1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * DMM IOMMU driver support functions for TI OMAP processors.
4  *
5  * Copyright (C) 2011 Texas Instruments Incorporated - https://www.ti.com/
6  * Author: Rob Clark <rob@ti.com>
7  *         Andy Gross <andy.gross@ti.com>
8  */
9 
10 #include <linux/completion.h>
11 #include <linux/delay.h>
12 #include <linux/dma-mapping.h>
13 #include <linux/dmaengine.h>
14 #include <linux/errno.h>
15 #include <linux/init.h>
16 #include <linux/interrupt.h>
17 #include <linux/list.h>
18 #include <linux/mm.h>
19 #include <linux/module.h>
20 #include <linux/of.h>
21 #include <linux/platform_device.h> /* platform_device() */
22 #include <linux/sched.h>
23 #include <linux/seq_file.h>
24 #include <linux/slab.h>
25 #include <linux/time.h>
26 #include <linux/vmalloc.h>
27 #include <linux/wait.h>
28 
29 #include "omap_dmm_tiler.h"
30 #include "omap_dmm_priv.h"
31 
32 #define DMM_DRIVER_NAME "dmm"
33 
34 /* mappings for associating views to luts */
35 static struct tcm *containers[TILFMT_NFORMATS];
36 static struct dmm *omap_dmm;
37 
38 #if defined(CONFIG_OF)
39 static const struct of_device_id dmm_of_match[];
40 #endif
41 
42 /* global spinlock for protecting lists */
43 static DEFINE_SPINLOCK(list_lock);
44 
45 /* Geometry table */
46 #define GEOM(xshift, yshift, bytes_per_pixel) { \
47 		.x_shft = (xshift), \
48 		.y_shft = (yshift), \
49 		.cpp    = (bytes_per_pixel), \
50 		.slot_w = 1 << (SLOT_WIDTH_BITS - (xshift)), \
51 		.slot_h = 1 << (SLOT_HEIGHT_BITS - (yshift)), \
52 	}
53 
54 static const struct {
55 	u32 x_shft;	/* unused X-bits (as part of bpp) */
56 	u32 y_shft;	/* unused Y-bits (as part of bpp) */
57 	u32 cpp;		/* bytes/chars per pixel */
58 	u32 slot_w;	/* width of each slot (in pixels) */
59 	u32 slot_h;	/* height of each slot (in pixels) */
60 } geom[TILFMT_NFORMATS] = {
61 	[TILFMT_8BIT]  = GEOM(0, 0, 1),
62 	[TILFMT_16BIT] = GEOM(0, 1, 2),
63 	[TILFMT_32BIT] = GEOM(1, 1, 4),
64 	[TILFMT_PAGE]  = GEOM(SLOT_WIDTH_BITS, SLOT_HEIGHT_BITS, 1),
65 };
66 
67 
68 /* lookup table for registers w/ per-engine instances */
69 static const u32 reg[][4] = {
70 	[PAT_STATUS] = {DMM_PAT_STATUS__0, DMM_PAT_STATUS__1,
71 			DMM_PAT_STATUS__2, DMM_PAT_STATUS__3},
72 	[PAT_DESCR]  = {DMM_PAT_DESCR__0, DMM_PAT_DESCR__1,
73 			DMM_PAT_DESCR__2, DMM_PAT_DESCR__3},
74 };
75 
76 static int dmm_dma_copy(struct dmm *dmm, dma_addr_t src, dma_addr_t dst)
77 {
78 	struct dma_async_tx_descriptor *tx;
79 	enum dma_status status;
80 	dma_cookie_t cookie;
81 
82 	tx = dmaengine_prep_dma_memcpy(dmm->wa_dma_chan, dst, src, 4, 0);
83 	if (!tx) {
84 		dev_err(dmm->dev, "Failed to prepare DMA memcpy\n");
85 		return -EIO;
86 	}
87 
88 	cookie = tx->tx_submit(tx);
89 	if (dma_submit_error(cookie)) {
90 		dev_err(dmm->dev, "Failed to do DMA tx_submit\n");
91 		return -EIO;
92 	}
93 
94 	status = dma_sync_wait(dmm->wa_dma_chan, cookie);
95 	if (status != DMA_COMPLETE)
96 		dev_err(dmm->dev, "i878 wa DMA copy failure\n");
97 
98 	dmaengine_terminate_all(dmm->wa_dma_chan);
99 	return 0;
100 }
101 
102 static u32 dmm_read_wa(struct dmm *dmm, u32 reg)
103 {
104 	dma_addr_t src, dst;
105 	int r;
106 
107 	src = dmm->phys_base + reg;
108 	dst = dmm->wa_dma_handle;
109 
110 	r = dmm_dma_copy(dmm, src, dst);
111 	if (r) {
112 		dev_err(dmm->dev, "sDMA read transfer timeout\n");
113 		return readl(dmm->base + reg);
114 	}
115 
116 	/*
117 	 * As per i878 workaround, the DMA is used to access the DMM registers.
118 	 * Make sure that the readl is not moved by the compiler or the CPU
119 	 * earlier than the DMA finished writing the value to memory.
120 	 */
121 	rmb();
122 	return readl(dmm->wa_dma_data);
123 }
124 
125 static void dmm_write_wa(struct dmm *dmm, u32 val, u32 reg)
126 {
127 	dma_addr_t src, dst;
128 	int r;
129 
130 	writel(val, dmm->wa_dma_data);
131 	/*
132 	 * As per i878 workaround, the DMA is used to access the DMM registers.
133 	 * Make sure that the writel is not moved by the compiler or the CPU, so
134 	 * the data will be in place before we start the DMA to do the actual
135 	 * register write.
136 	 */
137 	wmb();
138 
139 	src = dmm->wa_dma_handle;
140 	dst = dmm->phys_base + reg;
141 
142 	r = dmm_dma_copy(dmm, src, dst);
143 	if (r) {
144 		dev_err(dmm->dev, "sDMA write transfer timeout\n");
145 		writel(val, dmm->base + reg);
146 	}
147 }
148 
149 static u32 dmm_read(struct dmm *dmm, u32 reg)
150 {
151 	if (dmm->dmm_workaround) {
152 		u32 v;
153 		unsigned long flags;
154 
155 		spin_lock_irqsave(&dmm->wa_lock, flags);
156 		v = dmm_read_wa(dmm, reg);
157 		spin_unlock_irqrestore(&dmm->wa_lock, flags);
158 
159 		return v;
160 	} else {
161 		return readl(dmm->base + reg);
162 	}
163 }
164 
165 static void dmm_write(struct dmm *dmm, u32 val, u32 reg)
166 {
167 	if (dmm->dmm_workaround) {
168 		unsigned long flags;
169 
170 		spin_lock_irqsave(&dmm->wa_lock, flags);
171 		dmm_write_wa(dmm, val, reg);
172 		spin_unlock_irqrestore(&dmm->wa_lock, flags);
173 	} else {
174 		writel(val, dmm->base + reg);
175 	}
176 }
177 
178 static int dmm_workaround_init(struct dmm *dmm)
179 {
180 	dma_cap_mask_t mask;
181 
182 	spin_lock_init(&dmm->wa_lock);
183 
184 	dmm->wa_dma_data = dma_alloc_coherent(dmm->dev,  sizeof(u32),
185 					      &dmm->wa_dma_handle, GFP_KERNEL);
186 	if (!dmm->wa_dma_data)
187 		return -ENOMEM;
188 
189 	dma_cap_zero(mask);
190 	dma_cap_set(DMA_MEMCPY, mask);
191 
192 	dmm->wa_dma_chan = dma_request_channel(mask, NULL, NULL);
193 	if (!dmm->wa_dma_chan) {
194 		dma_free_coherent(dmm->dev, 4, dmm->wa_dma_data, dmm->wa_dma_handle);
195 		return -ENODEV;
196 	}
197 
198 	return 0;
199 }
200 
201 static void dmm_workaround_uninit(struct dmm *dmm)
202 {
203 	dma_release_channel(dmm->wa_dma_chan);
204 
205 	dma_free_coherent(dmm->dev, 4, dmm->wa_dma_data, dmm->wa_dma_handle);
206 }
207 
208 /* simple allocator to grab next 16 byte aligned memory from txn */
209 static void *alloc_dma(struct dmm_txn *txn, size_t sz, dma_addr_t *pa)
210 {
211 	void *ptr;
212 	struct refill_engine *engine = txn->engine_handle;
213 
214 	/* dmm programming requires 16 byte aligned addresses */
215 	txn->current_pa = round_up(txn->current_pa, 16);
216 	txn->current_va = (void *)round_up((long)txn->current_va, 16);
217 
218 	ptr = txn->current_va;
219 	*pa = txn->current_pa;
220 
221 	txn->current_pa += sz;
222 	txn->current_va += sz;
223 
224 	BUG_ON((txn->current_va - engine->refill_va) > REFILL_BUFFER_SIZE);
225 
226 	return ptr;
227 }
228 
229 /* check status and spin until wait_mask comes true */
230 static int wait_status(struct refill_engine *engine, u32 wait_mask)
231 {
232 	struct dmm *dmm = engine->dmm;
233 	u32 r = 0, err, i;
234 
235 	i = DMM_FIXED_RETRY_COUNT;
236 	while (true) {
237 		r = dmm_read(dmm, reg[PAT_STATUS][engine->id]);
238 		err = r & DMM_PATSTATUS_ERR;
239 		if (err) {
240 			dev_err(dmm->dev,
241 				"%s: error (engine%d). PAT_STATUS: 0x%08x\n",
242 				__func__, engine->id, r);
243 			return -EFAULT;
244 		}
245 
246 		if ((r & wait_mask) == wait_mask)
247 			break;
248 
249 		if (--i == 0) {
250 			dev_err(dmm->dev,
251 				"%s: timeout (engine%d). PAT_STATUS: 0x%08x\n",
252 				__func__, engine->id, r);
253 			return -ETIMEDOUT;
254 		}
255 
256 		udelay(1);
257 	}
258 
259 	return 0;
260 }
261 
262 static void release_engine(struct refill_engine *engine)
263 {
264 	unsigned long flags;
265 
266 	spin_lock_irqsave(&list_lock, flags);
267 	list_add(&engine->idle_node, &omap_dmm->idle_head);
268 	spin_unlock_irqrestore(&list_lock, flags);
269 
270 	atomic_inc(&omap_dmm->engine_counter);
271 	wake_up_interruptible(&omap_dmm->engine_queue);
272 }
273 
274 static irqreturn_t omap_dmm_irq_handler(int irq, void *arg)
275 {
276 	struct dmm *dmm = arg;
277 	u32 status = dmm_read(dmm, DMM_PAT_IRQSTATUS);
278 	int i;
279 
280 	/* ack IRQ */
281 	dmm_write(dmm, status, DMM_PAT_IRQSTATUS);
282 
283 	for (i = 0; i < dmm->num_engines; i++) {
284 		if (status & DMM_IRQSTAT_ERR_MASK)
285 			dev_err(dmm->dev,
286 				"irq error(engine%d): IRQSTAT 0x%02x\n",
287 				i, status & 0xff);
288 
289 		if (status & DMM_IRQSTAT_LST) {
290 			if (dmm->engines[i].async)
291 				release_engine(&dmm->engines[i]);
292 
293 			complete(&dmm->engines[i].compl);
294 		}
295 
296 		status >>= 8;
297 	}
298 
299 	return IRQ_HANDLED;
300 }
301 
302 /*
303  * Get a handle for a DMM transaction
304  */
305 static struct dmm_txn *dmm_txn_init(struct dmm *dmm, struct tcm *tcm)
306 {
307 	struct dmm_txn *txn = NULL;
308 	struct refill_engine *engine = NULL;
309 	int ret;
310 	unsigned long flags;
311 
312 
313 	/* wait until an engine is available */
314 	ret = wait_event_interruptible(omap_dmm->engine_queue,
315 		atomic_add_unless(&omap_dmm->engine_counter, -1, 0));
316 	if (ret)
317 		return ERR_PTR(ret);
318 
319 	/* grab an idle engine */
320 	spin_lock_irqsave(&list_lock, flags);
321 	if (!list_empty(&dmm->idle_head)) {
322 		engine = list_entry(dmm->idle_head.next, struct refill_engine,
323 					idle_node);
324 		list_del(&engine->idle_node);
325 	}
326 	spin_unlock_irqrestore(&list_lock, flags);
327 
328 	BUG_ON(!engine);
329 
330 	txn = &engine->txn;
331 	engine->tcm = tcm;
332 	txn->engine_handle = engine;
333 	txn->last_pat = NULL;
334 	txn->current_va = engine->refill_va;
335 	txn->current_pa = engine->refill_pa;
336 
337 	return txn;
338 }
339 
340 /*
341  * Add region to DMM transaction.  If pages or pages[i] is NULL, then the
342  * corresponding slot is cleared (ie. dummy_pa is programmed)
343  */
344 static void dmm_txn_append(struct dmm_txn *txn, struct pat_area *area,
345 		struct page **pages, u32 npages, u32 roll)
346 {
347 	dma_addr_t pat_pa = 0, data_pa = 0;
348 	u32 *data;
349 	struct pat *pat;
350 	struct refill_engine *engine = txn->engine_handle;
351 	int columns = (1 + area->x1 - area->x0);
352 	int rows = (1 + area->y1 - area->y0);
353 	int i = columns*rows;
354 
355 	pat = alloc_dma(txn, sizeof(*pat), &pat_pa);
356 
357 	if (txn->last_pat)
358 		txn->last_pat->next_pa = (u32)pat_pa;
359 
360 	pat->area = *area;
361 
362 	/* adjust Y coordinates based off of container parameters */
363 	pat->area.y0 += engine->tcm->y_offset;
364 	pat->area.y1 += engine->tcm->y_offset;
365 
366 	pat->ctrl = (struct pat_ctrl){
367 			.start = 1,
368 			.lut_id = engine->tcm->lut_id,
369 		};
370 
371 	data = alloc_dma(txn, 4*i, &data_pa);
372 	/* FIXME: what if data_pa is more than 32-bit ? */
373 	pat->data_pa = data_pa;
374 
375 	while (i--) {
376 		int n = i + roll;
377 		if (n >= npages)
378 			n -= npages;
379 		data[i] = (pages && pages[n]) ?
380 			page_to_phys(pages[n]) : engine->dmm->dummy_pa;
381 	}
382 
383 	txn->last_pat = pat;
384 
385 	return;
386 }
387 
388 /*
389  * Commit the DMM transaction.
390  */
391 static int dmm_txn_commit(struct dmm_txn *txn, bool wait)
392 {
393 	int ret = 0;
394 	struct refill_engine *engine = txn->engine_handle;
395 	struct dmm *dmm = engine->dmm;
396 
397 	if (!txn->last_pat) {
398 		dev_err(engine->dmm->dev, "need at least one txn\n");
399 		ret = -EINVAL;
400 		goto cleanup;
401 	}
402 
403 	txn->last_pat->next_pa = 0;
404 	/* ensure that the written descriptors are visible to DMM */
405 	wmb();
406 
407 	/*
408 	 * NOTE: the wmb() above should be enough, but there seems to be a bug
409 	 * in OMAP's memory barrier implementation, which in some rare cases may
410 	 * cause the writes not to be observable after wmb().
411 	 */
412 
413 	/* read back to ensure the data is in RAM */
414 	readl(&txn->last_pat->next_pa);
415 
416 	/* write to PAT_DESCR to clear out any pending transaction */
417 	dmm_write(dmm, 0x0, reg[PAT_DESCR][engine->id]);
418 
419 	/* wait for engine ready: */
420 	ret = wait_status(engine, DMM_PATSTATUS_READY);
421 	if (ret) {
422 		ret = -EFAULT;
423 		goto cleanup;
424 	}
425 
426 	/* mark whether it is async to denote list management in IRQ handler */
427 	engine->async = wait ? false : true;
428 	reinit_completion(&engine->compl);
429 	/* verify that the irq handler sees the 'async' and completion value */
430 	smp_mb();
431 
432 	/* kick reload */
433 	dmm_write(dmm, engine->refill_pa, reg[PAT_DESCR][engine->id]);
434 
435 	if (wait) {
436 		if (!wait_for_completion_timeout(&engine->compl,
437 				msecs_to_jiffies(100))) {
438 			dev_err(dmm->dev, "timed out waiting for done\n");
439 			ret = -ETIMEDOUT;
440 			goto cleanup;
441 		}
442 
443 		/* Check the engine status before continue */
444 		ret = wait_status(engine, DMM_PATSTATUS_READY |
445 				  DMM_PATSTATUS_VALID | DMM_PATSTATUS_DONE);
446 	}
447 
448 cleanup:
449 	/* only place engine back on list if we are done with it */
450 	if (ret || wait)
451 		release_engine(engine);
452 
453 	return ret;
454 }
455 
456 /*
457  * DMM programming
458  */
459 static int fill(struct tcm_area *area, struct page **pages,
460 		u32 npages, u32 roll, bool wait)
461 {
462 	int ret = 0;
463 	struct tcm_area slice, area_s;
464 	struct dmm_txn *txn;
465 
466 	/*
467 	 * FIXME
468 	 *
469 	 * Asynchronous fill does not work reliably, as the driver does not
470 	 * handle errors in the async code paths. The fill operation may
471 	 * silently fail, leading to leaking DMM engines, which may eventually
472 	 * lead to deadlock if we run out of DMM engines.
473 	 *
474 	 * For now, always set 'wait' so that we only use sync fills. Async
475 	 * fills should be fixed, or alternatively we could decide to only
476 	 * support sync fills and so the whole async code path could be removed.
477 	 */
478 
479 	wait = true;
480 
481 	txn = dmm_txn_init(omap_dmm, area->tcm);
482 	if (IS_ERR_OR_NULL(txn))
483 		return -ENOMEM;
484 
485 	tcm_for_each_slice(slice, *area, area_s) {
486 		struct pat_area p_area = {
487 				.x0 = slice.p0.x,  .y0 = slice.p0.y,
488 				.x1 = slice.p1.x,  .y1 = slice.p1.y,
489 		};
490 
491 		dmm_txn_append(txn, &p_area, pages, npages, roll);
492 
493 		roll += tcm_sizeof(slice);
494 	}
495 
496 	ret = dmm_txn_commit(txn, wait);
497 
498 	return ret;
499 }
500 
501 /*
502  * Pin/unpin
503  */
504 
505 /* note: slots for which pages[i] == NULL are filled w/ dummy page
506  */
507 int tiler_pin(struct tiler_block *block, struct page **pages,
508 		u32 npages, u32 roll, bool wait)
509 {
510 	int ret;
511 
512 	ret = fill(&block->area, pages, npages, roll, wait);
513 
514 	if (ret)
515 		tiler_unpin(block);
516 
517 	return ret;
518 }
519 
520 int tiler_unpin(struct tiler_block *block)
521 {
522 	return fill(&block->area, NULL, 0, 0, false);
523 }
524 
525 /*
526  * Reserve/release
527  */
528 struct tiler_block *tiler_reserve_2d(enum tiler_fmt fmt, u16 w,
529 		u16 h, u16 align)
530 {
531 	struct tiler_block *block;
532 	u32 min_align = 128;
533 	int ret;
534 	unsigned long flags;
535 	u32 slot_bytes;
536 
537 	block = kzalloc(sizeof(*block), GFP_KERNEL);
538 	if (!block)
539 		return ERR_PTR(-ENOMEM);
540 
541 	BUG_ON(!validfmt(fmt));
542 
543 	/* convert width/height to slots */
544 	w = DIV_ROUND_UP(w, geom[fmt].slot_w);
545 	h = DIV_ROUND_UP(h, geom[fmt].slot_h);
546 
547 	/* convert alignment to slots */
548 	slot_bytes = geom[fmt].slot_w * geom[fmt].cpp;
549 	min_align = max(min_align, slot_bytes);
550 	align = (align > min_align) ? ALIGN(align, min_align) : min_align;
551 	align /= slot_bytes;
552 
553 	block->fmt = fmt;
554 
555 	ret = tcm_reserve_2d(containers[fmt], w, h, align, -1, slot_bytes,
556 			&block->area);
557 	if (ret) {
558 		kfree(block);
559 		return ERR_PTR(-ENOMEM);
560 	}
561 
562 	/* add to allocation list */
563 	spin_lock_irqsave(&list_lock, flags);
564 	list_add(&block->alloc_node, &omap_dmm->alloc_head);
565 	spin_unlock_irqrestore(&list_lock, flags);
566 
567 	return block;
568 }
569 
570 struct tiler_block *tiler_reserve_1d(size_t size)
571 {
572 	struct tiler_block *block = kzalloc(sizeof(*block), GFP_KERNEL);
573 	int num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
574 	unsigned long flags;
575 
576 	if (!block)
577 		return ERR_PTR(-ENOMEM);
578 
579 	block->fmt = TILFMT_PAGE;
580 
581 	if (tcm_reserve_1d(containers[TILFMT_PAGE], num_pages,
582 				&block->area)) {
583 		kfree(block);
584 		return ERR_PTR(-ENOMEM);
585 	}
586 
587 	spin_lock_irqsave(&list_lock, flags);
588 	list_add(&block->alloc_node, &omap_dmm->alloc_head);
589 	spin_unlock_irqrestore(&list_lock, flags);
590 
591 	return block;
592 }
593 
594 /* note: if you have pin'd pages, you should have already unpin'd first! */
595 int tiler_release(struct tiler_block *block)
596 {
597 	int ret = tcm_free(&block->area);
598 	unsigned long flags;
599 
600 	if (block->area.tcm)
601 		dev_err(omap_dmm->dev, "failed to release block\n");
602 
603 	spin_lock_irqsave(&list_lock, flags);
604 	list_del(&block->alloc_node);
605 	spin_unlock_irqrestore(&list_lock, flags);
606 
607 	kfree(block);
608 	return ret;
609 }
610 
611 /*
612  * Utils
613  */
614 
615 /* calculate the tiler space address of a pixel in a view orientation...
616  * below description copied from the display subsystem section of TRM:
617  *
618  * When the TILER is addressed, the bits:
619  *   [28:27] = 0x0 for 8-bit tiled
620  *             0x1 for 16-bit tiled
621  *             0x2 for 32-bit tiled
622  *             0x3 for page mode
623  *   [31:29] = 0x0 for 0-degree view
624  *             0x1 for 180-degree view + mirroring
625  *             0x2 for 0-degree view + mirroring
626  *             0x3 for 180-degree view
627  *             0x4 for 270-degree view + mirroring
628  *             0x5 for 270-degree view
629  *             0x6 for 90-degree view
630  *             0x7 for 90-degree view + mirroring
631  * Otherwise the bits indicated the corresponding bit address to access
632  * the SDRAM.
633  */
634 static u32 tiler_get_address(enum tiler_fmt fmt, u32 orient, u32 x, u32 y)
635 {
636 	u32 x_bits, y_bits, tmp, x_mask, y_mask, alignment;
637 
638 	x_bits = CONT_WIDTH_BITS - geom[fmt].x_shft;
639 	y_bits = CONT_HEIGHT_BITS - geom[fmt].y_shft;
640 	alignment = geom[fmt].x_shft + geom[fmt].y_shft;
641 
642 	/* validate coordinate */
643 	x_mask = MASK(x_bits);
644 	y_mask = MASK(y_bits);
645 
646 	if (x < 0 || x > x_mask || y < 0 || y > y_mask) {
647 		DBG("invalid coords: %u < 0 || %u > %u || %u < 0 || %u > %u",
648 				x, x, x_mask, y, y, y_mask);
649 		return 0;
650 	}
651 
652 	/* account for mirroring */
653 	if (orient & MASK_X_INVERT)
654 		x ^= x_mask;
655 	if (orient & MASK_Y_INVERT)
656 		y ^= y_mask;
657 
658 	/* get coordinate address */
659 	if (orient & MASK_XY_FLIP)
660 		tmp = ((x << y_bits) + y);
661 	else
662 		tmp = ((y << x_bits) + x);
663 
664 	return TIL_ADDR((tmp << alignment), orient, fmt);
665 }
666 
667 dma_addr_t tiler_ssptr(struct tiler_block *block)
668 {
669 	BUG_ON(!validfmt(block->fmt));
670 
671 	return TILVIEW_8BIT + tiler_get_address(block->fmt, 0,
672 			block->area.p0.x * geom[block->fmt].slot_w,
673 			block->area.p0.y * geom[block->fmt].slot_h);
674 }
675 
676 dma_addr_t tiler_tsptr(struct tiler_block *block, u32 orient,
677 		u32 x, u32 y)
678 {
679 	struct tcm_pt *p = &block->area.p0;
680 	BUG_ON(!validfmt(block->fmt));
681 
682 	return tiler_get_address(block->fmt, orient,
683 			(p->x * geom[block->fmt].slot_w) + x,
684 			(p->y * geom[block->fmt].slot_h) + y);
685 }
686 
687 void tiler_align(enum tiler_fmt fmt, u16 *w, u16 *h)
688 {
689 	BUG_ON(!validfmt(fmt));
690 	*w = round_up(*w, geom[fmt].slot_w);
691 	*h = round_up(*h, geom[fmt].slot_h);
692 }
693 
694 u32 tiler_stride(enum tiler_fmt fmt, u32 orient)
695 {
696 	BUG_ON(!validfmt(fmt));
697 
698 	if (orient & MASK_XY_FLIP)
699 		return 1 << (CONT_HEIGHT_BITS + geom[fmt].x_shft);
700 	else
701 		return 1 << (CONT_WIDTH_BITS + geom[fmt].y_shft);
702 }
703 
704 size_t tiler_size(enum tiler_fmt fmt, u16 w, u16 h)
705 {
706 	tiler_align(fmt, &w, &h);
707 	return geom[fmt].cpp * w * h;
708 }
709 
710 size_t tiler_vsize(enum tiler_fmt fmt, u16 w, u16 h)
711 {
712 	BUG_ON(!validfmt(fmt));
713 	return round_up(geom[fmt].cpp * w, PAGE_SIZE) * h;
714 }
715 
716 u32 tiler_get_cpu_cache_flags(void)
717 {
718 	return omap_dmm->plat_data->cpu_cache_flags;
719 }
720 
721 bool dmm_is_available(void)
722 {
723 	return omap_dmm ? true : false;
724 }
725 
726 static int omap_dmm_remove(struct platform_device *dev)
727 {
728 	struct tiler_block *block, *_block;
729 	int i;
730 	unsigned long flags;
731 
732 	if (omap_dmm) {
733 		/* Disable all enabled interrupts */
734 		dmm_write(omap_dmm, 0x7e7e7e7e, DMM_PAT_IRQENABLE_CLR);
735 		free_irq(omap_dmm->irq, omap_dmm);
736 
737 		/* free all area regions */
738 		spin_lock_irqsave(&list_lock, flags);
739 		list_for_each_entry_safe(block, _block, &omap_dmm->alloc_head,
740 					alloc_node) {
741 			list_del(&block->alloc_node);
742 			kfree(block);
743 		}
744 		spin_unlock_irqrestore(&list_lock, flags);
745 
746 		for (i = 0; i < omap_dmm->num_lut; i++)
747 			if (omap_dmm->tcm && omap_dmm->tcm[i])
748 				omap_dmm->tcm[i]->deinit(omap_dmm->tcm[i]);
749 		kfree(omap_dmm->tcm);
750 
751 		kfree(omap_dmm->engines);
752 		if (omap_dmm->refill_va)
753 			dma_free_wc(omap_dmm->dev,
754 				    REFILL_BUFFER_SIZE * omap_dmm->num_engines,
755 				    omap_dmm->refill_va, omap_dmm->refill_pa);
756 		if (omap_dmm->dummy_page)
757 			__free_page(omap_dmm->dummy_page);
758 
759 		if (omap_dmm->dmm_workaround)
760 			dmm_workaround_uninit(omap_dmm);
761 
762 		iounmap(omap_dmm->base);
763 		kfree(omap_dmm);
764 		omap_dmm = NULL;
765 	}
766 
767 	return 0;
768 }
769 
770 static int omap_dmm_probe(struct platform_device *dev)
771 {
772 	int ret = -EFAULT, i;
773 	struct tcm_area area = {0};
774 	u32 hwinfo, pat_geom;
775 	struct resource *mem;
776 
777 	omap_dmm = kzalloc(sizeof(*omap_dmm), GFP_KERNEL);
778 	if (!omap_dmm)
779 		goto fail;
780 
781 	/* initialize lists */
782 	INIT_LIST_HEAD(&omap_dmm->alloc_head);
783 	INIT_LIST_HEAD(&omap_dmm->idle_head);
784 
785 	init_waitqueue_head(&omap_dmm->engine_queue);
786 
787 	if (dev->dev.of_node) {
788 		const struct of_device_id *match;
789 
790 		match = of_match_node(dmm_of_match, dev->dev.of_node);
791 		if (!match) {
792 			dev_err(&dev->dev, "failed to find matching device node\n");
793 			ret = -ENODEV;
794 			goto fail;
795 		}
796 
797 		omap_dmm->plat_data = match->data;
798 	}
799 
800 	/* lookup hwmod data - base address and irq */
801 	mem = platform_get_resource(dev, IORESOURCE_MEM, 0);
802 	if (!mem) {
803 		dev_err(&dev->dev, "failed to get base address resource\n");
804 		goto fail;
805 	}
806 
807 	omap_dmm->phys_base = mem->start;
808 	omap_dmm->base = ioremap(mem->start, SZ_2K);
809 
810 	if (!omap_dmm->base) {
811 		dev_err(&dev->dev, "failed to get dmm base address\n");
812 		goto fail;
813 	}
814 
815 	omap_dmm->irq = platform_get_irq(dev, 0);
816 	if (omap_dmm->irq < 0) {
817 		dev_err(&dev->dev, "failed to get IRQ resource\n");
818 		goto fail;
819 	}
820 
821 	omap_dmm->dev = &dev->dev;
822 
823 	if (of_machine_is_compatible("ti,dra7")) {
824 		/*
825 		 * DRA7 Errata i878 says that MPU should not be used to access
826 		 * RAM and DMM at the same time. As it's not possible to prevent
827 		 * MPU accessing RAM, we need to access DMM via a proxy.
828 		 */
829 		if (!dmm_workaround_init(omap_dmm)) {
830 			omap_dmm->dmm_workaround = true;
831 			dev_info(&dev->dev,
832 				"workaround for errata i878 in use\n");
833 		} else {
834 			dev_warn(&dev->dev,
835 				 "failed to initialize work-around for i878\n");
836 		}
837 	}
838 
839 	hwinfo = dmm_read(omap_dmm, DMM_PAT_HWINFO);
840 	omap_dmm->num_engines = (hwinfo >> 24) & 0x1F;
841 	omap_dmm->num_lut = (hwinfo >> 16) & 0x1F;
842 	omap_dmm->container_width = 256;
843 	omap_dmm->container_height = 128;
844 
845 	atomic_set(&omap_dmm->engine_counter, omap_dmm->num_engines);
846 
847 	/* read out actual LUT width and height */
848 	pat_geom = dmm_read(omap_dmm, DMM_PAT_GEOMETRY);
849 	omap_dmm->lut_width = ((pat_geom >> 16) & 0xF) << 5;
850 	omap_dmm->lut_height = ((pat_geom >> 24) & 0xF) << 5;
851 
852 	/* increment LUT by one if on OMAP5 */
853 	/* LUT has twice the height, and is split into a separate container */
854 	if (omap_dmm->lut_height != omap_dmm->container_height)
855 		omap_dmm->num_lut++;
856 
857 	/* initialize DMM registers */
858 	dmm_write(omap_dmm, 0x88888888, DMM_PAT_VIEW__0);
859 	dmm_write(omap_dmm, 0x88888888, DMM_PAT_VIEW__1);
860 	dmm_write(omap_dmm, 0x80808080, DMM_PAT_VIEW_MAP__0);
861 	dmm_write(omap_dmm, 0x80000000, DMM_PAT_VIEW_MAP_BASE);
862 	dmm_write(omap_dmm, 0x88888888, DMM_TILER_OR__0);
863 	dmm_write(omap_dmm, 0x88888888, DMM_TILER_OR__1);
864 
865 	omap_dmm->dummy_page = alloc_page(GFP_KERNEL | __GFP_DMA32);
866 	if (!omap_dmm->dummy_page) {
867 		dev_err(&dev->dev, "could not allocate dummy page\n");
868 		ret = -ENOMEM;
869 		goto fail;
870 	}
871 
872 	/* set dma mask for device */
873 	ret = dma_set_coherent_mask(&dev->dev, DMA_BIT_MASK(32));
874 	if (ret)
875 		goto fail;
876 
877 	omap_dmm->dummy_pa = page_to_phys(omap_dmm->dummy_page);
878 
879 	/* alloc refill memory */
880 	omap_dmm->refill_va = dma_alloc_wc(&dev->dev,
881 					   REFILL_BUFFER_SIZE * omap_dmm->num_engines,
882 					   &omap_dmm->refill_pa, GFP_KERNEL);
883 	if (!omap_dmm->refill_va) {
884 		dev_err(&dev->dev, "could not allocate refill memory\n");
885 		ret = -ENOMEM;
886 		goto fail;
887 	}
888 
889 	/* alloc engines */
890 	omap_dmm->engines = kcalloc(omap_dmm->num_engines,
891 				    sizeof(*omap_dmm->engines), GFP_KERNEL);
892 	if (!omap_dmm->engines) {
893 		ret = -ENOMEM;
894 		goto fail;
895 	}
896 
897 	for (i = 0; i < omap_dmm->num_engines; i++) {
898 		omap_dmm->engines[i].id = i;
899 		omap_dmm->engines[i].dmm = omap_dmm;
900 		omap_dmm->engines[i].refill_va = omap_dmm->refill_va +
901 						(REFILL_BUFFER_SIZE * i);
902 		omap_dmm->engines[i].refill_pa = omap_dmm->refill_pa +
903 						(REFILL_BUFFER_SIZE * i);
904 		init_completion(&omap_dmm->engines[i].compl);
905 
906 		list_add(&omap_dmm->engines[i].idle_node, &omap_dmm->idle_head);
907 	}
908 
909 	omap_dmm->tcm = kcalloc(omap_dmm->num_lut, sizeof(*omap_dmm->tcm),
910 				GFP_KERNEL);
911 	if (!omap_dmm->tcm) {
912 		ret = -ENOMEM;
913 		goto fail;
914 	}
915 
916 	/* init containers */
917 	/* Each LUT is associated with a TCM (container manager).  We use the
918 	   lut_id to denote the lut_id used to identify the correct LUT for
919 	   programming during reill operations */
920 	for (i = 0; i < omap_dmm->num_lut; i++) {
921 		omap_dmm->tcm[i] = sita_init(omap_dmm->container_width,
922 						omap_dmm->container_height);
923 
924 		if (!omap_dmm->tcm[i]) {
925 			dev_err(&dev->dev, "failed to allocate container\n");
926 			ret = -ENOMEM;
927 			goto fail;
928 		}
929 
930 		omap_dmm->tcm[i]->lut_id = i;
931 	}
932 
933 	/* assign access mode containers to applicable tcm container */
934 	/* OMAP 4 has 1 container for all 4 views */
935 	/* OMAP 5 has 2 containers, 1 for 2D and 1 for 1D */
936 	containers[TILFMT_8BIT] = omap_dmm->tcm[0];
937 	containers[TILFMT_16BIT] = omap_dmm->tcm[0];
938 	containers[TILFMT_32BIT] = omap_dmm->tcm[0];
939 
940 	if (omap_dmm->container_height != omap_dmm->lut_height) {
941 		/* second LUT is used for PAGE mode.  Programming must use
942 		   y offset that is added to all y coordinates.  LUT id is still
943 		   0, because it is the same LUT, just the upper 128 lines */
944 		containers[TILFMT_PAGE] = omap_dmm->tcm[1];
945 		omap_dmm->tcm[1]->y_offset = OMAP5_LUT_OFFSET;
946 		omap_dmm->tcm[1]->lut_id = 0;
947 	} else {
948 		containers[TILFMT_PAGE] = omap_dmm->tcm[0];
949 	}
950 
951 	area = (struct tcm_area) {
952 		.tcm = NULL,
953 		.p1.x = omap_dmm->container_width - 1,
954 		.p1.y = omap_dmm->container_height - 1,
955 	};
956 
957 	ret = request_irq(omap_dmm->irq, omap_dmm_irq_handler, IRQF_SHARED,
958 				"omap_dmm_irq_handler", omap_dmm);
959 
960 	if (ret) {
961 		dev_err(&dev->dev, "couldn't register IRQ %d, error %d\n",
962 			omap_dmm->irq, ret);
963 		omap_dmm->irq = -1;
964 		goto fail;
965 	}
966 
967 	/* Enable all interrupts for each refill engine except
968 	 * ERR_LUT_MISS<n> (which is just advisory, and we don't care
969 	 * about because we want to be able to refill live scanout
970 	 * buffers for accelerated pan/scroll) and FILL_DSC<n> which
971 	 * we just generally don't care about.
972 	 */
973 	dmm_write(omap_dmm, 0x7e7e7e7e, DMM_PAT_IRQENABLE_SET);
974 
975 	/* initialize all LUTs to dummy page entries */
976 	for (i = 0; i < omap_dmm->num_lut; i++) {
977 		area.tcm = omap_dmm->tcm[i];
978 		if (fill(&area, NULL, 0, 0, true))
979 			dev_err(omap_dmm->dev, "refill failed");
980 	}
981 
982 	dev_info(omap_dmm->dev, "initialized all PAT entries\n");
983 
984 	return 0;
985 
986 fail:
987 	if (omap_dmm_remove(dev))
988 		dev_err(&dev->dev, "cleanup failed\n");
989 	return ret;
990 }
991 
992 /*
993  * debugfs support
994  */
995 
996 #ifdef CONFIG_DEBUG_FS
997 
998 static const char *alphabet = "abcdefghijklmnopqrstuvwxyz"
999 				"ABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789";
1000 static const char *special = ".,:;'\"`~!^-+";
1001 
1002 static void fill_map(char **map, int xdiv, int ydiv, struct tcm_area *a,
1003 							char c, bool ovw)
1004 {
1005 	int x, y;
1006 	for (y = a->p0.y / ydiv; y <= a->p1.y / ydiv; y++)
1007 		for (x = a->p0.x / xdiv; x <= a->p1.x / xdiv; x++)
1008 			if (map[y][x] == ' ' || ovw)
1009 				map[y][x] = c;
1010 }
1011 
1012 static void fill_map_pt(char **map, int xdiv, int ydiv, struct tcm_pt *p,
1013 									char c)
1014 {
1015 	map[p->y / ydiv][p->x / xdiv] = c;
1016 }
1017 
1018 static char read_map_pt(char **map, int xdiv, int ydiv, struct tcm_pt *p)
1019 {
1020 	return map[p->y / ydiv][p->x / xdiv];
1021 }
1022 
1023 static int map_width(int xdiv, int x0, int x1)
1024 {
1025 	return (x1 / xdiv) - (x0 / xdiv) + 1;
1026 }
1027 
1028 static void text_map(char **map, int xdiv, char *nice, int yd, int x0, int x1)
1029 {
1030 	char *p = map[yd] + (x0 / xdiv);
1031 	int w = (map_width(xdiv, x0, x1) - strlen(nice)) / 2;
1032 	if (w >= 0) {
1033 		p += w;
1034 		while (*nice)
1035 			*p++ = *nice++;
1036 	}
1037 }
1038 
1039 static void map_1d_info(char **map, int xdiv, int ydiv, char *nice,
1040 							struct tcm_area *a)
1041 {
1042 	sprintf(nice, "%dK", tcm_sizeof(*a) * 4);
1043 	if (a->p0.y + 1 < a->p1.y) {
1044 		text_map(map, xdiv, nice, (a->p0.y + a->p1.y) / 2 / ydiv, 0,
1045 							256 - 1);
1046 	} else if (a->p0.y < a->p1.y) {
1047 		if (strlen(nice) < map_width(xdiv, a->p0.x, 256 - 1))
1048 			text_map(map, xdiv, nice, a->p0.y / ydiv,
1049 					a->p0.x + xdiv,	256 - 1);
1050 		else if (strlen(nice) < map_width(xdiv, 0, a->p1.x))
1051 			text_map(map, xdiv, nice, a->p1.y / ydiv,
1052 					0, a->p1.y - xdiv);
1053 	} else if (strlen(nice) + 1 < map_width(xdiv, a->p0.x, a->p1.x)) {
1054 		text_map(map, xdiv, nice, a->p0.y / ydiv, a->p0.x, a->p1.x);
1055 	}
1056 }
1057 
1058 static void map_2d_info(char **map, int xdiv, int ydiv, char *nice,
1059 							struct tcm_area *a)
1060 {
1061 	sprintf(nice, "(%d*%d)", tcm_awidth(*a), tcm_aheight(*a));
1062 	if (strlen(nice) + 1 < map_width(xdiv, a->p0.x, a->p1.x))
1063 		text_map(map, xdiv, nice, (a->p0.y + a->p1.y) / 2 / ydiv,
1064 							a->p0.x, a->p1.x);
1065 }
1066 
1067 int tiler_map_show(struct seq_file *s, void *arg)
1068 {
1069 	int xdiv = 2, ydiv = 1;
1070 	char **map = NULL, *global_map;
1071 	struct tiler_block *block;
1072 	struct tcm_area a, p;
1073 	int i;
1074 	const char *m2d = alphabet;
1075 	const char *a2d = special;
1076 	const char *m2dp = m2d, *a2dp = a2d;
1077 	char nice[128];
1078 	int h_adj;
1079 	int w_adj;
1080 	unsigned long flags;
1081 	int lut_idx;
1082 
1083 
1084 	if (!omap_dmm) {
1085 		/* early return if dmm/tiler device is not initialized */
1086 		return 0;
1087 	}
1088 
1089 	h_adj = omap_dmm->container_height / ydiv;
1090 	w_adj = omap_dmm->container_width / xdiv;
1091 
1092 	map = kmalloc_array(h_adj, sizeof(*map), GFP_KERNEL);
1093 	global_map = kmalloc_array(w_adj + 1, h_adj, GFP_KERNEL);
1094 
1095 	if (!map || !global_map)
1096 		goto error;
1097 
1098 	for (lut_idx = 0; lut_idx < omap_dmm->num_lut; lut_idx++) {
1099 		memset(map, 0, h_adj * sizeof(*map));
1100 		memset(global_map, ' ', (w_adj + 1) * h_adj);
1101 
1102 		for (i = 0; i < omap_dmm->container_height; i++) {
1103 			map[i] = global_map + i * (w_adj + 1);
1104 			map[i][w_adj] = 0;
1105 		}
1106 
1107 		spin_lock_irqsave(&list_lock, flags);
1108 
1109 		list_for_each_entry(block, &omap_dmm->alloc_head, alloc_node) {
1110 			if (block->area.tcm == omap_dmm->tcm[lut_idx]) {
1111 				if (block->fmt != TILFMT_PAGE) {
1112 					fill_map(map, xdiv, ydiv, &block->area,
1113 						*m2dp, true);
1114 					if (!*++a2dp)
1115 						a2dp = a2d;
1116 					if (!*++m2dp)
1117 						m2dp = m2d;
1118 					map_2d_info(map, xdiv, ydiv, nice,
1119 							&block->area);
1120 				} else {
1121 					bool start = read_map_pt(map, xdiv,
1122 						ydiv, &block->area.p0) == ' ';
1123 					bool end = read_map_pt(map, xdiv, ydiv,
1124 							&block->area.p1) == ' ';
1125 
1126 					tcm_for_each_slice(a, block->area, p)
1127 						fill_map(map, xdiv, ydiv, &a,
1128 							'=', true);
1129 					fill_map_pt(map, xdiv, ydiv,
1130 							&block->area.p0,
1131 							start ? '<' : 'X');
1132 					fill_map_pt(map, xdiv, ydiv,
1133 							&block->area.p1,
1134 							end ? '>' : 'X');
1135 					map_1d_info(map, xdiv, ydiv, nice,
1136 							&block->area);
1137 				}
1138 			}
1139 		}
1140 
1141 		spin_unlock_irqrestore(&list_lock, flags);
1142 
1143 		if (s) {
1144 			seq_printf(s, "CONTAINER %d DUMP BEGIN\n", lut_idx);
1145 			for (i = 0; i < 128; i++)
1146 				seq_printf(s, "%03d:%s\n", i, map[i]);
1147 			seq_printf(s, "CONTAINER %d DUMP END\n", lut_idx);
1148 		} else {
1149 			dev_dbg(omap_dmm->dev, "CONTAINER %d DUMP BEGIN\n",
1150 				lut_idx);
1151 			for (i = 0; i < 128; i++)
1152 				dev_dbg(omap_dmm->dev, "%03d:%s\n", i, map[i]);
1153 			dev_dbg(omap_dmm->dev, "CONTAINER %d DUMP END\n",
1154 				lut_idx);
1155 		}
1156 	}
1157 
1158 error:
1159 	kfree(map);
1160 	kfree(global_map);
1161 
1162 	return 0;
1163 }
1164 #endif
1165 
1166 #ifdef CONFIG_PM_SLEEP
1167 static int omap_dmm_resume(struct device *dev)
1168 {
1169 	struct tcm_area area;
1170 	int i;
1171 
1172 	if (!omap_dmm)
1173 		return -ENODEV;
1174 
1175 	area = (struct tcm_area) {
1176 		.tcm = NULL,
1177 		.p1.x = omap_dmm->container_width - 1,
1178 		.p1.y = omap_dmm->container_height - 1,
1179 	};
1180 
1181 	/* initialize all LUTs to dummy page entries */
1182 	for (i = 0; i < omap_dmm->num_lut; i++) {
1183 		area.tcm = omap_dmm->tcm[i];
1184 		if (fill(&area, NULL, 0, 0, true))
1185 			dev_err(dev, "refill failed");
1186 	}
1187 
1188 	return 0;
1189 }
1190 #endif
1191 
1192 static SIMPLE_DEV_PM_OPS(omap_dmm_pm_ops, NULL, omap_dmm_resume);
1193 
1194 #if defined(CONFIG_OF)
1195 static const struct dmm_platform_data dmm_omap4_platform_data = {
1196 	.cpu_cache_flags = OMAP_BO_WC,
1197 };
1198 
1199 static const struct dmm_platform_data dmm_omap5_platform_data = {
1200 	.cpu_cache_flags = OMAP_BO_UNCACHED,
1201 };
1202 
1203 static const struct of_device_id dmm_of_match[] = {
1204 	{
1205 		.compatible = "ti,omap4-dmm",
1206 		.data = &dmm_omap4_platform_data,
1207 	},
1208 	{
1209 		.compatible = "ti,omap5-dmm",
1210 		.data = &dmm_omap5_platform_data,
1211 	},
1212 	{},
1213 };
1214 #endif
1215 
1216 struct platform_driver omap_dmm_driver = {
1217 	.probe = omap_dmm_probe,
1218 	.remove = omap_dmm_remove,
1219 	.driver = {
1220 		.owner = THIS_MODULE,
1221 		.name = DMM_DRIVER_NAME,
1222 		.of_match_table = of_match_ptr(dmm_of_match),
1223 		.pm = &omap_dmm_pm_ops,
1224 	},
1225 };
1226 
1227 MODULE_LICENSE("GPL v2");
1228 MODULE_AUTHOR("Andy Gross <andy.gross@ti.com>");
1229 MODULE_DESCRIPTION("OMAP DMM/Tiler Driver");
1230