Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * offload engine driver for the Marvell XOR engine
4 * Copyright (C) 2007, 2008, Marvell International Ltd.
5 */
6
7#include <linux/init.h>
8#include <linux/slab.h>
9#include <linux/delay.h>
10#include <linux/dma-mapping.h>
11#include <linux/spinlock.h>
12#include <linux/interrupt.h>
13#include <linux/of_device.h>
14#include <linux/platform_device.h>
15#include <linux/memory.h>
16#include <linux/clk.h>
17#include <linux/of.h>
18#include <linux/of_irq.h>
19#include <linux/irqdomain.h>
20#include <linux/cpumask.h>
21#include <linux/platform_data/dma-mv_xor.h>
22
23#include "dmaengine.h"
24#include "mv_xor.h"
25
26enum mv_xor_type {
27 XOR_ORION,
28 XOR_ARMADA_38X,
29 XOR_ARMADA_37XX,
30};
31
32enum mv_xor_mode {
33 XOR_MODE_IN_REG,
34 XOR_MODE_IN_DESC,
35};
36
37static void mv_xor_issue_pending(struct dma_chan *chan);
38
39#define to_mv_xor_chan(chan) \
40 container_of(chan, struct mv_xor_chan, dmachan)
41
42#define to_mv_xor_slot(tx) \
43 container_of(tx, struct mv_xor_desc_slot, async_tx)
44
45#define mv_chan_to_devp(chan) \
46 ((chan)->dmadev.dev)
47
48static void mv_desc_init(struct mv_xor_desc_slot *desc,
49 dma_addr_t addr, u32 byte_count,
50 enum dma_ctrl_flags flags)
51{
52 struct mv_xor_desc *hw_desc = desc->hw_desc;
53
54 hw_desc->status = XOR_DESC_DMA_OWNED;
55 hw_desc->phy_next_desc = 0;
56 /* Enable end-of-descriptor interrupts only for DMA_PREP_INTERRUPT */
57 hw_desc->desc_command = (flags & DMA_PREP_INTERRUPT) ?
58 XOR_DESC_EOD_INT_EN : 0;
59 hw_desc->phy_dest_addr = addr;
60 hw_desc->byte_count = byte_count;
61}
62
63static void mv_desc_set_mode(struct mv_xor_desc_slot *desc)
64{
65 struct mv_xor_desc *hw_desc = desc->hw_desc;
66
67 switch (desc->type) {
68 case DMA_XOR:
69 case DMA_INTERRUPT:
70 hw_desc->desc_command |= XOR_DESC_OPERATION_XOR;
71 break;
72 case DMA_MEMCPY:
73 hw_desc->desc_command |= XOR_DESC_OPERATION_MEMCPY;
74 break;
75 default:
76 BUG();
77 return;
78 }
79}
80
81static void mv_desc_set_next_desc(struct mv_xor_desc_slot *desc,
82 u32 next_desc_addr)
83{
84 struct mv_xor_desc *hw_desc = desc->hw_desc;
85 BUG_ON(hw_desc->phy_next_desc);
86 hw_desc->phy_next_desc = next_desc_addr;
87}
88
89static void mv_desc_set_src_addr(struct mv_xor_desc_slot *desc,
90 int index, dma_addr_t addr)
91{
92 struct mv_xor_desc *hw_desc = desc->hw_desc;
93 hw_desc->phy_src_addr[mv_phy_src_idx(index)] = addr;
94 if (desc->type == DMA_XOR)
95 hw_desc->desc_command |= (1 << index);
96}
97
98static u32 mv_chan_get_current_desc(struct mv_xor_chan *chan)
99{
100 return readl_relaxed(XOR_CURR_DESC(chan));
101}
102
103static void mv_chan_set_next_descriptor(struct mv_xor_chan *chan,
104 u32 next_desc_addr)
105{
106 writel_relaxed(next_desc_addr, XOR_NEXT_DESC(chan));
107}
108
109static void mv_chan_unmask_interrupts(struct mv_xor_chan *chan)
110{
111 u32 val = readl_relaxed(XOR_INTR_MASK(chan));
112 val |= XOR_INTR_MASK_VALUE << (chan->idx * 16);
113 writel_relaxed(val, XOR_INTR_MASK(chan));
114}
115
116static u32 mv_chan_get_intr_cause(struct mv_xor_chan *chan)
117{
118 u32 intr_cause = readl_relaxed(XOR_INTR_CAUSE(chan));
119 intr_cause = (intr_cause >> (chan->idx * 16)) & 0xFFFF;
120 return intr_cause;
121}
122
123static void mv_chan_clear_eoc_cause(struct mv_xor_chan *chan)
124{
125 u32 val;
126
127 val = XOR_INT_END_OF_DESC | XOR_INT_END_OF_CHAIN | XOR_INT_STOPPED;
128 val = ~(val << (chan->idx * 16));
129 dev_dbg(mv_chan_to_devp(chan), "%s, val 0x%08x\n", __func__, val);
130 writel_relaxed(val, XOR_INTR_CAUSE(chan));
131}
132
133static void mv_chan_clear_err_status(struct mv_xor_chan *chan)
134{
135 u32 val = 0xFFFF0000 >> (chan->idx * 16);
136 writel_relaxed(val, XOR_INTR_CAUSE(chan));
137}
138
139static void mv_chan_set_mode(struct mv_xor_chan *chan,
140 u32 op_mode)
141{
142 u32 config = readl_relaxed(XOR_CONFIG(chan));
143
144 config &= ~0x7;
145 config |= op_mode;
146
147#if defined(__BIG_ENDIAN)
148 config |= XOR_DESCRIPTOR_SWAP;
149#else
150 config &= ~XOR_DESCRIPTOR_SWAP;
151#endif
152
153 writel_relaxed(config, XOR_CONFIG(chan));
154}
155
156static void mv_chan_activate(struct mv_xor_chan *chan)
157{
158 dev_dbg(mv_chan_to_devp(chan), " activate chan.\n");
159
160 /* writel ensures all descriptors are flushed before activation */
161 writel(BIT(0), XOR_ACTIVATION(chan));
162}
163
164static char mv_chan_is_busy(struct mv_xor_chan *chan)
165{
166 u32 state = readl_relaxed(XOR_ACTIVATION(chan));
167
168 state = (state >> 4) & 0x3;
169
170 return (state == 1) ? 1 : 0;
171}
172
173/*
174 * mv_chan_start_new_chain - program the engine to operate on new
175 * chain headed by sw_desc
176 * Caller must hold &mv_chan->lock while calling this function
177 */
178static void mv_chan_start_new_chain(struct mv_xor_chan *mv_chan,
179 struct mv_xor_desc_slot *sw_desc)
180{
181 dev_dbg(mv_chan_to_devp(mv_chan), "%s %d: sw_desc %p\n",
182 __func__, __LINE__, sw_desc);
183
184 /* set the hardware chain */
185 mv_chan_set_next_descriptor(mv_chan, sw_desc->async_tx.phys);
186
187 mv_chan->pending++;
188 mv_xor_issue_pending(&mv_chan->dmachan);
189}
190
191static dma_cookie_t
192mv_desc_run_tx_complete_actions(struct mv_xor_desc_slot *desc,
193 struct mv_xor_chan *mv_chan,
194 dma_cookie_t cookie)
195{
196 BUG_ON(desc->async_tx.cookie < 0);
197
198 if (desc->async_tx.cookie > 0) {
199 cookie = desc->async_tx.cookie;
200
201 dma_descriptor_unmap(&desc->async_tx);
202 /* call the callback (must not sleep or submit new
203 * operations to this channel)
204 */
205 dmaengine_desc_get_callback_invoke(&desc->async_tx, NULL);
206 }
207
208 /* run dependent operations */
209 dma_run_dependencies(&desc->async_tx);
210
211 return cookie;
212}
213
214static int
215mv_chan_clean_completed_slots(struct mv_xor_chan *mv_chan)
216{
217 struct mv_xor_desc_slot *iter, *_iter;
218
219 dev_dbg(mv_chan_to_devp(mv_chan), "%s %d\n", __func__, __LINE__);
220 list_for_each_entry_safe(iter, _iter, &mv_chan->completed_slots,
221 node) {
222
223 if (async_tx_test_ack(&iter->async_tx)) {
224 list_move_tail(&iter->node, &mv_chan->free_slots);
225 if (!list_empty(&iter->sg_tx_list)) {
226 list_splice_tail_init(&iter->sg_tx_list,
227 &mv_chan->free_slots);
228 }
229 }
230 }
231 return 0;
232}
233
234static int
235mv_desc_clean_slot(struct mv_xor_desc_slot *desc,
236 struct mv_xor_chan *mv_chan)
237{
238 dev_dbg(mv_chan_to_devp(mv_chan), "%s %d: desc %p flags %d\n",
239 __func__, __LINE__, desc, desc->async_tx.flags);
240
241 /* the client is allowed to attach dependent operations
242 * until 'ack' is set
243 */
244 if (!async_tx_test_ack(&desc->async_tx)) {
245 /* move this slot to the completed_slots */
246 list_move_tail(&desc->node, &mv_chan->completed_slots);
247 if (!list_empty(&desc->sg_tx_list)) {
248 list_splice_tail_init(&desc->sg_tx_list,
249 &mv_chan->completed_slots);
250 }
251 } else {
252 list_move_tail(&desc->node, &mv_chan->free_slots);
253 if (!list_empty(&desc->sg_tx_list)) {
254 list_splice_tail_init(&desc->sg_tx_list,
255 &mv_chan->free_slots);
256 }
257 }
258
259 return 0;
260}
261
262/* This function must be called with the mv_xor_chan spinlock held */
263static void mv_chan_slot_cleanup(struct mv_xor_chan *mv_chan)
264{
265 struct mv_xor_desc_slot *iter, *_iter;
266 dma_cookie_t cookie = 0;
267 int busy = mv_chan_is_busy(mv_chan);
268 u32 current_desc = mv_chan_get_current_desc(mv_chan);
269 int current_cleaned = 0;
270 struct mv_xor_desc *hw_desc;
271
272 dev_dbg(mv_chan_to_devp(mv_chan), "%s %d\n", __func__, __LINE__);
273 dev_dbg(mv_chan_to_devp(mv_chan), "current_desc %x\n", current_desc);
274 mv_chan_clean_completed_slots(mv_chan);
275
276 /* free completed slots from the chain starting with
277 * the oldest descriptor
278 */
279
280 list_for_each_entry_safe(iter, _iter, &mv_chan->chain,
281 node) {
282
283 /* clean finished descriptors */
284 hw_desc = iter->hw_desc;
285 if (hw_desc->status & XOR_DESC_SUCCESS) {
286 cookie = mv_desc_run_tx_complete_actions(iter, mv_chan,
287 cookie);
288
289 /* done processing desc, clean slot */
290 mv_desc_clean_slot(iter, mv_chan);
291
292 /* break if we did cleaned the current */
293 if (iter->async_tx.phys == current_desc) {
294 current_cleaned = 1;
295 break;
296 }
297 } else {
298 if (iter->async_tx.phys == current_desc) {
299 current_cleaned = 0;
300 break;
301 }
302 }
303 }
304
305 if ((busy == 0) && !list_empty(&mv_chan->chain)) {
306 if (current_cleaned) {
307 /*
308 * current descriptor cleaned and removed, run
309 * from list head
310 */
311 iter = list_entry(mv_chan->chain.next,
312 struct mv_xor_desc_slot,
313 node);
314 mv_chan_start_new_chain(mv_chan, iter);
315 } else {
316 if (!list_is_last(&iter->node, &mv_chan->chain)) {
317 /*
318 * descriptors are still waiting after
319 * current, trigger them
320 */
321 iter = list_entry(iter->node.next,
322 struct mv_xor_desc_slot,
323 node);
324 mv_chan_start_new_chain(mv_chan, iter);
325 } else {
326 /*
327 * some descriptors are still waiting
328 * to be cleaned
329 */
330 tasklet_schedule(&mv_chan->irq_tasklet);
331 }
332 }
333 }
334
335 if (cookie > 0)
336 mv_chan->dmachan.completed_cookie = cookie;
337}
338
339static void mv_xor_tasklet(unsigned long data)
340{
341 struct mv_xor_chan *chan = (struct mv_xor_chan *) data;
342
343 spin_lock(&chan->lock);
344 mv_chan_slot_cleanup(chan);
345 spin_unlock(&chan->lock);
346}
347
348static struct mv_xor_desc_slot *
349mv_chan_alloc_slot(struct mv_xor_chan *mv_chan)
350{
351 struct mv_xor_desc_slot *iter;
352
353 spin_lock_bh(&mv_chan->lock);
354
355 if (!list_empty(&mv_chan->free_slots)) {
356 iter = list_first_entry(&mv_chan->free_slots,
357 struct mv_xor_desc_slot,
358 node);
359
360 list_move_tail(&iter->node, &mv_chan->allocated_slots);
361
362 spin_unlock_bh(&mv_chan->lock);
363
364 /* pre-ack descriptor */
365 async_tx_ack(&iter->async_tx);
366 iter->async_tx.cookie = -EBUSY;
367
368 return iter;
369
370 }
371
372 spin_unlock_bh(&mv_chan->lock);
373
374 /* try to free some slots if the allocation fails */
375 tasklet_schedule(&mv_chan->irq_tasklet);
376
377 return NULL;
378}
379
380/************************ DMA engine API functions ****************************/
381static dma_cookie_t
382mv_xor_tx_submit(struct dma_async_tx_descriptor *tx)
383{
384 struct mv_xor_desc_slot *sw_desc = to_mv_xor_slot(tx);
385 struct mv_xor_chan *mv_chan = to_mv_xor_chan(tx->chan);
386 struct mv_xor_desc_slot *old_chain_tail;
387 dma_cookie_t cookie;
388 int new_hw_chain = 1;
389
390 dev_dbg(mv_chan_to_devp(mv_chan),
391 "%s sw_desc %p: async_tx %p\n",
392 __func__, sw_desc, &sw_desc->async_tx);
393
394 spin_lock_bh(&mv_chan->lock);
395 cookie = dma_cookie_assign(tx);
396
397 if (list_empty(&mv_chan->chain))
398 list_move_tail(&sw_desc->node, &mv_chan->chain);
399 else {
400 new_hw_chain = 0;
401
402 old_chain_tail = list_entry(mv_chan->chain.prev,
403 struct mv_xor_desc_slot,
404 node);
405 list_move_tail(&sw_desc->node, &mv_chan->chain);
406
407 dev_dbg(mv_chan_to_devp(mv_chan), "Append to last desc %pa\n",
408 &old_chain_tail->async_tx.phys);
409
410 /* fix up the hardware chain */
411 mv_desc_set_next_desc(old_chain_tail, sw_desc->async_tx.phys);
412
413 /* if the channel is not busy */
414 if (!mv_chan_is_busy(mv_chan)) {
415 u32 current_desc = mv_chan_get_current_desc(mv_chan);
416 /*
417 * and the curren desc is the end of the chain before
418 * the append, then we need to start the channel
419 */
420 if (current_desc == old_chain_tail->async_tx.phys)
421 new_hw_chain = 1;
422 }
423 }
424
425 if (new_hw_chain)
426 mv_chan_start_new_chain(mv_chan, sw_desc);
427
428 spin_unlock_bh(&mv_chan->lock);
429
430 return cookie;
431}
432
433/* returns the number of allocated descriptors */
434static int mv_xor_alloc_chan_resources(struct dma_chan *chan)
435{
436 void *virt_desc;
437 dma_addr_t dma_desc;
438 int idx;
439 struct mv_xor_chan *mv_chan = to_mv_xor_chan(chan);
440 struct mv_xor_desc_slot *slot = NULL;
441 int num_descs_in_pool = MV_XOR_POOL_SIZE/MV_XOR_SLOT_SIZE;
442
443 /* Allocate descriptor slots */
444 idx = mv_chan->slots_allocated;
445 while (idx < num_descs_in_pool) {
446 slot = kzalloc(sizeof(*slot), GFP_KERNEL);
447 if (!slot) {
448 dev_info(mv_chan_to_devp(mv_chan),
449 "channel only initialized %d descriptor slots",
450 idx);
451 break;
452 }
453 virt_desc = mv_chan->dma_desc_pool_virt;
454 slot->hw_desc = virt_desc + idx * MV_XOR_SLOT_SIZE;
455
456 dma_async_tx_descriptor_init(&slot->async_tx, chan);
457 slot->async_tx.tx_submit = mv_xor_tx_submit;
458 INIT_LIST_HEAD(&slot->node);
459 INIT_LIST_HEAD(&slot->sg_tx_list);
460 dma_desc = mv_chan->dma_desc_pool;
461 slot->async_tx.phys = dma_desc + idx * MV_XOR_SLOT_SIZE;
462 slot->idx = idx++;
463
464 spin_lock_bh(&mv_chan->lock);
465 mv_chan->slots_allocated = idx;
466 list_add_tail(&slot->node, &mv_chan->free_slots);
467 spin_unlock_bh(&mv_chan->lock);
468 }
469
470 dev_dbg(mv_chan_to_devp(mv_chan),
471 "allocated %d descriptor slots\n",
472 mv_chan->slots_allocated);
473
474 return mv_chan->slots_allocated ? : -ENOMEM;
475}
476
477/*
478 * Check if source or destination is an PCIe/IO address (non-SDRAM) and add
479 * a new MBus window if necessary. Use a cache for these check so that
480 * the MMIO mapped registers don't have to be accessed for this check
481 * to speed up this process.
482 */
483static int mv_xor_add_io_win(struct mv_xor_chan *mv_chan, u32 addr)
484{
485 struct mv_xor_device *xordev = mv_chan->xordev;
486 void __iomem *base = mv_chan->mmr_high_base;
487 u32 win_enable;
488 u32 size;
489 u8 target, attr;
490 int ret;
491 int i;
492
493 /* Nothing needs to get done for the Armada 3700 */
494 if (xordev->xor_type == XOR_ARMADA_37XX)
495 return 0;
496
497 /*
498 * Loop over the cached windows to check, if the requested area
499 * is already mapped. If this the case, nothing needs to be done
500 * and we can return.
501 */
502 for (i = 0; i < WINDOW_COUNT; i++) {
503 if (addr >= xordev->win_start[i] &&
504 addr <= xordev->win_end[i]) {
505 /* Window is already mapped */
506 return 0;
507 }
508 }
509
510 /*
511 * The window is not mapped, so we need to create the new mapping
512 */
513
514 /* If no IO window is found that addr has to be located in SDRAM */
515 ret = mvebu_mbus_get_io_win_info(addr, &size, &target, &attr);
516 if (ret < 0)
517 return 0;
518
519 /*
520 * Mask the base addr 'addr' according to 'size' read back from the
521 * MBus window. Otherwise we might end up with an address located
522 * somewhere in the middle of this area here.
523 */
524 size -= 1;
525 addr &= ~size;
526
527 /*
528 * Reading one of both enabled register is enough, as they are always
529 * programmed to the identical values
530 */
531 win_enable = readl(base + WINDOW_BAR_ENABLE(0));
532
533 /* Set 'i' to the first free window to write the new values to */
534 i = ffs(~win_enable) - 1;
535 if (i >= WINDOW_COUNT)
536 return -ENOMEM;
537
538 writel((addr & 0xffff0000) | (attr << 8) | target,
539 base + WINDOW_BASE(i));
540 writel(size & 0xffff0000, base + WINDOW_SIZE(i));
541
542 /* Fill the caching variables for later use */
543 xordev->win_start[i] = addr;
544 xordev->win_end[i] = addr + size;
545
546 win_enable |= (1 << i);
547 win_enable |= 3 << (16 + (2 * i));
548 writel(win_enable, base + WINDOW_BAR_ENABLE(0));
549 writel(win_enable, base + WINDOW_BAR_ENABLE(1));
550
551 return 0;
552}
553
554static struct dma_async_tx_descriptor *
555mv_xor_prep_dma_xor(struct dma_chan *chan, dma_addr_t dest, dma_addr_t *src,
556 unsigned int src_cnt, size_t len, unsigned long flags)
557{
558 struct mv_xor_chan *mv_chan = to_mv_xor_chan(chan);
559 struct mv_xor_desc_slot *sw_desc;
560 int ret;
561
562 if (unlikely(len < MV_XOR_MIN_BYTE_COUNT))
563 return NULL;
564
565 BUG_ON(len > MV_XOR_MAX_BYTE_COUNT);
566
567 dev_dbg(mv_chan_to_devp(mv_chan),
568 "%s src_cnt: %d len: %zu dest %pad flags: %ld\n",
569 __func__, src_cnt, len, &dest, flags);
570
571 /* Check if a new window needs to get added for 'dest' */
572 ret = mv_xor_add_io_win(mv_chan, dest);
573 if (ret)
574 return NULL;
575
576 sw_desc = mv_chan_alloc_slot(mv_chan);
577 if (sw_desc) {
578 sw_desc->type = DMA_XOR;
579 sw_desc->async_tx.flags = flags;
580 mv_desc_init(sw_desc, dest, len, flags);
581 if (mv_chan->op_in_desc == XOR_MODE_IN_DESC)
582 mv_desc_set_mode(sw_desc);
583 while (src_cnt--) {
584 /* Check if a new window needs to get added for 'src' */
585 ret = mv_xor_add_io_win(mv_chan, src[src_cnt]);
586 if (ret)
587 return NULL;
588 mv_desc_set_src_addr(sw_desc, src_cnt, src[src_cnt]);
589 }
590 }
591
592 dev_dbg(mv_chan_to_devp(mv_chan),
593 "%s sw_desc %p async_tx %p \n",
594 __func__, sw_desc, &sw_desc->async_tx);
595 return sw_desc ? &sw_desc->async_tx : NULL;
596}
597
598static struct dma_async_tx_descriptor *
599mv_xor_prep_dma_memcpy(struct dma_chan *chan, dma_addr_t dest, dma_addr_t src,
600 size_t len, unsigned long flags)
601{
602 /*
603 * A MEMCPY operation is identical to an XOR operation with only
604 * a single source address.
605 */
606 return mv_xor_prep_dma_xor(chan, dest, &src, 1, len, flags);
607}
608
609static struct dma_async_tx_descriptor *
610mv_xor_prep_dma_interrupt(struct dma_chan *chan, unsigned long flags)
611{
612 struct mv_xor_chan *mv_chan = to_mv_xor_chan(chan);
613 dma_addr_t src, dest;
614 size_t len;
615
616 src = mv_chan->dummy_src_addr;
617 dest = mv_chan->dummy_dst_addr;
618 len = MV_XOR_MIN_BYTE_COUNT;
619
620 /*
621 * We implement the DMA_INTERRUPT operation as a minimum sized
622 * XOR operation with a single dummy source address.
623 */
624 return mv_xor_prep_dma_xor(chan, dest, &src, 1, len, flags);
625}
626
627static void mv_xor_free_chan_resources(struct dma_chan *chan)
628{
629 struct mv_xor_chan *mv_chan = to_mv_xor_chan(chan);
630 struct mv_xor_desc_slot *iter, *_iter;
631 int in_use_descs = 0;
632
633 spin_lock_bh(&mv_chan->lock);
634
635 mv_chan_slot_cleanup(mv_chan);
636
637 list_for_each_entry_safe(iter, _iter, &mv_chan->chain,
638 node) {
639 in_use_descs++;
640 list_move_tail(&iter->node, &mv_chan->free_slots);
641 }
642 list_for_each_entry_safe(iter, _iter, &mv_chan->completed_slots,
643 node) {
644 in_use_descs++;
645 list_move_tail(&iter->node, &mv_chan->free_slots);
646 }
647 list_for_each_entry_safe(iter, _iter, &mv_chan->allocated_slots,
648 node) {
649 in_use_descs++;
650 list_move_tail(&iter->node, &mv_chan->free_slots);
651 }
652 list_for_each_entry_safe_reverse(
653 iter, _iter, &mv_chan->free_slots, node) {
654 list_del(&iter->node);
655 kfree(iter);
656 mv_chan->slots_allocated--;
657 }
658
659 dev_dbg(mv_chan_to_devp(mv_chan), "%s slots_allocated %d\n",
660 __func__, mv_chan->slots_allocated);
661 spin_unlock_bh(&mv_chan->lock);
662
663 if (in_use_descs)
664 dev_err(mv_chan_to_devp(mv_chan),
665 "freeing %d in use descriptors!\n", in_use_descs);
666}
667
668/**
669 * mv_xor_status - poll the status of an XOR transaction
670 * @chan: XOR channel handle
671 * @cookie: XOR transaction identifier
672 * @txstate: XOR transactions state holder (or NULL)
673 */
674static enum dma_status mv_xor_status(struct dma_chan *chan,
675 dma_cookie_t cookie,
676 struct dma_tx_state *txstate)
677{
678 struct mv_xor_chan *mv_chan = to_mv_xor_chan(chan);
679 enum dma_status ret;
680
681 ret = dma_cookie_status(chan, cookie, txstate);
682 if (ret == DMA_COMPLETE)
683 return ret;
684
685 spin_lock_bh(&mv_chan->lock);
686 mv_chan_slot_cleanup(mv_chan);
687 spin_unlock_bh(&mv_chan->lock);
688
689 return dma_cookie_status(chan, cookie, txstate);
690}
691
692static void mv_chan_dump_regs(struct mv_xor_chan *chan)
693{
694 u32 val;
695
696 val = readl_relaxed(XOR_CONFIG(chan));
697 dev_err(mv_chan_to_devp(chan), "config 0x%08x\n", val);
698
699 val = readl_relaxed(XOR_ACTIVATION(chan));
700 dev_err(mv_chan_to_devp(chan), "activation 0x%08x\n", val);
701
702 val = readl_relaxed(XOR_INTR_CAUSE(chan));
703 dev_err(mv_chan_to_devp(chan), "intr cause 0x%08x\n", val);
704
705 val = readl_relaxed(XOR_INTR_MASK(chan));
706 dev_err(mv_chan_to_devp(chan), "intr mask 0x%08x\n", val);
707
708 val = readl_relaxed(XOR_ERROR_CAUSE(chan));
709 dev_err(mv_chan_to_devp(chan), "error cause 0x%08x\n", val);
710
711 val = readl_relaxed(XOR_ERROR_ADDR(chan));
712 dev_err(mv_chan_to_devp(chan), "error addr 0x%08x\n", val);
713}
714
715static void mv_chan_err_interrupt_handler(struct mv_xor_chan *chan,
716 u32 intr_cause)
717{
718 if (intr_cause & XOR_INT_ERR_DECODE) {
719 dev_dbg(mv_chan_to_devp(chan), "ignoring address decode error\n");
720 return;
721 }
722
723 dev_err(mv_chan_to_devp(chan), "error on chan %d. intr cause 0x%08x\n",
724 chan->idx, intr_cause);
725
726 mv_chan_dump_regs(chan);
727 WARN_ON(1);
728}
729
730static irqreturn_t mv_xor_interrupt_handler(int irq, void *data)
731{
732 struct mv_xor_chan *chan = data;
733 u32 intr_cause = mv_chan_get_intr_cause(chan);
734
735 dev_dbg(mv_chan_to_devp(chan), "intr cause %x\n", intr_cause);
736
737 if (intr_cause & XOR_INTR_ERRORS)
738 mv_chan_err_interrupt_handler(chan, intr_cause);
739
740 tasklet_schedule(&chan->irq_tasklet);
741
742 mv_chan_clear_eoc_cause(chan);
743
744 return IRQ_HANDLED;
745}
746
747static void mv_xor_issue_pending(struct dma_chan *chan)
748{
749 struct mv_xor_chan *mv_chan = to_mv_xor_chan(chan);
750
751 if (mv_chan->pending >= MV_XOR_THRESHOLD) {
752 mv_chan->pending = 0;
753 mv_chan_activate(mv_chan);
754 }
755}
756
757/*
758 * Perform a transaction to verify the HW works.
759 */
760
761static int mv_chan_memcpy_self_test(struct mv_xor_chan *mv_chan)
762{
763 int i, ret;
764 void *src, *dest;
765 dma_addr_t src_dma, dest_dma;
766 struct dma_chan *dma_chan;
767 dma_cookie_t cookie;
768 struct dma_async_tx_descriptor *tx;
769 struct dmaengine_unmap_data *unmap;
770 int err = 0;
771
772 src = kmalloc(PAGE_SIZE, GFP_KERNEL);
773 if (!src)
774 return -ENOMEM;
775
776 dest = kzalloc(PAGE_SIZE, GFP_KERNEL);
777 if (!dest) {
778 kfree(src);
779 return -ENOMEM;
780 }
781
782 /* Fill in src buffer */
783 for (i = 0; i < PAGE_SIZE; i++)
784 ((u8 *) src)[i] = (u8)i;
785
786 dma_chan = &mv_chan->dmachan;
787 if (mv_xor_alloc_chan_resources(dma_chan) < 1) {
788 err = -ENODEV;
789 goto out;
790 }
791
792 unmap = dmaengine_get_unmap_data(dma_chan->device->dev, 2, GFP_KERNEL);
793 if (!unmap) {
794 err = -ENOMEM;
795 goto free_resources;
796 }
797
798 src_dma = dma_map_page(dma_chan->device->dev, virt_to_page(src),
799 offset_in_page(src), PAGE_SIZE,
800 DMA_TO_DEVICE);
801 unmap->addr[0] = src_dma;
802
803 ret = dma_mapping_error(dma_chan->device->dev, src_dma);
804 if (ret) {
805 err = -ENOMEM;
806 goto free_resources;
807 }
808 unmap->to_cnt = 1;
809
810 dest_dma = dma_map_page(dma_chan->device->dev, virt_to_page(dest),
811 offset_in_page(dest), PAGE_SIZE,
812 DMA_FROM_DEVICE);
813 unmap->addr[1] = dest_dma;
814
815 ret = dma_mapping_error(dma_chan->device->dev, dest_dma);
816 if (ret) {
817 err = -ENOMEM;
818 goto free_resources;
819 }
820 unmap->from_cnt = 1;
821 unmap->len = PAGE_SIZE;
822
823 tx = mv_xor_prep_dma_memcpy(dma_chan, dest_dma, src_dma,
824 PAGE_SIZE, 0);
825 if (!tx) {
826 dev_err(dma_chan->device->dev,
827 "Self-test cannot prepare operation, disabling\n");
828 err = -ENODEV;
829 goto free_resources;
830 }
831
832 cookie = mv_xor_tx_submit(tx);
833 if (dma_submit_error(cookie)) {
834 dev_err(dma_chan->device->dev,
835 "Self-test submit error, disabling\n");
836 err = -ENODEV;
837 goto free_resources;
838 }
839
840 mv_xor_issue_pending(dma_chan);
841 async_tx_ack(tx);
842 msleep(1);
843
844 if (mv_xor_status(dma_chan, cookie, NULL) !=
845 DMA_COMPLETE) {
846 dev_err(dma_chan->device->dev,
847 "Self-test copy timed out, disabling\n");
848 err = -ENODEV;
849 goto free_resources;
850 }
851
852 dma_sync_single_for_cpu(dma_chan->device->dev, dest_dma,
853 PAGE_SIZE, DMA_FROM_DEVICE);
854 if (memcmp(src, dest, PAGE_SIZE)) {
855 dev_err(dma_chan->device->dev,
856 "Self-test copy failed compare, disabling\n");
857 err = -ENODEV;
858 goto free_resources;
859 }
860
861free_resources:
862 dmaengine_unmap_put(unmap);
863 mv_xor_free_chan_resources(dma_chan);
864out:
865 kfree(src);
866 kfree(dest);
867 return err;
868}
869
870#define MV_XOR_NUM_SRC_TEST 4 /* must be <= 15 */
871static int
872mv_chan_xor_self_test(struct mv_xor_chan *mv_chan)
873{
874 int i, src_idx, ret;
875 struct page *dest;
876 struct page *xor_srcs[MV_XOR_NUM_SRC_TEST];
877 dma_addr_t dma_srcs[MV_XOR_NUM_SRC_TEST];
878 dma_addr_t dest_dma;
879 struct dma_async_tx_descriptor *tx;
880 struct dmaengine_unmap_data *unmap;
881 struct dma_chan *dma_chan;
882 dma_cookie_t cookie;
883 u8 cmp_byte = 0;
884 u32 cmp_word;
885 int err = 0;
886 int src_count = MV_XOR_NUM_SRC_TEST;
887
888 for (src_idx = 0; src_idx < src_count; src_idx++) {
889 xor_srcs[src_idx] = alloc_page(GFP_KERNEL);
890 if (!xor_srcs[src_idx]) {
891 while (src_idx--)
892 __free_page(xor_srcs[src_idx]);
893 return -ENOMEM;
894 }
895 }
896
897 dest = alloc_page(GFP_KERNEL);
898 if (!dest) {
899 while (src_idx--)
900 __free_page(xor_srcs[src_idx]);
901 return -ENOMEM;
902 }
903
904 /* Fill in src buffers */
905 for (src_idx = 0; src_idx < src_count; src_idx++) {
906 u8 *ptr = page_address(xor_srcs[src_idx]);
907 for (i = 0; i < PAGE_SIZE; i++)
908 ptr[i] = (1 << src_idx);
909 }
910
911 for (src_idx = 0; src_idx < src_count; src_idx++)
912 cmp_byte ^= (u8) (1 << src_idx);
913
914 cmp_word = (cmp_byte << 24) | (cmp_byte << 16) |
915 (cmp_byte << 8) | cmp_byte;
916
917 memset(page_address(dest), 0, PAGE_SIZE);
918
919 dma_chan = &mv_chan->dmachan;
920 if (mv_xor_alloc_chan_resources(dma_chan) < 1) {
921 err = -ENODEV;
922 goto out;
923 }
924
925 unmap = dmaengine_get_unmap_data(dma_chan->device->dev, src_count + 1,
926 GFP_KERNEL);
927 if (!unmap) {
928 err = -ENOMEM;
929 goto free_resources;
930 }
931
932 /* test xor */
933 for (i = 0; i < src_count; i++) {
934 unmap->addr[i] = dma_map_page(dma_chan->device->dev, xor_srcs[i],
935 0, PAGE_SIZE, DMA_TO_DEVICE);
936 dma_srcs[i] = unmap->addr[i];
937 ret = dma_mapping_error(dma_chan->device->dev, unmap->addr[i]);
938 if (ret) {
939 err = -ENOMEM;
940 goto free_resources;
941 }
942 unmap->to_cnt++;
943 }
944
945 unmap->addr[src_count] = dma_map_page(dma_chan->device->dev, dest, 0, PAGE_SIZE,
946 DMA_FROM_DEVICE);
947 dest_dma = unmap->addr[src_count];
948 ret = dma_mapping_error(dma_chan->device->dev, unmap->addr[src_count]);
949 if (ret) {
950 err = -ENOMEM;
951 goto free_resources;
952 }
953 unmap->from_cnt = 1;
954 unmap->len = PAGE_SIZE;
955
956 tx = mv_xor_prep_dma_xor(dma_chan, dest_dma, dma_srcs,
957 src_count, PAGE_SIZE, 0);
958 if (!tx) {
959 dev_err(dma_chan->device->dev,
960 "Self-test cannot prepare operation, disabling\n");
961 err = -ENODEV;
962 goto free_resources;
963 }
964
965 cookie = mv_xor_tx_submit(tx);
966 if (dma_submit_error(cookie)) {
967 dev_err(dma_chan->device->dev,
968 "Self-test submit error, disabling\n");
969 err = -ENODEV;
970 goto free_resources;
971 }
972
973 mv_xor_issue_pending(dma_chan);
974 async_tx_ack(tx);
975 msleep(8);
976
977 if (mv_xor_status(dma_chan, cookie, NULL) !=
978 DMA_COMPLETE) {
979 dev_err(dma_chan->device->dev,
980 "Self-test xor timed out, disabling\n");
981 err = -ENODEV;
982 goto free_resources;
983 }
984
985 dma_sync_single_for_cpu(dma_chan->device->dev, dest_dma,
986 PAGE_SIZE, DMA_FROM_DEVICE);
987 for (i = 0; i < (PAGE_SIZE / sizeof(u32)); i++) {
988 u32 *ptr = page_address(dest);
989 if (ptr[i] != cmp_word) {
990 dev_err(dma_chan->device->dev,
991 "Self-test xor failed compare, disabling. index %d, data %x, expected %x\n",
992 i, ptr[i], cmp_word);
993 err = -ENODEV;
994 goto free_resources;
995 }
996 }
997
998free_resources:
999 dmaengine_unmap_put(unmap);
1000 mv_xor_free_chan_resources(dma_chan);
1001out:
1002 src_idx = src_count;
1003 while (src_idx--)
1004 __free_page(xor_srcs[src_idx]);
1005 __free_page(dest);
1006 return err;
1007}
1008
1009static int mv_xor_channel_remove(struct mv_xor_chan *mv_chan)
1010{
1011 struct dma_chan *chan, *_chan;
1012 struct device *dev = mv_chan->dmadev.dev;
1013
1014 dma_async_device_unregister(&mv_chan->dmadev);
1015
1016 dma_free_coherent(dev, MV_XOR_POOL_SIZE,
1017 mv_chan->dma_desc_pool_virt, mv_chan->dma_desc_pool);
1018 dma_unmap_single(dev, mv_chan->dummy_src_addr,
1019 MV_XOR_MIN_BYTE_COUNT, DMA_FROM_DEVICE);
1020 dma_unmap_single(dev, mv_chan->dummy_dst_addr,
1021 MV_XOR_MIN_BYTE_COUNT, DMA_TO_DEVICE);
1022
1023 list_for_each_entry_safe(chan, _chan, &mv_chan->dmadev.channels,
1024 device_node) {
1025 list_del(&chan->device_node);
1026 }
1027
1028 free_irq(mv_chan->irq, mv_chan);
1029
1030 return 0;
1031}
1032
1033static struct mv_xor_chan *
1034mv_xor_channel_add(struct mv_xor_device *xordev,
1035 struct platform_device *pdev,
1036 int idx, dma_cap_mask_t cap_mask, int irq)
1037{
1038 int ret = 0;
1039 struct mv_xor_chan *mv_chan;
1040 struct dma_device *dma_dev;
1041
1042 mv_chan = devm_kzalloc(&pdev->dev, sizeof(*mv_chan), GFP_KERNEL);
1043 if (!mv_chan)
1044 return ERR_PTR(-ENOMEM);
1045
1046 mv_chan->idx = idx;
1047 mv_chan->irq = irq;
1048 if (xordev->xor_type == XOR_ORION)
1049 mv_chan->op_in_desc = XOR_MODE_IN_REG;
1050 else
1051 mv_chan->op_in_desc = XOR_MODE_IN_DESC;
1052
1053 dma_dev = &mv_chan->dmadev;
1054 dma_dev->dev = &pdev->dev;
1055 mv_chan->xordev = xordev;
1056
1057 /*
1058 * These source and destination dummy buffers are used to implement
1059 * a DMA_INTERRUPT operation as a minimum-sized XOR operation.
1060 * Hence, we only need to map the buffers at initialization-time.
1061 */
1062 mv_chan->dummy_src_addr = dma_map_single(dma_dev->dev,
1063 mv_chan->dummy_src, MV_XOR_MIN_BYTE_COUNT, DMA_FROM_DEVICE);
1064 mv_chan->dummy_dst_addr = dma_map_single(dma_dev->dev,
1065 mv_chan->dummy_dst, MV_XOR_MIN_BYTE_COUNT, DMA_TO_DEVICE);
1066
1067 /* allocate coherent memory for hardware descriptors
1068 * note: writecombine gives slightly better performance, but
1069 * requires that we explicitly flush the writes
1070 */
1071 mv_chan->dma_desc_pool_virt =
1072 dma_alloc_wc(&pdev->dev, MV_XOR_POOL_SIZE, &mv_chan->dma_desc_pool,
1073 GFP_KERNEL);
1074 if (!mv_chan->dma_desc_pool_virt)
1075 return ERR_PTR(-ENOMEM);
1076
1077 /* discover transaction capabilites from the platform data */
1078 dma_dev->cap_mask = cap_mask;
1079
1080 INIT_LIST_HEAD(&dma_dev->channels);
1081
1082 /* set base routines */
1083 dma_dev->device_alloc_chan_resources = mv_xor_alloc_chan_resources;
1084 dma_dev->device_free_chan_resources = mv_xor_free_chan_resources;
1085 dma_dev->device_tx_status = mv_xor_status;
1086 dma_dev->device_issue_pending = mv_xor_issue_pending;
1087
1088 /* set prep routines based on capability */
1089 if (dma_has_cap(DMA_INTERRUPT, dma_dev->cap_mask))
1090 dma_dev->device_prep_dma_interrupt = mv_xor_prep_dma_interrupt;
1091 if (dma_has_cap(DMA_MEMCPY, dma_dev->cap_mask))
1092 dma_dev->device_prep_dma_memcpy = mv_xor_prep_dma_memcpy;
1093 if (dma_has_cap(DMA_XOR, dma_dev->cap_mask)) {
1094 dma_dev->max_xor = 8;
1095 dma_dev->device_prep_dma_xor = mv_xor_prep_dma_xor;
1096 }
1097
1098 mv_chan->mmr_base = xordev->xor_base;
1099 mv_chan->mmr_high_base = xordev->xor_high_base;
1100 tasklet_init(&mv_chan->irq_tasklet, mv_xor_tasklet, (unsigned long)
1101 mv_chan);
1102
1103 /* clear errors before enabling interrupts */
1104 mv_chan_clear_err_status(mv_chan);
1105
1106 ret = request_irq(mv_chan->irq, mv_xor_interrupt_handler,
1107 0, dev_name(&pdev->dev), mv_chan);
1108 if (ret)
1109 goto err_free_dma;
1110
1111 mv_chan_unmask_interrupts(mv_chan);
1112
1113 if (mv_chan->op_in_desc == XOR_MODE_IN_DESC)
1114 mv_chan_set_mode(mv_chan, XOR_OPERATION_MODE_IN_DESC);
1115 else
1116 mv_chan_set_mode(mv_chan, XOR_OPERATION_MODE_XOR);
1117
1118 spin_lock_init(&mv_chan->lock);
1119 INIT_LIST_HEAD(&mv_chan->chain);
1120 INIT_LIST_HEAD(&mv_chan->completed_slots);
1121 INIT_LIST_HEAD(&mv_chan->free_slots);
1122 INIT_LIST_HEAD(&mv_chan->allocated_slots);
1123 mv_chan->dmachan.device = dma_dev;
1124 dma_cookie_init(&mv_chan->dmachan);
1125
1126 list_add_tail(&mv_chan->dmachan.device_node, &dma_dev->channels);
1127
1128 if (dma_has_cap(DMA_MEMCPY, dma_dev->cap_mask)) {
1129 ret = mv_chan_memcpy_self_test(mv_chan);
1130 dev_dbg(&pdev->dev, "memcpy self test returned %d\n", ret);
1131 if (ret)
1132 goto err_free_irq;
1133 }
1134
1135 if (dma_has_cap(DMA_XOR, dma_dev->cap_mask)) {
1136 ret = mv_chan_xor_self_test(mv_chan);
1137 dev_dbg(&pdev->dev, "xor self test returned %d\n", ret);
1138 if (ret)
1139 goto err_free_irq;
1140 }
1141
1142 dev_info(&pdev->dev, "Marvell XOR (%s): ( %s%s%s)\n",
1143 mv_chan->op_in_desc ? "Descriptor Mode" : "Registers Mode",
1144 dma_has_cap(DMA_XOR, dma_dev->cap_mask) ? "xor " : "",
1145 dma_has_cap(DMA_MEMCPY, dma_dev->cap_mask) ? "cpy " : "",
1146 dma_has_cap(DMA_INTERRUPT, dma_dev->cap_mask) ? "intr " : "");
1147
1148 ret = dma_async_device_register(dma_dev);
1149 if (ret)
1150 goto err_free_irq;
1151
1152 return mv_chan;
1153
1154err_free_irq:
1155 free_irq(mv_chan->irq, mv_chan);
1156err_free_dma:
1157 dma_free_coherent(&pdev->dev, MV_XOR_POOL_SIZE,
1158 mv_chan->dma_desc_pool_virt, mv_chan->dma_desc_pool);
1159 return ERR_PTR(ret);
1160}
1161
1162static void
1163mv_xor_conf_mbus_windows(struct mv_xor_device *xordev,
1164 const struct mbus_dram_target_info *dram)
1165{
1166 void __iomem *base = xordev->xor_high_base;
1167 u32 win_enable = 0;
1168 int i;
1169
1170 for (i = 0; i < 8; i++) {
1171 writel(0, base + WINDOW_BASE(i));
1172 writel(0, base + WINDOW_SIZE(i));
1173 if (i < 4)
1174 writel(0, base + WINDOW_REMAP_HIGH(i));
1175 }
1176
1177 for (i = 0; i < dram->num_cs; i++) {
1178 const struct mbus_dram_window *cs = dram->cs + i;
1179
1180 writel((cs->base & 0xffff0000) |
1181 (cs->mbus_attr << 8) |
1182 dram->mbus_dram_target_id, base + WINDOW_BASE(i));
1183 writel((cs->size - 1) & 0xffff0000, base + WINDOW_SIZE(i));
1184
1185 /* Fill the caching variables for later use */
1186 xordev->win_start[i] = cs->base;
1187 xordev->win_end[i] = cs->base + cs->size - 1;
1188
1189 win_enable |= (1 << i);
1190 win_enable |= 3 << (16 + (2 * i));
1191 }
1192
1193 writel(win_enable, base + WINDOW_BAR_ENABLE(0));
1194 writel(win_enable, base + WINDOW_BAR_ENABLE(1));
1195 writel(0, base + WINDOW_OVERRIDE_CTRL(0));
1196 writel(0, base + WINDOW_OVERRIDE_CTRL(1));
1197}
1198
1199static void
1200mv_xor_conf_mbus_windows_a3700(struct mv_xor_device *xordev)
1201{
1202 void __iomem *base = xordev->xor_high_base;
1203 u32 win_enable = 0;
1204 int i;
1205
1206 for (i = 0; i < 8; i++) {
1207 writel(0, base + WINDOW_BASE(i));
1208 writel(0, base + WINDOW_SIZE(i));
1209 if (i < 4)
1210 writel(0, base + WINDOW_REMAP_HIGH(i));
1211 }
1212 /*
1213 * For Armada3700 open default 4GB Mbus window. The dram
1214 * related configuration are done at AXIS level.
1215 */
1216 writel(0xffff0000, base + WINDOW_SIZE(0));
1217 win_enable |= 1;
1218 win_enable |= 3 << 16;
1219
1220 writel(win_enable, base + WINDOW_BAR_ENABLE(0));
1221 writel(win_enable, base + WINDOW_BAR_ENABLE(1));
1222 writel(0, base + WINDOW_OVERRIDE_CTRL(0));
1223 writel(0, base + WINDOW_OVERRIDE_CTRL(1));
1224}
1225
1226/*
1227 * Since this XOR driver is basically used only for RAID5, we don't
1228 * need to care about synchronizing ->suspend with DMA activity,
1229 * because the DMA engine will naturally be quiet due to the block
1230 * devices being suspended.
1231 */
1232static int mv_xor_suspend(struct platform_device *pdev, pm_message_t state)
1233{
1234 struct mv_xor_device *xordev = platform_get_drvdata(pdev);
1235 int i;
1236
1237 for (i = 0; i < MV_XOR_MAX_CHANNELS; i++) {
1238 struct mv_xor_chan *mv_chan = xordev->channels[i];
1239
1240 if (!mv_chan)
1241 continue;
1242
1243 mv_chan->saved_config_reg =
1244 readl_relaxed(XOR_CONFIG(mv_chan));
1245 mv_chan->saved_int_mask_reg =
1246 readl_relaxed(XOR_INTR_MASK(mv_chan));
1247 }
1248
1249 return 0;
1250}
1251
1252static int mv_xor_resume(struct platform_device *dev)
1253{
1254 struct mv_xor_device *xordev = platform_get_drvdata(dev);
1255 const struct mbus_dram_target_info *dram;
1256 int i;
1257
1258 for (i = 0; i < MV_XOR_MAX_CHANNELS; i++) {
1259 struct mv_xor_chan *mv_chan = xordev->channels[i];
1260
1261 if (!mv_chan)
1262 continue;
1263
1264 writel_relaxed(mv_chan->saved_config_reg,
1265 XOR_CONFIG(mv_chan));
1266 writel_relaxed(mv_chan->saved_int_mask_reg,
1267 XOR_INTR_MASK(mv_chan));
1268 }
1269
1270 if (xordev->xor_type == XOR_ARMADA_37XX) {
1271 mv_xor_conf_mbus_windows_a3700(xordev);
1272 return 0;
1273 }
1274
1275 dram = mv_mbus_dram_info();
1276 if (dram)
1277 mv_xor_conf_mbus_windows(xordev, dram);
1278
1279 return 0;
1280}
1281
1282static const struct of_device_id mv_xor_dt_ids[] = {
1283 { .compatible = "marvell,orion-xor", .data = (void *)XOR_ORION },
1284 { .compatible = "marvell,armada-380-xor", .data = (void *)XOR_ARMADA_38X },
1285 { .compatible = "marvell,armada-3700-xor", .data = (void *)XOR_ARMADA_37XX },
1286 {},
1287};
1288
1289static unsigned int mv_xor_engine_count;
1290
1291static int mv_xor_probe(struct platform_device *pdev)
1292{
1293 const struct mbus_dram_target_info *dram;
1294 struct mv_xor_device *xordev;
1295 struct mv_xor_platform_data *pdata = dev_get_platdata(&pdev->dev);
1296 struct resource *res;
1297 unsigned int max_engines, max_channels;
1298 int i, ret;
1299
1300 dev_notice(&pdev->dev, "Marvell shared XOR driver\n");
1301
1302 xordev = devm_kzalloc(&pdev->dev, sizeof(*xordev), GFP_KERNEL);
1303 if (!xordev)
1304 return -ENOMEM;
1305
1306 res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
1307 if (!res)
1308 return -ENODEV;
1309
1310 xordev->xor_base = devm_ioremap(&pdev->dev, res->start,
1311 resource_size(res));
1312 if (!xordev->xor_base)
1313 return -EBUSY;
1314
1315 res = platform_get_resource(pdev, IORESOURCE_MEM, 1);
1316 if (!res)
1317 return -ENODEV;
1318
1319 xordev->xor_high_base = devm_ioremap(&pdev->dev, res->start,
1320 resource_size(res));
1321 if (!xordev->xor_high_base)
1322 return -EBUSY;
1323
1324 platform_set_drvdata(pdev, xordev);
1325
1326
1327 /*
1328 * We need to know which type of XOR device we use before
1329 * setting up. In non-dt case it can only be the legacy one.
1330 */
1331 xordev->xor_type = XOR_ORION;
1332 if (pdev->dev.of_node) {
1333 const struct of_device_id *of_id =
1334 of_match_device(mv_xor_dt_ids,
1335 &pdev->dev);
1336
1337 xordev->xor_type = (uintptr_t)of_id->data;
1338 }
1339
1340 /*
1341 * (Re-)program MBUS remapping windows if we are asked to.
1342 */
1343 if (xordev->xor_type == XOR_ARMADA_37XX) {
1344 mv_xor_conf_mbus_windows_a3700(xordev);
1345 } else {
1346 dram = mv_mbus_dram_info();
1347 if (dram)
1348 mv_xor_conf_mbus_windows(xordev, dram);
1349 }
1350
1351 /* Not all platforms can gate the clock, so it is not
1352 * an error if the clock does not exists.
1353 */
1354 xordev->clk = clk_get(&pdev->dev, NULL);
1355 if (!IS_ERR(xordev->clk))
1356 clk_prepare_enable(xordev->clk);
1357
1358 /*
1359 * We don't want to have more than one channel per CPU in
1360 * order for async_tx to perform well. So we limit the number
1361 * of engines and channels so that we take into account this
1362 * constraint. Note that we also want to use channels from
1363 * separate engines when possible. For dual-CPU Armada 3700
1364 * SoC with single XOR engine allow using its both channels.
1365 */
1366 max_engines = num_present_cpus();
1367 if (xordev->xor_type == XOR_ARMADA_37XX)
1368 max_channels = num_present_cpus();
1369 else
1370 max_channels = min_t(unsigned int,
1371 MV_XOR_MAX_CHANNELS,
1372 DIV_ROUND_UP(num_present_cpus(), 2));
1373
1374 if (mv_xor_engine_count >= max_engines)
1375 return 0;
1376
1377 if (pdev->dev.of_node) {
1378 struct device_node *np;
1379 int i = 0;
1380
1381 for_each_child_of_node(pdev->dev.of_node, np) {
1382 struct mv_xor_chan *chan;
1383 dma_cap_mask_t cap_mask;
1384 int irq;
1385
1386 if (i >= max_channels)
1387 continue;
1388
1389 dma_cap_zero(cap_mask);
1390 dma_cap_set(DMA_MEMCPY, cap_mask);
1391 dma_cap_set(DMA_XOR, cap_mask);
1392 dma_cap_set(DMA_INTERRUPT, cap_mask);
1393
1394 irq = irq_of_parse_and_map(np, 0);
1395 if (!irq) {
1396 ret = -ENODEV;
1397 goto err_channel_add;
1398 }
1399
1400 chan = mv_xor_channel_add(xordev, pdev, i,
1401 cap_mask, irq);
1402 if (IS_ERR(chan)) {
1403 ret = PTR_ERR(chan);
1404 irq_dispose_mapping(irq);
1405 goto err_channel_add;
1406 }
1407
1408 xordev->channels[i] = chan;
1409 i++;
1410 }
1411 } else if (pdata && pdata->channels) {
1412 for (i = 0; i < max_channels; i++) {
1413 struct mv_xor_channel_data *cd;
1414 struct mv_xor_chan *chan;
1415 int irq;
1416
1417 cd = &pdata->channels[i];
1418 irq = platform_get_irq(pdev, i);
1419 if (irq < 0) {
1420 ret = irq;
1421 goto err_channel_add;
1422 }
1423
1424 chan = mv_xor_channel_add(xordev, pdev, i,
1425 cd->cap_mask, irq);
1426 if (IS_ERR(chan)) {
1427 ret = PTR_ERR(chan);
1428 goto err_channel_add;
1429 }
1430
1431 xordev->channels[i] = chan;
1432 }
1433 }
1434
1435 return 0;
1436
1437err_channel_add:
1438 for (i = 0; i < MV_XOR_MAX_CHANNELS; i++)
1439 if (xordev->channels[i]) {
1440 mv_xor_channel_remove(xordev->channels[i]);
1441 if (pdev->dev.of_node)
1442 irq_dispose_mapping(xordev->channels[i]->irq);
1443 }
1444
1445 if (!IS_ERR(xordev->clk)) {
1446 clk_disable_unprepare(xordev->clk);
1447 clk_put(xordev->clk);
1448 }
1449
1450 return ret;
1451}
1452
1453static struct platform_driver mv_xor_driver = {
1454 .probe = mv_xor_probe,
1455 .suspend = mv_xor_suspend,
1456 .resume = mv_xor_resume,
1457 .driver = {
1458 .name = MV_XOR_NAME,
1459 .of_match_table = of_match_ptr(mv_xor_dt_ids),
1460 },
1461};
1462
1463builtin_platform_driver(mv_xor_driver);
1464
1465/*
1466MODULE_AUTHOR("Saeed Bishara <saeed@marvell.com>");
1467MODULE_DESCRIPTION("DMA engine driver for Marvell's XOR engine");
1468MODULE_LICENSE("GPL");
1469*/
1/*
2 * offload engine driver for the Marvell XOR engine
3 * Copyright (C) 2007, 2008, Marvell International Ltd.
4 *
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms and conditions of the GNU General Public License,
7 * version 2, as published by the Free Software Foundation.
8 *
9 * This program is distributed in the hope 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
15#include <linux/init.h>
16#include <linux/slab.h>
17#include <linux/delay.h>
18#include <linux/dma-mapping.h>
19#include <linux/spinlock.h>
20#include <linux/interrupt.h>
21#include <linux/of_device.h>
22#include <linux/platform_device.h>
23#include <linux/memory.h>
24#include <linux/clk.h>
25#include <linux/of.h>
26#include <linux/of_irq.h>
27#include <linux/irqdomain.h>
28#include <linux/cpumask.h>
29#include <linux/platform_data/dma-mv_xor.h>
30
31#include "dmaengine.h"
32#include "mv_xor.h"
33
34enum mv_xor_type {
35 XOR_ORION,
36 XOR_ARMADA_38X,
37 XOR_ARMADA_37XX,
38};
39
40enum mv_xor_mode {
41 XOR_MODE_IN_REG,
42 XOR_MODE_IN_DESC,
43};
44
45static void mv_xor_issue_pending(struct dma_chan *chan);
46
47#define to_mv_xor_chan(chan) \
48 container_of(chan, struct mv_xor_chan, dmachan)
49
50#define to_mv_xor_slot(tx) \
51 container_of(tx, struct mv_xor_desc_slot, async_tx)
52
53#define mv_chan_to_devp(chan) \
54 ((chan)->dmadev.dev)
55
56static void mv_desc_init(struct mv_xor_desc_slot *desc,
57 dma_addr_t addr, u32 byte_count,
58 enum dma_ctrl_flags flags)
59{
60 struct mv_xor_desc *hw_desc = desc->hw_desc;
61
62 hw_desc->status = XOR_DESC_DMA_OWNED;
63 hw_desc->phy_next_desc = 0;
64 /* Enable end-of-descriptor interrupts only for DMA_PREP_INTERRUPT */
65 hw_desc->desc_command = (flags & DMA_PREP_INTERRUPT) ?
66 XOR_DESC_EOD_INT_EN : 0;
67 hw_desc->phy_dest_addr = addr;
68 hw_desc->byte_count = byte_count;
69}
70
71static void mv_desc_set_mode(struct mv_xor_desc_slot *desc)
72{
73 struct mv_xor_desc *hw_desc = desc->hw_desc;
74
75 switch (desc->type) {
76 case DMA_XOR:
77 case DMA_INTERRUPT:
78 hw_desc->desc_command |= XOR_DESC_OPERATION_XOR;
79 break;
80 case DMA_MEMCPY:
81 hw_desc->desc_command |= XOR_DESC_OPERATION_MEMCPY;
82 break;
83 default:
84 BUG();
85 return;
86 }
87}
88
89static void mv_desc_set_next_desc(struct mv_xor_desc_slot *desc,
90 u32 next_desc_addr)
91{
92 struct mv_xor_desc *hw_desc = desc->hw_desc;
93 BUG_ON(hw_desc->phy_next_desc);
94 hw_desc->phy_next_desc = next_desc_addr;
95}
96
97static void mv_desc_set_src_addr(struct mv_xor_desc_slot *desc,
98 int index, dma_addr_t addr)
99{
100 struct mv_xor_desc *hw_desc = desc->hw_desc;
101 hw_desc->phy_src_addr[mv_phy_src_idx(index)] = addr;
102 if (desc->type == DMA_XOR)
103 hw_desc->desc_command |= (1 << index);
104}
105
106static u32 mv_chan_get_current_desc(struct mv_xor_chan *chan)
107{
108 return readl_relaxed(XOR_CURR_DESC(chan));
109}
110
111static void mv_chan_set_next_descriptor(struct mv_xor_chan *chan,
112 u32 next_desc_addr)
113{
114 writel_relaxed(next_desc_addr, XOR_NEXT_DESC(chan));
115}
116
117static void mv_chan_unmask_interrupts(struct mv_xor_chan *chan)
118{
119 u32 val = readl_relaxed(XOR_INTR_MASK(chan));
120 val |= XOR_INTR_MASK_VALUE << (chan->idx * 16);
121 writel_relaxed(val, XOR_INTR_MASK(chan));
122}
123
124static u32 mv_chan_get_intr_cause(struct mv_xor_chan *chan)
125{
126 u32 intr_cause = readl_relaxed(XOR_INTR_CAUSE(chan));
127 intr_cause = (intr_cause >> (chan->idx * 16)) & 0xFFFF;
128 return intr_cause;
129}
130
131static void mv_chan_clear_eoc_cause(struct mv_xor_chan *chan)
132{
133 u32 val;
134
135 val = XOR_INT_END_OF_DESC | XOR_INT_END_OF_CHAIN | XOR_INT_STOPPED;
136 val = ~(val << (chan->idx * 16));
137 dev_dbg(mv_chan_to_devp(chan), "%s, val 0x%08x\n", __func__, val);
138 writel_relaxed(val, XOR_INTR_CAUSE(chan));
139}
140
141static void mv_chan_clear_err_status(struct mv_xor_chan *chan)
142{
143 u32 val = 0xFFFF0000 >> (chan->idx * 16);
144 writel_relaxed(val, XOR_INTR_CAUSE(chan));
145}
146
147static void mv_chan_set_mode(struct mv_xor_chan *chan,
148 u32 op_mode)
149{
150 u32 config = readl_relaxed(XOR_CONFIG(chan));
151
152 config &= ~0x7;
153 config |= op_mode;
154
155#if defined(__BIG_ENDIAN)
156 config |= XOR_DESCRIPTOR_SWAP;
157#else
158 config &= ~XOR_DESCRIPTOR_SWAP;
159#endif
160
161 writel_relaxed(config, XOR_CONFIG(chan));
162}
163
164static void mv_chan_activate(struct mv_xor_chan *chan)
165{
166 dev_dbg(mv_chan_to_devp(chan), " activate chan.\n");
167
168 /* writel ensures all descriptors are flushed before activation */
169 writel(BIT(0), XOR_ACTIVATION(chan));
170}
171
172static char mv_chan_is_busy(struct mv_xor_chan *chan)
173{
174 u32 state = readl_relaxed(XOR_ACTIVATION(chan));
175
176 state = (state >> 4) & 0x3;
177
178 return (state == 1) ? 1 : 0;
179}
180
181/*
182 * mv_chan_start_new_chain - program the engine to operate on new
183 * chain headed by sw_desc
184 * Caller must hold &mv_chan->lock while calling this function
185 */
186static void mv_chan_start_new_chain(struct mv_xor_chan *mv_chan,
187 struct mv_xor_desc_slot *sw_desc)
188{
189 dev_dbg(mv_chan_to_devp(mv_chan), "%s %d: sw_desc %p\n",
190 __func__, __LINE__, sw_desc);
191
192 /* set the hardware chain */
193 mv_chan_set_next_descriptor(mv_chan, sw_desc->async_tx.phys);
194
195 mv_chan->pending++;
196 mv_xor_issue_pending(&mv_chan->dmachan);
197}
198
199static dma_cookie_t
200mv_desc_run_tx_complete_actions(struct mv_xor_desc_slot *desc,
201 struct mv_xor_chan *mv_chan,
202 dma_cookie_t cookie)
203{
204 BUG_ON(desc->async_tx.cookie < 0);
205
206 if (desc->async_tx.cookie > 0) {
207 cookie = desc->async_tx.cookie;
208
209 dma_descriptor_unmap(&desc->async_tx);
210 /* call the callback (must not sleep or submit new
211 * operations to this channel)
212 */
213 dmaengine_desc_get_callback_invoke(&desc->async_tx, NULL);
214 }
215
216 /* run dependent operations */
217 dma_run_dependencies(&desc->async_tx);
218
219 return cookie;
220}
221
222static int
223mv_chan_clean_completed_slots(struct mv_xor_chan *mv_chan)
224{
225 struct mv_xor_desc_slot *iter, *_iter;
226
227 dev_dbg(mv_chan_to_devp(mv_chan), "%s %d\n", __func__, __LINE__);
228 list_for_each_entry_safe(iter, _iter, &mv_chan->completed_slots,
229 node) {
230
231 if (async_tx_test_ack(&iter->async_tx)) {
232 list_move_tail(&iter->node, &mv_chan->free_slots);
233 if (!list_empty(&iter->sg_tx_list)) {
234 list_splice_tail_init(&iter->sg_tx_list,
235 &mv_chan->free_slots);
236 }
237 }
238 }
239 return 0;
240}
241
242static int
243mv_desc_clean_slot(struct mv_xor_desc_slot *desc,
244 struct mv_xor_chan *mv_chan)
245{
246 dev_dbg(mv_chan_to_devp(mv_chan), "%s %d: desc %p flags %d\n",
247 __func__, __LINE__, desc, desc->async_tx.flags);
248
249 /* the client is allowed to attach dependent operations
250 * until 'ack' is set
251 */
252 if (!async_tx_test_ack(&desc->async_tx)) {
253 /* move this slot to the completed_slots */
254 list_move_tail(&desc->node, &mv_chan->completed_slots);
255 if (!list_empty(&desc->sg_tx_list)) {
256 list_splice_tail_init(&desc->sg_tx_list,
257 &mv_chan->completed_slots);
258 }
259 } else {
260 list_move_tail(&desc->node, &mv_chan->free_slots);
261 if (!list_empty(&desc->sg_tx_list)) {
262 list_splice_tail_init(&desc->sg_tx_list,
263 &mv_chan->free_slots);
264 }
265 }
266
267 return 0;
268}
269
270/* This function must be called with the mv_xor_chan spinlock held */
271static void mv_chan_slot_cleanup(struct mv_xor_chan *mv_chan)
272{
273 struct mv_xor_desc_slot *iter, *_iter;
274 dma_cookie_t cookie = 0;
275 int busy = mv_chan_is_busy(mv_chan);
276 u32 current_desc = mv_chan_get_current_desc(mv_chan);
277 int current_cleaned = 0;
278 struct mv_xor_desc *hw_desc;
279
280 dev_dbg(mv_chan_to_devp(mv_chan), "%s %d\n", __func__, __LINE__);
281 dev_dbg(mv_chan_to_devp(mv_chan), "current_desc %x\n", current_desc);
282 mv_chan_clean_completed_slots(mv_chan);
283
284 /* free completed slots from the chain starting with
285 * the oldest descriptor
286 */
287
288 list_for_each_entry_safe(iter, _iter, &mv_chan->chain,
289 node) {
290
291 /* clean finished descriptors */
292 hw_desc = iter->hw_desc;
293 if (hw_desc->status & XOR_DESC_SUCCESS) {
294 cookie = mv_desc_run_tx_complete_actions(iter, mv_chan,
295 cookie);
296
297 /* done processing desc, clean slot */
298 mv_desc_clean_slot(iter, mv_chan);
299
300 /* break if we did cleaned the current */
301 if (iter->async_tx.phys == current_desc) {
302 current_cleaned = 1;
303 break;
304 }
305 } else {
306 if (iter->async_tx.phys == current_desc) {
307 current_cleaned = 0;
308 break;
309 }
310 }
311 }
312
313 if ((busy == 0) && !list_empty(&mv_chan->chain)) {
314 if (current_cleaned) {
315 /*
316 * current descriptor cleaned and removed, run
317 * from list head
318 */
319 iter = list_entry(mv_chan->chain.next,
320 struct mv_xor_desc_slot,
321 node);
322 mv_chan_start_new_chain(mv_chan, iter);
323 } else {
324 if (!list_is_last(&iter->node, &mv_chan->chain)) {
325 /*
326 * descriptors are still waiting after
327 * current, trigger them
328 */
329 iter = list_entry(iter->node.next,
330 struct mv_xor_desc_slot,
331 node);
332 mv_chan_start_new_chain(mv_chan, iter);
333 } else {
334 /*
335 * some descriptors are still waiting
336 * to be cleaned
337 */
338 tasklet_schedule(&mv_chan->irq_tasklet);
339 }
340 }
341 }
342
343 if (cookie > 0)
344 mv_chan->dmachan.completed_cookie = cookie;
345}
346
347static void mv_xor_tasklet(unsigned long data)
348{
349 struct mv_xor_chan *chan = (struct mv_xor_chan *) data;
350
351 spin_lock_bh(&chan->lock);
352 mv_chan_slot_cleanup(chan);
353 spin_unlock_bh(&chan->lock);
354}
355
356static struct mv_xor_desc_slot *
357mv_chan_alloc_slot(struct mv_xor_chan *mv_chan)
358{
359 struct mv_xor_desc_slot *iter;
360
361 spin_lock_bh(&mv_chan->lock);
362
363 if (!list_empty(&mv_chan->free_slots)) {
364 iter = list_first_entry(&mv_chan->free_slots,
365 struct mv_xor_desc_slot,
366 node);
367
368 list_move_tail(&iter->node, &mv_chan->allocated_slots);
369
370 spin_unlock_bh(&mv_chan->lock);
371
372 /* pre-ack descriptor */
373 async_tx_ack(&iter->async_tx);
374 iter->async_tx.cookie = -EBUSY;
375
376 return iter;
377
378 }
379
380 spin_unlock_bh(&mv_chan->lock);
381
382 /* try to free some slots if the allocation fails */
383 tasklet_schedule(&mv_chan->irq_tasklet);
384
385 return NULL;
386}
387
388/************************ DMA engine API functions ****************************/
389static dma_cookie_t
390mv_xor_tx_submit(struct dma_async_tx_descriptor *tx)
391{
392 struct mv_xor_desc_slot *sw_desc = to_mv_xor_slot(tx);
393 struct mv_xor_chan *mv_chan = to_mv_xor_chan(tx->chan);
394 struct mv_xor_desc_slot *old_chain_tail;
395 dma_cookie_t cookie;
396 int new_hw_chain = 1;
397
398 dev_dbg(mv_chan_to_devp(mv_chan),
399 "%s sw_desc %p: async_tx %p\n",
400 __func__, sw_desc, &sw_desc->async_tx);
401
402 spin_lock_bh(&mv_chan->lock);
403 cookie = dma_cookie_assign(tx);
404
405 if (list_empty(&mv_chan->chain))
406 list_move_tail(&sw_desc->node, &mv_chan->chain);
407 else {
408 new_hw_chain = 0;
409
410 old_chain_tail = list_entry(mv_chan->chain.prev,
411 struct mv_xor_desc_slot,
412 node);
413 list_move_tail(&sw_desc->node, &mv_chan->chain);
414
415 dev_dbg(mv_chan_to_devp(mv_chan), "Append to last desc %pa\n",
416 &old_chain_tail->async_tx.phys);
417
418 /* fix up the hardware chain */
419 mv_desc_set_next_desc(old_chain_tail, sw_desc->async_tx.phys);
420
421 /* if the channel is not busy */
422 if (!mv_chan_is_busy(mv_chan)) {
423 u32 current_desc = mv_chan_get_current_desc(mv_chan);
424 /*
425 * and the curren desc is the end of the chain before
426 * the append, then we need to start the channel
427 */
428 if (current_desc == old_chain_tail->async_tx.phys)
429 new_hw_chain = 1;
430 }
431 }
432
433 if (new_hw_chain)
434 mv_chan_start_new_chain(mv_chan, sw_desc);
435
436 spin_unlock_bh(&mv_chan->lock);
437
438 return cookie;
439}
440
441/* returns the number of allocated descriptors */
442static int mv_xor_alloc_chan_resources(struct dma_chan *chan)
443{
444 void *virt_desc;
445 dma_addr_t dma_desc;
446 int idx;
447 struct mv_xor_chan *mv_chan = to_mv_xor_chan(chan);
448 struct mv_xor_desc_slot *slot = NULL;
449 int num_descs_in_pool = MV_XOR_POOL_SIZE/MV_XOR_SLOT_SIZE;
450
451 /* Allocate descriptor slots */
452 idx = mv_chan->slots_allocated;
453 while (idx < num_descs_in_pool) {
454 slot = kzalloc(sizeof(*slot), GFP_KERNEL);
455 if (!slot) {
456 dev_info(mv_chan_to_devp(mv_chan),
457 "channel only initialized %d descriptor slots",
458 idx);
459 break;
460 }
461 virt_desc = mv_chan->dma_desc_pool_virt;
462 slot->hw_desc = virt_desc + idx * MV_XOR_SLOT_SIZE;
463
464 dma_async_tx_descriptor_init(&slot->async_tx, chan);
465 slot->async_tx.tx_submit = mv_xor_tx_submit;
466 INIT_LIST_HEAD(&slot->node);
467 INIT_LIST_HEAD(&slot->sg_tx_list);
468 dma_desc = mv_chan->dma_desc_pool;
469 slot->async_tx.phys = dma_desc + idx * MV_XOR_SLOT_SIZE;
470 slot->idx = idx++;
471
472 spin_lock_bh(&mv_chan->lock);
473 mv_chan->slots_allocated = idx;
474 list_add_tail(&slot->node, &mv_chan->free_slots);
475 spin_unlock_bh(&mv_chan->lock);
476 }
477
478 dev_dbg(mv_chan_to_devp(mv_chan),
479 "allocated %d descriptor slots\n",
480 mv_chan->slots_allocated);
481
482 return mv_chan->slots_allocated ? : -ENOMEM;
483}
484
485/*
486 * Check if source or destination is an PCIe/IO address (non-SDRAM) and add
487 * a new MBus window if necessary. Use a cache for these check so that
488 * the MMIO mapped registers don't have to be accessed for this check
489 * to speed up this process.
490 */
491static int mv_xor_add_io_win(struct mv_xor_chan *mv_chan, u32 addr)
492{
493 struct mv_xor_device *xordev = mv_chan->xordev;
494 void __iomem *base = mv_chan->mmr_high_base;
495 u32 win_enable;
496 u32 size;
497 u8 target, attr;
498 int ret;
499 int i;
500
501 /* Nothing needs to get done for the Armada 3700 */
502 if (xordev->xor_type == XOR_ARMADA_37XX)
503 return 0;
504
505 /*
506 * Loop over the cached windows to check, if the requested area
507 * is already mapped. If this the case, nothing needs to be done
508 * and we can return.
509 */
510 for (i = 0; i < WINDOW_COUNT; i++) {
511 if (addr >= xordev->win_start[i] &&
512 addr <= xordev->win_end[i]) {
513 /* Window is already mapped */
514 return 0;
515 }
516 }
517
518 /*
519 * The window is not mapped, so we need to create the new mapping
520 */
521
522 /* If no IO window is found that addr has to be located in SDRAM */
523 ret = mvebu_mbus_get_io_win_info(addr, &size, &target, &attr);
524 if (ret < 0)
525 return 0;
526
527 /*
528 * Mask the base addr 'addr' according to 'size' read back from the
529 * MBus window. Otherwise we might end up with an address located
530 * somewhere in the middle of this area here.
531 */
532 size -= 1;
533 addr &= ~size;
534
535 /*
536 * Reading one of both enabled register is enough, as they are always
537 * programmed to the identical values
538 */
539 win_enable = readl(base + WINDOW_BAR_ENABLE(0));
540
541 /* Set 'i' to the first free window to write the new values to */
542 i = ffs(~win_enable) - 1;
543 if (i >= WINDOW_COUNT)
544 return -ENOMEM;
545
546 writel((addr & 0xffff0000) | (attr << 8) | target,
547 base + WINDOW_BASE(i));
548 writel(size & 0xffff0000, base + WINDOW_SIZE(i));
549
550 /* Fill the caching variables for later use */
551 xordev->win_start[i] = addr;
552 xordev->win_end[i] = addr + size;
553
554 win_enable |= (1 << i);
555 win_enable |= 3 << (16 + (2 * i));
556 writel(win_enable, base + WINDOW_BAR_ENABLE(0));
557 writel(win_enable, base + WINDOW_BAR_ENABLE(1));
558
559 return 0;
560}
561
562static struct dma_async_tx_descriptor *
563mv_xor_prep_dma_xor(struct dma_chan *chan, dma_addr_t dest, dma_addr_t *src,
564 unsigned int src_cnt, size_t len, unsigned long flags)
565{
566 struct mv_xor_chan *mv_chan = to_mv_xor_chan(chan);
567 struct mv_xor_desc_slot *sw_desc;
568 int ret;
569
570 if (unlikely(len < MV_XOR_MIN_BYTE_COUNT))
571 return NULL;
572
573 BUG_ON(len > MV_XOR_MAX_BYTE_COUNT);
574
575 dev_dbg(mv_chan_to_devp(mv_chan),
576 "%s src_cnt: %d len: %zu dest %pad flags: %ld\n",
577 __func__, src_cnt, len, &dest, flags);
578
579 /* Check if a new window needs to get added for 'dest' */
580 ret = mv_xor_add_io_win(mv_chan, dest);
581 if (ret)
582 return NULL;
583
584 sw_desc = mv_chan_alloc_slot(mv_chan);
585 if (sw_desc) {
586 sw_desc->type = DMA_XOR;
587 sw_desc->async_tx.flags = flags;
588 mv_desc_init(sw_desc, dest, len, flags);
589 if (mv_chan->op_in_desc == XOR_MODE_IN_DESC)
590 mv_desc_set_mode(sw_desc);
591 while (src_cnt--) {
592 /* Check if a new window needs to get added for 'src' */
593 ret = mv_xor_add_io_win(mv_chan, src[src_cnt]);
594 if (ret)
595 return NULL;
596 mv_desc_set_src_addr(sw_desc, src_cnt, src[src_cnt]);
597 }
598 }
599
600 dev_dbg(mv_chan_to_devp(mv_chan),
601 "%s sw_desc %p async_tx %p \n",
602 __func__, sw_desc, &sw_desc->async_tx);
603 return sw_desc ? &sw_desc->async_tx : NULL;
604}
605
606static struct dma_async_tx_descriptor *
607mv_xor_prep_dma_memcpy(struct dma_chan *chan, dma_addr_t dest, dma_addr_t src,
608 size_t len, unsigned long flags)
609{
610 /*
611 * A MEMCPY operation is identical to an XOR operation with only
612 * a single source address.
613 */
614 return mv_xor_prep_dma_xor(chan, dest, &src, 1, len, flags);
615}
616
617static struct dma_async_tx_descriptor *
618mv_xor_prep_dma_interrupt(struct dma_chan *chan, unsigned long flags)
619{
620 struct mv_xor_chan *mv_chan = to_mv_xor_chan(chan);
621 dma_addr_t src, dest;
622 size_t len;
623
624 src = mv_chan->dummy_src_addr;
625 dest = mv_chan->dummy_dst_addr;
626 len = MV_XOR_MIN_BYTE_COUNT;
627
628 /*
629 * We implement the DMA_INTERRUPT operation as a minimum sized
630 * XOR operation with a single dummy source address.
631 */
632 return mv_xor_prep_dma_xor(chan, dest, &src, 1, len, flags);
633}
634
635static void mv_xor_free_chan_resources(struct dma_chan *chan)
636{
637 struct mv_xor_chan *mv_chan = to_mv_xor_chan(chan);
638 struct mv_xor_desc_slot *iter, *_iter;
639 int in_use_descs = 0;
640
641 spin_lock_bh(&mv_chan->lock);
642
643 mv_chan_slot_cleanup(mv_chan);
644
645 list_for_each_entry_safe(iter, _iter, &mv_chan->chain,
646 node) {
647 in_use_descs++;
648 list_move_tail(&iter->node, &mv_chan->free_slots);
649 }
650 list_for_each_entry_safe(iter, _iter, &mv_chan->completed_slots,
651 node) {
652 in_use_descs++;
653 list_move_tail(&iter->node, &mv_chan->free_slots);
654 }
655 list_for_each_entry_safe(iter, _iter, &mv_chan->allocated_slots,
656 node) {
657 in_use_descs++;
658 list_move_tail(&iter->node, &mv_chan->free_slots);
659 }
660 list_for_each_entry_safe_reverse(
661 iter, _iter, &mv_chan->free_slots, node) {
662 list_del(&iter->node);
663 kfree(iter);
664 mv_chan->slots_allocated--;
665 }
666
667 dev_dbg(mv_chan_to_devp(mv_chan), "%s slots_allocated %d\n",
668 __func__, mv_chan->slots_allocated);
669 spin_unlock_bh(&mv_chan->lock);
670
671 if (in_use_descs)
672 dev_err(mv_chan_to_devp(mv_chan),
673 "freeing %d in use descriptors!\n", in_use_descs);
674}
675
676/**
677 * mv_xor_status - poll the status of an XOR transaction
678 * @chan: XOR channel handle
679 * @cookie: XOR transaction identifier
680 * @txstate: XOR transactions state holder (or NULL)
681 */
682static enum dma_status mv_xor_status(struct dma_chan *chan,
683 dma_cookie_t cookie,
684 struct dma_tx_state *txstate)
685{
686 struct mv_xor_chan *mv_chan = to_mv_xor_chan(chan);
687 enum dma_status ret;
688
689 ret = dma_cookie_status(chan, cookie, txstate);
690 if (ret == DMA_COMPLETE)
691 return ret;
692
693 spin_lock_bh(&mv_chan->lock);
694 mv_chan_slot_cleanup(mv_chan);
695 spin_unlock_bh(&mv_chan->lock);
696
697 return dma_cookie_status(chan, cookie, txstate);
698}
699
700static void mv_chan_dump_regs(struct mv_xor_chan *chan)
701{
702 u32 val;
703
704 val = readl_relaxed(XOR_CONFIG(chan));
705 dev_err(mv_chan_to_devp(chan), "config 0x%08x\n", val);
706
707 val = readl_relaxed(XOR_ACTIVATION(chan));
708 dev_err(mv_chan_to_devp(chan), "activation 0x%08x\n", val);
709
710 val = readl_relaxed(XOR_INTR_CAUSE(chan));
711 dev_err(mv_chan_to_devp(chan), "intr cause 0x%08x\n", val);
712
713 val = readl_relaxed(XOR_INTR_MASK(chan));
714 dev_err(mv_chan_to_devp(chan), "intr mask 0x%08x\n", val);
715
716 val = readl_relaxed(XOR_ERROR_CAUSE(chan));
717 dev_err(mv_chan_to_devp(chan), "error cause 0x%08x\n", val);
718
719 val = readl_relaxed(XOR_ERROR_ADDR(chan));
720 dev_err(mv_chan_to_devp(chan), "error addr 0x%08x\n", val);
721}
722
723static void mv_chan_err_interrupt_handler(struct mv_xor_chan *chan,
724 u32 intr_cause)
725{
726 if (intr_cause & XOR_INT_ERR_DECODE) {
727 dev_dbg(mv_chan_to_devp(chan), "ignoring address decode error\n");
728 return;
729 }
730
731 dev_err(mv_chan_to_devp(chan), "error on chan %d. intr cause 0x%08x\n",
732 chan->idx, intr_cause);
733
734 mv_chan_dump_regs(chan);
735 WARN_ON(1);
736}
737
738static irqreturn_t mv_xor_interrupt_handler(int irq, void *data)
739{
740 struct mv_xor_chan *chan = data;
741 u32 intr_cause = mv_chan_get_intr_cause(chan);
742
743 dev_dbg(mv_chan_to_devp(chan), "intr cause %x\n", intr_cause);
744
745 if (intr_cause & XOR_INTR_ERRORS)
746 mv_chan_err_interrupt_handler(chan, intr_cause);
747
748 tasklet_schedule(&chan->irq_tasklet);
749
750 mv_chan_clear_eoc_cause(chan);
751
752 return IRQ_HANDLED;
753}
754
755static void mv_xor_issue_pending(struct dma_chan *chan)
756{
757 struct mv_xor_chan *mv_chan = to_mv_xor_chan(chan);
758
759 if (mv_chan->pending >= MV_XOR_THRESHOLD) {
760 mv_chan->pending = 0;
761 mv_chan_activate(mv_chan);
762 }
763}
764
765/*
766 * Perform a transaction to verify the HW works.
767 */
768
769static int mv_chan_memcpy_self_test(struct mv_xor_chan *mv_chan)
770{
771 int i, ret;
772 void *src, *dest;
773 dma_addr_t src_dma, dest_dma;
774 struct dma_chan *dma_chan;
775 dma_cookie_t cookie;
776 struct dma_async_tx_descriptor *tx;
777 struct dmaengine_unmap_data *unmap;
778 int err = 0;
779
780 src = kmalloc(sizeof(u8) * PAGE_SIZE, GFP_KERNEL);
781 if (!src)
782 return -ENOMEM;
783
784 dest = kzalloc(sizeof(u8) * PAGE_SIZE, GFP_KERNEL);
785 if (!dest) {
786 kfree(src);
787 return -ENOMEM;
788 }
789
790 /* Fill in src buffer */
791 for (i = 0; i < PAGE_SIZE; i++)
792 ((u8 *) src)[i] = (u8)i;
793
794 dma_chan = &mv_chan->dmachan;
795 if (mv_xor_alloc_chan_resources(dma_chan) < 1) {
796 err = -ENODEV;
797 goto out;
798 }
799
800 unmap = dmaengine_get_unmap_data(dma_chan->device->dev, 2, GFP_KERNEL);
801 if (!unmap) {
802 err = -ENOMEM;
803 goto free_resources;
804 }
805
806 src_dma = dma_map_page(dma_chan->device->dev, virt_to_page(src),
807 offset_in_page(src), PAGE_SIZE,
808 DMA_TO_DEVICE);
809 unmap->addr[0] = src_dma;
810
811 ret = dma_mapping_error(dma_chan->device->dev, src_dma);
812 if (ret) {
813 err = -ENOMEM;
814 goto free_resources;
815 }
816 unmap->to_cnt = 1;
817
818 dest_dma = dma_map_page(dma_chan->device->dev, virt_to_page(dest),
819 offset_in_page(dest), PAGE_SIZE,
820 DMA_FROM_DEVICE);
821 unmap->addr[1] = dest_dma;
822
823 ret = dma_mapping_error(dma_chan->device->dev, dest_dma);
824 if (ret) {
825 err = -ENOMEM;
826 goto free_resources;
827 }
828 unmap->from_cnt = 1;
829 unmap->len = PAGE_SIZE;
830
831 tx = mv_xor_prep_dma_memcpy(dma_chan, dest_dma, src_dma,
832 PAGE_SIZE, 0);
833 if (!tx) {
834 dev_err(dma_chan->device->dev,
835 "Self-test cannot prepare operation, disabling\n");
836 err = -ENODEV;
837 goto free_resources;
838 }
839
840 cookie = mv_xor_tx_submit(tx);
841 if (dma_submit_error(cookie)) {
842 dev_err(dma_chan->device->dev,
843 "Self-test submit error, disabling\n");
844 err = -ENODEV;
845 goto free_resources;
846 }
847
848 mv_xor_issue_pending(dma_chan);
849 async_tx_ack(tx);
850 msleep(1);
851
852 if (mv_xor_status(dma_chan, cookie, NULL) !=
853 DMA_COMPLETE) {
854 dev_err(dma_chan->device->dev,
855 "Self-test copy timed out, disabling\n");
856 err = -ENODEV;
857 goto free_resources;
858 }
859
860 dma_sync_single_for_cpu(dma_chan->device->dev, dest_dma,
861 PAGE_SIZE, DMA_FROM_DEVICE);
862 if (memcmp(src, dest, PAGE_SIZE)) {
863 dev_err(dma_chan->device->dev,
864 "Self-test copy failed compare, disabling\n");
865 err = -ENODEV;
866 goto free_resources;
867 }
868
869free_resources:
870 dmaengine_unmap_put(unmap);
871 mv_xor_free_chan_resources(dma_chan);
872out:
873 kfree(src);
874 kfree(dest);
875 return err;
876}
877
878#define MV_XOR_NUM_SRC_TEST 4 /* must be <= 15 */
879static int
880mv_chan_xor_self_test(struct mv_xor_chan *mv_chan)
881{
882 int i, src_idx, ret;
883 struct page *dest;
884 struct page *xor_srcs[MV_XOR_NUM_SRC_TEST];
885 dma_addr_t dma_srcs[MV_XOR_NUM_SRC_TEST];
886 dma_addr_t dest_dma;
887 struct dma_async_tx_descriptor *tx;
888 struct dmaengine_unmap_data *unmap;
889 struct dma_chan *dma_chan;
890 dma_cookie_t cookie;
891 u8 cmp_byte = 0;
892 u32 cmp_word;
893 int err = 0;
894 int src_count = MV_XOR_NUM_SRC_TEST;
895
896 for (src_idx = 0; src_idx < src_count; src_idx++) {
897 xor_srcs[src_idx] = alloc_page(GFP_KERNEL);
898 if (!xor_srcs[src_idx]) {
899 while (src_idx--)
900 __free_page(xor_srcs[src_idx]);
901 return -ENOMEM;
902 }
903 }
904
905 dest = alloc_page(GFP_KERNEL);
906 if (!dest) {
907 while (src_idx--)
908 __free_page(xor_srcs[src_idx]);
909 return -ENOMEM;
910 }
911
912 /* Fill in src buffers */
913 for (src_idx = 0; src_idx < src_count; src_idx++) {
914 u8 *ptr = page_address(xor_srcs[src_idx]);
915 for (i = 0; i < PAGE_SIZE; i++)
916 ptr[i] = (1 << src_idx);
917 }
918
919 for (src_idx = 0; src_idx < src_count; src_idx++)
920 cmp_byte ^= (u8) (1 << src_idx);
921
922 cmp_word = (cmp_byte << 24) | (cmp_byte << 16) |
923 (cmp_byte << 8) | cmp_byte;
924
925 memset(page_address(dest), 0, PAGE_SIZE);
926
927 dma_chan = &mv_chan->dmachan;
928 if (mv_xor_alloc_chan_resources(dma_chan) < 1) {
929 err = -ENODEV;
930 goto out;
931 }
932
933 unmap = dmaengine_get_unmap_data(dma_chan->device->dev, src_count + 1,
934 GFP_KERNEL);
935 if (!unmap) {
936 err = -ENOMEM;
937 goto free_resources;
938 }
939
940 /* test xor */
941 for (i = 0; i < src_count; i++) {
942 unmap->addr[i] = dma_map_page(dma_chan->device->dev, xor_srcs[i],
943 0, PAGE_SIZE, DMA_TO_DEVICE);
944 dma_srcs[i] = unmap->addr[i];
945 ret = dma_mapping_error(dma_chan->device->dev, unmap->addr[i]);
946 if (ret) {
947 err = -ENOMEM;
948 goto free_resources;
949 }
950 unmap->to_cnt++;
951 }
952
953 unmap->addr[src_count] = dma_map_page(dma_chan->device->dev, dest, 0, PAGE_SIZE,
954 DMA_FROM_DEVICE);
955 dest_dma = unmap->addr[src_count];
956 ret = dma_mapping_error(dma_chan->device->dev, unmap->addr[src_count]);
957 if (ret) {
958 err = -ENOMEM;
959 goto free_resources;
960 }
961 unmap->from_cnt = 1;
962 unmap->len = PAGE_SIZE;
963
964 tx = mv_xor_prep_dma_xor(dma_chan, dest_dma, dma_srcs,
965 src_count, PAGE_SIZE, 0);
966 if (!tx) {
967 dev_err(dma_chan->device->dev,
968 "Self-test cannot prepare operation, disabling\n");
969 err = -ENODEV;
970 goto free_resources;
971 }
972
973 cookie = mv_xor_tx_submit(tx);
974 if (dma_submit_error(cookie)) {
975 dev_err(dma_chan->device->dev,
976 "Self-test submit error, disabling\n");
977 err = -ENODEV;
978 goto free_resources;
979 }
980
981 mv_xor_issue_pending(dma_chan);
982 async_tx_ack(tx);
983 msleep(8);
984
985 if (mv_xor_status(dma_chan, cookie, NULL) !=
986 DMA_COMPLETE) {
987 dev_err(dma_chan->device->dev,
988 "Self-test xor timed out, disabling\n");
989 err = -ENODEV;
990 goto free_resources;
991 }
992
993 dma_sync_single_for_cpu(dma_chan->device->dev, dest_dma,
994 PAGE_SIZE, DMA_FROM_DEVICE);
995 for (i = 0; i < (PAGE_SIZE / sizeof(u32)); i++) {
996 u32 *ptr = page_address(dest);
997 if (ptr[i] != cmp_word) {
998 dev_err(dma_chan->device->dev,
999 "Self-test xor failed compare, disabling. index %d, data %x, expected %x\n",
1000 i, ptr[i], cmp_word);
1001 err = -ENODEV;
1002 goto free_resources;
1003 }
1004 }
1005
1006free_resources:
1007 dmaengine_unmap_put(unmap);
1008 mv_xor_free_chan_resources(dma_chan);
1009out:
1010 src_idx = src_count;
1011 while (src_idx--)
1012 __free_page(xor_srcs[src_idx]);
1013 __free_page(dest);
1014 return err;
1015}
1016
1017static int mv_xor_channel_remove(struct mv_xor_chan *mv_chan)
1018{
1019 struct dma_chan *chan, *_chan;
1020 struct device *dev = mv_chan->dmadev.dev;
1021
1022 dma_async_device_unregister(&mv_chan->dmadev);
1023
1024 dma_free_coherent(dev, MV_XOR_POOL_SIZE,
1025 mv_chan->dma_desc_pool_virt, mv_chan->dma_desc_pool);
1026 dma_unmap_single(dev, mv_chan->dummy_src_addr,
1027 MV_XOR_MIN_BYTE_COUNT, DMA_FROM_DEVICE);
1028 dma_unmap_single(dev, mv_chan->dummy_dst_addr,
1029 MV_XOR_MIN_BYTE_COUNT, DMA_TO_DEVICE);
1030
1031 list_for_each_entry_safe(chan, _chan, &mv_chan->dmadev.channels,
1032 device_node) {
1033 list_del(&chan->device_node);
1034 }
1035
1036 free_irq(mv_chan->irq, mv_chan);
1037
1038 return 0;
1039}
1040
1041static struct mv_xor_chan *
1042mv_xor_channel_add(struct mv_xor_device *xordev,
1043 struct platform_device *pdev,
1044 int idx, dma_cap_mask_t cap_mask, int irq)
1045{
1046 int ret = 0;
1047 struct mv_xor_chan *mv_chan;
1048 struct dma_device *dma_dev;
1049
1050 mv_chan = devm_kzalloc(&pdev->dev, sizeof(*mv_chan), GFP_KERNEL);
1051 if (!mv_chan)
1052 return ERR_PTR(-ENOMEM);
1053
1054 mv_chan->idx = idx;
1055 mv_chan->irq = irq;
1056 if (xordev->xor_type == XOR_ORION)
1057 mv_chan->op_in_desc = XOR_MODE_IN_REG;
1058 else
1059 mv_chan->op_in_desc = XOR_MODE_IN_DESC;
1060
1061 dma_dev = &mv_chan->dmadev;
1062 mv_chan->xordev = xordev;
1063
1064 /*
1065 * These source and destination dummy buffers are used to implement
1066 * a DMA_INTERRUPT operation as a minimum-sized XOR operation.
1067 * Hence, we only need to map the buffers at initialization-time.
1068 */
1069 mv_chan->dummy_src_addr = dma_map_single(dma_dev->dev,
1070 mv_chan->dummy_src, MV_XOR_MIN_BYTE_COUNT, DMA_FROM_DEVICE);
1071 mv_chan->dummy_dst_addr = dma_map_single(dma_dev->dev,
1072 mv_chan->dummy_dst, MV_XOR_MIN_BYTE_COUNT, DMA_TO_DEVICE);
1073
1074 /* allocate coherent memory for hardware descriptors
1075 * note: writecombine gives slightly better performance, but
1076 * requires that we explicitly flush the writes
1077 */
1078 mv_chan->dma_desc_pool_virt =
1079 dma_alloc_wc(&pdev->dev, MV_XOR_POOL_SIZE, &mv_chan->dma_desc_pool,
1080 GFP_KERNEL);
1081 if (!mv_chan->dma_desc_pool_virt)
1082 return ERR_PTR(-ENOMEM);
1083
1084 /* discover transaction capabilites from the platform data */
1085 dma_dev->cap_mask = cap_mask;
1086
1087 INIT_LIST_HEAD(&dma_dev->channels);
1088
1089 /* set base routines */
1090 dma_dev->device_alloc_chan_resources = mv_xor_alloc_chan_resources;
1091 dma_dev->device_free_chan_resources = mv_xor_free_chan_resources;
1092 dma_dev->device_tx_status = mv_xor_status;
1093 dma_dev->device_issue_pending = mv_xor_issue_pending;
1094 dma_dev->dev = &pdev->dev;
1095
1096 /* set prep routines based on capability */
1097 if (dma_has_cap(DMA_INTERRUPT, dma_dev->cap_mask))
1098 dma_dev->device_prep_dma_interrupt = mv_xor_prep_dma_interrupt;
1099 if (dma_has_cap(DMA_MEMCPY, dma_dev->cap_mask))
1100 dma_dev->device_prep_dma_memcpy = mv_xor_prep_dma_memcpy;
1101 if (dma_has_cap(DMA_XOR, dma_dev->cap_mask)) {
1102 dma_dev->max_xor = 8;
1103 dma_dev->device_prep_dma_xor = mv_xor_prep_dma_xor;
1104 }
1105
1106 mv_chan->mmr_base = xordev->xor_base;
1107 mv_chan->mmr_high_base = xordev->xor_high_base;
1108 tasklet_init(&mv_chan->irq_tasklet, mv_xor_tasklet, (unsigned long)
1109 mv_chan);
1110
1111 /* clear errors before enabling interrupts */
1112 mv_chan_clear_err_status(mv_chan);
1113
1114 ret = request_irq(mv_chan->irq, mv_xor_interrupt_handler,
1115 0, dev_name(&pdev->dev), mv_chan);
1116 if (ret)
1117 goto err_free_dma;
1118
1119 mv_chan_unmask_interrupts(mv_chan);
1120
1121 if (mv_chan->op_in_desc == XOR_MODE_IN_DESC)
1122 mv_chan_set_mode(mv_chan, XOR_OPERATION_MODE_IN_DESC);
1123 else
1124 mv_chan_set_mode(mv_chan, XOR_OPERATION_MODE_XOR);
1125
1126 spin_lock_init(&mv_chan->lock);
1127 INIT_LIST_HEAD(&mv_chan->chain);
1128 INIT_LIST_HEAD(&mv_chan->completed_slots);
1129 INIT_LIST_HEAD(&mv_chan->free_slots);
1130 INIT_LIST_HEAD(&mv_chan->allocated_slots);
1131 mv_chan->dmachan.device = dma_dev;
1132 dma_cookie_init(&mv_chan->dmachan);
1133
1134 list_add_tail(&mv_chan->dmachan.device_node, &dma_dev->channels);
1135
1136 if (dma_has_cap(DMA_MEMCPY, dma_dev->cap_mask)) {
1137 ret = mv_chan_memcpy_self_test(mv_chan);
1138 dev_dbg(&pdev->dev, "memcpy self test returned %d\n", ret);
1139 if (ret)
1140 goto err_free_irq;
1141 }
1142
1143 if (dma_has_cap(DMA_XOR, dma_dev->cap_mask)) {
1144 ret = mv_chan_xor_self_test(mv_chan);
1145 dev_dbg(&pdev->dev, "xor self test returned %d\n", ret);
1146 if (ret)
1147 goto err_free_irq;
1148 }
1149
1150 dev_info(&pdev->dev, "Marvell XOR (%s): ( %s%s%s)\n",
1151 mv_chan->op_in_desc ? "Descriptor Mode" : "Registers Mode",
1152 dma_has_cap(DMA_XOR, dma_dev->cap_mask) ? "xor " : "",
1153 dma_has_cap(DMA_MEMCPY, dma_dev->cap_mask) ? "cpy " : "",
1154 dma_has_cap(DMA_INTERRUPT, dma_dev->cap_mask) ? "intr " : "");
1155
1156 dma_async_device_register(dma_dev);
1157 return mv_chan;
1158
1159err_free_irq:
1160 free_irq(mv_chan->irq, mv_chan);
1161err_free_dma:
1162 dma_free_coherent(&pdev->dev, MV_XOR_POOL_SIZE,
1163 mv_chan->dma_desc_pool_virt, mv_chan->dma_desc_pool);
1164 return ERR_PTR(ret);
1165}
1166
1167static void
1168mv_xor_conf_mbus_windows(struct mv_xor_device *xordev,
1169 const struct mbus_dram_target_info *dram)
1170{
1171 void __iomem *base = xordev->xor_high_base;
1172 u32 win_enable = 0;
1173 int i;
1174
1175 for (i = 0; i < 8; i++) {
1176 writel(0, base + WINDOW_BASE(i));
1177 writel(0, base + WINDOW_SIZE(i));
1178 if (i < 4)
1179 writel(0, base + WINDOW_REMAP_HIGH(i));
1180 }
1181
1182 for (i = 0; i < dram->num_cs; i++) {
1183 const struct mbus_dram_window *cs = dram->cs + i;
1184
1185 writel((cs->base & 0xffff0000) |
1186 (cs->mbus_attr << 8) |
1187 dram->mbus_dram_target_id, base + WINDOW_BASE(i));
1188 writel((cs->size - 1) & 0xffff0000, base + WINDOW_SIZE(i));
1189
1190 /* Fill the caching variables for later use */
1191 xordev->win_start[i] = cs->base;
1192 xordev->win_end[i] = cs->base + cs->size - 1;
1193
1194 win_enable |= (1 << i);
1195 win_enable |= 3 << (16 + (2 * i));
1196 }
1197
1198 writel(win_enable, base + WINDOW_BAR_ENABLE(0));
1199 writel(win_enable, base + WINDOW_BAR_ENABLE(1));
1200 writel(0, base + WINDOW_OVERRIDE_CTRL(0));
1201 writel(0, base + WINDOW_OVERRIDE_CTRL(1));
1202}
1203
1204static void
1205mv_xor_conf_mbus_windows_a3700(struct mv_xor_device *xordev)
1206{
1207 void __iomem *base = xordev->xor_high_base;
1208 u32 win_enable = 0;
1209 int i;
1210
1211 for (i = 0; i < 8; i++) {
1212 writel(0, base + WINDOW_BASE(i));
1213 writel(0, base + WINDOW_SIZE(i));
1214 if (i < 4)
1215 writel(0, base + WINDOW_REMAP_HIGH(i));
1216 }
1217 /*
1218 * For Armada3700 open default 4GB Mbus window. The dram
1219 * related configuration are done at AXIS level.
1220 */
1221 writel(0xffff0000, base + WINDOW_SIZE(0));
1222 win_enable |= 1;
1223 win_enable |= 3 << 16;
1224
1225 writel(win_enable, base + WINDOW_BAR_ENABLE(0));
1226 writel(win_enable, base + WINDOW_BAR_ENABLE(1));
1227 writel(0, base + WINDOW_OVERRIDE_CTRL(0));
1228 writel(0, base + WINDOW_OVERRIDE_CTRL(1));
1229}
1230
1231/*
1232 * Since this XOR driver is basically used only for RAID5, we don't
1233 * need to care about synchronizing ->suspend with DMA activity,
1234 * because the DMA engine will naturally be quiet due to the block
1235 * devices being suspended.
1236 */
1237static int mv_xor_suspend(struct platform_device *pdev, pm_message_t state)
1238{
1239 struct mv_xor_device *xordev = platform_get_drvdata(pdev);
1240 int i;
1241
1242 for (i = 0; i < MV_XOR_MAX_CHANNELS; i++) {
1243 struct mv_xor_chan *mv_chan = xordev->channels[i];
1244
1245 if (!mv_chan)
1246 continue;
1247
1248 mv_chan->saved_config_reg =
1249 readl_relaxed(XOR_CONFIG(mv_chan));
1250 mv_chan->saved_int_mask_reg =
1251 readl_relaxed(XOR_INTR_MASK(mv_chan));
1252 }
1253
1254 return 0;
1255}
1256
1257static int mv_xor_resume(struct platform_device *dev)
1258{
1259 struct mv_xor_device *xordev = platform_get_drvdata(dev);
1260 const struct mbus_dram_target_info *dram;
1261 int i;
1262
1263 for (i = 0; i < MV_XOR_MAX_CHANNELS; i++) {
1264 struct mv_xor_chan *mv_chan = xordev->channels[i];
1265
1266 if (!mv_chan)
1267 continue;
1268
1269 writel_relaxed(mv_chan->saved_config_reg,
1270 XOR_CONFIG(mv_chan));
1271 writel_relaxed(mv_chan->saved_int_mask_reg,
1272 XOR_INTR_MASK(mv_chan));
1273 }
1274
1275 if (xordev->xor_type == XOR_ARMADA_37XX) {
1276 mv_xor_conf_mbus_windows_a3700(xordev);
1277 return 0;
1278 }
1279
1280 dram = mv_mbus_dram_info();
1281 if (dram)
1282 mv_xor_conf_mbus_windows(xordev, dram);
1283
1284 return 0;
1285}
1286
1287static const struct of_device_id mv_xor_dt_ids[] = {
1288 { .compatible = "marvell,orion-xor", .data = (void *)XOR_ORION },
1289 { .compatible = "marvell,armada-380-xor", .data = (void *)XOR_ARMADA_38X },
1290 { .compatible = "marvell,armada-3700-xor", .data = (void *)XOR_ARMADA_37XX },
1291 {},
1292};
1293
1294static unsigned int mv_xor_engine_count;
1295
1296static int mv_xor_probe(struct platform_device *pdev)
1297{
1298 const struct mbus_dram_target_info *dram;
1299 struct mv_xor_device *xordev;
1300 struct mv_xor_platform_data *pdata = dev_get_platdata(&pdev->dev);
1301 struct resource *res;
1302 unsigned int max_engines, max_channels;
1303 int i, ret;
1304
1305 dev_notice(&pdev->dev, "Marvell shared XOR driver\n");
1306
1307 xordev = devm_kzalloc(&pdev->dev, sizeof(*xordev), GFP_KERNEL);
1308 if (!xordev)
1309 return -ENOMEM;
1310
1311 res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
1312 if (!res)
1313 return -ENODEV;
1314
1315 xordev->xor_base = devm_ioremap(&pdev->dev, res->start,
1316 resource_size(res));
1317 if (!xordev->xor_base)
1318 return -EBUSY;
1319
1320 res = platform_get_resource(pdev, IORESOURCE_MEM, 1);
1321 if (!res)
1322 return -ENODEV;
1323
1324 xordev->xor_high_base = devm_ioremap(&pdev->dev, res->start,
1325 resource_size(res));
1326 if (!xordev->xor_high_base)
1327 return -EBUSY;
1328
1329 platform_set_drvdata(pdev, xordev);
1330
1331
1332 /*
1333 * We need to know which type of XOR device we use before
1334 * setting up. In non-dt case it can only be the legacy one.
1335 */
1336 xordev->xor_type = XOR_ORION;
1337 if (pdev->dev.of_node) {
1338 const struct of_device_id *of_id =
1339 of_match_device(mv_xor_dt_ids,
1340 &pdev->dev);
1341
1342 xordev->xor_type = (uintptr_t)of_id->data;
1343 }
1344
1345 /*
1346 * (Re-)program MBUS remapping windows if we are asked to.
1347 */
1348 if (xordev->xor_type == XOR_ARMADA_37XX) {
1349 mv_xor_conf_mbus_windows_a3700(xordev);
1350 } else {
1351 dram = mv_mbus_dram_info();
1352 if (dram)
1353 mv_xor_conf_mbus_windows(xordev, dram);
1354 }
1355
1356 /* Not all platforms can gate the clock, so it is not
1357 * an error if the clock does not exists.
1358 */
1359 xordev->clk = clk_get(&pdev->dev, NULL);
1360 if (!IS_ERR(xordev->clk))
1361 clk_prepare_enable(xordev->clk);
1362
1363 /*
1364 * We don't want to have more than one channel per CPU in
1365 * order for async_tx to perform well. So we limit the number
1366 * of engines and channels so that we take into account this
1367 * constraint. Note that we also want to use channels from
1368 * separate engines when possible. For dual-CPU Armada 3700
1369 * SoC with single XOR engine allow using its both channels.
1370 */
1371 max_engines = num_present_cpus();
1372 if (xordev->xor_type == XOR_ARMADA_37XX)
1373 max_channels = num_present_cpus();
1374 else
1375 max_channels = min_t(unsigned int,
1376 MV_XOR_MAX_CHANNELS,
1377 DIV_ROUND_UP(num_present_cpus(), 2));
1378
1379 if (mv_xor_engine_count >= max_engines)
1380 return 0;
1381
1382 if (pdev->dev.of_node) {
1383 struct device_node *np;
1384 int i = 0;
1385
1386 for_each_child_of_node(pdev->dev.of_node, np) {
1387 struct mv_xor_chan *chan;
1388 dma_cap_mask_t cap_mask;
1389 int irq;
1390
1391 if (i >= max_channels)
1392 continue;
1393
1394 dma_cap_zero(cap_mask);
1395 dma_cap_set(DMA_MEMCPY, cap_mask);
1396 dma_cap_set(DMA_XOR, cap_mask);
1397 dma_cap_set(DMA_INTERRUPT, cap_mask);
1398
1399 irq = irq_of_parse_and_map(np, 0);
1400 if (!irq) {
1401 ret = -ENODEV;
1402 goto err_channel_add;
1403 }
1404
1405 chan = mv_xor_channel_add(xordev, pdev, i,
1406 cap_mask, irq);
1407 if (IS_ERR(chan)) {
1408 ret = PTR_ERR(chan);
1409 irq_dispose_mapping(irq);
1410 goto err_channel_add;
1411 }
1412
1413 xordev->channels[i] = chan;
1414 i++;
1415 }
1416 } else if (pdata && pdata->channels) {
1417 for (i = 0; i < max_channels; i++) {
1418 struct mv_xor_channel_data *cd;
1419 struct mv_xor_chan *chan;
1420 int irq;
1421
1422 cd = &pdata->channels[i];
1423 irq = platform_get_irq(pdev, i);
1424 if (irq < 0) {
1425 ret = irq;
1426 goto err_channel_add;
1427 }
1428
1429 chan = mv_xor_channel_add(xordev, pdev, i,
1430 cd->cap_mask, irq);
1431 if (IS_ERR(chan)) {
1432 ret = PTR_ERR(chan);
1433 goto err_channel_add;
1434 }
1435
1436 xordev->channels[i] = chan;
1437 }
1438 }
1439
1440 return 0;
1441
1442err_channel_add:
1443 for (i = 0; i < MV_XOR_MAX_CHANNELS; i++)
1444 if (xordev->channels[i]) {
1445 mv_xor_channel_remove(xordev->channels[i]);
1446 if (pdev->dev.of_node)
1447 irq_dispose_mapping(xordev->channels[i]->irq);
1448 }
1449
1450 if (!IS_ERR(xordev->clk)) {
1451 clk_disable_unprepare(xordev->clk);
1452 clk_put(xordev->clk);
1453 }
1454
1455 return ret;
1456}
1457
1458static struct platform_driver mv_xor_driver = {
1459 .probe = mv_xor_probe,
1460 .suspend = mv_xor_suspend,
1461 .resume = mv_xor_resume,
1462 .driver = {
1463 .name = MV_XOR_NAME,
1464 .of_match_table = of_match_ptr(mv_xor_dt_ids),
1465 },
1466};
1467
1468builtin_platform_driver(mv_xor_driver);
1469
1470/*
1471MODULE_AUTHOR("Saeed Bishara <saeed@marvell.com>");
1472MODULE_DESCRIPTION("DMA engine driver for Marvell's XOR engine");
1473MODULE_LICENSE("GPL");
1474*/