On Wed, Oct 20, 2021 at 08:15:25AM -0600, Jens Axboe wrote: > I just want to be very clear that neither mine nor Pavel's work is > trying to get into benchmarketing. There are very tangible performance > benefits, and they are backed up by code analysis and testing. That > said, the point is of course not to make this any harder to follow or > maintain, but we are doing suboptimal things in various places and those > should get cleaned up, particularly when they impact performance. Are > some a big eager? Perhaps, but let's not that that cloud the perception > of the effort as a whole. A lot of it seems generally useful. But optimizing for BIO_NO_PAGE_REF is really special. Besides a few in-kernel special cases this is all about the io_uring pre-registered buffers, which are very much a special case and not a normal workload at all. In this case it is just an extra conditional moved into a few callers so I think we're ok, but I'm really worried about optimizing this benchmark fast path over the code everyone actually uses.
