On Thu, Sep 15, 2022 at 6:07 PM Anshuman Khandual
<anshuman.khandual@xxxxxxx> wrote:
>
>
>
> On 9/9/22 11:05, Barry Song wrote:
> > On Fri, Sep 9, 2022 at 5:24 PM Anshuman Khandual
> > <anshuman.khandual@xxxxxxx> wrote:
> >>
> >>
> >>
> >> On 8/22/22 13:51, Yicong Yang wrote:
> >>> From: Barry Song <v-songbaohua@xxxxxxxx>
> >>>
> >>> on x86, batched and deferred tlb shootdown has lead to 90%
> >>> performance increase on tlb shootdown. on arm64, HW can do
> >>> tlb shootdown without software IPI. But sync tlbi is still
> >>> quite expensive.
> >>>
> >>> Even running a simplest program which requires swapout can
> >>> prove this is true,
> >>> #include <sys/types.h>
> >>> #include <unistd.h>
> >>> #include <sys/mman.h>
> >>> #include <string.h>
> >>>
> >>> int main()
> >>> {
> >>> #define SIZE (1 * 1024 * 1024)
> >>> 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 += 4096) {
> >>> (void)p[i];
> >>> }
> >>>
> >>> /* swap out */
> >>> madvise(p, SIZE, MADV_PAGEOUT);
> >>> }
> >>> }
> >>>
> >>> Perf result on snapdragon 888 with 8 cores by using zRAM
> >>> as the swap block device.
> >>>
> >>> ~ # perf record taskset -c 4 ./a.out
> >>> [ perf record: Woken up 10 times to write data ]
> >>> [ perf record: Captured and wrote 2.297 MB perf.data (60084 samples) ]
> >>> ~ # perf report
> >>> # To display the perf.data header info, please use --header/--header-only options.
> >>> # To display the perf.data header info, please use --header/--header-only options.
> >>> #
> >>> #
> >>> # Total Lost Samples: 0
> >>> #
> >>> # Samples: 60K of event 'cycles'
> >>> # Event count (approx.): 35706225414
> >>> #
> >>> # Overhead Command Shared Object Symbol
> >>> # ........ ....... ................. .............................................................................
> >>> #
> >>> 21.07% a.out [kernel.kallsyms] [k] _raw_spin_unlock_irq
> >>> 8.23% a.out [kernel.kallsyms] [k] _raw_spin_unlock_irqrestore
> >>> 6.67% a.out [kernel.kallsyms] [k] filemap_map_pages
> >>> 6.16% a.out [kernel.kallsyms] [k] __zram_bvec_write
> >>> 5.36% a.out [kernel.kallsyms] [k] ptep_clear_flush
> >>> 3.71% a.out [kernel.kallsyms] [k] _raw_spin_lock
> >>> 3.49% a.out [kernel.kallsyms] [k] memset64
> >>> 1.63% a.out [kernel.kallsyms] [k] clear_page
> >>> 1.42% a.out [kernel.kallsyms] [k] _raw_spin_unlock
> >>> 1.26% a.out [kernel.kallsyms] [k] mod_zone_state.llvm.8525150236079521930
> >>> 1.23% a.out [kernel.kallsyms] [k] xas_load
> >>> 1.15% a.out [kernel.kallsyms] [k] zram_slot_lock
> >>>
> >>> ptep_clear_flush() takes 5.36% CPU in the micro-benchmark
> >>> swapping in/out a page mapped by only one process. If the
> >>> page is mapped by multiple processes, typically, like more
> >>> than 100 on a phone, the overhead would be much higher as
> >>> we have to run tlb flush 100 times for one single page.
> >>> Plus, tlb flush overhead will increase with the number
> >>> of CPU cores due to the bad scalability of tlb shootdown
> >>> in HW, so those ARM64 servers should expect much higher
> >>> overhead.
> >>>
> >>> Further perf annonate shows 95% cpu time of ptep_clear_flush
> >>> is actually used by the final dsb() to wait for the completion
> >>> of tlb flush. This provides us a very good chance to leverage
> >>> the existing batched tlb in kernel. The minimum modification
> >>> is that we only send async tlbi in the first stage and we send
> >>> dsb while we have to sync in the second stage.
> >>>
> >>> With the above simplest micro benchmark, collapsed time to
> >>> finish the program decreases around 5%.
> >>>
> >>> Typical collapsed time w/o patch:
> >>> ~ # time taskset -c 4 ./a.out
> >>> 0.21user 14.34system 0:14.69elapsed
> >>> w/ patch:
> >>> ~ # time taskset -c 4 ./a.out
> >>> 0.22user 13.45system 0:13.80elapsed
> >>>
> >>> Also, Yicong Yang added the following observation.
> >>> Tested with benchmark in the commit on Kunpeng920 arm64 server,
> >>> observed an improvement around 12.5% with command
> >>> `time ./swap_bench`.
> >>> w/o w/
> >>> real 0m13.460s 0m11.771s
> >>> user 0m0.248s 0m0.279s
> >>> sys 0m12.039s 0m11.458s
> >>>
> >>> Originally it's noticed a 16.99% overhead of ptep_clear_flush()
> >>> which has been eliminated by this patch:
> >>>
> >>> [root@localhost yang]# perf record -- ./swap_bench && perf report
> >>> [...]
> >>> 16.99% swap_bench [kernel.kallsyms] [k] ptep_clear_flush
> >>>
> >>> Cc: Jonathan Corbet <corbet@xxxxxxx>
> >>> Cc: Nadav Amit <namit@xxxxxxxxxx>
> >>> Cc: Mel Gorman <mgorman@xxxxxxx>
> >>> Tested-by: Yicong Yang <yangyicong@xxxxxxxxxxxxx>
> >>> Tested-by: Xin Hao <xhao@xxxxxxxxxxxxxxxxx>
> >>> Signed-off-by: Barry Song <v-songbaohua@xxxxxxxx>
> >>> Signed-off-by: Yicong Yang <yangyicong@xxxxxxxxxxxxx>
> >>> ---
> >>> .../features/vm/TLB/arch-support.txt | 2 +-
> >>> arch/arm64/Kconfig | 1 +
> >>> arch/arm64/include/asm/tlbbatch.h | 12 ++++++++
> >>> arch/arm64/include/asm/tlbflush.h | 28 +++++++++++++++++--
> >>> 4 files changed, 40 insertions(+), 3 deletions(-)
> >>> create mode 100644 arch/arm64/include/asm/tlbbatch.h
> >>>
> >>> diff --git a/Documentation/features/vm/TLB/arch-support.txt b/Documentation/features/vm/TLB/arch-support.txt
> >>> index 1c009312b9c1..2caf815d7c6c 100644
> >>> --- a/Documentation/features/vm/TLB/arch-support.txt
> >>> +++ b/Documentation/features/vm/TLB/arch-support.txt
> >>> @@ -9,7 +9,7 @@
> >>> | alpha: | TODO |
> >>> | arc: | TODO |
> >>> | arm: | TODO |
> >>> - | arm64: | TODO |
> >>> + | arm64: | ok |
> >>> | csky: | TODO |
> >>> | hexagon: | TODO |
> >>> | ia64: | TODO |
> >>> diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
> >>> index 571cc234d0b3..09d45cd6d665 100644
> >>> --- a/arch/arm64/Kconfig
> >>> +++ b/arch/arm64/Kconfig
> >>> @@ -93,6 +93,7 @@ config ARM64
> >>> select ARCH_SUPPORTS_INT128 if CC_HAS_INT128
> >>> select ARCH_SUPPORTS_NUMA_BALANCING
> >>> select ARCH_SUPPORTS_PAGE_TABLE_CHECK
> >>> + select ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
> >>> select ARCH_WANT_COMPAT_IPC_PARSE_VERSION if COMPAT
> >>> select ARCH_WANT_DEFAULT_BPF_JIT
> >>> select ARCH_WANT_DEFAULT_TOPDOWN_MMAP_LAYOUT
> >>> diff --git a/arch/arm64/include/asm/tlbbatch.h b/arch/arm64/include/asm/tlbbatch.h
> >>> new file mode 100644
> >>> index 000000000000..fedb0b87b8db
> >>> --- /dev/null
> >>> +++ b/arch/arm64/include/asm/tlbbatch.h
> >>> @@ -0,0 +1,12 @@
> >>> +/* SPDX-License-Identifier: GPL-2.0 */
> >>> +#ifndef _ARCH_ARM64_TLBBATCH_H
> >>> +#define _ARCH_ARM64_TLBBATCH_H
> >>> +
> >>> +struct arch_tlbflush_unmap_batch {
> >>> + /*
> >>> + * For arm64, HW can do tlb shootdown, so we don't
> >>> + * need to record cpumask for sending IPI
> >>> + */
> >>> +};
> >>> +
> >>> +#endif /* _ARCH_ARM64_TLBBATCH_H */
> >>> diff --git a/arch/arm64/include/asm/tlbflush.h b/arch/arm64/include/asm/tlbflush.h
> >>> index 412a3b9a3c25..23cbc987321a 100644
> >>> --- a/arch/arm64/include/asm/tlbflush.h
> >>> +++ b/arch/arm64/include/asm/tlbflush.h
> >>> @@ -254,17 +254,24 @@ static inline void flush_tlb_mm(struct mm_struct *mm)
> >>> dsb(ish);
> >>> }
> >>>
> >>> -static inline void flush_tlb_page_nosync(struct vm_area_struct *vma,
> >>> +
> >>> +static inline void __flush_tlb_page_nosync(struct mm_struct *mm,
> >>> unsigned long uaddr)
> >>> {
> >>> unsigned long addr;
> >>>
> >>> dsb(ishst);
> >>> - addr = __TLBI_VADDR(uaddr, ASID(vma->vm_mm));
> >>> + addr = __TLBI_VADDR(uaddr, ASID(mm));
> >>> __tlbi(vale1is, addr);
> >>> __tlbi_user(vale1is, addr);
> >>> }
> >>>
> >>> +static inline void flush_tlb_page_nosync(struct vm_area_struct *vma,
> >>> + unsigned long uaddr)
> >>> +{
> >>> + return __flush_tlb_page_nosync(vma->vm_mm, uaddr);
> >>> +}
> >>> +
> >>> static inline void flush_tlb_page(struct vm_area_struct *vma,
> >>> unsigned long uaddr)
> >>> {
> >>> @@ -272,6 +279,23 @@ static inline void flush_tlb_page(struct vm_area_struct *vma,
> >>> dsb(ish);
> >>> }
> >>>
> >>> +static inline bool arch_tlbbatch_should_defer(struct mm_struct *mm)
> >>> +{
> >>> + return true;
> >>> +}
> >>
> >> Always defer and batch up TLB flush, unconditionally ?
> >
> > My understanding is we actually don't need tlbbatch for a machine with one
> > or two cores as the tlb flush is not expensive. even for a system with four
> > cortex-a55 cores, i didn't see obvious cost. it was less than 1%.
> > when we have 8 cores, we see the obvious cost of tlb flush. for a server with
> > 100 crores, the cost is incredibly huge.
>
> Although dsb(ish) is deferred via arch_tlbbatch_flush(), there is still
> one dsb(isht) instruction left in __flush_tlb_page_nosync(). Is not that
> expensive as well, while queuing up individual TLB flushes ?
This one is much much cheaper as it is not waiting for the
completion of tlbi. waiting for the completion of tlbi is a big
deal in arm64, thus, similar optimization can be seen here
3403e56b41c1("arm64: mm: Don't wait for completion of TLB invalidation
when page aging").
>
> The very idea behind TLB deferral is the opportunity it (might) provide
> to accumulate address ranges and cpu masks so that individual TLB flush
> can be replaced with a more cost effective range based TLB flush. Hence
> I guess unless address range or cpumask based cost effective TLB flush
> is available, deferral does not improve the unmap performance as much.
After sending tlbi, if we wait for the completion of tlbi, we have to get Ack
from all cpus in the system, tlbi is not scalable. The point here is that we
avoid waiting for each individual TLBi. Alternatively, they are batched. If
you read the benchmark in the commit log, you can find the great decline
in the cost to swap out a page.
Thanks
Barry
