[RFC PATCH 0/9] powerpc/64s: fast interrupt exit

[Christophe Leroy]

Le 10/11/2020 à 09:49, Nicholas Piggin a écrit :
> Excerpts from Christophe Leroy's message of November 7, 2020 8:35 pm:
>>
>>
>> Le 06/11/2020 à 16:59, Nicholas Piggin a écrit :
>>> This series attempts to improve the speed of interrupts and system calls
>>> in two major ways.
>>>
>>> Firstly, the SRR/HSRR registers do not need to be reloaded if they were
>>> not used or clobbered fur the duration of the interrupt.
>>>
>>> Secondly, an alternate return location facility is added for soft-masked
>>> asynchronous interrupts and then that's used to set everything up for
>>> return without having to disable MSR RI or EE.
>>>
>>> After this series, the entire system call / interrupt handler fast path
>>> executes no mtsprs and one mtmsrd to enable interrupts initially, and
>>> the system call vectored path doesn't even need to do that.
>>
>> Interesting series.
>>
>> Unfortunately, can't be done on PPC32 (at least on non bookE), because it would mean mapping kernel
>> at 0 instead of 0xC0000000. Not sure libc would like it, and anyway it would be an issue for
>> catching NULL pointer dereferencing, unless we use page tables instead of BATs to map kernel mem,
>> which would be serious performance cut.
> 
> Hmm, why would you have to map at 0?

In real mode, physical mem is at 0. If we want to switch from real to virtual mode by just writing 
to MSR, then we need virtuel addresses match with real addresses ?
We could play with chip selects to put RAM at 0xc0000000, but then we'd have a problem with 
exception as they have to be at 0. Or we could play with MSR[IP] and get the exceptions at 
0xfff00000, but that would not be so easy I guess.

> 
> PPC32 doesn't have soft mask interrupts, but you could still test all
> MSR[PR]=0 interrupts to see if they land inside some region to see if
> they hit in the restart table I think?

Yes and this is already what is done at exit for the ones that handle MSR[RI] I think.

> 
> Could PPC32 skip the SRR reload at least? That's simpler.

I think that would only be possible if real addresses where matching virtual addresses, or am I 
missing something ?

Christophe