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