On Fri, 28 Apr 2017 08:13:01 +0200 (CEST)
Christophe Leroy <christophe.le...@c-s.fr> wrote:
> Commit a7a9dcd882a67 ("powerpc: Avoid taking a data miss on every
> userspace instruction miss") has shown that limiting the read of
> faulting instruction to likely cases improves performance.
>
> This patch goes further into this direction by limiting the read
> of the faulting instruction to the only cases where it is definitly
> needed.
>
> On an MPC885, with the same benchmark app as in the commit referred
> above, we see a reduction of 4000 dTLB misses (approx 3%):
>
> Before the patch:
> Performance counter stats for './fault 500' (10 runs):
>
> 720495838 cpu-cycles
> ( +- 0.04% )
> 141769 dTLB-load-misses
> ( +- 0.02% )
> 52722 iTLB-load-misses
> ( +- 0.01% )
> 19611 faults
> ( +- 0.02% )
>
> 5.750535176 seconds time elapsed
> ( +- 0.16% )
>
> With the patch:
> Performance counter stats for './fault 500' (10 runs):
>
> 717669123 cpu-cycles
> ( +- 0.02% )
> 137344 dTLB-load-misses
> ( +- 0.03% )
> 52731 iTLB-load-misses
> ( +- 0.01% )
> 19614 faults
> ( +- 0.03% )
>
> 5.728423115 seconds time elapsed
> ( +- 0.14% )
>
> Signed-off-by: Christophe Leroy <christophe.le...@c-s.fr>
> ---
> v2: Changes 'if (cond1) if (cond2)' by 'if (cond1 && cond2)'
>
> In case the instruction we read has value 0, store_update_sp() will
> return false, so it will bail out.
>
> This patch applies after the serie "powerpc/mm: some cleanup of
> do_page_fault()"
>
> arch/powerpc/mm/fault.c | 22 ++++++++++++----------
> 1 file changed, 12 insertions(+), 10 deletions(-)
>
> diff --git a/arch/powerpc/mm/fault.c b/arch/powerpc/mm/fault.c
> index 400f2d0d42f8..2ec82a279d28 100644
> --- a/arch/powerpc/mm/fault.c
> +++ b/arch/powerpc/mm/fault.c
> @@ -280,14 +280,6 @@ int do_page_fault(struct pt_regs *regs, unsigned long
> address,
>
> perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
>
> - /*
> - * We want to do this outside mmap_sem, because reading code around nip
> - * can result in fault, which will cause a deadlock when called with
> - * mmap_sem held
> - */
> - if (is_write && is_user)
> - __get_user(inst, (unsigned int __user *)regs->nip);
> -
> if (is_user)
> flags |= FAULT_FLAG_USER;
>
> @@ -356,8 +348,18 @@ int do_page_fault(struct pt_regs *regs, unsigned long
> address,
> * between the last mapped region and the stack will
> * expand the stack rather than segfaulting.
> */
> - if (address + 2048 < uregs->gpr[1] && !store_updates_sp(inst))
> - goto bad_area;
> + if (address + 2048 < uregs->gpr[1] && !inst) {
> + /*
> + * We want to do this outside mmap_sem, because reading
> + * code around nip can result in fault, which will cause
> + * a deadlock when called with mmap_sem held
> + */
> + up_read(&mm->mmap_sem);
> + __get_user(inst, (unsigned int __user *)regs->nip);
> + if (!store_updates_sp(inst))
> + goto bad_area_nosemaphore;
> + goto retry;
> + }
Yes, nice patch. I wonder if you can do __get_user first as non-faulting to
avoid retaking the mmap_sem and retrying? Along the lines of:
+ nip = (unsigned int __user *)regs->nip;
+ pagefault_disable();
+ if (unlikely(__get_user_inatomic(inst, nip))) {
+ pagefault_enable();
+ up_read(&mm->mmap_sem);
+ if (get_user(inst, nip)) {
...
goto retry;
The user instruction should practically always have a Linux pte, so a
fault there should be exceedingly rare, I think?
Thanks,
Nick
