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 bool use_msi;
  97 #ifdef CONFIG_NTB_MSI
  98 module_param(use_msi, bool, 0644);
  99 MODULE_PARM_DESC(use_msi, "Use MSI interrupts instead of doorbells");
 100 #endif
 101
 102 static struct dentry *nt_debugfs_dir;
 103
 104 /* Only two-ports NTB devices are supported */
 105 #define PIDX            NTB_DEF_PEER_IDX
 106
 107 struct ntb_queue_entry {
 108         /* ntb_queue list reference */
 109         struct list_head entry;
 110         /* pointers to data to be transferred */
 111         void *cb_data;
 112         void *buf;
 113         unsigned int len;
 114         unsigned int flags;
 115         int retries;
 116         int errors;
 117         unsigned int tx_index;
 118         unsigned int rx_index;
 119
 120         struct ntb_transport_qp *qp;
 121         union {
 122                 struct ntb_payload_header __iomem *tx_hdr;
 123                 struct ntb_payload_header *rx_hdr;
 124         };
 125 };
 126
 127 struct ntb_rx_info {
 128         unsigned int entry;
 129 };
 130
 131 struct ntb_transport_qp {
 132         struct ntb_transport_ctx *transport;
 133         struct ntb_dev *ndev;
 134         void *cb_data;
 135         struct dma_chan *tx_dma_chan;
 136         struct dma_chan *rx_dma_chan;
 137
 138         bool client_ready;
 139         bool link_is_up;
 140         bool active;
 141
 142         u8 qp_num;      /* Only 64 QP's are allowed.  0-63 */
 143         u64 qp_bit;
 144
 145         struct ntb_rx_info __iomem *rx_info;
 146         struct ntb_rx_info *remote_rx_info;
 147
 148         void (*tx_handler)(struct ntb_transport_qp *qp, void *qp_data,
 149                            void *data, int len);
 150         struct list_head tx_free_q;
 151         spinlock_t ntb_tx_free_q_lock;
 152         void __iomem *tx_mw;
 153         phys_addr_t tx_mw_phys;
 154         size_t tx_mw_size;
 155         dma_addr_t tx_mw_dma_addr;
 156         unsigned int tx_index;
 157         unsigned int tx_max_entry;
 158         unsigned int tx_max_frame;
 159
 160         void (*rx_handler)(struct ntb_transport_qp *qp, void *qp_data,
 161                            void *data, int len);
 162         struct list_head rx_post_q;
 163         struct list_head rx_pend_q;
 164         struct list_head rx_free_q;
 165         /* ntb_rx_q_lock: synchronize access to rx_XXXX_q */
 166         spinlock_t ntb_rx_q_lock;
 167         void *rx_buff;
 168         unsigned int rx_index;
 169         unsigned int rx_max_entry;
 170         unsigned int rx_max_frame;
 171         unsigned int rx_alloc_entry;
 172         dma_cookie_t last_cookie;
 173         struct tasklet_struct rxc_db_work;
 174
 175         void (*event_handler)(void *data, int status);
 176         struct delayed_work link_work;
 177         struct work_struct link_cleanup;
 178
 179         struct dentry *debugfs_dir;
 180         struct dentry *debugfs_stats;
 181
 182         /* Stats */
 183         u64 rx_bytes;
 184         u64 rx_pkts;
 185         u64 rx_ring_empty;
 186         u64 rx_err_no_buf;
 187         u64 rx_err_oflow;
 188         u64 rx_err_ver;
 189         u64 rx_memcpy;
 190         u64 rx_async;
 191         u64 tx_bytes;
 192         u64 tx_pkts;
 193         u64 tx_ring_full;
 194         u64 tx_err_no_buf;
 195         u64 tx_memcpy;
 196         u64 tx_async;
 197
 198         bool use_msi;
 199         int msi_irq;
 200         struct ntb_msi_desc msi_desc;
 201         struct ntb_msi_desc peer_msi_desc;
 202 };
 203
 204 struct ntb_transport_mw {
 205         phys_addr_t phys_addr;
 206         resource_size_t phys_size;
 207         void __iomem *vbase;
 208         size_t xlat_size;
 209         size_t buff_size;
 210         size_t alloc_size;
 211         void *alloc_addr;
 212         void *virt_addr;
 213         dma_addr_t dma_addr;
 214 };
 215
 216 struct ntb_transport_client_dev {
 217         struct list_head entry;
 218         struct ntb_transport_ctx *nt;
 219         struct device dev;
 220 };
 221
 222 struct ntb_transport_ctx {
 223         struct list_head entry;
 224         struct list_head client_devs;
 225
 226         struct ntb_dev *ndev;
 227
 228         struct ntb_transport_mw *mw_vec;
 229         struct ntb_transport_qp *qp_vec;
 230         unsigned int mw_count;
 231         unsigned int qp_count;
 232         u64 qp_bitmap;
 233         u64 qp_bitmap_free;
 234
 235         bool use_msi;
 236         unsigned int msi_spad_offset;
 237         u64 msi_db_mask;
 238
 239         bool link_is_up;
 240         struct delayed_work link_work;
 241         struct work_struct link_cleanup;
 242
 243         struct dentry *debugfs_node_dir;
 244 };
 245
 246 enum {
 247         DESC_DONE_FLAG = BIT(0),
 248         LINK_DOWN_FLAG = BIT(1),
 249 };
 250
 251 struct ntb_payload_header {
 252         unsigned int ver;
 253         unsigned int len;
 254         unsigned int flags;
 255 };
 256
 257 enum {
 258         VERSION = 0,
 259         QP_LINKS,
 260         NUM_QPS,
 261         NUM_MWS,
 262         MW0_SZ_HIGH,
 263         MW0_SZ_LOW,
 264 };
 265
 266 #define dev_client_dev(__dev) \
 267         container_of((__dev), struct ntb_transport_client_dev, dev)
 268
 269 #define drv_client(__drv) \
 270         container_of((__drv), struct ntb_transport_client, driver)
 271
 272 #define QP_TO_MW(nt, qp)        ((qp) % nt->mw_count)
 273 #define NTB_QP_DEF_NUM_ENTRIES  100
 274 #define NTB_LINK_DOWN_TIMEOUT   10
 275
 276 static void ntb_transport_rxc_db(unsigned long data);
 277 static const struct ntb_ctx_ops ntb_transport_ops;
 278 static struct ntb_client ntb_transport_client;
 279 static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
 280                                struct ntb_queue_entry *entry);
 281 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset);
 282 static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset);
 283 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset);
 284
 285
 286 static int ntb_transport_bus_match(struct device *dev,
 287                                    struct device_driver *drv)
 288 {
 289         return !strncmp(dev_name(dev), drv->name, strlen(drv->name));
 290 }
 291
 292 static int ntb_transport_bus_probe(struct device *dev)
 293 {
 294         const struct ntb_transport_client *client;
 295         int rc;
 296
 297         get_device(dev);
 298
 299         client = drv_client(dev->driver);
 300         rc = client->probe(dev);
 301         if (rc)
 302                 put_device(dev);
 303
 304         return rc;
 305 }
 306
 307 static int ntb_transport_bus_remove(struct device *dev)
 308 {
 309         const struct ntb_transport_client *client;
 310
 311         client = drv_client(dev->driver);
 312         client->remove(dev);
 313
 314         put_device(dev);
 315
 316         return 0;
 317 }
 318
 319 static struct bus_type ntb_transport_bus = {
 320         .name = "ntb_transport",
 321         .match = ntb_transport_bus_match,
 322         .probe = ntb_transport_bus_probe,
 323         .remove = ntb_transport_bus_remove,
 324 };
 325
 326 static LIST_HEAD(ntb_transport_list);
 327
 328 static int ntb_bus_init(struct ntb_transport_ctx *nt)
 329 {
 330         list_add_tail(&nt->entry, &ntb_transport_list);
 331         return 0;
 332 }
 333
 334 static void ntb_bus_remove(struct ntb_transport_ctx *nt)
 335 {
 336         struct ntb_transport_client_dev *client_dev, *cd;
 337
 338         list_for_each_entry_safe(client_dev, cd, &nt->client_devs, entry) {
 339                 dev_err(client_dev->dev.parent, "%s still attached to bus, removing\n",
 340                         dev_name(&client_dev->dev));
 341                 list_del(&client_dev->entry);
 342                 device_unregister(&client_dev->dev);
 343         }
 344
 345         list_del(&nt->entry);
 346 }
 347
 348 static void ntb_transport_client_release(struct device *dev)
 349 {
 350         struct ntb_transport_client_dev *client_dev;
 351
 352         client_dev = dev_client_dev(dev);
 353         kfree(client_dev);
 354 }
 355
 356 /**
 357  * ntb_transport_unregister_client_dev - Unregister NTB client device
 358  * @device_name: Name of NTB client device
 359  *
 360  * Unregister an NTB client device with the NTB transport layer
 361  */
 362 void ntb_transport_unregister_client_dev(char *device_name)
 363 {
 364         struct ntb_transport_client_dev *client, *cd;
 365         struct ntb_transport_ctx *nt;
 366
 367         list_for_each_entry(nt, &ntb_transport_list, entry)
 368                 list_for_each_entry_safe(client, cd, &nt->client_devs, entry)
 369                         if (!strncmp(dev_name(&client->dev), device_name,
 370                                      strlen(device_name))) {
 371                                 list_del(&client->entry);
 372                                 device_unregister(&client->dev);
 373                         }
 374 }
 375 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client_dev);
 376
 377 /**
 378  * ntb_transport_register_client_dev - Register NTB client device
 379  * @device_name: Name of NTB client device
 380  *
 381  * Register an NTB client device with the NTB transport layer
 382  */
 383 int ntb_transport_register_client_dev(char *device_name)
 384 {
 385         struct ntb_transport_client_dev *client_dev;
 386         struct ntb_transport_ctx *nt;
 387         int node;
 388         int rc, i = 0;
 389
 390         if (list_empty(&ntb_transport_list))
 391                 return -ENODEV;
 392
 393         list_for_each_entry(nt, &ntb_transport_list, entry) {
 394                 struct device *dev;
 395
 396                 node = dev_to_node(&nt->ndev->dev);
 397
 398                 client_dev = kzalloc_node(sizeof(*client_dev),
 399                                           GFP_KERNEL, node);
 400                 if (!client_dev) {
 401                         rc = -ENOMEM;
 402                         goto err;
 403                 }
 404
 405                 dev = &client_dev->dev;
 406
 407                 /* setup and register client devices */
 408                 dev_set_name(dev, "%s%d", device_name, i);
 409                 dev->bus = &ntb_transport_bus;
 410                 dev->release = ntb_transport_client_release;
 411                 dev->parent = &nt->ndev->dev;
 412
 413                 rc = device_register(dev);
 414                 if (rc) {
 415                         put_device(dev);
 416                         goto err;
 417                 }
 418
 419                 list_add_tail(&client_dev->entry, &nt->client_devs);
 420                 i++;
 421         }
 422
 423         return 0;
 424
 425 err:
 426         ntb_transport_unregister_client_dev(device_name);
 427
 428         return rc;
 429 }
 430 EXPORT_SYMBOL_GPL(ntb_transport_register_client_dev);
 431
 432 /**
 433  * ntb_transport_register_client - Register NTB client driver
 434  * @drv: NTB client driver to be registered
 435  *
 436  * Register an NTB client driver with the NTB transport layer
 437  *
 438  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 439  */
 440 int ntb_transport_register_client(struct ntb_transport_client *drv)
 441 {
 442         drv->driver.bus = &ntb_transport_bus;
 443
 444         if (list_empty(&ntb_transport_list))
 445                 return -ENODEV;
 446
 447         return driver_register(&drv->driver);
 448 }
 449 EXPORT_SYMBOL_GPL(ntb_transport_register_client);
 450
 451 /**
 452  * ntb_transport_unregister_client - Unregister NTB client driver
 453  * @drv: NTB client driver to be unregistered
 454  *
 455  * Unregister an NTB client driver with the NTB transport layer
 456  *
 457  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
 458  */
 459 void ntb_transport_unregister_client(struct ntb_transport_client *drv)
 460 {
 461         driver_unregister(&drv->driver);
 462 }
 463 EXPORT_SYMBOL_GPL(ntb_transport_unregister_client);
 464
 465 static ssize_t debugfs_read(struct file *filp, char __user *ubuf, size_t count,
 466                             loff_t *offp)
 467 {
 468         struct ntb_transport_qp *qp;
 469         char *buf;
 470         ssize_t ret, out_offset, out_count;
 471
 472         qp = filp->private_data;
 473
 474         if (!qp || !qp->link_is_up)
 475                 return 0;
 476
 477         out_count = 1000;
 478
 479         buf = kmalloc(out_count, GFP_KERNEL);
 480         if (!buf)
 481                 return -ENOMEM;
 482
 483         out_offset = 0;
 484         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 485                                "\nNTB QP stats:\n\n");
 486         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 487                                "rx_bytes - \t%llu\n", qp->rx_bytes);
 488         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 489                                "rx_pkts - \t%llu\n", qp->rx_pkts);
 490         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 491                                "rx_memcpy - \t%llu\n", qp->rx_memcpy);
 492         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 493                                "rx_async - \t%llu\n", qp->rx_async);
 494         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 495                                "rx_ring_empty - %llu\n", qp->rx_ring_empty);
 496         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 497                                "rx_err_no_buf - %llu\n", qp->rx_err_no_buf);
 498         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 499                                "rx_err_oflow - \t%llu\n", qp->rx_err_oflow);
 500         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 501                                "rx_err_ver - \t%llu\n", qp->rx_err_ver);
 502         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 503                                "rx_buff - \t0x%p\n", qp->rx_buff);
 504         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 505                                "rx_index - \t%u\n", qp->rx_index);
 506         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 507                                "rx_max_entry - \t%u\n", qp->rx_max_entry);
 508         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 509                                "rx_alloc_entry - \t%u\n\n", qp->rx_alloc_entry);
 510
 511         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 512                                "tx_bytes - \t%llu\n", qp->tx_bytes);
 513         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 514                                "tx_pkts - \t%llu\n", qp->tx_pkts);
 515         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 516                                "tx_memcpy - \t%llu\n", qp->tx_memcpy);
 517         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 518                                "tx_async - \t%llu\n", qp->tx_async);
 519         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 520                                "tx_ring_full - \t%llu\n", qp->tx_ring_full);
 521         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 522                                "tx_err_no_buf - %llu\n", qp->tx_err_no_buf);
 523         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 524                                "tx_mw - \t0x%p\n", qp->tx_mw);
 525         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 526                                "tx_index (H) - \t%u\n", qp->tx_index);
 527         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 528                                "RRI (T) - \t%u\n",
 529                                qp->remote_rx_info->entry);
 530         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 531                                "tx_max_entry - \t%u\n", qp->tx_max_entry);
 532         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 533                                "free tx - \t%u\n",
 534                                ntb_transport_tx_free_entry(qp));
 535
 536         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 537                                "\n");
 538         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 539                                "Using TX DMA - \t%s\n",
 540                                qp->tx_dma_chan ? "Yes" : "No");
 541         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 542                                "Using RX DMA - \t%s\n",
 543                                qp->rx_dma_chan ? "Yes" : "No");
 544         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 545                                "QP Link - \t%s\n",
 546                                qp->link_is_up ? "Up" : "Down");
 547         out_offset += scnprintf(buf + out_offset, out_count - out_offset,
 548                                "\n");
 549
 550         if (out_offset > out_count)
 551                 out_offset = out_count;
 552
 553         ret = simple_read_from_buffer(ubuf, count, offp, buf, out_offset);
 554         kfree(buf);
 555         return ret;
 556 }
 557
 558 static const struct file_operations ntb_qp_debugfs_stats = {
 559         .owner = THIS_MODULE,
 560         .open = simple_open,
 561         .read = debugfs_read,
 562 };
 563
 564 static void ntb_list_add(spinlock_t *lock, struct list_head *entry,
 565                          struct list_head *list)
 566 {
 567         unsigned long flags;
 568
 569         spin_lock_irqsave(lock, flags);
 570         list_add_tail(entry, list);
 571         spin_unlock_irqrestore(lock, flags);
 572 }
 573
 574 static struct ntb_queue_entry *ntb_list_rm(spinlock_t *lock,
 575                                            struct list_head *list)
 576 {
 577         struct ntb_queue_entry *entry;
 578         unsigned long flags;
 579
 580         spin_lock_irqsave(lock, flags);
 581         if (list_empty(list)) {
 582                 entry = NULL;
 583                 goto out;
 584         }
 585         entry = list_first_entry(list, struct ntb_queue_entry, entry);
 586         list_del(&entry->entry);
 587
 588 out:
 589         spin_unlock_irqrestore(lock, flags);
 590
 591         return entry;
 592 }
 593
 594 static struct ntb_queue_entry *ntb_list_mv(spinlock_t *lock,
 595                                            struct list_head *list,
 596                                            struct list_head *to_list)
 597 {
 598         struct ntb_queue_entry *entry;
 599         unsigned long flags;
 600
 601         spin_lock_irqsave(lock, flags);
 602
 603         if (list_empty(list)) {
 604                 entry = NULL;
 605         } else {
 606                 entry = list_first_entry(list, struct ntb_queue_entry, entry);
 607                 list_move_tail(&entry->entry, to_list);
 608         }
 609
 610         spin_unlock_irqrestore(lock, flags);
 611
 612         return entry;
 613 }
 614
 615 static int ntb_transport_setup_qp_mw(struct ntb_transport_ctx *nt,
 616                                      unsigned int qp_num)
 617 {
 618         struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
 619         struct ntb_transport_mw *mw;
 620         struct ntb_dev *ndev = nt->ndev;
 621         struct ntb_queue_entry *entry;
 622         unsigned int rx_size, num_qps_mw;
 623         unsigned int mw_num, mw_count, qp_count;
 624         unsigned int i;
 625         int node;
 626
 627         mw_count = nt->mw_count;
 628         qp_count = nt->qp_count;
 629
 630         mw_num = QP_TO_MW(nt, qp_num);
 631         mw = &nt->mw_vec[mw_num];
 632
 633         if (!mw->virt_addr)
 634                 return -ENOMEM;
 635
 636         if (mw_num < qp_count % mw_count)
 637                 num_qps_mw = qp_count / mw_count + 1;
 638         else
 639                 num_qps_mw = qp_count / mw_count;
 640
 641         rx_size = (unsigned int)mw->xlat_size / num_qps_mw;
 642         qp->rx_buff = mw->virt_addr + rx_size * (qp_num / mw_count);
 643         rx_size -= sizeof(struct ntb_rx_info);
 644
 645         qp->remote_rx_info = qp->rx_buff + rx_size;
 646
 647         /* Due to housekeeping, there must be atleast 2 buffs */
 648         qp->rx_max_frame = min(transport_mtu, rx_size / 2);
 649         qp->rx_max_entry = rx_size / qp->rx_max_frame;
 650         qp->rx_index = 0;
 651
 652         /*
 653          * Checking to see if we have more entries than the default.
 654          * We should add additional entries if that is the case so we
 655          * can be in sync with the transport frames.
 656          */
 657         node = dev_to_node(&ndev->dev);
 658         for (i = qp->rx_alloc_entry; i < qp->rx_max_entry; i++) {
 659                 entry = kzalloc_node(sizeof(*entry), GFP_KERNEL, node);
 660                 if (!entry)
 661                         return -ENOMEM;
 662
 663                 entry->qp = qp;
 664                 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
 665                              &qp->rx_free_q);
 666                 qp->rx_alloc_entry++;
 667         }
 668
 669         qp->remote_rx_info->entry = qp->rx_max_entry - 1;
 670
 671         /* setup the hdr offsets with 0's */
 672         for (i = 0; i < qp->rx_max_entry; i++) {
 673                 void *offset = (qp->rx_buff + qp->rx_max_frame * (i + 1) -
 674                                 sizeof(struct ntb_payload_header));
 675                 memset(offset, 0, sizeof(struct ntb_payload_header));
 676         }
 677
 678         qp->rx_pkts = 0;
 679         qp->tx_pkts = 0;
 680         qp->tx_index = 0;
 681
 682         return 0;
 683 }
 684
 685 static irqreturn_t ntb_transport_isr(int irq, void *dev)
 686 {
 687         struct ntb_transport_qp *qp = dev;
 688
 689         tasklet_schedule(&qp->rxc_db_work);
 690
 691         return IRQ_HANDLED;
 692 }
 693
 694 static void ntb_transport_setup_qp_peer_msi(struct ntb_transport_ctx *nt,
 695                                             unsigned int qp_num)
 696 {
 697         struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
 698         int spad = qp_num * 2 + nt->msi_spad_offset;
 699
 700         if (!nt->use_msi)
 701                 return;
 702
 703         if (spad >= ntb_spad_count(nt->ndev))
 704                 return;
 705
 706         qp->peer_msi_desc.addr_offset =
 707                 ntb_peer_spad_read(qp->ndev, PIDX, spad);
 708         qp->peer_msi_desc.data =
 709                 ntb_peer_spad_read(qp->ndev, PIDX, spad + 1);
 710
 711         dev_dbg(&qp->ndev->pdev->dev, "QP%d Peer MSI addr=%x data=%x\n",
 712                 qp_num, qp->peer_msi_desc.addr_offset, qp->peer_msi_desc.data);
 713
 714         if (qp->peer_msi_desc.addr_offset) {
 715                 qp->use_msi = true;
 716                 dev_info(&qp->ndev->pdev->dev,
 717                          "Using MSI interrupts for QP%d\n", qp_num);
 718         }
 719 }
 720
 721 static void ntb_transport_setup_qp_msi(struct ntb_transport_ctx *nt,
 722                                        unsigned int qp_num)
 723 {
 724         struct ntb_transport_qp *qp = &nt->qp_vec[qp_num];
 725         int spad = qp_num * 2 + nt->msi_spad_offset;
 726         int rc;
 727
 728         if (!nt->use_msi)
 729                 return;
 730
 731         if (spad >= ntb_spad_count(nt->ndev)) {
 732                 dev_warn_once(&qp->ndev->pdev->dev,
 733                               "Not enough SPADS to use MSI interrupts\n");
 734                 return;
 735         }
 736
 737         ntb_spad_write(qp->ndev, spad, 0);
 738         ntb_spad_write(qp->ndev, spad + 1, 0);
 739
 740         if (!qp->msi_irq) {
 741                 qp->msi_irq = ntbm_msi_request_irq(qp->ndev, ntb_transport_isr,
 742                                                    KBUILD_MODNAME, qp,
 743                                                    &qp->msi_desc);
 744                 if (qp->msi_irq < 0) {
 745                         dev_warn(&qp->ndev->pdev->dev,
 746                                  "Unable to allocate MSI interrupt for qp%d\n",
 747                                  qp_num);
 748                         return;
 749                 }
 750         }
 751
 752         rc = ntb_spad_write(qp->ndev, spad, qp->msi_desc.addr_offset);
 753         if (rc)
 754                 goto err_free_interrupt;
 755
 756         rc = ntb_spad_write(qp->ndev, spad + 1, qp->msi_desc.data);
 757         if (rc)
 758                 goto err_free_interrupt;
 759
 760         dev_dbg(&qp->ndev->pdev->dev, "QP%d MSI %d addr=%x data=%x\n",
 761                 qp_num, qp->msi_irq, qp->msi_desc.addr_offset,
 762                 qp->msi_desc.data);
 763
 764         return;
 765
 766 err_free_interrupt:
 767         devm_free_irq(&nt->ndev->dev, qp->msi_irq, qp);
 768 }
 769
 770 static void ntb_transport_msi_peer_desc_changed(struct ntb_transport_ctx *nt)
 771 {
 772         int i;
 773
 774         dev_dbg(&nt->ndev->pdev->dev, "Peer MSI descriptors changed");
 775
 776         for (i = 0; i < nt->qp_count; i++)
 777                 ntb_transport_setup_qp_peer_msi(nt, i);
 778 }
 779
 780 static void ntb_transport_msi_desc_changed(void *data)
 781 {
 782         struct ntb_transport_ctx *nt = data;
 783         int i;
 784
 785         dev_dbg(&nt->ndev->pdev->dev, "MSI descriptors changed");
 786
 787         for (i = 0; i < nt->qp_count; i++)
 788                 ntb_transport_setup_qp_msi(nt, i);
 789
 790         ntb_peer_db_set(nt->ndev, nt->msi_db_mask);
 791 }
 792
 793 static void ntb_free_mw(struct ntb_transport_ctx *nt, int num_mw)
 794 {
 795         struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
 796         struct pci_dev *pdev = nt->ndev->pdev;
 797
 798         if (!mw->virt_addr)
 799                 return;
 800
 801         ntb_mw_clear_trans(nt->ndev, PIDX, num_mw);
 802         dma_free_coherent(&pdev->dev, mw->alloc_size,
 803                           mw->alloc_addr, mw->dma_addr);
 804         mw->xlat_size = 0;
 805         mw->buff_size = 0;
 806         mw->alloc_size = 0;
 807         mw->alloc_addr = NULL;
 808         mw->virt_addr = NULL;
 809 }
 810
 811 static int ntb_alloc_mw_buffer(struct ntb_transport_mw *mw,
 812                                struct device *dma_dev, size_t align)
 813 {
 814         dma_addr_t dma_addr;
 815         void *alloc_addr, *virt_addr;
 816         int rc;
 817
 818         alloc_addr = dma_alloc_coherent(dma_dev, mw->alloc_size,
 819                                         &dma_addr, GFP_KERNEL);
 820         if (!alloc_addr) {
 821                 dev_err(dma_dev, "Unable to alloc MW buff of size %zu\n",
 822                         mw->alloc_size);
 823                 return -ENOMEM;
 824         }
 825         virt_addr = alloc_addr;
 826
 827         /*
 828          * we must ensure that the memory address allocated is BAR size
 829          * aligned in order for the XLAT register to take the value. This
 830          * is a requirement of the hardware. It is recommended to setup CMA
 831          * for BAR sizes equal or greater than 4MB.
 832          */
 833         if (!IS_ALIGNED(dma_addr, align)) {
 834                 if (mw->alloc_size > mw->buff_size) {
 835                         virt_addr = PTR_ALIGN(alloc_addr, align);
 836                         dma_addr = ALIGN(dma_addr, align);
 837                 } else {
 838                         rc = -ENOMEM;
 839                         goto err;
 840                 }
 841         }
 842
 843         mw->alloc_addr = alloc_addr;
 844         mw->virt_addr = virt_addr;
 845         mw->dma_addr = dma_addr;
 846
 847         return 0;
 848
 849 err:
 850         dma_free_coherent(dma_dev, mw->alloc_size, alloc_addr, dma_addr);
 851
 852         return rc;
 853 }
 854
 855 static int ntb_set_mw(struct ntb_transport_ctx *nt, int num_mw,
 856                       resource_size_t size)
 857 {
 858         struct ntb_transport_mw *mw = &nt->mw_vec[num_mw];
 859         struct pci_dev *pdev = nt->ndev->pdev;
 860         size_t xlat_size, buff_size;
 861         resource_size_t xlat_align;
 862         resource_size_t xlat_align_size;
 863         int rc;
 864
 865         if (!size)
 866                 return -EINVAL;
 867
 868         rc = ntb_mw_get_align(nt->ndev, PIDX, num_mw, &xlat_align,
 869                               &xlat_align_size, NULL);
 870         if (rc)
 871                 return rc;
 872
 873         xlat_size = round_up(size, xlat_align_size);
 874         buff_size = round_up(size, xlat_align);
 875
 876         /* No need to re-setup */
 877         if (mw->xlat_size == xlat_size)
 878                 return 0;
 879
 880         if (mw->buff_size)
 881                 ntb_free_mw(nt, num_mw);
 882
 883         /* Alloc memory for receiving data.  Must be aligned */
 884         mw->xlat_size = xlat_size;
 885         mw->buff_size = buff_size;
 886         mw->alloc_size = buff_size;
 887
 888         rc = ntb_alloc_mw_buffer(mw, &pdev->dev, xlat_align);
 889         if (rc) {
 890                 mw->alloc_size *= 2;
 891                 rc = ntb_alloc_mw_buffer(mw, &pdev->dev, xlat_align);
 892                 if (rc) {
 893                         dev_err(&pdev->dev,
 894                                 "Unable to alloc aligned MW buff\n");
 895                         mw->xlat_size = 0;
 896                         mw->buff_size = 0;
 897                         mw->alloc_size = 0;
 898                         return rc;
 899                 }
 900         }
 901
 902         /* Notify HW the memory location of the receive buffer */
 903         rc = ntb_mw_set_trans(nt->ndev, PIDX, num_mw, mw->dma_addr,
 904                               mw->xlat_size);
 905         if (rc) {
 906                 dev_err(&pdev->dev, "Unable to set mw%d translation", num_mw);
 907                 ntb_free_mw(nt, num_mw);
 908                 return -EIO;
 909         }
 910
 911         return 0;
 912 }
 913
 914 static void ntb_qp_link_context_reset(struct ntb_transport_qp *qp)
 915 {
 916         qp->link_is_up = false;
 917         qp->active = false;
 918
 919         qp->tx_index = 0;
 920         qp->rx_index = 0;
 921         qp->rx_bytes = 0;
 922         qp->rx_pkts = 0;
 923         qp->rx_ring_empty = 0;
 924         qp->rx_err_no_buf = 0;
 925         qp->rx_err_oflow = 0;
 926         qp->rx_err_ver = 0;
 927         qp->rx_memcpy = 0;
 928         qp->rx_async = 0;
 929         qp->tx_bytes = 0;
 930         qp->tx_pkts = 0;
 931         qp->tx_ring_full = 0;
 932         qp->tx_err_no_buf = 0;
 933         qp->tx_memcpy = 0;
 934         qp->tx_async = 0;
 935 }
 936
 937 static void ntb_qp_link_down_reset(struct ntb_transport_qp *qp)
 938 {
 939         ntb_qp_link_context_reset(qp);
 940         if (qp->remote_rx_info)
 941                 qp->remote_rx_info->entry = qp->rx_max_entry - 1;
 942 }
 943
 944 static void ntb_qp_link_cleanup(struct ntb_transport_qp *qp)
 945 {
 946         struct ntb_transport_ctx *nt = qp->transport;
 947         struct pci_dev *pdev = nt->ndev->pdev;
 948
 949         dev_info(&pdev->dev, "qp %d: Link Cleanup\n", qp->qp_num);
 950
 951         cancel_delayed_work_sync(&qp->link_work);
 952         ntb_qp_link_down_reset(qp);
 953
 954         if (qp->event_handler)
 955                 qp->event_handler(qp->cb_data, qp->link_is_up);
 956 }
 957
 958 static void ntb_qp_link_cleanup_work(struct work_struct *work)
 959 {
 960         struct ntb_transport_qp *qp = container_of(work,
 961                                                    struct ntb_transport_qp,
 962                                                    link_cleanup);
 963         struct ntb_transport_ctx *nt = qp->transport;
 964
 965         ntb_qp_link_cleanup(qp);
 966
 967         if (nt->link_is_up)
 968                 schedule_delayed_work(&qp->link_work,
 969                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
 970 }
 971
 972 static void ntb_qp_link_down(struct ntb_transport_qp *qp)
 973 {
 974         schedule_work(&qp->link_cleanup);
 975 }
 976
 977 static void ntb_transport_link_cleanup(struct ntb_transport_ctx *nt)
 978 {
 979         struct ntb_transport_qp *qp;
 980         u64 qp_bitmap_alloc;
 981         unsigned int i, count;
 982
 983         qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
 984
 985         /* Pass along the info to any clients */
 986         for (i = 0; i < nt->qp_count; i++)
 987                 if (qp_bitmap_alloc & BIT_ULL(i)) {
 988                         qp = &nt->qp_vec[i];
 989                         ntb_qp_link_cleanup(qp);
 990                         cancel_work_sync(&qp->link_cleanup);
 991                         cancel_delayed_work_sync(&qp->link_work);
 992                 }
 993
 994         if (!nt->link_is_up)
 995                 cancel_delayed_work_sync(&nt->link_work);
 996
 997         for (i = 0; i < nt->mw_count; i++)
 998                 ntb_free_mw(nt, i);
 999
1000         /* The scratchpad registers keep the values if the remote side
1001          * goes down, blast them now to give them a sane value the next
1002          * time they are accessed
1003          */
1004         count = ntb_spad_count(nt->ndev);
1005         for (i = 0; i < count; i++)
1006                 ntb_spad_write(nt->ndev, i, 0);
1007 }
1008
1009 static void ntb_transport_link_cleanup_work(struct work_struct *work)
1010 {
1011         struct ntb_transport_ctx *nt =
1012                 container_of(work, struct ntb_transport_ctx, link_cleanup);
1013
1014         ntb_transport_link_cleanup(nt);
1015 }
1016
1017 static void ntb_transport_event_callback(void *data)
1018 {
1019         struct ntb_transport_ctx *nt = data;
1020
1021         if (ntb_link_is_up(nt->ndev, NULL, NULL) == 1)
1022                 schedule_delayed_work(&nt->link_work, 0);
1023         else
1024                 schedule_work(&nt->link_cleanup);
1025 }
1026
1027 static void ntb_transport_link_work(struct work_struct *work)
1028 {
1029         struct ntb_transport_ctx *nt =
1030                 container_of(work, struct ntb_transport_ctx, link_work.work);
1031         struct ntb_dev *ndev = nt->ndev;
1032         struct pci_dev *pdev = ndev->pdev;
1033         resource_size_t size;
1034         u32 val;
1035         int rc = 0, i, spad;
1036
1037         /* send the local info, in the opposite order of the way we read it */
1038
1039         if (nt->use_msi) {
1040                 rc = ntb_msi_setup_mws(ndev);
1041                 if (rc) {
1042                         dev_warn(&pdev->dev,
1043                                  "Failed to register MSI memory window: %d\n",
1044                                  rc);
1045                         nt->use_msi = false;
1046                 }
1047         }
1048
1049         for (i = 0; i < nt->qp_count; i++)
1050                 ntb_transport_setup_qp_msi(nt, i);
1051
1052         for (i = 0; i < nt->mw_count; i++) {
1053                 size = nt->mw_vec[i].phys_size;
1054
1055                 if (max_mw_size && size > max_mw_size)
1056                         size = max_mw_size;
1057
1058                 spad = MW0_SZ_HIGH + (i * 2);
1059                 ntb_peer_spad_write(ndev, PIDX, spad, upper_32_bits(size));
1060
1061                 spad = MW0_SZ_LOW + (i * 2);
1062                 ntb_peer_spad_write(ndev, PIDX, spad, lower_32_bits(size));
1063         }
1064
1065         ntb_peer_spad_write(ndev, PIDX, NUM_MWS, nt->mw_count);
1066
1067         ntb_peer_spad_write(ndev, PIDX, NUM_QPS, nt->qp_count);
1068
1069         ntb_peer_spad_write(ndev, PIDX, VERSION, NTB_TRANSPORT_VERSION);
1070
1071         /* Query the remote side for its info */
1072         val = ntb_spad_read(ndev, VERSION);
1073         dev_dbg(&pdev->dev, "Remote version = %d\n", val);
1074         if (val != NTB_TRANSPORT_VERSION)
1075                 goto out;
1076
1077         val = ntb_spad_read(ndev, NUM_QPS);
1078         dev_dbg(&pdev->dev, "Remote max number of qps = %d\n", val);
1079         if (val != nt->qp_count)
1080                 goto out;
1081
1082         val = ntb_spad_read(ndev, NUM_MWS);
1083         dev_dbg(&pdev->dev, "Remote number of mws = %d\n", val);
1084         if (val != nt->mw_count)
1085                 goto out;
1086
1087         for (i = 0; i < nt->mw_count; i++) {
1088                 u64 val64;
1089
1090                 val = ntb_spad_read(ndev, MW0_SZ_HIGH + (i * 2));
1091                 val64 = (u64)val << 32;
1092
1093                 val = ntb_spad_read(ndev, MW0_SZ_LOW + (i * 2));
1094                 val64 |= val;
1095
1096                 dev_dbg(&pdev->dev, "Remote MW%d size = %#llx\n", i, val64);
1097
1098                 rc = ntb_set_mw(nt, i, val64);
1099                 if (rc)
1100                         goto out1;
1101         }
1102
1103         nt->link_is_up = true;
1104
1105         for (i = 0; i < nt->qp_count; i++) {
1106                 struct ntb_transport_qp *qp = &nt->qp_vec[i];
1107
1108                 ntb_transport_setup_qp_mw(nt, i);
1109                 ntb_transport_setup_qp_peer_msi(nt, i);
1110
1111                 if (qp->client_ready)
1112                         schedule_delayed_work(&qp->link_work, 0);
1113         }
1114
1115         return;
1116
1117 out1:
1118         for (i = 0; i < nt->mw_count; i++)
1119                 ntb_free_mw(nt, i);
1120
1121         /* if there's an actual failure, we should just bail */
1122         if (rc < 0)
1123                 return;
1124
1125 out:
1126         if (ntb_link_is_up(ndev, NULL, NULL) == 1)
1127                 schedule_delayed_work(&nt->link_work,
1128                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
1129 }
1130
1131 static void ntb_qp_link_work(struct work_struct *work)
1132 {
1133         struct ntb_transport_qp *qp = container_of(work,
1134                                                    struct ntb_transport_qp,
1135                                                    link_work.work);
1136         struct pci_dev *pdev = qp->ndev->pdev;
1137         struct ntb_transport_ctx *nt = qp->transport;
1138         int val;
1139
1140         WARN_ON(!nt->link_is_up);
1141
1142         val = ntb_spad_read(nt->ndev, QP_LINKS);
1143
1144         ntb_peer_spad_write(nt->ndev, PIDX, QP_LINKS, val | BIT(qp->qp_num));
1145
1146         /* query remote spad for qp ready bits */
1147         dev_dbg_ratelimited(&pdev->dev, "Remote QP link status = %x\n", val);
1148
1149         /* See if the remote side is up */
1150         if (val & BIT(qp->qp_num)) {
1151                 dev_info(&pdev->dev, "qp %d: Link Up\n", qp->qp_num);
1152                 qp->link_is_up = true;
1153                 qp->active = true;
1154
1155                 if (qp->event_handler)
1156                         qp->event_handler(qp->cb_data, qp->link_is_up);
1157
1158                 if (qp->active)
1159                         tasklet_schedule(&qp->rxc_db_work);
1160         } else if (nt->link_is_up)
1161                 schedule_delayed_work(&qp->link_work,
1162                                       msecs_to_jiffies(NTB_LINK_DOWN_TIMEOUT));
1163 }
1164
1165 static int ntb_transport_init_queue(struct ntb_transport_ctx *nt,
1166                                     unsigned int qp_num)
1167 {
1168         struct ntb_transport_qp *qp;
1169         phys_addr_t mw_base;
1170         resource_size_t mw_size;
1171         unsigned int num_qps_mw, tx_size;
1172         unsigned int mw_num, mw_count, qp_count;
1173         u64 qp_offset;
1174
1175         mw_count = nt->mw_count;
1176         qp_count = nt->qp_count;
1177
1178         mw_num = QP_TO_MW(nt, qp_num);
1179
1180         qp = &nt->qp_vec[qp_num];
1181         qp->qp_num = qp_num;
1182         qp->transport = nt;
1183         qp->ndev = nt->ndev;
1184         qp->client_ready = false;
1185         qp->event_handler = NULL;
1186         ntb_qp_link_context_reset(qp);
1187
1188         if (mw_num < qp_count % mw_count)
1189                 num_qps_mw = qp_count / mw_count + 1;
1190         else
1191                 num_qps_mw = qp_count / mw_count;
1192
1193         mw_base = nt->mw_vec[mw_num].phys_addr;
1194         mw_size = nt->mw_vec[mw_num].phys_size;
1195
1196         if (max_mw_size && mw_size > max_mw_size)
1197                 mw_size = max_mw_size;
1198
1199         tx_size = (unsigned int)mw_size / num_qps_mw;
1200         qp_offset = tx_size * (qp_num / mw_count);
1201
1202         qp->tx_mw_size = tx_size;
1203         qp->tx_mw = nt->mw_vec[mw_num].vbase + qp_offset;
1204         if (!qp->tx_mw)
1205                 return -EINVAL;
1206
1207         qp->tx_mw_phys = mw_base + qp_offset;
1208         if (!qp->tx_mw_phys)
1209                 return -EINVAL;
1210
1211         tx_size -= sizeof(struct ntb_rx_info);
1212         qp->rx_info = qp->tx_mw + tx_size;
1213
1214         /* Due to housekeeping, there must be atleast 2 buffs */
1215         qp->tx_max_frame = min(transport_mtu, tx_size / 2);
1216         qp->tx_max_entry = tx_size / qp->tx_max_frame;
1217
1218         if (nt->debugfs_node_dir) {
1219                 char debugfs_name[4];
1220
1221                 snprintf(debugfs_name, 4, "qp%d", qp_num);
1222                 qp->debugfs_dir = debugfs_create_dir(debugfs_name,
1223                                                      nt->debugfs_node_dir);
1224
1225                 qp->debugfs_stats = debugfs_create_file("stats", S_IRUSR,
1226                                                         qp->debugfs_dir, qp,
1227                                                         &ntb_qp_debugfs_stats);
1228         } else {
1229                 qp->debugfs_dir = NULL;
1230                 qp->debugfs_stats = NULL;
1231         }
1232
1233         INIT_DELAYED_WORK(&qp->link_work, ntb_qp_link_work);
1234         INIT_WORK(&qp->link_cleanup, ntb_qp_link_cleanup_work);
1235
1236         spin_lock_init(&qp->ntb_rx_q_lock);
1237         spin_lock_init(&qp->ntb_tx_free_q_lock);
1238
1239         INIT_LIST_HEAD(&qp->rx_post_q);
1240         INIT_LIST_HEAD(&qp->rx_pend_q);
1241         INIT_LIST_HEAD(&qp->rx_free_q);
1242         INIT_LIST_HEAD(&qp->tx_free_q);
1243
1244         tasklet_init(&qp->rxc_db_work, ntb_transport_rxc_db,
1245                      (unsigned long)qp);
1246
1247         return 0;
1248 }
1249
1250 static int ntb_transport_probe(struct ntb_client *self, struct ntb_dev *ndev)
1251 {
1252         struct ntb_transport_ctx *nt;
1253         struct ntb_transport_mw *mw;
1254         unsigned int mw_count, qp_count, spad_count, max_mw_count_for_spads;
1255         u64 qp_bitmap;
1256         int node;
1257         int rc, i;
1258
1259         mw_count = ntb_peer_mw_count(ndev);
1260
1261         if (!ndev->ops->mw_set_trans) {
1262                 dev_err(&ndev->dev, "Inbound MW based NTB API is required\n");
1263                 return -EINVAL;
1264         }
1265
1266         if (ntb_db_is_unsafe(ndev))
1267                 dev_dbg(&ndev->dev,
1268                         "doorbell is unsafe, proceed anyway...\n");
1269         if (ntb_spad_is_unsafe(ndev))
1270                 dev_dbg(&ndev->dev,
1271                         "scratchpad is unsafe, proceed anyway...\n");
1272
1273         if (ntb_peer_port_count(ndev) != NTB_DEF_PEER_CNT)
1274                 dev_warn(&ndev->dev, "Multi-port NTB devices unsupported\n");
1275
1276         node = dev_to_node(&ndev->dev);
1277
1278         nt = kzalloc_node(sizeof(*nt), GFP_KERNEL, node);
1279         if (!nt)
1280                 return -ENOMEM;
1281
1282         nt->ndev = ndev;
1283
1284         /*
1285          * If we are using MSI, and have at least one extra memory window,
1286          * we will reserve the last MW for the MSI window.
1287          */
1288         if (use_msi && mw_count > 1) {
1289                 rc = ntb_msi_init(ndev, ntb_transport_msi_desc_changed);
1290                 if (!rc) {
1291                         mw_count -= 1;
1292                         nt->use_msi = true;
1293                 }
1294         }
1295
1296         spad_count = ntb_spad_count(ndev);
1297
1298         /* Limit the MW's based on the availability of scratchpads */
1299
1300         if (spad_count < NTB_TRANSPORT_MIN_SPADS) {
1301                 nt->mw_count = 0;
1302                 rc = -EINVAL;
1303                 goto err;
1304         }
1305
1306         max_mw_count_for_spads = (spad_count - MW0_SZ_HIGH) / 2;
1307         nt->mw_count = min(mw_count, max_mw_count_for_spads);
1308
1309         nt->msi_spad_offset = nt->mw_count * 2 + MW0_SZ_HIGH;
1310
1311         nt->mw_vec = kcalloc_node(mw_count, sizeof(*nt->mw_vec),
1312                                   GFP_KERNEL, node);
1313         if (!nt->mw_vec) {
1314                 rc = -ENOMEM;
1315                 goto err;
1316         }
1317
1318         for (i = 0; i < mw_count; i++) {
1319                 mw = &nt->mw_vec[i];
1320
1321                 rc = ntb_peer_mw_get_addr(ndev, i, &mw->phys_addr,
1322                                           &mw->phys_size);
1323                 if (rc)
1324                         goto err1;
1325
1326                 mw->vbase = ioremap_wc(mw->phys_addr, mw->phys_size);
1327                 if (!mw->vbase) {
1328                         rc = -ENOMEM;
1329                         goto err1;
1330                 }
1331
1332                 mw->buff_size = 0;
1333                 mw->xlat_size = 0;
1334                 mw->virt_addr = NULL;
1335                 mw->dma_addr = 0;
1336         }
1337
1338         qp_bitmap = ntb_db_valid_mask(ndev);
1339
1340         qp_count = ilog2(qp_bitmap);
1341         if (nt->use_msi) {
1342                 qp_count -= 1;
1343                 nt->msi_db_mask = 1 << qp_count;
1344                 ntb_db_clear_mask(ndev, nt->msi_db_mask);
1345         }
1346
1347         if (max_num_clients && max_num_clients < qp_count)
1348                 qp_count = max_num_clients;
1349         else if (nt->mw_count < qp_count)
1350                 qp_count = nt->mw_count;
1351
1352         qp_bitmap &= BIT_ULL(qp_count) - 1;
1353
1354         nt->qp_count = qp_count;
1355         nt->qp_bitmap = qp_bitmap;
1356         nt->qp_bitmap_free = qp_bitmap;
1357
1358         nt->qp_vec = kcalloc_node(qp_count, sizeof(*nt->qp_vec),
1359                                   GFP_KERNEL, node);
1360         if (!nt->qp_vec) {
1361                 rc = -ENOMEM;
1362                 goto err1;
1363         }
1364
1365         if (nt_debugfs_dir) {
1366                 nt->debugfs_node_dir =
1367                         debugfs_create_dir(pci_name(ndev->pdev),
1368                                            nt_debugfs_dir);
1369         }
1370
1371         for (i = 0; i < qp_count; i++) {
1372                 rc = ntb_transport_init_queue(nt, i);
1373                 if (rc)
1374                         goto err2;
1375         }
1376
1377         INIT_DELAYED_WORK(&nt->link_work, ntb_transport_link_work);
1378         INIT_WORK(&nt->link_cleanup, ntb_transport_link_cleanup_work);
1379
1380         rc = ntb_set_ctx(ndev, nt, &ntb_transport_ops);
1381         if (rc)
1382                 goto err2;
1383
1384         INIT_LIST_HEAD(&nt->client_devs);
1385         rc = ntb_bus_init(nt);
1386         if (rc)
1387                 goto err3;
1388
1389         nt->link_is_up = false;
1390         ntb_link_enable(ndev, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
1391         ntb_link_event(ndev);
1392
1393         return 0;
1394
1395 err3:
1396         ntb_clear_ctx(ndev);
1397 err2:
1398         kfree(nt->qp_vec);
1399 err1:
1400         while (i--) {
1401                 mw = &nt->mw_vec[i];
1402                 iounmap(mw->vbase);
1403         }
1404         kfree(nt->mw_vec);
1405 err:
1406         kfree(nt);
1407         return rc;
1408 }
1409
1410 static void ntb_transport_free(struct ntb_client *self, struct ntb_dev *ndev)
1411 {
1412         struct ntb_transport_ctx *nt = ndev->ctx;
1413         struct ntb_transport_qp *qp;
1414         u64 qp_bitmap_alloc;
1415         int i;
1416
1417         ntb_transport_link_cleanup(nt);
1418         cancel_work_sync(&nt->link_cleanup);
1419         cancel_delayed_work_sync(&nt->link_work);
1420
1421         qp_bitmap_alloc = nt->qp_bitmap & ~nt->qp_bitmap_free;
1422
1423         /* verify that all the qp's are freed */
1424         for (i = 0; i < nt->qp_count; i++) {
1425                 qp = &nt->qp_vec[i];
1426                 if (qp_bitmap_alloc & BIT_ULL(i))
1427                         ntb_transport_free_queue(qp);
1428                 debugfs_remove_recursive(qp->debugfs_dir);
1429         }
1430
1431         ntb_link_disable(ndev);
1432         ntb_clear_ctx(ndev);
1433
1434         ntb_bus_remove(nt);
1435
1436         for (i = nt->mw_count; i--; ) {
1437                 ntb_free_mw(nt, i);
1438                 iounmap(nt->mw_vec[i].vbase);
1439         }
1440
1441         kfree(nt->qp_vec);
1442         kfree(nt->mw_vec);
1443         kfree(nt);
1444 }
1445
1446 static void ntb_complete_rxc(struct ntb_transport_qp *qp)
1447 {
1448         struct ntb_queue_entry *entry;
1449         void *cb_data;
1450         unsigned int len;
1451         unsigned long irqflags;
1452
1453         spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1454
1455         while (!list_empty(&qp->rx_post_q)) {
1456                 entry = list_first_entry(&qp->rx_post_q,
1457                                          struct ntb_queue_entry, entry);
1458                 if (!(entry->flags & DESC_DONE_FLAG))
1459                         break;
1460
1461                 entry->rx_hdr->flags = 0;
1462                 iowrite32(entry->rx_index, &qp->rx_info->entry);
1463
1464                 cb_data = entry->cb_data;
1465                 len = entry->len;
1466
1467                 list_move_tail(&entry->entry, &qp->rx_free_q);
1468
1469                 spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1470
1471                 if (qp->rx_handler && qp->client_ready)
1472                         qp->rx_handler(qp, qp->cb_data, cb_data, len);
1473
1474                 spin_lock_irqsave(&qp->ntb_rx_q_lock, irqflags);
1475         }
1476
1477         spin_unlock_irqrestore(&qp->ntb_rx_q_lock, irqflags);
1478 }
1479
1480 static void ntb_rx_copy_callback(void *data,
1481                                  const struct dmaengine_result *res)
1482 {
1483         struct ntb_queue_entry *entry = data;
1484
1485         /* we need to check DMA results if we are using DMA */
1486         if (res) {
1487                 enum dmaengine_tx_result dma_err = res->result;
1488
1489                 switch (dma_err) {
1490                 case DMA_TRANS_READ_FAILED:
1491                 case DMA_TRANS_WRITE_FAILED:
1492                         entry->errors++;
1493                         fallthrough;
1494                 case DMA_TRANS_ABORTED:
1495                 {
1496                         struct ntb_transport_qp *qp = entry->qp;
1497                         void *offset = qp->rx_buff + qp->rx_max_frame *
1498                                         qp->rx_index;
1499
1500                         ntb_memcpy_rx(entry, offset);
1501                         qp->rx_memcpy++;
1502                         return;
1503                 }
1504
1505                 case DMA_TRANS_NOERROR:
1506                 default:
1507                         break;
1508                 }
1509         }
1510
1511         entry->flags |= DESC_DONE_FLAG;
1512
1513         ntb_complete_rxc(entry->qp);
1514 }
1515
1516 static void ntb_memcpy_rx(struct ntb_queue_entry *entry, void *offset)
1517 {
1518         void *buf = entry->buf;
1519         size_t len = entry->len;
1520
1521         memcpy(buf, offset, len);
1522
1523         /* Ensure that the data is fully copied out before clearing the flag */
1524         wmb();
1525
1526         ntb_rx_copy_callback(entry, NULL);
1527 }
1528
1529 static int ntb_async_rx_submit(struct ntb_queue_entry *entry, void *offset)
1530 {
1531         struct dma_async_tx_descriptor *txd;
1532         struct ntb_transport_qp *qp = entry->qp;
1533         struct dma_chan *chan = qp->rx_dma_chan;
1534         struct dma_device *device;
1535         size_t pay_off, buff_off, len;
1536         struct dmaengine_unmap_data *unmap;
1537         dma_cookie_t cookie;
1538         void *buf = entry->buf;
1539
1540         len = entry->len;
1541         device = chan->device;
1542         pay_off = (size_t)offset & ~PAGE_MASK;
1543         buff_off = (size_t)buf & ~PAGE_MASK;
1544
1545         if (!is_dma_copy_aligned(device, pay_off, buff_off, len))
1546                 goto err;
1547
1548         unmap = dmaengine_get_unmap_data(device->dev, 2, GFP_NOWAIT);
1549         if (!unmap)
1550                 goto err;
1551
1552         unmap->len = len;
1553         unmap->addr[0] = dma_map_page(device->dev, virt_to_page(offset),
1554                                       pay_off, len, DMA_TO_DEVICE);
1555         if (dma_mapping_error(device->dev, unmap->addr[0]))
1556                 goto err_get_unmap;
1557
1558         unmap->to_cnt = 1;
1559
1560         unmap->addr[1] = dma_map_page(device->dev, virt_to_page(buf),
1561                                       buff_off, len, DMA_FROM_DEVICE);
1562         if (dma_mapping_error(device->dev, unmap->addr[1]))
1563                 goto err_get_unmap;
1564
1565         unmap->from_cnt = 1;
1566
1567         txd = device->device_prep_dma_memcpy(chan, unmap->addr[1],
1568                                              unmap->addr[0], len,
1569                                              DMA_PREP_INTERRUPT);
1570         if (!txd)
1571                 goto err_get_unmap;
1572
1573         txd->callback_result = ntb_rx_copy_callback;
1574         txd->callback_param = entry;
1575         dma_set_unmap(txd, unmap);
1576
1577         cookie = dmaengine_submit(txd);
1578         if (dma_submit_error(cookie))
1579                 goto err_set_unmap;
1580
1581         dmaengine_unmap_put(unmap);
1582
1583         qp->last_cookie = cookie;
1584
1585         qp->rx_async++;
1586
1587         return 0;
1588
1589 err_set_unmap:
1590         dmaengine_unmap_put(unmap);
1591 err_get_unmap:
1592         dmaengine_unmap_put(unmap);
1593 err:
1594         return -ENXIO;
1595 }
1596
1597 static void ntb_async_rx(struct ntb_queue_entry *entry, void *offset)
1598 {
1599         struct ntb_transport_qp *qp = entry->qp;
1600         struct dma_chan *chan = qp->rx_dma_chan;
1601         int res;
1602
1603         if (!chan)
1604                 goto err;
1605
1606         if (entry->len < copy_bytes)
1607                 goto err;
1608
1609         res = ntb_async_rx_submit(entry, offset);
1610         if (res < 0)
1611                 goto err;
1612
1613         if (!entry->retries)
1614                 qp->rx_async++;
1615
1616         return;
1617
1618 err:
1619         ntb_memcpy_rx(entry, offset);
1620         qp->rx_memcpy++;
1621 }
1622
1623 static int ntb_process_rxc(struct ntb_transport_qp *qp)
1624 {
1625         struct ntb_payload_header *hdr;
1626         struct ntb_queue_entry *entry;
1627         void *offset;
1628
1629         offset = qp->rx_buff + qp->rx_max_frame * qp->rx_index;
1630         hdr = offset + qp->rx_max_frame - sizeof(struct ntb_payload_header);
1631
1632         dev_dbg(&qp->ndev->pdev->dev, "qp %d: RX ver %u len %d flags %x\n",
1633                 qp->qp_num, hdr->ver, hdr->len, hdr->flags);
1634
1635         if (!(hdr->flags & DESC_DONE_FLAG)) {
1636                 dev_dbg(&qp->ndev->pdev->dev, "done flag not set\n");
1637                 qp->rx_ring_empty++;
1638                 return -EAGAIN;
1639         }
1640
1641         if (hdr->flags & LINK_DOWN_FLAG) {
1642                 dev_dbg(&qp->ndev->pdev->dev, "link down flag set\n");
1643                 ntb_qp_link_down(qp);
1644                 hdr->flags = 0;
1645                 return -EAGAIN;
1646         }
1647
1648         if (hdr->ver != (u32)qp->rx_pkts) {
1649                 dev_dbg(&qp->ndev->pdev->dev,
1650                         "version mismatch, expected %llu - got %u\n",
1651                         qp->rx_pkts, hdr->ver);
1652                 qp->rx_err_ver++;
1653                 return -EIO;
1654         }
1655
1656         entry = ntb_list_mv(&qp->ntb_rx_q_lock, &qp->rx_pend_q, &qp->rx_post_q);
1657         if (!entry) {
1658                 dev_dbg(&qp->ndev->pdev->dev, "no receive buffer\n");
1659                 qp->rx_err_no_buf++;
1660                 return -EAGAIN;
1661         }
1662
1663         entry->rx_hdr = hdr;
1664         entry->rx_index = qp->rx_index;
1665
1666         if (hdr->len > entry->len) {
1667                 dev_dbg(&qp->ndev->pdev->dev,
1668                         "receive buffer overflow! Wanted %d got %d\n",
1669                         hdr->len, entry->len);
1670                 qp->rx_err_oflow++;
1671
1672                 entry->len = -EIO;
1673                 entry->flags |= DESC_DONE_FLAG;
1674
1675                 ntb_complete_rxc(qp);
1676         } else {
1677                 dev_dbg(&qp->ndev->pdev->dev,
1678                         "RX OK index %u ver %u size %d into buf size %d\n",
1679                         qp->rx_index, hdr->ver, hdr->len, entry->len);
1680
1681                 qp->rx_bytes += hdr->len;
1682                 qp->rx_pkts++;
1683
1684                 entry->len = hdr->len;
1685
1686                 ntb_async_rx(entry, offset);
1687         }
1688
1689         qp->rx_index++;
1690         qp->rx_index %= qp->rx_max_entry;
1691
1692         return 0;
1693 }
1694
1695 static void ntb_transport_rxc_db(unsigned long data)
1696 {
1697         struct ntb_transport_qp *qp = (void *)data;
1698         int rc, i;
1699
1700         dev_dbg(&qp->ndev->pdev->dev, "%s: doorbell %d received\n",
1701                 __func__, qp->qp_num);
1702
1703         /* Limit the number of packets processed in a single interrupt to
1704          * provide fairness to others
1705          */
1706         for (i = 0; i < qp->rx_max_entry; i++) {
1707                 rc = ntb_process_rxc(qp);
1708                 if (rc)
1709                         break;
1710         }
1711
1712         if (i && qp->rx_dma_chan)
1713                 dma_async_issue_pending(qp->rx_dma_chan);
1714
1715         if (i == qp->rx_max_entry) {
1716                 /* there is more work to do */
1717                 if (qp->active)
1718                         tasklet_schedule(&qp->rxc_db_work);
1719         } else if (ntb_db_read(qp->ndev) & BIT_ULL(qp->qp_num)) {
1720                 /* the doorbell bit is set: clear it */
1721                 ntb_db_clear(qp->ndev, BIT_ULL(qp->qp_num));
1722                 /* ntb_db_read ensures ntb_db_clear write is committed */
1723                 ntb_db_read(qp->ndev);
1724
1725                 /* an interrupt may have arrived between finishing
1726                  * ntb_process_rxc and clearing the doorbell bit:
1727                  * there might be some more work to do.
1728                  */
1729                 if (qp->active)
1730                         tasklet_schedule(&qp->rxc_db_work);
1731         }
1732 }
1733
1734 static void ntb_tx_copy_callback(void *data,
1735                                  const struct dmaengine_result *res)
1736 {
1737         struct ntb_queue_entry *entry = data;
1738         struct ntb_transport_qp *qp = entry->qp;
1739         struct ntb_payload_header __iomem *hdr = entry->tx_hdr;
1740
1741         /* we need to check DMA results if we are using DMA */
1742         if (res) {
1743                 enum dmaengine_tx_result dma_err = res->result;
1744
1745                 switch (dma_err) {
1746                 case DMA_TRANS_READ_FAILED:
1747                 case DMA_TRANS_WRITE_FAILED:
1748                         entry->errors++;
1749                         fallthrough;
1750                 case DMA_TRANS_ABORTED:
1751                 {
1752                         void __iomem *offset =
1753                                 qp->tx_mw + qp->tx_max_frame *
1754                                 entry->tx_index;
1755
1756                         /* resubmit via CPU */
1757                         ntb_memcpy_tx(entry, offset);
1758                         qp->tx_memcpy++;
1759                         return;
1760                 }
1761
1762                 case DMA_TRANS_NOERROR:
1763                 default:
1764                         break;
1765                 }
1766         }
1767
1768         iowrite32(entry->flags | DESC_DONE_FLAG, &hdr->flags);
1769
1770         if (qp->use_msi)
1771                 ntb_msi_peer_trigger(qp->ndev, PIDX, &qp->peer_msi_desc);
1772         else
1773                 ntb_peer_db_set(qp->ndev, BIT_ULL(qp->qp_num));
1774
1775         /* The entry length can only be zero if the packet is intended to be a
1776          * "link down" or similar.  Since no payload is being sent in these
1777          * cases, there is nothing to add to the completion queue.
1778          */
1779         if (entry->len > 0) {
1780                 qp->tx_bytes += entry->len;
1781
1782                 if (qp->tx_handler)
1783                         qp->tx_handler(qp, qp->cb_data, entry->cb_data,
1784                                        entry->len);
1785         }
1786
1787         ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry, &qp->tx_free_q);
1788 }
1789
1790 static void ntb_memcpy_tx(struct ntb_queue_entry *entry, void __iomem *offset)
1791 {
1792 #ifdef ARCH_HAS_NOCACHE_UACCESS
1793         /*
1794          * Using non-temporal mov to improve performance on non-cached
1795          * writes, even though we aren't actually copying from user space.
1796          */
1797         __copy_from_user_inatomic_nocache(offset, entry->buf, entry->len);
1798 #else
1799         memcpy_toio(offset, entry->buf, entry->len);
1800 #endif
1801
1802         /* Ensure that the data is fully copied out before setting the flags */
1803         wmb();
1804
1805         ntb_tx_copy_callback(entry, NULL);
1806 }
1807
1808 static int ntb_async_tx_submit(struct ntb_transport_qp *qp,
1809                                struct ntb_queue_entry *entry)
1810 {
1811         struct dma_async_tx_descriptor *txd;
1812         struct dma_chan *chan = qp->tx_dma_chan;
1813         struct dma_device *device;
1814         size_t len = entry->len;
1815         void *buf = entry->buf;
1816         size_t dest_off, buff_off;
1817         struct dmaengine_unmap_data *unmap;
1818         dma_addr_t dest;
1819         dma_cookie_t cookie;
1820
1821         device = chan->device;
1822         dest = qp->tx_mw_dma_addr + qp->tx_max_frame * entry->tx_index;
1823         buff_off = (size_t)buf & ~PAGE_MASK;
1824         dest_off = (size_t)dest & ~PAGE_MASK;
1825
1826         if (!is_dma_copy_aligned(device, buff_off, dest_off, len))
1827                 goto err;
1828
1829         unmap = dmaengine_get_unmap_data(device->dev, 1, GFP_NOWAIT);
1830         if (!unmap)
1831                 goto err;
1832
1833         unmap->len = len;
1834         unmap->addr[0] = dma_map_page(device->dev, virt_to_page(buf),
1835                                       buff_off, len, DMA_TO_DEVICE);
1836         if (dma_mapping_error(device->dev, unmap->addr[0]))
1837                 goto err_get_unmap;
1838
1839         unmap->to_cnt = 1;
1840
1841         txd = device->device_prep_dma_memcpy(chan, dest, unmap->addr[0], len,
1842                                              DMA_PREP_INTERRUPT);
1843         if (!txd)
1844                 goto err_get_unmap;
1845
1846         txd->callback_result = ntb_tx_copy_callback;
1847         txd->callback_param = entry;
1848         dma_set_unmap(txd, unmap);
1849
1850         cookie = dmaengine_submit(txd);
1851         if (dma_submit_error(cookie))
1852                 goto err_set_unmap;
1853
1854         dmaengine_unmap_put(unmap);
1855
1856         dma_async_issue_pending(chan);
1857
1858         return 0;
1859 err_set_unmap:
1860         dmaengine_unmap_put(unmap);
1861 err_get_unmap:
1862         dmaengine_unmap_put(unmap);
1863 err:
1864         return -ENXIO;
1865 }
1866
1867 static void ntb_async_tx(struct ntb_transport_qp *qp,
1868                          struct ntb_queue_entry *entry)
1869 {
1870         struct ntb_payload_header __iomem *hdr;
1871         struct dma_chan *chan = qp->tx_dma_chan;
1872         void __iomem *offset;
1873         int res;
1874
1875         entry->tx_index = qp->tx_index;
1876         offset = qp->tx_mw + qp->tx_max_frame * entry->tx_index;
1877         hdr = offset + qp->tx_max_frame - sizeof(struct ntb_payload_header);
1878         entry->tx_hdr = hdr;
1879
1880         iowrite32(entry->len, &hdr->len);
1881         iowrite32((u32)qp->tx_pkts, &hdr->ver);
1882
1883         if (!chan)
1884                 goto err;
1885
1886         if (entry->len < copy_bytes)
1887                 goto err;
1888
1889         res = ntb_async_tx_submit(qp, entry);
1890         if (res < 0)
1891                 goto err;
1892
1893         if (!entry->retries)
1894                 qp->tx_async++;
1895
1896         return;
1897
1898 err:
1899         ntb_memcpy_tx(entry, offset);
1900         qp->tx_memcpy++;
1901 }
1902
1903 static int ntb_process_tx(struct ntb_transport_qp *qp,
1904                           struct ntb_queue_entry *entry)
1905 {
1906         if (qp->tx_index == qp->remote_rx_info->entry) {
1907                 qp->tx_ring_full++;
1908                 return -EAGAIN;
1909         }
1910
1911         if (entry->len > qp->tx_max_frame - sizeof(struct ntb_payload_header)) {
1912                 if (qp->tx_handler)
1913                         qp->tx_handler(qp, qp->cb_data, NULL, -EIO);
1914
1915                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
1916                              &qp->tx_free_q);
1917                 return 0;
1918         }
1919
1920         ntb_async_tx(qp, entry);
1921
1922         qp->tx_index++;
1923         qp->tx_index %= qp->tx_max_entry;
1924
1925         qp->tx_pkts++;
1926
1927         return 0;
1928 }
1929
1930 static void ntb_send_link_down(struct ntb_transport_qp *qp)
1931 {
1932         struct pci_dev *pdev = qp->ndev->pdev;
1933         struct ntb_queue_entry *entry;
1934         int i, rc;
1935
1936         if (!qp->link_is_up)
1937                 return;
1938
1939         dev_info(&pdev->dev, "qp %d: Send Link Down\n", qp->qp_num);
1940
1941         for (i = 0; i < NTB_LINK_DOWN_TIMEOUT; i++) {
1942                 entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
1943                 if (entry)
1944                         break;
1945                 msleep(100);
1946         }
1947
1948         if (!entry)
1949                 return;
1950
1951         entry->cb_data = NULL;
1952         entry->buf = NULL;
1953         entry->len = 0;
1954         entry->flags = LINK_DOWN_FLAG;
1955
1956         rc = ntb_process_tx(qp, entry);
1957         if (rc)
1958                 dev_err(&pdev->dev, "ntb: QP%d unable to send linkdown msg\n",
1959                         qp->qp_num);
1960
1961         ntb_qp_link_down_reset(qp);
1962 }
1963
1964 static bool ntb_dma_filter_fn(struct dma_chan *chan, void *node)
1965 {
1966         return dev_to_node(&chan->dev->device) == (int)(unsigned long)node;
1967 }
1968
1969 /**
1970  * ntb_transport_create_queue - Create a new NTB transport layer queue
1971  * @rx_handler: receive callback function
1972  * @tx_handler: transmit callback function
1973  * @event_handler: event callback function
1974  *
1975  * Create a new NTB transport layer queue and provide the queue with a callback
1976  * routine for both transmit and receive.  The receive callback routine will be
1977  * used to pass up data when the transport has received it on the queue.   The
1978  * transmit callback routine will be called when the transport has completed the
1979  * transmission of the data on the queue and the data is ready to be freed.
1980  *
1981  * RETURNS: pointer to newly created ntb_queue, NULL on error.
1982  */
1983 struct ntb_transport_qp *
1984 ntb_transport_create_queue(void *data, struct device *client_dev,
1985                            const struct ntb_queue_handlers *handlers)
1986 {
1987         struct ntb_dev *ndev;
1988         struct pci_dev *pdev;
1989         struct ntb_transport_ctx *nt;
1990         struct ntb_queue_entry *entry;
1991         struct ntb_transport_qp *qp;
1992         u64 qp_bit;
1993         unsigned int free_queue;
1994         dma_cap_mask_t dma_mask;
1995         int node;
1996         int i;
1997
1998         ndev = dev_ntb(client_dev->parent);
1999         pdev = ndev->pdev;
2000         nt = ndev->ctx;
2001
2002         node = dev_to_node(&ndev->dev);
2003
2004         free_queue = ffs(nt->qp_bitmap_free);
2005         if (!free_queue)
2006                 goto err;
2007
2008         /* decrement free_queue to make it zero based */
2009         free_queue--;
2010
2011         qp = &nt->qp_vec[free_queue];
2012         qp_bit = BIT_ULL(qp->qp_num);
2013
2014         nt->qp_bitmap_free &= ~qp_bit;
2015
2016         qp->cb_data = data;
2017         qp->rx_handler = handlers->rx_handler;
2018         qp->tx_handler = handlers->tx_handler;
2019         qp->event_handler = handlers->event_handler;
2020
2021         dma_cap_zero(dma_mask);
2022         dma_cap_set(DMA_MEMCPY, dma_mask);
2023
2024         if (use_dma) {
2025                 qp->tx_dma_chan =
2026                         dma_request_channel(dma_mask, ntb_dma_filter_fn,
2027                                             (void *)(unsigned long)node);
2028                 if (!qp->tx_dma_chan)
2029                         dev_info(&pdev->dev, "Unable to allocate TX DMA channel\n");
2030
2031                 qp->rx_dma_chan =
2032                         dma_request_channel(dma_mask, ntb_dma_filter_fn,
2033                                             (void *)(unsigned long)node);
2034                 if (!qp->rx_dma_chan)
2035                         dev_info(&pdev->dev, "Unable to allocate RX DMA channel\n");
2036         } else {
2037                 qp->tx_dma_chan = NULL;
2038                 qp->rx_dma_chan = NULL;
2039         }
2040
2041         qp->tx_mw_dma_addr = 0;
2042         if (qp->tx_dma_chan) {
2043                 qp->tx_mw_dma_addr =
2044                         dma_map_resource(qp->tx_dma_chan->device->dev,
2045                                          qp->tx_mw_phys, qp->tx_mw_size,
2046                                          DMA_FROM_DEVICE, 0);
2047                 if (dma_mapping_error(qp->tx_dma_chan->device->dev,
2048                                       qp->tx_mw_dma_addr)) {
2049                         qp->tx_mw_dma_addr = 0;
2050                         goto err1;
2051                 }
2052         }
2053
2054         dev_dbg(&pdev->dev, "Using %s memcpy for TX\n",
2055                 qp->tx_dma_chan ? "DMA" : "CPU");
2056
2057         dev_dbg(&pdev->dev, "Using %s memcpy for RX\n",
2058                 qp->rx_dma_chan ? "DMA" : "CPU");
2059
2060         for (i = 0; i < NTB_QP_DEF_NUM_ENTRIES; i++) {
2061                 entry = kzalloc_node(sizeof(*entry), GFP_KERNEL, node);
2062                 if (!entry)
2063                         goto err1;
2064
2065                 entry->qp = qp;
2066                 ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry,
2067                              &qp->rx_free_q);
2068         }
2069         qp->rx_alloc_entry = NTB_QP_DEF_NUM_ENTRIES;
2070
2071         for (i = 0; i < qp->tx_max_entry; i++) {
2072                 entry = kzalloc_node(sizeof(*entry), GFP_KERNEL, node);
2073                 if (!entry)
2074                         goto err2;
2075
2076                 entry->qp = qp;
2077                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
2078                              &qp->tx_free_q);
2079         }
2080
2081         ntb_db_clear(qp->ndev, qp_bit);
2082         ntb_db_clear_mask(qp->ndev, qp_bit);
2083
2084         dev_info(&pdev->dev, "NTB Transport QP %d created\n", qp->qp_num);
2085
2086         return qp;
2087
2088 err2:
2089         while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
2090                 kfree(entry);
2091 err1:
2092         qp->rx_alloc_entry = 0;
2093         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
2094                 kfree(entry);
2095         if (qp->tx_mw_dma_addr)
2096                 dma_unmap_resource(qp->tx_dma_chan->device->dev,
2097                                    qp->tx_mw_dma_addr, qp->tx_mw_size,
2098                                    DMA_FROM_DEVICE, 0);
2099         if (qp->tx_dma_chan)
2100                 dma_release_channel(qp->tx_dma_chan);
2101         if (qp->rx_dma_chan)
2102                 dma_release_channel(qp->rx_dma_chan);
2103         nt->qp_bitmap_free |= qp_bit;
2104 err:
2105         return NULL;
2106 }
2107 EXPORT_SYMBOL_GPL(ntb_transport_create_queue);
2108
2109 /**
2110  * ntb_transport_free_queue - Frees NTB transport queue
2111  * @qp: NTB queue to be freed
2112  *
2113  * Frees NTB transport queue
2114  */
2115 void ntb_transport_free_queue(struct ntb_transport_qp *qp)
2116 {
2117         struct pci_dev *pdev;
2118         struct ntb_queue_entry *entry;
2119         u64 qp_bit;
2120
2121         if (!qp)
2122                 return;
2123
2124         pdev = qp->ndev->pdev;
2125
2126         qp->active = false;
2127
2128         if (qp->tx_dma_chan) {
2129                 struct dma_chan *chan = qp->tx_dma_chan;
2130                 /* Putting the dma_chan to NULL will force any new traffic to be
2131                  * processed by the CPU instead of the DAM engine
2132                  */
2133                 qp->tx_dma_chan = NULL;
2134
2135                 /* Try to be nice and wait for any queued DMA engine
2136                  * transactions to process before smashing it with a rock
2137                  */
2138                 dma_sync_wait(chan, qp->last_cookie);
2139                 dmaengine_terminate_all(chan);
2140
2141                 dma_unmap_resource(chan->device->dev,
2142                                    qp->tx_mw_dma_addr, qp->tx_mw_size,
2143                                    DMA_FROM_DEVICE, 0);
2144
2145                 dma_release_channel(chan);
2146         }
2147
2148         if (qp->rx_dma_chan) {
2149                 struct dma_chan *chan = qp->rx_dma_chan;
2150                 /* Putting the dma_chan to NULL will force any new traffic to be
2151                  * processed by the CPU instead of the DAM engine
2152                  */
2153                 qp->rx_dma_chan = NULL;
2154
2155                 /* Try to be nice and wait for any queued DMA engine
2156                  * transactions to process before smashing it with a rock
2157                  */
2158                 dma_sync_wait(chan, qp->last_cookie);
2159                 dmaengine_terminate_all(chan);
2160                 dma_release_channel(chan);
2161         }
2162
2163         qp_bit = BIT_ULL(qp->qp_num);
2164
2165         ntb_db_set_mask(qp->ndev, qp_bit);
2166         tasklet_kill(&qp->rxc_db_work);
2167
2168         cancel_delayed_work_sync(&qp->link_work);
2169
2170         qp->cb_data = NULL;
2171         qp->rx_handler = NULL;
2172         qp->tx_handler = NULL;
2173         qp->event_handler = NULL;
2174
2175         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q)))
2176                 kfree(entry);
2177
2178         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q))) {
2179                 dev_warn(&pdev->dev, "Freeing item from non-empty rx_pend_q\n");
2180                 kfree(entry);
2181         }
2182
2183         while ((entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_post_q))) {
2184                 dev_warn(&pdev->dev, "Freeing item from non-empty rx_post_q\n");
2185                 kfree(entry);
2186         }
2187
2188         while ((entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q)))
2189                 kfree(entry);
2190
2191         qp->transport->qp_bitmap_free |= qp_bit;
2192
2193         dev_info(&pdev->dev, "NTB Transport QP %d freed\n", qp->qp_num);
2194 }
2195 EXPORT_SYMBOL_GPL(ntb_transport_free_queue);
2196
2197 /**
2198  * ntb_transport_rx_remove - Dequeues enqueued rx packet
2199  * @qp: NTB queue to be freed
2200  * @len: pointer to variable to write enqueued buffers length
2201  *
2202  * Dequeues unused buffers from receive queue.  Should only be used during
2203  * shutdown of qp.
2204  *
2205  * RETURNS: NULL error value on error, or void* for success.
2206  */
2207 void *ntb_transport_rx_remove(struct ntb_transport_qp *qp, unsigned int *len)
2208 {
2209         struct ntb_queue_entry *entry;
2210         void *buf;
2211
2212         if (!qp || qp->client_ready)
2213                 return NULL;
2214
2215         entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_pend_q);
2216         if (!entry)
2217                 return NULL;
2218
2219         buf = entry->cb_data;
2220         *len = entry->len;
2221
2222         ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_free_q);
2223
2224         return buf;
2225 }
2226 EXPORT_SYMBOL_GPL(ntb_transport_rx_remove);
2227
2228 /**
2229  * ntb_transport_rx_enqueue - Enqueue a new NTB queue entry
2230  * @qp: NTB transport layer queue the entry is to be enqueued on
2231  * @cb: per buffer pointer for callback function to use
2232  * @data: pointer to data buffer that incoming packets will be copied into
2233  * @len: length of the data buffer
2234  *
2235  * Enqueue a new receive buffer onto the transport queue into which a NTB
2236  * payload can be received into.
2237  *
2238  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2239  */
2240 int ntb_transport_rx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
2241                              unsigned int len)
2242 {
2243         struct ntb_queue_entry *entry;
2244
2245         if (!qp)
2246                 return -EINVAL;
2247
2248         entry = ntb_list_rm(&qp->ntb_rx_q_lock, &qp->rx_free_q);
2249         if (!entry)
2250                 return -ENOMEM;
2251
2252         entry->cb_data = cb;
2253         entry->buf = data;
2254         entry->len = len;
2255         entry->flags = 0;
2256         entry->retries = 0;
2257         entry->errors = 0;
2258         entry->rx_index = 0;
2259
2260         ntb_list_add(&qp->ntb_rx_q_lock, &entry->entry, &qp->rx_pend_q);
2261
2262         if (qp->active)
2263                 tasklet_schedule(&qp->rxc_db_work);
2264
2265         return 0;
2266 }
2267 EXPORT_SYMBOL_GPL(ntb_transport_rx_enqueue);
2268
2269 /**
2270  * ntb_transport_tx_enqueue - Enqueue a new NTB queue entry
2271  * @qp: NTB transport layer queue the entry is to be enqueued on
2272  * @cb: per buffer pointer for callback function to use
2273  * @data: pointer to data buffer that will be sent
2274  * @len: length of the data buffer
2275  *
2276  * Enqueue a new transmit buffer onto the transport queue from which a NTB
2277  * payload will be transmitted.  This assumes that a lock is being held to
2278  * serialize access to the qp.
2279  *
2280  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
2281  */
2282 int ntb_transport_tx_enqueue(struct ntb_transport_qp *qp, void *cb, void *data,
2283                              unsigned int len)
2284 {
2285         struct ntb_queue_entry *entry;
2286         int rc;
2287
2288         if (!qp || !len)
2289                 return -EINVAL;
2290
2291         /* If the qp link is down already, just ignore. */
2292         if (!qp->link_is_up)
2293                 return 0;
2294
2295         entry = ntb_list_rm(&qp->ntb_tx_free_q_lock, &qp->tx_free_q);
2296         if (!entry) {
2297                 qp->tx_err_no_buf++;
2298                 return -EBUSY;
2299         }
2300
2301         entry->cb_data = cb;
2302         entry->buf = data;
2303         entry->len = len;
2304         entry->flags = 0;
2305         entry->errors = 0;
2306         entry->retries = 0;
2307         entry->tx_index = 0;
2308
2309         rc = ntb_process_tx(qp, entry);
2310         if (rc)
2311                 ntb_list_add(&qp->ntb_tx_free_q_lock, &entry->entry,
2312                              &qp->tx_free_q);
2313
2314         return rc;
2315 }
2316 EXPORT_SYMBOL_GPL(ntb_transport_tx_enqueue);
2317
2318 /**
2319  * ntb_transport_link_up - Notify NTB transport of client readiness to use queue
2320  * @qp: NTB transport layer queue to be enabled
2321  *
2322  * Notify NTB transport layer of client readiness to use queue
2323  */
2324 void ntb_transport_link_up(struct ntb_transport_qp *qp)
2325 {
2326         if (!qp)
2327                 return;
2328
2329         qp->client_ready = true;
2330
2331         if (qp->transport->link_is_up)
2332                 schedule_delayed_work(&qp->link_work, 0);
2333 }
2334 EXPORT_SYMBOL_GPL(ntb_transport_link_up);
2335
2336 /**
2337  * ntb_transport_link_down - Notify NTB transport to no longer enqueue data
2338  * @qp: NTB transport layer queue to be disabled
2339  *
2340  * Notify NTB transport layer of client's desire to no longer receive data on
2341  * transport queue specified.  It is the client's responsibility to ensure all
2342  * entries on queue are purged or otherwise handled appropriately.
2343  */
2344 void ntb_transport_link_down(struct ntb_transport_qp *qp)
2345 {
2346         int val;
2347
2348         if (!qp)
2349                 return;
2350
2351         qp->client_ready = false;
2352
2353         val = ntb_spad_read(qp->ndev, QP_LINKS);
2354
2355         ntb_peer_spad_write(qp->ndev, PIDX, QP_LINKS, val & ~BIT(qp->qp_num));
2356
2357         if (qp->link_is_up)
2358                 ntb_send_link_down(qp);
2359         else
2360                 cancel_delayed_work_sync(&qp->link_work);
2361 }
2362 EXPORT_SYMBOL_GPL(ntb_transport_link_down);
2363
2364 /**
2365  * ntb_transport_link_query - Query transport link state
2366  * @qp: NTB transport layer queue to be queried
2367  *
2368  * Query connectivity to the remote system of the NTB transport queue
2369  *
2370  * RETURNS: true for link up or false for link down
2371  */
2372 bool ntb_transport_link_query(struct ntb_transport_qp *qp)
2373 {
2374         if (!qp)
2375                 return false;
2376
2377         return qp->link_is_up;
2378 }
2379 EXPORT_SYMBOL_GPL(ntb_transport_link_query);
2380
2381 /**
2382  * ntb_transport_qp_num - Query the qp number
2383  * @qp: NTB transport layer queue to be queried
2384  *
2385  * Query qp number of the NTB transport queue
2386  *
2387  * RETURNS: a zero based number specifying the qp number
2388  */
2389 unsigned char ntb_transport_qp_num(struct ntb_transport_qp *qp)
2390 {
2391         if (!qp)
2392                 return 0;
2393
2394         return qp->qp_num;
2395 }
2396 EXPORT_SYMBOL_GPL(ntb_transport_qp_num);
2397
2398 /**
2399  * ntb_transport_max_size - Query the max payload size of a qp
2400  * @qp: NTB transport layer queue to be queried
2401  *
2402  * Query the maximum payload size permissible on the given qp
2403  *
2404  * RETURNS: the max payload size of a qp
2405  */
2406 unsigned int ntb_transport_max_size(struct ntb_transport_qp *qp)
2407 {
2408         unsigned int max_size;
2409         unsigned int copy_align;
2410         struct dma_chan *rx_chan, *tx_chan;
2411
2412         if (!qp)
2413                 return 0;
2414
2415         rx_chan = qp->rx_dma_chan;
2416         tx_chan = qp->tx_dma_chan;
2417
2418         copy_align = max(rx_chan ? rx_chan->device->copy_align : 0,
2419                          tx_chan ? tx_chan->device->copy_align : 0);
2420
2421         /* If DMA engine usage is possible, try to find the max size for that */
2422         max_size = qp->tx_max_frame - sizeof(struct ntb_payload_header);
2423         max_size = round_down(max_size, 1 << copy_align);
2424
2425         return max_size;
2426 }
2427 EXPORT_SYMBOL_GPL(ntb_transport_max_size);
2428
2429 unsigned int ntb_transport_tx_free_entry(struct ntb_transport_qp *qp)
2430 {
2431         unsigned int head = qp->tx_index;
2432         unsigned int tail = qp->remote_rx_info->entry;
2433
2434         return tail >= head ? tail - head : qp->tx_max_entry + tail - head;
2435 }
2436 EXPORT_SYMBOL_GPL(ntb_transport_tx_free_entry);
2437
2438 static void ntb_transport_doorbell_callback(void *data, int vector)
2439 {
2440         struct ntb_transport_ctx *nt = data;
2441         struct ntb_transport_qp *qp;
2442         u64 db_bits;
2443         unsigned int qp_num;
2444
2445         if (ntb_db_read(nt->ndev) & nt->msi_db_mask) {
2446                 ntb_transport_msi_peer_desc_changed(nt);
2447                 ntb_db_clear(nt->ndev, nt->msi_db_mask);
2448         }
2449
2450         db_bits = (nt->qp_bitmap & ~nt->qp_bitmap_free &
2451                    ntb_db_vector_mask(nt->ndev, vector));
2452
2453         while (db_bits) {
2454                 qp_num = __ffs(db_bits);
2455                 qp = &nt->qp_vec[qp_num];
2456
2457                 if (qp->active)
2458                         tasklet_schedule(&qp->rxc_db_work);
2459
2460                 db_bits &= ~BIT_ULL(qp_num);
2461         }
2462 }
2463
2464 static const struct ntb_ctx_ops ntb_transport_ops = {
2465         .link_event = ntb_transport_event_callback,
2466         .db_event = ntb_transport_doorbell_callback,
2467 };
2468
2469 static struct ntb_client ntb_transport_client = {
2470         .ops = {
2471                 .probe = ntb_transport_probe,
2472                 .remove = ntb_transport_free,
2473         },
2474 };
2475
2476 static int __init ntb_transport_init(void)
2477 {
2478         int rc;
2479
2480         pr_info("%s, version %s\n", NTB_TRANSPORT_DESC, NTB_TRANSPORT_VER);
2481
2482         if (debugfs_initialized())
2483                 nt_debugfs_dir = debugfs_create_dir(KBUILD_MODNAME, NULL);
2484
2485         rc = bus_register(&ntb_transport_bus);
2486         if (rc)
2487                 goto err_bus;
2488
2489         rc = ntb_register_client(&ntb_transport_client);
2490         if (rc)
2491                 goto err_client;
2492
2493         return 0;
2494
2495 err_client:
2496         bus_unregister(&ntb_transport_bus);
2497 err_bus:
2498         debugfs_remove_recursive(nt_debugfs_dir);
2499         return rc;
2500 }
2501 module_init(ntb_transport_init);
2502
2503 static void __exit ntb_transport_exit(void)
2504 {
2505         ntb_unregister_client(&ntb_transport_client);
2506         bus_unregister(&ntb_transport_bus);
2507         debugfs_remove_recursive(nt_debugfs_dir);
2508 }
2509 module_exit(ntb_transport_exit);