[Adeos-main] x86-32: Unpreemptible return from __ipipe_handle_irq

[Adeos-main] x86-32: Unpreemptible return from __ipipe_handle_irq

[Jan Kiszka]

Hi Philippe,

__ipipe_handle_irq:

#if defined(CONFIG_X86_32) && defined(CONFIG_SMP)
	/*
	 * Prevent a spurious rescheduling from being triggered on
	 * preemptible kernels along the way out through
	 * ret_from_intr.
	 */
	if ((long)regs->orig_ax < 0)
		__set_bit(IPIPE_STALL_FLAG, &ipipe_root_cpudom_var(status));
#endif	/* CONFIG_SMP */

But this defeats any preemption of kernel contexts on CONFIG_PREEMPT,
doesn't it? __ipipe_test_preemption should now always fail.

What kind of spurious rescheduling has to be prevented?

Jan

-- 
Siemens AG, Corporate Technology, CT T DE IT 1
Corporate Competence Center Embedded Linux

Re: [Adeos-main] x86-32: Unpreemptible return from __ipipe_handle_irq

[Philippe Gerum]

On Wed, 2010-09-08 at 18:53 +0200, Jan Kiszka wrote:
> Hi Philippe,
> 
> __ipipe_handle_irq:
> 
> #if defined(CONFIG_X86_32) && defined(CONFIG_SMP)
> 	/*
> 	 * Prevent a spurious rescheduling from being triggered on
> 	 * preemptible kernels along the way out through
> 	 * ret_from_intr.
> 	 */
> 	if ((long)regs->orig_ax < 0)
> 		__set_bit(IPIPE_STALL_FLAG, &ipipe_root_cpudom_var(status));
> #endif	/* CONFIG_SMP */
> 
> But this defeats any preemption of kernel contexts on CONFIG_PREEMPT,
> doesn't it? __ipipe_test_preemption should now always fail.
> 
> What kind of spurious rescheduling has to be prevented?

This hack was introduced in 2.6.9; disabling preemption on the IRQ
return path was working around a SMP issue causing a crash. I would not
be surprised if such code was only papering over a migration issue in
fact, which probably disappeared later in the pipeline implementation.
x86_64 never required this, and it is likely time to re-assess the need
for this hack, and fix the issue properly if ever it still makes sense.

> 
> Jan
> 

-- 
Philippe.