Re: [PATCH RFC] io_uring: support ioctl

[Hao Xu <haoxu@xxxxxxxxxxxxxxxxx>]

在 2020/11/2 下午7:07, Pavel Begunkov 写道:
> On 02/11/2020 07:44, Hao Xu wrote:
> > Async ioctl is necessary for some scenarios like nonblocking
> > single-threaded model
>
> Once I fell for it myself, see Jann explained why that's a bad idea.
Thanks, Pavel. The information is helpful and make sense to me.
It seems I have to rethink it. Btw, your version is more robust.
> https://lore.kernel.org/io-uring/CAG48ez0N_b+kjbddhHe+BUvSnOSvpm1vdfQ9cv+cgTLuCMXqug@xxxxxxxxxxxxxx/
>
> > Signed-off-by: Hao Xu <haoxu@xxxxxxxxxxxxxxxxx>
> > ---
> > I've written corresponding liburing tests for this feature. Currently
> > just a simple test for BLKGETSIZE operation. I'll release it later soon
> > when it gets better.
> >
> >   fs/io_uring.c                 | 56 +++++++++++++++++++++++++++++++++++++++++++
> >   fs/ioctl.c                    |  4 ++--
> >   include/linux/fs.h            |  3 ++-
> >   include/uapi/linux/io_uring.h |  1 +
> >   4 files changed, 61 insertions(+), 3 deletions(-)
> >
> > diff --git a/fs/io_uring.c b/fs/io_uring.c
> > index b42dfa0243bf..c8ab6b6d2d70 100644
> > --- a/fs/io_uring.c
> > +++ b/fs/io_uring.c
> > @@ -539,6 +539,13 @@ struct io_statx {
> >   	struct statx __user		*buffer;
> >   };
> >   
> > +struct io_ioctl {
> > +	struct file			*file;
> > +	unsigned int                    fd;
> > +	unsigned int                    cmd;
> > +	unsigned long                   arg;
> > +};
> > +
> >   struct io_completion {
> >   	struct file			*file;
> >   	struct list_head		list;
> > @@ -665,6 +672,7 @@ struct io_kiocb {
> >   		struct io_splice	splice;
> >   		struct io_provide_buf	pbuf;
> >   		struct io_statx		statx;
> > +		struct io_ioctl         ioctl;
> >   		/* use only after cleaning per-op data, see io_clean_op() */
> >   		struct io_completion	compl;
> >   	};
> > @@ -932,6 +940,10 @@ struct io_op_def {
> >   		.hash_reg_file		= 1,
> >   		.unbound_nonreg_file	= 1,
> >   	},
> > +	[IORING_OP_IOCTL] = {
> > +		.needs_file             = 1,
> > +		.work_flags             = IO_WQ_WORK_MM | IO_WQ_WORK_FILES
> > +	},
> >   };
> >   
> >   enum io_mem_account {
> > @@ -4819,6 +4831,45 @@ static int io_connect(struct io_kiocb *req, bool force_nonblock,
> >   }
> >   #endif /* CONFIG_NET */
> >   
> > +static int io_ioctl_prep(struct io_kiocb *req,
> > +			 const struct io_uring_sqe *sqe)
> > +{
> > +	if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
> > +		return -EINVAL;
> > +	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
> > +		return -EINVAL;
> > +
> > +	req->ioctl.fd = READ_ONCE(sqe->fd);
> > +	req->ioctl.cmd = READ_ONCE(sqe->len);
> > +	req->ioctl.arg = READ_ONCE(sqe->addr);
> > +	return 0;
> > +}
> > +
> > +static int io_ioctl(struct io_kiocb *req, bool force_nonblock)
> > +{
> > +	int ret;
> > +
> > +	if (force_nonblock)
> > +		return -EAGAIN;
> > +
> > +	if (!req->file)
> > +		return -EBADF;
> > +
> > +	ret = security_file_ioctl(req->file, req->ioctl.cmd, req->ioctl.arg);
> > +	if (ret)
> > +		goto out;
> > +
> > +	ret = do_vfs_ioctl(req->file, req->ioctl.fd, req->ioctl.cmd, req->ioctl.arg);
> > +	if (ret == -ENOIOCTLCMD)
> > +		ret = vfs_ioctl(req->file, req->ioctl.cmd, req->ioctl.arg);
> > +
> > +out:
> > +	if (ret)
> > +		req_set_fail_links(req);
> > +	io_req_complete(req, ret);
> > +	return 0;
> > +}
> > +
> >   struct io_poll_table {
> >   	struct poll_table_struct pt;
> >   	struct io_kiocb *req;
> > @@ -5742,6 +5793,8 @@ static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
> >   		return io_remove_buffers_prep(req, sqe);
> >   	case IORING_OP_TEE:
> >   		return io_tee_prep(req, sqe);
> > +	case IORING_OP_IOCTL:
> > +		return io_ioctl_prep(req, sqe);
> >   	}
> >   
> >   	printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
> > @@ -5985,6 +6038,9 @@ static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
> >   	case IORING_OP_TEE:
> >   		ret = io_tee(req, force_nonblock);
> >   		break;
> > +	case IORING_OP_IOCTL:
> > +		ret = io_ioctl(req, force_nonblock);
> > +		break;
> >   	default:
> >   		ret = -EINVAL;
> >   		break;
> > diff --git a/fs/ioctl.c b/fs/ioctl.c
> > index 4e6cc0a7d69c..4ff2eb0d8ee0 100644
> > --- a/fs/ioctl.c
> > +++ b/fs/ioctl.c
> > @@ -664,8 +664,8 @@ static int ioctl_file_dedupe_range(struct file *file,
> >    * When you add any new common ioctls to the switches above and below,
> >    * please ensure they have compatible arguments in compat mode.
> >    */
> > -static int do_vfs_ioctl(struct file *filp, unsigned int fd,
> > -			unsigned int cmd, unsigned long arg)
> > +int do_vfs_ioctl(struct file *filp, unsigned int fd,
> > +		unsigned int cmd, unsigned long arg)
> >   {
> >   	void __user *argp = (void __user *)arg;
> >   	struct inode *inode = file_inode(filp);
> > diff --git a/include/linux/fs.h b/include/linux/fs.h
> > index 0bd126418bb6..ad62aa6f6136 100644
> > --- a/include/linux/fs.h
> > +++ b/include/linux/fs.h
> > @@ -1732,7 +1732,8 @@ int vfs_mkobj(struct dentry *, umode_t,
> >   int vfs_utimes(const struct path *path, struct timespec64 *times);
> >   
> >   extern long vfs_ioctl(struct file *file, unsigned int cmd, unsigned long arg);
> > -
> > +extern int do_vfs_ioctl(struct file *filp, unsigned int fd,
> > +			unsigned int cmd, unsigned long arg);
> >   #ifdef CONFIG_COMPAT
> >   extern long compat_ptr_ioctl(struct file *file, unsigned int cmd,
> >   					unsigned long arg);
> > diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h
> > index 98d8e06dea22..4919b4e94c12 100644
> > --- a/include/uapi/linux/io_uring.h
> > +++ b/include/uapi/linux/io_uring.h
> > @@ -132,6 +132,7 @@ enum {
> >   	IORING_OP_PROVIDE_BUFFERS,
> >   	IORING_OP_REMOVE_BUFFERS,
> >   	IORING_OP_TEE,
> > +	IORING_OP_IOCTL,
> >   
> >   	/* this goes last, obviously */
> >   	IORING_OP_LAST,