On Fri, Feb 21, 2020 at 06:49:16AM -0800, Jens Axboe wrote:
> > Jens, what exactly is the benefit of running this on every random
> > schedule() vs in io_cqring_wait() ? Or even, since io_cqring_wait() is
> > the very last thing the syscall does, task_work.
>
> I took a step back and I think we can just use the task work, which
> makes this a lot less complicated in terms of locking and schedule
> state. Ran some quick testing with the below and it works for me.
>
> I'm going to re-spin based on this and just dump the sched_work
> addition.
Aswesome, simpler is better.
> diff --git a/fs/io_uring.c b/fs/io_uring.c
> index 81aa3959f326..413ac86d7882 100644
> --- a/fs/io_uring.c
> +++ b/fs/io_uring.c
> @@ -3529,7 +3529,7 @@ static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
> * the exit check will ultimately cancel these work items. Hence we
> * don't need to check here and handle it specifically.
> */
> - sched_work_add(tsk, &req->sched_work);
> + task_work_add(tsk, &req->sched_work, true);
> wake_up_process(tsk);
> return 1;
> }
> @@ -5367,9 +5367,9 @@ static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
> do {
> if (io_cqring_events(ctx, false) >= min_events)
> return 0;
> - if (!current->sched_work)
> + if (!current->task_works)
> break;
> - sched_work_run();
> + task_work_run();
> } while (1);
>
> if (sig) {
> @@ -5392,6 +5392,12 @@ static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
> TASK_INTERRUPTIBLE);
> if (io_should_wake(&iowq, false))
> break;
> + if (current->task_works) {
> + task_work_run();
> + if (io_should_wake(&iowq, false))
> + break;
> + continue;
> + }
if (current->task_works)
task_work_run();
if (io_should_wake(&iowq, false);
break;
doesn't work?
> schedule();
> if (signal_pending(current)) {
> ret = -EINTR;
Anyway, we need to be careful about the context where we call
task_work_run(), but afaict doing it here should be fine.
