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 list_head rx_pend_q;
 123         struct list_head rx_free_q;
 124         spinlock_t ntb_rx_pend_q_lock;
 125         spinlock_t ntb_rx_free_q_lock;
 126         void *rx_buff;
 127         unsigned int rx_index;
 128         unsigned int rx_max_entry;
 129         unsigned int rx_max_frame;
 130         dma_cookie_t last_cookie;
 131
 132         void (*event_handler) (void *data, int status);
 133         struct delayed_work link_work;
 134         struct work_struct link_cleanup;
 135
 136         struct dentry *debugfs_dir;
 137         struct dentry *debugfs_stats;
 138
 139         /* Stats */
 140         u64 rx_bytes;
 141         u64 rx_pkts;
 142         u64 rx_ring_empty;
 143         u64 rx_err_no_buf;
 144         u64 rx_err_oflow;
 145         u64 rx_err_ver;
 146         u64 rx_memcpy;
 147         u64 rx_async;
 148         u64 tx_bytes;
 149         u64 tx_pkts;
 150         u64 tx_ring_full;
 151         u64 tx_err_no_buf;
 152         u64 tx_memcpy;
 153         u64 tx_async;
 154 };
 155
 156 struct ntb_transport_mw {
 157         size_t size;
 158         void *virt_addr;
 159         dma_addr_t dma_addr;
 160 };
 161
 162 struct ntb_transport_client_dev {
 163         struct list_head entry;
 164         struct device dev;
 165 };
 166
 167 struct ntb_transport {
 168         struct list_head entry;
 169         struct list_head client_devs;
 170
 171         struct ntb_device *ndev;
 172         struct ntb_transport_mw *mw;
 173         struct ntb_transport_qp *qps;
 174         unsigned int max_qps;
 175         unsigned long qp_bitmap;
 176         bool transport_link;
 177         struct delayed_work link_work;
 178         struct work_struct link_cleanup;
 179 };
 180
 181 enum {
 182         DESC_DONE_FLAG = 1 << 0,
 183         LINK_DOWN_FLAG = 1 << 1,
 184 };
 185
 186 struct ntb_payload_header {
 187         unsigned int ver;
 188         unsigned int len;
 189         unsigned int flags;
 190 };
 191
 192 enum {
 193         VERSION = 0,
 194         QP_LINKS,
 195         NUM_QPS,
 196         NUM_MWS,
 197         MW0_SZ_HIGH,
 198         MW0_SZ_LOW,
 199         MW1_SZ_HIGH,
 200         MW1_SZ_LOW,
 201         MAX_SPAD,
 202 };
 203
 204 #define QP_TO_MW(ndev, qp)      ((qp) % ntb_max_mw(ndev))
 205 #define NTB_QP_DEF_NUM_ENTRIES  100
 206 #define NTB_LINK_DOWN_TIMEOUT   10
 207
 208 static int ntb_match_bus(struct device *dev, struct device_driver *drv)
 209 {
 210         return !strncmp(dev_name(dev), drv->name, strlen(drv->name));
 211 }
 212
 213 static int ntb_client_probe(struct device *dev)
 214 {
 215         const struct ntb_client *drv = container_of(dev->driver,
 216                                                     struct ntb_client, driver);
 217         struct pci_dev *pdev = container_of(dev->parent, struct pci_dev, dev);
 218         int rc = -EINVAL;
 219
 220         get_device(dev);
 221         if (drv && drv->probe)
 222                 rc = drv->probe(pdev);
 223         if (rc)
 224                 put_device(dev);
 225
 226         return rc;
 227 }
 228
 229 static int ntb_client_remove(struct device *dev)
 230 {
 231         const struct ntb_client *drv = container_of(dev->driver,
 232                                                     struct ntb_client, driver);
 233         struct pci_dev *pdev = container_of(dev->parent, struct pci_dev, dev);
 234
 235         if (drv && drv->remove)
 236                 drv->remove(pdev);
 237
 238         put_device(dev);
 239
 240         return 0;
 241 }
 242
 243 static struct bus_type ntb_bus_type = {
 244         .name = "ntb_bus",
 245         .match = ntb_match_bus,
 246         .probe = ntb_client_probe,
 247         .remove = ntb_client_remove,
 248 };
 249
 250 static LIST_HEAD(ntb_transport_list);
 251
 252 static int ntb_bus_init(struct ntb_transport *nt)
 253 {
 254         if (list_empty(&ntb_transport_list)) {
 255                 int rc = bus_register(&ntb_bus_type);
 256                 if (rc)
 257                         return rc;
 258         }
 259
 260         list_add(&nt->entry, &ntb_transport_list);
 261
 262         return 0;
 263 }
 264
 265 static void ntb_bus_remove(struct ntb_transport *nt)
 266 {
 267         struct ntb_transport_client_dev *client_dev, *cd;
 268
 269         list_for_each_entry_safe(client_dev, cd, &nt->client_devs, entry) {
 270                 dev_err(client_dev->dev.parent, "%s still attached to bus, removing\n",
 271                         dev_name(&client_dev->dev));
 272                 list_del(&client_dev->entry);
 273                 device_unregister(&client_dev->dev);
 274         }
 275
 276         list_del(&nt->entry);
 277
 278         if (list_empty(&ntb_transport_list))
 279                 bus_unregister(&ntb_bus_type);
 280 }
 281
 282 static void ntb_client_release(struct device *dev)
 283 {
 284         struct ntb_transport_client_dev *client_dev;
 285         client_dev = container_of(dev, struct ntb_transport_client_dev, dev);
 286
 287         kfree(client_dev);
 288 }
 289
 290 /**
 291  * ntb_unregister_client_dev - Unregister NTB client device
 292  * @device_name: Name of NTB client device
 293  *
 294  * Unregister an NTB client device with the NTB transport layer
 295  */
 296 void ntb_unregister_client_dev(char *device_name)
 297 {
 298         struct ntb_transport_client_dev *client, *cd;
 299         struct ntb_transport *nt;
 300
 301         list_for_each_entry(nt, &ntb_transport_list, entry)
 302                 list_for_each_entry_safe(client, cd, &nt->client_devs, entry)
 303                         if (!strncmp(dev_name(&client->dev), device_name,
 304                                      strlen(device_name))) {
 305                                 list_del(&client->entry);
 306                                 device_unregister(&client->dev);
 307                         }
 308 }
 309 EXPORT_SYMBOL_GPL(ntb_unregister_client_dev);
 310
 311 /**
 312  * ntb_register_client_dev - Register NTB client device
 313  * @device_name: Name of NTB client device
 314  *
 315  * Register an NTB client device with the NTB transport layer
 316  */
 317 int ntb_register_client_dev(char *device_name)
 318 {
 319         struct ntb_transport_client_dev *client_dev;
 320         struct ntb_transport *nt;
 321         int rc, i = 0;
 322
 323         if (list_empty(&ntb_transport_list))
 324                 return -ENODEV;
 325
 326         list_for_each_entry(nt, &ntb_transport_list, entry) {
 327                 struct device *dev;
 328
 329                 client_dev = kzalloc(sizeof(struct ntb_transport_client_dev),
 330                                      GFP_KERNEL);
 331                 if (!client_dev) {
 332                         rc = -ENOMEM;
 333                         goto err;
 334                 }
 335
 336                 dev = &client_dev->dev;
 337
 338                 /* setup and register client devices */
 339                 dev_set_name(dev, "%s%d", device_name, i);
 340                 dev->bus = &ntb_bus_type;
 341                 dev->release = ntb_client_release;
 342                 dev->parent = &ntb_query_pdev(nt->ndev)->dev;
 343
 344                 rc = device_register(dev);
 345                 if (rc) {
 346                         kfree(client_dev);
 347                         goto err;
 348                 }
 349
 350                 list_add_tail(&client_dev->entry, &nt->client_devs);
 351                 i++;
 352         }
 353
 354         return 0;
 355
 356 err:
 357         ntb_unregister_client_dev(device_name);
 358
 359         return rc;
 360 }
 361 EXPORT_SYMBOL_GPL(ntb_register_client_dev);
 362
 363 /**
 364  * ntb_register_client - Register NTB client driver
 365  * @drv: NTB client driver to be registered
 366  *
 367  * Register an NTB client driver with the NTB transport layer
 368  *
 369  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 370  */
 371 int ntb_register_client(struct ntb_client *drv)
 372 {
 373         drv->driver.bus = &ntb_bus_type;
 374
 375         if (list_empty(&ntb_transport_list))
 376                 return -ENODEV;
 377
 378         return driver_register(&drv->driver);
 379 }
 380 EXPORT_SYMBOL_GPL(ntb_register_client);
 381
 382 /**
 383  * ntb_unregister_client - Unregister NTB client driver
 384  * @drv: NTB client driver to be unregistered
 385  *
 386  * Unregister an NTB client driver with the NTB transport layer
 387  *
 388  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 389  */
 390 void ntb_unregister_client(struct ntb_client *drv)
 391 {
 392         driver_unregister(&drv->driver);
 393 }
 394 EXPORT_SYMBOL_GPL(ntb_unregister_client);
 395
 396 static ssize_t debugfs_read(struct file *filp, char __user *ubuf, size_t count,
 397                             loff_t *offp)
 398 {
 399         struct ntb_transport_qp *qp;
 400         char *buf;
 401         ssize_t ret, out_offset, out_count;
 402
 403         out_count = 1000;
 404
 405         buf = kmalloc(out_count, GFP_KERNEL);
 406         if (!buf)
 407                 return -ENOMEM;
 408
 409         qp = filp->private_data;
 410         out_offset = 0;
 411         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 412                                "NTB QP stats\n");
 413         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 414                                "rx_bytes - \t%llu\n", qp->rx_bytes);
 415         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 416                                "rx_pkts - \t%llu\n", qp->rx_pkts);
 417         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 418                                "rx_memcpy - \t%llu\n", qp->rx_memcpy);
 419         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 420                                "rx_async - \t%llu\n", qp->rx_async);
 421         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 422                                "rx_ring_empty - %llu\n", qp->rx_ring_empty);
 423         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 424                                "rx_err_no_buf - %llu\n", qp->rx_err_no_buf);
 425         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 426                                "rx_err_oflow - \t%llu\n", qp->rx_err_oflow);
 427         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 428                                "rx_err_ver - \t%llu\n", qp->rx_err_ver);
 429         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 430                                "rx_buff - \t%p\n", qp->rx_buff);
 431         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 432                                "rx_index - \t%u\n", qp->rx_index);
 433         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 434                                "rx_max_entry - \t%u\n", qp->rx_max_entry);
 435
 436         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 437                                "tx_bytes - \t%llu\n", qp->tx_bytes);
 438         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 439                                "tx_pkts - \t%llu\n", qp->tx_pkts);
 440         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 441                                "tx_memcpy - \t%llu\n", qp->tx_memcpy);
 442         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 443                                "tx_async - \t%llu\n", qp->tx_async);
 444         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 445                                "tx_ring_full - \t%llu\n", qp->tx_ring_full);
 446         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 447                                "tx_err_no_buf - %llu\n", qp->tx_err_no_buf);
 448         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 449                                "tx_mw - \t%p\n", qp->tx_mw);
 450         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 451                                "tx_index - \t%u\n", qp->tx_index);
 452         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 453                                "tx_max_entry - \t%u\n", qp->tx_max_entry);
 454
 455         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 456                                "\nQP Link %s\n", (qp->qp_link == NTB_LINK_UP) ?
 457                                "Up" : "Down");
 458         if (out_offset > out_count)
 459                 out_offset = out_count;
 460
 461         ret = simple_read_from_buffer(ubuf, count, offp, buf, out_offset);
 462         kfree(buf);
 463         return ret;
 464 }
 465
 466 static const struct file_operations ntb_qp_debugfs_stats = {
 467         .owner = THIS_MODULE,
 468         .open = simple_open,
 469         .read = debugfs_read,
 470 };
 471
 472 static void ntb_list_add(spinlock_t *lock, struct list_head *entry,
 473                          struct list_head *list)
 474 {
 475         unsigned long flags;
 476
 477         spin_lock_irqsave(lock, flags);
 478         list_add_tail(entry, list);
 479         spin_unlock_irqrestore(lock, flags);
 480 }
 481
 482 static struct ntb_queue_entry *ntb_list_rm(spinlock_t *lock,
 483                                                 struct list_head *list)
 484 {
 485         struct ntb_queue_entry *entry;
 486         unsigned long flags;
 487
 488         spin_lock_irqsave(lock, flags);
 489         if (list_empty(list)) {
 490                 entry = NULL;
 491                 goto out;
 492         }
 493         entry = list_first_entry(list, struct ntb_queue_entry, entry);
 494         list_del(&entry->entry);
 495 out:
 496         spin_unlock_irqrestore(lock, flags);
 497
 498         return entry;
 499 }
 500
 501 static void ntb_transport_setup_qp_mw(struct ntb_transport *nt,
 502                                       unsigned int qp_num)
 503 {
 504         struct ntb_transport_qp *qp = &nt->qps[qp_num];
 505         unsigned int rx_size, num_qps_mw;
 506         u8 mw_num, mw_max;
 507         unsigned int i;
 508
 509         mw_max = ntb_max_mw(nt->ndev);
 510         mw_num = QP_TO_MW(nt->ndev, qp_num);
 511
 512         WARN_ON(nt->mw[mw_num].virt_addr == NULL);
 513
 514         if (nt->max_qps % mw_max && mw_num < nt->max_qps % mw_max)
 515                 num_qps_mw = nt->max_qps / mw_max + 1;
 516         else
 517                 num_qps_mw = nt->max_qps / mw_max;
 518
 519         rx_size = (unsigned int) nt->mw[mw_num].size / num_qps_mw;
 520         qp->rx_buff = nt->mw[mw_num].virt_addr + qp_num / mw_max * rx_size;
 521         rx_size -= sizeof(struct ntb_rx_info);
 522
 523         qp->remote_rx_info = qp->rx_buff + rx_size;
 524
 525         /* Due to housekeeping, there must be atleast 2 buffs */
 526         qp->rx_max_frame = min(transport_mtu, rx_size / 2);
 527         qp->rx_max_entry = rx_size / qp->rx_max_frame;
 528         qp->rx_index = 0;
 529
 530         qp->remote_rx_info->entry = qp->rx_max_entry - 1;
 531
 532         /* setup the hdr offsets with 0's */
 533         for (i = 0; i < qp->rx_max_entry; i++) {
 534                 void *offset = qp->rx_buff + qp->rx_max_frame * (i + 1) -
 535                                sizeof(struct ntb_payload_header);
 536                 memset(offset, 0, sizeof(struct ntb_payload_header));
 537         }
 538
 539         qp->rx_pkts = 0;
 540         qp->tx_pkts = 0;
 541         qp->tx_index = 0;
 542 }
 543
 544 static void ntb_free_mw(struct ntb_transport *nt, int num_mw)
 545 {
 546         struct ntb_transport_mw *mw = &nt->mw[num_mw];
 547         struct pci_dev *pdev = ntb_query_pdev(nt->ndev);
 548
 549         if (!mw->virt_addr)
 550                 return;
 551
 552         dma_free_coherent(&pdev->dev, mw->size, mw->virt_addr, mw->dma_addr);
 553         mw->virt_addr = NULL;
 554 }
 555
 556 static int ntb_set_mw(struct ntb_transport *nt, int num_mw, unsigned int size)
 557 {
 558         struct ntb_transport_mw *mw = &nt->mw[num_mw];
 559         struct pci_dev *pdev = ntb_query_pdev(nt->ndev);
 560
 561         /* No need to re-setup */
 562         if (mw->size == ALIGN(size, 4096))
 563                 return 0;
 564
 565         if (mw->size != 0)
 566                 ntb_free_mw(nt, num_mw);
 567
 568         /* Alloc memory for receiving data.  Must be 4k aligned */
 569         mw->size = ALIGN(size, 4096);
 570
 571         mw->virt_addr = dma_alloc_coherent(&pdev->dev, mw->size, &mw->dma_addr,
 572                                            GFP_KERNEL);
 573         if (!mw->virt_addr) {
 574                 mw->size = 0;
 575                 dev_err(&pdev->dev, "Unable to allocate MW buffer of size %d\n",
 576                        (int) mw->size);
 577                 return -ENOMEM;
 578         }
 579
 580         /* Notify HW the memory location of the receive buffer */
 581         ntb_set_mw_addr(nt->ndev, num_mw, mw->dma_addr);
 582
 583         return 0;
 584 }
 585
 586 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
 587 {
 588         struct ntb_transport *nt = qp->transport;
 589         struct pci_dev *pdev = ntb_query_pdev(nt->ndev);
 590
 591         if (qp->qp_link == NTB_LINK_DOWN) {
 592                 cancel_delayed_work_sync(&qp->link_work);
 593                 return;
 594         }
 595
 596         if (qp->event_handler)
 597                 qp->event_handler(qp->cb_data, NTB_LINK_DOWN);
 598
 599         dev_info(&pdev->dev, "qp %d: Link Down\n", qp->qp_num);
 600         qp->qp_link = NTB_LINK_DOWN;
 601 }
 602
 603 static void ntb_qp_link_cleanup_work(struct work_struct *work)
 604 {
 605         struct ntb_transport_qp *qp = container_of(work,
 606                                                    struct ntb_transport_qp,
 607                                                    link_cleanup);
 608         struct ntb_transport *nt = qp->transport;
 609
 610         ntb_qp_link_cleanup(qp);
 611
 612         if (nt->transport_link == NTB_LINK_UP)
 613                 schedule_delayed_work(&qp->link_work,
 614                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 615 }
 616
 617 static void ntb_qp_link_down(struct ntb_transport_qp *qp)
 618 {
 619         schedule_work(&qp->link_cleanup);
 620 }
 621
 622 static void ntb_transport_link_cleanup(struct ntb_transport *nt)
 623 {
 624         int i;
 625
 626         /* Pass along the info to any clients */
 627         for (i = 0; i < nt->max_qps; i++)
 628                 if (!test_bit(i, &nt->qp_bitmap))
 629                         ntb_qp_link_cleanup(&nt->qps[i]);
 630
 631         if (nt->transport_link == NTB_LINK_DOWN)
 632                 cancel_delayed_work_sync(&nt->link_work);
 633         else
 634                 nt->transport_link = NTB_LINK_DOWN;
 635
 636         /* The scratchpad registers keep the values if the remote side
 637          * goes down, blast them now to give them a sane value the next
 638          * time they are accessed
 639          */
 640         for (i = 0; i < MAX_SPAD; i++)
 641                 ntb_write_local_spad(nt->ndev, i, 0);
 642 }
 643
 644 static void ntb_transport_link_cleanup_work(struct work_struct *work)
 645 {
 646         struct ntb_transport *nt = container_of(work, struct ntb_transport,
 647                                                 link_cleanup);
 648
 649         ntb_transport_link_cleanup(nt);
 650 }
 651
 652 static void ntb_transport_event_callback(void *data, enum ntb_hw_event event)
 653 {
 654         struct ntb_transport *nt = data;
 655
 656         switch (event) {
 657         case NTB_EVENT_HW_LINK_UP:
 658                 schedule_delayed_work(&nt->link_work, 0);
 659                 break;
 660         case NTB_EVENT_HW_LINK_DOWN:
 661                 schedule_work(&nt->link_cleanup);
 662                 break;
 663         default:
 664                 BUG();
 665         }
 666 }
 667
 668 static void ntb_transport_link_work(struct work_struct *work)
 669 {
 670         struct ntb_transport *nt = container_of(work, struct ntb_transport,
 671                                                 link_work.work);
 672         struct ntb_device *ndev = nt->ndev;
 673         struct pci_dev *pdev = ntb_query_pdev(ndev);
 674         u32 val;
 675         int rc, i;
 676
 677         /* send the local info, in the opposite order of the way we read it */
 678         for (i = 0; i < ntb_max_mw(ndev); i++) {
 679                 rc = ntb_write_remote_spad(ndev, MW0_SZ_HIGH + (i * 2),
 680                                            ntb_get_mw_size(ndev, i) >> 32);
 681                 if (rc) {
 682                         dev_err(&pdev->dev, "Error writing %u to remote spad %d\n",
 683                                 (u32)(ntb_get_mw_size(ndev, i) >> 32),
 684                                 MW0_SZ_HIGH + (i * 2));
 685                         goto out;
 686                 }
 687
 688                 rc = ntb_write_remote_spad(ndev, MW0_SZ_LOW + (i * 2),
 689                                            (u32) ntb_get_mw_size(ndev, i));
 690                 if (rc) {
 691                         dev_err(&pdev->dev, "Error writing %u to remote spad %d\n",
 692                                 (u32) ntb_get_mw_size(ndev, i),
 693                                 MW0_SZ_LOW + (i * 2));
 694                         goto out;
 695                 }
 696         }
 697
 698         rc = ntb_write_remote_spad(ndev, NUM_MWS, ntb_max_mw(ndev));
 699         if (rc) {
 700                 dev_err(&pdev->dev, "Error writing %x to remote spad %d\n",
 701                         ntb_max_mw(ndev), NUM_MWS);
 702                 goto out;
 703         }
 704
 705         rc = ntb_write_remote_spad(ndev, NUM_QPS, nt->max_qps);
 706         if (rc) {
 707                 dev_err(&pdev->dev, "Error writing %x to remote spad %d\n",
 708                         nt->max_qps, NUM_QPS);
 709                 goto out;
 710         }
 711
 712         rc = ntb_write_remote_spad(ndev, VERSION, NTB_TRANSPORT_VERSION);
 713         if (rc) {
 714                 dev_err(&pdev->dev, "Error writing %x to remote spad %d\n",
 715                         NTB_TRANSPORT_VERSION, VERSION);
 716                 goto out;
 717         }
 718
 719         /* Query the remote side for its info */
 720         rc = ntb_read_remote_spad(ndev, VERSION, &val);
 721         if (rc) {
 722                 dev_err(&pdev->dev, "Error reading remote spad %d\n", VERSION);
 723                 goto out;
 724         }
 725
 726         if (val != NTB_TRANSPORT_VERSION)
 727                 goto out;
 728         dev_dbg(&pdev->dev, "Remote version = %d\n", val);
 729
 730         rc = ntb_read_remote_spad(ndev, NUM_QPS, &val);
 731         if (rc) {
 732                 dev_err(&pdev->dev, "Error reading remote spad %d\n", NUM_QPS);
 733                 goto out;
 734         }
 735
 736         if (val != nt->max_qps)
 737                 goto out;
 738         dev_dbg(&pdev->dev, "Remote max number of qps = %d\n", val);
 739
 740         rc = ntb_read_remote_spad(ndev, NUM_MWS, &val);
 741         if (rc) {
 742                 dev_err(&pdev->dev, "Error reading remote spad %d\n", NUM_MWS);
 743                 goto out;
 744         }
 745
 746         if (val != ntb_max_mw(ndev))
 747                 goto out;
 748         dev_dbg(&pdev->dev, "Remote number of mws = %d\n", val);
 749
 750         for (i = 0; i < ntb_max_mw(ndev); i++) {
 751                 u64 val64;
 752
 753                 rc = ntb_read_remote_spad(ndev, MW0_SZ_HIGH + (i * 2), &val);
 754                 if (rc) {
 755                         dev_err(&pdev->dev, "Error reading remote spad %d\n",
 756                                 MW0_SZ_HIGH + (i * 2));
 757                         goto out1;
 758                 }
 759
 760                 val64 = (u64) val << 32;
 761
 762                 rc = ntb_read_remote_spad(ndev, MW0_SZ_LOW + (i * 2), &val);
 763                 if (rc) {
 764                         dev_err(&pdev->dev, "Error reading remote spad %d\n",
 765                                 MW0_SZ_LOW + (i * 2));
 766                         goto out1;
 767                 }
 768
 769                 val64 |= val;
 770
 771                 dev_dbg(&pdev->dev, "Remote MW%d size = %llu\n", i, val64);
 772
 773                 rc = ntb_set_mw(nt, i, val64);
 774                 if (rc)
 775                         goto out1;
 776         }
 777
 778         nt->transport_link = NTB_LINK_UP;
 779
 780         for (i = 0; i < nt->max_qps; i++) {
 781                 struct ntb_transport_qp *qp = &nt->qps[i];
 782
 783                 ntb_transport_setup_qp_mw(nt, i);
 784
 785                 if (qp->client_ready == NTB_LINK_UP)
 786                         schedule_delayed_work(&qp->link_work, 0);
 787         }
 788
 789         return;
 790
 791 out1:
 792         for (i = 0; i < ntb_max_mw(ndev); i++)
 793                 ntb_free_mw(nt, i);
 794 out:
 795         if (ntb_hw_link_status(ndev))
 796                 schedule_delayed_work(&nt->link_work,
 797                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 798 }
 799
 800 static void ntb_qp_link_work(struct work_struct *work)
 801 {
 802         struct ntb_transport_qp *qp = container_of(work,
 803                                                    struct ntb_transport_qp,
 804                                                    link_work.work);
 805         struct pci_dev *pdev = ntb_query_pdev(qp->ndev);
 806         struct ntb_transport *nt = qp->transport;
 807         int rc, val;
 808
 809         WARN_ON(nt->transport_link != NTB_LINK_UP);
 810
 811         rc = ntb_read_local_spad(nt->ndev, QP_LINKS, &val);
 812         if (rc) {
 813                 dev_err(&pdev->dev, "Error reading spad %d\n", QP_LINKS);
 814                 return;
 815         }
 816
 817         rc = ntb_write_remote_spad(nt->ndev, QP_LINKS, val | 1 << qp->qp_num);
 818         if (rc)
 819                 dev_err(&pdev->dev, "Error writing %x to remote spad %d\n",
 820                         val | 1 << qp->qp_num, QP_LINKS);
 821
 822         /* query remote spad for qp ready bits */
 823         rc = ntb_read_remote_spad(nt->ndev, QP_LINKS, &val);
 824         if (rc)
 825                 dev_err(&pdev->dev, "Error reading remote spad %d\n", QP_LINKS);
 826
 827         dev_dbg(&pdev->dev, "Remote QP link status = %x\n", val);
 828
 829         /* See if the remote side is up */
 830         if (1 << qp->qp_num & val) {
 831                 qp->qp_link = NTB_LINK_UP;
 832
 833                 dev_info(&pdev->dev, "qp %d: Link Up\n", qp->qp_num);
 834                 if (qp->event_handler)
 835                         qp->event_handler(qp->cb_data, NTB_LINK_UP);
 836         } else if (nt->transport_link == NTB_LINK_UP)
 837                 schedule_delayed_work(&qp->link_work,
 838                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 839 }
 840
 841 static int ntb_transport_init_queue(struct ntb_transport *nt,
 842                                      unsigned int qp_num)
 843 {
 844         struct ntb_transport_qp *qp;
 845         unsigned int num_qps_mw, tx_size;
 846         u8 mw_num, mw_max;
 847         u64 qp_offset;
 848
 849         mw_max = ntb_max_mw(nt->ndev);
 850         mw_num = QP_TO_MW(nt->ndev, qp_num);
 851
 852         qp = &nt->qps[qp_num];
 853         qp->qp_num = qp_num;
 854         qp->transport = nt;
 855         qp->ndev = nt->ndev;
 856         qp->qp_link = NTB_LINK_DOWN;
 857         qp->client_ready = NTB_LINK_DOWN;
 858         qp->event_handler = NULL;
 859
 860         if (nt->max_qps % mw_max && mw_num < nt->max_qps % mw_max)
 861                 num_qps_mw = nt->max_qps / mw_max + 1;
 862         else
 863                 num_qps_mw = nt->max_qps / mw_max;
 864
 865         tx_size = (unsigned int) ntb_get_mw_size(qp->ndev, mw_num) / num_qps_mw;
 866         qp_offset = qp_num / mw_max * tx_size;
 867         qp->tx_mw = ntb_get_mw_vbase(nt->ndev, mw_num) + qp_offset;
 868         if (!qp->tx_mw)
 869                 return -EINVAL;
 870
 871         qp->tx_mw_phys = ntb_get_mw_base(qp->ndev, mw_num) + qp_offset;
 872         if (!qp->tx_mw_phys)
 873                 return -EINVAL;
 874
 875         tx_size -= sizeof(struct ntb_rx_info);
 876         qp->rx_info = qp->tx_mw + tx_size;
 877
 878         /* Due to housekeeping, there must be atleast 2 buffs */
 879         qp->tx_max_frame = min(transport_mtu, tx_size / 2);
 880         qp->tx_max_entry = tx_size / qp->tx_max_frame;
 881
 882         if (ntb_query_debugfs(nt->ndev)) {
 883                 char debugfs_name[4];
 884
 885                 snprintf(debugfs_name, 4, "qp%d", qp_num);
 886                 qp->debugfs_dir = debugfs_create_dir(debugfs_name,
 887                                                  ntb_query_debugfs(nt->ndev));
 888
 889                 qp->debugfs_stats = debugfs_create_file("stats", S_IRUSR,
 890                                                         qp->debugfs_dir, qp,
 891                                                         &ntb_qp_debugfs_stats);
 892         }
 893
 894         INIT_DELAYED_WORK(&qp->link_work, ntb_qp_link_work);
 895         INIT_WORK(&qp->link_cleanup, ntb_qp_link_cleanup_work);
 896
 897         spin_lock_init(&qp->ntb_rx_pend_q_lock);
 898         spin_lock_init(&qp->ntb_rx_free_q_lock);
 899         spin_lock_init(&qp->ntb_tx_free_q_lock);
 900
 901         INIT_LIST_HEAD(&qp->rx_pend_q);
 902         INIT_LIST_HEAD(&qp->rx_free_q);
 903         INIT_LIST_HEAD(&qp->tx_free_q);
 904
 905         return 0;
 906 }
 907
 908 int ntb_transport_init(struct pci_dev *pdev)
 909 {
 910         struct ntb_transport *nt;
 911         int rc, i;
 912
 913         nt = kzalloc(sizeof(struct ntb_transport), GFP_KERNEL);
 914         if (!nt)
 915                 return -ENOMEM;
 916
 917         nt->ndev = ntb_register_transport(pdev, nt);
 918         if (!nt->ndev) {
 919                 rc = -EIO;
 920                 goto err;
 921         }
 922
 923         nt->mw = kcalloc(ntb_max_mw(nt->ndev), sizeof(struct ntb_transport_mw),
 924                          GFP_KERNEL);
 925         if (!nt->mw) {
 926                 rc = -ENOMEM;
 927                 goto err1;
 928         }
 929
 930         if (max_num_clients)
 931                 nt->max_qps = min(ntb_max_cbs(nt->ndev), max_num_clients);
 932         else
 933                 nt->max_qps = min(ntb_max_cbs(nt->ndev), ntb_max_mw(nt->ndev));
 934
 935         nt->qps = kcalloc(nt->max_qps, sizeof(struct ntb_transport_qp),
 936                           GFP_KERNEL);
 937         if (!nt->qps) {
 938                 rc = -ENOMEM;
 939                 goto err2;
 940         }
 941
 942         nt->qp_bitmap = ((u64) 1 << nt->max_qps) - 1;
 943
 944         for (i = 0; i < nt->max_qps; i++) {
 945                 rc = ntb_transport_init_queue(nt, i);
 946                 if (rc)
 947                         goto err3;
 948         }
 949
 950         INIT_DELAYED_WORK(&nt->link_work, ntb_transport_link_work);
 951         INIT_WORK(&nt->link_cleanup, ntb_transport_link_cleanup_work);
 952
 953         rc = ntb_register_event_callback(nt->ndev,
 954                                          ntb_transport_event_callback);
 955         if (rc)
 956                 goto err3;
 957
 958         INIT_LIST_HEAD(&nt->client_devs);
 959         rc = ntb_bus_init(nt);
 960         if (rc)
 961                 goto err4;
 962
 963         if (ntb_hw_link_status(nt->ndev))
 964                 schedule_delayed_work(&nt->link_work, 0);
 965
 966         return 0;
 967
 968 err4:
 969         ntb_unregister_event_callback(nt->ndev);
 970 err3:
 971         kfree(nt->qps);
 972 err2:
 973         kfree(nt->mw);
 974 err1:
 975         ntb_unregister_transport(nt->ndev);
 976 err:
 977         kfree(nt);
 978         return rc;
 979 }
 980
 981 void ntb_transport_free(void *transport)
 982 {
 983         struct ntb_transport *nt = transport;
 984         struct ntb_device *ndev = nt->ndev;
 985         int i;
 986
 987         ntb_transport_link_cleanup(nt);
 988
 989         /* verify that all the qp's are freed */
 990         for (i = 0; i < nt->max_qps; i++) {
 991                 if (!test_bit(i, &nt->qp_bitmap))
 992                         ntb_transport_free_queue(&nt->qps[i]);
 993                 debugfs_remove_recursive(nt->qps[i].debugfs_dir);
 994         }
 995
 996         ntb_bus_remove(nt);
 997
 998         cancel_delayed_work_sync(&nt->link_work);
 999
1000         ntb_unregister_event_callback(ndev);
1001
1002         for (i = 0; i < ntb_max_mw(ndev); i++)
1003                 ntb_free_mw(nt, i);
1004
1005         kfree(nt->qps);
1006         kfree(nt->mw);
1007         ntb_unregister_transport(ndev);
1008         kfree(nt);
1009 }
1010
1011 static void ntb_rx_copy_callback(void *data)
1012 {
1013         struct ntb_queue_entry *entry = data;
1014         struct ntb_transport_qp *qp = entry->qp;
1015         void *cb_data = entry->cb_data;
1016         unsigned int len = entry->len;
1017         struct ntb_payload_header *hdr = entry->rx_hdr;
1018
1019         /* Ensure that the data is fully copied out before clearing the flag */
1020         wmb();
1021         hdr->flags = 0;
1022
1023         iowrite32(entry->index, &qp->rx_info->entry);
1024
1025         ntb_list_add(&qp->ntb_rx_free_q_lock, &entry->entry, &qp->rx_free_q);
1026
1027         if (qp->rx_handler && qp->client_ready == NTB_LINK_UP)
1028                 qp->rx_handler(qp, qp->cb_data, cb_data, len);
1029 }
1030
1031 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset)
1032 {
1033         void *buf = entry->buf;
1034         size_t len = entry->len;
1035
1036         memcpy(buf, offset, len);
1037
1038         ntb_rx_copy_callback(entry);
1039 }
1040
1041 static void ntb_async_rx(struct ntb_queue_entry *entry, void *offset,
1042                          size_t len)
1043 {
1044         struct dma_async_tx_descriptor *txd;
1045         struct ntb_transport_qp *qp = entry->qp;
1046         struct dma_chan *chan = qp->dma_chan;
1047         struct dma_device *device;
1048         size_t pay_off, buff_off;
1049         struct dmaengine_unmap_data *unmap;
1050         dma_cookie_t cookie;
1051         void *buf = entry->buf;
1052
1053         entry->len = len;
1054
1055         if (!chan)
1056                 goto err;
1057
1058         if (len < copy_bytes)
1059                 goto err_wait;
1060
1061         device = chan->device;
1062         pay_off = (size_t) offset & ~PAGE_MASK;
1063         buff_off = (size_t) buf & ~PAGE_MASK;
1064
1065         if (!is_dma_copy_aligned(device, pay_off, buff_off, len))
1066                 goto err_wait;
1067
1068         unmap = dmaengine_get_unmap_data(device->dev, 2, GFP_NOWAIT);
1069         if (!unmap)
1070                 goto err_wait;
1071
1072         unmap->len = len;
1073         unmap->addr[0] = dma_map_page(device->dev, virt_to_page(offset),
1074                                       pay_off, len, DMA_TO_DEVICE);
1075         if (dma_mapping_error(device->dev, unmap->addr[0]))
1076                 goto err_get_unmap;
1077
1078         unmap->to_cnt = 1;
1079
1080         unmap->addr[1] = dma_map_page(device->dev, virt_to_page(buf),
1081                                       buff_off, len, DMA_FROM_DEVICE);
1082         if (dma_mapping_error(device->dev, unmap->addr[1]))
1083                 goto err_get_unmap;
1084
1085         unmap->from_cnt = 1;
1086
1087         txd = device->device_prep_dma_memcpy(chan, unmap->addr[1],
1088                                              unmap->addr[0], len,
1089                                              DMA_PREP_INTERRUPT);
1090         if (!txd)
1091                 goto err_get_unmap;
1092
1093         txd->callback = ntb_rx_copy_callback;
1094         txd->callback_param = entry;
1095         dma_set_unmap(txd, unmap);
1096
1097         cookie = dmaengine_submit(txd);
1098         if (dma_submit_error(cookie))
1099                 goto err_set_unmap;
1100
1101         dmaengine_unmap_put(unmap);
1102
1103         qp->last_cookie = cookie;
1104
1105         qp->rx_async++;
1106
1107         return;
1108
1109 err_set_unmap:
1110         dmaengine_unmap_put(unmap);
1111 err_get_unmap:
1112         dmaengine_unmap_put(unmap);
1113 err_wait:
1114         /* If the callbacks come out of order, the writing of the index to the
1115          * last completed will be out of order.  This may result in the
1116          * receive stalling forever.
1117          */
1118         dma_sync_wait(chan, qp->last_cookie);
1119 err:
1120         ntb_memcpy_rx(entry, offset);
1121         qp->rx_memcpy++;
1122 }
1123
1124 static int ntb_process_rxc(struct ntb_transport_qp *qp)
1125 {
1126         struct ntb_payload_header *hdr;
1127         struct ntb_queue_entry *entry;
1128         void *offset;
1129
1130         offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
1131         hdr = offset + qp->rx_max_frame - sizeof(struct ntb_payload_header);
1132
1133         entry = ntb_list_rm(&qp->ntb_rx_pend_q_lock, &qp->rx_pend_q);
1134         if (!entry) {
1135                 dev_dbg(&ntb_query_pdev(qp->ndev)->dev,
1136                         "no buffer - HDR ver %u, len %d, flags %x\n",
1137                         hdr->ver, hdr->len, hdr->flags);
1138                 qp->rx_err_no_buf++;
1139                 return -ENOMEM;
1140         }
1141
1142         if (!(hdr->flags & DESC_DONE_FLAG)) {
1143                 ntb_list_add(&qp->ntb_rx_pend_q_lock, &entry->entry,
1144                              &qp->rx_pend_q);
1145                 qp->rx_ring_empty++;
1146                 return -EAGAIN;
1147         }
1148
1149         if (hdr->ver != (u32) qp->rx_pkts) {
1150                 dev_dbg(&ntb_query_pdev(qp->ndev)->dev,
1151                         "qp %d: version mismatch, expected %llu - got %u\n",
1152                         qp->qp_num, qp->rx_pkts, hdr->ver);
1153                 ntb_list_add(&qp->ntb_rx_pend_q_lock, &entry->entry,
1154                              &qp->rx_pend_q);
1155                 qp->rx_err_ver++;
1156                 return -EIO;
1157         }
1158
1159         if (hdr->flags & LINK_DOWN_FLAG) {
1160                 ntb_qp_link_down(qp);
1161
1162                 goto err;
1163         }
1164
1165         dev_dbg(&ntb_query_pdev(qp->ndev)->dev,
1166                 "rx offset %u, ver %u - %d payload received, buf size %d\n",
1167                 qp->rx_index, hdr->ver, hdr->len, entry->len);
1168
1169         qp->rx_bytes += hdr->len;
1170         qp->rx_pkts++;
1171
1172         if (hdr->len > entry->len) {
1173                 qp->rx_err_oflow++;
1174                 dev_dbg(&ntb_query_pdev(qp->ndev)->dev,
1175                         "RX overflow! Wanted %d got %d\n",
1176                         hdr->len, entry->len);
1177
1178                 goto err;
1179         }
1180
1181         entry->index = qp->rx_index;
1182         entry->rx_hdr = hdr;
1183
1184         ntb_async_rx(entry, offset, hdr->len);
1185
1186 out:
1187         qp->rx_index++;
1188         qp->rx_index %= qp->rx_max_entry;
1189
1190         return 0;
1191
1192 err:
1193         ntb_list_add(&qp->ntb_rx_pend_q_lock, &entry->entry,
1194                      &qp->rx_pend_q);
1195         /* Ensure that the data is fully copied out before clearing the flag */
1196         wmb();
1197         hdr->flags = 0;
1198         iowrite32(qp->rx_index, &qp->rx_info->entry);
1199
1200         goto out;
1201 }
1202
1203 static int ntb_transport_rxc_db(void *data, int db_num)
1204 {
1205         struct ntb_transport_qp *qp = data;
1206         int rc, i;
1207
1208         dev_dbg(&ntb_query_pdev(qp->ndev)->dev, "%s: doorbell %d received\n",
1209                 __func__, db_num);
1210
1211         /* Limit the number of packets processed in a single interrupt to
1212          * provide fairness to others
1213          */
1214         for (i = 0; i < qp->rx_max_entry; i++) {
1215                 rc = ntb_process_rxc(qp);
1216                 if (rc)
1217                         break;
1218         }
1219
1220         if (qp->dma_chan)
1221                 dma_async_issue_pending(qp->dma_chan);
1222
1223         return i;
1224 }
1225
1226 static void ntb_tx_copy_callback(void *data)
1227 {
1228         struct ntb_queue_entry *entry = data;
1229         struct ntb_transport_qp *qp = entry->qp;
1230         struct ntb_payload_header __iomem *hdr = entry->tx_hdr;
1231
1232         /* Ensure that the data is fully copied out before setting the flags */
1233         wmb();
1234         iowrite32(entry->flags | DESC_DONE_FLAG, &hdr->flags);
1235
1236         ntb_ring_doorbell(qp->ndev, qp->qp_num);
1237
1238         /* The entry length can only be zero if the packet is intended to be a
1239          * "link down" or similar.  Since no payload is being sent in these
1240          * cases, there is nothing to add to the completion queue.
1241          */
1242         if (entry->len > 0) {
1243                 qp->tx_bytes += entry->len;
1244
1245                 if (qp->tx_handler)
1246                         qp->tx_handler(qp, qp->cb_data, entry->cb_data,
1247                                        entry->len);
1248         }
1249
1250         ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry, &qp->tx_free_q);
1251 }
1252
1253 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset)
1254 {
1255         memcpy_toio(offset, entry->buf, entry->len);
1256
1257         ntb_tx_copy_callback(entry);
1258 }
1259
1260 static void ntb_async_tx(struct ntb_transport_qp *qp,
1261                          struct ntb_queue_entry *entry)
1262 {
1263         struct ntb_payload_header __iomem *hdr;
1264         struct dma_async_tx_descriptor *txd;
1265         struct dma_chan *chan = qp->dma_chan;
1266         struct dma_device *device;
1267         size_t dest_off, buff_off;
1268         struct dmaengine_unmap_data *unmap;
1269         dma_addr_t dest;
1270         dma_cookie_t cookie;
1271         void __iomem *offset;
1272         size_t len = entry->len;
1273         void *buf = entry->buf;
1274
1275         offset = qp->tx_mw + qp->tx_max_frame * qp->tx_index;
1276         hdr = offset + qp->tx_max_frame - sizeof(struct ntb_payload_header);
1277         entry->tx_hdr = hdr;
1278
1279         iowrite32(entry->len, &hdr->len);
1280         iowrite32((u32) qp->tx_pkts, &hdr->ver);
1281
1282         if (!chan)
1283                 goto err;
1284
1285         if (len < copy_bytes)
1286                 goto err;
1287
1288         device = chan->device;
1289         dest = qp->tx_mw_phys + qp->tx_max_frame * qp->tx_index;
1290         buff_off = (size_t) buf & ~PAGE_MASK;
1291         dest_off = (size_t) dest & ~PAGE_MASK;
1292
1293         if (!is_dma_copy_aligned(device, buff_off, dest_off, len))
1294                 goto err;
1295
1296         unmap = dmaengine_get_unmap_data(device->dev, 1, GFP_NOWAIT);
1297         if (!unmap)
1298                 goto err;
1299
1300         unmap->len = len;
1301         unmap->addr[0] = dma_map_page(device->dev, virt_to_page(buf),
1302                                       buff_off, len, DMA_TO_DEVICE);
1303         if (dma_mapping_error(device->dev, unmap->addr[0]))
1304                 goto err_get_unmap;
1305
1306         unmap->to_cnt = 1;
1307
1308         txd = device->device_prep_dma_memcpy(chan, dest, unmap->addr[0], len,
1309                                              DMA_PREP_INTERRUPT);
1310         if (!txd)
1311                 goto err_get_unmap;
1312
1313         txd->callback = ntb_tx_copy_callback;
1314         txd->callback_param = entry;
1315         dma_set_unmap(txd, unmap);
1316
1317         cookie = dmaengine_submit(txd);
1318         if (dma_submit_error(cookie))
1319                 goto err_set_unmap;
1320
1321         dmaengine_unmap_put(unmap);
1322
1323         dma_async_issue_pending(chan);
1324         qp->tx_async++;
1325
1326         return;
1327 err_set_unmap:
1328         dmaengine_unmap_put(unmap);
1329 err_get_unmap:
1330         dmaengine_unmap_put(unmap);
1331 err:
1332         ntb_memcpy_tx(entry, offset);
1333         qp->tx_memcpy++;
1334 }
1335
1336 static int ntb_process_tx(struct ntb_transport_qp *qp,
1337                           struct ntb_queue_entry *entry)
1338 {
1339         dev_dbg(&ntb_query_pdev(qp->ndev)->dev, "%lld - tx %u, entry len %d flags %x buff %p\n",
1340                 qp->tx_pkts, qp->tx_index, entry->len, entry->flags,
1341                 entry->buf);
1342         if (qp->tx_index == qp->remote_rx_info->entry) {
1343                 qp->tx_ring_full++;
1344                 return -EAGAIN;
1345         }
1346
1347         if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
1348                 if (qp->tx_handler)
1349                         qp->tx_handler(qp->cb_data, qp, NULL, -EIO);
1350
1351                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1352                              &qp->tx_free_q);
1353                 return 0;
1354         }
1355
1356         ntb_async_tx(qp, entry);
1357
1358         qp->tx_index++;
1359         qp->tx_index %= qp->tx_max_entry;
1360
1361         qp->tx_pkts++;
1362
1363         return 0;
1364 }
1365
1366 static void ntb_send_link_down(struct ntb_transport_qp *qp)
1367 {
1368         struct pci_dev *pdev = ntb_query_pdev(qp->ndev);
1369         struct ntb_queue_entry *entry;
1370         int i, rc;
1371
1372         if (qp->qp_link == NTB_LINK_DOWN)
1373                 return;
1374
1375         qp->qp_link = NTB_LINK_DOWN;
1376         dev_info(&pdev->dev, "qp %d: Link Down\n", qp->qp_num);
1377
1378         for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1379                 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1380                 if (entry)
1381                         break;
1382                 msleep(100);
1383         }
1384
1385         if (!entry)
1386                 return;
1387
1388         entry->cb_data = NULL;
1389         entry->buf = NULL;
1390         entry->len = 0;
1391         entry->flags = LINK_DOWN_FLAG;
1392
1393         rc = ntb_process_tx(qp, entry);
1394         if (rc)
1395                 dev_err(&pdev->dev, "ntb: QP%d unable to send linkdown msg\n",
1396                         qp->qp_num);
1397 }
1398
1399 /**
1400  * ntb_transport_create_queue - Create a new NTB transport layer queue
1401  * @rx_handler: receive callback function
1402  * @tx_handler: transmit callback function
1403  * @event_handler: event callback function
1404  *
1405  * Create a new NTB transport layer queue and provide the queue with a callback
1406  * routine for both transmit and receive.  The receive callback routine will be
1407  * used to pass up data when the transport has received it on the queue.   The
1408  * transmit callback routine will be called when the transport has completed the
1409  * transmission of the data on the queue and the data is ready to be freed.
1410  *
1411  * RETURNS: pointer to newly created ntb_queue, NULL on error.
1412  */
1413 struct ntb_transport_qp *
1414 ntb_transport_create_queue(void *data, struct pci_dev *pdev,
1415                            const struct ntb_queue_handlers *handlers)
1416 {
1417         struct ntb_queue_entry *entry;
1418         struct ntb_transport_qp *qp;
1419         struct ntb_transport *nt;
1420         unsigned int free_queue;
1421         int rc, i;
1422
1423         nt = ntb_find_transport(pdev);
1424         if (!nt)
1425                 goto err;
1426
1427         free_queue = ffs(nt->qp_bitmap);
1428         if (!free_queue)
1429                 goto err;
1430
1431         /* decrement free_queue to make it zero based */
1432         free_queue--;
1433
1434         clear_bit(free_queue, &nt->qp_bitmap);
1435
1436         qp = &nt->qps[free_queue];
1437         qp->cb_data = data;
1438         qp->rx_handler = handlers->rx_handler;
1439         qp->tx_handler = handlers->tx_handler;
1440         qp->event_handler = handlers->event_handler;
1441
1442         dmaengine_get();
1443         qp->dma_chan = dma_find_channel(DMA_MEMCPY);
1444         if (!qp->dma_chan) {
1445                 dmaengine_put();
1446                 dev_info(&pdev->dev, "Unable to allocate DMA channel, using CPU instead\n");
1447         }
1448
1449         for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
1450                 entry = kzalloc(sizeof(struct ntb_queue_entry), GFP_ATOMIC);
1451                 if (!entry)
1452                         goto err1;
1453
1454                 entry->qp = qp;
1455                 ntb_list_add(&qp->ntb_rx_free_q_lock, &entry->entry,
1456                              &qp->rx_free_q);
1457         }
1458
1459         for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
1460                 entry = kzalloc(sizeof(struct ntb_queue_entry), GFP_ATOMIC);
1461                 if (!entry)
1462                         goto err2;
1463
1464                 entry->qp = qp;
1465                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1466                              &qp->tx_free_q);
1467         }
1468
1469         rc = ntb_register_db_callback(qp->ndev, free_queue, qp,
1470                                       ntb_transport_rxc_db);
1471         if (rc)
1472                 goto err2;
1473
1474         dev_info(&pdev->dev, "NTB Transport QP %d created\n", qp->qp_num);
1475
1476         return qp;
1477
1478 err2:
1479         while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1480                 kfree(entry);
1481 err1:
1482         while ((entry = ntb_list_rm(&qp->ntb_rx_free_q_lock, &qp->rx_free_q)))
1483                 kfree(entry);
1484         if (qp->dma_chan)
1485                 dmaengine_put();
1486         set_bit(free_queue, &nt->qp_bitmap);
1487 err:
1488         return NULL;
1489 }
1490 EXPORT_SYMBOL_GPL(ntb_transport_create_queue);
1491
1492 /**
1493  * ntb_transport_free_queue - Frees NTB transport queue
1494  * @qp: NTB queue to be freed
1495  *
1496  * Frees NTB transport queue
1497  */
1498 void ntb_transport_free_queue(struct ntb_transport_qp *qp)
1499 {
1500         struct pci_dev *pdev;
1501         struct ntb_queue_entry *entry;
1502
1503         if (!qp)
1504                 return;
1505
1506         pdev = ntb_query_pdev(qp->ndev);
1507
1508         if (qp->dma_chan) {
1509                 struct dma_chan *chan = qp->dma_chan;
1510                 /* Putting the dma_chan to NULL will force any new traffic to be
1511                  * processed by the CPU instead of the DAM engine
1512                  */
1513                 qp->dma_chan = NULL;
1514
1515                 /* Try to be nice and wait for any queued DMA engine
1516                  * transactions to process before smashing it with a rock
1517                  */
1518                 dma_sync_wait(chan, qp->last_cookie);
1519                 dmaengine_terminate_all(chan);
1520                 dmaengine_put();
1521         }
1522
1523         ntb_unregister_db_callback(qp->ndev, qp->qp_num);
1524
1525         cancel_delayed_work_sync(&qp->link_work);
1526
1527         while ((entry = ntb_list_rm(&qp->ntb_rx_free_q_lock, &qp->rx_free_q)))
1528                 kfree(entry);
1529
1530         while ((entry = ntb_list_rm(&qp->ntb_rx_pend_q_lock, &qp->rx_pend_q))) {
1531                 dev_warn(&pdev->dev, "Freeing item from a non-empty queue\n");
1532                 kfree(entry);
1533         }
1534
1535         while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1536                 kfree(entry);
1537
1538         set_bit(qp->qp_num, &qp->transport->qp_bitmap);
1539
1540         dev_info(&pdev->dev, "NTB Transport QP %d freed\n", qp->qp_num);
1541 }
1542 EXPORT_SYMBOL_GPL(ntb_transport_free_queue);
1543
1544 /**
1545  * ntb_transport_rx_remove - Dequeues enqueued rx packet
1546  * @qp: NTB queue to be freed
1547  * @len: pointer to variable to write enqueued buffers length
1548  *
1549  * Dequeues unused buffers from receive queue.  Should only be used during
1550  * shutdown of qp.
1551  *
1552  * RETURNS: NULL error value on error, or void* for success.
1553  */
1554 void *ntb_transport_rx_remove(struct ntb_transport_qp *qp, unsigned int *len)
1555 {
1556         struct ntb_queue_entry *entry;
1557         void *buf;
1558
1559         if (!qp || qp->client_ready == NTB_LINK_UP)
1560                 return NULL;
1561
1562         entry = ntb_list_rm(&qp->ntb_rx_pend_q_lock, &qp->rx_pend_q);
1563         if (!entry)
1564                 return NULL;
1565
1566         buf = entry->cb_data;
1567         *len = entry->len;
1568
1569         ntb_list_add(&qp->ntb_rx_free_q_lock, &entry->entry, &qp->rx_free_q);
1570
1571         return buf;
1572 }
1573 EXPORT_SYMBOL_GPL(ntb_transport_rx_remove);
1574
1575 /**
1576  * ntb_transport_rx_enqueue - Enqueue a new NTB queue entry
1577  * @qp: NTB transport layer queue the entry is to be enqueued on
1578  * @cb: per buffer pointer for callback function to use
1579  * @data: pointer to data buffer that incoming packets will be copied into
1580  * @len: length of the data buffer
1581  *
1582  * Enqueue a new receive buffer onto the transport queue into which a NTB
1583  * payload can be received into.
1584  *
1585  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
1586  */
1587 int ntb_transport_rx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
1588                              unsigned int len)
1589 {
1590         struct ntb_queue_entry *entry;
1591
1592         if (!qp)
1593                 return -EINVAL;
1594
1595         entry = ntb_list_rm(&qp->ntb_rx_free_q_lock, &qp->rx_free_q);
1596         if (!entry)
1597                 return -ENOMEM;
1598
1599         entry->cb_data = cb;
1600         entry->buf = data;
1601         entry->len = len;
1602
1603         ntb_list_add(&qp->ntb_rx_pend_q_lock, &entry->entry, &qp->rx_pend_q);
1604
1605         return 0;
1606 }
1607 EXPORT_SYMBOL_GPL(ntb_transport_rx_enqueue);
1608
1609 /**
1610  * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
1611  * @qp: NTB transport layer queue the entry is to be enqueued on
1612  * @cb: per buffer pointer for callback function to use
1613  * @data: pointer to data buffer that will be sent
1614  * @len: length of the data buffer
1615  *
1616  * Enqueue a new transmit buffer onto the transport queue from which a NTB
1617  * payload will be transmitted.  This assumes that a lock is being held to
1618  * serialize access to the qp.
1619  *
1620  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
1621  */
1622 int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
1623                              unsigned int len)
1624 {
1625         struct ntb_queue_entry *entry;
1626         int rc;
1627
1628         if (!qp || qp->qp_link != NTB_LINK_UP || !len)
1629                 return -EINVAL;
1630
1631         entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1632         if (!entry) {
1633                 qp->tx_err_no_buf++;
1634                 return -ENOMEM;
1635         }
1636
1637         entry->cb_data = cb;
1638         entry->buf = data;
1639         entry->len = len;
1640         entry->flags = 0;
1641
1642         rc = ntb_process_tx(qp, entry);
1643         if (rc)
1644                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1645                              &qp->tx_free_q);
1646
1647         return rc;
1648 }
1649 EXPORT_SYMBOL_GPL(ntb_transport_tx_enqueue);
1650
1651 /**
1652  * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
1653  * @qp: NTB transport layer queue to be enabled
1654  *
1655  * Notify NTB transport layer of client readiness to use queue
1656  */
1657 void ntb_transport_link_up(struct ntb_transport_qp *qp)
1658 {
1659         if (!qp)
1660                 return;
1661
1662         qp->client_ready = NTB_LINK_UP;
1663
1664         if (qp->transport->transport_link == NTB_LINK_UP)
1665                 schedule_delayed_work(&qp->link_work, 0);
1666 }
1667 EXPORT_SYMBOL_GPL(ntb_transport_link_up);
1668
1669 /**
1670  * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
1671  * @qp: NTB transport layer queue to be disabled
1672  *
1673  * Notify NTB transport layer of client's desire to no longer receive data on
1674  * transport queue specified.  It is the client's responsibility to ensure all
1675  * entries on queue are purged or otherwise handled appropriately.
1676  */
1677 void ntb_transport_link_down(struct ntb_transport_qp *qp)
1678 {
1679         struct pci_dev *pdev;
1680         int rc, val;
1681
1682         if (!qp)
1683                 return;
1684
1685         pdev = ntb_query_pdev(qp->ndev);
1686         qp->client_ready = NTB_LINK_DOWN;
1687
1688         rc = ntb_read_local_spad(qp->ndev, QP_LINKS, &val);
1689         if (rc) {
1690                 dev_err(&pdev->dev, "Error reading spad %d\n", QP_LINKS);
1691                 return;
1692         }
1693
1694         rc = ntb_write_remote_spad(qp->ndev, QP_LINKS,
1695                                    val & ~(1 << qp->qp_num));
1696         if (rc)
1697                 dev_err(&pdev->dev, "Error writing %x to remote spad %d\n",
1698                         val & ~(1 << qp->qp_num), QP_LINKS);
1699
1700         if (qp->qp_link == NTB_LINK_UP)
1701                 ntb_send_link_down(qp);
1702         else
1703                 cancel_delayed_work_sync(&qp->link_work);
1704 }
1705 EXPORT_SYMBOL_GPL(ntb_transport_link_down);
1706
1707 /**
1708  * ntb_transport_link_query - Query transport link state
1709  * @qp: NTB transport layer queue to be queried
1710  *
1711  * Query connectivity to the remote system of the NTB transport queue
1712  *
1713  * RETURNS: true for link up or false for link down
1714  */
1715 bool ntb_transport_link_query(struct ntb_transport_qp *qp)
1716 {
1717         if (!qp)
1718                 return false;
1719
1720         return qp->qp_link == NTB_LINK_UP;
1721 }
1722 EXPORT_SYMBOL_GPL(ntb_transport_link_query);
1723
1724 /**
1725  * ntb_transport_qp_num - Query the qp number
1726  * @qp: NTB transport layer queue to be queried
1727  *
1728  * Query qp number of the NTB transport queue
1729  *
1730  * RETURNS: a zero based number specifying the qp number
1731  */
1732 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
1733 {
1734         if (!qp)
1735                 return 0;
1736
1737         return qp->qp_num;
1738 }
1739 EXPORT_SYMBOL_GPL(ntb_transport_qp_num);
1740
1741 /**
1742  * ntb_transport_max_size - Query the max payload size of a qp
1743  * @qp: NTB transport layer queue to be queried
1744  *
1745  * Query the maximum payload size permissible on the given qp
1746  *
1747  * RETURNS: the max payload size of a qp
1748  */
1749 unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp)
1750 {
1751         unsigned int max;
1752
1753         if (!qp)
1754                 return 0;
1755
1756         if (!qp->dma_chan)
1757                 return qp->tx_max_frame - sizeof(struct ntb_payload_header);
1758
1759         /* If DMA engine usage is possible, try to find the max size for that */
1760         max = qp->tx_max_frame - sizeof(struct ntb_payload_header);
1761         max -= max % (1 << qp->dma_chan->device->copy_align);
1762
1763         return max;
1764 }
1765 EXPORT_SYMBOL_GPL(ntb_transport_max_size);