support for ZeroCopy Send

two changes:

1. use IORING_OP_SENDMSG_ZC instead of IORING_OP_SENDMSG for sending.
2. support for multislot completion notify.

the proactor nature in asio is very suitable for zero-copy send. since
the buffer is assumed to be valid until operation complete.

the only drawback is kernel requirement but this is not an issue. people
tent to upgrade their kernel to take advantage of io-uring anyway.

Author: microcai

asio/include/asio/detail/impl/io_uring_service.ipp

@@ -482,6 +482,18 @@ void io_uring_service::run(long usec, op_queue<operation>& ops)
         {
           --local_ops;
         }
+        else if (cqe->flags & IORING_CQE_F_MORE)
+        {
+          // std::printf("got IORING_CQE_F_MORE\n");
+          io_queue* io_q = static_cast<io_queue*>(ptr);
+          io_q->set_result(cqe->res);
+        }
+        else if (cqe->flags & IORING_CQE_F_NOTIF)
+        {
+          // std::printf("got IORING_CQE_F_NOTIF\n");
+          io_queue* io_q = static_cast<io_queue*>(ptr);
+          ops.push(io_q);
+        }
         else
         {
           io_queue* io_q = static_cast<io_queue*>(ptr);

asio/include/asio/detail/io_uring_socket_send_op.hpp

@@ -73,7 +73,7 @@ class io_uring_socket_send_op_base : public io_uring_operation
     }
     else
     {
-      ::io_uring_prep_sendmsg(sqe, o->socket_, &o->msghdr_, o->flags_);
+      ::io_uring_prep_sendmsg_zc(sqe, o->socket_, &o->msghdr_, o->flags_);
     }
   }
 

asio/include/asio/detail/io_uring_socket_sendto_op.hpp

@@ -70,7 +70,7 @@ class io_uring_socket_sendto_op_base : public io_uring_operation
     }
     else
     {
-      ::io_uring_prep_sendmsg(sqe, o->socket_, &o->msghdr_, o->flags_);
+      ::io_uring_prep_sendmsg_zc(sqe, o->socket_, &o->msghdr_, o->flags_);
     }
   }