2 * core routines for the asynchronous memory transfer/transform api
4 * Copyright © 2006, Intel Corporation.
6 * Dan Williams <dan.j.williams@intel.com>
8 * with architecture considerations by:
9 * Neil Brown <neilb@suse.de>
10 * Jeff Garzik <jeff@garzik.org>
12 * This program is free software; you can redistribute it and/or modify it
13 * under the terms and conditions of the GNU General Public License,
14 * version 2, as published by the Free Software Foundation.
16 * This program is distributed in the hope it will be useful, but WITHOUT
17 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
18 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
21 * You should have received a copy of the GNU General Public License along with
22 * this program; if not, write to the Free Software Foundation, Inc.,
23 * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
26 #include <linux/rculist.h>
27 #include <linux/kernel.h>
28 #include <linux/async_tx.h>
30 #ifdef CONFIG_DMA_ENGINE
31 static enum dma_state_client
32 dma_channel_add_remove(struct dma_client *client,
33 struct dma_chan *chan, enum dma_state state);
35 static struct dma_client async_tx_dma = {
36 .event_callback = dma_channel_add_remove,
37 /* .cap_mask == 0 defaults to all channels */
41 * dma_cap_mask_all - enable iteration over all operation types
43 static dma_cap_mask_t dma_cap_mask_all;
46 * chan_ref_percpu - tracks channel allocations per core/opertion
48 struct chan_ref_percpu {
49 struct dma_chan_ref *ref;
52 static int channel_table_initialized;
53 static struct chan_ref_percpu *channel_table[DMA_TX_TYPE_END];
56 * async_tx_lock - protect modification of async_tx_master_list and serialize
57 * rebalance operations
59 static spinlock_t async_tx_lock;
61 static LIST_HEAD(async_tx_master_list);
63 /* async_tx_issue_pending_all - start all transactions on all channels */
64 void async_tx_issue_pending_all(void)
66 struct dma_chan_ref *ref;
69 list_for_each_entry_rcu(ref, &async_tx_master_list, node)
70 ref->chan->device->device_issue_pending(ref->chan);
73 EXPORT_SYMBOL_GPL(async_tx_issue_pending_all);
75 /* dma_wait_for_async_tx - spin wait for a transcation to complete
76 * @tx: transaction to wait on
79 dma_wait_for_async_tx(struct dma_async_tx_descriptor *tx)
81 enum dma_status status;
82 struct dma_async_tx_descriptor *iter;
83 struct dma_async_tx_descriptor *parent;
88 /* poll through the dependency chain, return when tx is complete */
92 /* find the root of the unsubmitted dependency chain */
94 parent = iter->parent;
101 /* there is a small window for ->parent == NULL and
104 while (iter->cookie == -EBUSY)
107 status = dma_sync_wait(iter->chan, iter->cookie);
108 } while (status == DMA_IN_PROGRESS || (iter != tx));
112 EXPORT_SYMBOL_GPL(dma_wait_for_async_tx);
114 /* async_tx_run_dependencies - helper routine for dma drivers to process
115 * (start) dependent operations on their target channel
116 * @tx: transaction with dependencies
119 async_tx_run_dependencies(struct dma_async_tx_descriptor *tx)
121 struct dma_async_tx_descriptor *next = tx->next;
122 struct dma_chan *chan;
130 /* keep submitting up until a channel switch is detected
131 * in that case we will be called again as a result of
132 * processing the interrupt from async_tx_channel_switch
134 while (next && next->chan == chan) {
135 struct dma_async_tx_descriptor *_next;
137 spin_lock_bh(&next->lock);
140 if (_next && _next->chan == chan)
142 spin_unlock_bh(&next->lock);
144 next->tx_submit(next);
148 chan->device->device_issue_pending(chan);
150 EXPORT_SYMBOL_GPL(async_tx_run_dependencies);
153 free_dma_chan_ref(struct rcu_head *rcu)
155 struct dma_chan_ref *ref;
156 ref = container_of(rcu, struct dma_chan_ref, rcu);
161 init_dma_chan_ref(struct dma_chan_ref *ref, struct dma_chan *chan)
163 INIT_LIST_HEAD(&ref->node);
164 INIT_RCU_HEAD(&ref->rcu);
166 atomic_set(&ref->count, 0);
170 * get_chan_ref_by_cap - returns the nth channel of the given capability
171 * defaults to returning the channel with the desired capability and the
172 * lowest reference count if the index can not be satisfied
173 * @cap: capability to match
174 * @index: nth channel desired, passing -1 has the effect of forcing the
175 * default return value
177 static struct dma_chan_ref *
178 get_chan_ref_by_cap(enum dma_transaction_type cap, int index)
180 struct dma_chan_ref *ret_ref = NULL, *min_ref = NULL, *ref;
183 list_for_each_entry_rcu(ref, &async_tx_master_list, node)
184 if (dma_has_cap(cap, ref->chan->device->cap_mask)) {
187 else if (atomic_read(&ref->count) <
188 atomic_read(&min_ref->count))
202 atomic_inc(&ret_ref->count);
208 * async_tx_rebalance - redistribute the available channels, optimize
209 * for cpu isolation in the SMP case, and opertaion isolation in the
212 static void async_tx_rebalance(void)
214 int cpu, cap, cpu_idx = 0;
217 if (!channel_table_initialized)
220 spin_lock_irqsave(&async_tx_lock, flags);
222 /* undo the last distribution */
223 for_each_dma_cap_mask(cap, dma_cap_mask_all)
224 for_each_possible_cpu(cpu) {
225 struct dma_chan_ref *ref =
226 per_cpu_ptr(channel_table[cap], cpu)->ref;
228 atomic_set(&ref->count, 0);
229 per_cpu_ptr(channel_table[cap], cpu)->ref =
234 for_each_dma_cap_mask(cap, dma_cap_mask_all)
235 for_each_online_cpu(cpu) {
236 struct dma_chan_ref *new;
238 new = get_chan_ref_by_cap(cap, cpu_idx++);
240 new = get_chan_ref_by_cap(cap, -1);
242 per_cpu_ptr(channel_table[cap], cpu)->ref = new;
245 spin_unlock_irqrestore(&async_tx_lock, flags);
248 static enum dma_state_client
249 dma_channel_add_remove(struct dma_client *client,
250 struct dma_chan *chan, enum dma_state state)
252 unsigned long found, flags;
253 struct dma_chan_ref *master_ref, *ref;
254 enum dma_state_client ack = DMA_DUP; /* default: take no action */
257 case DMA_RESOURCE_AVAILABLE:
260 list_for_each_entry_rcu(ref, &async_tx_master_list, node)
261 if (ref->chan == chan) {
267 pr_debug("async_tx: dma resource available [%s]\n",
268 found ? "old" : "new");
275 /* add the channel to the generic management list */
276 master_ref = kmalloc(sizeof(*master_ref), GFP_KERNEL);
278 /* keep a reference until async_tx is unloaded */
280 init_dma_chan_ref(master_ref, chan);
281 spin_lock_irqsave(&async_tx_lock, flags);
282 list_add_tail_rcu(&master_ref->node,
283 &async_tx_master_list);
284 spin_unlock_irqrestore(&async_tx_lock,
287 printk(KERN_WARNING "async_tx: unable to create"
288 " new master entry in response to"
289 " a DMA_RESOURCE_ADDED event"
294 async_tx_rebalance();
296 case DMA_RESOURCE_REMOVED:
298 spin_lock_irqsave(&async_tx_lock, flags);
299 list_for_each_entry(ref, &async_tx_master_list, node)
300 if (ref->chan == chan) {
301 /* permit backing devices to go away */
302 dma_chan_put(ref->chan);
303 list_del_rcu(&ref->node);
304 call_rcu(&ref->rcu, free_dma_chan_ref);
308 spin_unlock_irqrestore(&async_tx_lock, flags);
310 pr_debug("async_tx: dma resource removed [%s]\n",
311 found ? "ours" : "not ours");
318 async_tx_rebalance();
320 case DMA_RESOURCE_SUSPEND:
321 case DMA_RESOURCE_RESUME:
322 printk(KERN_WARNING "async_tx: does not support dma channel"
323 " suspend/resume\n");
335 enum dma_transaction_type cap;
337 spin_lock_init(&async_tx_lock);
338 bitmap_fill(dma_cap_mask_all.bits, DMA_TX_TYPE_END);
340 /* an interrupt will never be an explicit operation type.
341 * clearing this bit prevents allocation to a slot in 'channel_table'
343 clear_bit(DMA_INTERRUPT, dma_cap_mask_all.bits);
345 for_each_dma_cap_mask(cap, dma_cap_mask_all) {
346 channel_table[cap] = alloc_percpu(struct chan_ref_percpu);
347 if (!channel_table[cap])
351 channel_table_initialized = 1;
352 dma_async_client_register(&async_tx_dma);
353 dma_async_client_chan_request(&async_tx_dma);
355 printk(KERN_INFO "async_tx: api initialized (async)\n");
359 printk(KERN_ERR "async_tx: initialization failure\n");
362 free_percpu(channel_table[cap]);
367 static void __exit async_tx_exit(void)
369 enum dma_transaction_type cap;
371 channel_table_initialized = 0;
373 for_each_dma_cap_mask(cap, dma_cap_mask_all)
374 if (channel_table[cap])
375 free_percpu(channel_table[cap]);
377 dma_async_client_unregister(&async_tx_dma);
381 * __async_tx_find_channel - find a channel to carry out the operation or let
382 * the transaction execute synchronously
383 * @depend_tx: transaction dependency
384 * @tx_type: transaction type
387 __async_tx_find_channel(struct dma_async_tx_descriptor *depend_tx,
388 enum dma_transaction_type tx_type)
390 /* see if we can keep the chain on one channel */
392 dma_has_cap(tx_type, depend_tx->chan->device->cap_mask))
393 return depend_tx->chan;
394 else if (likely(channel_table_initialized)) {
395 struct dma_chan_ref *ref;
397 ref = per_cpu_ptr(channel_table[tx_type], cpu)->ref;
399 return ref ? ref->chan : NULL;
403 EXPORT_SYMBOL_GPL(__async_tx_find_channel);
405 static int __init async_tx_init(void)
407 printk(KERN_INFO "async_tx: api initialized (sync-only)\n");
411 static void __exit async_tx_exit(void)
419 * async_tx_channel_switch - queue an interrupt descriptor with a dependency
421 * @depend_tx: the operation that must finish before the new operation runs
422 * @tx: the new operation
425 async_tx_channel_switch(struct dma_async_tx_descriptor *depend_tx,
426 struct dma_async_tx_descriptor *tx)
428 struct dma_chan *chan;
429 struct dma_device *device;
430 struct dma_async_tx_descriptor *intr_tx = (void *) ~0;
432 /* first check to see if we can still append to depend_tx */
433 spin_lock_bh(&depend_tx->lock);
434 if (depend_tx->parent && depend_tx->chan == tx->chan) {
435 tx->parent = depend_tx;
436 depend_tx->next = tx;
439 spin_unlock_bh(&depend_tx->lock);
444 chan = depend_tx->chan;
445 device = chan->device;
447 /* see if we can schedule an interrupt
448 * otherwise poll for completion
450 if (dma_has_cap(DMA_INTERRUPT, device->cap_mask))
451 intr_tx = device->device_prep_dma_interrupt(chan, 0);
456 intr_tx->callback = NULL;
457 intr_tx->callback_param = NULL;
458 tx->parent = intr_tx;
459 /* safe to set ->next outside the lock since we know we are
464 /* check if we need to append */
465 spin_lock_bh(&depend_tx->lock);
466 if (depend_tx->parent) {
467 intr_tx->parent = depend_tx;
468 depend_tx->next = intr_tx;
469 async_tx_ack(intr_tx);
472 spin_unlock_bh(&depend_tx->lock);
475 intr_tx->parent = NULL;
476 intr_tx->tx_submit(intr_tx);
477 async_tx_ack(intr_tx);
480 if (dma_wait_for_async_tx(depend_tx) == DMA_ERROR)
481 panic("%s: DMA_ERROR waiting for depend_tx\n",
489 * submit_disposition - while holding depend_tx->lock we must avoid submitting
490 * new operations to prevent a circular locking dependency with
491 * drivers that already hold a channel lock when calling
492 * async_tx_run_dependencies.
493 * @ASYNC_TX_SUBMITTED: we were able to append the new operation under the lock
494 * @ASYNC_TX_CHANNEL_SWITCH: when the lock is dropped schedule a channel switch
495 * @ASYNC_TX_DIRECT_SUBMIT: when the lock is dropped submit directly
497 enum submit_disposition {
499 ASYNC_TX_CHANNEL_SWITCH,
500 ASYNC_TX_DIRECT_SUBMIT,
504 async_tx_submit(struct dma_chan *chan, struct dma_async_tx_descriptor *tx,
505 enum async_tx_flags flags, struct dma_async_tx_descriptor *depend_tx,
506 dma_async_tx_callback cb_fn, void *cb_param)
508 tx->callback = cb_fn;
509 tx->callback_param = cb_param;
512 enum submit_disposition s;
514 /* sanity check the dependency chain:
515 * 1/ if ack is already set then we cannot be sure
516 * we are referring to the correct operation
517 * 2/ dependencies are 1:1 i.e. two transactions can
518 * not depend on the same parent
520 BUG_ON(async_tx_test_ack(depend_tx) || depend_tx->next ||
523 /* the lock prevents async_tx_run_dependencies from missing
524 * the setting of ->next when ->parent != NULL
526 spin_lock_bh(&depend_tx->lock);
527 if (depend_tx->parent) {
528 /* we have a parent so we can not submit directly
529 * if we are staying on the same channel: append
530 * else: channel switch
532 if (depend_tx->chan == chan) {
533 tx->parent = depend_tx;
534 depend_tx->next = tx;
535 s = ASYNC_TX_SUBMITTED;
537 s = ASYNC_TX_CHANNEL_SWITCH;
539 /* we do not have a parent so we may be able to submit
540 * directly if we are staying on the same channel
542 if (depend_tx->chan == chan)
543 s = ASYNC_TX_DIRECT_SUBMIT;
545 s = ASYNC_TX_CHANNEL_SWITCH;
547 spin_unlock_bh(&depend_tx->lock);
550 case ASYNC_TX_SUBMITTED:
552 case ASYNC_TX_CHANNEL_SWITCH:
553 async_tx_channel_switch(depend_tx, tx);
555 case ASYNC_TX_DIRECT_SUBMIT:
565 if (flags & ASYNC_TX_ACK)
568 if (depend_tx && (flags & ASYNC_TX_DEP_ACK))
569 async_tx_ack(depend_tx);
571 EXPORT_SYMBOL_GPL(async_tx_submit);
574 * async_trigger_callback - schedules the callback function to be run after
575 * any dependent operations have been completed.
576 * @flags: ASYNC_TX_ACK, ASYNC_TX_DEP_ACK
577 * @depend_tx: 'callback' requires the completion of this transaction
578 * @cb_fn: function to call after depend_tx completes
579 * @cb_param: parameter to pass to the callback routine
581 struct dma_async_tx_descriptor *
582 async_trigger_callback(enum async_tx_flags flags,
583 struct dma_async_tx_descriptor *depend_tx,
584 dma_async_tx_callback cb_fn, void *cb_param)
586 struct dma_chan *chan;
587 struct dma_device *device;
588 struct dma_async_tx_descriptor *tx;
591 chan = depend_tx->chan;
592 device = chan->device;
594 /* see if we can schedule an interrupt
595 * otherwise poll for completion
597 if (device && !dma_has_cap(DMA_INTERRUPT, device->cap_mask))
600 tx = device ? device->device_prep_dma_interrupt(chan, 0) : NULL;
605 pr_debug("%s: (async)\n", __func__);
607 async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param);
609 pr_debug("%s: (sync)\n", __func__);
611 /* wait for any prerequisite operations */
612 async_tx_quiesce(&depend_tx);
614 async_tx_sync_epilog(cb_fn, cb_param);
619 EXPORT_SYMBOL_GPL(async_trigger_callback);
622 * async_tx_quiesce - ensure tx is complete and freeable upon return
623 * @tx - transaction to quiesce
625 void async_tx_quiesce(struct dma_async_tx_descriptor **tx)
628 /* if ack is already set then we cannot be sure
629 * we are referring to the correct operation
631 BUG_ON(async_tx_test_ack(*tx));
632 if (dma_wait_for_async_tx(*tx) == DMA_ERROR)
633 panic("DMA_ERROR waiting for transaction\n");
638 EXPORT_SYMBOL_GPL(async_tx_quiesce);
640 module_init(async_tx_init);
641 module_exit(async_tx_exit);
643 MODULE_AUTHOR("Intel Corporation");
644 MODULE_DESCRIPTION("Asynchronous Bulk Memory Transactions API");
645 MODULE_LICENSE("GPL");