First of all, if you haven't already noticed, there's a new and shiny io_uring mailing list. It's: io-uring@xxxxxxxxxxxxxxx Subscribe if you are at all interested in io_uring development. It's meant to cover both kernel and user side, and everything from questions, bugs, development, etc. Anyway, this is support for IORING_OP_LINK_TIMEOUT. Unlike the timeouts we have now that work on purely the CQ ring, these timeouts are specifically tied to a specific command. They are meant to be used to auto-cancel a request, if it hasn't finished in X amount of time. The way to use then is to setup your command as you usually would, but then mark is IOSQE_IO_LINK and add an IORING_OP_LINK_TIMEOUT right after it. That's how linked commands work to begin with. The main difference here is that links are normally only started once the dependent request completes, but for IORING_OP_LINK_TIMEOUT they are armed as soon as we start the dependent request. If the dependent request finishes before the linked timeout, the timeout is canceled. If the timeout finishes before the dependent request, the dependent request is attempted canceled. IORING_OP_LINK_TIMEOUT is setup just like IORING_OP_TIMEOUT in terms of passing in the timespec associated with it. I added a bunch of test cases to liburing, currently residing in a link-timeout branch. View them here: https://git.kernel.dk/cgit/liburing/commit/?h=link-timeout&id=bc1bd5e97e2c758d6fd975bd35843b9b2c770c5a Patches are against for-5.5/io_uring, and can currently also be found in my for-5.5/io_uring-test branch. fs/io_uring.c | 203 +++++++++++++++++++++++++++++++--- include/uapi/linux/io_uring.h | 1 + 2 files changed, 187 insertions(+), 17 deletions(-) -- Jens Axboe
