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 #define NTB_TRANSPORT_MIN_SPADS (MW0_SZ_HIGH + 2)
  70
  71 MODULE_DESCRIPTION(NTB_TRANSPORT_DESC);
  72 MODULE_VERSION(NTB_TRANSPORT_VER);
  73 MODULE_LICENSE("Dual BSD/GPL");
  74 MODULE_AUTHOR("Intel Corporation");
  75
  76 static unsigned long max_mw_size;
  77 module_param(max_mw_size, ulong, 0644);
  78 MODULE_PARM_DESC(max_mw_size, "Limit size of large memory windows");
  79
  80 static unsigned int transport_mtu = 0x10000;
  81 module_param(transport_mtu, uint, 0644);
  82 MODULE_PARM_DESC(transport_mtu, "Maximum size of NTB transport packets");
  83
  84 static unsigned char max_num_clients;
  85 module_param(max_num_clients, byte, 0644);
  86 MODULE_PARM_DESC(max_num_clients, "Maximum number of NTB transport clients");
  87
  88 static unsigned int copy_bytes = 1024;
  89 module_param(copy_bytes, uint, 0644);
  90 MODULE_PARM_DESC(copy_bytes, "Threshold under which NTB will use the CPU to copy instead of DMA");
  91
  92 static bool use_dma;
  93 module_param(use_dma, bool, 0644);
  94 MODULE_PARM_DESC(use_dma, "Use DMA engine to perform large data copy");
  95
  96 static struct dentry *nt_debugfs_dir;
  97
  98 struct ntb_queue_entry {
  99         /* ntb_queue list reference */
 100         struct list_head entry;
 101         /* pointers to data to be transferred */
 102         void *cb_data;
 103         void *buf;
 104         unsigned int len;
 105         unsigned int flags;
 106         int retries;
 107         int errors;
 108         unsigned int tx_index;
 109         unsigned int rx_index;
 110
 111         struct ntb_transport_qp *qp;
 112         union {
 113                 struct ntb_payload_header __iomem *tx_hdr;
 114                 struct ntb_payload_header *rx_hdr;
 115         };
 116 };
 117
 118 struct ntb_rx_info {
 119         unsigned int entry;
 120 };
 121
 122 struct ntb_transport_qp {
 123         struct ntb_transport_ctx *transport;
 124         struct ntb_dev *ndev;
 125         void *cb_data;
 126         struct dma_chan *tx_dma_chan;
 127         struct dma_chan *rx_dma_chan;
 128
 129         bool client_ready;
 130         bool link_is_up;
 131         bool active;
 132
 133         u8 qp_num;      /* Only 64 QP's are allowed.  0-63 */
 134         u64 qp_bit;
 135
 136         struct ntb_rx_info __iomem *rx_info;
 137         struct ntb_rx_info *remote_rx_info;
 138
 139         void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data,
 140                            void *data, int len);
 141         struct list_head tx_free_q;
 142         spinlock_t ntb_tx_free_q_lock;
 143         void __iomem *tx_mw;
 144         dma_addr_t tx_mw_phys;
 145         unsigned int tx_index;
 146         unsigned int tx_max_entry;
 147         unsigned int tx_max_frame;
 148
 149         void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
 150                            void *data, int len);
 151         struct list_head rx_post_q;
 152         struct list_head rx_pend_q;
 153         struct list_head rx_free_q;
 154         /* ntb_rx_q_lock: synchronize access to rx_XXXX_q */
 155         spinlock_t ntb_rx_q_lock;
 156         void *rx_buff;
 157         unsigned int rx_index;
 158         unsigned int rx_max_entry;
 159         unsigned int rx_max_frame;
 160         unsigned int rx_alloc_entry;
 161         dma_cookie_t last_cookie;
 162         struct tasklet_struct rxc_db_work;
 163
 164         void (*event_handler)(void *data, int status);
 165         struct delayed_work link_work;
 166         struct work_struct link_cleanup;
 167
 168         struct dentry *debugfs_dir;
 169         struct dentry *debugfs_stats;
 170
 171         /* Stats */
 172         u64 rx_bytes;
 173         u64 rx_pkts;
 174         u64 rx_ring_empty;
 175         u64 rx_err_no_buf;
 176         u64 rx_err_oflow;
 177         u64 rx_err_ver;
 178         u64 rx_memcpy;
 179         u64 rx_async;
 180         u64 tx_bytes;
 181         u64 tx_pkts;
 182         u64 tx_ring_full;
 183         u64 tx_err_no_buf;
 184         u64 tx_memcpy;
 185         u64 tx_async;
 186 };
 187
 188 struct ntb_transport_mw {
 189         phys_addr_t phys_addr;
 190         resource_size_t phys_size;
 191         resource_size_t xlat_align;
 192         resource_size_t xlat_align_size;
 193         void __iomem *vbase;
 194         size_t xlat_size;
 195         size_t buff_size;
 196         void *virt_addr;
 197         dma_addr_t dma_addr;
 198 };
 199
 200 struct ntb_transport_client_dev {
 201         struct list_head entry;
 202         struct ntb_transport_ctx *nt;
 203         struct device dev;
 204 };
 205
 206 struct ntb_transport_ctx {
 207         struct list_head entry;
 208         struct list_head client_devs;
 209
 210         struct ntb_dev *ndev;
 211
 212         struct ntb_transport_mw *mw_vec;
 213         struct ntb_transport_qp *qp_vec;
 214         unsigned int mw_count;
 215         unsigned int qp_count;
 216         u64 qp_bitmap;
 217         u64 qp_bitmap_free;
 218
 219         bool link_is_up;
 220         struct delayed_work link_work;
 221         struct work_struct link_cleanup;
 222
 223         struct dentry *debugfs_node_dir;
 224 };
 225
 226 enum {
 227         DESC_DONE_FLAG = BIT(0),
 228         LINK_DOWN_FLAG = BIT(1),
 229 };
 230
 231 struct ntb_payload_header {
 232         unsigned int ver;
 233         unsigned int len;
 234         unsigned int flags;
 235 };
 236
 237 enum {
 238         VERSION = 0,
 239         QP_LINKS,
 240         NUM_QPS,
 241         NUM_MWS,
 242         MW0_SZ_HIGH,
 243         MW0_SZ_LOW,
 244 };
 245
 246 #define dev_client_dev(__dev) \
 247         container_of((__dev), struct ntb_transport_client_dev, dev)
 248
 249 #define drv_client(__drv) \
 250         container_of((__drv), struct ntb_transport_client, driver)
 251
 252 #define QP_TO_MW(nt, qp)        ((qp) % nt->mw_count)
 253 #define NTB_QP_DEF_NUM_ENTRIES  100
 254 #define NTB_LINK_DOWN_TIMEOUT   10
 255
 256 static void ntb_transport_rxc_db(unsigned long data);
 257 static const struct ntb_ctx_ops ntb_transport_ops;
 258 static struct ntb_client ntb_transport_client;
 259 static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
 260                                struct ntb_queue_entry *entry);
 261 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset);
 262 static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset);
 263 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset);
 264
 265
 266 static int ntb_transport_bus_match(struct device *dev,
 267                                    struct device_driver *drv)
 268 {
 269         return !strncmp(dev_name(dev), drv->name, strlen(drv->name));
 270 }
 271
 272 static int ntb_transport_bus_probe(struct device *dev)
 273 {
 274         const struct ntb_transport_client *client;
 275         int rc = -EINVAL;
 276
 277         get_device(dev);
 278
 279         client = drv_client(dev->driver);
 280         rc = client->probe(dev);
 281         if (rc)
 282                 put_device(dev);
 283
 284         return rc;
 285 }
 286
 287 static int ntb_transport_bus_remove(struct device *dev)
 288 {
 289         const struct ntb_transport_client *client;
 290
 291         client = drv_client(dev->driver);
 292         client->remove(dev);
 293
 294         put_device(dev);
 295
 296         return 0;
 297 }
 298
 299 static struct bus_type ntb_transport_bus = {
 300         .name = "ntb_transport",
 301         .match = ntb_transport_bus_match,
 302         .probe = ntb_transport_bus_probe,
 303         .remove = ntb_transport_bus_remove,
 304 };
 305
 306 static LIST_HEAD(ntb_transport_list);
 307
 308 static int ntb_bus_init(struct ntb_transport_ctx *nt)
 309 {
 310         list_add_tail(&nt->entry, &ntb_transport_list);
 311         return 0;
 312 }
 313
 314 static void ntb_bus_remove(struct ntb_transport_ctx *nt)
 315 {
 316         struct ntb_transport_client_dev *client_dev, *cd;
 317
 318         list_for_each_entry_safe(client_dev, cd, &nt->client_devs, entry) {
 319                 dev_err(client_dev->dev.parent, "%s still attached to bus, removing\n",
 320                         dev_name(&client_dev->dev));
 321                 list_del(&client_dev->entry);
 322                 device_unregister(&client_dev->dev);
 323         }
 324
 325         list_del(&nt->entry);
 326 }
 327
 328 static void ntb_transport_client_release(struct device *dev)
 329 {
 330         struct ntb_transport_client_dev *client_dev;
 331
 332         client_dev = dev_client_dev(dev);
 333         kfree(client_dev);
 334 }
 335
 336 /**
 337  * ntb_transport_unregister_client_dev - Unregister NTB client device
 338  * @device_name: Name of NTB client device
 339  *
 340  * Unregister an NTB client device with the NTB transport layer
 341  */
 342 void ntb_transport_unregister_client_dev(char *device_name)
 343 {
 344         struct ntb_transport_client_dev *client, *cd;
 345         struct ntb_transport_ctx *nt;
 346
 347         list_for_each_entry(nt, &ntb_transport_list, entry)
 348                 list_for_each_entry_safe(client, cd, &nt->client_devs, entry)
 349                         if (!strncmp(dev_name(&client->dev), device_name,
 350                                      strlen(device_name))) {
 351                                 list_del(&client->entry);
 352                                 device_unregister(&client->dev);
 353                         }
 354 }
 355 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client_dev);
 356
 357 /**
 358  * ntb_transport_register_client_dev - Register NTB client device
 359  * @device_name: Name of NTB client device
 360  *
 361  * Register an NTB client device with the NTB transport layer
 362  */
 363 int ntb_transport_register_client_dev(char *device_name)
 364 {
 365         struct ntb_transport_client_dev *client_dev;
 366         struct ntb_transport_ctx *nt;
 367         int node;
 368         int rc, i = 0;
 369
 370         if (list_empty(&ntb_transport_list))
 371                 return -ENODEV;
 372
 373         list_for_each_entry(nt, &ntb_transport_list, entry) {
 374                 struct device *dev;
 375
 376                 node = dev_to_node(&nt->ndev->dev);
 377
 378                 client_dev = kzalloc_node(sizeof(*client_dev),
 379                                           GFP_KERNEL, node);
 380                 if (!client_dev) {
 381                         rc = -ENOMEM;
 382                         goto err;
 383                 }
 384
 385                 dev = &client_dev->dev;
 386
 387                 /* setup and register client devices */
 388                 dev_set_name(dev, "%s%d", device_name, i);
 389                 dev->bus = &ntb_transport_bus;
 390                 dev->release = ntb_transport_client_release;
 391                 dev->parent = &nt->ndev->dev;
 392
 393                 rc = device_register(dev);
 394                 if (rc) {
 395                         kfree(client_dev);
 396                         goto err;
 397                 }
 398
 399                 list_add_tail(&client_dev->entry, &nt->client_devs);
 400                 i++;
 401         }
 402
 403         return 0;
 404
 405 err:
 406         ntb_transport_unregister_client_dev(device_name);
 407
 408         return rc;
 409 }
 410 EXPORT_SYMBOL_GPL(ntb_transport_register_client_dev);
 411
 412 /**
 413  * ntb_transport_register_client - Register NTB client driver
 414  * @drv: NTB client driver to be registered
 415  *
 416  * Register an NTB client driver with the NTB transport layer
 417  *
 418  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 419  */
 420 int ntb_transport_register_client(struct ntb_transport_client *drv)
 421 {
 422         drv->driver.bus = &ntb_transport_bus;
 423
 424         if (list_empty(&ntb_transport_list))
 425                 return -ENODEV;
 426
 427         return driver_register(&drv->driver);
 428 }
 429 EXPORT_SYMBOL_GPL(ntb_transport_register_client);
 430
 431 /**
 432  * ntb_transport_unregister_client - Unregister NTB client driver
 433  * @drv: NTB client driver to be unregistered
 434  *
 435  * Unregister an NTB client driver with the NTB transport layer
 436  *
 437  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 438  */
 439 void ntb_transport_unregister_client(struct ntb_transport_client *drv)
 440 {
 441         driver_unregister(&drv->driver);
 442 }
 443 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client);
 444
 445 static ssize_t debugfs_read(struct file *filp, char __user *ubuf, size_t count,
 446                             loff_t *offp)
 447 {
 448         struct ntb_transport_qp *qp;
 449         char *buf;
 450         ssize_t ret, out_offset, out_count;
 451
 452         qp = filp->private_data;
 453
 454         if (!qp || !qp->link_is_up)
 455                 return 0;
 456
 457         out_count = 1000;
 458
 459         buf = kmalloc(out_count, GFP_KERNEL);
 460         if (!buf)
 461                 return -ENOMEM;
 462
 463         out_offset = 0;
 464         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 465                                "\nNTB QP stats:\n\n");
 466         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 467                                "rx_bytes - \t%llu\n", qp->rx_bytes);
 468         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 469                                "rx_pkts - \t%llu\n", qp->rx_pkts);
 470         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 471                                "rx_memcpy - \t%llu\n", qp->rx_memcpy);
 472         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 473                                "rx_async - \t%llu\n", qp->rx_async);
 474         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 475                                "rx_ring_empty - %llu\n", qp->rx_ring_empty);
 476         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 477                                "rx_err_no_buf - %llu\n", qp->rx_err_no_buf);
 478         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 479                                "rx_err_oflow - \t%llu\n", qp->rx_err_oflow);
 480         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 481                                "rx_err_ver - \t%llu\n", qp->rx_err_ver);
 482         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 483                                "rx_buff - \t0x%p\n", qp->rx_buff);
 484         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 485                                "rx_index - \t%u\n", qp->rx_index);
 486         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 487                                "rx_max_entry - \t%u\n", qp->rx_max_entry);
 488         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 489                                "rx_alloc_entry - \t%u\n\n", qp->rx_alloc_entry);
 490
 491         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 492                                "tx_bytes - \t%llu\n", qp->tx_bytes);
 493         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 494                                "tx_pkts - \t%llu\n", qp->tx_pkts);
 495         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 496                                "tx_memcpy - \t%llu\n", qp->tx_memcpy);
 497         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 498                                "tx_async - \t%llu\n", qp->tx_async);
 499         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 500                                "tx_ring_full - \t%llu\n", qp->tx_ring_full);
 501         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 502                                "tx_err_no_buf - %llu\n", qp->tx_err_no_buf);
 503         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 504                                "tx_mw - \t0x%p\n", qp->tx_mw);
 505         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 506                                "tx_index (H) - \t%u\n", qp->tx_index);
 507         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 508                                "RRI (T) - \t%u\n",
 509                                qp->remote_rx_info->entry);
 510         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 511                                "tx_max_entry - \t%u\n", qp->tx_max_entry);
 512         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 513                                "free tx - \t%u\n",
 514                                ntb_transport_tx_free_entry(qp));
 515
 516         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 517                                "\n");
 518         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 519                                "Using TX DMA - \t%s\n",
 520                                qp->tx_dma_chan ? "Yes" : "No");
 521         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 522                                "Using RX DMA - \t%s\n",
 523                                qp->rx_dma_chan ? "Yes" : "No");
 524         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 525                                "QP Link - \t%s\n",
 526                                qp->link_is_up ? "Up" : "Down");
 527         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 528                                "\n");
 529
 530         if (out_offset > out_count)
 531                 out_offset = out_count;
 532
 533         ret = simple_read_from_buffer(ubuf, count, offp, buf, out_offset);
 534         kfree(buf);
 535         return ret;
 536 }
 537
 538 static const struct file_operations ntb_qp_debugfs_stats = {
 539         .owner = THIS_MODULE,
 540         .open = simple_open,
 541         .read = debugfs_read,
 542 };
 543
 544 static void ntb_list_add(spinlock_t *lock, struct list_head *entry,
 545                          struct list_head *list)
 546 {
 547         unsigned long flags;
 548
 549         spin_lock_irqsave(lock, flags);
 550         list_add_tail(entry, list);
 551         spin_unlock_irqrestore(lock, flags);
 552 }
 553
 554 static struct ntb_queue_entry *ntb_list_rm(spinlock_t *lock,
 555                                            struct list_head *list)
 556 {
 557         struct ntb_queue_entry *entry;
 558         unsigned long flags;
 559
 560         spin_lock_irqsave(lock, flags);
 561         if (list_empty(list)) {
 562                 entry = NULL;
 563                 goto out;
 564         }
 565         entry = list_first_entry(list, struct ntb_queue_entry, entry);
 566         list_del(&entry->entry);
 567
 568 out:
 569         spin_unlock_irqrestore(lock, flags);
 570
 571         return entry;
 572 }
 573
 574 static struct ntb_queue_entry *ntb_list_mv(spinlock_t *lock,
 575                                            struct list_head *list,
 576                                            struct list_head *to_list)
 577 {
 578         struct ntb_queue_entry *entry;
 579         unsigned long flags;
 580
 581         spin_lock_irqsave(lock, flags);
 582
 583         if (list_empty(list)) {
 584                 entry = NULL;
 585         } else {
 586                 entry = list_first_entry(list, struct ntb_queue_entry, entry);
 587                 list_move_tail(&entry->entry, to_list);
 588         }
 589
 590         spin_unlock_irqrestore(lock, flags);
 591
 592         return entry;
 593 }
 594
 595 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
 596                                      unsigned int qp_num)
 597 {
 598         struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
 599         struct ntb_transport_mw *mw;
 600         struct ntb_dev *ndev = nt->ndev;
 601         struct ntb_queue_entry *entry;
 602         unsigned int rx_size, num_qps_mw;
 603         unsigned int mw_num, mw_count, qp_count;
 604         unsigned int i;
 605         int node;
 606
 607         mw_count = nt->mw_count;
 608         qp_count = nt->qp_count;
 609
 610         mw_num = QP_TO_MW(nt, qp_num);
 611         mw = &nt->mw_vec[mw_num];
 612
 613         if (!mw->virt_addr)
 614                 return -ENOMEM;
 615
 616         if (mw_num < qp_count % mw_count)
 617                 num_qps_mw = qp_count / mw_count + 1;
 618         else
 619                 num_qps_mw = qp_count / mw_count;
 620
 621         rx_size = (unsigned int)mw->xlat_size / num_qps_mw;
 622         qp->rx_buff = mw->virt_addr + rx_size * (qp_num / mw_count);
 623         rx_size -= sizeof(struct ntb_rx_info);
 624
 625         qp->remote_rx_info = qp->rx_buff + rx_size;
 626
 627         /* Due to housekeeping, there must be atleast 2 buffs */
 628         qp->rx_max_frame = min(transport_mtu, rx_size / 2);
 629         qp->rx_max_entry = rx_size / qp->rx_max_frame;
 630         qp->rx_index = 0;
 631
 632         /*
 633          * Checking to see if we have more entries than the default.
 634          * We should add additional entries if that is the case so we
 635          * can be in sync with the transport frames.
 636          */
 637         node = dev_to_node(&ndev->dev);
 638         for (i = qp->rx_alloc_entry; i < qp->rx_max_entry; i++) {
 639                 entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
 640                 if (!entry)
 641                         return -ENOMEM;
 642
 643                 entry->qp = qp;
 644                 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
 645                              &qp->rx_free_q);
 646                 qp->rx_alloc_entry++;
 647         }
 648
 649         qp->remote_rx_info->entry = qp->rx_max_entry - 1;
 650
 651         /* setup the hdr offsets with 0's */
 652         for (i = 0; i < qp->rx_max_entry; i++) {
 653                 void *offset = (qp->rx_buff + qp->rx_max_frame * (i + 1) -
 654                                 sizeof(struct ntb_payload_header));
 655                 memset(offset, 0, sizeof(struct ntb_payload_header));
 656         }
 657
 658         qp->rx_pkts = 0;
 659         qp->tx_pkts = 0;
 660         qp->tx_index = 0;
 661
 662         return 0;
 663 }
 664
 665 static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
 666 {
 667         struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
 668         struct pci_dev *pdev = nt->ndev->pdev;
 669
 670         if (!mw->virt_addr)
 671                 return;
 672
 673         ntb_mw_clear_trans(nt->ndev, num_mw);
 674         dma_free_coherent(&pdev->dev, mw->buff_size,
 675                           mw->virt_addr, mw->dma_addr);
 676         mw->xlat_size = 0;
 677         mw->buff_size = 0;
 678         mw->virt_addr = NULL;
 679 }
 680
 681 static int ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw,
 682                       resource_size_t size)
 683 {
 684         struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
 685         struct pci_dev *pdev = nt->ndev->pdev;
 686         size_t xlat_size, buff_size;
 687         int rc;
 688
 689         if (!size)
 690                 return -EINVAL;
 691
 692         xlat_size = round_up(size, mw->xlat_align_size);
 693         buff_size = round_up(size, mw->xlat_align);
 694
 695         /* No need to re-setup */
 696         if (mw->xlat_size == xlat_size)
 697                 return 0;
 698
 699         if (mw->buff_size)
 700                 ntb_free_mw(nt, num_mw);
 701
 702         /* Alloc memory for receiving data.  Must be aligned */
 703         mw->xlat_size = xlat_size;
 704         mw->buff_size = buff_size;
 705
 706         mw->virt_addr = dma_alloc_coherent(&pdev->dev, buff_size,
 707                                            &mw->dma_addr, GFP_KERNEL);
 708         if (!mw->virt_addr) {
 709                 mw->xlat_size = 0;
 710                 mw->buff_size = 0;
 711                 dev_err(&pdev->dev, "Unable to alloc MW buff of size %zu\n",
 712                         buff_size);
 713                 return -ENOMEM;
 714         }
 715
 716         /*
 717          * we must ensure that the memory address allocated is BAR size
 718          * aligned in order for the XLAT register to take the value. This
 719          * is a requirement of the hardware. It is recommended to setup CMA
 720          * for BAR sizes equal or greater than 4MB.
 721          */
 722         if (!IS_ALIGNED(mw->dma_addr, mw->xlat_align)) {
 723                 dev_err(&pdev->dev, "DMA memory %pad is not aligned\n",
 724                         &mw->dma_addr);
 725                 ntb_free_mw(nt, num_mw);
 726                 return -ENOMEM;
 727         }
 728
 729         /* Notify HW the memory location of the receive buffer */
 730         rc = ntb_mw_set_trans(nt->ndev, num_mw, mw->dma_addr, mw->xlat_size);
 731         if (rc) {
 732                 dev_err(&pdev->dev, "Unable to set mw%d translation", num_mw);
 733                 ntb_free_mw(nt, num_mw);
 734                 return -EIO;
 735         }
 736
 737         return 0;
 738 }
 739
 740 static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
 741 {
 742         qp->link_is_up = false;
 743         qp->active = false;
 744
 745         qp->tx_index = 0;
 746         qp->rx_index = 0;
 747         qp->rx_bytes = 0;
 748         qp->rx_pkts = 0;
 749         qp->rx_ring_empty = 0;
 750         qp->rx_err_no_buf = 0;
 751         qp->rx_err_oflow = 0;
 752         qp->rx_err_ver = 0;
 753         qp->rx_memcpy = 0;
 754         qp->rx_async = 0;
 755         qp->tx_bytes = 0;
 756         qp->tx_pkts = 0;
 757         qp->tx_ring_full = 0;
 758         qp->tx_err_no_buf = 0;
 759         qp->tx_memcpy = 0;
 760         qp->tx_async = 0;
 761 }
 762
 763 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
 764 {
 765         struct ntb_transport_ctx *nt = qp->transport;
 766         struct pci_dev *pdev = nt->ndev->pdev;
 767
 768         dev_info(&pdev->dev, "qp %d: Link Cleanup\n", qp->qp_num);
 769
 770         cancel_delayed_work_sync(&qp->link_work);
 771         ntb_qp_link_down_reset(qp);
 772
 773         if (qp->event_handler)
 774                 qp->event_handler(qp->cb_data, qp->link_is_up);
 775 }
 776
 777 static void ntb_qp_link_cleanup_work(struct work_struct *work)
 778 {
 779         struct ntb_transport_qp *qp = container_of(work,
 780                                                    struct ntb_transport_qp,
 781                                                    link_cleanup);
 782         struct ntb_transport_ctx *nt = qp->transport;
 783
 784         ntb_qp_link_cleanup(qp);
 785
 786         if (nt->link_is_up)
 787                 schedule_delayed_work(&qp->link_work,
 788                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 789 }
 790
 791 static void ntb_qp_link_down(struct ntb_transport_qp *qp)
 792 {
 793         schedule_work(&qp->link_cleanup);
 794 }
 795
 796 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
 797 {
 798         struct ntb_transport_qp *qp;
 799         u64 qp_bitmap_alloc;
 800         unsigned int i, count;
 801
 802         qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
 803
 804         /* Pass along the info to any clients */
 805         for (i = 0; i < nt->qp_count; i++)
 806                 if (qp_bitmap_alloc & BIT_ULL(i)) {
 807                         qp = &nt->qp_vec[i];
 808                         ntb_qp_link_cleanup(qp);
 809                         cancel_work_sync(&qp->link_cleanup);
 810                         cancel_delayed_work_sync(&qp->link_work);
 811                 }
 812
 813         if (!nt->link_is_up)
 814                 cancel_delayed_work_sync(&nt->link_work);
 815
 816         /* The scratchpad registers keep the values if the remote side
 817          * goes down, blast them now to give them a sane value the next
 818          * time they are accessed
 819          */
 820         count = ntb_spad_count(nt->ndev);
 821         for (i = 0; i < count; i++)
 822                 ntb_spad_write(nt->ndev, i, 0);
 823 }
 824
 825 static void ntb_transport_link_cleanup_work(struct work_struct *work)
 826 {
 827         struct ntb_transport_ctx *nt =
 828                 container_of(work, struct ntb_transport_ctx, link_cleanup);
 829
 830         ntb_transport_link_cleanup(nt);
 831 }
 832
 833 static void ntb_transport_event_callback(void *data)
 834 {
 835         struct ntb_transport_ctx *nt = data;
 836
 837         if (ntb_link_is_up(nt->ndev, NULL, NULL) == 1)
 838                 schedule_delayed_work(&nt->link_work, 0);
 839         else
 840                 schedule_work(&nt->link_cleanup);
 841 }
 842
 843 static void ntb_transport_link_work(struct work_struct *work)
 844 {
 845         struct ntb_transport_ctx *nt =
 846                 container_of(work, struct ntb_transport_ctx, link_work.work);
 847         struct ntb_dev *ndev = nt->ndev;
 848         struct pci_dev *pdev = ndev->pdev;
 849         resource_size_t size;
 850         u32 val;
 851         int rc = 0, i, spad;
 852
 853         /* send the local info, in the opposite order of the way we read it */
 854         for (i = 0; i < nt->mw_count; i++) {
 855                 size = nt->mw_vec[i].phys_size;
 856
 857                 if (max_mw_size && size > max_mw_size)
 858                         size = max_mw_size;
 859
 860                 spad = MW0_SZ_HIGH + (i * 2);
 861                 ntb_peer_spad_write(ndev, spad, upper_32_bits(size));
 862
 863                 spad = MW0_SZ_LOW + (i * 2);
 864                 ntb_peer_spad_write(ndev, spad, lower_32_bits(size));
 865         }
 866
 867         ntb_peer_spad_write(ndev, NUM_MWS, nt->mw_count);
 868
 869         ntb_peer_spad_write(ndev, NUM_QPS, nt->qp_count);
 870
 871         ntb_peer_spad_write(ndev, VERSION, NTB_TRANSPORT_VERSION);
 872
 873         /* Query the remote side for its info */
 874         val = ntb_spad_read(ndev, VERSION);
 875         dev_dbg(&pdev->dev, "Remote version = %d\n", val);
 876         if (val != NTB_TRANSPORT_VERSION)
 877                 goto out;
 878
 879         val = ntb_spad_read(ndev, NUM_QPS);
 880         dev_dbg(&pdev->dev, "Remote max number of qps = %d\n", val);
 881         if (val != nt->qp_count)
 882                 goto out;
 883
 884         val = ntb_spad_read(ndev, NUM_MWS);
 885         dev_dbg(&pdev->dev, "Remote number of mws = %d\n", val);
 886         if (val != nt->mw_count)
 887                 goto out;
 888
 889         for (i = 0; i < nt->mw_count; i++) {
 890                 u64 val64;
 891
 892                 val = ntb_spad_read(ndev, MW0_SZ_HIGH + (i * 2));
 893                 val64 = (u64)val << 32;
 894
 895                 val = ntb_spad_read(ndev, MW0_SZ_LOW + (i * 2));
 896                 val64 |= val;
 897
 898                 dev_dbg(&pdev->dev, "Remote MW%d size = %#llx\n", i, val64);
 899
 900                 rc = ntb_set_mw(nt, i, val64);
 901                 if (rc)
 902                         goto out1;
 903         }
 904
 905         nt->link_is_up = true;
 906
 907         for (i = 0; i < nt->qp_count; i++) {
 908                 struct ntb_transport_qp *qp = &nt->qp_vec[i];
 909
 910                 ntb_transport_setup_qp_mw(nt, i);
 911
 912                 if (qp->client_ready)
 913                         schedule_delayed_work(&qp->link_work, 0);
 914         }
 915
 916         return;
 917
 918 out1:
 919         for (i = 0; i < nt->mw_count; i++)
 920                 ntb_free_mw(nt, i);
 921
 922         /* if there's an actual failure, we should just bail */
 923         if (rc < 0) {
 924                 ntb_link_disable(ndev);
 925                 return;
 926         }
 927
 928 out:
 929         if (ntb_link_is_up(ndev, NULL, NULL) == 1)
 930                 schedule_delayed_work(&nt->link_work,
 931                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 932 }
 933
 934 static void ntb_qp_link_work(struct work_struct *work)
 935 {
 936         struct ntb_transport_qp *qp = container_of(work,
 937                                                    struct ntb_transport_qp,
 938                                                    link_work.work);
 939         struct pci_dev *pdev = qp->ndev->pdev;
 940         struct ntb_transport_ctx *nt = qp->transport;
 941         int val;
 942
 943         WARN_ON(!nt->link_is_up);
 944
 945         val = ntb_spad_read(nt->ndev, QP_LINKS);
 946
 947         ntb_peer_spad_write(nt->ndev, QP_LINKS, val | BIT(qp->qp_num));
 948
 949         /* query remote spad for qp ready bits */
 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         tx_size = (unsigned int)mw_size / num_qps_mw;
1000         qp_offset = tx_size * (qp_num / mw_count);
1001
1002         qp->tx_mw = nt->mw_vec[mw_num].vbase + qp_offset;
1003         if (!qp->tx_mw)
1004                 return -EINVAL;
1005
1006         qp->tx_mw_phys = mw_base + qp_offset;
1007         if (!qp->tx_mw_phys)
1008                 return -EINVAL;
1009
1010         tx_size -= sizeof(struct ntb_rx_info);
1011         qp->rx_info = qp->tx_mw + tx_size;
1012
1013         /* Due to housekeeping, there must be atleast 2 buffs */
1014         qp->tx_max_frame = min(transport_mtu, tx_size / 2);
1015         qp->tx_max_entry = tx_size / qp->tx_max_frame;
1016
1017         if (nt->debugfs_node_dir) {
1018                 char debugfs_name[4];
1019
1020                 snprintf(debugfs_name, 4, "qp%d", qp_num);
1021                 qp->debugfs_dir = debugfs_create_dir(debugfs_name,
1022                                                      nt->debugfs_node_dir);
1023
1024                 qp->debugfs_stats = debugfs_create_file("stats", S_IRUSR,
1025                                                         qp->debugfs_dir, qp,
1026                                                         &ntb_qp_debugfs_stats);
1027         } else {
1028                 qp->debugfs_dir = NULL;
1029                 qp->debugfs_stats = NULL;
1030         }
1031
1032         INIT_DELAYED_WORK(&qp->link_work, ntb_qp_link_work);
1033         INIT_WORK(&qp->link_cleanup, ntb_qp_link_cleanup_work);
1034
1035         spin_lock_init(&qp->ntb_rx_q_lock);
1036         spin_lock_init(&qp->ntb_tx_free_q_lock);
1037
1038         INIT_LIST_HEAD(&qp->rx_post_q);
1039         INIT_LIST_HEAD(&qp->rx_pend_q);
1040         INIT_LIST_HEAD(&qp->rx_free_q);
1041         INIT_LIST_HEAD(&qp->tx_free_q);
1042
1043         tasklet_init(&qp->rxc_db_work, ntb_transport_rxc_db,
1044                      (unsigned long)qp);
1045
1046         return 0;
1047 }
1048
1049 static int ntb_transport_probe(struct ntb_client *self, struct ntb_dev *ndev)
1050 {
1051         struct ntb_transport_ctx *nt;
1052         struct ntb_transport_mw *mw;
1053         unsigned int mw_count, qp_count, spad_count, max_mw_count_for_spads;
1054         u64 qp_bitmap;
1055         int node;
1056         int rc, i;
1057
1058         mw_count = ntb_mw_count(ndev);
1059
1060         if (ntb_db_is_unsafe(ndev))
1061                 dev_dbg(&ndev->dev,
1062                         "doorbell is unsafe, proceed anyway...\n");
1063         if (ntb_spad_is_unsafe(ndev))
1064                 dev_dbg(&ndev->dev,
1065                         "scratchpad is unsafe, proceed anyway...\n");
1066
1067         node = dev_to_node(&ndev->dev);
1068
1069         nt = kzalloc_node(sizeof(*nt), GFP_KERNEL, node);
1070         if (!nt)
1071                 return -ENOMEM;
1072
1073         nt->ndev = ndev;
1074         spad_count = ntb_spad_count(ndev);
1075
1076         /* Limit the MW's based on the availability of scratchpads */
1077
1078         if (spad_count < NTB_TRANSPORT_MIN_SPADS) {
1079                 nt->mw_count = 0;
1080                 rc = -EINVAL;
1081                 goto err;
1082         }
1083
1084         max_mw_count_for_spads = (spad_count - MW0_SZ_HIGH) / 2;
1085         nt->mw_count = min(mw_count, max_mw_count_for_spads);
1086
1087         nt->mw_vec = kzalloc_node(mw_count * sizeof(*nt->mw_vec),
1088                                   GFP_KERNEL, node);
1089         if (!nt->mw_vec) {
1090                 rc = -ENOMEM;
1091                 goto err;
1092         }
1093
1094         for (i = 0; i < mw_count; i++) {
1095                 mw = &nt->mw_vec[i];
1096
1097                 rc = ntb_mw_get_range(ndev, i, &mw->phys_addr, &mw->phys_size,
1098                                       &mw->xlat_align, &mw->xlat_align_size);
1099                 if (rc)
1100                         goto err1;
1101
1102                 mw->vbase = ioremap_wc(mw->phys_addr, mw->phys_size);
1103                 if (!mw->vbase) {
1104                         rc = -ENOMEM;
1105                         goto err1;
1106                 }
1107
1108                 mw->buff_size = 0;
1109                 mw->xlat_size = 0;
1110                 mw->virt_addr = NULL;
1111                 mw->dma_addr = 0;
1112         }
1113
1114         qp_bitmap = ntb_db_valid_mask(ndev);
1115
1116         qp_count = ilog2(qp_bitmap);
1117         if (max_num_clients && max_num_clients < qp_count)
1118                 qp_count = max_num_clients;
1119         else if (nt->mw_count < qp_count)
1120                 qp_count = nt->mw_count;
1121
1122         qp_bitmap &= BIT_ULL(qp_count) - 1;
1123
1124         nt->qp_count = qp_count;
1125         nt->qp_bitmap = qp_bitmap;
1126         nt->qp_bitmap_free = qp_bitmap;
1127
1128         nt->qp_vec = kzalloc_node(qp_count * sizeof(*nt->qp_vec),
1129                                   GFP_KERNEL, node);
1130         if (!nt->qp_vec) {
1131                 rc = -ENOMEM;
1132                 goto err1;
1133         }
1134
1135         if (nt_debugfs_dir) {
1136                 nt->debugfs_node_dir =
1137                         debugfs_create_dir(pci_name(ndev->pdev),
1138                                            nt_debugfs_dir);
1139         }
1140
1141         for (i = 0; i < qp_count; i++) {
1142                 rc = ntb_transport_init_queue(nt, i);
1143                 if (rc)
1144                         goto err2;
1145         }
1146
1147         INIT_DELAYED_WORK(&nt->link_work, ntb_transport_link_work);
1148         INIT_WORK(&nt->link_cleanup, ntb_transport_link_cleanup_work);
1149
1150         rc = ntb_set_ctx(ndev, nt, &ntb_transport_ops);
1151         if (rc)
1152                 goto err2;
1153
1154         INIT_LIST_HEAD(&nt->client_devs);
1155         rc = ntb_bus_init(nt);
1156         if (rc)
1157                 goto err3;
1158
1159         nt->link_is_up = false;
1160         ntb_link_enable(ndev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
1161         ntb_link_event(ndev);
1162
1163         return 0;
1164
1165 err3:
1166         ntb_clear_ctx(ndev);
1167 err2:
1168         kfree(nt->qp_vec);
1169 err1:
1170         while (i--) {
1171                 mw = &nt->mw_vec[i];
1172                 iounmap(mw->vbase);
1173         }
1174         kfree(nt->mw_vec);
1175 err:
1176         kfree(nt);
1177         return rc;
1178 }
1179
1180 static void ntb_transport_free(struct ntb_client *self, struct ntb_dev *ndev)
1181 {
1182         struct ntb_transport_ctx *nt = ndev->ctx;
1183         struct ntb_transport_qp *qp;
1184         u64 qp_bitmap_alloc;
1185         int i;
1186
1187         ntb_transport_link_cleanup(nt);
1188         cancel_work_sync(&nt->link_cleanup);
1189         cancel_delayed_work_sync(&nt->link_work);
1190
1191         qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
1192
1193         /* verify that all the qp's are freed */
1194         for (i = 0; i < nt->qp_count; i++) {
1195                 qp = &nt->qp_vec[i];
1196                 if (qp_bitmap_alloc & BIT_ULL(i))
1197                         ntb_transport_free_queue(qp);
1198                 debugfs_remove_recursive(qp->debugfs_dir);
1199         }
1200
1201         ntb_link_disable(ndev);
1202         ntb_clear_ctx(ndev);
1203
1204         ntb_bus_remove(nt);
1205
1206         for (i = nt->mw_count; i--; ) {
1207                 ntb_free_mw(nt, i);
1208                 iounmap(nt->mw_vec[i].vbase);
1209         }
1210
1211         kfree(nt->qp_vec);
1212         kfree(nt->mw_vec);
1213         kfree(nt);
1214 }
1215
1216 static void ntb_complete_rxc(struct ntb_transport_qp *qp)
1217 {
1218         struct ntb_queue_entry *entry;
1219         void *cb_data;
1220         unsigned int len;
1221         unsigned long irqflags;
1222
1223         spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1224
1225         while (!list_empty(&qp->rx_post_q)) {
1226                 entry = list_first_entry(&qp->rx_post_q,
1227                                          struct ntb_queue_entry, entry);
1228                 if (!(entry->flags & DESC_DONE_FLAG))
1229                         break;
1230
1231                 entry->rx_hdr->flags = 0;
1232                 iowrite32(entry->rx_index, &qp->rx_info->entry);
1233
1234                 cb_data = entry->cb_data;
1235                 len = entry->len;
1236
1237                 list_move_tail(&entry->entry, &qp->rx_free_q);
1238
1239                 spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1240
1241                 if (qp->rx_handler && qp->client_ready)
1242                         qp->rx_handler(qp, qp->cb_data, cb_data, len);
1243
1244                 spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1245         }
1246
1247         spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1248 }
1249
1250 static void ntb_rx_copy_callback(void *data,
1251                                  const struct dmaengine_result *res)
1252 {
1253         struct ntb_queue_entry *entry = data;
1254
1255         /* we need to check DMA results if we are using DMA */
1256         if (res) {
1257                 enum dmaengine_tx_result dma_err = res->result;
1258
1259                 switch (dma_err) {
1260                 case DMA_TRANS_READ_FAILED:
1261                 case DMA_TRANS_WRITE_FAILED:
1262                         entry->errors++;
1263                 case DMA_TRANS_ABORTED:
1264                 {
1265                         struct ntb_transport_qp *qp = entry->qp;
1266                         void *offset = qp->rx_buff + qp->rx_max_frame *
1267                                         qp->rx_index;
1268
1269                         ntb_memcpy_rx(entry, offset);
1270                         qp->rx_memcpy++;
1271                         return;
1272                 }
1273
1274                 case DMA_TRANS_NOERROR:
1275                 default:
1276                         break;
1277                 }
1278         }
1279
1280         entry->flags |= DESC_DONE_FLAG;
1281
1282         ntb_complete_rxc(entry->qp);
1283 }
1284
1285 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset)
1286 {
1287         void *buf = entry->buf;
1288         size_t len = entry->len;
1289
1290         memcpy(buf, offset, len);
1291
1292         /* Ensure that the data is fully copied out before clearing the flag */
1293         wmb();
1294
1295         ntb_rx_copy_callback(entry, NULL);
1296 }
1297
1298 static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset)
1299 {
1300         struct dma_async_tx_descriptor *txd;
1301         struct ntb_transport_qp *qp = entry->qp;
1302         struct dma_chan *chan = qp->rx_dma_chan;
1303         struct dma_device *device;
1304         size_t pay_off, buff_off, len;
1305         struct dmaengine_unmap_data *unmap;
1306         dma_cookie_t cookie;
1307         void *buf = entry->buf;
1308
1309         len = entry->len;
1310         device = chan->device;
1311         pay_off = (size_t)offset & ~PAGE_MASK;
1312         buff_off = (size_t)buf & ~PAGE_MASK;
1313
1314         if (!is_dma_copy_aligned(device, pay_off, buff_off, len))
1315                 goto err;
1316
1317         unmap = dmaengine_get_unmap_data(device->dev, 2, GFP_NOWAIT);
1318         if (!unmap)
1319                 goto err;
1320
1321         unmap->len = len;
1322         unmap->addr[0] = dma_map_page(device->dev, virt_to_page(offset),
1323                                       pay_off, len, DMA_TO_DEVICE);
1324         if (dma_mapping_error(device->dev, unmap->addr[0]))
1325                 goto err_get_unmap;
1326
1327         unmap->to_cnt = 1;
1328
1329         unmap->addr[1] = dma_map_page(device->dev, virt_to_page(buf),
1330                                       buff_off, len, DMA_FROM_DEVICE);
1331         if (dma_mapping_error(device->dev, unmap->addr[1]))
1332                 goto err_get_unmap;
1333
1334         unmap->from_cnt = 1;
1335
1336         txd = device->device_prep_dma_memcpy(chan, unmap->addr[1],
1337                                              unmap->addr[0], len,
1338                                              DMA_PREP_INTERRUPT);
1339         if (!txd)
1340                 goto err_get_unmap;
1341
1342         txd->callback_result = ntb_rx_copy_callback;
1343         txd->callback_param = entry;
1344         dma_set_unmap(txd, unmap);
1345
1346         cookie = dmaengine_submit(txd);
1347         if (dma_submit_error(cookie))
1348                 goto err_set_unmap;
1349
1350         dmaengine_unmap_put(unmap);
1351
1352         qp->last_cookie = cookie;
1353
1354         qp->rx_async++;
1355
1356         return 0;
1357
1358 err_set_unmap:
1359         dmaengine_unmap_put(unmap);
1360 err_get_unmap:
1361         dmaengine_unmap_put(unmap);
1362 err:
1363         return -ENXIO;
1364 }
1365
1366 static void ntb_async_rx(struct ntb_queue_entry *entry, void *offset)
1367 {
1368         struct ntb_transport_qp *qp = entry->qp;
1369         struct dma_chan *chan = qp->rx_dma_chan;
1370         int res;
1371
1372         if (!chan)
1373                 goto err;
1374
1375         if (entry->len < copy_bytes)
1376                 goto err;
1377
1378         res = ntb_async_rx_submit(entry, offset);
1379         if (res < 0)
1380                 goto err;
1381
1382         if (!entry->retries)
1383                 qp->rx_async++;
1384
1385         return;
1386
1387 err:
1388         ntb_memcpy_rx(entry, offset);
1389         qp->rx_memcpy++;
1390 }
1391
1392 static int ntb_process_rxc(struct ntb_transport_qp *qp)
1393 {
1394         struct ntb_payload_header *hdr;
1395         struct ntb_queue_entry *entry;
1396         void *offset;
1397
1398         offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
1399         hdr = offset + qp->rx_max_frame - sizeof(struct ntb_payload_header);
1400
1401         dev_dbg(&qp->ndev->pdev->dev, "qp %d: RX ver %u len %d flags %x\n",
1402                 qp->qp_num, hdr->ver, hdr->len, hdr->flags);
1403
1404         if (!(hdr->flags & DESC_DONE_FLAG)) {
1405                 dev_dbg(&qp->ndev->pdev->dev, "done flag not set\n");
1406                 qp->rx_ring_empty++;
1407                 return -EAGAIN;
1408         }
1409
1410         if (hdr->flags & LINK_DOWN_FLAG) {
1411                 dev_dbg(&qp->ndev->pdev->dev, "link down flag set\n");
1412                 ntb_qp_link_down(qp);
1413                 hdr->flags = 0;
1414                 return -EAGAIN;
1415         }
1416
1417         if (hdr->ver != (u32)qp->rx_pkts) {
1418                 dev_dbg(&qp->ndev->pdev->dev,
1419                         "version mismatch, expected %llu - got %u\n",
1420                         qp->rx_pkts, hdr->ver);
1421                 qp->rx_err_ver++;
1422                 return -EIO;
1423         }
1424
1425         entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q);
1426         if (!entry) {
1427                 dev_dbg(&qp->ndev->pdev->dev, "no receive buffer\n");
1428                 qp->rx_err_no_buf++;
1429                 return -EAGAIN;
1430         }
1431
1432         entry->rx_hdr = hdr;
1433         entry->rx_index = qp->rx_index;
1434
1435         if (hdr->len > entry->len) {
1436                 dev_dbg(&qp->ndev->pdev->dev,
1437                         "receive buffer overflow! Wanted %d got %d\n",
1438                         hdr->len, entry->len);
1439                 qp->rx_err_oflow++;
1440
1441                 entry->len = -EIO;
1442                 entry->flags |= DESC_DONE_FLAG;
1443
1444                 ntb_complete_rxc(qp);
1445         } else {
1446                 dev_dbg(&qp->ndev->pdev->dev,
1447                         "RX OK index %u ver %u size %d into buf size %d\n",
1448                         qp->rx_index, hdr->ver, hdr->len, entry->len);
1449
1450                 qp->rx_bytes += hdr->len;
1451                 qp->rx_pkts++;
1452
1453                 entry->len = hdr->len;
1454
1455                 ntb_async_rx(entry, offset);
1456         }
1457
1458         qp->rx_index++;
1459         qp->rx_index %= qp->rx_max_entry;
1460
1461         return 0;
1462 }
1463
1464 static void ntb_transport_rxc_db(unsigned long data)
1465 {
1466         struct ntb_transport_qp *qp = (void *)data;
1467         int rc, i;
1468
1469         dev_dbg(&qp->ndev->pdev->dev, "%s: doorbell %d received\n",
1470                 __func__, qp->qp_num);
1471
1472         /* Limit the number of packets processed in a single interrupt to
1473          * provide fairness to others
1474          */
1475         for (i = 0; i < qp->rx_max_entry; i++) {
1476                 rc = ntb_process_rxc(qp);
1477                 if (rc)
1478                         break;
1479         }
1480
1481         if (i && qp->rx_dma_chan)
1482                 dma_async_issue_pending(qp->rx_dma_chan);
1483
1484         if (i == qp->rx_max_entry) {
1485                 /* there is more work to do */
1486                 if (qp->active)
1487                         tasklet_schedule(&qp->rxc_db_work);
1488         } else if (ntb_db_read(qp->ndev) & BIT_ULL(qp->qp_num)) {
1489                 /* the doorbell bit is set: clear it */
1490                 ntb_db_clear(qp->ndev, BIT_ULL(qp->qp_num));
1491                 /* ntb_db_read ensures ntb_db_clear write is committed */
1492                 ntb_db_read(qp->ndev);
1493
1494                 /* an interrupt may have arrived between finishing
1495                  * ntb_process_rxc and clearing the doorbell bit:
1496                  * there might be some more work to do.
1497                  */
1498                 if (qp->active)
1499                         tasklet_schedule(&qp->rxc_db_work);
1500         }
1501 }
1502
1503 static void ntb_tx_copy_callback(void *data,
1504                                  const struct dmaengine_result *res)
1505 {
1506         struct ntb_queue_entry *entry = data;
1507         struct ntb_transport_qp *qp = entry->qp;
1508         struct ntb_payload_header __iomem *hdr = entry->tx_hdr;
1509
1510         /* we need to check DMA results if we are using DMA */
1511         if (res) {
1512                 enum dmaengine_tx_result dma_err = res->result;
1513
1514                 switch (dma_err) {
1515                 case DMA_TRANS_READ_FAILED:
1516                 case DMA_TRANS_WRITE_FAILED:
1517                         entry->errors++;
1518                 case DMA_TRANS_ABORTED:
1519                 {
1520                         void __iomem *offset =
1521                                 qp->tx_mw + qp->tx_max_frame *
1522                                 entry->tx_index;
1523
1524                         /* resubmit via CPU */
1525                         ntb_memcpy_tx(entry, offset);
1526                         qp->tx_memcpy++;
1527                         return;
1528                 }
1529
1530                 case DMA_TRANS_NOERROR:
1531                 default:
1532                         break;
1533                 }
1534         }
1535
1536         iowrite32(entry->flags | DESC_DONE_FLAG, &hdr->flags);
1537
1538         ntb_peer_db_set(qp->ndev, BIT_ULL(qp->qp_num));
1539
1540         /* The entry length can only be zero if the packet is intended to be a
1541          * "link down" or similar.  Since no payload is being sent in these
1542          * cases, there is nothing to add to the completion queue.
1543          */
1544         if (entry->len > 0) {
1545                 qp->tx_bytes += entry->len;
1546
1547                 if (qp->tx_handler)
1548                         qp->tx_handler(qp, qp->cb_data, entry->cb_data,
1549                                        entry->len);
1550         }
1551
1552         ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry, &qp->tx_free_q);
1553 }
1554
1555 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset)
1556 {
1557 #ifdef ARCH_HAS_NOCACHE_UACCESS
1558         /*
1559          * Using non-temporal mov to improve performance on non-cached
1560          * writes, even though we aren't actually copying from user space.
1561          */
1562         __copy_from_user_inatomic_nocache(offset, entry->buf, entry->len);
1563 #else
1564         memcpy_toio(offset, entry->buf, entry->len);
1565 #endif
1566
1567         /* Ensure that the data is fully copied out before setting the flags */
1568         wmb();
1569
1570         ntb_tx_copy_callback(entry, NULL);
1571 }
1572
1573 static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
1574                                struct ntb_queue_entry *entry)
1575 {
1576         struct dma_async_tx_descriptor *txd;
1577         struct dma_chan *chan = qp->tx_dma_chan;
1578         struct dma_device *device;
1579         size_t len = entry->len;
1580         void *buf = entry->buf;
1581         size_t dest_off, buff_off;
1582         struct dmaengine_unmap_data *unmap;
1583         dma_addr_t dest;
1584         dma_cookie_t cookie;
1585
1586         device = chan->device;
1587         dest = qp->tx_mw_phys + qp->tx_max_frame * entry->tx_index;
1588         buff_off = (size_t)buf & ~PAGE_MASK;
1589         dest_off = (size_t)dest & ~PAGE_MASK;
1590
1591         if (!is_dma_copy_aligned(device, buff_off, dest_off, len))
1592                 goto err;
1593
1594         unmap = dmaengine_get_unmap_data(device->dev, 1, GFP_NOWAIT);
1595         if (!unmap)
1596                 goto err;
1597
1598         unmap->len = len;
1599         unmap->addr[0] = dma_map_page(device->dev, virt_to_page(buf),
1600                                       buff_off, len, DMA_TO_DEVICE);
1601         if (dma_mapping_error(device->dev, unmap->addr[0]))
1602                 goto err_get_unmap;
1603
1604         unmap->to_cnt = 1;
1605
1606         txd = device->device_prep_dma_memcpy(chan, dest, unmap->addr[0], len,
1607                                              DMA_PREP_INTERRUPT);
1608         if (!txd)
1609                 goto err_get_unmap;
1610
1611         txd->callback_result = ntb_tx_copy_callback;
1612         txd->callback_param = entry;
1613         dma_set_unmap(txd, unmap);
1614
1615         cookie = dmaengine_submit(txd);
1616         if (dma_submit_error(cookie))
1617                 goto err_set_unmap;
1618
1619         dmaengine_unmap_put(unmap);
1620
1621         dma_async_issue_pending(chan);
1622
1623         return 0;
1624 err_set_unmap:
1625         dmaengine_unmap_put(unmap);
1626 err_get_unmap:
1627         dmaengine_unmap_put(unmap);
1628 err:
1629         return -ENXIO;
1630 }
1631
1632 static void ntb_async_tx(struct ntb_transport_qp *qp,
1633                          struct ntb_queue_entry *entry)
1634 {
1635         struct ntb_payload_header __iomem *hdr;
1636         struct dma_chan *chan = qp->tx_dma_chan;
1637         void __iomem *offset;
1638         int res;
1639
1640         entry->tx_index = qp->tx_index;
1641         offset = qp->tx_mw + qp->tx_max_frame * entry->tx_index;
1642         hdr = offset + qp->tx_max_frame - sizeof(struct ntb_payload_header);
1643         entry->tx_hdr = hdr;
1644
1645         iowrite32(entry->len, &hdr->len);
1646         iowrite32((u32)qp->tx_pkts, &hdr->ver);
1647
1648         if (!chan)
1649                 goto err;
1650
1651         if (entry->len < copy_bytes)
1652                 goto err;
1653
1654         res = ntb_async_tx_submit(qp, entry);
1655         if (res < 0)
1656                 goto err;
1657
1658         if (!entry->retries)
1659                 qp->tx_async++;
1660
1661         return;
1662
1663 err:
1664         ntb_memcpy_tx(entry, offset);
1665         qp->tx_memcpy++;
1666 }
1667
1668 static int ntb_process_tx(struct ntb_transport_qp *qp,
1669                           struct ntb_queue_entry *entry)
1670 {
1671         if (qp->tx_index == qp->remote_rx_info->entry) {
1672                 qp->tx_ring_full++;
1673                 return -EAGAIN;
1674         }
1675
1676         if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
1677                 if (qp->tx_handler)
1678                         qp->tx_handler(qp, qp->cb_data, NULL, -EIO);
1679
1680                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1681                              &qp->tx_free_q);
1682                 return 0;
1683         }
1684
1685         ntb_async_tx(qp, entry);
1686
1687         qp->tx_index++;
1688         qp->tx_index %= qp->tx_max_entry;
1689
1690         qp->tx_pkts++;
1691
1692         return 0;
1693 }
1694
1695 static void ntb_send_link_down(struct ntb_transport_qp *qp)
1696 {
1697         struct pci_dev *pdev = qp->ndev->pdev;
1698         struct ntb_queue_entry *entry;
1699         int i, rc;
1700
1701         if (!qp->link_is_up)
1702                 return;
1703
1704         dev_info(&pdev->dev, "qp %d: Send Link Down\n", qp->qp_num);
1705
1706         for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1707                 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1708                 if (entry)
1709                         break;
1710                 msleep(100);
1711         }
1712
1713         if (!entry)
1714                 return;
1715
1716         entry->cb_data = NULL;
1717         entry->buf = NULL;
1718         entry->len = 0;
1719         entry->flags = LINK_DOWN_FLAG;
1720
1721         rc = ntb_process_tx(qp, entry);
1722         if (rc)
1723                 dev_err(&pdev->dev, "ntb: QP%d unable to send linkdown msg\n",
1724                         qp->qp_num);
1725
1726         ntb_qp_link_down_reset(qp);
1727 }
1728
1729 static bool ntb_dma_filter_fn(struct dma_chan *chan, void *node)
1730 {
1731         return dev_to_node(&chan->dev->device) == (int)(unsigned long)node;
1732 }
1733
1734 /**
1735  * ntb_transport_create_queue - Create a new NTB transport layer queue
1736  * @rx_handler: receive callback function
1737  * @tx_handler: transmit callback function
1738  * @event_handler: event callback function
1739  *
1740  * Create a new NTB transport layer queue and provide the queue with a callback
1741  * routine for both transmit and receive.  The receive callback routine will be
1742  * used to pass up data when the transport has received it on the queue.   The
1743  * transmit callback routine will be called when the transport has completed the
1744  * transmission of the data on the queue and the data is ready to be freed.
1745  *
1746  * RETURNS: pointer to newly created ntb_queue, NULL on error.
1747  */
1748 struct ntb_transport_qp *
1749 ntb_transport_create_queue(void *data, struct device *client_dev,
1750                            const struct ntb_queue_handlers *handlers)
1751 {
1752         struct ntb_dev *ndev;
1753         struct pci_dev *pdev;
1754         struct ntb_transport_ctx *nt;
1755         struct ntb_queue_entry *entry;
1756         struct ntb_transport_qp *qp;
1757         u64 qp_bit;
1758         unsigned int free_queue;
1759         dma_cap_mask_t dma_mask;
1760         int node;
1761         int i;
1762
1763         ndev = dev_ntb(client_dev->parent);
1764         pdev = ndev->pdev;
1765         nt = ndev->ctx;
1766
1767         node = dev_to_node(&ndev->dev);
1768
1769         free_queue = ffs(nt->qp_bitmap_free);
1770         if (!free_queue)
1771                 goto err;
1772
1773         /* decrement free_queue to make it zero based */
1774         free_queue--;
1775
1776         qp = &nt->qp_vec[free_queue];
1777         qp_bit = BIT_ULL(qp->qp_num);
1778
1779         nt->qp_bitmap_free &= ~qp_bit;
1780
1781         qp->cb_data = data;
1782         qp->rx_handler = handlers->rx_handler;
1783         qp->tx_handler = handlers->tx_handler;
1784         qp->event_handler = handlers->event_handler;
1785
1786         dma_cap_zero(dma_mask);
1787         dma_cap_set(DMA_MEMCPY, dma_mask);
1788
1789         if (use_dma) {
1790                 qp->tx_dma_chan =
1791                         dma_request_channel(dma_mask, ntb_dma_filter_fn,
1792                                             (void *)(unsigned long)node);
1793                 if (!qp->tx_dma_chan)
1794                         dev_info(&pdev->dev, "Unable to allocate TX DMA channel\n");
1795
1796                 qp->rx_dma_chan =
1797                         dma_request_channel(dma_mask, ntb_dma_filter_fn,
1798                                             (void *)(unsigned long)node);
1799                 if (!qp->rx_dma_chan)
1800                         dev_info(&pdev->dev, "Unable to allocate RX DMA channel\n");
1801         } else {
1802                 qp->tx_dma_chan = NULL;
1803                 qp->rx_dma_chan = NULL;
1804         }
1805
1806         dev_dbg(&pdev->dev, "Using %s memcpy for TX\n",
1807                 qp->tx_dma_chan ? "DMA" : "CPU");
1808
1809         dev_dbg(&pdev->dev, "Using %s memcpy for RX\n",
1810                 qp->rx_dma_chan ? "DMA" : "CPU");
1811
1812         for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
1813                 entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
1814                 if (!entry)
1815                         goto err1;
1816
1817                 entry->qp = qp;
1818                 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
1819                              &qp->rx_free_q);
1820         }
1821         qp->rx_alloc_entry = NTB_QP_DEF_NUM_ENTRIES;
1822
1823         for (i = 0; i < qp->tx_max_entry; i++) {
1824                 entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
1825                 if (!entry)
1826                         goto err2;
1827
1828                 entry->qp = qp;
1829                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1830                              &qp->tx_free_q);
1831         }
1832
1833         ntb_db_clear(qp->ndev, qp_bit);
1834         ntb_db_clear_mask(qp->ndev, qp_bit);
1835
1836         dev_info(&pdev->dev, "NTB Transport QP %d created\n", qp->qp_num);
1837
1838         return qp;
1839
1840 err2:
1841         while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1842                 kfree(entry);
1843 err1:
1844         qp->rx_alloc_entry = 0;
1845         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
1846                 kfree(entry);
1847         if (qp->tx_dma_chan)
1848                 dma_release_channel(qp->tx_dma_chan);
1849         if (qp->rx_dma_chan)
1850                 dma_release_channel(qp->rx_dma_chan);
1851         nt->qp_bitmap_free |= qp_bit;
1852 err:
1853         return NULL;
1854 }
1855 EXPORT_SYMBOL_GPL(ntb_transport_create_queue);
1856
1857 /**
1858  * ntb_transport_free_queue - Frees NTB transport queue
1859  * @qp: NTB queue to be freed
1860  *
1861  * Frees NTB transport queue
1862  */
1863 void ntb_transport_free_queue(struct ntb_transport_qp *qp)
1864 {
1865         struct pci_dev *pdev;
1866         struct ntb_queue_entry *entry;
1867         u64 qp_bit;
1868
1869         if (!qp)
1870                 return;
1871
1872         pdev = qp->ndev->pdev;
1873
1874         qp->active = false;
1875
1876         if (qp->tx_dma_chan) {
1877                 struct dma_chan *chan = qp->tx_dma_chan;
1878                 /* Putting the dma_chan to NULL will force any new traffic to be
1879                  * processed by the CPU instead of the DAM engine
1880                  */
1881                 qp->tx_dma_chan = NULL;
1882
1883                 /* Try to be nice and wait for any queued DMA engine
1884                  * transactions to process before smashing it with a rock
1885                  */
1886                 dma_sync_wait(chan, qp->last_cookie);
1887                 dmaengine_terminate_all(chan);
1888                 dma_release_channel(chan);
1889         }
1890
1891         if (qp->rx_dma_chan) {
1892                 struct dma_chan *chan = qp->rx_dma_chan;
1893                 /* Putting the dma_chan to NULL will force any new traffic to be
1894                  * processed by the CPU instead of the DAM engine
1895                  */
1896                 qp->rx_dma_chan = NULL;
1897
1898                 /* Try to be nice and wait for any queued DMA engine
1899                  * transactions to process before smashing it with a rock
1900                  */
1901                 dma_sync_wait(chan, qp->last_cookie);
1902                 dmaengine_terminate_all(chan);
1903                 dma_release_channel(chan);
1904         }
1905
1906         qp_bit = BIT_ULL(qp->qp_num);
1907
1908         ntb_db_set_mask(qp->ndev, qp_bit);
1909         tasklet_kill(&qp->rxc_db_work);
1910
1911         cancel_delayed_work_sync(&qp->link_work);
1912
1913         qp->cb_data = NULL;
1914         qp->rx_handler = NULL;
1915         qp->tx_handler = NULL;
1916         qp->event_handler = NULL;
1917
1918         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
1919                 kfree(entry);
1920
1921         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q))) {
1922                 dev_warn(&pdev->dev, "Freeing item from non-empty rx_pend_q\n");
1923                 kfree(entry);
1924         }
1925
1926         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q))) {
1927                 dev_warn(&pdev->dev, "Freeing item from non-empty rx_post_q\n");
1928                 kfree(entry);
1929         }
1930
1931         while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1932                 kfree(entry);
1933
1934         qp->transport->qp_bitmap_free |= qp_bit;
1935
1936         dev_info(&pdev->dev, "NTB Transport QP %d freed\n", qp->qp_num);
1937 }
1938 EXPORT_SYMBOL_GPL(ntb_transport_free_queue);
1939
1940 /**
1941  * ntb_transport_rx_remove - Dequeues enqueued rx packet
1942  * @qp: NTB queue to be freed
1943  * @len: pointer to variable to write enqueued buffers length
1944  *
1945  * Dequeues unused buffers from receive queue.  Should only be used during
1946  * shutdown of qp.
1947  *
1948  * RETURNS: NULL error value on error, or void* for success.
1949  */
1950 void *ntb_transport_rx_remove(struct ntb_transport_qp *qp, unsigned int *len)
1951 {
1952         struct ntb_queue_entry *entry;
1953         void *buf;
1954
1955         if (!qp || qp->client_ready)
1956                 return NULL;
1957
1958         entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q);
1959         if (!entry)
1960                 return NULL;
1961
1962         buf = entry->cb_data;
1963         *len = entry->len;
1964
1965         ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_free_q);
1966
1967         return buf;
1968 }
1969 EXPORT_SYMBOL_GPL(ntb_transport_rx_remove);
1970
1971 /**
1972  * ntb_transport_rx_enqueue - Enqueue a new NTB queue entry
1973  * @qp: NTB transport layer queue the entry is to be enqueued on
1974  * @cb: per buffer pointer for callback function to use
1975  * @data: pointer to data buffer that incoming packets will be copied into
1976  * @len: length of the data buffer
1977  *
1978  * Enqueue a new receive buffer onto the transport queue into which a NTB
1979  * payload can be received into.
1980  *
1981  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
1982  */
1983 int ntb_transport_rx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
1984                              unsigned int len)
1985 {
1986         struct ntb_queue_entry *entry;
1987
1988         if (!qp)
1989                 return -EINVAL;
1990
1991         entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q);
1992         if (!entry)
1993                 return -ENOMEM;
1994
1995         entry->cb_data = cb;
1996         entry->buf = data;
1997         entry->len = len;
1998         entry->flags = 0;
1999         entry->retries = 0;
2000         entry->errors = 0;
2001         entry->rx_index = 0;
2002
2003         ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_pend_q);
2004
2005         if (qp->active)
2006                 tasklet_schedule(&qp->rxc_db_work);
2007
2008         return 0;
2009 }
2010 EXPORT_SYMBOL_GPL(ntb_transport_rx_enqueue);
2011
2012 /**
2013  * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
2014  * @qp: NTB transport layer queue the entry is to be enqueued on
2015  * @cb: per buffer pointer for callback function to use
2016  * @data: pointer to data buffer that will be sent
2017  * @len: length of the data buffer
2018  *
2019  * Enqueue a new transmit buffer onto the transport queue from which a NTB
2020  * payload will be transmitted.  This assumes that a lock is being held to
2021  * serialize access to the qp.
2022  *
2023  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2024  */
2025 int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
2026                              unsigned int len)
2027 {
2028         struct ntb_queue_entry *entry;
2029         int rc;
2030
2031         if (!qp || !qp->link_is_up || !len)
2032                 return -EINVAL;
2033
2034         entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
2035         if (!entry) {
2036                 qp->tx_err_no_buf++;
2037                 return -EBUSY;
2038         }
2039
2040         entry->cb_data = cb;
2041         entry->buf = data;
2042         entry->len = len;
2043         entry->flags = 0;
2044         entry->errors = 0;
2045         entry->retries = 0;
2046         entry->tx_index = 0;
2047
2048         rc = ntb_process_tx(qp, entry);
2049         if (rc)
2050                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
2051                              &qp->tx_free_q);
2052
2053         return rc;
2054 }
2055 EXPORT_SYMBOL_GPL(ntb_transport_tx_enqueue);
2056
2057 /**
2058  * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
2059  * @qp: NTB transport layer queue to be enabled
2060  *
2061  * Notify NTB transport layer of client readiness to use queue
2062  */
2063 void ntb_transport_link_up(struct ntb_transport_qp *qp)
2064 {
2065         if (!qp)
2066                 return;
2067
2068         qp->client_ready = true;
2069
2070         if (qp->transport->link_is_up)
2071                 schedule_delayed_work(&qp->link_work, 0);
2072 }
2073 EXPORT_SYMBOL_GPL(ntb_transport_link_up);
2074
2075 /**
2076  * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
2077  * @qp: NTB transport layer queue to be disabled
2078  *
2079  * Notify NTB transport layer of client's desire to no longer receive data on
2080  * transport queue specified.  It is the client's responsibility to ensure all
2081  * entries on queue are purged or otherwise handled appropriately.
2082  */
2083 void ntb_transport_link_down(struct ntb_transport_qp *qp)
2084 {
2085         int val;
2086
2087         if (!qp)
2088                 return;
2089
2090         qp->client_ready = false;
2091
2092         val = ntb_spad_read(qp->ndev, QP_LINKS);
2093
2094         ntb_peer_spad_write(qp->ndev, QP_LINKS,
2095                             val & ~BIT(qp->qp_num));
2096
2097         if (qp->link_is_up)
2098                 ntb_send_link_down(qp);
2099         else
2100                 cancel_delayed_work_sync(&qp->link_work);
2101 }
2102 EXPORT_SYMBOL_GPL(ntb_transport_link_down);
2103
2104 /**
2105  * ntb_transport_link_query - Query transport link state
2106  * @qp: NTB transport layer queue to be queried
2107  *
2108  * Query connectivity to the remote system of the NTB transport queue
2109  *
2110  * RETURNS: true for link up or false for link down
2111  */
2112 bool ntb_transport_link_query(struct ntb_transport_qp *qp)
2113 {
2114         if (!qp)
2115                 return false;
2116
2117         return qp->link_is_up;
2118 }
2119 EXPORT_SYMBOL_GPL(ntb_transport_link_query);
2120
2121 /**
2122  * ntb_transport_qp_num - Query the qp number
2123  * @qp: NTB transport layer queue to be queried
2124  *
2125  * Query qp number of the NTB transport queue
2126  *
2127  * RETURNS: a zero based number specifying the qp number
2128  */
2129 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
2130 {
2131         if (!qp)
2132                 return 0;
2133
2134         return qp->qp_num;
2135 }
2136 EXPORT_SYMBOL_GPL(ntb_transport_qp_num);
2137
2138 /**
2139  * ntb_transport_max_size - Query the max payload size of a qp
2140  * @qp: NTB transport layer queue to be queried
2141  *
2142  * Query the maximum payload size permissible on the given qp
2143  *
2144  * RETURNS: the max payload size of a qp
2145  */
2146 unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp)
2147 {
2148         unsigned int max_size;
2149         unsigned int copy_align;
2150         struct dma_chan *rx_chan, *tx_chan;
2151
2152         if (!qp)
2153                 return 0;
2154
2155         rx_chan = qp->rx_dma_chan;
2156         tx_chan = qp->tx_dma_chan;
2157
2158         copy_align = max(rx_chan ? rx_chan->device->copy_align : 0,
2159                          tx_chan ? tx_chan->device->copy_align : 0);
2160
2161         /* If DMA engine usage is possible, try to find the max size for that */
2162         max_size = qp->tx_max_frame - sizeof(struct ntb_payload_header);
2163         max_size = round_down(max_size, 1 << copy_align);
2164
2165         return max_size;
2166 }
2167 EXPORT_SYMBOL_GPL(ntb_transport_max_size);
2168
2169 unsigned int ntb_transport_tx_free_entry(struct ntb_transport_qp *qp)
2170 {
2171         unsigned int head = qp->tx_index;
2172         unsigned int tail = qp->remote_rx_info->entry;
2173
2174         return tail > head ? tail - head : qp->tx_max_entry + tail - head;
2175 }
2176 EXPORT_SYMBOL_GPL(ntb_transport_tx_free_entry);
2177
2178 static void ntb_transport_doorbell_callback(void *data, int vector)
2179 {
2180         struct ntb_transport_ctx *nt = data;
2181         struct ntb_transport_qp *qp;
2182         u64 db_bits;
2183         unsigned int qp_num;
2184
2185         db_bits = (nt->qp_bitmap & ~nt->qp_bitmap_free &
2186                    ntb_db_vector_mask(nt->ndev, vector));
2187
2188         while (db_bits) {
2189                 qp_num = __ffs(db_bits);
2190                 qp = &nt->qp_vec[qp_num];
2191
2192                 if (qp->active)
2193                         tasklet_schedule(&qp->rxc_db_work);
2194
2195                 db_bits &= ~BIT_ULL(qp_num);
2196         }
2197 }
2198
2199 static const struct ntb_ctx_ops ntb_transport_ops = {
2200         .link_event = ntb_transport_event_callback,
2201         .db_event = ntb_transport_doorbell_callback,
2202 };
2203
2204 static struct ntb_client ntb_transport_client = {
2205         .ops = {
2206                 .probe = ntb_transport_probe,
2207                 .remove = ntb_transport_free,
2208         },
2209 };
2210
2211 static int __init ntb_transport_init(void)
2212 {
2213         int rc;
2214
2215         pr_info("%s, version %s\n", NTB_TRANSPORT_DESC, NTB_TRANSPORT_VER);
2216
2217         if (debugfs_initialized())
2218                 nt_debugfs_dir = debugfs_create_dir(KBUILD_MODNAME, NULL);
2219
2220         rc = bus_register(&ntb_transport_bus);
2221         if (rc)
2222                 goto err_bus;
2223
2224         rc = ntb_register_client(&ntb_transport_client);
2225         if (rc)
2226                 goto err_client;
2227
2228         return 0;
2229
2230 err_client:
2231         bus_unregister(&ntb_transport_bus);
2232 err_bus:
2233         debugfs_remove_recursive(nt_debugfs_dir);
2234         return rc;
2235 }
2236 module_init(ntb_transport_init);
2237
2238 static void __exit ntb_transport_exit(void)
2239 {
2240         ntb_unregister_client(&ntb_transport_client);
2241         bus_unregister(&ntb_transport_bus);
2242         debugfs_remove_recursive(nt_debugfs_dir);
2243 }
2244 module_exit(ntb_transport_exit);