[PATCH v6 21/23] powerpc/64s: Use contiguous PMD/PUD instead of HUGEPD

[LEROY Christophe]

Le 25/06/2024 à 06:49, Nicholas Piggin a écrit :
> On Tue Jun 25, 2024 at 12:45 AM AEST, Christophe Leroy wrote:
>> On book3s/64, the only user of hugepd is hash in 4k mode.
>>
>> All other setups (hash-64, radix-4, radix-64) use leaf PMD/PUD.
>>
>> Rework hash-4k to use contiguous PMD and PUD instead.
>>
>> In that setup there are only two huge page sizes: 16M and 16G.
>>
>> 16M sits at PMD level and 16G at PUD level.
>>
>> pte_update doesn't know page size, lets use the same trick as
>> hpte_need_flush() to get page size from segment properties. That's
>> not the most efficient way but let's do that until callers of
>> pte_update() provide page size instead of just a huge flag.
>>
>> Signed-off-by: Christophe Leroy <christophe.leroy at csgroup.eu>
> 
> [snip]
> 
>> +static inline unsigned long hash__pte_update(struct mm_struct *mm,
>> +					 unsigned long addr,
>> +					 pte_t *ptep, unsigned long clr,
>> +					 unsigned long set,
>> +					 int huge)
>> +{
>> +	unsigned long old;
>> +
>> +	old = hash__pte_update_one(ptep, clr, set);
>> +
>> +	if (IS_ENABLED(CONFIG_PPC_4K_PAGES) && huge) {
>> +		unsigned int psize = get_slice_psize(mm, addr);
>> +		int nb, i;
>> +
>> +		if (psize == MMU_PAGE_16M)
>> +			nb = SZ_16M / PMD_SIZE;
>> +		else if (psize == MMU_PAGE_16G)
>> +			nb = SZ_16G / PUD_SIZE;
>> +		else
>> +			nb = 1;
>> +
>> +		WARN_ON_ONCE(nb == 1);	/* Should never happen */
>> +
>> +		for (i = 1; i < nb; i++)
>> +			hash__pte_update_one(ptep + i, clr, set);
>> +	}
>>   	/* huge pages use the old page table lock */
>>   	if (!huge)
>>   		assert_pte_locked(mm, addr);
>>   
>> -	old = be64_to_cpu(old_be);
>>   	if (old & H_PAGE_HASHPTE)
>>   		hpte_need_flush(mm, addr, ptep, old, huge);
>>   
> 
> We definitely need a bit more comment and changelog about the atomicity
> issues here. I think the plan should be all hash-side access just
> operates on PTE[0], which should avoid that whole race. There could be
> some cases that don't follow that. Adding some warnings to catch such
> things could be good too.

That seems to be the case indeed, as we have the following in 
hash_page_mm():

#ifndef CONFIG_PPC_64K_PAGES
	/*
	 * If we use 4K pages and our psize is not 4K, then we might
	 * be hitting a special driver mapping, and need to align the
	 * address before we fetch the PTE.
	 *
	 * It could also be a hugepage mapping, in which case this is
	 * not necessary, but it's not harmful, either.
	 */
	if (psize != MMU_PAGE_4K)
		ea &= ~((1ul << mmu_psize_defs[psize].shift) - 1);
#endif /* CONFIG_PPC_64K_PAGES */


> 
> I'd been meaning to do more on this sooner, sorry. I've started
> tinkering with adding a bit of debug code. I'll see if I can help with
> adding a bit of comments.

Yes would we very welcome, I guess you'll send it as followup/fixup 
patch to the series ?

> 
> [snip]
> 
>> diff --git a/arch/powerpc/mm/book3s64/hugetlbpage.c b/arch/powerpc/mm/book3s64/hugetlbpage.c
>> index 5a2e512e96db..83c3361b358b 100644
>> --- a/arch/powerpc/mm/book3s64/hugetlbpage.c
>> +++ b/arch/powerpc/mm/book3s64/hugetlbpage.c
>> @@ -53,6 +53,16 @@ int __hash_page_huge(unsigned long ea, unsigned long access, unsigned long vsid,
>>   		/* If PTE permissions don't match, take page fault */
>>   		if (unlikely(!check_pte_access(access, old_pte)))
>>   			return 1;
>> +		/*
>> +		 * If hash-4k, hugepages use seeral contiguous PxD entries
>> +		 * so bail out and let mm make the page young or dirty
>> +		 */
>> +		if (IS_ENABLED(CONFIG_PPC_4K_PAGES)) {
>> +			if (!(old_pte & _PAGE_ACCESSED))
>> +				return 1;
>> +			if ((access & _PAGE_WRITE) && !(old_pte & _PAGE_DIRTY))
>> +				return 1;
>> +		}
>>   
>>   		/*
>>   		 * Try to lock the PTE, add ACCESSED and DIRTY if it was
> 
> I'm hoping we wouldn't have to do this, if we follow the PTE[0] rule.

But we still need all entries to be updated so that page walker which 
don't know they must use PTE[0] get the right information ?

> 
> I think is minor enough that should not prevent testing in -mm.

Thanks.

> 
> Thanks,
> Nick

Christophe