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