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
187 else if (ack
== DMA_NAK
)
193 enum dma_status
dma_sync_wait(struct dma_chan
*chan
, dma_cookie_t cookie
)
195 enum dma_status status
;
196 unsigned long dma_sync_wait_timeout
= jiffies
+ msecs_to_jiffies(5000);
198 dma_async_issue_pending(chan
);
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");
205 } while (status
== DMA_IN_PROGRESS
);
209 EXPORT_SYMBOL(dma_sync_wait
);
212 * dma_chan_cleanup - release a DMA channel's resources
213 * @kref: kernel reference structure that contains the DMA channel device
215 void dma_chan_cleanup(struct kref
*kref
)
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
);
221 EXPORT_SYMBOL(dma_chan_cleanup
);
223 static void dma_chan_free_rcu(struct rcu_head
*rcu
)
225 struct dma_chan
*chan
= container_of(rcu
, struct dma_chan
, rcu
);
226 int bias
= 0x7FFFFFFF;
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
);
234 static void dma_chan_release(struct dma_chan
*chan
)
236 atomic_add(0x7FFFFFFF, &chan
->refcount
.refcount
);
238 call_rcu(&chan
->rcu
, dma_chan_free_rcu
);
242 * dma_chans_notify_available - broadcast available channels to the clients
244 static void dma_clients_notify_available(void)
246 struct dma_client
*client
;
248 mutex_lock(&dma_list_mutex
);
250 list_for_each_entry(client
, &dma_client_list
, global_node
)
251 dma_client_chan_alloc(client
);
253 mutex_unlock(&dma_list_mutex
);
257 * dma_chans_notify_available - tell the clients that a channel is going away
258 * @chan: channel on its way out
260 static void dma_clients_notify_removed(struct dma_chan
*chan
)
262 struct dma_client
*client
;
263 enum dma_state_client ack
;
265 mutex_lock(&dma_list_mutex
);
267 list_for_each_entry(client
, &dma_client_list
, global_node
) {
268 ack
= client
->event_callback(client
, chan
,
269 DMA_RESOURCE_REMOVED
);
271 /* client was holding resources for this channel so
278 mutex_unlock(&dma_list_mutex
);
282 * dma_async_client_register - register a &dma_client
283 * @client: ptr to a client structure with valid 'event_callback' and 'cap_mask'
285 void dma_async_client_register(struct dma_client
*client
)
287 mutex_lock(&dma_list_mutex
);
288 list_add_tail(&client
->global_node
, &dma_client_list
);
289 mutex_unlock(&dma_list_mutex
);
291 EXPORT_SYMBOL(dma_async_client_register
);
294 * dma_async_client_unregister - unregister a client and free the &dma_client
295 * @client: &dma_client to free
297 * Force frees any allocated DMA channels, frees the &dma_client memory
299 void dma_async_client_unregister(struct dma_client
*client
)
301 struct dma_device
*device
;
302 struct dma_chan
*chan
;
303 enum dma_state_client ack
;
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
);
319 list_del(&client
->global_node
);
320 mutex_unlock(&dma_list_mutex
);
322 EXPORT_SYMBOL(dma_async_client_unregister
);
325 * dma_async_client_chan_request - send all available channels to the
326 * client that satisfy the capability mask
327 * @client - requester
329 void dma_async_client_chan_request(struct dma_client
*client
)
331 mutex_lock(&dma_list_mutex
);
332 dma_client_chan_alloc(client
);
333 mutex_unlock(&dma_list_mutex
);
335 EXPORT_SYMBOL(dma_async_client_chan_request
);
338 * dma_async_device_register - registers DMA devices found
339 * @device: &dma_device
341 int dma_async_device_register(struct dma_device
*device
)
345 struct dma_chan
* chan
;
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
);
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
);
369 init_completion(&device
->done
);
370 kref_init(&device
->refcount
);
371 device
->dev_id
= id
++;
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
)
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
);
385 rc
= class_device_register(&chan
->class_dev
);
388 free_percpu(chan
->local
);
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
);
398 INIT_RCU_HEAD(&chan
->rcu
);
401 mutex_lock(&dma_list_mutex
);
402 list_add_tail(&device
->global_node
, &dma_device_list
);
403 mutex_unlock(&dma_list_mutex
);
405 dma_clients_notify_available();
410 list_for_each_entry(chan
, &device
->channels
, device_node
) {
411 if (chan
->local
== NULL
)
413 kref_put(&device
->refcount
, dma_async_device_cleanup
);
414 class_device_unregister(&chan
->class_dev
);
416 free_percpu(chan
->local
);
420 EXPORT_SYMBOL(dma_async_device_register
);
423 * dma_async_device_cleanup - function called when all references are released
424 * @kref: kernel reference object
426 static void dma_async_device_cleanup(struct kref
*kref
)
428 struct dma_device
*device
;
430 device
= container_of(kref
, struct dma_device
, refcount
);
431 complete(&device
->done
);
435 * dma_async_device_unregister - unregisters DMA devices
436 * @device: &dma_device
438 void dma_async_device_unregister(struct dma_device
*device
)
440 struct dma_chan
*chan
;
442 mutex_lock(&dma_list_mutex
);
443 list_del(&device
->global_node
);
444 mutex_unlock(&dma_list_mutex
);
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
);
452 kref_put(&device
->refcount
, dma_async_device_cleanup
);
453 wait_for_completion(&device
->done
);
455 EXPORT_SYMBOL(dma_async_device_unregister
);
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)
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
470 dma_async_memcpy_buf_to_buf(struct dma_chan
*chan
, void *dest
,
471 void *src
, size_t len
)
473 struct dma_device
*dev
= chan
->device
;
474 struct dma_async_tx_descriptor
*tx
;
479 tx
= dev
->device_prep_dma_memcpy(chan
, len
, 0);
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
);
492 per_cpu_ptr(chan
->local
, cpu
)->bytes_transferred
+= len
;
493 per_cpu_ptr(chan
->local
, cpu
)->memcpy_count
++;
498 EXPORT_SYMBOL(dma_async_memcpy_buf_to_buf
);
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)
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)
514 dma_async_memcpy_buf_to_pg(struct dma_chan
*chan
, struct page
*page
,
515 unsigned int offset
, void *kdata
, size_t len
)
517 struct dma_device
*dev
= chan
->device
;
518 struct dma_async_tx_descriptor
*tx
;
523 tx
= dev
->device_prep_dma_memcpy(chan
, len
, 0);
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
);
536 per_cpu_ptr(chan
->local
, cpu
)->bytes_transferred
+= len
;
537 per_cpu_ptr(chan
->local
, cpu
)->memcpy_count
++;
542 EXPORT_SYMBOL(dma_async_memcpy_buf_to_pg
);
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
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).
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
,
563 struct dma_device
*dev
= chan
->device
;
564 struct dma_async_tx_descriptor
*tx
;
569 tx
= dev
->device_prep_dma_memcpy(chan
, len
, 0);
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
);
582 per_cpu_ptr(chan
->local
, cpu
)->bytes_transferred
+= len
;
583 per_cpu_ptr(chan
->local
, cpu
)->memcpy_count
++;
588 EXPORT_SYMBOL(dma_async_memcpy_pg_to_pg
);
590 void dma_async_tx_descriptor_init(struct dma_async_tx_descriptor
*tx
,
591 struct dma_chan
*chan
)
594 spin_lock_init(&tx
->lock
);
595 INIT_LIST_HEAD(&tx
->depend_node
);
596 INIT_LIST_HEAD(&tx
->depend_list
);
598 EXPORT_SYMBOL(dma_async_tx_descriptor_init
);
600 static int __init
dma_bus_init(void)
602 mutex_init(&dma_list_mutex
);
603 return class_register(&dma_devclass
);
605 subsys_initcall(dma_bus_init
);