[v3,00/11] Perf improvements for hugetlb and vmalloc on arm64

Message

Ryan Roberts March 4, 2025, 3:04 p.m. UTC

Hi All,

This is v3 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. 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.

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 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 v6.14-rc5, which already contains the fixes from [3]. All
mm selftests run and pass.

NOTE: Its possible that the changes in patch #10 may cause bugs I found in other
archs' lazy mmu implementations to become more likely to trigger. I've fixed all
those bugs in the series at [4], which is now in mm-unstable. But some
coordination when merging this may be required.

[1] https://lore.kernel.org/all/[e-mail]/
[2] https://lore.kernel.org/all/[e-mail]/
[3] https://lore.kernel.org/all/[e-mail]/
[4] https://lore.kernel.org/all/[e-mail]/

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     | 195 ++++++++++++++++++---------
 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          |  72 ++++------
 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, 315 insertions(+), 149 deletions(-)

--
2.43.0

Comments

Uladzislau Rezki March 27, 2025, 12:16 p.m. UTC | #1

On Tue, Mar 04, 2025 at 03:04:30PM +0000, Ryan Roberts wrote:
> Hi All,
> 
> This is v3 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. 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.
> 
I will have a look and review just give me some time :)

--
Uladzislau Rezki

Ryan Roberts March 27, 2025, 1:46 p.m. UTC | #2

On 27/03/2025 08:16, Uladzislau Rezki wrote:
> On Tue, Mar 04, 2025 at 03:04:30PM +0000, Ryan Roberts wrote:
>> Hi All,
>>
>> This is v3 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. 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.
>>
> I will have a look and review just give me some time :)

Thanks for the reviews - appreciate it!

> 
> --
> Uladzislau Rezki