On 8/22/22 12:33, Kanchan Joshi wrote:

On Mon, Aug 22, 2022 at 11:58:24AM +0100, Pavel Begunkov wrote:

[...]

quoted

quoted

diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h
index 1463cfecb56b..80ea35d1ed5c 100644
--- a/include/uapi/linux/io_uring.h
+++ b/include/uapi/linux/io_uring.h

@@ -203,6 +203,7 @@ enum io_uring_op {

     IORING_OP_SOCKET,
     IORING_OP_URING_CMD,
     IORING_OP_SENDZC_NOTIF,
+    IORING_OP_URING_CMD_FIXED,

I don't think it should be another opcode, is there any
control flags we can fit it in?

using sqe->rw_flags could be another way.

We also use ->ioprio for io_uring opcode specific flags,
e.g. like in io_sendmsg_prep() for IORING_RECVSEND_POLL_FIRST,
might be even better better.

But I think that may create bit of disharmony in user-space.
Current choice (IORING_OP_URING_CMD_FIXED) is along the same lines as
IORING_OP_READ/WRITE_FIXED.

And I still believe it was a bad choice, I don't like this encoding
of independent options/features by linearising toggles into opcodes.
A consistent way to add vectored fixed bufs would be to have a 4th
opcode, e.g. READV_FIXED, which is not great.

User-space uses new opcode, and sends the
buffer by filling sqe->buf_index. So must we take a different way?

I do think so

quoted

quoted

     /* this goes last, obviously */
     IORING_OP_LAST,

diff --git a/io_uring/opdef.c b/io_uring/opdef.c
index 9a0df19306fe..7d5731b84c92 100644
--- a/io_uring/opdef.c
+++ b/io_uring/opdef.c

@@ -472,6 +472,16 @@ const struct io_op_def io_op_defs[] = {

         .issue            = io_uring_cmd,
         .prep_async        = io_uring_cmd_prep_async,
     },
+    [IORING_OP_URING_CMD_FIXED] = {
+        .needs_file        = 1,
+        .plug            = 1,
+        .name            = "URING_CMD_FIXED",
+        .iopoll            = 1,
+        .async_size        = uring_cmd_pdu_size(1),
+        .prep            = io_uring_cmd_prep,
+        .issue            = io_uring_cmd,
+        .prep_async        = io_uring_cmd_prep_async,
+    },
     [IORING_OP_SENDZC_NOTIF] = {
         .name            = "SENDZC_NOTIF",
         .needs_file        = 1,

diff --git a/io_uring/rw.c b/io_uring/rw.c
index 1a4fb8a44b9a..3c7b94bffa62 100644
--- a/io_uring/rw.c
+++ b/io_uring/rw.c

@@ -1005,7 +1005,8 @@ int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin)

         if (READ_ONCE(req->iopoll_completed))
             break;
-        if (req->opcode == IORING_OP_URING_CMD) {
+        if (req->opcode == IORING_OP_URING_CMD ||
+                req->opcode == IORING_OP_URING_CMD_FIXED) {

I don't see the changed chunk upstream

Right, it is on top of iopoll support (plus one more series mentioned in
covered letter). Here is the link - https://lore.kernel.org/linux-block/20220807183607.352351-1-joshi.k@samsung.com/ (local)
It would be great if you could review that.

-- 
Pavel Begunkov