Re: [PATCH 2/7] arm64: mm: accelerate pagefault when VM_FAULT_BADACCESS
On Tue, Apr 02, 2024 at 03:51:37PM +0800, Kefeng Wang wrote: > The vm_flags of vma already checked under per-VMA lock, if it is a > bad access, directly set fault to VM_FAULT_BADACCESS and handle error, > no need to lock_mm_and_find_vma() and check vm_flags again, the latency > time reduce 34% in lmbench 'lat_sig -P 1 prot lat_sig'. > > Signed-off-by: Kefeng Wang > --- > arch/arm64/mm/fault.c | 4 +++- > 1 file changed, 3 insertions(+), 1 deletion(-) > > diff --git a/arch/arm64/mm/fault.c b/arch/arm64/mm/fault.c > index 9bb9f395351a..405f9aa831bd 100644 > --- a/arch/arm64/mm/fault.c > +++ b/arch/arm64/mm/fault.c > @@ -572,7 +572,9 @@ static int __kprobes do_page_fault(unsigned long far, > unsigned long esr, > > if (!(vma->vm_flags & vm_flags)) { > vma_end_read(vma); > - goto lock_mmap; > + fault = VM_FAULT_BADACCESS; > + count_vm_vma_lock_event(VMA_LOCK_SUCCESS); > + goto done; > } > fault = handle_mm_fault(vma, addr, mm_flags | FAULT_FLAG_VMA_LOCK, > regs); > if (!(fault & (VM_FAULT_RETRY | VM_FAULT_COMPLETED))) I think this makes sense. A concurrent modification of vma->vm_flags (e.g. mprotect()) would do a vma_start_write(), so no need to recheck again with the mmap lock held. Reviewed-by: Catalin Marinas
Re: [PATCH 2/7] arm64: mm: accelerate pagefault when VM_FAULT_BADACCESS
On 2024/4/3 13:30, Suren Baghdasaryan wrote: On Tue, Apr 2, 2024 at 10:19 PM Suren Baghdasaryan wrote: On Tue, Apr 2, 2024 at 12:53 AM Kefeng Wang wrote: The vm_flags of vma already checked under per-VMA lock, if it is a bad access, directly set fault to VM_FAULT_BADACCESS and handle error, no need to lock_mm_and_find_vma() and check vm_flags again, the latency time reduce 34% in lmbench 'lat_sig -P 1 prot lat_sig'. The change makes sense to me. Per-VMA lock is enough to keep vma->vm_flags stable, so no need to retry with mmap_lock. Signed-off-by: Kefeng Wang Reviewed-by: Suren Baghdasaryan --- arch/arm64/mm/fault.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/arch/arm64/mm/fault.c b/arch/arm64/mm/fault.c index 9bb9f395351a..405f9aa831bd 100644 --- a/arch/arm64/mm/fault.c +++ b/arch/arm64/mm/fault.c @@ -572,7 +572,9 @@ static int __kprobes do_page_fault(unsigned long far, unsigned long esr, if (!(vma->vm_flags & vm_flags)) { vma_end_read(vma); - goto lock_mmap; + fault = VM_FAULT_BADACCESS; + count_vm_vma_lock_event(VMA_LOCK_SUCCESS); nit: VMA_LOCK_SUCCESS accounting here seems correct to me but unrelated to the main change. Either splitting into a separate patch or mentioning this additional fixup in the changelog would be helpful. The above nit applies to all the patches after this one, so I won't comment on each one separately. If you decide to split or adjust the changelog please do that for each patch. I will update the change log for each patch, thank for your review and suggestion. + goto done; } fault = handle_mm_fault(vma, addr, mm_flags | FAULT_FLAG_VMA_LOCK, regs); if (!(fault & (VM_FAULT_RETRY | VM_FAULT_COMPLETED))) -- 2.27.0
Re: [PATCH 2/7] arm64: mm: accelerate pagefault when VM_FAULT_BADACCESS
On Tue, Apr 2, 2024 at 10:19 PM Suren Baghdasaryan wrote: > > On Tue, Apr 2, 2024 at 12:53 AM Kefeng Wang > wrote: > > > > The vm_flags of vma already checked under per-VMA lock, if it is a > > bad access, directly set fault to VM_FAULT_BADACCESS and handle error, > > no need to lock_mm_and_find_vma() and check vm_flags again, the latency > > time reduce 34% in lmbench 'lat_sig -P 1 prot lat_sig'. > > The change makes sense to me. Per-VMA lock is enough to keep > vma->vm_flags stable, so no need to retry with mmap_lock. > > > > > Signed-off-by: Kefeng Wang > > Reviewed-by: Suren Baghdasaryan > > > --- > > arch/arm64/mm/fault.c | 4 +++- > > 1 file changed, 3 insertions(+), 1 deletion(-) > > > > diff --git a/arch/arm64/mm/fault.c b/arch/arm64/mm/fault.c > > index 9bb9f395351a..405f9aa831bd 100644 > > --- a/arch/arm64/mm/fault.c > > +++ b/arch/arm64/mm/fault.c > > @@ -572,7 +572,9 @@ static int __kprobes do_page_fault(unsigned long far, > > unsigned long esr, > > > > if (!(vma->vm_flags & vm_flags)) { > > vma_end_read(vma); > > - goto lock_mmap; > > + fault = VM_FAULT_BADACCESS; > > + count_vm_vma_lock_event(VMA_LOCK_SUCCESS); > > nit: VMA_LOCK_SUCCESS accounting here seems correct to me but > unrelated to the main change. Either splitting into a separate patch > or mentioning this additional fixup in the changelog would be helpful. The above nit applies to all the patches after this one, so I won't comment on each one separately. If you decide to split or adjust the changelog please do that for each patch. > > > + goto done; > > } > > fault = handle_mm_fault(vma, addr, mm_flags | FAULT_FLAG_VMA_LOCK, > > regs); > > if (!(fault & (VM_FAULT_RETRY | VM_FAULT_COMPLETED))) > > -- > > 2.27.0 > >
Re: [PATCH 2/7] arm64: mm: accelerate pagefault when VM_FAULT_BADACCESS
On Tue, Apr 2, 2024 at 12:53 AM Kefeng Wang wrote: > > The vm_flags of vma already checked under per-VMA lock, if it is a > bad access, directly set fault to VM_FAULT_BADACCESS and handle error, > no need to lock_mm_and_find_vma() and check vm_flags again, the latency > time reduce 34% in lmbench 'lat_sig -P 1 prot lat_sig'. The change makes sense to me. Per-VMA lock is enough to keep vma->vm_flags stable, so no need to retry with mmap_lock. > > Signed-off-by: Kefeng Wang Reviewed-by: Suren Baghdasaryan > --- > arch/arm64/mm/fault.c | 4 +++- > 1 file changed, 3 insertions(+), 1 deletion(-) > > diff --git a/arch/arm64/mm/fault.c b/arch/arm64/mm/fault.c > index 9bb9f395351a..405f9aa831bd 100644 > --- a/arch/arm64/mm/fault.c > +++ b/arch/arm64/mm/fault.c > @@ -572,7 +572,9 @@ static int __kprobes do_page_fault(unsigned long far, > unsigned long esr, > > if (!(vma->vm_flags & vm_flags)) { > vma_end_read(vma); > - goto lock_mmap; > + fault = VM_FAULT_BADACCESS; > + count_vm_vma_lock_event(VMA_LOCK_SUCCESS); nit: VMA_LOCK_SUCCESS accounting here seems correct to me but unrelated to the main change. Either splitting into a separate patch or mentioning this additional fixup in the changelog would be helpful. > + goto done; > } > fault = handle_mm_fault(vma, addr, mm_flags | FAULT_FLAG_VMA_LOCK, > regs); > if (!(fault & (VM_FAULT_RETRY | VM_FAULT_COMPLETED))) > -- > 2.27.0 >