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 */
35static struct tcm *containers[TILFMT_NFORMATS];
36static struct dmm *omap_dmm;
37
38#if defined(CONFIG_OF)
39static const struct of_device_id dmm_of_match[];
40#endif
41
42/* global spinlock for protecting lists */
43static 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
54static 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 */
69static 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
76static 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(chan: dmm->wa_dma_chan, dest: dst, src, len: 4, flags: 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(chan: 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(chan: dmm->wa_dma_chan);
99 return 0;
100}
101
102static 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(addr: 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(addr: dmm->wa_dma_data);
123}
124
125static 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, addr: 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, addr: dmm->base + reg);
146 }
147}
148
149static 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(lock: &dmm->wa_lock, flags);
158
159 return v;
160 } else {
161 return readl(addr: dmm->base + reg);
162 }
163}
164
165static 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(lock: &dmm->wa_lock, flags);
173 } else {
174 writel(val, addr: dmm->base + reg);
175 }
176}
177
178static 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(dev: dmm->dev, size: sizeof(u32),
185 dma_handle: &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(dev: dmm->dev, size: 4, cpu_addr: dmm->wa_dma_data, dma_handle: dmm->wa_dma_handle);
195 return -ENODEV;
196 }
197
198 return 0;
199}
200
201static void dmm_workaround_uninit(struct dmm *dmm)
202{
203 dma_release_channel(chan: dmm->wa_dma_chan);
204
205 dma_free_coherent(dev: dmm->dev, size: 4, cpu_addr: dmm->wa_dma_data, dma_handle: dmm->wa_dma_handle);
206}
207
208/* simple allocator to grab next 16 byte aligned memory from txn */
209static 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 */
230static 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: 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
262static void release_engine(struct refill_engine *engine)
263{
264 unsigned long flags;
265
266 spin_lock_irqsave(&list_lock, flags);
267 list_add(new: &engine->idle_node, head: &omap_dmm->idle_head);
268 spin_unlock_irqrestore(lock: &list_lock, flags);
269
270 atomic_inc(v: &omap_dmm->engine_counter);
271 wake_up_interruptible(&omap_dmm->engine_queue);
272}
273
274static 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, val: 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(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 */
305static 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(error: ret);
318
319 /* grab an idle engine */
320 spin_lock_irqsave(&list_lock, flags);
321 if (!list_empty(head: &dmm->idle_head)) {
322 engine = list_entry(dmm->idle_head.next, struct refill_engine,
323 idle_node);
324 list_del(entry: &engine->idle_node);
325 }
326 spin_unlock_irqrestore(lock: &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 */
344static 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, sz: sizeof(*pat), pa: &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, sz: 4*i, pa: &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 */
391static 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(addr: &txn->last_pat->next_pa);
415
416 /* write to PAT_DESCR to clear out any pending transaction */
417 dmm_write(dmm, val: 0x0, reg: 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(x: &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, val: engine->refill_pa, reg: reg[PAT_DESCR][engine->id]);
434
435 if (wait) {
436 if (!wait_for_completion_timeout(x: &engine->compl,
437 timeout: msecs_to_jiffies(m: 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
448cleanup:
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 */
459static 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(dmm: omap_dmm, tcm: area->tcm);
482 if (IS_ERR_OR_NULL(ptr: 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, area: &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 */
507int tiler_pin(struct tiler_block *block, struct page **pages,
508 u32 npages, u32 roll, bool wait)
509{
510 int ret;
511
512 ret = fill(area: &block->area, pages, npages, roll, wait);
513
514 if (ret)
515 tiler_unpin(block);
516
517 return ret;
518}
519
520int tiler_unpin(struct tiler_block *block)
521{
522 return fill(area: &block->area, NULL, npages: 0, roll: 0, wait: false);
523}
524
525/*
526 * Reserve/release
527 */
528struct 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(size: sizeof(*block), GFP_KERNEL);
538 if (!block)
539 return ERR_PTR(error: -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(tcm: containers[fmt], width: w, height: h, align, offset: -1, slot_bytes,
556 area: &block->area);
557 if (ret) {
558 kfree(objp: block);
559 return ERR_PTR(error: -ENOMEM);
560 }
561
562 /* add to allocation list */
563 spin_lock_irqsave(&list_lock, flags);
564 list_add(new: &block->alloc_node, head: &omap_dmm->alloc_head);
565 spin_unlock_irqrestore(lock: &list_lock, flags);
566
567 return block;
568}
569
570struct tiler_block *tiler_reserve_1d(size_t size)
571{
572 struct tiler_block *block = kzalloc(size: 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(error: -ENOMEM);
578
579 block->fmt = TILFMT_PAGE;
580
581 if (tcm_reserve_1d(tcm: containers[TILFMT_PAGE], slots: num_pages,
582 area: &block->area)) {
583 kfree(objp: block);
584 return ERR_PTR(error: -ENOMEM);
585 }
586
587 spin_lock_irqsave(&list_lock, flags);
588 list_add(new: &block->alloc_node, head: &omap_dmm->alloc_head);
589 spin_unlock_irqrestore(lock: &list_lock, flags);
590
591 return block;
592}
593
594/* note: if you have pin'd pages, you should have already unpin'd first! */
595int tiler_release(struct tiler_block *block)
596{
597 int ret = tcm_free(area: &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(entry: &block->alloc_node);
605 spin_unlock_irqrestore(lock: &list_lock, flags);
606
607 kfree(objp: 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 */
634static 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
667dma_addr_t tiler_ssptr(struct tiler_block *block)
668{
669 BUG_ON(!validfmt(block->fmt));
670
671 return TILVIEW_8BIT + tiler_get_address(fmt: block->fmt, orient: 0,
672 x: block->area.p0.x * geom[block->fmt].slot_w,
673 y: block->area.p0.y * geom[block->fmt].slot_h);
674}
675
676dma_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(fmt: block->fmt, orient,
683 x: (p->x * geom[block->fmt].slot_w) + x,
684 y: (p->y * geom[block->fmt].slot_h) + y);
685}
686
687void 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
694u32 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
704size_t tiler_size(enum tiler_fmt fmt, u16 w, u16 h)
705{
706 tiler_align(fmt, w: &w, h: &h);
707 return geom[fmt].cpp * w * h;
708}
709
710size_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
716u32 tiler_get_cpu_cache_flags(void)
717{
718 return omap_dmm->plat_data->cpu_cache_flags;
719}
720
721bool dmm_is_available(void)
722{
723 return omap_dmm ? true : false;
724}
725
726static void 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(dmm: omap_dmm, val: 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(entry: &block->alloc_node);
742 kfree(objp: block);
743 }
744 spin_unlock_irqrestore(lock: &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(objp: omap_dmm->tcm);
750
751 kfree(objp: omap_dmm->engines);
752 if (omap_dmm->refill_va)
753 dma_free_wc(dev: omap_dmm->dev,
754 REFILL_BUFFER_SIZE * omap_dmm->num_engines,
755 cpu_addr: omap_dmm->refill_va, dma_addr: 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(dmm: omap_dmm);
761
762 iounmap(addr: omap_dmm->base);
763 kfree(objp: omap_dmm);
764 omap_dmm = NULL;
765 }
766}
767
768static int omap_dmm_probe(struct platform_device *dev)
769{
770 int ret = -EFAULT, i;
771 struct tcm_area area = {0};
772 u32 hwinfo, pat_geom;
773 struct resource *mem;
774
775 omap_dmm = kzalloc(size: sizeof(*omap_dmm), GFP_KERNEL);
776 if (!omap_dmm)
777 goto fail;
778
779 /* initialize lists */
780 INIT_LIST_HEAD(list: &omap_dmm->alloc_head);
781 INIT_LIST_HEAD(list: &omap_dmm->idle_head);
782
783 init_waitqueue_head(&omap_dmm->engine_queue);
784
785 if (dev->dev.of_node) {
786 const struct of_device_id *match;
787
788 match = of_match_node(matches: dmm_of_match, node: dev->dev.of_node);
789 if (!match) {
790 dev_err(&dev->dev, "failed to find matching device node\n");
791 ret = -ENODEV;
792 goto fail;
793 }
794
795 omap_dmm->plat_data = match->data;
796 }
797
798 /* lookup hwmod data - base address and irq */
799 mem = platform_get_resource(dev, IORESOURCE_MEM, 0);
800 if (!mem) {
801 dev_err(&dev->dev, "failed to get base address resource\n");
802 goto fail;
803 }
804
805 omap_dmm->phys_base = mem->start;
806 omap_dmm->base = ioremap(offset: mem->start, SZ_2K);
807
808 if (!omap_dmm->base) {
809 dev_err(&dev->dev, "failed to get dmm base address\n");
810 goto fail;
811 }
812
813 omap_dmm->irq = platform_get_irq(dev, 0);
814 if (omap_dmm->irq < 0)
815 goto fail;
816
817 omap_dmm->dev = &dev->dev;
818
819 if (of_machine_is_compatible(compat: "ti,dra7")) {
820 /*
821 * DRA7 Errata i878 says that MPU should not be used to access
822 * RAM and DMM at the same time. As it's not possible to prevent
823 * MPU accessing RAM, we need to access DMM via a proxy.
824 */
825 if (!dmm_workaround_init(dmm: omap_dmm)) {
826 omap_dmm->dmm_workaround = true;
827 dev_info(&dev->dev,
828 "workaround for errata i878 in use\n");
829 } else {
830 dev_warn(&dev->dev,
831 "failed to initialize work-around for i878\n");
832 }
833 }
834
835 hwinfo = dmm_read(dmm: omap_dmm, DMM_PAT_HWINFO);
836 omap_dmm->num_engines = (hwinfo >> 24) & 0x1F;
837 omap_dmm->num_lut = (hwinfo >> 16) & 0x1F;
838 omap_dmm->container_width = 256;
839 omap_dmm->container_height = 128;
840
841 atomic_set(v: &omap_dmm->engine_counter, i: omap_dmm->num_engines);
842
843 /* read out actual LUT width and height */
844 pat_geom = dmm_read(dmm: omap_dmm, DMM_PAT_GEOMETRY);
845 omap_dmm->lut_width = ((pat_geom >> 16) & 0xF) << 5;
846 omap_dmm->lut_height = ((pat_geom >> 24) & 0xF) << 5;
847
848 /* increment LUT by one if on OMAP5 */
849 /* LUT has twice the height, and is split into a separate container */
850 if (omap_dmm->lut_height != omap_dmm->container_height)
851 omap_dmm->num_lut++;
852
853 /* initialize DMM registers */
854 dmm_write(dmm: omap_dmm, val: 0x88888888, DMM_PAT_VIEW__0);
855 dmm_write(dmm: omap_dmm, val: 0x88888888, DMM_PAT_VIEW__1);
856 dmm_write(dmm: omap_dmm, val: 0x80808080, DMM_PAT_VIEW_MAP__0);
857 dmm_write(dmm: omap_dmm, val: 0x80000000, DMM_PAT_VIEW_MAP_BASE);
858 dmm_write(dmm: omap_dmm, val: 0x88888888, DMM_TILER_OR__0);
859 dmm_write(dmm: omap_dmm, val: 0x88888888, DMM_TILER_OR__1);
860
861 omap_dmm->dummy_page = alloc_page(GFP_KERNEL | __GFP_DMA32);
862 if (!omap_dmm->dummy_page) {
863 dev_err(&dev->dev, "could not allocate dummy page\n");
864 ret = -ENOMEM;
865 goto fail;
866 }
867
868 /* set dma mask for device */
869 ret = dma_set_coherent_mask(dev: &dev->dev, DMA_BIT_MASK(32));
870 if (ret)
871 goto fail;
872
873 omap_dmm->dummy_pa = page_to_phys(omap_dmm->dummy_page);
874
875 /* alloc refill memory */
876 omap_dmm->refill_va = dma_alloc_wc(dev: &dev->dev,
877 REFILL_BUFFER_SIZE * omap_dmm->num_engines,
878 dma_addr: &omap_dmm->refill_pa, GFP_KERNEL);
879 if (!omap_dmm->refill_va) {
880 dev_err(&dev->dev, "could not allocate refill memory\n");
881 ret = -ENOMEM;
882 goto fail;
883 }
884
885 /* alloc engines */
886 omap_dmm->engines = kcalloc(n: omap_dmm->num_engines,
887 size: sizeof(*omap_dmm->engines), GFP_KERNEL);
888 if (!omap_dmm->engines) {
889 ret = -ENOMEM;
890 goto fail;
891 }
892
893 for (i = 0; i < omap_dmm->num_engines; i++) {
894 omap_dmm->engines[i].id = i;
895 omap_dmm->engines[i].dmm = omap_dmm;
896 omap_dmm->engines[i].refill_va = omap_dmm->refill_va +
897 (REFILL_BUFFER_SIZE * i);
898 omap_dmm->engines[i].refill_pa = omap_dmm->refill_pa +
899 (REFILL_BUFFER_SIZE * i);
900 init_completion(x: &omap_dmm->engines[i].compl);
901
902 list_add(new: &omap_dmm->engines[i].idle_node, head: &omap_dmm->idle_head);
903 }
904
905 omap_dmm->tcm = kcalloc(n: omap_dmm->num_lut, size: sizeof(*omap_dmm->tcm),
906 GFP_KERNEL);
907 if (!omap_dmm->tcm) {
908 ret = -ENOMEM;
909 goto fail;
910 }
911
912 /* init containers */
913 /* Each LUT is associated with a TCM (container manager). We use the
914 lut_id to denote the lut_id used to identify the correct LUT for
915 programming during reill operations */
916 for (i = 0; i < omap_dmm->num_lut; i++) {
917 omap_dmm->tcm[i] = sita_init(width: omap_dmm->container_width,
918 height: omap_dmm->container_height);
919
920 if (!omap_dmm->tcm[i]) {
921 dev_err(&dev->dev, "failed to allocate container\n");
922 ret = -ENOMEM;
923 goto fail;
924 }
925
926 omap_dmm->tcm[i]->lut_id = i;
927 }
928
929 /* assign access mode containers to applicable tcm container */
930 /* OMAP 4 has 1 container for all 4 views */
931 /* OMAP 5 has 2 containers, 1 for 2D and 1 for 1D */
932 containers[TILFMT_8BIT] = omap_dmm->tcm[0];
933 containers[TILFMT_16BIT] = omap_dmm->tcm[0];
934 containers[TILFMT_32BIT] = omap_dmm->tcm[0];
935
936 if (omap_dmm->container_height != omap_dmm->lut_height) {
937 /* second LUT is used for PAGE mode. Programming must use
938 y offset that is added to all y coordinates. LUT id is still
939 0, because it is the same LUT, just the upper 128 lines */
940 containers[TILFMT_PAGE] = omap_dmm->tcm[1];
941 omap_dmm->tcm[1]->y_offset = OMAP5_LUT_OFFSET;
942 omap_dmm->tcm[1]->lut_id = 0;
943 } else {
944 containers[TILFMT_PAGE] = omap_dmm->tcm[0];
945 }
946
947 area = (struct tcm_area) {
948 .tcm = NULL,
949 .p1.x = omap_dmm->container_width - 1,
950 .p1.y = omap_dmm->container_height - 1,
951 };
952
953 ret = request_irq(irq: omap_dmm->irq, handler: omap_dmm_irq_handler, IRQF_SHARED,
954 name: "omap_dmm_irq_handler", dev: omap_dmm);
955
956 if (ret) {
957 dev_err(&dev->dev, "couldn't register IRQ %d, error %d\n",
958 omap_dmm->irq, ret);
959 omap_dmm->irq = -1;
960 goto fail;
961 }
962
963 /* Enable all interrupts for each refill engine except
964 * ERR_LUT_MISS<n> (which is just advisory, and we don't care
965 * about because we want to be able to refill live scanout
966 * buffers for accelerated pan/scroll) and FILL_DSC<n> which
967 * we just generally don't care about.
968 */
969 dmm_write(dmm: omap_dmm, val: 0x7e7e7e7e, DMM_PAT_IRQENABLE_SET);
970
971 /* initialize all LUTs to dummy page entries */
972 for (i = 0; i < omap_dmm->num_lut; i++) {
973 area.tcm = omap_dmm->tcm[i];
974 if (fill(area: &area, NULL, npages: 0, roll: 0, wait: true))
975 dev_err(omap_dmm->dev, "refill failed");
976 }
977
978 dev_info(omap_dmm->dev, "initialized all PAT entries\n");
979
980 return 0;
981
982fail:
983 omap_dmm_remove(dev);
984 return ret;
985}
986
987/*
988 * debugfs support
989 */
990
991#ifdef CONFIG_DEBUG_FS
992
993static const char *alphabet = "abcdefghijklmnopqrstuvwxyz"
994 "ABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789";
995static const char *special = ".,:;'\"`~!^-+";
996
997static void fill_map(char **map, int xdiv, int ydiv, struct tcm_area *a,
998 char c, bool ovw)
999{
1000 int x, y;
1001 for (y = a->p0.y / ydiv; y <= a->p1.y / ydiv; y++)
1002 for (x = a->p0.x / xdiv; x <= a->p1.x / xdiv; x++)
1003 if (map[y][x] == ' ' || ovw)
1004 map[y][x] = c;
1005}
1006
1007static void fill_map_pt(char **map, int xdiv, int ydiv, struct tcm_pt *p,
1008 char c)
1009{
1010 map[p->y / ydiv][p->x / xdiv] = c;
1011}
1012
1013static char read_map_pt(char **map, int xdiv, int ydiv, struct tcm_pt *p)
1014{
1015 return map[p->y / ydiv][p->x / xdiv];
1016}
1017
1018static int map_width(int xdiv, int x0, int x1)
1019{
1020 return (x1 / xdiv) - (x0 / xdiv) + 1;
1021}
1022
1023static void text_map(char **map, int xdiv, char *nice, int yd, int x0, int x1)
1024{
1025 char *p = map[yd] + (x0 / xdiv);
1026 int w = (map_width(xdiv, x0, x1) - strlen(nice)) / 2;
1027 if (w >= 0) {
1028 p += w;
1029 while (*nice)
1030 *p++ = *nice++;
1031 }
1032}
1033
1034static void map_1d_info(char **map, int xdiv, int ydiv, char *nice,
1035 struct tcm_area *a)
1036{
1037 sprintf(buf: nice, fmt: "%dK", tcm_sizeof(*a) * 4);
1038 if (a->p0.y + 1 < a->p1.y) {
1039 text_map(map, xdiv, nice, yd: (a->p0.y + a->p1.y) / 2 / ydiv, x0: 0,
1040 x1: 256 - 1);
1041 } else if (a->p0.y < a->p1.y) {
1042 if (strlen(nice) < map_width(xdiv, x0: a->p0.x, x1: 256 - 1))
1043 text_map(map, xdiv, nice, yd: a->p0.y / ydiv,
1044 x0: a->p0.x + xdiv, x1: 256 - 1);
1045 else if (strlen(nice) < map_width(xdiv, x0: 0, x1: a->p1.x))
1046 text_map(map, xdiv, nice, yd: a->p1.y / ydiv,
1047 x0: 0, x1: a->p1.y - xdiv);
1048 } else if (strlen(nice) + 1 < map_width(xdiv, x0: a->p0.x, x1: a->p1.x)) {
1049 text_map(map, xdiv, nice, yd: a->p0.y / ydiv, x0: a->p0.x, x1: a->p1.x);
1050 }
1051}
1052
1053static void map_2d_info(char **map, int xdiv, int ydiv, char *nice,
1054 struct tcm_area *a)
1055{
1056 sprintf(buf: nice, fmt: "(%d*%d)", tcm_awidth(*a), tcm_aheight(*a));
1057 if (strlen(nice) + 1 < map_width(xdiv, x0: a->p0.x, x1: a->p1.x))
1058 text_map(map, xdiv, nice, yd: (a->p0.y + a->p1.y) / 2 / ydiv,
1059 x0: a->p0.x, x1: a->p1.x);
1060}
1061
1062int tiler_map_show(struct seq_file *s, void *arg)
1063{
1064 int xdiv = 2, ydiv = 1;
1065 char **map = NULL, *global_map;
1066 struct tiler_block *block;
1067 struct tcm_area a, p;
1068 int i;
1069 const char *m2d = alphabet;
1070 const char *a2d = special;
1071 const char *m2dp = m2d, *a2dp = a2d;
1072 char nice[128];
1073 int h_adj;
1074 int w_adj;
1075 unsigned long flags;
1076 int lut_idx;
1077
1078
1079 if (!omap_dmm) {
1080 /* early return if dmm/tiler device is not initialized */
1081 return 0;
1082 }
1083
1084 h_adj = omap_dmm->container_height / ydiv;
1085 w_adj = omap_dmm->container_width / xdiv;
1086
1087 map = kmalloc_array(n: h_adj, size: sizeof(*map), GFP_KERNEL);
1088 global_map = kmalloc_array(n: w_adj + 1, size: h_adj, GFP_KERNEL);
1089
1090 if (!map || !global_map)
1091 goto error;
1092
1093 for (lut_idx = 0; lut_idx < omap_dmm->num_lut; lut_idx++) {
1094 memset(map, 0, h_adj * sizeof(*map));
1095 memset(global_map, ' ', (w_adj + 1) * h_adj);
1096
1097 for (i = 0; i < omap_dmm->container_height; i++) {
1098 map[i] = global_map + i * (w_adj + 1);
1099 map[i][w_adj] = 0;
1100 }
1101
1102 spin_lock_irqsave(&list_lock, flags);
1103
1104 list_for_each_entry(block, &omap_dmm->alloc_head, alloc_node) {
1105 if (block->area.tcm == omap_dmm->tcm[lut_idx]) {
1106 if (block->fmt != TILFMT_PAGE) {
1107 fill_map(map, xdiv, ydiv, a: &block->area,
1108 c: *m2dp, ovw: true);
1109 if (!*++a2dp)
1110 a2dp = a2d;
1111 if (!*++m2dp)
1112 m2dp = m2d;
1113 map_2d_info(map, xdiv, ydiv, nice,
1114 a: &block->area);
1115 } else {
1116 bool start = read_map_pt(map, xdiv,
1117 ydiv, p: &block->area.p0) == ' ';
1118 bool end = read_map_pt(map, xdiv, ydiv,
1119 p: &block->area.p1) == ' ';
1120
1121 tcm_for_each_slice(a, block->area, p)
1122 fill_map(map, xdiv, ydiv, a: &a,
1123 c: '=', ovw: true);
1124 fill_map_pt(map, xdiv, ydiv,
1125 p: &block->area.p0,
1126 c: start ? '<' : 'X');
1127 fill_map_pt(map, xdiv, ydiv,
1128 p: &block->area.p1,
1129 c: end ? '>' : 'X');
1130 map_1d_info(map, xdiv, ydiv, nice,
1131 a: &block->area);
1132 }
1133 }
1134 }
1135
1136 spin_unlock_irqrestore(lock: &list_lock, flags);
1137
1138 if (s) {
1139 seq_printf(m: s, fmt: "CONTAINER %d DUMP BEGIN\n", lut_idx);
1140 for (i = 0; i < 128; i++)
1141 seq_printf(m: s, fmt: "%03d:%s\n", i, map[i]);
1142 seq_printf(m: s, fmt: "CONTAINER %d DUMP END\n", lut_idx);
1143 } else {
1144 dev_dbg(omap_dmm->dev, "CONTAINER %d DUMP BEGIN\n",
1145 lut_idx);
1146 for (i = 0; i < 128; i++)
1147 dev_dbg(omap_dmm->dev, "%03d:%s\n", i, map[i]);
1148 dev_dbg(omap_dmm->dev, "CONTAINER %d DUMP END\n",
1149 lut_idx);
1150 }
1151 }
1152
1153error:
1154 kfree(objp: map);
1155 kfree(objp: global_map);
1156
1157 return 0;
1158}
1159#endif
1160
1161#ifdef CONFIG_PM_SLEEP
1162static int omap_dmm_resume(struct device *dev)
1163{
1164 struct tcm_area area;
1165 int i;
1166
1167 if (!omap_dmm)
1168 return -ENODEV;
1169
1170 area = (struct tcm_area) {
1171 .tcm = NULL,
1172 .p1.x = omap_dmm->container_width - 1,
1173 .p1.y = omap_dmm->container_height - 1,
1174 };
1175
1176 /* initialize all LUTs to dummy page entries */
1177 for (i = 0; i < omap_dmm->num_lut; i++) {
1178 area.tcm = omap_dmm->tcm[i];
1179 if (fill(area: &area, NULL, npages: 0, roll: 0, wait: true))
1180 dev_err(dev, "refill failed");
1181 }
1182
1183 return 0;
1184}
1185#endif
1186
1187static SIMPLE_DEV_PM_OPS(omap_dmm_pm_ops, NULL, omap_dmm_resume);
1188
1189#if defined(CONFIG_OF)
1190static const struct dmm_platform_data dmm_omap4_platform_data = {
1191 .cpu_cache_flags = OMAP_BO_WC,
1192};
1193
1194static const struct dmm_platform_data dmm_omap5_platform_data = {
1195 .cpu_cache_flags = OMAP_BO_UNCACHED,
1196};
1197
1198static const struct of_device_id dmm_of_match[] = {
1199 {
1200 .compatible = "ti,omap4-dmm",
1201 .data = &dmm_omap4_platform_data,
1202 },
1203 {
1204 .compatible = "ti,omap5-dmm",
1205 .data = &dmm_omap5_platform_data,
1206 },
1207 {},
1208};
1209#endif
1210
1211struct platform_driver omap_dmm_driver = {
1212 .probe = omap_dmm_probe,
1213 .remove_new = omap_dmm_remove,
1214 .driver = {
1215 .owner = THIS_MODULE,
1216 .name = DMM_DRIVER_NAME,
1217 .of_match_table = of_match_ptr(dmm_of_match),
1218 .pm = &omap_dmm_pm_ops,
1219 },
1220};
1221
1222MODULE_LICENSE("GPL v2");
1223MODULE_AUTHOR("Andy Gross <andy.gross@ti.com>");
1224MODULE_DESCRIPTION("OMAP DMM/Tiler Driver");
1225

source code of linux/drivers/gpu/drm/omapdrm/omap_dmm_tiler.c