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 /* Does the client require a specific DMA controller? */
174 if (client->slave && client->slave->dma_dev
175 && client->slave->dma_dev != device->dev)
178 list_for_each_entry(chan, &device->channels, device_node) {
179 if (!dma_chan_satisfies_mask(chan, client->cap_mask))
182 desc = chan->device->device_alloc_chan_resources(
185 ack = client->event_callback(client,
187 DMA_RESOURCE_AVAILABLE);
189 /* we are done once this client rejects
190 * an available resource
192 if (ack == DMA_ACK) {
194 chan->client_count++;
195 } else if (ack == DMA_NAK)
202 enum dma_status dma_sync_wait(struct dma_chan *chan, dma_cookie_t cookie)
204 enum dma_status status;
205 unsigned long dma_sync_wait_timeout = jiffies + msecs_to_jiffies(5000);
207 dma_async_issue_pending(chan);
209 status = dma_async_is_tx_complete(chan, cookie, NULL, NULL);
210 if (time_after_eq(jiffies, dma_sync_wait_timeout)) {
211 printk(KERN_ERR "dma_sync_wait_timeout!\n");
214 } while (status == DMA_IN_PROGRESS);
218 EXPORT_SYMBOL(dma_sync_wait);
221 * dma_chan_cleanup - release a DMA channel's resources
222 * @kref: kernel reference structure that contains the DMA channel device
224 void dma_chan_cleanup(struct kref *kref)
226 struct dma_chan *chan = container_of(kref, struct dma_chan, refcount);
227 chan->device->device_free_chan_resources(chan);
228 kref_put(&chan->device->refcount, dma_async_device_cleanup);
230 EXPORT_SYMBOL(dma_chan_cleanup);
232 static void dma_chan_free_rcu(struct rcu_head *rcu)
234 struct dma_chan *chan = container_of(rcu, struct dma_chan, rcu);
235 int bias = 0x7FFFFFFF;
237 for_each_possible_cpu(i)
238 bias -= local_read(&per_cpu_ptr(chan->local, i)->refcount);
239 atomic_sub(bias, &chan->refcount.refcount);
240 kref_put(&chan->refcount, dma_chan_cleanup);
243 static void dma_chan_release(struct dma_chan *chan)
245 atomic_add(0x7FFFFFFF, &chan->refcount.refcount);
247 call_rcu(&chan->rcu, dma_chan_free_rcu);
251 * dma_chans_notify_available - broadcast available channels to the clients
253 static void dma_clients_notify_available(void)
255 struct dma_client *client;
257 mutex_lock(&dma_list_mutex);
259 list_for_each_entry(client, &dma_client_list, global_node)
260 dma_client_chan_alloc(client);
262 mutex_unlock(&dma_list_mutex);
266 * dma_chans_notify_available - tell the clients that a channel is going away
267 * @chan: channel on its way out
269 static void dma_clients_notify_removed(struct dma_chan *chan)
271 struct dma_client *client;
272 enum dma_state_client ack;
274 mutex_lock(&dma_list_mutex);
276 list_for_each_entry(client, &dma_client_list, global_node) {
277 ack = client->event_callback(client, chan,
278 DMA_RESOURCE_REMOVED);
280 /* client was holding resources for this channel so
283 if (ack == DMA_ACK) {
285 chan->client_count--;
289 mutex_unlock(&dma_list_mutex);
293 * dma_async_client_register - register a &dma_client
294 * @client: ptr to a client structure with valid 'event_callback' and 'cap_mask'
296 void dma_async_client_register(struct dma_client *client)
298 /* validate client data */
299 BUG_ON(dma_has_cap(DMA_SLAVE, client->cap_mask) &&
302 mutex_lock(&dma_list_mutex);
303 list_add_tail(&client->global_node, &dma_client_list);
304 mutex_unlock(&dma_list_mutex);
306 EXPORT_SYMBOL(dma_async_client_register);
309 * dma_async_client_unregister - unregister a client and free the &dma_client
310 * @client: &dma_client to free
312 * Force frees any allocated DMA channels, frees the &dma_client memory
314 void dma_async_client_unregister(struct dma_client *client)
316 struct dma_device *device;
317 struct dma_chan *chan;
318 enum dma_state_client ack;
323 mutex_lock(&dma_list_mutex);
324 /* free all channels the client is holding */
325 list_for_each_entry(device, &dma_device_list, global_node)
326 list_for_each_entry(chan, &device->channels, device_node) {
327 ack = client->event_callback(client, chan,
328 DMA_RESOURCE_REMOVED);
330 if (ack == DMA_ACK) {
332 chan->client_count--;
336 list_del(&client->global_node);
337 mutex_unlock(&dma_list_mutex);
339 EXPORT_SYMBOL(dma_async_client_unregister);
342 * dma_async_client_chan_request - send all available channels to the
343 * client that satisfy the capability mask
344 * @client - requester
346 void dma_async_client_chan_request(struct dma_client *client)
348 mutex_lock(&dma_list_mutex);
349 dma_client_chan_alloc(client);
350 mutex_unlock(&dma_list_mutex);
352 EXPORT_SYMBOL(dma_async_client_chan_request);
355 * dma_async_device_register - registers DMA devices found
356 * @device: &dma_device
358 int dma_async_device_register(struct dma_device *device)
362 struct dma_chan* chan;
367 /* validate device routines */
368 BUG_ON(dma_has_cap(DMA_MEMCPY, device->cap_mask) &&
369 !device->device_prep_dma_memcpy);
370 BUG_ON(dma_has_cap(DMA_XOR, device->cap_mask) &&
371 !device->device_prep_dma_xor);
372 BUG_ON(dma_has_cap(DMA_ZERO_SUM, device->cap_mask) &&
373 !device->device_prep_dma_zero_sum);
374 BUG_ON(dma_has_cap(DMA_MEMSET, device->cap_mask) &&
375 !device->device_prep_dma_memset);
376 BUG_ON(dma_has_cap(DMA_INTERRUPT, device->cap_mask) &&
377 !device->device_prep_dma_interrupt);
378 BUG_ON(dma_has_cap(DMA_SLAVE, device->cap_mask) &&
379 !device->device_prep_slave_sg);
380 BUG_ON(dma_has_cap(DMA_SLAVE, device->cap_mask) &&
381 !device->device_terminate_all);
383 BUG_ON(!device->device_alloc_chan_resources);
384 BUG_ON(!device->device_free_chan_resources);
385 BUG_ON(!device->device_is_tx_complete);
386 BUG_ON(!device->device_issue_pending);
387 BUG_ON(!device->dev);
389 init_completion(&device->done);
390 kref_init(&device->refcount);
391 device->dev_id = id++;
393 /* represent channels in sysfs. Probably want devs too */
394 list_for_each_entry(chan, &device->channels, device_node) {
395 chan->local = alloc_percpu(typeof(*chan->local));
396 if (chan->local == NULL)
399 chan->chan_id = chancnt++;
400 chan->dev.class = &dma_devclass;
401 chan->dev.parent = device->dev;
402 dev_set_name(&chan->dev, "dma%dchan%d",
403 device->dev_id, chan->chan_id);
405 rc = device_register(&chan->dev);
408 free_percpu(chan->local);
413 /* One for the channel, one of the class device */
414 kref_get(&device->refcount);
415 kref_get(&device->refcount);
416 kref_init(&chan->refcount);
417 chan->client_count = 0;
419 INIT_RCU_HEAD(&chan->rcu);
422 mutex_lock(&dma_list_mutex);
423 list_add_tail(&device->global_node, &dma_device_list);
424 mutex_unlock(&dma_list_mutex);
426 dma_clients_notify_available();
431 list_for_each_entry(chan, &device->channels, device_node) {
432 if (chan->local == NULL)
434 kref_put(&device->refcount, dma_async_device_cleanup);
435 device_unregister(&chan->dev);
437 free_percpu(chan->local);
441 EXPORT_SYMBOL(dma_async_device_register);
444 * dma_async_device_cleanup - function called when all references are released
445 * @kref: kernel reference object
447 static void dma_async_device_cleanup(struct kref *kref)
449 struct dma_device *device;
451 device = container_of(kref, struct dma_device, refcount);
452 complete(&device->done);
456 * dma_async_device_unregister - unregisters DMA devices
457 * @device: &dma_device
459 void dma_async_device_unregister(struct dma_device *device)
461 struct dma_chan *chan;
463 mutex_lock(&dma_list_mutex);
464 list_del(&device->global_node);
465 mutex_unlock(&dma_list_mutex);
467 list_for_each_entry(chan, &device->channels, device_node) {
468 dma_clients_notify_removed(chan);
469 device_unregister(&chan->dev);
470 dma_chan_release(chan);
473 kref_put(&device->refcount, dma_async_device_cleanup);
474 wait_for_completion(&device->done);
476 EXPORT_SYMBOL(dma_async_device_unregister);
479 * dma_async_memcpy_buf_to_buf - offloaded copy between virtual addresses
480 * @chan: DMA channel to offload copy to
481 * @dest: destination address (virtual)
482 * @src: source address (virtual)
485 * Both @dest and @src must be mappable to a bus address according to the
486 * DMA mapping API rules for streaming mappings.
487 * Both @dest and @src must stay memory resident (kernel memory or locked
491 dma_async_memcpy_buf_to_buf(struct dma_chan *chan, void *dest,
492 void *src, size_t len)
494 struct dma_device *dev = chan->device;
495 struct dma_async_tx_descriptor *tx;
496 dma_addr_t dma_dest, dma_src;
500 dma_src = dma_map_single(dev->dev, src, len, DMA_TO_DEVICE);
501 dma_dest = dma_map_single(dev->dev, dest, len, DMA_FROM_DEVICE);
502 tx = dev->device_prep_dma_memcpy(chan, dma_dest, dma_src, len,
506 dma_unmap_single(dev->dev, dma_src, len, DMA_TO_DEVICE);
507 dma_unmap_single(dev->dev, dma_dest, len, DMA_FROM_DEVICE);
512 cookie = tx->tx_submit(tx);
515 per_cpu_ptr(chan->local, cpu)->bytes_transferred += len;
516 per_cpu_ptr(chan->local, cpu)->memcpy_count++;
521 EXPORT_SYMBOL(dma_async_memcpy_buf_to_buf);
524 * dma_async_memcpy_buf_to_pg - offloaded copy from address to page
525 * @chan: DMA channel to offload copy to
526 * @page: destination page
527 * @offset: offset in page to copy to
528 * @kdata: source address (virtual)
531 * Both @page/@offset and @kdata must be mappable to a bus address according
532 * to the DMA mapping API rules for streaming mappings.
533 * Both @page/@offset and @kdata must stay memory resident (kernel memory or
534 * locked user space pages)
537 dma_async_memcpy_buf_to_pg(struct dma_chan *chan, struct page *page,
538 unsigned int offset, void *kdata, size_t len)
540 struct dma_device *dev = chan->device;
541 struct dma_async_tx_descriptor *tx;
542 dma_addr_t dma_dest, dma_src;
546 dma_src = dma_map_single(dev->dev, kdata, len, DMA_TO_DEVICE);
547 dma_dest = dma_map_page(dev->dev, page, offset, len, DMA_FROM_DEVICE);
548 tx = dev->device_prep_dma_memcpy(chan, dma_dest, dma_src, len,
552 dma_unmap_single(dev->dev, dma_src, len, DMA_TO_DEVICE);
553 dma_unmap_page(dev->dev, dma_dest, len, DMA_FROM_DEVICE);
558 cookie = tx->tx_submit(tx);
561 per_cpu_ptr(chan->local, cpu)->bytes_transferred += len;
562 per_cpu_ptr(chan->local, cpu)->memcpy_count++;
567 EXPORT_SYMBOL(dma_async_memcpy_buf_to_pg);
570 * dma_async_memcpy_pg_to_pg - offloaded copy from page to page
571 * @chan: DMA channel to offload copy to
572 * @dest_pg: destination page
573 * @dest_off: offset in page to copy to
574 * @src_pg: source page
575 * @src_off: offset in page to copy from
578 * Both @dest_page/@dest_off and @src_page/@src_off must be mappable to a bus
579 * address according to the DMA mapping API rules for streaming mappings.
580 * Both @dest_page/@dest_off and @src_page/@src_off must stay memory resident
581 * (kernel memory or locked user space pages).
584 dma_async_memcpy_pg_to_pg(struct dma_chan *chan, struct page *dest_pg,
585 unsigned int dest_off, struct page *src_pg, unsigned int src_off,
588 struct dma_device *dev = chan->device;
589 struct dma_async_tx_descriptor *tx;
590 dma_addr_t dma_dest, dma_src;
594 dma_src = dma_map_page(dev->dev, src_pg, src_off, len, DMA_TO_DEVICE);
595 dma_dest = dma_map_page(dev->dev, dest_pg, dest_off, len,
597 tx = dev->device_prep_dma_memcpy(chan, dma_dest, dma_src, len,
601 dma_unmap_page(dev->dev, dma_src, len, DMA_TO_DEVICE);
602 dma_unmap_page(dev->dev, dma_dest, len, DMA_FROM_DEVICE);
607 cookie = tx->tx_submit(tx);
610 per_cpu_ptr(chan->local, cpu)->bytes_transferred += len;
611 per_cpu_ptr(chan->local, cpu)->memcpy_count++;
616 EXPORT_SYMBOL(dma_async_memcpy_pg_to_pg);
618 void dma_async_tx_descriptor_init(struct dma_async_tx_descriptor *tx,
619 struct dma_chan *chan)
622 spin_lock_init(&tx->lock);
624 EXPORT_SYMBOL(dma_async_tx_descriptor_init);
626 static int __init dma_bus_init(void)
628 mutex_init(&dma_list_mutex);
629 return class_register(&dma_devclass);
631 subsys_initcall(dma_bus_init);