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  *
   9  *   This program is free software; you can redistribute it and/or modify
  10  *   it under the terms of version 2 of the GNU General Public License as
  11  *   published by the Free Software Foundation.
  12  *
  13  *   BSD LICENSE
  14  *
  15  *   Copyright(c) 2012 Intel Corporation. All rights reserved.
  16  *
  17  *   Redistribution and use in source and binary forms, with or without
  18  *   modification, are permitted provided that the following conditions
  19  *   are met:
  20  *
  21  *     * Redistributions of source code must retain the above copyright
  22  *       notice, this list of conditions and the following disclaimer.
  23  *     * Redistributions in binary form must reproduce the above copy
  24  *       notice, this list of conditions and the following disclaimer in
  25  *       the documentation and/or other materials provided with the
  26  *       distribution.
  27  *     * Neither the name of Intel Corporation nor the names of its
  28  *       contributors may be used to endorse or promote products derived
  29  *       from this software without specific prior written permission.
  30  *
  31  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  32  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  33  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  34  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  35  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  36  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  37  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  38  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  39  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  40  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  41  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  42  *
  43  * Intel PCIe NTB Linux driver
  44  *
  45  * Contact Information:
  46  * Jon Mason <jon.mason@intel.com>
  47  */
  48 #include <linux/debugfs.h>
  49 #include <linux/delay.h>
  50 #include <linux/dmaengine.h>
  51 #include <linux/dma-mapping.h>
  52 #include <linux/errno.h>
  53 #include <linux/export.h>
  54 #include <linux/interrupt.h>
  55 #include <linux/module.h>
  56 #include <linux/pci.h>
  57 #include <linux/slab.h>
  58 #include <linux/types.h>
  59 #include <linux/ntb.h>
  60 #include "ntb_hw.h"
  61
  62 #define NTB_TRANSPORT_VERSION   3
  63
  64 static unsigned int transport_mtu = 0x401E;
  65 module_param(transport_mtu, uint, 0644);
  66 MODULE_PARM_DESC(transport_mtu, "Maximum size of NTB transport packets");
  67
  68 static unsigned char max_num_clients;
  69 module_param(max_num_clients, byte, 0644);
  70 MODULE_PARM_DESC(max_num_clients, "Maximum number of NTB transport clients");
  71
  72 static unsigned int copy_bytes = 1024;
  73 module_param(copy_bytes, uint, 0644);
  74 MODULE_PARM_DESC(copy_bytes, "Threshold under which NTB will use the CPU to copy instead of DMA");
  75
  76 struct ntb_queue_entry {
  77         /* ntb_queue list reference */
  78         struct list_head entry;
  79         /* pointers to data to be transfered */
  80         void *cb_data;
  81         void *buf;
  82         unsigned int len;
  83         unsigned int flags;
  84
  85         struct ntb_transport_qp *qp;
  86         union {
  87                 struct ntb_payload_header __iomem *tx_hdr;
  88                 struct ntb_payload_header *rx_hdr;
  89         };
  90         unsigned int index;
  91 };
  92
  93 struct ntb_rx_info {
  94         unsigned int entry;
  95 };
  96
  97 struct ntb_transport_qp {
  98         struct ntb_transport *transport;
  99         struct ntb_device *ndev;
 100         void *cb_data;
 101         struct dma_chan *dma_chan;
 102
 103         bool client_ready;
 104         bool qp_link;
 105         u8 qp_num;      /* Only 64 QP's are allowed.  0-63 */
 106
 107         struct ntb_rx_info __iomem *rx_info;
 108         struct ntb_rx_info *remote_rx_info;
 109
 110         void (*tx_handler) (struct ntb_transport_qp *qp, void *qp_data,
 111                             void *data, int len);
 112         struct list_head tx_free_q;
 113         spinlock_t ntb_tx_free_q_lock;
 114         void __iomem *tx_mw;
 115         dma_addr_t tx_mw_phys;
 116         unsigned int tx_index;
 117         unsigned int tx_max_entry;
 118         unsigned int tx_max_frame;
 119
 120         void (*rx_handler) (struct ntb_transport_qp *qp, void *qp_data,
 121                             void *data, int len);
 122         struct tasklet_struct rx_work;
 123         struct list_head rx_pend_q;
 124         struct list_head rx_free_q;
 125         spinlock_t ntb_rx_pend_q_lock;
 126         spinlock_t ntb_rx_free_q_lock;
 127         void *rx_buff;
 128         unsigned int rx_index;
 129         unsigned int rx_max_entry;
 130         unsigned int rx_max_frame;
 131         dma_cookie_t last_cookie;
 132
 133         void (*event_handler) (void *data, int status);
 134         struct delayed_work link_work;
 135         struct work_struct link_cleanup;
 136
 137         struct dentry *debugfs_dir;
 138         struct dentry *debugfs_stats;
 139
 140         /* Stats */
 141         u64 rx_bytes;
 142         u64 rx_pkts;
 143         u64 rx_ring_empty;
 144         u64 rx_err_no_buf;
 145         u64 rx_err_oflow;
 146         u64 rx_err_ver;
 147         u64 rx_memcpy;
 148         u64 rx_async;
 149         u64 tx_bytes;
 150         u64 tx_pkts;
 151         u64 tx_ring_full;
 152         u64 tx_err_no_buf;
 153         u64 tx_memcpy;
 154         u64 tx_async;
 155 };
 156
 157 struct ntb_transport_mw {
 158         size_t size;
 159         void *virt_addr;
 160         dma_addr_t dma_addr;
 161 };
 162
 163 struct ntb_transport_client_dev {
 164         struct list_head entry;
 165         struct device dev;
 166 };
 167
 168 struct ntb_transport {
 169         struct list_head entry;
 170         struct list_head client_devs;
 171
 172         struct ntb_device *ndev;
 173         struct ntb_transport_mw *mw;
 174         struct ntb_transport_qp *qps;
 175         unsigned int max_qps;
 176         unsigned long qp_bitmap;
 177         bool transport_link;
 178         struct delayed_work link_work;
 179         struct work_struct link_cleanup;
 180 };
 181
 182 enum {
 183         DESC_DONE_FLAG = 1 << 0,
 184         LINK_DOWN_FLAG = 1 << 1,
 185 };
 186
 187 struct ntb_payload_header {
 188         unsigned int ver;
 189         unsigned int len;
 190         unsigned int flags;
 191 };
 192
 193 enum {
 194         VERSION = 0,
 195         QP_LINKS,
 196         NUM_QPS,
 197         NUM_MWS,
 198         MW0_SZ_HIGH,
 199         MW0_SZ_LOW,
 200         MW1_SZ_HIGH,
 201         MW1_SZ_LOW,
 202         MAX_SPAD,
 203 };
 204
 205 #define QP_TO_MW(ndev, qp)      ((qp) % ntb_max_mw(ndev))
 206 #define NTB_QP_DEF_NUM_ENTRIES  100
 207 #define NTB_LINK_DOWN_TIMEOUT   10
 208
 209 static int ntb_match_bus(struct device *dev, struct device_driver *drv)
 210 {
 211         return !strncmp(dev_name(dev), drv->name, strlen(drv->name));
 212 }
 213
 214 static int ntb_client_probe(struct device *dev)
 215 {
 216         const struct ntb_client *drv = container_of(dev->driver,
 217                                                     struct ntb_client, driver);
 218         struct pci_dev *pdev = container_of(dev->parent, struct pci_dev, dev);
 219         int rc = -EINVAL;
 220
 221         get_device(dev);
 222         if (drv && drv->probe)
 223                 rc = drv->probe(pdev);
 224         if (rc)
 225                 put_device(dev);
 226
 227         return rc;
 228 }
 229
 230 static int ntb_client_remove(struct device *dev)
 231 {
 232         const struct ntb_client *drv = container_of(dev->driver,
 233                                                     struct ntb_client, driver);
 234         struct pci_dev *pdev = container_of(dev->parent, struct pci_dev, dev);
 235
 236         if (drv && drv->remove)
 237                 drv->remove(pdev);
 238
 239         put_device(dev);
 240
 241         return 0;
 242 }
 243
 244 static struct bus_type ntb_bus_type = {
 245         .name = "ntb_bus",
 246         .match = ntb_match_bus,
 247         .probe = ntb_client_probe,
 248         .remove = ntb_client_remove,
 249 };
 250
 251 static LIST_HEAD(ntb_transport_list);
 252
 253 static int ntb_bus_init(struct ntb_transport *nt)
 254 {
 255         if (list_empty(&ntb_transport_list)) {
 256                 int rc = bus_register(&ntb_bus_type);
 257                 if (rc)
 258                         return rc;
 259         }
 260
 261         list_add(&nt->entry, &ntb_transport_list);
 262
 263         return 0;
 264 }
 265
 266 static void ntb_bus_remove(struct ntb_transport *nt)
 267 {
 268         struct ntb_transport_client_dev *client_dev, *cd;
 269
 270         list_for_each_entry_safe(client_dev, cd, &nt->client_devs, entry) {
 271                 dev_err(client_dev->dev.parent, "%s still attached to bus, removing\n",
 272                         dev_name(&client_dev->dev));
 273                 list_del(&client_dev->entry);
 274                 device_unregister(&client_dev->dev);
 275         }
 276
 277         list_del(&nt->entry);
 278
 279         if (list_empty(&ntb_transport_list))
 280                 bus_unregister(&ntb_bus_type);
 281 }
 282
 283 static void ntb_client_release(struct device *dev)
 284 {
 285         struct ntb_transport_client_dev *client_dev;
 286         client_dev = container_of(dev, struct ntb_transport_client_dev, dev);
 287
 288         kfree(client_dev);
 289 }
 290
 291 /**
 292  * ntb_unregister_client_dev - Unregister NTB client device
 293  * @device_name: Name of NTB client device
 294  *
 295  * Unregister an NTB client device with the NTB transport layer
 296  */
 297 void ntb_unregister_client_dev(char *device_name)
 298 {
 299         struct ntb_transport_client_dev *client, *cd;
 300         struct ntb_transport *nt;
 301
 302         list_for_each_entry(nt, &ntb_transport_list, entry)
 303                 list_for_each_entry_safe(client, cd, &nt->client_devs, entry)
 304                         if (!strncmp(dev_name(&client->dev), device_name,
 305                                      strlen(device_name))) {
 306                                 list_del(&client->entry);
 307                                 device_unregister(&client->dev);
 308                         }
 309 }
 310 EXPORT_SYMBOL_GPL(ntb_unregister_client_dev);
 311
 312 /**
 313  * ntb_register_client_dev - Register NTB client device
 314  * @device_name: Name of NTB client device
 315  *
 316  * Register an NTB client device with the NTB transport layer
 317  */
 318 int ntb_register_client_dev(char *device_name)
 319 {
 320         struct ntb_transport_client_dev *client_dev;
 321         struct ntb_transport *nt;
 322         int rc, i = 0;
 323
 324         if (list_empty(&ntb_transport_list))
 325                 return -ENODEV;
 326
 327         list_for_each_entry(nt, &ntb_transport_list, entry) {
 328                 struct device *dev;
 329
 330                 client_dev = kzalloc(sizeof(struct ntb_transport_client_dev),
 331                                      GFP_KERNEL);
 332                 if (!client_dev) {
 333                         rc = -ENOMEM;
 334                         goto err;
 335                 }
 336
 337                 dev = &client_dev->dev;
 338
 339                 /* setup and register client devices */
 340                 dev_set_name(dev, "%s%d", device_name, i);
 341                 dev->bus = &ntb_bus_type;
 342                 dev->release = ntb_client_release;
 343                 dev->parent = &ntb_query_pdev(nt->ndev)->dev;
 344
 345                 rc = device_register(dev);
 346                 if (rc) {
 347                         kfree(client_dev);
 348                         goto err;
 349                 }
 350
 351                 list_add_tail(&client_dev->entry, &nt->client_devs);
 352                 i++;
 353         }
 354
 355         return 0;
 356
 357 err:
 358         ntb_unregister_client_dev(device_name);
 359
 360         return rc;
 361 }
 362 EXPORT_SYMBOL_GPL(ntb_register_client_dev);
 363
 364 /**
 365  * ntb_register_client - Register NTB client driver
 366  * @drv: NTB client driver to be registered
 367  *
 368  * Register an NTB client driver with the NTB transport layer
 369  *
 370  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 371  */
 372 int ntb_register_client(struct ntb_client *drv)
 373 {
 374         drv->driver.bus = &ntb_bus_type;
 375
 376         if (list_empty(&ntb_transport_list))
 377                 return -ENODEV;
 378
 379         return driver_register(&drv->driver);
 380 }
 381 EXPORT_SYMBOL_GPL(ntb_register_client);
 382
 383 /**
 384  * ntb_unregister_client - Unregister NTB client driver
 385  * @drv: NTB client driver to be unregistered
 386  *
 387  * Unregister an NTB client driver with the NTB transport layer
 388  *
 389  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 390  */
 391 void ntb_unregister_client(struct ntb_client *drv)
 392 {
 393         driver_unregister(&drv->driver);
 394 }
 395 EXPORT_SYMBOL_GPL(ntb_unregister_client);
 396
 397 static ssize_t debugfs_read(struct file *filp, char __user *ubuf, size_t count,
 398                             loff_t *offp)
 399 {
 400         struct ntb_transport_qp *qp;
 401         char *buf;
 402         ssize_t ret, out_offset, out_count;
 403
 404         out_count = 1000;
 405
 406         buf = kmalloc(out_count, GFP_KERNEL);
 407         if (!buf)
 408                 return -ENOMEM;
 409
 410         qp = filp->private_data;
 411         out_offset = 0;
 412         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 413                                "NTB QP stats\n");
 414         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 415                                "rx_bytes - \t%llu\n", qp->rx_bytes);
 416         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 417                                "rx_pkts - \t%llu\n", qp->rx_pkts);
 418         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 419                                "rx_memcpy - \t%llu\n", qp->rx_memcpy);
 420         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 421                                "rx_async - \t%llu\n", qp->rx_async);
 422         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 423                                "rx_ring_empty - %llu\n", qp->rx_ring_empty);
 424         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 425                                "rx_err_no_buf - %llu\n", qp->rx_err_no_buf);
 426         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 427                                "rx_err_oflow - \t%llu\n", qp->rx_err_oflow);
 428         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 429                                "rx_err_ver - \t%llu\n", qp->rx_err_ver);
 430         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 431                                "rx_buff - \t%p\n", qp->rx_buff);
 432         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 433                                "rx_index - \t%u\n", qp->rx_index);
 434         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 435                                "rx_max_entry - \t%u\n", qp->rx_max_entry);
 436
 437         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 438                                "tx_bytes - \t%llu\n", qp->tx_bytes);
 439         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 440                                "tx_pkts - \t%llu\n", qp->tx_pkts);
 441         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 442                                "tx_memcpy - \t%llu\n", qp->tx_memcpy);
 443         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 444                                "tx_async - \t%llu\n", qp->tx_async);
 445         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 446                                "tx_ring_full - \t%llu\n", qp->tx_ring_full);
 447         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 448                                "tx_err_no_buf - %llu\n", qp->tx_err_no_buf);
 449         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 450                                "tx_mw - \t%p\n", qp->tx_mw);
 451         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 452                                "tx_index - \t%u\n", qp->tx_index);
 453         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 454                                "tx_max_entry - \t%u\n", qp->tx_max_entry);
 455
 456         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 457                                "\nQP Link %s\n", (qp->qp_link == NTB_LINK_UP) ?
 458                                "Up" : "Down");
 459         if (out_offset > out_count)
 460                 out_offset = out_count;
 461
 462         ret = simple_read_from_buffer(ubuf, count, offp, buf, out_offset);
 463         kfree(buf);
 464         return ret;
 465 }
 466
 467 static const struct file_operations ntb_qp_debugfs_stats = {
 468         .owner = THIS_MODULE,
 469         .open = simple_open,
 470         .read = debugfs_read,
 471 };
 472
 473 static void ntb_list_add(spinlock_t *lock, struct list_head *entry,
 474                          struct list_head *list)
 475 {
 476         unsigned long flags;
 477
 478         spin_lock_irqsave(lock, flags);
 479         list_add_tail(entry, list);
 480         spin_unlock_irqrestore(lock, flags);
 481 }
 482
 483 static struct ntb_queue_entry *ntb_list_rm(spinlock_t *lock,
 484                                                 struct list_head *list)
 485 {
 486         struct ntb_queue_entry *entry;
 487         unsigned long flags;
 488
 489         spin_lock_irqsave(lock, flags);
 490         if (list_empty(list)) {
 491                 entry = NULL;
 492                 goto out;
 493         }
 494         entry = list_first_entry(list, struct ntb_queue_entry, entry);
 495         list_del(&entry->entry);
 496 out:
 497         spin_unlock_irqrestore(lock, flags);
 498
 499         return entry;
 500 }
 501
 502 static void ntb_transport_setup_qp_mw(struct ntb_transport *nt,
 503                                       unsigned int qp_num)
 504 {
 505         struct ntb_transport_qp *qp = &nt->qps[qp_num];
 506         unsigned int rx_size, num_qps_mw;
 507         u8 mw_num, mw_max;
 508         unsigned int i;
 509
 510         mw_max = ntb_max_mw(nt->ndev);
 511         mw_num = QP_TO_MW(nt->ndev, qp_num);
 512
 513         WARN_ON(nt->mw[mw_num].virt_addr == NULL);
 514
 515         if (nt->max_qps % mw_max && mw_num < nt->max_qps % mw_max)
 516                 num_qps_mw = nt->max_qps / mw_max + 1;
 517         else
 518                 num_qps_mw = nt->max_qps / mw_max;
 519
 520         rx_size = (unsigned int) nt->mw[mw_num].size / num_qps_mw;
 521         qp->rx_buff = nt->mw[mw_num].virt_addr + qp_num / mw_max * rx_size;
 522         rx_size -= sizeof(struct ntb_rx_info);
 523
 524         qp->remote_rx_info = qp->rx_buff + rx_size;
 525
 526         /* Due to housekeeping, there must be atleast 2 buffs */
 527         qp->rx_max_frame = min(transport_mtu, rx_size / 2);
 528         qp->rx_max_entry = rx_size / qp->rx_max_frame;
 529         qp->rx_index = 0;
 530
 531         qp->remote_rx_info->entry = qp->rx_max_entry - 1;
 532
 533         /* setup the hdr offsets with 0's */
 534         for (i = 0; i < qp->rx_max_entry; i++) {
 535                 void *offset = qp->rx_buff + qp->rx_max_frame * (i + 1) -
 536                                sizeof(struct ntb_payload_header);
 537                 memset(offset, 0, sizeof(struct ntb_payload_header));
 538         }
 539
 540         qp->rx_pkts = 0;
 541         qp->tx_pkts = 0;
 542         qp->tx_index = 0;
 543 }
 544
 545 static void ntb_free_mw(struct ntb_transport *nt, int num_mw)
 546 {
 547         struct ntb_transport_mw *mw = &nt->mw[num_mw];
 548         struct pci_dev *pdev = ntb_query_pdev(nt->ndev);
 549
 550         if (!mw->virt_addr)
 551                 return;
 552
 553         dma_free_coherent(&pdev->dev, mw->size, mw->virt_addr, mw->dma_addr);
 554         mw->virt_addr = NULL;
 555 }
 556
 557 static int ntb_set_mw(struct ntb_transport *nt, int num_mw, unsigned int size)
 558 {
 559         struct ntb_transport_mw *mw = &nt->mw[num_mw];
 560         struct pci_dev *pdev = ntb_query_pdev(nt->ndev);
 561
 562         /* No need to re-setup */
 563         if (mw->size == ALIGN(size, 4096))
 564                 return 0;
 565
 566         if (mw->size != 0)
 567                 ntb_free_mw(nt, num_mw);
 568
 569         /* Alloc memory for receiving data.  Must be 4k aligned */
 570         mw->size = ALIGN(size, 4096);
 571
 572         mw->virt_addr = dma_alloc_coherent(&pdev->dev, mw->size, &mw->dma_addr,
 573                                            GFP_KERNEL);
 574         if (!mw->virt_addr) {
 575                 mw->size = 0;
 576                 dev_err(&pdev->dev, "Unable to allocate MW buffer of size %d\n",
 577                        (int) mw->size);
 578                 return -ENOMEM;
 579         }
 580
 581         /* Notify HW the memory location of the receive buffer */
 582         ntb_set_mw_addr(nt->ndev, num_mw, mw->dma_addr);
 583
 584         return 0;
 585 }
 586
 587 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
 588 {
 589         struct ntb_transport *nt = qp->transport;
 590         struct pci_dev *pdev = ntb_query_pdev(nt->ndev);
 591
 592         if (qp->qp_link == NTB_LINK_DOWN) {
 593                 cancel_delayed_work_sync(&qp->link_work);
 594                 return;
 595         }
 596
 597         if (qp->event_handler)
 598                 qp->event_handler(qp->cb_data, NTB_LINK_DOWN);
 599
 600         dev_info(&pdev->dev, "qp %d: Link Down\n", qp->qp_num);
 601         qp->qp_link = NTB_LINK_DOWN;
 602 }
 603
 604 static void ntb_qp_link_cleanup_work(struct work_struct *work)
 605 {
 606         struct ntb_transport_qp *qp = container_of(work,
 607                                                    struct ntb_transport_qp,
 608                                                    link_cleanup);
 609         struct ntb_transport *nt = qp->transport;
 610
 611         ntb_qp_link_cleanup(qp);
 612
 613         if (nt->transport_link == NTB_LINK_UP)
 614                 schedule_delayed_work(&qp->link_work,
 615                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 616 }
 617
 618 static void ntb_qp_link_down(struct ntb_transport_qp *qp)
 619 {
 620         schedule_work(&qp->link_cleanup);
 621 }
 622
 623 static void ntb_transport_link_cleanup(struct ntb_transport *nt)
 624 {
 625         int i;
 626
 627         /* Pass along the info to any clients */
 628         for (i = 0; i < nt->max_qps; i++)
 629                 if (!test_bit(i, &nt->qp_bitmap))
 630                         ntb_qp_link_cleanup(&nt->qps[i]);
 631
 632         if (nt->transport_link == NTB_LINK_DOWN)
 633                 cancel_delayed_work_sync(&nt->link_work);
 634         else
 635                 nt->transport_link = NTB_LINK_DOWN;
 636
 637         /* The scratchpad registers keep the values if the remote side
 638          * goes down, blast them now to give them a sane value the next
 639          * time they are accessed
 640          */
 641         for (i = 0; i < MAX_SPAD; i++)
 642                 ntb_write_local_spad(nt->ndev, i, 0);
 643 }
 644
 645 static void ntb_transport_link_cleanup_work(struct work_struct *work)
 646 {
 647         struct ntb_transport *nt = container_of(work, struct ntb_transport,
 648                                                 link_cleanup);
 649
 650         ntb_transport_link_cleanup(nt);
 651 }
 652
 653 static void ntb_transport_event_callback(void *data, enum ntb_hw_event event)
 654 {
 655         struct ntb_transport *nt = data;
 656
 657         switch (event) {
 658         case NTB_EVENT_HW_LINK_UP:
 659                 schedule_delayed_work(&nt->link_work, 0);
 660                 break;
 661         case NTB_EVENT_HW_LINK_DOWN:
 662                 schedule_work(&nt->link_cleanup);
 663                 break;
 664         default:
 665                 BUG();
 666         }
 667 }
 668
 669 static void ntb_transport_link_work(struct work_struct *work)
 670 {
 671         struct ntb_transport *nt = container_of(work, struct ntb_transport,
 672                                                 link_work.work);
 673         struct ntb_device *ndev = nt->ndev;
 674         struct pci_dev *pdev = ntb_query_pdev(ndev);
 675         u32 val;
 676         int rc, i;
 677
 678         /* send the local info, in the opposite order of the way we read it */
 679         for (i = 0; i < ntb_max_mw(ndev); i++) {
 680                 rc = ntb_write_remote_spad(ndev, MW0_SZ_HIGH + (i * 2),
 681                                            ntb_get_mw_size(ndev, i) >> 32);
 682                 if (rc) {
 683                         dev_err(&pdev->dev, "Error writing %u to remote spad %d\n",
 684                                 (u32)(ntb_get_mw_size(ndev, i) >> 32),
 685                                 MW0_SZ_HIGH + (i * 2));
 686                         goto out;
 687                 }
 688
 689                 rc = ntb_write_remote_spad(ndev, MW0_SZ_LOW + (i * 2),
 690                                            (u32) ntb_get_mw_size(ndev, i));
 691                 if (rc) {
 692                         dev_err(&pdev->dev, "Error writing %u to remote spad %d\n",
 693                                 (u32) ntb_get_mw_size(ndev, i),
 694                                 MW0_SZ_LOW + (i * 2));
 695                         goto out;
 696                 }
 697         }
 698
 699         rc = ntb_write_remote_spad(ndev, NUM_MWS, ntb_max_mw(ndev));
 700         if (rc) {
 701                 dev_err(&pdev->dev, "Error writing %x to remote spad %d\n",
 702                         ntb_max_mw(ndev), NUM_MWS);
 703                 goto out;
 704         }
 705
 706         rc = ntb_write_remote_spad(ndev, NUM_QPS, nt->max_qps);
 707         if (rc) {
 708                 dev_err(&pdev->dev, "Error writing %x to remote spad %d\n",
 709                         nt->max_qps, NUM_QPS);
 710                 goto out;
 711         }
 712
 713         rc = ntb_write_remote_spad(ndev, VERSION, NTB_TRANSPORT_VERSION);
 714         if (rc) {
 715                 dev_err(&pdev->dev, "Error writing %x to remote spad %d\n",
 716                         NTB_TRANSPORT_VERSION, VERSION);
 717                 goto out;
 718         }
 719
 720         /* Query the remote side for its info */
 721         rc = ntb_read_remote_spad(ndev, VERSION, &val);
 722         if (rc) {
 723                 dev_err(&pdev->dev, "Error reading remote spad %d\n", VERSION);
 724                 goto out;
 725         }
 726
 727         if (val != NTB_TRANSPORT_VERSION)
 728                 goto out;
 729         dev_dbg(&pdev->dev, "Remote version = %d\n", val);
 730
 731         rc = ntb_read_remote_spad(ndev, NUM_QPS, &val);
 732         if (rc) {
 733                 dev_err(&pdev->dev, "Error reading remote spad %d\n", NUM_QPS);
 734                 goto out;
 735         }
 736
 737         if (val != nt->max_qps)
 738                 goto out;
 739         dev_dbg(&pdev->dev, "Remote max number of qps = %d\n", val);
 740
 741         rc = ntb_read_remote_spad(ndev, NUM_MWS, &val);
 742         if (rc) {
 743                 dev_err(&pdev->dev, "Error reading remote spad %d\n", NUM_MWS);
 744                 goto out;
 745         }
 746
 747         if (val != ntb_max_mw(ndev))
 748                 goto out;
 749         dev_dbg(&pdev->dev, "Remote number of mws = %d\n", val);
 750
 751         for (i = 0; i < ntb_max_mw(ndev); i++) {
 752                 u64 val64;
 753
 754                 rc = ntb_read_remote_spad(ndev, MW0_SZ_HIGH + (i * 2), &val);
 755                 if (rc) {
 756                         dev_err(&pdev->dev, "Error reading remote spad %d\n",
 757                                 MW0_SZ_HIGH + (i * 2));
 758                         goto out1;
 759                 }
 760
 761                 val64 = (u64) val << 32;
 762
 763                 rc = ntb_read_remote_spad(ndev, MW0_SZ_LOW + (i * 2), &val);
 764                 if (rc) {
 765                         dev_err(&pdev->dev, "Error reading remote spad %d\n",
 766                                 MW0_SZ_LOW + (i * 2));
 767                         goto out1;
 768                 }
 769
 770                 val64 |= val;
 771
 772                 dev_dbg(&pdev->dev, "Remote MW%d size = %llu\n", i, val64);
 773
 774                 rc = ntb_set_mw(nt, i, val64);
 775                 if (rc)
 776                         goto out1;
 777         }
 778
 779         nt->transport_link = NTB_LINK_UP;
 780
 781         for (i = 0; i < nt->max_qps; i++) {
 782                 struct ntb_transport_qp *qp = &nt->qps[i];
 783
 784                 ntb_transport_setup_qp_mw(nt, i);
 785
 786                 if (qp->client_ready == NTB_LINK_UP)
 787                         schedule_delayed_work(&qp->link_work, 0);
 788         }
 789
 790         return;
 791
 792 out1:
 793         for (i = 0; i < ntb_max_mw(ndev); i++)
 794                 ntb_free_mw(nt, i);
 795 out:
 796         if (ntb_hw_link_status(ndev))
 797                 schedule_delayed_work(&nt->link_work,
 798                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 799 }
 800
 801 static void ntb_qp_link_work(struct work_struct *work)
 802 {
 803         struct ntb_transport_qp *qp = container_of(work,
 804                                                    struct ntb_transport_qp,
 805                                                    link_work.work);
 806         struct pci_dev *pdev = ntb_query_pdev(qp->ndev);
 807         struct ntb_transport *nt = qp->transport;
 808         int rc, val;
 809
 810         WARN_ON(nt->transport_link != NTB_LINK_UP);
 811
 812         rc = ntb_read_local_spad(nt->ndev, QP_LINKS, &val);
 813         if (rc) {
 814                 dev_err(&pdev->dev, "Error reading spad %d\n", QP_LINKS);
 815                 return;
 816         }
 817
 818         rc = ntb_write_remote_spad(nt->ndev, QP_LINKS, val | 1 << qp->qp_num);
 819         if (rc)
 820                 dev_err(&pdev->dev, "Error writing %x to remote spad %d\n",
 821                         val | 1 << qp->qp_num, QP_LINKS);
 822
 823         /* query remote spad for qp ready bits */
 824         rc = ntb_read_remote_spad(nt->ndev, QP_LINKS, &val);
 825         if (rc)
 826                 dev_err(&pdev->dev, "Error reading remote spad %d\n", QP_LINKS);
 827
 828         dev_dbg(&pdev->dev, "Remote QP link status = %x\n", val);
 829
 830         /* See if the remote side is up */
 831         if (1 << qp->qp_num & val) {
 832                 qp->qp_link = NTB_LINK_UP;
 833
 834                 dev_info(&pdev->dev, "qp %d: Link Up\n", qp->qp_num);
 835                 if (qp->event_handler)
 836                         qp->event_handler(qp->cb_data, NTB_LINK_UP);
 837         } else if (nt->transport_link == NTB_LINK_UP)
 838                 schedule_delayed_work(&qp->link_work,
 839                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 840 }
 841
 842 static int ntb_transport_init_queue(struct ntb_transport *nt,
 843                                      unsigned int qp_num)
 844 {
 845         struct ntb_transport_qp *qp;
 846         unsigned int num_qps_mw, tx_size;
 847         u8 mw_num, mw_max;
 848         u64 qp_offset;
 849
 850         mw_max = ntb_max_mw(nt->ndev);
 851         mw_num = QP_TO_MW(nt->ndev, qp_num);
 852
 853         qp = &nt->qps[qp_num];
 854         qp->qp_num = qp_num;
 855         qp->transport = nt;
 856         qp->ndev = nt->ndev;
 857         qp->qp_link = NTB_LINK_DOWN;
 858         qp->client_ready = NTB_LINK_DOWN;
 859         qp->event_handler = NULL;
 860
 861         if (nt->max_qps % mw_max && mw_num < nt->max_qps % mw_max)
 862                 num_qps_mw = nt->max_qps / mw_max + 1;
 863         else
 864                 num_qps_mw = nt->max_qps / mw_max;
 865
 866         tx_size = (unsigned int) ntb_get_mw_size(qp->ndev, mw_num) / num_qps_mw;
 867         qp_offset = qp_num / mw_max * tx_size;
 868         qp->tx_mw = ntb_get_mw_vbase(nt->ndev, mw_num) + qp_offset;
 869         if (!qp->tx_mw)
 870                 return -EINVAL;
 871
 872         qp->tx_mw_phys = ntb_get_mw_base(qp->ndev, mw_num) + qp_offset;
 873         if (!qp->tx_mw_phys)
 874                 return -EINVAL;
 875
 876         tx_size -= sizeof(struct ntb_rx_info);
 877         qp->rx_info = qp->tx_mw + tx_size;
 878
 879         /* Due to housekeeping, there must be atleast 2 buffs */
 880         qp->tx_max_frame = min(transport_mtu, tx_size / 2);
 881         qp->tx_max_entry = tx_size / qp->tx_max_frame;
 882
 883         if (ntb_query_debugfs(nt->ndev)) {
 884                 char debugfs_name[4];
 885
 886                 snprintf(debugfs_name, 4, "qp%d", qp_num);
 887                 qp->debugfs_dir = debugfs_create_dir(debugfs_name,
 888                                                  ntb_query_debugfs(nt->ndev));
 889
 890                 qp->debugfs_stats = debugfs_create_file("stats", S_IRUSR,
 891                                                         qp->debugfs_dir, qp,
 892                                                         &ntb_qp_debugfs_stats);
 893         }
 894
 895         INIT_DELAYED_WORK(&qp->link_work, ntb_qp_link_work);
 896         INIT_WORK(&qp->link_cleanup, ntb_qp_link_cleanup_work);
 897
 898         spin_lock_init(&qp->ntb_rx_pend_q_lock);
 899         spin_lock_init(&qp->ntb_rx_free_q_lock);
 900         spin_lock_init(&qp->ntb_tx_free_q_lock);
 901
 902         INIT_LIST_HEAD(&qp->rx_pend_q);
 903         INIT_LIST_HEAD(&qp->rx_free_q);
 904         INIT_LIST_HEAD(&qp->tx_free_q);
 905
 906         return 0;
 907 }
 908
 909 int ntb_transport_init(struct pci_dev *pdev)
 910 {
 911         struct ntb_transport *nt;
 912         int rc, i;
 913
 914         nt = kzalloc(sizeof(struct ntb_transport), GFP_KERNEL);
 915         if (!nt)
 916                 return -ENOMEM;
 917
 918         nt->ndev = ntb_register_transport(pdev, nt);
 919         if (!nt->ndev) {
 920                 rc = -EIO;
 921                 goto err;
 922         }
 923
 924         nt->mw = kcalloc(ntb_max_mw(nt->ndev), sizeof(struct ntb_transport_mw),
 925                          GFP_KERNEL);
 926         if (!nt->mw) {
 927                 rc = -ENOMEM;
 928                 goto err1;
 929         }
 930
 931         if (max_num_clients)
 932                 nt->max_qps = min(ntb_max_cbs(nt->ndev), max_num_clients);
 933         else
 934                 nt->max_qps = min(ntb_max_cbs(nt->ndev), ntb_max_mw(nt->ndev));
 935
 936         nt->qps = kcalloc(nt->max_qps, sizeof(struct ntb_transport_qp),
 937                           GFP_KERNEL);
 938         if (!nt->qps) {
 939                 rc = -ENOMEM;
 940                 goto err2;
 941         }
 942
 943         nt->qp_bitmap = ((u64) 1 << nt->max_qps) - 1;
 944
 945         for (i = 0; i < nt->max_qps; i++) {
 946                 rc = ntb_transport_init_queue(nt, i);
 947                 if (rc)
 948                         goto err3;
 949         }
 950
 951         INIT_DELAYED_WORK(&nt->link_work, ntb_transport_link_work);
 952         INIT_WORK(&nt->link_cleanup, ntb_transport_link_cleanup_work);
 953
 954         rc = ntb_register_event_callback(nt->ndev,
 955                                          ntb_transport_event_callback);
 956         if (rc)
 957                 goto err3;
 958
 959         INIT_LIST_HEAD(&nt->client_devs);
 960         rc = ntb_bus_init(nt);
 961         if (rc)
 962                 goto err4;
 963
 964         if (ntb_hw_link_status(nt->ndev))
 965                 schedule_delayed_work(&nt->link_work, 0);
 966
 967         return 0;
 968
 969 err4:
 970         ntb_unregister_event_callback(nt->ndev);
 971 err3:
 972         kfree(nt->qps);
 973 err2:
 974         kfree(nt->mw);
 975 err1:
 976         ntb_unregister_transport(nt->ndev);
 977 err:
 978         kfree(nt);
 979         return rc;
 980 }
 981
 982 void ntb_transport_free(void *transport)
 983 {
 984         struct ntb_transport *nt = transport;
 985         struct ntb_device *ndev = nt->ndev;
 986         int i;
 987
 988         ntb_transport_link_cleanup(nt);
 989
 990         /* verify that all the qp's are freed */
 991         for (i = 0; i < nt->max_qps; i++) {
 992                 if (!test_bit(i, &nt->qp_bitmap))
 993                         ntb_transport_free_queue(&nt->qps[i]);
 994                 debugfs_remove_recursive(nt->qps[i].debugfs_dir);
 995         }
 996
 997         ntb_bus_remove(nt);
 998
 999         cancel_delayed_work_sync(&nt->link_work);
1000
1001         ntb_unregister_event_callback(ndev);
1002
1003         for (i = 0; i < ntb_max_mw(ndev); i++)
1004                 ntb_free_mw(nt, i);
1005
1006         kfree(nt->qps);
1007         kfree(nt->mw);
1008         ntb_unregister_transport(ndev);
1009         kfree(nt);
1010 }
1011
1012 static void ntb_rx_copy_callback(void *data)
1013 {
1014         struct ntb_queue_entry *entry = data;
1015         struct ntb_transport_qp *qp = entry->qp;
1016         void *cb_data = entry->cb_data;
1017         unsigned int len = entry->len;
1018         struct ntb_payload_header *hdr = entry->rx_hdr;
1019
1020         /* Ensure that the data is fully copied out before clearing the flag */
1021         wmb();
1022         hdr->flags = 0;
1023
1024         iowrite32(entry->index, &qp->rx_info->entry);
1025
1026         ntb_list_add(&qp->ntb_rx_free_q_lock, &entry->entry, &qp->rx_free_q);
1027
1028         if (qp->rx_handler && qp->client_ready == NTB_LINK_UP)
1029                 qp->rx_handler(qp, qp->cb_data, cb_data, len);
1030 }
1031
1032 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset)
1033 {
1034         void *buf = entry->buf;
1035         size_t len = entry->len;
1036
1037         memcpy(buf, offset, len);
1038
1039         ntb_rx_copy_callback(entry);
1040 }
1041
1042 static void ntb_async_rx(struct ntb_queue_entry *entry, void *offset,
1043                          size_t len)
1044 {
1045         struct dma_async_tx_descriptor *txd;
1046         struct ntb_transport_qp *qp = entry->qp;
1047         struct dma_chan *chan = qp->dma_chan;
1048         struct dma_device *device;
1049         size_t pay_off, buff_off;
1050         dma_addr_t src, dest;
1051         dma_cookie_t cookie;
1052         void *buf = entry->buf;
1053         unsigned long flags;
1054
1055         entry->len = len;
1056
1057         if (!chan)
1058                 goto err;
1059
1060         if (len < copy_bytes)
1061                 goto err1;
1062
1063         device = chan->device;
1064         pay_off = (size_t) offset & ~PAGE_MASK;
1065         buff_off = (size_t) buf & ~PAGE_MASK;
1066
1067         if (!is_dma_copy_aligned(device, pay_off, buff_off, len))
1068                 goto err1;
1069
1070         dest = dma_map_single(device->dev, buf, len, DMA_FROM_DEVICE);
1071         if (dma_mapping_error(device->dev, dest))
1072                 goto err1;
1073
1074         src = dma_map_single(device->dev, offset, len, DMA_TO_DEVICE);
1075         if (dma_mapping_error(device->dev, src))
1076                 goto err2;
1077
1078         flags = DMA_COMPL_DEST_UNMAP_SINGLE | DMA_COMPL_SRC_UNMAP_SINGLE |
1079                 DMA_PREP_INTERRUPT;
1080         txd = device->device_prep_dma_memcpy(chan, dest, src, len, flags);
1081         if (!txd)
1082                 goto err3;
1083
1084         txd->callback = ntb_rx_copy_callback;
1085         txd->callback_param = entry;
1086
1087         cookie = dmaengine_submit(txd);
1088         if (dma_submit_error(cookie))
1089                 goto err3;
1090
1091         qp->last_cookie = cookie;
1092
1093         qp->rx_async++;
1094
1095         return;
1096
1097 err3:
1098         dma_unmap_single(device->dev, src, len, DMA_TO_DEVICE);
1099 err2:
1100         dma_unmap_single(device->dev, dest, len, DMA_FROM_DEVICE);
1101 err1:
1102         /* If the callbacks come out of order, the writing of the index to the
1103          * last completed will be out of order.  This may result in the
1104          * receive stalling forever.
1105          */
1106         dma_sync_wait(chan, qp->last_cookie);
1107 err:
1108         ntb_memcpy_rx(entry, offset);
1109         qp->rx_memcpy++;
1110 }
1111
1112 static int ntb_process_rxc(struct ntb_transport_qp *qp)
1113 {
1114         struct ntb_payload_header *hdr;
1115         struct ntb_queue_entry *entry;
1116         void *offset;
1117
1118         offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
1119         hdr = offset + qp->rx_max_frame - sizeof(struct ntb_payload_header);
1120
1121         entry = ntb_list_rm(&qp->ntb_rx_pend_q_lock, &qp->rx_pend_q);
1122         if (!entry) {
1123                 dev_dbg(&ntb_query_pdev(qp->ndev)->dev,
1124                         "no buffer - HDR ver %u, len %d, flags %x\n",
1125                         hdr->ver, hdr->len, hdr->flags);
1126                 qp->rx_err_no_buf++;
1127                 return -ENOMEM;
1128         }
1129
1130         if (!(hdr->flags & DESC_DONE_FLAG)) {
1131                 ntb_list_add(&qp->ntb_rx_pend_q_lock, &entry->entry,
1132                              &qp->rx_pend_q);
1133                 qp->rx_ring_empty++;
1134                 return -EAGAIN;
1135         }
1136
1137         if (hdr->ver != (u32) qp->rx_pkts) {
1138                 dev_dbg(&ntb_query_pdev(qp->ndev)->dev,
1139                         "qp %d: version mismatch, expected %llu - got %u\n",
1140                         qp->qp_num, qp->rx_pkts, hdr->ver);
1141                 ntb_list_add(&qp->ntb_rx_pend_q_lock, &entry->entry,
1142                              &qp->rx_pend_q);
1143                 qp->rx_err_ver++;
1144                 return -EIO;
1145         }
1146
1147         if (hdr->flags & LINK_DOWN_FLAG) {
1148                 ntb_qp_link_down(qp);
1149
1150                 goto err;
1151         }
1152
1153         dev_dbg(&ntb_query_pdev(qp->ndev)->dev,
1154                 "rx offset %u, ver %u - %d payload received, buf size %d\n",
1155                 qp->rx_index, hdr->ver, hdr->len, entry->len);
1156
1157         qp->rx_bytes += hdr->len;
1158         qp->rx_pkts++;
1159
1160         if (hdr->len > entry->len) {
1161                 qp->rx_err_oflow++;
1162                 dev_dbg(&ntb_query_pdev(qp->ndev)->dev,
1163                         "RX overflow! Wanted %d got %d\n",
1164                         hdr->len, entry->len);
1165
1166                 goto err;
1167         }
1168
1169         entry->index = qp->rx_index;
1170         entry->rx_hdr = hdr;
1171
1172         ntb_async_rx(entry, offset, hdr->len);
1173
1174 out:
1175         qp->rx_index++;
1176         qp->rx_index %= qp->rx_max_entry;
1177
1178         return 0;
1179
1180 err:
1181         ntb_list_add(&qp->ntb_rx_pend_q_lock, &entry->entry,
1182                      &qp->rx_pend_q);
1183         /* Ensure that the data is fully copied out before clearing the flag */
1184         wmb();
1185         hdr->flags = 0;
1186         iowrite32(qp->rx_index, &qp->rx_info->entry);
1187
1188         goto out;
1189 }
1190
1191 static void ntb_transport_rx(unsigned long data)
1192 {
1193         struct ntb_transport_qp *qp = (struct ntb_transport_qp *)data;
1194         int rc, i;
1195
1196         /* Limit the number of packets processed in a single interrupt to
1197          * provide fairness to others
1198          */
1199         for (i = 0; i < qp->rx_max_entry; i++) {
1200                 rc = ntb_process_rxc(qp);
1201                 if (rc)
1202                         break;
1203         }
1204
1205         if (qp->dma_chan)
1206                 dma_async_issue_pending(qp->dma_chan);
1207 }
1208
1209 static void ntb_transport_rxc_db(void *data, int db_num)
1210 {
1211         struct ntb_transport_qp *qp = data;
1212
1213         dev_dbg(&ntb_query_pdev(qp->ndev)->dev, "%s: doorbell %d received\n",
1214                 __func__, db_num);
1215
1216         tasklet_schedule(&qp->rx_work);
1217 }
1218
1219 static void ntb_tx_copy_callback(void *data)
1220 {
1221         struct ntb_queue_entry *entry = data;
1222         struct ntb_transport_qp *qp = entry->qp;
1223         struct ntb_payload_header __iomem *hdr = entry->tx_hdr;
1224
1225         /* Ensure that the data is fully copied out before setting the flags */
1226         wmb();
1227         iowrite32(entry->flags | DESC_DONE_FLAG, &hdr->flags);
1228
1229         ntb_ring_doorbell(qp->ndev, qp->qp_num);
1230
1231         /* The entry length can only be zero if the packet is intended to be a
1232          * "link down" or similar.  Since no payload is being sent in these
1233          * cases, there is nothing to add to the completion queue.
1234          */
1235         if (entry->len > 0) {
1236                 qp->tx_bytes += entry->len;
1237
1238                 if (qp->tx_handler)
1239                         qp->tx_handler(qp, qp->cb_data, entry->cb_data,
1240                                        entry->len);
1241         }
1242
1243         ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry, &qp->tx_free_q);
1244 }
1245
1246 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset)
1247 {
1248         memcpy_toio(offset, entry->buf, entry->len);
1249
1250         ntb_tx_copy_callback(entry);
1251 }
1252
1253 static void ntb_async_tx(struct ntb_transport_qp *qp,
1254                          struct ntb_queue_entry *entry)
1255 {
1256         struct ntb_payload_header __iomem *hdr;
1257         struct dma_async_tx_descriptor *txd;
1258         struct dma_chan *chan = qp->dma_chan;
1259         struct dma_device *device;
1260         size_t dest_off, buff_off;
1261         dma_addr_t src, dest;
1262         dma_cookie_t cookie;
1263         void __iomem *offset;
1264         size_t len = entry->len;
1265         void *buf = entry->buf;
1266         unsigned long flags;
1267
1268         offset = qp->tx_mw + qp->tx_max_frame * qp->tx_index;
1269         hdr = offset + qp->tx_max_frame - sizeof(struct ntb_payload_header);
1270         entry->tx_hdr = hdr;
1271
1272         iowrite32(entry->len, &hdr->len);
1273         iowrite32((u32) qp->tx_pkts, &hdr->ver);
1274
1275         if (!chan)
1276                 goto err;
1277
1278         if (len < copy_bytes)
1279                 goto err;
1280
1281         device = chan->device;
1282         dest = qp->tx_mw_phys + qp->tx_max_frame * qp->tx_index;
1283         buff_off = (size_t) buf & ~PAGE_MASK;
1284         dest_off = (size_t) dest & ~PAGE_MASK;
1285
1286         if (!is_dma_copy_aligned(device, buff_off, dest_off, len))
1287                 goto err;
1288
1289         src = dma_map_single(device->dev, buf, len, DMA_TO_DEVICE);
1290         if (dma_mapping_error(device->dev, src))
1291                 goto err;
1292
1293         flags = DMA_COMPL_SRC_UNMAP_SINGLE | DMA_PREP_INTERRUPT;
1294         txd = device->device_prep_dma_memcpy(chan, dest, src, len, flags);
1295         if (!txd)
1296                 goto err1;
1297
1298         txd->callback = ntb_tx_copy_callback;
1299         txd->callback_param = entry;
1300
1301         cookie = dmaengine_submit(txd);
1302         if (dma_submit_error(cookie))
1303                 goto err1;
1304
1305         dma_async_issue_pending(chan);
1306         qp->tx_async++;
1307
1308         return;
1309 err1:
1310         dma_unmap_single(device->dev, src, len, DMA_TO_DEVICE);
1311 err:
1312         ntb_memcpy_tx(entry, offset);
1313         qp->tx_memcpy++;
1314 }
1315
1316 static int ntb_process_tx(struct ntb_transport_qp *qp,
1317                           struct ntb_queue_entry *entry)
1318 {
1319         dev_dbg(&ntb_query_pdev(qp->ndev)->dev, "%lld - tx %u, entry len %d flags %x buff %p\n",
1320                 qp->tx_pkts, qp->tx_index, entry->len, entry->flags,
1321                 entry->buf);
1322         if (qp->tx_index == qp->remote_rx_info->entry) {
1323                 qp->tx_ring_full++;
1324                 return -EAGAIN;
1325         }
1326
1327         if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
1328                 if (qp->tx_handler)
1329                         qp->tx_handler(qp->cb_data, qp, NULL, -EIO);
1330
1331                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1332                              &qp->tx_free_q);
1333                 return 0;
1334         }
1335
1336         ntb_async_tx(qp, entry);
1337
1338         qp->tx_index++;
1339         qp->tx_index %= qp->tx_max_entry;
1340
1341         qp->tx_pkts++;
1342
1343         return 0;
1344 }
1345
1346 static void ntb_send_link_down(struct ntb_transport_qp *qp)
1347 {
1348         struct pci_dev *pdev = ntb_query_pdev(qp->ndev);
1349         struct ntb_queue_entry *entry;
1350         int i, rc;
1351
1352         if (qp->qp_link == NTB_LINK_DOWN)
1353                 return;
1354
1355         qp->qp_link = NTB_LINK_DOWN;
1356         dev_info(&pdev->dev, "qp %d: Link Down\n", qp->qp_num);
1357
1358         for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1359                 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1360                 if (entry)
1361                         break;
1362                 msleep(100);
1363         }
1364
1365         if (!entry)
1366                 return;
1367
1368         entry->cb_data = NULL;
1369         entry->buf = NULL;
1370         entry->len = 0;
1371         entry->flags = LINK_DOWN_FLAG;
1372
1373         rc = ntb_process_tx(qp, entry);
1374         if (rc)
1375                 dev_err(&pdev->dev, "ntb: QP%d unable to send linkdown msg\n",
1376                         qp->qp_num);
1377 }
1378
1379 /**
1380  * ntb_transport_create_queue - Create a new NTB transport layer queue
1381  * @rx_handler: receive callback function
1382  * @tx_handler: transmit callback function
1383  * @event_handler: event callback function
1384  *
1385  * Create a new NTB transport layer queue and provide the queue with a callback
1386  * routine for both transmit and receive.  The receive callback routine will be
1387  * used to pass up data when the transport has received it on the queue.   The
1388  * transmit callback routine will be called when the transport has completed the
1389  * transmission of the data on the queue and the data is ready to be freed.
1390  *
1391  * RETURNS: pointer to newly created ntb_queue, NULL on error.
1392  */
1393 struct ntb_transport_qp *
1394 ntb_transport_create_queue(void *data, struct pci_dev *pdev,
1395                            const struct ntb_queue_handlers *handlers)
1396 {
1397         struct ntb_queue_entry *entry;
1398         struct ntb_transport_qp *qp;
1399         struct ntb_transport *nt;
1400         unsigned int free_queue;
1401         int rc, i;
1402
1403         nt = ntb_find_transport(pdev);
1404         if (!nt)
1405                 goto err;
1406
1407         free_queue = ffs(nt->qp_bitmap);
1408         if (!free_queue)
1409                 goto err;
1410
1411         /* decrement free_queue to make it zero based */
1412         free_queue--;
1413
1414         clear_bit(free_queue, &nt->qp_bitmap);
1415
1416         qp = &nt->qps[free_queue];
1417         qp->cb_data = data;
1418         qp->rx_handler = handlers->rx_handler;
1419         qp->tx_handler = handlers->tx_handler;
1420         qp->event_handler = handlers->event_handler;
1421
1422         dmaengine_get();
1423         qp->dma_chan = dma_find_channel(DMA_MEMCPY);
1424         if (!qp->dma_chan) {
1425                 dmaengine_put();
1426                 dev_info(&pdev->dev, "Unable to allocate DMA channel, using CPU instead\n");
1427         }
1428
1429         for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
1430                 entry = kzalloc(sizeof(struct ntb_queue_entry), GFP_ATOMIC);
1431                 if (!entry)
1432                         goto err1;
1433
1434                 entry->qp = qp;
1435                 ntb_list_add(&qp->ntb_rx_free_q_lock, &entry->entry,
1436                              &qp->rx_free_q);
1437         }
1438
1439         for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
1440                 entry = kzalloc(sizeof(struct ntb_queue_entry), GFP_ATOMIC);
1441                 if (!entry)
1442                         goto err2;
1443
1444                 entry->qp = qp;
1445                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1446                              &qp->tx_free_q);
1447         }
1448
1449         tasklet_init(&qp->rx_work, ntb_transport_rx, (unsigned long) qp);
1450
1451         rc = ntb_register_db_callback(qp->ndev, free_queue, qp,
1452                                       ntb_transport_rxc_db);
1453         if (rc)
1454                 goto err3;
1455
1456         dev_info(&pdev->dev, "NTB Transport QP %d created\n", qp->qp_num);
1457
1458         return qp;
1459
1460 err3:
1461         tasklet_disable(&qp->rx_work);
1462 err2:
1463         while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1464                 kfree(entry);
1465 err1:
1466         while ((entry = ntb_list_rm(&qp->ntb_rx_free_q_lock, &qp->rx_free_q)))
1467                 kfree(entry);
1468         if (qp->dma_chan)
1469                 dmaengine_put();
1470         set_bit(free_queue, &nt->qp_bitmap);
1471 err:
1472         return NULL;
1473 }
1474 EXPORT_SYMBOL_GPL(ntb_transport_create_queue);
1475
1476 /**
1477  * ntb_transport_free_queue - Frees NTB transport queue
1478  * @qp: NTB queue to be freed
1479  *
1480  * Frees NTB transport queue
1481  */
1482 void ntb_transport_free_queue(struct ntb_transport_qp *qp)
1483 {
1484         struct pci_dev *pdev;
1485         struct ntb_queue_entry *entry;
1486
1487         if (!qp)
1488                 return;
1489
1490         pdev = ntb_query_pdev(qp->ndev);
1491
1492         if (qp->dma_chan) {
1493                 struct dma_chan *chan = qp->dma_chan;
1494                 /* Putting the dma_chan to NULL will force any new traffic to be
1495                  * processed by the CPU instead of the DAM engine
1496                  */
1497                 qp->dma_chan = NULL;
1498
1499                 /* Try to be nice and wait for any queued DMA engine
1500                  * transactions to process before smashing it with a rock
1501                  */
1502                 dma_sync_wait(chan, qp->last_cookie);
1503                 dmaengine_terminate_all(chan);
1504                 dmaengine_put();
1505         }
1506
1507         ntb_unregister_db_callback(qp->ndev, qp->qp_num);
1508         tasklet_disable(&qp->rx_work);
1509
1510         cancel_delayed_work_sync(&qp->link_work);
1511
1512         while ((entry = ntb_list_rm(&qp->ntb_rx_free_q_lock, &qp->rx_free_q)))
1513                 kfree(entry);
1514
1515         while ((entry = ntb_list_rm(&qp->ntb_rx_pend_q_lock, &qp->rx_pend_q))) {
1516                 dev_warn(&pdev->dev, "Freeing item from a non-empty queue\n");
1517                 kfree(entry);
1518         }
1519
1520         while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1521                 kfree(entry);
1522
1523         set_bit(qp->qp_num, &qp->transport->qp_bitmap);
1524
1525         dev_info(&pdev->dev, "NTB Transport QP %d freed\n", qp->qp_num);
1526 }
1527 EXPORT_SYMBOL_GPL(ntb_transport_free_queue);
1528
1529 /**
1530  * ntb_transport_rx_remove - Dequeues enqueued rx packet
1531  * @qp: NTB queue to be freed
1532  * @len: pointer to variable to write enqueued buffers length
1533  *
1534  * Dequeues unused buffers from receive queue.  Should only be used during
1535  * shutdown of qp.
1536  *
1537  * RETURNS: NULL error value on error, or void* for success.
1538  */
1539 void *ntb_transport_rx_remove(struct ntb_transport_qp *qp, unsigned int *len)
1540 {
1541         struct ntb_queue_entry *entry;
1542         void *buf;
1543
1544         if (!qp || qp->client_ready == NTB_LINK_UP)
1545                 return NULL;
1546
1547         entry = ntb_list_rm(&qp->ntb_rx_pend_q_lock, &qp->rx_pend_q);
1548         if (!entry)
1549                 return NULL;
1550
1551         buf = entry->cb_data;
1552         *len = entry->len;
1553
1554         ntb_list_add(&qp->ntb_rx_free_q_lock, &entry->entry, &qp->rx_free_q);
1555
1556         return buf;
1557 }
1558 EXPORT_SYMBOL_GPL(ntb_transport_rx_remove);
1559
1560 /**
1561  * ntb_transport_rx_enqueue - Enqueue a new NTB queue entry
1562  * @qp: NTB transport layer queue the entry is to be enqueued on
1563  * @cb: per buffer pointer for callback function to use
1564  * @data: pointer to data buffer that incoming packets will be copied into
1565  * @len: length of the data buffer
1566  *
1567  * Enqueue a new receive buffer onto the transport queue into which a NTB
1568  * payload can be received into.
1569  *
1570  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
1571  */
1572 int ntb_transport_rx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
1573                              unsigned int len)
1574 {
1575         struct ntb_queue_entry *entry;
1576
1577         if (!qp)
1578                 return -EINVAL;
1579
1580         entry = ntb_list_rm(&qp->ntb_rx_free_q_lock, &qp->rx_free_q);
1581         if (!entry)
1582                 return -ENOMEM;
1583
1584         entry->cb_data = cb;
1585         entry->buf = data;
1586         entry->len = len;
1587
1588         ntb_list_add(&qp->ntb_rx_pend_q_lock, &entry->entry, &qp->rx_pend_q);
1589
1590         return 0;
1591 }
1592 EXPORT_SYMBOL_GPL(ntb_transport_rx_enqueue);
1593
1594 /**
1595  * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
1596  * @qp: NTB transport layer queue the entry is to be enqueued on
1597  * @cb: per buffer pointer for callback function to use
1598  * @data: pointer to data buffer that will be sent
1599  * @len: length of the data buffer
1600  *
1601  * Enqueue a new transmit buffer onto the transport queue from which a NTB
1602  * payload will be transmitted.  This assumes that a lock is being held to
1603  * serialize access to the qp.
1604  *
1605  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
1606  */
1607 int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
1608                              unsigned int len)
1609 {
1610         struct ntb_queue_entry *entry;
1611         int rc;
1612
1613         if (!qp || qp->qp_link != NTB_LINK_UP || !len)
1614                 return -EINVAL;
1615
1616         entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1617         if (!entry) {
1618                 qp->tx_err_no_buf++;
1619                 return -ENOMEM;
1620         }
1621
1622         entry->cb_data = cb;
1623         entry->buf = data;
1624         entry->len = len;
1625         entry->flags = 0;
1626
1627         rc = ntb_process_tx(qp, entry);
1628         if (rc)
1629                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1630                              &qp->tx_free_q);
1631
1632         return rc;
1633 }
1634 EXPORT_SYMBOL_GPL(ntb_transport_tx_enqueue);
1635
1636 /**
1637  * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
1638  * @qp: NTB transport layer queue to be enabled
1639  *
1640  * Notify NTB transport layer of client readiness to use queue
1641  */
1642 void ntb_transport_link_up(struct ntb_transport_qp *qp)
1643 {
1644         if (!qp)
1645                 return;
1646
1647         qp->client_ready = NTB_LINK_UP;
1648
1649         if (qp->transport->transport_link == NTB_LINK_UP)
1650                 schedule_delayed_work(&qp->link_work, 0);
1651 }
1652 EXPORT_SYMBOL_GPL(ntb_transport_link_up);
1653
1654 /**
1655  * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
1656  * @qp: NTB transport layer queue to be disabled
1657  *
1658  * Notify NTB transport layer of client's desire to no longer receive data on
1659  * transport queue specified.  It is the client's responsibility to ensure all
1660  * entries on queue are purged or otherwise handled appropriately.
1661  */
1662 void ntb_transport_link_down(struct ntb_transport_qp *qp)
1663 {
1664         struct pci_dev *pdev;
1665         int rc, val;
1666
1667         if (!qp)
1668                 return;
1669
1670         pdev = ntb_query_pdev(qp->ndev);
1671         qp->client_ready = NTB_LINK_DOWN;
1672
1673         rc = ntb_read_local_spad(qp->ndev, QP_LINKS, &val);
1674         if (rc) {
1675                 dev_err(&pdev->dev, "Error reading spad %d\n", QP_LINKS);
1676                 return;
1677         }
1678
1679         rc = ntb_write_remote_spad(qp->ndev, QP_LINKS,
1680                                    val & ~(1 << qp->qp_num));
1681         if (rc)
1682                 dev_err(&pdev->dev, "Error writing %x to remote spad %d\n",
1683                         val & ~(1 << qp->qp_num), QP_LINKS);
1684
1685         if (qp->qp_link == NTB_LINK_UP)
1686                 ntb_send_link_down(qp);
1687         else
1688                 cancel_delayed_work_sync(&qp->link_work);
1689 }
1690 EXPORT_SYMBOL_GPL(ntb_transport_link_down);
1691
1692 /**
1693  * ntb_transport_link_query - Query transport link state
1694  * @qp: NTB transport layer queue to be queried
1695  *
1696  * Query connectivity to the remote system of the NTB transport queue
1697  *
1698  * RETURNS: true for link up or false for link down
1699  */
1700 bool ntb_transport_link_query(struct ntb_transport_qp *qp)
1701 {
1702         if (!qp)
1703                 return false;
1704
1705         return qp->qp_link == NTB_LINK_UP;
1706 }
1707 EXPORT_SYMBOL_GPL(ntb_transport_link_query);
1708
1709 /**
1710  * ntb_transport_qp_num - Query the qp number
1711  * @qp: NTB transport layer queue to be queried
1712  *
1713  * Query qp number of the NTB transport queue
1714  *
1715  * RETURNS: a zero based number specifying the qp number
1716  */
1717 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
1718 {
1719         if (!qp)
1720                 return 0;
1721
1722         return qp->qp_num;
1723 }
1724 EXPORT_SYMBOL_GPL(ntb_transport_qp_num);
1725
1726 /**
1727  * ntb_transport_max_size - Query the max payload size of a qp
1728  * @qp: NTB transport layer queue to be queried
1729  *
1730  * Query the maximum payload size permissible on the given qp
1731  *
1732  * RETURNS: the max payload size of a qp
1733  */
1734 unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp)
1735 {
1736         unsigned int max;
1737
1738         if (!qp)
1739                 return 0;
1740
1741         if (!qp->dma_chan)
1742                 return qp->tx_max_frame - sizeof(struct ntb_payload_header);
1743
1744         /* If DMA engine usage is possible, try to find the max size for that */
1745         max = qp->tx_max_frame - sizeof(struct ntb_payload_header);
1746         max -= max % (1 << qp->dma_chan->device->copy_align);
1747
1748         return max;
1749 }
1750 EXPORT_SYMBOL_GPL(ntb_transport_max_size);