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 class_device registered. When the 45 * class_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, class_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 class_device *cd, char *buf) 81 { 82 struct dma_chan *chan = container_of(cd, struct dma_chan, class_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 class_device *cd, char *buf) 93 { 94 struct dma_chan *chan = container_of(cd, struct dma_chan, class_dev); 95 unsigned long count = 0; 96 int i; 97 98 for_each_possible_cpu(i) 99 count += per_cpu_ptr(chan->local, i)->bytes_transferred; 100 101 return sprintf(buf, "%lu\n", count); 102 } 103 104 static ssize_t show_in_use(struct class_device *cd, char *buf) 105 { 106 struct dma_chan *chan = container_of(cd, struct dma_chan, class_dev); 107 int in_use = 0; 108 109 if (unlikely(chan->slow_ref) && 110 atomic_read(&chan->refcount.refcount) > 1) 111 in_use = 1; 112 else { 113 if (local_read(&(per_cpu_ptr(chan->local, 114 get_cpu())->refcount)) > 0) 115 in_use = 1; 116 put_cpu(); 117 } 118 119 return sprintf(buf, "%d\n", in_use); 120 } 121 122 static struct class_device_attribute dma_class_attrs[] = { 123 __ATTR(memcpy_count, S_IRUGO, show_memcpy_count, NULL), 124 __ATTR(bytes_transferred, S_IRUGO, show_bytes_transferred, NULL), 125 __ATTR(in_use, S_IRUGO, show_in_use, NULL), 126 __ATTR_NULL 127 }; 128 129 static void dma_async_device_cleanup(struct kref *kref); 130 131 static void dma_class_dev_release(struct class_device *cd) 132 { 133 struct dma_chan *chan = container_of(cd, struct dma_chan, class_dev); 134 kref_put(&chan->device->refcount, dma_async_device_cleanup); 135 } 136 137 static struct class dma_devclass = { 138 .name = "dma", 139 .class_dev_attrs = dma_class_attrs, 140 .release = dma_class_dev_release, 141 }; 142 143 /* --- client and device registration --- */ 144 145 #define dma_chan_satisfies_mask(chan, mask) \ 146 __dma_chan_satisfies_mask((chan), &(mask)) 147 static int 148 __dma_chan_satisfies_mask(struct dma_chan *chan, dma_cap_mask_t *want) 149 { 150 dma_cap_mask_t has; 151 152 bitmap_and(has.bits, want->bits, chan->device->cap_mask.bits, 153 DMA_TX_TYPE_END); 154 return bitmap_equal(want->bits, has.bits, DMA_TX_TYPE_END); 155 } 156 157 /** 158 * dma_client_chan_alloc - try to allocate channels to a client 159 * @client: &dma_client 160 * 161 * Called with dma_list_mutex held. 162 */ 163 static void dma_client_chan_alloc(struct dma_client *client) 164 { 165 struct dma_device *device; 166 struct dma_chan *chan; 167 int desc; /* allocated descriptor count */ 168 enum dma_state_client ack; 169 170 /* Find a channel */ 171 list_for_each_entry(device, &dma_device_list, global_node) 172 list_for_each_entry(chan, &device->channels, device_node) { 173 if (!dma_chan_satisfies_mask(chan, client->cap_mask)) 174 continue; 175 176 desc = chan->device->device_alloc_chan_resources(chan); 177 if (desc >= 0) { 178 ack = client->event_callback(client, 179 chan, 180 DMA_RESOURCE_AVAILABLE); 181 182 /* we are done once this client rejects 183 * an available resource 184 */ 185 if (ack == DMA_ACK) 186 dma_chan_get(chan); 187 else if (ack == DMA_NAK) 188 return; 189 } 190 } 191 } 192 193 enum dma_status dma_sync_wait(struct dma_chan *chan, dma_cookie_t cookie) 194 { 195 enum dma_status status; 196 unsigned long dma_sync_wait_timeout = jiffies + msecs_to_jiffies(5000); 197 198 dma_async_issue_pending(chan); 199 do { 200 status = dma_async_is_tx_complete(chan, cookie, NULL, NULL); 201 if (time_after_eq(jiffies, dma_sync_wait_timeout)) { 202 printk(KERN_ERR "dma_sync_wait_timeout!\n"); 203 return DMA_ERROR; 204 } 205 } while (status == DMA_IN_PROGRESS); 206 207 return status; 208 } 209 EXPORT_SYMBOL(dma_sync_wait); 210 211 /** 212 * dma_chan_cleanup - release a DMA channel's resources 213 * @kref: kernel reference structure that contains the DMA channel device 214 */ 215 void dma_chan_cleanup(struct kref *kref) 216 { 217 struct dma_chan *chan = container_of(kref, struct dma_chan, refcount); 218 chan->device->device_free_chan_resources(chan); 219 kref_put(&chan->device->refcount, dma_async_device_cleanup); 220 } 221 EXPORT_SYMBOL(dma_chan_cleanup); 222 223 static void dma_chan_free_rcu(struct rcu_head *rcu) 224 { 225 struct dma_chan *chan = container_of(rcu, struct dma_chan, rcu); 226 int bias = 0x7FFFFFFF; 227 int i; 228 for_each_possible_cpu(i) 229 bias -= local_read(&per_cpu_ptr(chan->local, i)->refcount); 230 atomic_sub(bias, &chan->refcount.refcount); 231 kref_put(&chan->refcount, dma_chan_cleanup); 232 } 233 234 static void dma_chan_release(struct dma_chan *chan) 235 { 236 atomic_add(0x7FFFFFFF, &chan->refcount.refcount); 237 chan->slow_ref = 1; 238 call_rcu(&chan->rcu, dma_chan_free_rcu); 239 } 240 241 /** 242 * dma_chans_notify_available - broadcast available channels to the clients 243 */ 244 static void dma_clients_notify_available(void) 245 { 246 struct dma_client *client; 247 248 mutex_lock(&dma_list_mutex); 249 250 list_for_each_entry(client, &dma_client_list, global_node) 251 dma_client_chan_alloc(client); 252 253 mutex_unlock(&dma_list_mutex); 254 } 255 256 /** 257 * dma_chans_notify_available - tell the clients that a channel is going away 258 * @chan: channel on its way out 259 */ 260 static void dma_clients_notify_removed(struct dma_chan *chan) 261 { 262 struct dma_client *client; 263 enum dma_state_client ack; 264 265 mutex_lock(&dma_list_mutex); 266 267 list_for_each_entry(client, &dma_client_list, global_node) { 268 ack = client->event_callback(client, chan, 269 DMA_RESOURCE_REMOVED); 270 271 /* client was holding resources for this channel so 272 * free it 273 */ 274 if (ack == DMA_ACK) 275 dma_chan_put(chan); 276 } 277 278 mutex_unlock(&dma_list_mutex); 279 } 280 281 /** 282 * dma_async_client_register - register a &dma_client 283 * @client: ptr to a client structure with valid 'event_callback' and 'cap_mask' 284 */ 285 void dma_async_client_register(struct dma_client *client) 286 { 287 mutex_lock(&dma_list_mutex); 288 list_add_tail(&client->global_node, &dma_client_list); 289 mutex_unlock(&dma_list_mutex); 290 } 291 EXPORT_SYMBOL(dma_async_client_register); 292 293 /** 294 * dma_async_client_unregister - unregister a client and free the &dma_client 295 * @client: &dma_client to free 296 * 297 * Force frees any allocated DMA channels, frees the &dma_client memory 298 */ 299 void dma_async_client_unregister(struct dma_client *client) 300 { 301 struct dma_device *device; 302 struct dma_chan *chan; 303 enum dma_state_client ack; 304 305 if (!client) 306 return; 307 308 mutex_lock(&dma_list_mutex); 309 /* free all channels the client is holding */ 310 list_for_each_entry(device, &dma_device_list, global_node) 311 list_for_each_entry(chan, &device->channels, device_node) { 312 ack = client->event_callback(client, chan, 313 DMA_RESOURCE_REMOVED); 314 315 if (ack == DMA_ACK) 316 dma_chan_put(chan); 317 } 318 319 list_del(&client->global_node); 320 mutex_unlock(&dma_list_mutex); 321 } 322 EXPORT_SYMBOL(dma_async_client_unregister); 323 324 /** 325 * dma_async_client_chan_request - send all available channels to the 326 * client that satisfy the capability mask 327 * @client - requester 328 */ 329 void dma_async_client_chan_request(struct dma_client *client) 330 { 331 mutex_lock(&dma_list_mutex); 332 dma_client_chan_alloc(client); 333 mutex_unlock(&dma_list_mutex); 334 } 335 EXPORT_SYMBOL(dma_async_client_chan_request); 336 337 /** 338 * dma_async_device_register - registers DMA devices found 339 * @device: &dma_device 340 */ 341 int dma_async_device_register(struct dma_device *device) 342 { 343 static int id; 344 int chancnt = 0, rc; 345 struct dma_chan* chan; 346 347 if (!device) 348 return -ENODEV; 349 350 /* validate device routines */ 351 BUG_ON(dma_has_cap(DMA_MEMCPY, device->cap_mask) && 352 !device->device_prep_dma_memcpy); 353 BUG_ON(dma_has_cap(DMA_XOR, device->cap_mask) && 354 !device->device_prep_dma_xor); 355 BUG_ON(dma_has_cap(DMA_ZERO_SUM, device->cap_mask) && 356 !device->device_prep_dma_zero_sum); 357 BUG_ON(dma_has_cap(DMA_MEMSET, device->cap_mask) && 358 !device->device_prep_dma_memset); 359 BUG_ON(dma_has_cap(DMA_ZERO_SUM, device->cap_mask) && 360 !device->device_prep_dma_interrupt); 361 362 BUG_ON(!device->device_alloc_chan_resources); 363 BUG_ON(!device->device_free_chan_resources); 364 BUG_ON(!device->device_dependency_added); 365 BUG_ON(!device->device_is_tx_complete); 366 BUG_ON(!device->device_issue_pending); 367 BUG_ON(!device->dev); 368 369 init_completion(&device->done); 370 kref_init(&device->refcount); 371 device->dev_id = id++; 372 373 /* represent channels in sysfs. Probably want devs too */ 374 list_for_each_entry(chan, &device->channels, device_node) { 375 chan->local = alloc_percpu(typeof(*chan->local)); 376 if (chan->local == NULL) 377 continue; 378 379 chan->chan_id = chancnt++; 380 chan->class_dev.class = &dma_devclass; 381 chan->class_dev.dev = NULL; 382 snprintf(chan->class_dev.class_id, BUS_ID_SIZE, "dma%dchan%d", 383 device->dev_id, chan->chan_id); 384 385 rc = class_device_register(&chan->class_dev); 386 if (rc) { 387 chancnt--; 388 free_percpu(chan->local); 389 chan->local = NULL; 390 goto err_out; 391 } 392 393 /* One for the channel, one of the class device */ 394 kref_get(&device->refcount); 395 kref_get(&device->refcount); 396 kref_init(&chan->refcount); 397 chan->slow_ref = 0; 398 INIT_RCU_HEAD(&chan->rcu); 399 } 400 401 mutex_lock(&dma_list_mutex); 402 list_add_tail(&device->global_node, &dma_device_list); 403 mutex_unlock(&dma_list_mutex); 404 405 dma_clients_notify_available(); 406 407 return 0; 408 409 err_out: 410 list_for_each_entry(chan, &device->channels, device_node) { 411 if (chan->local == NULL) 412 continue; 413 kref_put(&device->refcount, dma_async_device_cleanup); 414 class_device_unregister(&chan->class_dev); 415 chancnt--; 416 free_percpu(chan->local); 417 } 418 return rc; 419 } 420 EXPORT_SYMBOL(dma_async_device_register); 421 422 /** 423 * dma_async_device_cleanup - function called when all references are released 424 * @kref: kernel reference object 425 */ 426 static void dma_async_device_cleanup(struct kref *kref) 427 { 428 struct dma_device *device; 429 430 device = container_of(kref, struct dma_device, refcount); 431 complete(&device->done); 432 } 433 434 /** 435 * dma_async_device_unregister - unregisters DMA devices 436 * @device: &dma_device 437 */ 438 void dma_async_device_unregister(struct dma_device *device) 439 { 440 struct dma_chan *chan; 441 442 mutex_lock(&dma_list_mutex); 443 list_del(&device->global_node); 444 mutex_unlock(&dma_list_mutex); 445 446 list_for_each_entry(chan, &device->channels, device_node) { 447 dma_clients_notify_removed(chan); 448 class_device_unregister(&chan->class_dev); 449 dma_chan_release(chan); 450 } 451 452 kref_put(&device->refcount, dma_async_device_cleanup); 453 wait_for_completion(&device->done); 454 } 455 EXPORT_SYMBOL(dma_async_device_unregister); 456 457 /** 458 * dma_async_memcpy_buf_to_buf - offloaded copy between virtual addresses 459 * @chan: DMA channel to offload copy to 460 * @dest: destination address (virtual) 461 * @src: source address (virtual) 462 * @len: length 463 * 464 * Both @dest and @src must be mappable to a bus address according to the 465 * DMA mapping API rules for streaming mappings. 466 * Both @dest and @src must stay memory resident (kernel memory or locked 467 * user space pages). 468 */ 469 dma_cookie_t 470 dma_async_memcpy_buf_to_buf(struct dma_chan *chan, void *dest, 471 void *src, size_t len) 472 { 473 struct dma_device *dev = chan->device; 474 struct dma_async_tx_descriptor *tx; 475 dma_addr_t addr; 476 dma_cookie_t cookie; 477 int cpu; 478 479 tx = dev->device_prep_dma_memcpy(chan, len, 0); 480 if (!tx) 481 return -ENOMEM; 482 483 tx->ack = 1; 484 tx->callback = NULL; 485 addr = dma_map_single(dev->dev, src, len, DMA_TO_DEVICE); 486 tx->tx_set_src(addr, tx, 0); 487 addr = dma_map_single(dev->dev, dest, len, DMA_FROM_DEVICE); 488 tx->tx_set_dest(addr, tx, 0); 489 cookie = tx->tx_submit(tx); 490 491 cpu = get_cpu(); 492 per_cpu_ptr(chan->local, cpu)->bytes_transferred += len; 493 per_cpu_ptr(chan->local, cpu)->memcpy_count++; 494 put_cpu(); 495 496 return cookie; 497 } 498 EXPORT_SYMBOL(dma_async_memcpy_buf_to_buf); 499 500 /** 501 * dma_async_memcpy_buf_to_pg - offloaded copy from address to page 502 * @chan: DMA channel to offload copy to 503 * @page: destination page 504 * @offset: offset in page to copy to 505 * @kdata: source address (virtual) 506 * @len: length 507 * 508 * Both @page/@offset and @kdata must be mappable to a bus address according 509 * to the DMA mapping API rules for streaming mappings. 510 * Both @page/@offset and @kdata must stay memory resident (kernel memory or 511 * locked user space pages) 512 */ 513 dma_cookie_t 514 dma_async_memcpy_buf_to_pg(struct dma_chan *chan, struct page *page, 515 unsigned int offset, void *kdata, size_t len) 516 { 517 struct dma_device *dev = chan->device; 518 struct dma_async_tx_descriptor *tx; 519 dma_addr_t addr; 520 dma_cookie_t cookie; 521 int cpu; 522 523 tx = dev->device_prep_dma_memcpy(chan, len, 0); 524 if (!tx) 525 return -ENOMEM; 526 527 tx->ack = 1; 528 tx->callback = NULL; 529 addr = dma_map_single(dev->dev, kdata, len, DMA_TO_DEVICE); 530 tx->tx_set_src(addr, tx, 0); 531 addr = dma_map_page(dev->dev, page, offset, len, DMA_FROM_DEVICE); 532 tx->tx_set_dest(addr, tx, 0); 533 cookie = tx->tx_submit(tx); 534 535 cpu = get_cpu(); 536 per_cpu_ptr(chan->local, cpu)->bytes_transferred += len; 537 per_cpu_ptr(chan->local, cpu)->memcpy_count++; 538 put_cpu(); 539 540 return cookie; 541 } 542 EXPORT_SYMBOL(dma_async_memcpy_buf_to_pg); 543 544 /** 545 * dma_async_memcpy_pg_to_pg - offloaded copy from page to page 546 * @chan: DMA channel to offload copy to 547 * @dest_pg: destination page 548 * @dest_off: offset in page to copy to 549 * @src_pg: source page 550 * @src_off: offset in page to copy from 551 * @len: length 552 * 553 * Both @dest_page/@dest_off and @src_page/@src_off must be mappable to a bus 554 * address according to the DMA mapping API rules for streaming mappings. 555 * Both @dest_page/@dest_off and @src_page/@src_off must stay memory resident 556 * (kernel memory or locked user space pages). 557 */ 558 dma_cookie_t 559 dma_async_memcpy_pg_to_pg(struct dma_chan *chan, struct page *dest_pg, 560 unsigned int dest_off, struct page *src_pg, unsigned int src_off, 561 size_t len) 562 { 563 struct dma_device *dev = chan->device; 564 struct dma_async_tx_descriptor *tx; 565 dma_addr_t addr; 566 dma_cookie_t cookie; 567 int cpu; 568 569 tx = dev->device_prep_dma_memcpy(chan, len, 0); 570 if (!tx) 571 return -ENOMEM; 572 573 tx->ack = 1; 574 tx->callback = NULL; 575 addr = dma_map_page(dev->dev, src_pg, src_off, len, DMA_TO_DEVICE); 576 tx->tx_set_src(addr, tx, 0); 577 addr = dma_map_page(dev->dev, dest_pg, dest_off, len, DMA_FROM_DEVICE); 578 tx->tx_set_dest(addr, tx, 0); 579 cookie = tx->tx_submit(tx); 580 581 cpu = get_cpu(); 582 per_cpu_ptr(chan->local, cpu)->bytes_transferred += len; 583 per_cpu_ptr(chan->local, cpu)->memcpy_count++; 584 put_cpu(); 585 586 return cookie; 587 } 588 EXPORT_SYMBOL(dma_async_memcpy_pg_to_pg); 589 590 void dma_async_tx_descriptor_init(struct dma_async_tx_descriptor *tx, 591 struct dma_chan *chan) 592 { 593 tx->chan = chan; 594 spin_lock_init(&tx->lock); 595 INIT_LIST_HEAD(&tx->depend_node); 596 INIT_LIST_HEAD(&tx->depend_list); 597 } 598 EXPORT_SYMBOL(dma_async_tx_descriptor_init); 599 600 static int __init dma_bus_init(void) 601 { 602 mutex_init(&dma_list_mutex); 603 return class_register(&dma_devclass); 604 } 605 subsys_initcall(dma_bus_init); 606 607