On 3/25/26 12:37, Pedro Falcato wrote:
> On Tue, Mar 24, 2026 at 09:18:42PM +0100, David Hildenbrand (Arm) wrote:
>> On 3/24/26 16:43, Pedro Falcato wrote:
>>> The common order-0 case is important enough to want its own branch, and
>>> avoids the hairy, large loop logic that the CPU does not seem to handle
>>> particularly well.
>>>
>>> While at it, encourage the compiler to inline batch PTE logic and resolve
>>> constant branches by adding __always_inline strategically.
>>>
>>> Reviewed-by: Lorenzo Stoakes (Oracle) <ljs@kernel.org>
>>> Signed-off-by: Pedro Falcato <pfalcato@suse.de>
>>> ---
>>>  mm/mprotect.c | 17 ++++++++++++-----
>>>  1 file changed, 12 insertions(+), 5 deletions(-)
>>>
>>> diff --git a/mm/mprotect.c b/mm/mprotect.c
>>> index 2eaf862e5734..2fda26107066 100644
>>> --- a/mm/mprotect.c
>>> +++ b/mm/mprotect.c
>>> @@ -103,7 +103,7 @@ bool can_change_pte_writable(struct vm_area_struct *vma, unsigned long addr,
>>>  	return can_change_shared_pte_writable(vma, pte);
>>>  }
>>> 
>>> -static int mprotect_folio_pte_batch(struct folio *folio, pte_t *ptep,
>>> +static __always_inline int mprotect_folio_pte_batch(struct folio *folio, pte_t *ptep,
>>>  				    pte_t pte, int max_nr_ptes, fpb_t flags)
>>>  {
>>>  	/* No underlying folio, so cannot batch */
>>> @@ -117,9 +117,9 @@ static int mprotect_folio_pte_batch(struct folio *folio, pte_t *ptep,
>>>  }
>>> 
>>>  /* Set nr_ptes number of ptes, starting from idx */
>>> -static void prot_commit_flush_ptes(struct vm_area_struct *vma, unsigned long addr,
>>> -		pte_t *ptep, pte_t oldpte, pte_t ptent, int nr_ptes,
>>> -		int idx, bool set_write, struct mmu_gather *tlb)
>>> +static __always_inline void prot_commit_flush_ptes(struct vm_area_struct *vma,
>>> +		unsigned long addr, pte_t *ptep, pte_t oldpte, pte_t ptent,
>>> +		int nr_ptes, int idx, bool set_write, struct mmu_gather *tlb)
>>>  {
>>>  	/*
>>>  	 * Advance the position in the batch by idx; note that if idx > 0,
>>> @@ -169,7 +169,7 @@ static int page_anon_exclusive_sub_batch(int start_idx, int max_len,
>>>   * pte of the batch. Therefore, we must individually check all pages and
>>>   * retrieve sub-batches.
>>>   */
>>> -static void commit_anon_folio_batch(struct vm_area_struct *vma,
>>> +static __always_inline void commit_anon_folio_batch(struct vm_area_struct *vma,
>>>  		struct folio *folio, struct page *first_page, unsigned long addr, pte_t *ptep,
>>>  		pte_t oldpte, pte_t ptent, int nr_ptes, struct mmu_gather *tlb)
>>>  {
>>> @@ -177,6 +177,13 @@ static void commit_anon_folio_batch(struct vm_area_struct *vma,
>>>  	int sub_batch_idx = 0;
>>>  	int len;
>>> 
>>> +	/* Optimize for the common order-0 case. */
>>> +	if (likely(nr_ptes == 1)) {
>>> +		prot_commit_flush_ptes(vma, addr, ptep, oldpte, ptent, 1,
>>> +				       0, PageAnonExclusive(first_page), tlb);
>>
>> To optimize that one, inlining prot_commit_flush_ptes() would be
>> sufficient. Does inlining the other two really help? I don't think we
>> can optimize out loops etc. for them?
> 
> Well, I'm getting meaningful (smaller) wins from adding those
> __always_inline's. (and I also get a small win for __always_inline on
> set_write_prot_commit_flush_ptes, but I didn't realize that until now).
> 
>>
>> I would have thought that specializing on nr_ptes==0 on an even higher
>> level--where we call
>> set_write_prot_commit_flush_ptes/prot_commit_flush_ptes() would allow
>> for optimizing the loops entirely for nr_ptes==0?
> 
> That could also work, but then set_write_prot_commit_flush_ptes (holy cow
> what a long name) would definitely need inlining. And might be a little uglier
> overall.

Right. The idea is that you __always__inline any code that has PTE
loops, such that all loops for nr_pages == 1 gets optimized out.

We do that for zap and fork logic.

> 
> This is the part where having data points other than my giga-fast-giga-powerful
> zen5 could prove handy :/
I just recently lost access to my reliably, well tunes, system ...

Is it just the following benchmark?

	https://gist.github.com/heatd/1450d273005aba91fa5744f44dfcd933

?


I can easily extending

	https://gitlab.com/davidhildenbrand/scratchspace/-/blob/main/pte-mapped-folio-benchmarks.c

to have an "mprotect" mode. I had that in the past bit discarded it.

Then, we can easily measure the effect on various folio sizes when
mprotect'ing a larger memory area.

With order-0 we can then benchmark small folios exclusively.



-- 
Cheers,

David