LKML Archive on lore.kernel.org
help / color / mirror / Atom feed
From: Sudeep Dutt <sudeep.dutt@intel.com>
To: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Arnd Bergmann <arnd@arndb.de>, Jonathan Corbet <corbet@lwn.net>,
	linux-kernel@vger.kernel.org, linux-doc@vger.kernel.org,
	Dave Jiang <dave.jiang@intel.com>,
	Nikhil Rao <nikhil.rao@intel.com>,
	Ashutosh Dixit <ashutosh.dixit@intel.com>,
	Sudeep Dutt <sudeep.dutt@intel.com>
Subject: [PATCH char-misc-next v4 07/13] misc: mic: SCIF node queue pair setup management
Date: Mon, 30 Mar 2015 18:36:21 -0700	[thread overview]
Message-ID: <421f4566df4e0200401e49845c631672603a9ba1.1427763968.git.sudeep.dutt@intel.com> (raw)
In-Reply-To: <cover.1427763968.git.sudeep.dutt@intel.com>
In-Reply-To: <cover.1427763968.git.sudeep.dutt@intel.com>

SCIF node queue pair setup creates the SCIF driver kernel
mode private node queue pairs between all the nodes to enable
internal control message communication once SCIF gets probed
by the SCIF hardware bus. Peer to peer communication between
MIC Coprocessor nodes is supported.

Reviewed-by: Nikhil Rao <nikhil.rao@intel.com>
Reviewed-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
Signed-off-by: Sudeep Dutt <sudeep.dutt@intel.com>
---
 drivers/misc/mic/scif/scif_nodeqp.h |  173 +++++
 drivers/misc/mic/scif/scif_nm.c     |  184 +++++
 drivers/misc/mic/scif/scif_nodeqp.c | 1264 +++++++++++++++++++++++++++++++++++
 3 files changed, 1621 insertions(+)
 create mode 100644 drivers/misc/mic/scif/scif_nodeqp.h
 create mode 100644 drivers/misc/mic/scif/scif_nm.c
 create mode 100644 drivers/misc/mic/scif/scif_nodeqp.c

diff --git a/drivers/misc/mic/scif/scif_nodeqp.h b/drivers/misc/mic/scif/scif_nodeqp.h
new file mode 100644
index 0000000..6200292
--- /dev/null
+++ b/drivers/misc/mic/scif/scif_nodeqp.h
@@ -0,0 +1,173 @@
+/*
+ * Intel MIC Platform Software Stack (MPSS)
+ *
+ * This file is provided under a dual BSD/GPLv2 license.  When using or
+ * redistributing this file, you may do so under either license.
+ *
+ * GPL LICENSE SUMMARY
+ *
+ * Copyright(c) 2014 Intel Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of version 2 of the GNU General Public License as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * BSD LICENSE
+ *
+ * Copyright(c) 2014 Intel Corporation.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *
+ * * Redistributions of source code must retain the above copyright
+ *   notice, this list of conditions and the following disclaimer.
+ * * Redistributions in binary form must reproduce the above copyright
+ *   notice, this list of conditions and the following disclaimer in
+ *   the documentation and/or other materials provided with the
+ *   distribution.
+ * * Neither the name of Intel Corporation nor the names of its
+ *   contributors may be used to endorse or promote products derived
+ *   from this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ * Intel SCIF driver.
+ *
+ */
+#ifndef SCIF_NODEQP
+#define SCIF_NODEQP
+
+#include "scif_rb.h"
+#include "scif_peer_bus.h"
+
+#define SCIF_INIT 1  /* First message sent to the peer node for discovery */
+#define SCIF_EXIT 2  /* Last message from the peer informing intent to exit */
+#define SCIF_EXIT_ACK 3 /* Response to SCIF_EXIT message */
+#define SCIF_NODE_ADD 4  /* Tell Online nodes a new node exits */
+#define SCIF_NODE_ADD_ACK 5  /* Confirm to mgmt node sequence is finished */
+#define SCIF_NODE_ADD_NACK 6 /* SCIF_NODE_ADD failed */
+#define SCIF_NODE_REMOVE 7 /* Request to deactivate a SCIF node */
+#define SCIF_NODE_REMOVE_ACK 8 /* Response to a SCIF_NODE_REMOVE message */
+#define SCIF_MAX_MSG SCIF_NODE_REMOVE_ACK
+
+/*
+ * struct scifmsg - Node QP message format
+ *
+ * @src: Source information
+ * @dst: Destination information
+ * @uop: The message opcode
+ * @payload: Unique payload format for each message
+ */
+struct scifmsg {
+	struct scif_port_id src;
+	struct scif_port_id dst;
+	u32 uop;
+	u64 payload[4];
+} __packed;
+
+/*
+ * struct scif_qp - Node Queue Pair
+ *
+ * Interesting structure -- a little difficult because we can only
+ * write across the PCIe, so any r/w pointer we need to read is
+ * local. We only need to read the read pointer on the inbound_q
+ * and read the write pointer in the outbound_q
+ *
+ * @magic: Magic value to ensure the peer sees the QP correctly
+ * @outbound_q: The outbound ring buffer for sending messages
+ * @inbound_q: The inbound ring buffer for receiving messages
+ * @local_write: Local write index
+ * @local_read: Local read index
+ * @remote_qp: The remote queue pair
+ * @local_buf: DMA address of local ring buffer
+ * @local_qp: DMA address of the local queue pair data structure
+ * @remote_buf: DMA address of remote ring buffer
+ * @qp_state: QP state i.e. online or offline used for P2P
+ * @send_lock: synchronize access to outbound queue
+ * @recv_lock: Synchronize access to inbound queue
+ */
+struct scif_qp {
+	u64 magic;
+#define SCIFEP_MAGIC 0x5c1f000000005c1fULL
+	struct scif_rb outbound_q;
+	struct scif_rb inbound_q;
+
+	u32 local_write __aligned(64);
+	u32 local_read __aligned(64);
+	struct scif_qp *remote_qp;
+	dma_addr_t local_buf;
+	dma_addr_t local_qp;
+	dma_addr_t remote_buf;
+	u32 qp_state;
+#define SCIF_QP_OFFLINE 0xdead
+#define SCIF_QP_ONLINE 0xc0de
+	spinlock_t send_lock;
+	spinlock_t recv_lock;
+};
+
+/*
+ * struct scif_loopb_msg - An element in the loopback Node QP message list.
+ *
+ * @msg - The SCIF node QP message
+ * @list - link in the list of messages
+ */
+struct scif_loopb_msg {
+	struct scifmsg msg;
+	struct list_head list;
+};
+
+int scif_nodeqp_send(struct scif_dev *scifdev, struct scifmsg *msg);
+int _scif_nodeqp_send(struct scif_dev *scifdev, struct scifmsg *msg);
+void scif_nodeqp_intrhandler(struct scif_dev *scifdev, struct scif_qp *qp);
+int scif_loopb_msg_handler(struct scif_dev *scifdev, struct scif_qp *qp);
+int scif_setup_qp(struct scif_dev *scifdev);
+int scif_qp_response(phys_addr_t phys, struct scif_dev *dev);
+int scif_setup_qp_connect(struct scif_qp *qp, dma_addr_t *qp_offset,
+			  int local_size, struct scif_dev *scifdev);
+int scif_setup_qp_accept(struct scif_qp *qp, dma_addr_t *qp_offset,
+			 dma_addr_t phys, int local_size,
+			 struct scif_dev *scifdev);
+int scif_setup_qp_connect_response(struct scif_dev *scifdev,
+				   struct scif_qp *qp, u64 payload);
+int scif_setup_loopback_qp(struct scif_dev *scifdev);
+int scif_destroy_loopback_qp(struct scif_dev *scifdev);
+void scif_poll_qp_state(struct work_struct *work);
+void scif_qp_response_ack(struct work_struct *work);
+void scif_destroy_p2p(struct scif_dev *scifdev);
+void scif_send_exit(struct scif_dev *scifdev);
+static inline struct device *scif_get_peer_dev(struct scif_dev *scifdev)
+{
+	struct scif_peer_dev *spdev;
+	struct device *spdev_ret;
+
+	rcu_read_lock();
+	spdev = rcu_dereference(scifdev->spdev);
+	if (spdev)
+		spdev_ret = get_device(&spdev->dev);
+	else
+		spdev_ret = ERR_PTR(-ENODEV);
+	rcu_read_unlock();
+	return spdev_ret;
+}
+
+static inline void scif_put_peer_dev(struct device *dev)
+{
+	put_device(dev);
+}
+#endif  /* SCIF_NODEQP */
diff --git a/drivers/misc/mic/scif/scif_nm.c b/drivers/misc/mic/scif/scif_nm.c
new file mode 100644
index 0000000..5fefacc
--- /dev/null
+++ b/drivers/misc/mic/scif/scif_nm.c
@@ -0,0 +1,184 @@
+/*
+ * Intel MIC Platform Software Stack (MPSS)
+ *
+ * Copyright(c) 2014 Intel Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License, version 2, as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * General Public License for more details.
+ *
+ * Intel SCIF driver.
+ *
+ */
+#include "scif_peer_bus.h"
+
+#include "scif_main.h"
+#include "scif_map.h"
+
+void scif_free_qp(struct scif_dev *scifdev)
+{
+	struct scif_qp *qp = scifdev->qpairs;
+
+	if (!qp)
+		return;
+	scif_free_coherent((void *)qp->inbound_q.rb_base,
+			   qp->local_buf, scifdev, qp->inbound_q.size);
+	scif_unmap_single(qp->local_qp, scifdev, sizeof(struct scif_qp));
+	kfree(scifdev->qpairs);
+	scifdev->qpairs = NULL;
+}
+
+static void scif_cleanup_qp(struct scif_dev *dev)
+{
+	struct scif_qp *qp = &dev->qpairs[0];
+
+	if (!qp)
+		return;
+	scif_iounmap((void *)qp->remote_qp, sizeof(struct scif_qp), dev);
+	scif_iounmap((void *)qp->outbound_q.rb_base,
+		     sizeof(struct scif_qp), dev);
+	qp->remote_qp = NULL;
+	qp->local_write = 0;
+	qp->inbound_q.current_write_offset = 0;
+	qp->inbound_q.current_read_offset = 0;
+	if (scifdev_is_p2p(dev))
+		scif_free_qp(dev);
+}
+
+void scif_send_acks(struct scif_dev *dev)
+{
+	struct scifmsg msg;
+
+	if (dev->node_remove_ack_pending) {
+		msg.uop = SCIF_NODE_REMOVE_ACK;
+		msg.src.node = scif_info.nodeid;
+		msg.dst.node = SCIF_MGMT_NODE;
+		msg.payload[0] = dev->node;
+		scif_nodeqp_send(&scif_dev[SCIF_MGMT_NODE], &msg);
+		dev->node_remove_ack_pending = false;
+	}
+	if (dev->exit_ack_pending) {
+		msg.uop = SCIF_EXIT_ACK;
+		msg.src.node = scif_info.nodeid;
+		msg.dst.node = dev->node;
+		scif_nodeqp_send(dev, &msg);
+		dev->exit_ack_pending = false;
+	}
+}
+
+/*
+ * scif_cleanup_scifdev
+ *
+ * @dev: Remote SCIF device.
+ * Uninitialize SCIF data structures for remote SCIF device.
+ */
+void scif_cleanup_scifdev(struct scif_dev *dev)
+{
+	struct scif_hw_dev *sdev = dev->sdev;
+
+	if (!dev->sdev)
+		return;
+	if (scifdev_is_p2p(dev)) {
+		if (dev->cookie) {
+			sdev->hw_ops->free_irq(sdev, dev->cookie, dev);
+			dev->cookie = NULL;
+		}
+		scif_destroy_intr_wq(dev);
+	}
+	scif_destroy_p2p(dev);
+	scif_send_acks(dev);
+	if (!dev->node && scif_info.card_initiated_exit) {
+		/*
+		 * Send an SCIF_EXIT message which is the last message from MIC
+		 * to the Host and wait for a SCIF_EXIT_ACK
+		 */
+		scif_send_exit(dev);
+		scif_info.card_initiated_exit = false;
+	}
+	scif_cleanup_qp(dev);
+}
+
+/*
+ * scif_remove_node:
+ *
+ * @node: Node to remove
+ */
+void scif_handle_remove_node(int node)
+{
+	struct scif_dev *scifdev = &scif_dev[node];
+	struct scif_peer_dev *spdev;
+
+	rcu_read_lock();
+	spdev = rcu_dereference(scifdev->spdev);
+	rcu_read_unlock();
+	if (spdev)
+		scif_peer_unregister_device(spdev);
+	else
+		scif_send_acks(scifdev);
+}
+
+static int scif_send_rmnode_msg(int node, int remove_node)
+{
+	struct scifmsg notif_msg;
+	struct scif_dev *dev = &scif_dev[node];
+
+	notif_msg.uop = SCIF_NODE_REMOVE;
+	notif_msg.src.node = scif_info.nodeid;
+	notif_msg.dst.node = node;
+	notif_msg.payload[0] = remove_node;
+	return scif_nodeqp_send(dev, &notif_msg);
+}
+
+/**
+ * scif_node_disconnect:
+ *
+ * @node_id[in]: source node id.
+ * @mgmt_initiated: Disconnection initiated from the mgmt node
+ *
+ * Disconnect a node from the scif network.
+ */
+void scif_disconnect_node(u32 node_id, bool mgmt_initiated)
+{
+	int ret;
+	int msg_cnt = 0;
+	u32 i = 0;
+	struct scif_dev *scifdev = &scif_dev[node_id];
+
+	if (!node_id)
+		return;
+
+	atomic_set(&scifdev->disconn_rescnt, 0);
+
+	/* Destroy p2p network */
+	for (i = 1; i <= scif_info.maxid; i++) {
+		if (i == node_id)
+			continue;
+		ret = scif_send_rmnode_msg(i, node_id);
+		if (!ret)
+			msg_cnt++;
+	}
+	/* Wait for the remote nodes to respond with SCIF_NODE_REMOVE_ACK */
+	ret = wait_event_timeout(scifdev->disconn_wq,
+				 (atomic_read(&scifdev->disconn_rescnt)
+				 == msg_cnt), SCIF_NODE_ALIVE_TIMEOUT);
+	/* Tell the card to clean up */
+	if (mgmt_initiated && _scifdev_alive(scifdev))
+		/*
+		 * Send an SCIF_EXIT message which is the last message from Host
+		 * to the MIC and wait for a SCIF_EXIT_ACK
+		 */
+		scif_send_exit(scifdev);
+	atomic_set(&scifdev->disconn_rescnt, 0);
+	/* Tell the mgmt node to clean up */
+	ret = scif_send_rmnode_msg(SCIF_MGMT_NODE, node_id);
+	if (!ret)
+		/* Wait for mgmt node to respond with SCIF_NODE_REMOVE_ACK */
+		wait_event_timeout(scifdev->disconn_wq,
+				   (atomic_read(&scifdev->disconn_rescnt) == 1),
+				   SCIF_NODE_ALIVE_TIMEOUT);
+}
diff --git a/drivers/misc/mic/scif/scif_nodeqp.c b/drivers/misc/mic/scif/scif_nodeqp.c
new file mode 100644
index 0000000..0e0cd6b
--- /dev/null
+++ b/drivers/misc/mic/scif/scif_nodeqp.c
@@ -0,0 +1,1264 @@
+/*
+ * Intel MIC Platform Software Stack (MPSS)
+ *
+ * Copyright(c) 2014 Intel Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License, version 2, as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * General Public License for more details.
+ *
+ * Intel SCIF driver.
+ *
+ */
+#include "../bus/scif_bus.h"
+#include "scif_peer_bus.h"
+#include "scif_main.h"
+#include "scif_nodeqp.h"
+#include "scif_map.h"
+
+/*
+ ************************************************************************
+ * SCIF node Queue Pair (QP) setup flow:
+ *
+ * 1) SCIF driver gets probed with a scif_hw_dev via the scif_hw_bus
+ * 2) scif_setup_qp(..) allocates the local qp and calls
+ *	scif_setup_qp_connect(..) which allocates and maps the local
+ *	buffer for the inbound QP
+ * 3) The local node updates the device page with the DMA address of the QP
+ * 4) A delayed work is scheduled (qp_dwork) which periodically reads if
+ *	the peer node has updated its QP DMA address
+ * 5) Once a valid non zero address is found in the QP DMA address field
+ *	in the device page, the local node maps the remote node's QP,
+ *	updates its outbound QP and sends a SCIF_INIT message to the peer
+ * 6) The SCIF_INIT message is received by the peer node QP interrupt bottom
+ *	half handler by calling scif_init(..)
+ * 7) scif_init(..) registers a new SCIF peer node by calling
+ *	scif_peer_register_device(..) which signifies the addition of a new
+ *	SCIF node
+ * 8) On the mgmt node, P2P network setup/teardown is initiated if all the
+ *	remote nodes are online via scif_p2p_setup(..)
+ * 9) For P2P setup, the host maps the remote nodes' aperture and memory
+ *	bars and sends a SCIF_NODE_ADD message to both nodes
+ * 10) As part of scif_nodeadd, both nodes set up their local inbound
+ *	QPs and send a SCIF_NODE_ADD_ACK to the mgmt node
+ * 11) As part of scif_node_add_ack(..) the mgmt node forwards the
+ *	SCIF_NODE_ADD_ACK to the remote nodes
+ * 12) As part of scif_node_add_ack(..) the remote nodes update their
+ *	outbound QPs, make sure they can access memory on the remote node
+ *	and then add a new SCIF peer node by calling
+ *	scif_peer_register_device(..) which signifies the addition of a new
+ *	SCIF node.
+ * 13) The SCIF network is now established across all nodes.
+ *
+ ************************************************************************
+ * SCIF node QP teardown flow (initiated by non mgmt node):
+ *
+ * 1) SCIF driver gets a remove callback with a scif_hw_dev via the scif_hw_bus
+ * 2) The device page QP DMA address field is updated with 0x0
+ * 3) A non mgmt node now cleans up all local data structures and sends a
+ *	SCIF_EXIT message to the peer and waits for a SCIF_EXIT_ACK
+ * 4) As part of scif_exit(..) handling scif_disconnect_node(..) is called
+ * 5) scif_disconnect_node(..) sends a SCIF_NODE_REMOVE message to all the
+ *	peers and waits for a SCIF_NODE_REMOVE_ACK
+ * 6) As part of scif_node_remove(..) a remote node unregisters the peer
+ *	node from the SCIF network and sends a SCIF_NODE_REMOVE_ACK
+ * 7) When the mgmt node has received all the SCIF_NODE_REMOVE_ACKs
+ *	it sends itself a node remove message whose handling cleans up local
+ *	data structures and unregisters the peer node from the SCIF network
+ * 8) The mgmt node sends a SCIF_EXIT_ACK
+ * 9) Upon receipt of the SCIF_EXIT_ACK the node initiating the teardown
+ *	completes the SCIF remove routine
+ * 10) The SCIF network is now torn down for the node initiating the
+ *	teardown sequence
+ *
+ ************************************************************************
+ * SCIF node QP teardown flow (initiated by mgmt node):
+ *
+ * 1) SCIF driver gets a remove callback with a scif_hw_dev via the scif_hw_bus
+ * 2) The device page QP DMA address field is updated with 0x0
+ * 3) The mgmt node calls scif_disconnect_node(..)
+ * 4) scif_disconnect_node(..) sends a SCIF_NODE_REMOVE message to all the peers
+ *	and waits for a SCIF_NODE_REMOVE_ACK
+ * 5) As part of scif_node_remove(..) a remote node unregisters the peer
+ *	node from the SCIF network and sends a SCIF_NODE_REMOVE_ACK
+ * 6) When the mgmt node has received all the SCIF_NODE_REMOVE_ACKs
+ *	it unregisters the peer node from the SCIF network
+ * 7) The mgmt node sends a SCIF_EXIT message and waits for a SCIF_EXIT_ACK.
+ * 8) A non mgmt node upon receipt of a SCIF_EXIT message calls scif_stop(..)
+ *	which would clean up local data structures for all SCIF nodes and
+ *	then send a SCIF_EXIT_ACK back to the mgmt node
+ * 9) Upon receipt of the SCIF_EXIT_ACK the the mgmt node sends itself a node
+ *	remove message whose handling cleans up local data structures and
+ *	destroys any P2P mappings.
+ * 10) The SCIF hardware device for which a remove callback was received is now
+ *	disconnected from the SCIF network.
+ */
+/*
+ * Initializes "local" data structures for the QP. Allocates the QP
+ * ring buffer (rb) and initializes the "in bound" queue.
+ */
+int scif_setup_qp_connect(struct scif_qp *qp, dma_addr_t *qp_offset,
+			  int local_size, struct scif_dev *scifdev)
+{
+	void *local_q = NULL;
+	int err = 0;
+	u32 tmp_rd = 0;
+
+	spin_lock_init(&qp->send_lock);
+	spin_lock_init(&qp->recv_lock);
+
+	local_q = kzalloc(local_size, GFP_KERNEL);
+	if (!local_q) {
+		err = -ENOMEM;
+		return err;
+	}
+	err = scif_map_single(&qp->local_buf, local_q, scifdev, local_size);
+	if (err)
+		goto kfree;
+	/*
+	 * To setup the inbound_q, the buffer lives locally, the read pointer
+	 * is remote and the write pointer is local.
+	 */
+	scif_rb_init(&qp->inbound_q,
+		     &tmp_rd,
+		     &qp->local_write,
+		     local_q, get_count_order(local_size));
+	/*
+	 * The read pointer is NULL initially and it is unsafe to use the ring
+	 * buffer til this changes!
+	 */
+	qp->inbound_q.read_ptr = NULL;
+	err = scif_map_single(qp_offset, qp,
+			      scifdev, sizeof(struct scif_qp));
+	if (err)
+		goto unmap;
+	qp->local_qp = *qp_offset;
+	return err;
+unmap:
+	scif_unmap_single(qp->local_buf, scifdev, local_size);
+	qp->local_buf = 0;
+kfree:
+	kfree(local_q);
+	return err;
+}
+
+/* When the other side has already done it's allocation, this is called */
+int scif_setup_qp_accept(struct scif_qp *qp, dma_addr_t *qp_offset,
+			 dma_addr_t phys, int local_size,
+			 struct scif_dev *scifdev)
+{
+	void *local_q;
+	void *remote_q;
+	struct scif_qp *remote_qp;
+	int remote_size;
+	int err = 0;
+
+	spin_lock_init(&qp->send_lock);
+	spin_lock_init(&qp->recv_lock);
+	/* Start by figuring out where we need to point */
+	remote_qp = scif_ioremap(phys, sizeof(struct scif_qp), scifdev);
+	if (!remote_qp)
+		return -EIO;
+	qp->remote_qp = remote_qp;
+	if (qp->remote_qp->magic != SCIFEP_MAGIC) {
+		err = -EIO;
+		goto iounmap;
+	}
+	qp->remote_buf = remote_qp->local_buf;
+	remote_size = qp->remote_qp->inbound_q.size;
+	remote_q = scif_ioremap(qp->remote_buf, remote_size, scifdev);
+	if (!remote_q) {
+		err = -EIO;
+		goto iounmap;
+	}
+	qp->remote_qp->local_write = 0;
+	/*
+	 * To setup the outbound_q, the buffer lives in remote memory,
+	 * the read pointer is local, the write pointer is remote
+	 */
+	scif_rb_init(&qp->outbound_q,
+		     &qp->local_read,
+		     &qp->remote_qp->local_write,
+		     remote_q,
+		     get_count_order(remote_size));
+	local_q = kzalloc(local_size, GFP_KERNEL);
+	if (!local_q) {
+		err = -ENOMEM;
+		goto iounmap_1;
+	}
+	err = scif_map_single(&qp->local_buf, local_q, scifdev, local_size);
+	if (err)
+		goto kfree;
+	qp->remote_qp->local_read = 0;
+	/*
+	 * To setup the inbound_q, the buffer lives locally, the read pointer
+	 * is remote and the write pointer is local
+	 */
+	scif_rb_init(&qp->inbound_q,
+		     &qp->remote_qp->local_read,
+		     &qp->local_write,
+		     local_q, get_count_order(local_size));
+	err = scif_map_single(qp_offset, qp, scifdev,
+			      sizeof(struct scif_qp));
+	if (err)
+		goto unmap;
+	qp->local_qp = *qp_offset;
+	return err;
+unmap:
+	scif_unmap_single(qp->local_buf, scifdev, local_size);
+	qp->local_buf = 0;
+kfree:
+	kfree(local_q);
+iounmap_1:
+	scif_iounmap(remote_q, remote_size, scifdev);
+	qp->outbound_q.rb_base = NULL;
+iounmap:
+	scif_iounmap(qp->remote_qp, sizeof(struct scif_qp), scifdev);
+	qp->remote_qp = NULL;
+	return err;
+}
+
+int scif_setup_qp_connect_response(struct scif_dev *scifdev,
+				   struct scif_qp *qp, u64 payload)
+{
+	int err = 0;
+	void *r_buf;
+	int remote_size;
+	phys_addr_t tmp_phys;
+
+	qp->remote_qp = scif_ioremap(payload, sizeof(struct scif_qp), scifdev);
+
+	if (!qp->remote_qp) {
+		err = -ENOMEM;
+		goto error;
+	}
+
+	if (qp->remote_qp->magic != SCIFEP_MAGIC) {
+		dev_err(&scifdev->sdev->dev,
+			"SCIFEP_MAGIC mismatch between self %d remote %d\n",
+			scif_dev[scif_info.nodeid].node, scifdev->node);
+		err = -ENODEV;
+		goto error;
+	}
+
+	tmp_phys = qp->remote_qp->local_buf;
+	remote_size = qp->remote_qp->inbound_q.size;
+	r_buf = scif_ioremap(tmp_phys, remote_size, scifdev);
+
+	if (!r_buf)
+		return -EIO;
+
+	qp->local_read = 0;
+	scif_rb_init(&qp->outbound_q,
+		     &qp->local_read,
+		     &qp->remote_qp->local_write,
+		     r_buf,
+		     get_count_order(remote_size));
+	/*
+	 * resetup the inbound_q now that we know where the
+	 * inbound_read really is.
+	 */
+	scif_rb_init(&qp->inbound_q,
+		     &qp->remote_qp->local_read,
+		     &qp->local_write,
+		     qp->inbound_q.rb_base,
+		     get_count_order(qp->inbound_q.size));
+error:
+	return err;
+}
+
+static __always_inline void
+scif_send_msg_intr(struct scif_dev *scifdev)
+{
+	struct scif_hw_dev *sdev = scifdev->sdev;
+
+	if (scifdev_is_p2p(scifdev))
+		sdev->hw_ops->send_p2p_intr(sdev, scifdev->rdb, &scifdev->mmio);
+	else
+		sdev->hw_ops->send_intr(sdev, scifdev->rdb);
+}
+
+int scif_qp_response(phys_addr_t phys, struct scif_dev *scifdev)
+{
+	int err = 0;
+	struct scifmsg msg;
+
+	err = scif_setup_qp_connect_response(scifdev, scifdev->qpairs, phys);
+	if (!err) {
+		/*
+		 * Now that everything is setup and mapped, we're ready
+		 * to tell the peer about our queue's location
+		 */
+		msg.uop = SCIF_INIT;
+		msg.dst.node = scifdev->node;
+		err = scif_nodeqp_send(scifdev, &msg);
+	}
+	return err;
+}
+
+void scif_send_exit(struct scif_dev *scifdev)
+{
+	struct scifmsg msg;
+	int ret;
+
+	scifdev->exit = OP_IN_PROGRESS;
+	msg.uop = SCIF_EXIT;
+	msg.src.node = scif_info.nodeid;
+	msg.dst.node = scifdev->node;
+	ret = scif_nodeqp_send(scifdev, &msg);
+	if (ret)
+		goto done;
+	/* Wait for a SCIF_EXIT_ACK message */
+	wait_event_timeout(scif_info.exitwq, scifdev->exit == OP_COMPLETED,
+			   SCIF_NODE_ALIVE_TIMEOUT);
+done:
+	scifdev->exit = OP_IDLE;
+}
+
+int scif_setup_qp(struct scif_dev *scifdev)
+{
+	int err = 0;
+	int local_size;
+	struct scif_qp *qp;
+
+	local_size = SCIF_NODE_QP_SIZE;
+
+	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
+	if (!qp) {
+		err = -ENOMEM;
+		return err;
+	}
+	qp->magic = SCIFEP_MAGIC;
+	scifdev->qpairs = qp;
+	err = scif_setup_qp_connect(qp, &scifdev->qp_dma_addr,
+				    local_size, scifdev);
+	if (err)
+		goto free_qp;
+	/*
+	 * We're as setup as we can be. The inbound_q is setup, w/o a usable
+	 * outbound q.  When we get a message, the read_ptr will be updated,
+	 * and we will pull the message.
+	 */
+	return err;
+free_qp:
+	kfree(scifdev->qpairs);
+	scifdev->qpairs = NULL;
+	return err;
+}
+
+static void scif_p2p_freesg(struct scatterlist *sg)
+{
+	kfree(sg);
+}
+
+static struct scatterlist *
+scif_p2p_setsg(void __iomem *va, int page_size, int page_cnt)
+{
+	struct scatterlist *sg;
+	struct page *page;
+	int i;
+
+	sg = kcalloc(page_cnt, sizeof(struct scatterlist), GFP_KERNEL);
+	if (!sg)
+		return NULL;
+	sg_init_table(sg, page_cnt);
+	for (i = 0; i < page_cnt; i++) {
+		page = vmalloc_to_page((void __force *)va);
+		if (!page)
+			goto p2p_sg_err;
+		sg_set_page(&sg[i], page, page_size, 0);
+		va += page_size;
+	}
+	return sg;
+p2p_sg_err:
+	kfree(sg);
+	return NULL;
+}
+
+/* Init p2p mappings required to access peerdev from scifdev */
+static struct scif_p2p_info *
+scif_init_p2p_info(struct scif_dev *scifdev, struct scif_dev *peerdev)
+{
+	struct scif_p2p_info *p2p;
+	int num_mmio_pages, num_aper_pages, sg_page_shift, err, num_aper_chunks;
+	struct scif_hw_dev *psdev = peerdev->sdev;
+	struct scif_hw_dev *sdev = scifdev->sdev;
+
+	num_mmio_pages = psdev->mmio->len >> PAGE_SHIFT;
+	num_aper_pages = psdev->aper->len >> PAGE_SHIFT;
+
+	p2p = kzalloc(sizeof(*p2p), GFP_KERNEL);
+	if (!p2p)
+		return NULL;
+	p2p->ppi_sg[SCIF_PPI_MMIO] = scif_p2p_setsg(psdev->mmio->va,
+						    PAGE_SIZE, num_mmio_pages);
+	if (!p2p->ppi_sg[SCIF_PPI_MMIO])
+		goto free_p2p;
+	p2p->sg_nentries[SCIF_PPI_MMIO] = num_mmio_pages;
+	sg_page_shift = get_order(min(psdev->aper->len, (u64)(1 << 30)));
+	num_aper_chunks = num_aper_pages >> (sg_page_shift - PAGE_SHIFT);
+	p2p->ppi_sg[SCIF_PPI_APER] = scif_p2p_setsg(psdev->aper->va,
+						    1 << sg_page_shift,
+						    num_aper_chunks);
+	p2p->sg_nentries[SCIF_PPI_APER] = num_aper_chunks;
+	err = dma_map_sg(&sdev->dev, p2p->ppi_sg[SCIF_PPI_MMIO],
+			 num_mmio_pages, PCI_DMA_BIDIRECTIONAL);
+	if (err != num_mmio_pages)
+		goto scif_p2p_free;
+	err = dma_map_sg(&sdev->dev, p2p->ppi_sg[SCIF_PPI_APER],
+			 num_aper_chunks, PCI_DMA_BIDIRECTIONAL);
+	if (err != num_aper_chunks)
+		goto dma_unmap;
+	p2p->ppi_da[SCIF_PPI_MMIO] = sg_dma_address(p2p->ppi_sg[SCIF_PPI_MMIO]);
+	p2p->ppi_da[SCIF_PPI_APER] = sg_dma_address(p2p->ppi_sg[SCIF_PPI_APER]);
+	p2p->ppi_len[SCIF_PPI_MMIO] = num_mmio_pages;
+	p2p->ppi_len[SCIF_PPI_APER] = num_aper_pages;
+	p2p->ppi_peer_id = peerdev->node;
+	return p2p;
+dma_unmap:
+	dma_unmap_sg(&sdev->dev, p2p->ppi_sg[SCIF_PPI_MMIO],
+		     p2p->sg_nentries[SCIF_PPI_MMIO], DMA_BIDIRECTIONAL);
+scif_p2p_free:
+	scif_p2p_freesg(p2p->ppi_sg[SCIF_PPI_MMIO]);
+	scif_p2p_freesg(p2p->ppi_sg[SCIF_PPI_APER]);
+free_p2p:
+	kfree(p2p);
+	return NULL;
+}
+
+/**
+ * scif_node_connect: Respond to SCIF_NODE_CONNECT interrupt message
+ * @dst: Destination node
+ *
+ * Connect the src and dst node by setting up the p2p connection
+ * between them. Management node here acts like a proxy.
+ */
+static void scif_node_connect(struct scif_dev *scifdev, int dst)
+{
+	struct scif_dev *dev_j = scifdev;
+	struct scif_dev *dev_i = NULL;
+	struct scif_p2p_info *p2p_ij = NULL;    /* bus addr for j from i */
+	struct scif_p2p_info *p2p_ji = NULL;    /* bus addr for i from j */
+	struct scif_p2p_info *p2p;
+	struct list_head *pos, *tmp;
+	struct scifmsg msg;
+	int err;
+	u64 tmppayload;
+
+	if (dst < 1 || dst > scif_info.maxid)
+		return;
+
+	dev_i = &scif_dev[dst];
+
+	if (!_scifdev_alive(dev_i))
+		return;
+	/*
+	 * If the p2p connection is already setup or in the process of setting
+	 * up then just ignore this request. The requested node will get
+	 * informed by SCIF_NODE_ADD_ACK or SCIF_NODE_ADD_NACK
+	 */
+	if (!list_empty(&dev_i->p2p)) {
+		list_for_each_safe(pos, tmp, &dev_i->p2p) {
+			p2p = list_entry(pos, struct scif_p2p_info, ppi_list);
+			if (p2p->ppi_peer_id == dev_j->node)
+				return;
+		}
+	}
+	p2p_ij = scif_init_p2p_info(dev_i, dev_j);
+	if (!p2p_ij)
+		return;
+	p2p_ji = scif_init_p2p_info(dev_j, dev_i);
+	if (!p2p_ji)
+		return;
+	list_add_tail(&p2p_ij->ppi_list, &dev_i->p2p);
+	list_add_tail(&p2p_ji->ppi_list, &dev_j->p2p);
+
+	/*
+	 * Send a SCIF_NODE_ADD to dev_i, pass it its bus address
+	 * as seen from dev_j
+	 */
+	msg.uop = SCIF_NODE_ADD;
+	msg.src.node = dev_j->node;
+	msg.dst.node = dev_i->node;
+
+	msg.payload[0] = p2p_ji->ppi_da[SCIF_PPI_APER];
+	msg.payload[1] = p2p_ij->ppi_da[SCIF_PPI_MMIO];
+	msg.payload[2] = p2p_ij->ppi_da[SCIF_PPI_APER];
+	msg.payload[3] = p2p_ij->ppi_len[SCIF_PPI_APER] << PAGE_SHIFT;
+
+	err = scif_nodeqp_send(dev_i,  &msg);
+	if (err) {
+		dev_err(&scifdev->sdev->dev,
+			"%s %d error %d\n", __func__, __LINE__, err);
+		return;
+	}
+
+	/* Same as above but to dev_j */
+	msg.uop = SCIF_NODE_ADD;
+	msg.src.node = dev_i->node;
+	msg.dst.node = dev_j->node;
+
+	tmppayload = msg.payload[0];
+	msg.payload[0] = msg.payload[2];
+	msg.payload[2] = tmppayload;
+	msg.payload[1] = p2p_ji->ppi_da[SCIF_PPI_MMIO];
+	msg.payload[3] = p2p_ji->ppi_len[SCIF_PPI_APER] << PAGE_SHIFT;
+
+	scif_nodeqp_send(dev_j, &msg);
+}
+
+static void scif_p2p_setup(void)
+{
+	int i, j;
+
+	if (!scif_info.p2p_enable)
+		return;
+
+	for (i = 1; i <= scif_info.maxid; i++)
+		if (!_scifdev_alive(&scif_dev[i]))
+			return;
+
+	for (i = 1; i <= scif_info.maxid; i++) {
+		for (j = 1; j <= scif_info.maxid; j++) {
+			struct scif_dev *scifdev = &scif_dev[i];
+
+			if (i == j)
+				continue;
+			scif_node_connect(scifdev, j);
+		}
+	}
+}
+
+void scif_qp_response_ack(struct work_struct *work)
+{
+	struct scif_dev *scifdev = container_of(work, struct scif_dev,
+						init_msg_work);
+	struct scif_peer_dev *spdev;
+
+	/* Drop the INIT message if it has already been received */
+	if (_scifdev_alive(scifdev))
+		return;
+
+	spdev = scif_peer_register_device(scifdev);
+	if (IS_ERR(spdev))
+		return;
+
+	if (scif_is_mgmt_node()) {
+		mutex_lock(&scif_info.conflock);
+		scif_p2p_setup();
+		mutex_unlock(&scif_info.conflock);
+	}
+}
+
+static char *message_types[] = {"BAD",
+				"INIT",
+				"EXIT",
+				"SCIF_EXIT_ACK",
+				"SCIF_NODE_ADD",
+				"SCIF_NODE_ADD_ACK",
+				"SCIF_NODE_ADD_NACK",
+				"REMOVE_NODE",
+				"REMOVE_NODE_ACK"};
+
+static void
+scif_display_message(struct scif_dev *scifdev, struct scifmsg *msg,
+		     const char *label)
+{
+	if (!scif_info.en_msg_log)
+		return;
+	if (msg->uop > SCIF_MAX_MSG) {
+		dev_err(&scifdev->sdev->dev,
+			"%s: unknown msg type %d\n", label, msg->uop);
+		return;
+	}
+	dev_info(&scifdev->sdev->dev,
+		 "%s: msg type %s, src %d:%d, dest %d:%d payload 0x%llx:0x%llx:0x%llx:0x%llx\n",
+		 label, message_types[msg->uop], msg->src.node, msg->src.port,
+		 msg->dst.node, msg->dst.port, msg->payload[0], msg->payload[1],
+		 msg->payload[2], msg->payload[3]);
+}
+
+int _scif_nodeqp_send(struct scif_dev *scifdev, struct scifmsg *msg)
+{
+	struct scif_qp *qp = scifdev->qpairs;
+	int err = -ENOMEM, loop_cnt = 0;
+
+	scif_display_message(scifdev, msg, "Sent");
+	if (!qp) {
+		err = -EINVAL;
+		goto error;
+	}
+	spin_lock(&qp->send_lock);
+
+	while ((err = scif_rb_write(&qp->outbound_q,
+				    msg, sizeof(struct scifmsg)))) {
+		mdelay(1);
+#define SCIF_NODEQP_SEND_TO_MSEC (3 * 1000)
+		if (loop_cnt++ > (SCIF_NODEQP_SEND_TO_MSEC)) {
+			err = -ENODEV;
+			break;
+		}
+	}
+	if (!err)
+		scif_rb_commit(&qp->outbound_q);
+	spin_unlock(&qp->send_lock);
+	if (!err) {
+		if (scifdev_self(scifdev))
+			/*
+			 * For loopback we need to emulate an interrupt by
+			 * queuing work for the queue handling real node
+			 * Qp interrupts.
+			 */
+			queue_work(scifdev->intr_wq, &scifdev->intr_bh);
+		else
+			scif_send_msg_intr(scifdev);
+	}
+error:
+	if (err)
+		dev_dbg(&scifdev->sdev->dev,
+			"%s %d error %d uop %d\n",
+			 __func__, __LINE__, err, msg->uop);
+	return err;
+}
+
+/**
+ * scif_nodeqp_send - Send a message on the node queue pair
+ * @scifdev: Scif Device.
+ * @msg: The message to be sent.
+ */
+int scif_nodeqp_send(struct scif_dev *scifdev, struct scifmsg *msg)
+{
+	int err;
+	struct device *spdev = NULL;
+
+	if (msg->uop > SCIF_EXIT_ACK) {
+		/* Dont send messages once the exit flow has begun */
+		if (OP_IDLE != scifdev->exit)
+			return -ENODEV;
+		spdev = scif_get_peer_dev(scifdev);
+		if (IS_ERR(spdev)) {
+			err = PTR_ERR(spdev);
+			return err;
+		}
+	}
+	err = _scif_nodeqp_send(scifdev, msg);
+	if (msg->uop > SCIF_EXIT_ACK)
+		scif_put_peer_dev(spdev);
+	return err;
+}
+
+/*
+ * scif_misc_handler:
+ *
+ * Work queue handler for servicing miscellaneous SCIF tasks.
+ * Examples include:
+ * 1) Cleanup of zombie endpoints.
+ */
+void scif_misc_handler(struct work_struct *work)
+{
+	scif_cleanup_zombie_epd();
+}
+
+/**
+ * scif_init() - Respond to SCIF_INIT interrupt message
+ * @scifdev:    Remote SCIF device node
+ * @msg:        Interrupt message
+ */
+static __always_inline void
+scif_init(struct scif_dev *scifdev, struct scifmsg *msg)
+{
+	/*
+	 * Allow the thread waiting for device page updates for the peer QP DMA
+	 * address to complete initializing the inbound_q.
+	 */
+	flush_delayed_work(&scifdev->qp_dwork);
+	/*
+	 * Delegate the peer device registration to a workqueue, otherwise if
+	 * SCIF client probe (called during peer device registration) calls
+	 * scif_connect(..), it will block the message processing thread causing
+	 * a deadlock.
+	 */
+	schedule_work(&scifdev->init_msg_work);
+}
+
+/**
+ * scif_exit() - Respond to SCIF_EXIT interrupt message
+ * @scifdev:    Remote SCIF device node
+ * @msg:        Interrupt message
+ *
+ * This function stops the SCIF interface for the node which sent
+ * the SCIF_EXIT message and starts waiting for that node to
+ * resetup the queue pair again.
+ */
+static __always_inline void
+scif_exit(struct scif_dev *scifdev, struct scifmsg *unused)
+{
+	scifdev->exit_ack_pending = true;
+	if (scif_is_mgmt_node())
+		scif_disconnect_node(scifdev->node, false);
+	else
+		scif_stop(scifdev);
+	schedule_delayed_work(&scifdev->qp_dwork,
+			      msecs_to_jiffies(1000));
+}
+
+/**
+ * scif_exitack() - Respond to SCIF_EXIT_ACK interrupt message
+ * @scifdev:    Remote SCIF device node
+ * @msg:        Interrupt message
+ *
+ */
+static __always_inline void
+scif_exit_ack(struct scif_dev *scifdev, struct scifmsg *unused)
+{
+	scifdev->exit = OP_COMPLETED;
+	wake_up(&scif_info.exitwq);
+}
+
+/**
+ * scif_node_add() - Respond to SCIF_NODE_ADD interrupt message
+ * @scifdev:    Remote SCIF device node
+ * @msg:        Interrupt message
+ *
+ * When the mgmt node driver has finished initializing a MIC node queue pair it
+ * marks the node as online. It then looks for all currently online MIC cards
+ * and send a SCIF_NODE_ADD message to identify the ID of the new card for
+ * peer to peer initialization
+ *
+ * The local node allocates its incoming queue and sends its address in the
+ * SCIF_NODE_ADD_ACK message back to the mgmt node, the mgmt node "reflects"
+ * this message to the new node
+ */
+static __always_inline void
+scif_node_add(struct scif_dev *scifdev, struct scifmsg *msg)
+{
+	struct scif_dev *newdev;
+	dma_addr_t qp_offset;
+	int qp_connect;
+	struct scif_hw_dev *sdev;
+
+	dev_dbg(&scifdev->sdev->dev,
+		"Scifdev %d:%d received NODE_ADD msg for node %d\n",
+		scifdev->node, msg->dst.node, msg->src.node);
+	dev_dbg(&scifdev->sdev->dev,
+		"Remote address for this node's aperture %llx\n",
+		msg->payload[0]);
+	newdev = &scif_dev[msg->src.node];
+	newdev->node = msg->src.node;
+	newdev->sdev = scif_dev[SCIF_MGMT_NODE].sdev;
+	sdev = newdev->sdev;
+
+	if (scif_setup_intr_wq(newdev)) {
+		dev_err(&scifdev->sdev->dev,
+			"failed to setup interrupts for %d\n", msg->src.node);
+		goto interrupt_setup_error;
+	}
+	newdev->mmio.va = ioremap_nocache(msg->payload[1], sdev->mmio->len);
+	if (!newdev->mmio.va) {
+		dev_err(&scifdev->sdev->dev,
+			"failed to map mmio for %d\n", msg->src.node);
+		goto mmio_map_error;
+	}
+	newdev->qpairs = kzalloc(sizeof(*newdev->qpairs), GFP_KERNEL);
+	if (!newdev->qpairs)
+		goto qp_alloc_error;
+	/*
+	 * Set the base address of the remote node's memory since it gets
+	 * added to qp_offset
+	 */
+	newdev->base_addr = msg->payload[0];
+
+	qp_connect = scif_setup_qp_connect(newdev->qpairs, &qp_offset,
+					   SCIF_NODE_QP_SIZE, newdev);
+	if (qp_connect) {
+		dev_err(&scifdev->sdev->dev,
+			"failed to setup qp_connect %d\n", qp_connect);
+		goto qp_connect_error;
+	}
+
+	newdev->db = sdev->hw_ops->next_db(sdev);
+	newdev->cookie = sdev->hw_ops->request_irq(sdev, scif_intr_handler,
+						   "SCIF_INTR", newdev,
+						   newdev->db);
+	if (IS_ERR(newdev->cookie))
+		goto qp_connect_error;
+	newdev->qpairs->magic = SCIFEP_MAGIC;
+	newdev->qpairs->qp_state = SCIF_QP_OFFLINE;
+
+	msg->uop = SCIF_NODE_ADD_ACK;
+	msg->dst.node = msg->src.node;
+	msg->src.node = scif_info.nodeid;
+	msg->payload[0] = qp_offset;
+	msg->payload[2] = newdev->db;
+	scif_nodeqp_send(&scif_dev[SCIF_MGMT_NODE], msg);
+	return;
+qp_connect_error:
+	kfree(newdev->qpairs);
+	newdev->qpairs = NULL;
+qp_alloc_error:
+	iounmap(newdev->mmio.va);
+	newdev->mmio.va = NULL;
+mmio_map_error:
+interrupt_setup_error:
+	dev_err(&scifdev->sdev->dev,
+		"node add failed for node %d\n", msg->src.node);
+	msg->uop = SCIF_NODE_ADD_NACK;
+	msg->dst.node = msg->src.node;
+	msg->src.node = scif_info.nodeid;
+	scif_nodeqp_send(&scif_dev[SCIF_MGMT_NODE], msg);
+}
+
+void scif_poll_qp_state(struct work_struct *work)
+{
+#define SCIF_NODE_QP_RETRY 100
+#define SCIF_NODE_QP_TIMEOUT 100
+	struct scif_dev *peerdev = container_of(work, struct scif_dev,
+							p2p_dwork.work);
+	struct scif_qp *qp = &peerdev->qpairs[0];
+
+	if (qp->qp_state != SCIF_QP_ONLINE ||
+	    qp->remote_qp->qp_state != SCIF_QP_ONLINE) {
+		if (peerdev->p2p_retry++ == SCIF_NODE_QP_RETRY) {
+			dev_err(&peerdev->sdev->dev,
+				"Warning: QP check timeout with state %d\n",
+				qp->qp_state);
+			goto timeout;
+		}
+		schedule_delayed_work(&peerdev->p2p_dwork,
+				      msecs_to_jiffies(SCIF_NODE_QP_TIMEOUT));
+		return;
+	}
+	scif_peer_register_device(peerdev);
+	return;
+timeout:
+	dev_err(&peerdev->sdev->dev,
+		"%s %d remote node %d offline,  state = 0x%x\n",
+		__func__, __LINE__, peerdev->node, qp->qp_state);
+	qp->remote_qp->qp_state = SCIF_QP_OFFLINE;
+	scif_cleanup_scifdev(peerdev);
+}
+
+/**
+ * scif_node_add_ack() - Respond to SCIF_NODE_ADD_ACK interrupt message
+ * @scifdev:    Remote SCIF device node
+ * @msg:        Interrupt message
+ *
+ * After a MIC node receives the SCIF_NODE_ADD_ACK message it send this
+ * message to the mgmt node to confirm the sequence is finished.
+ *
+ */
+static __always_inline void
+scif_node_add_ack(struct scif_dev *scifdev, struct scifmsg *msg)
+{
+	struct scif_dev *peerdev;
+	struct scif_qp *qp;
+	struct scif_dev *dst_dev = &scif_dev[msg->dst.node];
+
+	dev_dbg(&scifdev->sdev->dev,
+		"Scifdev %d received SCIF_NODE_ADD_ACK msg src %d dst %d\n",
+		scifdev->node, msg->src.node, msg->dst.node);
+	dev_dbg(&scifdev->sdev->dev,
+		"payload %llx %llx %llx %llx\n", msg->payload[0],
+		msg->payload[1], msg->payload[2], msg->payload[3]);
+	if (scif_is_mgmt_node()) {
+		/*
+		 * the lock serializes with scif_qp_response_ack. The mgmt node
+		 * is forwarding the NODE_ADD_ACK message from src to dst we
+		 * need to make sure that the dst has already received a
+		 * NODE_ADD for src and setup its end of the qp to dst
+		 */
+		mutex_lock(&scif_info.conflock);
+		msg->payload[1] = scif_info.maxid;
+		scif_nodeqp_send(dst_dev, msg);
+		mutex_unlock(&scif_info.conflock);
+		return;
+	}
+	peerdev = &scif_dev[msg->src.node];
+	peerdev->sdev = scif_dev[SCIF_MGMT_NODE].sdev;
+	peerdev->node = msg->src.node;
+
+	qp = &peerdev->qpairs[0];
+
+	if ((scif_setup_qp_connect_response(peerdev, &peerdev->qpairs[0],
+					    msg->payload[0])))
+		goto local_error;
+	peerdev->rdb = msg->payload[2];
+	qp->remote_qp->qp_state = SCIF_QP_ONLINE;
+	schedule_delayed_work(&peerdev->p2p_dwork, 0);
+	return;
+local_error:
+	scif_cleanup_scifdev(peerdev);
+}
+
+/**
+ * scif_node_add_nack: Respond to SCIF_NODE_ADD_NACK interrupt message
+ * @msg:        Interrupt message
+ *
+ * SCIF_NODE_ADD failed, so inform the waiting wq.
+ */
+static __always_inline void
+scif_node_add_nack(struct scif_dev *scifdev, struct scifmsg *msg)
+{
+	if (scif_is_mgmt_node()) {
+		struct scif_dev *dst_dev = &scif_dev[msg->dst.node];
+
+		dev_dbg(&scifdev->sdev->dev,
+			"SCIF_NODE_ADD_NACK received from %d\n", scifdev->node);
+		scif_nodeqp_send(dst_dev, msg);
+	}
+}
+
+/*
+ * scif_node_remove: Handle SCIF_NODE_REMOVE message
+ * @msg: Interrupt message
+ *
+ * Handle node removal.
+ */
+static __always_inline void
+scif_node_remove(struct scif_dev *scifdev, struct scifmsg *msg)
+{
+	int node = msg->payload[0];
+	struct scif_dev *scdev = &scif_dev[node];
+
+	scdev->node_remove_ack_pending = true;
+	scif_handle_remove_node(node);
+}
+
+/*
+ * scif_node_remove_ack: Handle SCIF_NODE_REMOVE_ACK message
+ * @msg: Interrupt message
+ *
+ * The peer has acked a SCIF_NODE_REMOVE message.
+ */
+static __always_inline void
+scif_node_remove_ack(struct scif_dev *scifdev, struct scifmsg *msg)
+{
+	struct scif_dev *sdev = &scif_dev[msg->payload[0]];
+
+	atomic_inc(&sdev->disconn_rescnt);
+	wake_up(&sdev->disconn_wq);
+}
+
+static void
+scif_msg_unknown(struct scif_dev *scifdev, struct scifmsg *msg)
+{
+	/* Bogus Node Qp Message? */
+	dev_err(&scifdev->sdev->dev,
+		"Unknown message 0x%xn scifdev->node 0x%x\n",
+		msg->uop, scifdev->node);
+}
+
+static void (*scif_intr_func[SCIF_MAX_MSG + 1])
+	    (struct scif_dev *, struct scifmsg *msg) = {
+	scif_msg_unknown,	/* Error */
+	scif_init,		/* SCIF_INIT */
+	scif_exit,		/* SCIF_EXIT */
+	scif_exit_ack,		/* SCIF_EXIT_ACK */
+	scif_node_add,		/* SCIF_NODE_ADD */
+	scif_node_add_ack,	/* SCIF_NODE_ADD_ACK */
+	scif_node_add_nack,	/* SCIF_NODE_ADD_NACK */
+	scif_node_remove,	/* SCIF_NODE_REMOVE */
+	scif_node_remove_ack,	/* SCIF_NODE_REMOVE_ACK */
+};
+
+/**
+ * scif_nodeqp_msg_handler() - Common handler for node messages
+ * @scifdev: Remote device to respond to
+ * @qp: Remote memory pointer
+ * @msg: The message to be handled.
+ *
+ * This routine calls the appropriate routine to handle a Node Qp
+ * message receipt
+ */
+static int scif_max_msg_id = SCIF_MAX_MSG;
+
+static void
+scif_nodeqp_msg_handler(struct scif_dev *scifdev,
+			struct scif_qp *qp, struct scifmsg *msg)
+{
+	scif_display_message(scifdev, msg, "Rcvd");
+
+	if (msg->uop > (u32)scif_max_msg_id) {
+		/* Bogus Node Qp Message? */
+		dev_err(&scifdev->sdev->dev,
+			"Unknown message 0x%xn scifdev->node 0x%x\n",
+			msg->uop, scifdev->node);
+		return;
+	}
+
+	scif_intr_func[msg->uop](scifdev, msg);
+}
+
+/**
+ * scif_nodeqp_intrhandler() - Interrupt handler for node messages
+ * @scifdev:    Remote device to respond to
+ * @qp:         Remote memory pointer
+ *
+ * This routine is triggered by the interrupt mechanism.  It reads
+ * messages from the node queue RB and calls the Node QP Message handling
+ * routine.
+ */
+void scif_nodeqp_intrhandler(struct scif_dev *scifdev, struct scif_qp *qp)
+{
+	struct scifmsg msg;
+	int read_size;
+
+	do {
+		read_size = scif_rb_get_next(&qp->inbound_q, &msg, sizeof(msg));
+		if (!read_size)
+			break;
+		scif_nodeqp_msg_handler(scifdev, qp, &msg);
+		/*
+		 * The node queue pair is unmapped so skip the read pointer
+		 * update after receipt of a SCIF_EXIT_ACK
+		 */
+		if (SCIF_EXIT_ACK == msg.uop)
+			break;
+		scif_rb_update_read_ptr(&qp->inbound_q);
+	} while (1);
+}
+
+/**
+ * scif_loopb_wq_handler - Loopback Workqueue Handler.
+ * @work: loop back work
+ *
+ * This work queue routine is invoked by the loopback work queue handler.
+ * It grabs the recv lock, dequeues any available messages from the head
+ * of the loopback message list, calls the node QP message handler,
+ * waits for it to return, then frees up this message and dequeues more
+ * elements of the list if available.
+ */
+static void scif_loopb_wq_handler(struct work_struct *unused)
+{
+	struct scif_dev *scifdev = scif_info.loopb_dev;
+	struct scif_qp *qp = scifdev->qpairs;
+	struct scif_loopb_msg *msg;
+
+	do {
+		msg = NULL;
+		spin_lock(&qp->recv_lock);
+		if (!list_empty(&scif_info.loopb_recv_q)) {
+			msg = list_first_entry(&scif_info.loopb_recv_q,
+					       struct scif_loopb_msg,
+					       list);
+			list_del(&msg->list);
+		}
+		spin_unlock(&qp->recv_lock);
+
+		if (msg) {
+			scif_nodeqp_msg_handler(scifdev, qp, &msg->msg);
+			kfree(msg);
+		}
+	} while (msg);
+}
+
+/**
+ * scif_loopb_msg_handler() - Workqueue handler for loopback messages.
+ * @scifdev: SCIF device
+ * @qp: Queue pair.
+ *
+ * This work queue routine is triggered when a loopback message is received.
+ *
+ * We need special handling for receiving Node Qp messages on a loopback SCIF
+ * device via two workqueues for receiving messages.
+ *
+ * The reason we need the extra workqueue which is not required with *normal*
+ * non-loopback SCIF devices is the potential classic deadlock described below:
+ *
+ * Thread A tries to send a message on a loopback SCIF device and blocks since
+ * there is no space in the RB while it has the send_lock held or another
+ * lock called lock X for example.
+ *
+ * Thread B: The Loopback Node QP message receive workqueue receives the message
+ * and tries to send a message (eg an ACK) to the loopback SCIF device. It tries
+ * to grab the send lock again or lock X and deadlocks with Thread A. The RB
+ * cannot be drained any further due to this classic deadlock.
+ *
+ * In order to avoid deadlocks as mentioned above we have an extra level of
+ * indirection achieved by having two workqueues.
+ * 1) The first workqueue whose handler is scif_loopb_msg_handler reads
+ * messages from the Node QP RB, adds them to a list and queues work for the
+ * second workqueue.
+ *
+ * 2) The second workqueue whose handler is scif_loopb_wq_handler dequeues
+ * messages from the list, handles them, frees up the memory and dequeues
+ * more elements from the list if possible.
+ */
+int
+scif_loopb_msg_handler(struct scif_dev *scifdev, struct scif_qp *qp)
+{
+	int read_size;
+	struct scif_loopb_msg *msg;
+
+	do {
+		msg = kmalloc(sizeof(*msg), GFP_KERNEL);
+		if (!msg)
+			return -ENOMEM;
+		read_size = scif_rb_get_next(&qp->inbound_q, &msg->msg,
+					     sizeof(struct scifmsg));
+		if (read_size != sizeof(struct scifmsg)) {
+			kfree(msg);
+			scif_rb_update_read_ptr(&qp->inbound_q);
+			break;
+		}
+		spin_lock(&qp->recv_lock);
+		list_add_tail(&msg->list, &scif_info.loopb_recv_q);
+		spin_unlock(&qp->recv_lock);
+		queue_work(scif_info.loopb_wq, &scif_info.loopb_work);
+		scif_rb_update_read_ptr(&qp->inbound_q);
+	} while (read_size == sizeof(struct scifmsg));
+	return read_size;
+}
+
+/**
+ * scif_setup_loopback_qp - One time setup work for Loopback Node Qp.
+ * @scifdev: SCIF device
+ *
+ * Sets up the required loopback workqueues, queue pairs and ring buffers
+ */
+int scif_setup_loopback_qp(struct scif_dev *scifdev)
+{
+	int err = 0;
+	void *local_q;
+	struct scif_qp *qp;
+	struct scif_peer_dev *spdev;
+
+	err = scif_setup_intr_wq(scifdev);
+	if (err)
+		goto exit;
+	INIT_LIST_HEAD(&scif_info.loopb_recv_q);
+	snprintf(scif_info.loopb_wqname, sizeof(scif_info.loopb_wqname),
+		 "SCIF LOOPB %d", scifdev->node);
+	scif_info.loopb_wq =
+		alloc_ordered_workqueue(scif_info.loopb_wqname, 0);
+	if (!scif_info.loopb_wq) {
+		err = -ENOMEM;
+		goto destroy_intr;
+	}
+	INIT_WORK(&scif_info.loopb_work, scif_loopb_wq_handler);
+	/* Allocate Self Qpair */
+	scifdev->qpairs = kzalloc(sizeof(*scifdev->qpairs), GFP_KERNEL);
+	if (!scifdev->qpairs) {
+		err = -ENOMEM;
+		goto destroy_loopb_wq;
+	}
+
+	qp = scifdev->qpairs;
+	qp->magic = SCIFEP_MAGIC;
+	spin_lock_init(&qp->send_lock);
+	spin_lock_init(&qp->recv_lock);
+
+	local_q = kzalloc(SCIF_NODE_QP_SIZE, GFP_KERNEL);
+	if (!local_q) {
+		err = -ENOMEM;
+		goto free_qpairs;
+	}
+	/*
+	 * For loopback the inbound_q and outbound_q are essentially the same
+	 * since the Node sends a message on the loopback interface to the
+	 * outbound_q which is then received on the inbound_q.
+	 */
+	scif_rb_init(&qp->outbound_q,
+		     &qp->local_read,
+		     &qp->local_write,
+		     local_q, get_count_order(SCIF_NODE_QP_SIZE));
+
+	scif_rb_init(&qp->inbound_q,
+		     &qp->local_read,
+		     &qp->local_write,
+		     local_q, get_count_order(SCIF_NODE_QP_SIZE));
+	scif_info.nodeid = scifdev->node;
+	spdev = scif_peer_register_device(scifdev);
+	if (IS_ERR(spdev)) {
+		err = PTR_ERR(spdev);
+		goto free_local_q;
+	}
+	scif_info.loopb_dev = scifdev;
+	return err;
+free_local_q:
+	kfree(local_q);
+free_qpairs:
+	kfree(scifdev->qpairs);
+destroy_loopb_wq:
+	destroy_workqueue(scif_info.loopb_wq);
+destroy_intr:
+	scif_destroy_intr_wq(scifdev);
+exit:
+	return err;
+}
+
+/**
+ * scif_destroy_loopback_qp - One time uninit work for Loopback Node Qp
+ * @scifdev: SCIF device
+ *
+ * Destroys the workqueues and frees up the Ring Buffer and Queue Pair memory.
+ */
+int scif_destroy_loopback_qp(struct scif_dev *scifdev)
+{
+	struct scif_peer_dev *spdev;
+
+	rcu_read_lock();
+	spdev = rcu_dereference(scifdev->spdev);
+	rcu_read_unlock();
+	if (spdev)
+		scif_peer_unregister_device(spdev);
+	destroy_workqueue(scif_info.loopb_wq);
+	scif_destroy_intr_wq(scifdev);
+	kfree(scifdev->qpairs->outbound_q.rb_base);
+	kfree(scifdev->qpairs);
+	scifdev->sdev = NULL;
+	scif_info.loopb_dev = NULL;
+	return 0;
+}
+
+void scif_destroy_p2p(struct scif_dev *scifdev)
+{
+	struct scif_dev *peer_dev;
+	struct scif_p2p_info *p2p;
+	struct list_head *pos, *tmp;
+	int bd;
+
+	mutex_lock(&scif_info.conflock);
+	/* Free P2P mappings in the given node for all its peer nodes */
+	list_for_each_safe(pos, tmp, &scifdev->p2p) {
+		p2p = list_entry(pos, struct scif_p2p_info, ppi_list);
+		dma_unmap_sg(&scifdev->sdev->dev, p2p->ppi_sg[SCIF_PPI_MMIO],
+			     p2p->sg_nentries[SCIF_PPI_MMIO],
+			     DMA_BIDIRECTIONAL);
+		dma_unmap_sg(&scifdev->sdev->dev, p2p->ppi_sg[SCIF_PPI_APER],
+			     p2p->sg_nentries[SCIF_PPI_APER],
+			     DMA_BIDIRECTIONAL);
+		scif_p2p_freesg(p2p->ppi_sg[SCIF_PPI_MMIO]);
+		scif_p2p_freesg(p2p->ppi_sg[SCIF_PPI_APER]);
+		list_del(pos);
+		kfree(p2p);
+	}
+
+	/* Free P2P mapping created in the peer nodes for the given node */
+	for (bd = SCIF_MGMT_NODE + 1; bd <= scif_info.maxid; bd++) {
+		peer_dev = &scif_dev[bd];
+		list_for_each_safe(pos, tmp, &peer_dev->p2p) {
+			p2p = list_entry(pos, struct scif_p2p_info, ppi_list);
+			if (p2p->ppi_peer_id == scifdev->node) {
+				dma_unmap_sg(&peer_dev->sdev->dev,
+					     p2p->ppi_sg[SCIF_PPI_MMIO],
+					     p2p->sg_nentries[SCIF_PPI_MMIO],
+					     DMA_BIDIRECTIONAL);
+				dma_unmap_sg(&peer_dev->sdev->dev,
+					     p2p->ppi_sg[SCIF_PPI_APER],
+					     p2p->sg_nentries[SCIF_PPI_APER],
+					     DMA_BIDIRECTIONAL);
+				scif_p2p_freesg(p2p->ppi_sg[SCIF_PPI_MMIO]);
+				scif_p2p_freesg(p2p->ppi_sg[SCIF_PPI_APER]);
+				list_del(pos);
+				kfree(p2p);
+			}
+		}
+	}
+	mutex_unlock(&scif_info.conflock);
+}
-- 
1.8.2.1


  parent reply	other threads:[~2015-03-31  1:39 UTC|newest]

Thread overview: 21+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2015-03-31  1:36 [PATCH char-misc-next v4 00/13] misc: mic: SCIF driver Sudeep Dutt
2015-03-31  1:36 ` [PATCH char-misc-next v4 01/13] misc: mic: SCIF header file and IOCTL interface Sudeep Dutt
2015-03-31  1:36 ` [PATCH char-misc-next v4 02/13] misc: mic: SCIF ring buffer infrastructure Sudeep Dutt
2015-03-31  1:36 ` [PATCH char-misc-next v4 03/13] misc: mic: SCIF Hardware Bus Sudeep Dutt
2015-03-31  8:36   ` Paul Bolle
2015-03-31  1:36 ` [PATCH char-misc-next v4 04/13] misc: mic: SCIF Peer Bus Sudeep Dutt
2015-03-31  1:36 ` [PATCH char-misc-next v4 05/13] misc: mic: Common MIC header file changes in preparation for SCIF Sudeep Dutt
2015-03-31  1:36 ` [PATCH char-misc-next v4 06/13] misc: mic: SCIF module initialization Sudeep Dutt
2015-03-31  1:36 ` Sudeep Dutt [this message]
2015-03-31  1:36 ` [PATCH char-misc-next v4 08/13] misc: mic: SCIF open close bind and listen APIs Sudeep Dutt
2015-03-31  8:39   ` Paul Bolle
2015-03-31 21:13     ` Sudeep Dutt
2015-03-31  1:36 ` [PATCH char-misc-next v4 09/13] misc: mic: SCIF connections APIs i.e. accept and connect Sudeep Dutt
2015-03-31  1:36 ` [PATCH char-misc-next v4 10/13] misc: mic: SCIF messaging and node enumeration APIs Sudeep Dutt
2015-03-31  1:36 ` [PATCH char-misc-next v4 11/13] misc: mic: MIC host driver specific changes to enable SCIF Sudeep Dutt
2015-03-31  8:42   ` Paul Bolle
2015-03-31  1:36 ` [PATCH char-misc-next v4 12/13] misc: mic: MIC card " Sudeep Dutt
2015-03-31  1:36 ` [PATCH char-misc-next v4 13/13] misc: mic: add support for loading/unloading SCIF driver Sudeep Dutt
2015-04-28 20:45 ` [PATCH char-misc-next v4 00/13] misc: mic: " Sudeep Dutt
2015-04-29 11:27   ` Greg Kroah-Hartman
2015-04-29 12:42     ` Sudeep Dutt

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=421f4566df4e0200401e49845c631672603a9ba1.1427763968.git.sudeep.dutt@intel.com \
    --to=sudeep.dutt@intel.com \
    --cc=arnd@arndb.de \
    --cc=ashutosh.dixit@intel.com \
    --cc=corbet@lwn.net \
    --cc=dave.jiang@intel.com \
    --cc=gregkh@linuxfoundation.org \
    --cc=linux-doc@vger.kernel.org \
    --cc=linux-kernel@vger.kernel.org \
    --cc=nikhil.rao@intel.com \
    --subject='Re: [PATCH char-misc-next v4 07/13] misc: mic: SCIF node queue pair setup management' \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).