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/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
  70 MODULE_DESCRIPTION(NTB_TRANSPORT_DESC);
  71 MODULE_VERSION(NTB_TRANSPORT_VER);
  72 MODULE_LICENSE("Dual BSD/GPL");
  73 MODULE_AUTHOR("Intel Corporation");
  74
  75 static unsigned long max_mw_size;
  76 module_param(max_mw_size, ulong, 0644);
  77 MODULE_PARM_DESC(max_mw_size, "Limit size of large memory windows");
  78
  79 static unsigned int transport_mtu = 0x401E;
  80 module_param(transport_mtu, uint, 0644);
  81 MODULE_PARM_DESC(transport_mtu, "Maximum size of NTB transport packets");
  82
  83 static unsigned char max_num_clients;
  84 module_param(max_num_clients, byte, 0644);
  85 MODULE_PARM_DESC(max_num_clients, "Maximum number of NTB transport clients");
  86
  87 static unsigned int copy_bytes = 1024;
  88 module_param(copy_bytes, uint, 0644);
  89 MODULE_PARM_DESC(copy_bytes, "Threshold under which NTB will use the CPU to copy instead of DMA");
  90
  91 static struct dentry *nt_debugfs_dir;
  92
  93 struct ntb_queue_entry {
  94         /* ntb_queue list reference */
  95         struct list_head entry;
  96         /* pointers to data to be transferred */
  97         void *cb_data;
  98         void *buf;
  99         unsigned int len;
 100         unsigned int flags;
 101
 102         struct ntb_transport_qp *qp;
 103         union {
 104                 struct ntb_payload_header __iomem *tx_hdr;
 105                 struct ntb_payload_header *rx_hdr;
 106         };
 107         unsigned int index;
 108 };
 109
 110 struct ntb_rx_info {
 111         unsigned int entry;
 112 };
 113
 114 struct ntb_transport_qp {
 115         struct ntb_transport_ctx *transport;
 116         struct ntb_dev *ndev;
 117         void *cb_data;
 118         struct dma_chan *dma_chan;
 119
 120         bool client_ready;
 121         bool link_is_up;
 122
 123         u8 qp_num;      /* Only 64 QP's are allowed.  0-63 */
 124         u64 qp_bit;
 125
 126         struct ntb_rx_info __iomem *rx_info;
 127         struct ntb_rx_info *remote_rx_info;
 128
 129         void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data,
 130                            void *data, int len);
 131         struct list_head tx_free_q;
 132         spinlock_t ntb_tx_free_q_lock;
 133         void __iomem *tx_mw;
 134         dma_addr_t tx_mw_phys;
 135         unsigned int tx_index;
 136         unsigned int tx_max_entry;
 137         unsigned int tx_max_frame;
 138
 139         void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
 140                            void *data, int len);
 141         struct list_head rx_pend_q;
 142         struct list_head rx_free_q;
 143         spinlock_t ntb_rx_pend_q_lock;
 144         spinlock_t ntb_rx_free_q_lock;
 145         void *rx_buff;
 146         unsigned int rx_index;
 147         unsigned int rx_max_entry;
 148         unsigned int rx_max_frame;
 149         dma_cookie_t last_cookie;
 150         struct tasklet_struct rxc_db_work;
 151
 152         void (*event_handler)(void *data, int status);
 153         struct delayed_work link_work;
 154         struct work_struct link_cleanup;
 155
 156         struct dentry *debugfs_dir;
 157         struct dentry *debugfs_stats;
 158
 159         /* Stats */
 160         u64 rx_bytes;
 161         u64 rx_pkts;
 162         u64 rx_ring_empty;
 163         u64 rx_err_no_buf;
 164         u64 rx_err_oflow;
 165         u64 rx_err_ver;
 166         u64 rx_memcpy;
 167         u64 rx_async;
 168         u64 tx_bytes;
 169         u64 tx_pkts;
 170         u64 tx_ring_full;
 171         u64 tx_err_no_buf;
 172         u64 tx_memcpy;
 173         u64 tx_async;
 174 };
 175
 176 struct ntb_transport_mw {
 177         phys_addr_t phys_addr;
 178         resource_size_t phys_size;
 179         resource_size_t xlat_align;
 180         resource_size_t xlat_align_size;
 181         void __iomem *vbase;
 182         size_t xlat_size;
 183         size_t buff_size;
 184         void *virt_addr;
 185         dma_addr_t dma_addr;
 186 };
 187
 188 struct ntb_transport_client_dev {
 189         struct list_head entry;
 190         struct ntb_transport_ctx *nt;
 191         struct device dev;
 192 };
 193
 194 struct ntb_transport_ctx {
 195         struct list_head entry;
 196         struct list_head client_devs;
 197
 198         struct ntb_dev *ndev;
 199
 200         struct ntb_transport_mw *mw_vec;
 201         struct ntb_transport_qp *qp_vec;
 202         unsigned int mw_count;
 203         unsigned int qp_count;
 204         u64 qp_bitmap;
 205         u64 qp_bitmap_free;
 206
 207         bool link_is_up;
 208         struct delayed_work link_work;
 209         struct work_struct link_cleanup;
 210 };
 211
 212 enum {
 213         DESC_DONE_FLAG = BIT(0),
 214         LINK_DOWN_FLAG = BIT(1),
 215 };
 216
 217 struct ntb_payload_header {
 218         unsigned int ver;
 219         unsigned int len;
 220         unsigned int flags;
 221 };
 222
 223 enum {
 224         VERSION = 0,
 225         QP_LINKS,
 226         NUM_QPS,
 227         NUM_MWS,
 228         MW0_SZ_HIGH,
 229         MW0_SZ_LOW,
 230         MW1_SZ_HIGH,
 231         MW1_SZ_LOW,
 232         MAX_SPAD,
 233 };
 234
 235 #define dev_client_dev(__dev) \
 236         container_of((__dev), struct ntb_transport_client_dev, dev)
 237
 238 #define drv_client(__drv) \
 239         container_of((__drv), struct ntb_transport_client, driver)
 240
 241 #define QP_TO_MW(nt, qp)        ((qp) % nt->mw_count)
 242 #define NTB_QP_DEF_NUM_ENTRIES  100
 243 #define NTB_LINK_DOWN_TIMEOUT   10
 244
 245 static void ntb_transport_rxc_db(unsigned long data);
 246 static const struct ntb_ctx_ops ntb_transport_ops;
 247 static struct ntb_client ntb_transport_client;
 248
 249 static int ntb_transport_bus_match(struct device *dev,
 250                                    struct device_driver *drv)
 251 {
 252         return !strncmp(dev_name(dev), drv->name, strlen(drv->name));
 253 }
 254
 255 static int ntb_transport_bus_probe(struct device *dev)
 256 {
 257         const struct ntb_transport_client *client;
 258         int rc = -EINVAL;
 259
 260         get_device(dev);
 261
 262         client = drv_client(dev->driver);
 263         rc = client->probe(dev);
 264         if (rc)
 265                 put_device(dev);
 266
 267         return rc;
 268 }
 269
 270 static int ntb_transport_bus_remove(struct device *dev)
 271 {
 272         const struct ntb_transport_client *client;
 273
 274         client = drv_client(dev->driver);
 275         client->remove(dev);
 276
 277         put_device(dev);
 278
 279         return 0;
 280 }
 281
 282 static struct bus_type ntb_transport_bus = {
 283         .name = "ntb_transport",
 284         .match = ntb_transport_bus_match,
 285         .probe = ntb_transport_bus_probe,
 286         .remove = ntb_transport_bus_remove,
 287 };
 288
 289 static LIST_HEAD(ntb_transport_list);
 290
 291 static int ntb_bus_init(struct ntb_transport_ctx *nt)
 292 {
 293         list_add(&nt->entry, &ntb_transport_list);
 294         return 0;
 295 }
 296
 297 static void ntb_bus_remove(struct ntb_transport_ctx *nt)
 298 {
 299         struct ntb_transport_client_dev *client_dev, *cd;
 300
 301         list_for_each_entry_safe(client_dev, cd, &nt->client_devs, entry) {
 302                 dev_err(client_dev->dev.parent, "%s still attached to bus, removing\n",
 303                         dev_name(&client_dev->dev));
 304                 list_del(&client_dev->entry);
 305                 device_unregister(&client_dev->dev);
 306         }
 307
 308         list_del(&nt->entry);
 309 }
 310
 311 static void ntb_transport_client_release(struct device *dev)
 312 {
 313         struct ntb_transport_client_dev *client_dev;
 314
 315         client_dev = dev_client_dev(dev);
 316         kfree(client_dev);
 317 }
 318
 319 /**
 320  * ntb_transport_unregister_client_dev - Unregister NTB client device
 321  * @device_name: Name of NTB client device
 322  *
 323  * Unregister an NTB client device with the NTB transport layer
 324  */
 325 void ntb_transport_unregister_client_dev(char *device_name)
 326 {
 327         struct ntb_transport_client_dev *client, *cd;
 328         struct ntb_transport_ctx *nt;
 329
 330         list_for_each_entry(nt, &ntb_transport_list, entry)
 331                 list_for_each_entry_safe(client, cd, &nt->client_devs, entry)
 332                         if (!strncmp(dev_name(&client->dev), device_name,
 333                                      strlen(device_name))) {
 334                                 list_del(&client->entry);
 335                                 device_unregister(&client->dev);
 336                         }
 337 }
 338 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client_dev);
 339
 340 /**
 341  * ntb_transport_register_client_dev - Register NTB client device
 342  * @device_name: Name of NTB client device
 343  *
 344  * Register an NTB client device with the NTB transport layer
 345  */
 346 int ntb_transport_register_client_dev(char *device_name)
 347 {
 348         struct ntb_transport_client_dev *client_dev;
 349         struct ntb_transport_ctx *nt;
 350         int node;
 351         int rc, i = 0;
 352
 353         if (list_empty(&ntb_transport_list))
 354                 return -ENODEV;
 355
 356         list_for_each_entry(nt, &ntb_transport_list, entry) {
 357                 struct device *dev;
 358
 359                 node = dev_to_node(&nt->ndev->dev);
 360
 361                 client_dev = kzalloc_node(sizeof(*client_dev),
 362                                           GFP_KERNEL, node);
 363                 if (!client_dev) {
 364                         rc = -ENOMEM;
 365                         goto err;
 366                 }
 367
 368                 dev = &client_dev->dev;
 369
 370                 /* setup and register client devices */
 371                 dev_set_name(dev, "%s%d", device_name, i);
 372                 dev->bus = &ntb_transport_bus;
 373                 dev->release = ntb_transport_client_release;
 374                 dev->parent = &nt->ndev->dev;
 375
 376                 rc = device_register(dev);
 377                 if (rc) {
 378                         kfree(client_dev);
 379                         goto err;
 380                 }
 381
 382                 list_add_tail(&client_dev->entry, &nt->client_devs);
 383                 i++;
 384         }
 385
 386         return 0;
 387
 388 err:
 389         ntb_transport_unregister_client_dev(device_name);
 390
 391         return rc;
 392 }
 393 EXPORT_SYMBOL_GPL(ntb_transport_register_client_dev);
 394
 395 /**
 396  * ntb_transport_register_client - Register NTB client driver
 397  * @drv: NTB client driver to be registered
 398  *
 399  * Register an NTB client driver with the NTB transport layer
 400  *
 401  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 402  */
 403 int ntb_transport_register_client(struct ntb_transport_client *drv)
 404 {
 405         drv->driver.bus = &ntb_transport_bus;
 406
 407         if (list_empty(&ntb_transport_list))
 408                 return -ENODEV;
 409
 410         return driver_register(&drv->driver);
 411 }
 412 EXPORT_SYMBOL_GPL(ntb_transport_register_client);
 413
 414 /**
 415  * ntb_transport_unregister_client - Unregister NTB client driver
 416  * @drv: NTB client driver to be unregistered
 417  *
 418  * Unregister an NTB client driver with the NTB transport layer
 419  *
 420  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 421  */
 422 void ntb_transport_unregister_client(struct ntb_transport_client *drv)
 423 {
 424         driver_unregister(&drv->driver);
 425 }
 426 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client);
 427
 428 static ssize_t debugfs_read(struct file *filp, char __user *ubuf, size_t count,
 429                             loff_t *offp)
 430 {
 431         struct ntb_transport_qp *qp;
 432         char *buf;
 433         ssize_t ret, out_offset, out_count;
 434
 435         out_count = 1000;
 436
 437         buf = kmalloc(out_count, GFP_KERNEL);
 438         if (!buf)
 439                 return -ENOMEM;
 440
 441         qp = filp->private_data;
 442         out_offset = 0;
 443         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 444                                "NTB QP stats\n");
 445         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 446                                "rx_bytes - \t%llu\n", qp->rx_bytes);
 447         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 448                                "rx_pkts - \t%llu\n", qp->rx_pkts);
 449         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 450                                "rx_memcpy - \t%llu\n", qp->rx_memcpy);
 451         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 452                                "rx_async - \t%llu\n", qp->rx_async);
 453         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 454                                "rx_ring_empty - %llu\n", qp->rx_ring_empty);
 455         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 456                                "rx_err_no_buf - %llu\n", qp->rx_err_no_buf);
 457         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 458                                "rx_err_oflow - \t%llu\n", qp->rx_err_oflow);
 459         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 460                                "rx_err_ver - \t%llu\n", qp->rx_err_ver);
 461         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 462                                "rx_buff - \t%p\n", qp->rx_buff);
 463         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 464                                "rx_index - \t%u\n", qp->rx_index);
 465         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 466                                "rx_max_entry - \t%u\n", qp->rx_max_entry);
 467
 468         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 469                                "tx_bytes - \t%llu\n", qp->tx_bytes);
 470         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 471                                "tx_pkts - \t%llu\n", qp->tx_pkts);
 472         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 473                                "tx_memcpy - \t%llu\n", qp->tx_memcpy);
 474         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 475                                "tx_async - \t%llu\n", qp->tx_async);
 476         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 477                                "tx_ring_full - \t%llu\n", qp->tx_ring_full);
 478         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 479                                "tx_err_no_buf - %llu\n", qp->tx_err_no_buf);
 480         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 481                                "tx_mw - \t%p\n", qp->tx_mw);
 482         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 483                                "tx_index - \t%u\n", qp->tx_index);
 484         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 485                                "tx_max_entry - \t%u\n", qp->tx_max_entry);
 486
 487         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 488                                "\nQP Link %s\n",
 489                                qp->link_is_up ? "Up" : "Down");
 490         if (out_offset > out_count)
 491                 out_offset = out_count;
 492
 493         ret = simple_read_from_buffer(ubuf, count, offp, buf, out_offset);
 494         kfree(buf);
 495         return ret;
 496 }
 497
 498 static const struct file_operations ntb_qp_debugfs_stats = {
 499         .owner = THIS_MODULE,
 500         .open = simple_open,
 501         .read = debugfs_read,
 502 };
 503
 504 static void ntb_list_add(spinlock_t *lock, struct list_head *entry,
 505                          struct list_head *list)
 506 {
 507         unsigned long flags;
 508
 509         spin_lock_irqsave(lock, flags);
 510         list_add_tail(entry, list);
 511         spin_unlock_irqrestore(lock, flags);
 512 }
 513
 514 static struct ntb_queue_entry *ntb_list_rm(spinlock_t *lock,
 515                                            struct list_head *list)
 516 {
 517         struct ntb_queue_entry *entry;
 518         unsigned long flags;
 519
 520         spin_lock_irqsave(lock, flags);
 521         if (list_empty(list)) {
 522                 entry = NULL;
 523                 goto out;
 524         }
 525         entry = list_first_entry(list, struct ntb_queue_entry, entry);
 526         list_del(&entry->entry);
 527 out:
 528         spin_unlock_irqrestore(lock, flags);
 529
 530         return entry;
 531 }
 532
 533 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
 534                                      unsigned int qp_num)
 535 {
 536         struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
 537         struct ntb_transport_mw *mw;
 538         unsigned int rx_size, num_qps_mw;
 539         unsigned int mw_num, mw_count, qp_count;
 540         unsigned int i;
 541
 542         mw_count = nt->mw_count;
 543         qp_count = nt->qp_count;
 544
 545         mw_num = QP_TO_MW(nt, qp_num);
 546         mw = &nt->mw_vec[mw_num];
 547
 548         if (!mw->virt_addr)
 549                 return -ENOMEM;
 550
 551         if (qp_count % mw_count && mw_num + 1 < qp_count / mw_count)
 552                 num_qps_mw = qp_count / mw_count + 1;
 553         else
 554                 num_qps_mw = qp_count / mw_count;
 555
 556         rx_size = (unsigned int)mw->xlat_size / num_qps_mw;
 557         qp->rx_buff = mw->virt_addr + rx_size * qp_num / mw_count;
 558         rx_size -= sizeof(struct ntb_rx_info);
 559
 560         qp->remote_rx_info = qp->rx_buff + rx_size;
 561
 562         /* Due to housekeeping, there must be atleast 2 buffs */
 563         qp->rx_max_frame = min(transport_mtu, rx_size / 2);
 564         qp->rx_max_entry = rx_size / qp->rx_max_frame;
 565         qp->rx_index = 0;
 566
 567         qp->remote_rx_info->entry = qp->rx_max_entry - 1;
 568
 569         /* setup the hdr offsets with 0's */
 570         for (i = 0; i < qp->rx_max_entry; i++) {
 571                 void *offset = (qp->rx_buff + qp->rx_max_frame * (i + 1) -
 572                                 sizeof(struct ntb_payload_header));
 573                 memset(offset, 0, sizeof(struct ntb_payload_header));
 574         }
 575
 576         qp->rx_pkts = 0;
 577         qp->tx_pkts = 0;
 578         qp->tx_index = 0;
 579
 580         return 0;
 581 }
 582
 583 static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
 584 {
 585         struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
 586         struct pci_dev *pdev = nt->ndev->pdev;
 587
 588         if (!mw->virt_addr)
 589                 return;
 590
 591         ntb_mw_clear_trans(nt->ndev, num_mw);
 592         dma_free_coherent(&pdev->dev, mw->buff_size,
 593                           mw->virt_addr, mw->dma_addr);
 594         mw->xlat_size = 0;
 595         mw->buff_size = 0;
 596         mw->virt_addr = NULL;
 597 }
 598
 599 static int ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw,
 600                       unsigned int size)
 601 {
 602         struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
 603         struct pci_dev *pdev = nt->ndev->pdev;
 604         unsigned int xlat_size, buff_size;
 605         int rc;
 606
 607         xlat_size = round_up(size, mw->xlat_align_size);
 608         buff_size = round_up(size, mw->xlat_align);
 609
 610         /* No need to re-setup */
 611         if (mw->xlat_size == xlat_size)
 612                 return 0;
 613
 614         if (mw->buff_size)
 615                 ntb_free_mw(nt, num_mw);
 616
 617         /* Alloc memory for receiving data.  Must be aligned */
 618         mw->xlat_size = xlat_size;
 619         mw->buff_size = buff_size;
 620
 621         mw->virt_addr = dma_alloc_coherent(&pdev->dev, buff_size,
 622                                            &mw->dma_addr, GFP_KERNEL);
 623         if (!mw->virt_addr) {
 624                 mw->xlat_size = 0;
 625                 mw->buff_size = 0;
 626                 dev_err(&pdev->dev, "Unable to alloc MW buff of size %d\n",
 627                         buff_size);
 628                 return -ENOMEM;
 629         }
 630
 631         /*
 632          * we must ensure that the memory address allocated is BAR size
 633          * aligned in order for the XLAT register to take the value. This
 634          * is a requirement of the hardware. It is recommended to setup CMA
 635          * for BAR sizes equal or greater than 4MB.
 636          */
 637         if (!IS_ALIGNED(mw->dma_addr, mw->xlat_align)) {
 638                 dev_err(&pdev->dev, "DMA memory %pad is not aligned\n",
 639                         &mw->dma_addr);
 640                 ntb_free_mw(nt, num_mw);
 641                 return -ENOMEM;
 642         }
 643
 644         /* Notify HW the memory location of the receive buffer */
 645         rc = ntb_mw_set_trans(nt->ndev, num_mw, mw->dma_addr, mw->xlat_size);
 646         if (rc) {
 647                 dev_err(&pdev->dev, "Unable to set mw%d translation", num_mw);
 648                 ntb_free_mw(nt, num_mw);
 649                 return -EIO;
 650         }
 651
 652         return 0;
 653 }
 654
 655 static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
 656 {
 657         qp->link_is_up = false;
 658
 659         qp->tx_index = 0;
 660         qp->rx_index = 0;
 661         qp->rx_bytes = 0;
 662         qp->rx_pkts = 0;
 663         qp->rx_ring_empty = 0;
 664         qp->rx_err_no_buf = 0;
 665         qp->rx_err_oflow = 0;
 666         qp->rx_err_ver = 0;
 667         qp->rx_memcpy = 0;
 668         qp->rx_async = 0;
 669         qp->tx_bytes = 0;
 670         qp->tx_pkts = 0;
 671         qp->tx_ring_full = 0;
 672         qp->tx_err_no_buf = 0;
 673         qp->tx_memcpy = 0;
 674         qp->tx_async = 0;
 675 }
 676
 677 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
 678 {
 679         struct ntb_transport_ctx *nt = qp->transport;
 680         struct pci_dev *pdev = nt->ndev->pdev;
 681
 682         dev_info(&pdev->dev, "qp %d: Link Cleanup\n", qp->qp_num);
 683
 684         cancel_delayed_work_sync(&qp->link_work);
 685         ntb_qp_link_down_reset(qp);
 686
 687         if (qp->event_handler)
 688                 qp->event_handler(qp->cb_data, qp->link_is_up);
 689 }
 690
 691 static void ntb_qp_link_cleanup_work(struct work_struct *work)
 692 {
 693         struct ntb_transport_qp *qp = container_of(work,
 694                                                    struct ntb_transport_qp,
 695                                                    link_cleanup);
 696         struct ntb_transport_ctx *nt = qp->transport;
 697
 698         ntb_qp_link_cleanup(qp);
 699
 700         if (nt->link_is_up)
 701                 schedule_delayed_work(&qp->link_work,
 702                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 703 }
 704
 705 static void ntb_qp_link_down(struct ntb_transport_qp *qp)
 706 {
 707         schedule_work(&qp->link_cleanup);
 708 }
 709
 710 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
 711 {
 712         struct ntb_transport_qp *qp;
 713         u64 qp_bitmap_alloc;
 714         int i;
 715
 716         qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
 717
 718         /* Pass along the info to any clients */
 719         for (i = 0; i < nt->qp_count; i++)
 720                 if (qp_bitmap_alloc & BIT_ULL(i)) {
 721                         qp = &nt->qp_vec[i];
 722                         ntb_qp_link_cleanup(qp);
 723                         cancel_work_sync(&qp->link_cleanup);
 724                         cancel_delayed_work_sync(&qp->link_work);
 725                 }
 726
 727         if (!nt->link_is_up)
 728                 cancel_delayed_work_sync(&nt->link_work);
 729
 730         /* The scratchpad registers keep the values if the remote side
 731          * goes down, blast them now to give them a sane value the next
 732          * time they are accessed
 733          */
 734         for (i = 0; i < MAX_SPAD; i++)
 735                 ntb_spad_write(nt->ndev, i, 0);
 736 }
 737
 738 static void ntb_transport_link_cleanup_work(struct work_struct *work)
 739 {
 740         struct ntb_transport_ctx *nt =
 741                 container_of(work, struct ntb_transport_ctx, link_cleanup);
 742
 743         ntb_transport_link_cleanup(nt);
 744 }
 745
 746 static void ntb_transport_event_callback(void *data)
 747 {
 748         struct ntb_transport_ctx *nt = data;
 749
 750         if (ntb_link_is_up(nt->ndev, NULL, NULL) == 1)
 751                 schedule_delayed_work(&nt->link_work, 0);
 752         else
 753                 schedule_work(&nt->link_cleanup);
 754 }
 755
 756 static void ntb_transport_link_work(struct work_struct *work)
 757 {
 758         struct ntb_transport_ctx *nt =
 759                 container_of(work, struct ntb_transport_ctx, link_work.work);
 760         struct ntb_dev *ndev = nt->ndev;
 761         struct pci_dev *pdev = ndev->pdev;
 762         resource_size_t size;
 763         u32 val;
 764         int rc, i, spad;
 765
 766         /* send the local info, in the opposite order of the way we read it */
 767         for (i = 0; i < nt->mw_count; i++) {
 768                 size = nt->mw_vec[i].phys_size;
 769
 770                 if (max_mw_size && size > max_mw_size)
 771                         size = max_mw_size;
 772
 773                 spad = MW0_SZ_HIGH + (i * 2);
 774                 ntb_peer_spad_write(ndev, spad, (u32)(size >> 32));
 775
 776                 spad = MW0_SZ_LOW + (i * 2);
 777                 ntb_peer_spad_write(ndev, spad, (u32)size);
 778         }
 779
 780         ntb_peer_spad_write(ndev, NUM_MWS, nt->mw_count);
 781
 782         ntb_peer_spad_write(ndev, NUM_QPS, nt->qp_count);
 783
 784         ntb_peer_spad_write(ndev, VERSION, NTB_TRANSPORT_VERSION);
 785
 786         /* Query the remote side for its info */
 787         val = ntb_spad_read(ndev, VERSION);
 788         dev_dbg(&pdev->dev, "Remote version = %d\n", val);
 789         if (val != NTB_TRANSPORT_VERSION)
 790                 goto out;
 791
 792         val = ntb_spad_read(ndev, NUM_QPS);
 793         dev_dbg(&pdev->dev, "Remote max number of qps = %d\n", val);
 794         if (val != nt->qp_count)
 795                 goto out;
 796
 797         val = ntb_spad_read(ndev, NUM_MWS);
 798         dev_dbg(&pdev->dev, "Remote number of mws = %d\n", val);
 799         if (val != nt->mw_count)
 800                 goto out;
 801
 802         for (i = 0; i < nt->mw_count; i++) {
 803                 u64 val64;
 804
 805                 val = ntb_spad_read(ndev, MW0_SZ_HIGH + (i * 2));
 806                 val64 = (u64)val << 32;
 807
 808                 val = ntb_spad_read(ndev, MW0_SZ_LOW + (i * 2));
 809                 val64 |= val;
 810
 811                 dev_dbg(&pdev->dev, "Remote MW%d size = %#llx\n", i, val64);
 812
 813                 rc = ntb_set_mw(nt, i, val64);
 814                 if (rc)
 815                         goto out1;
 816         }
 817
 818         nt->link_is_up = true;
 819
 820         for (i = 0; i < nt->qp_count; i++) {
 821                 struct ntb_transport_qp *qp = &nt->qp_vec[i];
 822
 823                 ntb_transport_setup_qp_mw(nt, i);
 824
 825                 if (qp->client_ready)
 826                         schedule_delayed_work(&qp->link_work, 0);
 827         }
 828
 829         return;
 830
 831 out1:
 832         for (i = 0; i < nt->mw_count; i++)
 833                 ntb_free_mw(nt, i);
 834 out:
 835         if (ntb_link_is_up(ndev, NULL, NULL) == 1)
 836                 schedule_delayed_work(&nt->link_work,
 837                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 838 }
 839
 840 static void ntb_qp_link_work(struct work_struct *work)
 841 {
 842         struct ntb_transport_qp *qp = container_of(work,
 843                                                    struct ntb_transport_qp,
 844                                                    link_work.work);
 845         struct pci_dev *pdev = qp->ndev->pdev;
 846         struct ntb_transport_ctx *nt = qp->transport;
 847         int val;
 848
 849         WARN_ON(!nt->link_is_up);
 850
 851         val = ntb_spad_read(nt->ndev, QP_LINKS);
 852
 853         ntb_peer_spad_write(nt->ndev, QP_LINKS, val | BIT(qp->qp_num));
 854
 855         /* query remote spad for qp ready bits */
 856         ntb_peer_spad_read(nt->ndev, QP_LINKS);
 857         dev_dbg_ratelimited(&pdev->dev, "Remote QP link status = %x\n", val);
 858
 859         /* See if the remote side is up */
 860         if (val & BIT(qp->qp_num)) {
 861                 dev_info(&pdev->dev, "qp %d: Link Up\n", qp->qp_num);
 862                 qp->link_is_up = true;
 863
 864                 if (qp->event_handler)
 865                         qp->event_handler(qp->cb_data, qp->link_is_up);
 866         } else if (nt->link_is_up)
 867                 schedule_delayed_work(&qp->link_work,
 868                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 869 }
 870
 871 static int ntb_transport_init_queue(struct ntb_transport_ctx *nt,
 872                                     unsigned int qp_num)
 873 {
 874         struct ntb_transport_qp *qp;
 875         struct ntb_transport_mw *mw;
 876         phys_addr_t mw_base;
 877         resource_size_t mw_size;
 878         unsigned int num_qps_mw, tx_size;
 879         unsigned int mw_num, mw_count, qp_count;
 880         u64 qp_offset;
 881
 882         mw_count = nt->mw_count;
 883         qp_count = nt->qp_count;
 884
 885         mw_num = QP_TO_MW(nt, qp_num);
 886         mw = &nt->mw_vec[mw_num];
 887
 888         qp = &nt->qp_vec[qp_num];
 889         qp->qp_num = qp_num;
 890         qp->transport = nt;
 891         qp->ndev = nt->ndev;
 892         qp->client_ready = false;
 893         qp->event_handler = NULL;
 894         ntb_qp_link_down_reset(qp);
 895
 896         if (qp_count % mw_count && mw_num + 1 < qp_count / mw_count)
 897                 num_qps_mw = qp_count / mw_count + 1;
 898         else
 899                 num_qps_mw = qp_count / mw_count;
 900
 901         mw_base = nt->mw_vec[mw_num].phys_addr;
 902         mw_size = nt->mw_vec[mw_num].phys_size;
 903
 904         tx_size = (unsigned int)mw_size / num_qps_mw;
 905         qp_offset = tx_size * qp_num / mw_count;
 906
 907         qp->tx_mw = nt->mw_vec[mw_num].vbase + qp_offset;
 908         if (!qp->tx_mw)
 909                 return -EINVAL;
 910
 911         qp->tx_mw_phys = mw_base + qp_offset;
 912         if (!qp->tx_mw_phys)
 913                 return -EINVAL;
 914
 915         tx_size -= sizeof(struct ntb_rx_info);
 916         qp->rx_info = qp->tx_mw + tx_size;
 917
 918         /* Due to housekeeping, there must be atleast 2 buffs */
 919         qp->tx_max_frame = min(transport_mtu, tx_size / 2);
 920         qp->tx_max_entry = tx_size / qp->tx_max_frame;
 921
 922         if (nt_debugfs_dir) {
 923                 char debugfs_name[4];
 924
 925                 snprintf(debugfs_name, 4, "qp%d", qp_num);
 926                 qp->debugfs_dir = debugfs_create_dir(debugfs_name,
 927                                                      nt_debugfs_dir);
 928
 929                 qp->debugfs_stats = debugfs_create_file("stats", S_IRUSR,
 930                                                         qp->debugfs_dir, qp,
 931                                                         &ntb_qp_debugfs_stats);
 932         } else {
 933                 qp->debugfs_dir = NULL;
 934                 qp->debugfs_stats = NULL;
 935         }
 936
 937         INIT_DELAYED_WORK(&qp->link_work, ntb_qp_link_work);
 938         INIT_WORK(&qp->link_cleanup, ntb_qp_link_cleanup_work);
 939
 940         spin_lock_init(&qp->ntb_rx_pend_q_lock);
 941         spin_lock_init(&qp->ntb_rx_free_q_lock);
 942         spin_lock_init(&qp->ntb_tx_free_q_lock);
 943
 944         INIT_LIST_HEAD(&qp->rx_pend_q);
 945         INIT_LIST_HEAD(&qp->rx_free_q);
 946         INIT_LIST_HEAD(&qp->tx_free_q);
 947
 948         tasklet_init(&qp->rxc_db_work, ntb_transport_rxc_db,
 949                      (unsigned long)qp);
 950
 951         return 0;
 952 }
 953
 954 static int ntb_transport_probe(struct ntb_client *self, struct ntb_dev *ndev)
 955 {
 956         struct ntb_transport_ctx *nt;
 957         struct ntb_transport_mw *mw;
 958         unsigned int mw_count, qp_count;
 959         u64 qp_bitmap;
 960         int node;
 961         int rc, i;
 962
 963         if (ntb_db_is_unsafe(ndev))
 964                 dev_dbg(&ndev->dev,
 965                         "doorbell is unsafe, proceed anyway...\n");
 966         if (ntb_spad_is_unsafe(ndev))
 967                 dev_dbg(&ndev->dev,
 968                         "scratchpad is unsafe, proceed anyway...\n");
 969
 970         node = dev_to_node(&ndev->dev);
 971
 972         nt = kzalloc_node(sizeof(*nt), GFP_KERNEL, node);
 973         if (!nt)
 974                 return -ENOMEM;
 975
 976         nt->ndev = ndev;
 977
 978         mw_count = ntb_mw_count(ndev);
 979
 980         nt->mw_count = mw_count;
 981
 982         nt->mw_vec = kzalloc_node(mw_count * sizeof(*nt->mw_vec),
 983                                   GFP_KERNEL, node);
 984         if (!nt->mw_vec) {
 985                 rc = -ENOMEM;
 986                 goto err;
 987         }
 988
 989         for (i = 0; i < mw_count; i++) {
 990                 mw = &nt->mw_vec[i];
 991
 992                 rc = ntb_mw_get_range(ndev, i, &mw->phys_addr, &mw->phys_size,
 993                                       &mw->xlat_align, &mw->xlat_align_size);
 994                 if (rc)
 995                         goto err1;
 996
 997                 mw->vbase = ioremap_wc(mw->phys_addr, mw->phys_size);
 998                 if (!mw->vbase) {
 999                         rc = -ENOMEM;
1000                         goto err1;
1001                 }
1002
1003                 mw->buff_size = 0;
1004                 mw->xlat_size = 0;
1005                 mw->virt_addr = NULL;
1006                 mw->dma_addr = 0;
1007         }
1008
1009         qp_bitmap = ntb_db_valid_mask(ndev);
1010
1011         qp_count = ilog2(qp_bitmap);
1012         if (max_num_clients && max_num_clients < qp_count)
1013                 qp_count = max_num_clients;
1014         else if (mw_count < qp_count)
1015                 qp_count = mw_count;
1016
1017         qp_bitmap &= BIT_ULL(qp_count) - 1;
1018
1019         nt->qp_count = qp_count;
1020         nt->qp_bitmap = qp_bitmap;
1021         nt->qp_bitmap_free = qp_bitmap;
1022
1023         nt->qp_vec = kzalloc_node(qp_count * sizeof(*nt->qp_vec),
1024                                   GFP_KERNEL, node);
1025         if (!nt->qp_vec) {
1026                 rc = -ENOMEM;
1027                 goto err2;
1028         }
1029
1030         for (i = 0; i < qp_count; i++) {
1031                 rc = ntb_transport_init_queue(nt, i);
1032                 if (rc)
1033                         goto err3;
1034         }
1035
1036         INIT_DELAYED_WORK(&nt->link_work, ntb_transport_link_work);
1037         INIT_WORK(&nt->link_cleanup, ntb_transport_link_cleanup_work);
1038
1039         rc = ntb_set_ctx(ndev, nt, &ntb_transport_ops);
1040         if (rc)
1041                 goto err3;
1042
1043         INIT_LIST_HEAD(&nt->client_devs);
1044         rc = ntb_bus_init(nt);
1045         if (rc)
1046                 goto err4;
1047
1048         nt->link_is_up = false;
1049         ntb_link_enable(ndev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
1050         ntb_link_event(ndev);
1051
1052         return 0;
1053
1054 err4:
1055         ntb_clear_ctx(ndev);
1056 err3:
1057         kfree(nt->qp_vec);
1058 err2:
1059         kfree(nt->mw_vec);
1060 err1:
1061         while (i--) {
1062                 mw = &nt->mw_vec[i];
1063                 iounmap(mw->vbase);
1064         }
1065 err:
1066         kfree(nt);
1067         return rc;
1068 }
1069
1070 static void ntb_transport_free(struct ntb_client *self, struct ntb_dev *ndev)
1071 {
1072         struct ntb_transport_ctx *nt = ndev->ctx;
1073         struct ntb_transport_qp *qp;
1074         u64 qp_bitmap_alloc;
1075         int i;
1076
1077         ntb_transport_link_cleanup(nt);
1078         cancel_work_sync(&nt->link_cleanup);
1079         cancel_delayed_work_sync(&nt->link_work);
1080
1081         qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
1082
1083         /* verify that all the qp's are freed */
1084         for (i = 0; i < nt->qp_count; i++) {
1085                 qp = &nt->qp_vec[i];
1086                 if (qp_bitmap_alloc & BIT_ULL(i))
1087                         ntb_transport_free_queue(qp);
1088                 debugfs_remove_recursive(qp->debugfs_dir);
1089         }
1090
1091         ntb_link_disable(ndev);
1092         ntb_clear_ctx(ndev);
1093
1094         ntb_bus_remove(nt);
1095
1096         for (i = nt->mw_count; i--; ) {
1097                 ntb_free_mw(nt, i);
1098                 iounmap(nt->mw_vec[i].vbase);
1099         }
1100
1101         kfree(nt->qp_vec);
1102         kfree(nt->mw_vec);
1103         kfree(nt);
1104 }
1105
1106 static void ntb_rx_copy_callback(void *data)
1107 {
1108         struct ntb_queue_entry *entry = data;
1109         struct ntb_transport_qp *qp = entry->qp;
1110         void *cb_data = entry->cb_data;
1111         unsigned int len = entry->len;
1112         struct ntb_payload_header *hdr = entry->rx_hdr;
1113
1114         hdr->flags = 0;
1115
1116         iowrite32(entry->index, &qp->rx_info->entry);
1117
1118         ntb_list_add(&qp->ntb_rx_free_q_lock, &entry->entry, &qp->rx_free_q);
1119
1120         if (qp->rx_handler && qp->client_ready)
1121                 qp->rx_handler(qp, qp->cb_data, cb_data, len);
1122 }
1123
1124 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset)
1125 {
1126         void *buf = entry->buf;
1127         size_t len = entry->len;
1128
1129         memcpy(buf, offset, len);
1130
1131         /* Ensure that the data is fully copied out before clearing the flag */
1132         wmb();
1133
1134         ntb_rx_copy_callback(entry);
1135 }
1136
1137 static void ntb_async_rx(struct ntb_queue_entry *entry, void *offset,
1138                          size_t len)
1139 {
1140         struct dma_async_tx_descriptor *txd;
1141         struct ntb_transport_qp *qp = entry->qp;
1142         struct dma_chan *chan = qp->dma_chan;
1143         struct dma_device *device;
1144         size_t pay_off, buff_off;
1145         struct dmaengine_unmap_data *unmap;
1146         dma_cookie_t cookie;
1147         void *buf = entry->buf;
1148
1149         entry->len = len;
1150
1151         if (!chan)
1152                 goto err;
1153
1154         if (len < copy_bytes)
1155                 goto err_wait;
1156
1157         device = chan->device;
1158         pay_off = (size_t)offset & ~PAGE_MASK;
1159         buff_off = (size_t)buf & ~PAGE_MASK;
1160
1161         if (!is_dma_copy_aligned(device, pay_off, buff_off, len))
1162                 goto err_wait;
1163
1164         unmap = dmaengine_get_unmap_data(device->dev, 2, GFP_NOWAIT);
1165         if (!unmap)
1166                 goto err_wait;
1167
1168         unmap->len = len;
1169         unmap->addr[0] = dma_map_page(device->dev, virt_to_page(offset),
1170                                       pay_off, len, DMA_TO_DEVICE);
1171         if (dma_mapping_error(device->dev, unmap->addr[0]))
1172                 goto err_get_unmap;
1173
1174         unmap->to_cnt = 1;
1175
1176         unmap->addr[1] = dma_map_page(device->dev, virt_to_page(buf),
1177                                       buff_off, len, DMA_FROM_DEVICE);
1178         if (dma_mapping_error(device->dev, unmap->addr[1]))
1179                 goto err_get_unmap;
1180
1181         unmap->from_cnt = 1;
1182
1183         txd = device->device_prep_dma_memcpy(chan, unmap->addr[1],
1184                                              unmap->addr[0], len,
1185                                              DMA_PREP_INTERRUPT);
1186         if (!txd)
1187                 goto err_get_unmap;
1188
1189         txd->callback = ntb_rx_copy_callback;
1190         txd->callback_param = entry;
1191         dma_set_unmap(txd, unmap);
1192
1193         cookie = dmaengine_submit(txd);
1194         if (dma_submit_error(cookie))
1195                 goto err_set_unmap;
1196
1197         dmaengine_unmap_put(unmap);
1198
1199         qp->last_cookie = cookie;
1200
1201         qp->rx_async++;
1202
1203         return;
1204
1205 err_set_unmap:
1206         dmaengine_unmap_put(unmap);
1207 err_get_unmap:
1208         dmaengine_unmap_put(unmap);
1209 err_wait:
1210         /* If the callbacks come out of order, the writing of the index to the
1211          * last completed will be out of order.  This may result in the
1212          * receive stalling forever.
1213          */
1214         dma_sync_wait(chan, qp->last_cookie);
1215 err:
1216         ntb_memcpy_rx(entry, offset);
1217         qp->rx_memcpy++;
1218 }
1219
1220 static int ntb_process_rxc(struct ntb_transport_qp *qp)
1221 {
1222         struct ntb_payload_header *hdr;
1223         struct ntb_queue_entry *entry;
1224         void *offset;
1225         int rc;
1226
1227         offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
1228         hdr = offset + qp->rx_max_frame - sizeof(struct ntb_payload_header);
1229
1230         dev_dbg(&qp->ndev->pdev->dev, "qp %d: RX ver %u len %d flags %x\n",
1231                 qp->qp_num, hdr->ver, hdr->len, hdr->flags);
1232
1233         if (!(hdr->flags & DESC_DONE_FLAG)) {
1234                 dev_dbg(&qp->ndev->pdev->dev, "done flag not set\n");
1235                 qp->rx_ring_empty++;
1236                 return -EAGAIN;
1237         }
1238
1239         if (hdr->flags & LINK_DOWN_FLAG) {
1240                 dev_dbg(&qp->ndev->pdev->dev, "link down flag set\n");
1241                 ntb_qp_link_down(qp);
1242                 hdr->flags = 0;
1243                 return -EAGAIN;
1244         }
1245
1246         if (hdr->ver != (u32)qp->rx_pkts) {
1247                 dev_dbg(&qp->ndev->pdev->dev,
1248                         "version mismatch, expected %llu - got %u\n",
1249                         qp->rx_pkts, hdr->ver);
1250                 qp->rx_err_ver++;
1251                 return -EIO;
1252         }
1253
1254         entry = ntb_list_rm(&qp->ntb_rx_pend_q_lock, &qp->rx_pend_q);
1255         if (!entry) {
1256                 dev_dbg(&qp->ndev->pdev->dev, "no receive buffer\n");
1257                 qp->rx_err_no_buf++;
1258
1259                 rc = -ENOMEM;
1260                 goto err;
1261         }
1262
1263         if (hdr->len > entry->len) {
1264                 dev_dbg(&qp->ndev->pdev->dev,
1265                         "receive buffer overflow! Wanted %d got %d\n",
1266                         hdr->len, entry->len);
1267                 qp->rx_err_oflow++;
1268
1269                 rc = -EIO;
1270                 goto err;
1271         }
1272
1273         dev_dbg(&qp->ndev->pdev->dev,
1274                 "RX OK index %u ver %u size %d into buf size %d\n",
1275                 qp->rx_index, hdr->ver, hdr->len, entry->len);
1276
1277         qp->rx_bytes += hdr->len;
1278         qp->rx_pkts++;
1279
1280         entry->index = qp->rx_index;
1281         entry->rx_hdr = hdr;
1282
1283         ntb_async_rx(entry, offset, hdr->len);
1284
1285         qp->rx_index++;
1286         qp->rx_index %= qp->rx_max_entry;
1287
1288         return 0;
1289
1290 err:
1291         /* FIXME: if this syncrhonous update of the rx_index gets ahead of
1292          * asyncrhonous ntb_rx_copy_callback of previous entry, there are three
1293          * scenarios:
1294          *
1295          * 1) The peer might miss this update, but observe the update
1296          * from the memcpy completion callback.  In this case, the buffer will
1297          * not be freed on the peer to be reused for a different packet.  The
1298          * successful rx of a later packet would clear the condition, but the
1299          * condition could persist if several rx fail in a row.
1300          *
1301          * 2) The peer may observe this update before the asyncrhonous copy of
1302          * prior packets is completed.  The peer may overwrite the buffers of
1303          * the prior packets before they are copied.
1304          *
1305          * 3) Both: the peer may observe the update, and then observe the index
1306          * decrement by the asynchronous completion callback.  Who knows what
1307          * badness that will cause.
1308          */
1309         hdr->flags = 0;
1310         iowrite32(qp->rx_index, &qp->rx_info->entry);
1311
1312         return rc;
1313 }
1314
1315 static void ntb_transport_rxc_db(unsigned long data)
1316 {
1317         struct ntb_transport_qp *qp = (void *)data;
1318         int rc, i;
1319
1320         dev_dbg(&qp->ndev->pdev->dev, "%s: doorbell %d received\n",
1321                 __func__, qp->qp_num);
1322
1323         /* Limit the number of packets processed in a single interrupt to
1324          * provide fairness to others
1325          */
1326         for (i = 0; i < qp->rx_max_entry; i++) {
1327                 rc = ntb_process_rxc(qp);
1328                 if (rc)
1329                         break;
1330         }
1331
1332         if (qp->dma_chan)
1333                 dma_async_issue_pending(qp->dma_chan);
1334
1335         if (i == qp->rx_max_entry) {
1336                 /* there is more work to do */
1337                 tasklet_schedule(&qp->rxc_db_work);
1338         } else if (ntb_db_read(qp->ndev) & BIT_ULL(qp->qp_num)) {
1339                 /* the doorbell bit is set: clear it */
1340                 ntb_db_clear(qp->ndev, BIT_ULL(qp->qp_num));
1341                 /* ntb_db_read ensures ntb_db_clear write is committed */
1342                 ntb_db_read(qp->ndev);
1343
1344                 /* an interrupt may have arrived between finishing
1345                  * ntb_process_rxc and clearing the doorbell bit:
1346                  * there might be some more work to do.
1347                  */
1348                 tasklet_schedule(&qp->rxc_db_work);
1349         }
1350 }
1351
1352 static void ntb_tx_copy_callback(void *data)
1353 {
1354         struct ntb_queue_entry *entry = data;
1355         struct ntb_transport_qp *qp = entry->qp;
1356         struct ntb_payload_header __iomem *hdr = entry->tx_hdr;
1357
1358         iowrite32(entry->flags | DESC_DONE_FLAG, &hdr->flags);
1359
1360         ntb_peer_db_set(qp->ndev, BIT_ULL(qp->qp_num));
1361
1362         /* The entry length can only be zero if the packet is intended to be a
1363          * "link down" or similar.  Since no payload is being sent in these
1364          * cases, there is nothing to add to the completion queue.
1365          */
1366         if (entry->len > 0) {
1367                 qp->tx_bytes += entry->len;
1368
1369                 if (qp->tx_handler)
1370                         qp->tx_handler(qp, qp->cb_data, entry->cb_data,
1371                                        entry->len);
1372         }
1373
1374         ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry, &qp->tx_free_q);
1375 }
1376
1377 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset)
1378 {
1379 #ifdef ARCH_HAS_NOCACHE_UACCESS
1380         /*
1381          * Using non-temporal mov to improve performance on non-cached
1382          * writes, even though we aren't actually copying from user space.
1383          */
1384         __copy_from_user_inatomic_nocache(offset, entry->buf, entry->len);
1385 #else
1386         memcpy_toio(offset, entry->buf, entry->len);
1387 #endif
1388
1389         /* Ensure that the data is fully copied out before setting the flags */
1390         wmb();
1391
1392         ntb_tx_copy_callback(entry);
1393 }
1394
1395 static void ntb_async_tx(struct ntb_transport_qp *qp,
1396                          struct ntb_queue_entry *entry)
1397 {
1398         struct ntb_payload_header __iomem *hdr;
1399         struct dma_async_tx_descriptor *txd;
1400         struct dma_chan *chan = qp->dma_chan;
1401         struct dma_device *device;
1402         size_t dest_off, buff_off;
1403         struct dmaengine_unmap_data *unmap;
1404         dma_addr_t dest;
1405         dma_cookie_t cookie;
1406         void __iomem *offset;
1407         size_t len = entry->len;
1408         void *buf = entry->buf;
1409
1410         offset = qp->tx_mw + qp->tx_max_frame * qp->tx_index;
1411         hdr = offset + qp->tx_max_frame - sizeof(struct ntb_payload_header);
1412         entry->tx_hdr = hdr;
1413
1414         iowrite32(entry->len, &hdr->len);
1415         iowrite32((u32)qp->tx_pkts, &hdr->ver);
1416
1417         if (!chan)
1418                 goto err;
1419
1420         if (len < copy_bytes)
1421                 goto err;
1422
1423         device = chan->device;
1424         dest = qp->tx_mw_phys + qp->tx_max_frame * qp->tx_index;
1425         buff_off = (size_t)buf & ~PAGE_MASK;
1426         dest_off = (size_t)dest & ~PAGE_MASK;
1427
1428         if (!is_dma_copy_aligned(device, buff_off, dest_off, len))
1429                 goto err;
1430
1431         unmap = dmaengine_get_unmap_data(device->dev, 1, GFP_NOWAIT);
1432         if (!unmap)
1433                 goto err;
1434
1435         unmap->len = len;
1436         unmap->addr[0] = dma_map_page(device->dev, virt_to_page(buf),
1437                                       buff_off, len, DMA_TO_DEVICE);
1438         if (dma_mapping_error(device->dev, unmap->addr[0]))
1439                 goto err_get_unmap;
1440
1441         unmap->to_cnt = 1;
1442
1443         txd = device->device_prep_dma_memcpy(chan, dest, unmap->addr[0], len,
1444                                              DMA_PREP_INTERRUPT);
1445         if (!txd)
1446                 goto err_get_unmap;
1447
1448         txd->callback = ntb_tx_copy_callback;
1449         txd->callback_param = entry;
1450         dma_set_unmap(txd, unmap);
1451
1452         cookie = dmaengine_submit(txd);
1453         if (dma_submit_error(cookie))
1454                 goto err_set_unmap;
1455
1456         dmaengine_unmap_put(unmap);
1457
1458         dma_async_issue_pending(chan);
1459         qp->tx_async++;
1460
1461         return;
1462 err_set_unmap:
1463         dmaengine_unmap_put(unmap);
1464 err_get_unmap:
1465         dmaengine_unmap_put(unmap);
1466 err:
1467         ntb_memcpy_tx(entry, offset);
1468         qp->tx_memcpy++;
1469 }
1470
1471 static int ntb_process_tx(struct ntb_transport_qp *qp,
1472                           struct ntb_queue_entry *entry)
1473 {
1474         if (qp->tx_index == qp->remote_rx_info->entry) {
1475                 qp->tx_ring_full++;
1476                 return -EAGAIN;
1477         }
1478
1479         if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
1480                 if (qp->tx_handler)
1481                         qp->tx_handler(qp->cb_data, qp, NULL, -EIO);
1482
1483                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1484                              &qp->tx_free_q);
1485                 return 0;
1486         }
1487
1488         ntb_async_tx(qp, entry);
1489
1490         qp->tx_index++;
1491         qp->tx_index %= qp->tx_max_entry;
1492
1493         qp->tx_pkts++;
1494
1495         return 0;
1496 }
1497
1498 static void ntb_send_link_down(struct ntb_transport_qp *qp)
1499 {
1500         struct pci_dev *pdev = qp->ndev->pdev;
1501         struct ntb_queue_entry *entry;
1502         int i, rc;
1503
1504         if (!qp->link_is_up)
1505                 return;
1506
1507         dev_info(&pdev->dev, "qp %d: Send Link Down\n", qp->qp_num);
1508
1509         for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1510                 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1511                 if (entry)
1512                         break;
1513                 msleep(100);
1514         }
1515
1516         if (!entry)
1517                 return;
1518
1519         entry->cb_data = NULL;
1520         entry->buf = NULL;
1521         entry->len = 0;
1522         entry->flags = LINK_DOWN_FLAG;
1523
1524         rc = ntb_process_tx(qp, entry);
1525         if (rc)
1526                 dev_err(&pdev->dev, "ntb: QP%d unable to send linkdown msg\n",
1527                         qp->qp_num);
1528
1529         ntb_qp_link_down_reset(qp);
1530 }
1531
1532 static bool ntb_dma_filter_fn(struct dma_chan *chan, void *node)
1533 {
1534         return dev_to_node(&chan->dev->device) == (int)(unsigned long)node;
1535 }
1536
1537 /**
1538  * ntb_transport_create_queue - Create a new NTB transport layer queue
1539  * @rx_handler: receive callback function
1540  * @tx_handler: transmit callback function
1541  * @event_handler: event callback function
1542  *
1543  * Create a new NTB transport layer queue and provide the queue with a callback
1544  * routine for both transmit and receive.  The receive callback routine will be
1545  * used to pass up data when the transport has received it on the queue.   The
1546  * transmit callback routine will be called when the transport has completed the
1547  * transmission of the data on the queue and the data is ready to be freed.
1548  *
1549  * RETURNS: pointer to newly created ntb_queue, NULL on error.
1550  */
1551 struct ntb_transport_qp *
1552 ntb_transport_create_queue(void *data, struct device *client_dev,
1553                            const struct ntb_queue_handlers *handlers)
1554 {
1555         struct ntb_dev *ndev;
1556         struct pci_dev *pdev;
1557         struct ntb_transport_ctx *nt;
1558         struct ntb_queue_entry *entry;
1559         struct ntb_transport_qp *qp;
1560         u64 qp_bit;
1561         unsigned int free_queue;
1562         dma_cap_mask_t dma_mask;
1563         int node;
1564         int i;
1565
1566         ndev = dev_ntb(client_dev->parent);
1567         pdev = ndev->pdev;
1568         nt = ndev->ctx;
1569
1570         node = dev_to_node(&ndev->dev);
1571
1572         free_queue = ffs(nt->qp_bitmap);
1573         if (!free_queue)
1574                 goto err;
1575
1576         /* decrement free_queue to make it zero based */
1577         free_queue--;
1578
1579         qp = &nt->qp_vec[free_queue];
1580         qp_bit = BIT_ULL(qp->qp_num);
1581
1582         nt->qp_bitmap_free &= ~qp_bit;
1583
1584         qp->cb_data = data;
1585         qp->rx_handler = handlers->rx_handler;
1586         qp->tx_handler = handlers->tx_handler;
1587         qp->event_handler = handlers->event_handler;
1588
1589         dma_cap_zero(dma_mask);
1590         dma_cap_set(DMA_MEMCPY, dma_mask);
1591
1592         qp->dma_chan = dma_request_channel(dma_mask, ntb_dma_filter_fn,
1593                                            (void *)(unsigned long)node);
1594         if (!qp->dma_chan)
1595                 dev_info(&pdev->dev, "Unable to allocate DMA channel, using CPU instead\n");
1596
1597         for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
1598                 entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
1599                 if (!entry)
1600                         goto err1;
1601
1602                 entry->qp = qp;
1603                 ntb_list_add(&qp->ntb_rx_free_q_lock, &entry->entry,
1604                              &qp->rx_free_q);
1605         }
1606
1607         for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
1608                 entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
1609                 if (!entry)
1610                         goto err2;
1611
1612                 entry->qp = qp;
1613                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1614                              &qp->tx_free_q);
1615         }
1616
1617         ntb_db_clear(qp->ndev, qp_bit);
1618         ntb_db_clear_mask(qp->ndev, qp_bit);
1619
1620         dev_info(&pdev->dev, "NTB Transport QP %d created\n", qp->qp_num);
1621
1622         return qp;
1623
1624 err2:
1625         while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1626                 kfree(entry);
1627 err1:
1628         while ((entry = ntb_list_rm(&qp->ntb_rx_free_q_lock, &qp->rx_free_q)))
1629                 kfree(entry);
1630         if (qp->dma_chan)
1631                 dma_release_channel(qp->dma_chan);
1632         nt->qp_bitmap_free |= qp_bit;
1633 err:
1634         return NULL;
1635 }
1636 EXPORT_SYMBOL_GPL(ntb_transport_create_queue);
1637
1638 /**
1639  * ntb_transport_free_queue - Frees NTB transport queue
1640  * @qp: NTB queue to be freed
1641  *
1642  * Frees NTB transport queue
1643  */
1644 void ntb_transport_free_queue(struct ntb_transport_qp *qp)
1645 {
1646         struct ntb_transport_ctx *nt = qp->transport;
1647         struct pci_dev *pdev;
1648         struct ntb_queue_entry *entry;
1649         u64 qp_bit;
1650
1651         if (!qp)
1652                 return;
1653
1654         pdev = qp->ndev->pdev;
1655
1656         if (qp->dma_chan) {
1657                 struct dma_chan *chan = qp->dma_chan;
1658                 /* Putting the dma_chan to NULL will force any new traffic to be
1659                  * processed by the CPU instead of the DAM engine
1660                  */
1661                 qp->dma_chan = NULL;
1662
1663                 /* Try to be nice and wait for any queued DMA engine
1664                  * transactions to process before smashing it with a rock
1665                  */
1666                 dma_sync_wait(chan, qp->last_cookie);
1667                 dmaengine_terminate_all(chan);
1668                 dma_release_channel(chan);
1669         }
1670
1671         qp_bit = BIT_ULL(qp->qp_num);
1672
1673         ntb_db_set_mask(qp->ndev, qp_bit);
1674         tasklet_disable(&qp->rxc_db_work);
1675
1676         cancel_delayed_work_sync(&qp->link_work);
1677
1678         qp->cb_data = NULL;
1679         qp->rx_handler = NULL;
1680         qp->tx_handler = NULL;
1681         qp->event_handler = NULL;
1682
1683         while ((entry = ntb_list_rm(&qp->ntb_rx_free_q_lock, &qp->rx_free_q)))
1684                 kfree(entry);
1685
1686         while ((entry = ntb_list_rm(&qp->ntb_rx_pend_q_lock, &qp->rx_pend_q))) {
1687                 dev_warn(&pdev->dev, "Freeing item from a non-empty queue\n");
1688                 kfree(entry);
1689         }
1690
1691         while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1692                 kfree(entry);
1693
1694         nt->qp_bitmap_free |= qp_bit;
1695
1696         dev_info(&pdev->dev, "NTB Transport QP %d freed\n", qp->qp_num);
1697 }
1698 EXPORT_SYMBOL_GPL(ntb_transport_free_queue);
1699
1700 /**
1701  * ntb_transport_rx_remove - Dequeues enqueued rx packet
1702  * @qp: NTB queue to be freed
1703  * @len: pointer to variable to write enqueued buffers length
1704  *
1705  * Dequeues unused buffers from receive queue.  Should only be used during
1706  * shutdown of qp.
1707  *
1708  * RETURNS: NULL error value on error, or void* for success.
1709  */
1710 void *ntb_transport_rx_remove(struct ntb_transport_qp *qp, unsigned int *len)
1711 {
1712         struct ntb_queue_entry *entry;
1713         void *buf;
1714
1715         if (!qp || qp->client_ready)
1716                 return NULL;
1717
1718         entry = ntb_list_rm(&qp->ntb_rx_pend_q_lock, &qp->rx_pend_q);
1719         if (!entry)
1720                 return NULL;
1721
1722         buf = entry->cb_data;
1723         *len = entry->len;
1724
1725         ntb_list_add(&qp->ntb_rx_free_q_lock, &entry->entry, &qp->rx_free_q);
1726
1727         return buf;
1728 }
1729 EXPORT_SYMBOL_GPL(ntb_transport_rx_remove);
1730
1731 /**
1732  * ntb_transport_rx_enqueue - Enqueue a new NTB queue entry
1733  * @qp: NTB transport layer queue the entry is to be enqueued on
1734  * @cb: per buffer pointer for callback function to use
1735  * @data: pointer to data buffer that incoming packets will be copied into
1736  * @len: length of the data buffer
1737  *
1738  * Enqueue a new receive buffer onto the transport queue into which a NTB
1739  * payload can be received into.
1740  *
1741  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
1742  */
1743 int ntb_transport_rx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
1744                              unsigned int len)
1745 {
1746         struct ntb_queue_entry *entry;
1747
1748         if (!qp)
1749                 return -EINVAL;
1750
1751         entry = ntb_list_rm(&qp->ntb_rx_free_q_lock, &qp->rx_free_q);
1752         if (!entry)
1753                 return -ENOMEM;
1754
1755         entry->cb_data = cb;
1756         entry->buf = data;
1757         entry->len = len;
1758
1759         ntb_list_add(&qp->ntb_rx_pend_q_lock, &entry->entry, &qp->rx_pend_q);
1760
1761         return 0;
1762 }
1763 EXPORT_SYMBOL_GPL(ntb_transport_rx_enqueue);
1764
1765 /**
1766  * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
1767  * @qp: NTB transport layer queue the entry is to be enqueued on
1768  * @cb: per buffer pointer for callback function to use
1769  * @data: pointer to data buffer that will be sent
1770  * @len: length of the data buffer
1771  *
1772  * Enqueue a new transmit buffer onto the transport queue from which a NTB
1773  * payload will be transmitted.  This assumes that a lock is being held to
1774  * serialize access to the qp.
1775  *
1776  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
1777  */
1778 int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
1779                              unsigned int len)
1780 {
1781         struct ntb_queue_entry *entry;
1782         int rc;
1783
1784         if (!qp || !qp->link_is_up || !len)
1785                 return -EINVAL;
1786
1787         entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1788         if (!entry) {
1789                 qp->tx_err_no_buf++;
1790                 return -ENOMEM;
1791         }
1792
1793         entry->cb_data = cb;
1794         entry->buf = data;
1795         entry->len = len;
1796         entry->flags = 0;
1797
1798         rc = ntb_process_tx(qp, entry);
1799         if (rc)
1800                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1801                              &qp->tx_free_q);
1802
1803         return rc;
1804 }
1805 EXPORT_SYMBOL_GPL(ntb_transport_tx_enqueue);
1806
1807 /**
1808  * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
1809  * @qp: NTB transport layer queue to be enabled
1810  *
1811  * Notify NTB transport layer of client readiness to use queue
1812  */
1813 void ntb_transport_link_up(struct ntb_transport_qp *qp)
1814 {
1815         if (!qp)
1816                 return;
1817
1818         qp->client_ready = true;
1819
1820         if (qp->transport->link_is_up)
1821                 schedule_delayed_work(&qp->link_work, 0);
1822 }
1823 EXPORT_SYMBOL_GPL(ntb_transport_link_up);
1824
1825 /**
1826  * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
1827  * @qp: NTB transport layer queue to be disabled
1828  *
1829  * Notify NTB transport layer of client's desire to no longer receive data on
1830  * transport queue specified.  It is the client's responsibility to ensure all
1831  * entries on queue are purged or otherwise handled appropriately.
1832  */
1833 void ntb_transport_link_down(struct ntb_transport_qp *qp)
1834 {
1835         struct pci_dev *pdev;
1836         int val;
1837
1838         if (!qp)
1839                 return;
1840
1841         pdev = qp->ndev->pdev;
1842         qp->client_ready = false;
1843
1844         val = ntb_spad_read(qp->ndev, QP_LINKS);
1845
1846         ntb_peer_spad_write(qp->ndev, QP_LINKS,
1847                             val & ~BIT(qp->qp_num));
1848
1849         if (qp->link_is_up)
1850                 ntb_send_link_down(qp);
1851         else
1852                 cancel_delayed_work_sync(&qp->link_work);
1853 }
1854 EXPORT_SYMBOL_GPL(ntb_transport_link_down);
1855
1856 /**
1857  * ntb_transport_link_query - Query transport link state
1858  * @qp: NTB transport layer queue to be queried
1859  *
1860  * Query connectivity to the remote system of the NTB transport queue
1861  *
1862  * RETURNS: true for link up or false for link down
1863  */
1864 bool ntb_transport_link_query(struct ntb_transport_qp *qp)
1865 {
1866         if (!qp)
1867                 return false;
1868
1869         return qp->link_is_up;
1870 }
1871 EXPORT_SYMBOL_GPL(ntb_transport_link_query);
1872
1873 /**
1874  * ntb_transport_qp_num - Query the qp number
1875  * @qp: NTB transport layer queue to be queried
1876  *
1877  * Query qp number of the NTB transport queue
1878  *
1879  * RETURNS: a zero based number specifying the qp number
1880  */
1881 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
1882 {
1883         if (!qp)
1884                 return 0;
1885
1886         return qp->qp_num;
1887 }
1888 EXPORT_SYMBOL_GPL(ntb_transport_qp_num);
1889
1890 /**
1891  * ntb_transport_max_size - Query the max payload size of a qp
1892  * @qp: NTB transport layer queue to be queried
1893  *
1894  * Query the maximum payload size permissible on the given qp
1895  *
1896  * RETURNS: the max payload size of a qp
1897  */
1898 unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp)
1899 {
1900         unsigned int max;
1901
1902         if (!qp)
1903                 return 0;
1904
1905         if (!qp->dma_chan)
1906                 return qp->tx_max_frame - sizeof(struct ntb_payload_header);
1907
1908         /* If DMA engine usage is possible, try to find the max size for that */
1909         max = qp->tx_max_frame - sizeof(struct ntb_payload_header);
1910         max -= max % (1 << qp->dma_chan->device->copy_align);
1911
1912         return max;
1913 }
1914 EXPORT_SYMBOL_GPL(ntb_transport_max_size);
1915
1916 static void ntb_transport_doorbell_callback(void *data, int vector)
1917 {
1918         struct ntb_transport_ctx *nt = data;
1919         struct ntb_transport_qp *qp;
1920         u64 db_bits;
1921         unsigned int qp_num;
1922
1923         db_bits = (nt->qp_bitmap & ~nt->qp_bitmap_free &
1924                    ntb_db_vector_mask(nt->ndev, vector));
1925
1926         while (db_bits) {
1927                 qp_num = __ffs(db_bits);
1928                 qp = &nt->qp_vec[qp_num];
1929
1930                 tasklet_schedule(&qp->rxc_db_work);
1931
1932                 db_bits &= ~BIT_ULL(qp_num);
1933         }
1934 }
1935
1936 static const struct ntb_ctx_ops ntb_transport_ops = {
1937         .link_event = ntb_transport_event_callback,
1938         .db_event = ntb_transport_doorbell_callback,
1939 };
1940
1941 static struct ntb_client ntb_transport_client = {
1942         .ops = {
1943                 .probe = ntb_transport_probe,
1944                 .remove = ntb_transport_free,
1945         },
1946 };
1947
1948 static int __init ntb_transport_init(void)
1949 {
1950         int rc;
1951
1952         if (debugfs_initialized())
1953                 nt_debugfs_dir = debugfs_create_dir(KBUILD_MODNAME, NULL);
1954
1955         rc = bus_register(&ntb_transport_bus);
1956         if (rc)
1957                 goto err_bus;
1958
1959         rc = ntb_register_client(&ntb_transport_client);
1960         if (rc)
1961                 goto err_client;
1962
1963         return 0;
1964
1965 err_client:
1966         bus_unregister(&ntb_transport_bus);
1967 err_bus:
1968         debugfs_remove_recursive(nt_debugfs_dir);
1969         return rc;
1970 }
1971 module_init(ntb_transport_init);
1972
1973 static void __exit ntb_transport_exit(void)
1974 {
1975         debugfs_remove_recursive(nt_debugfs_dir);
1976
1977         ntb_unregister_client(&ntb_transport_client);
1978         bus_unregister(&ntb_transport_bus);
1979 }
1980 module_exit(ntb_transport_exit);