On Tue, 22 May 2018 16:02:56 +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 likely > needed. > > On an MPC885, with the same benchmark app as in the commit referred > above, we see a reduction of about 3900 dTLB misses (approx 3%): > > Before the patch: > Performance counter stats for './fault 500' (10 runs): > > 683033312 cpu-cycles > ( +- 0.03% ) > 134538 dTLB-load-misses > ( +- 0.03% ) > 46099 iTLB-load-misses > ( +- 0.02% ) > 19681 faults > ( +- 0.02% ) > > 5.389747878 seconds time elapsed > ( +- 0.06% ) > > With the patch: > > Performance counter stats for './fault 500' (10 runs): > > 682112862 cpu-cycles > ( +- 0.03% ) > 130619 dTLB-load-misses > ( +- 0.03% ) > 46073 iTLB-load-misses > ( +- 0.05% ) > 19681 faults > ( +- 0.01% ) > > 5.381342641 seconds time elapsed > ( +- 0.07% ) > > The proper work of the huge stack expansion was tested with the > following app: > > int main(int argc, char **argv) > { > char buf[1024 * 1025]; > > sprintf(buf, "Hello world !\n"); > printf(buf); > > exit(0); > } > > Signed-off-by: Christophe Leroy <christophe.le...@c-s.fr> > --- > v7: Following comment from Nicholas on v6 on possibility of the page getting > removed from the pagetables > between the fault and the read, I have reworked the patch in order to do > the get_user() in > __do_page_fault() directly in order to reduce complexity compared to > version v5 This is looking better, thanks. > diff --git a/arch/powerpc/mm/fault.c b/arch/powerpc/mm/fault.c > index fcbb34431da2..dc64b8e06477 100644 > --- a/arch/powerpc/mm/fault.c > +++ b/arch/powerpc/mm/fault.c > @@ -450,9 +450,6 @@ static int __do_page_fault(struct pt_regs *regs, unsigned > long address, > * 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; > if (is_write) > @@ -498,6 +495,26 @@ static int __do_page_fault(struct pt_regs *regs, > unsigned long address, > if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) > return bad_area(regs, address); > > + if (unlikely(is_write && is_user && address + 0x100000 < vma->vm_end && > + !inst)) { > + unsigned int __user *nip = (unsigned int __user *)regs->nip; > + > + if (likely(access_ok(VERIFY_READ, nip, sizeof(inst)))) { > + int res; > + > + pagefault_disable(); > + res = __get_user_inatomic(inst, nip); > + pagefault_enable(); > + if (unlikely(res)) { > + up_read(&mm->mmap_sem); > + res = __get_user(inst, nip); > + if (!res && inst) > + goto retry; You're handling error here but the previous code did not? > + return bad_area_nosemaphore(regs, address); > + } > + } > + } Would it be nicer to move all this up into bad_stack_expansion(). It would need a way to handle the retry and insn, but I think it would still look better. Thanks, Nick