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                       resource_size_t size)
 633 {
 634         struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
 635         struct pci_dev *pdev = nt->ndev->pdev;
 636         size_t xlat_size, buff_size;
 637         int rc;
 638
 639         if (!size)
 640                 return -EINVAL;
 641
 642         xlat_size = round_up(size, mw->xlat_align_size);
 643         buff_size = round_up(size, mw->xlat_align);
 644
 645         /* No need to re-setup */
 646         if (mw->xlat_size == xlat_size)
 647                 return 0;
 648
 649         if (mw->buff_size)
 650                 ntb_free_mw(nt, num_mw);
 651
 652         /* Alloc memory for receiving data.  Must be aligned */
 653         mw->xlat_size = xlat_size;
 654         mw->buff_size = buff_size;
 655
 656         mw->virt_addr = dma_alloc_coherent(&pdev->dev, buff_size,
 657                                            &mw->dma_addr, GFP_KERNEL);
 658         if (!mw->virt_addr) {
 659                 mw->xlat_size = 0;
 660                 mw->buff_size = 0;
 661                 dev_err(&pdev->dev, "Unable to alloc MW buff of size %zu\n",
 662                         buff_size);
 663                 return -ENOMEM;
 664         }
 665
 666         /*
 667          * we must ensure that the memory address allocated is BAR size
 668          * aligned in order for the XLAT register to take the value. This
 669          * is a requirement of the hardware. It is recommended to setup CMA
 670          * for BAR sizes equal or greater than 4MB.
 671          */
 672         if (!IS_ALIGNED(mw->dma_addr, mw->xlat_align)) {
 673                 dev_err(&pdev->dev, "DMA memory %pad is not aligned\n",
 674                         &mw->dma_addr);
 675                 ntb_free_mw(nt, num_mw);
 676                 return -ENOMEM;
 677         }
 678
 679         /* Notify HW the memory location of the receive buffer */
 680         rc = ntb_mw_set_trans(nt->ndev, num_mw, mw->dma_addr, mw->xlat_size);
 681         if (rc) {
 682                 dev_err(&pdev->dev, "Unable to set mw%d translation", num_mw);
 683                 ntb_free_mw(nt, num_mw);
 684                 return -EIO;
 685         }
 686
 687         return 0;
 688 }
 689
 690 static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
 691 {
 692         qp->link_is_up = false;
 693
 694         qp->tx_index = 0;
 695         qp->rx_index = 0;
 696         qp->rx_bytes = 0;
 697         qp->rx_pkts = 0;
 698         qp->rx_ring_empty = 0;
 699         qp->rx_err_no_buf = 0;
 700         qp->rx_err_oflow = 0;
 701         qp->rx_err_ver = 0;
 702         qp->rx_memcpy = 0;
 703         qp->rx_async = 0;
 704         qp->tx_bytes = 0;
 705         qp->tx_pkts = 0;
 706         qp->tx_ring_full = 0;
 707         qp->tx_err_no_buf = 0;
 708         qp->tx_memcpy = 0;
 709         qp->tx_async = 0;
 710 }
 711
 712 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
 713 {
 714         struct ntb_transport_ctx *nt = qp->transport;
 715         struct pci_dev *pdev = nt->ndev->pdev;
 716
 717         dev_info(&pdev->dev, "qp %d: Link Cleanup\n", qp->qp_num);
 718
 719         cancel_delayed_work_sync(&qp->link_work);
 720         ntb_qp_link_down_reset(qp);
 721
 722         if (qp->event_handler)
 723                 qp->event_handler(qp->cb_data, qp->link_is_up);
 724 }
 725
 726 static void ntb_qp_link_cleanup_work(struct work_struct *work)
 727 {
 728         struct ntb_transport_qp *qp = container_of(work,
 729                                                    struct ntb_transport_qp,
 730                                                    link_cleanup);
 731         struct ntb_transport_ctx *nt = qp->transport;
 732
 733         ntb_qp_link_cleanup(qp);
 734
 735         if (nt->link_is_up)
 736                 schedule_delayed_work(&qp->link_work,
 737                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 738 }
 739
 740 static void ntb_qp_link_down(struct ntb_transport_qp *qp)
 741 {
 742         schedule_work(&qp->link_cleanup);
 743 }
 744
 745 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
 746 {
 747         struct ntb_transport_qp *qp;
 748         u64 qp_bitmap_alloc;
 749         int i;
 750
 751         qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
 752
 753         /* Pass along the info to any clients */
 754         for (i = 0; i < nt->qp_count; i++)
 755                 if (qp_bitmap_alloc & BIT_ULL(i)) {
 756                         qp = &nt->qp_vec[i];
 757                         ntb_qp_link_cleanup(qp);
 758                         cancel_work_sync(&qp->link_cleanup);
 759                         cancel_delayed_work_sync(&qp->link_work);
 760                 }
 761
 762         if (!nt->link_is_up)
 763                 cancel_delayed_work_sync(&nt->link_work);
 764
 765         /* The scratchpad registers keep the values if the remote side
 766          * goes down, blast them now to give them a sane value the next
 767          * time they are accessed
 768          */
 769         for (i = 0; i < MAX_SPAD; i++)
 770                 ntb_spad_write(nt->ndev, i, 0);
 771 }
 772
 773 static void ntb_transport_link_cleanup_work(struct work_struct *work)
 774 {
 775         struct ntb_transport_ctx *nt =
 776                 container_of(work, struct ntb_transport_ctx, link_cleanup);
 777
 778         ntb_transport_link_cleanup(nt);
 779 }
 780
 781 static void ntb_transport_event_callback(void *data)
 782 {
 783         struct ntb_transport_ctx *nt = data;
 784
 785         if (ntb_link_is_up(nt->ndev, NULL, NULL) == 1)
 786                 schedule_delayed_work(&nt->link_work, 0);
 787         else
 788                 schedule_work(&nt->link_cleanup);
 789 }
 790
 791 static void ntb_transport_link_work(struct work_struct *work)
 792 {
 793         struct ntb_transport_ctx *nt =
 794                 container_of(work, struct ntb_transport_ctx, link_work.work);
 795         struct ntb_dev *ndev = nt->ndev;
 796         struct pci_dev *pdev = ndev->pdev;
 797         resource_size_t size;
 798         u32 val;
 799         int rc, i, spad;
 800
 801         /* send the local info, in the opposite order of the way we read it */
 802         for (i = 0; i < nt->mw_count; i++) {
 803                 size = nt->mw_vec[i].phys_size;
 804
 805                 if (max_mw_size && size > max_mw_size)
 806                         size = max_mw_size;
 807
 808                 spad = MW0_SZ_HIGH + (i * 2);
 809                 ntb_peer_spad_write(ndev, spad, (u32)(size >> 32));
 810
 811                 spad = MW0_SZ_LOW + (i * 2);
 812                 ntb_peer_spad_write(ndev, spad, (u32)size);
 813         }
 814
 815         ntb_peer_spad_write(ndev, NUM_MWS, nt->mw_count);
 816
 817         ntb_peer_spad_write(ndev, NUM_QPS, nt->qp_count);
 818
 819         ntb_peer_spad_write(ndev, VERSION, NTB_TRANSPORT_VERSION);
 820
 821         /* Query the remote side for its info */
 822         val = ntb_spad_read(ndev, VERSION);
 823         dev_dbg(&pdev->dev, "Remote version = %d\n", val);
 824         if (val != NTB_TRANSPORT_VERSION)
 825                 goto out;
 826
 827         val = ntb_spad_read(ndev, NUM_QPS);
 828         dev_dbg(&pdev->dev, "Remote max number of qps = %d\n", val);
 829         if (val != nt->qp_count)
 830                 goto out;
 831
 832         val = ntb_spad_read(ndev, NUM_MWS);
 833         dev_dbg(&pdev->dev, "Remote number of mws = %d\n", val);
 834         if (val != nt->mw_count)
 835                 goto out;
 836
 837         for (i = 0; i < nt->mw_count; i++) {
 838                 u64 val64;
 839
 840                 val = ntb_spad_read(ndev, MW0_SZ_HIGH + (i * 2));
 841                 val64 = (u64)val << 32;
 842
 843                 val = ntb_spad_read(ndev, MW0_SZ_LOW + (i * 2));
 844                 val64 |= val;
 845
 846                 dev_dbg(&pdev->dev, "Remote MW%d size = %#llx\n", i, val64);
 847
 848                 rc = ntb_set_mw(nt, i, val64);
 849                 if (rc)
 850                         goto out1;
 851         }
 852
 853         nt->link_is_up = true;
 854
 855         for (i = 0; i < nt->qp_count; i++) {
 856                 struct ntb_transport_qp *qp = &nt->qp_vec[i];
 857
 858                 ntb_transport_setup_qp_mw(nt, i);
 859
 860                 if (qp->client_ready)
 861                         schedule_delayed_work(&qp->link_work, 0);
 862         }
 863
 864         return;
 865
 866 out1:
 867         for (i = 0; i < nt->mw_count; i++)
 868                 ntb_free_mw(nt, i);
 869 out:
 870         if (ntb_link_is_up(ndev, NULL, NULL) == 1)
 871                 schedule_delayed_work(&nt->link_work,
 872                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 873 }
 874
 875 static void ntb_qp_link_work(struct work_struct *work)
 876 {
 877         struct ntb_transport_qp *qp = container_of(work,
 878                                                    struct ntb_transport_qp,
 879                                                    link_work.work);
 880         struct pci_dev *pdev = qp->ndev->pdev;
 881         struct ntb_transport_ctx *nt = qp->transport;
 882         int val;
 883
 884         WARN_ON(!nt->link_is_up);
 885
 886         val = ntb_spad_read(nt->ndev, QP_LINKS);
 887
 888         ntb_peer_spad_write(nt->ndev, QP_LINKS, val | BIT(qp->qp_num));
 889
 890         /* query remote spad for qp ready bits */
 891         ntb_peer_spad_read(nt->ndev, QP_LINKS);
 892         dev_dbg_ratelimited(&pdev->dev, "Remote QP link status = %x\n", val);
 893
 894         /* See if the remote side is up */
 895         if (val & BIT(qp->qp_num)) {
 896                 dev_info(&pdev->dev, "qp %d: Link Up\n", qp->qp_num);
 897                 qp->link_is_up = true;
 898
 899                 if (qp->event_handler)
 900                         qp->event_handler(qp->cb_data, qp->link_is_up);
 901
 902                 tasklet_schedule(&qp->rxc_db_work);
 903         } else if (nt->link_is_up)
 904                 schedule_delayed_work(&qp->link_work,
 905                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 906 }
 907
 908 static int ntb_transport_init_queue(struct ntb_transport_ctx *nt,
 909                                     unsigned int qp_num)
 910 {
 911         struct ntb_transport_qp *qp;
 912         struct ntb_transport_mw *mw;
 913         phys_addr_t mw_base;
 914         resource_size_t mw_size;
 915         unsigned int num_qps_mw, tx_size;
 916         unsigned int mw_num, mw_count, qp_count;
 917         u64 qp_offset;
 918
 919         mw_count = nt->mw_count;
 920         qp_count = nt->qp_count;
 921
 922         mw_num = QP_TO_MW(nt, qp_num);
 923         mw = &nt->mw_vec[mw_num];
 924
 925         qp = &nt->qp_vec[qp_num];
 926         qp->qp_num = qp_num;
 927         qp->transport = nt;
 928         qp->ndev = nt->ndev;
 929         qp->client_ready = false;
 930         qp->event_handler = NULL;
 931         ntb_qp_link_down_reset(qp);
 932
 933         if (qp_count % mw_count && mw_num + 1 < qp_count / mw_count)
 934                 num_qps_mw = qp_count / mw_count + 1;
 935         else
 936                 num_qps_mw = qp_count / mw_count;
 937
 938         mw_base = nt->mw_vec[mw_num].phys_addr;
 939         mw_size = nt->mw_vec[mw_num].phys_size;
 940
 941         tx_size = (unsigned int)mw_size / num_qps_mw;
 942         qp_offset = tx_size * qp_num / mw_count;
 943
 944         qp->tx_mw = nt->mw_vec[mw_num].vbase + qp_offset;
 945         if (!qp->tx_mw)
 946                 return -EINVAL;
 947
 948         qp->tx_mw_phys = mw_base + qp_offset;
 949         if (!qp->tx_mw_phys)
 950                 return -EINVAL;
 951
 952         tx_size -= sizeof(struct ntb_rx_info);
 953         qp->rx_info = qp->tx_mw + tx_size;
 954
 955         /* Due to housekeeping, there must be atleast 2 buffs */
 956         qp->tx_max_frame = min(transport_mtu, tx_size / 2);
 957         qp->tx_max_entry = tx_size / qp->tx_max_frame;
 958
 959         if (nt->debugfs_node_dir) {
 960                 char debugfs_name[4];
 961
 962                 snprintf(debugfs_name, 4, "qp%d", qp_num);
 963                 qp->debugfs_dir = debugfs_create_dir(debugfs_name,
 964                                                      nt->debugfs_node_dir);
 965
 966                 qp->debugfs_stats = debugfs_create_file("stats", S_IRUSR,
 967                                                         qp->debugfs_dir, qp,
 968                                                         &ntb_qp_debugfs_stats);
 969         } else {
 970                 qp->debugfs_dir = NULL;
 971                 qp->debugfs_stats = NULL;
 972         }
 973
 974         INIT_DELAYED_WORK(&qp->link_work, ntb_qp_link_work);
 975         INIT_WORK(&qp->link_cleanup, ntb_qp_link_cleanup_work);
 976
 977         spin_lock_init(&qp->ntb_rx_q_lock);
 978         spin_lock_init(&qp->ntb_tx_free_q_lock);
 979
 980         INIT_LIST_HEAD(&qp->rx_post_q);
 981         INIT_LIST_HEAD(&qp->rx_pend_q);
 982         INIT_LIST_HEAD(&qp->rx_free_q);
 983         INIT_LIST_HEAD(&qp->tx_free_q);
 984
 985         tasklet_init(&qp->rxc_db_work, ntb_transport_rxc_db,
 986                      (unsigned long)qp);
 987
 988         return 0;
 989 }
 990
 991 static int ntb_transport_probe(struct ntb_client *self, struct ntb_dev *ndev)
 992 {
 993         struct ntb_transport_ctx *nt;
 994         struct ntb_transport_mw *mw;
 995         unsigned int mw_count, qp_count;
 996         u64 qp_bitmap;
 997         int node;
 998         int rc, i;
 999
1000         if (ntb_db_is_unsafe(ndev))
1001                 dev_dbg(&ndev->dev,
1002                         "doorbell is unsafe, proceed anyway...\n");
1003         if (ntb_spad_is_unsafe(ndev))
1004                 dev_dbg(&ndev->dev,
1005                         "scratchpad is unsafe, proceed anyway...\n");
1006
1007         node = dev_to_node(&ndev->dev);
1008
1009         nt = kzalloc_node(sizeof(*nt), GFP_KERNEL, node);
1010         if (!nt)
1011                 return -ENOMEM;
1012
1013         nt->ndev = ndev;
1014
1015         mw_count = ntb_mw_count(ndev);
1016
1017         nt->mw_count = mw_count;
1018
1019         nt->mw_vec = kzalloc_node(mw_count * sizeof(*nt->mw_vec),
1020                                   GFP_KERNEL, node);
1021         if (!nt->mw_vec) {
1022                 rc = -ENOMEM;
1023                 goto err;
1024         }
1025
1026         for (i = 0; i < mw_count; i++) {
1027                 mw = &nt->mw_vec[i];
1028
1029                 rc = ntb_mw_get_range(ndev, i, &mw->phys_addr, &mw->phys_size,
1030                                       &mw->xlat_align, &mw->xlat_align_size);
1031                 if (rc)
1032                         goto err1;
1033
1034                 mw->vbase = ioremap_wc(mw->phys_addr, mw->phys_size);
1035                 if (!mw->vbase) {
1036                         rc = -ENOMEM;
1037                         goto err1;
1038                 }
1039
1040                 mw->buff_size = 0;
1041                 mw->xlat_size = 0;
1042                 mw->virt_addr = NULL;
1043                 mw->dma_addr = 0;
1044         }
1045
1046         qp_bitmap = ntb_db_valid_mask(ndev);
1047
1048         qp_count = ilog2(qp_bitmap);
1049         if (max_num_clients && max_num_clients < qp_count)
1050                 qp_count = max_num_clients;
1051         else if (mw_count < qp_count)
1052                 qp_count = mw_count;
1053
1054         qp_bitmap &= BIT_ULL(qp_count) - 1;
1055
1056         nt->qp_count = qp_count;
1057         nt->qp_bitmap = qp_bitmap;
1058         nt->qp_bitmap_free = qp_bitmap;
1059
1060         nt->qp_vec = kzalloc_node(qp_count * sizeof(*nt->qp_vec),
1061                                   GFP_KERNEL, node);
1062         if (!nt->qp_vec) {
1063                 rc = -ENOMEM;
1064                 goto err2;
1065         }
1066
1067         if (nt_debugfs_dir) {
1068                 nt->debugfs_node_dir =
1069                         debugfs_create_dir(pci_name(ndev->pdev),
1070                                            nt_debugfs_dir);
1071         }
1072
1073         for (i = 0; i < qp_count; i++) {
1074                 rc = ntb_transport_init_queue(nt, i);
1075                 if (rc)
1076                         goto err3;
1077         }
1078
1079         INIT_DELAYED_WORK(&nt->link_work, ntb_transport_link_work);
1080         INIT_WORK(&nt->link_cleanup, ntb_transport_link_cleanup_work);
1081
1082         rc = ntb_set_ctx(ndev, nt, &ntb_transport_ops);
1083         if (rc)
1084                 goto err3;
1085
1086         INIT_LIST_HEAD(&nt->client_devs);
1087         rc = ntb_bus_init(nt);
1088         if (rc)
1089                 goto err4;
1090
1091         nt->link_is_up = false;
1092         ntb_link_enable(ndev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
1093         ntb_link_event(ndev);
1094
1095         return 0;
1096
1097 err4:
1098         ntb_clear_ctx(ndev);
1099 err3:
1100         kfree(nt->qp_vec);
1101 err2:
1102         kfree(nt->mw_vec);
1103 err1:
1104         while (i--) {
1105                 mw = &nt->mw_vec[i];
1106                 iounmap(mw->vbase);
1107         }
1108 err:
1109         kfree(nt);
1110         return rc;
1111 }
1112
1113 static void ntb_transport_free(struct ntb_client *self, struct ntb_dev *ndev)
1114 {
1115         struct ntb_transport_ctx *nt = ndev->ctx;
1116         struct ntb_transport_qp *qp;
1117         u64 qp_bitmap_alloc;
1118         int i;
1119
1120         ntb_transport_link_cleanup(nt);
1121         cancel_work_sync(&nt->link_cleanup);
1122         cancel_delayed_work_sync(&nt->link_work);
1123
1124         qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
1125
1126         /* verify that all the qp's are freed */
1127         for (i = 0; i < nt->qp_count; i++) {
1128                 qp = &nt->qp_vec[i];
1129                 if (qp_bitmap_alloc & BIT_ULL(i))
1130                         ntb_transport_free_queue(qp);
1131                 debugfs_remove_recursive(qp->debugfs_dir);
1132         }
1133
1134         ntb_link_disable(ndev);
1135         ntb_clear_ctx(ndev);
1136
1137         ntb_bus_remove(nt);
1138
1139         for (i = nt->mw_count; i--; ) {
1140                 ntb_free_mw(nt, i);
1141                 iounmap(nt->mw_vec[i].vbase);
1142         }
1143
1144         kfree(nt->qp_vec);
1145         kfree(nt->mw_vec);
1146         kfree(nt);
1147 }
1148
1149 static void ntb_complete_rxc(struct ntb_transport_qp *qp)
1150 {
1151         struct ntb_queue_entry *entry;
1152         void *cb_data;
1153         unsigned int len;
1154         unsigned long irqflags;
1155
1156         spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1157
1158         while (!list_empty(&qp->rx_post_q)) {
1159                 entry = list_first_entry(&qp->rx_post_q,
1160                                          struct ntb_queue_entry, entry);
1161                 if (!(entry->flags & DESC_DONE_FLAG))
1162                         break;
1163
1164                 entry->rx_hdr->flags = 0;
1165                 iowrite32(entry->index, &qp->rx_info->entry);
1166
1167                 cb_data = entry->cb_data;
1168                 len = entry->len;
1169
1170                 list_move_tail(&entry->entry, &qp->rx_free_q);
1171
1172                 spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1173
1174                 if (qp->rx_handler && qp->client_ready)
1175                         qp->rx_handler(qp, qp->cb_data, cb_data, len);
1176
1177                 spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1178         }
1179
1180         spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1181 }
1182
1183 static void ntb_rx_copy_callback(void *data)
1184 {
1185         struct ntb_queue_entry *entry = data;
1186
1187         entry->flags |= DESC_DONE_FLAG;
1188
1189         ntb_complete_rxc(entry->qp);
1190 }
1191
1192 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset)
1193 {
1194         void *buf = entry->buf;
1195         size_t len = entry->len;
1196
1197         memcpy(buf, offset, len);
1198
1199         /* Ensure that the data is fully copied out before clearing the flag */
1200         wmb();
1201
1202         ntb_rx_copy_callback(entry);
1203 }
1204
1205 static void ntb_async_rx(struct ntb_queue_entry *entry, void *offset)
1206 {
1207         struct dma_async_tx_descriptor *txd;
1208         struct ntb_transport_qp *qp = entry->qp;
1209         struct dma_chan *chan = qp->dma_chan;
1210         struct dma_device *device;
1211         size_t pay_off, buff_off, len;
1212         struct dmaengine_unmap_data *unmap;
1213         dma_cookie_t cookie;
1214         void *buf = entry->buf;
1215
1216         len = entry->len;
1217
1218         if (!chan)
1219                 goto err;
1220
1221         if (len < copy_bytes)
1222                 goto err_wait;
1223
1224         device = chan->device;
1225         pay_off = (size_t)offset & ~PAGE_MASK;
1226         buff_off = (size_t)buf & ~PAGE_MASK;
1227
1228         if (!is_dma_copy_aligned(device, pay_off, buff_off, len))
1229                 goto err_wait;
1230
1231         unmap = dmaengine_get_unmap_data(device->dev, 2, GFP_NOWAIT);
1232         if (!unmap)
1233                 goto err_wait;
1234
1235         unmap->len = len;
1236         unmap->addr[0] = dma_map_page(device->dev, virt_to_page(offset),
1237                                       pay_off, len, DMA_TO_DEVICE);
1238         if (dma_mapping_error(device->dev, unmap->addr[0]))
1239                 goto err_get_unmap;
1240
1241         unmap->to_cnt = 1;
1242
1243         unmap->addr[1] = dma_map_page(device->dev, virt_to_page(buf),
1244                                       buff_off, len, DMA_FROM_DEVICE);
1245         if (dma_mapping_error(device->dev, unmap->addr[1]))
1246                 goto err_get_unmap;
1247
1248         unmap->from_cnt = 1;
1249
1250         txd = device->device_prep_dma_memcpy(chan, unmap->addr[1],
1251                                              unmap->addr[0], len,
1252                                              DMA_PREP_INTERRUPT);
1253         if (!txd)
1254                 goto err_get_unmap;
1255
1256         txd->callback = ntb_rx_copy_callback;
1257         txd->callback_param = entry;
1258         dma_set_unmap(txd, unmap);
1259
1260         cookie = dmaengine_submit(txd);
1261         if (dma_submit_error(cookie))
1262                 goto err_set_unmap;
1263
1264         dmaengine_unmap_put(unmap);
1265
1266         qp->last_cookie = cookie;
1267
1268         qp->rx_async++;
1269
1270         return;
1271
1272 err_set_unmap:
1273         dmaengine_unmap_put(unmap);
1274 err_get_unmap:
1275         dmaengine_unmap_put(unmap);
1276 err_wait:
1277         /* If the callbacks come out of order, the writing of the index to the
1278          * last completed will be out of order.  This may result in the
1279          * receive stalling forever.
1280          */
1281         dma_sync_wait(chan, qp->last_cookie);
1282 err:
1283         ntb_memcpy_rx(entry, offset);
1284         qp->rx_memcpy++;
1285 }
1286
1287 static int ntb_process_rxc(struct ntb_transport_qp *qp)
1288 {
1289         struct ntb_payload_header *hdr;
1290         struct ntb_queue_entry *entry;
1291         void *offset;
1292
1293         offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
1294         hdr = offset + qp->rx_max_frame - sizeof(struct ntb_payload_header);
1295
1296         dev_dbg(&qp->ndev->pdev->dev, "qp %d: RX ver %u len %d flags %x\n",
1297                 qp->qp_num, hdr->ver, hdr->len, hdr->flags);
1298
1299         if (!(hdr->flags & DESC_DONE_FLAG)) {
1300                 dev_dbg(&qp->ndev->pdev->dev, "done flag not set\n");
1301                 qp->rx_ring_empty++;
1302                 return -EAGAIN;
1303         }
1304
1305         if (hdr->flags & LINK_DOWN_FLAG) {
1306                 dev_dbg(&qp->ndev->pdev->dev, "link down flag set\n");
1307                 ntb_qp_link_down(qp);
1308                 hdr->flags = 0;
1309                 return -EAGAIN;
1310         }
1311
1312         if (hdr->ver != (u32)qp->rx_pkts) {
1313                 dev_dbg(&qp->ndev->pdev->dev,
1314                         "version mismatch, expected %llu - got %u\n",
1315                         qp->rx_pkts, hdr->ver);
1316                 qp->rx_err_ver++;
1317                 return -EIO;
1318         }
1319
1320         entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q);
1321         if (!entry) {
1322                 dev_dbg(&qp->ndev->pdev->dev, "no receive buffer\n");
1323                 qp->rx_err_no_buf++;
1324                 return -EAGAIN;
1325         }
1326
1327         entry->rx_hdr = hdr;
1328         entry->index = qp->rx_index;
1329
1330         if (hdr->len > entry->len) {
1331                 dev_dbg(&qp->ndev->pdev->dev,
1332                         "receive buffer overflow! Wanted %d got %d\n",
1333                         hdr->len, entry->len);
1334                 qp->rx_err_oflow++;
1335
1336                 entry->len = -EIO;
1337                 entry->flags |= DESC_DONE_FLAG;
1338
1339                 ntb_complete_rxc(qp);
1340         } else {
1341                 dev_dbg(&qp->ndev->pdev->dev,
1342                         "RX OK index %u ver %u size %d into buf size %d\n",
1343                         qp->rx_index, hdr->ver, hdr->len, entry->len);
1344
1345                 qp->rx_bytes += hdr->len;
1346                 qp->rx_pkts++;
1347
1348                 entry->len = hdr->len;
1349
1350                 ntb_async_rx(entry, offset);
1351         }
1352
1353         qp->rx_index++;
1354         qp->rx_index %= qp->rx_max_entry;
1355
1356         return 0;
1357 }
1358
1359 static void ntb_transport_rxc_db(unsigned long data)
1360 {
1361         struct ntb_transport_qp *qp = (void *)data;
1362         int rc, i;
1363
1364         dev_dbg(&qp->ndev->pdev->dev, "%s: doorbell %d received\n",
1365                 __func__, qp->qp_num);
1366
1367         /* Limit the number of packets processed in a single interrupt to
1368          * provide fairness to others
1369          */
1370         for (i = 0; i < qp->rx_max_entry; i++) {
1371                 rc = ntb_process_rxc(qp);
1372                 if (rc)
1373                         break;
1374         }
1375
1376         if (i && qp->dma_chan)
1377                 dma_async_issue_pending(qp->dma_chan);
1378
1379         if (i == qp->rx_max_entry) {
1380                 /* there is more work to do */
1381                 tasklet_schedule(&qp->rxc_db_work);
1382         } else if (ntb_db_read(qp->ndev) & BIT_ULL(qp->qp_num)) {
1383                 /* the doorbell bit is set: clear it */
1384                 ntb_db_clear(qp->ndev, BIT_ULL(qp->qp_num));
1385                 /* ntb_db_read ensures ntb_db_clear write is committed */
1386                 ntb_db_read(qp->ndev);
1387
1388                 /* an interrupt may have arrived between finishing
1389                  * ntb_process_rxc and clearing the doorbell bit:
1390                  * there might be some more work to do.
1391                  */
1392                 tasklet_schedule(&qp->rxc_db_work);
1393         }
1394 }
1395
1396 static void ntb_tx_copy_callback(void *data)
1397 {
1398         struct ntb_queue_entry *entry = data;
1399         struct ntb_transport_qp *qp = entry->qp;
1400         struct ntb_payload_header __iomem *hdr = entry->tx_hdr;
1401
1402         iowrite32(entry->flags | DESC_DONE_FLAG, &hdr->flags);
1403
1404         ntb_peer_db_set(qp->ndev, BIT_ULL(qp->qp_num));
1405
1406         /* The entry length can only be zero if the packet is intended to be a
1407          * "link down" or similar.  Since no payload is being sent in these
1408          * cases, there is nothing to add to the completion queue.
1409          */
1410         if (entry->len > 0) {
1411                 qp->tx_bytes += entry->len;
1412
1413                 if (qp->tx_handler)
1414                         qp->tx_handler(qp, qp->cb_data, entry->cb_data,
1415                                        entry->len);
1416         }
1417
1418         ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry, &qp->tx_free_q);
1419 }
1420
1421 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset)
1422 {
1423 #ifdef ARCH_HAS_NOCACHE_UACCESS
1424         /*
1425          * Using non-temporal mov to improve performance on non-cached
1426          * writes, even though we aren't actually copying from user space.
1427          */
1428         __copy_from_user_inatomic_nocache(offset, entry->buf, entry->len);
1429 #else
1430         memcpy_toio(offset, entry->buf, entry->len);
1431 #endif
1432
1433         /* Ensure that the data is fully copied out before setting the flags */
1434         wmb();
1435
1436         ntb_tx_copy_callback(entry);
1437 }
1438
1439 static void ntb_async_tx(struct ntb_transport_qp *qp,
1440                          struct ntb_queue_entry *entry)
1441 {
1442         struct ntb_payload_header __iomem *hdr;
1443         struct dma_async_tx_descriptor *txd;
1444         struct dma_chan *chan = qp->dma_chan;
1445         struct dma_device *device;
1446         size_t dest_off, buff_off;
1447         struct dmaengine_unmap_data *unmap;
1448         dma_addr_t dest;
1449         dma_cookie_t cookie;
1450         void __iomem *offset;
1451         size_t len = entry->len;
1452         void *buf = entry->buf;
1453
1454         offset = qp->tx_mw + qp->tx_max_frame * qp->tx_index;
1455         hdr = offset + qp->tx_max_frame - sizeof(struct ntb_payload_header);
1456         entry->tx_hdr = hdr;
1457
1458         iowrite32(entry->len, &hdr->len);
1459         iowrite32((u32)qp->tx_pkts, &hdr->ver);
1460
1461         if (!chan)
1462                 goto err;
1463
1464         if (len < copy_bytes)
1465                 goto err;
1466
1467         device = chan->device;
1468         dest = qp->tx_mw_phys + qp->tx_max_frame * qp->tx_index;
1469         buff_off = (size_t)buf & ~PAGE_MASK;
1470         dest_off = (size_t)dest & ~PAGE_MASK;
1471
1472         if (!is_dma_copy_aligned(device, buff_off, dest_off, len))
1473                 goto err;
1474
1475         unmap = dmaengine_get_unmap_data(device->dev, 1, GFP_NOWAIT);
1476         if (!unmap)
1477                 goto err;
1478
1479         unmap->len = len;
1480         unmap->addr[0] = dma_map_page(device->dev, virt_to_page(buf),
1481                                       buff_off, len, DMA_TO_DEVICE);
1482         if (dma_mapping_error(device->dev, unmap->addr[0]))
1483                 goto err_get_unmap;
1484
1485         unmap->to_cnt = 1;
1486
1487         txd = device->device_prep_dma_memcpy(chan, dest, unmap->addr[0], len,
1488                                              DMA_PREP_INTERRUPT);
1489         if (!txd)
1490                 goto err_get_unmap;
1491
1492         txd->callback = ntb_tx_copy_callback;
1493         txd->callback_param = entry;
1494         dma_set_unmap(txd, unmap);
1495
1496         cookie = dmaengine_submit(txd);
1497         if (dma_submit_error(cookie))
1498                 goto err_set_unmap;
1499
1500         dmaengine_unmap_put(unmap);
1501
1502         dma_async_issue_pending(chan);
1503         qp->tx_async++;
1504
1505         return;
1506 err_set_unmap:
1507         dmaengine_unmap_put(unmap);
1508 err_get_unmap:
1509         dmaengine_unmap_put(unmap);
1510 err:
1511         ntb_memcpy_tx(entry, offset);
1512         qp->tx_memcpy++;
1513 }
1514
1515 static int ntb_process_tx(struct ntb_transport_qp *qp,
1516                           struct ntb_queue_entry *entry)
1517 {
1518         if (qp->tx_index == qp->remote_rx_info->entry) {
1519                 qp->tx_ring_full++;
1520                 return -EAGAIN;
1521         }
1522
1523         if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
1524                 if (qp->tx_handler)
1525                         qp->tx_handler(qp->cb_data, qp, NULL, -EIO);
1526
1527                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1528                              &qp->tx_free_q);
1529                 return 0;
1530         }
1531
1532         ntb_async_tx(qp, entry);
1533
1534         qp->tx_index++;
1535         qp->tx_index %= qp->tx_max_entry;
1536
1537         qp->tx_pkts++;
1538
1539         return 0;
1540 }
1541
1542 static void ntb_send_link_down(struct ntb_transport_qp *qp)
1543 {
1544         struct pci_dev *pdev = qp->ndev->pdev;
1545         struct ntb_queue_entry *entry;
1546         int i, rc;
1547
1548         if (!qp->link_is_up)
1549                 return;
1550
1551         dev_info(&pdev->dev, "qp %d: Send Link Down\n", qp->qp_num);
1552
1553         for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1554                 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1555                 if (entry)
1556                         break;
1557                 msleep(100);
1558         }
1559
1560         if (!entry)
1561                 return;
1562
1563         entry->cb_data = NULL;
1564         entry->buf = NULL;
1565         entry->len = 0;
1566         entry->flags = LINK_DOWN_FLAG;
1567
1568         rc = ntb_process_tx(qp, entry);
1569         if (rc)
1570                 dev_err(&pdev->dev, "ntb: QP%d unable to send linkdown msg\n",
1571                         qp->qp_num);
1572
1573         ntb_qp_link_down_reset(qp);
1574 }
1575
1576 static bool ntb_dma_filter_fn(struct dma_chan *chan, void *node)
1577 {
1578         return dev_to_node(&chan->dev->device) == (int)(unsigned long)node;
1579 }
1580
1581 /**
1582  * ntb_transport_create_queue - Create a new NTB transport layer queue
1583  * @rx_handler: receive callback function
1584  * @tx_handler: transmit callback function
1585  * @event_handler: event callback function
1586  *
1587  * Create a new NTB transport layer queue and provide the queue with a callback
1588  * routine for both transmit and receive.  The receive callback routine will be
1589  * used to pass up data when the transport has received it on the queue.   The
1590  * transmit callback routine will be called when the transport has completed the
1591  * transmission of the data on the queue and the data is ready to be freed.
1592  *
1593  * RETURNS: pointer to newly created ntb_queue, NULL on error.
1594  */
1595 struct ntb_transport_qp *
1596 ntb_transport_create_queue(void *data, struct device *client_dev,
1597                            const struct ntb_queue_handlers *handlers)
1598 {
1599         struct ntb_dev *ndev;
1600         struct pci_dev *pdev;
1601         struct ntb_transport_ctx *nt;
1602         struct ntb_queue_entry *entry;
1603         struct ntb_transport_qp *qp;
1604         u64 qp_bit;
1605         unsigned int free_queue;
1606         dma_cap_mask_t dma_mask;
1607         int node;
1608         int i;
1609
1610         ndev = dev_ntb(client_dev->parent);
1611         pdev = ndev->pdev;
1612         nt = ndev->ctx;
1613
1614         node = dev_to_node(&ndev->dev);
1615
1616         free_queue = ffs(nt->qp_bitmap);
1617         if (!free_queue)
1618                 goto err;
1619
1620         /* decrement free_queue to make it zero based */
1621         free_queue--;
1622
1623         qp = &nt->qp_vec[free_queue];
1624         qp_bit = BIT_ULL(qp->qp_num);
1625
1626         nt->qp_bitmap_free &= ~qp_bit;
1627
1628         qp->cb_data = data;
1629         qp->rx_handler = handlers->rx_handler;
1630         qp->tx_handler = handlers->tx_handler;
1631         qp->event_handler = handlers->event_handler;
1632
1633         dma_cap_zero(dma_mask);
1634         dma_cap_set(DMA_MEMCPY, dma_mask);
1635
1636         if (use_dma) {
1637                 qp->dma_chan = dma_request_channel(dma_mask, ntb_dma_filter_fn,
1638                                                    (void *)(unsigned long)node);
1639                 if (!qp->dma_chan)
1640                         dev_info(&pdev->dev, "Unable to allocate DMA channel\n");
1641         } else {
1642                 qp->dma_chan = NULL;
1643         }
1644         dev_dbg(&pdev->dev, "Using %s memcpy\n", qp->dma_chan ? "DMA" : "CPU");
1645
1646         for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
1647                 entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
1648                 if (!entry)
1649                         goto err1;
1650
1651                 entry->qp = qp;
1652                 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
1653                              &qp->rx_free_q);
1654         }
1655
1656         for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
1657                 entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
1658                 if (!entry)
1659                         goto err2;
1660
1661                 entry->qp = qp;
1662                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1663                              &qp->tx_free_q);
1664         }
1665
1666         ntb_db_clear(qp->ndev, qp_bit);
1667         ntb_db_clear_mask(qp->ndev, qp_bit);
1668
1669         dev_info(&pdev->dev, "NTB Transport QP %d created\n", qp->qp_num);
1670
1671         return qp;
1672
1673 err2:
1674         while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1675                 kfree(entry);
1676 err1:
1677         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
1678                 kfree(entry);
1679         if (qp->dma_chan)
1680                 dma_release_channel(qp->dma_chan);
1681         nt->qp_bitmap_free |= qp_bit;
1682 err:
1683         return NULL;
1684 }
1685 EXPORT_SYMBOL_GPL(ntb_transport_create_queue);
1686
1687 /**
1688  * ntb_transport_free_queue - Frees NTB transport queue
1689  * @qp: NTB queue to be freed
1690  *
1691  * Frees NTB transport queue
1692  */
1693 void ntb_transport_free_queue(struct ntb_transport_qp *qp)
1694 {
1695         struct pci_dev *pdev;
1696         struct ntb_queue_entry *entry;
1697         u64 qp_bit;
1698
1699         if (!qp)
1700                 return;
1701
1702         pdev = qp->ndev->pdev;
1703
1704         if (qp->dma_chan) {
1705                 struct dma_chan *chan = qp->dma_chan;
1706                 /* Putting the dma_chan to NULL will force any new traffic to be
1707                  * processed by the CPU instead of the DAM engine
1708                  */
1709                 qp->dma_chan = NULL;
1710
1711                 /* Try to be nice and wait for any queued DMA engine
1712                  * transactions to process before smashing it with a rock
1713                  */
1714                 dma_sync_wait(chan, qp->last_cookie);
1715                 dmaengine_terminate_all(chan);
1716                 dma_release_channel(chan);
1717         }
1718
1719         qp_bit = BIT_ULL(qp->qp_num);
1720
1721         ntb_db_set_mask(qp->ndev, qp_bit);
1722         tasklet_disable(&qp->rxc_db_work);
1723
1724         cancel_delayed_work_sync(&qp->link_work);
1725
1726         qp->cb_data = NULL;
1727         qp->rx_handler = NULL;
1728         qp->tx_handler = NULL;
1729         qp->event_handler = NULL;
1730
1731         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
1732                 kfree(entry);
1733
1734         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q))) {
1735                 dev_warn(&pdev->dev, "Freeing item from non-empty rx_pend_q\n");
1736                 kfree(entry);
1737         }
1738
1739         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q))) {
1740                 dev_warn(&pdev->dev, "Freeing item from non-empty rx_post_q\n");
1741                 kfree(entry);
1742         }
1743
1744         while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1745                 kfree(entry);
1746
1747         qp->transport->qp_bitmap_free |= qp_bit;
1748
1749         dev_info(&pdev->dev, "NTB Transport QP %d freed\n", qp->qp_num);
1750 }
1751 EXPORT_SYMBOL_GPL(ntb_transport_free_queue);
1752
1753 /**
1754  * ntb_transport_rx_remove - Dequeues enqueued rx packet
1755  * @qp: NTB queue to be freed
1756  * @len: pointer to variable to write enqueued buffers length
1757  *
1758  * Dequeues unused buffers from receive queue.  Should only be used during
1759  * shutdown of qp.
1760  *
1761  * RETURNS: NULL error value on error, or void* for success.
1762  */
1763 void *ntb_transport_rx_remove(struct ntb_transport_qp *qp, unsigned int *len)
1764 {
1765         struct ntb_queue_entry *entry;
1766         void *buf;
1767
1768         if (!qp || qp->client_ready)
1769                 return NULL;
1770
1771         entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q);
1772         if (!entry)
1773                 return NULL;
1774
1775         buf = entry->cb_data;
1776         *len = entry->len;
1777
1778         ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_free_q);
1779
1780         return buf;
1781 }
1782 EXPORT_SYMBOL_GPL(ntb_transport_rx_remove);
1783
1784 /**
1785  * ntb_transport_rx_enqueue - Enqueue a new NTB queue entry
1786  * @qp: NTB transport layer queue the entry is to be enqueued on
1787  * @cb: per buffer pointer for callback function to use
1788  * @data: pointer to data buffer that incoming packets will be copied into
1789  * @len: length of the data buffer
1790  *
1791  * Enqueue a new receive buffer onto the transport queue into which a NTB
1792  * payload can be received into.
1793  *
1794  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
1795  */
1796 int ntb_transport_rx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
1797                              unsigned int len)
1798 {
1799         struct ntb_queue_entry *entry;
1800
1801         if (!qp)
1802                 return -EINVAL;
1803
1804         entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q);
1805         if (!entry)
1806                 return -ENOMEM;
1807
1808         entry->cb_data = cb;
1809         entry->buf = data;
1810         entry->len = len;
1811         entry->flags = 0;
1812
1813         ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_pend_q);
1814
1815         tasklet_schedule(&qp->rxc_db_work);
1816
1817         return 0;
1818 }
1819 EXPORT_SYMBOL_GPL(ntb_transport_rx_enqueue);
1820
1821 /**
1822  * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
1823  * @qp: NTB transport layer queue the entry is to be enqueued on
1824  * @cb: per buffer pointer for callback function to use
1825  * @data: pointer to data buffer that will be sent
1826  * @len: length of the data buffer
1827  *
1828  * Enqueue a new transmit buffer onto the transport queue from which a NTB
1829  * payload will be transmitted.  This assumes that a lock is being held to
1830  * serialize access to the qp.
1831  *
1832  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
1833  */
1834 int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
1835                              unsigned int len)
1836 {
1837         struct ntb_queue_entry *entry;
1838         int rc;
1839
1840         if (!qp || !qp->link_is_up || !len)
1841                 return -EINVAL;
1842
1843         entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1844         if (!entry) {
1845                 qp->tx_err_no_buf++;
1846                 return -ENOMEM;
1847         }
1848
1849         entry->cb_data = cb;
1850         entry->buf = data;
1851         entry->len = len;
1852         entry->flags = 0;
1853
1854         rc = ntb_process_tx(qp, entry);
1855         if (rc)
1856                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1857                              &qp->tx_free_q);
1858
1859         return rc;
1860 }
1861 EXPORT_SYMBOL_GPL(ntb_transport_tx_enqueue);
1862
1863 /**
1864  * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
1865  * @qp: NTB transport layer queue to be enabled
1866  *
1867  * Notify NTB transport layer of client readiness to use queue
1868  */
1869 void ntb_transport_link_up(struct ntb_transport_qp *qp)
1870 {
1871         if (!qp)
1872                 return;
1873
1874         qp->client_ready = true;
1875
1876         if (qp->transport->link_is_up)
1877                 schedule_delayed_work(&qp->link_work, 0);
1878 }
1879 EXPORT_SYMBOL_GPL(ntb_transport_link_up);
1880
1881 /**
1882  * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
1883  * @qp: NTB transport layer queue to be disabled
1884  *
1885  * Notify NTB transport layer of client's desire to no longer receive data on
1886  * transport queue specified.  It is the client's responsibility to ensure all
1887  * entries on queue are purged or otherwise handled appropriately.
1888  */
1889 void ntb_transport_link_down(struct ntb_transport_qp *qp)
1890 {
1891         struct pci_dev *pdev;
1892         int val;
1893
1894         if (!qp)
1895                 return;
1896
1897         pdev = qp->ndev->pdev;
1898         qp->client_ready = false;
1899
1900         val = ntb_spad_read(qp->ndev, QP_LINKS);
1901
1902         ntb_peer_spad_write(qp->ndev, QP_LINKS,
1903                             val & ~BIT(qp->qp_num));
1904
1905         if (qp->link_is_up)
1906                 ntb_send_link_down(qp);
1907         else
1908                 cancel_delayed_work_sync(&qp->link_work);
1909 }
1910 EXPORT_SYMBOL_GPL(ntb_transport_link_down);
1911
1912 /**
1913  * ntb_transport_link_query - Query transport link state
1914  * @qp: NTB transport layer queue to be queried
1915  *
1916  * Query connectivity to the remote system of the NTB transport queue
1917  *
1918  * RETURNS: true for link up or false for link down
1919  */
1920 bool ntb_transport_link_query(struct ntb_transport_qp *qp)
1921 {
1922         if (!qp)
1923                 return false;
1924
1925         return qp->link_is_up;
1926 }
1927 EXPORT_SYMBOL_GPL(ntb_transport_link_query);
1928
1929 /**
1930  * ntb_transport_qp_num - Query the qp number
1931  * @qp: NTB transport layer queue to be queried
1932  *
1933  * Query qp number of the NTB transport queue
1934  *
1935  * RETURNS: a zero based number specifying the qp number
1936  */
1937 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
1938 {
1939         if (!qp)
1940                 return 0;
1941
1942         return qp->qp_num;
1943 }
1944 EXPORT_SYMBOL_GPL(ntb_transport_qp_num);
1945
1946 /**
1947  * ntb_transport_max_size - Query the max payload size of a qp
1948  * @qp: NTB transport layer queue to be queried
1949  *
1950  * Query the maximum payload size permissible on the given qp
1951  *
1952  * RETURNS: the max payload size of a qp
1953  */
1954 unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp)
1955 {
1956         unsigned int max;
1957
1958         if (!qp)
1959                 return 0;
1960
1961         if (!qp->dma_chan)
1962                 return qp->tx_max_frame - sizeof(struct ntb_payload_header);
1963
1964         /* If DMA engine usage is possible, try to find the max size for that */
1965         max = qp->tx_max_frame - sizeof(struct ntb_payload_header);
1966         max -= max % (1 << qp->dma_chan->device->copy_align);
1967
1968         return max;
1969 }
1970 EXPORT_SYMBOL_GPL(ntb_transport_max_size);
1971
1972 static void ntb_transport_doorbell_callback(void *data, int vector)
1973 {
1974         struct ntb_transport_ctx *nt = data;
1975         struct ntb_transport_qp *qp;
1976         u64 db_bits;
1977         unsigned int qp_num;
1978
1979         db_bits = (nt->qp_bitmap & ~nt->qp_bitmap_free &
1980                    ntb_db_vector_mask(nt->ndev, vector));
1981
1982         while (db_bits) {
1983                 qp_num = __ffs(db_bits);
1984                 qp = &nt->qp_vec[qp_num];
1985
1986                 tasklet_schedule(&qp->rxc_db_work);
1987
1988                 db_bits &= ~BIT_ULL(qp_num);
1989         }
1990 }
1991
1992 static const struct ntb_ctx_ops ntb_transport_ops = {
1993         .link_event = ntb_transport_event_callback,
1994         .db_event = ntb_transport_doorbell_callback,
1995 };
1996
1997 static struct ntb_client ntb_transport_client = {
1998         .ops = {
1999                 .probe = ntb_transport_probe,
2000                 .remove = ntb_transport_free,
2001         },
2002 };
2003
2004 static int __init ntb_transport_init(void)
2005 {
2006         int rc;
2007
2008         pr_info("%s, version %s\n", NTB_TRANSPORT_DESC, NTB_TRANSPORT_VER);
2009
2010         if (debugfs_initialized())
2011                 nt_debugfs_dir = debugfs_create_dir(KBUILD_MODNAME, NULL);
2012
2013         rc = bus_register(&ntb_transport_bus);
2014         if (rc)
2015                 goto err_bus;
2016
2017         rc = ntb_register_client(&ntb_transport_client);
2018         if (rc)
2019                 goto err_client;
2020
2021         return 0;
2022
2023 err_client:
2024         bus_unregister(&ntb_transport_bus);
2025 err_bus:
2026         debugfs_remove_recursive(nt_debugfs_dir);
2027         return rc;
2028 }
2029 module_init(ntb_transport_init);
2030
2031 static void __exit ntb_transport_exit(void)
2032 {
2033         debugfs_remove_recursive(nt_debugfs_dir);
2034
2035         ntb_unregister_client(&ntb_transport_client);
2036         bus_unregister(&ntb_transport_bus);
2037 }
2038 module_exit(ntb_transport_exit);