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         int retries;
 106         int errors;
 107         unsigned int tx_index;
 108         unsigned int rx_index;
 109
 110         struct ntb_transport_qp *qp;
 111         union {
 112                 struct ntb_payload_header __iomem *tx_hdr;
 113                 struct ntb_payload_header *rx_hdr;
 114         };
 115 };
 116
 117 struct ntb_rx_info {
 118         unsigned int entry;
 119 };
 120
 121 struct ntb_transport_qp {
 122         struct ntb_transport_ctx *transport;
 123         struct ntb_dev *ndev;
 124         void *cb_data;
 125         struct dma_chan *tx_dma_chan;
 126         struct dma_chan *rx_dma_chan;
 127
 128         bool client_ready;
 129         bool link_is_up;
 130         bool active;
 131
 132         u8 qp_num;      /* Only 64 QP's are allowed.  0-63 */
 133         u64 qp_bit;
 134
 135         struct ntb_rx_info __iomem *rx_info;
 136         struct ntb_rx_info *remote_rx_info;
 137
 138         void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data,
 139                            void *data, int len);
 140         struct list_head tx_free_q;
 141         spinlock_t ntb_tx_free_q_lock;
 142         void __iomem *tx_mw;
 143         dma_addr_t tx_mw_phys;
 144         unsigned int tx_index;
 145         unsigned int tx_max_entry;
 146         unsigned int tx_max_frame;
 147
 148         void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
 149                            void *data, int len);
 150         struct list_head rx_post_q;
 151         struct list_head rx_pend_q;
 152         struct list_head rx_free_q;
 153         /* ntb_rx_q_lock: synchronize access to rx_XXXX_q */
 154         spinlock_t ntb_rx_q_lock;
 155         void *rx_buff;
 156         unsigned int rx_index;
 157         unsigned int rx_max_entry;
 158         unsigned int rx_max_frame;
 159         unsigned int rx_alloc_entry;
 160         dma_cookie_t last_cookie;
 161         struct tasklet_struct rxc_db_work;
 162
 163         void (*event_handler)(void *data, int status);
 164         struct delayed_work link_work;
 165         struct work_struct link_cleanup;
 166
 167         struct dentry *debugfs_dir;
 168         struct dentry *debugfs_stats;
 169
 170         /* Stats */
 171         u64 rx_bytes;
 172         u64 rx_pkts;
 173         u64 rx_ring_empty;
 174         u64 rx_err_no_buf;
 175         u64 rx_err_oflow;
 176         u64 rx_err_ver;
 177         u64 rx_memcpy;
 178         u64 rx_async;
 179         u64 tx_bytes;
 180         u64 tx_pkts;
 181         u64 tx_ring_full;
 182         u64 tx_err_no_buf;
 183         u64 tx_memcpy;
 184         u64 tx_async;
 185 };
 186
 187 struct ntb_transport_mw {
 188         phys_addr_t phys_addr;
 189         resource_size_t phys_size;
 190         resource_size_t xlat_align;
 191         resource_size_t xlat_align_size;
 192         void __iomem *vbase;
 193         size_t xlat_size;
 194         size_t buff_size;
 195         void *virt_addr;
 196         dma_addr_t dma_addr;
 197 };
 198
 199 struct ntb_transport_client_dev {
 200         struct list_head entry;
 201         struct ntb_transport_ctx *nt;
 202         struct device dev;
 203 };
 204
 205 struct ntb_transport_ctx {
 206         struct list_head entry;
 207         struct list_head client_devs;
 208
 209         struct ntb_dev *ndev;
 210
 211         struct ntb_transport_mw *mw_vec;
 212         struct ntb_transport_qp *qp_vec;
 213         unsigned int mw_count;
 214         unsigned int qp_count;
 215         u64 qp_bitmap;
 216         u64 qp_bitmap_free;
 217
 218         bool link_is_up;
 219         struct delayed_work link_work;
 220         struct work_struct link_cleanup;
 221
 222         struct dentry *debugfs_node_dir;
 223 };
 224
 225 enum {
 226         DESC_DONE_FLAG = BIT(0),
 227         LINK_DOWN_FLAG = BIT(1),
 228 };
 229
 230 struct ntb_payload_header {
 231         unsigned int ver;
 232         unsigned int len;
 233         unsigned int flags;
 234 };
 235
 236 enum {
 237         VERSION = 0,
 238         QP_LINKS,
 239         NUM_QPS,
 240         NUM_MWS,
 241         MW0_SZ_HIGH,
 242         MW0_SZ_LOW,
 243         MW1_SZ_HIGH,
 244         MW1_SZ_LOW,
 245         MAX_SPAD,
 246 };
 247
 248 #define dev_client_dev(__dev) \
 249         container_of((__dev), struct ntb_transport_client_dev, dev)
 250
 251 #define drv_client(__drv) \
 252         container_of((__drv), struct ntb_transport_client, driver)
 253
 254 #define QP_TO_MW(nt, qp)        ((qp) % nt->mw_count)
 255 #define NTB_QP_DEF_NUM_ENTRIES  100
 256 #define NTB_LINK_DOWN_TIMEOUT   10
 257
 258 static void ntb_transport_rxc_db(unsigned long data);
 259 static const struct ntb_ctx_ops ntb_transport_ops;
 260 static struct ntb_client ntb_transport_client;
 261 static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
 262                                struct ntb_queue_entry *entry);
 263 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset);
 264 static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset);
 265 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset);
 266
 267
 268 static int ntb_transport_bus_match(struct device *dev,
 269                                    struct device_driver *drv)
 270 {
 271         return !strncmp(dev_name(dev), drv->name, strlen(drv->name));
 272 }
 273
 274 static int ntb_transport_bus_probe(struct device *dev)
 275 {
 276         const struct ntb_transport_client *client;
 277         int rc = -EINVAL;
 278
 279         get_device(dev);
 280
 281         client = drv_client(dev->driver);
 282         rc = client->probe(dev);
 283         if (rc)
 284                 put_device(dev);
 285
 286         return rc;
 287 }
 288
 289 static int ntb_transport_bus_remove(struct device *dev)
 290 {
 291         const struct ntb_transport_client *client;
 292
 293         client = drv_client(dev->driver);
 294         client->remove(dev);
 295
 296         put_device(dev);
 297
 298         return 0;
 299 }
 300
 301 static struct bus_type ntb_transport_bus = {
 302         .name = "ntb_transport",
 303         .match = ntb_transport_bus_match,
 304         .probe = ntb_transport_bus_probe,
 305         .remove = ntb_transport_bus_remove,
 306 };
 307
 308 static LIST_HEAD(ntb_transport_list);
 309
 310 static int ntb_bus_init(struct ntb_transport_ctx *nt)
 311 {
 312         list_add_tail(&nt->entry, &ntb_transport_list);
 313         return 0;
 314 }
 315
 316 static void ntb_bus_remove(struct ntb_transport_ctx *nt)
 317 {
 318         struct ntb_transport_client_dev *client_dev, *cd;
 319
 320         list_for_each_entry_safe(client_dev, cd, &nt->client_devs, entry) {
 321                 dev_err(client_dev->dev.parent, "%s still attached to bus, removing\n",
 322                         dev_name(&client_dev->dev));
 323                 list_del(&client_dev->entry);
 324                 device_unregister(&client_dev->dev);
 325         }
 326
 327         list_del(&nt->entry);
 328 }
 329
 330 static void ntb_transport_client_release(struct device *dev)
 331 {
 332         struct ntb_transport_client_dev *client_dev;
 333
 334         client_dev = dev_client_dev(dev);
 335         kfree(client_dev);
 336 }
 337
 338 /**
 339  * ntb_transport_unregister_client_dev - Unregister NTB client device
 340  * @device_name: Name of NTB client device
 341  *
 342  * Unregister an NTB client device with the NTB transport layer
 343  */
 344 void ntb_transport_unregister_client_dev(char *device_name)
 345 {
 346         struct ntb_transport_client_dev *client, *cd;
 347         struct ntb_transport_ctx *nt;
 348
 349         list_for_each_entry(nt, &ntb_transport_list, entry)
 350                 list_for_each_entry_safe(client, cd, &nt->client_devs, entry)
 351                         if (!strncmp(dev_name(&client->dev), device_name,
 352                                      strlen(device_name))) {
 353                                 list_del(&client->entry);
 354                                 device_unregister(&client->dev);
 355                         }
 356 }
 357 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client_dev);
 358
 359 /**
 360  * ntb_transport_register_client_dev - Register NTB client device
 361  * @device_name: Name of NTB client device
 362  *
 363  * Register an NTB client device with the NTB transport layer
 364  */
 365 int ntb_transport_register_client_dev(char *device_name)
 366 {
 367         struct ntb_transport_client_dev *client_dev;
 368         struct ntb_transport_ctx *nt;
 369         int node;
 370         int rc, i = 0;
 371
 372         if (list_empty(&ntb_transport_list))
 373                 return -ENODEV;
 374
 375         list_for_each_entry(nt, &ntb_transport_list, entry) {
 376                 struct device *dev;
 377
 378                 node = dev_to_node(&nt->ndev->dev);
 379
 380                 client_dev = kzalloc_node(sizeof(*client_dev),
 381                                           GFP_KERNEL, node);
 382                 if (!client_dev) {
 383                         rc = -ENOMEM;
 384                         goto err;
 385                 }
 386
 387                 dev = &client_dev->dev;
 388
 389                 /* setup and register client devices */
 390                 dev_set_name(dev, "%s%d", device_name, i);
 391                 dev->bus = &ntb_transport_bus;
 392                 dev->release = ntb_transport_client_release;
 393                 dev->parent = &nt->ndev->dev;
 394
 395                 rc = device_register(dev);
 396                 if (rc) {
 397                         kfree(client_dev);
 398                         goto err;
 399                 }
 400
 401                 list_add_tail(&client_dev->entry, &nt->client_devs);
 402                 i++;
 403         }
 404
 405         return 0;
 406
 407 err:
 408         ntb_transport_unregister_client_dev(device_name);
 409
 410         return rc;
 411 }
 412 EXPORT_SYMBOL_GPL(ntb_transport_register_client_dev);
 413
 414 /**
 415  * ntb_transport_register_client - Register NTB client driver
 416  * @drv: NTB client driver to be registered
 417  *
 418  * Register an NTB client driver with the NTB transport layer
 419  *
 420  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 421  */
 422 int ntb_transport_register_client(struct ntb_transport_client *drv)
 423 {
 424         drv->driver.bus = &ntb_transport_bus;
 425
 426         if (list_empty(&ntb_transport_list))
 427                 return -ENODEV;
 428
 429         return driver_register(&drv->driver);
 430 }
 431 EXPORT_SYMBOL_GPL(ntb_transport_register_client);
 432
 433 /**
 434  * ntb_transport_unregister_client - Unregister NTB client driver
 435  * @drv: NTB client driver to be unregistered
 436  *
 437  * Unregister an NTB client driver with the NTB transport layer
 438  *
 439  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 440  */
 441 void ntb_transport_unregister_client(struct ntb_transport_client *drv)
 442 {
 443         driver_unregister(&drv->driver);
 444 }
 445 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client);
 446
 447 static ssize_t debugfs_read(struct file *filp, char __user *ubuf, size_t count,
 448                             loff_t *offp)
 449 {
 450         struct ntb_transport_qp *qp;
 451         char *buf;
 452         ssize_t ret, out_offset, out_count;
 453
 454         qp = filp->private_data;
 455
 456         if (!qp || !qp->link_is_up)
 457                 return 0;
 458
 459         out_count = 1000;
 460
 461         buf = kmalloc(out_count, GFP_KERNEL);
 462         if (!buf)
 463                 return -ENOMEM;
 464
 465         out_offset = 0;
 466         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 467                                "\nNTB QP stats:\n\n");
 468         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 469                                "rx_bytes - \t%llu\n", qp->rx_bytes);
 470         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 471                                "rx_pkts - \t%llu\n", qp->rx_pkts);
 472         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 473                                "rx_memcpy - \t%llu\n", qp->rx_memcpy);
 474         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 475                                "rx_async - \t%llu\n", qp->rx_async);
 476         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 477                                "rx_ring_empty - %llu\n", qp->rx_ring_empty);
 478         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 479                                "rx_err_no_buf - %llu\n", qp->rx_err_no_buf);
 480         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 481                                "rx_err_oflow - \t%llu\n", qp->rx_err_oflow);
 482         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 483                                "rx_err_ver - \t%llu\n", qp->rx_err_ver);
 484         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 485                                "rx_buff - \t0x%p\n", qp->rx_buff);
 486         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 487                                "rx_index - \t%u\n", qp->rx_index);
 488         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 489                                "rx_max_entry - \t%u\n", qp->rx_max_entry);
 490         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 491                                "rx_alloc_entry - \t%u\n\n", qp->rx_alloc_entry);
 492
 493         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 494                                "tx_bytes - \t%llu\n", qp->tx_bytes);
 495         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 496                                "tx_pkts - \t%llu\n", qp->tx_pkts);
 497         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 498                                "tx_memcpy - \t%llu\n", qp->tx_memcpy);
 499         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 500                                "tx_async - \t%llu\n", qp->tx_async);
 501         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 502                                "tx_ring_full - \t%llu\n", qp->tx_ring_full);
 503         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 504                                "tx_err_no_buf - %llu\n", qp->tx_err_no_buf);
 505         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 506                                "tx_mw - \t0x%p\n", qp->tx_mw);
 507         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 508                                "tx_index (H) - \t%u\n", qp->tx_index);
 509         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 510                                "RRI (T) - \t%u\n",
 511                                qp->remote_rx_info->entry);
 512         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 513                                "tx_max_entry - \t%u\n", qp->tx_max_entry);
 514         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 515                                "free tx - \t%u\n",
 516                                ntb_transport_tx_free_entry(qp));
 517
 518         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 519                                "\n");
 520         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 521                                "Using TX DMA - \t%s\n",
 522                                qp->tx_dma_chan ? "Yes" : "No");
 523         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 524                                "Using RX DMA - \t%s\n",
 525                                qp->rx_dma_chan ? "Yes" : "No");
 526         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 527                                "QP Link - \t%s\n",
 528                                qp->link_is_up ? "Up" : "Down");
 529         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 530                                "\n");
 531
 532         if (out_offset > out_count)
 533                 out_offset = out_count;
 534
 535         ret = simple_read_from_buffer(ubuf, count, offp, buf, out_offset);
 536         kfree(buf);
 537         return ret;
 538 }
 539
 540 static const struct file_operations ntb_qp_debugfs_stats = {
 541         .owner = THIS_MODULE,
 542         .open = simple_open,
 543         .read = debugfs_read,
 544 };
 545
 546 static void ntb_list_add(spinlock_t *lock, struct list_head *entry,
 547                          struct list_head *list)
 548 {
 549         unsigned long flags;
 550
 551         spin_lock_irqsave(lock, flags);
 552         list_add_tail(entry, list);
 553         spin_unlock_irqrestore(lock, flags);
 554 }
 555
 556 static struct ntb_queue_entry *ntb_list_rm(spinlock_t *lock,
 557                                            struct list_head *list)
 558 {
 559         struct ntb_queue_entry *entry;
 560         unsigned long flags;
 561
 562         spin_lock_irqsave(lock, flags);
 563         if (list_empty(list)) {
 564                 entry = NULL;
 565                 goto out;
 566         }
 567         entry = list_first_entry(list, struct ntb_queue_entry, entry);
 568         list_del(&entry->entry);
 569
 570 out:
 571         spin_unlock_irqrestore(lock, flags);
 572
 573         return entry;
 574 }
 575
 576 static struct ntb_queue_entry *ntb_list_mv(spinlock_t *lock,
 577                                            struct list_head *list,
 578                                            struct list_head *to_list)
 579 {
 580         struct ntb_queue_entry *entry;
 581         unsigned long flags;
 582
 583         spin_lock_irqsave(lock, flags);
 584
 585         if (list_empty(list)) {
 586                 entry = NULL;
 587         } else {
 588                 entry = list_first_entry(list, struct ntb_queue_entry, entry);
 589                 list_move_tail(&entry->entry, to_list);
 590         }
 591
 592         spin_unlock_irqrestore(lock, flags);
 593
 594         return entry;
 595 }
 596
 597 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
 598                                      unsigned int qp_num)
 599 {
 600         struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
 601         struct ntb_transport_mw *mw;
 602         struct ntb_dev *ndev = nt->ndev;
 603         struct ntb_queue_entry *entry;
 604         unsigned int rx_size, num_qps_mw;
 605         unsigned int mw_num, mw_count, qp_count;
 606         unsigned int i;
 607         int node;
 608
 609         mw_count = nt->mw_count;
 610         qp_count = nt->qp_count;
 611
 612         mw_num = QP_TO_MW(nt, qp_num);
 613         mw = &nt->mw_vec[mw_num];
 614
 615         if (!mw->virt_addr)
 616                 return -ENOMEM;
 617
 618         if (mw_num < qp_count % mw_count)
 619                 num_qps_mw = qp_count / mw_count + 1;
 620         else
 621                 num_qps_mw = qp_count / mw_count;
 622
 623         rx_size = (unsigned int)mw->xlat_size / num_qps_mw;
 624         qp->rx_buff = mw->virt_addr + rx_size * (qp_num / mw_count);
 625         rx_size -= sizeof(struct ntb_rx_info);
 626
 627         qp->remote_rx_info = qp->rx_buff + rx_size;
 628
 629         /* Due to housekeeping, there must be atleast 2 buffs */
 630         qp->rx_max_frame = min(transport_mtu, rx_size / 2);
 631         qp->rx_max_entry = rx_size / qp->rx_max_frame;
 632         qp->rx_index = 0;
 633
 634         /*
 635          * Checking to see if we have more entries than the default.
 636          * We should add additional entries if that is the case so we
 637          * can be in sync with the transport frames.
 638          */
 639         node = dev_to_node(&ndev->dev);
 640         for (i = qp->rx_alloc_entry; i < qp->rx_max_entry; i++) {
 641                 entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
 642                 if (!entry)
 643                         return -ENOMEM;
 644
 645                 entry->qp = qp;
 646                 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
 647                              &qp->rx_free_q);
 648                 qp->rx_alloc_entry++;
 649         }
 650
 651         qp->remote_rx_info->entry = qp->rx_max_entry - 1;
 652
 653         /* setup the hdr offsets with 0's */
 654         for (i = 0; i < qp->rx_max_entry; i++) {
 655                 void *offset = (qp->rx_buff + qp->rx_max_frame * (i + 1) -
 656                                 sizeof(struct ntb_payload_header));
 657                 memset(offset, 0, sizeof(struct ntb_payload_header));
 658         }
 659
 660         qp->rx_pkts = 0;
 661         qp->tx_pkts = 0;
 662         qp->tx_index = 0;
 663
 664         return 0;
 665 }
 666
 667 static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
 668 {
 669         struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
 670         struct pci_dev *pdev = nt->ndev->pdev;
 671
 672         if (!mw->virt_addr)
 673                 return;
 674
 675         ntb_mw_clear_trans(nt->ndev, num_mw);
 676         dma_free_coherent(&pdev->dev, mw->buff_size,
 677                           mw->virt_addr, mw->dma_addr);
 678         mw->xlat_size = 0;
 679         mw->buff_size = 0;
 680         mw->virt_addr = NULL;
 681 }
 682
 683 static int ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw,
 684                       resource_size_t size)
 685 {
 686         struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
 687         struct pci_dev *pdev = nt->ndev->pdev;
 688         size_t xlat_size, buff_size;
 689         int rc;
 690
 691         if (!size)
 692                 return -EINVAL;
 693
 694         xlat_size = round_up(size, mw->xlat_align_size);
 695         buff_size = round_up(size, mw->xlat_align);
 696
 697         /* No need to re-setup */
 698         if (mw->xlat_size == xlat_size)
 699                 return 0;
 700
 701         if (mw->buff_size)
 702                 ntb_free_mw(nt, num_mw);
 703
 704         /* Alloc memory for receiving data.  Must be aligned */
 705         mw->xlat_size = xlat_size;
 706         mw->buff_size = buff_size;
 707
 708         mw->virt_addr = dma_alloc_coherent(&pdev->dev, buff_size,
 709                                            &mw->dma_addr, GFP_KERNEL);
 710         if (!mw->virt_addr) {
 711                 mw->xlat_size = 0;
 712                 mw->buff_size = 0;
 713                 dev_err(&pdev->dev, "Unable to alloc MW buff of size %zu\n",
 714                         buff_size);
 715                 return -ENOMEM;
 716         }
 717
 718         /*
 719          * we must ensure that the memory address allocated is BAR size
 720          * aligned in order for the XLAT register to take the value. This
 721          * is a requirement of the hardware. It is recommended to setup CMA
 722          * for BAR sizes equal or greater than 4MB.
 723          */
 724         if (!IS_ALIGNED(mw->dma_addr, mw->xlat_align)) {
 725                 dev_err(&pdev->dev, "DMA memory %pad is not aligned\n",
 726                         &mw->dma_addr);
 727                 ntb_free_mw(nt, num_mw);
 728                 return -ENOMEM;
 729         }
 730
 731         /* Notify HW the memory location of the receive buffer */
 732         rc = ntb_mw_set_trans(nt->ndev, num_mw, mw->dma_addr, mw->xlat_size);
 733         if (rc) {
 734                 dev_err(&pdev->dev, "Unable to set mw%d translation", num_mw);
 735                 ntb_free_mw(nt, num_mw);
 736                 return -EIO;
 737         }
 738
 739         return 0;
 740 }
 741
 742 static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
 743 {
 744         qp->link_is_up = false;
 745         qp->active = false;
 746
 747         qp->tx_index = 0;
 748         qp->rx_index = 0;
 749         qp->rx_bytes = 0;
 750         qp->rx_pkts = 0;
 751         qp->rx_ring_empty = 0;
 752         qp->rx_err_no_buf = 0;
 753         qp->rx_err_oflow = 0;
 754         qp->rx_err_ver = 0;
 755         qp->rx_memcpy = 0;
 756         qp->rx_async = 0;
 757         qp->tx_bytes = 0;
 758         qp->tx_pkts = 0;
 759         qp->tx_ring_full = 0;
 760         qp->tx_err_no_buf = 0;
 761         qp->tx_memcpy = 0;
 762         qp->tx_async = 0;
 763 }
 764
 765 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
 766 {
 767         struct ntb_transport_ctx *nt = qp->transport;
 768         struct pci_dev *pdev = nt->ndev->pdev;
 769
 770         dev_info(&pdev->dev, "qp %d: Link Cleanup\n", qp->qp_num);
 771
 772         cancel_delayed_work_sync(&qp->link_work);
 773         ntb_qp_link_down_reset(qp);
 774
 775         if (qp->event_handler)
 776                 qp->event_handler(qp->cb_data, qp->link_is_up);
 777 }
 778
 779 static void ntb_qp_link_cleanup_work(struct work_struct *work)
 780 {
 781         struct ntb_transport_qp *qp = container_of(work,
 782                                                    struct ntb_transport_qp,
 783                                                    link_cleanup);
 784         struct ntb_transport_ctx *nt = qp->transport;
 785
 786         ntb_qp_link_cleanup(qp);
 787
 788         if (nt->link_is_up)
 789                 schedule_delayed_work(&qp->link_work,
 790                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 791 }
 792
 793 static void ntb_qp_link_down(struct ntb_transport_qp *qp)
 794 {
 795         schedule_work(&qp->link_cleanup);
 796 }
 797
 798 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
 799 {
 800         struct ntb_transport_qp *qp;
 801         u64 qp_bitmap_alloc;
 802         int i;
 803
 804         qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
 805
 806         /* Pass along the info to any clients */
 807         for (i = 0; i < nt->qp_count; i++)
 808                 if (qp_bitmap_alloc & BIT_ULL(i)) {
 809                         qp = &nt->qp_vec[i];
 810                         ntb_qp_link_cleanup(qp);
 811                         cancel_work_sync(&qp->link_cleanup);
 812                         cancel_delayed_work_sync(&qp->link_work);
 813                 }
 814
 815         if (!nt->link_is_up)
 816                 cancel_delayed_work_sync(&nt->link_work);
 817
 818         /* The scratchpad registers keep the values if the remote side
 819          * goes down, blast them now to give them a sane value the next
 820          * time they are accessed
 821          */
 822         for (i = 0; i < MAX_SPAD; i++)
 823                 ntb_spad_write(nt->ndev, i, 0);
 824 }
 825
 826 static void ntb_transport_link_cleanup_work(struct work_struct *work)
 827 {
 828         struct ntb_transport_ctx *nt =
 829                 container_of(work, struct ntb_transport_ctx, link_cleanup);
 830
 831         ntb_transport_link_cleanup(nt);
 832 }
 833
 834 static void ntb_transport_event_callback(void *data)
 835 {
 836         struct ntb_transport_ctx *nt = data;
 837
 838         if (ntb_link_is_up(nt->ndev, NULL, NULL) == 1)
 839                 schedule_delayed_work(&nt->link_work, 0);
 840         else
 841                 schedule_work(&nt->link_cleanup);
 842 }
 843
 844 static void ntb_transport_link_work(struct work_struct *work)
 845 {
 846         struct ntb_transport_ctx *nt =
 847                 container_of(work, struct ntb_transport_ctx, link_work.work);
 848         struct ntb_dev *ndev = nt->ndev;
 849         struct pci_dev *pdev = ndev->pdev;
 850         resource_size_t size;
 851         u32 val;
 852         int rc = 0, i, spad;
 853
 854         /* send the local info, in the opposite order of the way we read it */
 855         for (i = 0; i < nt->mw_count; i++) {
 856                 size = nt->mw_vec[i].phys_size;
 857
 858                 if (max_mw_size && size > max_mw_size)
 859                         size = max_mw_size;
 860
 861                 spad = MW0_SZ_HIGH + (i * 2);
 862                 ntb_peer_spad_write(ndev, spad, upper_32_bits(size));
 863
 864                 spad = MW0_SZ_LOW + (i * 2);
 865                 ntb_peer_spad_write(ndev, spad, lower_32_bits(size));
 866         }
 867
 868         ntb_peer_spad_write(ndev, NUM_MWS, nt->mw_count);
 869
 870         ntb_peer_spad_write(ndev, NUM_QPS, nt->qp_count);
 871
 872         ntb_peer_spad_write(ndev, VERSION, NTB_TRANSPORT_VERSION);
 873
 874         /* Query the remote side for its info */
 875         val = ntb_spad_read(ndev, VERSION);
 876         dev_dbg(&pdev->dev, "Remote version = %d\n", val);
 877         if (val != NTB_TRANSPORT_VERSION)
 878                 goto out;
 879
 880         val = ntb_spad_read(ndev, NUM_QPS);
 881         dev_dbg(&pdev->dev, "Remote max number of qps = %d\n", val);
 882         if (val != nt->qp_count)
 883                 goto out;
 884
 885         val = ntb_spad_read(ndev, NUM_MWS);
 886         dev_dbg(&pdev->dev, "Remote number of mws = %d\n", val);
 887         if (val != nt->mw_count)
 888                 goto out;
 889
 890         for (i = 0; i < nt->mw_count; i++) {
 891                 u64 val64;
 892
 893                 val = ntb_spad_read(ndev, MW0_SZ_HIGH + (i * 2));
 894                 val64 = (u64)val << 32;
 895
 896                 val = ntb_spad_read(ndev, MW0_SZ_LOW + (i * 2));
 897                 val64 |= val;
 898
 899                 dev_dbg(&pdev->dev, "Remote MW%d size = %#llx\n", i, val64);
 900
 901                 rc = ntb_set_mw(nt, i, val64);
 902                 if (rc)
 903                         goto out1;
 904         }
 905
 906         nt->link_is_up = true;
 907
 908         for (i = 0; i < nt->qp_count; i++) {
 909                 struct ntb_transport_qp *qp = &nt->qp_vec[i];
 910
 911                 ntb_transport_setup_qp_mw(nt, i);
 912
 913                 if (qp->client_ready)
 914                         schedule_delayed_work(&qp->link_work, 0);
 915         }
 916
 917         return;
 918
 919 out1:
 920         for (i = 0; i < nt->mw_count; i++)
 921                 ntb_free_mw(nt, i);
 922
 923         /* if there's an actual failure, we should just bail */
 924         if (rc < 0)
 925                 return;
 926
 927 out:
 928         if (ntb_link_is_up(ndev, NULL, NULL) == 1)
 929                 schedule_delayed_work(&nt->link_work,
 930                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 931 }
 932
 933 static void ntb_qp_link_work(struct work_struct *work)
 934 {
 935         struct ntb_transport_qp *qp = container_of(work,
 936                                                    struct ntb_transport_qp,
 937                                                    link_work.work);
 938         struct pci_dev *pdev = qp->ndev->pdev;
 939         struct ntb_transport_ctx *nt = qp->transport;
 940         int val;
 941
 942         WARN_ON(!nt->link_is_up);
 943
 944         val = ntb_spad_read(nt->ndev, QP_LINKS);
 945
 946         ntb_peer_spad_write(nt->ndev, QP_LINKS, val | BIT(qp->qp_num));
 947
 948         /* query remote spad for qp ready bits */
 949         ntb_peer_spad_read(nt->ndev, QP_LINKS);
 950         dev_dbg_ratelimited(&pdev->dev, "Remote QP link status = %x\n", val);
 951
 952         /* See if the remote side is up */
 953         if (val & BIT(qp->qp_num)) {
 954                 dev_info(&pdev->dev, "qp %d: Link Up\n", qp->qp_num);
 955                 qp->link_is_up = true;
 956                 qp->active = true;
 957
 958                 if (qp->event_handler)
 959                         qp->event_handler(qp->cb_data, qp->link_is_up);
 960
 961                 if (qp->active)
 962                         tasklet_schedule(&qp->rxc_db_work);
 963         } else if (nt->link_is_up)
 964                 schedule_delayed_work(&qp->link_work,
 965                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 966 }
 967
 968 static int ntb_transport_init_queue(struct ntb_transport_ctx *nt,
 969                                     unsigned int qp_num)
 970 {
 971         struct ntb_transport_qp *qp;
 972         phys_addr_t mw_base;
 973         resource_size_t mw_size;
 974         unsigned int num_qps_mw, tx_size;
 975         unsigned int mw_num, mw_count, qp_count;
 976         u64 qp_offset;
 977
 978         mw_count = nt->mw_count;
 979         qp_count = nt->qp_count;
 980
 981         mw_num = QP_TO_MW(nt, qp_num);
 982
 983         qp = &nt->qp_vec[qp_num];
 984         qp->qp_num = qp_num;
 985         qp->transport = nt;
 986         qp->ndev = nt->ndev;
 987         qp->client_ready = false;
 988         qp->event_handler = NULL;
 989         ntb_qp_link_down_reset(qp);
 990
 991         if (mw_num < qp_count % mw_count)
 992                 num_qps_mw = qp_count / mw_count + 1;
 993         else
 994                 num_qps_mw = qp_count / mw_count;
 995
 996         mw_base = nt->mw_vec[mw_num].phys_addr;
 997         mw_size = nt->mw_vec[mw_num].phys_size;
 998
 999         if (max_mw_size && mw_size > max_mw_size)
1000                 mw_size = max_mw_size;
1001
1002         tx_size = (unsigned int)mw_size / num_qps_mw;
1003         qp_offset = tx_size * (qp_num / mw_count);
1004
1005         qp->tx_mw = nt->mw_vec[mw_num].vbase + qp_offset;
1006         if (!qp->tx_mw)
1007                 return -EINVAL;
1008
1009         qp->tx_mw_phys = mw_base + qp_offset;
1010         if (!qp->tx_mw_phys)
1011                 return -EINVAL;
1012
1013         tx_size -= sizeof(struct ntb_rx_info);
1014         qp->rx_info = qp->tx_mw + tx_size;
1015
1016         /* Due to housekeeping, there must be atleast 2 buffs */
1017         qp->tx_max_frame = min(transport_mtu, tx_size / 2);
1018         qp->tx_max_entry = tx_size / qp->tx_max_frame;
1019
1020         if (nt->debugfs_node_dir) {
1021                 char debugfs_name[4];
1022
1023                 snprintf(debugfs_name, 4, "qp%d", qp_num);
1024                 qp->debugfs_dir = debugfs_create_dir(debugfs_name,
1025                                                      nt->debugfs_node_dir);
1026
1027                 qp->debugfs_stats = debugfs_create_file("stats", S_IRUSR,
1028                                                         qp->debugfs_dir, qp,
1029                                                         &ntb_qp_debugfs_stats);
1030         } else {
1031                 qp->debugfs_dir = NULL;
1032                 qp->debugfs_stats = NULL;
1033         }
1034
1035         INIT_DELAYED_WORK(&qp->link_work, ntb_qp_link_work);
1036         INIT_WORK(&qp->link_cleanup, ntb_qp_link_cleanup_work);
1037
1038         spin_lock_init(&qp->ntb_rx_q_lock);
1039         spin_lock_init(&qp->ntb_tx_free_q_lock);
1040
1041         INIT_LIST_HEAD(&qp->rx_post_q);
1042         INIT_LIST_HEAD(&qp->rx_pend_q);
1043         INIT_LIST_HEAD(&qp->rx_free_q);
1044         INIT_LIST_HEAD(&qp->tx_free_q);
1045
1046         tasklet_init(&qp->rxc_db_work, ntb_transport_rxc_db,
1047                      (unsigned long)qp);
1048
1049         return 0;
1050 }
1051
1052 static int ntb_transport_probe(struct ntb_client *self, struct ntb_dev *ndev)
1053 {
1054         struct ntb_transport_ctx *nt;
1055         struct ntb_transport_mw *mw;
1056         unsigned int mw_count, qp_count;
1057         u64 qp_bitmap;
1058         int node;
1059         int rc, i;
1060
1061         mw_count = ntb_mw_count(ndev);
1062         if (ntb_spad_count(ndev) < (NUM_MWS + 1 + mw_count * 2)) {
1063                 dev_err(&ndev->dev, "Not enough scratch pad registers for %s",
1064                         NTB_TRANSPORT_NAME);
1065                 return -EIO;
1066         }
1067
1068         if (ntb_db_is_unsafe(ndev))
1069                 dev_dbg(&ndev->dev,
1070                         "doorbell is unsafe, proceed anyway...\n");
1071         if (ntb_spad_is_unsafe(ndev))
1072                 dev_dbg(&ndev->dev,
1073                         "scratchpad is unsafe, proceed anyway...\n");
1074
1075         node = dev_to_node(&ndev->dev);
1076
1077         nt = kzalloc_node(sizeof(*nt), GFP_KERNEL, node);
1078         if (!nt)
1079                 return -ENOMEM;
1080
1081         nt->ndev = ndev;
1082
1083         nt->mw_count = mw_count;
1084
1085         nt->mw_vec = kzalloc_node(mw_count * sizeof(*nt->mw_vec),
1086                                   GFP_KERNEL, node);
1087         if (!nt->mw_vec) {
1088                 rc = -ENOMEM;
1089                 goto err;
1090         }
1091
1092         for (i = 0; i < mw_count; i++) {
1093                 mw = &nt->mw_vec[i];
1094
1095                 rc = ntb_mw_get_range(ndev, i, &mw->phys_addr, &mw->phys_size,
1096                                       &mw->xlat_align, &mw->xlat_align_size);
1097                 if (rc)
1098                         goto err1;
1099
1100                 mw->vbase = ioremap_wc(mw->phys_addr, mw->phys_size);
1101                 if (!mw->vbase) {
1102                         rc = -ENOMEM;
1103                         goto err1;
1104                 }
1105
1106                 mw->buff_size = 0;
1107                 mw->xlat_size = 0;
1108                 mw->virt_addr = NULL;
1109                 mw->dma_addr = 0;
1110         }
1111
1112         qp_bitmap = ntb_db_valid_mask(ndev);
1113
1114         qp_count = ilog2(qp_bitmap);
1115         if (max_num_clients && max_num_clients < qp_count)
1116                 qp_count = max_num_clients;
1117         else if (nt->mw_count < qp_count)
1118                 qp_count = nt->mw_count;
1119
1120         qp_bitmap &= BIT_ULL(qp_count) - 1;
1121
1122         nt->qp_count = qp_count;
1123         nt->qp_bitmap = qp_bitmap;
1124         nt->qp_bitmap_free = qp_bitmap;
1125
1126         nt->qp_vec = kzalloc_node(qp_count * sizeof(*nt->qp_vec),
1127                                   GFP_KERNEL, node);
1128         if (!nt->qp_vec) {
1129                 rc = -ENOMEM;
1130                 goto err1;
1131         }
1132
1133         if (nt_debugfs_dir) {
1134                 nt->debugfs_node_dir =
1135                         debugfs_create_dir(pci_name(ndev->pdev),
1136                                            nt_debugfs_dir);
1137         }
1138
1139         for (i = 0; i < qp_count; i++) {
1140                 rc = ntb_transport_init_queue(nt, i);
1141                 if (rc)
1142                         goto err2;
1143         }
1144
1145         INIT_DELAYED_WORK(&nt->link_work, ntb_transport_link_work);
1146         INIT_WORK(&nt->link_cleanup, ntb_transport_link_cleanup_work);
1147
1148         rc = ntb_set_ctx(ndev, nt, &ntb_transport_ops);
1149         if (rc)
1150                 goto err2;
1151
1152         INIT_LIST_HEAD(&nt->client_devs);
1153         rc = ntb_bus_init(nt);
1154         if (rc)
1155                 goto err3;
1156
1157         nt->link_is_up = false;
1158         ntb_link_enable(ndev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
1159         ntb_link_event(ndev);
1160
1161         return 0;
1162
1163 err3:
1164         ntb_clear_ctx(ndev);
1165 err2:
1166         kfree(nt->qp_vec);
1167 err1:
1168         while (i--) {
1169                 mw = &nt->mw_vec[i];
1170                 iounmap(mw->vbase);
1171         }
1172         kfree(nt->mw_vec);
1173 err:
1174         kfree(nt);
1175         return rc;
1176 }
1177
1178 static void ntb_transport_free(struct ntb_client *self, struct ntb_dev *ndev)
1179 {
1180         struct ntb_transport_ctx *nt = ndev->ctx;
1181         struct ntb_transport_qp *qp;
1182         u64 qp_bitmap_alloc;
1183         int i;
1184
1185         ntb_transport_link_cleanup(nt);
1186         cancel_work_sync(&nt->link_cleanup);
1187         cancel_delayed_work_sync(&nt->link_work);
1188
1189         qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
1190
1191         /* verify that all the qp's are freed */
1192         for (i = 0; i < nt->qp_count; i++) {
1193                 qp = &nt->qp_vec[i];
1194                 if (qp_bitmap_alloc & BIT_ULL(i))
1195                         ntb_transport_free_queue(qp);
1196                 debugfs_remove_recursive(qp->debugfs_dir);
1197         }
1198
1199         ntb_link_disable(ndev);
1200         ntb_clear_ctx(ndev);
1201
1202         ntb_bus_remove(nt);
1203
1204         for (i = nt->mw_count; i--; ) {
1205                 ntb_free_mw(nt, i);
1206                 iounmap(nt->mw_vec[i].vbase);
1207         }
1208
1209         kfree(nt->qp_vec);
1210         kfree(nt->mw_vec);
1211         kfree(nt);
1212 }
1213
1214 static void ntb_complete_rxc(struct ntb_transport_qp *qp)
1215 {
1216         struct ntb_queue_entry *entry;
1217         void *cb_data;
1218         unsigned int len;
1219         unsigned long irqflags;
1220
1221         spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1222
1223         while (!list_empty(&qp->rx_post_q)) {
1224                 entry = list_first_entry(&qp->rx_post_q,
1225                                          struct ntb_queue_entry, entry);
1226                 if (!(entry->flags & DESC_DONE_FLAG))
1227                         break;
1228
1229                 entry->rx_hdr->flags = 0;
1230                 iowrite32(entry->rx_index, &qp->rx_info->entry);
1231
1232                 cb_data = entry->cb_data;
1233                 len = entry->len;
1234
1235                 list_move_tail(&entry->entry, &qp->rx_free_q);
1236
1237                 spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1238
1239                 if (qp->rx_handler && qp->client_ready)
1240                         qp->rx_handler(qp, qp->cb_data, cb_data, len);
1241
1242                 spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1243         }
1244
1245         spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1246 }
1247
1248 static void ntb_rx_copy_callback(void *data,
1249                                  const struct dmaengine_result *res)
1250 {
1251         struct ntb_queue_entry *entry = data;
1252
1253         /* we need to check DMA results if we are using DMA */
1254         if (res) {
1255                 enum dmaengine_tx_result dma_err = res->result;
1256
1257                 switch (dma_err) {
1258                 case DMA_TRANS_READ_FAILED:
1259                 case DMA_TRANS_WRITE_FAILED:
1260                         entry->errors++;
1261                 case DMA_TRANS_ABORTED:
1262                 {
1263                         struct ntb_transport_qp *qp = entry->qp;
1264                         void *offset = qp->rx_buff + qp->rx_max_frame *
1265                                         qp->rx_index;
1266
1267                         ntb_memcpy_rx(entry, offset);
1268                         qp->rx_memcpy++;
1269                         return;
1270                 }
1271
1272                 case DMA_TRANS_NOERROR:
1273                 default:
1274                         break;
1275                 }
1276         }
1277
1278         entry->flags |= DESC_DONE_FLAG;
1279
1280         ntb_complete_rxc(entry->qp);
1281 }
1282
1283 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset)
1284 {
1285         void *buf = entry->buf;
1286         size_t len = entry->len;
1287
1288         memcpy(buf, offset, len);
1289
1290         /* Ensure that the data is fully copied out before clearing the flag */
1291         wmb();
1292
1293         ntb_rx_copy_callback(entry, NULL);
1294 }
1295
1296 static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset)
1297 {
1298         struct dma_async_tx_descriptor *txd;
1299         struct ntb_transport_qp *qp = entry->qp;
1300         struct dma_chan *chan = qp->rx_dma_chan;
1301         struct dma_device *device;
1302         size_t pay_off, buff_off, len;
1303         struct dmaengine_unmap_data *unmap;
1304         dma_cookie_t cookie;
1305         void *buf = entry->buf;
1306
1307         len = entry->len;
1308         device = chan->device;
1309         pay_off = (size_t)offset & ~PAGE_MASK;
1310         buff_off = (size_t)buf & ~PAGE_MASK;
1311
1312         if (!is_dma_copy_aligned(device, pay_off, buff_off, len))
1313                 goto err;
1314
1315         unmap = dmaengine_get_unmap_data(device->dev, 2, GFP_NOWAIT);
1316         if (!unmap)
1317                 goto err;
1318
1319         unmap->len = len;
1320         unmap->addr[0] = dma_map_page(device->dev, virt_to_page(offset),
1321                                       pay_off, len, DMA_TO_DEVICE);
1322         if (dma_mapping_error(device->dev, unmap->addr[0]))
1323                 goto err_get_unmap;
1324
1325         unmap->to_cnt = 1;
1326
1327         unmap->addr[1] = dma_map_page(device->dev, virt_to_page(buf),
1328                                       buff_off, len, DMA_FROM_DEVICE);
1329         if (dma_mapping_error(device->dev, unmap->addr[1]))
1330                 goto err_get_unmap;
1331
1332         unmap->from_cnt = 1;
1333
1334         txd = device->device_prep_dma_memcpy(chan, unmap->addr[1],
1335                                              unmap->addr[0], len,
1336                                              DMA_PREP_INTERRUPT);
1337         if (!txd)
1338                 goto err_get_unmap;
1339
1340         txd->callback_result = ntb_rx_copy_callback;
1341         txd->callback_param = entry;
1342         dma_set_unmap(txd, unmap);
1343
1344         cookie = dmaengine_submit(txd);
1345         if (dma_submit_error(cookie))
1346                 goto err_set_unmap;
1347
1348         dmaengine_unmap_put(unmap);
1349
1350         qp->last_cookie = cookie;
1351
1352         qp->rx_async++;
1353
1354         return 0;
1355
1356 err_set_unmap:
1357         dmaengine_unmap_put(unmap);
1358 err_get_unmap:
1359         dmaengine_unmap_put(unmap);
1360 err:
1361         return -ENXIO;
1362 }
1363
1364 static void ntb_async_rx(struct ntb_queue_entry *entry, void *offset)
1365 {
1366         struct ntb_transport_qp *qp = entry->qp;
1367         struct dma_chan *chan = qp->rx_dma_chan;
1368         int res;
1369
1370         if (!chan)
1371                 goto err;
1372
1373         if (entry->len < copy_bytes)
1374                 goto err;
1375
1376         res = ntb_async_rx_submit(entry, offset);
1377         if (res < 0)
1378                 goto err;
1379
1380         if (!entry->retries)
1381                 qp->rx_async++;
1382
1383         return;
1384
1385 err:
1386         ntb_memcpy_rx(entry, offset);
1387         qp->rx_memcpy++;
1388 }
1389
1390 static int ntb_process_rxc(struct ntb_transport_qp *qp)
1391 {
1392         struct ntb_payload_header *hdr;
1393         struct ntb_queue_entry *entry;
1394         void *offset;
1395
1396         offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
1397         hdr = offset + qp->rx_max_frame - sizeof(struct ntb_payload_header);
1398
1399         dev_dbg(&qp->ndev->pdev->dev, "qp %d: RX ver %u len %d flags %x\n",
1400                 qp->qp_num, hdr->ver, hdr->len, hdr->flags);
1401
1402         if (!(hdr->flags & DESC_DONE_FLAG)) {
1403                 dev_dbg(&qp->ndev->pdev->dev, "done flag not set\n");
1404                 qp->rx_ring_empty++;
1405                 return -EAGAIN;
1406         }
1407
1408         if (hdr->flags & LINK_DOWN_FLAG) {
1409                 dev_dbg(&qp->ndev->pdev->dev, "link down flag set\n");
1410                 ntb_qp_link_down(qp);
1411                 hdr->flags = 0;
1412                 return -EAGAIN;
1413         }
1414
1415         if (hdr->ver != (u32)qp->rx_pkts) {
1416                 dev_dbg(&qp->ndev->pdev->dev,
1417                         "version mismatch, expected %llu - got %u\n",
1418                         qp->rx_pkts, hdr->ver);
1419                 qp->rx_err_ver++;
1420                 return -EIO;
1421         }
1422
1423         entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q);
1424         if (!entry) {
1425                 dev_dbg(&qp->ndev->pdev->dev, "no receive buffer\n");
1426                 qp->rx_err_no_buf++;
1427                 return -EAGAIN;
1428         }
1429
1430         entry->rx_hdr = hdr;
1431         entry->rx_index = qp->rx_index;
1432
1433         if (hdr->len > entry->len) {
1434                 dev_dbg(&qp->ndev->pdev->dev,
1435                         "receive buffer overflow! Wanted %d got %d\n",
1436                         hdr->len, entry->len);
1437                 qp->rx_err_oflow++;
1438
1439                 entry->len = -EIO;
1440                 entry->flags |= DESC_DONE_FLAG;
1441
1442                 ntb_complete_rxc(qp);
1443         } else {
1444                 dev_dbg(&qp->ndev->pdev->dev,
1445                         "RX OK index %u ver %u size %d into buf size %d\n",
1446                         qp->rx_index, hdr->ver, hdr->len, entry->len);
1447
1448                 qp->rx_bytes += hdr->len;
1449                 qp->rx_pkts++;
1450
1451                 entry->len = hdr->len;
1452
1453                 ntb_async_rx(entry, offset);
1454         }
1455
1456         qp->rx_index++;
1457         qp->rx_index %= qp->rx_max_entry;
1458
1459         return 0;
1460 }
1461
1462 static void ntb_transport_rxc_db(unsigned long data)
1463 {
1464         struct ntb_transport_qp *qp = (void *)data;
1465         int rc, i;
1466
1467         dev_dbg(&qp->ndev->pdev->dev, "%s: doorbell %d received\n",
1468                 __func__, qp->qp_num);
1469
1470         /* Limit the number of packets processed in a single interrupt to
1471          * provide fairness to others
1472          */
1473         for (i = 0; i < qp->rx_max_entry; i++) {
1474                 rc = ntb_process_rxc(qp);
1475                 if (rc)
1476                         break;
1477         }
1478
1479         if (i && qp->rx_dma_chan)
1480                 dma_async_issue_pending(qp->rx_dma_chan);
1481
1482         if (i == qp->rx_max_entry) {
1483                 /* there is more work to do */
1484                 if (qp->active)
1485                         tasklet_schedule(&qp->rxc_db_work);
1486         } else if (ntb_db_read(qp->ndev) & BIT_ULL(qp->qp_num)) {
1487                 /* the doorbell bit is set: clear it */
1488                 ntb_db_clear(qp->ndev, BIT_ULL(qp->qp_num));
1489                 /* ntb_db_read ensures ntb_db_clear write is committed */
1490                 ntb_db_read(qp->ndev);
1491
1492                 /* an interrupt may have arrived between finishing
1493                  * ntb_process_rxc and clearing the doorbell bit:
1494                  * there might be some more work to do.
1495                  */
1496                 if (qp->active)
1497                         tasklet_schedule(&qp->rxc_db_work);
1498         }
1499 }
1500
1501 static void ntb_tx_copy_callback(void *data,
1502                                  const struct dmaengine_result *res)
1503 {
1504         struct ntb_queue_entry *entry = data;
1505         struct ntb_transport_qp *qp = entry->qp;
1506         struct ntb_payload_header __iomem *hdr = entry->tx_hdr;
1507
1508         /* we need to check DMA results if we are using DMA */
1509         if (res) {
1510                 enum dmaengine_tx_result dma_err = res->result;
1511
1512                 switch (dma_err) {
1513                 case DMA_TRANS_READ_FAILED:
1514                 case DMA_TRANS_WRITE_FAILED:
1515                         entry->errors++;
1516                 case DMA_TRANS_ABORTED:
1517                 {
1518                         void __iomem *offset =
1519                                 qp->tx_mw + qp->tx_max_frame *
1520                                 entry->tx_index;
1521
1522                         /* resubmit via CPU */
1523                         ntb_memcpy_tx(entry, offset);
1524                         qp->tx_memcpy++;
1525                         return;
1526                 }
1527
1528                 case DMA_TRANS_NOERROR:
1529                 default:
1530                         break;
1531                 }
1532         }
1533
1534         iowrite32(entry->flags | DESC_DONE_FLAG, &hdr->flags);
1535
1536         ntb_peer_db_set(qp->ndev, BIT_ULL(qp->qp_num));
1537
1538         /* The entry length can only be zero if the packet is intended to be a
1539          * "link down" or similar.  Since no payload is being sent in these
1540          * cases, there is nothing to add to the completion queue.
1541          */
1542         if (entry->len > 0) {
1543                 qp->tx_bytes += entry->len;
1544
1545                 if (qp->tx_handler)
1546                         qp->tx_handler(qp, qp->cb_data, entry->cb_data,
1547                                        entry->len);
1548         }
1549
1550         ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry, &qp->tx_free_q);
1551 }
1552
1553 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset)
1554 {
1555 #ifdef ARCH_HAS_NOCACHE_UACCESS
1556         /*
1557          * Using non-temporal mov to improve performance on non-cached
1558          * writes, even though we aren't actually copying from user space.
1559          */
1560         __copy_from_user_inatomic_nocache(offset, entry->buf, entry->len);
1561 #else
1562         memcpy_toio(offset, entry->buf, entry->len);
1563 #endif
1564
1565         /* Ensure that the data is fully copied out before setting the flags */
1566         wmb();
1567
1568         ntb_tx_copy_callback(entry, NULL);
1569 }
1570
1571 static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
1572                                struct ntb_queue_entry *entry)
1573 {
1574         struct dma_async_tx_descriptor *txd;
1575         struct dma_chan *chan = qp->tx_dma_chan;
1576         struct dma_device *device;
1577         size_t len = entry->len;
1578         void *buf = entry->buf;
1579         size_t dest_off, buff_off;
1580         struct dmaengine_unmap_data *unmap;
1581         dma_addr_t dest;
1582         dma_cookie_t cookie;
1583
1584         device = chan->device;
1585         dest = qp->tx_mw_phys + qp->tx_max_frame * entry->tx_index;
1586         buff_off = (size_t)buf & ~PAGE_MASK;
1587         dest_off = (size_t)dest & ~PAGE_MASK;
1588
1589         if (!is_dma_copy_aligned(device, buff_off, dest_off, len))
1590                 goto err;
1591
1592         unmap = dmaengine_get_unmap_data(device->dev, 1, GFP_NOWAIT);
1593         if (!unmap)
1594                 goto err;
1595
1596         unmap->len = len;
1597         unmap->addr[0] = dma_map_page(device->dev, virt_to_page(buf),
1598                                       buff_off, len, DMA_TO_DEVICE);
1599         if (dma_mapping_error(device->dev, unmap->addr[0]))
1600                 goto err_get_unmap;
1601
1602         unmap->to_cnt = 1;
1603
1604         txd = device->device_prep_dma_memcpy(chan, dest, unmap->addr[0], len,
1605                                              DMA_PREP_INTERRUPT);
1606         if (!txd)
1607                 goto err_get_unmap;
1608
1609         txd->callback_result = ntb_tx_copy_callback;
1610         txd->callback_param = entry;
1611         dma_set_unmap(txd, unmap);
1612
1613         cookie = dmaengine_submit(txd);
1614         if (dma_submit_error(cookie))
1615                 goto err_set_unmap;
1616
1617         dmaengine_unmap_put(unmap);
1618
1619         dma_async_issue_pending(chan);
1620
1621         return 0;
1622 err_set_unmap:
1623         dmaengine_unmap_put(unmap);
1624 err_get_unmap:
1625         dmaengine_unmap_put(unmap);
1626 err:
1627         return -ENXIO;
1628 }
1629
1630 static void ntb_async_tx(struct ntb_transport_qp *qp,
1631                          struct ntb_queue_entry *entry)
1632 {
1633         struct ntb_payload_header __iomem *hdr;
1634         struct dma_chan *chan = qp->tx_dma_chan;
1635         void __iomem *offset;
1636         int res;
1637
1638         entry->tx_index = qp->tx_index;
1639         offset = qp->tx_mw + qp->tx_max_frame * entry->tx_index;
1640         hdr = offset + qp->tx_max_frame - sizeof(struct ntb_payload_header);
1641         entry->tx_hdr = hdr;
1642
1643         iowrite32(entry->len, &hdr->len);
1644         iowrite32((u32)qp->tx_pkts, &hdr->ver);
1645
1646         if (!chan)
1647                 goto err;
1648
1649         if (entry->len < copy_bytes)
1650                 goto err;
1651
1652         res = ntb_async_tx_submit(qp, entry);
1653         if (res < 0)
1654                 goto err;
1655
1656         if (!entry->retries)
1657                 qp->tx_async++;
1658
1659         return;
1660
1661 err:
1662         ntb_memcpy_tx(entry, offset);
1663         qp->tx_memcpy++;
1664 }
1665
1666 static int ntb_process_tx(struct ntb_transport_qp *qp,
1667                           struct ntb_queue_entry *entry)
1668 {
1669         if (qp->tx_index == qp->remote_rx_info->entry) {
1670                 qp->tx_ring_full++;
1671                 return -EAGAIN;
1672         }
1673
1674         if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
1675                 if (qp->tx_handler)
1676                         qp->tx_handler(qp, qp->cb_data, NULL, -EIO);
1677
1678                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1679                              &qp->tx_free_q);
1680                 return 0;
1681         }
1682
1683         ntb_async_tx(qp, entry);
1684
1685         qp->tx_index++;
1686         qp->tx_index %= qp->tx_max_entry;
1687
1688         qp->tx_pkts++;
1689
1690         return 0;
1691 }
1692
1693 static void ntb_send_link_down(struct ntb_transport_qp *qp)
1694 {
1695         struct pci_dev *pdev = qp->ndev->pdev;
1696         struct ntb_queue_entry *entry;
1697         int i, rc;
1698
1699         if (!qp->link_is_up)
1700                 return;
1701
1702         dev_info(&pdev->dev, "qp %d: Send Link Down\n", qp->qp_num);
1703
1704         for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1705                 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1706                 if (entry)
1707                         break;
1708                 msleep(100);
1709         }
1710
1711         if (!entry)
1712                 return;
1713
1714         entry->cb_data = NULL;
1715         entry->buf = NULL;
1716         entry->len = 0;
1717         entry->flags = LINK_DOWN_FLAG;
1718
1719         rc = ntb_process_tx(qp, entry);
1720         if (rc)
1721                 dev_err(&pdev->dev, "ntb: QP%d unable to send linkdown msg\n",
1722                         qp->qp_num);
1723
1724         ntb_qp_link_down_reset(qp);
1725 }
1726
1727 static bool ntb_dma_filter_fn(struct dma_chan *chan, void *node)
1728 {
1729         return dev_to_node(&chan->dev->device) == (int)(unsigned long)node;
1730 }
1731
1732 /**
1733  * ntb_transport_create_queue - Create a new NTB transport layer queue
1734  * @rx_handler: receive callback function
1735  * @tx_handler: transmit callback function
1736  * @event_handler: event callback function
1737  *
1738  * Create a new NTB transport layer queue and provide the queue with a callback
1739  * routine for both transmit and receive.  The receive callback routine will be
1740  * used to pass up data when the transport has received it on the queue.   The
1741  * transmit callback routine will be called when the transport has completed the
1742  * transmission of the data on the queue and the data is ready to be freed.
1743  *
1744  * RETURNS: pointer to newly created ntb_queue, NULL on error.
1745  */
1746 struct ntb_transport_qp *
1747 ntb_transport_create_queue(void *data, struct device *client_dev,
1748                            const struct ntb_queue_handlers *handlers)
1749 {
1750         struct ntb_dev *ndev;
1751         struct pci_dev *pdev;
1752         struct ntb_transport_ctx *nt;
1753         struct ntb_queue_entry *entry;
1754         struct ntb_transport_qp *qp;
1755         u64 qp_bit;
1756         unsigned int free_queue;
1757         dma_cap_mask_t dma_mask;
1758         int node;
1759         int i;
1760
1761         ndev = dev_ntb(client_dev->parent);
1762         pdev = ndev->pdev;
1763         nt = ndev->ctx;
1764
1765         node = dev_to_node(&ndev->dev);
1766
1767         free_queue = ffs(nt->qp_bitmap_free);
1768         if (!free_queue)
1769                 goto err;
1770
1771         /* decrement free_queue to make it zero based */
1772         free_queue--;
1773
1774         qp = &nt->qp_vec[free_queue];
1775         qp_bit = BIT_ULL(qp->qp_num);
1776
1777         nt->qp_bitmap_free &= ~qp_bit;
1778
1779         qp->cb_data = data;
1780         qp->rx_handler = handlers->rx_handler;
1781         qp->tx_handler = handlers->tx_handler;
1782         qp->event_handler = handlers->event_handler;
1783
1784         dma_cap_zero(dma_mask);
1785         dma_cap_set(DMA_MEMCPY, dma_mask);
1786
1787         if (use_dma) {
1788                 qp->tx_dma_chan =
1789                         dma_request_channel(dma_mask, ntb_dma_filter_fn,
1790                                             (void *)(unsigned long)node);
1791                 if (!qp->tx_dma_chan)
1792                         dev_info(&pdev->dev, "Unable to allocate TX DMA channel\n");
1793
1794                 qp->rx_dma_chan =
1795                         dma_request_channel(dma_mask, ntb_dma_filter_fn,
1796                                             (void *)(unsigned long)node);
1797                 if (!qp->rx_dma_chan)
1798                         dev_info(&pdev->dev, "Unable to allocate RX DMA channel\n");
1799         } else {
1800                 qp->tx_dma_chan = NULL;
1801                 qp->rx_dma_chan = NULL;
1802         }
1803
1804         dev_dbg(&pdev->dev, "Using %s memcpy for TX\n",
1805                 qp->tx_dma_chan ? "DMA" : "CPU");
1806
1807         dev_dbg(&pdev->dev, "Using %s memcpy for RX\n",
1808                 qp->rx_dma_chan ? "DMA" : "CPU");
1809
1810         for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
1811                 entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
1812                 if (!entry)
1813                         goto err1;
1814
1815                 entry->qp = qp;
1816                 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
1817                              &qp->rx_free_q);
1818         }
1819         qp->rx_alloc_entry = NTB_QP_DEF_NUM_ENTRIES;
1820
1821         for (i = 0; i < qp->tx_max_entry; i++) {
1822                 entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
1823                 if (!entry)
1824                         goto err2;
1825
1826                 entry->qp = qp;
1827                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1828                              &qp->tx_free_q);
1829         }
1830
1831         ntb_db_clear(qp->ndev, qp_bit);
1832         ntb_db_clear_mask(qp->ndev, qp_bit);
1833
1834         dev_info(&pdev->dev, "NTB Transport QP %d created\n", qp->qp_num);
1835
1836         return qp;
1837
1838 err2:
1839         while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1840                 kfree(entry);
1841 err1:
1842         qp->rx_alloc_entry = 0;
1843         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
1844                 kfree(entry);
1845         if (qp->tx_dma_chan)
1846                 dma_release_channel(qp->tx_dma_chan);
1847         if (qp->rx_dma_chan)
1848                 dma_release_channel(qp->rx_dma_chan);
1849         nt->qp_bitmap_free |= qp_bit;
1850 err:
1851         return NULL;
1852 }
1853 EXPORT_SYMBOL_GPL(ntb_transport_create_queue);
1854
1855 /**
1856  * ntb_transport_free_queue - Frees NTB transport queue
1857  * @qp: NTB queue to be freed
1858  *
1859  * Frees NTB transport queue
1860  */
1861 void ntb_transport_free_queue(struct ntb_transport_qp *qp)
1862 {
1863         struct pci_dev *pdev;
1864         struct ntb_queue_entry *entry;
1865         u64 qp_bit;
1866
1867         if (!qp)
1868                 return;
1869
1870         pdev = qp->ndev->pdev;
1871
1872         qp->active = false;
1873
1874         if (qp->tx_dma_chan) {
1875                 struct dma_chan *chan = qp->tx_dma_chan;
1876                 /* Putting the dma_chan to NULL will force any new traffic to be
1877                  * processed by the CPU instead of the DAM engine
1878                  */
1879                 qp->tx_dma_chan = NULL;
1880
1881                 /* Try to be nice and wait for any queued DMA engine
1882                  * transactions to process before smashing it with a rock
1883                  */
1884                 dma_sync_wait(chan, qp->last_cookie);
1885                 dmaengine_terminate_all(chan);
1886                 dma_release_channel(chan);
1887         }
1888
1889         if (qp->rx_dma_chan) {
1890                 struct dma_chan *chan = qp->rx_dma_chan;
1891                 /* Putting the dma_chan to NULL will force any new traffic to be
1892                  * processed by the CPU instead of the DAM engine
1893                  */
1894                 qp->rx_dma_chan = NULL;
1895
1896                 /* Try to be nice and wait for any queued DMA engine
1897                  * transactions to process before smashing it with a rock
1898                  */
1899                 dma_sync_wait(chan, qp->last_cookie);
1900                 dmaengine_terminate_all(chan);
1901                 dma_release_channel(chan);
1902         }
1903
1904         qp_bit = BIT_ULL(qp->qp_num);
1905
1906         ntb_db_set_mask(qp->ndev, qp_bit);
1907         tasklet_kill(&qp->rxc_db_work);
1908
1909         cancel_delayed_work_sync(&qp->link_work);
1910
1911         qp->cb_data = NULL;
1912         qp->rx_handler = NULL;
1913         qp->tx_handler = NULL;
1914         qp->event_handler = NULL;
1915
1916         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
1917                 kfree(entry);
1918
1919         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q))) {
1920                 dev_warn(&pdev->dev, "Freeing item from non-empty rx_pend_q\n");
1921                 kfree(entry);
1922         }
1923
1924         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q))) {
1925                 dev_warn(&pdev->dev, "Freeing item from non-empty rx_post_q\n");
1926                 kfree(entry);
1927         }
1928
1929         while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1930                 kfree(entry);
1931
1932         qp->transport->qp_bitmap_free |= qp_bit;
1933
1934         dev_info(&pdev->dev, "NTB Transport QP %d freed\n", qp->qp_num);
1935 }
1936 EXPORT_SYMBOL_GPL(ntb_transport_free_queue);
1937
1938 /**
1939  * ntb_transport_rx_remove - Dequeues enqueued rx packet
1940  * @qp: NTB queue to be freed
1941  * @len: pointer to variable to write enqueued buffers length
1942  *
1943  * Dequeues unused buffers from receive queue.  Should only be used during
1944  * shutdown of qp.
1945  *
1946  * RETURNS: NULL error value on error, or void* for success.
1947  */
1948 void *ntb_transport_rx_remove(struct ntb_transport_qp *qp, unsigned int *len)
1949 {
1950         struct ntb_queue_entry *entry;
1951         void *buf;
1952
1953         if (!qp || qp->client_ready)
1954                 return NULL;
1955
1956         entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q);
1957         if (!entry)
1958                 return NULL;
1959
1960         buf = entry->cb_data;
1961         *len = entry->len;
1962
1963         ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_free_q);
1964
1965         return buf;
1966 }
1967 EXPORT_SYMBOL_GPL(ntb_transport_rx_remove);
1968
1969 /**
1970  * ntb_transport_rx_enqueue - Enqueue a new NTB queue entry
1971  * @qp: NTB transport layer queue the entry is to be enqueued on
1972  * @cb: per buffer pointer for callback function to use
1973  * @data: pointer to data buffer that incoming packets will be copied into
1974  * @len: length of the data buffer
1975  *
1976  * Enqueue a new receive buffer onto the transport queue into which a NTB
1977  * payload can be received into.
1978  *
1979  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
1980  */
1981 int ntb_transport_rx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
1982                              unsigned int len)
1983 {
1984         struct ntb_queue_entry *entry;
1985
1986         if (!qp)
1987                 return -EINVAL;
1988
1989         entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q);
1990         if (!entry)
1991                 return -ENOMEM;
1992
1993         entry->cb_data = cb;
1994         entry->buf = data;
1995         entry->len = len;
1996         entry->flags = 0;
1997         entry->retries = 0;
1998         entry->errors = 0;
1999         entry->rx_index = 0;
2000
2001         ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_pend_q);
2002
2003         if (qp->active)
2004                 tasklet_schedule(&qp->rxc_db_work);
2005
2006         return 0;
2007 }
2008 EXPORT_SYMBOL_GPL(ntb_transport_rx_enqueue);
2009
2010 /**
2011  * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
2012  * @qp: NTB transport layer queue the entry is to be enqueued on
2013  * @cb: per buffer pointer for callback function to use
2014  * @data: pointer to data buffer that will be sent
2015  * @len: length of the data buffer
2016  *
2017  * Enqueue a new transmit buffer onto the transport queue from which a NTB
2018  * payload will be transmitted.  This assumes that a lock is being held to
2019  * serialize access to the qp.
2020  *
2021  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2022  */
2023 int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
2024                              unsigned int len)
2025 {
2026         struct ntb_queue_entry *entry;
2027         int rc;
2028
2029         if (!qp || !qp->link_is_up || !len)
2030                 return -EINVAL;
2031
2032         entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
2033         if (!entry) {
2034                 qp->tx_err_no_buf++;
2035                 return -EBUSY;
2036         }
2037
2038         entry->cb_data = cb;
2039         entry->buf = data;
2040         entry->len = len;
2041         entry->flags = 0;
2042         entry->errors = 0;
2043         entry->retries = 0;
2044         entry->tx_index = 0;
2045
2046         rc = ntb_process_tx(qp, entry);
2047         if (rc)
2048                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
2049                              &qp->tx_free_q);
2050
2051         return rc;
2052 }
2053 EXPORT_SYMBOL_GPL(ntb_transport_tx_enqueue);
2054
2055 /**
2056  * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
2057  * @qp: NTB transport layer queue to be enabled
2058  *
2059  * Notify NTB transport layer of client readiness to use queue
2060  */
2061 void ntb_transport_link_up(struct ntb_transport_qp *qp)
2062 {
2063         if (!qp)
2064                 return;
2065
2066         qp->client_ready = true;
2067
2068         if (qp->transport->link_is_up)
2069                 schedule_delayed_work(&qp->link_work, 0);
2070 }
2071 EXPORT_SYMBOL_GPL(ntb_transport_link_up);
2072
2073 /**
2074  * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
2075  * @qp: NTB transport layer queue to be disabled
2076  *
2077  * Notify NTB transport layer of client's desire to no longer receive data on
2078  * transport queue specified.  It is the client's responsibility to ensure all
2079  * entries on queue are purged or otherwise handled appropriately.
2080  */
2081 void ntb_transport_link_down(struct ntb_transport_qp *qp)
2082 {
2083         int val;
2084
2085         if (!qp)
2086                 return;
2087
2088         qp->client_ready = false;
2089
2090         val = ntb_spad_read(qp->ndev, QP_LINKS);
2091
2092         ntb_peer_spad_write(qp->ndev, QP_LINKS,
2093                             val & ~BIT(qp->qp_num));
2094
2095         if (qp->link_is_up)
2096                 ntb_send_link_down(qp);
2097         else
2098                 cancel_delayed_work_sync(&qp->link_work);
2099 }
2100 EXPORT_SYMBOL_GPL(ntb_transport_link_down);
2101
2102 /**
2103  * ntb_transport_link_query - Query transport link state
2104  * @qp: NTB transport layer queue to be queried
2105  *
2106  * Query connectivity to the remote system of the NTB transport queue
2107  *
2108  * RETURNS: true for link up or false for link down
2109  */
2110 bool ntb_transport_link_query(struct ntb_transport_qp *qp)
2111 {
2112         if (!qp)
2113                 return false;
2114
2115         return qp->link_is_up;
2116 }
2117 EXPORT_SYMBOL_GPL(ntb_transport_link_query);
2118
2119 /**
2120  * ntb_transport_qp_num - Query the qp number
2121  * @qp: NTB transport layer queue to be queried
2122  *
2123  * Query qp number of the NTB transport queue
2124  *
2125  * RETURNS: a zero based number specifying the qp number
2126  */
2127 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
2128 {
2129         if (!qp)
2130                 return 0;
2131
2132         return qp->qp_num;
2133 }
2134 EXPORT_SYMBOL_GPL(ntb_transport_qp_num);
2135
2136 /**
2137  * ntb_transport_max_size - Query the max payload size of a qp
2138  * @qp: NTB transport layer queue to be queried
2139  *
2140  * Query the maximum payload size permissible on the given qp
2141  *
2142  * RETURNS: the max payload size of a qp
2143  */
2144 unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp)
2145 {
2146         unsigned int max_size;
2147         unsigned int copy_align;
2148         struct dma_chan *rx_chan, *tx_chan;
2149
2150         if (!qp)
2151                 return 0;
2152
2153         rx_chan = qp->rx_dma_chan;
2154         tx_chan = qp->tx_dma_chan;
2155
2156         copy_align = max(rx_chan ? rx_chan->device->copy_align : 0,
2157                          tx_chan ? tx_chan->device->copy_align : 0);
2158
2159         /* If DMA engine usage is possible, try to find the max size for that */
2160         max_size = qp->tx_max_frame - sizeof(struct ntb_payload_header);
2161         max_size = round_down(max_size, 1 << copy_align);
2162
2163         return max_size;
2164 }
2165 EXPORT_SYMBOL_GPL(ntb_transport_max_size);
2166
2167 unsigned int ntb_transport_tx_free_entry(struct ntb_transport_qp *qp)
2168 {
2169         unsigned int head = qp->tx_index;
2170         unsigned int tail = qp->remote_rx_info->entry;
2171
2172         return tail > head ? tail - head : qp->tx_max_entry + tail - head;
2173 }
2174 EXPORT_SYMBOL_GPL(ntb_transport_tx_free_entry);
2175
2176 static void ntb_transport_doorbell_callback(void *data, int vector)
2177 {
2178         struct ntb_transport_ctx *nt = data;
2179         struct ntb_transport_qp *qp;
2180         u64 db_bits;
2181         unsigned int qp_num;
2182
2183         db_bits = (nt->qp_bitmap & ~nt->qp_bitmap_free &
2184                    ntb_db_vector_mask(nt->ndev, vector));
2185
2186         while (db_bits) {
2187                 qp_num = __ffs(db_bits);
2188                 qp = &nt->qp_vec[qp_num];
2189
2190                 if (qp->active)
2191                         tasklet_schedule(&qp->rxc_db_work);
2192
2193                 db_bits &= ~BIT_ULL(qp_num);
2194         }
2195 }
2196
2197 static const struct ntb_ctx_ops ntb_transport_ops = {
2198         .link_event = ntb_transport_event_callback,
2199         .db_event = ntb_transport_doorbell_callback,
2200 };
2201
2202 static struct ntb_client ntb_transport_client = {
2203         .ops = {
2204                 .probe = ntb_transport_probe,
2205                 .remove = ntb_transport_free,
2206         },
2207 };
2208
2209 static int __init ntb_transport_init(void)
2210 {
2211         int rc;
2212
2213         pr_info("%s, version %s\n", NTB_TRANSPORT_DESC, NTB_TRANSPORT_VER);
2214
2215         if (debugfs_initialized())
2216                 nt_debugfs_dir = debugfs_create_dir(KBUILD_MODNAME, NULL);
2217
2218         rc = bus_register(&ntb_transport_bus);
2219         if (rc)
2220                 goto err_bus;
2221
2222         rc = ntb_register_client(&ntb_transport_client);
2223         if (rc)
2224                 goto err_client;
2225
2226         return 0;
2227
2228 err_client:
2229         bus_unregister(&ntb_transport_bus);
2230 err_bus:
2231         debugfs_remove_recursive(nt_debugfs_dir);
2232         return rc;
2233 }
2234 module_init(ntb_transport_init);
2235
2236 static void __exit ntb_transport_exit(void)
2237 {
2238         ntb_unregister_client(&ntb_transport_client);
2239         bus_unregister(&ntb_transport_bus);
2240         debugfs_remove_recursive(nt_debugfs_dir);
2241 }
2242 module_exit(ntb_transport_exit);