drivers/ntb/ntb_transport.c

   1 /*
   2  * This file is provided under a dual BSD/GPLv2 license.  When using or
   3  *   redistributing this file, you may do so under either license.
   4  *
   5  *   GPL LICENSE SUMMARY
   6  *
   7  *   Copyright(c) 2012 Intel Corporation. All rights reserved.
   8  *   Copyright (C) 2015 EMC Corporation. All Rights Reserved.
   9  *
  10  *   This program is free software; you can redistribute it and/or modify
  11  *   it under the terms of version 2 of the GNU General Public License as
  12  *   published by the Free Software Foundation.
  13  *
  14  *   BSD LICENSE
  15  *
  16  *   Copyright(c) 2012 Intel Corporation. All rights reserved.
  17  *   Copyright (C) 2015 EMC Corporation. All Rights Reserved.
  18  *
  19  *   Redistribution and use in source and binary forms, with or without
  20  *   modification, are permitted provided that the following conditions
  21  *   are met:
  22  *
  23  *     * Redistributions of source code must retain the above copyright
  24  *       notice, this list of conditions and the following disclaimer.
  25  *     * Redistributions in binary form must reproduce the above copy
  26  *       notice, this list of conditions and the following disclaimer in
  27  *       the documentation and/or other materials provided with the
  28  *       distribution.
  29  *     * Neither the name of Intel Corporation nor the names of its
  30  *       contributors may be used to endorse or promote products derived
  31  *       from this software without specific prior written permission.
  32  *
  33  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  34  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  35  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  36  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  37  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  38  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  39  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  40  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  41  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  42  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  43  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  44  *
  45  * PCIe NTB Transport Linux driver
  46  *
  47  * Contact Information:
  48  * Jon Mason <jon.mason@intel.com>
  49  */
  50 #include <linux/debugfs.h>
  51 #include <linux/delay.h>
  52 #include <linux/dmaengine.h>
  53 #include <linux/dma-mapping.h>
  54 #include <linux/errno.h>
  55 #include <linux/export.h>
  56 #include <linux/interrupt.h>
  57 #include <linux/module.h>
  58 #include <linux/pci.h>
  59 #include <linux/slab.h>
  60 #include <linux/types.h>
  61 #include <linux/uaccess.h>
  62 #include "linux/ntb.h"
  63 #include "linux/ntb_transport.h"
  64
  65 #define NTB_TRANSPORT_VERSION   4
  66 #define NTB_TRANSPORT_VER       "4"
  67 #define NTB_TRANSPORT_NAME      "ntb_transport"
  68 #define NTB_TRANSPORT_DESC      "Software Queue-Pair Transport over NTB"
  69 #define NTB_TRANSPORT_MIN_SPADS (MW0_SZ_HIGH + 2)
  70
  71 MODULE_DESCRIPTION(NTB_TRANSPORT_DESC);
  72 MODULE_VERSION(NTB_TRANSPORT_VER);
  73 MODULE_LICENSE("Dual BSD/GPL");
  74 MODULE_AUTHOR("Intel Corporation");
  75
  76 static unsigned long max_mw_size;
  77 module_param(max_mw_size, ulong, 0644);
  78 MODULE_PARM_DESC(max_mw_size, "Limit size of large memory windows");
  79
  80 static unsigned int transport_mtu = 0x10000;
  81 module_param(transport_mtu, uint, 0644);
  82 MODULE_PARM_DESC(transport_mtu, "Maximum size of NTB transport packets");
  83
  84 static unsigned char max_num_clients;
  85 module_param(max_num_clients, byte, 0644);
  86 MODULE_PARM_DESC(max_num_clients, "Maximum number of NTB transport clients");
  87
  88 static unsigned int copy_bytes = 1024;
  89 module_param(copy_bytes, uint, 0644);
  90 MODULE_PARM_DESC(copy_bytes, "Threshold under which NTB will use the CPU to copy instead of DMA");
  91
  92 static bool use_dma;
  93 module_param(use_dma, bool, 0644);
  94 MODULE_PARM_DESC(use_dma, "Use DMA engine to perform large data copy");
  95
  96 static struct dentry *nt_debugfs_dir;
  97
  98 struct ntb_queue_entry {
  99         /* ntb_queue list reference */
 100         struct list_head entry;
 101         /* pointers to data to be transferred */
 102         void *cb_data;
 103         void *buf;
 104         unsigned int len;
 105         unsigned int flags;
 106         int retries;
 107         int errors;
 108         unsigned int tx_index;
 109         unsigned int rx_index;
 110
 111         struct ntb_transport_qp *qp;
 112         union {
 113                 struct ntb_payload_header __iomem *tx_hdr;
 114                 struct ntb_payload_header *rx_hdr;
 115         };
 116 };
 117
 118 struct ntb_rx_info {
 119         unsigned int entry;
 120 };
 121
 122 struct ntb_transport_qp {
 123         struct ntb_transport_ctx *transport;
 124         struct ntb_dev *ndev;
 125         void *cb_data;
 126         struct dma_chan *tx_dma_chan;
 127         struct dma_chan *rx_dma_chan;
 128
 129         bool client_ready;
 130         bool link_is_up;
 131         bool active;
 132
 133         u8 qp_num;      /* Only 64 QP's are allowed.  0-63 */
 134         u64 qp_bit;
 135
 136         struct ntb_rx_info __iomem *rx_info;
 137         struct ntb_rx_info *remote_rx_info;
 138
 139         void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data,
 140                            void *data, int len);
 141         struct list_head tx_free_q;
 142         spinlock_t ntb_tx_free_q_lock;
 143         void __iomem *tx_mw;
 144         dma_addr_t tx_mw_phys;
 145         unsigned int tx_index;
 146         unsigned int tx_max_entry;
 147         unsigned int tx_max_frame;
 148
 149         void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
 150                            void *data, int len);
 151         struct list_head rx_post_q;
 152         struct list_head rx_pend_q;
 153         struct list_head rx_free_q;
 154         /* ntb_rx_q_lock: synchronize access to rx_XXXX_q */
 155         spinlock_t ntb_rx_q_lock;
 156         void *rx_buff;
 157         unsigned int rx_index;
 158         unsigned int rx_max_entry;
 159         unsigned int rx_max_frame;
 160         unsigned int rx_alloc_entry;
 161         dma_cookie_t last_cookie;
 162         struct tasklet_struct rxc_db_work;
 163
 164         void (*event_handler)(void *data, int status);
 165         struct delayed_work link_work;
 166         struct work_struct link_cleanup;
 167
 168         struct dentry *debugfs_dir;
 169         struct dentry *debugfs_stats;
 170
 171         /* Stats */
 172         u64 rx_bytes;
 173         u64 rx_pkts;
 174         u64 rx_ring_empty;
 175         u64 rx_err_no_buf;
 176         u64 rx_err_oflow;
 177         u64 rx_err_ver;
 178         u64 rx_memcpy;
 179         u64 rx_async;
 180         u64 dma_rx_prep_err;
 181         u64 tx_bytes;
 182         u64 tx_pkts;
 183         u64 tx_ring_full;
 184         u64 tx_err_no_buf;
 185         u64 tx_memcpy;
 186         u64 tx_async;
 187         u64 dma_tx_prep_err;
 188 };
 189
 190 struct ntb_transport_mw {
 191         phys_addr_t phys_addr;
 192         resource_size_t phys_size;
 193         resource_size_t xlat_align;
 194         resource_size_t xlat_align_size;
 195         void __iomem *vbase;
 196         size_t xlat_size;
 197         size_t buff_size;
 198         void *virt_addr;
 199         dma_addr_t dma_addr;
 200 };
 201
 202 struct ntb_transport_client_dev {
 203         struct list_head entry;
 204         struct ntb_transport_ctx *nt;
 205         struct device dev;
 206 };
 207
 208 struct ntb_transport_ctx {
 209         struct list_head entry;
 210         struct list_head client_devs;
 211
 212         struct ntb_dev *ndev;
 213
 214         struct ntb_transport_mw *mw_vec;
 215         struct ntb_transport_qp *qp_vec;
 216         unsigned int mw_count;
 217         unsigned int qp_count;
 218         u64 qp_bitmap;
 219         u64 qp_bitmap_free;
 220
 221         bool link_is_up;
 222         struct delayed_work link_work;
 223         struct work_struct link_cleanup;
 224
 225         struct dentry *debugfs_node_dir;
 226 };
 227
 228 enum {
 229         DESC_DONE_FLAG = BIT(0),
 230         LINK_DOWN_FLAG = BIT(1),
 231 };
 232
 233 struct ntb_payload_header {
 234         unsigned int ver;
 235         unsigned int len;
 236         unsigned int flags;
 237 };
 238
 239 enum {
 240         VERSION = 0,
 241         QP_LINKS,
 242         NUM_QPS,
 243         NUM_MWS,
 244         MW0_SZ_HIGH,
 245         MW0_SZ_LOW,
 246 };
 247
 248 #define dev_client_dev(__dev) \
 249         container_of((__dev), struct ntb_transport_client_dev, dev)
 250
 251 #define drv_client(__drv) \
 252         container_of((__drv), struct ntb_transport_client, driver)
 253
 254 #define QP_TO_MW(nt, qp)        ((qp) % nt->mw_count)
 255 #define NTB_QP_DEF_NUM_ENTRIES  100
 256 #define NTB_LINK_DOWN_TIMEOUT   10
 257 #define DMA_RETRIES             20
 258 #define DMA_OUT_RESOURCE_TO     msecs_to_jiffies(50)
 259
 260 static void ntb_transport_rxc_db(unsigned long data);
 261 static const struct ntb_ctx_ops ntb_transport_ops;
 262 static struct ntb_client ntb_transport_client;
 263 static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
 264                                struct ntb_queue_entry *entry);
 265 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset);
 266 static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset);
 267 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset);
 268
 269
 270 static int ntb_transport_bus_match(struct device *dev,
 271                                    struct device_driver *drv)
 272 {
 273         return !strncmp(dev_name(dev), drv->name, strlen(drv->name));
 274 }
 275
 276 static int ntb_transport_bus_probe(struct device *dev)
 277 {
 278         const struct ntb_transport_client *client;
 279         int rc = -EINVAL;
 280
 281         get_device(dev);
 282
 283         client = drv_client(dev->driver);
 284         rc = client->probe(dev);
 285         if (rc)
 286                 put_device(dev);
 287
 288         return rc;
 289 }
 290
 291 static int ntb_transport_bus_remove(struct device *dev)
 292 {
 293         const struct ntb_transport_client *client;
 294
 295         client = drv_client(dev->driver);
 296         client->remove(dev);
 297
 298         put_device(dev);
 299
 300         return 0;
 301 }
 302
 303 static struct bus_type ntb_transport_bus = {
 304         .name = "ntb_transport",
 305         .match = ntb_transport_bus_match,
 306         .probe = ntb_transport_bus_probe,
 307         .remove = ntb_transport_bus_remove,
 308 };
 309
 310 static LIST_HEAD(ntb_transport_list);
 311
 312 static int ntb_bus_init(struct ntb_transport_ctx *nt)
 313 {
 314         list_add_tail(&nt->entry, &ntb_transport_list);
 315         return 0;
 316 }
 317
 318 static void ntb_bus_remove(struct ntb_transport_ctx *nt)
 319 {
 320         struct ntb_transport_client_dev *client_dev, *cd;
 321
 322         list_for_each_entry_safe(client_dev, cd, &nt->client_devs, entry) {
 323                 dev_err(client_dev->dev.parent, "%s still attached to bus, removing\n",
 324                         dev_name(&client_dev->dev));
 325                 list_del(&client_dev->entry);
 326                 device_unregister(&client_dev->dev);
 327         }
 328
 329         list_del(&nt->entry);
 330 }
 331
 332 static void ntb_transport_client_release(struct device *dev)
 333 {
 334         struct ntb_transport_client_dev *client_dev;
 335
 336         client_dev = dev_client_dev(dev);
 337         kfree(client_dev);
 338 }
 339
 340 /**
 341  * ntb_transport_unregister_client_dev - Unregister NTB client device
 342  * @device_name: Name of NTB client device
 343  *
 344  * Unregister an NTB client device with the NTB transport layer
 345  */
 346 void ntb_transport_unregister_client_dev(char *device_name)
 347 {
 348         struct ntb_transport_client_dev *client, *cd;
 349         struct ntb_transport_ctx *nt;
 350
 351         list_for_each_entry(nt, &ntb_transport_list, entry)
 352                 list_for_each_entry_safe(client, cd, &nt->client_devs, entry)
 353                         if (!strncmp(dev_name(&client->dev), device_name,
 354                                      strlen(device_name))) {
 355                                 list_del(&client->entry);
 356                                 device_unregister(&client->dev);
 357                         }
 358 }
 359 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client_dev);
 360
 361 /**
 362  * ntb_transport_register_client_dev - Register NTB client device
 363  * @device_name: Name of NTB client device
 364  *
 365  * Register an NTB client device with the NTB transport layer
 366  */
 367 int ntb_transport_register_client_dev(char *device_name)
 368 {
 369         struct ntb_transport_client_dev *client_dev;
 370         struct ntb_transport_ctx *nt;
 371         int node;
 372         int rc, i = 0;
 373
 374         if (list_empty(&ntb_transport_list))
 375                 return -ENODEV;
 376
 377         list_for_each_entry(nt, &ntb_transport_list, entry) {
 378                 struct device *dev;
 379
 380                 node = dev_to_node(&nt->ndev->dev);
 381
 382                 client_dev = kzalloc_node(sizeof(*client_dev),
 383                                           GFP_KERNEL, node);
 384                 if (!client_dev) {
 385                         rc = -ENOMEM;
 386                         goto err;
 387                 }
 388
 389                 dev = &client_dev->dev;
 390
 391                 /* setup and register client devices */
 392                 dev_set_name(dev, "%s%d", device_name, i);
 393                 dev->bus = &ntb_transport_bus;
 394                 dev->release = ntb_transport_client_release;
 395                 dev->parent = &nt->ndev->dev;
 396
 397                 rc = device_register(dev);
 398                 if (rc) {
 399                         kfree(client_dev);
 400                         goto err;
 401                 }
 402
 403                 list_add_tail(&client_dev->entry, &nt->client_devs);
 404                 i++;
 405         }
 406
 407         return 0;
 408
 409 err:
 410         ntb_transport_unregister_client_dev(device_name);
 411
 412         return rc;
 413 }
 414 EXPORT_SYMBOL_GPL(ntb_transport_register_client_dev);
 415
 416 /**
 417  * ntb_transport_register_client - Register NTB client driver
 418  * @drv: NTB client driver to be registered
 419  *
 420  * Register an NTB client driver with the NTB transport layer
 421  *
 422  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 423  */
 424 int ntb_transport_register_client(struct ntb_transport_client *drv)
 425 {
 426         drv->driver.bus = &ntb_transport_bus;
 427
 428         if (list_empty(&ntb_transport_list))
 429                 return -ENODEV;
 430
 431         return driver_register(&drv->driver);
 432 }
 433 EXPORT_SYMBOL_GPL(ntb_transport_register_client);
 434
 435 /**
 436  * ntb_transport_unregister_client - Unregister NTB client driver
 437  * @drv: NTB client driver to be unregistered
 438  *
 439  * Unregister an NTB client driver with the NTB transport layer
 440  *
 441  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 442  */
 443 void ntb_transport_unregister_client(struct ntb_transport_client *drv)
 444 {
 445         driver_unregister(&drv->driver);
 446 }
 447 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client);
 448
 449 static ssize_t debugfs_read(struct file *filp, char __user *ubuf, size_t count,
 450                             loff_t *offp)
 451 {
 452         struct ntb_transport_qp *qp;
 453         char *buf;
 454         ssize_t ret, out_offset, out_count;
 455
 456         qp = filp->private_data;
 457
 458         if (!qp || !qp->link_is_up)
 459                 return 0;
 460
 461         out_count = 1000;
 462
 463         buf = kmalloc(out_count, GFP_KERNEL);
 464         if (!buf)
 465                 return -ENOMEM;
 466
 467         out_offset = 0;
 468         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 469                                "\nNTB QP stats:\n\n");
 470         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 471                                "rx_bytes - \t%llu\n", qp->rx_bytes);
 472         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 473                                "rx_pkts - \t%llu\n", qp->rx_pkts);
 474         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 475                                "rx_memcpy - \t%llu\n", qp->rx_memcpy);
 476         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 477                                "rx_async - \t%llu\n", qp->rx_async);
 478         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 479                                "rx_ring_empty - %llu\n", qp->rx_ring_empty);
 480         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 481                                "rx_err_no_buf - %llu\n", qp->rx_err_no_buf);
 482         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 483                                "rx_err_oflow - \t%llu\n", qp->rx_err_oflow);
 484         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 485                                "rx_err_ver - \t%llu\n", qp->rx_err_ver);
 486         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 487                                "rx_buff - \t0x%p\n", qp->rx_buff);
 488         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 489                                "rx_index - \t%u\n", qp->rx_index);
 490         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 491                                "rx_max_entry - \t%u\n", qp->rx_max_entry);
 492         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 493                                "rx_alloc_entry - \t%u\n\n", qp->rx_alloc_entry);
 494
 495         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 496                                "tx_bytes - \t%llu\n", qp->tx_bytes);
 497         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 498                                "tx_pkts - \t%llu\n", qp->tx_pkts);
 499         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 500                                "tx_memcpy - \t%llu\n", qp->tx_memcpy);
 501         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 502                                "tx_async - \t%llu\n", qp->tx_async);
 503         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 504                                "tx_ring_full - \t%llu\n", qp->tx_ring_full);
 505         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 506                                "tx_err_no_buf - %llu\n", qp->tx_err_no_buf);
 507         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 508                                "tx_mw - \t0x%p\n", qp->tx_mw);
 509         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 510                                "tx_index (H) - \t%u\n", qp->tx_index);
 511         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 512                                "RRI (T) - \t%u\n",
 513                                qp->remote_rx_info->entry);
 514         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 515                                "tx_max_entry - \t%u\n", qp->tx_max_entry);
 516         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 517                                "free tx - \t%u\n",
 518                                ntb_transport_tx_free_entry(qp));
 519         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 520                                "DMA tx prep err - \t%llu\n",
 521                                qp->dma_tx_prep_err);
 522         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 523                                "DMA rx prep err - \t%llu\n",
 524                                qp->dma_rx_prep_err);
 525
 526         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 527                                "\n");
 528         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 529                                "Using TX DMA - \t%s\n",
 530                                qp->tx_dma_chan ? "Yes" : "No");
 531         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 532                                "Using RX DMA - \t%s\n",
 533                                qp->rx_dma_chan ? "Yes" : "No");
 534         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 535                                "QP Link - \t%s\n",
 536                                qp->link_is_up ? "Up" : "Down");
 537         out_offset += snprintf(buf + out_offset, out_count - out_offset,
 538                                "\n");
 539
 540         if (out_offset > out_count)
 541                 out_offset = out_count;
 542
 543         ret = simple_read_from_buffer(ubuf, count, offp, buf, out_offset);
 544         kfree(buf);
 545         return ret;
 546 }
 547
 548 static const struct file_operations ntb_qp_debugfs_stats = {
 549         .owner = THIS_MODULE,
 550         .open = simple_open,
 551         .read = debugfs_read,
 552 };
 553
 554 static void ntb_list_add(spinlock_t *lock, struct list_head *entry,
 555                          struct list_head *list)
 556 {
 557         unsigned long flags;
 558
 559         spin_lock_irqsave(lock, flags);
 560         list_add_tail(entry, list);
 561         spin_unlock_irqrestore(lock, flags);
 562 }
 563
 564 static struct ntb_queue_entry *ntb_list_rm(spinlock_t *lock,
 565                                            struct list_head *list)
 566 {
 567         struct ntb_queue_entry *entry;
 568         unsigned long flags;
 569
 570         spin_lock_irqsave(lock, flags);
 571         if (list_empty(list)) {
 572                 entry = NULL;
 573                 goto out;
 574         }
 575         entry = list_first_entry(list, struct ntb_queue_entry, entry);
 576         list_del(&entry->entry);
 577
 578 out:
 579         spin_unlock_irqrestore(lock, flags);
 580
 581         return entry;
 582 }
 583
 584 static struct ntb_queue_entry *ntb_list_mv(spinlock_t *lock,
 585                                            struct list_head *list,
 586                                            struct list_head *to_list)
 587 {
 588         struct ntb_queue_entry *entry;
 589         unsigned long flags;
 590
 591         spin_lock_irqsave(lock, flags);
 592
 593         if (list_empty(list)) {
 594                 entry = NULL;
 595         } else {
 596                 entry = list_first_entry(list, struct ntb_queue_entry, entry);
 597                 list_move_tail(&entry->entry, to_list);
 598         }
 599
 600         spin_unlock_irqrestore(lock, flags);
 601
 602         return entry;
 603 }
 604
 605 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
 606                                      unsigned int qp_num)
 607 {
 608         struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
 609         struct ntb_transport_mw *mw;
 610         struct ntb_dev *ndev = nt->ndev;
 611         struct ntb_queue_entry *entry;
 612         unsigned int rx_size, num_qps_mw;
 613         unsigned int mw_num, mw_count, qp_count;
 614         unsigned int i;
 615         int node;
 616
 617         mw_count = nt->mw_count;
 618         qp_count = nt->qp_count;
 619
 620         mw_num = QP_TO_MW(nt, qp_num);
 621         mw = &nt->mw_vec[mw_num];
 622
 623         if (!mw->virt_addr)
 624                 return -ENOMEM;
 625
 626         if (qp_count % mw_count && mw_num + 1 < qp_count / mw_count)
 627                 num_qps_mw = qp_count / mw_count + 1;
 628         else
 629                 num_qps_mw = qp_count / mw_count;
 630
 631         rx_size = (unsigned int)mw->xlat_size / num_qps_mw;
 632         qp->rx_buff = mw->virt_addr + rx_size * (qp_num / mw_count);
 633         rx_size -= sizeof(struct ntb_rx_info);
 634
 635         qp->remote_rx_info = qp->rx_buff + rx_size;
 636
 637         /* Due to housekeeping, there must be atleast 2 buffs */
 638         qp->rx_max_frame = min(transport_mtu, rx_size / 2);
 639         qp->rx_max_entry = rx_size / qp->rx_max_frame;
 640         qp->rx_index = 0;
 641
 642         /*
 643          * Checking to see if we have more entries than the default.
 644          * We should add additional entries if that is the case so we
 645          * can be in sync with the transport frames.
 646          */
 647         node = dev_to_node(&ndev->dev);
 648         for (i = qp->rx_alloc_entry; i < qp->rx_max_entry; i++) {
 649                 entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
 650                 if (!entry)
 651                         return -ENOMEM;
 652
 653                 entry->qp = qp;
 654                 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
 655                              &qp->rx_free_q);
 656                 qp->rx_alloc_entry++;
 657         }
 658
 659         qp->remote_rx_info->entry = qp->rx_max_entry - 1;
 660
 661         /* setup the hdr offsets with 0's */
 662         for (i = 0; i < qp->rx_max_entry; i++) {
 663                 void *offset = (qp->rx_buff + qp->rx_max_frame * (i + 1) -
 664                                 sizeof(struct ntb_payload_header));
 665                 memset(offset, 0, sizeof(struct ntb_payload_header));
 666         }
 667
 668         qp->rx_pkts = 0;
 669         qp->tx_pkts = 0;
 670         qp->tx_index = 0;
 671
 672         return 0;
 673 }
 674
 675 static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
 676 {
 677         struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
 678         struct pci_dev *pdev = nt->ndev->pdev;
 679
 680         if (!mw->virt_addr)
 681                 return;
 682
 683         ntb_mw_clear_trans(nt->ndev, num_mw);
 684         dma_free_coherent(&pdev->dev, mw->buff_size,
 685                           mw->virt_addr, mw->dma_addr);
 686         mw->xlat_size = 0;
 687         mw->buff_size = 0;
 688         mw->virt_addr = NULL;
 689 }
 690
 691 static int ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw,
 692                       resource_size_t size)
 693 {
 694         struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
 695         struct pci_dev *pdev = nt->ndev->pdev;
 696         size_t xlat_size, buff_size;
 697         int rc;
 698
 699         if (!size)
 700                 return -EINVAL;
 701
 702         xlat_size = round_up(size, mw->xlat_align_size);
 703         buff_size = round_up(size, mw->xlat_align);
 704
 705         /* No need to re-setup */
 706         if (mw->xlat_size == xlat_size)
 707                 return 0;
 708
 709         if (mw->buff_size)
 710                 ntb_free_mw(nt, num_mw);
 711
 712         /* Alloc memory for receiving data.  Must be aligned */
 713         mw->xlat_size = xlat_size;
 714         mw->buff_size = buff_size;
 715
 716         mw->virt_addr = dma_alloc_coherent(&pdev->dev, buff_size,
 717                                            &mw->dma_addr, GFP_KERNEL);
 718         if (!mw->virt_addr) {
 719                 mw->xlat_size = 0;
 720                 mw->buff_size = 0;
 721                 dev_err(&pdev->dev, "Unable to alloc MW buff of size %zu\n",
 722                         buff_size);
 723                 return -ENOMEM;
 724         }
 725
 726         /*
 727          * we must ensure that the memory address allocated is BAR size
 728          * aligned in order for the XLAT register to take the value. This
 729          * is a requirement of the hardware. It is recommended to setup CMA
 730          * for BAR sizes equal or greater than 4MB.
 731          */
 732         if (!IS_ALIGNED(mw->dma_addr, mw->xlat_align)) {
 733                 dev_err(&pdev->dev, "DMA memory %pad is not aligned\n",
 734                         &mw->dma_addr);
 735                 ntb_free_mw(nt, num_mw);
 736                 return -ENOMEM;
 737         }
 738
 739         /* Notify HW the memory location of the receive buffer */
 740         rc = ntb_mw_set_trans(nt->ndev, num_mw, mw->dma_addr, mw->xlat_size);
 741         if (rc) {
 742                 dev_err(&pdev->dev, "Unable to set mw%d translation", num_mw);
 743                 ntb_free_mw(nt, num_mw);
 744                 return -EIO;
 745         }
 746
 747         return 0;
 748 }
 749
 750 static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
 751 {
 752         qp->link_is_up = false;
 753         qp->active = false;
 754
 755         qp->tx_index = 0;
 756         qp->rx_index = 0;
 757         qp->rx_bytes = 0;
 758         qp->rx_pkts = 0;
 759         qp->rx_ring_empty = 0;
 760         qp->rx_err_no_buf = 0;
 761         qp->rx_err_oflow = 0;
 762         qp->rx_err_ver = 0;
 763         qp->rx_memcpy = 0;
 764         qp->rx_async = 0;
 765         qp->tx_bytes = 0;
 766         qp->tx_pkts = 0;
 767         qp->tx_ring_full = 0;
 768         qp->tx_err_no_buf = 0;
 769         qp->tx_memcpy = 0;
 770         qp->tx_async = 0;
 771         qp->dma_tx_prep_err = 0;
 772         qp->dma_rx_prep_err = 0;
 773 }
 774
 775 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
 776 {
 777         struct ntb_transport_ctx *nt = qp->transport;
 778         struct pci_dev *pdev = nt->ndev->pdev;
 779
 780         dev_info(&pdev->dev, "qp %d: Link Cleanup\n", qp->qp_num);
 781
 782         cancel_delayed_work_sync(&qp->link_work);
 783         ntb_qp_link_down_reset(qp);
 784
 785         if (qp->event_handler)
 786                 qp->event_handler(qp->cb_data, qp->link_is_up);
 787 }
 788
 789 static void ntb_qp_link_cleanup_work(struct work_struct *work)
 790 {
 791         struct ntb_transport_qp *qp = container_of(work,
 792                                                    struct ntb_transport_qp,
 793                                                    link_cleanup);
 794         struct ntb_transport_ctx *nt = qp->transport;
 795
 796         ntb_qp_link_cleanup(qp);
 797
 798         if (nt->link_is_up)
 799                 schedule_delayed_work(&qp->link_work,
 800                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 801 }
 802
 803 static void ntb_qp_link_down(struct ntb_transport_qp *qp)
 804 {
 805         schedule_work(&qp->link_cleanup);
 806 }
 807
 808 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
 809 {
 810         struct ntb_transport_qp *qp;
 811         u64 qp_bitmap_alloc;
 812         unsigned int i, count;
 813
 814         qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
 815
 816         /* Pass along the info to any clients */
 817         for (i = 0; i < nt->qp_count; i++)
 818                 if (qp_bitmap_alloc & BIT_ULL(i)) {
 819                         qp = &nt->qp_vec[i];
 820                         ntb_qp_link_cleanup(qp);
 821                         cancel_work_sync(&qp->link_cleanup);
 822                         cancel_delayed_work_sync(&qp->link_work);
 823                 }
 824
 825         if (!nt->link_is_up)
 826                 cancel_delayed_work_sync(&nt->link_work);
 827
 828         /* The scratchpad registers keep the values if the remote side
 829          * goes down, blast them now to give them a sane value the next
 830          * time they are accessed
 831          */
 832         count = ntb_spad_count(nt->ndev);
 833         for (i = 0; i < count; i++)
 834                 ntb_spad_write(nt->ndev, i, 0);
 835 }
 836
 837 static void ntb_transport_link_cleanup_work(struct work_struct *work)
 838 {
 839         struct ntb_transport_ctx *nt =
 840                 container_of(work, struct ntb_transport_ctx, link_cleanup);
 841
 842         ntb_transport_link_cleanup(nt);
 843 }
 844
 845 static void ntb_transport_event_callback(void *data)
 846 {
 847         struct ntb_transport_ctx *nt = data;
 848
 849         if (ntb_link_is_up(nt->ndev, NULL, NULL) == 1)
 850                 schedule_delayed_work(&nt->link_work, 0);
 851         else
 852                 schedule_work(&nt->link_cleanup);
 853 }
 854
 855 static void ntb_transport_link_work(struct work_struct *work)
 856 {
 857         struct ntb_transport_ctx *nt =
 858                 container_of(work, struct ntb_transport_ctx, link_work.work);
 859         struct ntb_dev *ndev = nt->ndev;
 860         struct pci_dev *pdev = ndev->pdev;
 861         resource_size_t size;
 862         u32 val;
 863         int rc = 0, i, spad;
 864
 865         /* send the local info, in the opposite order of the way we read it */
 866         for (i = 0; i < nt->mw_count; i++) {
 867                 size = nt->mw_vec[i].phys_size;
 868
 869                 if (max_mw_size && size > max_mw_size)
 870                         size = max_mw_size;
 871
 872                 spad = MW0_SZ_HIGH + (i * 2);
 873                 ntb_peer_spad_write(ndev, spad, upper_32_bits(size));
 874
 875                 spad = MW0_SZ_LOW + (i * 2);
 876                 ntb_peer_spad_write(ndev, spad, lower_32_bits(size));
 877         }
 878
 879         ntb_peer_spad_write(ndev, NUM_MWS, nt->mw_count);
 880
 881         ntb_peer_spad_write(ndev, NUM_QPS, nt->qp_count);
 882
 883         ntb_peer_spad_write(ndev, VERSION, NTB_TRANSPORT_VERSION);
 884
 885         /* Query the remote side for its info */
 886         val = ntb_spad_read(ndev, VERSION);
 887         dev_dbg(&pdev->dev, "Remote version = %d\n", val);
 888         if (val != NTB_TRANSPORT_VERSION)
 889                 goto out;
 890
 891         val = ntb_spad_read(ndev, NUM_QPS);
 892         dev_dbg(&pdev->dev, "Remote max number of qps = %d\n", val);
 893         if (val != nt->qp_count)
 894                 goto out;
 895
 896         val = ntb_spad_read(ndev, NUM_MWS);
 897         dev_dbg(&pdev->dev, "Remote number of mws = %d\n", val);
 898         if (val != nt->mw_count)
 899                 goto out;
 900
 901         for (i = 0; i < nt->mw_count; i++) {
 902                 u64 val64;
 903
 904                 val = ntb_spad_read(ndev, MW0_SZ_HIGH + (i * 2));
 905                 val64 = (u64)val << 32;
 906
 907                 val = ntb_spad_read(ndev, MW0_SZ_LOW + (i * 2));
 908                 val64 |= val;
 909
 910                 dev_dbg(&pdev->dev, "Remote MW%d size = %#llx\n", i, val64);
 911
 912                 rc = ntb_set_mw(nt, i, val64);
 913                 if (rc)
 914                         goto out1;
 915         }
 916
 917         nt->link_is_up = true;
 918
 919         for (i = 0; i < nt->qp_count; i++) {
 920                 struct ntb_transport_qp *qp = &nt->qp_vec[i];
 921
 922                 ntb_transport_setup_qp_mw(nt, i);
 923
 924                 if (qp->client_ready)
 925                         schedule_delayed_work(&qp->link_work, 0);
 926         }
 927
 928         return;
 929
 930 out1:
 931         for (i = 0; i < nt->mw_count; i++)
 932                 ntb_free_mw(nt, i);
 933
 934         /* if there's an actual failure, we should just bail */
 935         if (rc < 0) {
 936                 ntb_link_disable(ndev);
 937                 return;
 938         }
 939
 940 out:
 941         if (ntb_link_is_up(ndev, NULL, NULL) == 1)
 942                 schedule_delayed_work(&nt->link_work,
 943                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 944 }
 945
 946 static void ntb_qp_link_work(struct work_struct *work)
 947 {
 948         struct ntb_transport_qp *qp = container_of(work,
 949                                                    struct ntb_transport_qp,
 950                                                    link_work.work);
 951         struct pci_dev *pdev = qp->ndev->pdev;
 952         struct ntb_transport_ctx *nt = qp->transport;
 953         int val;
 954
 955         WARN_ON(!nt->link_is_up);
 956
 957         val = ntb_spad_read(nt->ndev, QP_LINKS);
 958
 959         ntb_peer_spad_write(nt->ndev, QP_LINKS, val | BIT(qp->qp_num));
 960
 961         /* query remote spad for qp ready bits */
 962         ntb_peer_spad_read(nt->ndev, QP_LINKS);
 963         dev_dbg_ratelimited(&pdev->dev, "Remote QP link status = %x\n", val);
 964
 965         /* See if the remote side is up */
 966         if (val & BIT(qp->qp_num)) {
 967                 dev_info(&pdev->dev, "qp %d: Link Up\n", qp->qp_num);
 968                 qp->link_is_up = true;
 969                 qp->active = true;
 970
 971                 if (qp->event_handler)
 972                         qp->event_handler(qp->cb_data, qp->link_is_up);
 973
 974                 if (qp->active)
 975                         tasklet_schedule(&qp->rxc_db_work);
 976         } else if (nt->link_is_up)
 977                 schedule_delayed_work(&qp->link_work,
 978                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 979 }
 980
 981 static int ntb_transport_init_queue(struct ntb_transport_ctx *nt,
 982                                     unsigned int qp_num)
 983 {
 984         struct ntb_transport_qp *qp;
 985         phys_addr_t mw_base;
 986         resource_size_t mw_size;
 987         unsigned int num_qps_mw, tx_size;
 988         unsigned int mw_num, mw_count, qp_count;
 989         u64 qp_offset;
 990
 991         mw_count = nt->mw_count;
 992         qp_count = nt->qp_count;
 993
 994         mw_num = QP_TO_MW(nt, qp_num);
 995
 996         qp = &nt->qp_vec[qp_num];
 997         qp->qp_num = qp_num;
 998         qp->transport = nt;
 999         qp->ndev = nt->ndev;
1000         qp->client_ready = false;
1001         qp->event_handler = NULL;
1002         ntb_qp_link_down_reset(qp);
1003
1004         if (qp_count % mw_count && mw_num + 1 < qp_count / mw_count)
1005                 num_qps_mw = qp_count / mw_count + 1;
1006         else
1007                 num_qps_mw = qp_count / mw_count;
1008
1009         mw_base = nt->mw_vec[mw_num].phys_addr;
1010         mw_size = nt->mw_vec[mw_num].phys_size;
1011
1012         tx_size = (unsigned int)mw_size / num_qps_mw;
1013         qp_offset = tx_size * (qp_num / mw_count);
1014
1015         qp->tx_mw = nt->mw_vec[mw_num].vbase + qp_offset;
1016         if (!qp->tx_mw)
1017                 return -EINVAL;
1018
1019         qp->tx_mw_phys = mw_base + qp_offset;
1020         if (!qp->tx_mw_phys)
1021                 return -EINVAL;
1022
1023         tx_size -= sizeof(struct ntb_rx_info);
1024         qp->rx_info = qp->tx_mw + tx_size;
1025
1026         /* Due to housekeeping, there must be atleast 2 buffs */
1027         qp->tx_max_frame = min(transport_mtu, tx_size / 2);
1028         qp->tx_max_entry = tx_size / qp->tx_max_frame;
1029
1030         if (nt->debugfs_node_dir) {
1031                 char debugfs_name[4];
1032
1033                 snprintf(debugfs_name, 4, "qp%d", qp_num);
1034                 qp->debugfs_dir = debugfs_create_dir(debugfs_name,
1035                                                      nt->debugfs_node_dir);
1036
1037                 qp->debugfs_stats = debugfs_create_file("stats", S_IRUSR,
1038                                                         qp->debugfs_dir, qp,
1039                                                         &ntb_qp_debugfs_stats);
1040         } else {
1041                 qp->debugfs_dir = NULL;
1042                 qp->debugfs_stats = NULL;
1043         }
1044
1045         INIT_DELAYED_WORK(&qp->link_work, ntb_qp_link_work);
1046         INIT_WORK(&qp->link_cleanup, ntb_qp_link_cleanup_work);
1047
1048         spin_lock_init(&qp->ntb_rx_q_lock);
1049         spin_lock_init(&qp->ntb_tx_free_q_lock);
1050
1051         INIT_LIST_HEAD(&qp->rx_post_q);
1052         INIT_LIST_HEAD(&qp->rx_pend_q);
1053         INIT_LIST_HEAD(&qp->rx_free_q);
1054         INIT_LIST_HEAD(&qp->tx_free_q);
1055
1056         tasklet_init(&qp->rxc_db_work, ntb_transport_rxc_db,
1057                      (unsigned long)qp);
1058
1059         return 0;
1060 }
1061
1062 static int ntb_transport_probe(struct ntb_client *self, struct ntb_dev *ndev)
1063 {
1064         struct ntb_transport_ctx *nt;
1065         struct ntb_transport_mw *mw;
1066         unsigned int mw_count, qp_count, spad_count, max_mw_count_for_spads;
1067         u64 qp_bitmap;
1068         int node;
1069         int rc, i;
1070
1071         mw_count = ntb_mw_count(ndev);
1072
1073         if (ntb_db_is_unsafe(ndev))
1074                 dev_dbg(&ndev->dev,
1075                         "doorbell is unsafe, proceed anyway...\n");
1076         if (ntb_spad_is_unsafe(ndev))
1077                 dev_dbg(&ndev->dev,
1078                         "scratchpad is unsafe, proceed anyway...\n");
1079
1080         node = dev_to_node(&ndev->dev);
1081
1082         nt = kzalloc_node(sizeof(*nt), GFP_KERNEL, node);
1083         if (!nt)
1084                 return -ENOMEM;
1085
1086         nt->ndev = ndev;
1087         spad_count = ntb_spad_count(ndev);
1088
1089         /* Limit the MW's based on the availability of scratchpads */
1090
1091         if (spad_count < NTB_TRANSPORT_MIN_SPADS) {
1092                 nt->mw_count = 0;
1093                 rc = -EINVAL;
1094                 goto err;
1095         }
1096
1097         max_mw_count_for_spads = (spad_count - MW0_SZ_HIGH) / 2;
1098         nt->mw_count = min(mw_count, max_mw_count_for_spads);
1099
1100         nt->mw_vec = kzalloc_node(mw_count * sizeof(*nt->mw_vec),
1101                                   GFP_KERNEL, node);
1102         if (!nt->mw_vec) {
1103                 rc = -ENOMEM;
1104                 goto err;
1105         }
1106
1107         for (i = 0; i < mw_count; i++) {
1108                 mw = &nt->mw_vec[i];
1109
1110                 rc = ntb_mw_get_range(ndev, i, &mw->phys_addr, &mw->phys_size,
1111                                       &mw->xlat_align, &mw->xlat_align_size);
1112                 if (rc)
1113                         goto err1;
1114
1115                 mw->vbase = ioremap_wc(mw->phys_addr, mw->phys_size);
1116                 if (!mw->vbase) {
1117                         rc = -ENOMEM;
1118                         goto err1;
1119                 }
1120
1121                 mw->buff_size = 0;
1122                 mw->xlat_size = 0;
1123                 mw->virt_addr = NULL;
1124                 mw->dma_addr = 0;
1125         }
1126
1127         qp_bitmap = ntb_db_valid_mask(ndev);
1128
1129         qp_count = ilog2(qp_bitmap);
1130         if (max_num_clients && max_num_clients < qp_count)
1131                 qp_count = max_num_clients;
1132         else if (mw_count < qp_count)
1133                 qp_count = mw_count;
1134
1135         qp_bitmap &= BIT_ULL(qp_count) - 1;
1136
1137         nt->qp_count = qp_count;
1138         nt->qp_bitmap = qp_bitmap;
1139         nt->qp_bitmap_free = qp_bitmap;
1140
1141         nt->qp_vec = kzalloc_node(qp_count * sizeof(*nt->qp_vec),
1142                                   GFP_KERNEL, node);
1143         if (!nt->qp_vec) {
1144                 rc = -ENOMEM;
1145                 goto err1;
1146         }
1147
1148         if (nt_debugfs_dir) {
1149                 nt->debugfs_node_dir =
1150                         debugfs_create_dir(pci_name(ndev->pdev),
1151                                            nt_debugfs_dir);
1152         }
1153
1154         for (i = 0; i < qp_count; i++) {
1155                 rc = ntb_transport_init_queue(nt, i);
1156                 if (rc)
1157                         goto err2;
1158         }
1159
1160         INIT_DELAYED_WORK(&nt->link_work, ntb_transport_link_work);
1161         INIT_WORK(&nt->link_cleanup, ntb_transport_link_cleanup_work);
1162
1163         rc = ntb_set_ctx(ndev, nt, &ntb_transport_ops);
1164         if (rc)
1165                 goto err2;
1166
1167         INIT_LIST_HEAD(&nt->client_devs);
1168         rc = ntb_bus_init(nt);
1169         if (rc)
1170                 goto err3;
1171
1172         nt->link_is_up = false;
1173         ntb_link_enable(ndev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
1174         ntb_link_event(ndev);
1175
1176         return 0;
1177
1178 err3:
1179         ntb_clear_ctx(ndev);
1180 err2:
1181         kfree(nt->qp_vec);
1182 err1:
1183         while (i--) {
1184                 mw = &nt->mw_vec[i];
1185                 iounmap(mw->vbase);
1186         }
1187         kfree(nt->mw_vec);
1188 err:
1189         kfree(nt);
1190         return rc;
1191 }
1192
1193 static void ntb_transport_free(struct ntb_client *self, struct ntb_dev *ndev)
1194 {
1195         struct ntb_transport_ctx *nt = ndev->ctx;
1196         struct ntb_transport_qp *qp;
1197         u64 qp_bitmap_alloc;
1198         int i;
1199
1200         ntb_transport_link_cleanup(nt);
1201         cancel_work_sync(&nt->link_cleanup);
1202         cancel_delayed_work_sync(&nt->link_work);
1203
1204         qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
1205
1206         /* verify that all the qp's are freed */
1207         for (i = 0; i < nt->qp_count; i++) {
1208                 qp = &nt->qp_vec[i];
1209                 if (qp_bitmap_alloc & BIT_ULL(i))
1210                         ntb_transport_free_queue(qp);
1211                 debugfs_remove_recursive(qp->debugfs_dir);
1212         }
1213
1214         ntb_link_disable(ndev);
1215         ntb_clear_ctx(ndev);
1216
1217         ntb_bus_remove(nt);
1218
1219         for (i = nt->mw_count; i--; ) {
1220                 ntb_free_mw(nt, i);
1221                 iounmap(nt->mw_vec[i].vbase);
1222         }
1223
1224         kfree(nt->qp_vec);
1225         kfree(nt->mw_vec);
1226         kfree(nt);
1227 }
1228
1229 static void ntb_complete_rxc(struct ntb_transport_qp *qp)
1230 {
1231         struct ntb_queue_entry *entry;
1232         void *cb_data;
1233         unsigned int len;
1234         unsigned long irqflags;
1235
1236         spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1237
1238         while (!list_empty(&qp->rx_post_q)) {
1239                 entry = list_first_entry(&qp->rx_post_q,
1240                                          struct ntb_queue_entry, entry);
1241                 if (!(entry->flags & DESC_DONE_FLAG))
1242                         break;
1243
1244                 entry->rx_hdr->flags = 0;
1245                 iowrite32(entry->rx_index, &qp->rx_info->entry);
1246
1247                 cb_data = entry->cb_data;
1248                 len = entry->len;
1249
1250                 list_move_tail(&entry->entry, &qp->rx_free_q);
1251
1252                 spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1253
1254                 if (qp->rx_handler && qp->client_ready)
1255                         qp->rx_handler(qp, qp->cb_data, cb_data, len);
1256
1257                 spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1258         }
1259
1260         spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1261 }
1262
1263 static void ntb_rx_copy_callback(void *data,
1264                                  const struct dmaengine_result *res)
1265 {
1266         struct ntb_queue_entry *entry = data;
1267
1268         /* we need to check DMA results if we are using DMA */
1269         if (res) {
1270                 enum dmaengine_tx_result dma_err = res->result;
1271
1272                 switch (dma_err) {
1273                 case DMA_TRANS_READ_FAILED:
1274                 case DMA_TRANS_WRITE_FAILED:
1275                         entry->errors++;
1276                 case DMA_TRANS_ABORTED:
1277                 {
1278                         struct ntb_transport_qp *qp = entry->qp;
1279                         void *offset = qp->rx_buff + qp->rx_max_frame *
1280                                         qp->rx_index;
1281
1282                         ntb_memcpy_rx(entry, offset);
1283                         qp->rx_memcpy++;
1284                         return;
1285                 }
1286
1287                 case DMA_TRANS_NOERROR:
1288                 default:
1289                         break;
1290                 }
1291         }
1292
1293         entry->flags |= DESC_DONE_FLAG;
1294
1295         ntb_complete_rxc(entry->qp);
1296 }
1297
1298 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset)
1299 {
1300         void *buf = entry->buf;
1301         size_t len = entry->len;
1302
1303         memcpy(buf, offset, len);
1304
1305         /* Ensure that the data is fully copied out before clearing the flag */
1306         wmb();
1307
1308         ntb_rx_copy_callback(entry, NULL);
1309 }
1310
1311 static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset)
1312 {
1313         struct dma_async_tx_descriptor *txd;
1314         struct ntb_transport_qp *qp = entry->qp;
1315         struct dma_chan *chan = qp->rx_dma_chan;
1316         struct dma_device *device;
1317         size_t pay_off, buff_off, len;
1318         struct dmaengine_unmap_data *unmap;
1319         dma_cookie_t cookie;
1320         void *buf = entry->buf;
1321         int retries = 0;
1322
1323         len = entry->len;
1324         device = chan->device;
1325         pay_off = (size_t)offset & ~PAGE_MASK;
1326         buff_off = (size_t)buf & ~PAGE_MASK;
1327
1328         if (!is_dma_copy_aligned(device, pay_off, buff_off, len))
1329                 goto err;
1330
1331         unmap = dmaengine_get_unmap_data(device->dev, 2, GFP_NOWAIT);
1332         if (!unmap)
1333                 goto err;
1334
1335         unmap->len = len;
1336         unmap->addr[0] = dma_map_page(device->dev, virt_to_page(offset),
1337                                       pay_off, len, DMA_TO_DEVICE);
1338         if (dma_mapping_error(device->dev, unmap->addr[0]))
1339                 goto err_get_unmap;
1340
1341         unmap->to_cnt = 1;
1342
1343         unmap->addr[1] = dma_map_page(device->dev, virt_to_page(buf),
1344                                       buff_off, len, DMA_FROM_DEVICE);
1345         if (dma_mapping_error(device->dev, unmap->addr[1]))
1346                 goto err_get_unmap;
1347
1348         unmap->from_cnt = 1;
1349
1350         for (retries = 0; retries < DMA_RETRIES; retries++) {
1351                 txd = device->device_prep_dma_memcpy(chan,
1352                                                      unmap->addr[1],
1353                                                      unmap->addr[0], len,
1354                                                      DMA_PREP_INTERRUPT);
1355                 if (txd)
1356                         break;
1357
1358                 set_current_state(TASK_INTERRUPTIBLE);
1359                 schedule_timeout(DMA_OUT_RESOURCE_TO);
1360         }
1361
1362         if (!txd) {
1363                 qp->dma_rx_prep_err++;
1364                 goto err_get_unmap;
1365         }
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         int retries = 0;
1611
1612         device = chan->device;
1613         dest = qp->tx_mw_phys + qp->tx_max_frame * entry->tx_index;
1614         buff_off = (size_t)buf & ~PAGE_MASK;
1615         dest_off = (size_t)dest & ~PAGE_MASK;
1616
1617         if (!is_dma_copy_aligned(device, buff_off, dest_off, len))
1618                 goto err;
1619
1620         unmap = dmaengine_get_unmap_data(device->dev, 1, GFP_NOWAIT);
1621         if (!unmap)
1622                 goto err;
1623
1624         unmap->len = len;
1625         unmap->addr[0] = dma_map_page(device->dev, virt_to_page(buf),
1626                                       buff_off, len, DMA_TO_DEVICE);
1627         if (dma_mapping_error(device->dev, unmap->addr[0]))
1628                 goto err_get_unmap;
1629
1630         unmap->to_cnt = 1;
1631
1632         for (retries = 0; retries < DMA_RETRIES; retries++) {
1633                 txd = device->device_prep_dma_memcpy(chan, dest,
1634                                                      unmap->addr[0], len,
1635                                                      DMA_PREP_INTERRUPT);
1636                 if (txd)
1637                         break;
1638
1639                 set_current_state(TASK_INTERRUPTIBLE);
1640                 schedule_timeout(DMA_OUT_RESOURCE_TO);
1641         }
1642
1643         if (!txd) {
1644                 qp->dma_tx_prep_err++;
1645                 goto err_get_unmap;
1646         }
1647
1648         txd->callback_result = ntb_tx_copy_callback;
1649         txd->callback_param = entry;
1650         dma_set_unmap(txd, unmap);
1651
1652         cookie = dmaengine_submit(txd);
1653         if (dma_submit_error(cookie))
1654                 goto err_set_unmap;
1655
1656         dmaengine_unmap_put(unmap);
1657
1658         dma_async_issue_pending(chan);
1659
1660         return 0;
1661 err_set_unmap:
1662         dmaengine_unmap_put(unmap);
1663 err_get_unmap:
1664         dmaengine_unmap_put(unmap);
1665 err:
1666         return -ENXIO;
1667 }
1668
1669 static void ntb_async_tx(struct ntb_transport_qp *qp,
1670                          struct ntb_queue_entry *entry)
1671 {
1672         struct ntb_payload_header __iomem *hdr;
1673         struct dma_chan *chan = qp->tx_dma_chan;
1674         void __iomem *offset;
1675         int res;
1676
1677         entry->tx_index = qp->tx_index;
1678         offset = qp->tx_mw + qp->tx_max_frame * entry->tx_index;
1679         hdr = offset + qp->tx_max_frame - sizeof(struct ntb_payload_header);
1680         entry->tx_hdr = hdr;
1681
1682         iowrite32(entry->len, &hdr->len);
1683         iowrite32((u32)qp->tx_pkts, &hdr->ver);
1684
1685         if (!chan)
1686                 goto err;
1687
1688         if (entry->len < copy_bytes)
1689                 goto err;
1690
1691         res = ntb_async_tx_submit(qp, entry);
1692         if (res < 0)
1693                 goto err;
1694
1695         if (!entry->retries)
1696                 qp->tx_async++;
1697
1698         return;
1699
1700 err:
1701         ntb_memcpy_tx(entry, offset);
1702         qp->tx_memcpy++;
1703 }
1704
1705 static int ntb_process_tx(struct ntb_transport_qp *qp,
1706                           struct ntb_queue_entry *entry)
1707 {
1708         if (qp->tx_index == qp->remote_rx_info->entry) {
1709                 qp->tx_ring_full++;
1710                 return -EAGAIN;
1711         }
1712
1713         if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
1714                 if (qp->tx_handler)
1715                         qp->tx_handler(qp, qp->cb_data, NULL, -EIO);
1716
1717                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1718                              &qp->tx_free_q);
1719                 return 0;
1720         }
1721
1722         ntb_async_tx(qp, entry);
1723
1724         qp->tx_index++;
1725         qp->tx_index %= qp->tx_max_entry;
1726
1727         qp->tx_pkts++;
1728
1729         return 0;
1730 }
1731
1732 static void ntb_send_link_down(struct ntb_transport_qp *qp)
1733 {
1734         struct pci_dev *pdev = qp->ndev->pdev;
1735         struct ntb_queue_entry *entry;
1736         int i, rc;
1737
1738         if (!qp->link_is_up)
1739                 return;
1740
1741         dev_info(&pdev->dev, "qp %d: Send Link Down\n", qp->qp_num);
1742
1743         for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1744                 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1745                 if (entry)
1746                         break;
1747                 msleep(100);
1748         }
1749
1750         if (!entry)
1751                 return;
1752
1753         entry->cb_data = NULL;
1754         entry->buf = NULL;
1755         entry->len = 0;
1756         entry->flags = LINK_DOWN_FLAG;
1757
1758         rc = ntb_process_tx(qp, entry);
1759         if (rc)
1760                 dev_err(&pdev->dev, "ntb: QP%d unable to send linkdown msg\n",
1761                         qp->qp_num);
1762
1763         ntb_qp_link_down_reset(qp);
1764 }
1765
1766 static bool ntb_dma_filter_fn(struct dma_chan *chan, void *node)
1767 {
1768         return dev_to_node(&chan->dev->device) == (int)(unsigned long)node;
1769 }
1770
1771 /**
1772  * ntb_transport_create_queue - Create a new NTB transport layer queue
1773  * @rx_handler: receive callback function
1774  * @tx_handler: transmit callback function
1775  * @event_handler: event callback function
1776  *
1777  * Create a new NTB transport layer queue and provide the queue with a callback
1778  * routine for both transmit and receive.  The receive callback routine will be
1779  * used to pass up data when the transport has received it on the queue.   The
1780  * transmit callback routine will be called when the transport has completed the
1781  * transmission of the data on the queue and the data is ready to be freed.
1782  *
1783  * RETURNS: pointer to newly created ntb_queue, NULL on error.
1784  */
1785 struct ntb_transport_qp *
1786 ntb_transport_create_queue(void *data, struct device *client_dev,
1787                            const struct ntb_queue_handlers *handlers)
1788 {
1789         struct ntb_dev *ndev;
1790         struct pci_dev *pdev;
1791         struct ntb_transport_ctx *nt;
1792         struct ntb_queue_entry *entry;
1793         struct ntb_transport_qp *qp;
1794         u64 qp_bit;
1795         unsigned int free_queue;
1796         dma_cap_mask_t dma_mask;
1797         int node;
1798         int i;
1799
1800         ndev = dev_ntb(client_dev->parent);
1801         pdev = ndev->pdev;
1802         nt = ndev->ctx;
1803
1804         node = dev_to_node(&ndev->dev);
1805
1806         free_queue = ffs(nt->qp_bitmap);
1807         if (!free_queue)
1808                 goto err;
1809
1810         /* decrement free_queue to make it zero based */
1811         free_queue--;
1812
1813         qp = &nt->qp_vec[free_queue];
1814         qp_bit = BIT_ULL(qp->qp_num);
1815
1816         nt->qp_bitmap_free &= ~qp_bit;
1817
1818         qp->cb_data = data;
1819         qp->rx_handler = handlers->rx_handler;
1820         qp->tx_handler = handlers->tx_handler;
1821         qp->event_handler = handlers->event_handler;
1822
1823         dma_cap_zero(dma_mask);
1824         dma_cap_set(DMA_MEMCPY, dma_mask);
1825
1826         if (use_dma) {
1827                 qp->tx_dma_chan =
1828                         dma_request_channel(dma_mask, ntb_dma_filter_fn,
1829                                             (void *)(unsigned long)node);
1830                 if (!qp->tx_dma_chan)
1831                         dev_info(&pdev->dev, "Unable to allocate TX DMA channel\n");
1832
1833                 qp->rx_dma_chan =
1834                         dma_request_channel(dma_mask, ntb_dma_filter_fn,
1835                                             (void *)(unsigned long)node);
1836                 if (!qp->rx_dma_chan)
1837                         dev_info(&pdev->dev, "Unable to allocate RX DMA channel\n");
1838         } else {
1839                 qp->tx_dma_chan = NULL;
1840                 qp->rx_dma_chan = NULL;
1841         }
1842
1843         dev_dbg(&pdev->dev, "Using %s memcpy for TX\n",
1844                 qp->tx_dma_chan ? "DMA" : "CPU");
1845
1846         dev_dbg(&pdev->dev, "Using %s memcpy for RX\n",
1847                 qp->rx_dma_chan ? "DMA" : "CPU");
1848
1849         for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
1850                 entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
1851                 if (!entry)
1852                         goto err1;
1853
1854                 entry->qp = qp;
1855                 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
1856                              &qp->rx_free_q);
1857         }
1858         qp->rx_alloc_entry = NTB_QP_DEF_NUM_ENTRIES;
1859
1860         for (i = 0; i < qp->tx_max_entry; i++) {
1861                 entry = kzalloc_node(sizeof(*entry), GFP_ATOMIC, node);
1862                 if (!entry)
1863                         goto err2;
1864
1865                 entry->qp = qp;
1866                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1867                              &qp->tx_free_q);
1868         }
1869
1870         ntb_db_clear(qp->ndev, qp_bit);
1871         ntb_db_clear_mask(qp->ndev, qp_bit);
1872
1873         dev_info(&pdev->dev, "NTB Transport QP %d created\n", qp->qp_num);
1874
1875         return qp;
1876
1877 err2:
1878         while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1879                 kfree(entry);
1880 err1:
1881         qp->rx_alloc_entry = 0;
1882         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
1883                 kfree(entry);
1884         if (qp->tx_dma_chan)
1885                 dma_release_channel(qp->tx_dma_chan);
1886         if (qp->rx_dma_chan)
1887                 dma_release_channel(qp->rx_dma_chan);
1888         nt->qp_bitmap_free |= qp_bit;
1889 err:
1890         return NULL;
1891 }
1892 EXPORT_SYMBOL_GPL(ntb_transport_create_queue);
1893
1894 /**
1895  * ntb_transport_free_queue - Frees NTB transport queue
1896  * @qp: NTB queue to be freed
1897  *
1898  * Frees NTB transport queue
1899  */
1900 void ntb_transport_free_queue(struct ntb_transport_qp *qp)
1901 {
1902         struct pci_dev *pdev;
1903         struct ntb_queue_entry *entry;
1904         u64 qp_bit;
1905
1906         if (!qp)
1907                 return;
1908
1909         pdev = qp->ndev->pdev;
1910
1911         qp->active = false;
1912
1913         if (qp->tx_dma_chan) {
1914                 struct dma_chan *chan = qp->tx_dma_chan;
1915                 /* Putting the dma_chan to NULL will force any new traffic to be
1916                  * processed by the CPU instead of the DAM engine
1917                  */
1918                 qp->tx_dma_chan = NULL;
1919
1920                 /* Try to be nice and wait for any queued DMA engine
1921                  * transactions to process before smashing it with a rock
1922                  */
1923                 dma_sync_wait(chan, qp->last_cookie);
1924                 dmaengine_terminate_all(chan);
1925                 dma_release_channel(chan);
1926         }
1927
1928         if (qp->rx_dma_chan) {
1929                 struct dma_chan *chan = qp->rx_dma_chan;
1930                 /* Putting the dma_chan to NULL will force any new traffic to be
1931                  * processed by the CPU instead of the DAM engine
1932                  */
1933                 qp->rx_dma_chan = NULL;
1934
1935                 /* Try to be nice and wait for any queued DMA engine
1936                  * transactions to process before smashing it with a rock
1937                  */
1938                 dma_sync_wait(chan, qp->last_cookie);
1939                 dmaengine_terminate_all(chan);
1940                 dma_release_channel(chan);
1941         }
1942
1943         qp_bit = BIT_ULL(qp->qp_num);
1944
1945         ntb_db_set_mask(qp->ndev, qp_bit);
1946         tasklet_kill(&qp->rxc_db_work);
1947
1948         cancel_delayed_work_sync(&qp->link_work);
1949
1950         qp->cb_data = NULL;
1951         qp->rx_handler = NULL;
1952         qp->tx_handler = NULL;
1953         qp->event_handler = NULL;
1954
1955         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
1956                 kfree(entry);
1957
1958         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q))) {
1959                 dev_warn(&pdev->dev, "Freeing item from non-empty rx_pend_q\n");
1960                 kfree(entry);
1961         }
1962
1963         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q))) {
1964                 dev_warn(&pdev->dev, "Freeing item from non-empty rx_post_q\n");
1965                 kfree(entry);
1966         }
1967
1968         while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
1969                 kfree(entry);
1970
1971         qp->transport->qp_bitmap_free |= qp_bit;
1972
1973         dev_info(&pdev->dev, "NTB Transport QP %d freed\n", qp->qp_num);
1974 }
1975 EXPORT_SYMBOL_GPL(ntb_transport_free_queue);
1976
1977 /**
1978  * ntb_transport_rx_remove - Dequeues enqueued rx packet
1979  * @qp: NTB queue to be freed
1980  * @len: pointer to variable to write enqueued buffers length
1981  *
1982  * Dequeues unused buffers from receive queue.  Should only be used during
1983  * shutdown of qp.
1984  *
1985  * RETURNS: NULL error value on error, or void* for success.
1986  */
1987 void *ntb_transport_rx_remove(struct ntb_transport_qp *qp, unsigned int *len)
1988 {
1989         struct ntb_queue_entry *entry;
1990         void *buf;
1991
1992         if (!qp || qp->client_ready)
1993                 return NULL;
1994
1995         entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q);
1996         if (!entry)
1997                 return NULL;
1998
1999         buf = entry->cb_data;
2000         *len = entry->len;
2001
2002         ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_free_q);
2003
2004         return buf;
2005 }
2006 EXPORT_SYMBOL_GPL(ntb_transport_rx_remove);
2007
2008 /**
2009  * ntb_transport_rx_enqueue - Enqueue a new NTB queue entry
2010  * @qp: NTB transport layer queue the entry is to be enqueued on
2011  * @cb: per buffer pointer for callback function to use
2012  * @data: pointer to data buffer that incoming packets will be copied into
2013  * @len: length of the data buffer
2014  *
2015  * Enqueue a new receive buffer onto the transport queue into which a NTB
2016  * payload can be received into.
2017  *
2018  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2019  */
2020 int ntb_transport_rx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
2021                              unsigned int len)
2022 {
2023         struct ntb_queue_entry *entry;
2024
2025         if (!qp)
2026                 return -EINVAL;
2027
2028         entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q);
2029         if (!entry)
2030                 return -ENOMEM;
2031
2032         entry->cb_data = cb;
2033         entry->buf = data;
2034         entry->len = len;
2035         entry->flags = 0;
2036         entry->retries = 0;
2037         entry->errors = 0;
2038         entry->rx_index = 0;
2039
2040         ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_pend_q);
2041
2042         if (qp->active)
2043                 tasklet_schedule(&qp->rxc_db_work);
2044
2045         return 0;
2046 }
2047 EXPORT_SYMBOL_GPL(ntb_transport_rx_enqueue);
2048
2049 /**
2050  * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
2051  * @qp: NTB transport layer queue the entry is to be enqueued on
2052  * @cb: per buffer pointer for callback function to use
2053  * @data: pointer to data buffer that will be sent
2054  * @len: length of the data buffer
2055  *
2056  * Enqueue a new transmit buffer onto the transport queue from which a NTB
2057  * payload will be transmitted.  This assumes that a lock is being held to
2058  * serialize access to the qp.
2059  *
2060  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2061  */
2062 int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
2063                              unsigned int len)
2064 {
2065         struct ntb_queue_entry *entry;
2066         int rc;
2067
2068         if (!qp || !qp->link_is_up || !len)
2069                 return -EINVAL;
2070
2071         entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
2072         if (!entry) {
2073                 qp->tx_err_no_buf++;
2074                 return -EBUSY;
2075         }
2076
2077         entry->cb_data = cb;
2078         entry->buf = data;
2079         entry->len = len;
2080         entry->flags = 0;
2081         entry->errors = 0;
2082         entry->retries = 0;
2083         entry->tx_index = 0;
2084
2085         rc = ntb_process_tx(qp, entry);
2086         if (rc)
2087                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
2088                              &qp->tx_free_q);
2089
2090         return rc;
2091 }
2092 EXPORT_SYMBOL_GPL(ntb_transport_tx_enqueue);
2093
2094 /**
2095  * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
2096  * @qp: NTB transport layer queue to be enabled
2097  *
2098  * Notify NTB transport layer of client readiness to use queue
2099  */
2100 void ntb_transport_link_up(struct ntb_transport_qp *qp)
2101 {
2102         if (!qp)
2103                 return;
2104
2105         qp->client_ready = true;
2106
2107         if (qp->transport->link_is_up)
2108                 schedule_delayed_work(&qp->link_work, 0);
2109 }
2110 EXPORT_SYMBOL_GPL(ntb_transport_link_up);
2111
2112 /**
2113  * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
2114  * @qp: NTB transport layer queue to be disabled
2115  *
2116  * Notify NTB transport layer of client's desire to no longer receive data on
2117  * transport queue specified.  It is the client's responsibility to ensure all
2118  * entries on queue are purged or otherwise handled appropriately.
2119  */
2120 void ntb_transport_link_down(struct ntb_transport_qp *qp)
2121 {
2122         int val;
2123
2124         if (!qp)
2125                 return;
2126
2127         qp->client_ready = false;
2128
2129         val = ntb_spad_read(qp->ndev, QP_LINKS);
2130
2131         ntb_peer_spad_write(qp->ndev, QP_LINKS,
2132                             val & ~BIT(qp->qp_num));
2133
2134         if (qp->link_is_up)
2135                 ntb_send_link_down(qp);
2136         else
2137                 cancel_delayed_work_sync(&qp->link_work);
2138 }
2139 EXPORT_SYMBOL_GPL(ntb_transport_link_down);
2140
2141 /**
2142  * ntb_transport_link_query - Query transport link state
2143  * @qp: NTB transport layer queue to be queried
2144  *
2145  * Query connectivity to the remote system of the NTB transport queue
2146  *
2147  * RETURNS: true for link up or false for link down
2148  */
2149 bool ntb_transport_link_query(struct ntb_transport_qp *qp)
2150 {
2151         if (!qp)
2152                 return false;
2153
2154         return qp->link_is_up;
2155 }
2156 EXPORT_SYMBOL_GPL(ntb_transport_link_query);
2157
2158 /**
2159  * ntb_transport_qp_num - Query the qp number
2160  * @qp: NTB transport layer queue to be queried
2161  *
2162  * Query qp number of the NTB transport queue
2163  *
2164  * RETURNS: a zero based number specifying the qp number
2165  */
2166 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
2167 {
2168         if (!qp)
2169                 return 0;
2170
2171         return qp->qp_num;
2172 }
2173 EXPORT_SYMBOL_GPL(ntb_transport_qp_num);
2174
2175 /**
2176  * ntb_transport_max_size - Query the max payload size of a qp
2177  * @qp: NTB transport layer queue to be queried
2178  *
2179  * Query the maximum payload size permissible on the given qp
2180  *
2181  * RETURNS: the max payload size of a qp
2182  */
2183 unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp)
2184 {
2185         unsigned int max_size;
2186         unsigned int copy_align;
2187         struct dma_chan *rx_chan, *tx_chan;
2188
2189         if (!qp)
2190                 return 0;
2191
2192         rx_chan = qp->rx_dma_chan;
2193         tx_chan = qp->tx_dma_chan;
2194
2195         copy_align = max(rx_chan ? rx_chan->device->copy_align : 0,
2196                          tx_chan ? tx_chan->device->copy_align : 0);
2197
2198         /* If DMA engine usage is possible, try to find the max size for that */
2199         max_size = qp->tx_max_frame - sizeof(struct ntb_payload_header);
2200         max_size = round_down(max_size, 1 << copy_align);
2201
2202         return max_size;
2203 }
2204 EXPORT_SYMBOL_GPL(ntb_transport_max_size);
2205
2206 unsigned int ntb_transport_tx_free_entry(struct ntb_transport_qp *qp)
2207 {
2208         unsigned int head = qp->tx_index;
2209         unsigned int tail = qp->remote_rx_info->entry;
2210
2211         return tail > head ? tail - head : qp->tx_max_entry + tail - head;
2212 }
2213 EXPORT_SYMBOL_GPL(ntb_transport_tx_free_entry);
2214
2215 static void ntb_transport_doorbell_callback(void *data, int vector)
2216 {
2217         struct ntb_transport_ctx *nt = data;
2218         struct ntb_transport_qp *qp;
2219         u64 db_bits;
2220         unsigned int qp_num;
2221
2222         db_bits = (nt->qp_bitmap & ~nt->qp_bitmap_free &
2223                    ntb_db_vector_mask(nt->ndev, vector));
2224
2225         while (db_bits) {
2226                 qp_num = __ffs(db_bits);
2227                 qp = &nt->qp_vec[qp_num];
2228
2229                 if (qp->active)
2230                         tasklet_schedule(&qp->rxc_db_work);
2231
2232                 db_bits &= ~BIT_ULL(qp_num);
2233         }
2234 }
2235
2236 static const struct ntb_ctx_ops ntb_transport_ops = {
2237         .link_event = ntb_transport_event_callback,
2238         .db_event = ntb_transport_doorbell_callback,
2239 };
2240
2241 static struct ntb_client ntb_transport_client = {
2242         .ops = {
2243                 .probe = ntb_transport_probe,
2244                 .remove = ntb_transport_free,
2245         },
2246 };
2247
2248 static int __init ntb_transport_init(void)
2249 {
2250         int rc;
2251
2252         pr_info("%s, version %s\n", NTB_TRANSPORT_DESC, NTB_TRANSPORT_VER);
2253
2254         if (debugfs_initialized())
2255                 nt_debugfs_dir = debugfs_create_dir(KBUILD_MODNAME, NULL);
2256
2257         rc = bus_register(&ntb_transport_bus);
2258         if (rc)
2259                 goto err_bus;
2260
2261         rc = ntb_register_client(&ntb_transport_client);
2262         if (rc)
2263                 goto err_client;
2264
2265         return 0;
2266
2267 err_client:
2268         bus_unregister(&ntb_transport_bus);
2269 err_bus:
2270         debugfs_remove_recursive(nt_debugfs_dir);
2271         return rc;
2272 }
2273 module_init(ntb_transport_init);
2274
2275 static void __exit ntb_transport_exit(void)
2276 {
2277         debugfs_remove_recursive(nt_debugfs_dir);
2278
2279         ntb_unregister_client(&ntb_transport_client);
2280         bus_unregister(&ntb_transport_bus);
2281 }
2282 module_exit(ntb_transport_exit);