On Wed, Sep 2, 2020 at 1:11 AM Jens Axboe <axboe@xxxxxxxxx> wrote:
> The restriction of needing fixed files for SQPOLL is problematic, and
> prevents/inhibits several valid uses cases.
>
> There's no real good reason for us not to allow it, except we need to
> have the sqpoll thread inherit current->files from the task that setup
> the ring. We can't easily do that, since we'd introduce a circular
> reference by holding on to our own file table.
>
> If we wait for the sqpoll thread to exit when the ring fd is closed,
> then we can safely reference the task files_struct without holding
> a reference to it. And once we inherit that in the SQPOLL thread, we
> can support non-fixed files for SQPOLL.
[...]
> diff --git a/fs/io_uring.c b/fs/io_uring.c
[...]
> + /*
> + * For SQPOLL usage - no reference is held to this file table, we
> + * rely on fops->flush() and our callback there waiting for the users
> + * to finish.
> + */
> + struct files_struct *sqo_files;
[...]
> @@ -6621,6 +6622,10 @@ static int io_sq_thread(void *data)
>
> old_cred = override_creds(ctx->creds);
>
> + task_lock(current);
> + current->files = ctx->sqo_files;
> + task_unlock(current);
[...]
> @@ -7549,6 +7557,13 @@ static int io_sq_offload_create(struct io_ring_ctx *ctx,
> if (!capable(CAP_SYS_ADMIN))
> goto err;
>
> + /*
> + * We will exit the sqthread before current exits, so we can
> + * avoid taking a reference here and introducing weird
> + * circular dependencies on the files table.
> + */
> + ctx->sqo_files = current->files;
[...]
> @@ -8239,8 +8254,10 @@ static int io_uring_flush(struct file *file, void *data)
> /*
> * If the task is going away, cancel work it may have pending
> */
> - if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
> + if (fatal_signal_pending(current) || (current->flags & PF_EXITING)) {
> + io_sq_thread_stop(ctx);
> io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, current, true);
> + }
What happens when the uring setup syscall fails around
io_uring_get_fd() (after the sq offload thread was started, but before
an fd was installed)? Will that also properly wait for the sq_thread
to go away before returning from the syscall (at which point the
files_struct could disappear)?
