Re: [PATCH RFC 11/14] arm64: Move the ASID allocator code in a separate file

[Guo Ren]

Thx Catalin,

On Fri, Jun 21, 2019 at 10:16 PM Catalin Marinas
<catalin.marinas@xxxxxxx> wrote:
>
> On Wed, Jun 19, 2019 at 07:51:03PM +0800, Guo Ren wrote:
> > On Wed, Jun 19, 2019 at 4:54 PM Julien Grall <julien.grall@xxxxxxx> wrote:
> > > On 6/19/19 9:07 AM, Guo Ren wrote:
> > > > Move arm asid allocator code in a generic one is a agood idea, I've
> > > > made a patchset for C-SKY and test is on processing, See:
> > > > https://lore.kernel.org/linux-csky/1560930553-26502-1-git-send-email-guoren@xxxxxxxxxx/
> > > >
> > > > If you plan to seperate it into generic one, I could co-work with you.
> > >
> > > Was the ASID allocator work out of box on C-Sky?
> >
> > Almost done, but one question:
> > arm64 remove the code in switch_mm:
> >   cpumask_clear_cpu(cpu, mm_cpumask(prev));
> >   cpumask_set_cpu(cpu, mm_cpumask(next));
> >
> > Why? Although arm64 cache operations could affect all harts with CTC
> > method of interconnect, I think we should keep these code for
> > primitive integrity in linux. Because cpu_bitmap is in mm_struct
> > instead of mm->context.
>
> We didn't have a use for this in the arm64 code, so no point in
> maintaining the mm_cpumask. On some arm32 systems (ARMv6) with no
> hardware broadcast of some TLB/cache operations, we use it to track
> where the task has run to issue IPI for TLB invalidation or some
> deferred I-cache invalidation.
The operation of set/clear mm_cpumask was removed in arm64 compared to
arm32. It seems no side effect on current arm64 system, but from
software meaning it's wrong.
I think we should keep mm_cpumask just like arm32.

>
> (there was also a potential optimisation on arm64 to avoid broadcast
> TLBI if the task only ran on a single CPU but Will found that was rarely
> the case on an SMP system because of rebalancing happening during
> execve(), ending up with two bits set in the mm_cpumask)
>
> The way you use it on csky is different from how it is done on arm. It
> seems to clear the mask for the scheduled out (prev) task but this
> wouldn't work on arm(64) since the TLB still contains prev entries
> tagged with the scheduled out ASID. Whether it matters, I guess it
> depends on the specifics of your hardware.
Sorry for the mistake quote, what I mean is what is done in arm32:
clear all bits of mm->cpu_mask in new_context(), and set back one by
one. Here is my patch:
https://lore.kernel.org/linux-csky/CAJF2gTQ0xQtQY1t-g9FgWaxfDXppMkFooCQzTFy7+ouwUfyA6w@xxxxxxxxxxxxxx/T/#m2ed464d2dfb45ac6f5547fb3936adf2da456cb65

>
> While the algorithm may seem fairly generic, the semantics have a few
> corner cases specific to each architecture. See [1] for a description of
> the semantics we need on arm64 (CnP is a feature where the hardware
> threads of the same core can share the TLB; the original algorithm
> violated the requirements when this feature was enabled).
C-SKY SMP is only one hart per core, but here is a patch [1] with my
thought on SMT duplicate tlb flush:
[1] https://lore.kernel.org/linux-csky/1561305869-18872-1-git-send-email-guoren@xxxxxxxxxx/T/#u

For TLA+ model, I still need some learning before I could talk with you.

>
> BTW, if you find the algorithm fairly straightforward ;), see this
> bug-fix which took a formal model to identify: a8ffaaa060b8 ("arm64:
> asid: Do not replace active_asids if already 0").
I think it's one fo the cases that other archs also could get benefit
from arm's asid allocator code.
Btw, Is this detected by arm's aisd allocator TLA+ model ? Or a real
bug report ?

--
Best Regards
 Guo Ren

ML: https://lore.kernel.org/linux-csky/