Hi All, This series contains some perf improvements for hugetlb and vmalloc on arm64, and is split out from v1 of a wider series at [1]. Although some of these patches are core-mm, advice from Andrew was to go via the arm64 tree. Hopefully I can get some ACKs from mm folks. 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. (perf results are against v1). 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) - mearements are times; smaller is better +--------------------------------------------------+-------------+-------------+ | Benchmark | | | | Result Class | Apple M2 | Ampere Alta | +==================================================+=============+=============+ | micromm/vmalloc | | | | fix_align_alloc_test: p:1, h:0 (usec) | (I) -12.93% | (I) -7.89% | | fix_size_alloc_test: p:1, h:0 (usec) | (R) 4.00% | 1.40% | | fix_size_alloc_test: p:1, h:1 (usec) | (R) 5.28% | 1.46% | | fix_size_alloc_test: p:2, h:0 (usec) | (I) -3.04% | -1.11% | | fix_size_alloc_test: p:2, h:1 (usec) | -3.24% | -2.86% | | fix_size_alloc_test: p:4, h:0 (usec) | (I) -11.77% | (I) -4.48% | | fix_size_alloc_test: p:4, h:1 (usec) | (I) -9.19% | (I) -4.45% | | fix_size_alloc_test: p:8, h:0 (usec) | (I) -19.79% | (I) -11.63% | | fix_size_alloc_test: p:8, h:1 (usec) | (I) -19.40% | (I) -11.11% | | fix_size_alloc_test: p:16, h:0 (usec) | (I) -24.89% | (I) -15.26% | | fix_size_alloc_test: p:16, h:1 (usec) | (I) -11.61% | (R) 6.00% | | fix_size_alloc_test: p:32, h:0 (usec) | (I) -26.54% | (I) -18.80% | | fix_size_alloc_test: p:32, h:1 (usec) | (I) -15.42% | (R) 5.82% | | fix_size_alloc_test: p:64, h:0 (usec) | (I) -30.25% | (I) -20.80% | | fix_size_alloc_test: p:64, h:1 (usec) | (I) -16.98% | (R) 6.54% | | fix_size_alloc_test: p:128, h:0 (usec) | (I) -32.56% | (I) -21.79% | | fix_size_alloc_test: p:128, h:1 (usec) | (I) -18.39% | (R) 5.91% | | fix_size_alloc_test: p:256, h:0 (usec) | (I) -33.33% | (I) -22.22% | | fix_size_alloc_test: p:256, h:1 (usec) | (I) -18.82% | (R) 5.79% | | fix_size_alloc_test: p:512, h:0 (usec) | (I) -33.27% | (I) -22.23% | | fix_size_alloc_test: p:512, h:1 (usec) | 0.86% | -0.71% | | full_fit_alloc_test: p:1, h:0 (usec) | 2.49% | -0.62% | | kvfree_rcu_1_arg_vmalloc_test: p:1, h:0 (usec) | 1.79% | -1.25% | | kvfree_rcu_2_arg_vmalloc_test: p:1, h:0 (usec) | -0.32% | 0.61% | | long_busy_list_alloc_test: p:1, h:0 (usec) | (I) -31.06% | (I) -19.62% | | pcpu_alloc_test: p:1, h:0 (usec) | 0.06% | 0.47% | | random_size_align_alloc_test: p:1, h:0 (usec) | (I) -14.94% | (I) -8.68% | | random_size_alloc_test: p:1, h:0 (usec) | (I) -30.22% | (I) -19.59% | | vm_map_ram_test: p:1, h:0 (usec) | 2.65% | (R) 7.22% | +--------------------------------------------------+-------------+-------------+ So there are some nice improvements but also some regressions to explain: First fix_size_alloc_test with h:1 and p:16,32,64,128,256 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. Next we have ~4% regression on M2 when vmalloc'ing a single page. (h is irrelevant because a single page is too small for contpte). I assume this is because there is some minor overhead in the barrier deferral mechanism and we are not getting to amortize it over multiple pages here. But I would assume vmalloc'ing 1 page is uncommon because it doesn't buy you anything over kmalloc? 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 Thanks to Anshuman for the review! Applies on top of my fixes series at [2], which applies on top of v6.14-rc3. All mm selftests run and pass. [1] https://lore.kernel.org/all/20250205151003.88959-1-ryan.roberts@xxxxxxx/ [2] https://lore.kernel.org/all/20250217140419.1702389-1-ryan.roberts@xxxxxxx/ Thanks, Ryan Ryan Roberts (14): 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 arm64/mm: Avoid barriers for invalid or userspace mappings 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: Batch arch_sync_kernel_mappings() more efficiently mm: Generalize arch_sync_kernel_mappings() mm: Only call arch_update_kernel_mappings_[begin|end]() for kernel mappings arm64/mm: Batch barriers when updating kernel mappings arch/arm64/include/asm/hugetlb.h | 29 ++-- arch/arm64/include/asm/pgtable.h | 207 +++++++++++++++++++-------- arch/arm64/include/asm/thread_info.h | 1 + arch/arm64/include/asm/vmalloc.h | 46 ++++++ arch/arm64/kernel/process.c | 9 +- arch/arm64/mm/hugetlbpage.c | 72 ++++------ include/linux/page_table_check.h | 30 ++-- include/linux/pgtable.h | 24 +--- include/linux/pgtable_modmask.h | 32 +++++ include/linux/vmalloc.h | 55 +++++++ mm/memory.c | 7 +- mm/page_table_check.c | 34 +++-- mm/vmalloc.c | 93 +++++++----- 13 files changed, 434 insertions(+), 205 deletions(-) create mode 100644 include/linux/pgtable_modmask.h -- 2.43.0
