io_uring: add support for ring mapped supplied buffers

Provided buffers allow an application to supply io_uring with buffers
that can then be grabbed for a read/receive request, when the data
source is ready to deliver data. The existing scheme relies on using
IORING_OP_PROVIDE_BUFFERS to do that, but it can be difficult to use
in real world applications. It's pretty efficient if the application
is able to supply back batches of provided buffers when they have been
consumed and the application is ready to recycle them, but if
fragmentation occurs in the buffer space, it can become difficult to
supply enough buffers at the time. This hurts efficiency.

Add a register op, IORING_REGISTER_PBUF_RING, which allows an application
to setup a shared queue for each buffer group of provided buffers. The
application can then supply buffers simply by adding them to this ring,
and the kernel can consume then just as easily. The ring shares the head
with the application, the tail remains private in the kernel.

Provided buffers setup with IORING_REGISTER_PBUF_RING cannot use
IORING_OP_{PROVIDE,REMOVE}_BUFFERS for adding or removing entries to the
ring, they must use the mapped ring. Mapped provided buffer rings can
co-exist with normal provided buffers, just not within the same group ID.

To gauge overhead of the existing scheme and evaluate the mapped ring
approach, a simple NOP benchmark was written. It uses a ring of 128
entries, and submits/completes 32 at the time. 'Replenish' is how
many buffers are provided back at the time after they have been
consumed:

Test			Replenish			NOPs/sec
================================================================
No provided buffers	NA				~30M
Provided buffers	32				~16M
Provided buffers	 1				~10M
Ring buffers		32				~27M
Ring buffers		 1				~27M

The ring mapped buffers perform almost as well as not using provided
buffers at all, and they don't care if you provided 1 or more back at
the same time. This means application can just replenish as they go,
rather than need to batch and compact, further reducing overhead in the
application. The NOP benchmark above doesn't need to do any compaction,
so that overhead isn't even reflected in the above test.
Co-developed-by: Dylan Yudaken <dylany@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

include/uapi/linux/io_uring.h

@@ -384,6 +384,10 @@ enum {
 	IORING_REGISTER_RING_FDS		= 20,
 	IORING_UNREGISTER_RING_FDS		= 21,
 
+	/* register ring based provide buffer group */
+	IORING_REGISTER_PBUF_RING		= 22,
+	IORING_UNREGISTER_PBUF_RING		= 23,
+
 	/* this goes last */
 	IORING_REGISTER_LAST
 };
@@ -461,6 +465,38 @@ struct io_uring_restriction {
 	__u32 resv2[3];
 };
 
+struct io_uring_buf {
+	__u64	addr;
+	__u32	len;
+	__u16	bid;
+	__u16	resv;
+};
+
+struct io_uring_buf_ring {
+	union {
+		/*
+		 * To avoid spilling into more pages than we need to, the
+		 * ring tail is overlaid with the io_uring_buf->resv field.
+		 */
+		struct {
+			__u64	resv1;
+			__u32	resv2;
+			__u16	resv3;
+			__u16	tail;
+		};
+		struct io_uring_buf	bufs[0];
+	};
+};
+
+/* argument for IORING_(UN)REGISTER_PBUF_RING */
+struct io_uring_buf_reg {
+	__u64	ring_addr;
+	__u32	ring_entries;
+	__u16	bgid;
+	__u16	pad;
+	__u64	resv[3];
+};
+
 /*
  * io_uring_restriction->opcode values
  */