drivers/ntb/ntb_transport.c

   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/ntb.h"
  62 #include "linux/ntb_transport.h"
  63
  64 #define NTB_TRANSPORT_VERSION   4
  65 #define NTB_TRANSPORT_VER       "4"
  66 #define NTB_TRANSPORT_NAME      "ntb_transport"
  67 #define NTB_TRANSPORT_DESC      "Software Queue-Pair Transport over NTB"
  68
  69 MODULE_DESCRIPTION(NTB_TRANSPORT_DESC);
  70 MODULE_VERSION(NTB_TRANSPORT_VER);
  71 MODULE_LICENSE("Dual BSD/GPL");
  72 MODULE_AUTHOR("Intel Corporation");
  73
  74 static unsigned long max_mw_size;
  75 module_param(max_mw_size, ulong, 0644);
  76 MODULE_PARM_DESC(max_mw_size, "Limit size of large memory windows");
  77
  78 static unsigned int transport_mtu = 0x401E;
  79 module_param(transport_mtu, uint, 0644);
  80 MODULE_PARM_DESC(transport_mtu, "Maximum size of NTB transport packets");
  81
  82 static unsigned char max_num_clients;
  83 module_param(max_num_clients, byte, 0644);
  84 MODULE_PARM_DESC(max_num_clients, "Maximum number of NTB transport clients");
  85
  86 static unsigned int copy_bytes = 1024;
  87 module_param(copy_bytes, uint, 0644);
  88 MODULE_PARM_DESC(copy_bytes, "Threshold under which NTB will use the CPU to copy instead of DMA");
  89
  90 static struct dentry *nt_debugfs_dir;
  91
  92 struct ntb_queue_entry {
  93         /* ntb_queue list reference */
  94         struct list_head entry;
  95         /* pointers to data to be transferred */
  96         void *cb_data;
  97         void *buf;
  98         unsigned int len;
  99         unsigned int flags;
 100
 101         struct ntb_transport_qp *qp;
 102         union {
 103                 struct ntb_payload_header __iomem *tx_hdr;
 104                 struct ntb_payload_header *rx_hdr;
 105         };
 106         unsigned int index;
 107 };
 108
 109 struct ntb_rx_info {
 110         unsigned int entry;
 111 };
 112
 113 struct ntb_transport_qp {
 114         struct ntb_transport_ctx *transport;
 115         struct ntb_dev *ndev;
 116         void *cb_data;
 117         struct dma_chan *dma_chan;
 118
 119         bool client_ready;
 120         bool link_is_up;
 121
 122         u8 qp_num;      /* Only 64 QP's are allowed.  0-63 */
 123         u64 qp_bit;
 124
 125         struct ntb_rx_info __iomem *rx_info;
 126         struct ntb_rx_info *remote_rx_info;
 127
 128         void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data,
 129                            void *data, int len);
 130         struct list_head tx_free_q;
 131         spinlock_t ntb_tx_free_q_lock;
 132         void __iomem *tx_mw;
 133         dma_addr_t tx_mw_phys;
 134         unsigned int tx_index;
 135         unsigned int tx_max_entry;
 136         unsigned int tx_max_frame;
 137
 138         void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
 139                            void *data, int len);
 140         struct list_head rx_pend_q;
 141         struct list_head rx_free_q;
 142         spinlock_t ntb_rx_pend_q_lock;
 143         spinlock_t ntb_rx_free_q_lock;
 144         void *rx_buff;
 145         unsigned int rx_index;
 146         unsigned int rx_max_entry;
 147         unsigned int rx_max_frame;
 148         dma_cookie_t last_cookie;
 149         struct tasklet_struct rxc_db_work;
 150
 151         void (*event_handler)(void *data, int status);
 152         struct delayed_work link_work;
 153         struct work_struct link_cleanup;
 154
 155         struct dentry *debugfs_dir;
 156         struct dentry *debugfs_stats;
 157
 158         /* Stats */
 159         u64 rx_bytes;
 160         u64 rx_pkts;
 161         u64 rx_ring_empty;
 162         u64 rx_err_no_buf;
 163         u64 rx_err_oflow;
 164         u64 rx_err_ver;
 165         u64 rx_memcpy;
 166         u64 rx_async;
 167         u64 tx_bytes;
 168         u64 tx_pkts;
 169         u64 tx_ring_full;
 170         u64 tx_err_no_buf;
 171         u64 tx_memcpy;
 172         u64 tx_async;
 173 };
 174
 175 struct ntb_transport_mw {
 176         phys_addr_t phys_addr;
 177         resource_size_t phys_size;
 178         resource_size_t xlat_align;
 179         resource_size_t xlat_align_size;
 180         void __iomem *vbase;
 181         size_t xlat_size;
 182         size_t buff_size;
 183         void *virt_addr;
 184         dma_addr_t dma_addr;
 185 };
 186
 187 struct ntb_transport_client_dev {
 188         struct list_head entry;
 189         struct ntb_transport_ctx *nt;
 190         struct device dev;
 191 };
 192
 193 struct ntb_transport_ctx {
 194         struct list_head entry;
 195         struct list_head client_devs;
 196
 197         struct ntb_dev *ndev;
 198
 199         struct ntb_transport_mw *mw_vec;
 200         struct ntb_transport_qp *qp_vec;
 201         unsigned int mw_count;
 202         unsigned int qp_count;
 203         u64 qp_bitmap;
 204         u64 qp_bitmap_free;
 205
 206         bool link_is_up;
 207         struct delayed_work link_work;
 208         struct work_struct link_cleanup;
 209 };
 210
 211 enum {
 212         DESC_DONE_FLAG = BIT(0),
 213         LINK_DOWN_FLAG = BIT(1),
 214 };
 215
 216 struct ntb_payload_header {
 217         unsigned int ver;
 218         unsigned int len;
 219         unsigned int flags;
 220 };
 221
 222 enum {
 223         VERSION = 0,
 224         QP_LINKS,
 225         NUM_QPS,
 226         NUM_MWS,
 227         MW0_SZ_HIGH,
 228         MW0_SZ_LOW,
 229         MW1_SZ_HIGH,
 230         MW1_SZ_LOW,
 231         MAX_SPAD,
 232 };
 233
 234 #define dev_client_dev(__dev) \
 235         container_of((__dev), struct ntb_transport_client_dev, dev)
 236
 237 #define drv_client(__drv) \
 238         container_of((__drv), struct ntb_transport_client, driver)
 239
 240 #define QP_TO_MW(nt, qp)        ((qp) % nt->mw_count)
 241 #define NTB_QP_DEF_NUM_ENTRIES  100
 242 #define NTB_LINK_DOWN_TIMEOUT   10
 243
 244 static void ntb_transport_rxc_db(unsigned long data);
 245 static const struct ntb_ctx_ops ntb_transport_ops;
 246 static struct ntb_client ntb_transport_client;
 247
 248 static int ntb_transport_bus_match(struct device *dev,
 249                                    struct device_driver *drv)
 250 {
 251         return !strncmp(dev_name(dev), drv->name, strlen(drv->name));
 252 }
 253
 254 static int ntb_transport_bus_probe(struct device *dev)
 255 {
 256         const struct ntb_transport_client *client;
 257         int rc = -EINVAL;
 258
 259         get_device(dev);
 260
 261         client = drv_client(dev->driver);
 262         rc = client->probe(dev);
 263         if (rc)
 264                 put_device(dev);
 265
 266         return rc;
 267 }
 268
 269 static int ntb_transport_bus_remove(struct device *dev)
 270 {
 271         const struct ntb_transport_client *client;
 272
 273         client = drv_client(dev->driver);
 274         client->remove(dev);
 275
 276         put_device(dev);
 277
 278         return 0;
 279 }
 280
 281 static struct bus_type ntb_transport_bus = {
 282         .name = "ntb_transport",
 283         .match = ntb_transport_bus_match,
 284         .probe = ntb_transport_bus_probe,
 285         .remove = ntb_transport_bus_remove,
 286 };
 287
 288 static LIST_HEAD(ntb_transport_list);
 289
 290 static int ntb_bus_init(struct ntb_transport_ctx *nt)
 291 {
 292         list_add(&nt->entry, &ntb_transport_list);
 293         return 0;
 294 }
 295
 296 static void ntb_bus_remove(struct ntb_transport_ctx *nt)
 297 {
 298         struct ntb_transport_client_dev *client_dev, *cd;
 299
 300         list_for_each_entry_safe(client_dev, cd, &nt->client_devs, entry) {
 301                 dev_err(client_dev->dev.parent, "%s still attached to bus, removing\n",
 302                         dev_name(&client_dev->dev));
 303                 list_del(&client_dev->entry);
 304                 device_unregister(&client_dev->dev);
 305         }
 306
 307         list_del(&nt->entry);
 308 }
 309
 310 static void ntb_transport_client_release(struct device *dev)
 311 {
 312         struct ntb_transport_client_dev *client_dev;
 313
 314         client_dev = dev_client_dev(dev);
 315         kfree(client_dev);
 316 }
 317
 318 /**
 319  * ntb_transport_unregister_client_dev - Unregister NTB client device
 320  * @device_name: Name of NTB client device
 321  *
 322  * Unregister an NTB client device with the NTB transport layer
 323  */
 324 void ntb_transport_unregister_client_dev(char *device_name)
 325 {
 326         struct ntb_transport_client_dev *client, *cd;
 327         struct ntb_transport_ctx *nt;
 328
 329         list_for_each_entry(nt, &ntb_transport_list, entry)
 330                 list_for_each_entry_safe(client, cd, &nt->client_devs, entry)
 331                         if (!strncmp(dev_name(&client->dev), device_name,
 332                                      strlen(device_name))) {
 333                                 list_del(&client->entry);
 334                                 device_unregister(&client->dev);
 335                         }
 336 }
 337 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client_dev);
 338
 339 /**
 340  * ntb_transport_register_client_dev - Register NTB client device
 341  * @device_name: Name of NTB client device
 342  *
 343  * Register an NTB client device with the NTB transport layer
 344  */
 345 int ntb_transport_register_client_dev(char *device_name)
 346 {
 347         struct ntb_transport_client_dev *client_dev;
 348         struct ntb_transport_ctx *nt;
 349         int rc, i = 0;
 350
 351         if (list_empty(&ntb_transport_list))
 352                 return -ENODEV;
 353
 354         list_for_each_entry(nt, &ntb_transport_list, entry) {
 355                 struct device *dev;
 356
 357                 client_dev = kzalloc(sizeof(*client_dev),
 358                                      GFP_KERNEL);
 359                 if (!client_dev) {
 360                         rc = -ENOMEM;
 361                         goto err;
 362                 }
 363
 364                 dev = &client_dev->dev;
 365
 366                 /* setup and register client devices */
 367                 dev_set_name(dev, "%s%d", device_name, i);
 368                 dev->bus = &ntb_transport_bus;
 369                 dev->release = ntb_transport_client_release;
 370                 dev->parent = &nt->ndev->dev;
 371
 372                 rc = device_register(dev);
 373                 if (rc) {
 374                         kfree(client_dev);
 375                         goto err;
 376                 }
 377
 378                 list_add_tail(&client_dev->entry, &nt->client_devs);
 379                 i++;
 380         }
 381
 382         return 0;
 383
 384 err:
 385         ntb_transport_unregister_client_dev(device_name);
 386
 387         return rc;
 388 }
 389 EXPORT_SYMBOL_GPL(ntb_transport_register_client_dev);
 390
 391 /**
 392  * ntb_transport_register_client - Register NTB client driver
 393  * @drv: NTB client driver to be registered
 394  *
 395  * Register an NTB client driver with the NTB transport layer
 396  *
 397  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 398  */
 399 int ntb_transport_register_client(struct ntb_transport_client *drv)
 400 {
 401         drv->driver.bus = &ntb_transport_bus;
 402
 403         if (list_empty(&ntb_transport_list))
 404                 return -ENODEV;
 405
 406         return driver_register(&drv->driver);
 407 }
 408 EXPORT_SYMBOL_GPL(ntb_transport_register_client);
 409
 410 /**
 411  * ntb_transport_unregister_client - Unregister NTB client driver
 412  * @drv: NTB client driver to be unregistered
 413  *
 414  * Unregister an NTB client driver with the NTB transport layer
 415  *
 416  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 417  */
 418 void ntb_transport_unregister_client(struct ntb_transport_client *drv)
 419 {
 420         driver_unregister(&drv->driver);
 421 }
 422 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client);
 423
 424 static ssize_t debugfs_read(struct file *filp, char __user *ubuf, size_t count,
 425                             loff_t *offp)
 426 {
 427         struct ntb_transport_qp *qp;
 428         char *buf;
 429         ssize_t ret, out_offset, out_count;
 430
 431         out_count = 1000;
 432
 433         buf = kmalloc(out_count, GFP_KERNEL);
 434         if (!buf)
 435                 return -ENOMEM;
 436
 437         qp = filp->private_data;
 438         out_offset = 0;
 439         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 440                                "NTB QP stats\n");
 441         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 442                                "rx_bytes - \t%llu\n", qp->rx_bytes);
 443         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 444                                "rx_pkts - \t%llu\n", qp->rx_pkts);
 445         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 446                                "rx_memcpy - \t%llu\n", qp->rx_memcpy);
 447         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 448                                "rx_async - \t%llu\n", qp->rx_async);
 449         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 450                                "rx_ring_empty - %llu\n", qp->rx_ring_empty);
 451         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 452                                "rx_err_no_buf - %llu\n", qp->rx_err_no_buf);
 453         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 454                                "rx_err_oflow - \t%llu\n", qp->rx_err_oflow);
 455         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 456                                "rx_err_ver - \t%llu\n", qp->rx_err_ver);
 457         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 458                                "rx_buff - \t%p\n", qp->rx_buff);
 459         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 460                                "rx_index - \t%u\n", qp->rx_index);
 461         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 462                                "rx_max_entry - \t%u\n", qp->rx_max_entry);
 463
 464         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 465                                "tx_bytes - \t%llu\n", qp->tx_bytes);
 466         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 467                                "tx_pkts - \t%llu\n", qp->tx_pkts);
 468         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 469                                "tx_memcpy - \t%llu\n", qp->tx_memcpy);
 470         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 471                                "tx_async - \t%llu\n", qp->tx_async);
 472         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 473                                "tx_ring_full - \t%llu\n", qp->tx_ring_full);
 474         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 475                                "tx_err_no_buf - %llu\n", qp->tx_err_no_buf);
 476         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 477                                "tx_mw - \t%p\n", qp->tx_mw);
 478         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 479                                "tx_index - \t%u\n", qp->tx_index);
 480         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 481                                "tx_max_entry - \t%u\n", qp->tx_max_entry);
 482
 483         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 484                                "\nQP Link %s\n",
 485                                qp->link_is_up ? "Up" : "Down");
 486         if (out_offset > out_count)
 487                 out_offset = out_count;
 488
 489         ret = simple_read_from_buffer(ubuf, count, offp, buf, out_offset);
 490         kfree(buf);
 491         return ret;
 492 }
 493
 494 static const struct file_operations ntb_qp_debugfs_stats = {
 495         .owner = THIS_MODULE,
 496         .open = simple_open,
 497         .read = debugfs_read,
 498 };
 499
 500 static void ntb_list_add(spinlock_t *lock, struct list_head *entry,
 501                          struct list_head *list)
 502 {
 503         unsigned long flags;
 504
 505         spin_lock_irqsave(lock, flags);
 506         list_add_tail(entry, list);
 507         spin_unlock_irqrestore(lock, flags);
 508 }
 509
 510 static struct ntb_queue_entry *ntb_list_rm(spinlock_t *lock,
 511                                            struct list_head *list)
 512 {
 513         struct ntb_queue_entry *entry;
 514         unsigned long flags;
 515
 516         spin_lock_irqsave(lock, flags);
 517         if (list_empty(list)) {
 518                 entry = NULL;
 519                 goto out;
 520         }
 521         entry = list_first_entry(list, struct ntb_queue_entry, entry);
 522         list_del(&entry->entry);
 523 out:
 524         spin_unlock_irqrestore(lock, flags);
 525
 526         return entry;
 527 }
 528
 529 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
 530                                      unsigned int qp_num)
 531 {
 532         struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
 533         struct ntb_transport_mw *mw;
 534         unsigned int rx_size, num_qps_mw;
 535         unsigned int mw_num, mw_count, qp_count;
 536         unsigned int i;
 537
 538         mw_count = nt->mw_count;
 539         qp_count = nt->qp_count;
 540
 541         mw_num = QP_TO_MW(nt, qp_num);
 542         mw = &nt->mw_vec[mw_num];
 543
 544         if (!mw->virt_addr)
 545                 return -ENOMEM;
 546
 547         if (qp_count % mw_count && mw_num + 1 < qp_count / mw_count)
 548                 num_qps_mw = qp_count / mw_count + 1;
 549         else
 550                 num_qps_mw = qp_count / mw_count;
 551
 552         rx_size = (unsigned int)mw->xlat_size / num_qps_mw;
 553         qp->rx_buff = mw->virt_addr + rx_size * qp_num / mw_count;
 554         rx_size -= sizeof(struct ntb_rx_info);
 555
 556         qp->remote_rx_info = qp->rx_buff + rx_size;
 557
 558         /* Due to housekeeping, there must be atleast 2 buffs */
 559         qp->rx_max_frame = min(transport_mtu, rx_size / 2);
 560         qp->rx_max_entry = rx_size / qp->rx_max_frame;
 561         qp->rx_index = 0;
 562
 563         qp->remote_rx_info->entry = qp->rx_max_entry - 1;
 564
 565         /* setup the hdr offsets with 0's */
 566         for (i = 0; i < qp->rx_max_entry; i++) {
 567                 void *offset = (qp->rx_buff + qp->rx_max_frame * (i + 1) -
 568                                 sizeof(struct ntb_payload_header));
 569                 memset(offset, 0, sizeof(struct ntb_payload_header));
 570         }
 571
 572         qp->rx_pkts = 0;
 573         qp->tx_pkts = 0;
 574         qp->tx_index = 0;
 575
 576         return 0;
 577 }
 578
 579 static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
 580 {
 581         struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
 582         struct pci_dev *pdev = nt->ndev->pdev;
 583
 584         if (!mw->virt_addr)
 585                 return;
 586
 587         ntb_mw_clear_trans(nt->ndev, num_mw);
 588         dma_free_coherent(&pdev->dev, mw->buff_size,
 589                           mw->virt_addr, mw->dma_addr);
 590         mw->xlat_size = 0;
 591         mw->buff_size = 0;
 592         mw->virt_addr = NULL;
 593 }
 594
 595 static int ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw,
 596                       unsigned int size)
 597 {
 598         struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
 599         struct pci_dev *pdev = nt->ndev->pdev;
 600         unsigned int xlat_size, buff_size;
 601         int rc;
 602
 603         xlat_size = round_up(size, mw->xlat_align_size);
 604         buff_size = round_up(size, mw->xlat_align);
 605
 606         /* No need to re-setup */
 607         if (mw->xlat_size == xlat_size)
 608                 return 0;
 609
 610         if (mw->buff_size)
 611                 ntb_free_mw(nt, num_mw);
 612
 613         /* Alloc memory for receiving data.  Must be aligned */
 614         mw->xlat_size = xlat_size;
 615         mw->buff_size = buff_size;
 616
 617         mw->virt_addr = dma_alloc_coherent(&pdev->dev, buff_size,
 618                                            &mw->dma_addr, GFP_KERNEL);
 619         if (!mw->virt_addr) {
 620                 mw->xlat_size = 0;
 621                 mw->buff_size = 0;
 622                 dev_err(&pdev->dev, "Unable to alloc MW buff of size %d\n",
 623                         buff_size);
 624                 return -ENOMEM;
 625         }
 626
 627         /*
 628          * we must ensure that the memory address allocated is BAR size
 629          * aligned in order for the XLAT register to take the value. This
 630          * is a requirement of the hardware. It is recommended to setup CMA
 631          * for BAR sizes equal or greater than 4MB.
 632          */
 633         if (!IS_ALIGNED(mw->dma_addr, mw->xlat_align)) {
 634                 dev_err(&pdev->dev, "DMA memory %pad is not aligned\n",
 635                         &mw->dma_addr);
 636                 ntb_free_mw(nt, num_mw);
 637                 return -ENOMEM;
 638         }
 639
 640         /* Notify HW the memory location of the receive buffer */
 641         rc = ntb_mw_set_trans(nt->ndev, num_mw, mw->dma_addr, mw->xlat_size);
 642         if (rc) {
 643                 dev_err(&pdev->dev, "Unable to set mw%d translation", num_mw);
 644                 ntb_free_mw(nt, num_mw);
 645                 return -EIO;
 646         }
 647
 648         return 0;
 649 }
 650
 651 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
 652 {
 653         struct ntb_transport_ctx *nt = qp->transport;
 654         struct pci_dev *pdev = nt->ndev->pdev;
 655
 656         if (qp->link_is_up) {
 657                 cancel_delayed_work_sync(&qp->link_work);
 658                 return;
 659         }
 660
 661         dev_info(&pdev->dev, "qp %d: Link Down\n", qp->qp_num);
 662         qp->link_is_up = false;
 663
 664         if (qp->event_handler)
 665                 qp->event_handler(qp->cb_data, qp->link_is_up);
 666 }
 667
 668 static void ntb_qp_link_cleanup_work(struct work_struct *work)
 669 {
 670         struct ntb_transport_qp *qp = container_of(work,
 671                                                    struct ntb_transport_qp,
 672                                                    link_cleanup);
 673         struct ntb_transport_ctx *nt = qp->transport;
 674
 675         ntb_qp_link_cleanup(qp);
 676
 677         if (nt->link_is_up)
 678                 schedule_delayed_work(&qp->link_work,
 679                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 680 }
 681
 682 static void ntb_qp_link_down(struct ntb_transport_qp *qp)
 683 {
 684         schedule_work(&qp->link_cleanup);
 685 }
 686
 687 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
 688 {
 689         struct ntb_transport_qp *qp;
 690         u64 qp_bitmap_alloc;
 691         int i;
 692
 693         qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
 694
 695         /* Pass along the info to any clients */
 696         for (i = 0; i < nt->qp_count; i++)
 697                 if (qp_bitmap_alloc & BIT_ULL(i)) {
 698                         qp = &nt->qp_vec[i];
 699                         ntb_qp_link_cleanup(qp);
 700                         cancel_work_sync(&qp->link_cleanup);
 701                         cancel_delayed_work_sync(&qp->link_work);
 702                 }
 703
 704         if (!nt->link_is_up)
 705                 cancel_delayed_work_sync(&nt->link_work);
 706
 707         /* The scratchpad registers keep the values if the remote side
 708          * goes down, blast them now to give them a sane value the next
 709          * time they are accessed
 710          */
 711         for (i = 0; i < MAX_SPAD; i++)
 712                 ntb_spad_write(nt->ndev, i, 0);
 713 }
 714
 715 static void ntb_transport_link_cleanup_work(struct work_struct *work)
 716 {
 717         struct ntb_transport_ctx *nt =
 718                 container_of(work, struct ntb_transport_ctx, link_cleanup);
 719
 720         ntb_transport_link_cleanup(nt);
 721 }
 722
 723 static void ntb_transport_event_callback(void *data)
 724 {
 725         struct ntb_transport_ctx *nt = data;
 726
 727         if (ntb_link_is_up(nt->ndev, NULL, NULL) == 1)
 728                 schedule_delayed_work(&nt->link_work, 0);
 729         else
 730                 schedule_work(&nt->link_cleanup);
 731 }
 732
 733 static void ntb_transport_link_work(struct work_struct *work)
 734 {
 735         struct ntb_transport_ctx *nt =
 736                 container_of(work, struct ntb_transport_ctx, link_work.work);
 737         struct ntb_dev *ndev = nt->ndev;
 738         struct pci_dev *pdev = ndev->pdev;
 739         resource_size_t size;
 740         u32 val;
 741         int rc, i, spad;
 742
 743         /* send the local info, in the opposite order of the way we read it */
 744         for (i = 0; i < nt->mw_count; i++) {
 745                 size = nt->mw_vec[i].phys_size;
 746
 747                 if (max_mw_size && size > max_mw_size)
 748                         size = max_mw_size;
 749
 750                 spad = MW0_SZ_HIGH + (i * 2);
 751                 ntb_peer_spad_write(ndev, spad, (u32)(size >> 32));
 752
 753                 spad = MW0_SZ_LOW + (i * 2);
 754                 ntb_peer_spad_write(ndev, spad, (u32)size);
 755         }
 756
 757         ntb_peer_spad_write(ndev, NUM_MWS, nt->mw_count);
 758
 759         ntb_peer_spad_write(ndev, NUM_QPS, nt->qp_count);
 760
 761         ntb_peer_spad_write(ndev, VERSION, NTB_TRANSPORT_VERSION);
 762
 763         /* Query the remote side for its info */
 764         val = ntb_peer_spad_read(ndev, VERSION);
 765         dev_dbg(&pdev->dev, "Remote version = %d\n", val);
 766         if (val != NTB_TRANSPORT_VERSION)
 767                 goto out;
 768
 769         val = ntb_peer_spad_read(ndev, NUM_QPS);
 770         dev_dbg(&pdev->dev, "Remote max number of qps = %d\n", val);
 771         if (val != nt->qp_count)
 772                 goto out;
 773
 774         val = ntb_peer_spad_read(ndev, NUM_MWS);
 775         dev_dbg(&pdev->dev, "Remote number of mws = %d\n", val);
 776         if (val != nt->mw_count)
 777                 goto out;
 778
 779         for (i = 0; i < nt->mw_count; i++) {
 780                 u64 val64;
 781
 782                 val = ntb_peer_spad_read(ndev, MW0_SZ_HIGH + (i * 2));
 783                 val64 = (u64)val << 32;
 784
 785                 val = ntb_peer_spad_read(ndev, MW0_SZ_LOW + (i * 2));
 786                 val64 |= val;
 787
 788                 dev_dbg(&pdev->dev, "Remote MW%d size = %#llx\n", i, val64);
 789
 790                 rc = ntb_set_mw(nt, i, val64);
 791                 if (rc)
 792                         goto out1;
 793         }
 794
 795         nt->link_is_up = true;
 796
 797         for (i = 0; i < nt->qp_count; i++) {
 798                 struct ntb_transport_qp *qp = &nt->qp_vec[i];
 799
 800                 ntb_transport_setup_qp_mw(nt, i);
 801
 802                 if (qp->client_ready)
 803                         schedule_delayed_work(&qp->link_work, 0);
 804         }
 805
 806         return;
 807
 808 out1:
 809         for (i = 0; i < nt->mw_count; i++)
 810                 ntb_free_mw(nt, i);
 811 out:
 812         if (ntb_link_is_up(ndev, NULL, NULL) == 1)
 813                 schedule_delayed_work(&nt->link_work,
 814                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 815 }
 816
 817 static void ntb_qp_link_work(struct work_struct *work)
 818 {
 819         struct ntb_transport_qp *qp = container_of(work,
 820                                                    struct ntb_transport_qp,
 821                                                    link_work.work);
 822         struct pci_dev *pdev = qp->ndev->pdev;
 823         struct ntb_transport_ctx *nt = qp->transport;
 824         int val;
 825
 826         WARN_ON(!nt->link_is_up);
 827
 828         val = ntb_spad_read(nt->ndev, QP_LINKS);
 829
 830         ntb_peer_spad_write(nt->ndev, QP_LINKS, val | BIT(qp->qp_num));
 831
 832         /* query remote spad for qp ready bits */
 833         ntb_peer_spad_read(nt->ndev, QP_LINKS);
 834         dev_dbg(&pdev->dev, "Remote QP link status = %x\n", val);
 835
 836         /* See if the remote side is up */
 837         if (val & BIT(qp->qp_num)) {
 838                 dev_info(&pdev->dev, "qp %d: Link Up\n", qp->qp_num);
 839                 qp->link_is_up = true;
 840
 841                 if (qp->event_handler)
 842                         qp->event_handler(qp->cb_data, qp->link_is_up);
 843         } else if (nt->link_is_up)
 844                 schedule_delayed_work(&qp->link_work,
 845                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 846 }
 847
 848 static int ntb_transport_init_queue(struct ntb_transport_ctx *nt,
 849                                     unsigned int qp_num)
 850 {
 851         struct ntb_transport_qp *qp;
 852         struct ntb_transport_mw *mw;
 853         phys_addr_t mw_base;
 854         resource_size_t mw_size;
 855         unsigned int num_qps_mw, tx_size;
 856         unsigned int mw_num, mw_count, qp_count;
 857         u64 qp_offset;
 858
 859         mw_count = nt->mw_count;
 860         qp_count = nt->qp_count;
 861
 862         mw_num = QP_TO_MW(nt, qp_num);
 863         mw = &nt->mw_vec[mw_num];
 864
 865         qp = &nt->qp_vec[qp_num];
 866         qp->qp_num = qp_num;
 867         qp->transport = nt;
 868         qp->ndev = nt->ndev;
 869         qp->link_is_up = false;
 870         qp->client_ready = false;
 871         qp->event_handler = NULL;
 872
 873         if (qp_count % mw_count && mw_num + 1 < qp_count / mw_count)
 874                 num_qps_mw = qp_count / mw_count + 1;
 875         else
 876                 num_qps_mw = qp_count / mw_count;
 877
 878         mw_base = nt->mw_vec[mw_num].phys_addr;
 879         mw_size = nt->mw_vec[mw_num].phys_size;
 880
 881         tx_size = (unsigned int)mw_size / num_qps_mw;
 882         qp_offset = tx_size * qp_num / mw_count;
 883
 884         qp->tx_mw = nt->mw_vec[mw_num].vbase + qp_offset;
 885         if (!qp->tx_mw)
 886                 return -EINVAL;
 887
 888         qp->tx_mw_phys = mw_base + qp_offset;
 889         if (!qp->tx_mw_phys)
 890                 return -EINVAL;
 891
 892         tx_size -= sizeof(struct ntb_rx_info);
 893         qp->rx_info = qp->tx_mw + tx_size;
 894
 895         /* Due to housekeeping, there must be atleast 2 buffs */
 896         qp->tx_max_frame = min(transport_mtu, tx_size / 2);
 897         qp->tx_max_entry = tx_size / qp->tx_max_frame;
 898
 899         if (nt_debugfs_dir) {
 900                 char debugfs_name[4];
 901
 902                 snprintf(debugfs_name, 4, "qp%d", qp_num);
 903                 qp->debugfs_dir = debugfs_create_dir(debugfs_name,
 904                                                      nt_debugfs_dir);
 905
 906                 qp->debugfs_stats = debugfs_create_file("stats", S_IRUSR,
 907                                                         qp->debugfs_dir, qp,
 908                                                         &ntb_qp_debugfs_stats);
 909         } else {
 910                 qp->debugfs_dir = NULL;
 911                 qp->debugfs_stats = NULL;
 912         }
 913
 914         INIT_DELAYED_WORK(&qp->link_work, ntb_qp_link_work);
 915         INIT_WORK(&qp->link_cleanup, ntb_qp_link_cleanup_work);
 916
 917         spin_lock_init(&qp->ntb_rx_pend_q_lock);
 918         spin_lock_init(&qp->ntb_rx_free_q_lock);
 919         spin_lock_init(&qp->ntb_tx_free_q_lock);
 920
 921         INIT_LIST_HEAD(&qp->rx_pend_q);
 922         INIT_LIST_HEAD(&qp->rx_free_q);
 923         INIT_LIST_HEAD(&qp->tx_free_q);
 924
 925         tasklet_init(&qp->rxc_db_work, ntb_transport_rxc_db,
 926                      (unsigned long)qp);
 927
 928         return 0;
 929 }
 930
 931 static int ntb_transport_probe(struct ntb_client *self, struct ntb_dev *ndev)
 932 {
 933         struct ntb_transport_ctx *nt;
 934         struct ntb_transport_mw *mw;
 935         unsigned int mw_count, qp_count;
 936         u64 qp_bitmap;
 937         int rc, i;
 938
 939         if (ntb_db_is_unsafe(ndev))
 940                 dev_dbg(&ndev->dev,
 941                         "doorbell is unsafe, proceed anyway...\n");
 942         if (ntb_spad_is_unsafe(ndev))
 943                 dev_dbg(&ndev->dev,
 944                         "scratchpad is unsafe, proceed anyway...\n");
 945
 946         nt = kzalloc(sizeof(*nt), GFP_KERNEL);
 947         if (!nt)
 948                 return -ENOMEM;
 949
 950         nt->ndev = ndev;
 951
 952         mw_count = ntb_mw_count(ndev);
 953
 954         nt->mw_count = mw_count;
 955
 956         nt->mw_vec = kcalloc(mw_count, sizeof(*nt->mw_vec), GFP_KERNEL);
 957         if (!nt->mw_vec) {
 958                 rc = -ENOMEM;
 959                 goto err;
 960         }
 961
 962         for (i = 0; i < mw_count; i++) {
 963                 mw = &nt->mw_vec[i];
 964
 965                 rc = ntb_mw_get_range(ndev, i, &mw->phys_addr, &mw->phys_size,
 966                                       &mw->xlat_align, &mw->xlat_align_size);
 967                 if (rc)
 968                         goto err1;
 969
 970                 mw->vbase = ioremap(mw->phys_addr, mw->phys_size);
 971                 if (!mw->vbase) {
 972                         rc = -ENOMEM;
 973                         goto err1;
 974                 }
 975
 976                 mw->buff_size = 0;
 977                 mw->xlat_size = 0;
 978                 mw->virt_addr = NULL;
 979                 mw->dma_addr = 0;
 980         }
 981
 982         qp_bitmap = ntb_db_valid_mask(ndev);
 983
 984         qp_count = ilog2(qp_bitmap);
 985         if (max_num_clients && max_num_clients < qp_count)
 986                 qp_count = max_num_clients;
 987         else if (mw_count < qp_count)
 988                 qp_count = mw_count;
 989
 990         qp_bitmap &= BIT_ULL(qp_count) - 1;
 991
 992         nt->qp_count = qp_count;
 993         nt->qp_bitmap = qp_bitmap;
 994         nt->qp_bitmap_free = qp_bitmap;
 995
 996         nt->qp_vec = kcalloc(qp_count, sizeof(*nt->qp_vec), GFP_KERNEL);
 997         if (!nt->qp_vec) {
 998                 rc = -ENOMEM;
 999                 goto err2;
1000         }
1001
1002         for (i = 0; i < qp_count; i++) {
1003                 rc = ntb_transport_init_queue(nt, i);
1004                 if (rc)
1005                         goto err3;
1006         }
1007
1008         INIT_DELAYED_WORK(&nt->link_work, ntb_transport_link_work);
1009         INIT_WORK(&nt->link_cleanup, ntb_transport_link_cleanup_work);
1010
1011         rc = ntb_set_ctx(ndev, nt, &ntb_transport_ops);
1012         if (rc)
1013                 goto err3;
1014
1015         INIT_LIST_HEAD(&nt->client_devs);
1016         rc = ntb_bus_init(nt);
1017         if (rc)
1018                 goto err4;
1019
1020         nt->link_is_up = false;
1021         ntb_link_enable(ndev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
1022         ntb_link_event(ndev);
1023
1024         return 0;
1025
1026 err4:
1027         ntb_clear_ctx(ndev);
1028 err3:
1029         kfree(nt->qp_vec);
1030 err2:
1031         kfree(nt->mw_vec);
1032 err1:
1033         while (i--) {
1034                 mw = &nt->mw_vec[i];
1035                 iounmap(mw->vbase);
1036         }
1037 err:
1038         kfree(nt);
1039         return rc;
1040 }
1041
1042 static void ntb_transport_free(struct ntb_client *self, struct ntb_dev *ndev)
1043 {
1044         struct ntb_transport_ctx *nt = ndev->ctx;
1045         struct ntb_transport_qp *qp;
1046         u64 qp_bitmap_alloc;
1047         int i;
1048
1049         ntb_transport_link_cleanup(nt);
1050         cancel_work_sync(&nt->link_cleanup);
1051         cancel_delayed_work_sync(&nt->link_work);
1052
1053         qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
1054
1055         /* verify that all the qp's are freed */
1056         for (i = 0; i < nt->qp_count; i++) {
1057                 qp = &nt->qp_vec[i];
1058                 if (qp_bitmap_alloc & BIT_ULL(i))
1059                         ntb_transport_free_queue(qp);
1060                 debugfs_remove_recursive(qp->debugfs_dir);
1061         }
1062
1063         ntb_link_disable(ndev);
1064         ntb_clear_ctx(ndev);
1065
1066         ntb_bus_remove(nt);
1067
1068         for (i = nt->mw_count; i--; ) {
1069                 ntb_free_mw(nt, i);
1070                 iounmap(nt->mw_vec[i].vbase);
1071         }
1072
1073         kfree(nt->qp_vec);
1074         kfree(nt->mw_vec);
1075         kfree(nt);
1076 }
1077
1078 static void ntb_rx_copy_callback(void *data)
1079 {
1080         struct ntb_queue_entry *entry = data;
1081         struct ntb_transport_qp *qp = entry->qp;
1082         void *cb_data = entry->cb_data;
1083         unsigned int len = entry->len;
1084         struct ntb_payload_header *hdr = entry->rx_hdr;
1085
1086         hdr->flags = 0;
1087
1088         iowrite32(entry->index, &qp->rx_info->entry);
1089
1090         ntb_list_add(&qp->ntb_rx_free_q_lock, &entry->entry, &qp->rx_free_q);
1091
1092         if (qp->rx_handler && qp->client_ready)
1093                 qp->rx_handler(qp, qp->cb_data, cb_data, len);
1094 }
1095
1096 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset)
1097 {
1098         void *buf = entry->buf;
1099         size_t len = entry->len;
1100
1101         memcpy(buf, offset, len);
1102
1103         /* Ensure that the data is fully copied out before clearing the flag */
1104         wmb();
1105
1106         ntb_rx_copy_callback(entry);
1107 }
1108
1109 static void ntb_async_rx(struct ntb_queue_entry *entry, void *offset,
1110                          size_t len)
1111 {
1112         struct dma_async_tx_descriptor *txd;
1113         struct ntb_transport_qp *qp = entry->qp;
1114         struct dma_chan *chan = qp->dma_chan;
1115         struct dma_device *device;
1116         size_t pay_off, buff_off;
1117         struct dmaengine_unmap_data *unmap;
1118         dma_cookie_t cookie;
1119         void *buf = entry->buf;
1120
1121         entry->len = len;
1122
1123         if (!chan)
1124                 goto err;
1125
1126         if (len < copy_bytes)
1127                 goto err_wait;
1128
1129         device = chan->device;
1130         pay_off = (size_t)offset & ~PAGE_MASK;
1131         buff_off = (size_t)buf & ~PAGE_MASK;
1132
1133         if (!is_dma_copy_aligned(device, pay_off, buff_off, len))
1134                 goto err_wait;
1135
1136         unmap = dmaengine_get_unmap_data(device->dev, 2, GFP_NOWAIT);
1137         if (!unmap)
1138                 goto err_wait;
1139
1140         unmap->len = len;
1141         unmap->addr[0] = dma_map_page(device->dev, virt_to_page(offset),
1142                                       pay_off, len, DMA_TO_DEVICE);
1143         if (dma_mapping_error(device->dev, unmap->addr[0]))
1144                 goto err_get_unmap;
1145
1146         unmap->to_cnt = 1;
1147
1148         unmap->addr[1] = dma_map_page(device->dev, virt_to_page(buf),
1149                                       buff_off, len, DMA_FROM_DEVICE);
1150         if (dma_mapping_error(device->dev, unmap->addr[1]))
1151                 goto err_get_unmap;
1152
1153         unmap->from_cnt = 1;
1154
1155         txd = device->device_prep_dma_memcpy(chan, unmap->addr[1],
1156                                              unmap->addr[0], len,
1157                                              DMA_PREP_INTERRUPT);
1158         if (!txd)
1159                 goto err_get_unmap;
1160
1161         txd->callback = ntb_rx_copy_callback;
1162         txd->callback_param = entry;
1163         dma_set_unmap(txd, unmap);
1164
1165         cookie = dmaengine_submit(txd);
1166         if (dma_submit_error(cookie))
1167                 goto err_set_unmap;
1168
1169         dmaengine_unmap_put(unmap);
1170
1171         qp->last_cookie = cookie;
1172
1173         qp->rx_async++;
1174
1175         return;
1176
1177 err_set_unmap:
1178         dmaengine_unmap_put(unmap);
1179 err_get_unmap:
1180         dmaengine_unmap_put(unmap);
1181 err_wait:
1182         /* If the callbacks come out of order, the writing of the index to the
1183          * last completed will be out of order.  This may result in the
1184          * receive stalling forever.
1185          */
1186         dma_sync_wait(chan, qp->last_cookie);
1187 err:
1188         ntb_memcpy_rx(entry, offset);
1189         qp->rx_memcpy++;
1190 }
1191
1192 static int ntb_process_rxc(struct ntb_transport_qp *qp)
1193 {
1194         struct ntb_payload_header *hdr;
1195         struct ntb_queue_entry *entry;
1196         void *offset;
1197         int rc;
1198
1199         offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
1200         hdr = offset + qp->rx_max_frame - sizeof(struct ntb_payload_header);
1201
1202         dev_dbg(&qp->ndev->pdev->dev, "qp %d: RX ver %u len %d flags %x\n",
1203                 qp->qp_num, hdr->ver, hdr->len, hdr->flags);
1204
1205         if (!(hdr->flags & DESC_DONE_FLAG)) {
1206                 dev_dbg(&qp->ndev->pdev->dev, "done flag not set\n");
1207                 qp->rx_ring_empty++;
1208                 return -EAGAIN;
1209         }
1210
1211         if (hdr->flags & LINK_DOWN_FLAG) {
1212                 dev_dbg(&qp->ndev->pdev->dev, "link down flag set\n");
1213                 ntb_qp_link_down(qp);
1214                 hdr->flags = 0;
1215                 iowrite32(qp->rx_index, &qp->rx_info->entry);
1216                 return 0;
1217         }
1218
1219         if (hdr->ver != (u32)qp->rx_pkts) {
1220                 dev_dbg(&qp->ndev->pdev->dev,
1221                         "version mismatch, expected %llu - got %u\n",
1222                         qp->rx_pkts, hdr->ver);
1223                 qp->rx_err_ver++;
1224                 return -EIO;
1225         }
1226
1227         entry = ntb_list_rm(&qp->ntb_rx_pend_q_lock, &qp->rx_pend_q);
1228         if (!entry) {
1229                 dev_dbg(&qp->ndev->pdev->dev, "no receive buffer\n");
1230                 qp->rx_err_no_buf++;
1231
1232                 rc = -ENOMEM;
1233                 goto err;
1234         }
1235
1236         if (hdr->len > entry->len) {
1237                 dev_dbg(&qp->ndev->pdev->dev,
1238                         "receive buffer overflow! Wanted %d got %d\n",
1239                         hdr->len, entry->len);
1240                 qp->rx_err_oflow++;
1241
1242                 rc = -EIO;
1243                 goto err;
1244         }
1245
1246         dev_dbg(&qp->ndev->pdev->dev,
1247                 "RX OK index %u ver %u size %d into buf size %d\n",
1248                 qp->rx_index, hdr->ver, hdr->len, entry->len);
1249
1250         qp->rx_bytes += hdr->len;
1251         qp->rx_pkts++;
1252
1253         entry->index = qp->rx_index;
1254         entry->rx_hdr = hdr;
1255
1256         ntb_async_rx(entry, offset, hdr->len);
1257
1258         qp->rx_index++;
1259         qp->rx_index %= qp->rx_max_entry;
1260
1261         return 0;
1262
1263 err:
1264         /* FIXME: if this syncrhonous update of the rx_index gets ahead of
1265          * asyncrhonous ntb_rx_copy_callback of previous entry, there are three
1266          * scenarios:
1267          *
1268          * 1) The peer might miss this update, but observe the update
1269          * from the memcpy completion callback.  In this case, the buffer will
1270          * not be freed on the peer to be reused for a different packet.  The
1271          * successful rx of a later packet would clear the condition, but the
1272          * condition could persist if several rx fail in a row.
1273          *
1274          * 2) The peer may observe this update before the asyncrhonous copy of
1275          * prior packets is completed.  The peer may overwrite the buffers of
1276          * the prior packets before they are copied.
1277          *
1278          * 3) Both: the peer may observe the update, and then observe the index
1279          * decrement by the asynchronous completion callback.  Who knows what
1280          * badness that will cause.
1281          */
1282         hdr->flags = 0;
1283         iowrite32(qp->rx_index, &qp->rx_info->entry);
1284
1285         return rc;
1286 }
1287
1288 static void ntb_transport_rxc_db(unsigned long data)
1289 {
1290         struct ntb_transport_qp *qp = (void *)data;
1291         int rc, i;
1292
1293         dev_dbg(&qp->ndev->pdev->dev, "%s: doorbell %d received\n",
1294                 __func__, qp->qp_num);
1295
1296         /* Limit the number of packets processed in a single interrupt to
1297          * provide fairness to others
1298          */
1299         for (i = 0; i < qp->rx_max_entry; i++) {
1300                 rc = ntb_process_rxc(qp);
1301                 if (rc)
1302                         break;
1303         }
1304
1305         if (qp->dma_chan)
1306                 dma_async_issue_pending(qp->dma_chan);
1307
1308         if (i == qp->rx_max_entry) {
1309                 /* there is more work to do */
1310                 tasklet_schedule(&qp->rxc_db_work);
1311         } else if (ntb_db_read(qp->ndev) & BIT_ULL(qp->qp_num)) {
1312                 /* the doorbell bit is set: clear it */
1313                 ntb_db_clear(qp->ndev, BIT_ULL(qp->qp_num));
1314                 /* ntb_db_read ensures ntb_db_clear write is committed */
1315                 ntb_db_read(qp->ndev);
1316
1317                 /* an interrupt may have arrived between finishing
1318                  * ntb_process_rxc and clearing the doorbell bit:
1319                  * there might be some more work to do.
1320                  */
1321                 tasklet_schedule(&qp->rxc_db_work);
1322         }
1323 }
1324
1325 static void ntb_tx_copy_callback(void *data)
1326 {
1327         struct ntb_queue_entry *entry = data;
1328         struct ntb_transport_qp *qp = entry->qp;
1329         struct ntb_payload_header __iomem *hdr = entry->tx_hdr;
1330
1331         iowrite32(entry->flags | DESC_DONE_FLAG, &hdr->flags);
1332
1333         ntb_peer_db_set(qp->ndev, BIT_ULL(qp->qp_num));
1334
1335         /* The entry length can only be zero if the packet is intended to be a
1336          * "link down" or similar.  Since no payload is being sent in these
1337          * cases, there is nothing to add to the completion queue.
1338          */
1339         if (entry->len > 0) {
1340                 qp->tx_bytes += entry->len;
1341
1342                 if (qp->tx_handler)
1343                         qp->tx_handler(qp, qp->cb_data, entry->cb_data,
1344                                        entry->len);
1345         }
1346
1347         ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry, &qp->tx_free_q);
1348 }
1349
1350 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset)
1351 {
1352         memcpy_toio(offset, entry->buf, entry->len);
1353
1354         /* Ensure that the data is fully copied out before setting the flags */
1355         wmb();
1356
1357         ntb_tx_copy_callback(entry);
1358 }
1359
1360 static void ntb_async_tx(struct ntb_transport_qp *qp,
1361                          struct ntb_queue_entry *entry)
1362 {
1363         struct ntb_payload_header __iomem *hdr;
1364         struct dma_async_tx_descriptor *txd;
1365         struct dma_chan *chan = qp->dma_chan;
1366         struct dma_device *device;
1367         size_t dest_off, buff_off;
1368         struct dmaengine_unmap_data *unmap;
1369         dma_addr_t dest;
1370         dma_cookie_t cookie;
1371         void __iomem *offset;
1372         size_t len = entry->len;
1373         void *buf = entry->buf;
1374
1375         offset = qp->tx_mw + qp->tx_max_frame * qp->tx_index;
1376         hdr = offset + qp->tx_max_frame - sizeof(struct ntb_payload_header);
1377         entry->tx_hdr = hdr;
1378
1379         iowrite32(entry->len, &hdr->len);
1380         iowrite32((u32)qp->tx_pkts, &hdr->ver);
1381
1382         if (!chan)
1383                 goto err;
1384
1385         if (len < copy_bytes)
1386                 goto err;
1387
1388         device = chan->device;
1389         dest = qp->tx_mw_phys + qp->tx_max_frame * qp->tx_index;
1390         buff_off = (size_t)buf & ~PAGE_MASK;
1391         dest_off = (size_t)dest & ~PAGE_MASK;
1392
1393         if (!is_dma_copy_aligned(device, buff_off, dest_off, len))
1394                 goto err;
1395
1396         unmap = dmaengine_get_unmap_data(device->dev, 1, GFP_NOWAIT);
1397         if (!unmap)
1398                 goto err;
1399
1400         unmap->len = len;
1401         unmap->addr[0] = dma_map_page(device->dev, virt_to_page(buf),
1402                                       buff_off, len, DMA_TO_DEVICE);
1403         if (dma_mapping_error(device->dev, unmap->addr[0]))
1404                 goto err_get_unmap;
1405
1406         unmap->to_cnt = 1;
1407
1408         txd = device->device_prep_dma_memcpy(chan, dest, unmap->addr[0], len,
1409                                              DMA_PREP_INTERRUPT);
1410         if (!txd)
1411                 goto err_get_unmap;
1412
1413         txd->callback = ntb_tx_copy_callback;
1414         txd->callback_param = entry;
1415         dma_set_unmap(txd, unmap);
1416
1417         cookie = dmaengine_submit(txd);
1418         if (dma_submit_error(cookie))
1419                 goto err_set_unmap;
1420
1421         dmaengine_unmap_put(unmap);
1422
1423         dma_async_issue_pending(chan);
1424         qp->tx_async++;
1425
1426         return;
1427 err_set_unmap:
1428         dmaengine_unmap_put(unmap);
1429 err_get_unmap:
1430         dmaengine_unmap_put(unmap);
1431 err:
1432         ntb_memcpy_tx(entry, offset);
1433         qp->tx_memcpy++;
1434 }
1435
1436 static int ntb_process_tx(struct ntb_transport_qp *qp,
1437                           struct ntb_queue_entry *entry)
1438 {
1439         if (qp->tx_index == qp->remote_rx_info->entry) {
1440                 qp->tx_ring_full++;
1441                 return -EAGAIN;
1442         }
1443
1444         if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
1445                 if (qp->tx_handler)
1446                         qp->tx_handler(qp->cb_data, qp, NULL, -EIO);
1447
1448                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1449                              &qp->tx_free_q);
1450                 return 0;
1451         }
1452
1453         ntb_async_tx(qp, entry);
1454
1455         qp->tx_index++;
1456         qp->tx_index %= qp->tx_max_entry;
1457
1458         qp->tx_pkts++;
1459
1460         return 0;
1461 }
1462
1463 static void ntb_send_link_down(struct ntb_transport_qp *qp)
1464 {
1465         struct pci_dev *pdev = qp->ndev->pdev;
1466         struct ntb_queue_entry *entry;
1467         int i, rc;
1468
1469         if (!qp->link_is_up)
1470                 return;
1471
1472         qp->link_is_up = false;
1473         dev_info(&pdev->dev, "qp %d: Link Down\n", qp->qp_num);
1474
1475         for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1476                 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1477                 if (entry)
1478                         break;
1479                 msleep(100);
1480         }
1481
1482         if (!entry)
1483                 return;
1484
1485         entry->cb_data = NULL;
1486         entry->buf = NULL;
1487         entry->len = 0;
1488         entry->flags = LINK_DOWN_FLAG;
1489
1490         rc = ntb_process_tx(qp, entry);
1491         if (rc)
1492                 dev_err(&pdev->dev, "ntb: QP%d unable to send linkdown msg\n",
1493                         qp->qp_num);
1494 }
1495
1496 /**
1497  * ntb_transport_create_queue - Create a new NTB transport layer queue
1498  * @rx_handler: receive callback function
1499  * @tx_handler: transmit callback function
1500  * @event_handler: event callback function
1501  *
1502  * Create a new NTB transport layer queue and provide the queue with a callback
1503  * routine for both transmit and receive.  The receive callback routine will be
1504  * used to pass up data when the transport has received it on the queue.   The
1505  * transmit callback routine will be called when the transport has completed the
1506  * transmission of the data on the queue and the data is ready to be freed.
1507  *
1508  * RETURNS: pointer to newly created ntb_queue, NULL on error.
1509  */
1510 struct ntb_transport_qp *
1511 ntb_transport_create_queue(void *data, struct device *client_dev,
1512                            const struct ntb_queue_handlers *handlers)
1513 {
1514         struct ntb_dev *ndev;
1515         struct pci_dev *pdev;
1516         struct ntb_transport_ctx *nt;
1517         struct ntb_queue_entry *entry;
1518         struct ntb_transport_qp *qp;
1519         u64 qp_bit;
1520         unsigned int free_queue;
1521         int i;
1522
1523         ndev = dev_ntb(client_dev->parent);
1524         pdev = ndev->pdev;
1525         nt = ndev->ctx;
1526
1527         free_queue = ffs(nt->qp_bitmap);
1528         if (!free_queue)
1529                 goto err;
1530
1531         /* decrement free_queue to make it zero based */
1532         free_queue--;
1533
1534         qp = &nt->qp_vec[free_queue];
1535         qp_bit = BIT_ULL(qp->qp_num);
1536
1537         nt->qp_bitmap_free &= ~qp_bit;
1538
1539         qp->cb_data = data;
1540         qp->rx_handler = handlers->rx_handler;
1541         qp->tx_handler = handlers->tx_handler;
1542         qp->event_handler = handlers->event_handler;
1543
1544         dmaengine_get();
1545         qp->dma_chan = dma_find_channel(DMA_MEMCPY);
1546         if (!qp->dma_chan) {
1547                 dmaengine_put();
1548                 dev_info(&pdev->dev, "Unable to allocate DMA channel, using CPU instead\n");
1549         }
1550
1551         for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
1552                 entry = kzalloc(sizeof(*entry), GFP_ATOMIC);
1553                 if (!entry)
1554                         goto err1;
1555
1556                 entry->qp = qp;
1557                 ntb_list_add(&qp->ntb_rx_free_q_lock, &entry->entry,
1558                              &qp->rx_free_q);
1559         }
1560
1561         for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
1562                 entry = kzalloc(sizeof(*entry), GFP_ATOMIC);
1563                 if (!entry)
1564                         goto err2;
1565
1566                 entry->qp = qp;
1567                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1568                              &qp->tx_free_q);
1569         }
1570
1571         ntb_db_clear(qp->ndev, qp_bit);
1572         ntb_db_clear_mask(qp->ndev, qp_bit);
1573
1574         dev_info(&pdev->dev, "NTB Transport QP %d created\n", qp->qp_num);
1575
1576         return qp;
1577
1578 err2:
1579         while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1580                 kfree(entry);
1581 err1:
1582         while ((entry = ntb_list_rm(&qp->ntb_rx_free_q_lock, &qp->rx_free_q)))
1583                 kfree(entry);
1584         if (qp->dma_chan)
1585                 dmaengine_put();
1586         nt->qp_bitmap_free |= qp_bit;
1587 err:
1588         return NULL;
1589 }
1590 EXPORT_SYMBOL_GPL(ntb_transport_create_queue);
1591
1592 /**
1593  * ntb_transport_free_queue - Frees NTB transport queue
1594  * @qp: NTB queue to be freed
1595  *
1596  * Frees NTB transport queue
1597  */
1598 void ntb_transport_free_queue(struct ntb_transport_qp *qp)
1599 {
1600         struct ntb_transport_ctx *nt = qp->transport;
1601         struct pci_dev *pdev;
1602         struct ntb_queue_entry *entry;
1603         u64 qp_bit;
1604
1605         if (!qp)
1606                 return;
1607
1608         pdev = qp->ndev->pdev;
1609
1610         if (qp->dma_chan) {
1611                 struct dma_chan *chan = qp->dma_chan;
1612                 /* Putting the dma_chan to NULL will force any new traffic to be
1613                  * processed by the CPU instead of the DAM engine
1614                  */
1615                 qp->dma_chan = NULL;
1616
1617                 /* Try to be nice and wait for any queued DMA engine
1618                  * transactions to process before smashing it with a rock
1619                  */
1620                 dma_sync_wait(chan, qp->last_cookie);
1621                 dmaengine_terminate_all(chan);
1622                 dmaengine_put();
1623         }
1624
1625         qp_bit = BIT_ULL(qp->qp_num);
1626
1627         ntb_db_set_mask(qp->ndev, qp_bit);
1628         tasklet_disable(&qp->rxc_db_work);
1629
1630         cancel_delayed_work_sync(&qp->link_work);
1631
1632         qp->cb_data = NULL;
1633         qp->rx_handler = NULL;
1634         qp->tx_handler = NULL;
1635         qp->event_handler = NULL;
1636
1637         while ((entry = ntb_list_rm(&qp->ntb_rx_free_q_lock, &qp->rx_free_q)))
1638                 kfree(entry);
1639
1640         while ((entry = ntb_list_rm(&qp->ntb_rx_pend_q_lock, &qp->rx_pend_q))) {
1641                 dev_warn(&pdev->dev, "Freeing item from a non-empty queue\n");
1642                 kfree(entry);
1643         }
1644
1645         while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1646                 kfree(entry);
1647
1648         nt->qp_bitmap_free |= qp_bit;
1649
1650         dev_info(&pdev->dev, "NTB Transport QP %d freed\n", qp->qp_num);
1651 }
1652 EXPORT_SYMBOL_GPL(ntb_transport_free_queue);
1653
1654 /**
1655  * ntb_transport_rx_remove - Dequeues enqueued rx packet
1656  * @qp: NTB queue to be freed
1657  * @len: pointer to variable to write enqueued buffers length
1658  *
1659  * Dequeues unused buffers from receive queue.  Should only be used during
1660  * shutdown of qp.
1661  *
1662  * RETURNS: NULL error value on error, or void* for success.
1663  */
1664 void *ntb_transport_rx_remove(struct ntb_transport_qp *qp, unsigned int *len)
1665 {
1666         struct ntb_queue_entry *entry;
1667         void *buf;
1668
1669         if (!qp || qp->client_ready)
1670                 return NULL;
1671
1672         entry = ntb_list_rm(&qp->ntb_rx_pend_q_lock, &qp->rx_pend_q);
1673         if (!entry)
1674                 return NULL;
1675
1676         buf = entry->cb_data;
1677         *len = entry->len;
1678
1679         ntb_list_add(&qp->ntb_rx_free_q_lock, &entry->entry, &qp->rx_free_q);
1680
1681         return buf;
1682 }
1683 EXPORT_SYMBOL_GPL(ntb_transport_rx_remove);
1684
1685 /**
1686  * ntb_transport_rx_enqueue - Enqueue a new NTB queue entry
1687  * @qp: NTB transport layer queue the entry is to be enqueued on
1688  * @cb: per buffer pointer for callback function to use
1689  * @data: pointer to data buffer that incoming packets will be copied into
1690  * @len: length of the data buffer
1691  *
1692  * Enqueue a new receive buffer onto the transport queue into which a NTB
1693  * payload can be received into.
1694  *
1695  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
1696  */
1697 int ntb_transport_rx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
1698                              unsigned int len)
1699 {
1700         struct ntb_queue_entry *entry;
1701
1702         if (!qp)
1703                 return -EINVAL;
1704
1705         entry = ntb_list_rm(&qp->ntb_rx_free_q_lock, &qp->rx_free_q);
1706         if (!entry)
1707                 return -ENOMEM;
1708
1709         entry->cb_data = cb;
1710         entry->buf = data;
1711         entry->len = len;
1712
1713         ntb_list_add(&qp->ntb_rx_pend_q_lock, &entry->entry, &qp->rx_pend_q);
1714
1715         return 0;
1716 }
1717 EXPORT_SYMBOL_GPL(ntb_transport_rx_enqueue);
1718
1719 /**
1720  * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
1721  * @qp: NTB transport layer queue the entry is to be enqueued on
1722  * @cb: per buffer pointer for callback function to use
1723  * @data: pointer to data buffer that will be sent
1724  * @len: length of the data buffer
1725  *
1726  * Enqueue a new transmit buffer onto the transport queue from which a NTB
1727  * payload will be transmitted.  This assumes that a lock is being held to
1728  * serialize access to the qp.
1729  *
1730  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
1731  */
1732 int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
1733                              unsigned int len)
1734 {
1735         struct ntb_queue_entry *entry;
1736         int rc;
1737
1738         if (!qp || !qp->link_is_up || !len)
1739                 return -EINVAL;
1740
1741         entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1742         if (!entry) {
1743                 qp->tx_err_no_buf++;
1744                 return -ENOMEM;
1745         }
1746
1747         entry->cb_data = cb;
1748         entry->buf = data;
1749         entry->len = len;
1750         entry->flags = 0;
1751
1752         rc = ntb_process_tx(qp, entry);
1753         if (rc)
1754                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1755                              &qp->tx_free_q);
1756
1757         return rc;
1758 }
1759 EXPORT_SYMBOL_GPL(ntb_transport_tx_enqueue);
1760
1761 /**
1762  * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
1763  * @qp: NTB transport layer queue to be enabled
1764  *
1765  * Notify NTB transport layer of client readiness to use queue
1766  */
1767 void ntb_transport_link_up(struct ntb_transport_qp *qp)
1768 {
1769         if (!qp)
1770                 return;
1771
1772         qp->client_ready = true;
1773
1774         if (qp->transport->link_is_up)
1775                 schedule_delayed_work(&qp->link_work, 0);
1776 }
1777 EXPORT_SYMBOL_GPL(ntb_transport_link_up);
1778
1779 /**
1780  * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
1781  * @qp: NTB transport layer queue to be disabled
1782  *
1783  * Notify NTB transport layer of client's desire to no longer receive data on
1784  * transport queue specified.  It is the client's responsibility to ensure all
1785  * entries on queue are purged or otherwise handled appropriately.
1786  */
1787 void ntb_transport_link_down(struct ntb_transport_qp *qp)
1788 {
1789         struct pci_dev *pdev;
1790         int val;
1791
1792         if (!qp)
1793                 return;
1794
1795         pdev = qp->ndev->pdev;
1796         qp->client_ready = false;
1797
1798         val = ntb_spad_read(qp->ndev, QP_LINKS);
1799
1800         ntb_peer_spad_write(qp->ndev, QP_LINKS,
1801                             val & ~BIT(qp->qp_num));
1802
1803         if (qp->link_is_up)
1804                 ntb_send_link_down(qp);
1805         else
1806                 cancel_delayed_work_sync(&qp->link_work);
1807 }
1808 EXPORT_SYMBOL_GPL(ntb_transport_link_down);
1809
1810 /**
1811  * ntb_transport_link_query - Query transport link state
1812  * @qp: NTB transport layer queue to be queried
1813  *
1814  * Query connectivity to the remote system of the NTB transport queue
1815  *
1816  * RETURNS: true for link up or false for link down
1817  */
1818 bool ntb_transport_link_query(struct ntb_transport_qp *qp)
1819 {
1820         if (!qp)
1821                 return false;
1822
1823         return qp->link_is_up;
1824 }
1825 EXPORT_SYMBOL_GPL(ntb_transport_link_query);
1826
1827 /**
1828  * ntb_transport_qp_num - Query the qp number
1829  * @qp: NTB transport layer queue to be queried
1830  *
1831  * Query qp number of the NTB transport queue
1832  *
1833  * RETURNS: a zero based number specifying the qp number
1834  */
1835 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
1836 {
1837         if (!qp)
1838                 return 0;
1839
1840         return qp->qp_num;
1841 }
1842 EXPORT_SYMBOL_GPL(ntb_transport_qp_num);
1843
1844 /**
1845  * ntb_transport_max_size - Query the max payload size of a qp
1846  * @qp: NTB transport layer queue to be queried
1847  *
1848  * Query the maximum payload size permissible on the given qp
1849  *
1850  * RETURNS: the max payload size of a qp
1851  */
1852 unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp)
1853 {
1854         unsigned int max;
1855
1856         if (!qp)
1857                 return 0;
1858
1859         if (!qp->dma_chan)
1860                 return qp->tx_max_frame - sizeof(struct ntb_payload_header);
1861
1862         /* If DMA engine usage is possible, try to find the max size for that */
1863         max = qp->tx_max_frame - sizeof(struct ntb_payload_header);
1864         max -= max % (1 << qp->dma_chan->device->copy_align);
1865
1866         return max;
1867 }
1868 EXPORT_SYMBOL_GPL(ntb_transport_max_size);
1869
1870 static void ntb_transport_doorbell_callback(void *data, int vector)
1871 {
1872         struct ntb_transport_ctx *nt = data;
1873         struct ntb_transport_qp *qp;
1874         u64 db_bits;
1875         unsigned int qp_num;
1876
1877         db_bits = (nt->qp_bitmap & ~nt->qp_bitmap_free &
1878                    ntb_db_vector_mask(nt->ndev, vector));
1879
1880         while (db_bits) {
1881                 qp_num = __ffs(db_bits);
1882                 qp = &nt->qp_vec[qp_num];
1883
1884                 tasklet_schedule(&qp->rxc_db_work);
1885
1886                 db_bits &= ~BIT_ULL(qp_num);
1887         }
1888 }
1889
1890 static const struct ntb_ctx_ops ntb_transport_ops = {
1891         .link_event = ntb_transport_event_callback,
1892         .db_event = ntb_transport_doorbell_callback,
1893 };
1894
1895 static struct ntb_client ntb_transport_client = {
1896         .ops = {
1897                 .probe = ntb_transport_probe,
1898                 .remove = ntb_transport_free,
1899         },
1900 };
1901
1902 static int __init ntb_transport_init(void)
1903 {
1904         int rc;
1905
1906         if (debugfs_initialized())
1907                 nt_debugfs_dir = debugfs_create_dir(KBUILD_MODNAME, NULL);
1908
1909         rc = bus_register(&ntb_transport_bus);
1910         if (rc)
1911                 goto err_bus;
1912
1913         rc = ntb_register_client(&ntb_transport_client);
1914         if (rc)
1915                 goto err_client;
1916
1917         return 0;
1918
1919 err_client:
1920         bus_unregister(&ntb_transport_bus);
1921 err_bus:
1922         debugfs_remove_recursive(nt_debugfs_dir);
1923         return rc;
1924 }
1925 module_init(ntb_transport_init);
1926
1927 static void __exit ntb_transport_exit(void)
1928 {
1929         debugfs_remove_recursive(nt_debugfs_dir);
1930
1931         ntb_unregister_client(&ntb_transport_client);
1932         bus_unregister(&ntb_transport_bus);
1933 }
1934 module_exit(ntb_transport_exit);