Skip to content

Perf improvements for hugetlb and vmalloc on arm64

Ryan Roberts linux-arm-kernel linux-mm lkml 3 versions – Applied

Status from this message: Applied to arm64 (for-next/mm), thanks!

Get series with b4
b4 am [email protected]

Versions newest first · linked by author and title

  1. v4 Latest11 patches 6 repliesApplied
  2. v311 patches 25 repliesSuperseded
  3. v214 patches 17 repliesSuperseded

Patches in v4 Applied

#Subject and filesSizeRepliesApplied
1/11arm64: hugetlb: Cleanup huge_pte size discovery mechanismsarch/arm64/mm/hugetlbpage.c+15 -50 replies
2/11arm64: hugetlb: Refine tlb maintenance scopearch/arm64/include/asm/hugetlb.h arch/arm64/mm/hugetlbpage.c+25 -130 replies
3/11mm/page_table_check: Batch-check pmds/puds just like ptesinclude/linux/page_table_check.h mm/page_table_check.c+38 -260 replies
4/11arm64/mm: Refactor __set_ptes() and __ptep_get_and_clear()arch/arm64/include/asm/pgtable.h+73 -411 reply
5/11arm64: hugetlb: Use __set_ptes_anysz() and __ptep_get_and_clear_anysz()arch/arm64/mm/hugetlbpage.c+10 -431 reply
6/11arm64/mm: Hoist barriers out of set_ptes_anysz() looparch/arm64/include/asm/pgtable.h+11 -50 replies
7/11mm/vmalloc: Warn on improper use of vunmap_range()mm/vmalloc.c+6 -20 replies
8/11mm/vmalloc: Gracefully unmap huge ptesinclude/linux/vmalloc.h mm/vmalloc.c+24 -20 replies
9/11arm64/mm: Support huge pte-mapped pages in vmaparch/arm64/include/asm/vmalloc.h arch/arm64/mm/hugetlbpage.c+49 -10 replies
10/11mm/vmalloc: Enter lazy mmu mode while manipulating vmalloc ptesmm/vmalloc.c+14 -00 replies
11/11arm64/mm: Batch barriers when updating kernel mappingsarch/arm64/include/asm/pgtable.h arch/arm64/include/asm/thread_info.h arch/arm64/kernel/process.c+72 -201 reply

Reviews trailers per patch

All 11 patches have a Reviewed-by or Acked-by.

PatchCatalin MarinasAnshuman KhandualPasha TatashinUladzislau Rezki (Sony)Luiz Capitulino
SeriesT-b
1/11R-b R-b
2/11R-b R-b
3/11R-b R-b R-b
4/11R-b R-b
5/11R-b R-b
6/11R-b R-b
7/11R-b R-b R-b
8/11R-b R-b R-b
9/11R-b R-b
10/11R-b R-b R-b
11/11R-b R-b

Trailers given in replies are filled; those carried in the patches from earlier versions are outlined with dashes. Scroll the table sideways for all 5 reviewers.

Read the v4 thread

Cover letter [PATCH v4 00/11] Perf improvements for hugetlb and vmalloc on arm64

Hi All,

This is v4 of a series to improve performance for hugetlb and vmalloc on arm64. Although some of these patches are core-mm, advice from Andrew was to go via the arm64 tree. All patches are now acked/reviewed by relevant maintainers so I believe this should be good-to-go.

The 2 key performance improvements are 1) enabling the use of contpte-mapped blocks in the vmalloc space when appropriate (which reduces TLB pressure). There were already hooks for this (used by powerpc) but they required some tidying and extending for arm64. And 2) batching up barriers when modifying the vmalloc address space for upto 30% reduction in time taken in vmalloc().

vmalloc() performance was measured using the test_vmalloc.ko module. Tested on Apple M2 and Ampere Altra. Each test had loop count set to 500000 and the whole test was repeated 10 times.

legend:
- p: nr_pages (pages to allocate)
- h: use_huge (vmalloc() vs vmalloc_huge())
- (I): statistically significant improvement (95% CI does not overlap)
- (R): statistically significant regression (95% CI does not overlap)
- measurements are times; smaller is better

+--------------------------------------------------+-------------+-------------+
| Benchmark                                        |             |             |
|   Result Class                                   |    Apple M2 | Ampere Alta |
+==================================================+=============+=============+
| micromm/vmalloc                                  |             |             |
|   fix_align_alloc_test: p:1, h:0 (usec)          | (I) -11.53% |      -2.57% |
|   fix_size_alloc_test: p:1, h:0 (usec)           |       2.14% |       1.79% |
|   fix_size_alloc_test: p:4, h:0 (usec)           |  (I) -9.93% |  (I) -4.80% |
|   fix_size_alloc_test: p:16, h:0 (usec)          | (I) -25.07% | (I) -14.24% |
|   fix_size_alloc_test: p:16, h:1 (usec)          | (I) -14.07% |   (R) 7.93% |
|   fix_size_alloc_test: p:64, h:0 (usec)          | (I) -29.43% | (I) -19.30% |
|   fix_size_alloc_test: p:64, h:1 (usec)          | (I) -16.39% |   (R) 6.71% |
|   fix_size_alloc_test: p:256, h:0 (usec)         | (I) -31.46% | (I) -20.60% |
|   fix_size_alloc_test: p:256, h:1 (usec)         | (I) -16.58% |   (R) 6.70% |
|   fix_size_alloc_test: p:512, h:0 (usec)         | (I) -31.96% | (I) -20.04% |
|   fix_size_alloc_test: p:512, h:1 (usec)         |       2.30% |       0.71% |
|   full_fit_alloc_test: p:1, h:0 (usec)           |      -2.94% |       1.77% |
|   kvfree_rcu_1_arg_vmalloc_test: p:1, h:0 (usec) |      -7.75% |       1.71% |
|   kvfree_rcu_2_arg_vmalloc_test: p:1, h:0 (usec) |      -9.07% |   (R) 2.34% |
|   long_busy_list_alloc_test: p:1, h:0 (usec)     | (I) -29.18% | (I) -17.91% |
|   pcpu_alloc_test: p:1, h:0 (usec)               |     -14.71% |      -3.14% |
|   random_size_align_alloc_test: p:1, h:0 (usec)  | (I) -11.08% |  (I) -4.62% |
|   random_size_alloc_test: p:1, h:0 (usec)        | (I) -30.25% | (I) -17.95% |
|   vm_map_ram_test: p:1, h:0 (usec)               |       5.06% |   (R) 6.63% |
+--------------------------------------------------+-------------+-------------+

So there are some nice improvements but also some regressions to explain:

fix_size_alloc_test with h:1 and p:16,64,256 regress by ~6% on Altra. The regression is actually introduced by enabling contpte-mapped 64K blocks in these tests, and that regression is reduced (from about 8% if memory serves) by doing the barrier batching. I don't have a definite conclusion on the root cause, but I've ruled out the differences in the mapping paths in vmalloc. I strongly believe this is likely due to the difference in the allocation path; 64K blocks are not cached per-cpu so we have to go all the way to the buddy. I'm not sure why this doesn't show up on M2 though. Regardless, I'm going to assert that it's better to choose 16x reduction in TLB pressure vs 6% on the vmalloc allocation call duration.

Changes since v3 [3]
====================
- Applied R-bs (thanks all!)
- Renamed set_ptes_anysz() -> __set_ptes_anysz() (Catalin)
- Renamed ptep_get_and_clear_anysz() -> __ptep_get_and_clear_anysz() (Catalin)
- Only set TIF_LAZY_MMU_PENDING if not already set to avoid atomic ops (Catalin)
- Fix commet typos (Anshuman)
- Fix build warnings when PMD is folded (buildbot)
- Reverse xmas tree for variables in __page_table_check_p[mu]ds_set() (Pasha)

Changes since v2 [2]
====================
- Removed the new arch_update_kernel_mappings_[begin|end]() API
- Switches to arch_[enter|leave]_lazy_mmu_mode() instead for barrier batching
- Removed clean up to avoid barriers for invalid or user mappings

Changes since v1 [1]
====================
- Split out the fixes into their own series
- Added Rbs from Anshuman - Thanks!
- Added patch to clean up the methods by which huge_pte size is determined
- Added "#ifndef __PAGETABLE_PMD_FOLDED" around PUD_SIZE in flush_hugetlb_tlb_range()
- Renamed ___set_ptes() -> set_ptes_anysz()
- Renamed ___ptep_get_and_clear() -> ptep_get_and_clear_anysz()
- Fixed typos in commit logs
- Refactored pXd_valid_not_user() for better reuse
- Removed TIF_KMAP_UPDATE_PENDING after concluding that single flag is sufficent
- Concluded the extra isb() in __switch_to() is not required
- Only call arch_update_kernel_mappings_[begin|end]() for kernel mappings

Applies on top of v6.15-rc3. All mm selftests run and no regressions observed.

[1] https://lore.kernel.org/all/[email protected]/
[2] https://lore.kernel.org/all/[email protected]/
[3] https://lore.kernel.org/all/[email protected]/

Thanks,
Ryan

Ryan Roberts (11):
arm64: hugetlb: Cleanup huge_pte size discovery mechanisms
arm64: hugetlb: Refine tlb maintenance scope mm/page_table_check: Batch-check pmds/puds just like ptes arm64/mm: Refactor __set_ptes() and __ptep_get_and_clear()
arm64: hugetlb: Use __set_ptes_anysz() and

    __ptep_get_and_clear_anysz()

arm64/mm: Hoist barriers out of set_ptes_anysz() loop mm/vmalloc: Warn on improper use of vunmap_range() mm/vmalloc: Gracefully unmap huge ptes
arm64/mm: Support huge pte-mapped pages in vmap mm/vmalloc: Enter lazy mmu mode while manipulating vmalloc ptes arm64/mm: Batch barriers when updating kernel mappings

 arch/arm64/include/asm/hugetlb.h     |  29 ++--
 arch/arm64/include/asm/pgtable.h     | 209 +++++++++++++++++++--------
 arch/arm64/include/asm/thread_info.h |   2 +
 arch/arm64/include/asm/vmalloc.h     |  45 ++++++
 arch/arm64/kernel/process.c          |   9 +-
 arch/arm64/mm/hugetlbpage.c          |  73 ++++------
 include/linux/page_table_check.h     |  30 ++--
 include/linux/vmalloc.h              |   8 +
 mm/page_table_check.c                |  34 +++--
 mm/vmalloc.c                         |  40 ++++-
 10 files changed, 329 insertions(+), 150 deletions(-)

--
2.43.0