On 1/9/24 10:30 AM, Christian Brauner wrote: > On Tue, Jan 09, 2024 at 09:08:40AM -0700, Jens Axboe wrote: >> This reverts commit d9e5d31084b024734e64307521414ef0ae1d5333. >> >> This commit added an extra fsnotify call from rw_verify_area(), which >> can be a very hot path. In my testing, this adds roughly 5-6% extra >> overhead per IO, which is quite a lot. As a result, throughput of >> said test also drops by 5-6%, as it is CPU bound. Looking at perf, >> it's apparent why: >> >> 3.36% [kernel.vmlinux] [k] fsnotify >> 2.32% [kernel.vmlinux] [k] __fsnotify_paren >> >> which directly correlates with performance lost. >> >> As the rationale for this patch isn't very strong, revert this commit >> for now and reclaim the performance. >> >> Signed-off-by: Jens Axboe <axboe@xxxxxxxxx> >> --- > > Thanks for tracking this down! I think Amir, you, and I came to the > conclusion that we can fix this in without having to revert. Amir is > sending out a new patch later tonight. I'll get that fixed by the end of > the week. Yep, either revert or that patch fixes things for me. -- Jens Axboe
