Re: [PATCH V2] io_uring: cancelable uring_cmd

[Jens Axboe <axboe@kernel.dk>]

On 9/22/23 8:28 AM, Ming Lei wrote:
> uring_cmd may never complete, such as ublk, in which uring cmd isn't
> completed until one new block request is coming from ublk block device.
> 
> Add cancelable uring_cmd to provide mechanism to driver for cancelling
> pending commands in its own way.
> 
> Add API of io_uring_cmd_mark_cancelable() for driver to mark one command as
> cancelable, then io_uring will cancel this command in
> io_uring_cancel_generic(). ->uring_cmd() callback is reused for canceling
> command in driver's way, then driver gets notified with the cancelling
> from io_uring.
> 
> Add API of io_uring_cmd_get_task() to help driver cancel handler
> deal with the canceling.

This looks better, a few comments:

> diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h
> index 8e61f8b7c2ce..29a7a7e71f57 100644
> --- a/include/uapi/linux/io_uring.h
> +++ b/include/uapi/linux/io_uring.h
> @@ -249,10 +249,13 @@ enum io_uring_op {
>   * sqe->uring_cmd_flags
>   * IORING_URING_CMD_FIXED	use registered buffer; pass this flag
>   *				along with setting sqe->buf_index.
> + * IORING_URING_CANCELABLE	not for userspace
>   * IORING_URING_CMD_POLLED	driver use only
>   */
> -#define IORING_URING_CMD_FIXED	(1U << 0)
> -#define IORING_URING_CMD_POLLED	(1U << 31)
> +#define IORING_URING_CMD_FIXED		(1U << 0)
> +/* set by driver, and handled by io_uring to cancel this cmd */
> +#define IORING_URING_CMD_CANCELABLE	(1U << 30)
> +#define IORING_URING_CMD_POLLED		(1U << 31)

If IORING_URING_CANCELABLE isn't UAPI, why stuff it in here? Should we
have a split where we retain the upper 8 bits for internal use, or
something like that?

> diff --git a/io_uring/io_uring.c b/io_uring/io_uring.c
> index 783ed0fff71b..a3135fd47a4e 100644
> --- a/io_uring/io_uring.c
> +++ b/io_uring/io_uring.c
> @@ -3256,6 +3256,40 @@ static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
>  	return ret;
>  }
>  
> +static bool io_uring_try_cancel_uring_cmd(struct io_ring_ctx *ctx,
> +		struct task_struct *task, bool cancel_all)
> +{
> +	struct hlist_node *tmp;
> +	struct io_kiocb *req;
> +	bool ret = false;
> +
> +	mutex_lock(&ctx->uring_lock);
> +	hlist_for_each_entry_safe(req, tmp, &ctx->cancelable_uring_cmd,
> +			hash_node) {
> +		struct io_uring_cmd *cmd = io_kiocb_to_cmd(req,
> +				struct io_uring_cmd);
> +		struct file *file = req->file;
> +
> +		if (WARN_ON_ONCE(!file->f_op->uring_cmd))
> +			continue;
> +
> +		if (!cancel_all && req->task != task)
> +			continue;
> +
> +		if (cmd->flags & IORING_URING_CMD_CANCELABLE) {
> +			/* ->sqe isn't available if no async data */
> +			if (!req_has_async_data(req))
> +				cmd->sqe = NULL;
> +			file->f_op->uring_cmd(cmd, IO_URING_F_CANCEL);
> +			ret = true;
> +		}
> +	}
> +	io_submit_flush_completions(ctx);
> +	mutex_unlock(&ctx->uring_lock);
> +
> +	return ret;
> +}

I think it'd be saner to drop uring_lock here, and then:

> @@ -3307,6 +3341,7 @@ static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
>  	ret |= io_kill_timeouts(ctx, task, cancel_all);
>  	if (task)
>  		ret |= io_run_task_work() > 0;
> +	ret |= io_uring_try_cancel_uring_cmd(ctx, task, cancel_all);
>  	return ret;
>  }

move this hunk into the uring_lock section. Also ensure that we do run
task_work for cancelation, should the uring_cmd side require that
(either now or eventually).

> diff --git a/io_uring/uring_cmd.c b/io_uring/uring_cmd.c
> index 537795fddc87..d6b200a0be33 100644
> --- a/io_uring/uring_cmd.c
> +++ b/io_uring/uring_cmd.c
> @@ -13,6 +13,52 @@
>  #include "rsrc.h"
>  #include "uring_cmd.h"
>  
> +static void io_uring_cmd_del_cancelable(struct io_uring_cmd *cmd,
> +		unsigned int issue_flags)
> +{
> +	if (cmd->flags & IORING_URING_CMD_CANCELABLE) {
> +		struct io_kiocb *req = cmd_to_io_kiocb(cmd);
> +		struct io_ring_ctx *ctx = req->ctx;
> +
> +		io_ring_submit_lock(ctx, issue_flags);
> +		cmd->flags &= ~IORING_URING_CMD_CANCELABLE;
> +		hlist_del(&req->hash_node);
> +		io_ring_submit_unlock(ctx, issue_flags);
> +	}
> +}

static void io_uring_cmd_del_cancelable(struct io_uring_cmd *cmd,
		unsigned int issue_flags)
{
	struct io_kiocb *req = cmd_to_io_kiocb(cmd);
	struct io_ring_ctx *ctx = req->ctx;

	if (!(cmd->flags & IORING_URING_CMD_CANCELABLE))
		return;

	io_ring_submit_lock(ctx, issue_flags);
	cmd->flags &= ~IORING_URING_CMD_CANCELABLE;
	hlist_del(&req->hash_node);
	io_ring_submit_unlock(ctx, issue_flags);
}

is cleaner imho. Minor nit.

> +
> +/*
> + * Mark this command as concelable, then io_uring_try_cancel_uring_cmd()
> + * will try to cancel this issued command by sending ->uring_cmd() with
> + * issue_flags of IO_URING_F_CANCEL.
> + *
> + * The command is guaranteed to not be done when calling ->uring_cmd()
> + * with IO_URING_F_CANCEL, but it is driver's responsibility to deal
> + * with race between io_uring canceling and normal completion.
> + */
> +int io_uring_cmd_mark_cancelable(struct io_uring_cmd *cmd,
> +		unsigned int issue_flags)
> +{
> +	struct io_kiocb *req = cmd_to_io_kiocb(cmd);
> +	struct io_ring_ctx *ctx = req->ctx;
> +
> +	io_ring_submit_lock(ctx, issue_flags);
> +	if (!(cmd->flags & IORING_URING_CMD_CANCELABLE)) {
> +		cmd->flags |= IORING_URING_CMD_CANCELABLE;
> +		hlist_add_head(&req->hash_node, &ctx->cancelable_uring_cmd);
> +	}
> +	io_ring_submit_unlock(ctx, issue_flags);
> +
> +	return 0;
> +}

A bit inconsistent here in terms of the locking. I'm assuming the
marking happens within issue, in which case it should be fine to do:

int io_uring_cmd_mark_cancelable(struct io_uring_cmd *cmd,
		unsigned int issue_flags)
{
	struct io_kiocb *req = cmd_to_io_kiocb(cmd);
	struct io_ring_ctx *ctx = req->ctx;

	if (!(cmd->flags & IORING_URING_CMD_CANCELABLE)) {
		cmd->flags |= IORING_URING_CMD_CANCELABLE;
		io_ring_submit_lock(ctx, issue_flags);
		hlist_add_head(&req->hash_node, &ctx->cancelable_uring_cmd);
		io_ring_submit_unlock(ctx, issue_flags);
	}

	return 0;
}

-- 
Jens Axboe