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 device registered. When the
45 * device is released, the corresponding 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 corresponding 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.
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 unregistered. 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 device *dev, struct device_attribute *attr, char *buf)
82 struct dma_chan *chan = to_dma_chan(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 device *dev, struct device_attribute *attr,
95 struct dma_chan *chan = to_dma_chan(dev);
96 unsigned long count = 0;
99 for_each_possible_cpu(i)
100 count += per_cpu_ptr(chan->local, i)->bytes_transferred;
102 return sprintf(buf, "%lu\n", count);
105 static ssize_t show_in_use(struct device *dev, struct device_attribute *attr, char *buf)
107 struct dma_chan *chan = to_dma_chan(dev);
110 if (unlikely(chan->slow_ref) &&
111 atomic_read(&chan->refcount.refcount) > 1)
114 if (local_read(&(per_cpu_ptr(chan->local,
115 get_cpu())->refcount)) > 0)
120 return sprintf(buf, "%d\n", in_use);
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),
130 static void dma_async_device_cleanup(struct kref *kref);
132 static void dma_dev_release(struct device *dev)
134 struct dma_chan *chan = to_dma_chan(dev);
135 kref_put(&chan->device->refcount, dma_async_device_cleanup);
138 static struct class dma_devclass = {
140 .dev_attrs = dma_attrs,
141 .dev_release = dma_dev_release,
144 /* --- client and device registration --- */
146 #define dma_chan_satisfies_mask(chan, mask) \
147 __dma_chan_satisfies_mask((chan), &(mask))
149 __dma_chan_satisfies_mask(struct dma_chan *chan, dma_cap_mask_t *want)
153 bitmap_and(has.bits, want->bits, chan->device->cap_mask.bits,
155 return bitmap_equal(want->bits, has.bits, DMA_TX_TYPE_END);
159 * dma_client_chan_alloc - try to allocate channels to a client
160 * @client: &dma_client
162 * Called with dma_list_mutex held.
164 static void dma_client_chan_alloc(struct dma_client *client)
166 struct dma_device *device;
167 struct dma_chan *chan;
168 int desc; /* allocated descriptor count */
169 enum dma_state_client ack;
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))
177 desc = chan->device->device_alloc_chan_resources(
180 ack = client->event_callback(client,
182 DMA_RESOURCE_AVAILABLE);
184 /* we are done once this client rejects
185 * an available resource
187 if (ack == DMA_ACK) {
189 chan->client_count++;
190 } else if (ack == DMA_NAK)
196 enum dma_status dma_sync_wait(struct dma_chan *chan, dma_cookie_t cookie)
198 enum dma_status status;
199 unsigned long dma_sync_wait_timeout = jiffies + msecs_to_jiffies(5000);
201 dma_async_issue_pending(chan);
203 status = dma_async_is_tx_complete(chan, cookie, NULL, NULL);
204 if (time_after_eq(jiffies, dma_sync_wait_timeout)) {
205 printk(KERN_ERR "dma_sync_wait_timeout!\n");
208 } while (status == DMA_IN_PROGRESS);
212 EXPORT_SYMBOL(dma_sync_wait);
215 * dma_chan_cleanup - release a DMA channel's resources
216 * @kref: kernel reference structure that contains the DMA channel device
218 void dma_chan_cleanup(struct kref *kref)
220 struct dma_chan *chan = container_of(kref, struct dma_chan, refcount);
221 chan->device->device_free_chan_resources(chan);
222 kref_put(&chan->device->refcount, dma_async_device_cleanup);
224 EXPORT_SYMBOL(dma_chan_cleanup);
226 static void dma_chan_free_rcu(struct rcu_head *rcu)
228 struct dma_chan *chan = container_of(rcu, struct dma_chan, rcu);
229 int bias = 0x7FFFFFFF;
231 for_each_possible_cpu(i)
232 bias -= local_read(&per_cpu_ptr(chan->local, i)->refcount);
233 atomic_sub(bias, &chan->refcount.refcount);
234 kref_put(&chan->refcount, dma_chan_cleanup);
237 static void dma_chan_release(struct dma_chan *chan)
239 atomic_add(0x7FFFFFFF, &chan->refcount.refcount);
241 call_rcu(&chan->rcu, dma_chan_free_rcu);
245 * dma_chans_notify_available - broadcast available channels to the clients
247 static void dma_clients_notify_available(void)
249 struct dma_client *client;
251 mutex_lock(&dma_list_mutex);
253 list_for_each_entry(client, &dma_client_list, global_node)
254 dma_client_chan_alloc(client);
256 mutex_unlock(&dma_list_mutex);
260 * dma_chans_notify_available - tell the clients that a channel is going away
261 * @chan: channel on its way out
263 static void dma_clients_notify_removed(struct dma_chan *chan)
265 struct dma_client *client;
266 enum dma_state_client ack;
268 mutex_lock(&dma_list_mutex);
270 list_for_each_entry(client, &dma_client_list, global_node) {
271 ack = client->event_callback(client, chan,
272 DMA_RESOURCE_REMOVED);
274 /* client was holding resources for this channel so
277 if (ack == DMA_ACK) {
279 chan->client_count--;
283 mutex_unlock(&dma_list_mutex);
287 * dma_async_client_register - register a &dma_client
288 * @client: ptr to a client structure with valid 'event_callback' and 'cap_mask'
290 void dma_async_client_register(struct dma_client *client)
292 mutex_lock(&dma_list_mutex);
293 list_add_tail(&client->global_node, &dma_client_list);
294 mutex_unlock(&dma_list_mutex);
296 EXPORT_SYMBOL(dma_async_client_register);
299 * dma_async_client_unregister - unregister a client and free the &dma_client
300 * @client: &dma_client to free
302 * Force frees any allocated DMA channels, frees the &dma_client memory
304 void dma_async_client_unregister(struct dma_client *client)
306 struct dma_device *device;
307 struct dma_chan *chan;
308 enum dma_state_client ack;
313 mutex_lock(&dma_list_mutex);
314 /* free all channels the client is holding */
315 list_for_each_entry(device, &dma_device_list, global_node)
316 list_for_each_entry(chan, &device->channels, device_node) {
317 ack = client->event_callback(client, chan,
318 DMA_RESOURCE_REMOVED);
320 if (ack == DMA_ACK) {
322 chan->client_count--;
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_INTERRUPT, 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_is_tx_complete);
372 BUG_ON(!device->device_issue_pending);
373 BUG_ON(!device->dev);
375 init_completion(&device->done);
376 kref_init(&device->refcount);
377 device->dev_id = id++;
379 /* represent channels in sysfs. Probably want devs too */
380 list_for_each_entry(chan, &device->channels, device_node) {
381 chan->local = alloc_percpu(typeof(*chan->local));
382 if (chan->local == NULL)
385 chan->chan_id = chancnt++;
386 chan->dev.class = &dma_devclass;
387 chan->dev.parent = device->dev;
388 snprintf(chan->dev.bus_id, BUS_ID_SIZE, "dma%dchan%d",
389 device->dev_id, chan->chan_id);
391 rc = device_register(&chan->dev);
394 free_percpu(chan->local);
399 /* One for the channel, one of the class device */
400 kref_get(&device->refcount);
401 kref_get(&device->refcount);
402 kref_init(&chan->refcount);
403 chan->client_count = 0;
405 INIT_RCU_HEAD(&chan->rcu);
408 mutex_lock(&dma_list_mutex);
409 list_add_tail(&device->global_node, &dma_device_list);
410 mutex_unlock(&dma_list_mutex);
412 dma_clients_notify_available();
417 list_for_each_entry(chan, &device->channels, device_node) {
418 if (chan->local == NULL)
420 kref_put(&device->refcount, dma_async_device_cleanup);
421 device_unregister(&chan->dev);
423 free_percpu(chan->local);
427 EXPORT_SYMBOL(dma_async_device_register);
430 * dma_async_device_cleanup - function called when all references are released
431 * @kref: kernel reference object
433 static void dma_async_device_cleanup(struct kref *kref)
435 struct dma_device *device;
437 device = container_of(kref, struct dma_device, refcount);
438 complete(&device->done);
442 * dma_async_device_unregister - unregisters DMA devices
443 * @device: &dma_device
445 void dma_async_device_unregister(struct dma_device *device)
447 struct dma_chan *chan;
449 mutex_lock(&dma_list_mutex);
450 list_del(&device->global_node);
451 mutex_unlock(&dma_list_mutex);
453 list_for_each_entry(chan, &device->channels, device_node) {
454 dma_clients_notify_removed(chan);
455 device_unregister(&chan->dev);
456 dma_chan_release(chan);
459 kref_put(&device->refcount, dma_async_device_cleanup);
460 wait_for_completion(&device->done);
462 EXPORT_SYMBOL(dma_async_device_unregister);
465 * dma_async_memcpy_buf_to_buf - offloaded copy between virtual addresses
466 * @chan: DMA channel to offload copy to
467 * @dest: destination address (virtual)
468 * @src: source address (virtual)
471 * Both @dest and @src must be mappable to a bus address according to the
472 * DMA mapping API rules for streaming mappings.
473 * Both @dest and @src must stay memory resident (kernel memory or locked
477 dma_async_memcpy_buf_to_buf(struct dma_chan *chan, void *dest,
478 void *src, size_t len)
480 struct dma_device *dev = chan->device;
481 struct dma_async_tx_descriptor *tx;
482 dma_addr_t dma_dest, dma_src;
486 dma_src = dma_map_single(dev->dev, src, len, DMA_TO_DEVICE);
487 dma_dest = dma_map_single(dev->dev, dest, len, DMA_FROM_DEVICE);
488 tx = dev->device_prep_dma_memcpy(chan, dma_dest, dma_src, len,
492 dma_unmap_single(dev->dev, dma_src, len, DMA_TO_DEVICE);
493 dma_unmap_single(dev->dev, dma_dest, len, DMA_FROM_DEVICE);
498 cookie = tx->tx_submit(tx);
501 per_cpu_ptr(chan->local, cpu)->bytes_transferred += len;
502 per_cpu_ptr(chan->local, cpu)->memcpy_count++;
507 EXPORT_SYMBOL(dma_async_memcpy_buf_to_buf);
510 * dma_async_memcpy_buf_to_pg - offloaded copy from address to page
511 * @chan: DMA channel to offload copy to
512 * @page: destination page
513 * @offset: offset in page to copy to
514 * @kdata: source address (virtual)
517 * Both @page/@offset and @kdata must be mappable to a bus address according
518 * to the DMA mapping API rules for streaming mappings.
519 * Both @page/@offset and @kdata must stay memory resident (kernel memory or
520 * locked user space pages)
523 dma_async_memcpy_buf_to_pg(struct dma_chan *chan, struct page *page,
524 unsigned int offset, void *kdata, size_t len)
526 struct dma_device *dev = chan->device;
527 struct dma_async_tx_descriptor *tx;
528 dma_addr_t dma_dest, dma_src;
532 dma_src = dma_map_single(dev->dev, kdata, len, DMA_TO_DEVICE);
533 dma_dest = dma_map_page(dev->dev, page, offset, len, DMA_FROM_DEVICE);
534 tx = dev->device_prep_dma_memcpy(chan, dma_dest, dma_src, len,
538 dma_unmap_single(dev->dev, dma_src, len, DMA_TO_DEVICE);
539 dma_unmap_page(dev->dev, dma_dest, len, DMA_FROM_DEVICE);
544 cookie = tx->tx_submit(tx);
547 per_cpu_ptr(chan->local, cpu)->bytes_transferred += len;
548 per_cpu_ptr(chan->local, cpu)->memcpy_count++;
553 EXPORT_SYMBOL(dma_async_memcpy_buf_to_pg);
556 * dma_async_memcpy_pg_to_pg - offloaded copy from page to page
557 * @chan: DMA channel to offload copy to
558 * @dest_pg: destination page
559 * @dest_off: offset in page to copy to
560 * @src_pg: source page
561 * @src_off: offset in page to copy from
564 * Both @dest_page/@dest_off and @src_page/@src_off must be mappable to a bus
565 * address according to the DMA mapping API rules for streaming mappings.
566 * Both @dest_page/@dest_off and @src_page/@src_off must stay memory resident
567 * (kernel memory or locked user space pages).
570 dma_async_memcpy_pg_to_pg(struct dma_chan *chan, struct page *dest_pg,
571 unsigned int dest_off, struct page *src_pg, unsigned int src_off,
574 struct dma_device *dev = chan->device;
575 struct dma_async_tx_descriptor *tx;
576 dma_addr_t dma_dest, dma_src;
580 dma_src = dma_map_page(dev->dev, src_pg, src_off, len, DMA_TO_DEVICE);
581 dma_dest = dma_map_page(dev->dev, dest_pg, dest_off, len,
583 tx = dev->device_prep_dma_memcpy(chan, dma_dest, dma_src, len,
587 dma_unmap_page(dev->dev, dma_src, len, DMA_TO_DEVICE);
588 dma_unmap_page(dev->dev, dma_dest, len, DMA_FROM_DEVICE);
593 cookie = tx->tx_submit(tx);
596 per_cpu_ptr(chan->local, cpu)->bytes_transferred += len;
597 per_cpu_ptr(chan->local, cpu)->memcpy_count++;
602 EXPORT_SYMBOL(dma_async_memcpy_pg_to_pg);
604 void dma_async_tx_descriptor_init(struct dma_async_tx_descriptor *tx,
605 struct dma_chan *chan)
608 spin_lock_init(&tx->lock);
610 EXPORT_SYMBOL(dma_async_tx_descriptor_init);
612 static int __init dma_bus_init(void)
614 mutex_init(&dma_list_mutex);
615 return class_register(&dma_devclass);
617 subsys_initcall(dma_bus_init);