drivers/dma/dmaengine.c

   1 /*
   2  * Copyright(c) 2004 - 2006 Intel Corporation. All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or modify it
   5  * under the terms of the GNU General Public License as published by the Free
   6  * Software Foundation; either version 2 of the License, or (at your option)
   7  * any later version.
   8  *
   9  * This program is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12  * more details.
  13  *
  14  * You should have received a copy of the GNU General Public License along with
  15  * this program; if not, write to the Free Software Foundation, Inc., 59
  16  * Temple Place - Suite 330, Boston, MA  02111-1307, USA.
  17  *
  18  * The full GNU General Public License is included in this distribution in the
  19  * file called COPYING.
  20  */
  21
  22 /*
  23  * This code implements the DMA subsystem. It provides a HW-neutral interface
  24  * for other kernel code to use asynchronous memory copy capabilities,
  25  * if present, and allows different HW DMA drivers to register as providing
  26  * this capability.
  27  *
  28  * Due to the fact we are accelerating what is already a relatively fast
  29  * operation, the code goes to great lengths to avoid additional overhead,
  30  * such as locking.
  31  *
  32  * LOCKING:
  33  *
  34  * The subsystem keeps two global lists, dma_device_list and dma_client_list.
  35  * Both of these are protected by a mutex, dma_list_mutex.
  36  *
  37  * Each device has a channels list, which runs unlocked but is never modified
  38  * once the device is registered, it's just setup by the driver.
  39  *
  40  * Each client is responsible for keeping track of the channels it uses.  See
  41  * the definition of dma_event_callback in dmaengine.h.
  42  *
  43  * Each device has a kref, which is initialized to 1 when the device is
  44  * registered. A kref_get is done for each device registered.  When the
  45  * device is released, the coresponding kref_put is done in the release
  46  * method. Every time one of the device's channels is allocated to a client,
  47  * a kref_get occurs.  When the channel is freed, the coresponding kref_put
  48  * happens. The device's release function does a completion, so
  49  * unregister_device does a remove event, device_unregister, a kref_put
  50  * for the first reference, then waits on the completion for all other
  51  * references to finish.
  52  *
  53  * Each channel has an open-coded implementation of Rusty Russell's "bigref,"
  54  * with a kref and a per_cpu local_t.  A dma_chan_get is called when a client
  55  * signals that it wants to use a channel, and dma_chan_put is called when
  56  * a channel is removed or a client using it is unregesitered.  A client can
  57  * take extra references per outstanding transaction, as is the case with
  58  * the NET DMA client.  The release function does a kref_put on the device.
  59  *      -ChrisL, DanW
  60  */
  61
  62 #include <linux/init.h>
  63 #include <linux/module.h>
  64 #include <linux/mm.h>
  65 #include <linux/device.h>
  66 #include <linux/dmaengine.h>
  67 #include <linux/hardirq.h>
  68 #include <linux/spinlock.h>
  69 #include <linux/percpu.h>
  70 #include <linux/rcupdate.h>
  71 #include <linux/mutex.h>
  72 #include <linux/jiffies.h>
  73
  74 static DEFINE_MUTEX(dma_list_mutex);
  75 static LIST_HEAD(dma_device_list);
  76 static LIST_HEAD(dma_client_list);
  77
  78 /* --- sysfs implementation --- */
  79
  80 static ssize_t show_memcpy_count(struct device *dev, struct device_attribute *attr, char *buf)
  81 {
  82         struct dma_chan *chan = to_dma_chan(dev);
  83         unsigned long count = 0;
  84         int i;
  85
  86         for_each_possible_cpu(i)
  87                 count += per_cpu_ptr(chan->local, i)->memcpy_count;
  88
  89         return sprintf(buf, "%lu\n", count);
  90 }
  91
  92 static ssize_t show_bytes_transferred(struct device *dev, struct device_attribute *attr,
  93                                       char *buf)
  94 {
  95         struct dma_chan *chan = to_dma_chan(dev);
  96         unsigned long count = 0;
  97         int i;
  98
  99         for_each_possible_cpu(i)
 100                 count += per_cpu_ptr(chan->local, i)->bytes_transferred;
 101
 102         return sprintf(buf, "%lu\n", count);
 103 }
 104
 105 static ssize_t show_in_use(struct device *dev, struct device_attribute *attr, char *buf)
 106 {
 107         struct dma_chan *chan = to_dma_chan(dev);
 108         int in_use = 0;
 109
 110         if (unlikely(chan->slow_ref) &&
 111                 atomic_read(&chan->refcount.refcount) > 1)
 112                 in_use = 1;
 113         else {
 114                 if (local_read(&(per_cpu_ptr(chan->local,
 115                         get_cpu())->refcount)) > 0)
 116                         in_use = 1;
 117                 put_cpu();
 118         }
 119
 120         return sprintf(buf, "%d\n", in_use);
 121 }
 122
 123 static struct device_attribute dma_attrs[] = {
 124         __ATTR(memcpy_count, S_IRUGO, show_memcpy_count, NULL),
 125         __ATTR(bytes_transferred, S_IRUGO, show_bytes_transferred, NULL),
 126         __ATTR(in_use, S_IRUGO, show_in_use, NULL),
 127         __ATTR_NULL
 128 };
 129
 130 static void dma_async_device_cleanup(struct kref *kref);
 131
 132 static void dma_dev_release(struct device *dev)
 133 {
 134         struct dma_chan *chan = to_dma_chan(dev);
 135         kref_put(&chan->device->refcount, dma_async_device_cleanup);
 136 }
 137
 138 static struct class dma_devclass = {
 139         .name           = "dma",
 140         .dev_attrs      = dma_attrs,
 141         .dev_release    = dma_dev_release,
 142 };
 143
 144 /* --- client and device registration --- */
 145
 146 #define dma_chan_satisfies_mask(chan, mask) \
 147         __dma_chan_satisfies_mask((chan), &(mask))
 148 static int
 149 __dma_chan_satisfies_mask(struct dma_chan *chan, dma_cap_mask_t *want)
 150 {
 151         dma_cap_mask_t has;
 152
 153         bitmap_and(has.bits, want->bits, chan->device->cap_mask.bits,
 154                 DMA_TX_TYPE_END);
 155         return bitmap_equal(want->bits, has.bits, DMA_TX_TYPE_END);
 156 }
 157
 158 /**
 159  * dma_client_chan_alloc - try to allocate channels to a client
 160  * @client: &dma_client
 161  *
 162  * Called with dma_list_mutex held.
 163  */
 164 static void dma_client_chan_alloc(struct dma_client *client)
 165 {
 166         struct dma_device *device;
 167         struct dma_chan *chan;
 168         int desc;       /* allocated descriptor count */
 169         enum dma_state_client ack;
 170
 171         /* Find a channel */
 172         list_for_each_entry(device, &dma_device_list, global_node)
 173                 list_for_each_entry(chan, &device->channels, device_node) {
 174                         if (!dma_chan_satisfies_mask(chan, client->cap_mask))
 175                                 continue;
 176
 177                         desc = chan->device->device_alloc_chan_resources(chan);
 178                         if (desc >= 0) {
 179                                 ack = client->event_callback(client,
 180                                                 chan,
 181                                                 DMA_RESOURCE_AVAILABLE);
 182
 183                                 /* we are done once this client rejects
 184                                  * an available resource
 185                                  */
 186                                 if (ack == DMA_ACK)
 187                                         dma_chan_get(chan);
 188                                 else if (ack == DMA_NAK)
 189                                         return;
 190                         }
 191                 }
 192 }
 193
 194 enum dma_status dma_sync_wait(struct dma_chan *chan, dma_cookie_t cookie)
 195 {
 196         enum dma_status status;
 197         unsigned long dma_sync_wait_timeout = jiffies + msecs_to_jiffies(5000);
 198
 199         dma_async_issue_pending(chan);
 200         do {
 201                 status = dma_async_is_tx_complete(chan, cookie, NULL, NULL);
 202                 if (time_after_eq(jiffies, dma_sync_wait_timeout)) {
 203                         printk(KERN_ERR "dma_sync_wait_timeout!\n");
 204                         return DMA_ERROR;
 205                 }
 206         } while (status == DMA_IN_PROGRESS);
 207
 208         return status;
 209 }
 210 EXPORT_SYMBOL(dma_sync_wait);
 211
 212 /**
 213  * dma_chan_cleanup - release a DMA channel's resources
 214  * @kref: kernel reference structure that contains the DMA channel device
 215  */
 216 void dma_chan_cleanup(struct kref *kref)
 217 {
 218         struct dma_chan *chan = container_of(kref, struct dma_chan, refcount);
 219         chan->device->device_free_chan_resources(chan);
 220         kref_put(&chan->device->refcount, dma_async_device_cleanup);
 221 }
 222 EXPORT_SYMBOL(dma_chan_cleanup);
 223
 224 static void dma_chan_free_rcu(struct rcu_head *rcu)
 225 {
 226         struct dma_chan *chan = container_of(rcu, struct dma_chan, rcu);
 227         int bias = 0x7FFFFFFF;
 228         int i;
 229         for_each_possible_cpu(i)
 230                 bias -= local_read(&per_cpu_ptr(chan->local, i)->refcount);
 231         atomic_sub(bias, &chan->refcount.refcount);
 232         kref_put(&chan->refcount, dma_chan_cleanup);
 233 }
 234
 235 static void dma_chan_release(struct dma_chan *chan)
 236 {
 237         atomic_add(0x7FFFFFFF, &chan->refcount.refcount);
 238         chan->slow_ref = 1;
 239         call_rcu(&chan->rcu, dma_chan_free_rcu);
 240 }
 241
 242 /**
 243  * dma_chans_notify_available - broadcast available channels to the clients
 244  */
 245 static void dma_clients_notify_available(void)
 246 {
 247         struct dma_client *client;
 248
 249         mutex_lock(&dma_list_mutex);
 250
 251         list_for_each_entry(client, &dma_client_list, global_node)
 252                 dma_client_chan_alloc(client);
 253
 254         mutex_unlock(&dma_list_mutex);
 255 }
 256
 257 /**
 258  * dma_chans_notify_available - tell the clients that a channel is going away
 259  * @chan: channel on its way out
 260  */
 261 static void dma_clients_notify_removed(struct dma_chan *chan)
 262 {
 263         struct dma_client *client;
 264         enum dma_state_client ack;
 265
 266         mutex_lock(&dma_list_mutex);
 267
 268         list_for_each_entry(client, &dma_client_list, global_node) {
 269                 ack = client->event_callback(client, chan,
 270                                 DMA_RESOURCE_REMOVED);
 271
 272                 /* client was holding resources for this channel so
 273                  * free it
 274                  */
 275                 if (ack == DMA_ACK)
 276                         dma_chan_put(chan);
 277         }
 278
 279         mutex_unlock(&dma_list_mutex);
 280 }
 281
 282 /**
 283  * dma_async_client_register - register a &dma_client
 284  * @client: ptr to a client structure with valid 'event_callback' and 'cap_mask'
 285  */
 286 void dma_async_client_register(struct dma_client *client)
 287 {
 288         mutex_lock(&dma_list_mutex);
 289         list_add_tail(&client->global_node, &dma_client_list);
 290         mutex_unlock(&dma_list_mutex);
 291 }
 292 EXPORT_SYMBOL(dma_async_client_register);
 293
 294 /**
 295  * dma_async_client_unregister - unregister a client and free the &dma_client
 296  * @client: &dma_client to free
 297  *
 298  * Force frees any allocated DMA channels, frees the &dma_client memory
 299  */
 300 void dma_async_client_unregister(struct dma_client *client)
 301 {
 302         struct dma_device *device;
 303         struct dma_chan *chan;
 304         enum dma_state_client ack;
 305
 306         if (!client)
 307                 return;
 308
 309         mutex_lock(&dma_list_mutex);
 310         /* free all channels the client is holding */
 311         list_for_each_entry(device, &dma_device_list, global_node)
 312                 list_for_each_entry(chan, &device->channels, device_node) {
 313                         ack = client->event_callback(client, chan,
 314                                 DMA_RESOURCE_REMOVED);
 315
 316                         if (ack == DMA_ACK)
 317                                 dma_chan_put(chan);
 318                 }
 319
 320         list_del(&client->global_node);
 321         mutex_unlock(&dma_list_mutex);
 322 }
 323 EXPORT_SYMBOL(dma_async_client_unregister);
 324
 325 /**
 326  * dma_async_client_chan_request - send all available channels to the
 327  * client that satisfy the capability mask
 328  * @client - requester
 329  */
 330 void dma_async_client_chan_request(struct dma_client *client)
 331 {
 332         mutex_lock(&dma_list_mutex);
 333         dma_client_chan_alloc(client);
 334         mutex_unlock(&dma_list_mutex);
 335 }
 336 EXPORT_SYMBOL(dma_async_client_chan_request);
 337
 338 /**
 339  * dma_async_device_register - registers DMA devices found
 340  * @device: &dma_device
 341  */
 342 int dma_async_device_register(struct dma_device *device)
 343 {
 344         static int id;
 345         int chancnt = 0, rc;
 346         struct dma_chan* chan;
 347
 348         if (!device)
 349                 return -ENODEV;
 350
 351         /* validate device routines */
 352         BUG_ON(dma_has_cap(DMA_MEMCPY, device->cap_mask) &&
 353                 !device->device_prep_dma_memcpy);
 354         BUG_ON(dma_has_cap(DMA_XOR, device->cap_mask) &&
 355                 !device->device_prep_dma_xor);
 356         BUG_ON(dma_has_cap(DMA_ZERO_SUM, device->cap_mask) &&
 357                 !device->device_prep_dma_zero_sum);
 358         BUG_ON(dma_has_cap(DMA_MEMSET, device->cap_mask) &&
 359                 !device->device_prep_dma_memset);
 360         BUG_ON(dma_has_cap(DMA_ZERO_SUM, device->cap_mask) &&
 361                 !device->device_prep_dma_interrupt);
 362
 363         BUG_ON(!device->device_alloc_chan_resources);
 364         BUG_ON(!device->device_free_chan_resources);
 365         BUG_ON(!device->device_dependency_added);
 366         BUG_ON(!device->device_is_tx_complete);
 367         BUG_ON(!device->device_issue_pending);
 368         BUG_ON(!device->dev);
 369
 370         init_completion(&device->done);
 371         kref_init(&device->refcount);
 372         device->dev_id = id++;
 373
 374         /* represent channels in sysfs. Probably want devs too */
 375         list_for_each_entry(chan, &device->channels, device_node) {
 376                 chan->local = alloc_percpu(typeof(*chan->local));
 377                 if (chan->local == NULL)
 378                         continue;
 379
 380                 chan->chan_id = chancnt++;
 381                 chan->dev.class = &dma_devclass;
 382                 chan->dev.parent = NULL;
 383                 snprintf(chan->dev.bus_id, BUS_ID_SIZE, "dma%dchan%d",
 384                          device->dev_id, chan->chan_id);
 385
 386                 rc = device_register(&chan->dev);
 387                 if (rc) {
 388                         chancnt--;
 389                         free_percpu(chan->local);
 390                         chan->local = NULL;
 391                         goto err_out;
 392                 }
 393
 394                 /* One for the channel, one of the class device */
 395                 kref_get(&device->refcount);
 396                 kref_get(&device->refcount);
 397                 kref_init(&chan->refcount);
 398                 chan->slow_ref = 0;
 399                 INIT_RCU_HEAD(&chan->rcu);
 400         }
 401
 402         mutex_lock(&dma_list_mutex);
 403         list_add_tail(&device->global_node, &dma_device_list);
 404         mutex_unlock(&dma_list_mutex);
 405
 406         dma_clients_notify_available();
 407
 408         return 0;
 409
 410 err_out:
 411         list_for_each_entry(chan, &device->channels, device_node) {
 412                 if (chan->local == NULL)
 413                         continue;
 414                 kref_put(&device->refcount, dma_async_device_cleanup);
 415                 device_unregister(&chan->dev);
 416                 chancnt--;
 417                 free_percpu(chan->local);
 418         }
 419         return rc;
 420 }
 421 EXPORT_SYMBOL(dma_async_device_register);
 422
 423 /**
 424  * dma_async_device_cleanup - function called when all references are released
 425  * @kref: kernel reference object
 426  */
 427 static void dma_async_device_cleanup(struct kref *kref)
 428 {
 429         struct dma_device *device;
 430
 431         device = container_of(kref, struct dma_device, refcount);
 432         complete(&device->done);
 433 }
 434
 435 /**
 436  * dma_async_device_unregister - unregisters DMA devices
 437  * @device: &dma_device
 438  */
 439 void dma_async_device_unregister(struct dma_device *device)
 440 {
 441         struct dma_chan *chan;
 442
 443         mutex_lock(&dma_list_mutex);
 444         list_del(&device->global_node);
 445         mutex_unlock(&dma_list_mutex);
 446
 447         list_for_each_entry(chan, &device->channels, device_node) {
 448                 dma_clients_notify_removed(chan);
 449                 device_unregister(&chan->dev);
 450                 dma_chan_release(chan);
 451         }
 452
 453         kref_put(&device->refcount, dma_async_device_cleanup);
 454         wait_for_completion(&device->done);
 455 }
 456 EXPORT_SYMBOL(dma_async_device_unregister);
 457
 458 /**
 459  * dma_async_memcpy_buf_to_buf - offloaded copy between virtual addresses
 460  * @chan: DMA channel to offload copy to
 461  * @dest: destination address (virtual)
 462  * @src: source address (virtual)
 463  * @len: length
 464  *
 465  * Both @dest and @src must be mappable to a bus address according to the
 466  * DMA mapping API rules for streaming mappings.
 467  * Both @dest and @src must stay memory resident (kernel memory or locked
 468  * user space pages).
 469  */
 470 dma_cookie_t
 471 dma_async_memcpy_buf_to_buf(struct dma_chan *chan, void *dest,
 472                         void *src, size_t len)
 473 {
 474         struct dma_device *dev = chan->device;
 475         struct dma_async_tx_descriptor *tx;
 476         dma_addr_t addr;
 477         dma_cookie_t cookie;
 478         int cpu;
 479
 480         tx = dev->device_prep_dma_memcpy(chan, len, 0);
 481         if (!tx)
 482                 return -ENOMEM;
 483
 484         tx->ack = 1;
 485         tx->callback = NULL;
 486         addr = dma_map_single(dev->dev, src, len, DMA_TO_DEVICE);
 487         tx->tx_set_src(addr, tx, 0);
 488         addr = dma_map_single(dev->dev, dest, len, DMA_FROM_DEVICE);
 489         tx->tx_set_dest(addr, tx, 0);
 490         cookie = tx->tx_submit(tx);
 491
 492         cpu = get_cpu();
 493         per_cpu_ptr(chan->local, cpu)->bytes_transferred += len;
 494         per_cpu_ptr(chan->local, cpu)->memcpy_count++;
 495         put_cpu();
 496
 497         return cookie;
 498 }
 499 EXPORT_SYMBOL(dma_async_memcpy_buf_to_buf);
 500
 501 /**
 502  * dma_async_memcpy_buf_to_pg - offloaded copy from address to page
 503  * @chan: DMA channel to offload copy to
 504  * @page: destination page
 505  * @offset: offset in page to copy to
 506  * @kdata: source address (virtual)
 507  * @len: length
 508  *
 509  * Both @page/@offset and @kdata must be mappable to a bus address according
 510  * to the DMA mapping API rules for streaming mappings.
 511  * Both @page/@offset and @kdata must stay memory resident (kernel memory or
 512  * locked user space pages)
 513  */
 514 dma_cookie_t
 515 dma_async_memcpy_buf_to_pg(struct dma_chan *chan, struct page *page,
 516                         unsigned int offset, void *kdata, size_t len)
 517 {
 518         struct dma_device *dev = chan->device;
 519         struct dma_async_tx_descriptor *tx;
 520         dma_addr_t addr;
 521         dma_cookie_t cookie;
 522         int cpu;
 523
 524         tx = dev->device_prep_dma_memcpy(chan, len, 0);
 525         if (!tx)
 526                 return -ENOMEM;
 527
 528         tx->ack = 1;
 529         tx->callback = NULL;
 530         addr = dma_map_single(dev->dev, kdata, len, DMA_TO_DEVICE);
 531         tx->tx_set_src(addr, tx, 0);
 532         addr = dma_map_page(dev->dev, page, offset, len, DMA_FROM_DEVICE);
 533         tx->tx_set_dest(addr, tx, 0);
 534         cookie = tx->tx_submit(tx);
 535
 536         cpu = get_cpu();
 537         per_cpu_ptr(chan->local, cpu)->bytes_transferred += len;
 538         per_cpu_ptr(chan->local, cpu)->memcpy_count++;
 539         put_cpu();
 540
 541         return cookie;
 542 }
 543 EXPORT_SYMBOL(dma_async_memcpy_buf_to_pg);
 544
 545 /**
 546  * dma_async_memcpy_pg_to_pg - offloaded copy from page to page
 547  * @chan: DMA channel to offload copy to
 548  * @dest_pg: destination page
 549  * @dest_off: offset in page to copy to
 550  * @src_pg: source page
 551  * @src_off: offset in page to copy from
 552  * @len: length
 553  *
 554  * Both @dest_page/@dest_off and @src_page/@src_off must be mappable to a bus
 555  * address according to the DMA mapping API rules for streaming mappings.
 556  * Both @dest_page/@dest_off and @src_page/@src_off must stay memory resident
 557  * (kernel memory or locked user space pages).
 558  */
 559 dma_cookie_t
 560 dma_async_memcpy_pg_to_pg(struct dma_chan *chan, struct page *dest_pg,
 561         unsigned int dest_off, struct page *src_pg, unsigned int src_off,
 562         size_t len)
 563 {
 564         struct dma_device *dev = chan->device;
 565         struct dma_async_tx_descriptor *tx;
 566         dma_addr_t addr;
 567         dma_cookie_t cookie;
 568         int cpu;
 569
 570         tx = dev->device_prep_dma_memcpy(chan, len, 0);
 571         if (!tx)
 572                 return -ENOMEM;
 573
 574         tx->ack = 1;
 575         tx->callback = NULL;
 576         addr = dma_map_page(dev->dev, src_pg, src_off, len, DMA_TO_DEVICE);
 577         tx->tx_set_src(addr, tx, 0);
 578         addr = dma_map_page(dev->dev, dest_pg, dest_off, len, DMA_FROM_DEVICE);
 579         tx->tx_set_dest(addr, tx, 0);
 580         cookie = tx->tx_submit(tx);
 581
 582         cpu = get_cpu();
 583         per_cpu_ptr(chan->local, cpu)->bytes_transferred += len;
 584         per_cpu_ptr(chan->local, cpu)->memcpy_count++;
 585         put_cpu();
 586
 587         return cookie;
 588 }
 589 EXPORT_SYMBOL(dma_async_memcpy_pg_to_pg);
 590
 591 void dma_async_tx_descriptor_init(struct dma_async_tx_descriptor *tx,
 592         struct dma_chan *chan)
 593 {
 594         tx->chan = chan;
 595         spin_lock_init(&tx->lock);
 596         INIT_LIST_HEAD(&tx->depend_node);
 597         INIT_LIST_HEAD(&tx->depend_list);
 598 }
 599 EXPORT_SYMBOL(dma_async_tx_descriptor_init);
 600
 601 static int __init dma_bus_init(void)
 602 {
 603         mutex_init(&dma_list_mutex);
 604         return class_register(&dma_devclass);
 605 }
 606 subsys_initcall(dma_bus_init);
 607