dmaengine: Driver for the Synopsys DesignWare DMA controller
[linux-2.6] / drivers / dma / dw_dmac.c
1 /*
2  * Driver for the Synopsys DesignWare DMA Controller (aka DMACA on
3  * AVR32 systems.)
4  *
5  * Copyright (C) 2007-2008 Atmel Corporation
6  *
7  * This program is free software; you can redistribute it and/or modify
8  * it under the terms of the GNU General Public License version 2 as
9  * published by the Free Software Foundation.
10  */
11 #include <linux/clk.h>
12 #include <linux/delay.h>
13 #include <linux/dmaengine.h>
14 #include <linux/dma-mapping.h>
15 #include <linux/init.h>
16 #include <linux/interrupt.h>
17 #include <linux/io.h>
18 #include <linux/mm.h>
19 #include <linux/module.h>
20 #include <linux/platform_device.h>
21 #include <linux/slab.h>
22
23 #include "dw_dmac_regs.h"
24
25 /*
26  * This supports the Synopsys "DesignWare AHB Central DMA Controller",
27  * (DW_ahb_dmac) which is used with various AMBA 2.0 systems (not all
28  * of which use ARM any more).  See the "Databook" from Synopsys for
29  * information beyond what licensees probably provide.
30  *
31  * The driver has currently been tested only with the Atmel AT32AP7000,
32  * which does not support descriptor writeback.
33  */
34
35 /* NOTE:  DMS+SMS is system-specific. We should get this information
36  * from the platform code somehow.
37  */
38 #define DWC_DEFAULT_CTLLO       (DWC_CTLL_DST_MSIZE(0)          \
39                                 | DWC_CTLL_SRC_MSIZE(0)         \
40                                 | DWC_CTLL_DMS(0)               \
41                                 | DWC_CTLL_SMS(1)               \
42                                 | DWC_CTLL_LLP_D_EN             \
43                                 | DWC_CTLL_LLP_S_EN)
44
45 /*
46  * This is configuration-dependent and usually a funny size like 4095.
47  * Let's round it down to the nearest power of two.
48  *
49  * Note that this is a transfer count, i.e. if we transfer 32-bit
50  * words, we can do 8192 bytes per descriptor.
51  *
52  * This parameter is also system-specific.
53  */
54 #define DWC_MAX_COUNT   2048U
55
56 /*
57  * Number of descriptors to allocate for each channel. This should be
58  * made configurable somehow; preferably, the clients (at least the
59  * ones using slave transfers) should be able to give us a hint.
60  */
61 #define NR_DESCS_PER_CHANNEL    64
62
63 /*----------------------------------------------------------------------*/
64
65 /*
66  * Because we're not relying on writeback from the controller (it may not
67  * even be configured into the core!) we don't need to use dma_pool.  These
68  * descriptors -- and associated data -- are cacheable.  We do need to make
69  * sure their dcache entries are written back before handing them off to
70  * the controller, though.
71  */
72
73 static struct dw_desc *dwc_first_active(struct dw_dma_chan *dwc)
74 {
75         return list_entry(dwc->active_list.next, struct dw_desc, desc_node);
76 }
77
78 static struct dw_desc *dwc_first_queued(struct dw_dma_chan *dwc)
79 {
80         return list_entry(dwc->queue.next, struct dw_desc, desc_node);
81 }
82
83 static struct dw_desc *dwc_desc_get(struct dw_dma_chan *dwc)
84 {
85         struct dw_desc *desc, *_desc;
86         struct dw_desc *ret = NULL;
87         unsigned int i = 0;
88
89         spin_lock_bh(&dwc->lock);
90         list_for_each_entry_safe(desc, _desc, &dwc->free_list, desc_node) {
91                 if (async_tx_test_ack(&desc->txd)) {
92                         list_del(&desc->desc_node);
93                         ret = desc;
94                         break;
95                 }
96                 dev_dbg(&dwc->chan.dev, "desc %p not ACKed\n", desc);
97                 i++;
98         }
99         spin_unlock_bh(&dwc->lock);
100
101         dev_vdbg(&dwc->chan.dev, "scanned %u descriptors on freelist\n", i);
102
103         return ret;
104 }
105
106 static void dwc_sync_desc_for_cpu(struct dw_dma_chan *dwc, struct dw_desc *desc)
107 {
108         struct dw_desc  *child;
109
110         list_for_each_entry(child, &desc->txd.tx_list, desc_node)
111                 dma_sync_single_for_cpu(dwc->chan.dev.parent,
112                                 child->txd.phys, sizeof(child->lli),
113                                 DMA_TO_DEVICE);
114         dma_sync_single_for_cpu(dwc->chan.dev.parent,
115                         desc->txd.phys, sizeof(desc->lli),
116                         DMA_TO_DEVICE);
117 }
118
119 /*
120  * Move a descriptor, including any children, to the free list.
121  * `desc' must not be on any lists.
122  */
123 static void dwc_desc_put(struct dw_dma_chan *dwc, struct dw_desc *desc)
124 {
125         if (desc) {
126                 struct dw_desc *child;
127
128                 dwc_sync_desc_for_cpu(dwc, desc);
129
130                 spin_lock_bh(&dwc->lock);
131                 list_for_each_entry(child, &desc->txd.tx_list, desc_node)
132                         dev_vdbg(&dwc->chan.dev,
133                                         "moving child desc %p to freelist\n",
134                                         child);
135                 list_splice_init(&desc->txd.tx_list, &dwc->free_list);
136                 dev_vdbg(&dwc->chan.dev, "moving desc %p to freelist\n", desc);
137                 list_add(&desc->desc_node, &dwc->free_list);
138                 spin_unlock_bh(&dwc->lock);
139         }
140 }
141
142 /* Called with dwc->lock held and bh disabled */
143 static dma_cookie_t
144 dwc_assign_cookie(struct dw_dma_chan *dwc, struct dw_desc *desc)
145 {
146         dma_cookie_t cookie = dwc->chan.cookie;
147
148         if (++cookie < 0)
149                 cookie = 1;
150
151         dwc->chan.cookie = cookie;
152         desc->txd.cookie = cookie;
153
154         return cookie;
155 }
156
157 /*----------------------------------------------------------------------*/
158
159 /* Called with dwc->lock held and bh disabled */
160 static void dwc_dostart(struct dw_dma_chan *dwc, struct dw_desc *first)
161 {
162         struct dw_dma   *dw = to_dw_dma(dwc->chan.device);
163
164         /* ASSERT:  channel is idle */
165         if (dma_readl(dw, CH_EN) & dwc->mask) {
166                 dev_err(&dwc->chan.dev,
167                         "BUG: Attempted to start non-idle channel\n");
168                 dev_err(&dwc->chan.dev,
169                         "  SAR: 0x%x DAR: 0x%x LLP: 0x%x CTL: 0x%x:%08x\n",
170                         channel_readl(dwc, SAR),
171                         channel_readl(dwc, DAR),
172                         channel_readl(dwc, LLP),
173                         channel_readl(dwc, CTL_HI),
174                         channel_readl(dwc, CTL_LO));
175
176                 /* The tasklet will hopefully advance the queue... */
177                 return;
178         }
179
180         channel_writel(dwc, LLP, first->txd.phys);
181         channel_writel(dwc, CTL_LO,
182                         DWC_CTLL_LLP_D_EN | DWC_CTLL_LLP_S_EN);
183         channel_writel(dwc, CTL_HI, 0);
184         channel_set_bit(dw, CH_EN, dwc->mask);
185 }
186
187 /*----------------------------------------------------------------------*/
188
189 static void
190 dwc_descriptor_complete(struct dw_dma_chan *dwc, struct dw_desc *desc)
191 {
192         dma_async_tx_callback           callback;
193         void                            *param;
194         struct dma_async_tx_descriptor  *txd = &desc->txd;
195
196         dev_vdbg(&dwc->chan.dev, "descriptor %u complete\n", txd->cookie);
197
198         dwc->completed = txd->cookie;
199         callback = txd->callback;
200         param = txd->callback_param;
201
202         dwc_sync_desc_for_cpu(dwc, desc);
203         list_splice_init(&txd->tx_list, &dwc->free_list);
204         list_move(&desc->desc_node, &dwc->free_list);
205
206         /*
207          * We use dma_unmap_page() regardless of how the buffers were
208          * mapped before they were submitted...
209          */
210         if (!(txd->flags & DMA_COMPL_SKIP_DEST_UNMAP))
211                 dma_unmap_page(dwc->chan.dev.parent, desc->lli.dar, desc->len,
212                                 DMA_FROM_DEVICE);
213         if (!(txd->flags & DMA_COMPL_SKIP_SRC_UNMAP))
214                 dma_unmap_page(dwc->chan.dev.parent, desc->lli.sar, desc->len,
215                                 DMA_TO_DEVICE);
216
217         /*
218          * The API requires that no submissions are done from a
219          * callback, so we don't need to drop the lock here
220          */
221         if (callback)
222                 callback(param);
223 }
224
225 static void dwc_complete_all(struct dw_dma *dw, struct dw_dma_chan *dwc)
226 {
227         struct dw_desc *desc, *_desc;
228         LIST_HEAD(list);
229
230         if (dma_readl(dw, CH_EN) & dwc->mask) {
231                 dev_err(&dwc->chan.dev,
232                         "BUG: XFER bit set, but channel not idle!\n");
233
234                 /* Try to continue after resetting the channel... */
235                 channel_clear_bit(dw, CH_EN, dwc->mask);
236                 while (dma_readl(dw, CH_EN) & dwc->mask)
237                         cpu_relax();
238         }
239
240         /*
241          * Submit queued descriptors ASAP, i.e. before we go through
242          * the completed ones.
243          */
244         if (!list_empty(&dwc->queue))
245                 dwc_dostart(dwc, dwc_first_queued(dwc));
246         list_splice_init(&dwc->active_list, &list);
247         list_splice_init(&dwc->queue, &dwc->active_list);
248
249         list_for_each_entry_safe(desc, _desc, &list, desc_node)
250                 dwc_descriptor_complete(dwc, desc);
251 }
252
253 static void dwc_scan_descriptors(struct dw_dma *dw, struct dw_dma_chan *dwc)
254 {
255         dma_addr_t llp;
256         struct dw_desc *desc, *_desc;
257         struct dw_desc *child;
258         u32 status_xfer;
259
260         /*
261          * Clear block interrupt flag before scanning so that we don't
262          * miss any, and read LLP before RAW_XFER to ensure it is
263          * valid if we decide to scan the list.
264          */
265         dma_writel(dw, CLEAR.BLOCK, dwc->mask);
266         llp = channel_readl(dwc, LLP);
267         status_xfer = dma_readl(dw, RAW.XFER);
268
269         if (status_xfer & dwc->mask) {
270                 /* Everything we've submitted is done */
271                 dma_writel(dw, CLEAR.XFER, dwc->mask);
272                 dwc_complete_all(dw, dwc);
273                 return;
274         }
275
276         dev_vdbg(&dwc->chan.dev, "scan_descriptors: llp=0x%x\n", llp);
277
278         list_for_each_entry_safe(desc, _desc, &dwc->active_list, desc_node) {
279                 if (desc->lli.llp == llp)
280                         /* This one is currently in progress */
281                         return;
282
283                 list_for_each_entry(child, &desc->txd.tx_list, desc_node)
284                         if (child->lli.llp == llp)
285                                 /* Currently in progress */
286                                 return;
287
288                 /*
289                  * No descriptors so far seem to be in progress, i.e.
290                  * this one must be done.
291                  */
292                 dwc_descriptor_complete(dwc, desc);
293         }
294
295         dev_err(&dwc->chan.dev,
296                 "BUG: All descriptors done, but channel not idle!\n");
297
298         /* Try to continue after resetting the channel... */
299         channel_clear_bit(dw, CH_EN, dwc->mask);
300         while (dma_readl(dw, CH_EN) & dwc->mask)
301                 cpu_relax();
302
303         if (!list_empty(&dwc->queue)) {
304                 dwc_dostart(dwc, dwc_first_queued(dwc));
305                 list_splice_init(&dwc->queue, &dwc->active_list);
306         }
307 }
308
309 static void dwc_dump_lli(struct dw_dma_chan *dwc, struct dw_lli *lli)
310 {
311         dev_printk(KERN_CRIT, &dwc->chan.dev,
312                         "  desc: s0x%x d0x%x l0x%x c0x%x:%x\n",
313                         lli->sar, lli->dar, lli->llp,
314                         lli->ctlhi, lli->ctllo);
315 }
316
317 static void dwc_handle_error(struct dw_dma *dw, struct dw_dma_chan *dwc)
318 {
319         struct dw_desc *bad_desc;
320         struct dw_desc *child;
321
322         dwc_scan_descriptors(dw, dwc);
323
324         /*
325          * The descriptor currently at the head of the active list is
326          * borked. Since we don't have any way to report errors, we'll
327          * just have to scream loudly and try to carry on.
328          */
329         bad_desc = dwc_first_active(dwc);
330         list_del_init(&bad_desc->desc_node);
331         list_splice_init(&dwc->queue, dwc->active_list.prev);
332
333         /* Clear the error flag and try to restart the controller */
334         dma_writel(dw, CLEAR.ERROR, dwc->mask);
335         if (!list_empty(&dwc->active_list))
336                 dwc_dostart(dwc, dwc_first_active(dwc));
337
338         /*
339          * KERN_CRITICAL may seem harsh, but since this only happens
340          * when someone submits a bad physical address in a
341          * descriptor, we should consider ourselves lucky that the
342          * controller flagged an error instead of scribbling over
343          * random memory locations.
344          */
345         dev_printk(KERN_CRIT, &dwc->chan.dev,
346                         "Bad descriptor submitted for DMA!\n");
347         dev_printk(KERN_CRIT, &dwc->chan.dev,
348                         "  cookie: %d\n", bad_desc->txd.cookie);
349         dwc_dump_lli(dwc, &bad_desc->lli);
350         list_for_each_entry(child, &bad_desc->txd.tx_list, desc_node)
351                 dwc_dump_lli(dwc, &child->lli);
352
353         /* Pretend the descriptor completed successfully */
354         dwc_descriptor_complete(dwc, bad_desc);
355 }
356
357 static void dw_dma_tasklet(unsigned long data)
358 {
359         struct dw_dma *dw = (struct dw_dma *)data;
360         struct dw_dma_chan *dwc;
361         u32 status_block;
362         u32 status_xfer;
363         u32 status_err;
364         int i;
365
366         status_block = dma_readl(dw, RAW.BLOCK);
367         status_xfer = dma_readl(dw, RAW.BLOCK);
368         status_err = dma_readl(dw, RAW.ERROR);
369
370         dev_vdbg(dw->dma.dev, "tasklet: status_block=%x status_err=%x\n",
371                         status_block, status_err);
372
373         for (i = 0; i < dw->dma.chancnt; i++) {
374                 dwc = &dw->chan[i];
375                 spin_lock(&dwc->lock);
376                 if (status_err & (1 << i))
377                         dwc_handle_error(dw, dwc);
378                 else if ((status_block | status_xfer) & (1 << i))
379                         dwc_scan_descriptors(dw, dwc);
380                 spin_unlock(&dwc->lock);
381         }
382
383         /*
384          * Re-enable interrupts. Block Complete interrupts are only
385          * enabled if the INT_EN bit in the descriptor is set. This
386          * will trigger a scan before the whole list is done.
387          */
388         channel_set_bit(dw, MASK.XFER, dw->all_chan_mask);
389         channel_set_bit(dw, MASK.BLOCK, dw->all_chan_mask);
390         channel_set_bit(dw, MASK.ERROR, dw->all_chan_mask);
391 }
392
393 static irqreturn_t dw_dma_interrupt(int irq, void *dev_id)
394 {
395         struct dw_dma *dw = dev_id;
396         u32 status;
397
398         dev_vdbg(dw->dma.dev, "interrupt: status=0x%x\n",
399                         dma_readl(dw, STATUS_INT));
400
401         /*
402          * Just disable the interrupts. We'll turn them back on in the
403          * softirq handler.
404          */
405         channel_clear_bit(dw, MASK.XFER, dw->all_chan_mask);
406         channel_clear_bit(dw, MASK.BLOCK, dw->all_chan_mask);
407         channel_clear_bit(dw, MASK.ERROR, dw->all_chan_mask);
408
409         status = dma_readl(dw, STATUS_INT);
410         if (status) {
411                 dev_err(dw->dma.dev,
412                         "BUG: Unexpected interrupts pending: 0x%x\n",
413                         status);
414
415                 /* Try to recover */
416                 channel_clear_bit(dw, MASK.XFER, (1 << 8) - 1);
417                 channel_clear_bit(dw, MASK.BLOCK, (1 << 8) - 1);
418                 channel_clear_bit(dw, MASK.SRC_TRAN, (1 << 8) - 1);
419                 channel_clear_bit(dw, MASK.DST_TRAN, (1 << 8) - 1);
420                 channel_clear_bit(dw, MASK.ERROR, (1 << 8) - 1);
421         }
422
423         tasklet_schedule(&dw->tasklet);
424
425         return IRQ_HANDLED;
426 }
427
428 /*----------------------------------------------------------------------*/
429
430 static dma_cookie_t dwc_tx_submit(struct dma_async_tx_descriptor *tx)
431 {
432         struct dw_desc          *desc = txd_to_dw_desc(tx);
433         struct dw_dma_chan      *dwc = to_dw_dma_chan(tx->chan);
434         dma_cookie_t            cookie;
435
436         spin_lock_bh(&dwc->lock);
437         cookie = dwc_assign_cookie(dwc, desc);
438
439         /*
440          * REVISIT: We should attempt to chain as many descriptors as
441          * possible, perhaps even appending to those already submitted
442          * for DMA. But this is hard to do in a race-free manner.
443          */
444         if (list_empty(&dwc->active_list)) {
445                 dev_vdbg(&tx->chan->dev, "tx_submit: started %u\n",
446                                 desc->txd.cookie);
447                 dwc_dostart(dwc, desc);
448                 list_add_tail(&desc->desc_node, &dwc->active_list);
449         } else {
450                 dev_vdbg(&tx->chan->dev, "tx_submit: queued %u\n",
451                                 desc->txd.cookie);
452
453                 list_add_tail(&desc->desc_node, &dwc->queue);
454         }
455
456         spin_unlock_bh(&dwc->lock);
457
458         return cookie;
459 }
460
461 static struct dma_async_tx_descriptor *
462 dwc_prep_dma_memcpy(struct dma_chan *chan, dma_addr_t dest, dma_addr_t src,
463                 size_t len, unsigned long flags)
464 {
465         struct dw_dma_chan      *dwc = to_dw_dma_chan(chan);
466         struct dw_desc          *desc;
467         struct dw_desc          *first;
468         struct dw_desc          *prev;
469         size_t                  xfer_count;
470         size_t                  offset;
471         unsigned int            src_width;
472         unsigned int            dst_width;
473         u32                     ctllo;
474
475         dev_vdbg(&chan->dev, "prep_dma_memcpy d0x%x s0x%x l0x%zx f0x%lx\n",
476                         dest, src, len, flags);
477
478         if (unlikely(!len)) {
479                 dev_dbg(&chan->dev, "prep_dma_memcpy: length is zero!\n");
480                 return NULL;
481         }
482
483         /*
484          * We can be a lot more clever here, but this should take care
485          * of the most common optimization.
486          */
487         if (!((src | dest  | len) & 3))
488                 src_width = dst_width = 2;
489         else if (!((src | dest | len) & 1))
490                 src_width = dst_width = 1;
491         else
492                 src_width = dst_width = 0;
493
494         ctllo = DWC_DEFAULT_CTLLO
495                         | DWC_CTLL_DST_WIDTH(dst_width)
496                         | DWC_CTLL_SRC_WIDTH(src_width)
497                         | DWC_CTLL_DST_INC
498                         | DWC_CTLL_SRC_INC
499                         | DWC_CTLL_FC_M2M;
500         prev = first = NULL;
501
502         for (offset = 0; offset < len; offset += xfer_count << src_width) {
503                 xfer_count = min_t(size_t, (len - offset) >> src_width,
504                                 DWC_MAX_COUNT);
505
506                 desc = dwc_desc_get(dwc);
507                 if (!desc)
508                         goto err_desc_get;
509
510                 desc->lli.sar = src + offset;
511                 desc->lli.dar = dest + offset;
512                 desc->lli.ctllo = ctllo;
513                 desc->lli.ctlhi = xfer_count;
514
515                 if (!first) {
516                         first = desc;
517                 } else {
518                         prev->lli.llp = desc->txd.phys;
519                         dma_sync_single_for_device(chan->dev.parent,
520                                         prev->txd.phys, sizeof(prev->lli),
521                                         DMA_TO_DEVICE);
522                         list_add_tail(&desc->desc_node,
523                                         &first->txd.tx_list);
524                 }
525                 prev = desc;
526         }
527
528
529         if (flags & DMA_PREP_INTERRUPT)
530                 /* Trigger interrupt after last block */
531                 prev->lli.ctllo |= DWC_CTLL_INT_EN;
532
533         prev->lli.llp = 0;
534         dma_sync_single_for_device(chan->dev.parent,
535                         prev->txd.phys, sizeof(prev->lli),
536                         DMA_TO_DEVICE);
537
538         first->txd.flags = flags;
539         first->len = len;
540
541         return &first->txd;
542
543 err_desc_get:
544         dwc_desc_put(dwc, first);
545         return NULL;
546 }
547
548 static struct dma_async_tx_descriptor *
549 dwc_prep_slave_sg(struct dma_chan *chan, struct scatterlist *sgl,
550                 unsigned int sg_len, enum dma_data_direction direction,
551                 unsigned long flags)
552 {
553         struct dw_dma_chan      *dwc = to_dw_dma_chan(chan);
554         struct dw_dma_slave     *dws = dwc->dws;
555         struct dw_desc          *prev;
556         struct dw_desc          *first;
557         u32                     ctllo;
558         dma_addr_t              reg;
559         unsigned int            reg_width;
560         unsigned int            mem_width;
561         unsigned int            i;
562         struct scatterlist      *sg;
563         size_t                  total_len = 0;
564
565         dev_vdbg(&chan->dev, "prep_dma_slave\n");
566
567         if (unlikely(!dws || !sg_len))
568                 return NULL;
569
570         reg_width = dws->slave.reg_width;
571         prev = first = NULL;
572
573         sg_len = dma_map_sg(chan->dev.parent, sgl, sg_len, direction);
574
575         switch (direction) {
576         case DMA_TO_DEVICE:
577                 ctllo = (DWC_DEFAULT_CTLLO
578                                 | DWC_CTLL_DST_WIDTH(reg_width)
579                                 | DWC_CTLL_DST_FIX
580                                 | DWC_CTLL_SRC_INC
581                                 | DWC_CTLL_FC_M2P);
582                 reg = dws->slave.tx_reg;
583                 for_each_sg(sgl, sg, sg_len, i) {
584                         struct dw_desc  *desc;
585                         u32             len;
586                         u32             mem;
587
588                         desc = dwc_desc_get(dwc);
589                         if (!desc) {
590                                 dev_err(&chan->dev,
591                                         "not enough descriptors available\n");
592                                 goto err_desc_get;
593                         }
594
595                         mem = sg_phys(sg);
596                         len = sg_dma_len(sg);
597                         mem_width = 2;
598                         if (unlikely(mem & 3 || len & 3))
599                                 mem_width = 0;
600
601                         desc->lli.sar = mem;
602                         desc->lli.dar = reg;
603                         desc->lli.ctllo = ctllo | DWC_CTLL_SRC_WIDTH(mem_width);
604                         desc->lli.ctlhi = len >> mem_width;
605
606                         if (!first) {
607                                 first = desc;
608                         } else {
609                                 prev->lli.llp = desc->txd.phys;
610                                 dma_sync_single_for_device(chan->dev.parent,
611                                                 prev->txd.phys,
612                                                 sizeof(prev->lli),
613                                                 DMA_TO_DEVICE);
614                                 list_add_tail(&desc->desc_node,
615                                                 &first->txd.tx_list);
616                         }
617                         prev = desc;
618                         total_len += len;
619                 }
620                 break;
621         case DMA_FROM_DEVICE:
622                 ctllo = (DWC_DEFAULT_CTLLO
623                                 | DWC_CTLL_SRC_WIDTH(reg_width)
624                                 | DWC_CTLL_DST_INC
625                                 | DWC_CTLL_SRC_FIX
626                                 | DWC_CTLL_FC_P2M);
627
628                 reg = dws->slave.rx_reg;
629                 for_each_sg(sgl, sg, sg_len, i) {
630                         struct dw_desc  *desc;
631                         u32             len;
632                         u32             mem;
633
634                         desc = dwc_desc_get(dwc);
635                         if (!desc) {
636                                 dev_err(&chan->dev,
637                                         "not enough descriptors available\n");
638                                 goto err_desc_get;
639                         }
640
641                         mem = sg_phys(sg);
642                         len = sg_dma_len(sg);
643                         mem_width = 2;
644                         if (unlikely(mem & 3 || len & 3))
645                                 mem_width = 0;
646
647                         desc->lli.sar = reg;
648                         desc->lli.dar = mem;
649                         desc->lli.ctllo = ctllo | DWC_CTLL_DST_WIDTH(mem_width);
650                         desc->lli.ctlhi = len >> reg_width;
651
652                         if (!first) {
653                                 first = desc;
654                         } else {
655                                 prev->lli.llp = desc->txd.phys;
656                                 dma_sync_single_for_device(chan->dev.parent,
657                                                 prev->txd.phys,
658                                                 sizeof(prev->lli),
659                                                 DMA_TO_DEVICE);
660                                 list_add_tail(&desc->desc_node,
661                                                 &first->txd.tx_list);
662                         }
663                         prev = desc;
664                         total_len += len;
665                 }
666                 break;
667         default:
668                 return NULL;
669         }
670
671         if (flags & DMA_PREP_INTERRUPT)
672                 /* Trigger interrupt after last block */
673                 prev->lli.ctllo |= DWC_CTLL_INT_EN;
674
675         prev->lli.llp = 0;
676         dma_sync_single_for_device(chan->dev.parent,
677                         prev->txd.phys, sizeof(prev->lli),
678                         DMA_TO_DEVICE);
679
680         first->len = total_len;
681
682         return &first->txd;
683
684 err_desc_get:
685         dwc_desc_put(dwc, first);
686         return NULL;
687 }
688
689 static void dwc_terminate_all(struct dma_chan *chan)
690 {
691         struct dw_dma_chan      *dwc = to_dw_dma_chan(chan);
692         struct dw_dma           *dw = to_dw_dma(chan->device);
693         struct dw_desc          *desc, *_desc;
694         LIST_HEAD(list);
695
696         /*
697          * This is only called when something went wrong elsewhere, so
698          * we don't really care about the data. Just disable the
699          * channel. We still have to poll the channel enable bit due
700          * to AHB/HSB limitations.
701          */
702         spin_lock_bh(&dwc->lock);
703
704         channel_clear_bit(dw, CH_EN, dwc->mask);
705
706         while (dma_readl(dw, CH_EN) & dwc->mask)
707                 cpu_relax();
708
709         /* active_list entries will end up before queued entries */
710         list_splice_init(&dwc->queue, &list);
711         list_splice_init(&dwc->active_list, &list);
712
713         spin_unlock_bh(&dwc->lock);
714
715         /* Flush all pending and queued descriptors */
716         list_for_each_entry_safe(desc, _desc, &list, desc_node)
717                 dwc_descriptor_complete(dwc, desc);
718 }
719
720 static enum dma_status
721 dwc_is_tx_complete(struct dma_chan *chan,
722                 dma_cookie_t cookie,
723                 dma_cookie_t *done, dma_cookie_t *used)
724 {
725         struct dw_dma_chan      *dwc = to_dw_dma_chan(chan);
726         dma_cookie_t            last_used;
727         dma_cookie_t            last_complete;
728         int                     ret;
729
730         last_complete = dwc->completed;
731         last_used = chan->cookie;
732
733         ret = dma_async_is_complete(cookie, last_complete, last_used);
734         if (ret != DMA_SUCCESS) {
735                 dwc_scan_descriptors(to_dw_dma(chan->device), dwc);
736
737                 last_complete = dwc->completed;
738                 last_used = chan->cookie;
739
740                 ret = dma_async_is_complete(cookie, last_complete, last_used);
741         }
742
743         if (done)
744                 *done = last_complete;
745         if (used)
746                 *used = last_used;
747
748         return ret;
749 }
750
751 static void dwc_issue_pending(struct dma_chan *chan)
752 {
753         struct dw_dma_chan      *dwc = to_dw_dma_chan(chan);
754
755         spin_lock_bh(&dwc->lock);
756         if (!list_empty(&dwc->queue))
757                 dwc_scan_descriptors(to_dw_dma(chan->device), dwc);
758         spin_unlock_bh(&dwc->lock);
759 }
760
761 static int dwc_alloc_chan_resources(struct dma_chan *chan,
762                 struct dma_client *client)
763 {
764         struct dw_dma_chan      *dwc = to_dw_dma_chan(chan);
765         struct dw_dma           *dw = to_dw_dma(chan->device);
766         struct dw_desc          *desc;
767         struct dma_slave        *slave;
768         struct dw_dma_slave     *dws;
769         int                     i;
770         u32                     cfghi;
771         u32                     cfglo;
772
773         dev_vdbg(&chan->dev, "alloc_chan_resources\n");
774
775         /* Channels doing slave DMA can only handle one client. */
776         if (dwc->dws || client->slave) {
777                 if (chan->client_count)
778                         return -EBUSY;
779         }
780
781         /* ASSERT:  channel is idle */
782         if (dma_readl(dw, CH_EN) & dwc->mask) {
783                 dev_dbg(&chan->dev, "DMA channel not idle?\n");
784                 return -EIO;
785         }
786
787         dwc->completed = chan->cookie = 1;
788
789         cfghi = DWC_CFGH_FIFO_MODE;
790         cfglo = 0;
791
792         slave = client->slave;
793         if (slave) {
794                 /*
795                  * We need controller-specific data to set up slave
796                  * transfers.
797                  */
798                 BUG_ON(!slave->dma_dev || slave->dma_dev != dw->dma.dev);
799
800                 dws = container_of(slave, struct dw_dma_slave, slave);
801
802                 dwc->dws = dws;
803                 cfghi = dws->cfg_hi;
804                 cfglo = dws->cfg_lo;
805         } else {
806                 dwc->dws = NULL;
807         }
808
809         channel_writel(dwc, CFG_LO, cfglo);
810         channel_writel(dwc, CFG_HI, cfghi);
811
812         /*
813          * NOTE: some controllers may have additional features that we
814          * need to initialize here, like "scatter-gather" (which
815          * doesn't mean what you think it means), and status writeback.
816          */
817
818         spin_lock_bh(&dwc->lock);
819         i = dwc->descs_allocated;
820         while (dwc->descs_allocated < NR_DESCS_PER_CHANNEL) {
821                 spin_unlock_bh(&dwc->lock);
822
823                 desc = kzalloc(sizeof(struct dw_desc), GFP_KERNEL);
824                 if (!desc) {
825                         dev_info(&chan->dev,
826                                 "only allocated %d descriptors\n", i);
827                         spin_lock_bh(&dwc->lock);
828                         break;
829                 }
830
831                 dma_async_tx_descriptor_init(&desc->txd, chan);
832                 desc->txd.tx_submit = dwc_tx_submit;
833                 desc->txd.flags = DMA_CTRL_ACK;
834                 INIT_LIST_HEAD(&desc->txd.tx_list);
835                 desc->txd.phys = dma_map_single(chan->dev.parent, &desc->lli,
836                                 sizeof(desc->lli), DMA_TO_DEVICE);
837                 dwc_desc_put(dwc, desc);
838
839                 spin_lock_bh(&dwc->lock);
840                 i = ++dwc->descs_allocated;
841         }
842
843         /* Enable interrupts */
844         channel_set_bit(dw, MASK.XFER, dwc->mask);
845         channel_set_bit(dw, MASK.BLOCK, dwc->mask);
846         channel_set_bit(dw, MASK.ERROR, dwc->mask);
847
848         spin_unlock_bh(&dwc->lock);
849
850         dev_dbg(&chan->dev,
851                 "alloc_chan_resources allocated %d descriptors\n", i);
852
853         return i;
854 }
855
856 static void dwc_free_chan_resources(struct dma_chan *chan)
857 {
858         struct dw_dma_chan      *dwc = to_dw_dma_chan(chan);
859         struct dw_dma           *dw = to_dw_dma(chan->device);
860         struct dw_desc          *desc, *_desc;
861         LIST_HEAD(list);
862
863         dev_dbg(&chan->dev, "free_chan_resources (descs allocated=%u)\n",
864                         dwc->descs_allocated);
865
866         /* ASSERT:  channel is idle */
867         BUG_ON(!list_empty(&dwc->active_list));
868         BUG_ON(!list_empty(&dwc->queue));
869         BUG_ON(dma_readl(to_dw_dma(chan->device), CH_EN) & dwc->mask);
870
871         spin_lock_bh(&dwc->lock);
872         list_splice_init(&dwc->free_list, &list);
873         dwc->descs_allocated = 0;
874         dwc->dws = NULL;
875
876         /* Disable interrupts */
877         channel_clear_bit(dw, MASK.XFER, dwc->mask);
878         channel_clear_bit(dw, MASK.BLOCK, dwc->mask);
879         channel_clear_bit(dw, MASK.ERROR, dwc->mask);
880
881         spin_unlock_bh(&dwc->lock);
882
883         list_for_each_entry_safe(desc, _desc, &list, desc_node) {
884                 dev_vdbg(&chan->dev, "  freeing descriptor %p\n", desc);
885                 dma_unmap_single(chan->dev.parent, desc->txd.phys,
886                                 sizeof(desc->lli), DMA_TO_DEVICE);
887                 kfree(desc);
888         }
889
890         dev_vdbg(&chan->dev, "free_chan_resources done\n");
891 }
892
893 /*----------------------------------------------------------------------*/
894
895 static void dw_dma_off(struct dw_dma *dw)
896 {
897         dma_writel(dw, CFG, 0);
898
899         channel_clear_bit(dw, MASK.XFER, dw->all_chan_mask);
900         channel_clear_bit(dw, MASK.BLOCK, dw->all_chan_mask);
901         channel_clear_bit(dw, MASK.SRC_TRAN, dw->all_chan_mask);
902         channel_clear_bit(dw, MASK.DST_TRAN, dw->all_chan_mask);
903         channel_clear_bit(dw, MASK.ERROR, dw->all_chan_mask);
904
905         while (dma_readl(dw, CFG) & DW_CFG_DMA_EN)
906                 cpu_relax();
907 }
908
909 static int __init dw_probe(struct platform_device *pdev)
910 {
911         struct dw_dma_platform_data *pdata;
912         struct resource         *io;
913         struct dw_dma           *dw;
914         size_t                  size;
915         int                     irq;
916         int                     err;
917         int                     i;
918
919         pdata = pdev->dev.platform_data;
920         if (!pdata || pdata->nr_channels > DW_DMA_MAX_NR_CHANNELS)
921                 return -EINVAL;
922
923         io = platform_get_resource(pdev, IORESOURCE_MEM, 0);
924         if (!io)
925                 return -EINVAL;
926
927         irq = platform_get_irq(pdev, 0);
928         if (irq < 0)
929                 return irq;
930
931         size = sizeof(struct dw_dma);
932         size += pdata->nr_channels * sizeof(struct dw_dma_chan);
933         dw = kzalloc(size, GFP_KERNEL);
934         if (!dw)
935                 return -ENOMEM;
936
937         if (!request_mem_region(io->start, DW_REGLEN, pdev->dev.driver->name)) {
938                 err = -EBUSY;
939                 goto err_kfree;
940         }
941
942         memset(dw, 0, sizeof *dw);
943
944         dw->regs = ioremap(io->start, DW_REGLEN);
945         if (!dw->regs) {
946                 err = -ENOMEM;
947                 goto err_release_r;
948         }
949
950         dw->clk = clk_get(&pdev->dev, "hclk");
951         if (IS_ERR(dw->clk)) {
952                 err = PTR_ERR(dw->clk);
953                 goto err_clk;
954         }
955         clk_enable(dw->clk);
956
957         /* force dma off, just in case */
958         dw_dma_off(dw);
959
960         err = request_irq(irq, dw_dma_interrupt, 0, "dw_dmac", dw);
961         if (err)
962                 goto err_irq;
963
964         platform_set_drvdata(pdev, dw);
965
966         tasklet_init(&dw->tasklet, dw_dma_tasklet, (unsigned long)dw);
967
968         dw->all_chan_mask = (1 << pdata->nr_channels) - 1;
969
970         INIT_LIST_HEAD(&dw->dma.channels);
971         for (i = 0; i < pdata->nr_channels; i++, dw->dma.chancnt++) {
972                 struct dw_dma_chan      *dwc = &dw->chan[i];
973
974                 dwc->chan.device = &dw->dma;
975                 dwc->chan.cookie = dwc->completed = 1;
976                 dwc->chan.chan_id = i;
977                 list_add_tail(&dwc->chan.device_node, &dw->dma.channels);
978
979                 dwc->ch_regs = &__dw_regs(dw)->CHAN[i];
980                 spin_lock_init(&dwc->lock);
981                 dwc->mask = 1 << i;
982
983                 INIT_LIST_HEAD(&dwc->active_list);
984                 INIT_LIST_HEAD(&dwc->queue);
985                 INIT_LIST_HEAD(&dwc->free_list);
986
987                 channel_clear_bit(dw, CH_EN, dwc->mask);
988         }
989
990         /* Clear/disable all interrupts on all channels. */
991         dma_writel(dw, CLEAR.XFER, dw->all_chan_mask);
992         dma_writel(dw, CLEAR.BLOCK, dw->all_chan_mask);
993         dma_writel(dw, CLEAR.SRC_TRAN, dw->all_chan_mask);
994         dma_writel(dw, CLEAR.DST_TRAN, dw->all_chan_mask);
995         dma_writel(dw, CLEAR.ERROR, dw->all_chan_mask);
996
997         channel_clear_bit(dw, MASK.XFER, dw->all_chan_mask);
998         channel_clear_bit(dw, MASK.BLOCK, dw->all_chan_mask);
999         channel_clear_bit(dw, MASK.SRC_TRAN, dw->all_chan_mask);
1000         channel_clear_bit(dw, MASK.DST_TRAN, dw->all_chan_mask);
1001         channel_clear_bit(dw, MASK.ERROR, dw->all_chan_mask);
1002
1003         dma_cap_set(DMA_MEMCPY, dw->dma.cap_mask);
1004         dma_cap_set(DMA_SLAVE, dw->dma.cap_mask);
1005         dw->dma.dev = &pdev->dev;
1006         dw->dma.device_alloc_chan_resources = dwc_alloc_chan_resources;
1007         dw->dma.device_free_chan_resources = dwc_free_chan_resources;
1008
1009         dw->dma.device_prep_dma_memcpy = dwc_prep_dma_memcpy;
1010
1011         dw->dma.device_prep_slave_sg = dwc_prep_slave_sg;
1012         dw->dma.device_terminate_all = dwc_terminate_all;
1013
1014         dw->dma.device_is_tx_complete = dwc_is_tx_complete;
1015         dw->dma.device_issue_pending = dwc_issue_pending;
1016
1017         dma_writel(dw, CFG, DW_CFG_DMA_EN);
1018
1019         printk(KERN_INFO "%s: DesignWare DMA Controller, %d channels\n",
1020                         pdev->dev.bus_id, dw->dma.chancnt);
1021
1022         dma_async_device_register(&dw->dma);
1023
1024         return 0;
1025
1026 err_irq:
1027         clk_disable(dw->clk);
1028         clk_put(dw->clk);
1029 err_clk:
1030         iounmap(dw->regs);
1031         dw->regs = NULL;
1032 err_release_r:
1033         release_resource(io);
1034 err_kfree:
1035         kfree(dw);
1036         return err;
1037 }
1038
1039 static int __exit dw_remove(struct platform_device *pdev)
1040 {
1041         struct dw_dma           *dw = platform_get_drvdata(pdev);
1042         struct dw_dma_chan      *dwc, *_dwc;
1043         struct resource         *io;
1044
1045         dw_dma_off(dw);
1046         dma_async_device_unregister(&dw->dma);
1047
1048         free_irq(platform_get_irq(pdev, 0), dw);
1049         tasklet_kill(&dw->tasklet);
1050
1051         list_for_each_entry_safe(dwc, _dwc, &dw->dma.channels,
1052                         chan.device_node) {
1053                 list_del(&dwc->chan.device_node);
1054                 channel_clear_bit(dw, CH_EN, dwc->mask);
1055         }
1056
1057         clk_disable(dw->clk);
1058         clk_put(dw->clk);
1059
1060         iounmap(dw->regs);
1061         dw->regs = NULL;
1062
1063         io = platform_get_resource(pdev, IORESOURCE_MEM, 0);
1064         release_mem_region(io->start, DW_REGLEN);
1065
1066         kfree(dw);
1067
1068         return 0;
1069 }
1070
1071 static void dw_shutdown(struct platform_device *pdev)
1072 {
1073         struct dw_dma   *dw = platform_get_drvdata(pdev);
1074
1075         dw_dma_off(platform_get_drvdata(pdev));
1076         clk_disable(dw->clk);
1077 }
1078
1079 static int dw_suspend_late(struct platform_device *pdev, pm_message_t mesg)
1080 {
1081         struct dw_dma   *dw = platform_get_drvdata(pdev);
1082
1083         dw_dma_off(platform_get_drvdata(pdev));
1084         clk_disable(dw->clk);
1085         return 0;
1086 }
1087
1088 static int dw_resume_early(struct platform_device *pdev)
1089 {
1090         struct dw_dma   *dw = platform_get_drvdata(pdev);
1091
1092         clk_enable(dw->clk);
1093         dma_writel(dw, CFG, DW_CFG_DMA_EN);
1094         return 0;
1095
1096 }
1097
1098 static struct platform_driver dw_driver = {
1099         .remove         = __exit_p(dw_remove),
1100         .shutdown       = dw_shutdown,
1101         .suspend_late   = dw_suspend_late,
1102         .resume_early   = dw_resume_early,
1103         .driver = {
1104                 .name   = "dw_dmac",
1105         },
1106 };
1107
1108 static int __init dw_init(void)
1109 {
1110         return platform_driver_probe(&dw_driver, dw_probe);
1111 }
1112 module_init(dw_init);
1113
1114 static void __exit dw_exit(void)
1115 {
1116         platform_driver_unregister(&dw_driver);
1117 }
1118 module_exit(dw_exit);
1119
1120 MODULE_LICENSE("GPL v2");
1121 MODULE_DESCRIPTION("Synopsys DesignWare DMA Controller driver");
1122 MODULE_AUTHOR("Haavard Skinnemoen <haavard.skinnemoen@atmel.com>");