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

[Alistair Popple <apopple@nvidia.com>]

On Fri, Dec 20, 2024 at 07:32:52PM +0100, David Hildenbrand wrote:
> 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@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.
> 
> :)

In the end there was only a single caller so I will leave this open-coded.

> > > 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!

Not a problem. The other main thing in this patch is also hiding them from
/proc/<PID>/pagemap. Based on this discussion I can't think of any good reason
why we would want to hide them there so will also remove the checks in the
pagemap walker.

> -- 
> Cheers,
> 
> David / dhildenb
>