[PATCH v4 19/25] proc/task_mmu: Ignore ZONE_DEVICE pages

[David Hildenbrand]

On 19.12.24 00:11, Alistair Popple wrote:
> On Tue, Dec 17, 2024 at 11:31:25PM +0100, David Hildenbrand wrote:
>> On 17.12.24 06:13, Alistair Popple wrote:
>>> The procfs mmu files such as smaps currently ignore device dax and fs
>>> dax pages because these pages are considered special. To maintain
>>> existing behaviour once these pages are treated as normal pages and
>>> returned from vm_normal_page() add tests to explicitly skip them.
>>>
>>> Signed-off-by: Alistair Popple <apopple at nvidia.com>
>>> ---
>>>    fs/proc/task_mmu.c | 18 ++++++++++++++----
>>>    1 file changed, 14 insertions(+), 4 deletions(-)
>>>
>>> diff --git a/fs/proc/task_mmu.c b/fs/proc/task_mmu.c
>>> index 38a5a3e..c9b227a 100644
>>> --- a/fs/proc/task_mmu.c
>>> +++ b/fs/proc/task_mmu.c
>>> @@ -801,6 +801,8 @@ static void smaps_pte_entry(pte_t *pte, unsigned long addr,
>>>    	if (pte_present(ptent)) {
>>>    		page = vm_normal_page(vma, addr, ptent);
>>> +		if (page && (is_device_dax_page(page) || is_fsdax_page(page)))
>>
>> This "is_device_dax_page(page) || is_fsdax_page(page)" is a common theme
>> here, likely we should have a special helper?
> 
> Sounds good, will add is_dax_page() if there are enough callers left after any
> review comments.

:)

>   
>> But, don't we actually want to include them in the smaps output now? I think
>> we want.
> 
> I'm not an expert in what callers of vm_normal_page() think of as a "normal"
> page. 

Yeah, it's tricky. It means "this is abnormal, don't look at the struct 
page". We're moving away from that, such that these folios/pages will be 
... mostly normal :)

> So my philosphy here was to ensure anything calling vm_normal_page()
> didn't accidentally start seeing DAX pages, either by checking existing filters
> (lots of callers already call vma_is_special_huge() or some equivalent) or
> explicitly filtering them out in the hope someone smarter than me could tell me
> it was unneccssary.
> 
> That stategy seems to have worked, and so I agree we likely do want them in
> smaps. I just didn't want to silently do it without this kind of discussion
> first.

Yes, absolutely.

> 
>> The rmap code will indicate these pages in /proc/meminfo, per-node info, in
>> the memcg ... as "Mapped:" etc.
>>
>> So likely we just want to also indicate them here, or is there any downsides
>> we know of?
> 
> I don't know of any, and I think it makes sense to also indicate them so will
> drop this check in the respin.

It will be easy to hide them later, at least we talked about it. Thanks 
for doing all this!

-- 
Cheers,

David / dhildenb