On 2024/4/19 0:32, Peter Xu wrote:
Hi, Kefeng, On Thu, Apr 18, 2024 at 08:06:41PM +0800, Kefeng Wang wrote:Add userfaultfd_wp() check in vmf_orig_pte_uffd_wp() to avoid the unnecessary pte_marker_entry_uffd_wp() in most pagefault, difference as shows below from perf data of lat_pagefault, note, the function vmf_orig_pte_uffd_wp() is not inlined in the two kernel versions. perf report -i perf.data.before | grep vmf 0.17% 0.13% lat_pagefault [kernel.kallsyms] [k] vmf_orig_pte_uffd_wp.part.0.isra.0 perf report -i perf.data.after | grep vmfAny real number to share too besides the perf greps? I meant, even if perf report will not report such function anymore, it doesn't mean it'll be faster, and how much it improves?
dd if=/dev/zero of=/tmp/XXX bs=512M count=1 ./lat_pagefault -W 5 -N 5 /tmp/XXX before after 1 0.2623 0.2605 2 0.2622 0.2598 3 0.2621 0.2595 4 0.2622 0.2600 5 0.2651 0.2598 6 0.2624 0.2594 7 0.2624 0.2605 8 0.2627 0.2608 average 0.262675 0.2600375 -0.0026375 The lat_pagefault does show some improvement(also I reboot and retest, the results are same).
Now we're switching from pte_marker_uffd_wp() check into a vma flag check. I think it makes more sense to compare the number rather than the perf reports, as the vma flag check instructions will be buried under other entries IIUC. Thanks,
