Re: [PATCH v2] mm: count zeromap read and set for swapout and swapin

[David Hildenbrand]

On 05.11.24 10:15, Barry Song wrote:
> On Tue, Nov 5, 2024 at 9:23 PM David Hildenbrand <david@xxxxxxxxxx> wrote:
> > On 05.11.24 04:40, Andrew Morton wrote:
> > > On Mon, 4 Nov 2024 13:32:55 +0100 David Hildenbrand <david@xxxxxxxxxx> wrote:
> > >
> > > > > As mentioned above, this isn't about fixing a bug; it's simply to ensure
> > > > > that swap-related metrics don't disappear.
> > > >
> > > > Documentation/process/submitting-patches.rst:
> > > >
> > > > "A Fixes: tag indicates that the patch fixes an issue in a previous
> > > > commit. It is used to make it easy to determine where a bug originated,
> > > > which can help review a bug fix."
> > > >
> > > > If there is no BUG, I'm afraid you are abusing that tag.
> > >
> > > I think the abuse is reasonable.  We have no Should-be-included-with:.
> >
> > A "Belongs-to:" might make sense, for this kind of stuff that is still
> > only in an RFC. Or we update the doc to explicitly spell out this
> > special case of using "Fixes" to sort out something into the RC.
> >
> > Because if this would be already in a released kernel, it would get a
> > bit trickier: stable rules explicitly spell out "fix a real bug".
> >
> > > 0ca0c24e3211 is only in 6.12-rcX so this is the time to make
> > > userspace-visible tweaks, so the 6.12 interfaces are the same as the
> > > 6.13+ interfaces (which is what I think is happening here?)
> >   > > And including the Fixes in this patch might be useful to someone who is
> > > backporting 0ca0c24e3211 into some earlier kernel for their own
> > > purposes.
> >
> > Just to be clear: adding new counters would hardly be fixing existing
> > tools that perform calculations based on existing counters. So we are
> > already changing the "userspace-visible" portion in some way, and I have
> > no idea what in vmstat we consider "stable".
> >
> > But I still don't think it's all that big of a deal except in some
> > handcrafted scenarios hardly anybody cares about; the cover letter is
> > also pretty clear on that.
>
> I may have been mistaken in the cover letter. According to the zswap data Usama
> provided for servers, zero-filled pages accounted for about 1%. So
> really doesn't
> matter too much, but I just checked with Hailong from our team—he has data
> on same-page-filled usage in Android apps, which on average show 3-4%
> same-page-filled, with over 85% being zero-filled. Some apps even reach
> 6-7% zero-filled pages. We previously used these counters to profile
> optimizations, but with zeromap now serving as the frontend for swap files,
> those counters will disappear entirely from both zRAM and pswpin/pswpout
> metrics, as folios are filtered earlier.
>
> Hailong essentially has a table that looks like the below which could be
> collected from the existing counters:
>
> com.xxx.app     5% same-page-filled.    88% zero
> com.yyy.app     6% same-page-filled.     85% zero
> com.zzz.map   6.7 same-page-filled.       88% zero
> ....
>
> Anyone on 6.12 will be unable to track zero-filled pages unless they
> backport this patch from a newer kernel version if it doesn’t make it
> into 6.12.
>
> Whether it's marked as 'Belongs-to:' or 'Fixes:', I'd prefer we aim to
> land it in
> 6.12 :-)

Thanks for sharing these numbers. Just for the records, I think the 
optimization is quite powerful. What I am not convinced is that the 
missing stat is a big deal.

If I upgrade my car and it suddenly consumes 5% less fuel/energy I (a) 
either won't notice at all or (b) am happy and don't care why; I won't 
start digging through the updated manual, looking for some hidden gages 
that might explain why :)

But I might be missing something important, and I do understand why it 
can be helpful to have it black-on-white how much this optimization 
gives you. So feel free to move forward with this as you like, having it 
in 6.12 sounds very reasonable to me.

--
Cheers,

David / dhildenb