1/*
   2 * This file is provided under a dual BSD/GPLv2 license.  When using or
   3 *   redistributing this file, you may do so under either license.
   4 *
   5 *   GPL LICENSE SUMMARY
   6 *
   7 *   Copyright(c) 2012 Intel Corporation. All rights reserved.
   8 *   Copyright (C) 2015 EMC Corporation. All Rights Reserved.
   9 *
  10 *   This program is free software; you can redistribute it and/or modify
  11 *   it under the terms of version 2 of the GNU General Public License as
  12 *   published by the Free Software Foundation.
  13 *
  14 *   BSD LICENSE
  15 *
  16 *   Copyright(c) 2012 Intel Corporation. All rights reserved.
  17 *   Copyright (C) 2015 EMC Corporation. All Rights Reserved.
  18 *
  19 *   Redistribution and use in source and binary forms, with or without
  20 *   modification, are permitted provided that the following conditions
  21 *   are met:
  22 *
  23 *     * Redistributions of source code must retain the above copyright
  24 *       notice, this list of conditions and the following disclaimer.
  25 *     * Redistributions in binary form must reproduce the above copy
  26 *       notice, this list of conditions and the following disclaimer in
  27 *       the documentation and/or other materials provided with the
  28 *       distribution.
  29 *     * Neither the name of Intel Corporation nor the names of its
  30 *       contributors may be used to endorse or promote products derived
  31 *       from this software without specific prior written permission.
  32 *
  33 *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  34 *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  35 *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  36 *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  37 *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  38 *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  39 *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  40 *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  41 *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  42 *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  43 *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  44 *
  45 * PCIe NTB Transport Linux driver
  46 *
  47 * Contact Information:
  48 * Jon Mason <jon.mason@intel.com>
  49 */
  50#include <linux/debugfs.h>
  51#include <linux/delay.h>
  52#include <linux/dmaengine.h>
  53#include <linux/dma-mapping.h>
  54#include <linux/errno.h>
  55#include <linux/export.h>
  56#include <linux/interrupt.h>
  57#include <linux/module.h>
  58#include <linux/pci.h>
  59#include <linux/slab.h>
  60#include <linux/types.h>
  61#include <linux/uaccess.h>
  62#include "linux/ntb.h"
  63#include "linux/ntb_transport.h"
  64
  65#define NTB_TRANSPORT_VERSION	4
  66#define NTB_TRANSPORT_VER	"4"
  67#define NTB_TRANSPORT_NAME	"ntb_transport"
  68#define NTB_TRANSPORT_DESC	"Software Queue-Pair Transport over NTB"
  69#define NTB_TRANSPORT_MIN_SPADS (MW0_SZ_HIGH + 2)
  70
  71MODULE_DESCRIPTION(NTB_TRANSPORT_DESC);
  72MODULE_VERSION(NTB_TRANSPORT_VER);
  73MODULE_LICENSE("Dual BSD/GPL");
  74MODULE_AUTHOR("Intel Corporation");
  75
  76static unsigned long max_mw_size;
  77module_param(max_mw_size, ulong, 0644);
  78MODULE_PARM_DESC(max_mw_size, "Limit size of large memory windows");
  79
  80static unsigned int transport_mtu = 0x10000;
  81module_param(transport_mtu, uint, 0644);
  82MODULE_PARM_DESC(transport_mtu, "Maximum size of NTB transport packets");
  83
  84static unsigned char max_num_clients;
  85module_param(max_num_clients, byte, 0644);
  86MODULE_PARM_DESC(max_num_clients, "Maximum number of NTB transport clients");
  87
  88static unsigned int copy_bytes = 1024;
  89module_param(copy_bytes, uint, 0644);
  90MODULE_PARM_DESC(copy_bytes, "Threshold under which NTB will use the CPU to copy instead of DMA");
  91
  92static bool use_dma;
  93module_param(use_dma, bool, 0644);
  94MODULE_PARM_DESC(use_dma, "Use DMA engine to perform large data copy");
  95
  96static bool use_msi;
  97#ifdef CONFIG_NTB_MSI
  98module_param(use_msi, bool, 0644);
  99MODULE_PARM_DESC(use_msi, "Use MSI interrupts instead of doorbells");
 100#endif
 101
 102static struct dentry *nt_debugfs_dir;
 103
 104/* Only two-ports NTB devices are supported */
 105#define PIDX		NTB_DEF_PEER_IDX
 106
 107struct ntb_queue_entry {
 108	/* ntb_queue list reference */
 109	struct list_head entry;
 110	/* pointers to data to be transferred */
 111	void *cb_data;
 112	void *buf;
 113	unsigned int len;
 114	unsigned int flags;
 115	int retries;
 116	int errors;
 117	unsigned int tx_index;
 118	unsigned int rx_index;
 119
 120	struct ntb_transport_qp *qp;
 121	union {
 122		struct ntb_payload_header __iomem *tx_hdr;
 123		struct ntb_payload_header *rx_hdr;
 124	};
 125};
 126
 127struct ntb_rx_info {
 128	unsigned int entry;
 129};
 130
 131struct ntb_transport_qp {
 132	struct ntb_transport_ctx *transport;
 133	struct ntb_dev *ndev;
 134	void *cb_data;
 135	struct dma_chan *tx_dma_chan;
 136	struct dma_chan *rx_dma_chan;
 137
 138	bool client_ready;
 139	bool link_is_up;
 140	bool active;
 141
 142	u8 qp_num;	/* Only 64 QP's are allowed.  0-63 */
 143	u64 qp_bit;
 144
 145	struct ntb_rx_info __iomem *rx_info;
 146	struct ntb_rx_info *remote_rx_info;
 147
 148	void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data,
 149			   void *data, int len);
 150	struct list_head tx_free_q;
 151	spinlock_t ntb_tx_free_q_lock;
 152	void __iomem *tx_mw;
 153	phys_addr_t tx_mw_phys;
 154	size_t tx_mw_size;
 155	dma_addr_t tx_mw_dma_addr;
 156	unsigned int tx_index;
 157	unsigned int tx_max_entry;
 158	unsigned int tx_max_frame;
 159
 160	void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
 161			   void *data, int len);
 162	struct list_head rx_post_q;
 163	struct list_head rx_pend_q;
 164	struct list_head rx_free_q;
 165	/* ntb_rx_q_lock: synchronize access to rx_XXXX_q */
 166	spinlock_t ntb_rx_q_lock;
 167	void *rx_buff;
 168	unsigned int rx_index;
 169	unsigned int rx_max_entry;
 170	unsigned int rx_max_frame;
 171	unsigned int rx_alloc_entry;
 172	dma_cookie_t last_cookie;
 173	struct tasklet_struct rxc_db_work;
 174
 175	void (*event_handler)(void *data, int status);
 176	struct delayed_work link_work;
 177	struct work_struct link_cleanup;
 178
 179	struct dentry *debugfs_dir;
 180	struct dentry *debugfs_stats;
 181
 182	/* Stats */
 183	u64 rx_bytes;
 184	u64 rx_pkts;
 185	u64 rx_ring_empty;
 186	u64 rx_err_no_buf;
 187	u64 rx_err_oflow;
 188	u64 rx_err_ver;
 189	u64 rx_memcpy;
 190	u64 rx_async;
 191	u64 tx_bytes;
 192	u64 tx_pkts;
 193	u64 tx_ring_full;
 194	u64 tx_err_no_buf;
 195	u64 tx_memcpy;
 196	u64 tx_async;
 197
 198	bool use_msi;
 199	int msi_irq;
 200	struct ntb_msi_desc msi_desc;
 201	struct ntb_msi_desc peer_msi_desc;
 202};
 203
 204struct ntb_transport_mw {
 205	phys_addr_t phys_addr;
 206	resource_size_t phys_size;
 207	void __iomem *vbase;
 208	size_t xlat_size;
 209	size_t buff_size;
 210	size_t alloc_size;
 211	void *alloc_addr;
 212	void *virt_addr;
 213	dma_addr_t dma_addr;
 214};
 215
 216struct ntb_transport_client_dev {
 217	struct list_head entry;
 218	struct ntb_transport_ctx *nt;
 219	struct device dev;
 220};
 221
 222struct ntb_transport_ctx {
 223	struct list_head entry;
 224	struct list_head client_devs;
 225
 226	struct ntb_dev *ndev;
 227
 228	struct ntb_transport_mw *mw_vec;
 229	struct ntb_transport_qp *qp_vec;
 230	unsigned int mw_count;
 231	unsigned int qp_count;
 232	u64 qp_bitmap;
 233	u64 qp_bitmap_free;
 234
 235	bool use_msi;
 236	unsigned int msi_spad_offset;
 237	u64 msi_db_mask;
 238
 239	bool link_is_up;
 240	struct delayed_work link_work;
 241	struct work_struct link_cleanup;
 242
 243	struct dentry *debugfs_node_dir;
 244};
 245
 246enum {
 247	DESC_DONE_FLAG = BIT(0),
 248	LINK_DOWN_FLAG = BIT(1),
 249};
 250
 251struct ntb_payload_header {
 252	unsigned int ver;
 253	unsigned int len;
 254	unsigned int flags;
 255};
 256
 257enum {
 258	VERSION = 0,
 259	QP_LINKS,
 260	NUM_QPS,
 261	NUM_MWS,
 262	MW0_SZ_HIGH,
 263	MW0_SZ_LOW,
 264};
 265
 266#define dev_client_dev(__dev) \
 267	container_of((__dev), struct ntb_transport_client_dev, dev)
 268
 269#define drv_client(__drv) \
 270	container_of((__drv), struct ntb_transport_client, driver)
 271
 272#define QP_TO_MW(nt, qp)	((qp) % nt->mw_count)
 273#define NTB_QP_DEF_NUM_ENTRIES	100
 274#define NTB_LINK_DOWN_TIMEOUT	10
 275
 276static void ntb_transport_rxc_db(unsigned long data);
 277static const struct ntb_ctx_ops ntb_transport_ops;
 278static struct ntb_client ntb_transport_client;
 279static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
 280			       struct ntb_queue_entry *entry);
 281static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset);
 282static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset);
 283static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset);
 284
 285
 286static int ntb_transport_bus_match(struct device *dev,
 287				   struct device_driver *drv)
 288{
 289	return !strncmp(dev_name(dev), drv->name, strlen(drv->name));
 290}
 291
 292static int ntb_transport_bus_probe(struct device *dev)
 293{
 294	const struct ntb_transport_client *client;
 295	int rc;
 296
 297	get_device(dev);
 298
 299	client = drv_client(dev->driver);
 300	rc = client->probe(dev);
 301	if (rc)
 302		put_device(dev);
 303
 304	return rc;
 305}
 306
 307static void ntb_transport_bus_remove(struct device *dev)
 308{
 309	const struct ntb_transport_client *client;
 310
 311	client = drv_client(dev->driver);
 312	client->remove(dev);
 313
 314	put_device(dev);
 
 
 315}
 316
 317static struct bus_type ntb_transport_bus = {
 318	.name = "ntb_transport",
 319	.match = ntb_transport_bus_match,
 320	.probe = ntb_transport_bus_probe,
 321	.remove = ntb_transport_bus_remove,
 322};
 323
 324static LIST_HEAD(ntb_transport_list);
 325
 326static int ntb_bus_init(struct ntb_transport_ctx *nt)
 327{
 328	list_add_tail(&nt->entry, &ntb_transport_list);
 329	return 0;
 330}
 331
 332static void ntb_bus_remove(struct ntb_transport_ctx *nt)
 333{
 334	struct ntb_transport_client_dev *client_dev, *cd;
 335
 336	list_for_each_entry_safe(client_dev, cd, &nt->client_devs, entry) {
 337		dev_err(client_dev->dev.parent, "%s still attached to bus, removing\n",
 338			dev_name(&client_dev->dev));
 339		list_del(&client_dev->entry);
 340		device_unregister(&client_dev->dev);
 341	}
 342
 343	list_del(&nt->entry);
 344}
 345
 346static void ntb_transport_client_release(struct device *dev)
 347{
 348	struct ntb_transport_client_dev *client_dev;
 349
 350	client_dev = dev_client_dev(dev);
 351	kfree(client_dev);
 352}
 353
 354/**
 355 * ntb_transport_unregister_client_dev - Unregister NTB client device
 356 * @device_name: Name of NTB client device
 357 *
 358 * Unregister an NTB client device with the NTB transport layer
 359 */
 360void ntb_transport_unregister_client_dev(char *device_name)
 361{
 362	struct ntb_transport_client_dev *client, *cd;
 363	struct ntb_transport_ctx *nt;
 364
 365	list_for_each_entry(nt, &ntb_transport_list, entry)
 366		list_for_each_entry_safe(client, cd, &nt->client_devs, entry)
 367			if (!strncmp(dev_name(&client->dev), device_name,
 368				     strlen(device_name))) {
 369				list_del(&client->entry);
 370				device_unregister(&client->dev);
 371			}
 372}
 373EXPORT_SYMBOL_GPL(ntb_transport_unregister_client_dev);
 374
 375/**
 376 * ntb_transport_register_client_dev - Register NTB client device
 377 * @device_name: Name of NTB client device
 378 *
 379 * Register an NTB client device with the NTB transport layer
 380 */
 381int ntb_transport_register_client_dev(char *device_name)
 382{
 383	struct ntb_transport_client_dev *client_dev;
 384	struct ntb_transport_ctx *nt;
 385	int node;
 386	int rc, i = 0;
 387
 388	if (list_empty(&ntb_transport_list))
 389		return -ENODEV;
 390
 391	list_for_each_entry(nt, &ntb_transport_list, entry) {
 392		struct device *dev;
 393
 394		node = dev_to_node(&nt->ndev->dev);
 395
 396		client_dev = kzalloc_node(sizeof(*client_dev),
 397					  GFP_KERNEL, node);
 398		if (!client_dev) {
 399			rc = -ENOMEM;
 400			goto err;
 401		}
 402
 403		dev = &client_dev->dev;
 404
 405		/* setup and register client devices */
 406		dev_set_name(dev, "%s%d", device_name, i);
 407		dev->bus = &ntb_transport_bus;
 408		dev->release = ntb_transport_client_release;
 409		dev->parent = &nt->ndev->dev;
 410
 411		rc = device_register(dev);
 412		if (rc) {
 413			kfree(client_dev);
 414			goto err;
 415		}
 416
 417		list_add_tail(&client_dev->entry, &nt->client_devs);
 418		i++;
 419	}
 420
 421	return 0;
 422
 423err:
 424	ntb_transport_unregister_client_dev(device_name);
 425
 426	return rc;
 427}
 428EXPORT_SYMBOL_GPL(ntb_transport_register_client_dev);
 429
 430/**
 431 * ntb_transport_register_client - Register NTB client driver
 432 * @drv: NTB client driver to be registered
 433 *
 434 * Register an NTB client driver with the NTB transport layer
 435 *
 436 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 437 */
 438int ntb_transport_register_client(struct ntb_transport_client *drv)
 439{
 440	drv->driver.bus = &ntb_transport_bus;
 441
 442	if (list_empty(&ntb_transport_list))
 443		return -ENODEV;
 444
 445	return driver_register(&drv->driver);
 446}
 447EXPORT_SYMBOL_GPL(ntb_transport_register_client);
 448
 449/**
 450 * ntb_transport_unregister_client - Unregister NTB client driver
 451 * @drv: NTB client driver to be unregistered
 452 *
 453 * Unregister an NTB client driver with the NTB transport layer
 454 *
 455 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 456 */
 457void ntb_transport_unregister_client(struct ntb_transport_client *drv)
 458{
 459	driver_unregister(&drv->driver);
 460}
 461EXPORT_SYMBOL_GPL(ntb_transport_unregister_client);
 462
 463static ssize_t debugfs_read(struct file *filp, char __user *ubuf, size_t count,
 464			    loff_t *offp)
 465{
 466	struct ntb_transport_qp *qp;
 467	char *buf;
 468	ssize_t ret, out_offset, out_count;
 469
 470	qp = filp->private_data;
 471
 472	if (!qp || !qp->link_is_up)
 473		return 0;
 474
 475	out_count = 1000;
 476
 477	buf = kmalloc(out_count, GFP_KERNEL);
 478	if (!buf)
 479		return -ENOMEM;
 480
 481	out_offset = 0;
 482	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 483			       "\nNTB QP stats:\n\n");
 484	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 485			       "rx_bytes - \t%llu\n", qp->rx_bytes);
 486	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 487			       "rx_pkts - \t%llu\n", qp->rx_pkts);
 488	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 489			       "rx_memcpy - \t%llu\n", qp->rx_memcpy);
 490	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 491			       "rx_async - \t%llu\n", qp->rx_async);
 492	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 493			       "rx_ring_empty - %llu\n", qp->rx_ring_empty);
 494	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 495			       "rx_err_no_buf - %llu\n", qp->rx_err_no_buf);
 496	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 497			       "rx_err_oflow - \t%llu\n", qp->rx_err_oflow);
 498	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 499			       "rx_err_ver - \t%llu\n", qp->rx_err_ver);
 500	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 501			       "rx_buff - \t0x%p\n", qp->rx_buff);
 502	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 503			       "rx_index - \t%u\n", qp->rx_index);
 504	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 505			       "rx_max_entry - \t%u\n", qp->rx_max_entry);
 506	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 507			       "rx_alloc_entry - \t%u\n\n", qp->rx_alloc_entry);
 508
 509	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 510			       "tx_bytes - \t%llu\n", qp->tx_bytes);
 511	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 512			       "tx_pkts - \t%llu\n", qp->tx_pkts);
 513	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 514			       "tx_memcpy - \t%llu\n", qp->tx_memcpy);
 515	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 516			       "tx_async - \t%llu\n", qp->tx_async);
 517	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 518			       "tx_ring_full - \t%llu\n", qp->tx_ring_full);
 519	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 520			       "tx_err_no_buf - %llu\n", qp->tx_err_no_buf);
 521	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 522			       "tx_mw - \t0x%p\n", qp->tx_mw);
 523	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 524			       "tx_index (H) - \t%u\n", qp->tx_index);
 525	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 526			       "RRI (T) - \t%u\n",
 527			       qp->remote_rx_info->entry);
 528	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 529			       "tx_max_entry - \t%u\n", qp->tx_max_entry);
 530	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 531			       "free tx - \t%u\n",
 532			       ntb_transport_tx_free_entry(qp));
 533
 534	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 535			       "\n");
 536	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 537			       "Using TX DMA - \t%s\n",
 538			       qp->tx_dma_chan ? "Yes" : "No");
 539	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 540			       "Using RX DMA - \t%s\n",
 541			       qp->rx_dma_chan ? "Yes" : "No");
 542	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 543			       "QP Link - \t%s\n",
 544			       qp->link_is_up ? "Up" : "Down");
 545	out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 546			       "\n");
 547
 548	if (out_offset > out_count)
 549		out_offset = out_count;
 550
 551	ret = simple_read_from_buffer(ubuf, count, offp, buf, out_offset);
 552	kfree(buf);
 553	return ret;
 554}
 555
 556static const struct file_operations ntb_qp_debugfs_stats = {
 557	.owner = THIS_MODULE,
 558	.open = simple_open,
 559	.read = debugfs_read,
 560};
 561
 562static void ntb_list_add(spinlock_t *lock, struct list_head *entry,
 563			 struct list_head *list)
 564{
 565	unsigned long flags;
 566
 567	spin_lock_irqsave(lock, flags);
 568	list_add_tail(entry, list);
 569	spin_unlock_irqrestore(lock, flags);
 570}
 571
 572static struct ntb_queue_entry *ntb_list_rm(spinlock_t *lock,
 573					   struct list_head *list)
 574{
 575	struct ntb_queue_entry *entry;
 576	unsigned long flags;
 577
 578	spin_lock_irqsave(lock, flags);
 579	if (list_empty(list)) {
 580		entry = NULL;
 581		goto out;
 582	}
 583	entry = list_first_entry(list, struct ntb_queue_entry, entry);
 584	list_del(&entry->entry);
 585
 586out:
 587	spin_unlock_irqrestore(lock, flags);
 588
 589	return entry;
 590}
 591
 592static struct ntb_queue_entry *ntb_list_mv(spinlock_t *lock,
 593					   struct list_head *list,
 594					   struct list_head *to_list)
 595{
 596	struct ntb_queue_entry *entry;
 597	unsigned long flags;
 598
 599	spin_lock_irqsave(lock, flags);
 600
 601	if (list_empty(list)) {
 602		entry = NULL;
 603	} else {
 604		entry = list_first_entry(list, struct ntb_queue_entry, entry);
 605		list_move_tail(&entry->entry, to_list);
 606	}
 607
 608	spin_unlock_irqrestore(lock, flags);
 609
 610	return entry;
 611}
 612
 613static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
 614				     unsigned int qp_num)
 615{
 616	struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
 617	struct ntb_transport_mw *mw;
 618	struct ntb_dev *ndev = nt->ndev;
 619	struct ntb_queue_entry *entry;
 620	unsigned int rx_size, num_qps_mw;
 621	unsigned int mw_num, mw_count, qp_count;
 622	unsigned int i;
 623	int node;
 624
 625	mw_count = nt->mw_count;
 626	qp_count = nt->qp_count;
 627
 628	mw_num = QP_TO_MW(nt, qp_num);
 629	mw = &nt->mw_vec[mw_num];
 630
 631	if (!mw->virt_addr)
 632		return -ENOMEM;
 633
 634	if (mw_num < qp_count % mw_count)
 635		num_qps_mw = qp_count / mw_count + 1;
 636	else
 637		num_qps_mw = qp_count / mw_count;
 638
 639	rx_size = (unsigned int)mw->xlat_size / num_qps_mw;
 640	qp->rx_buff = mw->virt_addr + rx_size * (qp_num / mw_count);
 641	rx_size -= sizeof(struct ntb_rx_info);
 642
 643	qp->remote_rx_info = qp->rx_buff + rx_size;
 644
 645	/* Due to housekeeping, there must be atleast 2 buffs */
 646	qp->rx_max_frame = min(transport_mtu, rx_size / 2);
 647	qp->rx_max_entry = rx_size / qp->rx_max_frame;
 648	qp->rx_index = 0;
 649
 650	/*
 651	 * Checking to see if we have more entries than the default.
 652	 * We should add additional entries if that is the case so we
 653	 * can be in sync with the transport frames.
 654	 */
 655	node = dev_to_node(&ndev->dev);
 656	for (i = qp->rx_alloc_entry; i < qp->rx_max_entry; i++) {
 657		entry = kzalloc_node(sizeof(*entry), GFP_KERNEL, node);
 658		if (!entry)
 659			return -ENOMEM;
 660
 661		entry->qp = qp;
 662		ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
 663			     &qp->rx_free_q);
 664		qp->rx_alloc_entry++;
 665	}
 666
 667	qp->remote_rx_info->entry = qp->rx_max_entry - 1;
 668
 669	/* setup the hdr offsets with 0's */
 670	for (i = 0; i < qp->rx_max_entry; i++) {
 671		void *offset = (qp->rx_buff + qp->rx_max_frame * (i + 1) -
 672				sizeof(struct ntb_payload_header));
 673		memset(offset, 0, sizeof(struct ntb_payload_header));
 674	}
 675
 676	qp->rx_pkts = 0;
 677	qp->tx_pkts = 0;
 678	qp->tx_index = 0;
 679
 680	return 0;
 681}
 682
 683static irqreturn_t ntb_transport_isr(int irq, void *dev)
 684{
 685	struct ntb_transport_qp *qp = dev;
 686
 687	tasklet_schedule(&qp->rxc_db_work);
 688
 689	return IRQ_HANDLED;
 690}
 691
 692static void ntb_transport_setup_qp_peer_msi(struct ntb_transport_ctx *nt,
 693					    unsigned int qp_num)
 694{
 695	struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
 696	int spad = qp_num * 2 + nt->msi_spad_offset;
 697
 698	if (!nt->use_msi)
 699		return;
 700
 701	if (spad >= ntb_spad_count(nt->ndev))
 702		return;
 703
 704	qp->peer_msi_desc.addr_offset =
 705		ntb_peer_spad_read(qp->ndev, PIDX, spad);
 706	qp->peer_msi_desc.data =
 707		ntb_peer_spad_read(qp->ndev, PIDX, spad + 1);
 708
 709	dev_dbg(&qp->ndev->pdev->dev, "QP%d Peer MSI addr=%x data=%x\n",
 710		qp_num, qp->peer_msi_desc.addr_offset, qp->peer_msi_desc.data);
 711
 712	if (qp->peer_msi_desc.addr_offset) {
 713		qp->use_msi = true;
 714		dev_info(&qp->ndev->pdev->dev,
 715			 "Using MSI interrupts for QP%d\n", qp_num);
 716	}
 717}
 718
 719static void ntb_transport_setup_qp_msi(struct ntb_transport_ctx *nt,
 720				       unsigned int qp_num)
 721{
 722	struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
 723	int spad = qp_num * 2 + nt->msi_spad_offset;
 724	int rc;
 725
 726	if (!nt->use_msi)
 727		return;
 728
 729	if (spad >= ntb_spad_count(nt->ndev)) {
 730		dev_warn_once(&qp->ndev->pdev->dev,
 731			      "Not enough SPADS to use MSI interrupts\n");
 732		return;
 733	}
 734
 735	ntb_spad_write(qp->ndev, spad, 0);
 736	ntb_spad_write(qp->ndev, spad + 1, 0);
 737
 738	if (!qp->msi_irq) {
 739		qp->msi_irq = ntbm_msi_request_irq(qp->ndev, ntb_transport_isr,
 740						   KBUILD_MODNAME, qp,
 741						   &qp->msi_desc);
 742		if (qp->msi_irq < 0) {
 743			dev_warn(&qp->ndev->pdev->dev,
 744				 "Unable to allocate MSI interrupt for qp%d\n",
 745				 qp_num);
 746			return;
 747		}
 748	}
 749
 750	rc = ntb_spad_write(qp->ndev, spad, qp->msi_desc.addr_offset);
 751	if (rc)
 752		goto err_free_interrupt;
 753
 754	rc = ntb_spad_write(qp->ndev, spad + 1, qp->msi_desc.data);
 755	if (rc)
 756		goto err_free_interrupt;
 757
 758	dev_dbg(&qp->ndev->pdev->dev, "QP%d MSI %d addr=%x data=%x\n",
 759		qp_num, qp->msi_irq, qp->msi_desc.addr_offset,
 760		qp->msi_desc.data);
 761
 762	return;
 763
 764err_free_interrupt:
 765	devm_free_irq(&nt->ndev->dev, qp->msi_irq, qp);
 766}
 767
 768static void ntb_transport_msi_peer_desc_changed(struct ntb_transport_ctx *nt)
 769{
 770	int i;
 771
 772	dev_dbg(&nt->ndev->pdev->dev, "Peer MSI descriptors changed");
 773
 774	for (i = 0; i < nt->qp_count; i++)
 775		ntb_transport_setup_qp_peer_msi(nt, i);
 776}
 777
 778static void ntb_transport_msi_desc_changed(void *data)
 779{
 780	struct ntb_transport_ctx *nt = data;
 781	int i;
 782
 783	dev_dbg(&nt->ndev->pdev->dev, "MSI descriptors changed");
 784
 785	for (i = 0; i < nt->qp_count; i++)
 786		ntb_transport_setup_qp_msi(nt, i);
 787
 788	ntb_peer_db_set(nt->ndev, nt->msi_db_mask);
 789}
 790
 791static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
 792{
 793	struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
 794	struct pci_dev *pdev = nt->ndev->pdev;
 795
 796	if (!mw->virt_addr)
 797		return;
 798
 799	ntb_mw_clear_trans(nt->ndev, PIDX, num_mw);
 800	dma_free_coherent(&pdev->dev, mw->alloc_size,
 801			  mw->alloc_addr, mw->dma_addr);
 802	mw->xlat_size = 0;
 803	mw->buff_size = 0;
 804	mw->alloc_size = 0;
 805	mw->alloc_addr = NULL;
 806	mw->virt_addr = NULL;
 807}
 808
 809static int ntb_alloc_mw_buffer(struct ntb_transport_mw *mw,
 810			       struct device *dma_dev, size_t align)
 811{
 812	dma_addr_t dma_addr;
 813	void *alloc_addr, *virt_addr;
 814	int rc;
 815
 816	alloc_addr = dma_alloc_coherent(dma_dev, mw->alloc_size,
 817					&dma_addr, GFP_KERNEL);
 818	if (!alloc_addr) {
 819		dev_err(dma_dev, "Unable to alloc MW buff of size %zu\n",
 820			mw->alloc_size);
 821		return -ENOMEM;
 822	}
 823	virt_addr = alloc_addr;
 824
 825	/*
 826	 * we must ensure that the memory address allocated is BAR size
 827	 * aligned in order for the XLAT register to take the value. This
 828	 * is a requirement of the hardware. It is recommended to setup CMA
 829	 * for BAR sizes equal or greater than 4MB.
 830	 */
 831	if (!IS_ALIGNED(dma_addr, align)) {
 832		if (mw->alloc_size > mw->buff_size) {
 833			virt_addr = PTR_ALIGN(alloc_addr, align);
 834			dma_addr = ALIGN(dma_addr, align);
 835		} else {
 836			rc = -ENOMEM;
 837			goto err;
 838		}
 839	}
 840
 841	mw->alloc_addr = alloc_addr;
 842	mw->virt_addr = virt_addr;
 843	mw->dma_addr = dma_addr;
 844
 845	return 0;
 846
 847err:
 848	dma_free_coherent(dma_dev, mw->alloc_size, alloc_addr, dma_addr);
 849
 850	return rc;
 851}
 852
 853static int ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw,
 854		      resource_size_t size)
 855{
 856	struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
 857	struct pci_dev *pdev = nt->ndev->pdev;
 858	size_t xlat_size, buff_size;
 859	resource_size_t xlat_align;
 860	resource_size_t xlat_align_size;
 861	int rc;
 862
 863	if (!size)
 864		return -EINVAL;
 865
 866	rc = ntb_mw_get_align(nt->ndev, PIDX, num_mw, &xlat_align,
 867			      &xlat_align_size, NULL);
 868	if (rc)
 869		return rc;
 870
 871	xlat_size = round_up(size, xlat_align_size);
 872	buff_size = round_up(size, xlat_align);
 873
 874	/* No need to re-setup */
 875	if (mw->xlat_size == xlat_size)
 876		return 0;
 877
 878	if (mw->buff_size)
 879		ntb_free_mw(nt, num_mw);
 880
 881	/* Alloc memory for receiving data.  Must be aligned */
 882	mw->xlat_size = xlat_size;
 883	mw->buff_size = buff_size;
 884	mw->alloc_size = buff_size;
 885
 886	rc = ntb_alloc_mw_buffer(mw, &pdev->dev, xlat_align);
 887	if (rc) {
 888		mw->alloc_size *= 2;
 889		rc = ntb_alloc_mw_buffer(mw, &pdev->dev, xlat_align);
 890		if (rc) {
 891			dev_err(&pdev->dev,
 892				"Unable to alloc aligned MW buff\n");
 893			mw->xlat_size = 0;
 894			mw->buff_size = 0;
 895			mw->alloc_size = 0;
 896			return rc;
 897		}
 
 
 
 
 
 
 
 
 
 898	}
 899
 900	/* Notify HW the memory location of the receive buffer */
 901	rc = ntb_mw_set_trans(nt->ndev, PIDX, num_mw, mw->dma_addr,
 902			      mw->xlat_size);
 903	if (rc) {
 904		dev_err(&pdev->dev, "Unable to set mw%d translation", num_mw);
 905		ntb_free_mw(nt, num_mw);
 906		return -EIO;
 907	}
 908
 909	return 0;
 910}
 911
 912static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
 913{
 914	qp->link_is_up = false;
 915	qp->active = false;
 916
 917	qp->tx_index = 0;
 918	qp->rx_index = 0;
 919	qp->rx_bytes = 0;
 920	qp->rx_pkts = 0;
 921	qp->rx_ring_empty = 0;
 922	qp->rx_err_no_buf = 0;
 923	qp->rx_err_oflow = 0;
 924	qp->rx_err_ver = 0;
 925	qp->rx_memcpy = 0;
 926	qp->rx_async = 0;
 927	qp->tx_bytes = 0;
 928	qp->tx_pkts = 0;
 929	qp->tx_ring_full = 0;
 930	qp->tx_err_no_buf = 0;
 931	qp->tx_memcpy = 0;
 932	qp->tx_async = 0;
 933}
 934
 935static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
 936{
 937	struct ntb_transport_ctx *nt = qp->transport;
 938	struct pci_dev *pdev = nt->ndev->pdev;
 939
 940	dev_info(&pdev->dev, "qp %d: Link Cleanup\n", qp->qp_num);
 941
 942	cancel_delayed_work_sync(&qp->link_work);
 943	ntb_qp_link_down_reset(qp);
 944
 945	if (qp->event_handler)
 946		qp->event_handler(qp->cb_data, qp->link_is_up);
 947}
 948
 949static void ntb_qp_link_cleanup_work(struct work_struct *work)
 950{
 951	struct ntb_transport_qp *qp = container_of(work,
 952						   struct ntb_transport_qp,
 953						   link_cleanup);
 954	struct ntb_transport_ctx *nt = qp->transport;
 955
 956	ntb_qp_link_cleanup(qp);
 957
 958	if (nt->link_is_up)
 959		schedule_delayed_work(&qp->link_work,
 960				      msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 961}
 962
 963static void ntb_qp_link_down(struct ntb_transport_qp *qp)
 964{
 965	schedule_work(&qp->link_cleanup);
 966}
 967
 968static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
 969{
 970	struct ntb_transport_qp *qp;
 971	u64 qp_bitmap_alloc;
 972	unsigned int i, count;
 973
 974	qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
 975
 976	/* Pass along the info to any clients */
 977	for (i = 0; i < nt->qp_count; i++)
 978		if (qp_bitmap_alloc & BIT_ULL(i)) {
 979			qp = &nt->qp_vec[i];
 980			ntb_qp_link_cleanup(qp);
 981			cancel_work_sync(&qp->link_cleanup);
 982			cancel_delayed_work_sync(&qp->link_work);
 983		}
 984
 985	if (!nt->link_is_up)
 986		cancel_delayed_work_sync(&nt->link_work);
 987
 988	for (i = 0; i < nt->mw_count; i++)
 989		ntb_free_mw(nt, i);
 990
 991	/* The scratchpad registers keep the values if the remote side
 992	 * goes down, blast them now to give them a sane value the next
 993	 * time they are accessed
 994	 */
 995	count = ntb_spad_count(nt->ndev);
 996	for (i = 0; i < count; i++)
 997		ntb_spad_write(nt->ndev, i, 0);
 998}
 999
1000static void ntb_transport_link_cleanup_work(struct work_struct *work)
1001{
1002	struct ntb_transport_ctx *nt =
1003		container_of(work, struct ntb_transport_ctx, link_cleanup);
1004
1005	ntb_transport_link_cleanup(nt);
1006}
1007
1008static void ntb_transport_event_callback(void *data)
1009{
1010	struct ntb_transport_ctx *nt = data;
1011
1012	if (ntb_link_is_up(nt->ndev, NULL, NULL) == 1)
1013		schedule_delayed_work(&nt->link_work, 0);
1014	else
1015		schedule_work(&nt->link_cleanup);
1016}
1017
1018static void ntb_transport_link_work(struct work_struct *work)
1019{
1020	struct ntb_transport_ctx *nt =
1021		container_of(work, struct ntb_transport_ctx, link_work.work);
1022	struct ntb_dev *ndev = nt->ndev;
1023	struct pci_dev *pdev = ndev->pdev;
1024	resource_size_t size;
1025	u32 val;
1026	int rc = 0, i, spad;
1027
1028	/* send the local info, in the opposite order of the way we read it */
1029
1030	if (nt->use_msi) {
1031		rc = ntb_msi_setup_mws(ndev);
1032		if (rc) {
1033			dev_warn(&pdev->dev,
1034				 "Failed to register MSI memory window: %d\n",
1035				 rc);
1036			nt->use_msi = false;
1037		}
1038	}
1039
1040	for (i = 0; i < nt->qp_count; i++)
1041		ntb_transport_setup_qp_msi(nt, i);
1042
1043	for (i = 0; i < nt->mw_count; i++) {
1044		size = nt->mw_vec[i].phys_size;
1045
1046		if (max_mw_size && size > max_mw_size)
1047			size = max_mw_size;
1048
1049		spad = MW0_SZ_HIGH + (i * 2);
1050		ntb_peer_spad_write(ndev, PIDX, spad, upper_32_bits(size));
1051
1052		spad = MW0_SZ_LOW + (i * 2);
1053		ntb_peer_spad_write(ndev, PIDX, spad, lower_32_bits(size));
1054	}
1055
1056	ntb_peer_spad_write(ndev, PIDX, NUM_MWS, nt->mw_count);
1057
1058	ntb_peer_spad_write(ndev, PIDX, NUM_QPS, nt->qp_count);
1059
1060	ntb_peer_spad_write(ndev, PIDX, VERSION, NTB_TRANSPORT_VERSION);
1061
1062	/* Query the remote side for its info */
1063	val = ntb_spad_read(ndev, VERSION);
1064	dev_dbg(&pdev->dev, "Remote version = %d\n", val);
1065	if (val != NTB_TRANSPORT_VERSION)
1066		goto out;
1067
1068	val = ntb_spad_read(ndev, NUM_QPS);
1069	dev_dbg(&pdev->dev, "Remote max number of qps = %d\n", val);
1070	if (val != nt->qp_count)
1071		goto out;
1072
1073	val = ntb_spad_read(ndev, NUM_MWS);
1074	dev_dbg(&pdev->dev, "Remote number of mws = %d\n", val);
1075	if (val != nt->mw_count)
1076		goto out;
1077
1078	for (i = 0; i < nt->mw_count; i++) {
1079		u64 val64;
1080
1081		val = ntb_spad_read(ndev, MW0_SZ_HIGH + (i * 2));
1082		val64 = (u64)val << 32;
1083
1084		val = ntb_spad_read(ndev, MW0_SZ_LOW + (i * 2));
1085		val64 |= val;
1086
1087		dev_dbg(&pdev->dev, "Remote MW%d size = %#llx\n", i, val64);
1088
1089		rc = ntb_set_mw(nt, i, val64);
1090		if (rc)
1091			goto out1;
1092	}
1093
1094	nt->link_is_up = true;
1095
1096	for (i = 0; i < nt->qp_count; i++) {
1097		struct ntb_transport_qp *qp = &nt->qp_vec[i];
1098
1099		ntb_transport_setup_qp_mw(nt, i);
1100		ntb_transport_setup_qp_peer_msi(nt, i);
1101
1102		if (qp->client_ready)
1103			schedule_delayed_work(&qp->link_work, 0);
1104	}
1105
1106	return;
1107
1108out1:
1109	for (i = 0; i < nt->mw_count; i++)
1110		ntb_free_mw(nt, i);
1111
1112	/* if there's an actual failure, we should just bail */
1113	if (rc < 0)
1114		return;
1115
1116out:
1117	if (ntb_link_is_up(ndev, NULL, NULL) == 1)
1118		schedule_delayed_work(&nt->link_work,
1119				      msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
1120}
1121
1122static void ntb_qp_link_work(struct work_struct *work)
1123{
1124	struct ntb_transport_qp *qp = container_of(work,
1125						   struct ntb_transport_qp,
1126						   link_work.work);
1127	struct pci_dev *pdev = qp->ndev->pdev;
1128	struct ntb_transport_ctx *nt = qp->transport;
1129	int val;
1130
1131	WARN_ON(!nt->link_is_up);
1132
1133	val = ntb_spad_read(nt->ndev, QP_LINKS);
1134
1135	ntb_peer_spad_write(nt->ndev, PIDX, QP_LINKS, val | BIT(qp->qp_num));
1136
1137	/* query remote spad for qp ready bits */
1138	dev_dbg_ratelimited(&pdev->dev, "Remote QP link status = %x\n", val);
1139
1140	/* See if the remote side is up */
1141	if (val & BIT(qp->qp_num)) {
1142		dev_info(&pdev->dev, "qp %d: Link Up\n", qp->qp_num);
1143		qp->link_is_up = true;
1144		qp->active = true;
1145
1146		if (qp->event_handler)
1147			qp->event_handler(qp->cb_data, qp->link_is_up);
1148
1149		if (qp->active)
1150			tasklet_schedule(&qp->rxc_db_work);
1151	} else if (nt->link_is_up)
1152		schedule_delayed_work(&qp->link_work,
1153				      msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
1154}
1155
1156static int ntb_transport_init_queue(struct ntb_transport_ctx *nt,
1157				    unsigned int qp_num)
1158{
1159	struct ntb_transport_qp *qp;
1160	phys_addr_t mw_base;
1161	resource_size_t mw_size;
1162	unsigned int num_qps_mw, tx_size;
1163	unsigned int mw_num, mw_count, qp_count;
1164	u64 qp_offset;
1165
1166	mw_count = nt->mw_count;
1167	qp_count = nt->qp_count;
1168
1169	mw_num = QP_TO_MW(nt, qp_num);
1170
1171	qp = &nt->qp_vec[qp_num];
1172	qp->qp_num = qp_num;
1173	qp->transport = nt;
1174	qp->ndev = nt->ndev;
1175	qp->client_ready = false;
1176	qp->event_handler = NULL;
1177	ntb_qp_link_down_reset(qp);
1178
1179	if (mw_num < qp_count % mw_count)
1180		num_qps_mw = qp_count / mw_count + 1;
1181	else
1182		num_qps_mw = qp_count / mw_count;
1183
1184	mw_base = nt->mw_vec[mw_num].phys_addr;
1185	mw_size = nt->mw_vec[mw_num].phys_size;
1186
1187	if (max_mw_size && mw_size > max_mw_size)
1188		mw_size = max_mw_size;
1189
1190	tx_size = (unsigned int)mw_size / num_qps_mw;
1191	qp_offset = tx_size * (qp_num / mw_count);
1192
1193	qp->tx_mw_size = tx_size;
1194	qp->tx_mw = nt->mw_vec[mw_num].vbase + qp_offset;
1195	if (!qp->tx_mw)
1196		return -EINVAL;
1197
1198	qp->tx_mw_phys = mw_base + qp_offset;
1199	if (!qp->tx_mw_phys)
1200		return -EINVAL;
1201
1202	tx_size -= sizeof(struct ntb_rx_info);
1203	qp->rx_info = qp->tx_mw + tx_size;
1204
1205	/* Due to housekeeping, there must be atleast 2 buffs */
1206	qp->tx_max_frame = min(transport_mtu, tx_size / 2);
1207	qp->tx_max_entry = tx_size / qp->tx_max_frame;
1208
1209	if (nt->debugfs_node_dir) {
1210		char debugfs_name[4];
1211
1212		snprintf(debugfs_name, 4, "qp%d", qp_num);
1213		qp->debugfs_dir = debugfs_create_dir(debugfs_name,
1214						     nt->debugfs_node_dir);
1215
1216		qp->debugfs_stats = debugfs_create_file("stats", S_IRUSR,
1217							qp->debugfs_dir, qp,
1218							&ntb_qp_debugfs_stats);
1219	} else {
1220		qp->debugfs_dir = NULL;
1221		qp->debugfs_stats = NULL;
1222	}
1223
1224	INIT_DELAYED_WORK(&qp->link_work, ntb_qp_link_work);
1225	INIT_WORK(&qp->link_cleanup, ntb_qp_link_cleanup_work);
1226
1227	spin_lock_init(&qp->ntb_rx_q_lock);
1228	spin_lock_init(&qp->ntb_tx_free_q_lock);
1229
1230	INIT_LIST_HEAD(&qp->rx_post_q);
1231	INIT_LIST_HEAD(&qp->rx_pend_q);
1232	INIT_LIST_HEAD(&qp->rx_free_q);
1233	INIT_LIST_HEAD(&qp->tx_free_q);
1234
1235	tasklet_init(&qp->rxc_db_work, ntb_transport_rxc_db,
1236		     (unsigned long)qp);
1237
1238	return 0;
1239}
1240
1241static int ntb_transport_probe(struct ntb_client *self, struct ntb_dev *ndev)
1242{
1243	struct ntb_transport_ctx *nt;
1244	struct ntb_transport_mw *mw;
1245	unsigned int mw_count, qp_count, spad_count, max_mw_count_for_spads;
1246	u64 qp_bitmap;
1247	int node;
1248	int rc, i;
1249
1250	mw_count = ntb_peer_mw_count(ndev);
1251
1252	if (!ndev->ops->mw_set_trans) {
1253		dev_err(&ndev->dev, "Inbound MW based NTB API is required\n");
1254		return -EINVAL;
1255	}
1256
1257	if (ntb_db_is_unsafe(ndev))
1258		dev_dbg(&ndev->dev,
1259			"doorbell is unsafe, proceed anyway...\n");
1260	if (ntb_spad_is_unsafe(ndev))
1261		dev_dbg(&ndev->dev,
1262			"scratchpad is unsafe, proceed anyway...\n");
1263
1264	if (ntb_peer_port_count(ndev) != NTB_DEF_PEER_CNT)
1265		dev_warn(&ndev->dev, "Multi-port NTB devices unsupported\n");
1266
1267	node = dev_to_node(&ndev->dev);
1268
1269	nt = kzalloc_node(sizeof(*nt), GFP_KERNEL, node);
1270	if (!nt)
1271		return -ENOMEM;
1272
1273	nt->ndev = ndev;
1274
1275	/*
1276	 * If we are using MSI, and have at least one extra memory window,
1277	 * we will reserve the last MW for the MSI window.
1278	 */
1279	if (use_msi && mw_count > 1) {
1280		rc = ntb_msi_init(ndev, ntb_transport_msi_desc_changed);
1281		if (!rc) {
1282			mw_count -= 1;
1283			nt->use_msi = true;
1284		}
1285	}
1286
1287	spad_count = ntb_spad_count(ndev);
1288
1289	/* Limit the MW's based on the availability of scratchpads */
1290
1291	if (spad_count < NTB_TRANSPORT_MIN_SPADS) {
1292		nt->mw_count = 0;
1293		rc = -EINVAL;
1294		goto err;
1295	}
1296
1297	max_mw_count_for_spads = (spad_count - MW0_SZ_HIGH) / 2;
1298	nt->mw_count = min(mw_count, max_mw_count_for_spads);
1299
1300	nt->msi_spad_offset = nt->mw_count * 2 + MW0_SZ_HIGH;
1301
1302	nt->mw_vec = kcalloc_node(mw_count, sizeof(*nt->mw_vec),
1303				  GFP_KERNEL, node);
1304	if (!nt->mw_vec) {
1305		rc = -ENOMEM;
1306		goto err;
1307	}
1308
1309	for (i = 0; i < mw_count; i++) {
1310		mw = &nt->mw_vec[i];
1311
1312		rc = ntb_peer_mw_get_addr(ndev, i, &mw->phys_addr,
1313					  &mw->phys_size);
1314		if (rc)
1315			goto err1;
1316
1317		mw->vbase = ioremap_wc(mw->phys_addr, mw->phys_size);
1318		if (!mw->vbase) {
1319			rc = -ENOMEM;
1320			goto err1;
1321		}
1322
1323		mw->buff_size = 0;
1324		mw->xlat_size = 0;
1325		mw->virt_addr = NULL;
1326		mw->dma_addr = 0;
1327	}
1328
1329	qp_bitmap = ntb_db_valid_mask(ndev);
1330
1331	qp_count = ilog2(qp_bitmap);
1332	if (nt->use_msi) {
1333		qp_count -= 1;
1334		nt->msi_db_mask = 1 << qp_count;
1335		ntb_db_clear_mask(ndev, nt->msi_db_mask);
1336	}
1337
1338	if (max_num_clients && max_num_clients < qp_count)
1339		qp_count = max_num_clients;
1340	else if (nt->mw_count < qp_count)
1341		qp_count = nt->mw_count;
1342
1343	qp_bitmap &= BIT_ULL(qp_count) - 1;
1344
1345	nt->qp_count = qp_count;
1346	nt->qp_bitmap = qp_bitmap;
1347	nt->qp_bitmap_free = qp_bitmap;
1348
1349	nt->qp_vec = kcalloc_node(qp_count, sizeof(*nt->qp_vec),
1350				  GFP_KERNEL, node);
1351	if (!nt->qp_vec) {
1352		rc = -ENOMEM;
1353		goto err1;
1354	}
1355
1356	if (nt_debugfs_dir) {
1357		nt->debugfs_node_dir =
1358			debugfs_create_dir(pci_name(ndev->pdev),
1359					   nt_debugfs_dir);
1360	}
1361
1362	for (i = 0; i < qp_count; i++) {
1363		rc = ntb_transport_init_queue(nt, i);
1364		if (rc)
1365			goto err2;
1366	}
1367
1368	INIT_DELAYED_WORK(&nt->link_work, ntb_transport_link_work);
1369	INIT_WORK(&nt->link_cleanup, ntb_transport_link_cleanup_work);
1370
1371	rc = ntb_set_ctx(ndev, nt, &ntb_transport_ops);
1372	if (rc)
1373		goto err2;
1374
1375	INIT_LIST_HEAD(&nt->client_devs);
1376	rc = ntb_bus_init(nt);
1377	if (rc)
1378		goto err3;
1379
1380	nt->link_is_up = false;
1381	ntb_link_enable(ndev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
1382	ntb_link_event(ndev);
1383
1384	return 0;
1385
1386err3:
1387	ntb_clear_ctx(ndev);
1388err2:
1389	kfree(nt->qp_vec);
1390err1:
1391	while (i--) {
1392		mw = &nt->mw_vec[i];
1393		iounmap(mw->vbase);
1394	}
1395	kfree(nt->mw_vec);
1396err:
1397	kfree(nt);
1398	return rc;
1399}
1400
1401static void ntb_transport_free(struct ntb_client *self, struct ntb_dev *ndev)
1402{
1403	struct ntb_transport_ctx *nt = ndev->ctx;
1404	struct ntb_transport_qp *qp;
1405	u64 qp_bitmap_alloc;
1406	int i;
1407
1408	ntb_transport_link_cleanup(nt);
1409	cancel_work_sync(&nt->link_cleanup);
1410	cancel_delayed_work_sync(&nt->link_work);
1411
1412	qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
1413
1414	/* verify that all the qp's are freed */
1415	for (i = 0; i < nt->qp_count; i++) {
1416		qp = &nt->qp_vec[i];
1417		if (qp_bitmap_alloc & BIT_ULL(i))
1418			ntb_transport_free_queue(qp);
1419		debugfs_remove_recursive(qp->debugfs_dir);
1420	}
1421
1422	ntb_link_disable(ndev);
1423	ntb_clear_ctx(ndev);
1424
1425	ntb_bus_remove(nt);
1426
1427	for (i = nt->mw_count; i--; ) {
1428		ntb_free_mw(nt, i);
1429		iounmap(nt->mw_vec[i].vbase);
1430	}
1431
1432	kfree(nt->qp_vec);
1433	kfree(nt->mw_vec);
1434	kfree(nt);
1435}
1436
1437static void ntb_complete_rxc(struct ntb_transport_qp *qp)
1438{
1439	struct ntb_queue_entry *entry;
1440	void *cb_data;
1441	unsigned int len;
1442	unsigned long irqflags;
1443
1444	spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1445
1446	while (!list_empty(&qp->rx_post_q)) {
1447		entry = list_first_entry(&qp->rx_post_q,
1448					 struct ntb_queue_entry, entry);
1449		if (!(entry->flags & DESC_DONE_FLAG))
1450			break;
1451
1452		entry->rx_hdr->flags = 0;
1453		iowrite32(entry->rx_index, &qp->rx_info->entry);
1454
1455		cb_data = entry->cb_data;
1456		len = entry->len;
1457
1458		list_move_tail(&entry->entry, &qp->rx_free_q);
1459
1460		spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1461
1462		if (qp->rx_handler && qp->client_ready)
1463			qp->rx_handler(qp, qp->cb_data, cb_data, len);
1464
1465		spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1466	}
1467
1468	spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1469}
1470
1471static void ntb_rx_copy_callback(void *data,
1472				 const struct dmaengine_result *res)
1473{
1474	struct ntb_queue_entry *entry = data;
1475
1476	/* we need to check DMA results if we are using DMA */
1477	if (res) {
1478		enum dmaengine_tx_result dma_err = res->result;
1479
1480		switch (dma_err) {
1481		case DMA_TRANS_READ_FAILED:
1482		case DMA_TRANS_WRITE_FAILED:
1483			entry->errors++;
1484			fallthrough;
1485		case DMA_TRANS_ABORTED:
1486		{
1487			struct ntb_transport_qp *qp = entry->qp;
1488			void *offset = qp->rx_buff + qp->rx_max_frame *
1489					qp->rx_index;
1490
1491			ntb_memcpy_rx(entry, offset);
1492			qp->rx_memcpy++;
1493			return;
1494		}
1495
1496		case DMA_TRANS_NOERROR:
1497		default:
1498			break;
1499		}
1500	}
1501
1502	entry->flags |= DESC_DONE_FLAG;
1503
1504	ntb_complete_rxc(entry->qp);
1505}
1506
1507static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset)
1508{
1509	void *buf = entry->buf;
1510	size_t len = entry->len;
1511
1512	memcpy(buf, offset, len);
1513
1514	/* Ensure that the data is fully copied out before clearing the flag */
1515	wmb();
1516
1517	ntb_rx_copy_callback(entry, NULL);
1518}
1519
1520static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset)
1521{
1522	struct dma_async_tx_descriptor *txd;
1523	struct ntb_transport_qp *qp = entry->qp;
1524	struct dma_chan *chan = qp->rx_dma_chan;
1525	struct dma_device *device;
1526	size_t pay_off, buff_off, len;
1527	struct dmaengine_unmap_data *unmap;
1528	dma_cookie_t cookie;
1529	void *buf = entry->buf;
1530
1531	len = entry->len;
1532	device = chan->device;
1533	pay_off = (size_t)offset & ~PAGE_MASK;
1534	buff_off = (size_t)buf & ~PAGE_MASK;
1535
1536	if (!is_dma_copy_aligned(device, pay_off, buff_off, len))
1537		goto err;
1538
1539	unmap = dmaengine_get_unmap_data(device->dev, 2, GFP_NOWAIT);
1540	if (!unmap)
1541		goto err;
1542
1543	unmap->len = len;
1544	unmap->addr[0] = dma_map_page(device->dev, virt_to_page(offset),
1545				      pay_off, len, DMA_TO_DEVICE);
1546	if (dma_mapping_error(device->dev, unmap->addr[0]))
1547		goto err_get_unmap;
1548
1549	unmap->to_cnt = 1;
1550
1551	unmap->addr[1] = dma_map_page(device->dev, virt_to_page(buf),
1552				      buff_off, len, DMA_FROM_DEVICE);
1553	if (dma_mapping_error(device->dev, unmap->addr[1]))
1554		goto err_get_unmap;
1555
1556	unmap->from_cnt = 1;
1557
1558	txd = device->device_prep_dma_memcpy(chan, unmap->addr[1],
1559					     unmap->addr[0], len,
1560					     DMA_PREP_INTERRUPT);
1561	if (!txd)
1562		goto err_get_unmap;
1563
1564	txd->callback_result = ntb_rx_copy_callback;
1565	txd->callback_param = entry;
1566	dma_set_unmap(txd, unmap);
1567
1568	cookie = dmaengine_submit(txd);
1569	if (dma_submit_error(cookie))
1570		goto err_set_unmap;
1571
1572	dmaengine_unmap_put(unmap);
1573
1574	qp->last_cookie = cookie;
1575
1576	qp->rx_async++;
1577
1578	return 0;
1579
1580err_set_unmap:
1581	dmaengine_unmap_put(unmap);
1582err_get_unmap:
1583	dmaengine_unmap_put(unmap);
1584err:
1585	return -ENXIO;
1586}
1587
1588static void ntb_async_rx(struct ntb_queue_entry *entry, void *offset)
1589{
1590	struct ntb_transport_qp *qp = entry->qp;
1591	struct dma_chan *chan = qp->rx_dma_chan;
1592	int res;
1593
1594	if (!chan)
1595		goto err;
1596
1597	if (entry->len < copy_bytes)
1598		goto err;
1599
1600	res = ntb_async_rx_submit(entry, offset);
1601	if (res < 0)
1602		goto err;
1603
1604	if (!entry->retries)
1605		qp->rx_async++;
1606
1607	return;
1608
1609err:
1610	ntb_memcpy_rx(entry, offset);
1611	qp->rx_memcpy++;
1612}
1613
1614static int ntb_process_rxc(struct ntb_transport_qp *qp)
1615{
1616	struct ntb_payload_header *hdr;
1617	struct ntb_queue_entry *entry;
1618	void *offset;
1619
1620	offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
1621	hdr = offset + qp->rx_max_frame - sizeof(struct ntb_payload_header);
1622
1623	dev_dbg(&qp->ndev->pdev->dev, "qp %d: RX ver %u len %d flags %x\n",
1624		qp->qp_num, hdr->ver, hdr->len, hdr->flags);
1625
1626	if (!(hdr->flags & DESC_DONE_FLAG)) {
1627		dev_dbg(&qp->ndev->pdev->dev, "done flag not set\n");
1628		qp->rx_ring_empty++;
1629		return -EAGAIN;
1630	}
1631
1632	if (hdr->flags & LINK_DOWN_FLAG) {
1633		dev_dbg(&qp->ndev->pdev->dev, "link down flag set\n");
1634		ntb_qp_link_down(qp);
1635		hdr->flags = 0;
1636		return -EAGAIN;
1637	}
1638
1639	if (hdr->ver != (u32)qp->rx_pkts) {
1640		dev_dbg(&qp->ndev->pdev->dev,
1641			"version mismatch, expected %llu - got %u\n",
1642			qp->rx_pkts, hdr->ver);
1643		qp->rx_err_ver++;
1644		return -EIO;
1645	}
1646
1647	entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q);
1648	if (!entry) {
1649		dev_dbg(&qp->ndev->pdev->dev, "no receive buffer\n");
1650		qp->rx_err_no_buf++;
1651		return -EAGAIN;
1652	}
1653
1654	entry->rx_hdr = hdr;
1655	entry->rx_index = qp->rx_index;
1656
1657	if (hdr->len > entry->len) {
1658		dev_dbg(&qp->ndev->pdev->dev,
1659			"receive buffer overflow! Wanted %d got %d\n",
1660			hdr->len, entry->len);
1661		qp->rx_err_oflow++;
1662
1663		entry->len = -EIO;
1664		entry->flags |= DESC_DONE_FLAG;
1665
1666		ntb_complete_rxc(qp);
1667	} else {
1668		dev_dbg(&qp->ndev->pdev->dev,
1669			"RX OK index %u ver %u size %d into buf size %d\n",
1670			qp->rx_index, hdr->ver, hdr->len, entry->len);
1671
1672		qp->rx_bytes += hdr->len;
1673		qp->rx_pkts++;
1674
1675		entry->len = hdr->len;
1676
1677		ntb_async_rx(entry, offset);
1678	}
1679
1680	qp->rx_index++;
1681	qp->rx_index %= qp->rx_max_entry;
1682
1683	return 0;
1684}
1685
1686static void ntb_transport_rxc_db(unsigned long data)
1687{
1688	struct ntb_transport_qp *qp = (void *)data;
1689	int rc, i;
1690
1691	dev_dbg(&qp->ndev->pdev->dev, "%s: doorbell %d received\n",
1692		__func__, qp->qp_num);
1693
1694	/* Limit the number of packets processed in a single interrupt to
1695	 * provide fairness to others
1696	 */
1697	for (i = 0; i < qp->rx_max_entry; i++) {
1698		rc = ntb_process_rxc(qp);
1699		if (rc)
1700			break;
1701	}
1702
1703	if (i && qp->rx_dma_chan)
1704		dma_async_issue_pending(qp->rx_dma_chan);
1705
1706	if (i == qp->rx_max_entry) {
1707		/* there is more work to do */
1708		if (qp->active)
1709			tasklet_schedule(&qp->rxc_db_work);
1710	} else if (ntb_db_read(qp->ndev) & BIT_ULL(qp->qp_num)) {
1711		/* the doorbell bit is set: clear it */
1712		ntb_db_clear(qp->ndev, BIT_ULL(qp->qp_num));
1713		/* ntb_db_read ensures ntb_db_clear write is committed */
1714		ntb_db_read(qp->ndev);
1715
1716		/* an interrupt may have arrived between finishing
1717		 * ntb_process_rxc and clearing the doorbell bit:
1718		 * there might be some more work to do.
1719		 */
1720		if (qp->active)
1721			tasklet_schedule(&qp->rxc_db_work);
1722	}
1723}
1724
1725static void ntb_tx_copy_callback(void *data,
1726				 const struct dmaengine_result *res)
1727{
1728	struct ntb_queue_entry *entry = data;
1729	struct ntb_transport_qp *qp = entry->qp;
1730	struct ntb_payload_header __iomem *hdr = entry->tx_hdr;
1731
1732	/* we need to check DMA results if we are using DMA */
1733	if (res) {
1734		enum dmaengine_tx_result dma_err = res->result;
1735
1736		switch (dma_err) {
1737		case DMA_TRANS_READ_FAILED:
1738		case DMA_TRANS_WRITE_FAILED:
1739			entry->errors++;
1740			fallthrough;
1741		case DMA_TRANS_ABORTED:
1742		{
1743			void __iomem *offset =
1744				qp->tx_mw + qp->tx_max_frame *
1745				entry->tx_index;
1746
1747			/* resubmit via CPU */
1748			ntb_memcpy_tx(entry, offset);
1749			qp->tx_memcpy++;
1750			return;
1751		}
1752
1753		case DMA_TRANS_NOERROR:
1754		default:
1755			break;
1756		}
1757	}
1758
1759	iowrite32(entry->flags | DESC_DONE_FLAG, &hdr->flags);
1760
1761	if (qp->use_msi)
1762		ntb_msi_peer_trigger(qp->ndev, PIDX, &qp->peer_msi_desc);
1763	else
1764		ntb_peer_db_set(qp->ndev, BIT_ULL(qp->qp_num));
1765
1766	/* The entry length can only be zero if the packet is intended to be a
1767	 * "link down" or similar.  Since no payload is being sent in these
1768	 * cases, there is nothing to add to the completion queue.
1769	 */
1770	if (entry->len > 0) {
1771		qp->tx_bytes += entry->len;
1772
1773		if (qp->tx_handler)
1774			qp->tx_handler(qp, qp->cb_data, entry->cb_data,
1775				       entry->len);
1776	}
1777
1778	ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry, &qp->tx_free_q);
1779}
1780
1781static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset)
1782{
1783#ifdef ARCH_HAS_NOCACHE_UACCESS
1784	/*
1785	 * Using non-temporal mov to improve performance on non-cached
1786	 * writes, even though we aren't actually copying from user space.
1787	 */
1788	__copy_from_user_inatomic_nocache(offset, entry->buf, entry->len);
1789#else
1790	memcpy_toio(offset, entry->buf, entry->len);
1791#endif
1792
1793	/* Ensure that the data is fully copied out before setting the flags */
1794	wmb();
1795
1796	ntb_tx_copy_callback(entry, NULL);
1797}
1798
1799static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
1800			       struct ntb_queue_entry *entry)
1801{
1802	struct dma_async_tx_descriptor *txd;
1803	struct dma_chan *chan = qp->tx_dma_chan;
1804	struct dma_device *device;
1805	size_t len = entry->len;
1806	void *buf = entry->buf;
1807	size_t dest_off, buff_off;
1808	struct dmaengine_unmap_data *unmap;
1809	dma_addr_t dest;
1810	dma_cookie_t cookie;
1811
1812	device = chan->device;
1813	dest = qp->tx_mw_dma_addr + qp->tx_max_frame * entry->tx_index;
1814	buff_off = (size_t)buf & ~PAGE_MASK;
1815	dest_off = (size_t)dest & ~PAGE_MASK;
1816
1817	if (!is_dma_copy_aligned(device, buff_off, dest_off, len))
1818		goto err;
1819
1820	unmap = dmaengine_get_unmap_data(device->dev, 1, GFP_NOWAIT);
1821	if (!unmap)
1822		goto err;
1823
1824	unmap->len = len;
1825	unmap->addr[0] = dma_map_page(device->dev, virt_to_page(buf),
1826				      buff_off, len, DMA_TO_DEVICE);
1827	if (dma_mapping_error(device->dev, unmap->addr[0]))
1828		goto err_get_unmap;
1829
1830	unmap->to_cnt = 1;
1831
1832	txd = device->device_prep_dma_memcpy(chan, dest, unmap->addr[0], len,
1833					     DMA_PREP_INTERRUPT);
1834	if (!txd)
1835		goto err_get_unmap;
1836
1837	txd->callback_result = ntb_tx_copy_callback;
1838	txd->callback_param = entry;
1839	dma_set_unmap(txd, unmap);
1840
1841	cookie = dmaengine_submit(txd);
1842	if (dma_submit_error(cookie))
1843		goto err_set_unmap;
1844
1845	dmaengine_unmap_put(unmap);
1846
1847	dma_async_issue_pending(chan);
1848
1849	return 0;
1850err_set_unmap:
1851	dmaengine_unmap_put(unmap);
1852err_get_unmap:
1853	dmaengine_unmap_put(unmap);
1854err:
1855	return -ENXIO;
1856}
1857
1858static void ntb_async_tx(struct ntb_transport_qp *qp,
1859			 struct ntb_queue_entry *entry)
1860{
1861	struct ntb_payload_header __iomem *hdr;
1862	struct dma_chan *chan = qp->tx_dma_chan;
1863	void __iomem *offset;
1864	int res;
1865
1866	entry->tx_index = qp->tx_index;
1867	offset = qp->tx_mw + qp->tx_max_frame * entry->tx_index;
1868	hdr = offset + qp->tx_max_frame - sizeof(struct ntb_payload_header);
1869	entry->tx_hdr = hdr;
1870
1871	iowrite32(entry->len, &hdr->len);
1872	iowrite32((u32)qp->tx_pkts, &hdr->ver);
1873
1874	if (!chan)
1875		goto err;
1876
1877	if (entry->len < copy_bytes)
1878		goto err;
1879
1880	res = ntb_async_tx_submit(qp, entry);
1881	if (res < 0)
1882		goto err;
1883
1884	if (!entry->retries)
1885		qp->tx_async++;
1886
1887	return;
1888
1889err:
1890	ntb_memcpy_tx(entry, offset);
1891	qp->tx_memcpy++;
1892}
1893
1894static int ntb_process_tx(struct ntb_transport_qp *qp,
1895			  struct ntb_queue_entry *entry)
1896{
1897	if (qp->tx_index == qp->remote_rx_info->entry) {
1898		qp->tx_ring_full++;
1899		return -EAGAIN;
1900	}
1901
1902	if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
1903		if (qp->tx_handler)
1904			qp->tx_handler(qp, qp->cb_data, NULL, -EIO);
1905
1906		ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1907			     &qp->tx_free_q);
1908		return 0;
1909	}
1910
1911	ntb_async_tx(qp, entry);
1912
1913	qp->tx_index++;
1914	qp->tx_index %= qp->tx_max_entry;
1915
1916	qp->tx_pkts++;
1917
1918	return 0;
1919}
1920
1921static void ntb_send_link_down(struct ntb_transport_qp *qp)
1922{
1923	struct pci_dev *pdev = qp->ndev->pdev;
1924	struct ntb_queue_entry *entry;
1925	int i, rc;
1926
1927	if (!qp->link_is_up)
1928		return;
1929
1930	dev_info(&pdev->dev, "qp %d: Send Link Down\n", qp->qp_num);
1931
1932	for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1933		entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1934		if (entry)
1935			break;
1936		msleep(100);
1937	}
1938
1939	if (!entry)
1940		return;
1941
1942	entry->cb_data = NULL;
1943	entry->buf = NULL;
1944	entry->len = 0;
1945	entry->flags = LINK_DOWN_FLAG;
1946
1947	rc = ntb_process_tx(qp, entry);
1948	if (rc)
1949		dev_err(&pdev->dev, "ntb: QP%d unable to send linkdown msg\n",
1950			qp->qp_num);
1951
1952	ntb_qp_link_down_reset(qp);
1953}
1954
1955static bool ntb_dma_filter_fn(struct dma_chan *chan, void *node)
1956{
1957	return dev_to_node(&chan->dev->device) == (int)(unsigned long)node;
1958}
1959
1960/**
1961 * ntb_transport_create_queue - Create a new NTB transport layer queue
1962 * @rx_handler: receive callback function
1963 * @tx_handler: transmit callback function
1964 * @event_handler: event callback function
1965 *
1966 * Create a new NTB transport layer queue and provide the queue with a callback
1967 * routine for both transmit and receive.  The receive callback routine will be
1968 * used to pass up data when the transport has received it on the queue.   The
1969 * transmit callback routine will be called when the transport has completed the
1970 * transmission of the data on the queue and the data is ready to be freed.
1971 *
1972 * RETURNS: pointer to newly created ntb_queue, NULL on error.
1973 */
1974struct ntb_transport_qp *
1975ntb_transport_create_queue(void *data, struct device *client_dev,
1976			   const struct ntb_queue_handlers *handlers)
1977{
1978	struct ntb_dev *ndev;
1979	struct pci_dev *pdev;
1980	struct ntb_transport_ctx *nt;
1981	struct ntb_queue_entry *entry;
1982	struct ntb_transport_qp *qp;
1983	u64 qp_bit;
1984	unsigned int free_queue;
1985	dma_cap_mask_t dma_mask;
1986	int node;
1987	int i;
1988
1989	ndev = dev_ntb(client_dev->parent);
1990	pdev = ndev->pdev;
1991	nt = ndev->ctx;
1992
1993	node = dev_to_node(&ndev->dev);
1994
1995	free_queue = ffs(nt->qp_bitmap_free);
1996	if (!free_queue)
1997		goto err;
1998
1999	/* decrement free_queue to make it zero based */
2000	free_queue--;
2001
2002	qp = &nt->qp_vec[free_queue];
2003	qp_bit = BIT_ULL(qp->qp_num);
2004
2005	nt->qp_bitmap_free &= ~qp_bit;
2006
2007	qp->cb_data = data;
2008	qp->rx_handler = handlers->rx_handler;
2009	qp->tx_handler = handlers->tx_handler;
2010	qp->event_handler = handlers->event_handler;
2011
2012	dma_cap_zero(dma_mask);
2013	dma_cap_set(DMA_MEMCPY, dma_mask);
2014
2015	if (use_dma) {
2016		qp->tx_dma_chan =
2017			dma_request_channel(dma_mask, ntb_dma_filter_fn,
2018					    (void *)(unsigned long)node);
2019		if (!qp->tx_dma_chan)
2020			dev_info(&pdev->dev, "Unable to allocate TX DMA channel\n");
2021
2022		qp->rx_dma_chan =
2023			dma_request_channel(dma_mask, ntb_dma_filter_fn,
2024					    (void *)(unsigned long)node);
2025		if (!qp->rx_dma_chan)
2026			dev_info(&pdev->dev, "Unable to allocate RX DMA channel\n");
2027	} else {
2028		qp->tx_dma_chan = NULL;
2029		qp->rx_dma_chan = NULL;
2030	}
2031
2032	qp->tx_mw_dma_addr = 0;
2033	if (qp->tx_dma_chan) {
2034		qp->tx_mw_dma_addr =
2035			dma_map_resource(qp->tx_dma_chan->device->dev,
2036					 qp->tx_mw_phys, qp->tx_mw_size,
2037					 DMA_FROM_DEVICE, 0);
2038		if (dma_mapping_error(qp->tx_dma_chan->device->dev,
2039				      qp->tx_mw_dma_addr)) {
2040			qp->tx_mw_dma_addr = 0;
2041			goto err1;
2042		}
2043	}
2044
2045	dev_dbg(&pdev->dev, "Using %s memcpy for TX\n",
2046		qp->tx_dma_chan ? "DMA" : "CPU");
2047
2048	dev_dbg(&pdev->dev, "Using %s memcpy for RX\n",
2049		qp->rx_dma_chan ? "DMA" : "CPU");
2050
2051	for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
2052		entry = kzalloc_node(sizeof(*entry), GFP_KERNEL, node);
2053		if (!entry)
2054			goto err1;
2055
2056		entry->qp = qp;
2057		ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
2058			     &qp->rx_free_q);
2059	}
2060	qp->rx_alloc_entry = NTB_QP_DEF_NUM_ENTRIES;
2061
2062	for (i = 0; i < qp->tx_max_entry; i++) {
2063		entry = kzalloc_node(sizeof(*entry), GFP_KERNEL, node);
2064		if (!entry)
2065			goto err2;
2066
2067		entry->qp = qp;
2068		ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
2069			     &qp->tx_free_q);
2070	}
2071
2072	ntb_db_clear(qp->ndev, qp_bit);
2073	ntb_db_clear_mask(qp->ndev, qp_bit);
2074
2075	dev_info(&pdev->dev, "NTB Transport QP %d created\n", qp->qp_num);
2076
2077	return qp;
2078
2079err2:
2080	while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
2081		kfree(entry);
2082err1:
2083	qp->rx_alloc_entry = 0;
2084	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
2085		kfree(entry);
2086	if (qp->tx_mw_dma_addr)
2087		dma_unmap_resource(qp->tx_dma_chan->device->dev,
2088				   qp->tx_mw_dma_addr, qp->tx_mw_size,
2089				   DMA_FROM_DEVICE, 0);
2090	if (qp->tx_dma_chan)
2091		dma_release_channel(qp->tx_dma_chan);
2092	if (qp->rx_dma_chan)
2093		dma_release_channel(qp->rx_dma_chan);
2094	nt->qp_bitmap_free |= qp_bit;
2095err:
2096	return NULL;
2097}
2098EXPORT_SYMBOL_GPL(ntb_transport_create_queue);
2099
2100/**
2101 * ntb_transport_free_queue - Frees NTB transport queue
2102 * @qp: NTB queue to be freed
2103 *
2104 * Frees NTB transport queue
2105 */
2106void ntb_transport_free_queue(struct ntb_transport_qp *qp)
2107{
2108	struct pci_dev *pdev;
2109	struct ntb_queue_entry *entry;
2110	u64 qp_bit;
2111
2112	if (!qp)
2113		return;
2114
2115	pdev = qp->ndev->pdev;
2116
2117	qp->active = false;
2118
2119	if (qp->tx_dma_chan) {
2120		struct dma_chan *chan = qp->tx_dma_chan;
2121		/* Putting the dma_chan to NULL will force any new traffic to be
2122		 * processed by the CPU instead of the DAM engine
2123		 */
2124		qp->tx_dma_chan = NULL;
2125
2126		/* Try to be nice and wait for any queued DMA engine
2127		 * transactions to process before smashing it with a rock
2128		 */
2129		dma_sync_wait(chan, qp->last_cookie);
2130		dmaengine_terminate_all(chan);
2131
2132		dma_unmap_resource(chan->device->dev,
2133				   qp->tx_mw_dma_addr, qp->tx_mw_size,
2134				   DMA_FROM_DEVICE, 0);
2135
2136		dma_release_channel(chan);
2137	}
2138
2139	if (qp->rx_dma_chan) {
2140		struct dma_chan *chan = qp->rx_dma_chan;
2141		/* Putting the dma_chan to NULL will force any new traffic to be
2142		 * processed by the CPU instead of the DAM engine
2143		 */
2144		qp->rx_dma_chan = NULL;
2145
2146		/* Try to be nice and wait for any queued DMA engine
2147		 * transactions to process before smashing it with a rock
2148		 */
2149		dma_sync_wait(chan, qp->last_cookie);
2150		dmaengine_terminate_all(chan);
2151		dma_release_channel(chan);
2152	}
2153
2154	qp_bit = BIT_ULL(qp->qp_num);
2155
2156	ntb_db_set_mask(qp->ndev, qp_bit);
2157	tasklet_kill(&qp->rxc_db_work);
2158
2159	cancel_delayed_work_sync(&qp->link_work);
2160
2161	qp->cb_data = NULL;
2162	qp->rx_handler = NULL;
2163	qp->tx_handler = NULL;
2164	qp->event_handler = NULL;
2165
2166	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
2167		kfree(entry);
2168
2169	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q))) {
2170		dev_warn(&pdev->dev, "Freeing item from non-empty rx_pend_q\n");
2171		kfree(entry);
2172	}
2173
2174	while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q))) {
2175		dev_warn(&pdev->dev, "Freeing item from non-empty rx_post_q\n");
2176		kfree(entry);
2177	}
2178
2179	while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
2180		kfree(entry);
2181
2182	qp->transport->qp_bitmap_free |= qp_bit;
2183
2184	dev_info(&pdev->dev, "NTB Transport QP %d freed\n", qp->qp_num);
2185}
2186EXPORT_SYMBOL_GPL(ntb_transport_free_queue);
2187
2188/**
2189 * ntb_transport_rx_remove - Dequeues enqueued rx packet
2190 * @qp: NTB queue to be freed
2191 * @len: pointer to variable to write enqueued buffers length
2192 *
2193 * Dequeues unused buffers from receive queue.  Should only be used during
2194 * shutdown of qp.
2195 *
2196 * RETURNS: NULL error value on error, or void* for success.
2197 */
2198void *ntb_transport_rx_remove(struct ntb_transport_qp *qp, unsigned int *len)
2199{
2200	struct ntb_queue_entry *entry;
2201	void *buf;
2202
2203	if (!qp || qp->client_ready)
2204		return NULL;
2205
2206	entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q);
2207	if (!entry)
2208		return NULL;
2209
2210	buf = entry->cb_data;
2211	*len = entry->len;
2212
2213	ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_free_q);
2214
2215	return buf;
2216}
2217EXPORT_SYMBOL_GPL(ntb_transport_rx_remove);
2218
2219/**
2220 * ntb_transport_rx_enqueue - Enqueue a new NTB queue entry
2221 * @qp: NTB transport layer queue the entry is to be enqueued on
2222 * @cb: per buffer pointer for callback function to use
2223 * @data: pointer to data buffer that incoming packets will be copied into
2224 * @len: length of the data buffer
2225 *
2226 * Enqueue a new receive buffer onto the transport queue into which a NTB
2227 * payload can be received into.
2228 *
2229 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2230 */
2231int ntb_transport_rx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
2232			     unsigned int len)
2233{
2234	struct ntb_queue_entry *entry;
2235
2236	if (!qp)
2237		return -EINVAL;
2238
2239	entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q);
2240	if (!entry)
2241		return -ENOMEM;
2242
2243	entry->cb_data = cb;
2244	entry->buf = data;
2245	entry->len = len;
2246	entry->flags = 0;
2247	entry->retries = 0;
2248	entry->errors = 0;
2249	entry->rx_index = 0;
2250
2251	ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_pend_q);
2252
2253	if (qp->active)
2254		tasklet_schedule(&qp->rxc_db_work);
2255
2256	return 0;
2257}
2258EXPORT_SYMBOL_GPL(ntb_transport_rx_enqueue);
2259
2260/**
2261 * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
2262 * @qp: NTB transport layer queue the entry is to be enqueued on
2263 * @cb: per buffer pointer for callback function to use
2264 * @data: pointer to data buffer that will be sent
2265 * @len: length of the data buffer
2266 *
2267 * Enqueue a new transmit buffer onto the transport queue from which a NTB
2268 * payload will be transmitted.  This assumes that a lock is being held to
2269 * serialize access to the qp.
2270 *
2271 * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2272 */
2273int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
2274			     unsigned int len)
2275{
2276	struct ntb_queue_entry *entry;
2277	int rc;
2278
2279	if (!qp || !qp->link_is_up || !len)
2280		return -EINVAL;
2281
2282	entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
2283	if (!entry) {
2284		qp->tx_err_no_buf++;
2285		return -EBUSY;
2286	}
2287
2288	entry->cb_data = cb;
2289	entry->buf = data;
2290	entry->len = len;
2291	entry->flags = 0;
2292	entry->errors = 0;
2293	entry->retries = 0;
2294	entry->tx_index = 0;
2295
2296	rc = ntb_process_tx(qp, entry);
2297	if (rc)
2298		ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
2299			     &qp->tx_free_q);
2300
2301	return rc;
2302}
2303EXPORT_SYMBOL_GPL(ntb_transport_tx_enqueue);
2304
2305/**
2306 * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
2307 * @qp: NTB transport layer queue to be enabled
2308 *
2309 * Notify NTB transport layer of client readiness to use queue
2310 */
2311void ntb_transport_link_up(struct ntb_transport_qp *qp)
2312{
2313	if (!qp)
2314		return;
2315
2316	qp->client_ready = true;
2317
2318	if (qp->transport->link_is_up)
2319		schedule_delayed_work(&qp->link_work, 0);
2320}
2321EXPORT_SYMBOL_GPL(ntb_transport_link_up);
2322
2323/**
2324 * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
2325 * @qp: NTB transport layer queue to be disabled
2326 *
2327 * Notify NTB transport layer of client's desire to no longer receive data on
2328 * transport queue specified.  It is the client's responsibility to ensure all
2329 * entries on queue are purged or otherwise handled appropriately.
2330 */
2331void ntb_transport_link_down(struct ntb_transport_qp *qp)
2332{
2333	int val;
2334
2335	if (!qp)
2336		return;
2337
2338	qp->client_ready = false;
2339
2340	val = ntb_spad_read(qp->ndev, QP_LINKS);
2341
2342	ntb_peer_spad_write(qp->ndev, PIDX, QP_LINKS, val & ~BIT(qp->qp_num));
2343
2344	if (qp->link_is_up)
2345		ntb_send_link_down(qp);
2346	else
2347		cancel_delayed_work_sync(&qp->link_work);
2348}
2349EXPORT_SYMBOL_GPL(ntb_transport_link_down);
2350
2351/**
2352 * ntb_transport_link_query - Query transport link state
2353 * @qp: NTB transport layer queue to be queried
2354 *
2355 * Query connectivity to the remote system of the NTB transport queue
2356 *
2357 * RETURNS: true for link up or false for link down
2358 */
2359bool ntb_transport_link_query(struct ntb_transport_qp *qp)
2360{
2361	if (!qp)
2362		return false;
2363
2364	return qp->link_is_up;
2365}
2366EXPORT_SYMBOL_GPL(ntb_transport_link_query);
2367
2368/**
2369 * ntb_transport_qp_num - Query the qp number
2370 * @qp: NTB transport layer queue to be queried
2371 *
2372 * Query qp number of the NTB transport queue
2373 *
2374 * RETURNS: a zero based number specifying the qp number
2375 */
2376unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
2377{
2378	if (!qp)
2379		return 0;
2380
2381	return qp->qp_num;
2382}
2383EXPORT_SYMBOL_GPL(ntb_transport_qp_num);
2384
2385/**
2386 * ntb_transport_max_size - Query the max payload size of a qp
2387 * @qp: NTB transport layer queue to be queried
2388 *
2389 * Query the maximum payload size permissible on the given qp
2390 *
2391 * RETURNS: the max payload size of a qp
2392 */
2393unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp)
2394{
2395	unsigned int max_size;
2396	unsigned int copy_align;
2397	struct dma_chan *rx_chan, *tx_chan;
2398
2399	if (!qp)
2400		return 0;
2401
2402	rx_chan = qp->rx_dma_chan;
2403	tx_chan = qp->tx_dma_chan;
2404
2405	copy_align = max(rx_chan ? rx_chan->device->copy_align : 0,
2406			 tx_chan ? tx_chan->device->copy_align : 0);
2407
2408	/* If DMA engine usage is possible, try to find the max size for that */
2409	max_size = qp->tx_max_frame - sizeof(struct ntb_payload_header);
2410	max_size = round_down(max_size, 1 << copy_align);
2411
2412	return max_size;
2413}
2414EXPORT_SYMBOL_GPL(ntb_transport_max_size);
2415
2416unsigned int ntb_transport_tx_free_entry(struct ntb_transport_qp *qp)
2417{
2418	unsigned int head = qp->tx_index;
2419	unsigned int tail = qp->remote_rx_info->entry;
2420
2421	return tail > head ? tail - head : qp->tx_max_entry + tail - head;
2422}
2423EXPORT_SYMBOL_GPL(ntb_transport_tx_free_entry);
2424
2425static void ntb_transport_doorbell_callback(void *data, int vector)
2426{
2427	struct ntb_transport_ctx *nt = data;
2428	struct ntb_transport_qp *qp;
2429	u64 db_bits;
2430	unsigned int qp_num;
2431
2432	if (ntb_db_read(nt->ndev) & nt->msi_db_mask) {
2433		ntb_transport_msi_peer_desc_changed(nt);
2434		ntb_db_clear(nt->ndev, nt->msi_db_mask);
2435	}
2436
2437	db_bits = (nt->qp_bitmap & ~nt->qp_bitmap_free &
2438		   ntb_db_vector_mask(nt->ndev, vector));
2439
2440	while (db_bits) {
2441		qp_num = __ffs(db_bits);
2442		qp = &nt->qp_vec[qp_num];
2443
2444		if (qp->active)
2445			tasklet_schedule(&qp->rxc_db_work);
2446
2447		db_bits &= ~BIT_ULL(qp_num);
2448	}
2449}
2450
2451static const struct ntb_ctx_ops ntb_transport_ops = {
2452	.link_event = ntb_transport_event_callback,
2453	.db_event = ntb_transport_doorbell_callback,
2454};
2455
2456static struct ntb_client ntb_transport_client = {
2457	.ops = {
2458		.probe = ntb_transport_probe,
2459		.remove = ntb_transport_free,
2460	},
2461};
2462
2463static int __init ntb_transport_init(void)
2464{
2465	int rc;
2466
2467	pr_info("%s, version %s\n", NTB_TRANSPORT_DESC, NTB_TRANSPORT_VER);
2468
2469	if (debugfs_initialized())
2470		nt_debugfs_dir = debugfs_create_dir(KBUILD_MODNAME, NULL);
2471
2472	rc = bus_register(&ntb_transport_bus);
2473	if (rc)
2474		goto err_bus;
2475
2476	rc = ntb_register_client(&ntb_transport_client);
2477	if (rc)
2478		goto err_client;
2479
2480	return 0;
2481
2482err_client:
2483	bus_unregister(&ntb_transport_bus);
2484err_bus:
2485	debugfs_remove_recursive(nt_debugfs_dir);
2486	return rc;
2487}
2488module_init(ntb_transport_init);
2489
2490static void __exit ntb_transport_exit(void)
2491{
2492	ntb_unregister_client(&ntb_transport_client);
2493	bus_unregister(&ntb_transport_bus);
2494	debugfs_remove_recursive(nt_debugfs_dir);
2495}
2496module_exit(ntb_transport_exit);