Splice via io_uring - bad result too

40% CPU according to perf is lost inside do_splice() -> unix_stream_sendpage()
without io_uring and in various exc_page_fault() with io_uring

src/nbd_proxy.cpp

@@ -26,6 +26,18 @@
 
 const char *exe_name = NULL;
 
+static inline void my_uring_prep_splice(struct io_uring_sqe *sqe,
+					int fd_in, int64_t off_in,
+					int fd_out, int64_t off_out,
+					unsigned int nbytes,
+					unsigned int splice_flags)
+{
+	my_uring_prep_rw(IORING_OP_SPLICE, sqe, fd_out, NULL, nbytes, (__u64) off_out);
+	sqe->splice_off_in = (__u64) off_in;
+	sqe->splice_fd_in = fd_in;
+	sqe->splice_flags = splice_flags;
+}
+
 struct buf_to_free_t
 {
     void *buf = NULL;
@@ -61,6 +73,7 @@ protected:
 
     mmap_manager_t mm;
     int pipe_fd[2];
+    int vmspliced = 0;
 
 public:
     static json11::Json::object parse_args(int narg, const char *args[])
@@ -571,12 +584,29 @@ protected:
         }
         else
         {
-            int res = vmsplice(pipe_fd[1], send_list.data(), 1, SPLICE_F_GIFT);
-            if (res < 0)
+            io_uring_sqe* sqe = ringloop->get_sqe();
+            if (!sqe)
             {
-                throw std::runtime_error(std::string("vmsplice: ")+strerror(errno));
+                return;
             }
-            int sent = res, spl = res;
+            if (vmspliced <= 0)
+            {
+                vmspliced = vmsplice(pipe_fd[1], send_list.data(), 1, SPLICE_F_GIFT);
+                if (vmspliced < 0)
+                {
+                    throw std::runtime_error(std::string("vmsplice: ")+strerror(errno));
+                }
+            }
+            send_msg.msg_iovlen = 1;
+            ring_data_t* data = ((ring_data_t*)sqe->user_data);
+            data->callback = [this](ring_data_t *data)
+            {
+                if (data->res > 0)
+                    vmspliced -= data->res;
+                handle_send(data->res);
+            };
+            my_uring_prep_splice(sqe, pipe_fd[0], -1l, nbd_fd, -1l, vmspliced, SPLICE_F_MOVE);
+            /*int sent = res, spl = res;
             while (spl > 0)
             {
                 res = splice(pipe_fd[0], NULL, nbd_fd, NULL, spl, SPLICE_F_MOVE);
@@ -590,7 +620,7 @@ protected:
                     spl -= res;
                 }
             }
-            handle_send(sent);
+            handle_send(sent);*/
         }
     }