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/kernel.h>
27 #include <linux/async_tx.h>
29 #ifdef CONFIG_DMA_ENGINE
30 static enum dma_state_client
31 dma_channel_add_remove(struct dma_client *client,
32 struct dma_chan *chan, enum dma_state state);
34 static struct dma_client async_tx_dma = {
35 .event_callback = dma_channel_add_remove,
36 /* .cap_mask == 0 defaults to all channels */
40 * dma_cap_mask_all - enable iteration over all operation types
42 static dma_cap_mask_t dma_cap_mask_all;
45 * chan_ref_percpu - tracks channel allocations per core/opertion
47 struct chan_ref_percpu {
48 struct dma_chan_ref *ref;
51 static int channel_table_initialized;
52 static struct chan_ref_percpu *channel_table[DMA_TX_TYPE_END];
55 * async_tx_lock - protect modification of async_tx_master_list and serialize
56 * rebalance operations
58 static spinlock_t async_tx_lock;
60 static struct list_head
61 async_tx_master_list = LIST_HEAD_INIT(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;
87 /* poll through the dependency chain, return when tx is complete */
90 while (iter->cookie == -EBUSY)
93 status = dma_sync_wait(iter->chan, iter->cookie);
94 } while (status == DMA_IN_PROGRESS || (iter != tx));
98 EXPORT_SYMBOL_GPL(dma_wait_for_async_tx);
100 /* async_tx_run_dependencies - helper routine for dma drivers to process
101 * (start) dependent operations on their target channel
102 * @tx: transaction with dependencies
105 async_tx_run_dependencies(struct dma_async_tx_descriptor *tx)
107 struct dma_async_tx_descriptor *dep_tx, *_dep_tx;
108 struct dma_device *dev;
109 struct dma_chan *chan;
111 list_for_each_entry_safe(dep_tx, _dep_tx, &tx->depend_list,
115 /* we can't depend on ourselves */
116 BUG_ON(chan == tx->chan);
117 list_del(&dep_tx->depend_node);
118 tx->tx_submit(dep_tx);
120 /* we need to poke the engine as client code does not
121 * know about dependency submission events
123 dev->device_issue_pending(chan);
126 EXPORT_SYMBOL_GPL(async_tx_run_dependencies);
129 free_dma_chan_ref(struct rcu_head *rcu)
131 struct dma_chan_ref *ref;
132 ref = container_of(rcu, struct dma_chan_ref, rcu);
137 init_dma_chan_ref(struct dma_chan_ref *ref, struct dma_chan *chan)
139 INIT_LIST_HEAD(&ref->node);
140 INIT_RCU_HEAD(&ref->rcu);
142 atomic_set(&ref->count, 0);
146 * get_chan_ref_by_cap - returns the nth channel of the given capability
147 * defaults to returning the channel with the desired capability and the
148 * lowest reference count if the index can not be satisfied
149 * @cap: capability to match
150 * @index: nth channel desired, passing -1 has the effect of forcing the
151 * default return value
153 static struct dma_chan_ref *
154 get_chan_ref_by_cap(enum dma_transaction_type cap, int index)
156 struct dma_chan_ref *ret_ref = NULL, *min_ref = NULL, *ref;
159 list_for_each_entry_rcu(ref, &async_tx_master_list, node)
160 if (dma_has_cap(cap, ref->chan->device->cap_mask)) {
163 else if (atomic_read(&ref->count) <
164 atomic_read(&min_ref->count))
178 atomic_inc(&ret_ref->count);
184 * async_tx_rebalance - redistribute the available channels, optimize
185 * for cpu isolation in the SMP case, and opertaion isolation in the
188 static void async_tx_rebalance(void)
190 int cpu, cap, cpu_idx = 0;
193 if (!channel_table_initialized)
196 spin_lock_irqsave(&async_tx_lock, flags);
198 /* undo the last distribution */
199 for_each_dma_cap_mask(cap, dma_cap_mask_all)
200 for_each_possible_cpu(cpu) {
201 struct dma_chan_ref *ref =
202 per_cpu_ptr(channel_table[cap], cpu)->ref;
204 atomic_set(&ref->count, 0);
205 per_cpu_ptr(channel_table[cap], cpu)->ref =
210 for_each_dma_cap_mask(cap, dma_cap_mask_all)
211 for_each_online_cpu(cpu) {
212 struct dma_chan_ref *new;
214 new = get_chan_ref_by_cap(cap, cpu_idx++);
216 new = get_chan_ref_by_cap(cap, -1);
218 per_cpu_ptr(channel_table[cap], cpu)->ref = new;
221 spin_unlock_irqrestore(&async_tx_lock, flags);
224 static enum dma_state_client
225 dma_channel_add_remove(struct dma_client *client,
226 struct dma_chan *chan, enum dma_state state)
228 unsigned long found, flags;
229 struct dma_chan_ref *master_ref, *ref;
230 enum dma_state_client ack = DMA_DUP; /* default: take no action */
233 case DMA_RESOURCE_AVAILABLE:
236 list_for_each_entry_rcu(ref, &async_tx_master_list, node)
237 if (ref->chan == chan) {
243 pr_debug("async_tx: dma resource available [%s]\n",
244 found ? "old" : "new");
251 /* add the channel to the generic management list */
252 master_ref = kmalloc(sizeof(*master_ref), GFP_KERNEL);
254 /* keep a reference until async_tx is unloaded */
256 init_dma_chan_ref(master_ref, chan);
257 spin_lock_irqsave(&async_tx_lock, flags);
258 list_add_tail_rcu(&master_ref->node,
259 &async_tx_master_list);
260 spin_unlock_irqrestore(&async_tx_lock,
263 printk(KERN_WARNING "async_tx: unable to create"
264 " new master entry in response to"
265 " a DMA_RESOURCE_ADDED event"
270 async_tx_rebalance();
272 case DMA_RESOURCE_REMOVED:
274 spin_lock_irqsave(&async_tx_lock, flags);
275 list_for_each_entry_rcu(ref, &async_tx_master_list, node)
276 if (ref->chan == chan) {
277 /* permit backing devices to go away */
278 dma_chan_put(ref->chan);
279 list_del_rcu(&ref->node);
280 call_rcu(&ref->rcu, free_dma_chan_ref);
284 spin_unlock_irqrestore(&async_tx_lock, flags);
286 pr_debug("async_tx: dma resource removed [%s]\n",
287 found ? "ours" : "not ours");
294 async_tx_rebalance();
296 case DMA_RESOURCE_SUSPEND:
297 case DMA_RESOURCE_RESUME:
298 printk(KERN_WARNING "async_tx: does not support dma channel"
299 " suspend/resume\n");
311 enum dma_transaction_type cap;
313 spin_lock_init(&async_tx_lock);
314 bitmap_fill(dma_cap_mask_all.bits, DMA_TX_TYPE_END);
316 /* an interrupt will never be an explicit operation type.
317 * clearing this bit prevents allocation to a slot in 'channel_table'
319 clear_bit(DMA_INTERRUPT, dma_cap_mask_all.bits);
321 for_each_dma_cap_mask(cap, dma_cap_mask_all) {
322 channel_table[cap] = alloc_percpu(struct chan_ref_percpu);
323 if (!channel_table[cap])
327 channel_table_initialized = 1;
328 dma_async_client_register(&async_tx_dma);
329 dma_async_client_chan_request(&async_tx_dma);
331 printk(KERN_INFO "async_tx: api initialized (async)\n");
335 printk(KERN_ERR "async_tx: initialization failure\n");
338 free_percpu(channel_table[cap]);
343 static void __exit async_tx_exit(void)
345 enum dma_transaction_type cap;
347 channel_table_initialized = 0;
349 for_each_dma_cap_mask(cap, dma_cap_mask_all)
350 if (channel_table[cap])
351 free_percpu(channel_table[cap]);
353 dma_async_client_unregister(&async_tx_dma);
357 * async_tx_find_channel - find a channel to carry out the operation or let
358 * the transaction execute synchronously
359 * @depend_tx: transaction dependency
360 * @tx_type: transaction type
363 async_tx_find_channel(struct dma_async_tx_descriptor *depend_tx,
364 enum dma_transaction_type tx_type)
366 /* see if we can keep the chain on one channel */
368 dma_has_cap(tx_type, depend_tx->chan->device->cap_mask))
369 return depend_tx->chan;
370 else if (likely(channel_table_initialized)) {
371 struct dma_chan_ref *ref;
373 ref = per_cpu_ptr(channel_table[tx_type], cpu)->ref;
375 return ref ? ref->chan : NULL;
379 EXPORT_SYMBOL_GPL(async_tx_find_channel);
381 static int __init async_tx_init(void)
383 printk(KERN_INFO "async_tx: api initialized (sync-only)\n");
387 static void __exit async_tx_exit(void)
394 async_tx_submit(struct dma_chan *chan, struct dma_async_tx_descriptor *tx,
395 enum async_tx_flags flags, struct dma_async_tx_descriptor *depend_tx,
396 dma_async_tx_callback cb_fn, void *cb_param)
398 tx->callback = cb_fn;
399 tx->callback_param = cb_param;
401 /* set this new tx to run after depend_tx if:
402 * 1/ a dependency exists (depend_tx is !NULL)
403 * 2/ the tx can not be submitted to the current channel
405 if (depend_tx && depend_tx->chan != chan) {
406 /* if ack is already set then we cannot be sure
407 * we are referring to the correct operation
409 BUG_ON(depend_tx->ack);
411 tx->parent = depend_tx;
412 spin_lock_bh(&depend_tx->lock);
413 list_add_tail(&tx->depend_node, &depend_tx->depend_list);
414 if (depend_tx->cookie == 0) {
415 struct dma_chan *dep_chan = depend_tx->chan;
416 struct dma_device *dep_dev = dep_chan->device;
417 dep_dev->device_dependency_added(dep_chan);
419 spin_unlock_bh(&depend_tx->lock);
421 /* schedule an interrupt to trigger the channel switch */
422 async_trigger_callback(ASYNC_TX_ACK, depend_tx, NULL, NULL);
428 if (flags & ASYNC_TX_ACK)
431 if (depend_tx && (flags & ASYNC_TX_DEP_ACK))
432 async_tx_ack(depend_tx);
434 EXPORT_SYMBOL_GPL(async_tx_submit);
437 * async_trigger_callback - schedules the callback function to be run after
438 * any dependent operations have been completed.
439 * @flags: ASYNC_TX_ACK, ASYNC_TX_DEP_ACK
440 * @depend_tx: 'callback' requires the completion of this transaction
441 * @cb_fn: function to call after depend_tx completes
442 * @cb_param: parameter to pass to the callback routine
444 struct dma_async_tx_descriptor *
445 async_trigger_callback(enum async_tx_flags flags,
446 struct dma_async_tx_descriptor *depend_tx,
447 dma_async_tx_callback cb_fn, void *cb_param)
449 struct dma_chan *chan;
450 struct dma_device *device;
451 struct dma_async_tx_descriptor *tx;
454 chan = depend_tx->chan;
455 device = chan->device;
457 /* see if we can schedule an interrupt
458 * otherwise poll for completion
460 if (device && !dma_has_cap(DMA_INTERRUPT, device->cap_mask))
463 tx = device ? device->device_prep_dma_interrupt(chan) : NULL;
468 pr_debug("%s: (async)\n", __FUNCTION__);
470 async_tx_submit(chan, tx, flags, depend_tx, cb_fn, cb_param);
472 pr_debug("%s: (sync)\n", __FUNCTION__);
474 /* wait for any prerequisite operations */
476 /* if ack is already set then we cannot be sure
477 * we are referring to the correct operation
479 BUG_ON(depend_tx->ack);
480 if (dma_wait_for_async_tx(depend_tx) == DMA_ERROR)
481 panic("%s: DMA_ERROR waiting for depend_tx\n",
485 async_tx_sync_epilog(flags, depend_tx, cb_fn, cb_param);
490 EXPORT_SYMBOL_GPL(async_trigger_callback);
492 module_init(async_tx_init);
493 module_exit(async_tx_exit);
495 MODULE_AUTHOR("Intel Corporation");
496 MODULE_DESCRIPTION("Asynchronous Bulk Memory Transactions API");
497 MODULE_LICENSE("GPL");