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