Hi Linus, There is 1 bigger change to percpu_counter's api allowing for init and destroy of multiple counters via percpu_counter_init_many() and percpu_counter_destroy_many(). This is used to help begin remediating a performance regression with percpu rss stats. Additionally, it seems larger core count machines are feeling the burden of the single threaded allocation of percpu. Mateusz is thinking about it and I will spend some time on it too. Thanks, Dennis The following changes since commit 5d0c230f1de8c7515b6567d9afba1f196fb4e2f4: Linux 6.5-rc4 (2023-07-30 13:23:47 -0700) are available in the Git repository at: git://git.kernel.org/pub/scm/linux/kernel/git/dennis/percpu.git tags/percpu-for-6.6 for you to fetch changes up to 14ef95be6f5558fb9e43aaf06ef9a1d6e0cae6c8: kernel/fork: group allocation/free of per-cpu counters for mm struct (2023-08-25 08:10:35 -0700) ---------------------------------------------------------------- percpu: changes for v6.6 percpu * A couple cleanups by Baoquan He and Bibo Mao. The only behavior change is to start printing messages if we're under the warn limit for failed atomic allocations. percpu_counter * Shakeel introduced percpu counters into mm_struct which caused percpu allocations be on the hot path [1]. Originally I spent some time trying to improve the percpu allocator, but instead preferred what Mateusz Guzik proposed grouping at the allocation site, percpu_counter_init_many(). This allows a single percpu allocation to be shared by the counters. I like this approach because it creates a shared lifetime by the allocations. Additionally, I believe many inits have higher level synchronization requirements, like percpu_counter does against HOTPLUG_CPU. Therefore we can group these optimizations together. [1] https://lore.kernel.org/linux-mm/20221024052841.3291983-1-shakeelb@xxxxxxxxxx/ ---------------------------------------------------------------- Baoquan He (3): mm/percpu.c: remove redundant check mm/percpu.c: optimize the code in pcpu_setup_first_chunk() a little bit mm/percpu.c: print error message too if atomic alloc failed Bibo Mao (1): mm/percpu: Remove some local variables in pcpu_populate_pte Mateusz Guzik (2): pcpcntr: add group allocation/free kernel/fork: group allocation/free of per-cpu counters for mm struct include/linux/percpu_counter.h | 41 ++++++++++++++++++++----- kernel/fork.c | 15 +++------ lib/percpu_counter.c | 62 +++++++++++++++++++++++++------------ mm/percpu.c | 69 +++++++++++++++++------------------------- 4 files changed, 109 insertions(+), 78 deletions(-)
