[PATCH 2/2] io_uring: Add vmsplice support

[arni@xxxxxxxx]

From: Árni Dagur <arni@xxxxxxxx>

* The `sqe->splice_flags` field is used to hold flags.
* We return -EAGAIN if force_nonblock is set.

Signed-off-by: Árni Dagur <arni@xxxxxxxx>
---
 fs/io_uring.c                 | 76 +++++++++++++++++++++++++++++++++++
 include/uapi/linux/io_uring.h |  1 +
 2 files changed, 77 insertions(+)

diff --git a/fs/io_uring.c b/fs/io_uring.c
index ca46f314640b..a99a89798386 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -531,6 +531,13 @@ struct io_splice {
 	unsigned int			flags;
 };
 
+struct io_vmsplice {
+	struct file			*file;
+	u64				addr;
+	u64				len;
+	unsigned int			flags;
+};
+
 struct io_provide_buf {
 	struct file			*file;
 	__u64				addr;
@@ -692,6 +699,7 @@ struct io_kiocb {
 		struct io_madvise	madvise;
 		struct io_epoll		epoll;
 		struct io_splice	splice;
+		struct io_vmsplice	vmsplice;
 		struct io_provide_buf	pbuf;
 		struct io_statx		statx;
 		struct io_shutdown	shutdown;
@@ -967,6 +975,12 @@ static const struct io_op_def io_op_defs[] = {
 		.unbound_nonreg_file	= 1,
 		.work_flags		= IO_WQ_WORK_BLKCG,
 	},
+	[IORING_OP_VMSPLICE] = {
+		.needs_file = 1,
+		.hash_reg_file		= 1,
+		.unbound_nonreg_file	= 1,
+		.work_flags		= IO_WQ_WORK_MM,
+	},
 	[IORING_OP_PROVIDE_BUFFERS] = {},
 	[IORING_OP_REMOVE_BUFFERS] = {},
 	[IORING_OP_TEE] = {
@@ -3884,6 +3898,63 @@ static int io_splice(struct io_kiocb *req, bool force_nonblock)
 	return 0;
 }
 
+static int io_vmsplice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
+{
+	struct io_vmsplice *sp = &req->vmsplice;
+
+	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
+		return -EINVAL;
+	if (unlikely(READ_ONCE(sqe->off)))
+		return -EINVAL;
+
+	sp->addr = READ_ONCE(sqe->addr);
+	sp->len = READ_ONCE(sqe->len);
+	sp->flags = READ_ONCE(sqe->splice_flags);
+
+	if (sp->flags & ~SPLICE_F_ALL)
+		return -EINVAL;
+
+	return 0;
+}
+
+static int io_vmsplice(struct io_kiocb *req, bool force_nonblock)
+{
+	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
+	struct io_vmsplice *sp = &req->vmsplice;
+	void __user *buf = u64_to_user_ptr(sp->addr);
+	struct iov_iter __iter, *iter = &__iter;
+	struct file *file = sp->file;
+	ssize_t io_size;
+	int type, ret;
+
+	if (force_nonblock)
+		return -EAGAIN;
+
+	if (file->f_mode & FMODE_WRITE)
+		type = WRITE;
+	else if (file->f_mode & FMODE_READ)
+		type = READ;
+	else {
+		ret = -EBADF;
+		goto err;
+	}
+
+	ret = __import_iovec(type, buf, sp->len, UIO_FASTIOV, &iovec, iter,
+				req->ctx->compat);
+	if (ret < 0)
+		goto err;
+	io_size = iov_iter_count(iter);
+
+	ret = do_vmsplice(file, iter, sp->flags);
+	if (ret != io_size) {
+err:
+		req_set_fail_links(req);
+	}
+	io_req_complete(req, ret);
+	kfree(iovec);
+	return 0;
+}
+
 /*
  * IORING_OP_NOP just posts a completion event, nothing else.
  */
@@ -6009,6 +6080,8 @@ static int io_req_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
 		return io_epoll_ctl_prep(req, sqe);
 	case IORING_OP_SPLICE:
 		return io_splice_prep(req, sqe);
+	case IORING_OP_VMSPLICE:
+		return io_vmsplice_prep(req, sqe);
 	case IORING_OP_PROVIDE_BUFFERS:
 		return io_provide_buffers_prep(req, sqe);
 	case IORING_OP_REMOVE_BUFFERS:
@@ -6262,6 +6335,9 @@ static int io_issue_sqe(struct io_kiocb *req, bool force_nonblock,
 	case IORING_OP_SPLICE:
 		ret = io_splice(req, force_nonblock);
 		break;
+	case IORING_OP_VMSPLICE:
+		ret = io_vmsplice(req, force_nonblock);
+		break;
 	case IORING_OP_PROVIDE_BUFFERS:
 		ret = io_provide_buffers(req, force_nonblock, cs);
 		break;
diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h
index d31a2a1e8ef9..6bc79f9bb123 100644
--- a/include/uapi/linux/io_uring.h
+++ b/include/uapi/linux/io_uring.h
@@ -137,6 +137,7 @@ enum {
 	IORING_OP_SHUTDOWN,
 	IORING_OP_RENAMEAT,
 	IORING_OP_UNLINKAT,
+	IORING_OP_VMSPLICE,
 
 	/* this goes last, obviously */
 	IORING_OP_LAST,
-- 
2.30.0