> On top of that, capabilities will also be reduced from root to > CAP_SYS_NICE instead, and sharing across rings for the SQPOLL thread > will be supported. So it'll be a lot more useful/flexible in general. oha that's nice, I'm pretty excited :) I'm just wondering if all op are supported when the SQPOLL flag is set? the accept op seems to fail with -EINVAL, when I enable SQPOLL to reproduce it: https://gist.github.com/1Jo1/accb91b737abb55d07487799739ad70a (just want to test a non blocking accept op in SQPOLL mode) --- Josef On Sun, 6 Sep 2020 at 18:25, Jens Axboe <axboe@xxxxxxxxx> wrote: > > On 9/6/20 10:24 AM, Josef wrote: > >> You're using the 'fd' as the file descriptor, for registered files > >> you want to use the index instead. Since it's the only fd you > >> registered, the index would be 0 and that's what you should use. > > > > oh..yeah it works, thanks :) > > Great! > > >> It's worth mentioning that for 5.10 and on, SQPOLL will no longer > >> require registered files. > > > > that's awesome, it would be really handy as I just implemented a kind > > of workaround in netty :) > > On top of that, capabilities will also be reduced from root to > CAP_SYS_NICE instead, and sharing across rings for the SQPOLL thread > will be supported. So it'll be a lot more useful/flexible in general. > > -- > Jens Axboe >
