Implement io_uring for supporting kernels.

Open jevolk opened this issue 6 years ago • 1 comments

Without RWF_NOWAIT (since Linux 4.14) the call to io_submit(2) in ircd/fs_aio.cc is liable to block the entire daemon under certain conditions. A non-blocking use of this call allows the daemon to continue making progress in other contexts during that time.

Mar 13 '19 00:03 jevolk

I'm starting to believe this isn't feasible. Even if we can get the -EAGAIN and start a readahead(2)/posix_fadvise(2) and then context-switch to something else, there is no facility I'm aware of for the kernel to indicate when we would no longer get -EAGAIN. Adding that fd into the epoll doesn't work (still?). So this might have to be redirected into just implementing io_uring on Linux 5.1.

Jun 04 '19 03:06 jevolk