1 /*
2  * DMM IOMMU driver support functions for TI OMAP processors.
3  *
4  * Copyright (C) 2011 Texas Instruments Incorporated - http://www.ti.com/
5  * Author: Rob Clark <rob@ti.com>
6  *         Andy Gross <andy.gross@ti.com>
7  *
8  * This program is free software; you can redistribute it and/or
9  * modify it under the terms of the GNU General Public License as
10  * published by the Free Software Foundation version 2.
11  *
12  * This program is distributed "as is" WITHOUT ANY WARRANTY of any
13  * kind, whether express or implied; without even the implied warranty
14  * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
15  * GNU General Public License for more details.
16  */
17 
18 #include <linux/completion.h>
19 #include <linux/delay.h>
20 #include <linux/dma-mapping.h>
21 #include <linux/dmaengine.h>
22 #include <linux/errno.h>
23 #include <linux/init.h>
24 #include <linux/interrupt.h>
25 #include <linux/list.h>
26 #include <linux/mm.h>
27 #include <linux/module.h>
28 #include <linux/platform_device.h> /* platform_device() */
29 #include <linux/sched.h>
30 #include <linux/seq_file.h>
31 #include <linux/slab.h>
32 #include <linux/time.h>
33 #include <linux/vmalloc.h>
34 #include <linux/wait.h>
35 
36 #include "omap_dmm_tiler.h"
37 #include "omap_dmm_priv.h"
38 
39 #define DMM_DRIVER_NAME "dmm"
40 
41 /* mappings for associating views to luts */
42 static struct tcm *containers[TILFMT_NFORMATS];
43 static struct dmm *omap_dmm;
44 
45 #if defined(CONFIG_OF)
46 static const struct of_device_id dmm_of_match[];
47 #endif
48 
49 /* global spinlock for protecting lists */
50 static DEFINE_SPINLOCK(list_lock);
51 
52 /* Geometry table */
53 #define GEOM(xshift, yshift, bytes_per_pixel) { \
54 		.x_shft = (xshift), \
55 		.y_shft = (yshift), \
56 		.cpp    = (bytes_per_pixel), \
57 		.slot_w = 1 << (SLOT_WIDTH_BITS - (xshift)), \
58 		.slot_h = 1 << (SLOT_HEIGHT_BITS - (yshift)), \
59 	}
60 
61 static const struct {
62 	u32 x_shft;	/* unused X-bits (as part of bpp) */
63 	u32 y_shft;	/* unused Y-bits (as part of bpp) */
64 	u32 cpp;		/* bytes/chars per pixel */
65 	u32 slot_w;	/* width of each slot (in pixels) */
66 	u32 slot_h;	/* height of each slot (in pixels) */
67 } geom[TILFMT_NFORMATS] = {
68 	[TILFMT_8BIT]  = GEOM(0, 0, 1),
69 	[TILFMT_16BIT] = GEOM(0, 1, 2),
70 	[TILFMT_32BIT] = GEOM(1, 1, 4),
71 	[TILFMT_PAGE]  = GEOM(SLOT_WIDTH_BITS, SLOT_HEIGHT_BITS, 1),
72 };
73 
74 
75 /* lookup table for registers w/ per-engine instances */
76 static const u32 reg[][4] = {
77 	[PAT_STATUS] = {DMM_PAT_STATUS__0, DMM_PAT_STATUS__1,
78 			DMM_PAT_STATUS__2, DMM_PAT_STATUS__3},
79 	[PAT_DESCR]  = {DMM_PAT_DESCR__0, DMM_PAT_DESCR__1,
80 			DMM_PAT_DESCR__2, DMM_PAT_DESCR__3},
81 };
82 
83 static int dmm_dma_copy(struct dmm *dmm, dma_addr_t src, dma_addr_t dst)
84 {
85 	struct dma_device *dma_dev = dmm->wa_dma_chan->device;
86 	struct dma_async_tx_descriptor *tx;
87 	enum dma_status status;
88 	dma_cookie_t cookie;
89 
90 	tx = dma_dev->device_prep_dma_memcpy(dmm->wa_dma_chan, dst, src, 4, 0);
91 	if (!tx) {
92 		dev_err(dmm->dev, "Failed to prepare DMA memcpy\n");
93 		return -EIO;
94 	}
95 
96 	cookie = tx->tx_submit(tx);
97 	if (dma_submit_error(cookie)) {
98 		dev_err(dmm->dev, "Failed to do DMA tx_submit\n");
99 		return -EIO;
100 	}
101 
102 	dma_async_issue_pending(dmm->wa_dma_chan);
103 	status = dma_sync_wait(dmm->wa_dma_chan, cookie);
104 	if (status != DMA_COMPLETE)
105 		dev_err(dmm->dev, "i878 wa DMA copy failure\n");
106 
107 	dmaengine_terminate_all(dmm->wa_dma_chan);
108 	return 0;
109 }
110 
111 static u32 dmm_read_wa(struct dmm *dmm, u32 reg)
112 {
113 	dma_addr_t src, dst;
114 	int r;
115 
116 	src = dmm->phys_base + reg;
117 	dst = dmm->wa_dma_handle;
118 
119 	r = dmm_dma_copy(dmm, src, dst);
120 	if (r) {
121 		dev_err(dmm->dev, "sDMA read transfer timeout\n");
122 		return readl(dmm->base + reg);
123 	}
124 
125 	/*
126 	 * As per i878 workaround, the DMA is used to access the DMM registers.
127 	 * Make sure that the readl is not moved by the compiler or the CPU
128 	 * earlier than the DMA finished writing the value to memory.
129 	 */
130 	rmb();
131 	return readl(dmm->wa_dma_data);
132 }
133 
134 static void dmm_write_wa(struct dmm *dmm, u32 val, u32 reg)
135 {
136 	dma_addr_t src, dst;
137 	int r;
138 
139 	writel(val, dmm->wa_dma_data);
140 	/*
141 	 * As per i878 workaround, the DMA is used to access the DMM registers.
142 	 * Make sure that the writel is not moved by the compiler or the CPU, so
143 	 * the data will be in place before we start the DMA to do the actual
144 	 * register write.
145 	 */
146 	wmb();
147 
148 	src = dmm->wa_dma_handle;
149 	dst = dmm->phys_base + reg;
150 
151 	r = dmm_dma_copy(dmm, src, dst);
152 	if (r) {
153 		dev_err(dmm->dev, "sDMA write transfer timeout\n");
154 		writel(val, dmm->base + reg);
155 	}
156 }
157 
158 static u32 dmm_read(struct dmm *dmm, u32 reg)
159 {
160 	if (dmm->dmm_workaround) {
161 		u32 v;
162 		unsigned long flags;
163 
164 		spin_lock_irqsave(&dmm->wa_lock, flags);
165 		v = dmm_read_wa(dmm, reg);
166 		spin_unlock_irqrestore(&dmm->wa_lock, flags);
167 
168 		return v;
169 	} else {
170 		return readl(dmm->base + reg);
171 	}
172 }
173 
174 static void dmm_write(struct dmm *dmm, u32 val, u32 reg)
175 {
176 	if (dmm->dmm_workaround) {
177 		unsigned long flags;
178 
179 		spin_lock_irqsave(&dmm->wa_lock, flags);
180 		dmm_write_wa(dmm, val, reg);
181 		spin_unlock_irqrestore(&dmm->wa_lock, flags);
182 	} else {
183 		writel(val, dmm->base + reg);
184 	}
185 }
186 
187 static int dmm_workaround_init(struct dmm *dmm)
188 {
189 	dma_cap_mask_t mask;
190 
191 	spin_lock_init(&dmm->wa_lock);
192 
193 	dmm->wa_dma_data = dma_alloc_coherent(dmm->dev,  sizeof(u32),
194 					      &dmm->wa_dma_handle, GFP_KERNEL);
195 	if (!dmm->wa_dma_data)
196 		return -ENOMEM;
197 
198 	dma_cap_zero(mask);
199 	dma_cap_set(DMA_MEMCPY, mask);
200 
201 	dmm->wa_dma_chan = dma_request_channel(mask, NULL, NULL);
202 	if (!dmm->wa_dma_chan) {
203 		dma_free_coherent(dmm->dev, 4, dmm->wa_dma_data, dmm->wa_dma_handle);
204 		return -ENODEV;
205 	}
206 
207 	return 0;
208 }
209 
210 static void dmm_workaround_uninit(struct dmm *dmm)
211 {
212 	dma_release_channel(dmm->wa_dma_chan);
213 
214 	dma_free_coherent(dmm->dev, 4, dmm->wa_dma_data, dmm->wa_dma_handle);
215 }
216 
217 /* simple allocator to grab next 16 byte aligned memory from txn */
218 static void *alloc_dma(struct dmm_txn *txn, size_t sz, dma_addr_t *pa)
219 {
220 	void *ptr;
221 	struct refill_engine *engine = txn->engine_handle;
222 
223 	/* dmm programming requires 16 byte aligned addresses */
224 	txn->current_pa = round_up(txn->current_pa, 16);
225 	txn->current_va = (void *)round_up((long)txn->current_va, 16);
226 
227 	ptr = txn->current_va;
228 	*pa = txn->current_pa;
229 
230 	txn->current_pa += sz;
231 	txn->current_va += sz;
232 
233 	BUG_ON((txn->current_va - engine->refill_va) > REFILL_BUFFER_SIZE);
234 
235 	return ptr;
236 }
237 
238 /* check status and spin until wait_mask comes true */
239 static int wait_status(struct refill_engine *engine, u32 wait_mask)
240 {
241 	struct dmm *dmm = engine->dmm;
242 	u32 r = 0, err, i;
243 
244 	i = DMM_FIXED_RETRY_COUNT;
245 	while (true) {
246 		r = dmm_read(dmm, reg[PAT_STATUS][engine->id]);
247 		err = r & DMM_PATSTATUS_ERR;
248 		if (err) {
249 			dev_err(dmm->dev,
250 				"%s: error (engine%d). PAT_STATUS: 0x%08x\n",
251 				__func__, engine->id, r);
252 			return -EFAULT;
253 		}
254 
255 		if ((r & wait_mask) == wait_mask)
256 			break;
257 
258 		if (--i == 0) {
259 			dev_err(dmm->dev,
260 				"%s: timeout (engine%d). PAT_STATUS: 0x%08x\n",
261 				__func__, engine->id, r);
262 			return -ETIMEDOUT;
263 		}
264 
265 		udelay(1);
266 	}
267 
268 	return 0;
269 }
270 
271 static void release_engine(struct refill_engine *engine)
272 {
273 	unsigned long flags;
274 
275 	spin_lock_irqsave(&list_lock, flags);
276 	list_add(&engine->idle_node, &omap_dmm->idle_head);
277 	spin_unlock_irqrestore(&list_lock, flags);
278 
279 	atomic_inc(&omap_dmm->engine_counter);
280 	wake_up_interruptible(&omap_dmm->engine_queue);
281 }
282 
283 static irqreturn_t omap_dmm_irq_handler(int irq, void *arg)
284 {
285 	struct dmm *dmm = arg;
286 	u32 status = dmm_read(dmm, DMM_PAT_IRQSTATUS);
287 	int i;
288 
289 	/* ack IRQ */
290 	dmm_write(dmm, status, DMM_PAT_IRQSTATUS);
291 
292 	for (i = 0; i < dmm->num_engines; i++) {
293 		if (status & DMM_IRQSTAT_ERR_MASK)
294 			dev_err(dmm->dev,
295 				"irq error(engine%d): IRQSTAT 0x%02x\n",
296 				i, status & 0xff);
297 
298 		if (status & DMM_IRQSTAT_LST) {
299 			if (dmm->engines[i].async)
300 				release_engine(&dmm->engines[i]);
301 
302 			complete(&dmm->engines[i].compl);
303 		}
304 
305 		status >>= 8;
306 	}
307 
308 	return IRQ_HANDLED;
309 }
310 
311 /**
312  * Get a handle for a DMM transaction
313  */
314 static struct dmm_txn *dmm_txn_init(struct dmm *dmm, struct tcm *tcm)
315 {
316 	struct dmm_txn *txn = NULL;
317 	struct refill_engine *engine = NULL;
318 	int ret;
319 	unsigned long flags;
320 
321 
322 	/* wait until an engine is available */
323 	ret = wait_event_interruptible(omap_dmm->engine_queue,
324 		atomic_add_unless(&omap_dmm->engine_counter, -1, 0));
325 	if (ret)
326 		return ERR_PTR(ret);
327 
328 	/* grab an idle engine */
329 	spin_lock_irqsave(&list_lock, flags);
330 	if (!list_empty(&dmm->idle_head)) {
331 		engine = list_entry(dmm->idle_head.next, struct refill_engine,
332 					idle_node);
333 		list_del(&engine->idle_node);
334 	}
335 	spin_unlock_irqrestore(&list_lock, flags);
336 
337 	BUG_ON(!engine);
338 
339 	txn = &engine->txn;
340 	engine->tcm = tcm;
341 	txn->engine_handle = engine;
342 	txn->last_pat = NULL;
343 	txn->current_va = engine->refill_va;
344 	txn->current_pa = engine->refill_pa;
345 
346 	return txn;
347 }
348 
349 /**
350  * Add region to DMM transaction.  If pages or pages[i] is NULL, then the
351  * corresponding slot is cleared (ie. dummy_pa is programmed)
352  */
353 static void dmm_txn_append(struct dmm_txn *txn, struct pat_area *area,
354 		struct page **pages, u32 npages, u32 roll)
355 {
356 	dma_addr_t pat_pa = 0, data_pa = 0;
357 	u32 *data;
358 	struct pat *pat;
359 	struct refill_engine *engine = txn->engine_handle;
360 	int columns = (1 + area->x1 - area->x0);
361 	int rows = (1 + area->y1 - area->y0);
362 	int i = columns*rows;
363 
364 	pat = alloc_dma(txn, sizeof(*pat), &pat_pa);
365 
366 	if (txn->last_pat)
367 		txn->last_pat->next_pa = (u32)pat_pa;
368 
369 	pat->area = *area;
370 
371 	/* adjust Y coordinates based off of container parameters */
372 	pat->area.y0 += engine->tcm->y_offset;
373 	pat->area.y1 += engine->tcm->y_offset;
374 
375 	pat->ctrl = (struct pat_ctrl){
376 			.start = 1,
377 			.lut_id = engine->tcm->lut_id,
378 		};
379 
380 	data = alloc_dma(txn, 4*i, &data_pa);
381 	/* FIXME: what if data_pa is more than 32-bit ? */
382 	pat->data_pa = data_pa;
383 
384 	while (i--) {
385 		int n = i + roll;
386 		if (n >= npages)
387 			n -= npages;
388 		data[i] = (pages && pages[n]) ?
389 			page_to_phys(pages[n]) : engine->dmm->dummy_pa;
390 	}
391 
392 	txn->last_pat = pat;
393 
394 	return;
395 }
396 
397 /**
398  * Commit the DMM transaction.
399  */
400 static int dmm_txn_commit(struct dmm_txn *txn, bool wait)
401 {
402 	int ret = 0;
403 	struct refill_engine *engine = txn->engine_handle;
404 	struct dmm *dmm = engine->dmm;
405 
406 	if (!txn->last_pat) {
407 		dev_err(engine->dmm->dev, "need at least one txn\n");
408 		ret = -EINVAL;
409 		goto cleanup;
410 	}
411 
412 	txn->last_pat->next_pa = 0;
413 	/* ensure that the written descriptors are visible to DMM */
414 	wmb();
415 
416 	/*
417 	 * NOTE: the wmb() above should be enough, but there seems to be a bug
418 	 * in OMAP's memory barrier implementation, which in some rare cases may
419 	 * cause the writes not to be observable after wmb().
420 	 */
421 
422 	/* read back to ensure the data is in RAM */
423 	readl(&txn->last_pat->next_pa);
424 
425 	/* write to PAT_DESCR to clear out any pending transaction */
426 	dmm_write(dmm, 0x0, reg[PAT_DESCR][engine->id]);
427 
428 	/* wait for engine ready: */
429 	ret = wait_status(engine, DMM_PATSTATUS_READY);
430 	if (ret) {
431 		ret = -EFAULT;
432 		goto cleanup;
433 	}
434 
435 	/* mark whether it is async to denote list management in IRQ handler */
436 	engine->async = wait ? false : true;
437 	reinit_completion(&engine->compl);
438 	/* verify that the irq handler sees the 'async' and completion value */
439 	smp_mb();
440 
441 	/* kick reload */
442 	dmm_write(dmm, engine->refill_pa, reg[PAT_DESCR][engine->id]);
443 
444 	if (wait) {
445 		if (!wait_for_completion_timeout(&engine->compl,
446 				msecs_to_jiffies(100))) {
447 			dev_err(dmm->dev, "timed out waiting for done\n");
448 			ret = -ETIMEDOUT;
449 			goto cleanup;
450 		}
451 
452 		/* Check the engine status before continue */
453 		ret = wait_status(engine, DMM_PATSTATUS_READY |
454 				  DMM_PATSTATUS_VALID | DMM_PATSTATUS_DONE);
455 	}
456 
457 cleanup:
458 	/* only place engine back on list if we are done with it */
459 	if (ret || wait)
460 		release_engine(engine);
461 
462 	return ret;
463 }
464 
465 /*
466  * DMM programming
467  */
468 static int fill(struct tcm_area *area, struct page **pages,
469 		u32 npages, u32 roll, bool wait)
470 {
471 	int ret = 0;
472 	struct tcm_area slice, area_s;
473 	struct dmm_txn *txn;
474 
475 	/*
476 	 * FIXME
477 	 *
478 	 * Asynchronous fill does not work reliably, as the driver does not
479 	 * handle errors in the async code paths. The fill operation may
480 	 * silently fail, leading to leaking DMM engines, which may eventually
481 	 * lead to deadlock if we run out of DMM engines.
482 	 *
483 	 * For now, always set 'wait' so that we only use sync fills. Async
484 	 * fills should be fixed, or alternatively we could decide to only
485 	 * support sync fills and so the whole async code path could be removed.
486 	 */
487 
488 	wait = true;
489 
490 	txn = dmm_txn_init(omap_dmm, area->tcm);
491 	if (IS_ERR_OR_NULL(txn))
492 		return -ENOMEM;
493 
494 	tcm_for_each_slice(slice, *area, area_s) {
495 		struct pat_area p_area = {
496 				.x0 = slice.p0.x,  .y0 = slice.p0.y,
497 				.x1 = slice.p1.x,  .y1 = slice.p1.y,
498 		};
499 
500 		dmm_txn_append(txn, &p_area, pages, npages, roll);
501 
502 		roll += tcm_sizeof(slice);
503 	}
504 
505 	ret = dmm_txn_commit(txn, wait);
506 
507 	return ret;
508 }
509 
510 /*
511  * Pin/unpin
512  */
513 
514 /* note: slots for which pages[i] == NULL are filled w/ dummy page
515  */
516 int tiler_pin(struct tiler_block *block, struct page **pages,
517 		u32 npages, u32 roll, bool wait)
518 {
519 	int ret;
520 
521 	ret = fill(&block->area, pages, npages, roll, wait);
522 
523 	if (ret)
524 		tiler_unpin(block);
525 
526 	return ret;
527 }
528 
529 int tiler_unpin(struct tiler_block *block)
530 {
531 	return fill(&block->area, NULL, 0, 0, false);
532 }
533 
534 /*
535  * Reserve/release
536  */
537 struct tiler_block *tiler_reserve_2d(enum tiler_fmt fmt, u16 w,
538 		u16 h, u16 align)
539 {
540 	struct tiler_block *block;
541 	u32 min_align = 128;
542 	int ret;
543 	unsigned long flags;
544 	u32 slot_bytes;
545 
546 	block = kzalloc(sizeof(*block), GFP_KERNEL);
547 	if (!block)
548 		return ERR_PTR(-ENOMEM);
549 
550 	BUG_ON(!validfmt(fmt));
551 
552 	/* convert width/height to slots */
553 	w = DIV_ROUND_UP(w, geom[fmt].slot_w);
554 	h = DIV_ROUND_UP(h, geom[fmt].slot_h);
555 
556 	/* convert alignment to slots */
557 	slot_bytes = geom[fmt].slot_w * geom[fmt].cpp;
558 	min_align = max(min_align, slot_bytes);
559 	align = (align > min_align) ? ALIGN(align, min_align) : min_align;
560 	align /= slot_bytes;
561 
562 	block->fmt = fmt;
563 
564 	ret = tcm_reserve_2d(containers[fmt], w, h, align, -1, slot_bytes,
565 			&block->area);
566 	if (ret) {
567 		kfree(block);
568 		return ERR_PTR(-ENOMEM);
569 	}
570 
571 	/* add to allocation list */
572 	spin_lock_irqsave(&list_lock, flags);
573 	list_add(&block->alloc_node, &omap_dmm->alloc_head);
574 	spin_unlock_irqrestore(&list_lock, flags);
575 
576 	return block;
577 }
578 
579 struct tiler_block *tiler_reserve_1d(size_t size)
580 {
581 	struct tiler_block *block = kzalloc(sizeof(*block), GFP_KERNEL);
582 	int num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
583 	unsigned long flags;
584 
585 	if (!block)
586 		return ERR_PTR(-ENOMEM);
587 
588 	block->fmt = TILFMT_PAGE;
589 
590 	if (tcm_reserve_1d(containers[TILFMT_PAGE], num_pages,
591 				&block->area)) {
592 		kfree(block);
593 		return ERR_PTR(-ENOMEM);
594 	}
595 
596 	spin_lock_irqsave(&list_lock, flags);
597 	list_add(&block->alloc_node, &omap_dmm->alloc_head);
598 	spin_unlock_irqrestore(&list_lock, flags);
599 
600 	return block;
601 }
602 
603 /* note: if you have pin'd pages, you should have already unpin'd first! */
604 int tiler_release(struct tiler_block *block)
605 {
606 	int ret = tcm_free(&block->area);
607 	unsigned long flags;
608 
609 	if (block->area.tcm)
610 		dev_err(omap_dmm->dev, "failed to release block\n");
611 
612 	spin_lock_irqsave(&list_lock, flags);
613 	list_del(&block->alloc_node);
614 	spin_unlock_irqrestore(&list_lock, flags);
615 
616 	kfree(block);
617 	return ret;
618 }
619 
620 /*
621  * Utils
622  */
623 
624 /* calculate the tiler space address of a pixel in a view orientation...
625  * below description copied from the display subsystem section of TRM:
626  *
627  * When the TILER is addressed, the bits:
628  *   [28:27] = 0x0 for 8-bit tiled
629  *             0x1 for 16-bit tiled
630  *             0x2 for 32-bit tiled
631  *             0x3 for page mode
632  *   [31:29] = 0x0 for 0-degree view
633  *             0x1 for 180-degree view + mirroring
634  *             0x2 for 0-degree view + mirroring
635  *             0x3 for 180-degree view
636  *             0x4 for 270-degree view + mirroring
637  *             0x5 for 270-degree view
638  *             0x6 for 90-degree view
639  *             0x7 for 90-degree view + mirroring
640  * Otherwise the bits indicated the corresponding bit address to access
641  * the SDRAM.
642  */
643 static u32 tiler_get_address(enum tiler_fmt fmt, u32 orient, u32 x, u32 y)
644 {
645 	u32 x_bits, y_bits, tmp, x_mask, y_mask, alignment;
646 
647 	x_bits = CONT_WIDTH_BITS - geom[fmt].x_shft;
648 	y_bits = CONT_HEIGHT_BITS - geom[fmt].y_shft;
649 	alignment = geom[fmt].x_shft + geom[fmt].y_shft;
650 
651 	/* validate coordinate */
652 	x_mask = MASK(x_bits);
653 	y_mask = MASK(y_bits);
654 
655 	if (x < 0 || x > x_mask || y < 0 || y > y_mask) {
656 		DBG("invalid coords: %u < 0 || %u > %u || %u < 0 || %u > %u",
657 				x, x, x_mask, y, y, y_mask);
658 		return 0;
659 	}
660 
661 	/* account for mirroring */
662 	if (orient & MASK_X_INVERT)
663 		x ^= x_mask;
664 	if (orient & MASK_Y_INVERT)
665 		y ^= y_mask;
666 
667 	/* get coordinate address */
668 	if (orient & MASK_XY_FLIP)
669 		tmp = ((x << y_bits) + y);
670 	else
671 		tmp = ((y << x_bits) + x);
672 
673 	return TIL_ADDR((tmp << alignment), orient, fmt);
674 }
675 
676 dma_addr_t tiler_ssptr(struct tiler_block *block)
677 {
678 	BUG_ON(!validfmt(block->fmt));
679 
680 	return TILVIEW_8BIT + tiler_get_address(block->fmt, 0,
681 			block->area.p0.x * geom[block->fmt].slot_w,
682 			block->area.p0.y * geom[block->fmt].slot_h);
683 }
684 
685 dma_addr_t tiler_tsptr(struct tiler_block *block, u32 orient,
686 		u32 x, u32 y)
687 {
688 	struct tcm_pt *p = &block->area.p0;
689 	BUG_ON(!validfmt(block->fmt));
690 
691 	return tiler_get_address(block->fmt, orient,
692 			(p->x * geom[block->fmt].slot_w) + x,
693 			(p->y * geom[block->fmt].slot_h) + y);
694 }
695 
696 void tiler_align(enum tiler_fmt fmt, u16 *w, u16 *h)
697 {
698 	BUG_ON(!validfmt(fmt));
699 	*w = round_up(*w, geom[fmt].slot_w);
700 	*h = round_up(*h, geom[fmt].slot_h);
701 }
702 
703 u32 tiler_stride(enum tiler_fmt fmt, u32 orient)
704 {
705 	BUG_ON(!validfmt(fmt));
706 
707 	if (orient & MASK_XY_FLIP)
708 		return 1 << (CONT_HEIGHT_BITS + geom[fmt].x_shft);
709 	else
710 		return 1 << (CONT_WIDTH_BITS + geom[fmt].y_shft);
711 }
712 
713 size_t tiler_size(enum tiler_fmt fmt, u16 w, u16 h)
714 {
715 	tiler_align(fmt, &w, &h);
716 	return geom[fmt].cpp * w * h;
717 }
718 
719 size_t tiler_vsize(enum tiler_fmt fmt, u16 w, u16 h)
720 {
721 	BUG_ON(!validfmt(fmt));
722 	return round_up(geom[fmt].cpp * w, PAGE_SIZE) * h;
723 }
724 
725 u32 tiler_get_cpu_cache_flags(void)
726 {
727 	return omap_dmm->plat_data->cpu_cache_flags;
728 }
729 
730 bool dmm_is_available(void)
731 {
732 	return omap_dmm ? true : false;
733 }
734 
735 static int omap_dmm_remove(struct platform_device *dev)
736 {
737 	struct tiler_block *block, *_block;
738 	int i;
739 	unsigned long flags;
740 
741 	if (omap_dmm) {
742 		/* Disable all enabled interrupts */
743 		dmm_write(omap_dmm, 0x7e7e7e7e, DMM_PAT_IRQENABLE_CLR);
744 		free_irq(omap_dmm->irq, omap_dmm);
745 
746 		/* free all area regions */
747 		spin_lock_irqsave(&list_lock, flags);
748 		list_for_each_entry_safe(block, _block, &omap_dmm->alloc_head,
749 					alloc_node) {
750 			list_del(&block->alloc_node);
751 			kfree(block);
752 		}
753 		spin_unlock_irqrestore(&list_lock, flags);
754 
755 		for (i = 0; i < omap_dmm->num_lut; i++)
756 			if (omap_dmm->tcm && omap_dmm->tcm[i])
757 				omap_dmm->tcm[i]->deinit(omap_dmm->tcm[i]);
758 		kfree(omap_dmm->tcm);
759 
760 		kfree(omap_dmm->engines);
761 		if (omap_dmm->refill_va)
762 			dma_free_wc(omap_dmm->dev,
763 				    REFILL_BUFFER_SIZE * omap_dmm->num_engines,
764 				    omap_dmm->refill_va, omap_dmm->refill_pa);
765 		if (omap_dmm->dummy_page)
766 			__free_page(omap_dmm->dummy_page);
767 
768 		if (omap_dmm->dmm_workaround)
769 			dmm_workaround_uninit(omap_dmm);
770 
771 		iounmap(omap_dmm->base);
772 		kfree(omap_dmm);
773 		omap_dmm = NULL;
774 	}
775 
776 	return 0;
777 }
778 
779 static int omap_dmm_probe(struct platform_device *dev)
780 {
781 	int ret = -EFAULT, i;
782 	struct tcm_area area = {0};
783 	u32 hwinfo, pat_geom;
784 	struct resource *mem;
785 
786 	omap_dmm = kzalloc(sizeof(*omap_dmm), GFP_KERNEL);
787 	if (!omap_dmm)
788 		goto fail;
789 
790 	/* initialize lists */
791 	INIT_LIST_HEAD(&omap_dmm->alloc_head);
792 	INIT_LIST_HEAD(&omap_dmm->idle_head);
793 
794 	init_waitqueue_head(&omap_dmm->engine_queue);
795 
796 	if (dev->dev.of_node) {
797 		const struct of_device_id *match;
798 
799 		match = of_match_node(dmm_of_match, dev->dev.of_node);
800 		if (!match) {
801 			dev_err(&dev->dev, "failed to find matching device node\n");
802 			ret = -ENODEV;
803 			goto fail;
804 		}
805 
806 		omap_dmm->plat_data = match->data;
807 	}
808 
809 	/* lookup hwmod data - base address and irq */
810 	mem = platform_get_resource(dev, IORESOURCE_MEM, 0);
811 	if (!mem) {
812 		dev_err(&dev->dev, "failed to get base address resource\n");
813 		goto fail;
814 	}
815 
816 	omap_dmm->phys_base = mem->start;
817 	omap_dmm->base = ioremap(mem->start, SZ_2K);
818 
819 	if (!omap_dmm->base) {
820 		dev_err(&dev->dev, "failed to get dmm base address\n");
821 		goto fail;
822 	}
823 
824 	omap_dmm->irq = platform_get_irq(dev, 0);
825 	if (omap_dmm->irq < 0) {
826 		dev_err(&dev->dev, "failed to get IRQ resource\n");
827 		goto fail;
828 	}
829 
830 	omap_dmm->dev = &dev->dev;
831 
832 	if (of_machine_is_compatible("ti,dra7")) {
833 		/*
834 		 * DRA7 Errata i878 says that MPU should not be used to access
835 		 * RAM and DMM at the same time. As it's not possible to prevent
836 		 * MPU accessing RAM, we need to access DMM via a proxy.
837 		 */
838 		if (!dmm_workaround_init(omap_dmm)) {
839 			omap_dmm->dmm_workaround = true;
840 			dev_info(&dev->dev,
841 				"workaround for errata i878 in use\n");
842 		} else {
843 			dev_warn(&dev->dev,
844 				 "failed to initialize work-around for i878\n");
845 		}
846 	}
847 
848 	hwinfo = dmm_read(omap_dmm, DMM_PAT_HWINFO);
849 	omap_dmm->num_engines = (hwinfo >> 24) & 0x1F;
850 	omap_dmm->num_lut = (hwinfo >> 16) & 0x1F;
851 	omap_dmm->container_width = 256;
852 	omap_dmm->container_height = 128;
853 
854 	atomic_set(&omap_dmm->engine_counter, omap_dmm->num_engines);
855 
856 	/* read out actual LUT width and height */
857 	pat_geom = dmm_read(omap_dmm, DMM_PAT_GEOMETRY);
858 	omap_dmm->lut_width = ((pat_geom >> 16) & 0xF) << 5;
859 	omap_dmm->lut_height = ((pat_geom >> 24) & 0xF) << 5;
860 
861 	/* increment LUT by one if on OMAP5 */
862 	/* LUT has twice the height, and is split into a separate container */
863 	if (omap_dmm->lut_height != omap_dmm->container_height)
864 		omap_dmm->num_lut++;
865 
866 	/* initialize DMM registers */
867 	dmm_write(omap_dmm, 0x88888888, DMM_PAT_VIEW__0);
868 	dmm_write(omap_dmm, 0x88888888, DMM_PAT_VIEW__1);
869 	dmm_write(omap_dmm, 0x80808080, DMM_PAT_VIEW_MAP__0);
870 	dmm_write(omap_dmm, 0x80000000, DMM_PAT_VIEW_MAP_BASE);
871 	dmm_write(omap_dmm, 0x88888888, DMM_TILER_OR__0);
872 	dmm_write(omap_dmm, 0x88888888, DMM_TILER_OR__1);
873 
874 	omap_dmm->dummy_page = alloc_page(GFP_KERNEL | __GFP_DMA32);
875 	if (!omap_dmm->dummy_page) {
876 		dev_err(&dev->dev, "could not allocate dummy page\n");
877 		ret = -ENOMEM;
878 		goto fail;
879 	}
880 
881 	/* set dma mask for device */
882 	ret = dma_set_coherent_mask(&dev->dev, DMA_BIT_MASK(32));
883 	if (ret)
884 		goto fail;
885 
886 	omap_dmm->dummy_pa = page_to_phys(omap_dmm->dummy_page);
887 
888 	/* alloc refill memory */
889 	omap_dmm->refill_va = dma_alloc_wc(&dev->dev,
890 					   REFILL_BUFFER_SIZE * omap_dmm->num_engines,
891 					   &omap_dmm->refill_pa, GFP_KERNEL);
892 	if (!omap_dmm->refill_va) {
893 		dev_err(&dev->dev, "could not allocate refill memory\n");
894 		goto fail;
895 	}
896 
897 	/* alloc engines */
898 	omap_dmm->engines = kcalloc(omap_dmm->num_engines,
899 				    sizeof(*omap_dmm->engines), GFP_KERNEL);
900 	if (!omap_dmm->engines) {
901 		ret = -ENOMEM;
902 		goto fail;
903 	}
904 
905 	for (i = 0; i < omap_dmm->num_engines; i++) {
906 		omap_dmm->engines[i].id = i;
907 		omap_dmm->engines[i].dmm = omap_dmm;
908 		omap_dmm->engines[i].refill_va = omap_dmm->refill_va +
909 						(REFILL_BUFFER_SIZE * i);
910 		omap_dmm->engines[i].refill_pa = omap_dmm->refill_pa +
911 						(REFILL_BUFFER_SIZE * i);
912 		init_completion(&omap_dmm->engines[i].compl);
913 
914 		list_add(&omap_dmm->engines[i].idle_node, &omap_dmm->idle_head);
915 	}
916 
917 	omap_dmm->tcm = kcalloc(omap_dmm->num_lut, sizeof(*omap_dmm->tcm),
918 				GFP_KERNEL);
919 	if (!omap_dmm->tcm) {
920 		ret = -ENOMEM;
921 		goto fail;
922 	}
923 
924 	/* init containers */
925 	/* Each LUT is associated with a TCM (container manager).  We use the
926 	   lut_id to denote the lut_id used to identify the correct LUT for
927 	   programming during reill operations */
928 	for (i = 0; i < omap_dmm->num_lut; i++) {
929 		omap_dmm->tcm[i] = sita_init(omap_dmm->container_width,
930 						omap_dmm->container_height);
931 
932 		if (!omap_dmm->tcm[i]) {
933 			dev_err(&dev->dev, "failed to allocate container\n");
934 			ret = -ENOMEM;
935 			goto fail;
936 		}
937 
938 		omap_dmm->tcm[i]->lut_id = i;
939 	}
940 
941 	/* assign access mode containers to applicable tcm container */
942 	/* OMAP 4 has 1 container for all 4 views */
943 	/* OMAP 5 has 2 containers, 1 for 2D and 1 for 1D */
944 	containers[TILFMT_8BIT] = omap_dmm->tcm[0];
945 	containers[TILFMT_16BIT] = omap_dmm->tcm[0];
946 	containers[TILFMT_32BIT] = omap_dmm->tcm[0];
947 
948 	if (omap_dmm->container_height != omap_dmm->lut_height) {
949 		/* second LUT is used for PAGE mode.  Programming must use
950 		   y offset that is added to all y coordinates.  LUT id is still
951 		   0, because it is the same LUT, just the upper 128 lines */
952 		containers[TILFMT_PAGE] = omap_dmm->tcm[1];
953 		omap_dmm->tcm[1]->y_offset = OMAP5_LUT_OFFSET;
954 		omap_dmm->tcm[1]->lut_id = 0;
955 	} else {
956 		containers[TILFMT_PAGE] = omap_dmm->tcm[0];
957 	}
958 
959 	area = (struct tcm_area) {
960 		.tcm = NULL,
961 		.p1.x = omap_dmm->container_width - 1,
962 		.p1.y = omap_dmm->container_height - 1,
963 	};
964 
965 	ret = request_irq(omap_dmm->irq, omap_dmm_irq_handler, IRQF_SHARED,
966 				"omap_dmm_irq_handler", omap_dmm);
967 
968 	if (ret) {
969 		dev_err(&dev->dev, "couldn't register IRQ %d, error %d\n",
970 			omap_dmm->irq, ret);
971 		omap_dmm->irq = -1;
972 		goto fail;
973 	}
974 
975 	/* Enable all interrupts for each refill engine except
976 	 * ERR_LUT_MISS<n> (which is just advisory, and we don't care
977 	 * about because we want to be able to refill live scanout
978 	 * buffers for accelerated pan/scroll) and FILL_DSC<n> which
979 	 * we just generally don't care about.
980 	 */
981 	dmm_write(omap_dmm, 0x7e7e7e7e, DMM_PAT_IRQENABLE_SET);
982 
983 	/* initialize all LUTs to dummy page entries */
984 	for (i = 0; i < omap_dmm->num_lut; i++) {
985 		area.tcm = omap_dmm->tcm[i];
986 		if (fill(&area, NULL, 0, 0, true))
987 			dev_err(omap_dmm->dev, "refill failed");
988 	}
989 
990 	dev_info(omap_dmm->dev, "initialized all PAT entries\n");
991 
992 	return 0;
993 
994 fail:
995 	if (omap_dmm_remove(dev))
996 		dev_err(&dev->dev, "cleanup failed\n");
997 	return ret;
998 }
999 
1000 /*
1001  * debugfs support
1002  */
1003 
1004 #ifdef CONFIG_DEBUG_FS
1005 
1006 static const char *alphabet = "abcdefghijklmnopqrstuvwxyz"
1007 				"ABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789";
1008 static const char *special = ".,:;'\"`~!^-+";
1009 
1010 static void fill_map(char **map, int xdiv, int ydiv, struct tcm_area *a,
1011 							char c, bool ovw)
1012 {
1013 	int x, y;
1014 	for (y = a->p0.y / ydiv; y <= a->p1.y / ydiv; y++)
1015 		for (x = a->p0.x / xdiv; x <= a->p1.x / xdiv; x++)
1016 			if (map[y][x] == ' ' || ovw)
1017 				map[y][x] = c;
1018 }
1019 
1020 static void fill_map_pt(char **map, int xdiv, int ydiv, struct tcm_pt *p,
1021 									char c)
1022 {
1023 	map[p->y / ydiv][p->x / xdiv] = c;
1024 }
1025 
1026 static char read_map_pt(char **map, int xdiv, int ydiv, struct tcm_pt *p)
1027 {
1028 	return map[p->y / ydiv][p->x / xdiv];
1029 }
1030 
1031 static int map_width(int xdiv, int x0, int x1)
1032 {
1033 	return (x1 / xdiv) - (x0 / xdiv) + 1;
1034 }
1035 
1036 static void text_map(char **map, int xdiv, char *nice, int yd, int x0, int x1)
1037 {
1038 	char *p = map[yd] + (x0 / xdiv);
1039 	int w = (map_width(xdiv, x0, x1) - strlen(nice)) / 2;
1040 	if (w >= 0) {
1041 		p += w;
1042 		while (*nice)
1043 			*p++ = *nice++;
1044 	}
1045 }
1046 
1047 static void map_1d_info(char **map, int xdiv, int ydiv, char *nice,
1048 							struct tcm_area *a)
1049 {
1050 	sprintf(nice, "%dK", tcm_sizeof(*a) * 4);
1051 	if (a->p0.y + 1 < a->p1.y) {
1052 		text_map(map, xdiv, nice, (a->p0.y + a->p1.y) / 2 / ydiv, 0,
1053 							256 - 1);
1054 	} else if (a->p0.y < a->p1.y) {
1055 		if (strlen(nice) < map_width(xdiv, a->p0.x, 256 - 1))
1056 			text_map(map, xdiv, nice, a->p0.y / ydiv,
1057 					a->p0.x + xdiv,	256 - 1);
1058 		else if (strlen(nice) < map_width(xdiv, 0, a->p1.x))
1059 			text_map(map, xdiv, nice, a->p1.y / ydiv,
1060 					0, a->p1.y - xdiv);
1061 	} else if (strlen(nice) + 1 < map_width(xdiv, a->p0.x, a->p1.x)) {
1062 		text_map(map, xdiv, nice, a->p0.y / ydiv, a->p0.x, a->p1.x);
1063 	}
1064 }
1065 
1066 static void map_2d_info(char **map, int xdiv, int ydiv, char *nice,
1067 							struct tcm_area *a)
1068 {
1069 	sprintf(nice, "(%d*%d)", tcm_awidth(*a), tcm_aheight(*a));
1070 	if (strlen(nice) + 1 < map_width(xdiv, a->p0.x, a->p1.x))
1071 		text_map(map, xdiv, nice, (a->p0.y + a->p1.y) / 2 / ydiv,
1072 							a->p0.x, a->p1.x);
1073 }
1074 
1075 int tiler_map_show(struct seq_file *s, void *arg)
1076 {
1077 	int xdiv = 2, ydiv = 1;
1078 	char **map = NULL, *global_map;
1079 	struct tiler_block *block;
1080 	struct tcm_area a, p;
1081 	int i;
1082 	const char *m2d = alphabet;
1083 	const char *a2d = special;
1084 	const char *m2dp = m2d, *a2dp = a2d;
1085 	char nice[128];
1086 	int h_adj;
1087 	int w_adj;
1088 	unsigned long flags;
1089 	int lut_idx;
1090 
1091 
1092 	if (!omap_dmm) {
1093 		/* early return if dmm/tiler device is not initialized */
1094 		return 0;
1095 	}
1096 
1097 	h_adj = omap_dmm->container_height / ydiv;
1098 	w_adj = omap_dmm->container_width / xdiv;
1099 
1100 	map = kmalloc_array(h_adj, sizeof(*map), GFP_KERNEL);
1101 	global_map = kmalloc_array(w_adj + 1, h_adj, GFP_KERNEL);
1102 
1103 	if (!map || !global_map)
1104 		goto error;
1105 
1106 	for (lut_idx = 0; lut_idx < omap_dmm->num_lut; lut_idx++) {
1107 		memset(map, 0, h_adj * sizeof(*map));
1108 		memset(global_map, ' ', (w_adj + 1) * h_adj);
1109 
1110 		for (i = 0; i < omap_dmm->container_height; i++) {
1111 			map[i] = global_map + i * (w_adj + 1);
1112 			map[i][w_adj] = 0;
1113 		}
1114 
1115 		spin_lock_irqsave(&list_lock, flags);
1116 
1117 		list_for_each_entry(block, &omap_dmm->alloc_head, alloc_node) {
1118 			if (block->area.tcm == omap_dmm->tcm[lut_idx]) {
1119 				if (block->fmt != TILFMT_PAGE) {
1120 					fill_map(map, xdiv, ydiv, &block->area,
1121 						*m2dp, true);
1122 					if (!*++a2dp)
1123 						a2dp = a2d;
1124 					if (!*++m2dp)
1125 						m2dp = m2d;
1126 					map_2d_info(map, xdiv, ydiv, nice,
1127 							&block->area);
1128 				} else {
1129 					bool start = read_map_pt(map, xdiv,
1130 						ydiv, &block->area.p0) == ' ';
1131 					bool end = read_map_pt(map, xdiv, ydiv,
1132 							&block->area.p1) == ' ';
1133 
1134 					tcm_for_each_slice(a, block->area, p)
1135 						fill_map(map, xdiv, ydiv, &a,
1136 							'=', true);
1137 					fill_map_pt(map, xdiv, ydiv,
1138 							&block->area.p0,
1139 							start ? '<' : 'X');
1140 					fill_map_pt(map, xdiv, ydiv,
1141 							&block->area.p1,
1142 							end ? '>' : 'X');
1143 					map_1d_info(map, xdiv, ydiv, nice,
1144 							&block->area);
1145 				}
1146 			}
1147 		}
1148 
1149 		spin_unlock_irqrestore(&list_lock, flags);
1150 
1151 		if (s) {
1152 			seq_printf(s, "CONTAINER %d DUMP BEGIN\n", lut_idx);
1153 			for (i = 0; i < 128; i++)
1154 				seq_printf(s, "%03d:%s\n", i, map[i]);
1155 			seq_printf(s, "CONTAINER %d DUMP END\n", lut_idx);
1156 		} else {
1157 			dev_dbg(omap_dmm->dev, "CONTAINER %d DUMP BEGIN\n",
1158 				lut_idx);
1159 			for (i = 0; i < 128; i++)
1160 				dev_dbg(omap_dmm->dev, "%03d:%s\n", i, map[i]);
1161 			dev_dbg(omap_dmm->dev, "CONTAINER %d DUMP END\n",
1162 				lut_idx);
1163 		}
1164 	}
1165 
1166 error:
1167 	kfree(map);
1168 	kfree(global_map);
1169 
1170 	return 0;
1171 }
1172 #endif
1173 
1174 #ifdef CONFIG_PM_SLEEP
1175 static int omap_dmm_resume(struct device *dev)
1176 {
1177 	struct tcm_area area;
1178 	int i;
1179 
1180 	if (!omap_dmm)
1181 		return -ENODEV;
1182 
1183 	area = (struct tcm_area) {
1184 		.tcm = NULL,
1185 		.p1.x = omap_dmm->container_width - 1,
1186 		.p1.y = omap_dmm->container_height - 1,
1187 	};
1188 
1189 	/* initialize all LUTs to dummy page entries */
1190 	for (i = 0; i < omap_dmm->num_lut; i++) {
1191 		area.tcm = omap_dmm->tcm[i];
1192 		if (fill(&area, NULL, 0, 0, true))
1193 			dev_err(dev, "refill failed");
1194 	}
1195 
1196 	return 0;
1197 }
1198 #endif
1199 
1200 static SIMPLE_DEV_PM_OPS(omap_dmm_pm_ops, NULL, omap_dmm_resume);
1201 
1202 #if defined(CONFIG_OF)
1203 static const struct dmm_platform_data dmm_omap4_platform_data = {
1204 	.cpu_cache_flags = OMAP_BO_WC,
1205 };
1206 
1207 static const struct dmm_platform_data dmm_omap5_platform_data = {
1208 	.cpu_cache_flags = OMAP_BO_UNCACHED,
1209 };
1210 
1211 static const struct of_device_id dmm_of_match[] = {
1212 	{
1213 		.compatible = "ti,omap4-dmm",
1214 		.data = &dmm_omap4_platform_data,
1215 	},
1216 	{
1217 		.compatible = "ti,omap5-dmm",
1218 		.data = &dmm_omap5_platform_data,
1219 	},
1220 	{},
1221 };
1222 #endif
1223 
1224 struct platform_driver omap_dmm_driver = {
1225 	.probe = omap_dmm_probe,
1226 	.remove = omap_dmm_remove,
1227 	.driver = {
1228 		.owner = THIS_MODULE,
1229 		.name = DMM_DRIVER_NAME,
1230 		.of_match_table = of_match_ptr(dmm_of_match),
1231 		.pm = &omap_dmm_pm_ops,
1232 	},
1233 };
1234 
1235 MODULE_LICENSE("GPL v2");
1236 MODULE_AUTHOR("Andy Gross <andy.gross@ti.com>");
1237 MODULE_DESCRIPTION("OMAP DMM/Tiler Driver");
1238