Re: [RFC v2 PATCH 0/16] Optimize this_cpu_*() ops for non-x86 (ARM64 for this series)
From: Yang Shi
Date: Wed Sep 02 2026 - 17:17:49 EST
On 9/2/26 4:51 AM, Usama Anjum wrote:
On 15/07/2026 7:04 pm, Yang Shi wrote:
Hi,tl;dr
This is v2 RFC. In v2 a lot problems found out by Sashiko were fixed and more
feature gaps were closed (please see the below changelog for the details).
Although there are still some open issues, for example, it just can support
48 bits VA (for 4K and 64K) and 47 bits VA (for 16K), KPTI support has not
been solved yet, etc, but I think the delta should be big enough and worth
a new RFC to gather comments in order to make sure I'm on the right track.
Some more benchmarks were done, for example, some latency related benchmarks
that I mentioned at LSFMM because I thought responsiveness should be improved
due to the removal of preempt_disable. Collected more PMU counters as well.
Please refer to the benchmark section for more details.
Look forward to comments.
Changelog
v2: * Added 3-level and 2-level page table support.
* Tested with 16K and 64K page size. But we just support 48 bits VA (4K
and 64K) and 47 bits VA with 16K for now. Please refer to the below
"known issue" section for the detail reason.
* Added support for memory hotplug.
* Added support for KASAN (generic).
* Treated percpu and local percpu area address as vmalloc address.
* Fixed build failure for x86.
* Fixed build failure for !CONFIG_NUMA.
* Added KASAN support for local percpu area.
* Some other misc bug fixes found out by Sashiko.
* More code refactor and cleanup.
* Regorganized the patches.
* More benchmarks, refer to benchmark section for more details.
* Rebased to v7.2-rc1.
Introduction
============
This patch series implemented the LSFMM 2026 proposal for optimizing
this_cpu_*() ops on ARM64. For the details of the proposal, Please refer to:
https://lore.kernel.org/linux-mm/CAHbLzkpcN-T8MH6=W3jCxcFj1gVZp8fRqe231yzZT-rV_E_org@xxxxxxxxxxxxxx/
I didn't repeat it in the cover letter because there is no change to the
proposal.
The series is based on 7.1-rc1. It is basically minimum viable patches.
There are still a few hacks in this series and it may break something,
for example, KPTI, SMT machines which shared TLB, etc. But it shoule be
good enough for now to demonstrate the core idea. The main purpose of the
RFC is to gather feedback, figure out missing parts and risks, and make sure
we are on the right track, as well as hopefully it can help the discussion
for the upcoming LSFMM.
I broke the patches down to arch-dependent and arch-independent parts so that
hopefully the interested persons can do experiments on other architectures,
for example, S390, easier.
A new kernel config is introduced, HAVE_LOCAL_PER_CPU_MAP. The architectures
which can support this feature will select it. Allocating and freeing percpu
local mapping is protected by this config so that others won't pay the cost.
Known Issues
============
1. KPTI
-------
We need determine what CPU we are on, then switch to the right page table.
Currently arm64 kernel fetches tramp_pg_dir via swapper_pg_dir - fixed_offset,
and fetches swapper_pg_dir from ttbr1. But ttbr1 may not hold swapper_pg_dir
anymore except CPU #0. So we need to figure out the other way to handle it.
Switching to tramp_pg_dir should be easy, but the reverse seems harder because
tramp_pg_dir just maps the trampoline vectors.
Maybe we can do two steps switch. Switch to swapper_pg_dir at the first step,
then switch to per cpu page table (for entry) or tramp page table (for exit).
Nobody should call this_cpu_*() at either userspace -> kernel entry stage or
kernel -> userspace exit stage.
2. SW PAN
---------
Has the similar issue as KPTI. It installs reserved_pg_dir to TTBR0 when running
in kernel space, but fetching reserved_pg_dir via swapper_pg_dir - fixed_offset.
Maybe we can save the physical address of swapper_pg_dir in a variable, then load
it from that variable instead of ttbr1.
3. Shared TLB machines
----------------------
Some machines may share TLB between CPUs, for example, SMT machines may share
TLB between the two hardware threads in one core.
The per cpu page table just can't work with it. Maybe we need a new
cpufeature to indicate whether per cpu page table is allowed or not. Then
just enable it for not-shared-TLB machines.
4. Don't support all VA bits
----------------------------
We just support 48 bits VA (4K and 64K) and 47 bits VA (16K) for now. For 4K
and 64K, supporting other VA bits is not hard, we just need to determine the
size for percpu and local percpu area.
But it is harder for supporting 48 bits VA + 16K page size. We just have two
top level kernel page table entries with this configuration, but we assume we
just need to sync up kernel page table at the top level for now. We need to
sync up kernel page table at the second level in order to support it. I'm not
sure whether it is worth it or not.
Benchmark
=========
The benchmarks are done on 160 core AmpereOne machine. The baseline is
v7.2-rc1 kernel.
1. Reduction of kernel text size
--------------------------------
The patchset can reduce at least 11 instructions for this_cpu_*() ops. Both
preempt_disable() and preempt_enable() need 4 instructions to manipulate
the preempt count, and preempt_enable() needs more instructions (compare +
READ + compare) to determine whether reschedule is needed or not.
Because this_cpu_*() ops are inlined and called in a lot of places so we
can save a lot of instructions.
The size of kernel text is reduced by ~184KB with default Fedora kernel
config. This also helps reduce kernel icache miss rate and stalled frontend
cycles as kernel build benchmark result showed.
2. Kernel Build
---------------
Run kernel build (make -j160) with the default Fedora kernel config in a
memcg.
13% - 18% sys time improvment
3% - 7% wall time improvement
5% fewer kernel icache miss, 5% fewer executed kernel instructions and
15% fewer stalled frontend cycles for kernel.
3. stress-ng vm ops
-------------------
stress-ng --vm 160 --vm-bytes 128M --vm-ops 100000000
8.5% improvement
4. stress-ng vm ops + fork
--------------------------
stress-ng --mmapfork 160 --mmapfork-bytes 128M --mmapfork-ops 500
15% improvement
5. Specjbb
----------
The specjbb test latency curves showed the patched kernel has consistently
lower p99 latency (the lower the better) than the baseline.
2.5% improvement on max-jOPS and 4% - 5% improvement on critical-jOPS.
The specjbb benchmark is quite sensitive to latency and responsiveness,
particularly critical-jOPS result. The patches are supposed to improve the
responsiveness due to the reduction of preempt-disabled critical sections.
6. MySQL
--------
1% - 2% gains on read-only test, 2% - 4% gains on write-only test. Also see
15% decrease on frontend cache stall.
Comparing GPR approach [1] with this series gives 8 improvements and 3 regressions.
Fastpath is a Linux kernel performance benchmarking service. The table
compares the both patched kernels: positive values are faster, negative
values are slower, and (I)/(R) indicate statistically significant
improvements/regressions. Unmarked differences are not significant after
accounting for confidence and noise thresholds.
+---------------------------------+--------------------+-----------------+-----------------------+
| Benchmark | percpu-pgtable [2] | gpr-fixup [3] | gpr-fixup [3] |
| | vs baseline [1] | vs baseline [1] | vs percpu-pgtable [2] |
+=================================+====================+=================+=======================+
| lmbench/lat-mem-rd | -0.26% | (I) 1.24% | (I) 1.50% |
| micromm/fork | (I) 1.10% | (I) 2.46% | (I) 1.35% |
| micromm/munmap | (I) 18.10% | (I) 19.78% | (I) 1.43% |
| micromm/vmalloc | (I) 11.92% | (I) 14.68% | (I) 2.47% |
| mmtests/hackbench | 0.27% | (I) 1.09% | 0.81% |
| mmtests/kernbench | (I) 1.19% | (I) 1.18% | -0.01% |
| mmtests/sysbench-cpu | 0.01% | 0.06% | 0.05% |
| mmtests/sysbench-mutex | -0.85% | -0.14% | 0.72% |
| mmtests/sysbench-thread | (R) -4.49% | (I) 3.02% | (I) 7.86% |
| perf/futex | 0.55% | (R) -2.80% | (R) -3.34% |
| perf/sched | (I) 1.61% | -0.39% | (R) -1.97% |
| perf/syscall | (I) 1.54% | 0.98% | -0.55% |
| pts/memtier-benchmark | (I) 2.37% | (I) 2.51% | 0.14% |
| pts/nginx | 0.42% | 0.96% | 0.54% |
| pts/perl-benchmark | (I) 1.83% | (I) 1.63% | -0.20% |
| pts/pgbench | -0.09% | 0.55% | 0.65% |
| pts/pybench | -0.05% | -0.07% | -0.02% |
| pts/redis | -0.00% | -0.04% | -0.04% |
| pts/sqlite-speedtest | 0.15% | 0.44% | 0.28% |
| repro-collection/mysql-workload | 0.73% | 0.25% | -0.48% |
| schbench/thread-contention | 0.42% | (I) 1.04% | 0.61% |
| sockperf/echo-lat-tcp | (I) 2.88% | (I) 1.69% | (R) -1.16% |
| sockperf/echo-lat-udp | (I) 3.93% | (I) 5.11% | (I) 1.14% |
| sockperf/packet-tp-tcp | -0.41% | (I) 1.37% | (I) 1.79% |
| sockperf/packet-tp-udp | (I) 1.45% | (I) 1.49% | 0.04% |
| specjbb/composite | (I) 1.39% | (I) 1.65% | 0.26% |
| speedometer/v2.0 | (R) -1.04% | (R) -1.04% | 0.00% |
| speedometer/v2.1 | -0.10% | 0.10% | 0.20% |
| syscall/getpid | -0.98% | 0.52% | (I) 1.51% |
| syscall/getppid | -0.43% | 0.14% | 0.57% |
| syscall/invalid | (I) 3.06% | (I) 3.56% | 0.48% |
+---------------------------------+--------------------+-----------------+-----------------------+
[1] v7.2
[2] v7.2-percpu-pgtable (this series)
[3] v7.2-percpu-gpr [1] (Based on email review, we fixed SDEI to restore x26
instead of corrupting x22, and adjusted this_cpu_write() helpers to
avoid GCC overflow warnings.)
The constraints/workarounds for the per-CPU page-table series were supplied as
this Kconfig fragment for all the different runs.
CONFIG_EXPERT=y
CONFIG_ARM64_4K_PAGES=y
CONFIG_ARM64_VA_BITS_48=y
CONFIG_ARM64_VA_BITS_39=n
CONFIG_ARM64_VA_BITS_52=n
CONFIG_ARM64_PA_BITS_48=y
CONFIG_ARM64_PA_BITS_52=n
CONFIG_UNMAP_KERNEL_AT_EL0=n
CONFIG_ARM64_SW_TTBR0_PAN=n
CONFIG_KASAN=n
[1]https://lore.kernel.org/all/20260804170503.3513916-19-mark.rutland@xxxxxxx/
Hi Usama,
Thank you for running the benchmarks. I'm done some benchmarks and wrapped up the results too. I did run some same benchmarks as yours, like perf bench and kernel build. I also did some profiling for some benchmarks and did some investigation about where the performance difference came from. This is why the work took longer than I expected in the first place.
Test Environment
===========
Hardware
--------
1P AmpereOne machine with 192 cores, 512GB memory.
Kernel
------
v7.2 based.
I applied the first 6 patches from Mark's v2 series, which are some common micro optimization IIRC, on top of my percpu page table series. I'm not sure whether they can make measurable difference or not, but I think I'd better do it because we were benchmarking some subtle difference (4 instructions difference in some hot path) so we can have some apple to apple comparison.
4K page size + 48 VA bits.
Result Analysis
-------------
I use ministat (https://github.com/leahneukirchen/ministat), which is a small tool to do the statistics legwork on benchmarks, to process the benchmark results. It should help us rule out the misleading noises.
The ministat output looks like:
N Min Max Median Avg Stddev
x 10 46998 48366 48192 47986.333 499.14714
+ 10 48438 49333 49030 48952.333 304.42711
Difference at 95.0% confidence
966 +/- 531.791
2.01307% +/- 1.10821%
(Student's t, pooled s = 413.415)
I use the percpu page table series as baseline, which is shown as "x" row in ministat output. The "+" row is Mark's series. The column "N" means the number of samples (basically equals to the number of run iterations). Other columns are quite self-explained. But I will just show the percentage reported by ministat in this email.
Benchmarks
========
I have done some different benchmarks which cover from macro to micro benchmarks and different type of applications (kernel build, Java, database, networking, etc). They are:
- Kernel text size (just compare the kernel text size for the two kernels)
- Kernel build
- Specjbb
- MySQL (sysbench)
- Memcached (via Mellanox NIC)
- iperf (localhost)
- perf bench
- will-it-scale (page fault and mmap)
Because we are benchmarking some subtle code difference, so all the benchmarks were run multiple times (the number of iterations depend on benchmarks duration and variance) and a couple of reboots in order to rule out reboot by reboot spread (I did see some significant reboot by reboot spread for some test cases). I also bind the tests to some specific CPUs if they don't use all CPUs in order to minimize the core-to-core latency variance and scheduler interference.
Basically I didn't see measurable difference for the most benchmarks listed above (ministat showed no difference), I think this aligned with your results too. But some benchmarks showed regressions for Mark's series. Please see the below for the details (just the regressed benchmarks).
Kernel text size
-------------
With the same kernel configuration (the default Fedora config), the kernel text size for the percpu page table series is 92K smaller than Mark's series. This is kind of expected because all this_cpu ops are basically inlined and are called quite frequently, so even small difference could make big impact.
Kernel build
----------
Ran kernel build with "make -j192" in a memcg with cold page cache. "Drop cache" is called for each run before building kernel.
I did 5 runs. The percpu page table series has roughly 2% better sys time (the smaller the better).
Difference at 95.0% confidence
966 +/- 531.791
2.01307% +/- 1.10821%
(Student's t, pooled s = 413.415)
And slightly better wall time.
Difference at 95.0% confidence
2.2 +/- 1.92934
0.553877% +/- 0.485735%
(Student's t, pooled s = 1.32288)
perf sched pipe
-------------
Bind two threads to CPU 78 and 79 (randomly picked) in order to avoid core-2-core latency variance and some scheduler overhead. I did 200 runs. The percpu page table series is better. I didn't see measurable difference with process mode.
Throughput (The bigger the better)
Difference at 95.0% confidence
-24549.2 +/- 4900.32
-9.73769% +/- 1.78314%
(Student's t, pooled s = 25001.6)
Latency (The smaller the better)
Difference at 95.0% confidence
0.377888 +/- 0.0753446
9.37686% +/- 2.01494%
(Student's t, pooled s = 0.384411)
perf futex wake-parallel (wake up one thread from 192 waiters)
------------------------------------------------------
The result shows the wake up latency. I did 1000 runs. It is the smaller the better.
Difference at 95.0% confidence
0.0009767 +/- 0.000134466
17.0749% +/- 2.53593%
(Student's t, pooled s = 0.00153406)
The histograms showed obvious long tail latency from Mark's series for both perf bench pipe and perf futex wake-parallel.
I also did some profiling for the above two benchmarks. The profiling showed the percpu page table series has lower overhead in schedule(). I got the profiling with pseudo NMI which has extra overhead, it may distort the profiling a little bit. The profiling showed finish_task_switch() as the hot spot. But it actually doesn't do too many things other than enabling IRQ. This is a typical misleading profiling due to not supported NMI on ARM64. But anyway both profiling results (w/ pseudo NMI and w/o it) show the overhead is in schedule().
perf bench fork
-------------
The result is the smaller the better. I did 20 runs. And the profiling showed the percpu page table series spent less time in __percpu_add_case_64().
Difference at 95.0% confidence
48.5027 +/- 4.16055
1.93944% +/- 0.167849%
(Student's t, pooled s = 6.50039)
will-it-scale/page fault
--------------------
All the page fault micro benchmarks from will-it-scale were run with one single thread and bound to the same CPU in order to minimize the overhead from page allocator and all the VM locks (for example, lru lock, zone lock, etc). I did 100 runs for each test case. The result is the bigger the better.
page_fault1 (Anon page fault)
-------------------------
The percpu page table series is slightly better.
Difference at 95.0% confidence
-13391.9 +/- 944.563
-1.67893% +/- 0.117849%
(Student's t, pooled s = 3407.69)
page fault2 (CoW page fault)
------------------------
The percpu page table series is slightly better.
Difference at 95.0% confidence
-4736.52 +/- 438.185
-1.16333% +/- 0.10686%
(Student's t, pooled s = 1580.83)
page fault3 (tmpfs page fault)
-------------------------
The percpu page table series is slightly better.
Difference at 95.0% confidence
-21667.9 +/- 1352.51
-1.91305% +/- 0.1186%
(Student's t, pooled s = 4879.44)
The profiling for all the page fault benchmarks showed the percpu page table series spent less time in __percpu_add_case_64().
will-it-scale/mmap
----------------
Like page fault tests, the mmap benchmarks were run with one single thread and bound to the same CPU in order to minimize the overhead from VM locks. I did 50 runs for each. The results are the bigger the better.
mmap1
------
The percpu page table series is better.
Difference at 95.0% confidence
-15916.4 +/- 6094.58
-3.53615% +/- 1.32214%
(Student's t, pooled s = 15509.2)
mmap2
------
The percpu page table series is better.
Difference at 95.0% confidence
-14662.4 +/- 9359.95
-3.91005% +/- 2.44462%
(Student's t, pooled s = 23588.6)
Thanks,
Yang
Regression test
===============
1. memcg creation
-----------------
Create 10K memcgs. Each memcg creation needs to allocate multiple percpu
variables, for example, percpu refcnt, rstat and objcg percpu refcnt.
Consumed 2112K more virtual memory for percpu “local mapping” and a few
more mega bytes consumed by per cpu page tables.
No noticeable regression was found for elapsed time.
2. fork test
------------
stress-ng --fork 160 --fork-ops 10000000
fork() needs to allocate multiple percpu variables, for example, rss
counters and mm_cid_cpu.
Roughly 1% regression was found. However stress-ng fork test has quites
small address space, the real life workloads typically have much larger
address space and do more complicated works. The stress-ng mmapfork
benchmark saw 15% improvement.
The organization of patches
===========================
The refactor and prepatory patches (patch 1 - patch 4)
Percpu page table support patches (patch 5 - patch 8)
Local percpu area support patches (patch 7 - patch 15)
Use local percpu area for this_cpu ops (patch 16)
Yang Shi (16):
drivers: arch_numa: move percpu set up code to arch
arm64: kconfig: make percpu related configs not depend on NUMA
mm: pgalloc: introduce {pud|pmd}_populate_sync()
vmalloc: pass in pgd pointer for vmap{__vunmap}_range_noflush()
arm64: mm: enable percpu kernel page table
arm64: mm: defined {pud|pmd}_populate_sync()
arm64: mm: sync percpu page table for memory hotplug/unplug
arm64: kasan: sync up kasan shadow area page table
arm64: mm: define percpu virtual space area
mm: percpu: prepare to use dedicated percpu area
arm64: mm: map local percpu first chunk
mm: percpu: set up first chunk and reserve chunk
arm64: mm: introduce __per_cpu_local_off
mm: percpu: allocate and free local percpu vm area
arm64: kconfig: select HAVE_LOCAL_PER_CPU_MAP
arm64: percpu: use local percpu for this_cpu_*() APIs
arch/arm64/Kconfig | 12 +++++++---
arch/arm64/include/asm/mmu.h | 5 ++++
arch/arm64/include/asm/mmu_context.h | 9 +++++++-
arch/arm64/include/asm/percpu.h | 37 ++++++++++++++++++++++++++++-
arch/arm64/include/asm/pgalloc.h | 24 +++++++++++++++++++
arch/arm64/include/asm/pgtable.h | 37 ++++++++++++++++++++++++++---
arch/arm64/kernel/setup.c | 3 +++
arch/arm64/kernel/smp.c | 44 +++++++++++++++++++++++++++++++++++
arch/arm64/mm/kasan_init.c | 47 +++++++++++++++++++++++--------------
arch/arm64/mm/mmu.c | 165 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++----------------------
arch/arm64/mm/ptdump.c | 4 ++++
arch/riscv/kernel/smp.c | 51 ++++++++++++++++++++++++++++++++++++++++
drivers/base/arch_numa.c | 51 +---------------------------------------
include/linux/mm.h | 11 +++++++++
include/linux/percpu.h | 4 +++-
include/linux/pgalloc.h | 13 +++++++++++
include/linux/vmalloc.h | 3 +++
mm/Kconfig | 9 ++++++++
mm/internal.h | 5 +++-
mm/kmsan/hooks.c | 14 +++++------
mm/percpu-internal.h | 14 +++++++++++
mm/percpu-vm.c | 94 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
mm/percpu.c | 58 +++++++++++++++++++++++++++++++++++++---------
mm/sparse-vmemmap.c | 4 ++--
mm/vmalloc.c | 138 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++---------------
25 files changed, 712 insertions(+), 144 deletions(-)
Thanks,
Yang