1/*
2 * Copyright(c) 2004 - 2006 Intel Corporation. All rights reserved.
3 *
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)
7 * any later version.
8 *
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
12 * more details.
13 *
14 * The full GNU General Public License is included in this distribution in the
15 * file called COPYING.
16 */
17
18/*
19 * This code implements the DMA subsystem. It provides a HW-neutral interface
20 * for other kernel code to use asynchronous memory copy capabilities,
21 * if present, and allows different HW DMA drivers to register as providing
22 * this capability.
23 *
24 * Due to the fact we are accelerating what is already a relatively fast
25 * operation, the code goes to great lengths to avoid additional overhead,
26 * such as locking.
27 *
28 * LOCKING:
29 *
30 * The subsystem keeps a global list of dma_device structs it is protected by a
31 * mutex, dma_list_mutex.
32 *
33 * A subsystem can get access to a channel by calling dmaengine_get() followed
34 * by dma_find_channel(), or if it has need for an exclusive channel it can call
35 * dma_request_channel().  Once a channel is allocated a reference is taken
36 * against its corresponding driver to disable removal.
37 *
38 * Each device has a channels list, which runs unlocked but is never modified
39 * once the device is registered, it's just setup by the driver.
40 *
41 * See Documentation/dmaengine.txt for more details
42 */
43
44#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
45
46#include <linux/dma-mapping.h>
47#include <linux/init.h>
48#include <linux/module.h>
49#include <linux/mm.h>
50#include <linux/device.h>
51#include <linux/dmaengine.h>
52#include <linux/hardirq.h>
53#include <linux/spinlock.h>
54#include <linux/percpu.h>
55#include <linux/rcupdate.h>
56#include <linux/mutex.h>
57#include <linux/jiffies.h>
58#include <linux/rculist.h>
59#include <linux/idr.h>
60#include <linux/slab.h>
61#include <linux/acpi.h>
62#include <linux/acpi_dma.h>
63#include <linux/of_dma.h>
64#include <linux/mempool.h>
65
66static DEFINE_MUTEX(dma_list_mutex);
67static DEFINE_IDR(dma_idr);
68static LIST_HEAD(dma_device_list);
69static long dmaengine_ref_count;
70
71/* --- sysfs implementation --- */
72
73/**
74 * dev_to_dma_chan - convert a device pointer to the its sysfs container object
75 * @dev - device node
76 *
77 * Must be called under dma_list_mutex
78 */
79static struct dma_chan *dev_to_dma_chan(struct device *dev)
80{
81	struct dma_chan_dev *chan_dev;
82
83	chan_dev = container_of(dev, typeof(*chan_dev), device);
84	return chan_dev->chan;
85}
86
87static ssize_t memcpy_count_show(struct device *dev,
88				 struct device_attribute *attr, char *buf)
89{
90	struct dma_chan *chan;
91	unsigned long count = 0;
92	int i;
93	int err;
94
95	mutex_lock(&dma_list_mutex);
96	chan = dev_to_dma_chan(dev);
97	if (chan) {
98		for_each_possible_cpu(i)
99			count += per_cpu_ptr(chan->local, i)->memcpy_count;
100		err = sprintf(buf, "%lu\n", count);
101	} else
102		err = -ENODEV;
103	mutex_unlock(&dma_list_mutex);
104
105	return err;
106}
107static DEVICE_ATTR_RO(memcpy_count);
108
109static ssize_t bytes_transferred_show(struct device *dev,
110				      struct device_attribute *attr, char *buf)
111{
112	struct dma_chan *chan;
113	unsigned long count = 0;
114	int i;
115	int err;
116
117	mutex_lock(&dma_list_mutex);
118	chan = dev_to_dma_chan(dev);
119	if (chan) {
120		for_each_possible_cpu(i)
121			count += per_cpu_ptr(chan->local, i)->bytes_transferred;
122		err = sprintf(buf, "%lu\n", count);
123	} else
124		err = -ENODEV;
125	mutex_unlock(&dma_list_mutex);
126
127	return err;
128}
129static DEVICE_ATTR_RO(bytes_transferred);
130
131static ssize_t in_use_show(struct device *dev, struct device_attribute *attr,
132			   char *buf)
133{
134	struct dma_chan *chan;
135	int err;
136
137	mutex_lock(&dma_list_mutex);
138	chan = dev_to_dma_chan(dev);
139	if (chan)
140		err = sprintf(buf, "%d\n", chan->client_count);
141	else
142		err = -ENODEV;
143	mutex_unlock(&dma_list_mutex);
144
145	return err;
146}
147static DEVICE_ATTR_RO(in_use);
148
149static struct attribute *dma_dev_attrs[] = {
150	&dev_attr_memcpy_count.attr,
151	&dev_attr_bytes_transferred.attr,
152	&dev_attr_in_use.attr,
153	NULL,
154};
155ATTRIBUTE_GROUPS(dma_dev);
156
157static void chan_dev_release(struct device *dev)
158{
159	struct dma_chan_dev *chan_dev;
160
161	chan_dev = container_of(dev, typeof(*chan_dev), device);
162	if (atomic_dec_and_test(chan_dev->idr_ref)) {
163		mutex_lock(&dma_list_mutex);
164		idr_remove(&dma_idr, chan_dev->dev_id);
165		mutex_unlock(&dma_list_mutex);
166		kfree(chan_dev->idr_ref);
167	}
168	kfree(chan_dev);
169}
170
171static struct class dma_devclass = {
172	.name		= "dma",
173	.dev_groups	= dma_dev_groups,
174	.dev_release	= chan_dev_release,
175};
176
177/* --- client and device registration --- */
178
179#define dma_device_satisfies_mask(device, mask) \
180	__dma_device_satisfies_mask((device), &(mask))
181static int
182__dma_device_satisfies_mask(struct dma_device *device,
183			    const dma_cap_mask_t *want)
184{
185	dma_cap_mask_t has;
186
187	bitmap_and(has.bits, want->bits, device->cap_mask.bits,
188		DMA_TX_TYPE_END);
189	return bitmap_equal(want->bits, has.bits, DMA_TX_TYPE_END);
190}
191
192static struct module *dma_chan_to_owner(struct dma_chan *chan)
193{
194	return chan->device->dev->driver->owner;
195}
196
197/**
198 * balance_ref_count - catch up the channel reference count
199 * @chan - channel to balance ->client_count versus dmaengine_ref_count
200 *
201 * balance_ref_count must be called under dma_list_mutex
202 */
203static void balance_ref_count(struct dma_chan *chan)
204{
205	struct module *owner = dma_chan_to_owner(chan);
206
207	while (chan->client_count < dmaengine_ref_count) {
208		__module_get(owner);
209		chan->client_count++;
210	}
211}
212
213/**
214 * dma_chan_get - try to grab a dma channel's parent driver module
215 * @chan - channel to grab
216 *
217 * Must be called under dma_list_mutex
218 */
219static int dma_chan_get(struct dma_chan *chan)
220{
221	struct module *owner = dma_chan_to_owner(chan);
222	int ret;
223
224	/* The channel is already in use, update client count */
225	if (chan->client_count) {
226		__module_get(owner);
227		goto out;
228	}
229
230	if (!try_module_get(owner))
231		return -ENODEV;
232
233	/* allocate upon first client reference */
234	if (chan->device->device_alloc_chan_resources) {
235		ret = chan->device->device_alloc_chan_resources(chan);
236		if (ret < 0)
237			goto err_out;
238	}
239
240	if (!dma_has_cap(DMA_PRIVATE, chan->device->cap_mask))
241		balance_ref_count(chan);
242
243out:
244	chan->client_count++;
245	return 0;
246
247err_out:
248	module_put(owner);
249	return ret;
250}
251
252/**
253 * dma_chan_put - drop a reference to a dma channel's parent driver module
254 * @chan - channel to release
255 *
256 * Must be called under dma_list_mutex
257 */
258static void dma_chan_put(struct dma_chan *chan)
259{
260	/* This channel is not in use, bail out */
261	if (!chan->client_count)
262		return;
263
264	chan->client_count--;
265	module_put(dma_chan_to_owner(chan));
266
267	/* This channel is not in use anymore, free it */
268	if (!chan->client_count && chan->device->device_free_chan_resources)
269		chan->device->device_free_chan_resources(chan);
270}
271
272enum dma_status dma_sync_wait(struct dma_chan *chan, dma_cookie_t cookie)
273{
274	enum dma_status status;
275	unsigned long dma_sync_wait_timeout = jiffies + msecs_to_jiffies(5000);
276
277	dma_async_issue_pending(chan);
278	do {
279		status = dma_async_is_tx_complete(chan, cookie, NULL, NULL);
280		if (time_after_eq(jiffies, dma_sync_wait_timeout)) {
281			pr_err("%s: timeout!\n", __func__);
282			return DMA_ERROR;
283		}
284		if (status != DMA_IN_PROGRESS)
285			break;
286		cpu_relax();
287	} while (1);
288
289	return status;
290}
291EXPORT_SYMBOL(dma_sync_wait);
292
293/**
294 * dma_cap_mask_all - enable iteration over all operation types
295 */
296static dma_cap_mask_t dma_cap_mask_all;
297
298/**
299 * dma_chan_tbl_ent - tracks channel allocations per core/operation
300 * @chan - associated channel for this entry
301 */
302struct dma_chan_tbl_ent {
303	struct dma_chan *chan;
304};
305
306/**
307 * channel_table - percpu lookup table for memory-to-memory offload providers
308 */
309static struct dma_chan_tbl_ent __percpu *channel_table[DMA_TX_TYPE_END];
310
311static int __init dma_channel_table_init(void)
312{
313	enum dma_transaction_type cap;
314	int err = 0;
315
316	bitmap_fill(dma_cap_mask_all.bits, DMA_TX_TYPE_END);
317
318	/* 'interrupt', 'private', and 'slave' are channel capabilities,
319	 * but are not associated with an operation so they do not need
320	 * an entry in the channel_table
321	 */
322	clear_bit(DMA_INTERRUPT, dma_cap_mask_all.bits);
323	clear_bit(DMA_PRIVATE, dma_cap_mask_all.bits);
324	clear_bit(DMA_SLAVE, dma_cap_mask_all.bits);
325
326	for_each_dma_cap_mask(cap, dma_cap_mask_all) {
327		channel_table[cap] = alloc_percpu(struct dma_chan_tbl_ent);
328		if (!channel_table[cap]) {
329			err = -ENOMEM;
330			break;
331		}
332	}
333
334	if (err) {
335		pr_err("initialization failure\n");
336		for_each_dma_cap_mask(cap, dma_cap_mask_all)
337			free_percpu(channel_table[cap]);
338	}
339
340	return err;
341}
342arch_initcall(dma_channel_table_init);
343
344/**
345 * dma_find_channel - find a channel to carry out the operation
346 * @tx_type: transaction type
347 */
348struct dma_chan *dma_find_channel(enum dma_transaction_type tx_type)
349{
350	return this_cpu_read(channel_table[tx_type]->chan);
351}
352EXPORT_SYMBOL(dma_find_channel);
353
354/**
355 * dma_issue_pending_all - flush all pending operations across all channels
356 */
357void dma_issue_pending_all(void)
358{
359	struct dma_device *device;
360	struct dma_chan *chan;
361
362	rcu_read_lock();
363	list_for_each_entry_rcu(device, &dma_device_list, global_node) {
364		if (dma_has_cap(DMA_PRIVATE, device->cap_mask))
365			continue;
366		list_for_each_entry(chan, &device->channels, device_node)
367			if (chan->client_count)
368				device->device_issue_pending(chan);
369	}
370	rcu_read_unlock();
371}
372EXPORT_SYMBOL(dma_issue_pending_all);
373
374/**
375 * dma_chan_is_local - returns true if the channel is in the same numa-node as the cpu
376 */
377static bool dma_chan_is_local(struct dma_chan *chan, int cpu)
378{
379	int node = dev_to_node(chan->device->dev);
380	return node == -1 || cpumask_test_cpu(cpu, cpumask_of_node(node));
381}
382
383/**
384 * min_chan - returns the channel with min count and in the same numa-node as the cpu
385 * @cap: capability to match
386 * @cpu: cpu index which the channel should be close to
387 *
388 * If some channels are close to the given cpu, the one with the lowest
389 * reference count is returned. Otherwise, cpu is ignored and only the
390 * reference count is taken into account.
391 * Must be called under dma_list_mutex.
392 */
393static struct dma_chan *min_chan(enum dma_transaction_type cap, int cpu)
394{
395	struct dma_device *device;
396	struct dma_chan *chan;
397	struct dma_chan *min = NULL;
398	struct dma_chan *localmin = NULL;
399
400	list_for_each_entry(device, &dma_device_list, global_node) {
401		if (!dma_has_cap(cap, device->cap_mask) ||
402		    dma_has_cap(DMA_PRIVATE, device->cap_mask))
403			continue;
404		list_for_each_entry(chan, &device->channels, device_node) {
405			if (!chan->client_count)
406				continue;
407			if (!min || chan->table_count < min->table_count)
408				min = chan;
409
410			if (dma_chan_is_local(chan, cpu))
411				if (!localmin ||
412				    chan->table_count < localmin->table_count)
413					localmin = chan;
414		}
415	}
416
417	chan = localmin ? localmin : min;
418
419	if (chan)
420		chan->table_count++;
421
422	return chan;
423}
424
425/**
426 * dma_channel_rebalance - redistribute the available channels
427 *
428 * Optimize for cpu isolation (each cpu gets a dedicated channel for an
429 * operation type) in the SMP case,  and operation isolation (avoid
430 * multi-tasking channels) in the non-SMP case.  Must be called under
431 * dma_list_mutex.
432 */
433static void dma_channel_rebalance(void)
434{
435	struct dma_chan *chan;
436	struct dma_device *device;
437	int cpu;
438	int cap;
439
440	/* undo the last distribution */
441	for_each_dma_cap_mask(cap, dma_cap_mask_all)
442		for_each_possible_cpu(cpu)
443			per_cpu_ptr(channel_table[cap], cpu)->chan = NULL;
444
445	list_for_each_entry(device, &dma_device_list, global_node) {
446		if (dma_has_cap(DMA_PRIVATE, device->cap_mask))
447			continue;
448		list_for_each_entry(chan, &device->channels, device_node)
449			chan->table_count = 0;
450	}
451
452	/* don't populate the channel_table if no clients are available */
453	if (!dmaengine_ref_count)
454		return;
455
456	/* redistribute available channels */
457	for_each_dma_cap_mask(cap, dma_cap_mask_all)
458		for_each_online_cpu(cpu) {
459			chan = min_chan(cap, cpu);
460			per_cpu_ptr(channel_table[cap], cpu)->chan = chan;
461		}
462}
463
464int dma_get_slave_caps(struct dma_chan *chan, struct dma_slave_caps *caps)
465{
466	struct dma_device *device;
467
468	if (!chan || !caps)
469		return -EINVAL;
470
471	device = chan->device;
472
473	/* check if the channel supports slave transactions */
474	if (!test_bit(DMA_SLAVE, device->cap_mask.bits))
475		return -ENXIO;
476
477	/*
478	 * Check whether it reports it uses the generic slave
479	 * capabilities, if not, that means it doesn't support any
480	 * kind of slave capabilities reporting.
481	 */
482	if (!device->directions)
483		return -ENXIO;
484
485	caps->src_addr_widths = device->src_addr_widths;
486	caps->dst_addr_widths = device->dst_addr_widths;
487	caps->directions = device->directions;
488	caps->residue_granularity = device->residue_granularity;
489
490	/*
491	 * Some devices implement only pause (e.g. to get residuum) but no
492	 * resume. However cmd_pause is advertised as pause AND resume.
493	 */
494	caps->cmd_pause = !!(device->device_pause && device->device_resume);
495	caps->cmd_terminate = !!device->device_terminate_all;
496
497	return 0;
498}
499EXPORT_SYMBOL_GPL(dma_get_slave_caps);
500
501static struct dma_chan *private_candidate(const dma_cap_mask_t *mask,
502					  struct dma_device *dev,
503					  dma_filter_fn fn, void *fn_param)
504{
505	struct dma_chan *chan;
506
507	if (!__dma_device_satisfies_mask(dev, mask)) {
508		pr_debug("%s: wrong capabilities\n", __func__);
509		return NULL;
510	}
511	/* devices with multiple channels need special handling as we need to
512	 * ensure that all channels are either private or public.
513	 */
514	if (dev->chancnt > 1 && !dma_has_cap(DMA_PRIVATE, dev->cap_mask))
515		list_for_each_entry(chan, &dev->channels, device_node) {
516			/* some channels are already publicly allocated */
517			if (chan->client_count)
518				return NULL;
519		}
520
521	list_for_each_entry(chan, &dev->channels, device_node) {
522		if (chan->client_count) {
523			pr_debug("%s: %s busy\n",
524				 __func__, dma_chan_name(chan));
525			continue;
526		}
527		if (fn && !fn(chan, fn_param)) {
528			pr_debug("%s: %s filter said false\n",
529				 __func__, dma_chan_name(chan));
530			continue;
531		}
532		return chan;
533	}
534
535	return NULL;
536}
537
538/**
539 * dma_request_slave_channel - try to get specific channel exclusively
540 * @chan: target channel
541 */
542struct dma_chan *dma_get_slave_channel(struct dma_chan *chan)
543{
544	int err = -EBUSY;
545
546	/* lock against __dma_request_channel */
547	mutex_lock(&dma_list_mutex);
548
549	if (chan->client_count == 0) {
550		err = dma_chan_get(chan);
551		if (err)
552			pr_debug("%s: failed to get %s: (%d)\n",
553				__func__, dma_chan_name(chan), err);
554	} else
555		chan = NULL;
556
557	mutex_unlock(&dma_list_mutex);
558
559
560	return chan;
561}
562EXPORT_SYMBOL_GPL(dma_get_slave_channel);
563
564struct dma_chan *dma_get_any_slave_channel(struct dma_device *device)
565{
566	dma_cap_mask_t mask;
567	struct dma_chan *chan;
568	int err;
569
570	dma_cap_zero(mask);
571	dma_cap_set(DMA_SLAVE, mask);
572
573	/* lock against __dma_request_channel */
574	mutex_lock(&dma_list_mutex);
575
576	chan = private_candidate(&mask, device, NULL, NULL);
577	if (chan) {
578		dma_cap_set(DMA_PRIVATE, device->cap_mask);
579		device->privatecnt++;
580		err = dma_chan_get(chan);
581		if (err) {
582			pr_debug("%s: failed to get %s: (%d)\n",
583				__func__, dma_chan_name(chan), err);
584			chan = NULL;
585			if (--device->privatecnt == 0)
586				dma_cap_clear(DMA_PRIVATE, device->cap_mask);
587		}
588	}
589
590	mutex_unlock(&dma_list_mutex);
591
592	return chan;
593}
594EXPORT_SYMBOL_GPL(dma_get_any_slave_channel);
595
596/**
597 * __dma_request_channel - try to allocate an exclusive channel
598 * @mask: capabilities that the channel must satisfy
599 * @fn: optional callback to disposition available channels
600 * @fn_param: opaque parameter to pass to dma_filter_fn
601 *
602 * Returns pointer to appropriate DMA channel on success or NULL.
603 */
604struct dma_chan *__dma_request_channel(const dma_cap_mask_t *mask,
605				       dma_filter_fn fn, void *fn_param)
606{
607	struct dma_device *device, *_d;
608	struct dma_chan *chan = NULL;
609	int err;
610
611	/* Find a channel */
612	mutex_lock(&dma_list_mutex);
613	list_for_each_entry_safe(device, _d, &dma_device_list, global_node) {
614		chan = private_candidate(mask, device, fn, fn_param);
615		if (chan) {
616			/* Found a suitable channel, try to grab, prep, and
617			 * return it.  We first set DMA_PRIVATE to disable
618			 * balance_ref_count as this channel will not be
619			 * published in the general-purpose allocator
620			 */
621			dma_cap_set(DMA_PRIVATE, device->cap_mask);
622			device->privatecnt++;
623			err = dma_chan_get(chan);
624
625			if (err == -ENODEV) {
626				pr_debug("%s: %s module removed\n",
627					 __func__, dma_chan_name(chan));
628				list_del_rcu(&device->global_node);
629			} else if (err)
630				pr_debug("%s: failed to get %s: (%d)\n",
631					 __func__, dma_chan_name(chan), err);
632			else
633				break;
634			if (--device->privatecnt == 0)
635				dma_cap_clear(DMA_PRIVATE, device->cap_mask);
636			chan = NULL;
637		}
638	}
639	mutex_unlock(&dma_list_mutex);
640
641	pr_debug("%s: %s (%s)\n",
642		 __func__,
643		 chan ? "success" : "fail",
644		 chan ? dma_chan_name(chan) : NULL);
645
646	return chan;
647}
648EXPORT_SYMBOL_GPL(__dma_request_channel);
649
650/**
651 * dma_request_slave_channel - try to allocate an exclusive slave channel
652 * @dev:	pointer to client device structure
653 * @name:	slave channel name
654 *
655 * Returns pointer to appropriate DMA channel on success or an error pointer.
656 */
657struct dma_chan *dma_request_slave_channel_reason(struct device *dev,
658						  const char *name)
659{
660	/* If device-tree is present get slave info from here */
661	if (dev->of_node)
662		return of_dma_request_slave_channel(dev->of_node, name);
663
664	/* If device was enumerated by ACPI get slave info from here */
665	if (ACPI_HANDLE(dev))
666		return acpi_dma_request_slave_chan_by_name(dev, name);
667
668	return ERR_PTR(-ENODEV);
669}
670EXPORT_SYMBOL_GPL(dma_request_slave_channel_reason);
671
672/**
673 * dma_request_slave_channel - try to allocate an exclusive slave channel
674 * @dev:	pointer to client device structure
675 * @name:	slave channel name
676 *
677 * Returns pointer to appropriate DMA channel on success or NULL.
678 */
679struct dma_chan *dma_request_slave_channel(struct device *dev,
680					   const char *name)
681{
682	struct dma_chan *ch = dma_request_slave_channel_reason(dev, name);
683	if (IS_ERR(ch))
684		return NULL;
685	return ch;
686}
687EXPORT_SYMBOL_GPL(dma_request_slave_channel);
688
689void dma_release_channel(struct dma_chan *chan)
690{
691	mutex_lock(&dma_list_mutex);
692	WARN_ONCE(chan->client_count != 1,
693		  "chan reference count %d != 1\n", chan->client_count);
694	dma_chan_put(chan);
695	/* drop PRIVATE cap enabled by __dma_request_channel() */
696	if (--chan->device->privatecnt == 0)
697		dma_cap_clear(DMA_PRIVATE, chan->device->cap_mask);
698	mutex_unlock(&dma_list_mutex);
699}
700EXPORT_SYMBOL_GPL(dma_release_channel);
701
702/**
703 * dmaengine_get - register interest in dma_channels
704 */
705void dmaengine_get(void)
706{
707	struct dma_device *device, *_d;
708	struct dma_chan *chan;
709	int err;
710
711	mutex_lock(&dma_list_mutex);
712	dmaengine_ref_count++;
713
714	/* try to grab channels */
715	list_for_each_entry_safe(device, _d, &dma_device_list, global_node) {
716		if (dma_has_cap(DMA_PRIVATE, device->cap_mask))
717			continue;
718		list_for_each_entry(chan, &device->channels, device_node) {
719			err = dma_chan_get(chan);
720			if (err == -ENODEV) {
721				/* module removed before we could use it */
722				list_del_rcu(&device->global_node);
723				break;
724			} else if (err)
725				pr_debug("%s: failed to get %s: (%d)\n",
726				       __func__, dma_chan_name(chan), err);
727		}
728	}
729
730	/* if this is the first reference and there were channels
731	 * waiting we need to rebalance to get those channels
732	 * incorporated into the channel table
733	 */
734	if (dmaengine_ref_count == 1)
735		dma_channel_rebalance();
736	mutex_unlock(&dma_list_mutex);
737}
738EXPORT_SYMBOL(dmaengine_get);
739
740/**
741 * dmaengine_put - let dma drivers be removed when ref_count == 0
742 */
743void dmaengine_put(void)
744{
745	struct dma_device *device;
746	struct dma_chan *chan;
747
748	mutex_lock(&dma_list_mutex);
749	dmaengine_ref_count--;
750	BUG_ON(dmaengine_ref_count < 0);
751	/* drop channel references */
752	list_for_each_entry(device, &dma_device_list, global_node) {
753		if (dma_has_cap(DMA_PRIVATE, device->cap_mask))
754			continue;
755		list_for_each_entry(chan, &device->channels, device_node)
756			dma_chan_put(chan);
757	}
758	mutex_unlock(&dma_list_mutex);
759}
760EXPORT_SYMBOL(dmaengine_put);
761
762static bool device_has_all_tx_types(struct dma_device *device)
763{
764	/* A device that satisfies this test has channels that will never cause
765	 * an async_tx channel switch event as all possible operation types can
766	 * be handled.
767	 */
768	#ifdef CONFIG_ASYNC_TX_DMA
769	if (!dma_has_cap(DMA_INTERRUPT, device->cap_mask))
770		return false;
771	#endif
772
773	#if defined(CONFIG_ASYNC_MEMCPY) || defined(CONFIG_ASYNC_MEMCPY_MODULE)
774	if (!dma_has_cap(DMA_MEMCPY, device->cap_mask))
775		return false;
776	#endif
777
778	#if defined(CONFIG_ASYNC_XOR) || defined(CONFIG_ASYNC_XOR_MODULE)
779	if (!dma_has_cap(DMA_XOR, device->cap_mask))
780		return false;
781
782	#ifndef CONFIG_ASYNC_TX_DISABLE_XOR_VAL_DMA
783	if (!dma_has_cap(DMA_XOR_VAL, device->cap_mask))
784		return false;
785	#endif
786	#endif
787
788	#if defined(CONFIG_ASYNC_PQ) || defined(CONFIG_ASYNC_PQ_MODULE)
789	if (!dma_has_cap(DMA_PQ, device->cap_mask))
790		return false;
791
792	#ifndef CONFIG_ASYNC_TX_DISABLE_PQ_VAL_DMA
793	if (!dma_has_cap(DMA_PQ_VAL, device->cap_mask))
794		return false;
795	#endif
796	#endif
797
798	return true;
799}
800
801static int get_dma_id(struct dma_device *device)
802{
803	int rc;
804
805	mutex_lock(&dma_list_mutex);
806
807	rc = idr_alloc(&dma_idr, NULL, 0, 0, GFP_KERNEL);
808	if (rc >= 0)
809		device->dev_id = rc;
810
811	mutex_unlock(&dma_list_mutex);
812	return rc < 0 ? rc : 0;
813}
814
815/**
816 * dma_async_device_register - registers DMA devices found
817 * @device: &dma_device
818 */
819int dma_async_device_register(struct dma_device *device)
820{
821	int chancnt = 0, rc;
822	struct dma_chan* chan;
823	atomic_t *idr_ref;
824
825	if (!device)
826		return -ENODEV;
827
828	/* validate device routines */
829	BUG_ON(dma_has_cap(DMA_MEMCPY, device->cap_mask) &&
830		!device->device_prep_dma_memcpy);
831	BUG_ON(dma_has_cap(DMA_XOR, device->cap_mask) &&
832		!device->device_prep_dma_xor);
833	BUG_ON(dma_has_cap(DMA_XOR_VAL, device->cap_mask) &&
834		!device->device_prep_dma_xor_val);
835	BUG_ON(dma_has_cap(DMA_PQ, device->cap_mask) &&
836		!device->device_prep_dma_pq);
837	BUG_ON(dma_has_cap(DMA_PQ_VAL, device->cap_mask) &&
838		!device->device_prep_dma_pq_val);
839	BUG_ON(dma_has_cap(DMA_INTERRUPT, device->cap_mask) &&
840		!device->device_prep_dma_interrupt);
841	BUG_ON(dma_has_cap(DMA_SG, device->cap_mask) &&
842		!device->device_prep_dma_sg);
843	BUG_ON(dma_has_cap(DMA_CYCLIC, device->cap_mask) &&
844		!device->device_prep_dma_cyclic);
845	BUG_ON(dma_has_cap(DMA_INTERLEAVE, device->cap_mask) &&
846		!device->device_prep_interleaved_dma);
847
848	BUG_ON(!device->device_tx_status);
849	BUG_ON(!device->device_issue_pending);
850	BUG_ON(!device->dev);
851
852	/* note: this only matters in the
853	 * CONFIG_ASYNC_TX_ENABLE_CHANNEL_SWITCH=n case
854	 */
855	if (device_has_all_tx_types(device))
856		dma_cap_set(DMA_ASYNC_TX, device->cap_mask);
857
858	idr_ref = kmalloc(sizeof(*idr_ref), GFP_KERNEL);
859	if (!idr_ref)
860		return -ENOMEM;
861	rc = get_dma_id(device);
862	if (rc != 0) {
863		kfree(idr_ref);
864		return rc;
865	}
866
867	atomic_set(idr_ref, 0);
868
869	/* represent channels in sysfs. Probably want devs too */
870	list_for_each_entry(chan, &device->channels, device_node) {
871		rc = -ENOMEM;
872		chan->local = alloc_percpu(typeof(*chan->local));
873		if (chan->local == NULL)
874			goto err_out;
875		chan->dev = kzalloc(sizeof(*chan->dev), GFP_KERNEL);
876		if (chan->dev == NULL) {
877			free_percpu(chan->local);
878			chan->local = NULL;
879			goto err_out;
880		}
881
882		chan->chan_id = chancnt++;
883		chan->dev->device.class = &dma_devclass;
884		chan->dev->device.parent = device->dev;
885		chan->dev->chan = chan;
886		chan->dev->idr_ref = idr_ref;
887		chan->dev->dev_id = device->dev_id;
888		atomic_inc(idr_ref);
889		dev_set_name(&chan->dev->device, "dma%dchan%d",
890			     device->dev_id, chan->chan_id);
891
892		rc = device_register(&chan->dev->device);
893		if (rc) {
894			free_percpu(chan->local);
895			chan->local = NULL;
896			kfree(chan->dev);
897			atomic_dec(idr_ref);
898			goto err_out;
899		}
900		chan->client_count = 0;
901	}
902	device->chancnt = chancnt;
903
904	mutex_lock(&dma_list_mutex);
905	/* take references on public channels */
906	if (dmaengine_ref_count && !dma_has_cap(DMA_PRIVATE, device->cap_mask))
907		list_for_each_entry(chan, &device->channels, device_node) {
908			/* if clients are already waiting for channels we need
909			 * to take references on their behalf
910			 */
911			if (dma_chan_get(chan) == -ENODEV) {
912				/* note we can only get here for the first
913				 * channel as the remaining channels are
914				 * guaranteed to get a reference
915				 */
916				rc = -ENODEV;
917				mutex_unlock(&dma_list_mutex);
918				goto err_out;
919			}
920		}
921	list_add_tail_rcu(&device->global_node, &dma_device_list);
922	if (dma_has_cap(DMA_PRIVATE, device->cap_mask))
923		device->privatecnt++;	/* Always private */
924	dma_channel_rebalance();
925	mutex_unlock(&dma_list_mutex);
926
927	return 0;
928
929err_out:
930	/* if we never registered a channel just release the idr */
931	if (atomic_read(idr_ref) == 0) {
932		mutex_lock(&dma_list_mutex);
933		idr_remove(&dma_idr, device->dev_id);
934		mutex_unlock(&dma_list_mutex);
935		kfree(idr_ref);
936		return rc;
937	}
938
939	list_for_each_entry(chan, &device->channels, device_node) {
940		if (chan->local == NULL)
941			continue;
942		mutex_lock(&dma_list_mutex);
943		chan->dev->chan = NULL;
944		mutex_unlock(&dma_list_mutex);
945		device_unregister(&chan->dev->device);
946		free_percpu(chan->local);
947	}
948	return rc;
949}
950EXPORT_SYMBOL(dma_async_device_register);
951
952/**
953 * dma_async_device_unregister - unregister a DMA device
954 * @device: &dma_device
955 *
956 * This routine is called by dma driver exit routines, dmaengine holds module
957 * references to prevent it being called while channels are in use.
958 */
959void dma_async_device_unregister(struct dma_device *device)
960{
961	struct dma_chan *chan;
962
963	mutex_lock(&dma_list_mutex);
964	list_del_rcu(&device->global_node);
965	dma_channel_rebalance();
966	mutex_unlock(&dma_list_mutex);
967
968	list_for_each_entry(chan, &device->channels, device_node) {
969		WARN_ONCE(chan->client_count,
970			  "%s called while %d clients hold a reference\n",
971			  __func__, chan->client_count);
972		mutex_lock(&dma_list_mutex);
973		chan->dev->chan = NULL;
974		mutex_unlock(&dma_list_mutex);
975		device_unregister(&chan->dev->device);
976		free_percpu(chan->local);
977	}
978}
979EXPORT_SYMBOL(dma_async_device_unregister);
980
981struct dmaengine_unmap_pool {
982	struct kmem_cache *cache;
983	const char *name;
984	mempool_t *pool;
985	size_t size;
986};
987
988#define __UNMAP_POOL(x) { .size = x, .name = "dmaengine-unmap-" __stringify(x) }
989static struct dmaengine_unmap_pool unmap_pool[] = {
990	__UNMAP_POOL(2),
991	#if IS_ENABLED(CONFIG_DMA_ENGINE_RAID)
992	__UNMAP_POOL(16),
993	__UNMAP_POOL(128),
994	__UNMAP_POOL(256),
995	#endif
996};
997
998static struct dmaengine_unmap_pool *__get_unmap_pool(int nr)
999{
1000	int order = get_count_order(nr);
1001
1002	switch (order) {
1003	case 0 ... 1:
1004		return &unmap_pool[0];
1005	case 2 ... 4:
1006		return &unmap_pool[1];
1007	case 5 ... 7:
1008		return &unmap_pool[2];
1009	case 8:
1010		return &unmap_pool[3];
1011	default:
1012		BUG();
1013		return NULL;
1014	}
1015}
1016
1017static void dmaengine_unmap(struct kref *kref)
1018{
1019	struct dmaengine_unmap_data *unmap = container_of(kref, typeof(*unmap), kref);
1020	struct device *dev = unmap->dev;
1021	int cnt, i;
1022
1023	cnt = unmap->to_cnt;
1024	for (i = 0; i < cnt; i++)
1025		dma_unmap_page(dev, unmap->addr[i], unmap->len,
1026			       DMA_TO_DEVICE);
1027	cnt += unmap->from_cnt;
1028	for (; i < cnt; i++)
1029		dma_unmap_page(dev, unmap->addr[i], unmap->len,
1030			       DMA_FROM_DEVICE);
1031	cnt += unmap->bidi_cnt;
1032	for (; i < cnt; i++) {
1033		if (unmap->addr[i] == 0)
1034			continue;
1035		dma_unmap_page(dev, unmap->addr[i], unmap->len,
1036			       DMA_BIDIRECTIONAL);
1037	}
1038	cnt = unmap->map_cnt;
1039	mempool_free(unmap, __get_unmap_pool(cnt)->pool);
1040}
1041
1042void dmaengine_unmap_put(struct dmaengine_unmap_data *unmap)
1043{
1044	if (unmap)
1045		kref_put(&unmap->kref, dmaengine_unmap);
1046}
1047EXPORT_SYMBOL_GPL(dmaengine_unmap_put);
1048
1049static void dmaengine_destroy_unmap_pool(void)
1050{
1051	int i;
1052
1053	for (i = 0; i < ARRAY_SIZE(unmap_pool); i++) {
1054		struct dmaengine_unmap_pool *p = &unmap_pool[i];
1055
1056		if (p->pool)
1057			mempool_destroy(p->pool);
1058		p->pool = NULL;
1059		if (p->cache)
1060			kmem_cache_destroy(p->cache);
1061		p->cache = NULL;
1062	}
1063}
1064
1065static int __init dmaengine_init_unmap_pool(void)
1066{
1067	int i;
1068
1069	for (i = 0; i < ARRAY_SIZE(unmap_pool); i++) {
1070		struct dmaengine_unmap_pool *p = &unmap_pool[i];
1071		size_t size;
1072
1073		size = sizeof(struct dmaengine_unmap_data) +
1074		       sizeof(dma_addr_t) * p->size;
1075
1076		p->cache = kmem_cache_create(p->name, size, 0,
1077					     SLAB_HWCACHE_ALIGN, NULL);
1078		if (!p->cache)
1079			break;
1080		p->pool = mempool_create_slab_pool(1, p->cache);
1081		if (!p->pool)
1082			break;
1083	}
1084
1085	if (i == ARRAY_SIZE(unmap_pool))
1086		return 0;
1087
1088	dmaengine_destroy_unmap_pool();
1089	return -ENOMEM;
1090}
1091
1092struct dmaengine_unmap_data *
1093dmaengine_get_unmap_data(struct device *dev, int nr, gfp_t flags)
1094{
1095	struct dmaengine_unmap_data *unmap;
1096
1097	unmap = mempool_alloc(__get_unmap_pool(nr)->pool, flags);
1098	if (!unmap)
1099		return NULL;
1100
1101	memset(unmap, 0, sizeof(*unmap));
1102	kref_init(&unmap->kref);
1103	unmap->dev = dev;
1104	unmap->map_cnt = nr;
1105
1106	return unmap;
1107}
1108EXPORT_SYMBOL(dmaengine_get_unmap_data);
1109
1110void dma_async_tx_descriptor_init(struct dma_async_tx_descriptor *tx,
1111	struct dma_chan *chan)
1112{
1113	tx->chan = chan;
1114	#ifdef CONFIG_ASYNC_TX_ENABLE_CHANNEL_SWITCH
1115	spin_lock_init(&tx->lock);
1116	#endif
1117}
1118EXPORT_SYMBOL(dma_async_tx_descriptor_init);
1119
1120/* dma_wait_for_async_tx - spin wait for a transaction to complete
1121 * @tx: in-flight transaction to wait on
1122 */
1123enum dma_status
1124dma_wait_for_async_tx(struct dma_async_tx_descriptor *tx)
1125{
1126	unsigned long dma_sync_wait_timeout = jiffies + msecs_to_jiffies(5000);
1127
1128	if (!tx)
1129		return DMA_COMPLETE;
1130
1131	while (tx->cookie == -EBUSY) {
1132		if (time_after_eq(jiffies, dma_sync_wait_timeout)) {
1133			pr_err("%s timeout waiting for descriptor submission\n",
1134			       __func__);
1135			return DMA_ERROR;
1136		}
1137		cpu_relax();
1138	}
1139	return dma_sync_wait(tx->chan, tx->cookie);
1140}
1141EXPORT_SYMBOL_GPL(dma_wait_for_async_tx);
1142
1143/* dma_run_dependencies - helper routine for dma drivers to process
1144 *	(start) dependent operations on their target channel
1145 * @tx: transaction with dependencies
1146 */
1147void dma_run_dependencies(struct dma_async_tx_descriptor *tx)
1148{
1149	struct dma_async_tx_descriptor *dep = txd_next(tx);
1150	struct dma_async_tx_descriptor *dep_next;
1151	struct dma_chan *chan;
1152
1153	if (!dep)
1154		return;
1155
1156	/* we'll submit tx->next now, so clear the link */
1157	txd_clear_next(tx);
1158	chan = dep->chan;
1159
1160	/* keep submitting up until a channel switch is detected
1161	 * in that case we will be called again as a result of
1162	 * processing the interrupt from async_tx_channel_switch
1163	 */
1164	for (; dep; dep = dep_next) {
1165		txd_lock(dep);
1166		txd_clear_parent(dep);
1167		dep_next = txd_next(dep);
1168		if (dep_next && dep_next->chan == chan)
1169			txd_clear_next(dep); /* ->next will be submitted */
1170		else
1171			dep_next = NULL; /* submit current dep and terminate */
1172		txd_unlock(dep);
1173
1174		dep->tx_submit(dep);
1175	}
1176
1177	chan->device->device_issue_pending(chan);
1178}
1179EXPORT_SYMBOL_GPL(dma_run_dependencies);
1180
1181static int __init dma_bus_init(void)
1182{
1183	int err = dmaengine_init_unmap_pool();
1184
1185	if (err)
1186		return err;
1187	return class_register(&dma_devclass);
1188}
1189arch_initcall(dma_bus_init);
1190
1191
1192