[PATCHv10 00/36] THP refcounting redesign

From: Kirill A. Shutemov
Date: Thu Sep 03 2015 - 11:26:03 EST

Hello everybody,

The THP refcounting has been rebased onto newer -mm tree with
compound_head patchset applied. This caused few conflicts.

Since page->lru in all tail pages occupied by page->compound_head, I had
to adjust the code to use storage of ->mapping and ->index in the second
tail page as list_head to link pages into deferred split queue.

The goal of patchset is to make refcounting on THP pages cheaper with
simpler semantics and allow the same THP compound page to be mapped with
PMD and PTEs. This is required to get reasonable THP-pagecache
implementation.

With the new refcounting design it's much easier to protect against
split_huge_page(): simple reference on a page will make you the deal.
It makes gup_fast() implementation simpler and doesn't require
special-case in futex code to handle tail THP pages.

It should improve THP utilization over the system since splitting THP in
one process doesn't necessary lead to splitting the page in all other
processes have the page mapped.

The patchset drastically lower complexity of get_page()/put_page()
codepaths. I encourage people look on this code before-and-after to
justify time budget on reviewing this patchset.

= Changelog =

v10:
- rebased onto newer -mm tree with compound_head patchset applied.

v9:
- rebased to mmotm-2015-07-15-16-46: fix conflicts with DAX and pagemap
patchsets;
- simplify PG_double_map handling in __split_huge_pmd_locked();
- fix arm64 typo (Suzuki K. Poulose);
- fix build on !THP;
- checkpatch fixes;
- Tested-by/Reviewed-by Aneesh Kumar K.V;

v8:
- rebased to since-4.1;
- fix mmap10 from LTP: make check and clear PG_double_map atomic.

v7:
- avoid situation during split_huge_pmd() where we can temporarily drop
page_mapcount() to zero. It can lead to races e.g. with unmap code;
- update documentation;
- fix NR_ANON_PAGES accounting in page_remove_rmap();
- fix page_mapped();
- optimize page_mapped() and page_mapcount();
- fix PSS calculation for non-shared pages;

v6:
- rebase to since-4.0;
- optimize mapcount handling: significantely reduce overhead for most
common cases.
- split pages on migrate_pages();
- remove infrastructure for handling splitting PMDs on all architectures;
- fix page_mapcount() for hugetlb pages;

v5:
- Tested-by: Sasha Levin!â
- re-split patchset in hope to improve readability;
- rebased on top of page flags and ->mapping sanitizing patchset;
- uncharge compound_mapcount rather than mapcount for hugetlb pages
during removing from rmap;
- differentiate page_mapped() from page_mapcount() for compound pages;
- rework deferred_split_huge_page() to use shrinker interface;
- fix race in page_remove_rmap();
- get rid of __get_page_tail();
- few random bug fixes;
v4:
- fix sizes reported in smaps;
- defines instead of enum for RMAP_{EXCLUSIVE,COMPOUND};
- skip THP pages on munlock_vma_pages_range(): they are never mlocked;
- properly handle huge zero page on FOLL_SPLIT;
- fix lock_page() slow path on tail pages;
- account page_get_anon_vma() fail to THP_SPLIT_PAGE_FAILED;
- fix split_huge_page() on huge page with unmapped head page;
- fix transfering 'write' and 'young' from pmd to ptes on split_huge_pmd;
- call page_remove_rmap() in unfreeze_page under ptl.

= Design overview =

The main reason why we can't map THP with 4k is how refcounting on THP
designed. It built around two requirements:

- split of huge page should never fail;
- we can't change interface of get_user_page();

To be able to split huge page at any point we have to track which tail
page was pinned. It leads to tricky and expensive get_page() on tail pages
and also occupy tail_page->_mapcount.

Most split_huge_page*() users want PMD to be split into table of PTEs and
don't care whether compound page is going to be split or not.

The plan is:

- allow split_huge_page() to fail if the page is pinned. It's trivial to
split non-pinned page and it doesn't require tail page refcounting, so
tail_page->_mapcount is free to be reused.

- introduce new routine -- split_huge_pmd() -- to split PMD into table of
PTEs. It splits only one PMD, not touching other PMDs the page is
mapped with or underlying compound page. Unlike new split_huge_page(),
split_huge_pmd() never fails.

Fortunately, we have only few places where split_huge_page() is needed:
swap out, memory failure, migration, KSM. And all of them can handle
split_huge_page() fail.

In new scheme we use page->_mapcount is used to account how many time
the page is mapped with PTEs. We have separate compound_mapcount() to
count mappings with PMD. page_mapcount() returns sum of PTE and PMD
mappings of the page.

Introducing split_huge_pmd() effectively allows THP to be mapped with 4k.
It may be a surprise to some code to see a PTE which points to tail page
or VMA start/end in the middle of compound page.

munmap() part of THP will split PMD, but doesn't split the huge page. In
order to take memory consumption under control we put partially unmapped
huge page on list. The pages will be split by shrinker if memory pressure
comes. This way we also avoid unnecessary split_huge_page() on exit(2) if
a THP belong to more than one VMA.

= Refcounts and transparent huge pages =

- get_page() and put_page() work *only* on head page's ->_count.
We don't touch tail pages at all for these oprations. We stopped
touching ->_mapcount in tail pages to store it's pins.

- ->_count in tail pages is always zero: get_page_unless_zero() never
succeed on tail pages. Nothing changed in this respect.

- map/unmap of the pages with PTE entry increment/decrement ->_mapcount
on relevent sub-page of the compound page.

- map/unmap of the whole compound page accounted in compound_mapcount
(stored in first tail page).

- PageDoubleMap() indicates that ->_mapcount in all subpages is offset
up by one. This additional reference is required to get race-free
detection of unmap of subpages when we have them mapped with both PMDs
and PTEs.

This is optimization required to lower overhead of per-subpage
mapcount tracking. The alternative is alter ->_mapcount in all
subpages on each map/unmap of the whole compound page.

We set PG_double_map when a PMD of the page got split for the first
time, but still have PMD mapping. The addtional references go away
with last compound_mapcount.

= Benchmarks =

Kernel build benchmark:

baseline v6

Amean user-2 447.76 ( 0.00%) 451.24 ( -0.78%)
Amean user-4 314.94 ( 0.00%) 310.10 ( 1.54%)
Amean user-8 388.91 ( 0.00%) 388.95 ( -0.01%)
Amean user-16 518.68 ( 0.00%) 518.60 ( 0.02%)
Amean user-24 533.58 ( 0.00%) 535.35 ( -0.33%)
Amean syst-2 77.52 ( 0.00%) 70.16 ( 9.49%)
Amean syst-4 51.21 ( 0.00%) 44.55 ( 13.00%)
Amean syst-8 42.12 ( 0.00%) 42.59 ( -1.12%)
Amean syst-16 50.29 ( 0.00%) 50.14 ( 0.30%)
Amean syst-24 49.36 ( 0.00%) 48.54 ( 1.65%)
Amean elsp-2 242.64 ( 0.00%) 244.46 ( -0.75%)
Amean elsp-4 93.78 ( 0.00%) 92.89 ( 0.95%)
Amean elsp-8 61.51 ( 0.00%) 61.92 ( -0.66%)
Amean elsp-16 53.95 ( 0.00%) 53.80 ( 0.29%)
Amean elsp-24 52.75 ( 0.00%) 53.14 ( -0.74%)
Stddev user-2 15.49 ( 0.00%) 13.75 ( 11.24%)
Stddev user-4 7.85 ( 0.00%) 4.42 ( 43.68%)
Stddev user-8 1.29 ( 0.00%) 2.77 (-114.28%)
Stddev user-16 2.56 ( 0.00%) 1.54 ( 39.89%)
Stddev user-24 1.75 ( 0.00%) 1.06 ( 39.75%)
Stddev syst-2 3.02 ( 0.00%) 2.00 ( 33.86%)
Stddev syst-4 1.23 ( 0.00%) 0.91 ( 26.65%)
Stddev syst-8 0.41 ( 0.00%) 0.30 ( 28.32%)
Stddev syst-16 0.51 ( 0.00%) 0.71 (-38.07%)
Stddev syst-24 0.92 ( 0.00%) 0.70 ( 23.86%)
Stddev elsp-2 8.70 ( 0.00%) 7.99 ( 8.13%)
Stddev elsp-4 1.74 ( 0.00%) 0.59 ( 66.08%)
Stddev elsp-8 0.40 ( 0.00%) 0.30 ( 25.11%)
Stddev elsp-16 0.45 ( 0.00%) 0.37 ( 17.73%)
Stddev elsp-24 0.57 ( 0.00%) 0.38 ( 33.64%)

Changes are mostly non-significant. The only noticble part is reduction of
system time for -j2 and -j4: 9.49% and 13.00%.

specjvm
base v6
Ops compiler 569.52 ( 0.00%) 618.74 ( 8.64%)
Ops compress 456.32 ( 0.00%) 469.20 ( 2.82%)
Ops crypto 424.34 ( 0.00%) 413.64 ( -2.52%)
Ops derby 535.15 ( 0.00%) 536.96 ( 0.34%)
Ops mpegaudio 291.03 ( 0.00%) 286.35 ( -1.61%)
Ops scimark.large 75.91 ( 0.00%) 77.21 ( 1.71%)
Ops scimark.small 529.19 ( 0.00%) 527.07 ( -0.40%)
Ops serial 316.13 ( 0.00%) 316.40 ( 0.09%)
Ops sunflow 154.90 ( 0.00%) 154.85 ( -0.03%)
Ops xml 612.94 ( 0.00%) 575.20 ( -6.16%)
Ops compiler.compiler 770.11 ( 0.00%) 878.22 ( 14.04%)
Ops compiler.sunflow 421.17 ( 0.00%) 435.92 ( 3.50%)
Ops compress 456.32 ( 0.00%) 469.20 ( 2.82%)
Ops crypto.aes 153.17 ( 0.00%) 151.30 ( -1.22%)
Ops crypto.rsa 607.43 ( 0.00%) 564.65 ( -7.04%)
Ops crypto.signverify 821.23 ( 0.00%) 828.40 ( 0.87%)
Ops derby 535.15 ( 0.00%) 536.96 ( 0.34%)
Ops mpegaudio 291.03 ( 0.00%) 286.35 ( -1.61%)
Ops scimark.fft.large 69.59 ( 0.00%) 69.81 ( 0.32%)
Ops scimark.lu.large 20.31 ( 0.00%) 20.32 ( 0.05%)
Ops scimark.sor.large 114.57 ( 0.00%) 113.99 ( -0.51%)
Ops scimark.sparse.large 55.56 ( 0.00%) 61.71 ( 11.07%)
Ops scimark.monte_carlo 280.13 ( 0.00%) 275.09 ( -1.80%)
Ops scimark.fft.small 815.19 ( 0.00%) 819.55 ( 0.53%)
Ops scimark.lu.small 1072.62 ( 0.00%) 1081.47 ( 0.83%)
Ops scimark.sor.small 674.47 ( 0.00%) 674.24 ( -0.03%)
Ops scimark.sparse.small 251.20 ( 0.00%) 247.43 ( -1.50%)
Ops serial 316.13 ( 0.00%) 316.40 ( 0.09%)
Ops sunflow 154.90 ( 0.00%) 154.85 ( -0.03%)
Ops xml.transform 538.16 ( 0.00%) 519.64 ( -3.44%)
Ops xml.validation 698.10 ( 0.00%) 636.70 ( -8.80%)

Results are mixed.

= Patches overview =

Patch 1:
We need to look on all subpages of compound page to calculate
correct PSS, because they can have different mapcount.

Patch 2:
With PTE-mapeed THP, rmap cannot rely on PageTransHuge() check to
decide if map small page or THP. We need to get the info from
caller.

Patch 3:
Make memcg aware about new refcounting. Validation needed.

Patch 4:
Adjust conditions when we can re-use the page on write-protection
fault.

Patch 5:
FOLL_SPLIT should be handled on PTE level too.

Patch 6:
Make generic fast GUP implementation aware about PTE-mapped huge
pages.

Patch 7:
Split all pages in mlocked VMA. That should be good enough for
now.

Patch 8:
Make khugepaged aware about PTE-mapped huge pages.

Patch 9:
Rename split_huge_page_pmd() to split_huge_pmd() to reflect that
page is not going to be split, only PMD.

Patch 10:
New THP_SPLIT_* vmstats.

Patch 11:
Up to this point we tried to keep patchset bisectable, but next
patches are going to change how core of THP refcounting work.
That's easier to review change if we would disable THP temporally
and bring it back once everything is ready.

Patch 12:
Remove all split_huge_page()-related code. It also remove need in
tail page refcounting.

Patch 13:
Drop tail page refcounting. Diffstat is nice! :)

Patch 14:
Remove ugly special case if futex happened to be in tail THP page.
With new refcounting it much easier to protect against split.

Patch 15:
Simplify KSM code which handle THP.

Patch 16:
No need in compound_lock anymore.

Patches 17-25:
Drop infrastructure for handling PMD splitting. We don't use it
anymore in split_huge_page().

Patch 26:
Store mapcount for compound pages separately: in the first tail
page ->mapping.

Patch 27:
Let's define page_mapped() to be true for compound pages if any
sub-pages of the compound page is mapped (with PMD or PTE).

Patch 28:
Make numabalancing aware about PTE-mapped THP.

Patch 29:
Implement new split_huge_pmd().

Patch 30-32:
Implement new split_huge_page().

Patch 33:
Split pages instaed of PMDs on migrate_pages.

Patch 34:
Handle partial unmap of THP. We put partially unmapped huge page
list. Pages from list will split via shrinker if memory pressure
comes. This way we also avoid unnecessary split_huge_page() on
exit(2) if a THP belong to more than one VMA.

Patch 35:
Everything is in place. Re-enable THP.

Patch 36:
Documentation update.

The patchset also available on git:

git://git.kernel.org/pub/scm/linux/kernel/git/kas/linux.git thp/refcounting/v10

Please review and consider applying.
Kirill A. Shutemov (36):
mm, proc: adjust PSS calculation
rmap: add argument to charge compound page
memcg: adjust to support new THP refcounting
mm, thp: adjust conditions when we can reuse the page on WP fault
mm: adjust FOLL_SPLIT for new refcounting
mm: handle PTE-mapped tail pages in gerneric fast gup implementaiton
thp, mlock: do not allow huge pages in mlocked area
khugepaged: ignore pmd tables with THP mapped with ptes
thp: rename split_huge_page_pmd() to split_huge_pmd()
mm, vmstats: new THP splitting event
mm: temporally mark THP broken
thp: drop all split_huge_page()-related code
mm: drop tail page refcounting
futex, thp: remove special case for THP in get_futex_key
ksm: prepare to new THP semantics
mm, thp: remove compound_lock
arm64, thp: remove infrastructure for handling splitting PMDs
arm, thp: remove infrastructure for handling splitting PMDs
mips, thp: remove infrastructure for handling splitting PMDs
powerpc, thp: remove infrastructure for handling splitting PMDs
s390, thp: remove infrastructure for handling splitting PMDs
sparc, thp: remove infrastructure for handling splitting PMDs
tile, thp: remove infrastructure for handling splitting PMDs
x86, thp: remove infrastructure for handling splitting PMDs
mm, thp: remove infrastructure for handling splitting PMDs
mm: rework mapcount accounting to enable 4k mapping of THPs
mm: differentiate page_mapped() from page_mapcount() for compound
pages
mm, numa: skip PTE-mapped THP on numa fault
thp: implement split_huge_pmd()
thp: add option to setup migration entries during PMD split
thp, mm: split_huge_page(): caller need to lock page
thp: reintroduce split_huge_page()
migrate_pages: try to split pages on qeueuing
thp: introduce deferred_split_huge_page()
mm: re-enable THP
thp: update documentation

Documentation/vm/transhuge.txt | 151 ++--
arch/arc/mm/cache.c | 4 +-
arch/arm/include/asm/pgtable-3level.h | 10 -
arch/arm/lib/uaccess_with_memcpy.c | 5 +-
arch/arm/mm/flush.c | 17 +-
arch/arm64/include/asm/pgtable.h | 8 -
arch/arm64/mm/flush.c | 16 -
arch/mips/include/asm/pgtable-bits.h | 6 +-
arch/mips/include/asm/pgtable.h | 18 -
arch/mips/mm/c-r4k.c | 3 +-
arch/mips/mm/cache.c | 2 +-
arch/mips/mm/gup.c | 17 +-
arch/mips/mm/init.c | 6 +-
arch/mips/mm/pgtable-64.c | 14 -
arch/mips/mm/tlbex.c | 1 -
arch/powerpc/include/asm/pgtable-ppc64.h | 25 +-
arch/powerpc/mm/hugepage-hash64.c | 3 -
arch/powerpc/mm/hugetlbpage.c | 17 +-
arch/powerpc/mm/pgtable_64.c | 49 --
arch/powerpc/mm/subpage-prot.c | 2 +-
arch/s390/include/asm/pgtable.h | 15 +-
arch/s390/mm/gup.c | 24 +-
arch/s390/mm/pgtable.c | 16 -
arch/sh/mm/cache-sh4.c | 2 +-
arch/sh/mm/cache.c | 8 +-
arch/sparc/include/asm/pgtable_64.h | 16 -
arch/sparc/mm/fault_64.c | 3 -
arch/sparc/mm/gup.c | 16 +-
arch/tile/include/asm/pgtable.h | 10 -
arch/x86/include/asm/pgtable.h | 9 -
arch/x86/include/asm/pgtable_types.h | 2 -
arch/x86/kernel/vm86_32.c | 6 +-
arch/x86/mm/gup.c | 17 +-
arch/x86/mm/pgtable.c | 14 -
arch/xtensa/mm/tlb.c | 2 +-
fs/proc/page.c | 4 +-
fs/proc/task_mmu.c | 55 +-
include/asm-generic/pgtable.h | 9 -
include/linux/huge_mm.h | 56 +-
include/linux/memcontrol.h | 16 +-
include/linux/mm.h | 116 ++--
include/linux/mm_types.h | 20 +-
include/linux/page-flags.h | 49 +-
include/linux/pagemap.h | 13 +-
include/linux/rmap.h | 16 +-
include/linux/swap.h | 3 +-
include/linux/vm_event_item.h | 4 +-
kernel/events/uprobes.c | 11 +-
kernel/futex.c | 61 +-
mm/debug.c | 8 +-
mm/filemap.c | 10 +-
mm/gup.c | 112 ++-
mm/huge_memory.c | 1100 +++++++++++++++++-------------
mm/hugetlb.c | 10 +-
mm/internal.h | 70 +-
mm/ksm.c | 61 +-
mm/madvise.c | 2 +-
mm/memcontrol.c | 84 +--
mm/memory-failure.c | 8 +-
mm/memory.c | 73 +-
mm/mempolicy.c | 37 +-
mm/migrate.c | 19 +-
mm/mincore.c | 2 +-
mm/mlock.c | 51 +-
mm/mprotect.c | 2 +-
mm/mremap.c | 15 +-
mm/page_alloc.c | 34 +-
mm/pagewalk.c | 2 +-
mm/pgtable-generic.c | 14 -
mm/rmap.c | 163 +++--
mm/shmem.c | 21 +-
mm/swap.c | 273 +-------
mm/swapfile.c | 16 +-
mm/userfaultfd.c | 8 +-
mm/vmstat.c | 4 +-
75 files changed, 1367 insertions(+), 1799 deletions(-)

--
2.5.0

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/