This is a fantastic analysis and writeup of the problem. Is there a practical reason that system-journald doesn't do O_DIRECT and non-blocking (using io_submit

This is a fantastic analysis and writeup of the problem. Is there a practical reason that system-journald doesn't do O_DIRECT and non-blocking (using io_submit family of syscalls) writes/fsync? If minimizing blocking is such an important issue then given the state of disk IO in Linux that your only solution. As long as you pre-allocated the space and are not required to do metadata updates you will get mostly non-blocking operations. Since this is done for logging which I'm presuming is not going to be read right back there's no downside of using O_DIRECT, in fact not polluting the page cache might be a net benefit. Btw, is there anyway to get in touch with the author to talk about the issue? Can't find his contact details.