mbox series

[RFC,/,PoC,v1,0/1] powerpc: Add support for batched unmap TLB flush

Message ID cover.1727001426.git.ritesh.list@gmail.com (mailing list archive)
Headers show
Series powerpc: Add support for batched unmap TLB flush | expand

Message

Ritesh Harjani (IBM) Sept. 22, 2024, 12:46 p.m. UTC 
Hello All,

This is a quick PoC to add ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH support to
powerpc for book3s64. The ISA in 6.10 of "Translation Table Update
Synchronization Requirements" says that the architecture allows for optimizing
the translation cache invalidation by doing it in bulk later after the PTE
change has been done.
That means if we can add ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH support, it will be
possible to utilize optimizations in reclaim and migrate pages path which can
defer the tlb invalidations to be done in bulk after all the page unmap
operations has been completed.

This a quick PoC for the same. Note that this may not be a complete patch yet,
TLB on Power is already complex from the hardware side :) and then many
optimizations done in the software (e.g. exit_lazy_flush_tlb to avoid tlbies).
But since the current patch looked somewhat sane to me, I wanted to share to get
an early feedback from people who are well versed with this side of code.

Meanwhile I have many TODOs to look into which I am working in parallel for this
work. Later will also get some benchmarks w.r.t promotion / demotion.

I ran a micro-benchmark which was shared in other commits that adds this
support on other archs. I can see some good initial improvements.

without patch (perf report showing 7% in radix__flush_tlb_page_psize, even with
single thread)
==================
root# time ./a.out
real    0m23.538s
user    0m0.191s
sys     0m5.270s

# Overhead  Command  Shared Object               Symbol
# ........  .......  ..........................  .............................................
#
     7.19%  a.out    [kernel.vmlinux]            [k] radix__flush_tlb_page_psize
     5.63%  a.out    [kernel.vmlinux]            [k] _raw_spin_lock
     3.21%  a.out    a.out                       [.] main
     2.93%  a.out    [kernel.vmlinux]            [k] page_counter_cancel
     2.58%  a.out    [kernel.vmlinux]            [k] page_counter_try_charge
     2.56%  a.out    [kernel.vmlinux]            [k] _raw_spin_lock_irq
     2.30%  a.out    [kernel.vmlinux]            [k] try_to_unmap_one

with patch
============
root# time ./a.out
real    0m8.593s
user    0m0.064s
sys     0m1.610s

# Overhead  Command  Shared Object               Symbol
# ........  .......  ..........................  .............................................
#
     5.10%  a.out    [kernel.vmlinux]            [k] _raw_spin_lock
     3.55%  a.out    [kernel.vmlinux]            [k] __mod_memcg_lruvec_state
     3.13%  a.out    a.out                       [.] main
     3.00%  a.out    [kernel.vmlinux]            [k] page_counter_try_charge
     2.62%  a.out    [kernel.vmlinux]            [k] _raw_spin_lock_irq
     2.58%  a.out    [kernel.vmlinux]            [k] page_counter_cancel
     2.22%  a.out    [kernel.vmlinux]            [k] try_to_unmap_one


<micro-benchmark>
====================
#define PAGESIZE 65536
#define SIZE (1 * 1024 * 1024 * 10)
int main()
{
        volatile unsigned char *p = mmap(NULL, SIZE, PROT_READ | PROT_WRITE,
                                         MAP_SHARED | MAP_ANONYMOUS, -1, 0);

        memset(p, 0x88, SIZE);

        for (int k = 0; k < 10000; k++) {
                /* swap in */
                for (int i = 0; i < SIZE; i += PAGESIZE) {
                        (void)p[i];
                }

                /* swap out */
                madvise(p, SIZE, MADV_PAGEOUT);
        }
}



Ritesh Harjani (IBM) (1):
  powerpc: Add support for batched unmap TLB flush

 arch/powerpc/Kconfig                          |  1 +
 arch/powerpc/include/asm/book3s/64/tlbflush.h |  5 +++
 arch/powerpc/include/asm/tlbbatch.h           | 14 ++++++++
 arch/powerpc/mm/book3s64/radix_tlb.c          | 32 +++++++++++++++++++
 4 files changed, 52 insertions(+)
 create mode 100644 arch/powerpc/include/asm/tlbbatch.h

--
2.46.0