Re: [External] [PATCH] riscv: Fix wrong size passed to local_flush_tlb_range_asid()

[Alexandre Ghiti <alex@ghiti.fr>]

Hi Dennis, Yunhui,

On 24/01/2024 09:19, Dennis Zhou wrote:
> Hello,
>
> On Wed, Jan 24, 2024 at 10:44:12AM +0800, yunhui cui wrote:
>> Hi Alexandre,
>>
>> On Tue, Jan 23, 2024 at 9:31 PM Alexandre Ghiti <alexghiti@rivosinc.com> wrote:
>>> local_flush_tlb_range_asid() takes the size as argument, not the end of
>>> the range to flush, so fix this by computing the size from the end and
>>> the start of the range.
>>>
>>> Fixes: 7a92fc8b4d20 ("mm: Introduce flush_cache_vmap_early()")
>>> Signed-off-by: Alexandre Ghiti <alexghiti@rivosinc.com>
>>> ---
>>>   arch/riscv/mm/tlbflush.c | 2 +-
>>>   1 file changed, 1 insertion(+), 1 deletion(-)
>>>
>>> diff --git a/arch/riscv/mm/tlbflush.c b/arch/riscv/mm/tlbflush.c
>>> index 8d12b26f5ac3..9619965f6501 100644
>>> --- a/arch/riscv/mm/tlbflush.c
>>> +++ b/arch/riscv/mm/tlbflush.c
>>> @@ -68,7 +68,7 @@ static inline void local_flush_tlb_range_asid(unsigned long start,
>>>
>>>   void local_flush_tlb_kernel_range(unsigned long start, unsigned long end)
>>>   {
>>> -       local_flush_tlb_range_asid(start, end, PAGE_SIZE, FLUSH_TLB_NO_ASID);
>>> +       local_flush_tlb_range_asid(start, end - start, PAGE_SIZE, FLUSH_TLB_NO_ASID);
>>>   }
> Well this was a miss during code review.. I'm going to take another look
> tomorrow and then likely pull this into a fixes branch.
>
>> What makes me curious is that this patch has not been tested?
>> BTW, It is best to keep the parameter order of all functions in
>> tlbflush.c consistent: cpumask, start, size, stride, asid.
>>
> I can't speak to the riscv communities testing/regression suites, but
> this would only be caught in a performance regression test.
>
> That being said, Alexandre, can you please lmk what level of testing
> this has gone through?


All my patches go through the same level of testing:

* Build/boot an Ubuntu kernel with and without KASAN + a few simple 
testsuites (libhugetlbfs, riscv kselftests and other)
* Build/boot a simple rootfs on ~40 different rv64 configs
* Build/boot a simple rootfs on ~30 different rv32 configs

And I run LTP/full kselftests/perf testsuite on a weekly basis on every 
rc. All this validation is done on qemu.

The patch is functional, it "simply" flushes the whole TLB instead of a 
few entries, so the only way to catch that would have been a performance 
regression. But given it only runs on qemu, it would have been hard to 
catch any performance regression since that involves the TLB.

@Yunhui: Please let me know how I should validate my patches better.

Thanks,

Alex


> Thanks,
> Dennis
>
> _______________________________________________
> linux-riscv mailing list
> linux-riscv@lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/linux-riscv