drm/amd/display: Try RGB before YCbCr 4:4:4 in stream validationgrafted

amdgpu_dm_create_validate_stream_for_sink() walks encoding_order[] and
uses the first encoding that validates. YCbCr 4:4:4 is listed before
RGB, so an HDMI sink that advertises 4:4:4 gets YCbCr 4:4:4 whenever the
"color format" property is left at AUTO, even though RGB fits the same
link.

That contradicts the documented AUTO behaviour for HDMI in enum
drm_connector_color_format (RGB, falling back to YCbCr 4:2:0 only when
the bandwidth is not available or the mode is 4:2:0-only), which the
amdgpu implementation of the property also describes. It also leaves
the "Broadcast RGB" property without effect on such sinks, since the
quantization range it selects only applies to RGB output.

Try RGB first. The mask still holds every encoding the sink supports,
so a mode that cannot carry RGB falls back exactly as before.

For reference, v7.2 picked RGB here unless YCbCr 4:4:4 was forced
through debugfs, while earlier kernels picked YCbCr 4:4:4 for any HDMI
sink that advertised it.

Fixes: 0b0ff65d3ca1 ("drm/amd/display: Refactor stream validation")
Suggested-by: Adolfo Rodrigues <adolfotregosa@gmail.com>
Assisted-by: Claude Code:claude-fable-5-1
Signed-off-by: Adrian Betschart <adrian.betschart@cinemaone.ch>
Reviewed-by: Fangzhi Zuo <jerry.zuo@amd.com>
Tested-by: Adolfo Rodrigues <adolfotregosa@gmail.com>
Tested-by: Daniel Wheeler <daniel.wheeler@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
(cherry picked from commit 7de432bc753133764dc40d37f9388da56de251e7)

6 files changed, 4158 insertions, 0 deletions

diff --git a/drivers/ntb/test/Kconfig b/drivers/ntb/test/Kconfig
new file mode 100644
index 000000000..516b991f3
--- /dev/null
+++ b/drivers/ntb/test/Kconfig
@@ -0,0 +1,37 @@
+# SPDX-License-Identifier: GPL-2.0-only
+config NTB_PINGPONG
+	tristate "NTB Ping Pong Test Client"
+	help
+	 This is a simple ping pong driver that exercises the scratchpads and
+	 doorbells of the ntb hardware.  This driver may be used to test that
+	 your ntb hardware and drivers are functioning at a basic level.
+
+	 If unsure, say N.
+
+config NTB_TOOL
+	tristate "NTB Debugging Tool Test Client"
+	help
+	 This is a simple debugging driver that enables the doorbell and
+	 scratchpad registers to be read and written from the debugfs.  This
+	 enables more complicated debugging to be scripted from user space.
+	 This driver may be used to test that your ntb hardware and drivers are
+	 functioning at a basic level.
+
+	 If unsure, say N.
+
+config NTB_PERF
+	tristate "NTB RAW Perf Measuring Tool"
+	help
+	 This is a tool to measure raw NTB performance by transferring data
+	 to and from the window without additional software interaction.
+
+	 If unsure, say N.
+
+config NTB_MSI_TEST
+	tristate "NTB MSI Test Client"
+	depends on NTB_MSI
+	help
+	  This tool demonstrates the use of the NTB MSI library to
+	  send MSI interrupts between peers.
+
+	  If unsure, say N.
diff --git a/drivers/ntb/test/Makefile b/drivers/ntb/test/Makefile
new file mode 100644
index 000000000..19ed91d8a
--- /dev/null
+++ b/drivers/ntb/test/Makefile
@@ -0,0 +1,5 @@
+# SPDX-License-Identifier: GPL-2.0-only
+obj-$(CONFIG_NTB_PINGPONG) += ntb_pingpong.o
+obj-$(CONFIG_NTB_TOOL) += ntb_tool.o
+obj-$(CONFIG_NTB_PERF) += ntb_perf.o
+obj-$(CONFIG_NTB_MSI_TEST) += ntb_msi_test.o
diff --git a/drivers/ntb/test/ntb_msi_test.c b/drivers/ntb/test/ntb_msi_test.c
new file mode 100644
index 000000000..f52d409ba
--- /dev/null
+++ b/drivers/ntb/test/ntb_msi_test.c
@@ -0,0 +1,432 @@
+// SPDX-License-Identifier: (GPL-2.0 OR BSD-3-Clause)
+
+#include <linux/module.h>
+#include <linux/debugfs.h>
+#include <linux/ntb.h>
+#include <linux/pci.h>
+#include <linux/radix-tree.h>
+#include <linux/workqueue.h>
+
+MODULE_LICENSE("Dual BSD/GPL");
+MODULE_VERSION("0.1");
+MODULE_AUTHOR("Logan Gunthorpe <logang@deltatee.com>");
+MODULE_DESCRIPTION("Test for sending MSI interrupts over an NTB memory window");
+
+static int num_irqs = 4;
+module_param(num_irqs, int, 0644);
+MODULE_PARM_DESC(num_irqs, "number of irqs to use");
+
+struct ntb_msit_ctx {
+	struct ntb_dev *ntb;
+	struct dentry *dbgfs_dir;
+	struct work_struct setup_work;
+
+	struct ntb_msit_isr_ctx {
+		int irq_idx;
+		int irq_num;
+		int occurrences;
+		struct ntb_msit_ctx *nm;
+		struct ntb_msi_desc desc;
+	} *isr_ctx;
+
+	struct ntb_msit_peer {
+		struct ntb_msit_ctx *nm;
+		int pidx;
+		int num_irqs;
+		struct completion init_comp;
+		struct ntb_msi_desc *msi_desc;
+	} peers[];
+};
+
+static struct dentry *ntb_msit_dbgfs_topdir;
+
+static irqreturn_t ntb_msit_isr(int irq, void *dev)
+{
+	struct ntb_msit_isr_ctx *isr_ctx = dev;
+	struct ntb_msit_ctx *nm = isr_ctx->nm;
+
+	dev_dbg(&nm->ntb->dev, "Interrupt Occurred: %d",
+		isr_ctx->irq_idx);
+
+	isr_ctx->occurrences++;
+
+	return IRQ_HANDLED;
+}
+
+static void ntb_msit_setup_work(struct work_struct *work)
+{
+	struct ntb_msit_ctx *nm = container_of(work, struct ntb_msit_ctx,
+					       setup_work);
+	int irq_count = 0;
+	int irq;
+	int ret;
+	uintptr_t i;
+
+	ret = ntb_msi_setup_mws(nm->ntb);
+	if (ret) {
+		dev_err(&nm->ntb->dev, "Unable to setup MSI windows: %d\n",
+			ret);
+		return;
+	}
+
+	for (i = 0; i < num_irqs; i++) {
+		nm->isr_ctx[i].irq_idx = i;
+		nm->isr_ctx[i].nm = nm;
+
+		if (!nm->isr_ctx[i].irq_num) {
+			irq = ntbm_msi_request_irq(nm->ntb, ntb_msit_isr,
+						   KBUILD_MODNAME,
+						   &nm->isr_ctx[i],
+						   &nm->isr_ctx[i].desc);
+			if (irq < 0)
+				break;
+
+			nm->isr_ctx[i].irq_num = irq;
+		}
+
+		ret = ntb_spad_write(nm->ntb, 2 * i + 1,
+				     nm->isr_ctx[i].desc.addr_offset);
+		if (ret)
+			break;
+
+		ret = ntb_spad_write(nm->ntb, 2 * i + 2,
+				     nm->isr_ctx[i].desc.data);
+		if (ret)
+			break;
+
+		irq_count++;
+	}
+
+	ntb_spad_write(nm->ntb, 0, irq_count);
+	ntb_peer_db_set(nm->ntb, BIT(ntb_port_number(nm->ntb)));
+}
+
+static void ntb_msit_desc_changed(void *ctx)
+{
+	struct ntb_msit_ctx *nm = ctx;
+	int i;
+
+	dev_dbg(&nm->ntb->dev, "MSI Descriptors Changed\n");
+
+	for (i = 0; i < num_irqs; i++) {
+		ntb_spad_write(nm->ntb, 2 * i + 1,
+			       nm->isr_ctx[i].desc.addr_offset);
+		ntb_spad_write(nm->ntb, 2 * i + 2,
+			       nm->isr_ctx[i].desc.data);
+	}
+
+	ntb_peer_db_set(nm->ntb, BIT(ntb_port_number(nm->ntb)));
+}
+
+static void ntb_msit_link_event(void *ctx)
+{
+	struct ntb_msit_ctx *nm = ctx;
+
+	if (!ntb_link_is_up(nm->ntb, NULL, NULL))
+		return;
+
+	schedule_work(&nm->setup_work);
+}
+
+static void ntb_msit_copy_peer_desc(struct ntb_msit_ctx *nm, int peer)
+{
+	int i;
+	struct ntb_msi_desc *desc = nm->peers[peer].msi_desc;
+	int irq_count = nm->peers[peer].num_irqs;
+
+	for (i = 0; i < irq_count; i++) {
+		desc[i].addr_offset = ntb_peer_spad_read(nm->ntb, peer,
+							 2 * i + 1);
+		desc[i].data = ntb_peer_spad_read(nm->ntb, peer, 2 * i + 2);
+	}
+
+	dev_info(&nm->ntb->dev, "Found %d interrupts on peer %d\n",
+		 irq_count, peer);
+
+	complete_all(&nm->peers[peer].init_comp);
+}
+
+static void ntb_msit_db_event(void *ctx, int vec)
+{
+	struct ntb_msit_ctx *nm = ctx;
+	struct ntb_msi_desc *desc;
+	u64 peer_mask = ntb_db_read(nm->ntb);
+	u32 irq_count;
+	int peer;
+
+	ntb_db_clear(nm->ntb, peer_mask);
+
+	for (peer = 0; peer < sizeof(peer_mask) * 8; peer++) {
+		if (!(peer_mask & BIT(peer)))
+			continue;
+
+		irq_count = ntb_peer_spad_read(nm->ntb, peer, 0);
+		if (irq_count == -1)
+			continue;
+
+		desc = kzalloc_objs(*desc, irq_count, GFP_ATOMIC);
+		if (!desc)
+			continue;
+
+		kfree(nm->peers[peer].msi_desc);
+		nm->peers[peer].msi_desc = desc;
+		nm->peers[peer].num_irqs = irq_count;
+
+		ntb_msit_copy_peer_desc(nm, peer);
+	}
+}
+
+static const struct ntb_ctx_ops ntb_msit_ops = {
+	.link_event = ntb_msit_link_event,
+	.db_event = ntb_msit_db_event,
+};
+
+static int ntb_msit_dbgfs_trigger(void *data, u64 idx)
+{
+	struct ntb_msit_peer *peer = data;
+
+	if (idx >= peer->num_irqs)
+		return -EINVAL;
+
+	dev_dbg(&peer->nm->ntb->dev, "trigger irq %llu on peer %u\n",
+		idx, peer->pidx);
+
+	return ntb_msi_peer_trigger(peer->nm->ntb, peer->pidx,
+				    &peer->msi_desc[idx]);
+}
+
+DEFINE_DEBUGFS_ATTRIBUTE(ntb_msit_trigger_fops, NULL,
+			 ntb_msit_dbgfs_trigger, "%llu\n");
+
+static int ntb_msit_dbgfs_port_get(void *data, u64 *port)
+{
+	struct ntb_msit_peer *peer = data;
+
+	*port = ntb_peer_port_number(peer->nm->ntb, peer->pidx);
+
+	return 0;
+}
+
+DEFINE_DEBUGFS_ATTRIBUTE(ntb_msit_port_fops, ntb_msit_dbgfs_port_get,
+			 NULL, "%llu\n");
+
+static int ntb_msit_dbgfs_count_get(void *data, u64 *count)
+{
+	struct ntb_msit_peer *peer = data;
+
+	*count = peer->num_irqs;
+
+	return 0;
+}
+
+DEFINE_DEBUGFS_ATTRIBUTE(ntb_msit_count_fops, ntb_msit_dbgfs_count_get,
+			 NULL, "%llu\n");
+
+static int ntb_msit_dbgfs_ready_get(void *data, u64 *ready)
+{
+	struct ntb_msit_peer *peer = data;
+
+	*ready = try_wait_for_completion(&peer->init_comp);
+
+	return 0;
+}
+
+static int ntb_msit_dbgfs_ready_set(void *data, u64 ready)
+{
+	struct ntb_msit_peer *peer = data;
+
+	return wait_for_completion_interruptible(&peer->init_comp);
+}
+
+DEFINE_DEBUGFS_ATTRIBUTE(ntb_msit_ready_fops, ntb_msit_dbgfs_ready_get,
+			 ntb_msit_dbgfs_ready_set, "%llu\n");
+
+static int ntb_msit_dbgfs_occurrences_get(void *data, u64 *occurrences)
+{
+	struct ntb_msit_isr_ctx *isr_ctx = data;
+
+	*occurrences = isr_ctx->occurrences;
+
+	return 0;
+}
+
+DEFINE_DEBUGFS_ATTRIBUTE(ntb_msit_occurrences_fops,
+			 ntb_msit_dbgfs_occurrences_get,
+			 NULL, "%llu\n");
+
+static int ntb_msit_dbgfs_local_port_get(void *data, u64 *port)
+{
+	struct ntb_msit_ctx *nm = data;
+
+	*port = ntb_port_number(nm->ntb);
+
+	return 0;
+}
+
+DEFINE_DEBUGFS_ATTRIBUTE(ntb_msit_local_port_fops,
+			 ntb_msit_dbgfs_local_port_get,
+			 NULL, "%llu\n");
+
+static void ntb_msit_create_dbgfs(struct ntb_msit_ctx *nm)
+{
+	struct pci_dev *pdev = nm->ntb->pdev;
+	char buf[32];
+	int i;
+	struct dentry *peer_dir;
+
+	nm->dbgfs_dir = debugfs_create_dir(pci_name(pdev),
+					   ntb_msit_dbgfs_topdir);
+	debugfs_create_file("port", 0400, nm->dbgfs_dir, nm,
+			    &ntb_msit_local_port_fops);
+
+	for (i = 0; i < ntb_peer_port_count(nm->ntb); i++) {
+		nm->peers[i].pidx = i;
+		nm->peers[i].nm = nm;
+		init_completion(&nm->peers[i].init_comp);
+
+		snprintf(buf, sizeof(buf), "peer%d", i);
+		peer_dir = debugfs_create_dir(buf, nm->dbgfs_dir);
+
+		debugfs_create_file_unsafe("trigger", 0200, peer_dir,
+					   &nm->peers[i],
+					   &ntb_msit_trigger_fops);
+
+		debugfs_create_file_unsafe("port", 0400, peer_dir,
+					   &nm->peers[i], &ntb_msit_port_fops);
+
+		debugfs_create_file_unsafe("count", 0400, peer_dir,
+					   &nm->peers[i],
+					   &ntb_msit_count_fops);
+
+		debugfs_create_file_unsafe("ready", 0600, peer_dir,
+					   &nm->peers[i],
+					   &ntb_msit_ready_fops);
+	}
+
+	for (i = 0; i < num_irqs; i++) {
+		snprintf(buf, sizeof(buf), "irq%d_occurrences", i);
+		debugfs_create_file_unsafe(buf, 0400, nm->dbgfs_dir,
+					   &nm->isr_ctx[i],
+					   &ntb_msit_occurrences_fops);
+	}
+}
+
+static void ntb_msit_remove_dbgfs(struct ntb_msit_ctx *nm)
+{
+	debugfs_remove_recursive(nm->dbgfs_dir);
+}
+
+static int ntb_msit_probe(struct ntb_client *client, struct ntb_dev *ntb)
+{
+	struct ntb_msit_ctx *nm;
+	int peers;
+	int ret;
+
+	peers = ntb_peer_port_count(ntb);
+	if (peers <= 0)
+		return -EINVAL;
+
+	if (ntb_spad_is_unsafe(ntb) || ntb_spad_count(ntb) < 2 * num_irqs + 1) {
+		dev_err(&ntb->dev, "NTB MSI test requires at least %d spads for %d irqs\n",
+			2 * num_irqs + 1, num_irqs);
+		return -EFAULT;
+	}
+
+	ret = ntb_spad_write(ntb, 0, -1);
+	if (ret) {
+		dev_err(&ntb->dev, "Unable to write spads: %d\n", ret);
+		return ret;
+	}
+
+	ret = ntb_db_clear_mask(ntb, GENMASK(peers - 1, 0));
+	if (ret) {
+		dev_err(&ntb->dev, "Unable to clear doorbell mask: %d\n", ret);
+		return ret;
+	}
+
+	ret = ntb_msi_init(ntb, ntb_msit_desc_changed);
+	if (ret) {
+		dev_err(&ntb->dev, "Unable to initialize MSI library: %d\n",
+			ret);
+		return ret;
+	}
+
+	nm = devm_kzalloc(&ntb->dev, struct_size(nm, peers, peers), GFP_KERNEL);
+	if (!nm)
+		return -ENOMEM;
+
+	nm->isr_ctx = devm_kcalloc(&ntb->dev, num_irqs, sizeof(*nm->isr_ctx),
+				   GFP_KERNEL);
+	if (!nm->isr_ctx)
+		return -ENOMEM;
+
+	INIT_WORK(&nm->setup_work, ntb_msit_setup_work);
+	nm->ntb = ntb;
+
+	ntb_msit_create_dbgfs(nm);
+
+	ret = ntb_set_ctx(ntb, nm, &ntb_msit_ops);
+	if (ret)
+		goto remove_dbgfs;
+
+	if (!nm->isr_ctx) {
+		ret = -ENOMEM;
+		goto remove_dbgfs;
+	}
+
+	ntb_link_enable(ntb, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
+
+	return 0;
+
+remove_dbgfs:
+	ntb_msit_remove_dbgfs(nm);
+	devm_kfree(&ntb->dev, nm->isr_ctx);
+	devm_kfree(&ntb->dev, nm);
+	return ret;
+}
+
+static void ntb_msit_remove(struct ntb_client *client, struct ntb_dev *ntb)
+{
+	struct ntb_msit_ctx *nm = ntb->ctx;
+	int i;
+
+	ntb_link_disable(ntb);
+	ntb_db_set_mask(ntb, ntb_db_valid_mask(ntb));
+	ntb_msi_clear_mws(ntb);
+
+	for (i = 0; i < ntb_peer_port_count(ntb); i++)
+		kfree(nm->peers[i].msi_desc);
+
+	ntb_clear_ctx(ntb);
+	ntb_msit_remove_dbgfs(nm);
+}
+
+static struct ntb_client ntb_msit_client = {
+	.ops = {
+		.probe = ntb_msit_probe,
+		.remove = ntb_msit_remove
+	}
+};
+
+static int __init ntb_msit_init(void)
+{
+	int ret;
+
+	if (debugfs_initialized())
+		ntb_msit_dbgfs_topdir = debugfs_create_dir(KBUILD_MODNAME,
+							   NULL);
+
+	ret = ntb_register_client(&ntb_msit_client);
+	if (ret)
+		debugfs_remove_recursive(ntb_msit_dbgfs_topdir);
+
+	return ret;
+}
+module_init(ntb_msit_init);
+
+static void __exit ntb_msit_exit(void)
+{
+	ntb_unregister_client(&ntb_msit_client);
+	debugfs_remove_recursive(ntb_msit_dbgfs_topdir);
+}
+module_exit(ntb_msit_exit);
diff --git a/drivers/ntb/test/ntb_perf.c b/drivers/ntb/test/ntb_perf.c
new file mode 100644
index 000000000..dfd175f79
--- /dev/null
+++ b/drivers/ntb/test/ntb_perf.c
@@ -0,0 +1,1565 @@
+/*
+ * 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) 2015 Intel Corporation. All rights reserved.
+ *   Copyright(c) 2017 T-Platforms. All Rights Reserved.
+ *
+ *   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.
+ *
+ *   BSD LICENSE
+ *
+ *   Copyright(c) 2015 Intel Corporation. All rights reserved.
+ *   Copyright(c) 2017 T-Platforms. All Rights Reserved.
+ *
+ *   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 copy
+ *       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.
+ *
+ * PCIe NTB Perf Linux driver
+ */
+
+/*
+ * How to use this tool, by example.
+ *
+ * Assuming $DBG_DIR is something like:
+ * '/sys/kernel/debug/ntb_perf/0000:00:03.0'
+ * Suppose aside from local device there is at least one remote device
+ * connected to NTB with index 0.
+ *-----------------------------------------------------------------------------
+ * Eg: install driver with specified chunk/total orders and dma-enabled flag
+ *
+ * root@self# insmod ntb_perf.ko chunk_order=19 total_order=28 use_dma
+ *-----------------------------------------------------------------------------
+ * Eg: check NTB ports (index) and MW mapping information
+ *
+ * root@self# cat $DBG_DIR/info
+ *-----------------------------------------------------------------------------
+ * Eg: start performance test with peer (index 0) and get the test metrics
+ *
+ * root@self# echo 0 > $DBG_DIR/run
+ * root@self# cat $DBG_DIR/run
+ */
+
+#include <linux/init.h>
+#include <linux/kernel.h>
+#include <linux/module.h>
+#include <linux/sched.h>
+#include <linux/wait.h>
+#include <linux/dma-mapping.h>
+#include <linux/dmaengine.h>
+#include <linux/pci.h>
+#include <linux/ktime.h>
+#include <linux/slab.h>
+#include <linux/delay.h>
+#include <linux/sizes.h>
+#include <linux/workqueue.h>
+#include <linux/debugfs.h>
+#include <linux/random.h>
+#include <linux/ntb.h>
+
+#define DRIVER_NAME		"ntb_perf"
+#define DRIVER_VERSION		"2.0"
+
+MODULE_LICENSE("Dual BSD/GPL");
+MODULE_VERSION(DRIVER_VERSION);
+MODULE_AUTHOR("Dave Jiang <dave.jiang@intel.com>");
+MODULE_DESCRIPTION("PCIe NTB Performance Measurement Tool");
+
+#define MAX_THREADS_CNT		32
+#define DEF_THREADS_CNT		1
+#define MAX_CHUNK_SIZE		SZ_1M
+#define MAX_CHUNK_ORDER		20 /* no larger than 1M */
+
+#define DMA_TRIES		100
+#define DMA_MDELAY		10
+
+#define MSG_TRIES		1000
+#define MSG_UDELAY_LOW		1000000
+#define MSG_UDELAY_HIGH		2000000
+
+#define PERF_BUF_LEN 1024
+
+static unsigned long max_mw_size;
+module_param(max_mw_size, ulong, 0644);
+MODULE_PARM_DESC(max_mw_size, "Upper limit of memory window size");
+
+static unsigned char chunk_order = 19; /* 512K */
+module_param(chunk_order, byte, 0644);
+MODULE_PARM_DESC(chunk_order, "Data chunk order [2^n] to transfer");
+
+static unsigned char total_order = 30; /* 1G */
+module_param(total_order, byte, 0644);
+MODULE_PARM_DESC(total_order, "Total data order [2^n] to transfer");
+
+static bool use_dma; /* default to 0 */
+module_param(use_dma, bool, 0644);
+MODULE_PARM_DESC(use_dma, "Use DMA engine to measure performance");
+
+/*==============================================================================
+ *                         Perf driver data definition
+ *==============================================================================
+ */
+
+enum perf_cmd {
+	PERF_CMD_INVAL = -1,/* invalid spad command */
+	PERF_CMD_SSIZE = 0, /* send out buffer size */
+	PERF_CMD_RSIZE = 1, /* recv in  buffer size */
+	PERF_CMD_SXLAT = 2, /* send in  buffer xlat */
+	PERF_CMD_RXLAT = 3, /* recv out buffer xlat */
+	PERF_CMD_CLEAR = 4, /* clear allocated memory */
+	PERF_STS_DONE  = 5, /* init is done */
+	PERF_STS_LNKUP = 6, /* link up state flag */
+};
+
+struct perf_ctx;
+
+struct perf_peer {
+	struct perf_ctx	*perf;
+	int pidx;
+	int gidx;
+
+	/* Outbound MW params */
+	u64 outbuf_xlat;
+	resource_size_t outbuf_size;
+	void __iomem *outbuf;
+	phys_addr_t out_phys_addr;
+	dma_addr_t dma_dst_addr;
+	/* Inbound MW params */
+	dma_addr_t inbuf_xlat;
+	resource_size_t inbuf_size;
+	void		*inbuf;
+
+	/* NTB connection setup service */
+	struct work_struct	service;
+	unsigned long		sts;
+
+	struct completion init_comp;
+};
+#define to_peer_service(__work) \
+	container_of(__work, struct perf_peer, service)
+
+struct perf_thread {
+	struct perf_ctx *perf;
+	int tidx;
+
+	/* DMA-based test sync parameters */
+	atomic_t dma_sync;
+	wait_queue_head_t dma_wait;
+	struct dma_chan *dma_chan;
+
+	/* Data source and measured statistics */
+	void *src;
+	u64 copied;
+	ktime_t duration;
+	int status;
+	struct work_struct work;
+};
+#define to_thread_work(__work) \
+	container_of(__work, struct perf_thread, work)
+
+struct perf_ctx {
+	struct ntb_dev *ntb;
+
+	/* Global device index and peers descriptors */
+	int gidx;
+	int pcnt;
+	struct perf_peer *peers;
+
+	/* Performance measuring work-threads interface */
+	unsigned long busy_flag;
+	wait_queue_head_t twait;
+	atomic_t tsync;
+	u8 tcnt;
+	struct perf_peer *test_peer;
+	struct perf_thread threads[MAX_THREADS_CNT];
+
+	/* Scratchpad/Message IO operations */
+	int (*cmd_send)(struct perf_peer *peer, enum perf_cmd cmd, u64 data);
+	int (*cmd_recv)(struct perf_ctx *perf, int *pidx, enum perf_cmd *cmd,
+			u64 *data);
+
+	struct dentry *dbgfs_dir;
+};
+
+/*
+ * Scratchpads-base commands interface
+ */
+#define PERF_SPAD_CNT(_pcnt) \
+	(3*((_pcnt) + 1))
+#define PERF_SPAD_CMD(_gidx) \
+	(3*(_gidx))
+#define PERF_SPAD_LDATA(_gidx) \
+	(3*(_gidx) + 1)
+#define PERF_SPAD_HDATA(_gidx) \
+	(3*(_gidx) + 2)
+#define PERF_SPAD_NOTIFY(_gidx) \
+	(BIT_ULL(_gidx))
+
+/*
+ * Messages-base commands interface
+ */
+#define PERF_MSG_CNT		3
+#define PERF_MSG_CMD		0
+#define PERF_MSG_LDATA		1
+#define PERF_MSG_HDATA		2
+
+/*==============================================================================
+ *                           Static data declarations
+ *==============================================================================
+ */
+
+static struct dentry *perf_dbgfs_topdir;
+
+static struct workqueue_struct *perf_wq __read_mostly;
+
+/*==============================================================================
+ *                  NTB cross-link commands execution service
+ *==============================================================================
+ */
+
+static void perf_terminate_test(struct perf_ctx *perf);
+
+static inline bool perf_link_is_up(struct perf_peer *peer)
+{
+	u64 link;
+
+	link = ntb_link_is_up(peer->perf->ntb, NULL, NULL);
+	return !!(link & BIT_ULL_MASK(peer->pidx));
+}
+
+static int perf_spad_cmd_send(struct perf_peer *peer, enum perf_cmd cmd,
+			      u64 data)
+{
+	struct perf_ctx *perf = peer->perf;
+	int try;
+	u32 sts;
+
+	dev_dbg(&perf->ntb->dev, "CMD send: %d 0x%llx\n", cmd, data);
+
+	/*
+	 * Perform predefined number of attempts before give up.
+	 * We are sending the data to the port specific scratchpad, so
+	 * to prevent a multi-port access race-condition. Additionally
+	 * there is no need in local locking since only thread-safe
+	 * service work is using this method.
+	 */
+	for (try = 0; try < MSG_TRIES; try++) {
+		if (!perf_link_is_up(peer))
+			return -ENOLINK;
+
+		sts = ntb_peer_spad_read(perf->ntb, peer->pidx,
+					 PERF_SPAD_CMD(perf->gidx));
+		if (sts != PERF_CMD_INVAL) {
+			usleep_range(MSG_UDELAY_LOW, MSG_UDELAY_HIGH);
+			continue;
+		}
+
+		ntb_peer_spad_write(perf->ntb, peer->pidx,
+				    PERF_SPAD_LDATA(perf->gidx),
+				    lower_32_bits(data));
+		ntb_peer_spad_write(perf->ntb, peer->pidx,
+				    PERF_SPAD_HDATA(perf->gidx),
+				    upper_32_bits(data));
+		ntb_peer_spad_write(perf->ntb, peer->pidx,
+				    PERF_SPAD_CMD(perf->gidx),
+				    cmd);
+		ntb_peer_db_set(perf->ntb, PERF_SPAD_NOTIFY(peer->gidx));
+
+		dev_dbg(&perf->ntb->dev, "DB ring peer %#llx\n",
+			PERF_SPAD_NOTIFY(peer->gidx));
+
+		break;
+	}
+
+	return try < MSG_TRIES ? 0 : -EAGAIN;
+}
+
+static int perf_spad_cmd_recv(struct perf_ctx *perf, int *pidx,
+			      enum perf_cmd *cmd, u64 *data)
+{
+	struct perf_peer *peer;
+	u32 val;
+
+	ntb_db_clear(perf->ntb, PERF_SPAD_NOTIFY(perf->gidx));
+
+	/*
+	 * We start scanning all over, since cleared DB may have been set
+	 * by any peer. Yes, it makes peer with smaller index being
+	 * serviced with greater priority, but it's convenient for spad
+	 * and message code unification and simplicity.
+	 */
+	for (*pidx = 0; *pidx < perf->pcnt; (*pidx)++) {
+		peer = &perf->peers[*pidx];
+
+		if (!perf_link_is_up(peer))
+			continue;
+
+		val = ntb_spad_read(perf->ntb, PERF_SPAD_CMD(peer->gidx));
+		if (val == PERF_CMD_INVAL)
+			continue;
+
+		*cmd = val;
+
+		val = ntb_spad_read(perf->ntb, PERF_SPAD_LDATA(peer->gidx));
+		*data = val;
+
+		val = ntb_spad_read(perf->ntb, PERF_SPAD_HDATA(peer->gidx));
+		*data |= (u64)val << 32;
+
+		/* Next command can be retrieved from now */
+		ntb_spad_write(perf->ntb, PERF_SPAD_CMD(peer->gidx),
+			       PERF_CMD_INVAL);
+
+		dev_dbg(&perf->ntb->dev, "CMD recv: %d 0x%llx\n", *cmd, *data);
+
+		return 0;
+	}
+
+	return -ENODATA;
+}
+
+static int perf_msg_cmd_send(struct perf_peer *peer, enum perf_cmd cmd,
+			     u64 data)
+{
+	struct perf_ctx *perf = peer->perf;
+	int try, ret;
+	u64 outbits;
+
+	dev_dbg(&perf->ntb->dev, "CMD send: %d 0x%llx\n", cmd, data);
+
+	/*
+	 * Perform predefined number of attempts before give up. Message
+	 * registers are free of race-condition problem when accessed
+	 * from different ports, so we don't need splitting registers
+	 * by global device index. We also won't have local locking,
+	 * since the method is used from service work only.
+	 */
+	outbits = ntb_msg_outbits(perf->ntb);
+	for (try = 0; try < MSG_TRIES; try++) {
+		if (!perf_link_is_up(peer))
+			return -ENOLINK;
+
+		ret = ntb_msg_clear_sts(perf->ntb, outbits);
+		if (ret)
+			return ret;
+
+		ntb_peer_msg_write(perf->ntb, peer->pidx, PERF_MSG_LDATA,
+				   lower_32_bits(data));
+
+		if (ntb_msg_read_sts(perf->ntb) & outbits) {
+			usleep_range(MSG_UDELAY_LOW, MSG_UDELAY_HIGH);
+			continue;
+		}
+
+		ntb_peer_msg_write(perf->ntb, peer->pidx, PERF_MSG_HDATA,
+				   upper_32_bits(data));
+
+		/* This call shall trigger peer message event */
+		ntb_peer_msg_write(perf->ntb, peer->pidx, PERF_MSG_CMD, cmd);
+
+		break;
+	}
+
+	return try < MSG_TRIES ? 0 : -EAGAIN;
+}
+
+static int perf_msg_cmd_recv(struct perf_ctx *perf, int *pidx,
+			     enum perf_cmd *cmd, u64 *data)
+{
+	u64 inbits;
+	u32 val;
+
+	inbits = ntb_msg_inbits(perf->ntb);
+
+	if (hweight64(ntb_msg_read_sts(perf->ntb) & inbits) < 3)
+		return -ENODATA;
+
+	val = ntb_msg_read(perf->ntb, pidx, PERF_MSG_CMD);
+	*cmd = val;
+
+	val = ntb_msg_read(perf->ntb, pidx, PERF_MSG_LDATA);
+	*data = val;
+
+	val = ntb_msg_read(perf->ntb, pidx, PERF_MSG_HDATA);
+	*data |= (u64)val << 32;
+
+	/* Next command can be retrieved from now */
+	ntb_msg_clear_sts(perf->ntb, inbits);
+
+	dev_dbg(&perf->ntb->dev, "CMD recv: %d 0x%llx\n", *cmd, *data);
+
+	return 0;
+}
+
+static int perf_cmd_send(struct perf_peer *peer, enum perf_cmd cmd, u64 data)
+{
+	struct perf_ctx *perf = peer->perf;
+
+	if (cmd == PERF_CMD_SSIZE || cmd == PERF_CMD_SXLAT)
+		return perf->cmd_send(peer, cmd, data);
+
+	dev_err(&perf->ntb->dev, "Send invalid command\n");
+	return -EINVAL;
+}
+
+static int perf_cmd_exec(struct perf_peer *peer, enum perf_cmd cmd)
+{
+	switch (cmd) {
+	case PERF_CMD_SSIZE:
+	case PERF_CMD_RSIZE:
+	case PERF_CMD_SXLAT:
+	case PERF_CMD_RXLAT:
+	case PERF_CMD_CLEAR:
+		break;
+	default:
+		dev_err(&peer->perf->ntb->dev, "Exec invalid command\n");
+		return -EINVAL;
+	}
+
+	/* No need of memory barrier, since bit ops have invernal lock */
+	set_bit(cmd, &peer->sts);
+
+	dev_dbg(&peer->perf->ntb->dev, "CMD exec: %d\n", cmd);
+
+	(void)queue_work(system_highpri_wq, &peer->service);
+
+	return 0;
+}
+
+static int perf_cmd_recv(struct perf_ctx *perf)
+{
+	struct perf_peer *peer;
+	int ret, pidx, cmd;
+	u64 data;
+
+	while (!(ret = perf->cmd_recv(perf, &pidx, &cmd, &data))) {
+		peer = &perf->peers[pidx];
+
+		switch (cmd) {
+		case PERF_CMD_SSIZE:
+			peer->inbuf_size = data;
+			return perf_cmd_exec(peer, PERF_CMD_RSIZE);
+		case PERF_CMD_SXLAT:
+			peer->outbuf_xlat = data;
+			return perf_cmd_exec(peer, PERF_CMD_RXLAT);
+		default:
+			dev_err(&perf->ntb->dev, "Recv invalid command\n");
+			return -EINVAL;
+		}
+	}
+
+	/* Return 0 if no data left to process, otherwise an error */
+	return ret == -ENODATA ? 0 : ret;
+}
+
+static void perf_link_event(void *ctx)
+{
+	struct perf_ctx *perf = ctx;
+	struct perf_peer *peer;
+	bool lnk_up;
+	int pidx;
+
+	for (pidx = 0; pidx < perf->pcnt; pidx++) {
+		peer = &perf->peers[pidx];
+
+		lnk_up = perf_link_is_up(peer);
+
+		if (lnk_up &&
+		    !test_and_set_bit(PERF_STS_LNKUP, &peer->sts)) {
+			perf_cmd_exec(peer, PERF_CMD_SSIZE);
+		} else if (!lnk_up &&
+			   test_and_clear_bit(PERF_STS_LNKUP, &peer->sts)) {
+			perf_cmd_exec(peer, PERF_CMD_CLEAR);
+		}
+	}
+}
+
+static void perf_db_event(void *ctx, int vec)
+{
+	struct perf_ctx *perf = ctx;
+
+	dev_dbg(&perf->ntb->dev, "DB vec %d mask %#llx bits %#llx\n", vec,
+		ntb_db_vector_mask(perf->ntb, vec), ntb_db_read(perf->ntb));
+
+	/* Just receive all available commands */
+	(void)perf_cmd_recv(perf);
+}
+
+static void perf_msg_event(void *ctx)
+{
+	struct perf_ctx *perf = ctx;
+
+	dev_dbg(&perf->ntb->dev, "Msg status bits %#llx\n",
+		ntb_msg_read_sts(perf->ntb));
+
+	/* Messages are only sent one-by-one */
+	(void)perf_cmd_recv(perf);
+}
+
+static const struct ntb_ctx_ops perf_ops = {
+	.link_event = perf_link_event,
+	.db_event = perf_db_event,
+	.msg_event = perf_msg_event
+};
+
+static void perf_free_outbuf(struct perf_peer *peer)
+{
+	(void)ntb_peer_mw_clear_trans(peer->perf->ntb, peer->pidx, peer->gidx);
+}
+
+static int perf_setup_outbuf(struct perf_peer *peer)
+{
+	struct perf_ctx *perf = peer->perf;
+	int ret;
+
+	/* Outbuf size can be unaligned due to custom max_mw_size */
+	ret = ntb_peer_mw_set_trans(perf->ntb, peer->pidx, peer->gidx,
+				    peer->outbuf_xlat, peer->outbuf_size);
+	if (ret) {
+		dev_err(&perf->ntb->dev, "Failed to set outbuf translation\n");
+		return ret;
+	}
+
+	/* Initialization is finally done */
+	set_bit(PERF_STS_DONE, &peer->sts);
+	complete_all(&peer->init_comp);
+
+	return 0;
+}
+
+static void perf_free_inbuf(struct perf_peer *peer)
+{
+	if (!peer->inbuf)
+		return;
+
+	(void)ntb_mw_clear_trans(peer->perf->ntb, peer->pidx, peer->gidx);
+	dma_free_coherent(&peer->perf->ntb->pdev->dev, peer->inbuf_size,
+			  peer->inbuf, peer->inbuf_xlat);
+	peer->inbuf = NULL;
+}
+
+static int perf_setup_inbuf(struct perf_peer *peer)
+{
+	resource_size_t xlat_align, size_align, size_max;
+	struct perf_ctx *perf = peer->perf;
+	int ret;
+
+	/* Get inbound MW parameters */
+	ret = ntb_mw_get_align(perf->ntb, peer->pidx, perf->gidx,
+			       &xlat_align, &size_align, &size_max);
+	if (ret) {
+		dev_err(&perf->ntb->dev, "Couldn't get inbuf restrictions\n");
+		return ret;
+	}
+
+	if (peer->inbuf_size > size_max) {
+		dev_err(&perf->ntb->dev, "Too big inbuf size %pa > %pa\n",
+			&peer->inbuf_size, &size_max);
+		return -EINVAL;
+	}
+
+	peer->inbuf_size = round_up(peer->inbuf_size, size_align);
+
+	perf_free_inbuf(peer);
+
+	peer->inbuf = dma_alloc_coherent(&perf->ntb->pdev->dev,
+					 peer->inbuf_size, &peer->inbuf_xlat,
+					 GFP_KERNEL);
+	if (!peer->inbuf) {
+		dev_err(&perf->ntb->dev, "Failed to alloc inbuf of %pa\n",
+			&peer->inbuf_size);
+		return -ENOMEM;
+	}
+	if (!IS_ALIGNED(peer->inbuf_xlat, xlat_align)) {
+		ret = -EINVAL;
+		dev_err(&perf->ntb->dev, "Unaligned inbuf allocated\n");
+		goto err_free_inbuf;
+	}
+
+	ret = ntb_mw_set_trans(perf->ntb, peer->pidx, peer->gidx,
+			       peer->inbuf_xlat, peer->inbuf_size);
+	if (ret) {
+		dev_err(&perf->ntb->dev, "Failed to set inbuf translation\n");
+		goto err_free_inbuf;
+	}
+
+	/*
+	 * We submit inbuf xlat transmission cmd for execution here to follow
+	 * the code architecture, even though this method is called from service
+	 * work itself so the command will be executed right after it returns.
+	 */
+	(void)perf_cmd_exec(peer, PERF_CMD_SXLAT);
+
+	return 0;
+
+err_free_inbuf:
+	perf_free_inbuf(peer);
+
+	return ret;
+}
+
+static void perf_service_work(struct work_struct *work)
+{
+	struct perf_peer *peer = to_peer_service(work);
+
+	if (test_and_clear_bit(PERF_CMD_SSIZE, &peer->sts))
+		perf_cmd_send(peer, PERF_CMD_SSIZE, peer->outbuf_size);
+
+	if (test_and_clear_bit(PERF_CMD_RSIZE, &peer->sts))
+		perf_setup_inbuf(peer);
+
+	if (test_and_clear_bit(PERF_CMD_SXLAT, &peer->sts))
+		perf_cmd_send(peer, PERF_CMD_SXLAT, peer->inbuf_xlat);
+
+	if (test_and_clear_bit(PERF_CMD_RXLAT, &peer->sts))
+		perf_setup_outbuf(peer);
+
+	if (test_and_clear_bit(PERF_CMD_CLEAR, &peer->sts)) {
+		init_completion(&peer->init_comp);
+		clear_bit(PERF_STS_DONE, &peer->sts);
+		if (test_bit(0, &peer->perf->busy_flag) &&
+		    peer == peer->perf->test_peer) {
+			dev_warn(&peer->perf->ntb->dev,
+				"Freeing while test on-fly\n");
+			perf_terminate_test(peer->perf);
+		}
+		perf_free_outbuf(peer);
+		perf_free_inbuf(peer);
+	}
+}
+
+static int perf_init_service(struct perf_ctx *perf)
+{
+	u64 mask;
+
+	if (ntb_peer_mw_count(perf->ntb) < perf->pcnt) {
+		dev_err(&perf->ntb->dev, "Not enough memory windows\n");
+		return -EINVAL;
+	}
+
+	if (ntb_msg_count(perf->ntb) >= PERF_MSG_CNT) {
+		perf->cmd_send = perf_msg_cmd_send;
+		perf->cmd_recv = perf_msg_cmd_recv;
+
+		dev_dbg(&perf->ntb->dev, "Message service initialized\n");
+
+		return 0;
+	}
+
+	dev_dbg(&perf->ntb->dev, "Message service unsupported\n");
+
+	mask = GENMASK_ULL(perf->pcnt, 0);
+	if (ntb_spad_count(perf->ntb) >= PERF_SPAD_CNT(perf->pcnt) &&
+	    (ntb_db_valid_mask(perf->ntb) & mask) == mask) {
+		perf->cmd_send = perf_spad_cmd_send;
+		perf->cmd_recv = perf_spad_cmd_recv;
+
+		dev_dbg(&perf->ntb->dev, "Scratchpad service initialized\n");
+
+		return 0;
+	}
+
+	dev_dbg(&perf->ntb->dev, "Scratchpad service unsupported\n");
+
+	dev_err(&perf->ntb->dev, "Command services unsupported\n");
+
+	return -EINVAL;
+}
+
+static int perf_enable_service(struct perf_ctx *perf)
+{
+	u64 mask, incmd_bit;
+	int ret, sidx, scnt;
+
+	mask = ntb_db_valid_mask(perf->ntb);
+	(void)ntb_db_set_mask(perf->ntb, mask);
+
+	ret = ntb_set_ctx(perf->ntb, perf, &perf_ops);
+	if (ret)
+		return ret;
+
+	if (perf->cmd_send == perf_msg_cmd_send) {
+		u64 inbits, outbits;
+
+		inbits = ntb_msg_inbits(perf->ntb);
+		outbits = ntb_msg_outbits(perf->ntb);
+		(void)ntb_msg_set_mask(perf->ntb, inbits | outbits);
+
+		incmd_bit = BIT_ULL(__ffs64(inbits));
+		ret = ntb_msg_clear_mask(perf->ntb, incmd_bit);
+
+		dev_dbg(&perf->ntb->dev, "MSG sts unmasked %#llx\n", incmd_bit);
+	} else {
+		scnt = ntb_spad_count(perf->ntb);
+		for (sidx = 0; sidx < scnt; sidx++)
+			ntb_spad_write(perf->ntb, sidx, PERF_CMD_INVAL);
+		incmd_bit = PERF_SPAD_NOTIFY(perf->gidx);
+		ret = ntb_db_clear_mask(perf->ntb, incmd_bit);
+
+		dev_dbg(&perf->ntb->dev, "DB bits unmasked %#llx\n", incmd_bit);
+	}
+	if (ret) {
+		ntb_clear_ctx(perf->ntb);
+		return ret;
+	}
+
+	ntb_link_enable(perf->ntb, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
+	/* Might be not necessary */
+	ntb_link_event(perf->ntb);
+
+	return 0;
+}
+
+static void perf_disable_service(struct perf_ctx *perf)
+{
+	int pidx;
+
+	if (perf->cmd_send == perf_msg_cmd_send) {
+		u64 inbits;
+
+		inbits = ntb_msg_inbits(perf->ntb);
+		(void)ntb_msg_set_mask(perf->ntb, inbits);
+	} else {
+		(void)ntb_db_set_mask(perf->ntb, PERF_SPAD_NOTIFY(perf->gidx));
+	}
+
+	ntb_clear_ctx(perf->ntb);
+
+	for (pidx = 0; pidx < perf->pcnt; pidx++)
+		perf_cmd_exec(&perf->peers[pidx], PERF_CMD_CLEAR);
+
+	for (pidx = 0; pidx < perf->pcnt; pidx++)
+		flush_work(&perf->peers[pidx].service);
+
+	for (pidx = 0; pidx < perf->pcnt; pidx++) {
+		struct perf_peer *peer = &perf->peers[pidx];
+
+		ntb_spad_write(perf->ntb, PERF_SPAD_CMD(peer->gidx), 0);
+	}
+
+	ntb_db_clear(perf->ntb, PERF_SPAD_NOTIFY(perf->gidx));
+
+	ntb_link_disable(perf->ntb);
+}
+
+/*==============================================================================
+ *                      Performance measuring work-thread
+ *==============================================================================
+ */
+
+static void perf_dma_copy_callback(void *data)
+{
+	struct perf_thread *pthr = data;
+
+	atomic_dec(&pthr->dma_sync);
+	wake_up(&pthr->dma_wait);
+}
+
+static int perf_copy_chunk(struct perf_thread *pthr,
+			   void __iomem *dst, void *src, size_t len)
+{
+	struct dma_async_tx_descriptor *tx;
+	struct dmaengine_unmap_data *unmap;
+	struct device *dma_dev;
+	int try = 0, ret = 0;
+	struct perf_peer *peer = pthr->perf->test_peer;
+	void __iomem *vbase;
+	void __iomem *dst_vaddr;
+	dma_addr_t dst_dma_addr;
+
+	if (!use_dma) {
+		memcpy_toio(dst, src, len);
+		goto ret_check_tsync;
+	}
+
+	dma_dev = pthr->dma_chan->device->dev;
+
+	if (!is_dma_copy_aligned(pthr->dma_chan->device, offset_in_page(src),
+				 offset_in_page(dst), len))
+		return -EIO;
+
+	vbase = peer->outbuf;
+	dst_vaddr = dst;
+	dst_dma_addr = peer->dma_dst_addr + (dst_vaddr - vbase);
+
+	unmap = dmaengine_get_unmap_data(dma_dev, 1, GFP_NOWAIT);
+	if (!unmap)
+		return -ENOMEM;
+
+	unmap->len = len;
+	unmap->addr[0] = dma_map_page(dma_dev, virt_to_page(src),
+		offset_in_page(src), len, DMA_TO_DEVICE);
+	if (dma_mapping_error(dma_dev, unmap->addr[0])) {
+		ret = -EIO;
+		goto err_free_resource;
+	}
+	unmap->to_cnt = 1;
+
+	do {
+		tx = dmaengine_prep_dma_memcpy(pthr->dma_chan, dst_dma_addr,
+			unmap->addr[0], len, DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
+		if (!tx)
+			msleep(DMA_MDELAY);
+	} while (!tx && (try++ < DMA_TRIES));
+
+	if (!tx) {
+		ret = -EIO;
+		goto err_free_resource;
+	}
+
+	tx->callback = perf_dma_copy_callback;
+	tx->callback_param = pthr;
+	dma_set_unmap(tx, unmap);
+
+	ret = dma_submit_error(dmaengine_submit(tx));
+	if (ret)
+		goto err_free_resource;
+
+	dmaengine_unmap_put(unmap);
+
+	atomic_inc(&pthr->dma_sync);
+	dma_async_issue_pending(pthr->dma_chan);
+
+ret_check_tsync:
+	return likely(atomic_read(&pthr->perf->tsync) > 0) ? 0 : -EINTR;
+
+err_free_resource:
+	dmaengine_unmap_put(unmap);
+
+	return ret;
+}
+
+static bool perf_dma_filter(struct dma_chan *chan, void *data)
+{
+	struct perf_ctx *perf = data;
+	int node;
+
+	node = dev_to_node(&perf->ntb->dev);
+
+	return node == NUMA_NO_NODE || node == dev_to_node(chan->device->dev);
+}
+
+static int perf_init_test(struct perf_thread *pthr)
+{
+	struct perf_ctx *perf = pthr->perf;
+	dma_cap_mask_t dma_mask;
+	struct perf_peer *peer = pthr->perf->test_peer;
+
+	pthr->src = kmalloc_node(perf->test_peer->outbuf_size, GFP_KERNEL,
+				 dev_to_node(&perf->ntb->dev));
+	if (!pthr->src)
+		return -ENOMEM;
+
+	get_random_bytes(pthr->src, perf->test_peer->outbuf_size);
+
+	if (!use_dma)
+		return 0;
+
+	dma_cap_zero(dma_mask);
+	dma_cap_set(DMA_MEMCPY, dma_mask);
+	pthr->dma_chan = dma_request_channel(dma_mask, perf_dma_filter, perf);
+	if (!pthr->dma_chan) {
+		dev_err(&perf->ntb->dev, "%d: Failed to get DMA channel\n",
+			pthr->tidx);
+		goto err_free;
+	}
+	peer->dma_dst_addr =
+		dma_map_resource(pthr->dma_chan->device->dev,
+				 peer->out_phys_addr, peer->outbuf_size,
+				 DMA_FROM_DEVICE, 0);
+	if (dma_mapping_error(pthr->dma_chan->device->dev,
+			      peer->dma_dst_addr)) {
+		dev_err(pthr->dma_chan->device->dev, "%d: Failed to map DMA addr\n",
+			pthr->tidx);
+		peer->dma_dst_addr = 0;
+		dma_release_channel(pthr->dma_chan);
+		goto err_free;
+	}
+	dev_dbg(pthr->dma_chan->device->dev, "%d: Map MMIO %pa to DMA addr %pad\n",
+			pthr->tidx,
+			&peer->out_phys_addr,
+			&peer->dma_dst_addr);
+
+	atomic_set(&pthr->dma_sync, 0);
+	return 0;
+
+err_free:
+	atomic_dec(&perf->tsync);
+	wake_up(&perf->twait);
+	kfree(pthr->src);
+	return -ENODEV;
+}
+
+static int perf_run_test(struct perf_thread *pthr)
+{
+	struct perf_peer *peer = pthr->perf->test_peer;
+	struct perf_ctx *perf = pthr->perf;
+	void __iomem *flt_dst, *bnd_dst;
+	u64 total_size, chunk_size;
+	void *flt_src;
+	int ret = 0;
+
+	total_size = 1ULL << total_order;
+	chunk_size = 1ULL << chunk_order;
+	chunk_size = min_t(u64, peer->outbuf_size, chunk_size);
+
+	flt_src = pthr->src;
+	bnd_dst = peer->outbuf + peer->outbuf_size;
+	flt_dst = peer->outbuf;
+
+	pthr->duration = ktime_get();
+
+	/* Copied field is cleared on test launch stage */
+	while (pthr->copied < total_size) {
+		ret = perf_copy_chunk(pthr, flt_dst, flt_src, chunk_size);
+		if (ret) {
+			dev_err(&perf->ntb->dev, "%d: Got error %d on test\n",
+				pthr->tidx, ret);
+			return ret;
+		}
+
+		pthr->copied += chunk_size;
+
+		flt_dst += chunk_size;
+		flt_src += chunk_size;
+		if (flt_dst >= bnd_dst || flt_dst < peer->outbuf) {
+			flt_dst = peer->outbuf;
+			flt_src = pthr->src;
+		}
+
+		/* Give up CPU to give a chance for other threads to use it */
+		schedule();
+	}
+
+	return 0;
+}
+
+static int perf_sync_test(struct perf_thread *pthr)
+{
+	struct perf_ctx *perf = pthr->perf;
+
+	if (!use_dma)
+		goto no_dma_ret;
+
+	wait_event(pthr->dma_wait,
+		   (atomic_read(&pthr->dma_sync) == 0 ||
+		    atomic_read(&perf->tsync) < 0));
+
+	if (atomic_read(&perf->tsync) < 0)
+		return -EINTR;
+
+no_dma_ret:
+	pthr->duration = ktime_sub(ktime_get(), pthr->duration);
+
+	dev_dbg(&perf->ntb->dev, "%d: copied %llu bytes\n",
+		pthr->tidx, pthr->copied);
+
+	dev_dbg(&perf->ntb->dev, "%d: lasted %llu usecs\n",
+		pthr->tidx, ktime_to_us(pthr->duration));
+
+	dev_dbg(&perf->ntb->dev, "%d: %llu MBytes/s\n", pthr->tidx,
+		div64_u64(pthr->copied, ktime_to_us(pthr->duration)));
+
+	return 0;
+}
+
+static void perf_clear_test(struct perf_thread *pthr)
+{
+	struct perf_ctx *perf = pthr->perf;
+
+	if (!use_dma)
+		goto no_dma_notify;
+
+	/*
+	 * If test finished without errors, termination isn't needed.
+	 * We call it anyway just to be sure of the transfers completion.
+	 */
+	(void)dmaengine_terminate_sync(pthr->dma_chan);
+	if (pthr->perf->test_peer->dma_dst_addr)
+		dma_unmap_resource(pthr->dma_chan->device->dev,
+				   pthr->perf->test_peer->dma_dst_addr,
+				   pthr->perf->test_peer->outbuf_size,
+				   DMA_FROM_DEVICE, 0);
+
+	dma_release_channel(pthr->dma_chan);
+
+no_dma_notify:
+	atomic_dec(&perf->tsync);
+	wake_up(&perf->twait);
+	kfree(pthr->src);
+}
+
+static void perf_thread_work(struct work_struct *work)
+{
+	struct perf_thread *pthr = to_thread_work(work);
+	int ret;
+
+	/*
+	 * Perform stages in compliance with use_dma flag value.
+	 * Test status is changed only if error happened, otherwise
+	 * status -ENODATA is kept while test is on-fly. Results
+	 * synchronization is performed only if test fininshed
+	 * without an error or interruption.
+	 */
+	ret = perf_init_test(pthr);
+	if (ret) {
+		pthr->status = ret;
+		return;
+	}
+
+	ret = perf_run_test(pthr);
+	if (ret) {
+		pthr->status = ret;
+		goto err_clear_test;
+	}
+
+	pthr->status = perf_sync_test(pthr);
+
+err_clear_test:
+	perf_clear_test(pthr);
+}
+
+static int perf_set_tcnt(struct perf_ctx *perf, u8 tcnt)
+{
+	if (tcnt == 0 || tcnt > MAX_THREADS_CNT)
+		return -EINVAL;
+
+	if (test_and_set_bit_lock(0, &perf->busy_flag))
+		return -EBUSY;
+
+	perf->tcnt = tcnt;
+
+	clear_bit_unlock(0, &perf->busy_flag);
+
+	return 0;
+}
+
+static void perf_terminate_test(struct perf_ctx *perf)
+{
+	int tidx;
+
+	atomic_set(&perf->tsync, -1);
+	wake_up(&perf->twait);
+
+	for (tidx = 0; tidx < MAX_THREADS_CNT; tidx++) {
+		wake_up(&perf->threads[tidx].dma_wait);
+		cancel_work_sync(&perf->threads[tidx].work);
+	}
+}
+
+static int perf_submit_test(struct perf_peer *peer)
+{
+	struct perf_ctx *perf = peer->perf;
+	struct perf_thread *pthr;
+	int tidx, ret;
+
+	ret = wait_for_completion_interruptible(&peer->init_comp);
+	if (ret < 0)
+		return ret;
+
+	if (test_and_set_bit_lock(0, &perf->busy_flag))
+		return -EBUSY;
+
+	perf->test_peer = peer;
+	atomic_set(&perf->tsync, perf->tcnt);
+
+	for (tidx = 0; tidx < MAX_THREADS_CNT; tidx++) {
+		pthr = &perf->threads[tidx];
+
+		pthr->status = -ENODATA;
+		pthr->copied = 0;
+		pthr->duration = ktime_set(0, 0);
+		if (tidx < perf->tcnt)
+			(void)queue_work(perf_wq, &pthr->work);
+	}
+
+	ret = wait_event_interruptible(perf->twait,
+				       atomic_read(&perf->tsync) <= 0);
+	if (ret == -ERESTARTSYS) {
+		perf_terminate_test(perf);
+		ret = -EINTR;
+	}
+
+	clear_bit_unlock(0, &perf->busy_flag);
+
+	return ret;
+}
+
+static int perf_read_stats(struct perf_ctx *perf, char *buf,
+			   size_t size, ssize_t *pos)
+{
+	struct perf_thread *pthr;
+	int tidx;
+
+	if (test_and_set_bit_lock(0, &perf->busy_flag))
+		return -EBUSY;
+
+	(*pos) += scnprintf(buf + *pos, size - *pos,
+		"    Peer %d test statistics:\n", perf->test_peer->pidx);
+
+	for (tidx = 0; tidx < MAX_THREADS_CNT; tidx++) {
+		pthr = &perf->threads[tidx];
+
+		if (pthr->status == -ENODATA)
+			continue;
+
+		if (pthr->status) {
+			(*pos) += scnprintf(buf + *pos, size - *pos,
+				"%d: error status %d\n", tidx, pthr->status);
+			continue;
+		}
+
+		(*pos) += scnprintf(buf + *pos, size - *pos,
+			"%d: copied %llu bytes in %llu usecs, %llu MBytes/s\n",
+			tidx, pthr->copied, ktime_to_us(pthr->duration),
+			div64_u64(pthr->copied, ktime_to_us(pthr->duration)));
+	}
+
+	clear_bit_unlock(0, &perf->busy_flag);
+
+	return 0;
+}
+
+static void perf_init_threads(struct perf_ctx *perf)
+{
+	struct perf_thread *pthr;
+	int tidx;
+
+	perf->tcnt = DEF_THREADS_CNT;
+	perf->test_peer = &perf->peers[0];
+	init_waitqueue_head(&perf->twait);
+
+	for (tidx = 0; tidx < MAX_THREADS_CNT; tidx++) {
+		pthr = &perf->threads[tidx];
+
+		pthr->perf = perf;
+		pthr->tidx = tidx;
+		pthr->status = -ENODATA;
+		init_waitqueue_head(&pthr->dma_wait);
+		INIT_WORK(&pthr->work, perf_thread_work);
+	}
+}
+
+static void perf_clear_threads(struct perf_ctx *perf)
+{
+	perf_terminate_test(perf);
+}
+
+/*==============================================================================
+ *                               DebugFS nodes
+ *==============================================================================
+ */
+
+static ssize_t perf_dbgfs_read_info(struct file *filep, char __user *ubuf,
+				    size_t size, loff_t *offp)
+{
+	struct perf_ctx *perf = filep->private_data;
+	struct perf_peer *peer;
+	size_t buf_size;
+	ssize_t pos = 0;
+	int ret, pidx;
+	char *buf;
+
+	buf_size = min_t(size_t, size, 0x1000U);
+
+	buf = kmalloc(buf_size, GFP_KERNEL);
+	if (!buf)
+		return -ENOMEM;
+
+	pos += scnprintf(buf + pos, buf_size - pos,
+		"    Performance measuring tool info:\n\n");
+
+	pos += scnprintf(buf + pos, buf_size - pos,
+		"Local port %d, Global index %d\n", ntb_port_number(perf->ntb),
+		perf->gidx);
+	pos += scnprintf(buf + pos, buf_size - pos, "Test status: ");
+	if (test_bit(0, &perf->busy_flag)) {
+		pos += scnprintf(buf + pos, buf_size - pos,
+			"on-fly with port %d (%d)\n",
+			ntb_peer_port_number(perf->ntb, perf->test_peer->pidx),
+			perf->test_peer->pidx);
+	} else {
+		pos += scnprintf(buf + pos, buf_size - pos, "idle\n");
+	}
+
+	for (pidx = 0; pidx < perf->pcnt; pidx++) {
+		peer = &perf->peers[pidx];
+
+		pos += scnprintf(buf + pos, buf_size - pos,
+			"Port %d (%d), Global index %d:\n",
+			ntb_peer_port_number(perf->ntb, peer->pidx), peer->pidx,
+			peer->gidx);
+
+		pos += scnprintf(buf + pos, buf_size - pos,
+			"\tLink status: %s\n",
+			test_bit(PERF_STS_LNKUP, &peer->sts) ? "up" : "down");
+
+		pos += scnprintf(buf + pos, buf_size - pos,
+			"\tOut buffer addr 0x%pK\n", peer->outbuf);
+
+		pos += scnprintf(buf + pos, buf_size - pos,
+			"\tOut buff phys addr %pap\n", &peer->out_phys_addr);
+
+		pos += scnprintf(buf + pos, buf_size - pos,
+			"\tOut buffer size %pa\n", &peer->outbuf_size);
+
+		pos += scnprintf(buf + pos, buf_size - pos,
+			"\tOut buffer xlat 0x%016llx[p]\n", peer->outbuf_xlat);
+
+		if (!peer->inbuf) {
+			pos += scnprintf(buf + pos, buf_size - pos,
+				"\tIn buffer addr: unallocated\n");
+			continue;
+		}
+
+		pos += scnprintf(buf + pos, buf_size - pos,
+			"\tIn buffer addr 0x%pK\n", peer->inbuf);
+
+		pos += scnprintf(buf + pos, buf_size - pos,
+			"\tIn buffer size %pa\n", &peer->inbuf_size);
+
+		pos += scnprintf(buf + pos, buf_size - pos,
+			"\tIn buffer xlat %pad[p]\n", &peer->inbuf_xlat);
+	}
+
+	ret = simple_read_from_buffer(ubuf, size, offp, buf, pos);
+	kfree(buf);
+
+	return ret;
+}
+
+static const struct file_operations perf_dbgfs_info = {
+	.open = simple_open,
+	.read = perf_dbgfs_read_info
+};
+
+static ssize_t perf_dbgfs_read_run(struct file *filep, char __user *ubuf,
+				   size_t size, loff_t *offp)
+{
+	struct perf_ctx *perf = filep->private_data;
+	ssize_t ret, pos = 0;
+	char *buf;
+
+	buf = kmalloc(PERF_BUF_LEN, GFP_KERNEL);
+	if (!buf)
+		return -ENOMEM;
+
+	ret = perf_read_stats(perf, buf, PERF_BUF_LEN, &pos);
+	if (ret)
+		goto err_free;
+
+	ret = simple_read_from_buffer(ubuf, size, offp, buf, pos);
+err_free:
+	kfree(buf);
+
+	return ret;
+}
+
+static ssize_t perf_dbgfs_write_run(struct file *filep, const char __user *ubuf,
+				    size_t size, loff_t *offp)
+{
+	struct perf_ctx *perf = filep->private_data;
+	struct perf_peer *peer;
+	int pidx, ret;
+
+	ret = kstrtoint_from_user(ubuf, size, 0, &pidx);
+	if (ret)
+		return ret;
+
+	if (pidx < 0 || pidx >= perf->pcnt)
+		return -EINVAL;
+
+	peer = &perf->peers[pidx];
+
+	ret = perf_submit_test(peer);
+	if (ret)
+		return ret;
+
+	return size;
+}
+
+static const struct file_operations perf_dbgfs_run = {
+	.open = simple_open,
+	.read = perf_dbgfs_read_run,
+	.write = perf_dbgfs_write_run
+};
+
+static ssize_t perf_dbgfs_read_tcnt(struct file *filep, char __user *ubuf,
+				    size_t size, loff_t *offp)
+{
+	struct perf_ctx *perf = filep->private_data;
+	char buf[8];
+	ssize_t pos;
+
+	pos = scnprintf(buf, sizeof(buf), "%hhu\n", perf->tcnt);
+
+	return simple_read_from_buffer(ubuf, size, offp, buf, pos);
+}
+
+static ssize_t perf_dbgfs_write_tcnt(struct file *filep,
+				     const char __user *ubuf,
+				     size_t size, loff_t *offp)
+{
+	struct perf_ctx *perf = filep->private_data;
+	int ret;
+	u8 val;
+
+	ret = kstrtou8_from_user(ubuf, size, 0, &val);
+	if (ret)
+		return ret;
+
+	ret = perf_set_tcnt(perf, val);
+	if (ret)
+		return ret;
+
+	return size;
+}
+
+static const struct file_operations perf_dbgfs_tcnt = {
+	.open = simple_open,
+	.read = perf_dbgfs_read_tcnt,
+	.write = perf_dbgfs_write_tcnt
+};
+
+static void perf_setup_dbgfs(struct perf_ctx *perf)
+{
+	struct pci_dev *pdev = perf->ntb->pdev;
+
+	perf->dbgfs_dir = debugfs_create_dir(pci_name(pdev), perf_dbgfs_topdir);
+	if (IS_ERR(perf->dbgfs_dir)) {
+		dev_warn(&perf->ntb->dev, "DebugFS unsupported\n");
+		return;
+	}
+
+	debugfs_create_file("info", 0600, perf->dbgfs_dir, perf,
+			    &perf_dbgfs_info);
+
+	debugfs_create_file("run", 0600, perf->dbgfs_dir, perf,
+			    &perf_dbgfs_run);
+
+	debugfs_create_file("threads_count", 0600, perf->dbgfs_dir, perf,
+			    &perf_dbgfs_tcnt);
+
+	/* They are made read-only for test exec safety and integrity */
+	debugfs_create_u8("chunk_order", 0500, perf->dbgfs_dir, &chunk_order);
+
+	debugfs_create_u8("total_order", 0500, perf->dbgfs_dir, &total_order);
+
+	debugfs_create_bool("use_dma", 0500, perf->dbgfs_dir, &use_dma);
+}
+
+static void perf_clear_dbgfs(struct perf_ctx *perf)
+{
+	debugfs_remove_recursive(perf->dbgfs_dir);
+}
+
+/*==============================================================================
+ *                        Basic driver initialization
+ *==============================================================================
+ */
+
+static struct perf_ctx *perf_create_data(struct ntb_dev *ntb)
+{
+	struct perf_ctx *perf;
+
+	perf = devm_kzalloc(&ntb->dev, sizeof(*perf), GFP_KERNEL);
+	if (!perf)
+		return ERR_PTR(-ENOMEM);
+
+	perf->pcnt = ntb_peer_port_count(ntb);
+	perf->peers = devm_kcalloc(&ntb->dev, perf->pcnt, sizeof(*perf->peers),
+				  GFP_KERNEL);
+	if (!perf->peers)
+		return ERR_PTR(-ENOMEM);
+
+	perf->ntb = ntb;
+
+	return perf;
+}
+
+static int perf_setup_peer_mw(struct perf_peer *peer)
+{
+	struct perf_ctx *perf = peer->perf;
+	phys_addr_t phys_addr;
+	int ret;
+
+	/* Get outbound MW parameters and map it */
+	ret = ntb_peer_mw_get_addr(perf->ntb, perf->gidx, &phys_addr,
+				   &peer->outbuf_size);
+	if (ret)
+		return ret;
+
+	peer->outbuf = devm_ioremap_wc(&perf->ntb->dev, phys_addr,
+					peer->outbuf_size);
+	if (!peer->outbuf)
+		return -ENOMEM;
+
+	peer->out_phys_addr = phys_addr;
+
+	if (max_mw_size && peer->outbuf_size > max_mw_size) {
+		peer->outbuf_size = max_mw_size;
+		dev_warn(&peer->perf->ntb->dev,
+			"Peer %d outbuf reduced to %pa\n", peer->pidx,
+			&peer->outbuf_size);
+	}
+
+	return 0;
+}
+
+static int perf_init_peers(struct perf_ctx *perf)
+{
+	struct perf_peer *peer;
+	int pidx, lport, ret;
+
+	lport = ntb_port_number(perf->ntb);
+	perf->gidx = -1;
+	for (pidx = 0; pidx < perf->pcnt; pidx++) {
+		peer = &perf->peers[pidx];
+
+		peer->perf = perf;
+		peer->pidx = pidx;
+		if (lport < ntb_peer_port_number(perf->ntb, pidx)) {
+			if (perf->gidx == -1)
+				perf->gidx = pidx;
+			peer->gidx = pidx + 1;
+		} else {
+			peer->gidx = pidx;
+		}
+		INIT_WORK(&peer->service, perf_service_work);
+		init_completion(&peer->init_comp);
+	}
+	if (perf->gidx == -1)
+		perf->gidx = pidx;
+
+	/*
+	 * Hardware with only two ports may not have unique port
+	 * numbers. In this case, the gidxs should all be zero.
+	 */
+	if (perf->pcnt == 1 &&  ntb_port_number(perf->ntb) == 0 &&
+	    ntb_peer_port_number(perf->ntb, 0) == 0) {
+		perf->gidx = 0;
+		perf->peers[0].gidx = 0;
+	}
+
+	for (pidx = 0; pidx < perf->pcnt; pidx++) {
+		ret = perf_setup_peer_mw(&perf->peers[pidx]);
+		if (ret)
+			return ret;
+	}
+
+	dev_dbg(&perf->ntb->dev, "Global port index %d\n", perf->gidx);
+
+	return 0;
+}
+
+static int perf_probe(struct ntb_client *client, struct ntb_dev *ntb)
+{
+	struct perf_ctx *perf;
+	int ret;
+
+	perf = perf_create_data(ntb);
+	if (IS_ERR(perf))
+		return PTR_ERR(perf);
+
+	ret = perf_init_peers(perf);
+	if (ret)
+		return ret;
+
+	perf_init_threads(perf);
+
+	ret = perf_init_service(perf);
+	if (ret)
+		return ret;
+
+	ret = perf_enable_service(perf);
+	if (ret)
+		return ret;
+
+	perf_setup_dbgfs(perf);
+
+	return 0;
+}
+
+static void perf_remove(struct ntb_client *client, struct ntb_dev *ntb)
+{
+	struct perf_ctx *perf = ntb->ctx;
+
+	perf_clear_dbgfs(perf);
+
+	perf_disable_service(perf);
+
+	perf_clear_threads(perf);
+}
+
+static struct ntb_client perf_client = {
+	.ops = {
+		.probe = perf_probe,
+		.remove = perf_remove
+	}
+};
+
+static int __init perf_init(void)
+{
+	int ret;
+
+	if (chunk_order > MAX_CHUNK_ORDER) {
+		chunk_order = MAX_CHUNK_ORDER;
+		pr_info("Chunk order reduced to %hhu\n", chunk_order);
+	}
+
+	if (total_order < chunk_order) {
+		total_order = chunk_order;
+		pr_info("Total data order reduced to %hhu\n", total_order);
+	}
+
+	perf_wq = alloc_workqueue("perf_wq", WQ_UNBOUND | WQ_SYSFS, 0);
+	if (!perf_wq)
+		return -ENOMEM;
+
+	if (debugfs_initialized())
+		perf_dbgfs_topdir = debugfs_create_dir(KBUILD_MODNAME, NULL);
+
+	ret = ntb_register_client(&perf_client);
+	if (ret) {
+		debugfs_remove_recursive(perf_dbgfs_topdir);
+		destroy_workqueue(perf_wq);
+	}
+
+	return ret;
+}
+module_init(perf_init);
+
+static void __exit perf_exit(void)
+{
+	ntb_unregister_client(&perf_client);
+	debugfs_remove_recursive(perf_dbgfs_topdir);
+	destroy_workqueue(perf_wq);
+}
+module_exit(perf_exit);
diff --git a/drivers/ntb/test/ntb_pingpong.c b/drivers/ntb/test/ntb_pingpong.c
new file mode 100644
index 000000000..1c1c74f4f
--- /dev/null
+++ b/drivers/ntb/test/ntb_pingpong.c
@@ -0,0 +1,432 @@
+/*
+ *   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) 2015 EMC Corporation. All Rights Reserved.
+ *   Copyright (C) 2017 T-Platforms. All Rights Reserved.
+ *
+ *   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) 2015 EMC Corporation. All Rights Reserved.
+ *   Copyright (C) 2017 T-Platforms. All Rights Reserved.
+ *
+ *   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 copy
+ *       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.
+ *
+ * PCIe NTB Pingpong Linux driver
+ */
+
+/*
+ * How to use this tool, by example.
+ *
+ * Assuming $DBG_DIR is something like:
+ * '/sys/kernel/debug/ntb_perf/0000:00:03.0'
+ * Suppose aside from local device there is at least one remote device
+ * connected to NTB with index 0.
+ *-----------------------------------------------------------------------------
+ * Eg: install driver with specified delay between doorbell event and response
+ *
+ * root@self# insmod ntb_pingpong.ko delay_ms=1000
+ *-----------------------------------------------------------------------------
+ * Eg: get number of ping-pong cycles performed
+ *
+ * root@self# cat $DBG_DIR/count
+ */
+
+#include <linux/init.h>
+#include <linux/kernel.h>
+#include <linux/module.h>
+#include <linux/device.h>
+#include <linux/bitops.h>
+
+#include <linux/pci.h>
+#include <linux/slab.h>
+#include <linux/hrtimer.h>
+#include <linux/debugfs.h>
+
+#include <linux/ntb.h>
+
+#define DRIVER_NAME		"ntb_pingpong"
+#define DRIVER_VERSION		"2.0"
+
+MODULE_LICENSE("Dual BSD/GPL");
+MODULE_VERSION(DRIVER_VERSION);
+MODULE_AUTHOR("Allen Hubbe <Allen.Hubbe@emc.com>");
+MODULE_DESCRIPTION("PCIe NTB Simple Pingpong Client");
+
+static unsigned int unsafe;
+module_param(unsafe, uint, 0644);
+MODULE_PARM_DESC(unsafe, "Run even though ntb operations may be unsafe");
+
+static unsigned int delay_ms = 1000;
+module_param(delay_ms, uint, 0644);
+MODULE_PARM_DESC(delay_ms, "Milliseconds to delay the response to peer");
+
+struct pp_ctx {
+	struct ntb_dev *ntb;
+	struct hrtimer timer;
+	u64 in_db;
+	u64 out_db;
+	int out_pidx;
+	u64 nmask;
+	u64 pmask;
+	atomic_t count;
+	spinlock_t lock;
+	struct dentry *dbgfs_dir;
+};
+#define to_pp_timer(__timer) \
+	container_of(__timer, struct pp_ctx, timer)
+
+static struct dentry *pp_dbgfs_topdir;
+
+static int pp_find_next_peer(struct pp_ctx *pp)
+{
+	u64 link, out_db;
+	int pidx;
+
+	link = ntb_link_is_up(pp->ntb, NULL, NULL);
+
+	/* Find next available peer */
+	if (link & pp->nmask)
+		pidx = __ffs64(link & pp->nmask);
+	else if (link & pp->pmask)
+		pidx = __ffs64(link & pp->pmask);
+	else
+		return -ENODEV;
+
+	out_db = BIT_ULL(ntb_peer_port_number(pp->ntb, pidx));
+
+	spin_lock(&pp->lock);
+	pp->out_pidx = pidx;
+	pp->out_db = out_db;
+	spin_unlock(&pp->lock);
+
+	return 0;
+}
+
+static void pp_setup(struct pp_ctx *pp)
+{
+	int ret;
+
+	ntb_db_set_mask(pp->ntb, pp->in_db);
+
+	hrtimer_cancel(&pp->timer);
+
+	ret = pp_find_next_peer(pp);
+	if (ret == -ENODEV) {
+		dev_dbg(&pp->ntb->dev, "Got no peers, so cancel\n");
+		return;
+	}
+
+	dev_dbg(&pp->ntb->dev, "Ping-pong started with port %d, db %#llx\n",
+		ntb_peer_port_number(pp->ntb, pp->out_pidx), pp->out_db);
+
+	hrtimer_start(&pp->timer, ms_to_ktime(delay_ms), HRTIMER_MODE_REL);
+}
+
+static void pp_clear(struct pp_ctx *pp)
+{
+	hrtimer_cancel(&pp->timer);
+
+	ntb_db_set_mask(pp->ntb, pp->in_db);
+
+	dev_dbg(&pp->ntb->dev, "Ping-pong cancelled\n");
+}
+
+static void pp_ping(struct pp_ctx *pp)
+{
+	u32 count;
+
+	count = atomic_read(&pp->count);
+
+	spin_lock(&pp->lock);
+	ntb_peer_spad_write(pp->ntb, pp->out_pidx, 0, count);
+	ntb_peer_msg_write(pp->ntb, pp->out_pidx, 0, count);
+
+	dev_dbg(&pp->ntb->dev, "Ping port %d spad %#x, msg %#x\n",
+		ntb_peer_port_number(pp->ntb, pp->out_pidx), count, count);
+
+	ntb_peer_db_set(pp->ntb, pp->out_db);
+	ntb_db_clear_mask(pp->ntb, pp->in_db);
+	spin_unlock(&pp->lock);
+}
+
+static void pp_pong(struct pp_ctx *pp)
+{
+	u32 msg_data, spad_data;
+	int pidx = 0;
+
+	/* Read pong data */
+	spad_data = ntb_spad_read(pp->ntb, 0);
+	msg_data = ntb_msg_read(pp->ntb, &pidx, 0);
+	ntb_msg_clear_sts(pp->ntb, -1);
+
+	/*
+	 * Scratchpad and message data may differ, since message register can't
+	 * be rewritten unless status is cleared. Additionally either of them
+	 * might be unsupported
+	 */
+	dev_dbg(&pp->ntb->dev, "Pong spad %#x, msg %#x (port %d)\n",
+		spad_data, msg_data, ntb_peer_port_number(pp->ntb, pidx));
+
+	atomic_inc(&pp->count);
+
+	ntb_db_set_mask(pp->ntb, pp->in_db);
+	ntb_db_clear(pp->ntb, pp->in_db);
+
+	hrtimer_start(&pp->timer, ms_to_ktime(delay_ms), HRTIMER_MODE_REL);
+}
+
+static enum hrtimer_restart pp_timer_func(struct hrtimer *t)
+{
+	struct pp_ctx *pp = to_pp_timer(t);
+
+	pp_ping(pp);
+
+	return HRTIMER_NORESTART;
+}
+
+static void pp_link_event(void *ctx)
+{
+	struct pp_ctx *pp = ctx;
+
+	pp_setup(pp);
+}
+
+static void pp_db_event(void *ctx, int vec)
+{
+	struct pp_ctx *pp = ctx;
+
+	pp_pong(pp);
+}
+
+static const struct ntb_ctx_ops pp_ops = {
+	.link_event = pp_link_event,
+	.db_event = pp_db_event
+};
+
+static int pp_check_ntb(struct ntb_dev *ntb)
+{
+	u64 pmask;
+
+	if (ntb_db_is_unsafe(ntb)) {
+		dev_dbg(&ntb->dev, "Doorbell is unsafe\n");
+		if (!unsafe)
+			return -EINVAL;
+	}
+
+	if (ntb_spad_is_unsafe(ntb)) {
+		dev_dbg(&ntb->dev, "Scratchpad is unsafe\n");
+		if (!unsafe)
+			return -EINVAL;
+	}
+
+	pmask = GENMASK_ULL(ntb_peer_port_count(ntb), 0);
+	if ((ntb_db_valid_mask(ntb) & pmask) != pmask) {
+		dev_err(&ntb->dev, "Unsupported DB configuration\n");
+		return -EINVAL;
+	}
+
+	if (ntb_spad_count(ntb) < 1 && ntb_msg_count(ntb) < 1) {
+		dev_err(&ntb->dev, "Scratchpads and messages unsupported\n");
+		return -EINVAL;
+	} else if (ntb_spad_count(ntb) < 1) {
+		dev_dbg(&ntb->dev, "Scratchpads unsupported\n");
+	} else if (ntb_msg_count(ntb) < 1) {
+		dev_dbg(&ntb->dev, "Messages unsupported\n");
+	}
+
+	return 0;
+}
+
+static struct pp_ctx *pp_create_data(struct ntb_dev *ntb)
+{
+	struct pp_ctx *pp;
+
+	pp = devm_kzalloc(&ntb->dev, sizeof(*pp), GFP_KERNEL);
+	if (!pp)
+		return ERR_PTR(-ENOMEM);
+
+	pp->ntb = ntb;
+	atomic_set(&pp->count, 0);
+	spin_lock_init(&pp->lock);
+	hrtimer_setup(&pp->timer, pp_timer_func, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
+
+	return pp;
+}
+
+static void pp_init_flds(struct pp_ctx *pp)
+{
+	int pidx, lport, pcnt;
+
+	/* Find global port index */
+	lport = ntb_port_number(pp->ntb);
+	pcnt = ntb_peer_port_count(pp->ntb);
+	for (pidx = 0; pidx < pcnt; pidx++) {
+		if (lport < ntb_peer_port_number(pp->ntb, pidx))
+			break;
+	}
+
+	pp->in_db = BIT_ULL(lport);
+	pp->pmask = GENMASK_ULL(pidx, 0) >> 1;
+	pp->nmask = GENMASK_ULL(pcnt - 1, pidx);
+
+	dev_dbg(&pp->ntb->dev, "Inbound db %#llx, prev %#llx, next %#llx\n",
+		pp->in_db, pp->pmask, pp->nmask);
+}
+
+static int pp_mask_events(struct pp_ctx *pp)
+{
+	u64 db_mask, msg_mask;
+	int ret;
+
+	db_mask = ntb_db_valid_mask(pp->ntb);
+	ret = ntb_db_set_mask(pp->ntb, db_mask);
+	if (ret)
+		return ret;
+
+	/* Skip message events masking if unsupported */
+	if (ntb_msg_count(pp->ntb) < 1)
+		return 0;
+
+	msg_mask = ntb_msg_outbits(pp->ntb) | ntb_msg_inbits(pp->ntb);
+	return ntb_msg_set_mask(pp->ntb, msg_mask);
+}
+
+static int pp_setup_ctx(struct pp_ctx *pp)
+{
+	int ret;
+
+	ret = ntb_set_ctx(pp->ntb, pp, &pp_ops);
+	if (ret)
+		return ret;
+
+	ntb_link_enable(pp->ntb, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
+	/* Might be not necessary */
+	ntb_link_event(pp->ntb);
+
+	return 0;
+}
+
+static void pp_clear_ctx(struct pp_ctx *pp)
+{
+	ntb_link_disable(pp->ntb);
+
+	ntb_clear_ctx(pp->ntb);
+}
+
+static void pp_setup_dbgfs(struct pp_ctx *pp)
+{
+	struct pci_dev *pdev = pp->ntb->pdev;
+
+	pp->dbgfs_dir = debugfs_create_dir(pci_name(pdev), pp_dbgfs_topdir);
+
+	debugfs_create_atomic_t("count", 0600, pp->dbgfs_dir, &pp->count);
+}
+
+static void pp_clear_dbgfs(struct pp_ctx *pp)
+{
+	debugfs_remove_recursive(pp->dbgfs_dir);
+}
+
+static int pp_probe(struct ntb_client *client, struct ntb_dev *ntb)
+{
+	struct pp_ctx *pp;
+	int ret;
+
+	ret = pp_check_ntb(ntb);
+	if (ret)
+		return ret;
+
+	pp = pp_create_data(ntb);
+	if (IS_ERR(pp))
+		return PTR_ERR(pp);
+
+	pp_init_flds(pp);
+
+	ret = pp_mask_events(pp);
+	if (ret)
+		return ret;
+
+	ret = pp_setup_ctx(pp);
+	if (ret)
+		return ret;
+
+	pp_setup_dbgfs(pp);
+
+	return 0;
+}
+
+static void pp_remove(struct ntb_client *client, struct ntb_dev *ntb)
+{
+	struct pp_ctx *pp = ntb->ctx;
+
+	pp_clear_dbgfs(pp);
+
+	pp_clear_ctx(pp);
+
+	pp_clear(pp);
+}
+
+static struct ntb_client pp_client = {
+	.ops = {
+		.probe = pp_probe,
+		.remove = pp_remove
+	}
+};
+
+static int __init pp_init(void)
+{
+	int ret;
+
+	if (debugfs_initialized())
+		pp_dbgfs_topdir = debugfs_create_dir(KBUILD_MODNAME, NULL);
+
+	ret = ntb_register_client(&pp_client);
+	if (ret)
+		debugfs_remove_recursive(pp_dbgfs_topdir);
+
+	return ret;
+}
+module_init(pp_init);
+
+static void __exit pp_exit(void)
+{
+	ntb_unregister_client(&pp_client);
+	debugfs_remove_recursive(pp_dbgfs_topdir);
+}
+module_exit(pp_exit);
diff --git a/drivers/ntb/test/ntb_tool.c b/drivers/ntb/test/ntb_tool.c
new file mode 100644
index 000000000..06881047f
--- /dev/null
+++ b/drivers/ntb/test/ntb_tool.c
@@ -0,0 +1,1687 @@
+/*
+ * 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) 2015 EMC Corporation. All Rights Reserved.
+ *   Copyright (C) 2017 T-Platforms All Rights Reserved.
+ *
+ *   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) 2015 EMC Corporation. All Rights Reserved.
+ *   Copyright (C) 2017 T-Platforms All Rights Reserved.
+ *
+ *   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 copy
+ *       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.
+ *
+ * PCIe NTB Debugging Tool Linux driver
+ */
+
+/*
+ * How to use this tool, by example.
+ *
+ * Assuming $DBG_DIR is something like:
+ * '/sys/kernel/debug/ntb_tool/0000:00:03.0'
+ * Suppose aside from local device there is at least one remote device
+ * connected to NTB with index 0.
+ *-----------------------------------------------------------------------------
+ * Eg: check local/peer device information.
+ *
+ * # Get local device port number
+ * root@self# cat $DBG_DIR/port
+ *
+ * # Check local device functionality
+ * root@self# ls $DBG_DIR
+ * db            msg1          msg_sts     peer4/        port
+ * db_event      msg2          peer0/      peer5/        spad0
+ * db_mask       msg3          peer1/      peer_db       spad1
+ * link          msg_event     peer2/      peer_db_mask  spad2
+ * msg0          msg_mask      peer3/      peer_spad     spad3
+ * # As one can see it supports:
+ * # 1) four inbound message registers
+ * # 2) four inbound scratchpads
+ * # 3) up to six peer devices
+ *
+ * # Check peer device port number
+ * root@self# cat $DBG_DIR/peer0/port
+ *
+ * # Check peer device(s) functionality to be used
+ * root@self# ls $DBG_DIR/peer0
+ * link             mw_trans0       mw_trans6        port
+ * link_event       mw_trans1       mw_trans7        spad0
+ * msg0             mw_trans2       peer_mw_trans0   spad1
+ * msg1             mw_trans3       peer_mw_trans1   spad2
+ * msg2             mw_trans4       peer_mw_trans2   spad3
+ * msg3             mw_trans5       peer_mw_trans3
+ * # As one can see we got:
+ * # 1) four outbound message registers
+ * # 2) four outbound scratchpads
+ * # 3) eight inbound memory windows
+ * # 4) four outbound memory windows
+ *-----------------------------------------------------------------------------
+ * Eg: NTB link tests
+ *
+ * # Set local link up/down
+ * root@self# echo Y > $DBG_DIR/link
+ * root@self# echo N > $DBG_DIR/link
+ *
+ * # Check if link with peer device is up/down:
+ * root@self# cat $DBG_DIR/peer0/link
+ *
+ * # Block until the link is up/down
+ * root@self# echo Y > $DBG_DIR/peer0/link_event
+ * root@self# echo N > $DBG_DIR/peer0/link_event
+ *-----------------------------------------------------------------------------
+ * Eg: Doorbell registers tests (some functionality might be absent)
+ *
+ * # Set/clear/get local doorbell
+ * root@self# echo 's 1' > $DBG_DIR/db
+ * root@self# echo 'c 1' > $DBG_DIR/db
+ * root@self# cat  $DBG_DIR/db
+ *
+ * # Set/clear/get local doorbell mask
+ * root@self# echo 's 1' > $DBG_DIR/db_mask
+ * root@self# echo 'c 1' > $DBG_DIR/db_mask
+ * root@self# cat $DBG_DIR/db_mask
+ *
+ * # Ring/clear/get peer doorbell
+ * root@peer# echo 's 1' > $DBG_DIR/peer_db
+ * root@peer# echo 'c 1' > $DBG_DIR/peer_db
+ * root@peer# cat $DBG_DIR/peer_db
+ *
+ * # Set/clear/get peer doorbell mask
+ * root@self# echo 's 1' > $DBG_DIR/peer_db_mask
+ * root@self# echo 'c 1' > $DBG_DIR/peer_db_mask
+ * root@self# cat $DBG_DIR/peer_db_mask
+ *
+ * # Block until local doorbell is set with specified value
+ * root@self# echo 1 > $DBG_DIR/db_event
+ *-----------------------------------------------------------------------------
+ * Eg: Message registers tests (functionality might be absent)
+ *
+ * # Set/clear/get in/out message registers status
+ * root@self# echo 's 1' > $DBG_DIR/msg_sts
+ * root@self# echo 'c 1' > $DBG_DIR/msg_sts
+ * root@self# cat $DBG_DIR/msg_sts
+ *
+ * # Set/clear in/out message registers mask
+ * root@self# echo 's 1' > $DBG_DIR/msg_mask
+ * root@self# echo 'c 1' > $DBG_DIR/msg_mask
+ *
+ * # Get inbound message register #0 value and source of port index
+ * root@self# cat  $DBG_DIR/msg0
+ *
+ * # Send some data to peer over outbound message register #0
+ * root@self# echo 0x01020304 > $DBG_DIR/peer0/msg0
+ *-----------------------------------------------------------------------------
+ * Eg: Scratchpad registers tests (functionality might be absent)
+ *
+ * # Write/read to/from local scratchpad register #0
+ * root@peer# echo 0x01020304 > $DBG_DIR/spad0
+ * root@peer# cat $DBG_DIR/spad0
+ *
+ * # Write/read to/from peer scratchpad register #0
+ * root@peer# echo 0x01020304 > $DBG_DIR/peer0/spad0
+ * root@peer# cat $DBG_DIR/peer0/spad0
+ *-----------------------------------------------------------------------------
+ * Eg: Memory windows tests
+ *
+ * # Create inbound memory window buffer of specified size/get its base address
+ * root@peer# echo 16384 > $DBG_DIR/peer0/mw_trans0
+ * root@peer# cat $DBG_DIR/peer0/mw_trans0
+ *
+ * # Write/read data to/from inbound memory window
+ * root@peer# echo Hello > $DBG_DIR/peer0/mw0
+ * root@peer# head -c 7 $DBG_DIR/peer0/mw0
+ *
+ * # Map outbound memory window/check it settings (on peer device)
+ * root@peer# echo 0xADD0BA5E:16384 > $DBG_DIR/peer0/peer_mw_trans0
+ * root@peer# cat $DBG_DIR/peer0/peer_mw_trans0
+ *
+ * # Write/read data to/from outbound memory window (on peer device)
+ * root@peer# echo olleH > $DBG_DIR/peer0/peer_mw0
+ * root@peer# head -c 7 $DBG_DIR/peer0/peer_mw0
+ */
+
+#include <linux/init.h>
+#include <linux/kernel.h>
+#include <linux/module.h>
+
+#include <linux/debugfs.h>
+#include <linux/dma-mapping.h>
+#include <linux/pci.h>
+#include <linux/slab.h>
+#include <linux/uaccess.h>
+
+#include <linux/ntb.h>
+
+#define DRIVER_NAME		"ntb_tool"
+#define DRIVER_VERSION		"2.0"
+
+MODULE_LICENSE("Dual BSD/GPL");
+MODULE_VERSION(DRIVER_VERSION);
+MODULE_AUTHOR("Allen Hubbe <Allen.Hubbe@emc.com>");
+MODULE_DESCRIPTION("PCIe NTB Debugging Tool");
+
+/*
+ * Inbound and outbound memory windows descriptor. Union members selection
+ * depends on the MW type the structure describes. mm_base/dma_base are the
+ * virtual and DMA address of an inbound MW. io_base/tr_base are the MMIO
+ * mapped virtual and xlat addresses of an outbound MW respectively.
+ */
+struct tool_mw {
+	int widx;
+	int pidx;
+	struct tool_ctx *tc;
+	union {
+		u8 *mm_base;
+		u8 __iomem *io_base;
+	};
+	union {
+		dma_addr_t dma_base;
+		u64 tr_base;
+	};
+	resource_size_t size;
+	struct dentry *dbgfs_file;
+};
+
+/*
+ * Wrapper structure is used to distinguish the outbound MW peers reference
+ * within the corresponding DebugFS directory IO operation.
+ */
+struct tool_mw_wrap {
+	int pidx;
+	struct tool_mw *mw;
+};
+
+struct tool_msg {
+	int midx;
+	int pidx;
+	struct tool_ctx *tc;
+};
+
+struct tool_spad {
+	int sidx;
+	int pidx;
+	struct tool_ctx *tc;
+};
+
+struct tool_peer {
+	int pidx;
+	struct tool_ctx *tc;
+	int inmw_cnt;
+	struct tool_mw *inmws;
+	int outmw_cnt;
+	struct tool_mw_wrap *outmws;
+	int outmsg_cnt;
+	struct tool_msg *outmsgs;
+	int outspad_cnt;
+	struct tool_spad *outspads;
+	struct dentry *dbgfs_dir;
+};
+
+struct tool_ctx {
+	struct ntb_dev *ntb;
+	wait_queue_head_t link_wq;
+	wait_queue_head_t db_wq;
+	wait_queue_head_t msg_wq;
+	int outmw_cnt;
+	struct tool_mw *outmws;
+	int peer_cnt;
+	struct tool_peer *peers;
+	int inmsg_cnt;
+	struct tool_msg *inmsgs;
+	int inspad_cnt;
+	struct tool_spad *inspads;
+	struct dentry *dbgfs_dir;
+};
+
+#define TOOL_FOPS_RDWR(__name, __read, __write) \
+	const struct file_operations __name = {	\
+		.owner = THIS_MODULE,		\
+		.open = simple_open,		\
+		.read = __read,			\
+		.write = __write,		\
+	}
+
+#define TOOL_BUF_LEN 32
+
+static struct dentry *tool_dbgfs_topdir;
+
+/*==============================================================================
+ *                               NTB events handlers
+ *==============================================================================
+ */
+
+static void tool_link_event(void *ctx)
+{
+	struct tool_ctx *tc = ctx;
+	enum ntb_speed speed;
+	enum ntb_width width;
+	int up;
+
+	up = ntb_link_is_up(tc->ntb, &speed, &width);
+
+	dev_dbg(&tc->ntb->dev, "link is %s speed %d width %d\n",
+		up ? "up" : "down", speed, width);
+
+	wake_up(&tc->link_wq);
+}
+
+static void tool_db_event(void *ctx, int vec)
+{
+	struct tool_ctx *tc = ctx;
+	u64 db_bits, db_mask;
+
+	db_mask = ntb_db_vector_mask(tc->ntb, vec);
+	db_bits = ntb_db_read(tc->ntb);
+
+	dev_dbg(&tc->ntb->dev, "doorbell vec %d mask %#llx bits %#llx\n",
+		vec, db_mask, db_bits);
+
+	wake_up(&tc->db_wq);
+}
+
+static void tool_msg_event(void *ctx)
+{
+	struct tool_ctx *tc = ctx;
+	u64 msg_sts;
+
+	msg_sts = ntb_msg_read_sts(tc->ntb);
+
+	dev_dbg(&tc->ntb->dev, "message bits %#llx\n", msg_sts);
+
+	wake_up(&tc->msg_wq);
+}
+
+static const struct ntb_ctx_ops tool_ops = {
+	.link_event = tool_link_event,
+	.db_event = tool_db_event,
+	.msg_event = tool_msg_event
+};
+
+/*==============================================================================
+ *                        Common read/write methods
+ *==============================================================================
+ */
+
+static ssize_t tool_fn_read(struct tool_ctx *tc, char __user *ubuf,
+			    size_t size, loff_t *offp,
+			    u64 (*fn_read)(struct ntb_dev *))
+{
+	size_t buf_size;
+	char buf[TOOL_BUF_LEN];
+	ssize_t pos;
+
+	if (!fn_read)
+		return -EINVAL;
+
+	buf_size = min(size, sizeof(buf));
+
+	pos = scnprintf(buf, buf_size, "%#llx\n", fn_read(tc->ntb));
+
+	return simple_read_from_buffer(ubuf, size, offp, buf, pos);
+}
+
+static ssize_t tool_fn_write(struct tool_ctx *tc,
+			     const char __user *ubuf,
+			     size_t size, loff_t *offp,
+			     int (*fn_set)(struct ntb_dev *, u64),
+			     int (*fn_clear)(struct ntb_dev *, u64))
+{
+	char *buf, cmd;
+	ssize_t ret;
+	u64 bits;
+	int n;
+
+	if (*offp)
+		return 0;
+
+	buf = memdup_user_nul(ubuf, size);
+	if (IS_ERR(buf))
+		return PTR_ERR(buf);
+
+	n = sscanf(buf, "%c %lli", &cmd, &bits);
+
+	kfree(buf);
+
+	if (n != 2) {
+		ret = -EINVAL;
+	} else if (cmd == 's') {
+		if (!fn_set)
+			ret = -EINVAL;
+		else
+			ret = fn_set(tc->ntb, bits);
+	} else if (cmd == 'c') {
+		if (!fn_clear)
+			ret = -EINVAL;
+		else
+			ret = fn_clear(tc->ntb, bits);
+	} else {
+		ret = -EINVAL;
+	}
+
+	return ret ? : size;
+}
+
+/*==============================================================================
+ *                            Port read/write methods
+ *==============================================================================
+ */
+
+static ssize_t tool_port_read(struct file *filep, char __user *ubuf,
+			      size_t size, loff_t *offp)
+{
+	struct tool_ctx *tc = filep->private_data;
+	char buf[TOOL_BUF_LEN];
+	int pos;
+
+	pos = scnprintf(buf, sizeof(buf), "%d\n", ntb_port_number(tc->ntb));
+
+	return simple_read_from_buffer(ubuf, size, offp, buf, pos);
+}
+
+static TOOL_FOPS_RDWR(tool_port_fops,
+		      tool_port_read,
+		      NULL);
+
+static ssize_t tool_peer_port_read(struct file *filep, char __user *ubuf,
+				   size_t size, loff_t *offp)
+{
+	struct tool_peer *peer = filep->private_data;
+	struct tool_ctx *tc = peer->tc;
+	char buf[TOOL_BUF_LEN];
+	int pos;
+
+	pos = scnprintf(buf, sizeof(buf), "%d\n",
+		ntb_peer_port_number(tc->ntb, peer->pidx));
+
+	return simple_read_from_buffer(ubuf, size, offp, buf, pos);
+}
+
+static TOOL_FOPS_RDWR(tool_peer_port_fops,
+		      tool_peer_port_read,
+		      NULL);
+
+static int tool_init_peers(struct tool_ctx *tc)
+{
+	int pidx;
+
+	tc->peer_cnt = ntb_peer_port_count(tc->ntb);
+	tc->peers = devm_kcalloc(&tc->ntb->dev, tc->peer_cnt,
+				 sizeof(*tc->peers), GFP_KERNEL);
+	if (tc->peers == NULL)
+		return -ENOMEM;
+
+	for (pidx = 0; pidx < tc->peer_cnt; pidx++) {
+		tc->peers[pidx].pidx = pidx;
+		tc->peers[pidx].tc = tc;
+	}
+
+	return 0;
+}
+
+/*==============================================================================
+ *                       Link state read/write methods
+ *==============================================================================
+ */
+
+static ssize_t tool_link_write(struct file *filep, const char __user *ubuf,
+			       size_t size, loff_t *offp)
+{
+	struct tool_ctx *tc = filep->private_data;
+	bool val;
+	int ret;
+
+	ret = kstrtobool_from_user(ubuf, size, &val);
+	if (ret)
+		return ret;
+
+	if (val)
+		ret = ntb_link_enable(tc->ntb, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
+	else
+		ret = ntb_link_disable(tc->ntb);
+
+	if (ret)
+		return ret;
+
+	return size;
+}
+
+static TOOL_FOPS_RDWR(tool_link_fops,
+		      NULL,
+		      tool_link_write);
+
+static ssize_t tool_peer_link_read(struct file *filep, char __user *ubuf,
+				   size_t size, loff_t *offp)
+{
+	struct tool_peer *peer = filep->private_data;
+	struct tool_ctx *tc = peer->tc;
+	char buf[3];
+
+	if (ntb_link_is_up(tc->ntb, NULL, NULL) & BIT(peer->pidx))
+		buf[0] = 'Y';
+	else
+		buf[0] = 'N';
+	buf[1] = '\n';
+	buf[2] = '\0';
+
+	return simple_read_from_buffer(ubuf, size, offp, buf, 2);
+}
+
+static TOOL_FOPS_RDWR(tool_peer_link_fops,
+		      tool_peer_link_read,
+		      NULL);
+
+static ssize_t tool_peer_link_event_write(struct file *filep,
+					  const char __user *ubuf,
+					  size_t size, loff_t *offp)
+{
+	struct tool_peer *peer = filep->private_data;
+	struct tool_ctx *tc = peer->tc;
+	u64 link_msk;
+	bool val;
+	int ret;
+
+	ret = kstrtobool_from_user(ubuf, size, &val);
+	if (ret)
+		return ret;
+
+	link_msk = BIT_ULL_MASK(peer->pidx);
+
+	if (wait_event_interruptible(tc->link_wq,
+		!!(ntb_link_is_up(tc->ntb, NULL, NULL) & link_msk) == val))
+		return -ERESTART;
+
+	return size;
+}
+
+static TOOL_FOPS_RDWR(tool_peer_link_event_fops,
+		      NULL,
+		      tool_peer_link_event_write);
+
+/*==============================================================================
+ *                  Memory windows read/write/setting methods
+ *==============================================================================
+ */
+
+static ssize_t tool_mw_read(struct file *filep, char __user *ubuf,
+			    size_t size, loff_t *offp)
+{
+	struct tool_mw *inmw = filep->private_data;
+
+	if (inmw->mm_base == NULL)
+		return -ENXIO;
+
+	return simple_read_from_buffer(ubuf, size, offp,
+				       inmw->mm_base, inmw->size);
+}
+
+static ssize_t tool_mw_write(struct file *filep, const char __user *ubuf,
+			     size_t size, loff_t *offp)
+{
+	struct tool_mw *inmw = filep->private_data;
+
+	if (inmw->mm_base == NULL)
+		return -ENXIO;
+
+	return simple_write_to_buffer(inmw->mm_base, inmw->size, offp,
+				      ubuf, size);
+}
+
+static TOOL_FOPS_RDWR(tool_mw_fops,
+		      tool_mw_read,
+		      tool_mw_write);
+
+static int tool_setup_mw(struct tool_ctx *tc, int pidx, int widx,
+			 size_t req_size)
+{
+	resource_size_t size, addr_align, size_align;
+	struct tool_mw *inmw = &tc->peers[pidx].inmws[widx];
+	char buf[TOOL_BUF_LEN];
+	int ret;
+
+	if (inmw->mm_base != NULL)
+		return 0;
+
+	ret = ntb_mw_get_align(tc->ntb, pidx, widx, &addr_align,
+				&size_align, &size);
+	if (ret)
+		return ret;
+
+	inmw->size = min_t(resource_size_t, req_size, size);
+	inmw->size = round_up(inmw->size, addr_align);
+	inmw->size = round_up(inmw->size, size_align);
+	inmw->mm_base = dma_alloc_coherent(&tc->ntb->pdev->dev, inmw->size,
+					   &inmw->dma_base, GFP_KERNEL);
+	if (!inmw->mm_base)
+		return -ENOMEM;
+
+	if (!IS_ALIGNED(inmw->dma_base, addr_align)) {
+		ret = -ENOMEM;
+		goto err_free_dma;
+	}
+
+	ret = ntb_mw_set_trans(tc->ntb, pidx, widx, inmw->dma_base, inmw->size);
+	if (ret)
+		goto err_free_dma;
+
+	snprintf(buf, sizeof(buf), "mw%d", widx);
+	inmw->dbgfs_file = debugfs_create_file(buf, 0600,
+					       tc->peers[pidx].dbgfs_dir, inmw,
+					       &tool_mw_fops);
+
+	return 0;
+
+err_free_dma:
+	dma_free_coherent(&tc->ntb->pdev->dev, inmw->size, inmw->mm_base,
+			  inmw->dma_base);
+	inmw->mm_base = NULL;
+	inmw->dma_base = 0;
+	inmw->size = 0;
+
+	return ret;
+}
+
+static void tool_free_mw(struct tool_ctx *tc, int pidx, int widx)
+{
+	struct tool_mw *inmw = &tc->peers[pidx].inmws[widx];
+
+	debugfs_remove(inmw->dbgfs_file);
+
+	if (inmw->mm_base != NULL) {
+		ntb_mw_clear_trans(tc->ntb, pidx, widx);
+		dma_free_coherent(&tc->ntb->pdev->dev, inmw->size,
+				  inmw->mm_base, inmw->dma_base);
+	}
+
+	inmw->mm_base = NULL;
+	inmw->dma_base = 0;
+	inmw->size = 0;
+	inmw->dbgfs_file = NULL;
+}
+
+static ssize_t tool_mw_trans_read(struct file *filep, char __user *ubuf,
+				  size_t size, loff_t *offp)
+{
+	struct tool_mw *inmw = filep->private_data;
+	resource_size_t addr_align;
+	resource_size_t size_align;
+	resource_size_t size_max;
+	ssize_t ret, off = 0;
+	size_t buf_size;
+	char *buf;
+
+	buf_size = min_t(size_t, size, 512);
+
+	buf = kmalloc(buf_size, GFP_KERNEL);
+	if (!buf)
+		return -ENOMEM;
+
+	ret = ntb_mw_get_align(inmw->tc->ntb, inmw->pidx, inmw->widx,
+			       &addr_align, &size_align, &size_max);
+	if (ret)
+		goto err;
+
+	off += scnprintf(buf + off, buf_size - off,
+			 "Inbound MW     \t%d\n",
+			 inmw->widx);
+
+	off += scnprintf(buf + off, buf_size - off,
+			 "Port           \t%d (%d)\n",
+			 ntb_peer_port_number(inmw->tc->ntb, inmw->pidx),
+			 inmw->pidx);
+
+	off += scnprintf(buf + off, buf_size - off,
+			 "Window Address \t0x%pK\n", inmw->mm_base);
+
+	off += scnprintf(buf + off, buf_size - off,
+			 "DMA Address    \t%pad\n",
+			 &inmw->dma_base);
+
+	off += scnprintf(buf + off, buf_size - off,
+			 "Window Size    \t%pap\n",
+			 &inmw->size);
+
+	off += scnprintf(buf + off, buf_size - off,
+			 "Alignment      \t%pap\n",
+			 &addr_align);
+
+	off += scnprintf(buf + off, buf_size - off,
+			 "Size Alignment \t%pap\n",
+			 &size_align);
+
+	off += scnprintf(buf + off, buf_size - off,
+			 "Size Max       \t%pap\n",
+			 &size_max);
+
+	ret = simple_read_from_buffer(ubuf, size, offp, buf, off);
+
+err:
+	kfree(buf);
+
+	return ret;
+}
+
+static ssize_t tool_mw_trans_write(struct file *filep, const char __user *ubuf,
+				   size_t size, loff_t *offp)
+{
+	struct tool_mw *inmw = filep->private_data;
+	unsigned int val;
+	int ret;
+
+	ret = kstrtouint_from_user(ubuf, size, 0, &val);
+	if (ret)
+		return ret;
+
+	tool_free_mw(inmw->tc, inmw->pidx, inmw->widx);
+	if (val) {
+		ret = tool_setup_mw(inmw->tc, inmw->pidx, inmw->widx, val);
+		if (ret)
+			return ret;
+	}
+
+	return size;
+}
+
+static TOOL_FOPS_RDWR(tool_mw_trans_fops,
+		      tool_mw_trans_read,
+		      tool_mw_trans_write);
+
+static ssize_t tool_peer_mw_read(struct file *filep, char __user *ubuf,
+				 size_t size, loff_t *offp)
+{
+	struct tool_mw *outmw = filep->private_data;
+	loff_t pos = *offp;
+	ssize_t ret;
+	void *buf;
+
+	if (outmw->io_base == NULL)
+		return -EIO;
+
+	if (pos >= outmw->size || !size)
+		return 0;
+
+	if (size > outmw->size - pos)
+		size = outmw->size - pos;
+
+	buf = kmalloc(size, GFP_KERNEL);
+	if (!buf)
+		return -ENOMEM;
+
+	memcpy_fromio(buf, outmw->io_base + pos, size);
+	ret = copy_to_user(ubuf, buf, size);
+	if (ret == size) {
+		ret = -EFAULT;
+		goto err_free;
+	}
+
+	size -= ret;
+	*offp = pos + size;
+	ret = size;
+
+err_free:
+	kfree(buf);
+
+	return ret;
+}
+
+static ssize_t tool_peer_mw_write(struct file *filep, const char __user *ubuf,
+				  size_t size, loff_t *offp)
+{
+	struct tool_mw *outmw = filep->private_data;
+	ssize_t ret;
+	loff_t pos = *offp;
+	void *buf;
+
+	if (outmw->io_base == NULL)
+		return -EIO;
+
+	if (pos >= outmw->size || !size)
+		return 0;
+	if (size > outmw->size - pos)
+		size = outmw->size - pos;
+
+	buf = kmalloc(size, GFP_KERNEL);
+	if (!buf)
+		return -ENOMEM;
+
+	ret = copy_from_user(buf, ubuf, size);
+	if (ret == size) {
+		ret = -EFAULT;
+		goto err_free;
+	}
+
+	size -= ret;
+	*offp = pos + size;
+	ret = size;
+
+	memcpy_toio(outmw->io_base + pos, buf, size);
+
+err_free:
+	kfree(buf);
+
+	return ret;
+}
+
+static TOOL_FOPS_RDWR(tool_peer_mw_fops,
+		      tool_peer_mw_read,
+		      tool_peer_mw_write);
+
+static int tool_setup_peer_mw(struct tool_ctx *tc, int pidx, int widx,
+			      u64 req_addr, size_t req_size)
+{
+	struct tool_mw *outmw = &tc->outmws[widx];
+	resource_size_t map_size;
+	phys_addr_t map_base;
+	char buf[TOOL_BUF_LEN];
+	int ret;
+
+	if (outmw->io_base != NULL)
+		return 0;
+
+	ret = ntb_peer_mw_get_addr(tc->ntb, widx, &map_base, &map_size);
+	if (ret)
+		return ret;
+
+	ret = ntb_peer_mw_set_trans(tc->ntb, pidx, widx, req_addr, req_size);
+	if (ret)
+		return ret;
+
+	outmw->io_base = ioremap_wc(map_base, map_size);
+	if (outmw->io_base == NULL) {
+		ret = -EFAULT;
+		goto err_clear_trans;
+	}
+
+	outmw->tr_base = req_addr;
+	outmw->size = req_size;
+	outmw->pidx = pidx;
+
+	snprintf(buf, sizeof(buf), "peer_mw%d", widx);
+	outmw->dbgfs_file = debugfs_create_file(buf, 0600,
+					       tc->peers[pidx].dbgfs_dir, outmw,
+					       &tool_peer_mw_fops);
+
+	return 0;
+
+err_clear_trans:
+	ntb_peer_mw_clear_trans(tc->ntb, pidx, widx);
+
+	return ret;
+}
+
+static void tool_free_peer_mw(struct tool_ctx *tc, int widx)
+{
+	struct tool_mw *outmw = &tc->outmws[widx];
+
+	debugfs_remove(outmw->dbgfs_file);
+
+	if (outmw->io_base != NULL) {
+		iounmap(tc->outmws[widx].io_base);
+		ntb_peer_mw_clear_trans(tc->ntb, outmw->pidx, widx);
+	}
+
+	outmw->io_base = NULL;
+	outmw->tr_base = 0;
+	outmw->size = 0;
+	outmw->pidx = -1;
+	outmw->dbgfs_file = NULL;
+}
+
+static ssize_t tool_peer_mw_trans_read(struct file *filep, char __user *ubuf,
+					size_t size, loff_t *offp)
+{
+	struct tool_mw_wrap *outmw_wrap = filep->private_data;
+	struct tool_mw *outmw = outmw_wrap->mw;
+	resource_size_t map_size;
+	phys_addr_t map_base;
+	ssize_t off = 0;
+	size_t buf_size;
+	char *buf;
+	int ret;
+
+	ret = ntb_peer_mw_get_addr(outmw->tc->ntb, outmw->widx,
+				  &map_base, &map_size);
+	if (ret)
+		return ret;
+
+	buf_size = min_t(size_t, size, 512);
+
+	buf = kmalloc(buf_size, GFP_KERNEL);
+	if (!buf)
+		return -ENOMEM;
+
+	off += scnprintf(buf + off, buf_size - off,
+			 "Outbound MW:        \t%d\n", outmw->widx);
+
+	if (outmw->io_base != NULL) {
+		off += scnprintf(buf + off, buf_size - off,
+			"Port attached       \t%d (%d)\n",
+			ntb_peer_port_number(outmw->tc->ntb, outmw->pidx),
+			outmw->pidx);
+	} else {
+		off += scnprintf(buf + off, buf_size - off,
+				 "Port attached       \t-1 (-1)\n");
+	}
+
+	off += scnprintf(buf + off, buf_size - off,
+			 "Virtual address     \t0x%pK\n", outmw->io_base);
+
+	off += scnprintf(buf + off, buf_size - off,
+			 "Phys Address        \t%pap\n", &map_base);
+
+	off += scnprintf(buf + off, buf_size - off,
+			 "Mapping Size        \t%pap\n", &map_size);
+
+	off += scnprintf(buf + off, buf_size - off,
+			 "Translation Address \t0x%016llx\n", outmw->tr_base);
+
+	off += scnprintf(buf + off, buf_size - off,
+			 "Window Size         \t%pap\n", &outmw->size);
+
+	ret = simple_read_from_buffer(ubuf, size, offp, buf, off);
+	kfree(buf);
+
+	return ret;
+}
+
+static ssize_t tool_peer_mw_trans_write(struct file *filep,
+					const char __user *ubuf,
+					size_t size, loff_t *offp)
+{
+	struct tool_mw_wrap *outmw_wrap = filep->private_data;
+	struct tool_mw *outmw = outmw_wrap->mw;
+	size_t buf_size, wsize;
+	char buf[TOOL_BUF_LEN];
+	int ret, n;
+	u64 addr;
+
+	buf_size = min(size, (sizeof(buf) - 1));
+	if (copy_from_user(buf, ubuf, buf_size))
+		return -EFAULT;
+
+	buf[buf_size] = '\0';
+
+	n = sscanf(buf, "%llu:%zu", &addr, &wsize);
+	if (n != 2)
+		return -EINVAL;
+
+	tool_free_peer_mw(outmw->tc, outmw->widx);
+	if (wsize) {
+		ret = tool_setup_peer_mw(outmw->tc, outmw_wrap->pidx,
+					 outmw->widx, addr, wsize);
+		if (ret)
+			return ret;
+	}
+
+	return size;
+}
+
+static TOOL_FOPS_RDWR(tool_peer_mw_trans_fops,
+		      tool_peer_mw_trans_read,
+		      tool_peer_mw_trans_write);
+
+static int tool_init_mws(struct tool_ctx *tc)
+{
+	int widx, pidx;
+
+	/* Initialize outbound memory windows */
+	tc->outmw_cnt = ntb_peer_mw_count(tc->ntb);
+	tc->outmws = devm_kcalloc(&tc->ntb->dev, tc->outmw_cnt,
+				  sizeof(*tc->outmws), GFP_KERNEL);
+	if (tc->outmws == NULL)
+		return -ENOMEM;
+
+	for (widx = 0; widx < tc->outmw_cnt; widx++) {
+		tc->outmws[widx].widx = widx;
+		tc->outmws[widx].pidx = -1;
+		tc->outmws[widx].tc = tc;
+	}
+
+	/* Initialize inbound memory windows and outbound MWs wrapper */
+	for (pidx = 0; pidx < tc->peer_cnt; pidx++) {
+		tc->peers[pidx].inmw_cnt = ntb_mw_count(tc->ntb, pidx);
+		tc->peers[pidx].inmws =
+			devm_kcalloc(&tc->ntb->dev, tc->peers[pidx].inmw_cnt,
+				    sizeof(*tc->peers[pidx].inmws), GFP_KERNEL);
+		if (tc->peers[pidx].inmws == NULL)
+			return -ENOMEM;
+
+		for (widx = 0; widx < tc->peers[pidx].inmw_cnt; widx++) {
+			tc->peers[pidx].inmws[widx].widx = widx;
+			tc->peers[pidx].inmws[widx].pidx = pidx;
+			tc->peers[pidx].inmws[widx].tc = tc;
+		}
+
+		tc->peers[pidx].outmw_cnt = ntb_peer_mw_count(tc->ntb);
+		tc->peers[pidx].outmws =
+			devm_kcalloc(&tc->ntb->dev, tc->peers[pidx].outmw_cnt,
+				   sizeof(*tc->peers[pidx].outmws), GFP_KERNEL);
+		if (tc->peers[pidx].outmws == NULL)
+			return -ENOMEM;
+
+		for (widx = 0; widx < tc->peers[pidx].outmw_cnt; widx++) {
+			tc->peers[pidx].outmws[widx].pidx = pidx;
+			tc->peers[pidx].outmws[widx].mw = &tc->outmws[widx];
+		}
+	}
+
+	return 0;
+}
+
+static void tool_clear_mws(struct tool_ctx *tc)
+{
+	int widx, pidx;
+
+	/* Free outbound memory windows */
+	for (widx = 0; widx < tc->outmw_cnt; widx++)
+		tool_free_peer_mw(tc, widx);
+
+	/* Free outbound memory windows */
+	for (pidx = 0; pidx < tc->peer_cnt; pidx++)
+		for (widx = 0; widx < tc->peers[pidx].inmw_cnt; widx++)
+			tool_free_mw(tc, pidx, widx);
+}
+
+/*==============================================================================
+ *                       Doorbell read/write methods
+ *==============================================================================
+ */
+
+static ssize_t tool_db_read(struct file *filep, char __user *ubuf,
+			    size_t size, loff_t *offp)
+{
+	struct tool_ctx *tc = filep->private_data;
+
+	return tool_fn_read(tc, ubuf, size, offp, tc->ntb->ops->db_read);
+}
+
+static ssize_t tool_db_write(struct file *filep, const char __user *ubuf,
+			     size_t size, loff_t *offp)
+{
+	struct tool_ctx *tc = filep->private_data;
+
+	return tool_fn_write(tc, ubuf, size, offp, tc->ntb->ops->db_set,
+			     tc->ntb->ops->db_clear);
+}
+
+static TOOL_FOPS_RDWR(tool_db_fops,
+		      tool_db_read,
+		      tool_db_write);
+
+static ssize_t tool_db_valid_mask_read(struct file *filep, char __user *ubuf,
+				       size_t size, loff_t *offp)
+{
+	struct tool_ctx *tc = filep->private_data;
+
+	return tool_fn_read(tc, ubuf, size, offp, tc->ntb->ops->db_valid_mask);
+}
+
+static TOOL_FOPS_RDWR(tool_db_valid_mask_fops,
+		      tool_db_valid_mask_read,
+		      NULL);
+
+static ssize_t tool_db_mask_read(struct file *filep, char __user *ubuf,
+				 size_t size, loff_t *offp)
+{
+	struct tool_ctx *tc = filep->private_data;
+
+	return tool_fn_read(tc, ubuf, size, offp, tc->ntb->ops->db_read_mask);
+}
+
+static ssize_t tool_db_mask_write(struct file *filep, const char __user *ubuf,
+			       size_t size, loff_t *offp)
+{
+	struct tool_ctx *tc = filep->private_data;
+
+	return tool_fn_write(tc, ubuf, size, offp, tc->ntb->ops->db_set_mask,
+			     tc->ntb->ops->db_clear_mask);
+}
+
+static TOOL_FOPS_RDWR(tool_db_mask_fops,
+		      tool_db_mask_read,
+		      tool_db_mask_write);
+
+static ssize_t tool_peer_db_read(struct file *filep, char __user *ubuf,
+				 size_t size, loff_t *offp)
+{
+	struct tool_ctx *tc = filep->private_data;
+
+	return tool_fn_read(tc, ubuf, size, offp, tc->ntb->ops->peer_db_read);
+}
+
+static ssize_t tool_peer_db_write(struct file *filep, const char __user *ubuf,
+				  size_t size, loff_t *offp)
+{
+	struct tool_ctx *tc = filep->private_data;
+
+	return tool_fn_write(tc, ubuf, size, offp, tc->ntb->ops->peer_db_set,
+			     tc->ntb->ops->peer_db_clear);
+}
+
+static TOOL_FOPS_RDWR(tool_peer_db_fops,
+		      tool_peer_db_read,
+		      tool_peer_db_write);
+
+static ssize_t tool_peer_db_mask_read(struct file *filep, char __user *ubuf,
+				   size_t size, loff_t *offp)
+{
+	struct tool_ctx *tc = filep->private_data;
+
+	return tool_fn_read(tc, ubuf, size, offp,
+			    tc->ntb->ops->peer_db_read_mask);
+}
+
+static ssize_t tool_peer_db_mask_write(struct file *filep,
+				       const char __user *ubuf,
+				       size_t size, loff_t *offp)
+{
+	struct tool_ctx *tc = filep->private_data;
+
+	return tool_fn_write(tc, ubuf, size, offp,
+			     tc->ntb->ops->peer_db_set_mask,
+			     tc->ntb->ops->peer_db_clear_mask);
+}
+
+static TOOL_FOPS_RDWR(tool_peer_db_mask_fops,
+		      tool_peer_db_mask_read,
+		      tool_peer_db_mask_write);
+
+static ssize_t tool_db_event_write(struct file *filep,
+				   const char __user *ubuf,
+				   size_t size, loff_t *offp)
+{
+	struct tool_ctx *tc = filep->private_data;
+	u64 val;
+	int ret;
+
+	ret = kstrtou64_from_user(ubuf, size, 0, &val);
+	if (ret)
+		return ret;
+
+	if (wait_event_interruptible(tc->db_wq, ntb_db_read(tc->ntb) == val))
+		return -ERESTART;
+
+	return size;
+}
+
+static TOOL_FOPS_RDWR(tool_db_event_fops,
+		      NULL,
+		      tool_db_event_write);
+
+/*==============================================================================
+ *                       Scratchpads read/write methods
+ *==============================================================================
+ */
+
+static ssize_t tool_spad_read(struct file *filep, char __user *ubuf,
+			      size_t size, loff_t *offp)
+{
+	struct tool_spad *spad = filep->private_data;
+	char buf[TOOL_BUF_LEN];
+	ssize_t pos;
+
+	if (!spad->tc->ntb->ops->spad_read)
+		return -EINVAL;
+
+	pos = scnprintf(buf, sizeof(buf), "%#x\n",
+		ntb_spad_read(spad->tc->ntb, spad->sidx));
+
+	return simple_read_from_buffer(ubuf, size, offp, buf, pos);
+}
+
+static ssize_t tool_spad_write(struct file *filep, const char __user *ubuf,
+			       size_t size, loff_t *offp)
+{
+	struct tool_spad *spad = filep->private_data;
+	u32 val;
+	int ret;
+
+	if (!spad->tc->ntb->ops->spad_write) {
+		dev_dbg(&spad->tc->ntb->dev, "no spad write fn\n");
+		return -EINVAL;
+	}
+
+	ret = kstrtou32_from_user(ubuf, size, 0, &val);
+	if (ret)
+		return ret;
+
+	ret = ntb_spad_write(spad->tc->ntb, spad->sidx, val);
+
+	return ret ?: size;
+}
+
+static TOOL_FOPS_RDWR(tool_spad_fops,
+		      tool_spad_read,
+		      tool_spad_write);
+
+static ssize_t tool_peer_spad_read(struct file *filep, char __user *ubuf,
+				   size_t size, loff_t *offp)
+{
+	struct tool_spad *spad = filep->private_data;
+	char buf[TOOL_BUF_LEN];
+	ssize_t pos;
+
+	if (!spad->tc->ntb->ops->peer_spad_read)
+		return -EINVAL;
+
+	pos = scnprintf(buf, sizeof(buf), "%#x\n",
+		ntb_peer_spad_read(spad->tc->ntb, spad->pidx, spad->sidx));
+
+	return simple_read_from_buffer(ubuf, size, offp, buf, pos);
+}
+
+static ssize_t tool_peer_spad_write(struct file *filep, const char __user *ubuf,
+				    size_t size, loff_t *offp)
+{
+	struct tool_spad *spad = filep->private_data;
+	u32 val;
+	int ret;
+
+	if (!spad->tc->ntb->ops->peer_spad_write) {
+		dev_dbg(&spad->tc->ntb->dev, "no spad write fn\n");
+		return -EINVAL;
+	}
+
+	ret = kstrtou32_from_user(ubuf, size, 0, &val);
+	if (ret)
+		return ret;
+
+	ret = ntb_peer_spad_write(spad->tc->ntb, spad->pidx, spad->sidx, val);
+
+	return ret ?: size;
+}
+
+static TOOL_FOPS_RDWR(tool_peer_spad_fops,
+		      tool_peer_spad_read,
+		      tool_peer_spad_write);
+
+static int tool_init_spads(struct tool_ctx *tc)
+{
+	int sidx, pidx;
+
+	/* Initialize inbound scratchpad structures */
+	tc->inspad_cnt = ntb_spad_count(tc->ntb);
+	tc->inspads = devm_kcalloc(&tc->ntb->dev, tc->inspad_cnt,
+				   sizeof(*tc->inspads), GFP_KERNEL);
+	if (tc->inspads == NULL)
+		return -ENOMEM;
+
+	for (sidx = 0; sidx < tc->inspad_cnt; sidx++) {
+		tc->inspads[sidx].sidx = sidx;
+		tc->inspads[sidx].pidx = -1;
+		tc->inspads[sidx].tc = tc;
+	}
+
+	/* Initialize outbound scratchpad structures */
+	for (pidx = 0; pidx < tc->peer_cnt; pidx++) {
+		tc->peers[pidx].outspad_cnt = ntb_spad_count(tc->ntb);
+		tc->peers[pidx].outspads =
+			devm_kcalloc(&tc->ntb->dev, tc->peers[pidx].outspad_cnt,
+				sizeof(*tc->peers[pidx].outspads), GFP_KERNEL);
+		if (tc->peers[pidx].outspads == NULL)
+			return -ENOMEM;
+
+		for (sidx = 0; sidx < tc->peers[pidx].outspad_cnt; sidx++) {
+			tc->peers[pidx].outspads[sidx].sidx = sidx;
+			tc->peers[pidx].outspads[sidx].pidx = pidx;
+			tc->peers[pidx].outspads[sidx].tc = tc;
+		}
+	}
+
+	return 0;
+}
+
+/*==============================================================================
+ *                       Messages read/write methods
+ *==============================================================================
+ */
+
+static ssize_t tool_inmsg_read(struct file *filep, char __user *ubuf,
+			       size_t size, loff_t *offp)
+{
+	struct tool_msg *msg = filep->private_data;
+	char buf[TOOL_BUF_LEN];
+	ssize_t pos;
+	u32 data;
+	int pidx;
+
+	data = ntb_msg_read(msg->tc->ntb, &pidx, msg->midx);
+
+	pos = scnprintf(buf, sizeof(buf), "0x%08x<-%d\n", data, pidx);
+
+	return simple_read_from_buffer(ubuf, size, offp, buf, pos);
+}
+
+static TOOL_FOPS_RDWR(tool_inmsg_fops,
+		      tool_inmsg_read,
+		      NULL);
+
+static ssize_t tool_outmsg_write(struct file *filep,
+				 const char __user *ubuf,
+				 size_t size, loff_t *offp)
+{
+	struct tool_msg *msg = filep->private_data;
+	u32 val;
+	int ret;
+
+	ret = kstrtou32_from_user(ubuf, size, 0, &val);
+	if (ret)
+		return ret;
+
+	ret = ntb_peer_msg_write(msg->tc->ntb, msg->pidx, msg->midx, val);
+
+	return ret ? : size;
+}
+
+static TOOL_FOPS_RDWR(tool_outmsg_fops,
+		      NULL,
+		      tool_outmsg_write);
+
+static ssize_t tool_msg_sts_read(struct file *filep, char __user *ubuf,
+				 size_t size, loff_t *offp)
+{
+	struct tool_ctx *tc = filep->private_data;
+
+	return tool_fn_read(tc, ubuf, size, offp, tc->ntb->ops->msg_read_sts);
+}
+
+static ssize_t tool_msg_sts_write(struct file *filep, const char __user *ubuf,
+				  size_t size, loff_t *offp)
+{
+	struct tool_ctx *tc = filep->private_data;
+
+	return tool_fn_write(tc, ubuf, size, offp, NULL,
+			     tc->ntb->ops->msg_clear_sts);
+}
+
+static TOOL_FOPS_RDWR(tool_msg_sts_fops,
+		      tool_msg_sts_read,
+		      tool_msg_sts_write);
+
+static ssize_t tool_msg_inbits_read(struct file *filep, char __user *ubuf,
+				    size_t size, loff_t *offp)
+{
+	struct tool_ctx *tc = filep->private_data;
+
+	return tool_fn_read(tc, ubuf, size, offp, tc->ntb->ops->msg_inbits);
+}
+
+static TOOL_FOPS_RDWR(tool_msg_inbits_fops,
+		      tool_msg_inbits_read,
+		      NULL);
+
+static ssize_t tool_msg_outbits_read(struct file *filep, char __user *ubuf,
+				     size_t size, loff_t *offp)
+{
+	struct tool_ctx *tc = filep->private_data;
+
+	return tool_fn_read(tc, ubuf, size, offp, tc->ntb->ops->msg_outbits);
+}
+
+static TOOL_FOPS_RDWR(tool_msg_outbits_fops,
+		      tool_msg_outbits_read,
+		      NULL);
+
+static ssize_t tool_msg_mask_write(struct file *filep, const char __user *ubuf,
+				   size_t size, loff_t *offp)
+{
+	struct tool_ctx *tc = filep->private_data;
+
+	return tool_fn_write(tc, ubuf, size, offp,
+			     tc->ntb->ops->msg_set_mask,
+			     tc->ntb->ops->msg_clear_mask);
+}
+
+static TOOL_FOPS_RDWR(tool_msg_mask_fops,
+		      NULL,
+		      tool_msg_mask_write);
+
+static ssize_t tool_msg_event_write(struct file *filep,
+				    const char __user *ubuf,
+				    size_t size, loff_t *offp)
+{
+	struct tool_ctx *tc = filep->private_data;
+	u64 val;
+	int ret;
+
+	ret = kstrtou64_from_user(ubuf, size, 0, &val);
+	if (ret)
+		return ret;
+
+	if (wait_event_interruptible(tc->msg_wq,
+		ntb_msg_read_sts(tc->ntb) == val))
+		return -ERESTART;
+
+	return size;
+}
+
+static TOOL_FOPS_RDWR(tool_msg_event_fops,
+		      NULL,
+		      tool_msg_event_write);
+
+static int tool_init_msgs(struct tool_ctx *tc)
+{
+	int midx, pidx;
+
+	/* Initialize inbound message structures */
+	tc->inmsg_cnt = ntb_msg_count(tc->ntb);
+	tc->inmsgs = devm_kcalloc(&tc->ntb->dev, tc->inmsg_cnt,
+				   sizeof(*tc->inmsgs), GFP_KERNEL);
+	if (tc->inmsgs == NULL)
+		return -ENOMEM;
+
+	for (midx = 0; midx < tc->inmsg_cnt; midx++) {
+		tc->inmsgs[midx].midx = midx;
+		tc->inmsgs[midx].pidx = -1;
+		tc->inmsgs[midx].tc = tc;
+	}
+
+	/* Initialize outbound message structures */
+	for (pidx = 0; pidx < tc->peer_cnt; pidx++) {
+		tc->peers[pidx].outmsg_cnt = ntb_msg_count(tc->ntb);
+		tc->peers[pidx].outmsgs =
+			devm_kcalloc(&tc->ntb->dev, tc->peers[pidx].outmsg_cnt,
+				sizeof(*tc->peers[pidx].outmsgs), GFP_KERNEL);
+		if (tc->peers[pidx].outmsgs == NULL)
+			return -ENOMEM;
+
+		for (midx = 0; midx < tc->peers[pidx].outmsg_cnt; midx++) {
+			tc->peers[pidx].outmsgs[midx].midx = midx;
+			tc->peers[pidx].outmsgs[midx].pidx = pidx;
+			tc->peers[pidx].outmsgs[midx].tc = tc;
+		}
+	}
+
+	return 0;
+}
+
+/*==============================================================================
+ *                          Initialization methods
+ *==============================================================================
+ */
+
+static struct tool_ctx *tool_create_data(struct ntb_dev *ntb)
+{
+	struct tool_ctx *tc;
+
+	tc = devm_kzalloc(&ntb->dev, sizeof(*tc), GFP_KERNEL);
+	if (tc == NULL)
+		return ERR_PTR(-ENOMEM);
+
+	tc->ntb = ntb;
+	init_waitqueue_head(&tc->link_wq);
+	init_waitqueue_head(&tc->db_wq);
+	init_waitqueue_head(&tc->msg_wq);
+
+	if (ntb_db_is_unsafe(ntb))
+		dev_dbg(&ntb->dev, "doorbell is unsafe\n");
+
+	if (ntb_spad_is_unsafe(ntb))
+		dev_dbg(&ntb->dev, "scratchpad is unsafe\n");
+
+	return tc;
+}
+
+static void tool_clear_data(struct tool_ctx *tc)
+{
+	wake_up(&tc->link_wq);
+	wake_up(&tc->db_wq);
+	wake_up(&tc->msg_wq);
+}
+
+static int tool_init_ntb(struct tool_ctx *tc)
+{
+	return ntb_set_ctx(tc->ntb, tc, &tool_ops);
+}
+
+static void tool_clear_ntb(struct tool_ctx *tc)
+{
+	ntb_clear_ctx(tc->ntb);
+	ntb_link_disable(tc->ntb);
+}
+
+static void tool_setup_dbgfs(struct tool_ctx *tc)
+{
+	int pidx, widx, sidx, midx;
+	char buf[TOOL_BUF_LEN];
+
+	/* This modules is useless without dbgfs... */
+	if (!tool_dbgfs_topdir) {
+		tc->dbgfs_dir = NULL;
+		return;
+	}
+
+	tc->dbgfs_dir = debugfs_create_dir(dev_name(&tc->ntb->dev),
+					   tool_dbgfs_topdir);
+
+	debugfs_create_file("port", 0600, tc->dbgfs_dir,
+			    tc, &tool_port_fops);
+
+	debugfs_create_file("link", 0600, tc->dbgfs_dir,
+			    tc, &tool_link_fops);
+
+	debugfs_create_file("db", 0600, tc->dbgfs_dir,
+			    tc, &tool_db_fops);
+
+	debugfs_create_file("db_valid_mask", 0600, tc->dbgfs_dir,
+			    tc, &tool_db_valid_mask_fops);
+
+	debugfs_create_file("db_mask", 0600, tc->dbgfs_dir,
+			    tc, &tool_db_mask_fops);
+
+	debugfs_create_file("db_event", 0600, tc->dbgfs_dir,
+			    tc, &tool_db_event_fops);
+
+	debugfs_create_file("peer_db", 0600, tc->dbgfs_dir,
+			    tc, &tool_peer_db_fops);
+
+	debugfs_create_file("peer_db_mask", 0600, tc->dbgfs_dir,
+			    tc, &tool_peer_db_mask_fops);
+
+	if (tc->inspad_cnt != 0) {
+		for (sidx = 0; sidx < tc->inspad_cnt; sidx++) {
+			snprintf(buf, sizeof(buf), "spad%d", sidx);
+
+			debugfs_create_file(buf, 0600, tc->dbgfs_dir,
+					   &tc->inspads[sidx], &tool_spad_fops);
+		}
+	}
+
+	if (tc->inmsg_cnt != 0) {
+		for (midx = 0; midx < tc->inmsg_cnt; midx++) {
+			snprintf(buf, sizeof(buf), "msg%d", midx);
+			debugfs_create_file(buf, 0600, tc->dbgfs_dir,
+					   &tc->inmsgs[midx], &tool_inmsg_fops);
+		}
+
+		debugfs_create_file("msg_sts", 0600, tc->dbgfs_dir,
+				    tc, &tool_msg_sts_fops);
+
+		debugfs_create_file("msg_inbits", 0600, tc->dbgfs_dir,
+				    tc, &tool_msg_inbits_fops);
+
+		debugfs_create_file("msg_outbits", 0600, tc->dbgfs_dir,
+				    tc, &tool_msg_outbits_fops);
+
+		debugfs_create_file("msg_mask", 0600, tc->dbgfs_dir,
+				    tc, &tool_msg_mask_fops);
+
+		debugfs_create_file("msg_event", 0600, tc->dbgfs_dir,
+				    tc, &tool_msg_event_fops);
+	}
+
+	for (pidx = 0; pidx < tc->peer_cnt; pidx++) {
+		snprintf(buf, sizeof(buf), "peer%d", pidx);
+		tc->peers[pidx].dbgfs_dir =
+			debugfs_create_dir(buf, tc->dbgfs_dir);
+
+		debugfs_create_file("port", 0600,
+				    tc->peers[pidx].dbgfs_dir,
+				    &tc->peers[pidx], &tool_peer_port_fops);
+
+		debugfs_create_file("link", 0200,
+				    tc->peers[pidx].dbgfs_dir,
+				    &tc->peers[pidx], &tool_peer_link_fops);
+
+		debugfs_create_file("link_event", 0200,
+				  tc->peers[pidx].dbgfs_dir,
+				  &tc->peers[pidx], &tool_peer_link_event_fops);
+
+		for (widx = 0; widx < tc->peers[pidx].inmw_cnt; widx++) {
+			snprintf(buf, sizeof(buf), "mw_trans%d", widx);
+			debugfs_create_file(buf, 0600,
+					    tc->peers[pidx].dbgfs_dir,
+					    &tc->peers[pidx].inmws[widx],
+					    &tool_mw_trans_fops);
+		}
+
+		for (widx = 0; widx < tc->peers[pidx].outmw_cnt; widx++) {
+			snprintf(buf, sizeof(buf), "peer_mw_trans%d", widx);
+			debugfs_create_file(buf, 0600,
+					    tc->peers[pidx].dbgfs_dir,
+					    &tc->peers[pidx].outmws[widx],
+					    &tool_peer_mw_trans_fops);
+		}
+
+		for (sidx = 0; sidx < tc->peers[pidx].outspad_cnt; sidx++) {
+			snprintf(buf, sizeof(buf), "spad%d", sidx);
+
+			debugfs_create_file(buf, 0600,
+					    tc->peers[pidx].dbgfs_dir,
+					    &tc->peers[pidx].outspads[sidx],
+					    &tool_peer_spad_fops);
+		}
+
+		for (midx = 0; midx < tc->peers[pidx].outmsg_cnt; midx++) {
+			snprintf(buf, sizeof(buf), "msg%d", midx);
+			debugfs_create_file(buf, 0600,
+					    tc->peers[pidx].dbgfs_dir,
+					    &tc->peers[pidx].outmsgs[midx],
+					    &tool_outmsg_fops);
+		}
+	}
+}
+
+static void tool_clear_dbgfs(struct tool_ctx *tc)
+{
+	debugfs_remove_recursive(tc->dbgfs_dir);
+}
+
+static int tool_probe(struct ntb_client *self, struct ntb_dev *ntb)
+{
+	struct tool_ctx *tc;
+	int ret;
+
+	tc = tool_create_data(ntb);
+	if (IS_ERR(tc))
+		return PTR_ERR(tc);
+
+	ret = tool_init_peers(tc);
+	if (ret != 0)
+		goto err_clear_data;
+
+	ret = tool_init_mws(tc);
+	if (ret != 0)
+		goto err_clear_data;
+
+	ret = tool_init_spads(tc);
+	if (ret != 0)
+		goto err_clear_mws;
+
+	ret = tool_init_msgs(tc);
+	if (ret != 0)
+		goto err_clear_mws;
+
+	ret = tool_init_ntb(tc);
+	if (ret != 0)
+		goto err_clear_mws;
+
+	tool_setup_dbgfs(tc);
+
+	return 0;
+
+err_clear_mws:
+	tool_clear_mws(tc);
+
+err_clear_data:
+	tool_clear_data(tc);
+
+	return ret;
+}
+
+static void tool_remove(struct ntb_client *self, struct ntb_dev *ntb)
+{
+	struct tool_ctx *tc = ntb->ctx;
+
+	tool_clear_dbgfs(tc);
+
+	tool_clear_ntb(tc);
+
+	tool_clear_mws(tc);
+
+	tool_clear_data(tc);
+}
+
+static struct ntb_client tool_client = {
+	.ops = {
+		.probe = tool_probe,
+		.remove = tool_remove,
+	}
+};
+
+static int __init tool_init(void)
+{
+	int ret;
+
+	if (debugfs_initialized())
+		tool_dbgfs_topdir = debugfs_create_dir(KBUILD_MODNAME, NULL);
+
+	ret = ntb_register_client(&tool_client);
+	if (ret)
+		debugfs_remove_recursive(tool_dbgfs_topdir);
+
+	return ret;
+}
+module_init(tool_init);
+
+static void __exit tool_exit(void)
+{
+	ntb_unregister_client(&tool_client);
+	debugfs_remove_recursive(tool_dbgfs_topdir);
+}
+module_exit(tool_exit);