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);
77 static long dmaengine_ref_count;
79 /* --- sysfs implementation --- */
81 static ssize_t show_memcpy_count(struct device *dev, struct device_attribute *attr, char *buf)
83 struct dma_chan *chan = to_dma_chan(dev);
84 unsigned long count = 0;
87 for_each_possible_cpu(i)
88 count += per_cpu_ptr(chan->local, i)->memcpy_count;
90 return sprintf(buf, "%lu\n", count);
93 static ssize_t show_bytes_transferred(struct device *dev, struct device_attribute *attr,
96 struct dma_chan *chan = to_dma_chan(dev);
97 unsigned long count = 0;
100 for_each_possible_cpu(i)
101 count += per_cpu_ptr(chan->local, i)->bytes_transferred;
103 return sprintf(buf, "%lu\n", count);
106 static ssize_t show_in_use(struct device *dev, struct device_attribute *attr, char *buf)
108 struct dma_chan *chan = to_dma_chan(dev);
110 return sprintf(buf, "%d\n", chan->client_count);
113 static struct device_attribute dma_attrs[] = {
114 __ATTR(memcpy_count, S_IRUGO, show_memcpy_count, NULL),
115 __ATTR(bytes_transferred, S_IRUGO, show_bytes_transferred, NULL),
116 __ATTR(in_use, S_IRUGO, show_in_use, NULL),
120 static void dma_async_device_cleanup(struct kref *kref);
122 static void dma_dev_release(struct device *dev)
124 struct dma_chan *chan = to_dma_chan(dev);
125 kref_put(&chan->device->refcount, dma_async_device_cleanup);
128 static struct class dma_devclass = {
130 .dev_attrs = dma_attrs,
131 .dev_release = dma_dev_release,
134 /* --- client and device registration --- */
136 #define dma_chan_satisfies_mask(chan, mask) \
137 __dma_chan_satisfies_mask((chan), &(mask))
139 __dma_chan_satisfies_mask(struct dma_chan *chan, dma_cap_mask_t *want)
143 bitmap_and(has.bits, want->bits, chan->device->cap_mask.bits,
145 return bitmap_equal(want->bits, has.bits, DMA_TX_TYPE_END);
148 static struct module *dma_chan_to_owner(struct dma_chan *chan)
150 return chan->device->dev->driver->owner;
154 * balance_ref_count - catch up the channel reference count
155 * @chan - channel to balance ->client_count versus dmaengine_ref_count
157 * balance_ref_count must be called under dma_list_mutex
159 static void balance_ref_count(struct dma_chan *chan)
161 struct module *owner = dma_chan_to_owner(chan);
163 while (chan->client_count < dmaengine_ref_count) {
165 chan->client_count++;
170 * dma_chan_get - try to grab a dma channel's parent driver module
171 * @chan - channel to grab
173 * Must be called under dma_list_mutex
175 static int dma_chan_get(struct dma_chan *chan)
178 struct module *owner = dma_chan_to_owner(chan);
180 if (chan->client_count) {
183 } else if (try_module_get(owner))
187 chan->client_count++;
189 /* allocate upon first client reference */
190 if (chan->client_count == 1 && err == 0) {
191 int desc_cnt = chan->device->device_alloc_chan_resources(chan, NULL);
195 chan->client_count = 0;
198 balance_ref_count(chan);
205 * dma_chan_put - drop a reference to a dma channel's parent driver module
206 * @chan - channel to release
208 * Must be called under dma_list_mutex
210 static void dma_chan_put(struct dma_chan *chan)
212 if (!chan->client_count)
213 return; /* this channel failed alloc_chan_resources */
214 chan->client_count--;
215 module_put(dma_chan_to_owner(chan));
216 if (chan->client_count == 0)
217 chan->device->device_free_chan_resources(chan);
221 * dma_client_chan_alloc - try to allocate channels to a client
222 * @client: &dma_client
224 * Called with dma_list_mutex held.
226 static void dma_client_chan_alloc(struct dma_client *client)
228 struct dma_device *device;
229 struct dma_chan *chan;
230 enum dma_state_client ack;
233 list_for_each_entry(device, &dma_device_list, global_node) {
234 /* Does the client require a specific DMA controller? */
235 if (client->slave && client->slave->dma_dev
236 && client->slave->dma_dev != device->dev)
239 list_for_each_entry(chan, &device->channels, device_node) {
240 if (!dma_chan_satisfies_mask(chan, client->cap_mask))
242 if (!chan->client_count)
244 ack = client->event_callback(client, chan,
245 DMA_RESOURCE_AVAILABLE);
247 /* we are done once this client rejects
248 * an available resource
256 enum dma_status dma_sync_wait(struct dma_chan *chan, dma_cookie_t cookie)
258 enum dma_status status;
259 unsigned long dma_sync_wait_timeout = jiffies + msecs_to_jiffies(5000);
261 dma_async_issue_pending(chan);
263 status = dma_async_is_tx_complete(chan, cookie, NULL, NULL);
264 if (time_after_eq(jiffies, dma_sync_wait_timeout)) {
265 printk(KERN_ERR "dma_sync_wait_timeout!\n");
268 } while (status == DMA_IN_PROGRESS);
272 EXPORT_SYMBOL(dma_sync_wait);
275 * dma_chan_cleanup - release a DMA channel's resources
276 * @kref: kernel reference structure that contains the DMA channel device
278 void dma_chan_cleanup(struct kref *kref)
280 struct dma_chan *chan = container_of(kref, struct dma_chan, refcount);
281 kref_put(&chan->device->refcount, dma_async_device_cleanup);
283 EXPORT_SYMBOL(dma_chan_cleanup);
285 static void dma_chan_free_rcu(struct rcu_head *rcu)
287 struct dma_chan *chan = container_of(rcu, struct dma_chan, rcu);
289 kref_put(&chan->refcount, dma_chan_cleanup);
292 static void dma_chan_release(struct dma_chan *chan)
294 call_rcu(&chan->rcu, dma_chan_free_rcu);
298 * dma_chans_notify_available - broadcast available channels to the clients
300 static void dma_clients_notify_available(void)
302 struct dma_client *client;
304 mutex_lock(&dma_list_mutex);
306 list_for_each_entry(client, &dma_client_list, global_node)
307 dma_client_chan_alloc(client);
309 mutex_unlock(&dma_list_mutex);
313 * dma_async_client_register - register a &dma_client
314 * @client: ptr to a client structure with valid 'event_callback' and 'cap_mask'
316 void dma_async_client_register(struct dma_client *client)
318 struct dma_device *device, *_d;
319 struct dma_chan *chan;
322 /* validate client data */
323 BUG_ON(dma_has_cap(DMA_SLAVE, client->cap_mask) &&
326 mutex_lock(&dma_list_mutex);
327 dmaengine_ref_count++;
329 /* try to grab channels */
330 list_for_each_entry_safe(device, _d, &dma_device_list, global_node)
331 list_for_each_entry(chan, &device->channels, device_node) {
332 err = dma_chan_get(chan);
333 if (err == -ENODEV) {
334 /* module removed before we could use it */
335 list_del_init(&device->global_node);
338 pr_err("dmaengine: failed to get %s: (%d)\n",
339 dev_name(&chan->dev), err);
343 list_add_tail(&client->global_node, &dma_client_list);
344 mutex_unlock(&dma_list_mutex);
346 EXPORT_SYMBOL(dma_async_client_register);
349 * dma_async_client_unregister - unregister a client and free the &dma_client
350 * @client: &dma_client to free
352 * Force frees any allocated DMA channels, frees the &dma_client memory
354 void dma_async_client_unregister(struct dma_client *client)
356 struct dma_device *device;
357 struct dma_chan *chan;
362 mutex_lock(&dma_list_mutex);
363 dmaengine_ref_count--;
364 BUG_ON(dmaengine_ref_count < 0);
365 /* drop channel references */
366 list_for_each_entry(device, &dma_device_list, global_node)
367 list_for_each_entry(chan, &device->channels, device_node)
370 list_del(&client->global_node);
371 mutex_unlock(&dma_list_mutex);
373 EXPORT_SYMBOL(dma_async_client_unregister);
376 * dma_async_client_chan_request - send all available channels to the
377 * client that satisfy the capability mask
378 * @client - requester
380 void dma_async_client_chan_request(struct dma_client *client)
382 mutex_lock(&dma_list_mutex);
383 dma_client_chan_alloc(client);
384 mutex_unlock(&dma_list_mutex);
386 EXPORT_SYMBOL(dma_async_client_chan_request);
389 * dma_async_device_register - registers DMA devices found
390 * @device: &dma_device
392 int dma_async_device_register(struct dma_device *device)
396 struct dma_chan* chan;
401 /* validate device routines */
402 BUG_ON(dma_has_cap(DMA_MEMCPY, device->cap_mask) &&
403 !device->device_prep_dma_memcpy);
404 BUG_ON(dma_has_cap(DMA_XOR, device->cap_mask) &&
405 !device->device_prep_dma_xor);
406 BUG_ON(dma_has_cap(DMA_ZERO_SUM, device->cap_mask) &&
407 !device->device_prep_dma_zero_sum);
408 BUG_ON(dma_has_cap(DMA_MEMSET, device->cap_mask) &&
409 !device->device_prep_dma_memset);
410 BUG_ON(dma_has_cap(DMA_INTERRUPT, device->cap_mask) &&
411 !device->device_prep_dma_interrupt);
412 BUG_ON(dma_has_cap(DMA_SLAVE, device->cap_mask) &&
413 !device->device_prep_slave_sg);
414 BUG_ON(dma_has_cap(DMA_SLAVE, device->cap_mask) &&
415 !device->device_terminate_all);
417 BUG_ON(!device->device_alloc_chan_resources);
418 BUG_ON(!device->device_free_chan_resources);
419 BUG_ON(!device->device_is_tx_complete);
420 BUG_ON(!device->device_issue_pending);
421 BUG_ON(!device->dev);
423 init_completion(&device->done);
424 kref_init(&device->refcount);
426 mutex_lock(&dma_list_mutex);
427 device->dev_id = id++;
428 mutex_unlock(&dma_list_mutex);
430 /* represent channels in sysfs. Probably want devs too */
431 list_for_each_entry(chan, &device->channels, device_node) {
432 chan->local = alloc_percpu(typeof(*chan->local));
433 if (chan->local == NULL)
436 chan->chan_id = chancnt++;
437 chan->dev.class = &dma_devclass;
438 chan->dev.parent = device->dev;
439 dev_set_name(&chan->dev, "dma%dchan%d",
440 device->dev_id, chan->chan_id);
442 rc = device_register(&chan->dev);
445 free_percpu(chan->local);
450 /* One for the channel, one of the class device */
451 kref_get(&device->refcount);
452 kref_get(&device->refcount);
453 kref_init(&chan->refcount);
454 chan->client_count = 0;
456 INIT_RCU_HEAD(&chan->rcu);
459 mutex_lock(&dma_list_mutex);
460 if (dmaengine_ref_count)
461 list_for_each_entry(chan, &device->channels, device_node) {
462 /* if clients are already waiting for channels we need
463 * to take references on their behalf
465 if (dma_chan_get(chan) == -ENODEV) {
466 /* note we can only get here for the first
467 * channel as the remaining channels are
468 * guaranteed to get a reference
471 mutex_unlock(&dma_list_mutex);
475 list_add_tail(&device->global_node, &dma_device_list);
476 mutex_unlock(&dma_list_mutex);
478 dma_clients_notify_available();
483 list_for_each_entry(chan, &device->channels, device_node) {
484 if (chan->local == NULL)
486 kref_put(&device->refcount, dma_async_device_cleanup);
487 device_unregister(&chan->dev);
489 free_percpu(chan->local);
493 EXPORT_SYMBOL(dma_async_device_register);
496 * dma_async_device_cleanup - function called when all references are released
497 * @kref: kernel reference object
499 static void dma_async_device_cleanup(struct kref *kref)
501 struct dma_device *device;
503 device = container_of(kref, struct dma_device, refcount);
504 complete(&device->done);
508 * dma_async_device_unregister - unregister a DMA device
509 * @device: &dma_device
511 void dma_async_device_unregister(struct dma_device *device)
513 struct dma_chan *chan;
515 mutex_lock(&dma_list_mutex);
516 list_del(&device->global_node);
517 mutex_unlock(&dma_list_mutex);
519 list_for_each_entry(chan, &device->channels, device_node) {
520 WARN_ONCE(chan->client_count,
521 "%s called while %d clients hold a reference\n",
522 __func__, chan->client_count);
523 device_unregister(&chan->dev);
524 dma_chan_release(chan);
527 kref_put(&device->refcount, dma_async_device_cleanup);
528 wait_for_completion(&device->done);
530 EXPORT_SYMBOL(dma_async_device_unregister);
533 * dma_async_memcpy_buf_to_buf - offloaded copy between virtual addresses
534 * @chan: DMA channel to offload copy to
535 * @dest: destination address (virtual)
536 * @src: source address (virtual)
539 * Both @dest and @src must be mappable to a bus address according to the
540 * DMA mapping API rules for streaming mappings.
541 * Both @dest and @src must stay memory resident (kernel memory or locked
545 dma_async_memcpy_buf_to_buf(struct dma_chan *chan, void *dest,
546 void *src, size_t len)
548 struct dma_device *dev = chan->device;
549 struct dma_async_tx_descriptor *tx;
550 dma_addr_t dma_dest, dma_src;
554 dma_src = dma_map_single(dev->dev, src, len, DMA_TO_DEVICE);
555 dma_dest = dma_map_single(dev->dev, dest, len, DMA_FROM_DEVICE);
556 tx = dev->device_prep_dma_memcpy(chan, dma_dest, dma_src, len,
560 dma_unmap_single(dev->dev, dma_src, len, DMA_TO_DEVICE);
561 dma_unmap_single(dev->dev, dma_dest, len, DMA_FROM_DEVICE);
566 cookie = tx->tx_submit(tx);
569 per_cpu_ptr(chan->local, cpu)->bytes_transferred += len;
570 per_cpu_ptr(chan->local, cpu)->memcpy_count++;
575 EXPORT_SYMBOL(dma_async_memcpy_buf_to_buf);
578 * dma_async_memcpy_buf_to_pg - offloaded copy from address to page
579 * @chan: DMA channel to offload copy to
580 * @page: destination page
581 * @offset: offset in page to copy to
582 * @kdata: source address (virtual)
585 * Both @page/@offset and @kdata must be mappable to a bus address according
586 * to the DMA mapping API rules for streaming mappings.
587 * Both @page/@offset and @kdata must stay memory resident (kernel memory or
588 * locked user space pages)
591 dma_async_memcpy_buf_to_pg(struct dma_chan *chan, struct page *page,
592 unsigned int offset, void *kdata, size_t len)
594 struct dma_device *dev = chan->device;
595 struct dma_async_tx_descriptor *tx;
596 dma_addr_t dma_dest, dma_src;
600 dma_src = dma_map_single(dev->dev, kdata, len, DMA_TO_DEVICE);
601 dma_dest = dma_map_page(dev->dev, page, offset, len, DMA_FROM_DEVICE);
602 tx = dev->device_prep_dma_memcpy(chan, dma_dest, dma_src, len,
606 dma_unmap_single(dev->dev, dma_src, len, DMA_TO_DEVICE);
607 dma_unmap_page(dev->dev, dma_dest, len, DMA_FROM_DEVICE);
612 cookie = tx->tx_submit(tx);
615 per_cpu_ptr(chan->local, cpu)->bytes_transferred += len;
616 per_cpu_ptr(chan->local, cpu)->memcpy_count++;
621 EXPORT_SYMBOL(dma_async_memcpy_buf_to_pg);
624 * dma_async_memcpy_pg_to_pg - offloaded copy from page to page
625 * @chan: DMA channel to offload copy to
626 * @dest_pg: destination page
627 * @dest_off: offset in page to copy to
628 * @src_pg: source page
629 * @src_off: offset in page to copy from
632 * Both @dest_page/@dest_off and @src_page/@src_off must be mappable to a bus
633 * address according to the DMA mapping API rules for streaming mappings.
634 * Both @dest_page/@dest_off and @src_page/@src_off must stay memory resident
635 * (kernel memory or locked user space pages).
638 dma_async_memcpy_pg_to_pg(struct dma_chan *chan, struct page *dest_pg,
639 unsigned int dest_off, struct page *src_pg, unsigned int src_off,
642 struct dma_device *dev = chan->device;
643 struct dma_async_tx_descriptor *tx;
644 dma_addr_t dma_dest, dma_src;
648 dma_src = dma_map_page(dev->dev, src_pg, src_off, len, DMA_TO_DEVICE);
649 dma_dest = dma_map_page(dev->dev, dest_pg, dest_off, len,
651 tx = dev->device_prep_dma_memcpy(chan, dma_dest, dma_src, len,
655 dma_unmap_page(dev->dev, dma_src, len, DMA_TO_DEVICE);
656 dma_unmap_page(dev->dev, dma_dest, len, DMA_FROM_DEVICE);
661 cookie = tx->tx_submit(tx);
664 per_cpu_ptr(chan->local, cpu)->bytes_transferred += len;
665 per_cpu_ptr(chan->local, cpu)->memcpy_count++;
670 EXPORT_SYMBOL(dma_async_memcpy_pg_to_pg);
672 void dma_async_tx_descriptor_init(struct dma_async_tx_descriptor *tx,
673 struct dma_chan *chan)
676 spin_lock_init(&tx->lock);
678 EXPORT_SYMBOL(dma_async_tx_descriptor_init);
680 /* dma_wait_for_async_tx - spin wait for a transaction to complete
681 * @tx: in-flight transaction to wait on
683 * This routine assumes that tx was obtained from a call to async_memcpy,
684 * async_xor, async_memset, etc which ensures that tx is "in-flight" (prepped
685 * and submitted). Walking the parent chain is only meant to cover for DMA
686 * drivers that do not implement the DMA_INTERRUPT capability and may race with
687 * the driver's descriptor cleanup routine.
690 dma_wait_for_async_tx(struct dma_async_tx_descriptor *tx)
692 enum dma_status status;
693 struct dma_async_tx_descriptor *iter;
694 struct dma_async_tx_descriptor *parent;
699 WARN_ONCE(tx->parent, "%s: speculatively walking dependency chain for"
700 " %s\n", __func__, dev_name(&tx->chan->dev));
702 /* poll through the dependency chain, return when tx is complete */
706 /* find the root of the unsubmitted dependency chain */
708 parent = iter->parent;
715 /* there is a small window for ->parent == NULL and
718 while (iter->cookie == -EBUSY)
721 status = dma_sync_wait(iter->chan, iter->cookie);
722 } while (status == DMA_IN_PROGRESS || (iter != tx));
726 EXPORT_SYMBOL_GPL(dma_wait_for_async_tx);
728 /* dma_run_dependencies - helper routine for dma drivers to process
729 * (start) dependent operations on their target channel
730 * @tx: transaction with dependencies
732 void dma_run_dependencies(struct dma_async_tx_descriptor *tx)
734 struct dma_async_tx_descriptor *dep = tx->next;
735 struct dma_async_tx_descriptor *dep_next;
736 struct dma_chan *chan;
743 /* keep submitting up until a channel switch is detected
744 * in that case we will be called again as a result of
745 * processing the interrupt from async_tx_channel_switch
747 for (; dep; dep = dep_next) {
748 spin_lock_bh(&dep->lock);
750 dep_next = dep->next;
751 if (dep_next && dep_next->chan == chan)
752 dep->next = NULL; /* ->next will be submitted */
754 dep_next = NULL; /* submit current dep and terminate */
755 spin_unlock_bh(&dep->lock);
760 chan->device->device_issue_pending(chan);
762 EXPORT_SYMBOL_GPL(dma_run_dependencies);
764 static int __init dma_bus_init(void)
766 mutex_init(&dma_list_mutex);
767 return class_register(&dma_devclass);
769 subsys_initcall(dma_bus_init);