kernel/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 "ntb_hw.h"
  60
  61 #define NTB_TRANSPORT_VERSION   3
  62
  63 static unsigned int transport_mtu = 0x401E;
  64 module_param(transport_mtu, uint, 0644);
  65 MODULE_PARM_DESC(transport_mtu, "Maximum size of NTB transport packets");
  66
  67 static unsigned char max_num_clients;
  68 module_param(max_num_clients, byte, 0644);
  69 MODULE_PARM_DESC(max_num_clients, "Maximum number of NTB transport clients");
  70
  71 static unsigned int copy_bytes = 1024;
  72 module_param(copy_bytes, uint, 0644);
  73 MODULE_PARM_DESC(copy_bytes, "Threshold under which NTB will use the CPU to copy instead of DMA");
  74
  75 struct ntb_queue_entry {
  76         /* ntb_queue list reference */
  77         struct list_head entry;
  78         /* pointers to data to be transfered */
  79         void *cb_data;
  80         void *buf;
  81         unsigned int len;
  82         unsigned int flags;
  83
  84         struct ntb_transport_qp *qp;
  85         union {
  86                 struct ntb_payload_header __iomem *tx_hdr;
  87                 struct ntb_payload_header *rx_hdr;
  88         };
  89         unsigned int index;
  90 };
  91
  92 struct ntb_rx_info {
  93         unsigned int entry;
  94 };
  95
  96 struct ntb_transport_qp {
  97         struct ntb_transport *transport;
  98         struct ntb_device *ndev;
  99         void *cb_data;
 100         struct dma_chan *dma_chan;
 101
 102         bool client_ready;
 103         bool qp_link;
 104         u8 qp_num;      /* Only 64 QP's are allowed.  0-63 */
 105
 106         struct ntb_rx_info __iomem *rx_info;
 107         struct ntb_rx_info *remote_rx_info;
 108
 109         void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data,
 110                            void *data, int len);
 111         struct list_head tx_free_q;
 112         spinlock_t ntb_tx_free_q_lock;
 113         void __iomem *tx_mw;
 114         dma_addr_t tx_mw_phys;
 115         unsigned int tx_index;
 116         unsigned int tx_max_entry;
 117         unsigned int tx_max_frame;
 118
 119         void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
 120                            void *data, int len);
 121         struct list_head rx_pend_q;
 122         struct list_head rx_free_q;
 123         spinlock_t ntb_rx_pend_q_lock;
 124         spinlock_t ntb_rx_free_q_lock;
 125         void *rx_buff;
 126         unsigned int rx_index;
 127         unsigned int rx_max_entry;
 128         unsigned int rx_max_frame;
 129         dma_cookie_t last_cookie;
 130
 131         void (*event_handler)(void *data, int status);
 132         struct delayed_work link_work;
 133         struct work_struct link_cleanup;
 134
 135         struct dentry *debugfs_dir;
 136         struct dentry *debugfs_stats;
 137
 138         /* Stats */
 139         u64 rx_bytes;
 140         u64 rx_pkts;
 141         u64 rx_ring_empty;
 142         u64 rx_err_no_buf;
 143         u64 rx_err_oflow;
 144         u64 rx_err_ver;
 145         u64 rx_memcpy;
 146         u64 rx_async;
 147         u64 tx_bytes;
 148         u64 tx_pkts;
 149         u64 tx_ring_full;
 150         u64 tx_err_no_buf;
 151         u64 tx_memcpy;
 152         u64 tx_async;
 153 };
 154
 155 struct ntb_transport_mw {
 156         size_t size;
 157         void *virt_addr;
 158         dma_addr_t dma_addr;
 159 };
 160
 161 struct ntb_transport_client_dev {
 162         struct list_head entry;
 163         struct device dev;
 164 };
 165
 166 struct ntb_transport {
 167         struct list_head entry;
 168         struct list_head client_devs;
 169
 170         struct ntb_device *ndev;
 171         struct ntb_transport_mw *mw;
 172         struct ntb_transport_qp *qps;
 173         unsigned int max_qps;
 174         unsigned long qp_bitmap;
 175         bool transport_link;
 176         struct delayed_work link_work;
 177         struct work_struct link_cleanup;
 178 };
 179
 180 enum {
 181         DESC_DONE_FLAG = 1 << 0,
 182         LINK_DOWN_FLAG = 1 << 1,
 183 };
 184
 185 struct ntb_payload_header {
 186         unsigned int ver;
 187         unsigned int len;
 188         unsigned int flags;
 189 };
 190
 191 enum {
 192         VERSION = 0,
 193         QP_LINKS,
 194         NUM_QPS,
 195         NUM_MWS,
 196         MW0_SZ_HIGH,
 197         MW0_SZ_LOW,
 198         MW1_SZ_HIGH,
 199         MW1_SZ_LOW,
 200         MAX_SPAD,
 201 };
 202
 203 #define QP_TO_MW(ndev, qp)      ((qp) % ntb_max_mw(ndev))
 204 #define NTB_QP_DEF_NUM_ENTRIES  100
 205 #define NTB_LINK_DOWN_TIMEOUT   10
 206
 207 static int ntb_match_bus(struct device *dev, struct device_driver *drv)
 208 {
 209         return !strncmp(dev_name(dev), drv->name, strlen(drv->name));
 210 }
 211
 212 static int ntb_client_probe(struct device *dev)
 213 {
 214         const struct ntb_client *drv = container_of(dev->driver,
 215                                                     struct ntb_client, driver);
 216         struct pci_dev *pdev = container_of(dev->parent, struct pci_dev, dev);
 217         int rc = -EINVAL;
 218
 219         get_device(dev);
 220         if (drv && drv->probe)
 221                 rc = drv->probe(pdev);
 222         if (rc)
 223                 put_device(dev);
 224
 225         return rc;
 226 }
 227
 228 static int ntb_client_remove(struct device *dev)
 229 {
 230         const struct ntb_client *drv = container_of(dev->driver,
 231                                                     struct ntb_client, driver);
 232         struct pci_dev *pdev = container_of(dev->parent, struct pci_dev, dev);
 233
 234         if (drv && drv->remove)
 235                 drv->remove(pdev);
 236
 237         put_device(dev);
 238
 239         return 0;
 240 }
 241
 242 static struct bus_type ntb_bus_type = {
 243         .name = "ntb_bus",
 244         .match = ntb_match_bus,
 245         .probe = ntb_client_probe,
 246         .remove = ntb_client_remove,
 247 };
 248
 249 static LIST_HEAD(ntb_transport_list);
 250
 251 static int ntb_bus_init(struct ntb_transport *nt)
 252 {
 253         if (list_empty(&ntb_transport_list)) {
 254                 int rc = bus_register(&ntb_bus_type);
 255                 if (rc)
 256                         return rc;
 257         }
 258
 259         list_add(&nt->entry, &ntb_transport_list);
 260
 261         return 0;
 262 }
 263
 264 static void ntb_bus_remove(struct ntb_transport *nt)
 265 {
 266         struct ntb_transport_client_dev *client_dev, *cd;
 267
 268         list_for_each_entry_safe(client_dev, cd, &nt->client_devs, entry) {
 269                 dev_err(client_dev->dev.parent, "%s still attached to bus, removing\n",
 270                         dev_name(&client_dev->dev));
 271                 list_del(&client_dev->entry);
 272                 device_unregister(&client_dev->dev);
 273         }
 274
 275         list_del(&nt->entry);
 276
 277         if (list_empty(&ntb_transport_list))
 278                 bus_unregister(&ntb_bus_type);
 279 }
 280
 281 static void ntb_client_release(struct device *dev)
 282 {
 283         struct ntb_transport_client_dev *client_dev;
 284         client_dev = container_of(dev, struct ntb_transport_client_dev, dev);
 285
 286         kfree(client_dev);
 287 }
 288
 289 /**
 290  * ntb_unregister_client_dev - Unregister NTB client device
 291  * @device_name: Name of NTB client device
 292  *
 293  * Unregister an NTB client device with the NTB transport layer
 294  */
 295 void ntb_unregister_client_dev(char *device_name)
 296 {
 297         struct ntb_transport_client_dev *client, *cd;
 298         struct ntb_transport *nt;
 299
 300         list_for_each_entry(nt, &ntb_transport_list, entry)
 301                 list_for_each_entry_safe(client, cd, &nt->client_devs, entry)
 302                         if (!strncmp(dev_name(&client->dev), device_name,
 303                                      strlen(device_name))) {
 304                                 list_del(&client->entry);
 305                                 device_unregister(&client->dev);
 306                         }
 307 }
 308 EXPORT_SYMBOL_GPL(ntb_unregister_client_dev);
 309
 310 /**
 311  * ntb_register_client_dev - Register NTB client device
 312  * @device_name: Name of NTB client device
 313  *
 314  * Register an NTB client device with the NTB transport layer
 315  */
 316 int ntb_register_client_dev(char *device_name)
 317 {
 318         struct ntb_transport_client_dev *client_dev;
 319         struct ntb_transport *nt;
 320         int rc, i = 0;
 321
 322         if (list_empty(&ntb_transport_list))
 323                 return -ENODEV;
 324
 325         list_for_each_entry(nt, &ntb_transport_list, entry) {
 326                 struct device *dev;
 327
 328                 client_dev = kzalloc(sizeof(struct ntb_transport_client_dev),
 329                                      GFP_KERNEL);
 330                 if (!client_dev) {
 331                         rc = -ENOMEM;
 332                         goto err;
 333                 }
 334
 335                 dev = &client_dev->dev;
 336
 337                 /* setup and register client devices */
 338                 dev_set_name(dev, "%s%d", device_name, i);
 339                 dev->bus = &ntb_bus_type;
 340                 dev->release = ntb_client_release;
 341                 dev->parent = &ntb_query_pdev(nt->ndev)->dev;
 342
 343                 rc = device_register(dev);
 344                 if (rc) {
 345                         kfree(client_dev);
 346                         goto err;
 347                 }
 348
 349                 list_add_tail(&client_dev->entry, &nt->client_devs);
 350                 i++;
 351         }
 352
 353         return 0;
 354
 355 err:
 356         ntb_unregister_client_dev(device_name);
 357
 358         return rc;
 359 }
 360 EXPORT_SYMBOL_GPL(ntb_register_client_dev);
 361
 362 /**
 363  * ntb_register_client - Register NTB client driver
 364  * @drv: NTB client driver to be registered
 365  *
 366  * Register an NTB client driver with the NTB transport layer
 367  *
 368  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 369  */
 370 int ntb_register_client(struct ntb_client *drv)
 371 {
 372         drv->driver.bus = &ntb_bus_type;
 373
 374         if (list_empty(&ntb_transport_list))
 375                 return -ENODEV;
 376
 377         return driver_register(&drv->driver);
 378 }
 379 EXPORT_SYMBOL_GPL(ntb_register_client);
 380
 381 /**
 382  * ntb_unregister_client - Unregister NTB client driver
 383  * @drv: NTB client driver to be unregistered
 384  *
 385  * Unregister an NTB client driver with the NTB transport layer
 386  *
 387  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 388  */
 389 void ntb_unregister_client(struct ntb_client *drv)
 390 {
 391         driver_unregister(&drv->driver);
 392 }
 393 EXPORT_SYMBOL_GPL(ntb_unregister_client);
 394
 395 static ssize_t debugfs_read(struct file *filp, char __user *ubuf, size_t count,
 396                             loff_t *offp)
 397 {
 398         struct ntb_transport_qp *qp;
 399         char *buf;
 400         ssize_t ret, out_offset, out_count;
 401
 402         out_count = 1000;
 403
 404         buf = kmalloc(out_count, GFP_KERNEL);
 405         if (!buf)
 406                 return -ENOMEM;
 407
 408         qp = filp->private_data;
 409         out_offset = 0;
 410         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 411                                "NTB QP stats\n");
 412         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 413                                "rx_bytes - \t%llu\n", qp->rx_bytes);
 414         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 415                                "rx_pkts - \t%llu\n", qp->rx_pkts);
 416         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 417                                "rx_memcpy - \t%llu\n", qp->rx_memcpy);
 418         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 419                                "rx_async - \t%llu\n", qp->rx_async);
 420         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 421                                "rx_ring_empty - %llu\n", qp->rx_ring_empty);
 422         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 423                                "rx_err_no_buf - %llu\n", qp->rx_err_no_buf);
 424         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 425                                "rx_err_oflow - \t%llu\n", qp->rx_err_oflow);
 426         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 427                                "rx_err_ver - \t%llu\n", qp->rx_err_ver);
 428         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 429                                "rx_buff - \t%p\n", qp->rx_buff);
 430         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 431                                "rx_index - \t%u\n", qp->rx_index);
 432         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 433                                "rx_max_entry - \t%u\n", qp->rx_max_entry);
 434
 435         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 436                                "tx_bytes - \t%llu\n", qp->tx_bytes);
 437         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 438                                "tx_pkts - \t%llu\n", qp->tx_pkts);
 439         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 440                                "tx_memcpy - \t%llu\n", qp->tx_memcpy);
 441         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 442                                "tx_async - \t%llu\n", qp->tx_async);
 443         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 444                                "tx_ring_full - \t%llu\n", qp->tx_ring_full);
 445         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 446                                "tx_err_no_buf - %llu\n", qp->tx_err_no_buf);
 447         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 448                                "tx_mw - \t%p\n", qp->tx_mw);
 449         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 450                                "tx_index - \t%u\n", qp->tx_index);
 451         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 452                                "tx_max_entry - \t%u\n", qp->tx_max_entry);
 453
 454         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 455                                "\nQP Link %s\n", (qp->qp_link == NTB_LINK_UP) ?
 456                                "Up" : "Down");
 457         if (out_offset > out_count)
 458                 out_offset = out_count;
 459
 460         ret = simple_read_from_buffer(ubuf, count, offp, buf, out_offset);
 461         kfree(buf);
 462         return ret;
 463 }
 464
 465 static const struct file_operations ntb_qp_debugfs_stats = {
 466         .owner = THIS_MODULE,
 467         .open = simple_open,
 468         .read = debugfs_read,
 469 };
 470
 471 static void ntb_list_add(spinlock_t *lock, struct list_head *entry,
 472                          struct list_head *list)
 473 {
 474         unsigned long flags;
 475
 476         spin_lock_irqsave(lock, flags);
 477         list_add_tail(entry, list);
 478         spin_unlock_irqrestore(lock, flags);
 479 }
 480
 481 static struct ntb_queue_entry *ntb_list_rm(spinlock_t *lock,
 482                                            struct list_head *list)
 483 {
 484         struct ntb_queue_entry *entry;
 485         unsigned long flags;
 486
 487         spin_lock_irqsave(lock, flags);
 488         if (list_empty(list)) {
 489                 entry = NULL;
 490                 goto out;
 491         }
 492         entry = list_first_entry(list, struct ntb_queue_entry, entry);
 493         list_del(&entry->entry);
 494 out:
 495         spin_unlock_irqrestore(lock, flags);
 496
 497         return entry;
 498 }
 499
 500 static void ntb_transport_setup_qp_mw(struct ntb_transport *nt,
 501                                       unsigned int qp_num)
 502 {
 503         struct ntb_transport_qp *qp = &nt->qps[qp_num];
 504         unsigned int rx_size, num_qps_mw;
 505         u8 mw_num, mw_max;
 506         unsigned int i;
 507
 508         mw_max = ntb_max_mw(nt->ndev);
 509         mw_num = QP_TO_MW(nt->ndev, qp_num);
 510
 511         WARN_ON(nt->mw[mw_num].virt_addr == NULL);
 512
 513         if (nt->max_qps % mw_max && mw_num + 1 < nt->max_qps / mw_max)
 514                 num_qps_mw = nt->max_qps / mw_max + 1;
 515         else
 516                 num_qps_mw = nt->max_qps / mw_max;
 517
 518         rx_size = (unsigned int) nt->mw[mw_num].size / num_qps_mw;
 519         qp->rx_buff = nt->mw[mw_num].virt_addr + qp_num / mw_max * rx_size;
 520         rx_size -= sizeof(struct ntb_rx_info);
 521
 522         qp->remote_rx_info = qp->rx_buff + rx_size;
 523
 524         /* Due to housekeeping, there must be atleast 2 buffs */
 525         qp->rx_max_frame = min(transport_mtu, rx_size / 2);
 526         qp->rx_max_entry = rx_size / qp->rx_max_frame;
 527         qp->rx_index = 0;
 528
 529         qp->remote_rx_info->entry = qp->rx_max_entry - 1;
 530
 531         /* setup the hdr offsets with 0's */
 532         for (i = 0; i < qp->rx_max_entry; i++) {
 533                 void *offset = qp->rx_buff + qp->rx_max_frame * (i + 1) -
 534                                sizeof(struct ntb_payload_header);
 535                 memset(offset, 0, sizeof(struct ntb_payload_header));
 536         }
 537
 538         qp->rx_pkts = 0;
 539         qp->tx_pkts = 0;
 540         qp->tx_index = 0;
 541 }
 542
 543 static void ntb_free_mw(struct ntb_transport *nt, int num_mw)
 544 {
 545         struct ntb_transport_mw *mw = &nt->mw[num_mw];
 546         struct pci_dev *pdev = ntb_query_pdev(nt->ndev);
 547
 548         if (!mw->virt_addr)
 549                 return;
 550
 551         dma_free_coherent(&pdev->dev, mw->size, mw->virt_addr, mw->dma_addr);
 552         mw->virt_addr = NULL;
 553 }
 554
 555 static int ntb_set_mw(struct ntb_transport *nt, int num_mw, unsigned int size)
 556 {
 557         struct ntb_transport_mw *mw = &nt->mw[num_mw];
 558         struct pci_dev *pdev = ntb_query_pdev(nt->ndev);
 559
 560         /* No need to re-setup */
 561         if (mw->size == ALIGN(size, 4096))
 562                 return 0;
 563
 564         if (mw->size != 0)
 565                 ntb_free_mw(nt, num_mw);
 566
 567         /* Alloc memory for receiving data.  Must be 4k aligned */
 568         mw->size = ALIGN(size, 4096);
 569
 570         mw->virt_addr = dma_alloc_coherent(&pdev->dev, mw->size, &mw->dma_addr,
 571                                            GFP_KERNEL);
 572         if (!mw->virt_addr) {
 573                 mw->size = 0;
 574                 dev_err(&pdev->dev, "Unable to allocate MW buffer of size %d\n",
 575                        (int) mw->size);
 576                 return -ENOMEM;
 577         }
 578
 579         /*
 580          * we must ensure that the memory address allocated is BAR size
 581          * aligned in order for the XLAT register to take the value. This
 582          * is a requirement of the hardware. It is recommended to setup CMA
 583          * for BAR sizes equal or greater than 4MB.
 584          */
 585         if (!IS_ALIGNED(mw->dma_addr, mw->size)) {
 586                 dev_err(&pdev->dev, "DMA memory %pad not aligned to BAR size\n",
 587                         &mw->dma_addr);
 588                 ntb_free_mw(nt, num_mw);
 589                 return -ENOMEM;
 590         }
 591
 592         /* Notify HW the memory location of the receive buffer */
 593         ntb_set_mw_addr(nt->ndev, num_mw, mw->dma_addr);
 594
 595         return 0;
 596 }
 597
 598 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
 599 {
 600         struct ntb_transport *nt = qp->transport;
 601         struct pci_dev *pdev = ntb_query_pdev(nt->ndev);
 602
 603         if (qp->qp_link == NTB_LINK_DOWN) {
 604                 cancel_delayed_work_sync(&qp->link_work);
 605                 return;
 606         }
 607
 608         if (qp->event_handler)
 609                 qp->event_handler(qp->cb_data, NTB_LINK_DOWN);
 610
 611         dev_info(&pdev->dev, "qp %d: Link Down\n", qp->qp_num);
 612         qp->qp_link = NTB_LINK_DOWN;
 613 }
 614
 615 static void ntb_qp_link_cleanup_work(struct work_struct *work)
 616 {
 617         struct ntb_transport_qp *qp = container_of(work,
 618                                                    struct ntb_transport_qp,
 619                                                    link_cleanup);
 620         struct ntb_transport *nt = qp->transport;
 621
 622         ntb_qp_link_cleanup(qp);
 623
 624         if (nt->transport_link == NTB_LINK_UP)
 625                 schedule_delayed_work(&qp->link_work,
 626                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 627 }
 628
 629 static void ntb_qp_link_down(struct ntb_transport_qp *qp)
 630 {
 631         schedule_work(&qp->link_cleanup);
 632 }
 633
 634 static void ntb_transport_link_cleanup(struct ntb_transport *nt)
 635 {
 636         int i;
 637
 638         /* Pass along the info to any clients */
 639         for (i = 0; i < nt->max_qps; i++)
 640                 if (!test_bit(i, &nt->qp_bitmap))
 641                         ntb_qp_link_cleanup(&nt->qps[i]);
 642
 643         if (nt->transport_link == NTB_LINK_DOWN)
 644                 cancel_delayed_work_sync(&nt->link_work);
 645         else
 646                 nt->transport_link = NTB_LINK_DOWN;
 647
 648         /* The scratchpad registers keep the values if the remote side
 649          * goes down, blast them now to give them a sane value the next
 650          * time they are accessed
 651          */
 652         for (i = 0; i < MAX_SPAD; i++)
 653                 ntb_write_local_spad(nt->ndev, i, 0);
 654 }
 655
 656 static void ntb_transport_link_cleanup_work(struct work_struct *work)
 657 {
 658         struct ntb_transport *nt = container_of(work, struct ntb_transport,
 659                                                 link_cleanup);
 660
 661         ntb_transport_link_cleanup(nt);
 662 }
 663
 664 static void ntb_transport_event_callback(void *data, enum ntb_hw_event event)
 665 {
 666         struct ntb_transport *nt = data;
 667
 668         switch (event) {
 669         case NTB_EVENT_HW_LINK_UP:
 670                 schedule_delayed_work(&nt->link_work, 0);
 671                 break;
 672         case NTB_EVENT_HW_LINK_DOWN:
 673                 schedule_work(&nt->link_cleanup);
 674                 break;
 675         default:
 676                 BUG();
 677         }
 678 }
 679
 680 static void ntb_transport_link_work(struct work_struct *work)
 681 {
 682         struct ntb_transport *nt = container_of(work, struct ntb_transport,
 683                                                 link_work.work);
 684         struct ntb_device *ndev = nt->ndev;
 685         struct pci_dev *pdev = ntb_query_pdev(ndev);
 686         u32 val;
 687         int rc, i;
 688
 689         /* send the local info, in the opposite order of the way we read it */
 690         for (i = 0; i < ntb_max_mw(ndev); i++) {
 691                 rc = ntb_write_remote_spad(ndev, MW0_SZ_HIGH + (i * 2),
 692                                            ntb_get_mw_size(ndev, i) >> 32);
 693                 if (rc) {
 694                         dev_err(&pdev->dev, "Error writing %u to remote spad %d\n",
 695                                 (u32)(ntb_get_mw_size(ndev, i) >> 32),
 696                                 MW0_SZ_HIGH + (i * 2));
 697                         goto out;
 698                 }
 699
 700                 rc = ntb_write_remote_spad(ndev, MW0_SZ_LOW + (i * 2),
 701                                            (u32) ntb_get_mw_size(ndev, i));
 702                 if (rc) {
 703                         dev_err(&pdev->dev, "Error writing %u to remote spad %d\n",
 704                                 (u32) ntb_get_mw_size(ndev, i),
 705                                 MW0_SZ_LOW + (i * 2));
 706                         goto out;
 707                 }
 708         }
 709
 710         rc = ntb_write_remote_spad(ndev, NUM_MWS, ntb_max_mw(ndev));
 711         if (rc) {
 712                 dev_err(&pdev->dev, "Error writing %x to remote spad %d\n",
 713                         ntb_max_mw(ndev), NUM_MWS);
 714                 goto out;
 715         }
 716
 717         rc = ntb_write_remote_spad(ndev, NUM_QPS, nt->max_qps);
 718         if (rc) {
 719                 dev_err(&pdev->dev, "Error writing %x to remote spad %d\n",
 720                         nt->max_qps, NUM_QPS);
 721                 goto out;
 722         }
 723
 724         rc = ntb_write_remote_spad(ndev, VERSION, NTB_TRANSPORT_VERSION);
 725         if (rc) {
 726                 dev_err(&pdev->dev, "Error writing %x to remote spad %d\n",
 727                         NTB_TRANSPORT_VERSION, VERSION);
 728                 goto out;
 729         }
 730
 731         /* Query the remote side for its info */
 732         rc = ntb_read_remote_spad(ndev, VERSION, &val);
 733         if (rc) {
 734                 dev_err(&pdev->dev, "Error reading remote spad %d\n", VERSION);
 735                 goto out;
 736         }
 737
 738         if (val != NTB_TRANSPORT_VERSION)
 739                 goto out;
 740         dev_dbg(&pdev->dev, "Remote version = %d\n", val);
 741
 742         rc = ntb_read_remote_spad(ndev, NUM_QPS, &val);
 743         if (rc) {
 744                 dev_err(&pdev->dev, "Error reading remote spad %d\n", NUM_QPS);
 745                 goto out;
 746         }
 747
 748         if (val != nt->max_qps)
 749                 goto out;
 750         dev_dbg(&pdev->dev, "Remote max number of qps = %d\n", val);
 751
 752         rc = ntb_read_remote_spad(ndev, NUM_MWS, &val);
 753         if (rc) {
 754                 dev_err(&pdev->dev, "Error reading remote spad %d\n", NUM_MWS);
 755                 goto out;
 756         }
 757
 758         if (val != ntb_max_mw(ndev))
 759                 goto out;
 760         dev_dbg(&pdev->dev, "Remote number of mws = %d\n", val);
 761
 762         for (i = 0; i < ntb_max_mw(ndev); i++) {
 763                 u64 val64;
 764
 765                 rc = ntb_read_remote_spad(ndev, MW0_SZ_HIGH + (i * 2), &val);
 766                 if (rc) {
 767                         dev_err(&pdev->dev, "Error reading remote spad %d\n",
 768                                 MW0_SZ_HIGH + (i * 2));
 769                         goto out1;
 770                 }
 771
 772                 val64 = (u64) val << 32;
 773
 774                 rc = ntb_read_remote_spad(ndev, MW0_SZ_LOW + (i * 2), &val);
 775                 if (rc) {
 776                         dev_err(&pdev->dev, "Error reading remote spad %d\n",
 777                                 MW0_SZ_LOW + (i * 2));
 778                         goto out1;
 779                 }
 780
 781                 val64 |= val;
 782
 783                 dev_dbg(&pdev->dev, "Remote MW%d size = %llu\n", i, val64);
 784
 785                 rc = ntb_set_mw(nt, i, val64);
 786                 if (rc)
 787                         goto out1;
 788         }
 789
 790         nt->transport_link = NTB_LINK_UP;
 791
 792         for (i = 0; i < nt->max_qps; i++) {
 793                 struct ntb_transport_qp *qp = &nt->qps[i];
 794
 795                 ntb_transport_setup_qp_mw(nt, i);
 796
 797                 if (qp->client_ready == NTB_LINK_UP)
 798                         schedule_delayed_work(&qp->link_work, 0);
 799         }
 800
 801         return;
 802
 803 out1:
 804         for (i = 0; i < ntb_max_mw(ndev); i++)
 805                 ntb_free_mw(nt, i);
 806 out:
 807         if (ntb_hw_link_status(ndev))
 808                 schedule_delayed_work(&nt->link_work,
 809                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 810 }
 811
 812 static void ntb_qp_link_work(struct work_struct *work)
 813 {
 814         struct ntb_transport_qp *qp = container_of(work,
 815                                                    struct ntb_transport_qp,
 816                                                    link_work.work);
 817         struct pci_dev *pdev = ntb_query_pdev(qp->ndev);
 818         struct ntb_transport *nt = qp->transport;
 819         int rc, val;
 820
 821         WARN_ON(nt->transport_link != NTB_LINK_UP);
 822
 823         rc = ntb_read_local_spad(nt->ndev, QP_LINKS, &val);
 824         if (rc) {
 825                 dev_err(&pdev->dev, "Error reading spad %d\n", QP_LINKS);
 826                 return;
 827         }
 828
 829         rc = ntb_write_remote_spad(nt->ndev, QP_LINKS, val | 1 << qp->qp_num);
 830         if (rc)
 831                 dev_err(&pdev->dev, "Error writing %x to remote spad %d\n",
 832                         val | 1 << qp->qp_num, QP_LINKS);
 833
 834         /* query remote spad for qp ready bits */
 835         rc = ntb_read_remote_spad(nt->ndev, QP_LINKS, &val);
 836         if (rc)
 837                 dev_err(&pdev->dev, "Error reading remote spad %d\n", QP_LINKS);
 838
 839         dev_dbg(&pdev->dev, "Remote QP link status = %x\n", val);
 840
 841         /* See if the remote side is up */
 842         if (1 << qp->qp_num & val) {
 843                 qp->qp_link = NTB_LINK_UP;
 844
 845                 dev_info(&pdev->dev, "qp %d: Link Up\n", qp->qp_num);
 846                 if (qp->event_handler)
 847                         qp->event_handler(qp->cb_data, NTB_LINK_UP);
 848         } else if (nt->transport_link == NTB_LINK_UP)
 849                 schedule_delayed_work(&qp->link_work,
 850                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 851 }
 852
 853 static int ntb_transport_init_queue(struct ntb_transport *nt,
 854                                     unsigned int qp_num)
 855 {
 856         struct ntb_transport_qp *qp;
 857         unsigned int num_qps_mw, tx_size;
 858         u8 mw_num, mw_max;
 859         u64 qp_offset;
 860
 861         mw_max = ntb_max_mw(nt->ndev);
 862         mw_num = QP_TO_MW(nt->ndev, qp_num);
 863
 864         qp = &nt->qps[qp_num];
 865         qp->qp_num = qp_num;
 866         qp->transport = nt;
 867         qp->ndev = nt->ndev;
 868         qp->qp_link = NTB_LINK_DOWN;
 869         qp->client_ready = NTB_LINK_DOWN;
 870         qp->event_handler = NULL;
 871
 872         if (nt->max_qps % mw_max && mw_num + 1 < nt->max_qps / mw_max)
 873                 num_qps_mw = nt->max_qps / mw_max + 1;
 874         else
 875                 num_qps_mw = nt->max_qps / mw_max;
 876
 877         tx_size = (unsigned int) ntb_get_mw_size(qp->ndev, mw_num) / num_qps_mw;
 878         qp_offset = qp_num / mw_max * tx_size;
 879         qp->tx_mw = ntb_get_mw_vbase(nt->ndev, mw_num) + qp_offset;
 880         if (!qp->tx_mw)
 881                 return -EINVAL;
 882
 883         qp->tx_mw_phys = ntb_get_mw_base(qp->ndev, mw_num) + qp_offset;
 884         if (!qp->tx_mw_phys)
 885                 return -EINVAL;
 886
 887         tx_size -= sizeof(struct ntb_rx_info);
 888         qp->rx_info = qp->tx_mw + tx_size;
 889
 890         /* Due to housekeeping, there must be atleast 2 buffs */
 891         qp->tx_max_frame = min(transport_mtu, tx_size / 2);
 892         qp->tx_max_entry = tx_size / qp->tx_max_frame;
 893
 894         if (ntb_query_debugfs(nt->ndev)) {
 895                 char debugfs_name[4];
 896
 897                 snprintf(debugfs_name, 4, "qp%d", qp_num);
 898                 qp->debugfs_dir = debugfs_create_dir(debugfs_name,
 899                                                  ntb_query_debugfs(nt->ndev));
 900
 901                 qp->debugfs_stats = debugfs_create_file("stats", S_IRUSR,
 902                                                         qp->debugfs_dir, qp,
 903                                                         &ntb_qp_debugfs_stats);
 904         }
 905
 906         INIT_DELAYED_WORK(&qp->link_work, ntb_qp_link_work);
 907         INIT_WORK(&qp->link_cleanup, ntb_qp_link_cleanup_work);
 908
 909         spin_lock_init(&qp->ntb_rx_pend_q_lock);
 910         spin_lock_init(&qp->ntb_rx_free_q_lock);
 911         spin_lock_init(&qp->ntb_tx_free_q_lock);
 912
 913         INIT_LIST_HEAD(&qp->rx_pend_q);
 914         INIT_LIST_HEAD(&qp->rx_free_q);
 915         INIT_LIST_HEAD(&qp->tx_free_q);
 916
 917         return 0;
 918 }
 919
 920 int ntb_transport_init(struct pci_dev *pdev)
 921 {
 922         struct ntb_transport *nt;
 923         int rc, i;
 924
 925         nt = kzalloc(sizeof(struct ntb_transport), GFP_KERNEL);
 926         if (!nt)
 927                 return -ENOMEM;
 928
 929         nt->ndev = ntb_register_transport(pdev, nt);
 930         if (!nt->ndev) {
 931                 rc = -EIO;
 932                 goto err;
 933         }
 934
 935         nt->mw = kcalloc(ntb_max_mw(nt->ndev), sizeof(struct ntb_transport_mw),
 936                          GFP_KERNEL);
 937         if (!nt->mw) {
 938                 rc = -ENOMEM;
 939                 goto err1;
 940         }
 941
 942         if (max_num_clients)
 943                 nt->max_qps = min(ntb_max_cbs(nt->ndev), max_num_clients);
 944         else
 945                 nt->max_qps = min(ntb_max_cbs(nt->ndev), ntb_max_mw(nt->ndev));
 946
 947         nt->qps = kcalloc(nt->max_qps, sizeof(struct ntb_transport_qp),
 948                           GFP_KERNEL);
 949         if (!nt->qps) {
 950                 rc = -ENOMEM;
 951                 goto err2;
 952         }
 953
 954         nt->qp_bitmap = ((u64) 1 << nt->max_qps) - 1;
 955
 956         for (i = 0; i < nt->max_qps; i++) {
 957                 rc = ntb_transport_init_queue(nt, i);
 958                 if (rc)
 959                         goto err3;
 960         }
 961
 962         INIT_DELAYED_WORK(&nt->link_work, ntb_transport_link_work);
 963         INIT_WORK(&nt->link_cleanup, ntb_transport_link_cleanup_work);
 964
 965         rc = ntb_register_event_callback(nt->ndev,
 966                                          ntb_transport_event_callback);
 967         if (rc)
 968                 goto err3;
 969
 970         INIT_LIST_HEAD(&nt->client_devs);
 971         rc = ntb_bus_init(nt);
 972         if (rc)
 973                 goto err4;
 974
 975         if (ntb_hw_link_status(nt->ndev))
 976                 schedule_delayed_work(&nt->link_work, 0);
 977
 978         return 0;
 979
 980 err4:
 981         ntb_unregister_event_callback(nt->ndev);
 982 err3:
 983         kfree(nt->qps);
 984 err2:
 985         kfree(nt->mw);
 986 err1:
 987         ntb_unregister_transport(nt->ndev);
 988 err:
 989         kfree(nt);
 990         return rc;
 991 }
 992
 993 void ntb_transport_free(void *transport)
 994 {
 995         struct ntb_transport *nt = transport;
 996         struct ntb_device *ndev = nt->ndev;
 997         int i;
 998
 999         ntb_transport_link_cleanup(nt);
1000
1001         /* verify that all the qp's are freed */
1002         for (i = 0; i < nt->max_qps; i++) {
1003                 if (!test_bit(i, &nt->qp_bitmap))
1004                         ntb_transport_free_queue(&nt->qps[i]);
1005                 debugfs_remove_recursive(nt->qps[i].debugfs_dir);
1006         }
1007
1008         ntb_bus_remove(nt);
1009
1010         cancel_delayed_work_sync(&nt->link_work);
1011
1012         ntb_unregister_event_callback(ndev);
1013
1014         for (i = 0; i < ntb_max_mw(ndev); i++)
1015                 ntb_free_mw(nt, i);
1016
1017         kfree(nt->qps);
1018         kfree(nt->mw);
1019         ntb_unregister_transport(ndev);
1020         kfree(nt);
1021 }
1022
1023 static void ntb_rx_copy_callback(void *data)
1024 {
1025         struct ntb_queue_entry *entry = data;
1026         struct ntb_transport_qp *qp = entry->qp;
1027         void *cb_data = entry->cb_data;
1028         unsigned int len = entry->len;
1029         struct ntb_payload_header *hdr = entry->rx_hdr;
1030
1031         /* Ensure that the data is fully copied out before clearing the flag */
1032         wmb();
1033         hdr->flags = 0;
1034
1035         iowrite32(entry->index, &qp->rx_info->entry);
1036
1037         ntb_list_add(&qp->ntb_rx_free_q_lock, &entry->entry, &qp->rx_free_q);
1038
1039         if (qp->rx_handler && qp->client_ready == NTB_LINK_UP)
1040                 qp->rx_handler(qp, qp->cb_data, cb_data, len);
1041 }
1042
1043 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset)
1044 {
1045         void *buf = entry->buf;
1046         size_t len = entry->len;
1047
1048         memcpy(buf, offset, len);
1049
1050         ntb_rx_copy_callback(entry);
1051 }
1052
1053 static void ntb_async_rx(struct ntb_queue_entry *entry, void *offset,
1054                          size_t len)
1055 {
1056         struct dma_async_tx_descriptor *txd;
1057         struct ntb_transport_qp *qp = entry->qp;
1058         struct dma_chan *chan = qp->dma_chan;
1059         struct dma_device *device;
1060         size_t pay_off, buff_off;
1061         struct dmaengine_unmap_data *unmap;
1062         dma_cookie_t cookie;
1063         void *buf = entry->buf;
1064
1065         entry->len = len;
1066
1067         if (!chan)
1068                 goto err;
1069
1070         if (len < copy_bytes)
1071                 goto err_wait;
1072
1073         device = chan->device;
1074         pay_off = (size_t) offset & ~PAGE_MASK;
1075         buff_off = (size_t) buf & ~PAGE_MASK;
1076
1077         if (!is_dma_copy_aligned(device, pay_off, buff_off, len))
1078                 goto err_wait;
1079
1080         unmap = dmaengine_get_unmap_data(device->dev, 2, GFP_NOWAIT);
1081         if (!unmap)
1082                 goto err_wait;
1083
1084         unmap->len = len;
1085         unmap->addr[0] = dma_map_page(device->dev, virt_to_page(offset),
1086                                       pay_off, len, DMA_TO_DEVICE);
1087         if (dma_mapping_error(device->dev, unmap->addr[0]))
1088                 goto err_get_unmap;
1089
1090         unmap->to_cnt = 1;
1091
1092         unmap->addr[1] = dma_map_page(device->dev, virt_to_page(buf),
1093                                       buff_off, len, DMA_FROM_DEVICE);
1094         if (dma_mapping_error(device->dev, unmap->addr[1]))
1095                 goto err_get_unmap;
1096
1097         unmap->from_cnt = 1;
1098
1099         txd = device->device_prep_dma_memcpy(chan, unmap->addr[1],
1100                                              unmap->addr[0], len,
1101                                              DMA_PREP_INTERRUPT);
1102         if (!txd)
1103                 goto err_get_unmap;
1104
1105         txd->callback = ntb_rx_copy_callback;
1106         txd->callback_param = entry;
1107         dma_set_unmap(txd, unmap);
1108
1109         cookie = dmaengine_submit(txd);
1110         if (dma_submit_error(cookie))
1111                 goto err_set_unmap;
1112
1113         dmaengine_unmap_put(unmap);
1114
1115         qp->last_cookie = cookie;
1116
1117         qp->rx_async++;
1118
1119         return;
1120
1121 err_set_unmap:
1122         dmaengine_unmap_put(unmap);
1123 err_get_unmap:
1124         dmaengine_unmap_put(unmap);
1125 err_wait:
1126         /* If the callbacks come out of order, the writing of the index to the
1127          * last completed will be out of order.  This may result in the
1128          * receive stalling forever.
1129          */
1130         dma_sync_wait(chan, qp->last_cookie);
1131 err:
1132         ntb_memcpy_rx(entry, offset);
1133         qp->rx_memcpy++;
1134 }
1135
1136 static int ntb_process_rxc(struct ntb_transport_qp *qp)
1137 {
1138         struct ntb_payload_header *hdr;
1139         struct ntb_queue_entry *entry;
1140         void *offset;
1141
1142         offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
1143         hdr = offset + qp->rx_max_frame - sizeof(struct ntb_payload_header);
1144
1145         entry = ntb_list_rm(&qp->ntb_rx_pend_q_lock, &qp->rx_pend_q);
1146         if (!entry) {
1147                 dev_dbg(&ntb_query_pdev(qp->ndev)->dev,
1148                         "no buffer - HDR ver %u, len %d, flags %x\n",
1149                         hdr->ver, hdr->len, hdr->flags);
1150                 qp->rx_err_no_buf++;
1151                 return -ENOMEM;
1152         }
1153
1154         if (!(hdr->flags & DESC_DONE_FLAG)) {
1155                 ntb_list_add(&qp->ntb_rx_pend_q_lock, &entry->entry,
1156                              &qp->rx_pend_q);
1157                 qp->rx_ring_empty++;
1158                 return -EAGAIN;
1159         }
1160
1161         if (hdr->ver != (u32) qp->rx_pkts) {
1162                 dev_dbg(&ntb_query_pdev(qp->ndev)->dev,
1163                         "qp %d: version mismatch, expected %llu - got %u\n",
1164                         qp->qp_num, qp->rx_pkts, hdr->ver);
1165                 ntb_list_add(&qp->ntb_rx_pend_q_lock, &entry->entry,
1166                              &qp->rx_pend_q);
1167                 qp->rx_err_ver++;
1168                 return -EIO;
1169         }
1170
1171         if (hdr->flags & LINK_DOWN_FLAG) {
1172                 ntb_qp_link_down(qp);
1173
1174                 goto err;
1175         }
1176
1177         dev_dbg(&ntb_query_pdev(qp->ndev)->dev,
1178                 "rx offset %u, ver %u - %d payload received, buf size %d\n",
1179                 qp->rx_index, hdr->ver, hdr->len, entry->len);
1180
1181         qp->rx_bytes += hdr->len;
1182         qp->rx_pkts++;
1183
1184         if (hdr->len > entry->len) {
1185                 qp->rx_err_oflow++;
1186                 dev_dbg(&ntb_query_pdev(qp->ndev)->dev,
1187                         "RX overflow! Wanted %d got %d\n",
1188                         hdr->len, entry->len);
1189
1190                 goto err;
1191         }
1192
1193         entry->index = qp->rx_index;
1194         entry->rx_hdr = hdr;
1195
1196         ntb_async_rx(entry, offset, hdr->len);
1197
1198 out:
1199         qp->rx_index++;
1200         qp->rx_index %= qp->rx_max_entry;
1201
1202         return 0;
1203
1204 err:
1205         ntb_list_add(&qp->ntb_rx_pend_q_lock, &entry->entry, &qp->rx_pend_q);
1206         /* Ensure that the data is fully copied out before clearing the flag */
1207         wmb();
1208         hdr->flags = 0;
1209         iowrite32(qp->rx_index, &qp->rx_info->entry);
1210
1211         goto out;
1212 }
1213
1214 static int ntb_transport_rxc_db(void *data, int db_num)
1215 {
1216         struct ntb_transport_qp *qp = data;
1217         int rc, i;
1218
1219         dev_dbg(&ntb_query_pdev(qp->ndev)->dev, "%s: doorbell %d received\n",
1220                 __func__, db_num);
1221
1222         /* Limit the number of packets processed in a single interrupt to
1223          * provide fairness to others
1224          */
1225         for (i = 0; i < qp->rx_max_entry; i++) {
1226                 rc = ntb_process_rxc(qp);
1227                 if (rc)
1228                         break;
1229         }
1230
1231         if (qp->dma_chan)
1232                 dma_async_issue_pending(qp->dma_chan);
1233
1234         return i;
1235 }
1236
1237 static void ntb_tx_copy_callback(void *data)
1238 {
1239         struct ntb_queue_entry *entry = data;
1240         struct ntb_transport_qp *qp = entry->qp;
1241         struct ntb_payload_header __iomem *hdr = entry->tx_hdr;
1242
1243         /* Ensure that the data is fully copied out before setting the flags */
1244         wmb();
1245         iowrite32(entry->flags | DESC_DONE_FLAG, &hdr->flags);
1246
1247         ntb_ring_doorbell(qp->ndev, qp->qp_num);
1248
1249         /* The entry length can only be zero if the packet is intended to be a
1250          * "link down" or similar.  Since no payload is being sent in these
1251          * cases, there is nothing to add to the completion queue.
1252          */
1253         if (entry->len > 0) {
1254                 qp->tx_bytes += entry->len;
1255
1256                 if (qp->tx_handler)
1257                         qp->tx_handler(qp, qp->cb_data, entry->cb_data,
1258                                        entry->len);
1259         }
1260
1261         ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry, &qp->tx_free_q);
1262 }
1263
1264 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset)
1265 {
1266         memcpy_toio(offset, entry->buf, entry->len);
1267
1268         ntb_tx_copy_callback(entry);
1269 }
1270
1271 static void ntb_async_tx(struct ntb_transport_qp *qp,
1272                          struct ntb_queue_entry *entry)
1273 {
1274         struct ntb_payload_header __iomem *hdr;
1275         struct dma_async_tx_descriptor *txd;
1276         struct dma_chan *chan = qp->dma_chan;
1277         struct dma_device *device;
1278         size_t dest_off, buff_off;
1279         struct dmaengine_unmap_data *unmap;
1280         dma_addr_t dest;
1281         dma_cookie_t cookie;
1282         void __iomem *offset;
1283         size_t len = entry->len;
1284         void *buf = entry->buf;
1285
1286         offset = qp->tx_mw + qp->tx_max_frame * qp->tx_index;
1287         hdr = offset + qp->tx_max_frame - sizeof(struct ntb_payload_header);
1288         entry->tx_hdr = hdr;
1289
1290         iowrite32(entry->len, &hdr->len);
1291         iowrite32((u32) qp->tx_pkts, &hdr->ver);
1292
1293         if (!chan)
1294                 goto err;
1295
1296         if (len < copy_bytes)
1297                 goto err;
1298
1299         device = chan->device;
1300         dest = qp->tx_mw_phys + qp->tx_max_frame * qp->tx_index;
1301         buff_off = (size_t) buf & ~PAGE_MASK;
1302         dest_off = (size_t) dest & ~PAGE_MASK;
1303
1304         if (!is_dma_copy_aligned(device, buff_off, dest_off, len))
1305                 goto err;
1306
1307         unmap = dmaengine_get_unmap_data(device->dev, 1, GFP_NOWAIT);
1308         if (!unmap)
1309                 goto err;
1310
1311         unmap->len = len;
1312         unmap->addr[0] = dma_map_page(device->dev, virt_to_page(buf),
1313                                       buff_off, len, DMA_TO_DEVICE);
1314         if (dma_mapping_error(device->dev, unmap->addr[0]))
1315                 goto err_get_unmap;
1316
1317         unmap->to_cnt = 1;
1318
1319         txd = device->device_prep_dma_memcpy(chan, dest, unmap->addr[0], len,
1320                                              DMA_PREP_INTERRUPT);
1321         if (!txd)
1322                 goto err_get_unmap;
1323
1324         txd->callback = ntb_tx_copy_callback;
1325         txd->callback_param = entry;
1326         dma_set_unmap(txd, unmap);
1327
1328         cookie = dmaengine_submit(txd);
1329         if (dma_submit_error(cookie))
1330                 goto err_set_unmap;
1331
1332         dmaengine_unmap_put(unmap);
1333
1334         dma_async_issue_pending(chan);
1335         qp->tx_async++;
1336
1337         return;
1338 err_set_unmap:
1339         dmaengine_unmap_put(unmap);
1340 err_get_unmap:
1341         dmaengine_unmap_put(unmap);
1342 err:
1343         ntb_memcpy_tx(entry, offset);
1344         qp->tx_memcpy++;
1345 }
1346
1347 static int ntb_process_tx(struct ntb_transport_qp *qp,
1348                           struct ntb_queue_entry *entry)
1349 {
1350         dev_dbg(&ntb_query_pdev(qp->ndev)->dev, "%lld - tx %u, entry len %d flags %x buff %p\n",
1351                 qp->tx_pkts, qp->tx_index, entry->len, entry->flags,
1352                 entry->buf);
1353         if (qp->tx_index == qp->remote_rx_info->entry) {
1354                 qp->tx_ring_full++;
1355                 return -EAGAIN;
1356         }
1357
1358         if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
1359                 if (qp->tx_handler)
1360                         qp->tx_handler(qp->cb_data, qp, NULL, -EIO);
1361
1362                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1363                              &qp->tx_free_q);
1364                 return 0;
1365         }
1366
1367         ntb_async_tx(qp, entry);
1368
1369         qp->tx_index++;
1370         qp->tx_index %= qp->tx_max_entry;
1371
1372         qp->tx_pkts++;
1373
1374         return 0;
1375 }
1376
1377 static void ntb_send_link_down(struct ntb_transport_qp *qp)
1378 {
1379         struct pci_dev *pdev = ntb_query_pdev(qp->ndev);
1380         struct ntb_queue_entry *entry;
1381         int i, rc;
1382
1383         if (qp->qp_link == NTB_LINK_DOWN)
1384                 return;
1385
1386         qp->qp_link = NTB_LINK_DOWN;
1387         dev_info(&pdev->dev, "qp %d: Link Down\n", qp->qp_num);
1388
1389         for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1390                 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1391                 if (entry)
1392                         break;
1393                 msleep(100);
1394         }
1395
1396         if (!entry)
1397                 return;
1398
1399         entry->cb_data = NULL;
1400         entry->buf = NULL;
1401         entry->len = 0;
1402         entry->flags = LINK_DOWN_FLAG;
1403
1404         rc = ntb_process_tx(qp, entry);
1405         if (rc)
1406                 dev_err(&pdev->dev, "ntb: QP%d unable to send linkdown msg\n",
1407                         qp->qp_num);
1408 }
1409
1410 /**
1411  * ntb_transport_create_queue - Create a new NTB transport layer queue
1412  * @rx_handler: receive callback function
1413  * @tx_handler: transmit callback function
1414  * @event_handler: event callback function
1415  *
1416  * Create a new NTB transport layer queue and provide the queue with a callback
1417  * routine for both transmit and receive.  The receive callback routine will be
1418  * used to pass up data when the transport has received it on the queue.   The
1419  * transmit callback routine will be called when the transport has completed the
1420  * transmission of the data on the queue and the data is ready to be freed.
1421  *
1422  * RETURNS: pointer to newly created ntb_queue, NULL on error.
1423  */
1424 struct ntb_transport_qp *
1425 ntb_transport_create_queue(void *data, struct pci_dev *pdev,
1426                            const struct ntb_queue_handlers *handlers)
1427 {
1428         struct ntb_queue_entry *entry;
1429         struct ntb_transport_qp *qp;
1430         struct ntb_transport *nt;
1431         unsigned int free_queue;
1432         int rc, i;
1433
1434         nt = ntb_find_transport(pdev);
1435         if (!nt)
1436                 goto err;
1437
1438         free_queue = ffs(nt->qp_bitmap);
1439         if (!free_queue)
1440                 goto err;
1441
1442         /* decrement free_queue to make it zero based */
1443         free_queue--;
1444
1445         clear_bit(free_queue, &nt->qp_bitmap);
1446
1447         qp = &nt->qps[free_queue];
1448         qp->cb_data = data;
1449         qp->rx_handler = handlers->rx_handler;
1450         qp->tx_handler = handlers->tx_handler;
1451         qp->event_handler = handlers->event_handler;
1452
1453         dmaengine_get();
1454         qp->dma_chan = dma_find_channel(DMA_MEMCPY);
1455         if (!qp->dma_chan) {
1456                 dmaengine_put();
1457                 dev_info(&pdev->dev, "Unable to allocate DMA channel, using CPU instead\n");
1458         }
1459
1460         for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
1461                 entry = kzalloc(sizeof(struct ntb_queue_entry), GFP_ATOMIC);
1462                 if (!entry)
1463                         goto err1;
1464
1465                 entry->qp = qp;
1466                 ntb_list_add(&qp->ntb_rx_free_q_lock, &entry->entry,
1467                              &qp->rx_free_q);
1468         }
1469
1470         for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
1471                 entry = kzalloc(sizeof(struct ntb_queue_entry), GFP_ATOMIC);
1472                 if (!entry)
1473                         goto err2;
1474
1475                 entry->qp = qp;
1476                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1477                              &qp->tx_free_q);
1478         }
1479
1480         rc = ntb_register_db_callback(qp->ndev, free_queue, qp,
1481                                       ntb_transport_rxc_db);
1482         if (rc)
1483                 goto err2;
1484
1485         dev_info(&pdev->dev, "NTB Transport QP %d created\n", qp->qp_num);
1486
1487         return qp;
1488
1489 err2:
1490         while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1491                 kfree(entry);
1492 err1:
1493         while ((entry = ntb_list_rm(&qp->ntb_rx_free_q_lock, &qp->rx_free_q)))
1494                 kfree(entry);
1495         if (qp->dma_chan)
1496                 dmaengine_put();
1497         set_bit(free_queue, &nt->qp_bitmap);
1498 err:
1499         return NULL;
1500 }
1501 EXPORT_SYMBOL_GPL(ntb_transport_create_queue);
1502
1503 /**
1504  * ntb_transport_free_queue - Frees NTB transport queue
1505  * @qp: NTB queue to be freed
1506  *
1507  * Frees NTB transport queue
1508  */
1509 void ntb_transport_free_queue(struct ntb_transport_qp *qp)
1510 {
1511         struct pci_dev *pdev;
1512         struct ntb_queue_entry *entry;
1513
1514         if (!qp)
1515                 return;
1516
1517         pdev = ntb_query_pdev(qp->ndev);
1518
1519         if (qp->dma_chan) {
1520                 struct dma_chan *chan = qp->dma_chan;
1521                 /* Putting the dma_chan to NULL will force any new traffic to be
1522                  * processed by the CPU instead of the DAM engine
1523                  */
1524                 qp->dma_chan = NULL;
1525
1526                 /* Try to be nice and wait for any queued DMA engine
1527                  * transactions to process before smashing it with a rock
1528                  */
1529                 dma_sync_wait(chan, qp->last_cookie);
1530                 dmaengine_terminate_all(chan);
1531                 dmaengine_put();
1532         }
1533
1534         ntb_unregister_db_callback(qp->ndev, qp->qp_num);
1535
1536         cancel_delayed_work_sync(&qp->link_work);
1537
1538         while ((entry = ntb_list_rm(&qp->ntb_rx_free_q_lock, &qp->rx_free_q)))
1539                 kfree(entry);
1540
1541         while ((entry = ntb_list_rm(&qp->ntb_rx_pend_q_lock, &qp->rx_pend_q))) {
1542                 dev_warn(&pdev->dev, "Freeing item from a non-empty queue\n");
1543                 kfree(entry);
1544         }
1545
1546         while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1547                 kfree(entry);
1548
1549         set_bit(qp->qp_num, &qp->transport->qp_bitmap);
1550
1551         dev_info(&pdev->dev, "NTB Transport QP %d freed\n", qp->qp_num);
1552 }
1553 EXPORT_SYMBOL_GPL(ntb_transport_free_queue);
1554
1555 /**
1556  * ntb_transport_rx_remove - Dequeues enqueued rx packet
1557  * @qp: NTB queue to be freed
1558  * @len: pointer to variable to write enqueued buffers length
1559  *
1560  * Dequeues unused buffers from receive queue.  Should only be used during
1561  * shutdown of qp.
1562  *
1563  * RETURNS: NULL error value on error, or void* for success.
1564  */
1565 void *ntb_transport_rx_remove(struct ntb_transport_qp *qp, unsigned int *len)
1566 {
1567         struct ntb_queue_entry *entry;
1568         void *buf;
1569
1570         if (!qp || qp->client_ready == NTB_LINK_UP)
1571                 return NULL;
1572
1573         entry = ntb_list_rm(&qp->ntb_rx_pend_q_lock, &qp->rx_pend_q);
1574         if (!entry)
1575                 return NULL;
1576
1577         buf = entry->cb_data;
1578         *len = entry->len;
1579
1580         ntb_list_add(&qp->ntb_rx_free_q_lock, &entry->entry, &qp->rx_free_q);
1581
1582         return buf;
1583 }
1584 EXPORT_SYMBOL_GPL(ntb_transport_rx_remove);
1585
1586 /**
1587  * ntb_transport_rx_enqueue - Enqueue a new NTB queue entry
1588  * @qp: NTB transport layer queue the entry is to be enqueued on
1589  * @cb: per buffer pointer for callback function to use
1590  * @data: pointer to data buffer that incoming packets will be copied into
1591  * @len: length of the data buffer
1592  *
1593  * Enqueue a new receive buffer onto the transport queue into which a NTB
1594  * payload can be received into.
1595  *
1596  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
1597  */
1598 int ntb_transport_rx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
1599                              unsigned int len)
1600 {
1601         struct ntb_queue_entry *entry;
1602
1603         if (!qp)
1604                 return -EINVAL;
1605
1606         entry = ntb_list_rm(&qp->ntb_rx_free_q_lock, &qp->rx_free_q);
1607         if (!entry)
1608                 return -ENOMEM;
1609
1610         entry->cb_data = cb;
1611         entry->buf = data;
1612         entry->len = len;
1613
1614         ntb_list_add(&qp->ntb_rx_pend_q_lock, &entry->entry, &qp->rx_pend_q);
1615
1616         return 0;
1617 }
1618 EXPORT_SYMBOL_GPL(ntb_transport_rx_enqueue);
1619
1620 /**
1621  * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
1622  * @qp: NTB transport layer queue the entry is to be enqueued on
1623  * @cb: per buffer pointer for callback function to use
1624  * @data: pointer to data buffer that will be sent
1625  * @len: length of the data buffer
1626  *
1627  * Enqueue a new transmit buffer onto the transport queue from which a NTB
1628  * payload will be transmitted.  This assumes that a lock is being held to
1629  * serialize access to the qp.
1630  *
1631  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
1632  */
1633 int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
1634                              unsigned int len)
1635 {
1636         struct ntb_queue_entry *entry;
1637         int rc;
1638
1639         if (!qp || qp->qp_link != NTB_LINK_UP || !len)
1640                 return -EINVAL;
1641
1642         entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1643         if (!entry) {
1644                 qp->tx_err_no_buf++;
1645                 return -ENOMEM;
1646         }
1647
1648         entry->cb_data = cb;
1649         entry->buf = data;
1650         entry->len = len;
1651         entry->flags = 0;
1652
1653         rc = ntb_process_tx(qp, entry);
1654         if (rc)
1655                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1656                              &qp->tx_free_q);
1657
1658         return rc;
1659 }
1660 EXPORT_SYMBOL_GPL(ntb_transport_tx_enqueue);
1661
1662 /**
1663  * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
1664  * @qp: NTB transport layer queue to be enabled
1665  *
1666  * Notify NTB transport layer of client readiness to use queue
1667  */
1668 void ntb_transport_link_up(struct ntb_transport_qp *qp)
1669 {
1670         if (!qp)
1671                 return;
1672
1673         qp->client_ready = NTB_LINK_UP;
1674
1675         if (qp->transport->transport_link == NTB_LINK_UP)
1676                 schedule_delayed_work(&qp->link_work, 0);
1677 }
1678 EXPORT_SYMBOL_GPL(ntb_transport_link_up);
1679
1680 /**
1681  * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
1682  * @qp: NTB transport layer queue to be disabled
1683  *
1684  * Notify NTB transport layer of client's desire to no longer receive data on
1685  * transport queue specified.  It is the client's responsibility to ensure all
1686  * entries on queue are purged or otherwise handled appropriately.
1687  */
1688 void ntb_transport_link_down(struct ntb_transport_qp *qp)
1689 {
1690         struct pci_dev *pdev;
1691         int rc, val;
1692
1693         if (!qp)
1694                 return;
1695
1696         pdev = ntb_query_pdev(qp->ndev);
1697         qp->client_ready = NTB_LINK_DOWN;
1698
1699         rc = ntb_read_local_spad(qp->ndev, QP_LINKS, &val);
1700         if (rc) {
1701                 dev_err(&pdev->dev, "Error reading spad %d\n", QP_LINKS);
1702                 return;
1703         }
1704
1705         rc = ntb_write_remote_spad(qp->ndev, QP_LINKS,
1706                                    val & ~(1 << qp->qp_num));
1707         if (rc)
1708                 dev_err(&pdev->dev, "Error writing %x to remote spad %d\n",
1709                         val & ~(1 << qp->qp_num), QP_LINKS);
1710
1711         if (qp->qp_link == NTB_LINK_UP)
1712                 ntb_send_link_down(qp);
1713         else
1714                 cancel_delayed_work_sync(&qp->link_work);
1715 }
1716 EXPORT_SYMBOL_GPL(ntb_transport_link_down);
1717
1718 /**
1719  * ntb_transport_link_query - Query transport link state
1720  * @qp: NTB transport layer queue to be queried
1721  *
1722  * Query connectivity to the remote system of the NTB transport queue
1723  *
1724  * RETURNS: true for link up or false for link down
1725  */
1726 bool ntb_transport_link_query(struct ntb_transport_qp *qp)
1727 {
1728         if (!qp)
1729                 return false;
1730
1731         return qp->qp_link == NTB_LINK_UP;
1732 }
1733 EXPORT_SYMBOL_GPL(ntb_transport_link_query);
1734
1735 /**
1736  * ntb_transport_qp_num - Query the qp number
1737  * @qp: NTB transport layer queue to be queried
1738  *
1739  * Query qp number of the NTB transport queue
1740  *
1741  * RETURNS: a zero based number specifying the qp number
1742  */
1743 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
1744 {
1745         if (!qp)
1746                 return 0;
1747
1748         return qp->qp_num;
1749 }
1750 EXPORT_SYMBOL_GPL(ntb_transport_qp_num);
1751
1752 /**
1753  * ntb_transport_max_size - Query the max payload size of a qp
1754  * @qp: NTB transport layer queue to be queried
1755  *
1756  * Query the maximum payload size permissible on the given qp
1757  *
1758  * RETURNS: the max payload size of a qp
1759  */
1760 unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp)
1761 {
1762         unsigned int max;
1763
1764         if (!qp)
1765                 return 0;
1766
1767         if (!qp->dma_chan)
1768                 return qp->tx_max_frame - sizeof(struct ntb_payload_header);
1769
1770         /* If DMA engine usage is possible, try to find the max size for that */
1771         max = qp->tx_max_frame - sizeof(struct ntb_payload_header);
1772         max -= max % (1 << qp->dma_chan->device->copy_align);
1773
1774         return max;
1775 }
1776 EXPORT_SYMBOL_GPL(ntb_transport_max_size);