Re: [PATCH V2] mm/thp: Make ALLOC_SPLIT_PTLOCKS dependent on USE_SPLIT_PTE_PTLOCKS

[Anshuman Khandual]

On 7/1/21 6:27 PM, Matthew Wilcox wrote:
> On Thu, Jul 01, 2021 at 10:51:27AM +0530, Anshuman Khandual wrote:
>>
>>
>> On 5/20/21 4:47 PM, Matthew Wilcox wrote:
>>> On Wed, May 19, 2021 at 01:03:06PM +0530, Anshuman Khandual wrote:
>>>> Split ptlocks need not be defined and allocated unless they are being used.
>>>> ALLOC_SPLIT_PTLOCKS is inherently dependent on USE_SPLIT_PTE_PTLOCKS. This
>>>> just makes it explicit and clear. While here drop the spinlock_t element
>>>> from the struct page when USE_SPLIT_PTE_PTLOCKS is not enabled.
>>>
>>> I didn't spot this email yesterday.  I'm not a fan.  Isn't struct page
>>> already complicated enough without adding another ifdef to it?  Surely
>>> there's a better way than this.
>>
>> This discussion thread just got dropped off the radar, sorry about it.
>> None of the spinlock_t elements are required unless split ptlocks are
>> in use. I understand your concern regarding yet another #ifdef in the
>> struct page definition. But this change is simple and minimal. Do you
>> have any other particular alternative in mind which I could explore ?
> 
> Do nothing?  I don't understand what problem you're trying to solve.

Currently there is an element (spinlock_t ptl) in the struct page for page
table lock. Although a struct page based spinlock is not even required in
case USE_SPLIT_PTE_PTLOCKS evaluates to be false. Is not that something to
be fixed here i.e drop the splinlock_t element if not required ?

The problem is USE_SPLIT_PTE_PTLOCKS and ALLOC_SPLIT_PTLOCKS get evaluated
independently, although they are inherently dependent. ALLOC_SPLIT_PTLOCKS
could just be set to 0, when USE_SPLIT_PTE_PTLOCKS evaluates to be 0. This
patch makes that dependency explicit and also fixes the above situation.