Hi Will,
On 03/06/2018 12:48 PM, Shanker Donthineni wrote:
> Hi Will,
>
> On 03/06/2018 09:23 AM, Will Deacon wrote:
>> Hi Shanker,
>>
>> On Tue, Mar 06, 2018 at 08:47:27AM -0600, Shanker Donthineni wrote:
>>> On 03/06/2018 07:44 AM, Will Deacon wrote:
>>>> I think this is a slight asymmetry with the code for the I-side. On the
>>>> I-side, you hook into invalidate_icache_by_line, whereas on the D-side you
>>>> hook into the callers of dcache_by_line_op. Why is that?
>>>>
>>>
>>> There is no particular reason other than complexity of the macro with
>>> another
>>> alternative. I tried to avoid this change by updating
>>> __clean_dcache_area_pou().
>>> I can change if you're interested to see both I-Side and D-Side changes are
>>> symmetric some thing like this...
>>>
>>> .macro dcache_by_line_op op, domain, kaddr, size, tmp1, tmp2
>>>
>>> .if (\op == cvau)
>>> alternative_if ARM64_HAS_CACHE_IDC
>>> dsb ishst
>>> b 9997f
>>> alternative_else_nop_endif
>>> .endif
>>>
>>> dcache_line_size \tmp1, \tmp2
>>> add \size, \kaddr, \size
>>> sub \tmp2, \tmp1, #1
>>> bic \kaddr, \kaddr, \tmp2
>>> 9998:
>>> .if (\op == cvau || \op == cvac)
>>> alternative_if_not ARM64_WORKAROUND_CLEAN_CACHE
>>> dc \op, \kaddr
>>> alternative_else
>>> dc civac, \kaddr
>>> alternative_endif
>>> .elseif (\op == cvap)
>>> alternative_if ARM64_HAS_DCPOP
>>> sys 3, c7, c12, 1, \kaddr // dc cvap
>>> alternative_else
>>> dc cvac, \kaddr
>>> alternative_endif
>>> .else
>>> dc \op, \kaddr
>>> .endif
>>> add \kaddr, \kaddr, \tmp1
>>> cmp \kaddr, \size
>>> b.lo 9998b
>>> dsb \domain
>>> 9997:
>>> .endm
>>
>> I think it would be cleaner the other way round, actually -- move the check
>> out of invalidate_icache_by_line and into its two callers.
>>
>
> Sure, I'll send out the next patch with your suggestions.
>
>>>> I notice that the only user other than
>>>> flush_icache_range/__flush_cache_user_range or invalidate_icache_by_line
>>>> is in KVM, via invalidate_icache_range. If you want to hook in there, why
>>>> aren't you also patching __flush_icache_all? If so, I'd rather have the
>>>> I-side code consistent with the D-side code and do this in the handful of
>>>> callers. We might even be able to elide a branch or two that way.
>>>>
>>>
>>> Agree with you, it saves function calls overhead. I'll do this change...
>>>
>>> static void invalidate_icache_guest_page(kvm_pfn_t pfn, unsigned long size)
>>> {
>>> if (cpus_have_const_cap(ARM64_HAS_CACHE_DIC)
>>> __invalidate_icache_guest_page(pfn, size);
>>> }
>>>
>>>
>>>> I'm going to assume that I-cache aliases are all coherent if DIC=1, so it's
>>>> safe to elide our alias sync code.
>>>>
>>>
>>> I'm not sure about I-cache whether aliases are all coherent if DIC=1 ot not.
>>> Unfortunately I don't have any hardware to test DIC=1. I've verified IDC=1.
>>
>> I checked with our architects and aliases don't pose a problem here, so you
>> can ignore me :)
>>
>
> I also confirmed with Thomas Speier, we can skip __flush_icache_all() if
> DIC=1.
>
>
Planning to patch __flush_icache_all() itself instead of changing the callers.
This
way we can avoid "ic ialluis" completely. Is this okay for you?
static inline void __flush_icache_all(void)
{
/* Instruction cache invalidation is not required for I/D coherence? */
if (!cpus_have_const_cap(ARM64_HAS_CACHE_DIC)) {
asm("ic ialluis");
dsb(ish);
}
}
>> Will
>>
>> _______________________________________________
>> linux-arm-kernel mailing list
>> linux-arm-ker...@lists.infradead.org
>> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
>>
>
--
Shanker Donthineni
Qualcomm Datacenter Technologies, Inc. as an affiliate of Qualcomm
Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the Code Aurora Forum, a Linux
Foundation Collaborative Project.
