[PATCH v4 08/15] arm64: mm: Avoid TLBI when marking pages as valid

[Steven Price]

On 09/07/2024 12:57, Will Deacon wrote:
> On Mon, Jul 01, 2024 at 10:54:58AM +0100, Steven Price wrote:
>> When __change_memory_common() is purely setting the valid bit on a PTE
>> (e.g. via the set_memory_valid() call) there is no need for a TLBI as
>> either the entry isn't changing (the valid bit was already set) or the
>> entry was invalid and so should not have been cached in the TLB.
>>
>> Signed-off-by: Steven Price <steven.price at arm.com>
>> ---
>> v4: New patch
>> ---
>>  arch/arm64/mm/pageattr.c | 8 +++++++-
>>  1 file changed, 7 insertions(+), 1 deletion(-)
>>
>> diff --git a/arch/arm64/mm/pageattr.c b/arch/arm64/mm/pageattr.c
>> index 0e270a1c51e6..547a9e0b46c2 100644
>> --- a/arch/arm64/mm/pageattr.c
>> +++ b/arch/arm64/mm/pageattr.c
>> @@ -60,7 +60,13 @@ static int __change_memory_common(unsigned long start, unsigned long size,
>>  	ret = apply_to_page_range(&init_mm, start, size, change_page_range,
>>  					&data);
>>  
>> -	flush_tlb_kernel_range(start, start + size);
>> +	/*
>> +	 * If the memory is being made valid without changing any other bits
>> +	 * then a TLBI isn't required as a non-valid entry cannot be cached in
>> +	 * the TLB.
>> +	 */
>> +	if (pgprot_val(set_mask) != PTE_VALID || pgprot_val(clear_mask))
>> +		flush_tlb_kernel_range(start, start + size);
>>  	return ret;
> 
> Can you elaborate on when this actually happens, please? It feels like a
> case of "Doctor, it hurts when I do this" rather than something we should
> be trying to short-circuit in the low-level code.

This is for the benefit of the following patch. When transitioning a
page between shared and private we need to change the IPA (to set/clear
the top bit). This requires a break-before-make - see
__set_memory_enc_dec() in the following patch.

The easiest way of implementing the code was to just call
__change_memory_common() twice - once to make the entry invalid and then
again to make it valid with the new IPA. But that led to a double TLBI
which Catalin didn't like[1].

So this patch removes the unnecessary second TLBI by detecting that
we're just making the entry valid. Or at least it would if I hadn't
screwed up...

I should have changed the following patch to ensure that the second call
to __change_memory_common was only setting the PTE_VALID bit and not
changing anything else (so that the TLBI could be skipped) and forgot to
do that. So thanks for making me take a look at this again!

Thanks,
Steve

[1] https://lore.kernel.org/lkml/Zmc3euO2YGh-g9Th@arm.com/