RDS: Implement silent atomics

Signed-off-by: Andy Grover <andy.grover@oracle.com>

net/rds/ib.h

@@ -336,7 +336,7 @@ void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits);
 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted);
 int rds_ib_send_grab_credits(struct rds_ib_connection *ic, u32 wanted,
 			     u32 *adv_credits, int need_posted, int max_posted);
-int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op);
+int rds_ib_xmit_atomic(struct rds_connection *conn, struct rds_message *rm);
 
 /* ib_stats.c */
 DECLARE_PER_CPU(struct rds_ib_statistics, rds_ib_stats);

net/rds/ib_send.c

@@ -225,15 +225,12 @@ void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context)
 			/* In the error case, wc.opcode sometimes contains garbage */
 			switch (send->s_wr.opcode) {
 			case IB_WR_SEND:
-				if (send->s_rm)
-					rds_ib_send_unmap_rm(ic, send, wc.status);
-				break;
 			case IB_WR_RDMA_WRITE:
 			case IB_WR_RDMA_READ:
 			case IB_WR_ATOMIC_FETCH_AND_ADD:
 			case IB_WR_ATOMIC_CMP_AND_SWP:
-				/* Nothing to be done - the SG list will be unmapped
+				if (send->s_rm)
-				 * when the SEND completes. */
+					rds_ib_send_unmap_rm(ic, send, wc.status);
 				break;
 			default:
 				if (printk_ratelimit())
@@ -425,6 +422,21 @@ void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted)
 		set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
 }
 
+static inline void rds_ib_set_wr_signal_state(struct rds_ib_connection *ic,
+					      struct rds_ib_send_work *send,
+					      bool notify)
+{
+	/*
+	 * We want to delay signaling completions just enough to get
+	 * the batching benefits but not so much that we create dead time
+	 * on the wire.
+	 */
+	if (ic->i_unsignaled_wrs-- == 0 || notify) {
+		ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
+		send->s_wr.send_flags |= IB_SEND_SIGNALED;
+	}
+}
+
 /*
  * This can be called multiple times for a given message.  The first time
  * we see a message we map its scatterlist into the IB device so that
@@ -517,7 +529,6 @@ int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 			rm->data.m_count = 0;
 		}
 
-		ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
 		rds_message_addref(rm);
 		ic->i_rm = rm;
 
@@ -608,15 +619,7 @@ int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 			}
 		}
 
-		/*
+		rds_ib_set_wr_signal_state(ic, send, 0);
-		 * We want to delay signaling completions just enough to get
-		 * the batching benefits but not so much that we create dead time
-		 * on the wire.
-		 */
-		if (ic->i_unsignaled_wrs-- == 0) {
-			ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
-			send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
-		}
 
 		/*
 		 * Always signal the last one if we're stopping due to flow control.
@@ -656,7 +659,7 @@ int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 	/* if we finished the message then send completion owns it */
 	if (scat == &rm->data.m_sg[rm->data.m_count]) {
 		prev->s_rm = ic->i_rm;
-		prev->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
+		prev->s_wr.send_flags |= IB_SEND_SOLICITED;
 		ic->i_rm = NULL;
 	}
 
@@ -698,9 +701,10 @@ out:
  * A simplified version of the rdma case, we always map 1 SG, and
  * only 8 bytes, for the return value from the atomic operation.
  */
-int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op)
+int rds_ib_xmit_atomic(struct rds_connection *conn, struct rds_message *rm)
 {
 	struct rds_ib_connection *ic = conn->c_transport_data;
+	struct rm_atomic_op *op = &rm->atomic;
 	struct rds_ib_send_work *send = NULL;
 	struct ib_send_wr *failed_wr;
 	struct rds_ib_device *rds_ibdev;
@@ -731,12 +735,20 @@ int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op)
 		send->s_wr.wr.atomic.compare_add = op->op_swap_add;
 		send->s_wr.wr.atomic.swap = 0;
 	}
-	send->s_wr.send_flags = IB_SEND_SIGNALED;
+	rds_ib_set_wr_signal_state(ic, send, op->op_notify);
 	send->s_wr.num_sge = 1;
 	send->s_wr.next = NULL;
 	send->s_wr.wr.atomic.remote_addr = op->op_remote_addr;
 	send->s_wr.wr.atomic.rkey = op->op_rkey;
 
+	/*
+	 * If there is no data or rdma ops in the message, then
+	 * we must fill in s_rm ourselves, so we properly clean up
+	 * on completion.
+	 */
+	if (!rm->rdma.m_rdma_op.r_active && !rm->data.op_active)
+		send->s_rm = rm;
+
 	/* map 8 byte retval buffer to the device */
 	ret = ib_dma_map_sg(ic->i_cm_id->device, op->op_sg, 1, DMA_FROM_DEVICE);
 	rdsdebug("ic %p mapping atomic op %p. mapped %d pg\n", ic, op, ret);
@@ -836,14 +848,8 @@ int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
 	for (i = 0; i < work_alloc && scat != &op->r_sg[op->r_count]; i++) {
 		send->s_wr.send_flags = 0;
 		send->s_queued = jiffies;
-		/*
+
-		 * We want to delay signaling completions just enough to get
+		rds_ib_set_wr_signal_state(ic, send, op->r_notify);
-		 * the batching benefits but not so much that we create dead time on the wire.
-		 */
-		if (ic->i_unsignaled_wrs-- == 0) {
-			ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
-			send->s_wr.send_flags = IB_SEND_SIGNALED;
-		}
 
 		send->s_wr.opcode = op->r_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ;
 		send->s_wr.wr.rdma.remote_addr = remote_addr;
@@ -884,10 +890,6 @@ int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
 			send = ic->i_sends;
 	}
 
-	/* if we finished the message then send completion owns it */
-	if (scat == &op->r_sg[op->r_count])
-		prev->s_wr.send_flags = IB_SEND_SIGNALED;
-
 	if (i < work_alloc) {
 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
 		work_alloc = i;

net/rds/message.c

@@ -325,6 +325,8 @@ int rds_message_copy_from_user(struct rds_message *rm, struct iovec *first_iov,
 			sg++;
 	}
 
+	rm->data.op_active = 1;
+
 out:
 	return ret;
 }

net/rds/rds.h

@@ -341,6 +341,7 @@ struct rds_message {
 			struct rds_mr		*m_rdma_mr;
 		} rdma;
 		struct rm_data_op {
+			unsigned int		op_active:1;
 			unsigned int		m_nents;
 			unsigned int		m_count;
 			struct scatterlist	*m_sg;
@@ -418,7 +419,7 @@ struct rds_transport {
 	int (*xmit_cong_map)(struct rds_connection *conn,
 			     struct rds_cong_map *map, unsigned long offset);
 	int (*xmit_rdma)(struct rds_connection *conn, struct rds_rdma_op *op);
-	int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op);
+	int (*xmit_atomic)(struct rds_connection *conn, struct rds_message *rm);
 	int (*recv)(struct rds_connection *conn);
 	int (*inc_copy_to_user)(struct rds_incoming *inc, struct iovec *iov,
 				size_t size);

net/rds/send.c

@@ -266,7 +266,7 @@ int rds_send_xmit(struct rds_connection *conn)
 
 
 		if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
-			ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
+			ret = conn->c_trans->xmit_atomic(conn, rm);
 			if (ret)
 				break;
 			conn->c_xmit_atomic_sent = 1;
@@ -285,13 +285,18 @@ int rds_send_xmit(struct rds_connection *conn)
 			if (ret)
 				break;
 			conn->c_xmit_rdma_sent = 1;
+
+			/* rdmas need data sent, even if just the header */
+			rm->data.op_active = 1;
+
 			/* The transport owns the mapped memory for now.
 			 * You can't unmap it while it's on the send queue */
 			set_bit(RDS_MSG_MAPPED, &rm->m_flags);
 		}
 
-		if (conn->c_xmit_hdr_off < sizeof(struct rds_header) ||
+		if (rm->data.op_active
-		    conn->c_xmit_sg < rm->data.m_nents) {
+		    && (conn->c_xmit_hdr_off < sizeof(struct rds_header) ||
+			conn->c_xmit_sg < rm->data.m_nents)) {
 			ret = conn->c_trans->xmit(conn, rm,
 						  conn->c_xmit_hdr_off,
 						  conn->c_xmit_sg,