On 25.04.24 07:40, John Hubbard wrote:
On 4/24/24 9:17 PM, Matthew Wilcox wrote:
On Wed, Apr 24, 2024 at 09:00:50PM -0700, John Hubbard wrote:
We want to limit the use of page_mapcount() to places where absolutely
required, to prepare for kernel configs where we won't keep track of
per-page mapcounts in large folios.
Just curious, can you elaborate on the motivation? I probably missed
the discussions that explained why page_mapcount() in large folios
is not desirable. Are we getting rid of a field in struct page/folio?
Some other reason?
Thanks for looking into this!
Two reasons. One is that, regardless of anything else, folio_mapcount()
is expensive on large folios as it has to walk every page in the folio
summing the mapcounts. The more important reason is that when we move
to separately allocated folios, we don't want to allocate an array of
mapcounts in order to maintain a per-page mapcount.
So we're looking for a more compact scheme to avoid maintaining a
per-page mapcount.
I see. Thanks for explaining the story.
We'll talk more about all that at LSF/MM in the mapcount session. A spoiler:
page_mapcount() in the context of large folios:
* Is a misunderstood function (e.g., page_mapcount() vs page_count()
checks, mapped = !page_mapcount() checks).
* Is a misleading function (e.g., page_mapped() == folio_mapped() but
page_mapcount() != folio_mapcount())
We could just rename it to "folio_precise_page_mapcount()", but then,
once we tackle the subpage mapcount optimizations (initially using a
separate kernel config toggle), we'll have to teach each caller about an
alternative that gets the job done, and it's not that easy to prevent
further reuse around the kernel.
If you look at linux-next, we're down to 5 page_mapcount() calls in
fs/proc/, so I'll relocate it to fs/proc/internal.h to prevent any
further use - once the s390x change lands in the next merge window.
Regarding the subpage mapcount optimizations, I can further add:
* (un)map performance improvements for PTE-mapped THP
* Preparation for folio_order() > PMD_ORDER, where the current scheme
won't scale and needs further adjustments/complexity to even keep it
working
* Preparation for hugetlb-like vmemmap optimizations until we have
memdescs / dynamically allocated folios
* (Paving the way for partially mapping hugetlb folios that faced
similar issues? Not sure if that ever gets real, though)
Is this patch ahead of its time? LSF/MM is just around the corner, and
I'm planning on posting the other relevant patches in the next months.
The khugepage MM selftests keep working as expected, including:
Run test: collapse_max_ptes_shared (khugepaged:anon)
Allocate huge page... OK
Share huge page over fork()... OK
Trigger CoW on page 255 of 512... OK
Maybe collapse with max_ptes_shared exceeded.... OK
Trigger CoW on page 256 of 512... OK
Collapse with max_ptes_shared PTEs shared.... OK
Check if parent still has huge page... OK
Well, a word of caution! These tests do not (yet) cover either of
the interesting new cases that folio_likely_mapped_shared() presents:
KSM or hugetlbfs interactions. In other words, false positives.
Hmm ... KSM never uses large folios and hugetlbfs is disjoint from
khugepaged?
Right, folio_likely_mapped_shared() behaves exactly like page_mapcount()
would for (small) KSM folios, so no change for them.
Thankfully, hugetlb is out of the picture this time.
Oh good. I thought we might have had a testing hole, but no.
Thanks for having a look!
I'm only a bit concerned about folio_likely_mapped_shared() "false
negatives" (detecting exclusive although shared), until we have a more
precise folio_likely_mapped_shared() variant to not unexpectedly waste
memory.
Imagine someone would be setting "khugepaged_max_ptes_shared=0", and
then we have an area where (I think this is the extreme case):
* We map 256 subpages of a 2M folio that are shared 256 times with a
child process.
* We don't map the first subpage.
* One PTE maps another page and is pte_write().
* 255 PTEs are pte_none().
folio_likely_mapped_shared() would return "false".
But then my thinking is:
* We are already wasting 256 subpages that are free in the 2M folio.
Sure, we might be able to relaim it when deferred splitting.
* Why would someone set khugepaged_max_ptes_shared=0 but leave
khugepaged_max_ptes_none set that high that we would allow 255
pte_none?
* If the child is a short-living subprocess, we don't really care
* Any futher writes to unbacked/R/O PTEs in that PMD area would COW and
consume memory.
So I had to use more and more "ifs" to construct a scenario where we
might end up wasting 1M of memory, at which point I decided "this is
really a corner case" and likely not worth the worry.
If we run into real issues, though, it will be easy to just inline
page_mapcount() here to resolve it; but the less special-casing the better.
--
Cheers,
David / dhildenb
