Re: [Adeos-main] [PATCH+HACK] optimise root stalling

[Jan Kiszka <jan.kiszka@domain.hid>]

[-- Attachment #1: Type: text/plain, Size: 1828 bytes --]

Philippe Gerum wrote:
> On Thu, 2006-09-28 at 11:23 +0200, Jan Kiszka wrote:
>> Philippe Gerum wrote:
>>> On Thu, 2006-09-28 at 11:15 +0200, Jan Kiszka wrote:
>>>> Philippe Gerum wrote:
>>>>> On Mon, 2006-09-25 at 09:35 +0200, Jan Kiszka wrote:
>>>>>
>>>>> This one won't work. We need to forcibly re-enable the hw IRQs upon root
>>>>> unstall requests, regardless of the fact that interrupts are pending in
>>>>> the log; some code rely on this. E.g. Adeos/ppc over 2.4 would remain
>>>>> stuck in the delay calibration routine, but there are other more tricky
>>>>> places where this would bite too.
>>>> But this would be ok?
>>>>
>>>> #else  /* !CONFIG_SMP */
>>>>         __clear_bit(IPIPE_STALL_FLAG,
>>>>                     &ipipe_root_domain->cpudata[cpuid].status);
>>>>
>>>>        if (unlikely(ipipe_root_domain->cpudata[cpuid].irq_pending_hi !=
>>>>            0)) {
>>>>                local_irq_disable_hw();
>>>>                __ipipe_sync_pipeline(IPIPE_IRQMASK_ANY);
>>>>        }
>>>>        local_irq_enable_hw();
>>>> #endif /* CONFIG_SMP */
>>>>
>>>> So we can still save one disable IRQ in the fastpath.
>>>>
>>> But in such a case, you would have to use clear_bit() to keep atomicity.
>> This is for UP, not SMP.
>>
> 
> This is not a UP vs SMP issue, the stall bit is strictly CPU local
> anyway. The issue is that some archs do _not_ have intrinsically atomic
> bitops; they need to emulate them by masking interrupts around the bit
> setting operation. Using __clear_bit() in this context would make this
> code preemptible by an interrupt in the middle of the bitop. Disabling
> hw IRQs on entry allows us to use non-atomic bitops.

Yeah, I meanwhile realised my x86-restricted view as well. :)

Does it still make sense then to optimise like I proposed?


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 250 bytes --]