drivers/ntb/ntb_transport.c

   1 /*
   2  * This file is provided under a dual BSD/GPLv2 license.  When using or
   3  *   redistributing this file, you may do so under either license.
   4  *
   5  *   GPL LICENSE SUMMARY
   6  *
   7  *   Copyright(c) 2012 Intel Corporation. All rights reserved.
   8  *   Copyright (C) 2015 EMC Corporation. All Rights Reserved.
   9  *
  10  *   This program is free software; you can redistribute it and/or modify
  11  *   it under the terms of version 2 of the GNU General Public License as
  12  *   published by the Free Software Foundation.
  13  *
  14  *   BSD LICENSE
  15  *
  16  *   Copyright(c) 2012 Intel Corporation. All rights reserved.
  17  *   Copyright (C) 2015 EMC Corporation. All Rights Reserved.
  18  *
  19  *   Redistribution and use in source and binary forms, with or without
  20  *   modification, are permitted provided that the following conditions
  21  *   are met:
  22  *
  23  *     * Redistributions of source code must retain the above copyright
  24  *       notice, this list of conditions and the following disclaimer.
  25  *     * Redistributions in binary form must reproduce the above copy
  26  *       notice, this list of conditions and the following disclaimer in
  27  *       the documentation and/or other materials provided with the
  28  *       distribution.
  29  *     * Neither the name of Intel Corporation nor the names of its
  30  *       contributors may be used to endorse or promote products derived
  31  *       from this software without specific prior written permission.
  32  *
  33  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  34  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  35  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  36  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  37  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  38  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  39  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  40  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  41  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  42  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  43  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  44  *
  45  * PCIe NTB Transport Linux driver
  46  *
  47  * Contact Information:
  48  * Jon Mason <jon.mason@intel.com>
  49  */
  50 #include <linux/debugfs.h>
  51 #include <linux/delay.h>
  52 #include <linux/dmaengine.h>
  53 #include <linux/dma-mapping.h>
  54 #include <linux/errno.h>
  55 #include <linux/export.h>
  56 #include <linux/interrupt.h>
  57 #include <linux/module.h>
  58 #include <linux/pci.h>
  59 #include <linux/slab.h>
  60 #include <linux/types.h>
  61 #include <linux/uaccess.h>
  62 #include "linux/ntb.h"
  63 #include "linux/ntb_transport.h"
  64
  65 #define NTB_TRANSPORT_VERSION   4
  66 #define NTB_TRANSPORT_VER       "4"
  67 #define NTB_TRANSPORT_NAME      "ntb_transport"
  68 #define NTB_TRANSPORT_DESC      "Software Queue-Pair Transport over NTB"
  69 #define NTB_TRANSPORT_MIN_SPADS (MW0_SZ_HIGH + 2)
  70
  71 MODULE_DESCRIPTION(NTB_TRANSPORT_DESC);
  72 MODULE_VERSION(NTB_TRANSPORT_VER);
  73 MODULE_LICENSE("Dual BSD/GPL");
  74 MODULE_AUTHOR("Intel Corporation");
  75
  76 static unsigned long max_mw_size;
  77 module_param(max_mw_size, ulong, 0644);
  78 MODULE_PARM_DESC(max_mw_size, "Limit size of large memory windows");
  79
  80 static unsigned int transport_mtu = 0x10000;
  81 module_param(transport_mtu, uint, 0644);
  82 MODULE_PARM_DESC(transport_mtu, "Maximum size of NTB transport packets");
  83
  84 static unsigned char max_num_clients;
  85 module_param(max_num_clients, byte, 0644);
  86 MODULE_PARM_DESC(max_num_clients, "Maximum number of NTB transport clients");
  87
  88 static unsigned int copy_bytes = 1024;
  89 module_param(copy_bytes, uint, 0644);
  90 MODULE_PARM_DESC(copy_bytes, "Threshold under which NTB will use the CPU to copy instead of DMA");
  91
  92 static bool use_dma;
  93 module_param(use_dma, bool, 0644);
  94 MODULE_PARM_DESC(use_dma, "Use DMA engine to perform large data copy");
  95
  96 static struct dentry *nt_debugfs_dir;
  97
  98 /* Only two-ports NTB devices are supported */
  99 #define PIDX            NTB_DEF_PEER_IDX
 100
 101 struct ntb_queue_entry {
 102         /* ntb_queue list reference */
 103         struct list_head entry;
 104         /* pointers to data to be transferred */
 105         void *cb_data;
 106         void *buf;
 107         unsigned int len;
 108         unsigned int flags;
 109         int retries;
 110         int errors;
 111         unsigned int tx_index;
 112         unsigned int rx_index;
 113
 114         struct ntb_transport_qp *qp;
 115         union {
 116                 struct ntb_payload_header __iomem *tx_hdr;
 117                 struct ntb_payload_header *rx_hdr;
 118         };
 119 };
 120
 121 struct ntb_rx_info {
 122         unsigned int entry;
 123 };
 124
 125 struct ntb_transport_qp {
 126         struct ntb_transport_ctx *transport;
 127         struct ntb_dev *ndev;
 128         void *cb_data;
 129         struct dma_chan *tx_dma_chan;
 130         struct dma_chan *rx_dma_chan;
 131
 132         bool client_ready;
 133         bool link_is_up;
 134         bool active;
 135
 136         u8 qp_num;      /* Only 64 QP's are allowed.  0-63 */
 137         u64 qp_bit;
 138
 139         struct ntb_rx_info __iomem *rx_info;
 140         struct ntb_rx_info *remote_rx_info;
 141
 142         void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data,
 143                            void *data, int len);
 144         struct list_head tx_free_q;
 145         spinlock_t ntb_tx_free_q_lock;
 146         void __iomem *tx_mw;
 147         dma_addr_t tx_mw_phys;
 148         unsigned int tx_index;
 149         unsigned int tx_max_entry;
 150         unsigned int tx_max_frame;
 151
 152         void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
 153                            void *data, int len);
 154         struct list_head rx_post_q;
 155         struct list_head rx_pend_q;
 156         struct list_head rx_free_q;
 157         /* ntb_rx_q_lock: synchronize access to rx_XXXX_q */
 158         spinlock_t ntb_rx_q_lock;
 159         void *rx_buff;
 160         unsigned int rx_index;
 161         unsigned int rx_max_entry;
 162         unsigned int rx_max_frame;
 163         unsigned int rx_alloc_entry;
 164         dma_cookie_t last_cookie;
 165         struct tasklet_struct rxc_db_work;
 166
 167         void (*event_handler)(void *data, int status);
 168         struct delayed_work link_work;
 169         struct work_struct link_cleanup;
 170
 171         struct dentry *debugfs_dir;
 172         struct dentry *debugfs_stats;
 173
 174         /* Stats */
 175         u64 rx_bytes;
 176         u64 rx_pkts;
 177         u64 rx_ring_empty;
 178         u64 rx_err_no_buf;
 179         u64 rx_err_oflow;
 180         u64 rx_err_ver;
 181         u64 rx_memcpy;
 182         u64 rx_async;
 183         u64 tx_bytes;
 184         u64 tx_pkts;
 185         u64 tx_ring_full;
 186         u64 tx_err_no_buf;
 187         u64 tx_memcpy;
 188         u64 tx_async;
 189 };
 190
 191 struct ntb_transport_mw {
 192         phys_addr_t phys_addr;
 193         resource_size_t phys_size;
 194         resource_size_t xlat_align;
 195         resource_size_t xlat_align_size;
 196         void __iomem *vbase;
 197         size_t xlat_size;
 198         size_t buff_size;
 199         void *virt_addr;
 200         dma_addr_t dma_addr;
 201 };
 202
 203 struct ntb_transport_client_dev {
 204         struct list_head entry;
 205         struct ntb_transport_ctx *nt;
 206         struct device dev;
 207 };
 208
 209 struct ntb_transport_ctx {
 210         struct list_head entry;
 211         struct list_head client_devs;
 212
 213         struct ntb_dev *ndev;
 214
 215         struct ntb_transport_mw *mw_vec;
 216         struct ntb_transport_qp *qp_vec;
 217         unsigned int mw_count;
 218         unsigned int qp_count;
 219         u64 qp_bitmap;
 220         u64 qp_bitmap_free;
 221
 222         bool link_is_up;
 223         struct delayed_work link_work;
 224         struct work_struct link_cleanup;
 225
 226         struct dentry *debugfs_node_dir;
 227 };
 228
 229 enum {
 230         DESC_DONE_FLAG = BIT(0),
 231         LINK_DOWN_FLAG = BIT(1),
 232 };
 233
 234 struct ntb_payload_header {
 235         unsigned int ver;
 236         unsigned int len;
 237         unsigned int flags;
 238 };
 239
 240 enum {
 241         VERSION = 0,
 242         QP_LINKS,
 243         NUM_QPS,
 244         NUM_MWS,
 245         MW0_SZ_HIGH,
 246         MW0_SZ_LOW,
 247 };
 248
 249 #define dev_client_dev(__dev) \
 250         container_of((__dev), struct ntb_transport_client_dev, dev)
 251
 252 #define drv_client(__drv) \
 253         container_of((__drv), struct ntb_transport_client, driver)
 254
 255 #define QP_TO_MW(nt, qp)        ((qp) % nt->mw_count)
 256 #define NTB_QP_DEF_NUM_ENTRIES  100
 257 #define NTB_LINK_DOWN_TIMEOUT   10
 258
 259 static void ntb_transport_rxc_db(unsigned long data);
 260 static const struct ntb_ctx_ops ntb_transport_ops;
 261 static struct ntb_client ntb_transport_client;
 262 static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
 263                                struct ntb_queue_entry *entry);
 264 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset);
 265 static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset);
 266 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset);
 267
 268
 269 static int ntb_transport_bus_match(struct device *dev,
 270                                    struct device_driver *drv)
 271 {
 272         return !strncmp(dev_name(dev), drv->name, strlen(drv->name));
 273 }
 274
 275 static int ntb_transport_bus_probe(struct device *dev)
 276 {
 277         const struct ntb_transport_client *client;
 278         int rc = -EINVAL;
 279
 280         get_device(dev);
 281
 282         client = drv_client(dev->driver);
 283         rc = client->probe(dev);
 284         if (rc)
 285                 put_device(dev);
 286
 287         return rc;
 288 }
 289
 290 static int ntb_transport_bus_remove(struct device *dev)
 291 {
 292         const struct ntb_transport_client *client;
 293
 294         client = drv_client(dev->driver);
 295         client->remove(dev);
 296
 297         put_device(dev);
 298
 299         return 0;
 300 }
 301
 302 static struct bus_type ntb_transport_bus = {
 303         .name = "ntb_transport",
 304         .match = ntb_transport_bus_match,
 305         .probe = ntb_transport_bus_probe,
 306         .remove = ntb_transport_bus_remove,
 307 };
 308
 309 static LIST_HEAD(ntb_transport_list);
 310
 311 static int ntb_bus_init(struct ntb_transport_ctx *nt)
 312 {
 313         list_add_tail(&nt->entry, &ntb_transport_list);
 314         return 0;
 315 }
 316
 317 static void ntb_bus_remove(struct ntb_transport_ctx *nt)
 318 {
 319         struct ntb_transport_client_dev *client_dev, *cd;
 320
 321         list_for_each_entry_safe(client_dev, cd, &nt->client_devs, entry) {
 322                 dev_err(client_dev->dev.parent, "%s still attached to bus, removing\n",
 323                         dev_name(&client_dev->dev));
 324                 list_del(&client_dev->entry);
 325                 device_unregister(&client_dev->dev);
 326         }
 327
 328         list_del(&nt->entry);
 329 }
 330
 331 static void ntb_transport_client_release(struct device *dev)
 332 {
 333         struct ntb_transport_client_dev *client_dev;
 334
 335         client_dev = dev_client_dev(dev);
 336         kfree(client_dev);
 337 }
 338
 339 /**
 340  * ntb_transport_unregister_client_dev - Unregister NTB client device
 341  * @device_name: Name of NTB client device
 342  *
 343  * Unregister an NTB client device with the NTB transport layer
 344  */
 345 void ntb_transport_unregister_client_dev(char *device_name)
 346 {
 347         struct ntb_transport_client_dev *client, *cd;
 348         struct ntb_transport_ctx *nt;
 349
 350         list_for_each_entry(nt, &ntb_transport_list, entry)
 351                 list_for_each_entry_safe(client, cd, &nt->client_devs, entry)
 352                         if (!strncmp(dev_name(&client->dev), device_name,
 353                                      strlen(device_name))) {
 354                                 list_del(&client->entry);
 355                                 device_unregister(&client->dev);
 356                         }
 357 }
 358 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client_dev);
 359
 360 /**
 361  * ntb_transport_register_client_dev - Register NTB client device
 362  * @device_name: Name of NTB client device
 363  *
 364  * Register an NTB client device with the NTB transport layer
 365  */
 366 int ntb_transport_register_client_dev(char *device_name)
 367 {
 368         struct ntb_transport_client_dev *client_dev;
 369         struct ntb_transport_ctx *nt;
 370         int node;
 371         int rc, i = 0;
 372
 373         if (list_empty(&ntb_transport_list))
 374                 return -ENODEV;
 375
 376         list_for_each_entry(nt, &ntb_transport_list, entry) {
 377                 struct device *dev;
 378
 379                 node = dev_to_node(&nt->ndev->dev);
 380
 381                 client_dev = kzalloc_node(sizeof(*client_dev),
 382                                           GFP_KERNEL, node);
 383                 if (!client_dev) {
 384                         rc = -ENOMEM;
 385                         goto err;
 386                 }
 387
 388                 dev = &client_dev->dev;
 389
 390                 /* setup and register client devices */
 391                 dev_set_name(dev, "%s%d", device_name, i);
 392                 dev->bus = &ntb_transport_bus;
 393                 dev->release = ntb_transport_client_release;
 394                 dev->parent = &nt->ndev->dev;
 395
 396                 rc = device_register(dev);
 397                 if (rc) {
 398                         put_device(dev);
 399                         goto err;
 400                 }
 401
 402                 list_add_tail(&client_dev->entry, &nt->client_devs);
 403                 i++;
 404         }
 405
 406         return 0;
 407
 408 err:
 409         ntb_transport_unregister_client_dev(device_name);
 410
 411         return rc;
 412 }
 413 EXPORT_SYMBOL_GPL(ntb_transport_register_client_dev);
 414
 415 /**
 416  * ntb_transport_register_client - Register NTB client driver
 417  * @drv: NTB client driver to be registered
 418  *
 419  * Register an NTB client driver with the NTB transport layer
 420  *
 421  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 422  */
 423 int ntb_transport_register_client(struct ntb_transport_client *drv)
 424 {
 425         drv->driver.bus = &ntb_transport_bus;
 426
 427         if (list_empty(&ntb_transport_list))
 428                 return -ENODEV;
 429
 430         return driver_register(&drv->driver);
 431 }
 432 EXPORT_SYMBOL_GPL(ntb_transport_register_client);
 433
 434 /**
 435  * ntb_transport_unregister_client - Unregister NTB client driver
 436  * @drv: NTB client driver to be unregistered
 437  *
 438  * Unregister an NTB client driver with the NTB transport layer
 439  *
 440  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 441  */
 442 void ntb_transport_unregister_client(struct ntb_transport_client *drv)
 443 {
 444         driver_unregister(&drv->driver);
 445 }
 446 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client);
 447
 448 static ssize_t debugfs_read(struct file *filp, char __user *ubuf, size_t count,
 449                             loff_t *offp)
 450 {
 451         struct ntb_transport_qp *qp;
 452         char *buf;
 453         ssize_t ret, out_offset, out_count;
 454
 455         qp = filp->private_data;
 456
 457         if (!qp || !qp->link_is_up)
 458                 return 0;
 459
 460         out_count = 1000;
 461
 462         buf = kmalloc(out_count, GFP_KERNEL);
 463         if (!buf)
 464                 return -ENOMEM;
 465
 466         out_offset = 0;
 467         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 468                                "\nNTB QP stats:\n\n");
 469         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 470                                "rx_bytes - \t%llu\n", qp->rx_bytes);
 471         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 472                                "rx_pkts - \t%llu\n", qp->rx_pkts);
 473         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 474                                "rx_memcpy - \t%llu\n", qp->rx_memcpy);
 475         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 476                                "rx_async - \t%llu\n", qp->rx_async);
 477         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 478                                "rx_ring_empty - %llu\n", qp->rx_ring_empty);
 479         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 480                                "rx_err_no_buf - %llu\n", qp->rx_err_no_buf);
 481         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 482                                "rx_err_oflow - \t%llu\n", qp->rx_err_oflow);
 483         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 484                                "rx_err_ver - \t%llu\n", qp->rx_err_ver);
 485         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 486                                "rx_buff - \t0x%p\n", qp->rx_buff);
 487         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 488                                "rx_index - \t%u\n", qp->rx_index);
 489         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 490                                "rx_max_entry - \t%u\n", qp->rx_max_entry);
 491         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 492                                "rx_alloc_entry - \t%u\n\n", qp->rx_alloc_entry);
 493
 494         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 495                                "tx_bytes - \t%llu\n", qp->tx_bytes);
 496         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 497                                "tx_pkts - \t%llu\n", qp->tx_pkts);
 498         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 499                                "tx_memcpy - \t%llu\n", qp->tx_memcpy);
 500         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 501                                "tx_async - \t%llu\n", qp->tx_async);
 502         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 503                                "tx_ring_full - \t%llu\n", qp->tx_ring_full);
 504         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 505                                "tx_err_no_buf - %llu\n", qp->tx_err_no_buf);
 506         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 507                                "tx_mw - \t0x%p\n", qp->tx_mw);
 508         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 509                                "tx_index (H) - \t%u\n", qp->tx_index);
 510         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 511                                "RRI (T) - \t%u\n",
 512                                qp->remote_rx_info->entry);
 513         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 514                                "tx_max_entry - \t%u\n", qp->tx_max_entry);
 515         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 516                                "free tx - \t%u\n",
 517                                ntb_transport_tx_free_entry(qp));
 518
 519         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 520                                "\n");
 521         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 522                                "Using TX DMA - \t%s\n",
 523                                qp->tx_dma_chan ? "Yes" : "No");
 524         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 525                                "Using RX DMA - \t%s\n",
 526                                qp->rx_dma_chan ? "Yes" : "No");
 527         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 528                                "QP Link - \t%s\n",
 529                                qp->link_is_up ? "Up" : "Down");
 530         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 531                                "\n");
 532
 533         if (out_offset > out_count)
 534                 out_offset = out_count;
 535
 536         ret = simple_read_from_buffer(ubuf, count, offp, buf, out_offset);
 537         kfree(buf);
 538         return ret;
 539 }
 540
 541 static const struct file_operations ntb_qp_debugfs_stats = {
 542         .owner = THIS_MODULE,
 543         .open = simple_open,
 544         .read = debugfs_read,
 545 };
 546
 547 static void ntb_list_add(spinlock_t *lock, struct list_head *entry,
 548                          struct list_head *list)
 549 {
 550         unsigned long flags;
 551
 552         spin_lock_irqsave(lock, flags);
 553         list_add_tail(entry, list);
 554         spin_unlock_irqrestore(lock, flags);
 555 }
 556
 557 static struct ntb_queue_entry *ntb_list_rm(spinlock_t *lock,
 558                                            struct list_head *list)
 559 {
 560         struct ntb_queue_entry *entry;
 561         unsigned long flags;
 562
 563         spin_lock_irqsave(lock, flags);
 564         if (list_empty(list)) {
 565                 entry = NULL;
 566                 goto out;
 567         }
 568         entry = list_first_entry(list, struct ntb_queue_entry, entry);
 569         list_del(&entry->entry);
 570
 571 out:
 572         spin_unlock_irqrestore(lock, flags);
 573
 574         return entry;
 575 }
 576
 577 static struct ntb_queue_entry *ntb_list_mv(spinlock_t *lock,
 578                                            struct list_head *list,
 579                                            struct list_head *to_list)
 580 {
 581         struct ntb_queue_entry *entry;
 582         unsigned long flags;
 583
 584         spin_lock_irqsave(lock, flags);
 585
 586         if (list_empty(list)) {
 587                 entry = NULL;
 588         } else {
 589                 entry = list_first_entry(list, struct ntb_queue_entry, entry);
 590                 list_move_tail(&entry->entry, to_list);
 591         }
 592
 593         spin_unlock_irqrestore(lock, flags);
 594
 595         return entry;
 596 }
 597
 598 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
 599                                      unsigned int qp_num)
 600 {
 601         struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
 602         struct ntb_transport_mw *mw;
 603         struct ntb_dev *ndev = nt->ndev;
 604         struct ntb_queue_entry *entry;
 605         unsigned int rx_size, num_qps_mw;
 606         unsigned int mw_num, mw_count, qp_count;
 607         unsigned int i;
 608         int node;
 609
 610         mw_count = nt->mw_count;
 611         qp_count = nt->qp_count;
 612
 613         mw_num = QP_TO_MW(nt, qp_num);
 614         mw = &nt->mw_vec[mw_num];
 615
 616         if (!mw->virt_addr)
 617                 return -ENOMEM;
 618
 619         if (mw_num < qp_count % mw_count)
 620                 num_qps_mw = qp_count / mw_count + 1;
 621         else
 622                 num_qps_mw = qp_count / mw_count;
 623
 624         rx_size = (unsigned int)mw->xlat_size / num_qps_mw;
 625         qp->rx_buff = mw->virt_addr + rx_size * (qp_num / mw_count);
 626         rx_size -= sizeof(struct ntb_rx_info);
 627
 628         qp->remote_rx_info = qp->rx_buff + rx_size;
 629
 630         /* Due to housekeeping, there must be atleast 2 buffs */
 631         qp->rx_max_frame = min(transport_mtu, rx_size / 2);
 632         qp->rx_max_entry = rx_size / qp->rx_max_frame;
 633         qp->rx_index = 0;
 634
 635         /*
 636          * Checking to see if we have more entries than the default.
 637          * We should add additional entries if that is the case so we
 638          * can be in sync with the transport frames.
 639          */
 640         node = dev_to_node(&ndev->dev);
 641         for (i = qp->rx_alloc_entry; i < qp->rx_max_entry; i++) {
 642                 entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
 643                 if (!entry)
 644                         return -ENOMEM;
 645
 646                 entry->qp = qp;
 647                 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
 648                              &qp->rx_free_q);
 649                 qp->rx_alloc_entry++;
 650         }
 651
 652         qp->remote_rx_info->entry = qp->rx_max_entry - 1;
 653
 654         /* setup the hdr offsets with 0's */
 655         for (i = 0; i < qp->rx_max_entry; i++) {
 656                 void *offset = (qp->rx_buff + qp->rx_max_frame * (i + 1) -
 657                                 sizeof(struct ntb_payload_header));
 658                 memset(offset, 0, sizeof(struct ntb_payload_header));
 659         }
 660
 661         qp->rx_pkts = 0;
 662         qp->tx_pkts = 0;
 663         qp->tx_index = 0;
 664
 665         return 0;
 666 }
 667
 668 static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
 669 {
 670         struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
 671         struct pci_dev *pdev = nt->ndev->pdev;
 672
 673         if (!mw->virt_addr)
 674                 return;
 675
 676         ntb_mw_clear_trans(nt->ndev, PIDX, num_mw);
 677         dma_free_coherent(&pdev->dev, mw->buff_size,
 678                           mw->virt_addr, mw->dma_addr);
 679         mw->xlat_size = 0;
 680         mw->buff_size = 0;
 681         mw->virt_addr = NULL;
 682 }
 683
 684 static int ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw,
 685                       resource_size_t size)
 686 {
 687         struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
 688         struct pci_dev *pdev = nt->ndev->pdev;
 689         size_t xlat_size, buff_size;
 690         int rc;
 691
 692         if (!size)
 693                 return -EINVAL;
 694
 695         xlat_size = round_up(size, mw->xlat_align_size);
 696         buff_size = round_up(size, mw->xlat_align);
 697
 698         /* No need to re-setup */
 699         if (mw->xlat_size == xlat_size)
 700                 return 0;
 701
 702         if (mw->buff_size)
 703                 ntb_free_mw(nt, num_mw);
 704
 705         /* Alloc memory for receiving data.  Must be aligned */
 706         mw->xlat_size = xlat_size;
 707         mw->buff_size = buff_size;
 708
 709         mw->virt_addr = dma_alloc_coherent(&pdev->dev, buff_size,
 710                                            &mw->dma_addr, GFP_KERNEL);
 711         if (!mw->virt_addr) {
 712                 mw->xlat_size = 0;
 713                 mw->buff_size = 0;
 714                 dev_err(&pdev->dev, "Unable to alloc MW buff of size %zu\n",
 715                         buff_size);
 716                 return -ENOMEM;
 717         }
 718
 719         /*
 720          * we must ensure that the memory address allocated is BAR size
 721          * aligned in order for the XLAT register to take the value. This
 722          * is a requirement of the hardware. It is recommended to setup CMA
 723          * for BAR sizes equal or greater than 4MB.
 724          */
 725         if (!IS_ALIGNED(mw->dma_addr, mw->xlat_align)) {
 726                 dev_err(&pdev->dev, "DMA memory %pad is not aligned\n",
 727                         &mw->dma_addr);
 728                 ntb_free_mw(nt, num_mw);
 729                 return -ENOMEM;
 730         }
 731
 732         /* Notify HW the memory location of the receive buffer */
 733         rc = ntb_mw_set_trans(nt->ndev, PIDX, num_mw, mw->dma_addr,
 734                               mw->xlat_size);
 735         if (rc) {
 736                 dev_err(&pdev->dev, "Unable to set mw%d translation", num_mw);
 737                 ntb_free_mw(nt, num_mw);
 738                 return -EIO;
 739         }
 740
 741         return 0;
 742 }
 743
 744 static void ntb_qp_link_context_reset(struct ntb_transport_qp *qp)
 745 {
 746         qp->link_is_up = false;
 747         qp->active = false;
 748
 749         qp->tx_index = 0;
 750         qp->rx_index = 0;
 751         qp->rx_bytes = 0;
 752         qp->rx_pkts = 0;
 753         qp->rx_ring_empty = 0;
 754         qp->rx_err_no_buf = 0;
 755         qp->rx_err_oflow = 0;
 756         qp->rx_err_ver = 0;
 757         qp->rx_memcpy = 0;
 758         qp->rx_async = 0;
 759         qp->tx_bytes = 0;
 760         qp->tx_pkts = 0;
 761         qp->tx_ring_full = 0;
 762         qp->tx_err_no_buf = 0;
 763         qp->tx_memcpy = 0;
 764         qp->tx_async = 0;
 765 }
 766
 767 static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
 768 {
 769         ntb_qp_link_context_reset(qp);
 770         if (qp->remote_rx_info)
 771                 qp->remote_rx_info->entry = qp->rx_max_entry - 1;
 772 }
 773
 774 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
 775 {
 776         struct ntb_transport_ctx *nt = qp->transport;
 777         struct pci_dev *pdev = nt->ndev->pdev;
 778
 779         dev_info(&pdev->dev, "qp %d: Link Cleanup\n", qp->qp_num);
 780
 781         cancel_delayed_work_sync(&qp->link_work);
 782         ntb_qp_link_down_reset(qp);
 783
 784         if (qp->event_handler)
 785                 qp->event_handler(qp->cb_data, qp->link_is_up);
 786 }
 787
 788 static void ntb_qp_link_cleanup_work(struct work_struct *work)
 789 {
 790         struct ntb_transport_qp *qp = container_of(work,
 791                                                    struct ntb_transport_qp,
 792                                                    link_cleanup);
 793         struct ntb_transport_ctx *nt = qp->transport;
 794
 795         ntb_qp_link_cleanup(qp);
 796
 797         if (nt->link_is_up)
 798                 schedule_delayed_work(&qp->link_work,
 799                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 800 }
 801
 802 static void ntb_qp_link_down(struct ntb_transport_qp *qp)
 803 {
 804         schedule_work(&qp->link_cleanup);
 805 }
 806
 807 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
 808 {
 809         struct ntb_transport_qp *qp;
 810         u64 qp_bitmap_alloc;
 811         unsigned int i, count;
 812
 813         qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
 814
 815         /* Pass along the info to any clients */
 816         for (i = 0; i < nt->qp_count; i++)
 817                 if (qp_bitmap_alloc & BIT_ULL(i)) {
 818                         qp = &nt->qp_vec[i];
 819                         ntb_qp_link_cleanup(qp);
 820                         cancel_work_sync(&qp->link_cleanup);
 821                         cancel_delayed_work_sync(&qp->link_work);
 822                 }
 823
 824         if (!nt->link_is_up)
 825                 cancel_delayed_work_sync(&nt->link_work);
 826
 827         /* The scratchpad registers keep the values if the remote side
 828          * goes down, blast them now to give them a sane value the next
 829          * time they are accessed
 830          */
 831         count = ntb_spad_count(nt->ndev);
 832         for (i = 0; i < count; i++)
 833                 ntb_spad_write(nt->ndev, i, 0);
 834 }
 835
 836 static void ntb_transport_link_cleanup_work(struct work_struct *work)
 837 {
 838         struct ntb_transport_ctx *nt =
 839                 container_of(work, struct ntb_transport_ctx, link_cleanup);
 840
 841         ntb_transport_link_cleanup(nt);
 842 }
 843
 844 static void ntb_transport_event_callback(void *data)
 845 {
 846         struct ntb_transport_ctx *nt = data;
 847
 848         if (ntb_link_is_up(nt->ndev, NULL, NULL) == 1)
 849                 schedule_delayed_work(&nt->link_work, 0);
 850         else
 851                 schedule_work(&nt->link_cleanup);
 852 }
 853
 854 static void ntb_transport_link_work(struct work_struct *work)
 855 {
 856         struct ntb_transport_ctx *nt =
 857                 container_of(work, struct ntb_transport_ctx, link_work.work);
 858         struct ntb_dev *ndev = nt->ndev;
 859         struct pci_dev *pdev = ndev->pdev;
 860         resource_size_t size;
 861         u32 val;
 862         int rc = 0, i, spad;
 863
 864         /* send the local info, in the opposite order of the way we read it */
 865         for (i = 0; i < nt->mw_count; i++) {
 866                 size = nt->mw_vec[i].phys_size;
 867
 868                 if (max_mw_size && size > max_mw_size)
 869                         size = max_mw_size;
 870
 871                 spad = MW0_SZ_HIGH + (i * 2);
 872                 ntb_peer_spad_write(ndev, PIDX, spad, upper_32_bits(size));
 873
 874                 spad = MW0_SZ_LOW + (i * 2);
 875                 ntb_peer_spad_write(ndev, PIDX, spad, lower_32_bits(size));
 876         }
 877
 878         ntb_peer_spad_write(ndev, PIDX, NUM_MWS, nt->mw_count);
 879
 880         ntb_peer_spad_write(ndev, PIDX, NUM_QPS, nt->qp_count);
 881
 882         ntb_peer_spad_write(ndev, PIDX, VERSION, NTB_TRANSPORT_VERSION);
 883
 884         /* Query the remote side for its info */
 885         val = ntb_spad_read(ndev, VERSION);
 886         dev_dbg(&pdev->dev, "Remote version = %d\n", val);
 887         if (val != NTB_TRANSPORT_VERSION)
 888                 goto out;
 889
 890         val = ntb_spad_read(ndev, NUM_QPS);
 891         dev_dbg(&pdev->dev, "Remote max number of qps = %d\n", val);
 892         if (val != nt->qp_count)
 893                 goto out;
 894
 895         val = ntb_spad_read(ndev, NUM_MWS);
 896         dev_dbg(&pdev->dev, "Remote number of mws = %d\n", val);
 897         if (val != nt->mw_count)
 898                 goto out;
 899
 900         for (i = 0; i < nt->mw_count; i++) {
 901                 u64 val64;
 902
 903                 val = ntb_spad_read(ndev, MW0_SZ_HIGH + (i * 2));
 904                 val64 = (u64)val << 32;
 905
 906                 val = ntb_spad_read(ndev, MW0_SZ_LOW + (i * 2));
 907                 val64 |= val;
 908
 909                 dev_dbg(&pdev->dev, "Remote MW%d size = %#llx\n", i, val64);
 910
 911                 rc = ntb_set_mw(nt, i, val64);
 912                 if (rc)
 913                         goto out1;
 914         }
 915
 916         nt->link_is_up = true;
 917
 918         for (i = 0; i < nt->qp_count; i++) {
 919                 struct ntb_transport_qp *qp = &nt->qp_vec[i];
 920
 921                 ntb_transport_setup_qp_mw(nt, i);
 922
 923                 if (qp->client_ready)
 924                         schedule_delayed_work(&qp->link_work, 0);
 925         }
 926
 927         return;
 928
 929 out1:
 930         for (i = 0; i < nt->mw_count; i++)
 931                 ntb_free_mw(nt, i);
 932
 933         /* if there's an actual failure, we should just bail */
 934         if (rc < 0)
 935                 return;
 936
 937 out:
 938         if (ntb_link_is_up(ndev, NULL, NULL) == 1)
 939                 schedule_delayed_work(&nt->link_work,
 940                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 941 }
 942
 943 static void ntb_qp_link_work(struct work_struct *work)
 944 {
 945         struct ntb_transport_qp *qp = container_of(work,
 946                                                    struct ntb_transport_qp,
 947                                                    link_work.work);
 948         struct pci_dev *pdev = qp->ndev->pdev;
 949         struct ntb_transport_ctx *nt = qp->transport;
 950         int val;
 951
 952         WARN_ON(!nt->link_is_up);
 953
 954         val = ntb_spad_read(nt->ndev, QP_LINKS);
 955
 956         ntb_peer_spad_write(nt->ndev, PIDX, QP_LINKS, val | BIT(qp->qp_num));
 957
 958         /* query remote spad for qp ready bits */
 959         dev_dbg_ratelimited(&pdev->dev, "Remote QP link status = %x\n", val);
 960
 961         /* See if the remote side is up */
 962         if (val & BIT(qp->qp_num)) {
 963                 dev_info(&pdev->dev, "qp %d: Link Up\n", qp->qp_num);
 964                 qp->link_is_up = true;
 965                 qp->active = true;
 966
 967                 if (qp->event_handler)
 968                         qp->event_handler(qp->cb_data, qp->link_is_up);
 969
 970                 if (qp->active)
 971                         tasklet_schedule(&qp->rxc_db_work);
 972         } else if (nt->link_is_up)
 973                 schedule_delayed_work(&qp->link_work,
 974                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 975 }
 976
 977 static int ntb_transport_init_queue(struct ntb_transport_ctx *nt,
 978                                     unsigned int qp_num)
 979 {
 980         struct ntb_transport_qp *qp;
 981         phys_addr_t mw_base;
 982         resource_size_t mw_size;
 983         unsigned int num_qps_mw, tx_size;
 984         unsigned int mw_num, mw_count, qp_count;
 985         u64 qp_offset;
 986
 987         mw_count = nt->mw_count;
 988         qp_count = nt->qp_count;
 989
 990         mw_num = QP_TO_MW(nt, qp_num);
 991
 992         qp = &nt->qp_vec[qp_num];
 993         qp->qp_num = qp_num;
 994         qp->transport = nt;
 995         qp->ndev = nt->ndev;
 996         qp->client_ready = false;
 997         qp->event_handler = NULL;
 998         ntb_qp_link_context_reset(qp);
 999
1000         if (mw_num < qp_count % mw_count)
1001                 num_qps_mw = qp_count / mw_count + 1;
1002         else
1003                 num_qps_mw = qp_count / mw_count;
1004
1005         mw_base = nt->mw_vec[mw_num].phys_addr;
1006         mw_size = nt->mw_vec[mw_num].phys_size;
1007
1008         if (max_mw_size && mw_size > max_mw_size)
1009                 mw_size = max_mw_size;
1010
1011         tx_size = (unsigned int)mw_size / num_qps_mw;
1012         qp_offset = tx_size * (qp_num / mw_count);
1013
1014         qp->tx_mw = nt->mw_vec[mw_num].vbase + qp_offset;
1015         if (!qp->tx_mw)
1016                 return -EINVAL;
1017
1018         qp->tx_mw_phys = mw_base + qp_offset;
1019         if (!qp->tx_mw_phys)
1020                 return -EINVAL;
1021
1022         tx_size -= sizeof(struct ntb_rx_info);
1023         qp->rx_info = qp->tx_mw + tx_size;
1024
1025         /* Due to housekeeping, there must be atleast 2 buffs */
1026         qp->tx_max_frame = min(transport_mtu, tx_size / 2);
1027         qp->tx_max_entry = tx_size / qp->tx_max_frame;
1028
1029         if (nt->debugfs_node_dir) {
1030                 char debugfs_name[4];
1031
1032                 snprintf(debugfs_name, 4, "qp%d", qp_num);
1033                 qp->debugfs_dir = debugfs_create_dir(debugfs_name,
1034                                                      nt->debugfs_node_dir);
1035
1036                 qp->debugfs_stats = debugfs_create_file("stats", S_IRUSR,
1037                                                         qp->debugfs_dir, qp,
1038                                                         &ntb_qp_debugfs_stats);
1039         } else {
1040                 qp->debugfs_dir = NULL;
1041                 qp->debugfs_stats = NULL;
1042         }
1043
1044         INIT_DELAYED_WORK(&qp->link_work, ntb_qp_link_work);
1045         INIT_WORK(&qp->link_cleanup, ntb_qp_link_cleanup_work);
1046
1047         spin_lock_init(&qp->ntb_rx_q_lock);
1048         spin_lock_init(&qp->ntb_tx_free_q_lock);
1049
1050         INIT_LIST_HEAD(&qp->rx_post_q);
1051         INIT_LIST_HEAD(&qp->rx_pend_q);
1052         INIT_LIST_HEAD(&qp->rx_free_q);
1053         INIT_LIST_HEAD(&qp->tx_free_q);
1054
1055         tasklet_init(&qp->rxc_db_work, ntb_transport_rxc_db,
1056                      (unsigned long)qp);
1057
1058         return 0;
1059 }
1060
1061 static int ntb_transport_probe(struct ntb_client *self, struct ntb_dev *ndev)
1062 {
1063         struct ntb_transport_ctx *nt;
1064         struct ntb_transport_mw *mw;
1065         unsigned int mw_count, qp_count, spad_count, max_mw_count_for_spads;
1066         u64 qp_bitmap;
1067         int node;
1068         int rc, i;
1069
1070         mw_count = ntb_peer_mw_count(ndev);
1071
1072         if (!ndev->ops->mw_set_trans) {
1073                 dev_err(&ndev->dev, "Inbound MW based NTB API is required\n");
1074                 return -EINVAL;
1075         }
1076
1077         if (ntb_db_is_unsafe(ndev))
1078                 dev_dbg(&ndev->dev,
1079                         "doorbell is unsafe, proceed anyway...\n");
1080         if (ntb_spad_is_unsafe(ndev))
1081                 dev_dbg(&ndev->dev,
1082                         "scratchpad is unsafe, proceed anyway...\n");
1083
1084         if (ntb_peer_port_count(ndev) != NTB_DEF_PEER_CNT)
1085                 dev_warn(&ndev->dev, "Multi-port NTB devices unsupported\n");
1086
1087         node = dev_to_node(&ndev->dev);
1088
1089         nt = kzalloc_node(sizeof(*nt), GFP_KERNEL, node);
1090         if (!nt)
1091                 return -ENOMEM;
1092
1093         nt->ndev = ndev;
1094         spad_count = ntb_spad_count(ndev);
1095
1096         /* Limit the MW's based on the availability of scratchpads */
1097
1098         if (spad_count < NTB_TRANSPORT_MIN_SPADS) {
1099                 nt->mw_count = 0;
1100                 rc = -EINVAL;
1101                 goto err;
1102         }
1103
1104         max_mw_count_for_spads = (spad_count - MW0_SZ_HIGH) / 2;
1105         nt->mw_count = min(mw_count, max_mw_count_for_spads);
1106
1107         nt->mw_vec = kzalloc_node(mw_count * sizeof(*nt->mw_vec),
1108                                   GFP_KERNEL, node);
1109         if (!nt->mw_vec) {
1110                 rc = -ENOMEM;
1111                 goto err;
1112         }
1113
1114         for (i = 0; i < mw_count; i++) {
1115                 mw = &nt->mw_vec[i];
1116
1117                 rc = ntb_mw_get_align(ndev, PIDX, i, &mw->xlat_align,
1118                                       &mw->xlat_align_size, NULL);
1119                 if (rc)
1120                         goto err1;
1121
1122                 rc = ntb_peer_mw_get_addr(ndev, i, &mw->phys_addr,
1123                                           &mw->phys_size);
1124                 if (rc)
1125                         goto err1;
1126
1127                 mw->vbase = ioremap_wc(mw->phys_addr, mw->phys_size);
1128                 if (!mw->vbase) {
1129                         rc = -ENOMEM;
1130                         goto err1;
1131                 }
1132
1133                 mw->buff_size = 0;
1134                 mw->xlat_size = 0;
1135                 mw->virt_addr = NULL;
1136                 mw->dma_addr = 0;
1137         }
1138
1139         qp_bitmap = ntb_db_valid_mask(ndev);
1140
1141         qp_count = ilog2(qp_bitmap);
1142         if (max_num_clients && max_num_clients < qp_count)
1143                 qp_count = max_num_clients;
1144         else if (nt->mw_count < qp_count)
1145                 qp_count = nt->mw_count;
1146
1147         qp_bitmap &= BIT_ULL(qp_count) - 1;
1148
1149         nt->qp_count = qp_count;
1150         nt->qp_bitmap = qp_bitmap;
1151         nt->qp_bitmap_free = qp_bitmap;
1152
1153         nt->qp_vec = kzalloc_node(qp_count * sizeof(*nt->qp_vec),
1154                                   GFP_KERNEL, node);
1155         if (!nt->qp_vec) {
1156                 rc = -ENOMEM;
1157                 goto err1;
1158         }
1159
1160         if (nt_debugfs_dir) {
1161                 nt->debugfs_node_dir =
1162                         debugfs_create_dir(pci_name(ndev->pdev),
1163                                            nt_debugfs_dir);
1164         }
1165
1166         for (i = 0; i < qp_count; i++) {
1167                 rc = ntb_transport_init_queue(nt, i);
1168                 if (rc)
1169                         goto err2;
1170         }
1171
1172         INIT_DELAYED_WORK(&nt->link_work, ntb_transport_link_work);
1173         INIT_WORK(&nt->link_cleanup, ntb_transport_link_cleanup_work);
1174
1175         rc = ntb_set_ctx(ndev, nt, &ntb_transport_ops);
1176         if (rc)
1177                 goto err2;
1178
1179         INIT_LIST_HEAD(&nt->client_devs);
1180         rc = ntb_bus_init(nt);
1181         if (rc)
1182                 goto err3;
1183
1184         nt->link_is_up = false;
1185         ntb_link_enable(ndev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
1186         ntb_link_event(ndev);
1187
1188         return 0;
1189
1190 err3:
1191         ntb_clear_ctx(ndev);
1192 err2:
1193         kfree(nt->qp_vec);
1194 err1:
1195         while (i--) {
1196                 mw = &nt->mw_vec[i];
1197                 iounmap(mw->vbase);
1198         }
1199         kfree(nt->mw_vec);
1200 err:
1201         kfree(nt);
1202         return rc;
1203 }
1204
1205 static void ntb_transport_free(struct ntb_client *self, struct ntb_dev *ndev)
1206 {
1207         struct ntb_transport_ctx *nt = ndev->ctx;
1208         struct ntb_transport_qp *qp;
1209         u64 qp_bitmap_alloc;
1210         int i;
1211
1212         ntb_transport_link_cleanup(nt);
1213         cancel_work_sync(&nt->link_cleanup);
1214         cancel_delayed_work_sync(&nt->link_work);
1215
1216         qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
1217
1218         /* verify that all the qp's are freed */
1219         for (i = 0; i < nt->qp_count; i++) {
1220                 qp = &nt->qp_vec[i];
1221                 if (qp_bitmap_alloc & BIT_ULL(i))
1222                         ntb_transport_free_queue(qp);
1223                 debugfs_remove_recursive(qp->debugfs_dir);
1224         }
1225
1226         ntb_link_disable(ndev);
1227         ntb_clear_ctx(ndev);
1228
1229         ntb_bus_remove(nt);
1230
1231         for (i = nt->mw_count; i--; ) {
1232                 ntb_free_mw(nt, i);
1233                 iounmap(nt->mw_vec[i].vbase);
1234         }
1235
1236         kfree(nt->qp_vec);
1237         kfree(nt->mw_vec);
1238         kfree(nt);
1239 }
1240
1241 static void ntb_complete_rxc(struct ntb_transport_qp *qp)
1242 {
1243         struct ntb_queue_entry *entry;
1244         void *cb_data;
1245         unsigned int len;
1246         unsigned long irqflags;
1247
1248         spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1249
1250         while (!list_empty(&qp->rx_post_q)) {
1251                 entry = list_first_entry(&qp->rx_post_q,
1252                                          struct ntb_queue_entry, entry);
1253                 if (!(entry->flags & DESC_DONE_FLAG))
1254                         break;
1255
1256                 entry->rx_hdr->flags = 0;
1257                 iowrite32(entry->rx_index, &qp->rx_info->entry);
1258
1259                 cb_data = entry->cb_data;
1260                 len = entry->len;
1261
1262                 list_move_tail(&entry->entry, &qp->rx_free_q);
1263
1264                 spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1265
1266                 if (qp->rx_handler && qp->client_ready)
1267                         qp->rx_handler(qp, qp->cb_data, cb_data, len);
1268
1269                 spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1270         }
1271
1272         spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1273 }
1274
1275 static void ntb_rx_copy_callback(void *data,
1276                                  const struct dmaengine_result *res)
1277 {
1278         struct ntb_queue_entry *entry = data;
1279
1280         /* we need to check DMA results if we are using DMA */
1281         if (res) {
1282                 enum dmaengine_tx_result dma_err = res->result;
1283
1284                 switch (dma_err) {
1285                 case DMA_TRANS_READ_FAILED:
1286                 case DMA_TRANS_WRITE_FAILED:
1287                         entry->errors++;
1288                 case DMA_TRANS_ABORTED:
1289                 {
1290                         struct ntb_transport_qp *qp = entry->qp;
1291                         void *offset = qp->rx_buff + qp->rx_max_frame *
1292                                         qp->rx_index;
1293
1294                         ntb_memcpy_rx(entry, offset);
1295                         qp->rx_memcpy++;
1296                         return;
1297                 }
1298
1299                 case DMA_TRANS_NOERROR:
1300                 default:
1301                         break;
1302                 }
1303         }
1304
1305         entry->flags |= DESC_DONE_FLAG;
1306
1307         ntb_complete_rxc(entry->qp);
1308 }
1309
1310 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset)
1311 {
1312         void *buf = entry->buf;
1313         size_t len = entry->len;
1314
1315         memcpy(buf, offset, len);
1316
1317         /* Ensure that the data is fully copied out before clearing the flag */
1318         wmb();
1319
1320         ntb_rx_copy_callback(entry, NULL);
1321 }
1322
1323 static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset)
1324 {
1325         struct dma_async_tx_descriptor *txd;
1326         struct ntb_transport_qp *qp = entry->qp;
1327         struct dma_chan *chan = qp->rx_dma_chan;
1328         struct dma_device *device;
1329         size_t pay_off, buff_off, len;
1330         struct dmaengine_unmap_data *unmap;
1331         dma_cookie_t cookie;
1332         void *buf = entry->buf;
1333
1334         len = entry->len;
1335         device = chan->device;
1336         pay_off = (size_t)offset & ~PAGE_MASK;
1337         buff_off = (size_t)buf & ~PAGE_MASK;
1338
1339         if (!is_dma_copy_aligned(device, pay_off, buff_off, len))
1340                 goto err;
1341
1342         unmap = dmaengine_get_unmap_data(device->dev, 2, GFP_NOWAIT);
1343         if (!unmap)
1344                 goto err;
1345
1346         unmap->len = len;
1347         unmap->addr[0] = dma_map_page(device->dev, virt_to_page(offset),
1348                                       pay_off, len, DMA_TO_DEVICE);
1349         if (dma_mapping_error(device->dev, unmap->addr[0]))
1350                 goto err_get_unmap;
1351
1352         unmap->to_cnt = 1;
1353
1354         unmap->addr[1] = dma_map_page(device->dev, virt_to_page(buf),
1355                                       buff_off, len, DMA_FROM_DEVICE);
1356         if (dma_mapping_error(device->dev, unmap->addr[1]))
1357                 goto err_get_unmap;
1358
1359         unmap->from_cnt = 1;
1360
1361         txd = device->device_prep_dma_memcpy(chan, unmap->addr[1],
1362                                              unmap->addr[0], len,
1363                                              DMA_PREP_INTERRUPT);
1364         if (!txd)
1365                 goto err_get_unmap;
1366
1367         txd->callback_result = ntb_rx_copy_callback;
1368         txd->callback_param = entry;
1369         dma_set_unmap(txd, unmap);
1370
1371         cookie = dmaengine_submit(txd);
1372         if (dma_submit_error(cookie))
1373                 goto err_set_unmap;
1374
1375         dmaengine_unmap_put(unmap);
1376
1377         qp->last_cookie = cookie;
1378
1379         qp->rx_async++;
1380
1381         return 0;
1382
1383 err_set_unmap:
1384         dmaengine_unmap_put(unmap);
1385 err_get_unmap:
1386         dmaengine_unmap_put(unmap);
1387 err:
1388         return -ENXIO;
1389 }
1390
1391 static void ntb_async_rx(struct ntb_queue_entry *entry, void *offset)
1392 {
1393         struct ntb_transport_qp *qp = entry->qp;
1394         struct dma_chan *chan = qp->rx_dma_chan;
1395         int res;
1396
1397         if (!chan)
1398                 goto err;
1399
1400         if (entry->len < copy_bytes)
1401                 goto err;
1402
1403         res = ntb_async_rx_submit(entry, offset);
1404         if (res < 0)
1405                 goto err;
1406
1407         if (!entry->retries)
1408                 qp->rx_async++;
1409
1410         return;
1411
1412 err:
1413         ntb_memcpy_rx(entry, offset);
1414         qp->rx_memcpy++;
1415 }
1416
1417 static int ntb_process_rxc(struct ntb_transport_qp *qp)
1418 {
1419         struct ntb_payload_header *hdr;
1420         struct ntb_queue_entry *entry;
1421         void *offset;
1422
1423         offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
1424         hdr = offset + qp->rx_max_frame - sizeof(struct ntb_payload_header);
1425
1426         dev_dbg(&qp->ndev->pdev->dev, "qp %d: RX ver %u len %d flags %x\n",
1427                 qp->qp_num, hdr->ver, hdr->len, hdr->flags);
1428
1429         if (!(hdr->flags & DESC_DONE_FLAG)) {
1430                 dev_dbg(&qp->ndev->pdev->dev, "done flag not set\n");
1431                 qp->rx_ring_empty++;
1432                 return -EAGAIN;
1433         }
1434
1435         if (hdr->flags & LINK_DOWN_FLAG) {
1436                 dev_dbg(&qp->ndev->pdev->dev, "link down flag set\n");
1437                 ntb_qp_link_down(qp);
1438                 hdr->flags = 0;
1439                 return -EAGAIN;
1440         }
1441
1442         if (hdr->ver != (u32)qp->rx_pkts) {
1443                 dev_dbg(&qp->ndev->pdev->dev,
1444                         "version mismatch, expected %llu - got %u\n",
1445                         qp->rx_pkts, hdr->ver);
1446                 qp->rx_err_ver++;
1447                 return -EIO;
1448         }
1449
1450         entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q);
1451         if (!entry) {
1452                 dev_dbg(&qp->ndev->pdev->dev, "no receive buffer\n");
1453                 qp->rx_err_no_buf++;
1454                 return -EAGAIN;
1455         }
1456
1457         entry->rx_hdr = hdr;
1458         entry->rx_index = qp->rx_index;
1459
1460         if (hdr->len > entry->len) {
1461                 dev_dbg(&qp->ndev->pdev->dev,
1462                         "receive buffer overflow! Wanted %d got %d\n",
1463                         hdr->len, entry->len);
1464                 qp->rx_err_oflow++;
1465
1466                 entry->len = -EIO;
1467                 entry->flags |= DESC_DONE_FLAG;
1468
1469                 ntb_complete_rxc(qp);
1470         } else {
1471                 dev_dbg(&qp->ndev->pdev->dev,
1472                         "RX OK index %u ver %u size %d into buf size %d\n",
1473                         qp->rx_index, hdr->ver, hdr->len, entry->len);
1474
1475                 qp->rx_bytes += hdr->len;
1476                 qp->rx_pkts++;
1477
1478                 entry->len = hdr->len;
1479
1480                 ntb_async_rx(entry, offset);
1481         }
1482
1483         qp->rx_index++;
1484         qp->rx_index %= qp->rx_max_entry;
1485
1486         return 0;
1487 }
1488
1489 static void ntb_transport_rxc_db(unsigned long data)
1490 {
1491         struct ntb_transport_qp *qp = (void *)data;
1492         int rc, i;
1493
1494         dev_dbg(&qp->ndev->pdev->dev, "%s: doorbell %d received\n",
1495                 __func__, qp->qp_num);
1496
1497         /* Limit the number of packets processed in a single interrupt to
1498          * provide fairness to others
1499          */
1500         for (i = 0; i < qp->rx_max_entry; i++) {
1501                 rc = ntb_process_rxc(qp);
1502                 if (rc)
1503                         break;
1504         }
1505
1506         if (i && qp->rx_dma_chan)
1507                 dma_async_issue_pending(qp->rx_dma_chan);
1508
1509         if (i == qp->rx_max_entry) {
1510                 /* there is more work to do */
1511                 if (qp->active)
1512                         tasklet_schedule(&qp->rxc_db_work);
1513         } else if (ntb_db_read(qp->ndev) & BIT_ULL(qp->qp_num)) {
1514                 /* the doorbell bit is set: clear it */
1515                 ntb_db_clear(qp->ndev, BIT_ULL(qp->qp_num));
1516                 /* ntb_db_read ensures ntb_db_clear write is committed */
1517                 ntb_db_read(qp->ndev);
1518
1519                 /* an interrupt may have arrived between finishing
1520                  * ntb_process_rxc and clearing the doorbell bit:
1521                  * there might be some more work to do.
1522                  */
1523                 if (qp->active)
1524                         tasklet_schedule(&qp->rxc_db_work);
1525         }
1526 }
1527
1528 static void ntb_tx_copy_callback(void *data,
1529                                  const struct dmaengine_result *res)
1530 {
1531         struct ntb_queue_entry *entry = data;
1532         struct ntb_transport_qp *qp = entry->qp;
1533         struct ntb_payload_header __iomem *hdr = entry->tx_hdr;
1534
1535         /* we need to check DMA results if we are using DMA */
1536         if (res) {
1537                 enum dmaengine_tx_result dma_err = res->result;
1538
1539                 switch (dma_err) {
1540                 case DMA_TRANS_READ_FAILED:
1541                 case DMA_TRANS_WRITE_FAILED:
1542                         entry->errors++;
1543                 case DMA_TRANS_ABORTED:
1544                 {
1545                         void __iomem *offset =
1546                                 qp->tx_mw + qp->tx_max_frame *
1547                                 entry->tx_index;
1548
1549                         /* resubmit via CPU */
1550                         ntb_memcpy_tx(entry, offset);
1551                         qp->tx_memcpy++;
1552                         return;
1553                 }
1554
1555                 case DMA_TRANS_NOERROR:
1556                 default:
1557                         break;
1558                 }
1559         }
1560
1561         iowrite32(entry->flags | DESC_DONE_FLAG, &hdr->flags);
1562
1563         ntb_peer_db_set(qp->ndev, BIT_ULL(qp->qp_num));
1564
1565         /* The entry length can only be zero if the packet is intended to be a
1566          * "link down" or similar.  Since no payload is being sent in these
1567          * cases, there is nothing to add to the completion queue.
1568          */
1569         if (entry->len > 0) {
1570                 qp->tx_bytes += entry->len;
1571
1572                 if (qp->tx_handler)
1573                         qp->tx_handler(qp, qp->cb_data, entry->cb_data,
1574                                        entry->len);
1575         }
1576
1577         ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry, &qp->tx_free_q);
1578 }
1579
1580 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset)
1581 {
1582 #ifdef ARCH_HAS_NOCACHE_UACCESS
1583         /*
1584          * Using non-temporal mov to improve performance on non-cached
1585          * writes, even though we aren't actually copying from user space.
1586          */
1587         __copy_from_user_inatomic_nocache(offset, entry->buf, entry->len);
1588 #else
1589         memcpy_toio(offset, entry->buf, entry->len);
1590 #endif
1591
1592         /* Ensure that the data is fully copied out before setting the flags */
1593         wmb();
1594
1595         ntb_tx_copy_callback(entry, NULL);
1596 }
1597
1598 static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
1599                                struct ntb_queue_entry *entry)
1600 {
1601         struct dma_async_tx_descriptor *txd;
1602         struct dma_chan *chan = qp->tx_dma_chan;
1603         struct dma_device *device;
1604         size_t len = entry->len;
1605         void *buf = entry->buf;
1606         size_t dest_off, buff_off;
1607         struct dmaengine_unmap_data *unmap;
1608         dma_addr_t dest;
1609         dma_cookie_t cookie;
1610
1611         device = chan->device;
1612         dest = qp->tx_mw_phys + qp->tx_max_frame * entry->tx_index;
1613         buff_off = (size_t)buf & ~PAGE_MASK;
1614         dest_off = (size_t)dest & ~PAGE_MASK;
1615
1616         if (!is_dma_copy_aligned(device, buff_off, dest_off, len))
1617                 goto err;
1618
1619         unmap = dmaengine_get_unmap_data(device->dev, 1, GFP_NOWAIT);
1620         if (!unmap)
1621                 goto err;
1622
1623         unmap->len = len;
1624         unmap->addr[0] = dma_map_page(device->dev, virt_to_page(buf),
1625                                       buff_off, len, DMA_TO_DEVICE);
1626         if (dma_mapping_error(device->dev, unmap->addr[0]))
1627                 goto err_get_unmap;
1628
1629         unmap->to_cnt = 1;
1630
1631         txd = device->device_prep_dma_memcpy(chan, dest, unmap->addr[0], len,
1632                                              DMA_PREP_INTERRUPT);
1633         if (!txd)
1634                 goto err_get_unmap;
1635
1636         txd->callback_result = ntb_tx_copy_callback;
1637         txd->callback_param = entry;
1638         dma_set_unmap(txd, unmap);
1639
1640         cookie = dmaengine_submit(txd);
1641         if (dma_submit_error(cookie))
1642                 goto err_set_unmap;
1643
1644         dmaengine_unmap_put(unmap);
1645
1646         dma_async_issue_pending(chan);
1647
1648         return 0;
1649 err_set_unmap:
1650         dmaengine_unmap_put(unmap);
1651 err_get_unmap:
1652         dmaengine_unmap_put(unmap);
1653 err:
1654         return -ENXIO;
1655 }
1656
1657 static void ntb_async_tx(struct ntb_transport_qp *qp,
1658                          struct ntb_queue_entry *entry)
1659 {
1660         struct ntb_payload_header __iomem *hdr;
1661         struct dma_chan *chan = qp->tx_dma_chan;
1662         void __iomem *offset;
1663         int res;
1664
1665         entry->tx_index = qp->tx_index;
1666         offset = qp->tx_mw + qp->tx_max_frame * entry->tx_index;
1667         hdr = offset + qp->tx_max_frame - sizeof(struct ntb_payload_header);
1668         entry->tx_hdr = hdr;
1669
1670         iowrite32(entry->len, &hdr->len);
1671         iowrite32((u32)qp->tx_pkts, &hdr->ver);
1672
1673         if (!chan)
1674                 goto err;
1675
1676         if (entry->len < copy_bytes)
1677                 goto err;
1678
1679         res = ntb_async_tx_submit(qp, entry);
1680         if (res < 0)
1681                 goto err;
1682
1683         if (!entry->retries)
1684                 qp->tx_async++;
1685
1686         return;
1687
1688 err:
1689         ntb_memcpy_tx(entry, offset);
1690         qp->tx_memcpy++;
1691 }
1692
1693 static int ntb_process_tx(struct ntb_transport_qp *qp,
1694                           struct ntb_queue_entry *entry)
1695 {
1696         if (qp->tx_index == qp->remote_rx_info->entry) {
1697                 qp->tx_ring_full++;
1698                 return -EAGAIN;
1699         }
1700
1701         if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
1702                 if (qp->tx_handler)
1703                         qp->tx_handler(qp, qp->cb_data, NULL, -EIO);
1704
1705                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1706                              &qp->tx_free_q);
1707                 return 0;
1708         }
1709
1710         ntb_async_tx(qp, entry);
1711
1712         qp->tx_index++;
1713         qp->tx_index %= qp->tx_max_entry;
1714
1715         qp->tx_pkts++;
1716
1717         return 0;
1718 }
1719
1720 static void ntb_send_link_down(struct ntb_transport_qp *qp)
1721 {
1722         struct pci_dev *pdev = qp->ndev->pdev;
1723         struct ntb_queue_entry *entry;
1724         int i, rc;
1725
1726         if (!qp->link_is_up)
1727                 return;
1728
1729         dev_info(&pdev->dev, "qp %d: Send Link Down\n", qp->qp_num);
1730
1731         for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1732                 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1733                 if (entry)
1734                         break;
1735                 msleep(100);
1736         }
1737
1738         if (!entry)
1739                 return;
1740
1741         entry->cb_data = NULL;
1742         entry->buf = NULL;
1743         entry->len = 0;
1744         entry->flags = LINK_DOWN_FLAG;
1745
1746         rc = ntb_process_tx(qp, entry);
1747         if (rc)
1748                 dev_err(&pdev->dev, "ntb: QP%d unable to send linkdown msg\n",
1749                         qp->qp_num);
1750
1751         ntb_qp_link_down_reset(qp);
1752 }
1753
1754 static bool ntb_dma_filter_fn(struct dma_chan *chan, void *node)
1755 {
1756         return dev_to_node(&chan->dev->device) == (int)(unsigned long)node;
1757 }
1758
1759 /**
1760  * ntb_transport_create_queue - Create a new NTB transport layer queue
1761  * @rx_handler: receive callback function
1762  * @tx_handler: transmit callback function
1763  * @event_handler: event callback function
1764  *
1765  * Create a new NTB transport layer queue and provide the queue with a callback
1766  * routine for both transmit and receive.  The receive callback routine will be
1767  * used to pass up data when the transport has received it on the queue.   The
1768  * transmit callback routine will be called when the transport has completed the
1769  * transmission of the data on the queue and the data is ready to be freed.
1770  *
1771  * RETURNS: pointer to newly created ntb_queue, NULL on error.
1772  */
1773 struct ntb_transport_qp *
1774 ntb_transport_create_queue(void *data, struct device *client_dev,
1775                            const struct ntb_queue_handlers *handlers)
1776 {
1777         struct ntb_dev *ndev;
1778         struct pci_dev *pdev;
1779         struct ntb_transport_ctx *nt;
1780         struct ntb_queue_entry *entry;
1781         struct ntb_transport_qp *qp;
1782         u64 qp_bit;
1783         unsigned int free_queue;
1784         dma_cap_mask_t dma_mask;
1785         int node;
1786         int i;
1787
1788         ndev = dev_ntb(client_dev->parent);
1789         pdev = ndev->pdev;
1790         nt = ndev->ctx;
1791
1792         node = dev_to_node(&ndev->dev);
1793
1794         free_queue = ffs(nt->qp_bitmap_free);
1795         if (!free_queue)
1796                 goto err;
1797
1798         /* decrement free_queue to make it zero based */
1799         free_queue--;
1800
1801         qp = &nt->qp_vec[free_queue];
1802         qp_bit = BIT_ULL(qp->qp_num);
1803
1804         nt->qp_bitmap_free &= ~qp_bit;
1805
1806         qp->cb_data = data;
1807         qp->rx_handler = handlers->rx_handler;
1808         qp->tx_handler = handlers->tx_handler;
1809         qp->event_handler = handlers->event_handler;
1810
1811         dma_cap_zero(dma_mask);
1812         dma_cap_set(DMA_MEMCPY, dma_mask);
1813
1814         if (use_dma) {
1815                 qp->tx_dma_chan =
1816                         dma_request_channel(dma_mask, ntb_dma_filter_fn,
1817                                             (void *)(unsigned long)node);
1818                 if (!qp->tx_dma_chan)
1819                         dev_info(&pdev->dev, "Unable to allocate TX DMA channel\n");
1820
1821                 qp->rx_dma_chan =
1822                         dma_request_channel(dma_mask, ntb_dma_filter_fn,
1823                                             (void *)(unsigned long)node);
1824                 if (!qp->rx_dma_chan)
1825                         dev_info(&pdev->dev, "Unable to allocate RX DMA channel\n");
1826         } else {
1827                 qp->tx_dma_chan = NULL;
1828                 qp->rx_dma_chan = NULL;
1829         }
1830
1831         dev_dbg(&pdev->dev, "Using %s memcpy for TX\n",
1832                 qp->tx_dma_chan ? "DMA" : "CPU");
1833
1834         dev_dbg(&pdev->dev, "Using %s memcpy for RX\n",
1835                 qp->rx_dma_chan ? "DMA" : "CPU");
1836
1837         for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
1838                 entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
1839                 if (!entry)
1840                         goto err1;
1841
1842                 entry->qp = qp;
1843                 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
1844                              &qp->rx_free_q);
1845         }
1846         qp->rx_alloc_entry = NTB_QP_DEF_NUM_ENTRIES;
1847
1848         for (i = 0; i < qp->tx_max_entry; i++) {
1849                 entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
1850                 if (!entry)
1851                         goto err2;
1852
1853                 entry->qp = qp;
1854                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1855                              &qp->tx_free_q);
1856         }
1857
1858         ntb_db_clear(qp->ndev, qp_bit);
1859         ntb_db_clear_mask(qp->ndev, qp_bit);
1860
1861         dev_info(&pdev->dev, "NTB Transport QP %d created\n", qp->qp_num);
1862
1863         return qp;
1864
1865 err2:
1866         while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1867                 kfree(entry);
1868 err1:
1869         qp->rx_alloc_entry = 0;
1870         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
1871                 kfree(entry);
1872         if (qp->tx_dma_chan)
1873                 dma_release_channel(qp->tx_dma_chan);
1874         if (qp->rx_dma_chan)
1875                 dma_release_channel(qp->rx_dma_chan);
1876         nt->qp_bitmap_free |= qp_bit;
1877 err:
1878         return NULL;
1879 }
1880 EXPORT_SYMBOL_GPL(ntb_transport_create_queue);
1881
1882 /**
1883  * ntb_transport_free_queue - Frees NTB transport queue
1884  * @qp: NTB queue to be freed
1885  *
1886  * Frees NTB transport queue
1887  */
1888 void ntb_transport_free_queue(struct ntb_transport_qp *qp)
1889 {
1890         struct pci_dev *pdev;
1891         struct ntb_queue_entry *entry;
1892         u64 qp_bit;
1893
1894         if (!qp)
1895                 return;
1896
1897         pdev = qp->ndev->pdev;
1898
1899         qp->active = false;
1900
1901         if (qp->tx_dma_chan) {
1902                 struct dma_chan *chan = qp->tx_dma_chan;
1903                 /* Putting the dma_chan to NULL will force any new traffic to be
1904                  * processed by the CPU instead of the DAM engine
1905                  */
1906                 qp->tx_dma_chan = NULL;
1907
1908                 /* Try to be nice and wait for any queued DMA engine
1909                  * transactions to process before smashing it with a rock
1910                  */
1911                 dma_sync_wait(chan, qp->last_cookie);
1912                 dmaengine_terminate_all(chan);
1913                 dma_release_channel(chan);
1914         }
1915
1916         if (qp->rx_dma_chan) {
1917                 struct dma_chan *chan = qp->rx_dma_chan;
1918                 /* Putting the dma_chan to NULL will force any new traffic to be
1919                  * processed by the CPU instead of the DAM engine
1920                  */
1921                 qp->rx_dma_chan = NULL;
1922
1923                 /* Try to be nice and wait for any queued DMA engine
1924                  * transactions to process before smashing it with a rock
1925                  */
1926                 dma_sync_wait(chan, qp->last_cookie);
1927                 dmaengine_terminate_all(chan);
1928                 dma_release_channel(chan);
1929         }
1930
1931         qp_bit = BIT_ULL(qp->qp_num);
1932
1933         ntb_db_set_mask(qp->ndev, qp_bit);
1934         tasklet_kill(&qp->rxc_db_work);
1935
1936         cancel_delayed_work_sync(&qp->link_work);
1937
1938         qp->cb_data = NULL;
1939         qp->rx_handler = NULL;
1940         qp->tx_handler = NULL;
1941         qp->event_handler = NULL;
1942
1943         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
1944                 kfree(entry);
1945
1946         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q))) {
1947                 dev_warn(&pdev->dev, "Freeing item from non-empty rx_pend_q\n");
1948                 kfree(entry);
1949         }
1950
1951         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q))) {
1952                 dev_warn(&pdev->dev, "Freeing item from non-empty rx_post_q\n");
1953                 kfree(entry);
1954         }
1955
1956         while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1957                 kfree(entry);
1958
1959         qp->transport->qp_bitmap_free |= qp_bit;
1960
1961         dev_info(&pdev->dev, "NTB Transport QP %d freed\n", qp->qp_num);
1962 }
1963 EXPORT_SYMBOL_GPL(ntb_transport_free_queue);
1964
1965 /**
1966  * ntb_transport_rx_remove - Dequeues enqueued rx packet
1967  * @qp: NTB queue to be freed
1968  * @len: pointer to variable to write enqueued buffers length
1969  *
1970  * Dequeues unused buffers from receive queue.  Should only be used during
1971  * shutdown of qp.
1972  *
1973  * RETURNS: NULL error value on error, or void* for success.
1974  */
1975 void *ntb_transport_rx_remove(struct ntb_transport_qp *qp, unsigned int *len)
1976 {
1977         struct ntb_queue_entry *entry;
1978         void *buf;
1979
1980         if (!qp || qp->client_ready)
1981                 return NULL;
1982
1983         entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q);
1984         if (!entry)
1985                 return NULL;
1986
1987         buf = entry->cb_data;
1988         *len = entry->len;
1989
1990         ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_free_q);
1991
1992         return buf;
1993 }
1994 EXPORT_SYMBOL_GPL(ntb_transport_rx_remove);
1995
1996 /**
1997  * ntb_transport_rx_enqueue - Enqueue a new NTB queue entry
1998  * @qp: NTB transport layer queue the entry is to be enqueued on
1999  * @cb: per buffer pointer for callback function to use
2000  * @data: pointer to data buffer that incoming packets will be copied into
2001  * @len: length of the data buffer
2002  *
2003  * Enqueue a new receive buffer onto the transport queue into which a NTB
2004  * payload can be received into.
2005  *
2006  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2007  */
2008 int ntb_transport_rx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
2009                              unsigned int len)
2010 {
2011         struct ntb_queue_entry *entry;
2012
2013         if (!qp)
2014                 return -EINVAL;
2015
2016         entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q);
2017         if (!entry)
2018                 return -ENOMEM;
2019
2020         entry->cb_data = cb;
2021         entry->buf = data;
2022         entry->len = len;
2023         entry->flags = 0;
2024         entry->retries = 0;
2025         entry->errors = 0;
2026         entry->rx_index = 0;
2027
2028         ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_pend_q);
2029
2030         if (qp->active)
2031                 tasklet_schedule(&qp->rxc_db_work);
2032
2033         return 0;
2034 }
2035 EXPORT_SYMBOL_GPL(ntb_transport_rx_enqueue);
2036
2037 /**
2038  * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
2039  * @qp: NTB transport layer queue the entry is to be enqueued on
2040  * @cb: per buffer pointer for callback function to use
2041  * @data: pointer to data buffer that will be sent
2042  * @len: length of the data buffer
2043  *
2044  * Enqueue a new transmit buffer onto the transport queue from which a NTB
2045  * payload will be transmitted.  This assumes that a lock is being held to
2046  * serialize access to the qp.
2047  *
2048  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2049  */
2050 int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
2051                              unsigned int len)
2052 {
2053         struct ntb_queue_entry *entry;
2054         int rc;
2055
2056         if (!qp || !len)
2057                 return -EINVAL;
2058
2059         /* If the qp link is down already, just ignore. */
2060         if (!qp->link_is_up)
2061                 return 0;
2062
2063         entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
2064         if (!entry) {
2065                 qp->tx_err_no_buf++;
2066                 return -EBUSY;
2067         }
2068
2069         entry->cb_data = cb;
2070         entry->buf = data;
2071         entry->len = len;
2072         entry->flags = 0;
2073         entry->errors = 0;
2074         entry->retries = 0;
2075         entry->tx_index = 0;
2076
2077         rc = ntb_process_tx(qp, entry);
2078         if (rc)
2079                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
2080                              &qp->tx_free_q);
2081
2082         return rc;
2083 }
2084 EXPORT_SYMBOL_GPL(ntb_transport_tx_enqueue);
2085
2086 /**
2087  * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
2088  * @qp: NTB transport layer queue to be enabled
2089  *
2090  * Notify NTB transport layer of client readiness to use queue
2091  */
2092 void ntb_transport_link_up(struct ntb_transport_qp *qp)
2093 {
2094         if (!qp)
2095                 return;
2096
2097         qp->client_ready = true;
2098
2099         if (qp->transport->link_is_up)
2100                 schedule_delayed_work(&qp->link_work, 0);
2101 }
2102 EXPORT_SYMBOL_GPL(ntb_transport_link_up);
2103
2104 /**
2105  * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
2106  * @qp: NTB transport layer queue to be disabled
2107  *
2108  * Notify NTB transport layer of client's desire to no longer receive data on
2109  * transport queue specified.  It is the client's responsibility to ensure all
2110  * entries on queue are purged or otherwise handled appropriately.
2111  */
2112 void ntb_transport_link_down(struct ntb_transport_qp *qp)
2113 {
2114         int val;
2115
2116         if (!qp)
2117                 return;
2118
2119         qp->client_ready = false;
2120
2121         val = ntb_spad_read(qp->ndev, QP_LINKS);
2122
2123         ntb_peer_spad_write(qp->ndev, PIDX, QP_LINKS, val & ~BIT(qp->qp_num));
2124
2125         if (qp->link_is_up)
2126                 ntb_send_link_down(qp);
2127         else
2128                 cancel_delayed_work_sync(&qp->link_work);
2129 }
2130 EXPORT_SYMBOL_GPL(ntb_transport_link_down);
2131
2132 /**
2133  * ntb_transport_link_query - Query transport link state
2134  * @qp: NTB transport layer queue to be queried
2135  *
2136  * Query connectivity to the remote system of the NTB transport queue
2137  *
2138  * RETURNS: true for link up or false for link down
2139  */
2140 bool ntb_transport_link_query(struct ntb_transport_qp *qp)
2141 {
2142         if (!qp)
2143                 return false;
2144
2145         return qp->link_is_up;
2146 }
2147 EXPORT_SYMBOL_GPL(ntb_transport_link_query);
2148
2149 /**
2150  * ntb_transport_qp_num - Query the qp number
2151  * @qp: NTB transport layer queue to be queried
2152  *
2153  * Query qp number of the NTB transport queue
2154  *
2155  * RETURNS: a zero based number specifying the qp number
2156  */
2157 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
2158 {
2159         if (!qp)
2160                 return 0;
2161
2162         return qp->qp_num;
2163 }
2164 EXPORT_SYMBOL_GPL(ntb_transport_qp_num);
2165
2166 /**
2167  * ntb_transport_max_size - Query the max payload size of a qp
2168  * @qp: NTB transport layer queue to be queried
2169  *
2170  * Query the maximum payload size permissible on the given qp
2171  *
2172  * RETURNS: the max payload size of a qp
2173  */
2174 unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp)
2175 {
2176         unsigned int max_size;
2177         unsigned int copy_align;
2178         struct dma_chan *rx_chan, *tx_chan;
2179
2180         if (!qp)
2181                 return 0;
2182
2183         rx_chan = qp->rx_dma_chan;
2184         tx_chan = qp->tx_dma_chan;
2185
2186         copy_align = max(rx_chan ? rx_chan->device->copy_align : 0,
2187                          tx_chan ? tx_chan->device->copy_align : 0);
2188
2189         /* If DMA engine usage is possible, try to find the max size for that */
2190         max_size = qp->tx_max_frame - sizeof(struct ntb_payload_header);
2191         max_size = round_down(max_size, 1 << copy_align);
2192
2193         return max_size;
2194 }
2195 EXPORT_SYMBOL_GPL(ntb_transport_max_size);
2196
2197 unsigned int ntb_transport_tx_free_entry(struct ntb_transport_qp *qp)
2198 {
2199         unsigned int head = qp->tx_index;
2200         unsigned int tail = qp->remote_rx_info->entry;
2201
2202         return tail >= head ? tail - head : qp->tx_max_entry + tail - head;
2203 }
2204 EXPORT_SYMBOL_GPL(ntb_transport_tx_free_entry);
2205
2206 static void ntb_transport_doorbell_callback(void *data, int vector)
2207 {
2208         struct ntb_transport_ctx *nt = data;
2209         struct ntb_transport_qp *qp;
2210         u64 db_bits;
2211         unsigned int qp_num;
2212
2213         db_bits = (nt->qp_bitmap & ~nt->qp_bitmap_free &
2214                    ntb_db_vector_mask(nt->ndev, vector));
2215
2216         while (db_bits) {
2217                 qp_num = __ffs(db_bits);
2218                 qp = &nt->qp_vec[qp_num];
2219
2220                 if (qp->active)
2221                         tasklet_schedule(&qp->rxc_db_work);
2222
2223                 db_bits &= ~BIT_ULL(qp_num);
2224         }
2225 }
2226
2227 static const struct ntb_ctx_ops ntb_transport_ops = {
2228         .link_event = ntb_transport_event_callback,
2229         .db_event = ntb_transport_doorbell_callback,
2230 };
2231
2232 static struct ntb_client ntb_transport_client = {
2233         .ops = {
2234                 .probe = ntb_transport_probe,
2235                 .remove = ntb_transport_free,
2236         },
2237 };
2238
2239 static int __init ntb_transport_init(void)
2240 {
2241         int rc;
2242
2243         pr_info("%s, version %s\n", NTB_TRANSPORT_DESC, NTB_TRANSPORT_VER);
2244
2245         if (debugfs_initialized())
2246                 nt_debugfs_dir = debugfs_create_dir(KBUILD_MODNAME, NULL);
2247
2248         rc = bus_register(&ntb_transport_bus);
2249         if (rc)
2250                 goto err_bus;
2251
2252         rc = ntb_register_client(&ntb_transport_client);
2253         if (rc)
2254                 goto err_client;
2255
2256         return 0;
2257
2258 err_client:
2259         bus_unregister(&ntb_transport_bus);
2260 err_bus:
2261         debugfs_remove_recursive(nt_debugfs_dir);
2262         return rc;
2263 }
2264 module_init(ntb_transport_init);
2265
2266 static void __exit ntb_transport_exit(void)
2267 {
2268         ntb_unregister_client(&ntb_transport_client);
2269         bus_unregister(&ntb_transport_bus);
2270         debugfs_remove_recursive(nt_debugfs_dir);
2271 }
2272 module_exit(ntb_transport_exit);