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 = 0x10000;
  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 bool use_dma;
  92 module_param(use_dma, bool, 0644);
  93 MODULE_PARM_DESC(use_dma, "Use DMA engine to perform large data copy");
  94
  95 static struct dentry *nt_debugfs_dir;
  96
  97 struct ntb_queue_entry {
  98         /* ntb_queue list reference */
  99         struct list_head entry;
 100         /* pointers to data to be transferred */
 101         void *cb_data;
 102         void *buf;
 103         unsigned int len;
 104         unsigned int flags;
 105
 106         struct ntb_transport_qp *qp;
 107         union {
 108                 struct ntb_payload_header __iomem *tx_hdr;
 109                 struct ntb_payload_header *rx_hdr;
 110         };
 111         unsigned int index;
 112 };
 113
 114 struct ntb_rx_info {
 115         unsigned int entry;
 116 };
 117
 118 struct ntb_transport_qp {
 119         struct ntb_transport_ctx *transport;
 120         struct ntb_dev *ndev;
 121         void *cb_data;
 122         struct dma_chan *dma_chan;
 123
 124         bool client_ready;
 125         bool link_is_up;
 126
 127         u8 qp_num;      /* Only 64 QP's are allowed.  0-63 */
 128         u64 qp_bit;
 129
 130         struct ntb_rx_info __iomem *rx_info;
 131         struct ntb_rx_info *remote_rx_info;
 132
 133         void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data,
 134                            void *data, int len);
 135         struct list_head tx_free_q;
 136         spinlock_t ntb_tx_free_q_lock;
 137         void __iomem *tx_mw;
 138         dma_addr_t tx_mw_phys;
 139         unsigned int tx_index;
 140         unsigned int tx_max_entry;
 141         unsigned int tx_max_frame;
 142
 143         void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
 144                            void *data, int len);
 145         struct list_head rx_post_q;
 146         struct list_head rx_pend_q;
 147         struct list_head rx_free_q;
 148         /* ntb_rx_q_lock: synchronize access to rx_XXXX_q */
 149         spinlock_t ntb_rx_q_lock;
 150         void *rx_buff;
 151         unsigned int rx_index;
 152         unsigned int rx_max_entry;
 153         unsigned int rx_max_frame;
 154         dma_cookie_t last_cookie;
 155         struct tasklet_struct rxc_db_work;
 156
 157         void (*event_handler)(void *data, int status);
 158         struct delayed_work link_work;
 159         struct work_struct link_cleanup;
 160
 161         struct dentry *debugfs_dir;
 162         struct dentry *debugfs_stats;
 163
 164         /* Stats */
 165         u64 rx_bytes;
 166         u64 rx_pkts;
 167         u64 rx_ring_empty;
 168         u64 rx_err_no_buf;
 169         u64 rx_err_oflow;
 170         u64 rx_err_ver;
 171         u64 rx_memcpy;
 172         u64 rx_async;
 173         u64 tx_bytes;
 174         u64 tx_pkts;
 175         u64 tx_ring_full;
 176         u64 tx_err_no_buf;
 177         u64 tx_memcpy;
 178         u64 tx_async;
 179 };
 180
 181 struct ntb_transport_mw {
 182         phys_addr_t phys_addr;
 183         resource_size_t phys_size;
 184         resource_size_t xlat_align;
 185         resource_size_t xlat_align_size;
 186         void __iomem *vbase;
 187         size_t xlat_size;
 188         size_t buff_size;
 189         void *virt_addr;
 190         dma_addr_t dma_addr;
 191 };
 192
 193 struct ntb_transport_client_dev {
 194         struct list_head entry;
 195         struct ntb_transport_ctx *nt;
 196         struct device dev;
 197 };
 198
 199 struct ntb_transport_ctx {
 200         struct list_head entry;
 201         struct list_head client_devs;
 202
 203         struct ntb_dev *ndev;
 204
 205         struct ntb_transport_mw *mw_vec;
 206         struct ntb_transport_qp *qp_vec;
 207         unsigned int mw_count;
 208         unsigned int qp_count;
 209         u64 qp_bitmap;
 210         u64 qp_bitmap_free;
 211
 212         bool link_is_up;
 213         struct delayed_work link_work;
 214         struct work_struct link_cleanup;
 215
 216         struct dentry *debugfs_node_dir;
 217 };
 218
 219 enum {
 220         DESC_DONE_FLAG = BIT(0),
 221         LINK_DOWN_FLAG = BIT(1),
 222 };
 223
 224 struct ntb_payload_header {
 225         unsigned int ver;
 226         unsigned int len;
 227         unsigned int flags;
 228 };
 229
 230 enum {
 231         VERSION = 0,
 232         QP_LINKS,
 233         NUM_QPS,
 234         NUM_MWS,
 235         MW0_SZ_HIGH,
 236         MW0_SZ_LOW,
 237         MW1_SZ_HIGH,
 238         MW1_SZ_LOW,
 239         MAX_SPAD,
 240 };
 241
 242 #define dev_client_dev(__dev) \
 243         container_of((__dev), struct ntb_transport_client_dev, dev)
 244
 245 #define drv_client(__drv) \
 246         container_of((__drv), struct ntb_transport_client, driver)
 247
 248 #define QP_TO_MW(nt, qp)        ((qp) % nt->mw_count)
 249 #define NTB_QP_DEF_NUM_ENTRIES  100
 250 #define NTB_LINK_DOWN_TIMEOUT   10
 251
 252 static void ntb_transport_rxc_db(unsigned long data);
 253 static const struct ntb_ctx_ops ntb_transport_ops;
 254 static struct ntb_client ntb_transport_client;
 255
 256 static int ntb_transport_bus_match(struct device *dev,
 257                                    struct device_driver *drv)
 258 {
 259         return !strncmp(dev_name(dev), drv->name, strlen(drv->name));
 260 }
 261
 262 static int ntb_transport_bus_probe(struct device *dev)
 263 {
 264         const struct ntb_transport_client *client;
 265         int rc = -EINVAL;
 266
 267         get_device(dev);
 268
 269         client = drv_client(dev->driver);
 270         rc = client->probe(dev);
 271         if (rc)
 272                 put_device(dev);
 273
 274         return rc;
 275 }
 276
 277 static int ntb_transport_bus_remove(struct device *dev)
 278 {
 279         const struct ntb_transport_client *client;
 280
 281         client = drv_client(dev->driver);
 282         client->remove(dev);
 283
 284         put_device(dev);
 285
 286         return 0;
 287 }
 288
 289 static struct bus_type ntb_transport_bus = {
 290         .name = "ntb_transport",
 291         .match = ntb_transport_bus_match,
 292         .probe = ntb_transport_bus_probe,
 293         .remove = ntb_transport_bus_remove,
 294 };
 295
 296 static LIST_HEAD(ntb_transport_list);
 297
 298 static int ntb_bus_init(struct ntb_transport_ctx *nt)
 299 {
 300         list_add(&nt->entry, &ntb_transport_list);
 301         return 0;
 302 }
 303
 304 static void ntb_bus_remove(struct ntb_transport_ctx *nt)
 305 {
 306         struct ntb_transport_client_dev *client_dev, *cd;
 307
 308         list_for_each_entry_safe(client_dev, cd, &nt->client_devs, entry) {
 309                 dev_err(client_dev->dev.parent, "%s still attached to bus, removing\n",
 310                         dev_name(&client_dev->dev));
 311                 list_del(&client_dev->entry);
 312                 device_unregister(&client_dev->dev);
 313         }
 314
 315         list_del(&nt->entry);
 316 }
 317
 318 static void ntb_transport_client_release(struct device *dev)
 319 {
 320         struct ntb_transport_client_dev *client_dev;
 321
 322         client_dev = dev_client_dev(dev);
 323         kfree(client_dev);
 324 }
 325
 326 /**
 327  * ntb_transport_unregister_client_dev - Unregister NTB client device
 328  * @device_name: Name of NTB client device
 329  *
 330  * Unregister an NTB client device with the NTB transport layer
 331  */
 332 void ntb_transport_unregister_client_dev(char *device_name)
 333 {
 334         struct ntb_transport_client_dev *client, *cd;
 335         struct ntb_transport_ctx *nt;
 336
 337         list_for_each_entry(nt, &ntb_transport_list, entry)
 338                 list_for_each_entry_safe(client, cd, &nt->client_devs, entry)
 339                         if (!strncmp(dev_name(&client->dev), device_name,
 340                                      strlen(device_name))) {
 341                                 list_del(&client->entry);
 342                                 device_unregister(&client->dev);
 343                         }
 344 }
 345 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client_dev);
 346
 347 /**
 348  * ntb_transport_register_client_dev - Register NTB client device
 349  * @device_name: Name of NTB client device
 350  *
 351  * Register an NTB client device with the NTB transport layer
 352  */
 353 int ntb_transport_register_client_dev(char *device_name)
 354 {
 355         struct ntb_transport_client_dev *client_dev;
 356         struct ntb_transport_ctx *nt;
 357         int node;
 358         int rc, i = 0;
 359
 360         if (list_empty(&ntb_transport_list))
 361                 return -ENODEV;
 362
 363         list_for_each_entry(nt, &ntb_transport_list, entry) {
 364                 struct device *dev;
 365
 366                 node = dev_to_node(&nt->ndev->dev);
 367
 368                 client_dev = kzalloc_node(sizeof(*client_dev),
 369                                           GFP_KERNEL, node);
 370                 if (!client_dev) {
 371                         rc = -ENOMEM;
 372                         goto err;
 373                 }
 374
 375                 dev = &client_dev->dev;
 376
 377                 /* setup and register client devices */
 378                 dev_set_name(dev, "%s%d", device_name, i);
 379                 dev->bus = &ntb_transport_bus;
 380                 dev->release = ntb_transport_client_release;
 381                 dev->parent = &nt->ndev->dev;
 382
 383                 rc = device_register(dev);
 384                 if (rc) {
 385                         kfree(client_dev);
 386                         goto err;
 387                 }
 388
 389                 list_add_tail(&client_dev->entry, &nt->client_devs);
 390                 i++;
 391         }
 392
 393         return 0;
 394
 395 err:
 396         ntb_transport_unregister_client_dev(device_name);
 397
 398         return rc;
 399 }
 400 EXPORT_SYMBOL_GPL(ntb_transport_register_client_dev);
 401
 402 /**
 403  * ntb_transport_register_client - Register NTB client driver
 404  * @drv: NTB client driver to be registered
 405  *
 406  * Register an NTB client driver with the NTB transport layer
 407  *
 408  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 409  */
 410 int ntb_transport_register_client(struct ntb_transport_client *drv)
 411 {
 412         drv->driver.bus = &ntb_transport_bus;
 413
 414         if (list_empty(&ntb_transport_list))
 415                 return -ENODEV;
 416
 417         return driver_register(&drv->driver);
 418 }
 419 EXPORT_SYMBOL_GPL(ntb_transport_register_client);
 420
 421 /**
 422  * ntb_transport_unregister_client - Unregister NTB client driver
 423  * @drv: NTB client driver to be unregistered
 424  *
 425  * Unregister an NTB client driver with the NTB transport layer
 426  *
 427  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 428  */
 429 void ntb_transport_unregister_client(struct ntb_transport_client *drv)
 430 {
 431         driver_unregister(&drv->driver);
 432 }
 433 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client);
 434
 435 static ssize_t debugfs_read(struct file *filp, char __user *ubuf, size_t count,
 436                             loff_t *offp)
 437 {
 438         struct ntb_transport_qp *qp;
 439         char *buf;
 440         ssize_t ret, out_offset, out_count;
 441
 442         qp = filp->private_data;
 443
 444         if (!qp || !qp->link_is_up)
 445                 return 0;
 446
 447         out_count = 1000;
 448
 449         buf = kmalloc(out_count, GFP_KERNEL);
 450         if (!buf)
 451                 return -ENOMEM;
 452
 453         out_offset = 0;
 454         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 455                                "NTB QP stats\n");
 456         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 457                                "rx_bytes - \t%llu\n", qp->rx_bytes);
 458         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 459                                "rx_pkts - \t%llu\n", qp->rx_pkts);
 460         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 461                                "rx_memcpy - \t%llu\n", qp->rx_memcpy);
 462         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 463                                "rx_async - \t%llu\n", qp->rx_async);
 464         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 465                                "rx_ring_empty - %llu\n", qp->rx_ring_empty);
 466         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 467                                "rx_err_no_buf - %llu\n", qp->rx_err_no_buf);
 468         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 469                                "rx_err_oflow - \t%llu\n", qp->rx_err_oflow);
 470         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 471                                "rx_err_ver - \t%llu\n", qp->rx_err_ver);
 472         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 473                                "rx_buff - \t%p\n", qp->rx_buff);
 474         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 475                                "rx_index - \t%u\n", qp->rx_index);
 476         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 477                                "rx_max_entry - \t%u\n", qp->rx_max_entry);
 478
 479         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 480                                "tx_bytes - \t%llu\n", qp->tx_bytes);
 481         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 482                                "tx_pkts - \t%llu\n", qp->tx_pkts);
 483         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 484                                "tx_memcpy - \t%llu\n", qp->tx_memcpy);
 485         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 486                                "tx_async - \t%llu\n", qp->tx_async);
 487         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 488                                "tx_ring_full - \t%llu\n", qp->tx_ring_full);
 489         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 490                                "tx_err_no_buf - %llu\n", qp->tx_err_no_buf);
 491         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 492                                "tx_mw - \t%p\n", qp->tx_mw);
 493         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 494                                "tx_index - \t%u\n", qp->tx_index);
 495         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 496                                "tx_max_entry - \t%u\n", qp->tx_max_entry);
 497         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 498                                "qp->remote_rx_info->entry - \t%u\n",
 499                                qp->remote_rx_info->entry);
 500         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 501                                "free tx - \t%u\n",
 502                                ntb_transport_tx_free_entry(qp));
 503
 504         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 505                                "\nQP Link %s\n",
 506                                qp->link_is_up ? "Up" : "Down");
 507         if (out_offset > out_count)
 508                 out_offset = out_count;
 509
 510         ret = simple_read_from_buffer(ubuf, count, offp, buf, out_offset);
 511         kfree(buf);
 512         return ret;
 513 }
 514
 515 static const struct file_operations ntb_qp_debugfs_stats = {
 516         .owner = THIS_MODULE,
 517         .open = simple_open,
 518         .read = debugfs_read,
 519 };
 520
 521 static void ntb_list_add(spinlock_t *lock, struct list_head *entry,
 522                          struct list_head *list)
 523 {
 524         unsigned long flags;
 525
 526         spin_lock_irqsave(lock, flags);
 527         list_add_tail(entry, list);
 528         spin_unlock_irqrestore(lock, flags);
 529 }
 530
 531 static struct ntb_queue_entry *ntb_list_rm(spinlock_t *lock,
 532                                            struct list_head *list)
 533 {
 534         struct ntb_queue_entry *entry;
 535         unsigned long flags;
 536
 537         spin_lock_irqsave(lock, flags);
 538         if (list_empty(list)) {
 539                 entry = NULL;
 540                 goto out;
 541         }
 542         entry = list_first_entry(list, struct ntb_queue_entry, entry);
 543         list_del(&entry->entry);
 544
 545 out:
 546         spin_unlock_irqrestore(lock, flags);
 547
 548         return entry;
 549 }
 550
 551 static struct ntb_queue_entry *ntb_list_mv(spinlock_t *lock,
 552                                            struct list_head *list,
 553                                            struct list_head *to_list)
 554 {
 555         struct ntb_queue_entry *entry;
 556         unsigned long flags;
 557
 558         spin_lock_irqsave(lock, flags);
 559
 560         if (list_empty(list)) {
 561                 entry = NULL;
 562         } else {
 563                 entry = list_first_entry(list, struct ntb_queue_entry, entry);
 564                 list_move_tail(&entry->entry, to_list);
 565         }
 566
 567         spin_unlock_irqrestore(lock, flags);
 568
 569         return entry;
 570 }
 571
 572 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
 573                                      unsigned int qp_num)
 574 {
 575         struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
 576         struct ntb_transport_mw *mw;
 577         unsigned int rx_size, num_qps_mw;
 578         unsigned int mw_num, mw_count, qp_count;
 579         unsigned int i;
 580
 581         mw_count = nt->mw_count;
 582         qp_count = nt->qp_count;
 583
 584         mw_num = QP_TO_MW(nt, qp_num);
 585         mw = &nt->mw_vec[mw_num];
 586
 587         if (!mw->virt_addr)
 588                 return -ENOMEM;
 589
 590         if (qp_count % mw_count && mw_num + 1 < qp_count / mw_count)
 591                 num_qps_mw = qp_count / mw_count + 1;
 592         else
 593                 num_qps_mw = qp_count / mw_count;
 594
 595         rx_size = (unsigned int)mw->xlat_size / num_qps_mw;
 596         qp->rx_buff = mw->virt_addr + rx_size * qp_num / mw_count;
 597         rx_size -= sizeof(struct ntb_rx_info);
 598
 599         qp->remote_rx_info = qp->rx_buff + rx_size;
 600
 601         /* Due to housekeeping, there must be atleast 2 buffs */
 602         qp->rx_max_frame = min(transport_mtu, rx_size / 2);
 603         qp->rx_max_entry = rx_size / qp->rx_max_frame;
 604         qp->rx_index = 0;
 605
 606         qp->remote_rx_info->entry = qp->rx_max_entry - 1;
 607
 608         /* setup the hdr offsets with 0's */
 609         for (i = 0; i < qp->rx_max_entry; i++) {
 610                 void *offset = (qp->rx_buff + qp->rx_max_frame * (i + 1) -
 611                                 sizeof(struct ntb_payload_header));
 612                 memset(offset, 0, sizeof(struct ntb_payload_header));
 613         }
 614
 615         qp->rx_pkts = 0;
 616         qp->tx_pkts = 0;
 617         qp->tx_index = 0;
 618
 619         return 0;
 620 }
 621
 622 static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
 623 {
 624         struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
 625         struct pci_dev *pdev = nt->ndev->pdev;
 626
 627         if (!mw->virt_addr)
 628                 return;
 629
 630         ntb_mw_clear_trans(nt->ndev, num_mw);
 631         dma_free_coherent(&pdev->dev, mw->buff_size,
 632                           mw->virt_addr, mw->dma_addr);
 633         mw->xlat_size = 0;
 634         mw->buff_size = 0;
 635         mw->virt_addr = NULL;
 636 }
 637
 638 static int ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw,
 639                       resource_size_t size)
 640 {
 641         struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
 642         struct pci_dev *pdev = nt->ndev->pdev;
 643         size_t xlat_size, buff_size;
 644         int rc;
 645
 646         if (!size)
 647                 return -EINVAL;
 648
 649         xlat_size = round_up(size, mw->xlat_align_size);
 650         buff_size = round_up(size, mw->xlat_align);
 651
 652         /* No need to re-setup */
 653         if (mw->xlat_size == xlat_size)
 654                 return 0;
 655
 656         if (mw->buff_size)
 657                 ntb_free_mw(nt, num_mw);
 658
 659         /* Alloc memory for receiving data.  Must be aligned */
 660         mw->xlat_size = xlat_size;
 661         mw->buff_size = buff_size;
 662
 663         mw->virt_addr = dma_alloc_coherent(&pdev->dev, buff_size,
 664                                            &mw->dma_addr, GFP_KERNEL);
 665         if (!mw->virt_addr) {
 666                 mw->xlat_size = 0;
 667                 mw->buff_size = 0;
 668                 dev_err(&pdev->dev, "Unable to alloc MW buff of size %zu\n",
 669                         buff_size);
 670                 return -ENOMEM;
 671         }
 672
 673         /*
 674          * we must ensure that the memory address allocated is BAR size
 675          * aligned in order for the XLAT register to take the value. This
 676          * is a requirement of the hardware. It is recommended to setup CMA
 677          * for BAR sizes equal or greater than 4MB.
 678          */
 679         if (!IS_ALIGNED(mw->dma_addr, mw->xlat_align)) {
 680                 dev_err(&pdev->dev, "DMA memory %pad is not aligned\n",
 681                         &mw->dma_addr);
 682                 ntb_free_mw(nt, num_mw);
 683                 return -ENOMEM;
 684         }
 685
 686         /* Notify HW the memory location of the receive buffer */
 687         rc = ntb_mw_set_trans(nt->ndev, num_mw, mw->dma_addr, mw->xlat_size);
 688         if (rc) {
 689                 dev_err(&pdev->dev, "Unable to set mw%d translation", num_mw);
 690                 ntb_free_mw(nt, num_mw);
 691                 return -EIO;
 692         }
 693
 694         return 0;
 695 }
 696
 697 static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
 698 {
 699         qp->link_is_up = false;
 700
 701         qp->tx_index = 0;
 702         qp->rx_index = 0;
 703         qp->rx_bytes = 0;
 704         qp->rx_pkts = 0;
 705         qp->rx_ring_empty = 0;
 706         qp->rx_err_no_buf = 0;
 707         qp->rx_err_oflow = 0;
 708         qp->rx_err_ver = 0;
 709         qp->rx_memcpy = 0;
 710         qp->rx_async = 0;
 711         qp->tx_bytes = 0;
 712         qp->tx_pkts = 0;
 713         qp->tx_ring_full = 0;
 714         qp->tx_err_no_buf = 0;
 715         qp->tx_memcpy = 0;
 716         qp->tx_async = 0;
 717 }
 718
 719 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
 720 {
 721         struct ntb_transport_ctx *nt = qp->transport;
 722         struct pci_dev *pdev = nt->ndev->pdev;
 723
 724         dev_info(&pdev->dev, "qp %d: Link Cleanup\n", qp->qp_num);
 725
 726         cancel_delayed_work_sync(&qp->link_work);
 727         ntb_qp_link_down_reset(qp);
 728
 729         if (qp->event_handler)
 730                 qp->event_handler(qp->cb_data, qp->link_is_up);
 731 }
 732
 733 static void ntb_qp_link_cleanup_work(struct work_struct *work)
 734 {
 735         struct ntb_transport_qp *qp = container_of(work,
 736                                                    struct ntb_transport_qp,
 737                                                    link_cleanup);
 738         struct ntb_transport_ctx *nt = qp->transport;
 739
 740         ntb_qp_link_cleanup(qp);
 741
 742         if (nt->link_is_up)
 743                 schedule_delayed_work(&qp->link_work,
 744                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 745 }
 746
 747 static void ntb_qp_link_down(struct ntb_transport_qp *qp)
 748 {
 749         schedule_work(&qp->link_cleanup);
 750 }
 751
 752 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
 753 {
 754         struct ntb_transport_qp *qp;
 755         u64 qp_bitmap_alloc;
 756         int i;
 757
 758         qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
 759
 760         /* Pass along the info to any clients */
 761         for (i = 0; i < nt->qp_count; i++)
 762                 if (qp_bitmap_alloc & BIT_ULL(i)) {
 763                         qp = &nt->qp_vec[i];
 764                         ntb_qp_link_cleanup(qp);
 765                         cancel_work_sync(&qp->link_cleanup);
 766                         cancel_delayed_work_sync(&qp->link_work);
 767                 }
 768
 769         if (!nt->link_is_up)
 770                 cancel_delayed_work_sync(&nt->link_work);
 771
 772         /* The scratchpad registers keep the values if the remote side
 773          * goes down, blast them now to give them a sane value the next
 774          * time they are accessed
 775          */
 776         for (i = 0; i < MAX_SPAD; i++)
 777                 ntb_spad_write(nt->ndev, i, 0);
 778 }
 779
 780 static void ntb_transport_link_cleanup_work(struct work_struct *work)
 781 {
 782         struct ntb_transport_ctx *nt =
 783                 container_of(work, struct ntb_transport_ctx, link_cleanup);
 784
 785         ntb_transport_link_cleanup(nt);
 786 }
 787
 788 static void ntb_transport_event_callback(void *data)
 789 {
 790         struct ntb_transport_ctx *nt = data;
 791
 792         if (ntb_link_is_up(nt->ndev, NULL, NULL) == 1)
 793                 schedule_delayed_work(&nt->link_work, 0);
 794         else
 795                 schedule_work(&nt->link_cleanup);
 796 }
 797
 798 static void ntb_transport_link_work(struct work_struct *work)
 799 {
 800         struct ntb_transport_ctx *nt =
 801                 container_of(work, struct ntb_transport_ctx, link_work.work);
 802         struct ntb_dev *ndev = nt->ndev;
 803         struct pci_dev *pdev = ndev->pdev;
 804         resource_size_t size;
 805         u32 val;
 806         int rc, i, spad;
 807
 808         /* send the local info, in the opposite order of the way we read it */
 809         for (i = 0; i < nt->mw_count; i++) {
 810                 size = nt->mw_vec[i].phys_size;
 811
 812                 if (max_mw_size && size > max_mw_size)
 813                         size = max_mw_size;
 814
 815                 spad = MW0_SZ_HIGH + (i * 2);
 816                 ntb_peer_spad_write(ndev, spad, (u32)(size >> 32));
 817
 818                 spad = MW0_SZ_LOW + (i * 2);
 819                 ntb_peer_spad_write(ndev, spad, (u32)size);
 820         }
 821
 822         ntb_peer_spad_write(ndev, NUM_MWS, nt->mw_count);
 823
 824         ntb_peer_spad_write(ndev, NUM_QPS, nt->qp_count);
 825
 826         ntb_peer_spad_write(ndev, VERSION, NTB_TRANSPORT_VERSION);
 827
 828         /* Query the remote side for its info */
 829         val = ntb_spad_read(ndev, VERSION);
 830         dev_dbg(&pdev->dev, "Remote version = %d\n", val);
 831         if (val != NTB_TRANSPORT_VERSION)
 832                 goto out;
 833
 834         val = ntb_spad_read(ndev, NUM_QPS);
 835         dev_dbg(&pdev->dev, "Remote max number of qps = %d\n", val);
 836         if (val != nt->qp_count)
 837                 goto out;
 838
 839         val = ntb_spad_read(ndev, NUM_MWS);
 840         dev_dbg(&pdev->dev, "Remote number of mws = %d\n", val);
 841         if (val != nt->mw_count)
 842                 goto out;
 843
 844         for (i = 0; i < nt->mw_count; i++) {
 845                 u64 val64;
 846
 847                 val = ntb_spad_read(ndev, MW0_SZ_HIGH + (i * 2));
 848                 val64 = (u64)val << 32;
 849
 850                 val = ntb_spad_read(ndev, MW0_SZ_LOW + (i * 2));
 851                 val64 |= val;
 852
 853                 dev_dbg(&pdev->dev, "Remote MW%d size = %#llx\n", i, val64);
 854
 855                 rc = ntb_set_mw(nt, i, val64);
 856                 if (rc)
 857                         goto out1;
 858         }
 859
 860         nt->link_is_up = true;
 861
 862         for (i = 0; i < nt->qp_count; i++) {
 863                 struct ntb_transport_qp *qp = &nt->qp_vec[i];
 864
 865                 ntb_transport_setup_qp_mw(nt, i);
 866
 867                 if (qp->client_ready)
 868                         schedule_delayed_work(&qp->link_work, 0);
 869         }
 870
 871         return;
 872
 873 out1:
 874         for (i = 0; i < nt->mw_count; i++)
 875                 ntb_free_mw(nt, i);
 876 out:
 877         if (ntb_link_is_up(ndev, NULL, NULL) == 1)
 878                 schedule_delayed_work(&nt->link_work,
 879                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 880 }
 881
 882 static void ntb_qp_link_work(struct work_struct *work)
 883 {
 884         struct ntb_transport_qp *qp = container_of(work,
 885                                                    struct ntb_transport_qp,
 886                                                    link_work.work);
 887         struct pci_dev *pdev = qp->ndev->pdev;
 888         struct ntb_transport_ctx *nt = qp->transport;
 889         int val;
 890
 891         WARN_ON(!nt->link_is_up);
 892
 893         val = ntb_spad_read(nt->ndev, QP_LINKS);
 894
 895         ntb_peer_spad_write(nt->ndev, QP_LINKS, val | BIT(qp->qp_num));
 896
 897         /* query remote spad for qp ready bits */
 898         ntb_peer_spad_read(nt->ndev, QP_LINKS);
 899         dev_dbg_ratelimited(&pdev->dev, "Remote QP link status = %x\n", val);
 900
 901         /* See if the remote side is up */
 902         if (val & BIT(qp->qp_num)) {
 903                 dev_info(&pdev->dev, "qp %d: Link Up\n", qp->qp_num);
 904                 qp->link_is_up = true;
 905
 906                 if (qp->event_handler)
 907                         qp->event_handler(qp->cb_data, qp->link_is_up);
 908
 909                 tasklet_schedule(&qp->rxc_db_work);
 910         } else if (nt->link_is_up)
 911                 schedule_delayed_work(&qp->link_work,
 912                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 913 }
 914
 915 static int ntb_transport_init_queue(struct ntb_transport_ctx *nt,
 916                                     unsigned int qp_num)
 917 {
 918         struct ntb_transport_qp *qp;
 919         struct ntb_transport_mw *mw;
 920         phys_addr_t mw_base;
 921         resource_size_t mw_size;
 922         unsigned int num_qps_mw, tx_size;
 923         unsigned int mw_num, mw_count, qp_count;
 924         u64 qp_offset;
 925
 926         mw_count = nt->mw_count;
 927         qp_count = nt->qp_count;
 928
 929         mw_num = QP_TO_MW(nt, qp_num);
 930         mw = &nt->mw_vec[mw_num];
 931
 932         qp = &nt->qp_vec[qp_num];
 933         qp->qp_num = qp_num;
 934         qp->transport = nt;
 935         qp->ndev = nt->ndev;
 936         qp->client_ready = false;
 937         qp->event_handler = NULL;
 938         ntb_qp_link_down_reset(qp);
 939
 940         if (qp_count % mw_count && mw_num + 1 < qp_count / mw_count)
 941                 num_qps_mw = qp_count / mw_count + 1;
 942         else
 943                 num_qps_mw = qp_count / mw_count;
 944
 945         mw_base = nt->mw_vec[mw_num].phys_addr;
 946         mw_size = nt->mw_vec[mw_num].phys_size;
 947
 948         tx_size = (unsigned int)mw_size / num_qps_mw;
 949         qp_offset = tx_size * qp_num / mw_count;
 950
 951         qp->tx_mw = nt->mw_vec[mw_num].vbase + qp_offset;
 952         if (!qp->tx_mw)
 953                 return -EINVAL;
 954
 955         qp->tx_mw_phys = mw_base + qp_offset;
 956         if (!qp->tx_mw_phys)
 957                 return -EINVAL;
 958
 959         tx_size -= sizeof(struct ntb_rx_info);
 960         qp->rx_info = qp->tx_mw + tx_size;
 961
 962         /* Due to housekeeping, there must be atleast 2 buffs */
 963         qp->tx_max_frame = min(transport_mtu, tx_size / 2);
 964         qp->tx_max_entry = tx_size / qp->tx_max_frame;
 965
 966         if (nt->debugfs_node_dir) {
 967                 char debugfs_name[4];
 968
 969                 snprintf(debugfs_name, 4, "qp%d", qp_num);
 970                 qp->debugfs_dir = debugfs_create_dir(debugfs_name,
 971                                                      nt->debugfs_node_dir);
 972
 973                 qp->debugfs_stats = debugfs_create_file("stats", S_IRUSR,
 974                                                         qp->debugfs_dir, qp,
 975                                                         &ntb_qp_debugfs_stats);
 976         } else {
 977                 qp->debugfs_dir = NULL;
 978                 qp->debugfs_stats = NULL;
 979         }
 980
 981         INIT_DELAYED_WORK(&qp->link_work, ntb_qp_link_work);
 982         INIT_WORK(&qp->link_cleanup, ntb_qp_link_cleanup_work);
 983
 984         spin_lock_init(&qp->ntb_rx_q_lock);
 985         spin_lock_init(&qp->ntb_tx_free_q_lock);
 986
 987         INIT_LIST_HEAD(&qp->rx_post_q);
 988         INIT_LIST_HEAD(&qp->rx_pend_q);
 989         INIT_LIST_HEAD(&qp->rx_free_q);
 990         INIT_LIST_HEAD(&qp->tx_free_q);
 991
 992         tasklet_init(&qp->rxc_db_work, ntb_transport_rxc_db,
 993                      (unsigned long)qp);
 994
 995         return 0;
 996 }
 997
 998 static int ntb_transport_probe(struct ntb_client *self, struct ntb_dev *ndev)
 999 {
1000         struct ntb_transport_ctx *nt;
1001         struct ntb_transport_mw *mw;
1002         unsigned int mw_count, qp_count;
1003         u64 qp_bitmap;
1004         int node;
1005         int rc, i;
1006
1007         if (ntb_db_is_unsafe(ndev))
1008                 dev_dbg(&ndev->dev,
1009                         "doorbell is unsafe, proceed anyway...\n");
1010         if (ntb_spad_is_unsafe(ndev))
1011                 dev_dbg(&ndev->dev,
1012                         "scratchpad is unsafe, proceed anyway...\n");
1013
1014         node = dev_to_node(&ndev->dev);
1015
1016         nt = kzalloc_node(sizeof(*nt), GFP_KERNEL, node);
1017         if (!nt)
1018                 return -ENOMEM;
1019
1020         nt->ndev = ndev;
1021
1022         mw_count = ntb_mw_count(ndev);
1023
1024         nt->mw_count = mw_count;
1025
1026         nt->mw_vec = kzalloc_node(mw_count * sizeof(*nt->mw_vec),
1027                                   GFP_KERNEL, node);
1028         if (!nt->mw_vec) {
1029                 rc = -ENOMEM;
1030                 goto err;
1031         }
1032
1033         for (i = 0; i < mw_count; i++) {
1034                 mw = &nt->mw_vec[i];
1035
1036                 rc = ntb_mw_get_range(ndev, i, &mw->phys_addr, &mw->phys_size,
1037                                       &mw->xlat_align, &mw->xlat_align_size);
1038                 if (rc)
1039                         goto err1;
1040
1041                 mw->vbase = ioremap_wc(mw->phys_addr, mw->phys_size);
1042                 if (!mw->vbase) {
1043                         rc = -ENOMEM;
1044                         goto err1;
1045                 }
1046
1047                 mw->buff_size = 0;
1048                 mw->xlat_size = 0;
1049                 mw->virt_addr = NULL;
1050                 mw->dma_addr = 0;
1051         }
1052
1053         qp_bitmap = ntb_db_valid_mask(ndev);
1054
1055         qp_count = ilog2(qp_bitmap);
1056         if (max_num_clients && max_num_clients < qp_count)
1057                 qp_count = max_num_clients;
1058         else if (mw_count < qp_count)
1059                 qp_count = mw_count;
1060
1061         qp_bitmap &= BIT_ULL(qp_count) - 1;
1062
1063         nt->qp_count = qp_count;
1064         nt->qp_bitmap = qp_bitmap;
1065         nt->qp_bitmap_free = qp_bitmap;
1066
1067         nt->qp_vec = kzalloc_node(qp_count * sizeof(*nt->qp_vec),
1068                                   GFP_KERNEL, node);
1069         if (!nt->qp_vec) {
1070                 rc = -ENOMEM;
1071                 goto err2;
1072         }
1073
1074         if (nt_debugfs_dir) {
1075                 nt->debugfs_node_dir =
1076                         debugfs_create_dir(pci_name(ndev->pdev),
1077                                            nt_debugfs_dir);
1078         }
1079
1080         for (i = 0; i < qp_count; i++) {
1081                 rc = ntb_transport_init_queue(nt, i);
1082                 if (rc)
1083                         goto err3;
1084         }
1085
1086         INIT_DELAYED_WORK(&nt->link_work, ntb_transport_link_work);
1087         INIT_WORK(&nt->link_cleanup, ntb_transport_link_cleanup_work);
1088
1089         rc = ntb_set_ctx(ndev, nt, &ntb_transport_ops);
1090         if (rc)
1091                 goto err3;
1092
1093         INIT_LIST_HEAD(&nt->client_devs);
1094         rc = ntb_bus_init(nt);
1095         if (rc)
1096                 goto err4;
1097
1098         nt->link_is_up = false;
1099         ntb_link_enable(ndev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
1100         ntb_link_event(ndev);
1101
1102         return 0;
1103
1104 err4:
1105         ntb_clear_ctx(ndev);
1106 err3:
1107         kfree(nt->qp_vec);
1108 err2:
1109         kfree(nt->mw_vec);
1110 err1:
1111         while (i--) {
1112                 mw = &nt->mw_vec[i];
1113                 iounmap(mw->vbase);
1114         }
1115 err:
1116         kfree(nt);
1117         return rc;
1118 }
1119
1120 static void ntb_transport_free(struct ntb_client *self, struct ntb_dev *ndev)
1121 {
1122         struct ntb_transport_ctx *nt = ndev->ctx;
1123         struct ntb_transport_qp *qp;
1124         u64 qp_bitmap_alloc;
1125         int i;
1126
1127         ntb_transport_link_cleanup(nt);
1128         cancel_work_sync(&nt->link_cleanup);
1129         cancel_delayed_work_sync(&nt->link_work);
1130
1131         qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
1132
1133         /* verify that all the qp's are freed */
1134         for (i = 0; i < nt->qp_count; i++) {
1135                 qp = &nt->qp_vec[i];
1136                 if (qp_bitmap_alloc & BIT_ULL(i))
1137                         ntb_transport_free_queue(qp);
1138                 debugfs_remove_recursive(qp->debugfs_dir);
1139         }
1140
1141         ntb_link_disable(ndev);
1142         ntb_clear_ctx(ndev);
1143
1144         ntb_bus_remove(nt);
1145
1146         for (i = nt->mw_count; i--; ) {
1147                 ntb_free_mw(nt, i);
1148                 iounmap(nt->mw_vec[i].vbase);
1149         }
1150
1151         kfree(nt->qp_vec);
1152         kfree(nt->mw_vec);
1153         kfree(nt);
1154 }
1155
1156 static void ntb_complete_rxc(struct ntb_transport_qp *qp)
1157 {
1158         struct ntb_queue_entry *entry;
1159         void *cb_data;
1160         unsigned int len;
1161         unsigned long irqflags;
1162
1163         spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1164
1165         while (!list_empty(&qp->rx_post_q)) {
1166                 entry = list_first_entry(&qp->rx_post_q,
1167                                          struct ntb_queue_entry, entry);
1168                 if (!(entry->flags & DESC_DONE_FLAG))
1169                         break;
1170
1171                 entry->rx_hdr->flags = 0;
1172                 iowrite32(entry->index, &qp->rx_info->entry);
1173
1174                 cb_data = entry->cb_data;
1175                 len = entry->len;
1176
1177                 list_move_tail(&entry->entry, &qp->rx_free_q);
1178
1179                 spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1180
1181                 if (qp->rx_handler && qp->client_ready)
1182                         qp->rx_handler(qp, qp->cb_data, cb_data, len);
1183
1184                 spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1185         }
1186
1187         spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1188 }
1189
1190 static void ntb_rx_copy_callback(void *data)
1191 {
1192         struct ntb_queue_entry *entry = data;
1193
1194         entry->flags |= DESC_DONE_FLAG;
1195
1196         ntb_complete_rxc(entry->qp);
1197 }
1198
1199 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset)
1200 {
1201         void *buf = entry->buf;
1202         size_t len = entry->len;
1203
1204         memcpy(buf, offset, len);
1205
1206         /* Ensure that the data is fully copied out before clearing the flag */
1207         wmb();
1208
1209         ntb_rx_copy_callback(entry);
1210 }
1211
1212 static void ntb_async_rx(struct ntb_queue_entry *entry, void *offset)
1213 {
1214         struct dma_async_tx_descriptor *txd;
1215         struct ntb_transport_qp *qp = entry->qp;
1216         struct dma_chan *chan = qp->dma_chan;
1217         struct dma_device *device;
1218         size_t pay_off, buff_off, len;
1219         struct dmaengine_unmap_data *unmap;
1220         dma_cookie_t cookie;
1221         void *buf = entry->buf;
1222
1223         len = entry->len;
1224
1225         if (!chan)
1226                 goto err;
1227
1228         if (len < copy_bytes)
1229                 goto err_wait;
1230
1231         device = chan->device;
1232         pay_off = (size_t)offset & ~PAGE_MASK;
1233         buff_off = (size_t)buf & ~PAGE_MASK;
1234
1235         if (!is_dma_copy_aligned(device, pay_off, buff_off, len))
1236                 goto err_wait;
1237
1238         unmap = dmaengine_get_unmap_data(device->dev, 2, GFP_NOWAIT);
1239         if (!unmap)
1240                 goto err_wait;
1241
1242         unmap->len = len;
1243         unmap->addr[0] = dma_map_page(device->dev, virt_to_page(offset),
1244                                       pay_off, len, DMA_TO_DEVICE);
1245         if (dma_mapping_error(device->dev, unmap->addr[0]))
1246                 goto err_get_unmap;
1247
1248         unmap->to_cnt = 1;
1249
1250         unmap->addr[1] = dma_map_page(device->dev, virt_to_page(buf),
1251                                       buff_off, len, DMA_FROM_DEVICE);
1252         if (dma_mapping_error(device->dev, unmap->addr[1]))
1253                 goto err_get_unmap;
1254
1255         unmap->from_cnt = 1;
1256
1257         txd = device->device_prep_dma_memcpy(chan, unmap->addr[1],
1258                                              unmap->addr[0], len,
1259                                              DMA_PREP_INTERRUPT);
1260         if (!txd)
1261                 goto err_get_unmap;
1262
1263         txd->callback = ntb_rx_copy_callback;
1264         txd->callback_param = entry;
1265         dma_set_unmap(txd, unmap);
1266
1267         cookie = dmaengine_submit(txd);
1268         if (dma_submit_error(cookie))
1269                 goto err_set_unmap;
1270
1271         dmaengine_unmap_put(unmap);
1272
1273         qp->last_cookie = cookie;
1274
1275         qp->rx_async++;
1276
1277         return;
1278
1279 err_set_unmap:
1280         dmaengine_unmap_put(unmap);
1281 err_get_unmap:
1282         dmaengine_unmap_put(unmap);
1283 err_wait:
1284         /* If the callbacks come out of order, the writing of the index to the
1285          * last completed will be out of order.  This may result in the
1286          * receive stalling forever.
1287          */
1288         dma_sync_wait(chan, qp->last_cookie);
1289 err:
1290         ntb_memcpy_rx(entry, offset);
1291         qp->rx_memcpy++;
1292 }
1293
1294 static int ntb_process_rxc(struct ntb_transport_qp *qp)
1295 {
1296         struct ntb_payload_header *hdr;
1297         struct ntb_queue_entry *entry;
1298         void *offset;
1299
1300         offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
1301         hdr = offset + qp->rx_max_frame - sizeof(struct ntb_payload_header);
1302
1303         dev_dbg(&qp->ndev->pdev->dev, "qp %d: RX ver %u len %d flags %x\n",
1304                 qp->qp_num, hdr->ver, hdr->len, hdr->flags);
1305
1306         if (!(hdr->flags & DESC_DONE_FLAG)) {
1307                 dev_dbg(&qp->ndev->pdev->dev, "done flag not set\n");
1308                 qp->rx_ring_empty++;
1309                 return -EAGAIN;
1310         }
1311
1312         if (hdr->flags & LINK_DOWN_FLAG) {
1313                 dev_dbg(&qp->ndev->pdev->dev, "link down flag set\n");
1314                 ntb_qp_link_down(qp);
1315                 hdr->flags = 0;
1316                 return -EAGAIN;
1317         }
1318
1319         if (hdr->ver != (u32)qp->rx_pkts) {
1320                 dev_dbg(&qp->ndev->pdev->dev,
1321                         "version mismatch, expected %llu - got %u\n",
1322                         qp->rx_pkts, hdr->ver);
1323                 qp->rx_err_ver++;
1324                 return -EIO;
1325         }
1326
1327         entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q);
1328         if (!entry) {
1329                 dev_dbg(&qp->ndev->pdev->dev, "no receive buffer\n");
1330                 qp->rx_err_no_buf++;
1331                 return -EAGAIN;
1332         }
1333
1334         entry->rx_hdr = hdr;
1335         entry->index = qp->rx_index;
1336
1337         if (hdr->len > entry->len) {
1338                 dev_dbg(&qp->ndev->pdev->dev,
1339                         "receive buffer overflow! Wanted %d got %d\n",
1340                         hdr->len, entry->len);
1341                 qp->rx_err_oflow++;
1342
1343                 entry->len = -EIO;
1344                 entry->flags |= DESC_DONE_FLAG;
1345
1346                 ntb_complete_rxc(qp);
1347         } else {
1348                 dev_dbg(&qp->ndev->pdev->dev,
1349                         "RX OK index %u ver %u size %d into buf size %d\n",
1350                         qp->rx_index, hdr->ver, hdr->len, entry->len);
1351
1352                 qp->rx_bytes += hdr->len;
1353                 qp->rx_pkts++;
1354
1355                 entry->len = hdr->len;
1356
1357                 ntb_async_rx(entry, offset);
1358         }
1359
1360         qp->rx_index++;
1361         qp->rx_index %= qp->rx_max_entry;
1362
1363         return 0;
1364 }
1365
1366 static void ntb_transport_rxc_db(unsigned long data)
1367 {
1368         struct ntb_transport_qp *qp = (void *)data;
1369         int rc, i;
1370
1371         dev_dbg(&qp->ndev->pdev->dev, "%s: doorbell %d received\n",
1372                 __func__, qp->qp_num);
1373
1374         /* Limit the number of packets processed in a single interrupt to
1375          * provide fairness to others
1376          */
1377         for (i = 0; i < qp->rx_max_entry; i++) {
1378                 rc = ntb_process_rxc(qp);
1379                 if (rc)
1380                         break;
1381         }
1382
1383         if (i && qp->dma_chan)
1384                 dma_async_issue_pending(qp->dma_chan);
1385
1386         if (i == qp->rx_max_entry) {
1387                 /* there is more work to do */
1388                 tasklet_schedule(&qp->rxc_db_work);
1389         } else if (ntb_db_read(qp->ndev) & BIT_ULL(qp->qp_num)) {
1390                 /* the doorbell bit is set: clear it */
1391                 ntb_db_clear(qp->ndev, BIT_ULL(qp->qp_num));
1392                 /* ntb_db_read ensures ntb_db_clear write is committed */
1393                 ntb_db_read(qp->ndev);
1394
1395                 /* an interrupt may have arrived between finishing
1396                  * ntb_process_rxc and clearing the doorbell bit:
1397                  * there might be some more work to do.
1398                  */
1399                 tasklet_schedule(&qp->rxc_db_work);
1400         }
1401 }
1402
1403 static void ntb_tx_copy_callback(void *data)
1404 {
1405         struct ntb_queue_entry *entry = data;
1406         struct ntb_transport_qp *qp = entry->qp;
1407         struct ntb_payload_header __iomem *hdr = entry->tx_hdr;
1408
1409         iowrite32(entry->flags | DESC_DONE_FLAG, &hdr->flags);
1410
1411         ntb_peer_db_set(qp->ndev, BIT_ULL(qp->qp_num));
1412
1413         /* The entry length can only be zero if the packet is intended to be a
1414          * "link down" or similar.  Since no payload is being sent in these
1415          * cases, there is nothing to add to the completion queue.
1416          */
1417         if (entry->len > 0) {
1418                 qp->tx_bytes += entry->len;
1419
1420                 if (qp->tx_handler)
1421                         qp->tx_handler(qp, qp->cb_data, entry->cb_data,
1422                                        entry->len);
1423         }
1424
1425         ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry, &qp->tx_free_q);
1426 }
1427
1428 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset)
1429 {
1430 #ifdef ARCH_HAS_NOCACHE_UACCESS
1431         /*
1432          * Using non-temporal mov to improve performance on non-cached
1433          * writes, even though we aren't actually copying from user space.
1434          */
1435         __copy_from_user_inatomic_nocache(offset, entry->buf, entry->len);
1436 #else
1437         memcpy_toio(offset, entry->buf, entry->len);
1438 #endif
1439
1440         /* Ensure that the data is fully copied out before setting the flags */
1441         wmb();
1442
1443         ntb_tx_copy_callback(entry);
1444 }
1445
1446 static void ntb_async_tx(struct ntb_transport_qp *qp,
1447                          struct ntb_queue_entry *entry)
1448 {
1449         struct ntb_payload_header __iomem *hdr;
1450         struct dma_async_tx_descriptor *txd;
1451         struct dma_chan *chan = qp->dma_chan;
1452         struct dma_device *device;
1453         size_t dest_off, buff_off;
1454         struct dmaengine_unmap_data *unmap;
1455         dma_addr_t dest;
1456         dma_cookie_t cookie;
1457         void __iomem *offset;
1458         size_t len = entry->len;
1459         void *buf = entry->buf;
1460
1461         offset = qp->tx_mw + qp->tx_max_frame * qp->tx_index;
1462         hdr = offset + qp->tx_max_frame - sizeof(struct ntb_payload_header);
1463         entry->tx_hdr = hdr;
1464
1465         iowrite32(entry->len, &hdr->len);
1466         iowrite32((u32)qp->tx_pkts, &hdr->ver);
1467
1468         if (!chan)
1469                 goto err;
1470
1471         if (len < copy_bytes)
1472                 goto err;
1473
1474         device = chan->device;
1475         dest = qp->tx_mw_phys + qp->tx_max_frame * qp->tx_index;
1476         buff_off = (size_t)buf & ~PAGE_MASK;
1477         dest_off = (size_t)dest & ~PAGE_MASK;
1478
1479         if (!is_dma_copy_aligned(device, buff_off, dest_off, len))
1480                 goto err;
1481
1482         unmap = dmaengine_get_unmap_data(device->dev, 1, GFP_NOWAIT);
1483         if (!unmap)
1484                 goto err;
1485
1486         unmap->len = len;
1487         unmap->addr[0] = dma_map_page(device->dev, virt_to_page(buf),
1488                                       buff_off, len, DMA_TO_DEVICE);
1489         if (dma_mapping_error(device->dev, unmap->addr[0]))
1490                 goto err_get_unmap;
1491
1492         unmap->to_cnt = 1;
1493
1494         txd = device->device_prep_dma_memcpy(chan, dest, unmap->addr[0], len,
1495                                              DMA_PREP_INTERRUPT);
1496         if (!txd)
1497                 goto err_get_unmap;
1498
1499         txd->callback = ntb_tx_copy_callback;
1500         txd->callback_param = entry;
1501         dma_set_unmap(txd, unmap);
1502
1503         cookie = dmaengine_submit(txd);
1504         if (dma_submit_error(cookie))
1505                 goto err_set_unmap;
1506
1507         dmaengine_unmap_put(unmap);
1508
1509         dma_async_issue_pending(chan);
1510         qp->tx_async++;
1511
1512         return;
1513 err_set_unmap:
1514         dmaengine_unmap_put(unmap);
1515 err_get_unmap:
1516         dmaengine_unmap_put(unmap);
1517 err:
1518         ntb_memcpy_tx(entry, offset);
1519         qp->tx_memcpy++;
1520 }
1521
1522 static int ntb_process_tx(struct ntb_transport_qp *qp,
1523                           struct ntb_queue_entry *entry)
1524 {
1525         if (qp->tx_index == qp->remote_rx_info->entry) {
1526                 qp->tx_ring_full++;
1527                 return -EAGAIN;
1528         }
1529
1530         if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
1531                 if (qp->tx_handler)
1532                         qp->tx_handler(qp->cb_data, qp, NULL, -EIO);
1533
1534                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1535                              &qp->tx_free_q);
1536                 return 0;
1537         }
1538
1539         ntb_async_tx(qp, entry);
1540
1541         qp->tx_index++;
1542         qp->tx_index %= qp->tx_max_entry;
1543
1544         qp->tx_pkts++;
1545
1546         return 0;
1547 }
1548
1549 static void ntb_send_link_down(struct ntb_transport_qp *qp)
1550 {
1551         struct pci_dev *pdev = qp->ndev->pdev;
1552         struct ntb_queue_entry *entry;
1553         int i, rc;
1554
1555         if (!qp->link_is_up)
1556                 return;
1557
1558         dev_info(&pdev->dev, "qp %d: Send Link Down\n", qp->qp_num);
1559
1560         for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1561                 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1562                 if (entry)
1563                         break;
1564                 msleep(100);
1565         }
1566
1567         if (!entry)
1568                 return;
1569
1570         entry->cb_data = NULL;
1571         entry->buf = NULL;
1572         entry->len = 0;
1573         entry->flags = LINK_DOWN_FLAG;
1574
1575         rc = ntb_process_tx(qp, entry);
1576         if (rc)
1577                 dev_err(&pdev->dev, "ntb: QP%d unable to send linkdown msg\n",
1578                         qp->qp_num);
1579
1580         ntb_qp_link_down_reset(qp);
1581 }
1582
1583 static bool ntb_dma_filter_fn(struct dma_chan *chan, void *node)
1584 {
1585         return dev_to_node(&chan->dev->device) == (int)(unsigned long)node;
1586 }
1587
1588 /**
1589  * ntb_transport_create_queue - Create a new NTB transport layer queue
1590  * @rx_handler: receive callback function
1591  * @tx_handler: transmit callback function
1592  * @event_handler: event callback function
1593  *
1594  * Create a new NTB transport layer queue and provide the queue with a callback
1595  * routine for both transmit and receive.  The receive callback routine will be
1596  * used to pass up data when the transport has received it on the queue.   The
1597  * transmit callback routine will be called when the transport has completed the
1598  * transmission of the data on the queue and the data is ready to be freed.
1599  *
1600  * RETURNS: pointer to newly created ntb_queue, NULL on error.
1601  */
1602 struct ntb_transport_qp *
1603 ntb_transport_create_queue(void *data, struct device *client_dev,
1604                            const struct ntb_queue_handlers *handlers)
1605 {
1606         struct ntb_dev *ndev;
1607         struct pci_dev *pdev;
1608         struct ntb_transport_ctx *nt;
1609         struct ntb_queue_entry *entry;
1610         struct ntb_transport_qp *qp;
1611         u64 qp_bit;
1612         unsigned int free_queue;
1613         dma_cap_mask_t dma_mask;
1614         int node;
1615         int i;
1616
1617         ndev = dev_ntb(client_dev->parent);
1618         pdev = ndev->pdev;
1619         nt = ndev->ctx;
1620
1621         node = dev_to_node(&ndev->dev);
1622
1623         free_queue = ffs(nt->qp_bitmap);
1624         if (!free_queue)
1625                 goto err;
1626
1627         /* decrement free_queue to make it zero based */
1628         free_queue--;
1629
1630         qp = &nt->qp_vec[free_queue];
1631         qp_bit = BIT_ULL(qp->qp_num);
1632
1633         nt->qp_bitmap_free &= ~qp_bit;
1634
1635         qp->cb_data = data;
1636         qp->rx_handler = handlers->rx_handler;
1637         qp->tx_handler = handlers->tx_handler;
1638         qp->event_handler = handlers->event_handler;
1639
1640         dma_cap_zero(dma_mask);
1641         dma_cap_set(DMA_MEMCPY, dma_mask);
1642
1643         if (use_dma) {
1644                 qp->dma_chan = dma_request_channel(dma_mask, ntb_dma_filter_fn,
1645                                                    (void *)(unsigned long)node);
1646                 if (!qp->dma_chan)
1647                         dev_info(&pdev->dev, "Unable to allocate DMA channel\n");
1648         } else {
1649                 qp->dma_chan = NULL;
1650         }
1651         dev_dbg(&pdev->dev, "Using %s memcpy\n", qp->dma_chan ? "DMA" : "CPU");
1652
1653         for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
1654                 entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
1655                 if (!entry)
1656                         goto err1;
1657
1658                 entry->qp = qp;
1659                 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
1660                              &qp->rx_free_q);
1661         }
1662
1663         for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
1664                 entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
1665                 if (!entry)
1666                         goto err2;
1667
1668                 entry->qp = qp;
1669                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1670                              &qp->tx_free_q);
1671         }
1672
1673         ntb_db_clear(qp->ndev, qp_bit);
1674         ntb_db_clear_mask(qp->ndev, qp_bit);
1675
1676         dev_info(&pdev->dev, "NTB Transport QP %d created\n", qp->qp_num);
1677
1678         return qp;
1679
1680 err2:
1681         while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1682                 kfree(entry);
1683 err1:
1684         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
1685                 kfree(entry);
1686         if (qp->dma_chan)
1687                 dma_release_channel(qp->dma_chan);
1688         nt->qp_bitmap_free |= qp_bit;
1689 err:
1690         return NULL;
1691 }
1692 EXPORT_SYMBOL_GPL(ntb_transport_create_queue);
1693
1694 /**
1695  * ntb_transport_free_queue - Frees NTB transport queue
1696  * @qp: NTB queue to be freed
1697  *
1698  * Frees NTB transport queue
1699  */
1700 void ntb_transport_free_queue(struct ntb_transport_qp *qp)
1701 {
1702         struct pci_dev *pdev;
1703         struct ntb_queue_entry *entry;
1704         u64 qp_bit;
1705
1706         if (!qp)
1707                 return;
1708
1709         pdev = qp->ndev->pdev;
1710
1711         if (qp->dma_chan) {
1712                 struct dma_chan *chan = qp->dma_chan;
1713                 /* Putting the dma_chan to NULL will force any new traffic to be
1714                  * processed by the CPU instead of the DAM engine
1715                  */
1716                 qp->dma_chan = NULL;
1717
1718                 /* Try to be nice and wait for any queued DMA engine
1719                  * transactions to process before smashing it with a rock
1720                  */
1721                 dma_sync_wait(chan, qp->last_cookie);
1722                 dmaengine_terminate_all(chan);
1723                 dma_release_channel(chan);
1724         }
1725
1726         qp_bit = BIT_ULL(qp->qp_num);
1727
1728         ntb_db_set_mask(qp->ndev, qp_bit);
1729         tasklet_disable(&qp->rxc_db_work);
1730
1731         cancel_delayed_work_sync(&qp->link_work);
1732
1733         qp->cb_data = NULL;
1734         qp->rx_handler = NULL;
1735         qp->tx_handler = NULL;
1736         qp->event_handler = NULL;
1737
1738         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
1739                 kfree(entry);
1740
1741         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q))) {
1742                 dev_warn(&pdev->dev, "Freeing item from non-empty rx_pend_q\n");
1743                 kfree(entry);
1744         }
1745
1746         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q))) {
1747                 dev_warn(&pdev->dev, "Freeing item from non-empty rx_post_q\n");
1748                 kfree(entry);
1749         }
1750
1751         while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1752                 kfree(entry);
1753
1754         qp->transport->qp_bitmap_free |= qp_bit;
1755
1756         dev_info(&pdev->dev, "NTB Transport QP %d freed\n", qp->qp_num);
1757 }
1758 EXPORT_SYMBOL_GPL(ntb_transport_free_queue);
1759
1760 /**
1761  * ntb_transport_rx_remove - Dequeues enqueued rx packet
1762  * @qp: NTB queue to be freed
1763  * @len: pointer to variable to write enqueued buffers length
1764  *
1765  * Dequeues unused buffers from receive queue.  Should only be used during
1766  * shutdown of qp.
1767  *
1768  * RETURNS: NULL error value on error, or void* for success.
1769  */
1770 void *ntb_transport_rx_remove(struct ntb_transport_qp *qp, unsigned int *len)
1771 {
1772         struct ntb_queue_entry *entry;
1773         void *buf;
1774
1775         if (!qp || qp->client_ready)
1776                 return NULL;
1777
1778         entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q);
1779         if (!entry)
1780                 return NULL;
1781
1782         buf = entry->cb_data;
1783         *len = entry->len;
1784
1785         ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_free_q);
1786
1787         return buf;
1788 }
1789 EXPORT_SYMBOL_GPL(ntb_transport_rx_remove);
1790
1791 /**
1792  * ntb_transport_rx_enqueue - Enqueue a new NTB queue entry
1793  * @qp: NTB transport layer queue the entry is to be enqueued on
1794  * @cb: per buffer pointer for callback function to use
1795  * @data: pointer to data buffer that incoming packets will be copied into
1796  * @len: length of the data buffer
1797  *
1798  * Enqueue a new receive buffer onto the transport queue into which a NTB
1799  * payload can be received into.
1800  *
1801  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
1802  */
1803 int ntb_transport_rx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
1804                              unsigned int len)
1805 {
1806         struct ntb_queue_entry *entry;
1807
1808         if (!qp)
1809                 return -EINVAL;
1810
1811         entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q);
1812         if (!entry)
1813                 return -ENOMEM;
1814
1815         entry->cb_data = cb;
1816         entry->buf = data;
1817         entry->len = len;
1818         entry->flags = 0;
1819
1820         ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_pend_q);
1821
1822         tasklet_schedule(&qp->rxc_db_work);
1823
1824         return 0;
1825 }
1826 EXPORT_SYMBOL_GPL(ntb_transport_rx_enqueue);
1827
1828 /**
1829  * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
1830  * @qp: NTB transport layer queue the entry is to be enqueued on
1831  * @cb: per buffer pointer for callback function to use
1832  * @data: pointer to data buffer that will be sent
1833  * @len: length of the data buffer
1834  *
1835  * Enqueue a new transmit buffer onto the transport queue from which a NTB
1836  * payload will be transmitted.  This assumes that a lock is being held to
1837  * serialize access to the qp.
1838  *
1839  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
1840  */
1841 int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
1842                              unsigned int len)
1843 {
1844         struct ntb_queue_entry *entry;
1845         int rc;
1846
1847         if (!qp || !qp->link_is_up || !len)
1848                 return -EINVAL;
1849
1850         entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1851         if (!entry) {
1852                 qp->tx_err_no_buf++;
1853                 return -EBUSY;
1854         }
1855
1856         entry->cb_data = cb;
1857         entry->buf = data;
1858         entry->len = len;
1859         entry->flags = 0;
1860
1861         rc = ntb_process_tx(qp, entry);
1862         if (rc)
1863                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1864                              &qp->tx_free_q);
1865
1866         return rc;
1867 }
1868 EXPORT_SYMBOL_GPL(ntb_transport_tx_enqueue);
1869
1870 /**
1871  * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
1872  * @qp: NTB transport layer queue to be enabled
1873  *
1874  * Notify NTB transport layer of client readiness to use queue
1875  */
1876 void ntb_transport_link_up(struct ntb_transport_qp *qp)
1877 {
1878         if (!qp)
1879                 return;
1880
1881         qp->client_ready = true;
1882
1883         if (qp->transport->link_is_up)
1884                 schedule_delayed_work(&qp->link_work, 0);
1885 }
1886 EXPORT_SYMBOL_GPL(ntb_transport_link_up);
1887
1888 /**
1889  * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
1890  * @qp: NTB transport layer queue to be disabled
1891  *
1892  * Notify NTB transport layer of client's desire to no longer receive data on
1893  * transport queue specified.  It is the client's responsibility to ensure all
1894  * entries on queue are purged or otherwise handled appropriately.
1895  */
1896 void ntb_transport_link_down(struct ntb_transport_qp *qp)
1897 {
1898         struct pci_dev *pdev;
1899         int val;
1900
1901         if (!qp)
1902                 return;
1903
1904         pdev = qp->ndev->pdev;
1905         qp->client_ready = false;
1906
1907         val = ntb_spad_read(qp->ndev, QP_LINKS);
1908
1909         ntb_peer_spad_write(qp->ndev, QP_LINKS,
1910                             val & ~BIT(qp->qp_num));
1911
1912         if (qp->link_is_up)
1913                 ntb_send_link_down(qp);
1914         else
1915                 cancel_delayed_work_sync(&qp->link_work);
1916 }
1917 EXPORT_SYMBOL_GPL(ntb_transport_link_down);
1918
1919 /**
1920  * ntb_transport_link_query - Query transport link state
1921  * @qp: NTB transport layer queue to be queried
1922  *
1923  * Query connectivity to the remote system of the NTB transport queue
1924  *
1925  * RETURNS: true for link up or false for link down
1926  */
1927 bool ntb_transport_link_query(struct ntb_transport_qp *qp)
1928 {
1929         if (!qp)
1930                 return false;
1931
1932         return qp->link_is_up;
1933 }
1934 EXPORT_SYMBOL_GPL(ntb_transport_link_query);
1935
1936 /**
1937  * ntb_transport_qp_num - Query the qp number
1938  * @qp: NTB transport layer queue to be queried
1939  *
1940  * Query qp number of the NTB transport queue
1941  *
1942  * RETURNS: a zero based number specifying the qp number
1943  */
1944 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
1945 {
1946         if (!qp)
1947                 return 0;
1948
1949         return qp->qp_num;
1950 }
1951 EXPORT_SYMBOL_GPL(ntb_transport_qp_num);
1952
1953 /**
1954  * ntb_transport_max_size - Query the max payload size of a qp
1955  * @qp: NTB transport layer queue to be queried
1956  *
1957  * Query the maximum payload size permissible on the given qp
1958  *
1959  * RETURNS: the max payload size of a qp
1960  */
1961 unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp)
1962 {
1963         unsigned int max;
1964
1965         if (!qp)
1966                 return 0;
1967
1968         if (!qp->dma_chan)
1969                 return qp->tx_max_frame - sizeof(struct ntb_payload_header);
1970
1971         /* If DMA engine usage is possible, try to find the max size for that */
1972         max = qp->tx_max_frame - sizeof(struct ntb_payload_header);
1973         max -= max % (1 << qp->dma_chan->device->copy_align);
1974
1975         return max;
1976 }
1977 EXPORT_SYMBOL_GPL(ntb_transport_max_size);
1978
1979 unsigned int ntb_transport_tx_free_entry(struct ntb_transport_qp *qp)
1980 {
1981         unsigned int head = qp->tx_index;
1982         unsigned int tail = qp->remote_rx_info->entry;
1983
1984         return tail > head ? tail - head : qp->tx_max_entry + tail - head;
1985 }
1986 EXPORT_SYMBOL_GPL(ntb_transport_tx_free_entry);
1987
1988 static void ntb_transport_doorbell_callback(void *data, int vector)
1989 {
1990         struct ntb_transport_ctx *nt = data;
1991         struct ntb_transport_qp *qp;
1992         u64 db_bits;
1993         unsigned int qp_num;
1994
1995         db_bits = (nt->qp_bitmap & ~nt->qp_bitmap_free &
1996                    ntb_db_vector_mask(nt->ndev, vector));
1997
1998         while (db_bits) {
1999                 qp_num = __ffs(db_bits);
2000                 qp = &nt->qp_vec[qp_num];
2001
2002                 tasklet_schedule(&qp->rxc_db_work);
2003
2004                 db_bits &= ~BIT_ULL(qp_num);
2005         }
2006 }
2007
2008 static const struct ntb_ctx_ops ntb_transport_ops = {
2009         .link_event = ntb_transport_event_callback,
2010         .db_event = ntb_transport_doorbell_callback,
2011 };
2012
2013 static struct ntb_client ntb_transport_client = {
2014         .ops = {
2015                 .probe = ntb_transport_probe,
2016                 .remove = ntb_transport_free,
2017         },
2018 };
2019
2020 static int __init ntb_transport_init(void)
2021 {
2022         int rc;
2023
2024         pr_info("%s, version %s\n", NTB_TRANSPORT_DESC, NTB_TRANSPORT_VER);
2025
2026         if (debugfs_initialized())
2027                 nt_debugfs_dir = debugfs_create_dir(KBUILD_MODNAME, NULL);
2028
2029         rc = bus_register(&ntb_transport_bus);
2030         if (rc)
2031                 goto err_bus;
2032
2033         rc = ntb_register_client(&ntb_transport_client);
2034         if (rc)
2035                 goto err_client;
2036
2037         return 0;
2038
2039 err_client:
2040         bus_unregister(&ntb_transport_bus);
2041 err_bus:
2042         debugfs_remove_recursive(nt_debugfs_dir);
2043         return rc;
2044 }
2045 module_init(ntb_transport_init);
2046
2047 static void __exit ntb_transport_exit(void)
2048 {
2049         debugfs_remove_recursive(nt_debugfs_dir);
2050
2051         ntb_unregister_client(&ntb_transport_client);
2052         bus_unregister(&ntb_transport_bus);
2053 }
2054 module_exit(ntb_transport_exit);