > 0);

it's no better.

i've also tried similar tricks setting the EFLAGS.ID bit... but the popf
hurts in that case.

i think a general vsyscall restart mechanism would be useful (for more
than just the time functions), but still haven't found one which is
cheap enough.

-dean
-
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/