[v5,0/2] arm64: support batched/deferred tlb shootdown during page reclamation

Message

Yicong Yang Oct. 28, 2022, 8:12 a.m. UTC

From: Yicong Yang <[e-mail]>

Though ARM64 has the hardware to do tlb shootdown, the hardware
broadcasting is not free.
A simplest micro benchmark shows even on snapdragon 888 with only
8 cores, the overhead for ptep_clear_flush is huge even for paging
out one page mapped by only one process:
5.36%  a.out    [kernel.kallsyms]  [k] ptep_clear_flush

While pages are mapped by multiple processes or HW has more CPUs,
the cost should become even higher due to the bad scalability of
tlb shootdown.

The same benchmark can result in 16.99% CPU consumption on ARM64
server with around 100 cores according to Yicong's test on patch
4/4.

This patchset leverages the existing BATCHED_UNMAP_TLB_FLUSH by
1. only send tlbi instructions in the first stage -
	arch_tlbbatch_add_mm()
2. wait for the completion of tlbi by dsb while doing tlbbatch
	sync in arch_tlbbatch_flush()
Testing on snapdragon shows the overhead of ptep_clear_flush
is removed by the patchset. The micro benchmark becomes 5% faster
even for one page mapped by single process on snapdragon 888.

With this support we're possible to do more optimization for memory
reclamation and migration[*].

[*] https://lore.kernel.org/lkml/[e-mail]/

-v5:
1. Make ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH depends on EXPERT for this stage on arm64.
2. Make a threshhold of CPU numbers for enabling batched TLP flush on arm64
Link: https://lore.kernel.org/linux-arm-kernel/[e-mail]/T/

-v4:
1. Add tags from Kefeng and Anshuman, Thanks.
2. Limit the TLB batch/defer on systems with >4 CPUs, per Anshuman
3. Merge previous Patch 1,2-3 into one, per Anshuman
Link: https://lore.kernel.org/linux-mm/[e-mail]/

-v3:
1. Declare arch's tlbbatch defer support by arch_tlbbatch_should_defer() instead
   of ARCH_HAS_MM_CPUMASK, per Barry and Kefeng
2. Add Tested-by from Xin Hao
Link: https://lore.kernel.org/linux-mm/[e-mail]/

-v2:
1. Collected Yicong's test result on kunpeng920 ARM64 server;
2. Removed the redundant vma parameter in arch_tlbbatch_add_mm()
   according to the comments of Peter Zijlstra and Dave Hansen
3. Added ARCH_HAS_MM_CPUMASK rather than checking if mm_cpumask
   is empty according to the comments of Nadav Amit

Thanks, Peter, Dave and Nadav for your testing or reviewing
, and comments.

-v1:
https://lore.kernel.org/lkml/[e-mail]/

Anshuman Khandual (1):
  mm/tlbbatch: Introduce arch_tlbbatch_should_defer()

Barry Song (1):
  arm64: support batched/deferred tlb shootdown during page reclamation

 .../features/vm/TLB/arch-support.txt          |  2 +-
 arch/arm64/Kconfig                            |  6 +++
 arch/arm64/include/asm/tlbbatch.h             | 12 +++++
 arch/arm64/include/asm/tlbflush.h             | 46 ++++++++++++++++++-
 arch/x86/include/asm/tlbflush.h               | 15 +++++-
 mm/rmap.c                                     | 19 +++-----
 6 files changed, 84 insertions(+), 16 deletions(-)
 create mode 100644 arch/arm64/include/asm/tlbbatch.h

Comments

Punit Agrawal Nov. 11, 2022, 10:17 a.m. UTC | #1

Yicong Yang <[e-mail]> writes:

> From: Yicong Yang <[e-mail]>
>
> Though ARM64 has the hardware to do tlb shootdown, the hardware
> broadcasting is not free.
> A simplest micro benchmark shows even on snapdragon 888 with only
> 8 cores, the overhead for ptep_clear_flush is huge even for paging
> out one page mapped by only one process:
> 5.36%  a.out    [kernel.kallsyms]  [k] ptep_clear_flush
>
> While pages are mapped by multiple processes or HW has more CPUs,
> the cost should become even higher due to the bad scalability of
> tlb shootdown.
>
> The same benchmark can result in 16.99% CPU consumption on ARM64
> server with around 100 cores according to Yicong's test on patch
> 4/4.
>
> This patchset leverages the existing BATCHED_UNMAP_TLB_FLUSH by
> 1. only send tlbi instructions in the first stage -
> 	arch_tlbbatch_add_mm()
> 2. wait for the completion of tlbi by dsb while doing tlbbatch
> 	sync in arch_tlbbatch_flush()
> Testing on snapdragon shows the overhead of ptep_clear_flush
> is removed by the patchset. The micro benchmark becomes 5% faster
> even for one page mapped by single process on snapdragon 888.
>
> With this support we're possible to do more optimization for memory
> reclamation and migration[*].

I applied the patches on v6.1-rc4 and was able to see the drop in
ptep_clear_flush() in the perf report when running the test program from
Patch 2. The tests were done on a rk3399 based system with benefits
visible when running the tests on either of the clusters. 

So, for the series,

Tested-by: Punit Agrawal <[e-mail]>

Thanks,
Punit

[...]