2 * Copyright(c) 2004 - 2006 Intel Corporation. All rights reserved.
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)
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
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.
18 * The full GNU General Public License is included in this distribution in the
19 * file called COPYING.
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
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,
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.
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.
40 * Each client is responsible for keeping track of the channels it uses. See
41 * the definition of dma_event_callback in dmaengine.h.
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.
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.
62 #include <linux/init.h>
63 #include <linux/module.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>
74 static DEFINE_MUTEX(dma_list_mutex);
75 static LIST_HEAD(dma_device_list);
76 static LIST_HEAD(dma_client_list);
78 /* --- sysfs implementation --- */
80 static ssize_t show_memcpy_count(struct class_device *cd, char *buf)
82 struct dma_chan *chan = container_of(cd, struct dma_chan, class_dev);
83 unsigned long count = 0;
86 for_each_possible_cpu(i)
87 count += per_cpu_ptr(chan->local, i)->memcpy_count;
89 return sprintf(buf, "%lu\n", count);
92 static ssize_t show_bytes_transferred(struct class_device *cd, char *buf)
94 struct dma_chan *chan = container_of(cd, struct dma_chan, class_dev);
95 unsigned long count = 0;
98 for_each_possible_cpu(i)
99 count += per_cpu_ptr(chan->local, i)->bytes_transferred;
101 return sprintf(buf, "%lu\n", count);
104 static ssize_t show_in_use(struct class_device *cd, char *buf)
106 struct dma_chan *chan = container_of(cd, struct dma_chan, class_dev);
109 if (unlikely(chan->slow_ref) &&
110 atomic_read(&chan->refcount.refcount) > 1)
113 if (local_read(&(per_cpu_ptr(chan->local,
114 get_cpu())->refcount)) > 0)
119 return sprintf(buf, "%d\n", in_use);
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),
129 static void dma_async_device_cleanup(struct kref *kref);
131 static void dma_class_dev_release(struct class_device *cd)
133 struct dma_chan *chan = container_of(cd, struct dma_chan, class_dev);
134 kref_put(&chan->device->refcount, dma_async_device_cleanup);
137 static struct class dma_devclass = {
139 .class_dev_attrs = dma_class_attrs,
140 .release = dma_class_dev_release,
143 /* --- client and device registration --- */
145 #define dma_chan_satisfies_mask(chan, mask) \
146 __dma_chan_satisfies_mask((chan), &(mask))
148 __dma_chan_satisfies_mask(struct dma_chan *chan, dma_cap_mask_t *want)
152 bitmap_and(has.bits, want->bits, chan->device->cap_mask.bits,
154 return bitmap_equal(want->bits, has.bits, DMA_TX_TYPE_END);
158 * dma_client_chan_alloc - try to allocate channels to a client
159 * @client: &dma_client
161 * Called with dma_list_mutex held.
163 static void dma_client_chan_alloc(struct dma_client *client)
165 struct dma_device *device;
166 struct dma_chan *chan;
167 int desc; /* allocated descriptor count */
168 enum dma_state_client ack;
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))
176 desc = chan->device->device_alloc_chan_resources(chan);
178 ack = client->event_callback(client,
180 DMA_RESOURCE_AVAILABLE);
182 /* we are done once this client rejects
183 * an available resource
185 if (ack == DMA_ACK) {
187 kref_get(&device->refcount);
188 } else if (ack == DMA_NAK)
194 enum dma_status dma_sync_wait(struct dma_chan *chan, dma_cookie_t cookie)
196 enum dma_status status;
197 unsigned long dma_sync_wait_timeout = jiffies + msecs_to_jiffies(5000);
199 dma_async_issue_pending(chan);
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");
206 } while (status == DMA_IN_PROGRESS);
210 EXPORT_SYMBOL(dma_sync_wait);
213 * dma_chan_cleanup - release a DMA channel's resources
214 * @kref: kernel reference structure that contains the DMA channel device
216 void dma_chan_cleanup(struct kref *kref)
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);
222 EXPORT_SYMBOL(dma_chan_cleanup);
224 static void dma_chan_free_rcu(struct rcu_head *rcu)
226 struct dma_chan *chan = container_of(rcu, struct dma_chan, rcu);
227 int bias = 0x7FFFFFFF;
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);
235 static void dma_chan_release(struct dma_chan *chan)
237 atomic_add(0x7FFFFFFF, &chan->refcount.refcount);
239 call_rcu(&chan->rcu, dma_chan_free_rcu);
243 * dma_chans_notify_available - broadcast available channels to the clients
245 static void dma_clients_notify_available(void)
247 struct dma_client *client;
249 mutex_lock(&dma_list_mutex);
251 list_for_each_entry(client, &dma_client_list, global_node)
252 dma_client_chan_alloc(client);
254 mutex_unlock(&dma_list_mutex);
258 * dma_chans_notify_available - tell the clients that a channel is going away
259 * @chan: channel on its way out
261 static void dma_clients_notify_removed(struct dma_chan *chan)
263 struct dma_client *client;
264 enum dma_state_client ack;
266 mutex_lock(&dma_list_mutex);
268 list_for_each_entry(client, &dma_client_list, global_node) {
269 ack = client->event_callback(client, chan,
270 DMA_RESOURCE_REMOVED);
272 /* client was holding resources for this channel so
275 if (ack == DMA_ACK) {
277 kref_put(&chan->device->refcount,
278 dma_async_device_cleanup);
282 mutex_unlock(&dma_list_mutex);
286 * dma_async_client_register - register a &dma_client
287 * @client: ptr to a client structure with valid 'event_callback' and 'cap_mask'
289 void dma_async_client_register(struct dma_client *client)
291 mutex_lock(&dma_list_mutex);
292 list_add_tail(&client->global_node, &dma_client_list);
293 mutex_unlock(&dma_list_mutex);
295 EXPORT_SYMBOL(dma_async_client_register);
298 * dma_async_client_unregister - unregister a client and free the &dma_client
299 * @client: &dma_client to free
301 * Force frees any allocated DMA channels, frees the &dma_client memory
303 void dma_async_client_unregister(struct dma_client *client)
305 struct dma_device *device;
306 struct dma_chan *chan;
307 enum dma_state_client ack;
312 mutex_lock(&dma_list_mutex);
313 /* free all channels the client is holding */
314 list_for_each_entry(device, &dma_device_list, global_node)
315 list_for_each_entry(chan, &device->channels, device_node) {
316 ack = client->event_callback(client, chan,
317 DMA_RESOURCE_REMOVED);
319 if (ack == DMA_ACK) {
321 kref_put(&chan->device->refcount,
322 dma_async_device_cleanup);
326 list_del(&client->global_node);
327 mutex_unlock(&dma_list_mutex);
329 EXPORT_SYMBOL(dma_async_client_unregister);
332 * dma_async_client_chan_request - send all available channels to the
333 * client that satisfy the capability mask
334 * @client - requester
336 void dma_async_client_chan_request(struct dma_client *client)
338 mutex_lock(&dma_list_mutex);
339 dma_client_chan_alloc(client);
340 mutex_unlock(&dma_list_mutex);
342 EXPORT_SYMBOL(dma_async_client_chan_request);
345 * dma_async_device_register - registers DMA devices found
346 * @device: &dma_device
348 int dma_async_device_register(struct dma_device *device)
352 struct dma_chan* chan;
357 /* validate device routines */
358 BUG_ON(dma_has_cap(DMA_MEMCPY, device->cap_mask) &&
359 !device->device_prep_dma_memcpy);
360 BUG_ON(dma_has_cap(DMA_XOR, device->cap_mask) &&
361 !device->device_prep_dma_xor);
362 BUG_ON(dma_has_cap(DMA_ZERO_SUM, device->cap_mask) &&
363 !device->device_prep_dma_zero_sum);
364 BUG_ON(dma_has_cap(DMA_MEMSET, device->cap_mask) &&
365 !device->device_prep_dma_memset);
366 BUG_ON(dma_has_cap(DMA_ZERO_SUM, device->cap_mask) &&
367 !device->device_prep_dma_interrupt);
369 BUG_ON(!device->device_alloc_chan_resources);
370 BUG_ON(!device->device_free_chan_resources);
371 BUG_ON(!device->device_dependency_added);
372 BUG_ON(!device->device_is_tx_complete);
373 BUG_ON(!device->device_issue_pending);
374 BUG_ON(!device->dev);
376 init_completion(&device->done);
377 kref_init(&device->refcount);
378 device->dev_id = id++;
380 /* represent channels in sysfs. Probably want devs too */
381 list_for_each_entry(chan, &device->channels, device_node) {
382 chan->local = alloc_percpu(typeof(*chan->local));
383 if (chan->local == NULL)
386 chan->chan_id = chancnt++;
387 chan->class_dev.class = &dma_devclass;
388 chan->class_dev.dev = NULL;
389 snprintf(chan->class_dev.class_id, BUS_ID_SIZE, "dma%dchan%d",
390 device->dev_id, chan->chan_id);
392 rc = class_device_register(&chan->class_dev);
395 free_percpu(chan->local);
400 kref_get(&device->refcount);
401 kref_init(&chan->refcount);
403 INIT_RCU_HEAD(&chan->rcu);
406 mutex_lock(&dma_list_mutex);
407 list_add_tail(&device->global_node, &dma_device_list);
408 mutex_unlock(&dma_list_mutex);
410 dma_clients_notify_available();
415 list_for_each_entry(chan, &device->channels, device_node) {
416 if (chan->local == NULL)
418 kref_put(&device->refcount, dma_async_device_cleanup);
419 class_device_unregister(&chan->class_dev);
421 free_percpu(chan->local);
425 EXPORT_SYMBOL(dma_async_device_register);
428 * dma_async_device_cleanup - function called when all references are released
429 * @kref: kernel reference object
431 static void dma_async_device_cleanup(struct kref *kref)
433 struct dma_device *device;
435 device = container_of(kref, struct dma_device, refcount);
436 complete(&device->done);
440 * dma_async_device_unregister - unregisters DMA devices
441 * @device: &dma_device
443 void dma_async_device_unregister(struct dma_device *device)
445 struct dma_chan *chan;
447 mutex_lock(&dma_list_mutex);
448 list_del(&device->global_node);
449 mutex_unlock(&dma_list_mutex);
451 list_for_each_entry(chan, &device->channels, device_node) {
452 dma_clients_notify_removed(chan);
453 class_device_unregister(&chan->class_dev);
454 dma_chan_release(chan);
457 kref_put(&device->refcount, dma_async_device_cleanup);
458 wait_for_completion(&device->done);
460 EXPORT_SYMBOL(dma_async_device_unregister);
463 * dma_async_memcpy_buf_to_buf - offloaded copy between virtual addresses
464 * @chan: DMA channel to offload copy to
465 * @dest: destination address (virtual)
466 * @src: source address (virtual)
469 * Both @dest and @src must be mappable to a bus address according to the
470 * DMA mapping API rules for streaming mappings.
471 * Both @dest and @src must stay memory resident (kernel memory or locked
475 dma_async_memcpy_buf_to_buf(struct dma_chan *chan, void *dest,
476 void *src, size_t len)
478 struct dma_device *dev = chan->device;
479 struct dma_async_tx_descriptor *tx;
484 tx = dev->device_prep_dma_memcpy(chan, len, 0);
490 addr = dma_map_single(dev->dev, src, len, DMA_TO_DEVICE);
491 tx->tx_set_src(addr, tx, 0);
492 addr = dma_map_single(dev->dev, dest, len, DMA_FROM_DEVICE);
493 tx->tx_set_dest(addr, tx, 0);
494 cookie = tx->tx_submit(tx);
497 per_cpu_ptr(chan->local, cpu)->bytes_transferred += len;
498 per_cpu_ptr(chan->local, cpu)->memcpy_count++;
503 EXPORT_SYMBOL(dma_async_memcpy_buf_to_buf);
506 * dma_async_memcpy_buf_to_pg - offloaded copy from address to page
507 * @chan: DMA channel to offload copy to
508 * @page: destination page
509 * @offset: offset in page to copy to
510 * @kdata: source address (virtual)
513 * Both @page/@offset and @kdata must be mappable to a bus address according
514 * to the DMA mapping API rules for streaming mappings.
515 * Both @page/@offset and @kdata must stay memory resident (kernel memory or
516 * locked user space pages)
519 dma_async_memcpy_buf_to_pg(struct dma_chan *chan, struct page *page,
520 unsigned int offset, void *kdata, size_t len)
522 struct dma_device *dev = chan->device;
523 struct dma_async_tx_descriptor *tx;
528 tx = dev->device_prep_dma_memcpy(chan, len, 0);
534 addr = dma_map_single(dev->dev, kdata, len, DMA_TO_DEVICE);
535 tx->tx_set_src(addr, tx, 0);
536 addr = dma_map_page(dev->dev, page, offset, len, DMA_FROM_DEVICE);
537 tx->tx_set_dest(addr, tx, 0);
538 cookie = tx->tx_submit(tx);
541 per_cpu_ptr(chan->local, cpu)->bytes_transferred += len;
542 per_cpu_ptr(chan->local, cpu)->memcpy_count++;
547 EXPORT_SYMBOL(dma_async_memcpy_buf_to_pg);
550 * dma_async_memcpy_pg_to_pg - offloaded copy from page to page
551 * @chan: DMA channel to offload copy to
552 * @dest_pg: destination page
553 * @dest_off: offset in page to copy to
554 * @src_pg: source page
555 * @src_off: offset in page to copy from
558 * Both @dest_page/@dest_off and @src_page/@src_off must be mappable to a bus
559 * address according to the DMA mapping API rules for streaming mappings.
560 * Both @dest_page/@dest_off and @src_page/@src_off must stay memory resident
561 * (kernel memory or locked user space pages).
564 dma_async_memcpy_pg_to_pg(struct dma_chan *chan, struct page *dest_pg,
565 unsigned int dest_off, struct page *src_pg, unsigned int src_off,
568 struct dma_device *dev = chan->device;
569 struct dma_async_tx_descriptor *tx;
574 tx = dev->device_prep_dma_memcpy(chan, len, 0);
580 addr = dma_map_page(dev->dev, src_pg, src_off, len, DMA_TO_DEVICE);
581 tx->tx_set_src(addr, tx, 0);
582 addr = dma_map_page(dev->dev, dest_pg, dest_off, len, DMA_FROM_DEVICE);
583 tx->tx_set_dest(addr, tx, 0);
584 cookie = tx->tx_submit(tx);
587 per_cpu_ptr(chan->local, cpu)->bytes_transferred += len;
588 per_cpu_ptr(chan->local, cpu)->memcpy_count++;
593 EXPORT_SYMBOL(dma_async_memcpy_pg_to_pg);
595 void dma_async_tx_descriptor_init(struct dma_async_tx_descriptor *tx,
596 struct dma_chan *chan)
599 spin_lock_init(&tx->lock);
600 INIT_LIST_HEAD(&tx->depend_node);
601 INIT_LIST_HEAD(&tx->depend_list);
603 EXPORT_SYMBOL(dma_async_tx_descriptor_init);
605 static int __init dma_bus_init(void)
607 mutex_init(&dma_list_mutex);
608 return class_register(&dma_devclass);
610 subsys_initcall(dma_bus_init);