Linux Audio

Check our new training course

Loading...
v3.1
   1/*
   2 * Copyright 2007 Dave Airlied
   3 * All Rights Reserved.
   4 *
   5 * Permission is hereby granted, free of charge, to any person obtaining a
   6 * copy of this software and associated documentation files (the "Software"),
   7 * to deal in the Software without restriction, including without limitation
   8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9 * and/or sell copies of the Software, and to permit persons to whom the
  10 * Software is furnished to do so, subject to the following conditions:
  11 *
  12 * The above copyright notice and this permission notice (including the next
  13 * paragraph) shall be included in all copies or substantial portions of the
  14 * Software.
  15 *
  16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19 * VA LINUX SYSTEMS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
  20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  22 * OTHER DEALINGS IN THE SOFTWARE.
  23 */
  24/*
  25 * Authors: Dave Airlied <airlied@linux.ie>
  26 *	    Ben Skeggs   <darktama@iinet.net.au>
  27 *	    Jeremy Kolb  <jkolb@brandeis.edu>
  28 */
  29
  30#include "drmP.h"
 
  31
  32#include "nouveau_drm.h"
  33#include "nouveau_drv.h"
  34#include "nouveau_dma.h"
  35#include "nouveau_mm.h"
  36#include "nouveau_vm.h"
  37
  38#include <linux/log2.h>
  39#include <linux/slab.h>
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  40
  41static void
  42nouveau_bo_del_ttm(struct ttm_buffer_object *bo)
  43{
  44	struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
  45	struct drm_device *dev = dev_priv->dev;
  46	struct nouveau_bo *nvbo = nouveau_bo(bo);
  47
  48	if (unlikely(nvbo->gem))
  49		DRM_ERROR("bo %p still attached to GEM object\n", bo);
  50
  51	nv10_mem_put_tile_region(dev, nvbo->tile, NULL);
  52	kfree(nvbo);
  53}
  54
  55static void
  56nouveau_bo_fixup_align(struct nouveau_bo *nvbo, u32 flags,
  57		       int *align, int *size)
  58{
  59	struct drm_nouveau_private *dev_priv = nouveau_bdev(nvbo->bo.bdev);
 
  60
  61	if (dev_priv->card_type < NV_50) {
  62		if (nvbo->tile_mode) {
  63			if (dev_priv->chipset >= 0x40) {
  64				*align = 65536;
  65				*size = roundup(*size, 64 * nvbo->tile_mode);
  66
  67			} else if (dev_priv->chipset >= 0x30) {
  68				*align = 32768;
  69				*size = roundup(*size, 64 * nvbo->tile_mode);
  70
  71			} else if (dev_priv->chipset >= 0x20) {
  72				*align = 16384;
  73				*size = roundup(*size, 64 * nvbo->tile_mode);
  74
  75			} else if (dev_priv->chipset >= 0x10) {
  76				*align = 16384;
  77				*size = roundup(*size, 32 * nvbo->tile_mode);
  78			}
  79		}
  80	} else {
  81		*size = roundup(*size, (1 << nvbo->page_shift));
  82		*align = max((1 <<  nvbo->page_shift), *align);
  83	}
  84
  85	*size = roundup(*size, PAGE_SIZE);
  86}
  87
  88int
  89nouveau_bo_new(struct drm_device *dev, int size, int align,
  90	       uint32_t flags, uint32_t tile_mode, uint32_t tile_flags,
 
  91	       struct nouveau_bo **pnvbo)
  92{
  93	struct drm_nouveau_private *dev_priv = dev->dev_private;
  94	struct nouveau_bo *nvbo;
 
  95	int ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  96
  97	nvbo = kzalloc(sizeof(struct nouveau_bo), GFP_KERNEL);
  98	if (!nvbo)
  99		return -ENOMEM;
 100	INIT_LIST_HEAD(&nvbo->head);
 101	INIT_LIST_HEAD(&nvbo->entry);
 102	INIT_LIST_HEAD(&nvbo->vma_list);
 103	nvbo->tile_mode = tile_mode;
 104	nvbo->tile_flags = tile_flags;
 105	nvbo->bo.bdev = &dev_priv->ttm.bdev;
 
 
 
 106
 107	nvbo->page_shift = 12;
 108	if (dev_priv->bar1_vm) {
 109		if (!(flags & TTM_PL_FLAG_TT) && size > 256 * 1024)
 110			nvbo->page_shift = dev_priv->bar1_vm->lpg_shift;
 111	}
 112
 113	nouveau_bo_fixup_align(nvbo, flags, &align, &size);
 114	nvbo->bo.mem.num_pages = size >> PAGE_SHIFT;
 115	nouveau_bo_placement_set(nvbo, flags, 0);
 116
 117	ret = ttm_bo_init(&dev_priv->ttm.bdev, &nvbo->bo, size,
 118			  ttm_bo_type_device, &nvbo->placement,
 119			  align >> PAGE_SHIFT, 0, false, NULL, size,
 120			  nouveau_bo_del_ttm);
 
 
 
 121	if (ret) {
 122		/* ttm will call nouveau_bo_del_ttm if it fails.. */
 123		return ret;
 124	}
 125
 126	*pnvbo = nvbo;
 127	return 0;
 128}
 129
 130static void
 131set_placement_list(uint32_t *pl, unsigned *n, uint32_t type, uint32_t flags)
 132{
 133	*n = 0;
 134
 135	if (type & TTM_PL_FLAG_VRAM)
 136		pl[(*n)++] = TTM_PL_FLAG_VRAM | flags;
 137	if (type & TTM_PL_FLAG_TT)
 138		pl[(*n)++] = TTM_PL_FLAG_TT | flags;
 139	if (type & TTM_PL_FLAG_SYSTEM)
 140		pl[(*n)++] = TTM_PL_FLAG_SYSTEM | flags;
 141}
 142
 143static void
 144set_placement_range(struct nouveau_bo *nvbo, uint32_t type)
 145{
 146	struct drm_nouveau_private *dev_priv = nouveau_bdev(nvbo->bo.bdev);
 147	int vram_pages = dev_priv->vram_size >> PAGE_SHIFT;
 
 148
 149	if (dev_priv->card_type == NV_10 &&
 150	    nvbo->tile_mode && (type & TTM_PL_FLAG_VRAM) &&
 151	    nvbo->bo.mem.num_pages < vram_pages / 2) {
 152		/*
 153		 * Make sure that the color and depth buffers are handled
 154		 * by independent memory controller units. Up to a 9x
 155		 * speed up when alpha-blending and depth-test are enabled
 156		 * at the same time.
 157		 */
 158		if (nvbo->tile_flags & NOUVEAU_GEM_TILE_ZETA) {
 159			nvbo->placement.fpfn = vram_pages / 2;
 160			nvbo->placement.lpfn = ~0;
 161		} else {
 162			nvbo->placement.fpfn = 0;
 163			nvbo->placement.lpfn = vram_pages / 2;
 
 
 
 
 
 
 
 
 164		}
 165	}
 166}
 167
 168void
 169nouveau_bo_placement_set(struct nouveau_bo *nvbo, uint32_t type, uint32_t busy)
 170{
 171	struct ttm_placement *pl = &nvbo->placement;
 172	uint32_t flags = TTM_PL_MASK_CACHING |
 173		(nvbo->pin_refcnt ? TTM_PL_FLAG_NO_EVICT : 0);
 
 174
 175	pl->placement = nvbo->placements;
 176	set_placement_list(nvbo->placements, &pl->num_placement,
 177			   type, flags);
 178
 179	pl->busy_placement = nvbo->busy_placements;
 180	set_placement_list(nvbo->busy_placements, &pl->num_busy_placement,
 181			   type | busy, flags);
 182
 183	set_placement_range(nvbo, type);
 184}
 185
 186int
 187nouveau_bo_pin(struct nouveau_bo *nvbo, uint32_t memtype)
 188{
 189	struct drm_nouveau_private *dev_priv = nouveau_bdev(nvbo->bo.bdev);
 190	struct ttm_buffer_object *bo = &nvbo->bo;
 
 191	int ret;
 192
 193	if (nvbo->pin_refcnt && !(memtype & (1 << bo->mem.mem_type))) {
 194		NV_ERROR(nouveau_bdev(bo->bdev)->dev,
 195			 "bo %p pinned elsewhere: 0x%08x vs 0x%08x\n", bo,
 196			 1 << bo->mem.mem_type, memtype);
 197		return -EINVAL;
 198	}
 199
 200	if (nvbo->pin_refcnt++)
 201		return 0;
 
 
 
 
 
 
 
 
 
 
 202
 203	ret = ttm_bo_reserve(bo, false, false, false, 0);
 204	if (ret)
 
 
 
 
 
 
 205		goto out;
 
 206
 
 
 
 
 
 
 
 
 207	nouveau_bo_placement_set(nvbo, memtype, 0);
 208
 209	ret = nouveau_bo_validate(nvbo, false, false, false);
 210	if (ret == 0) {
 211		switch (bo->mem.mem_type) {
 212		case TTM_PL_VRAM:
 213			dev_priv->fb_aper_free -= bo->mem.size;
 214			break;
 215		case TTM_PL_TT:
 216			dev_priv->gart_info.aper_free -= bo->mem.size;
 217			break;
 218		default:
 219			break;
 220		}
 
 
 
 
 
 
 
 221	}
 222	ttm_bo_unreserve(bo);
 223out:
 224	if (unlikely(ret))
 225		nvbo->pin_refcnt--;
 
 226	return ret;
 227}
 228
 229int
 230nouveau_bo_unpin(struct nouveau_bo *nvbo)
 231{
 232	struct drm_nouveau_private *dev_priv = nouveau_bdev(nvbo->bo.bdev);
 233	struct ttm_buffer_object *bo = &nvbo->bo;
 234	int ret;
 235
 236	if (--nvbo->pin_refcnt)
 237		return 0;
 238
 239	ret = ttm_bo_reserve(bo, false, false, false, 0);
 240	if (ret)
 241		return ret;
 242
 
 
 
 
 
 243	nouveau_bo_placement_set(nvbo, bo->mem.placement, 0);
 244
 245	ret = nouveau_bo_validate(nvbo, false, false, false);
 246	if (ret == 0) {
 247		switch (bo->mem.mem_type) {
 248		case TTM_PL_VRAM:
 249			dev_priv->fb_aper_free += bo->mem.size;
 250			break;
 251		case TTM_PL_TT:
 252			dev_priv->gart_info.aper_free += bo->mem.size;
 253			break;
 254		default:
 255			break;
 256		}
 257	}
 258
 
 259	ttm_bo_unreserve(bo);
 260	return ret;
 261}
 262
 263int
 264nouveau_bo_map(struct nouveau_bo *nvbo)
 265{
 266	int ret;
 267
 268	ret = ttm_bo_reserve(&nvbo->bo, false, false, false, 0);
 269	if (ret)
 270		return ret;
 271
 272	ret = ttm_bo_kmap(&nvbo->bo, 0, nvbo->bo.mem.num_pages, &nvbo->kmap);
 
 
 
 
 
 
 
 273	ttm_bo_unreserve(&nvbo->bo);
 274	return ret;
 275}
 276
 277void
 278nouveau_bo_unmap(struct nouveau_bo *nvbo)
 279{
 280	if (nvbo)
 
 
 
 
 
 
 
 281		ttm_bo_kunmap(&nvbo->kmap);
 282}
 283
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 284int
 285nouveau_bo_validate(struct nouveau_bo *nvbo, bool interruptible,
 286		    bool no_wait_reserve, bool no_wait_gpu)
 287{
 288	int ret;
 289
 290	ret = ttm_bo_validate(&nvbo->bo, &nvbo->placement, interruptible,
 291			      no_wait_reserve, no_wait_gpu);
 292	if (ret)
 293		return ret;
 294
 
 
 295	return 0;
 296}
 297
 298u16
 299nouveau_bo_rd16(struct nouveau_bo *nvbo, unsigned index)
 300{
 301	bool is_iomem;
 302	u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
 303	mem = &mem[index];
 304	if (is_iomem)
 305		return ioread16_native((void __force __iomem *)mem);
 306	else
 307		return *mem;
 
 
 
 
 
 
 
 
 
 308}
 
 309
 310void
 311nouveau_bo_wr16(struct nouveau_bo *nvbo, unsigned index, u16 val)
 312{
 313	bool is_iomem;
 314	u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
 315	mem = &mem[index];
 
 
 316	if (is_iomem)
 317		iowrite16_native(val, (void __force __iomem *)mem);
 318	else
 319		*mem = val;
 320}
 321
 322u32
 323nouveau_bo_rd32(struct nouveau_bo *nvbo, unsigned index)
 324{
 325	bool is_iomem;
 326	u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
 327	mem = &mem[index];
 
 
 328	if (is_iomem)
 329		return ioread32_native((void __force __iomem *)mem);
 330	else
 331		return *mem;
 332}
 333
 334void
 335nouveau_bo_wr32(struct nouveau_bo *nvbo, unsigned index, u32 val)
 336{
 337	bool is_iomem;
 338	u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
 339	mem = &mem[index];
 
 
 340	if (is_iomem)
 341		iowrite32_native(val, (void __force __iomem *)mem);
 342	else
 343		*mem = val;
 344}
 345
 346static struct ttm_backend *
 347nouveau_bo_create_ttm_backend_entry(struct ttm_bo_device *bdev)
 348{
 349	struct drm_nouveau_private *dev_priv = nouveau_bdev(bdev);
 350	struct drm_device *dev = dev_priv->dev;
 351
 352	switch (dev_priv->gart_info.type) {
 353#if __OS_HAS_AGP
 354	case NOUVEAU_GART_AGP:
 355		return ttm_agp_backend_init(bdev, dev->agp->bridge);
 356#endif
 357	case NOUVEAU_GART_PDMA:
 358	case NOUVEAU_GART_HW:
 359		return nouveau_sgdma_init_ttm(dev);
 360	default:
 361		NV_ERROR(dev, "Unknown GART type %d\n",
 362			 dev_priv->gart_info.type);
 363		break;
 364	}
 
 365
 366	return NULL;
 367}
 368
 369static int
 370nouveau_bo_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
 371{
 372	/* We'll do this from user space. */
 373	return 0;
 374}
 375
 376static int
 377nouveau_bo_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
 378			 struct ttm_mem_type_manager *man)
 379{
 380	struct drm_nouveau_private *dev_priv = nouveau_bdev(bdev);
 381	struct drm_device *dev = dev_priv->dev;
 382
 383	switch (type) {
 384	case TTM_PL_SYSTEM:
 385		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
 386		man->available_caching = TTM_PL_MASK_CACHING;
 387		man->default_caching = TTM_PL_FLAG_CACHED;
 388		break;
 389	case TTM_PL_VRAM:
 390		if (dev_priv->card_type >= NV_50) {
 
 
 
 
 
 
 
 
 
 
 
 
 391			man->func = &nouveau_vram_manager;
 392			man->io_reserve_fastpath = false;
 393			man->use_io_reserve_lru = true;
 394		} else {
 395			man->func = &ttm_bo_manager_func;
 396		}
 397		man->flags = TTM_MEMTYPE_FLAG_FIXED |
 398			     TTM_MEMTYPE_FLAG_MAPPABLE;
 399		man->available_caching = TTM_PL_FLAG_UNCACHED |
 400					 TTM_PL_FLAG_WC;
 401		man->default_caching = TTM_PL_FLAG_WC;
 402		break;
 403	case TTM_PL_TT:
 404		if (dev_priv->card_type >= NV_50)
 405			man->func = &nouveau_gart_manager;
 406		else
 
 
 
 407			man->func = &ttm_bo_manager_func;
 408		switch (dev_priv->gart_info.type) {
 409		case NOUVEAU_GART_AGP:
 410			man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
 411			man->available_caching = TTM_PL_FLAG_UNCACHED |
 412				TTM_PL_FLAG_WC;
 413			man->default_caching = TTM_PL_FLAG_WC;
 414			break;
 415		case NOUVEAU_GART_PDMA:
 416		case NOUVEAU_GART_HW:
 417			man->flags = TTM_MEMTYPE_FLAG_MAPPABLE |
 418				     TTM_MEMTYPE_FLAG_CMA;
 419			man->available_caching = TTM_PL_MASK_CACHING;
 420			man->default_caching = TTM_PL_FLAG_CACHED;
 421			break;
 422		default:
 423			NV_ERROR(dev, "Unknown GART type: %d\n",
 424				 dev_priv->gart_info.type);
 425			return -EINVAL;
 426		}
 
 427		break;
 428	default:
 429		NV_ERROR(dev, "Unsupported memory type %u\n", (unsigned)type);
 430		return -EINVAL;
 431	}
 432	return 0;
 433}
 434
 435static void
 436nouveau_bo_evict_flags(struct ttm_buffer_object *bo, struct ttm_placement *pl)
 437{
 438	struct nouveau_bo *nvbo = nouveau_bo(bo);
 439
 440	switch (bo->mem.mem_type) {
 441	case TTM_PL_VRAM:
 442		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_TT,
 443					 TTM_PL_FLAG_SYSTEM);
 444		break;
 445	default:
 446		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_SYSTEM, 0);
 447		break;
 448	}
 449
 450	*pl = nvbo->placement;
 451}
 452
 453
 454/* GPU-assisted copy using NV_MEMORY_TO_MEMORY_FORMAT, can access
 455 * TTM_PL_{VRAM,TT} directly.
 456 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 457
 458static int
 459nouveau_bo_move_accel_cleanup(struct nouveau_channel *chan,
 460			      struct nouveau_bo *nvbo, bool evict,
 461			      bool no_wait_reserve, bool no_wait_gpu,
 462			      struct ttm_mem_reg *new_mem)
 463{
 464	struct nouveau_fence *fence = NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 465	int ret;
 466
 467	ret = nouveau_fence_new(chan, &fence, true);
 468	if (ret)
 469		return ret;
 470
 471	ret = ttm_bo_move_accel_cleanup(&nvbo->bo, fence, NULL, evict,
 472					no_wait_reserve, no_wait_gpu, new_mem);
 473	nouveau_fence_unref(&fence);
 474	return ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 475}
 476
 477static int
 478nvc0_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
 479		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
 480{
 481	struct nouveau_mem *node = old_mem->mm_node;
 482	u64 src_offset = node->vma[0].offset;
 483	u64 dst_offset = node->vma[1].offset;
 484	u32 page_count = new_mem->num_pages;
 485	int ret;
 486
 487	page_count = new_mem->num_pages;
 488	while (page_count) {
 489		int line_count = (page_count > 2047) ? 2047 : page_count;
 490
 491		ret = RING_SPACE(chan, 12);
 492		if (ret)
 493			return ret;
 494
 495		BEGIN_NVC0(chan, 2, NvSubM2MF, 0x0238, 2);
 496		OUT_RING  (chan, upper_32_bits(dst_offset));
 497		OUT_RING  (chan, lower_32_bits(dst_offset));
 498		BEGIN_NVC0(chan, 2, NvSubM2MF, 0x030c, 6);
 499		OUT_RING  (chan, upper_32_bits(src_offset));
 500		OUT_RING  (chan, lower_32_bits(src_offset));
 501		OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
 502		OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
 503		OUT_RING  (chan, PAGE_SIZE); /* line_length */
 504		OUT_RING  (chan, line_count);
 505		BEGIN_NVC0(chan, 2, NvSubM2MF, 0x0300, 1);
 506		OUT_RING  (chan, 0x00100110);
 507
 508		page_count -= line_count;
 509		src_offset += (PAGE_SIZE * line_count);
 510		dst_offset += (PAGE_SIZE * line_count);
 511	}
 512
 513	return 0;
 514}
 515
 516static int
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 517nv50_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
 518		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
 519{
 520	struct nouveau_mem *node = old_mem->mm_node;
 521	struct nouveau_bo *nvbo = nouveau_bo(bo);
 522	u64 length = (new_mem->num_pages << PAGE_SHIFT);
 523	u64 src_offset = node->vma[0].offset;
 524	u64 dst_offset = node->vma[1].offset;
 
 
 525	int ret;
 526
 527	while (length) {
 528		u32 amount, stride, height;
 529
 
 
 
 
 530		amount  = min(length, (u64)(4 * 1024 * 1024));
 531		stride  = 16 * 4;
 532		height  = amount / stride;
 533
 534		if (new_mem->mem_type == TTM_PL_VRAM &&
 535		    nouveau_bo_tile_layout(nvbo)) {
 536			ret = RING_SPACE(chan, 8);
 537			if (ret)
 538				return ret;
 539
 540			BEGIN_RING(chan, NvSubM2MF, 0x0200, 7);
 541			OUT_RING  (chan, 0);
 542			OUT_RING  (chan, 0);
 543			OUT_RING  (chan, stride);
 544			OUT_RING  (chan, height);
 545			OUT_RING  (chan, 1);
 546			OUT_RING  (chan, 0);
 547			OUT_RING  (chan, 0);
 548		} else {
 549			ret = RING_SPACE(chan, 2);
 550			if (ret)
 551				return ret;
 552
 553			BEGIN_RING(chan, NvSubM2MF, 0x0200, 1);
 554			OUT_RING  (chan, 1);
 555		}
 556		if (old_mem->mem_type == TTM_PL_VRAM &&
 557		    nouveau_bo_tile_layout(nvbo)) {
 558			ret = RING_SPACE(chan, 8);
 559			if (ret)
 560				return ret;
 561
 562			BEGIN_RING(chan, NvSubM2MF, 0x021c, 7);
 563			OUT_RING  (chan, 0);
 564			OUT_RING  (chan, 0);
 565			OUT_RING  (chan, stride);
 566			OUT_RING  (chan, height);
 567			OUT_RING  (chan, 1);
 568			OUT_RING  (chan, 0);
 569			OUT_RING  (chan, 0);
 570		} else {
 571			ret = RING_SPACE(chan, 2);
 572			if (ret)
 573				return ret;
 574
 575			BEGIN_RING(chan, NvSubM2MF, 0x021c, 1);
 576			OUT_RING  (chan, 1);
 577		}
 578
 579		ret = RING_SPACE(chan, 14);
 580		if (ret)
 581			return ret;
 582
 583		BEGIN_RING(chan, NvSubM2MF, 0x0238, 2);
 584		OUT_RING  (chan, upper_32_bits(src_offset));
 585		OUT_RING  (chan, upper_32_bits(dst_offset));
 586		BEGIN_RING(chan, NvSubM2MF, 0x030c, 8);
 587		OUT_RING  (chan, lower_32_bits(src_offset));
 588		OUT_RING  (chan, lower_32_bits(dst_offset));
 589		OUT_RING  (chan, stride);
 590		OUT_RING  (chan, stride);
 591		OUT_RING  (chan, stride);
 592		OUT_RING  (chan, height);
 593		OUT_RING  (chan, 0x00000101);
 594		OUT_RING  (chan, 0x00000000);
 595		BEGIN_RING(chan, NvSubM2MF, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
 596		OUT_RING  (chan, 0);
 597
 598		length -= amount;
 599		src_offset += amount;
 600		dst_offset += amount;
 601	}
 602
 603	return 0;
 604}
 605
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 606static inline uint32_t
 607nouveau_bo_mem_ctxdma(struct ttm_buffer_object *bo,
 608		      struct nouveau_channel *chan, struct ttm_mem_reg *mem)
 609{
 610	if (mem->mem_type == TTM_PL_TT)
 611		return chan->gart_handle;
 612	return chan->vram_handle;
 613}
 614
 615static int
 616nv04_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
 617		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
 618{
 619	u32 src_offset = old_mem->start << PAGE_SHIFT;
 620	u32 dst_offset = new_mem->start << PAGE_SHIFT;
 621	u32 page_count = new_mem->num_pages;
 622	int ret;
 623
 624	ret = RING_SPACE(chan, 3);
 625	if (ret)
 626		return ret;
 627
 628	BEGIN_RING(chan, NvSubM2MF, NV_MEMORY_TO_MEMORY_FORMAT_DMA_SOURCE, 2);
 629	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, old_mem));
 630	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, new_mem));
 631
 632	page_count = new_mem->num_pages;
 633	while (page_count) {
 634		int line_count = (page_count > 2047) ? 2047 : page_count;
 635
 636		ret = RING_SPACE(chan, 11);
 637		if (ret)
 638			return ret;
 639
 640		BEGIN_RING(chan, NvSubM2MF,
 641				 NV_MEMORY_TO_MEMORY_FORMAT_OFFSET_IN, 8);
 642		OUT_RING  (chan, src_offset);
 643		OUT_RING  (chan, dst_offset);
 644		OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
 645		OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
 646		OUT_RING  (chan, PAGE_SIZE); /* line_length */
 647		OUT_RING  (chan, line_count);
 648		OUT_RING  (chan, 0x00000101);
 649		OUT_RING  (chan, 0x00000000);
 650		BEGIN_RING(chan, NvSubM2MF, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
 651		OUT_RING  (chan, 0);
 652
 653		page_count -= line_count;
 654		src_offset += (PAGE_SIZE * line_count);
 655		dst_offset += (PAGE_SIZE * line_count);
 656	}
 657
 658	return 0;
 659}
 660
 661static int
 662nouveau_vma_getmap(struct nouveau_channel *chan, struct nouveau_bo *nvbo,
 663		   struct ttm_mem_reg *mem, struct nouveau_vma *vma)
 664{
 665	struct nouveau_mem *node = mem->mm_node;
 
 
 666	int ret;
 667
 668	ret = nouveau_vm_get(chan->vm, mem->num_pages << PAGE_SHIFT,
 669			     node->page_shift, NV_MEM_ACCESS_RO, vma);
 670	if (ret)
 671		return ret;
 672
 673	if (mem->mem_type == TTM_PL_VRAM)
 674		nouveau_vm_map(vma, node);
 675	else
 676		nouveau_vm_map_sg(vma, 0, mem->num_pages << PAGE_SHIFT,
 677				  node, node->pages);
 
 678
 
 
 679	return 0;
 680}
 681
 682static int
 683nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict, bool intr,
 684		     bool no_wait_reserve, bool no_wait_gpu,
 685		     struct ttm_mem_reg *new_mem)
 686{
 687	struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
 688	struct nouveau_bo *nvbo = nouveau_bo(bo);
 689	struct ttm_mem_reg *old_mem = &bo->mem;
 690	struct nouveau_channel *chan;
 691	int ret;
 692
 693	chan = nvbo->channel;
 694	if (!chan) {
 695		chan = dev_priv->channel;
 696		mutex_lock_nested(&chan->mutex, NOUVEAU_KCHANNEL_MUTEX);
 697	}
 698
 699	/* create temporary vmas for the transfer and attach them to the
 700	 * old nouveau_mem node, these will get cleaned up after ttm has
 701	 * destroyed the ttm_mem_reg
 702	 */
 703	if (dev_priv->card_type >= NV_50) {
 704		struct nouveau_mem *node = old_mem->mm_node;
 705
 706		ret = nouveau_vma_getmap(chan, nvbo, old_mem, &node->vma[0]);
 707		if (ret)
 708			goto out;
 709
 710		ret = nouveau_vma_getmap(chan, nvbo, new_mem, &node->vma[1]);
 711		if (ret)
 712			goto out;
 713	}
 714
 715	if (dev_priv->card_type < NV_50)
 716		ret = nv04_bo_move_m2mf(chan, bo, &bo->mem, new_mem);
 717	else
 718	if (dev_priv->card_type < NV_C0)
 719		ret = nv50_bo_move_m2mf(chan, bo, &bo->mem, new_mem);
 720	else
 721		ret = nvc0_bo_move_m2mf(chan, bo, &bo->mem, new_mem);
 722	if (ret == 0) {
 723		ret = nouveau_bo_move_accel_cleanup(chan, nvbo, evict,
 724						    no_wait_reserve,
 725						    no_wait_gpu, new_mem);
 
 
 
 
 
 
 
 
 
 726	}
 727
 728out:
 729	if (chan == dev_priv->channel)
 730		mutex_unlock(&chan->mutex);
 731	return ret;
 732}
 733
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 734static int
 735nouveau_bo_move_flipd(struct ttm_buffer_object *bo, bool evict, bool intr,
 736		      bool no_wait_reserve, bool no_wait_gpu,
 737		      struct ttm_mem_reg *new_mem)
 738{
 739	u32 placement_memtype = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING;
 
 
 
 
 740	struct ttm_placement placement;
 741	struct ttm_mem_reg tmp_mem;
 742	int ret;
 743
 744	placement.fpfn = placement.lpfn = 0;
 745	placement.num_placement = placement.num_busy_placement = 1;
 746	placement.placement = placement.busy_placement = &placement_memtype;
 747
 748	tmp_mem = *new_mem;
 749	tmp_mem.mm_node = NULL;
 750	ret = ttm_bo_mem_space(bo, &placement, &tmp_mem, intr, no_wait_reserve, no_wait_gpu);
 751	if (ret)
 752		return ret;
 753
 754	ret = ttm_tt_bind(bo->ttm, &tmp_mem);
 755	if (ret)
 756		goto out;
 757
 758	ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_reserve, no_wait_gpu, &tmp_mem);
 759	if (ret)
 760		goto out;
 761
 762	ret = ttm_bo_move_ttm(bo, true, no_wait_reserve, no_wait_gpu, new_mem);
 763out:
 764	ttm_bo_mem_put(bo, &tmp_mem);
 765	return ret;
 766}
 767
 768static int
 769nouveau_bo_move_flips(struct ttm_buffer_object *bo, bool evict, bool intr,
 770		      bool no_wait_reserve, bool no_wait_gpu,
 771		      struct ttm_mem_reg *new_mem)
 772{
 773	u32 placement_memtype = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING;
 
 
 
 
 774	struct ttm_placement placement;
 775	struct ttm_mem_reg tmp_mem;
 776	int ret;
 777
 778	placement.fpfn = placement.lpfn = 0;
 779	placement.num_placement = placement.num_busy_placement = 1;
 780	placement.placement = placement.busy_placement = &placement_memtype;
 781
 782	tmp_mem = *new_mem;
 783	tmp_mem.mm_node = NULL;
 784	ret = ttm_bo_mem_space(bo, &placement, &tmp_mem, intr, no_wait_reserve, no_wait_gpu);
 785	if (ret)
 786		return ret;
 787
 788	ret = ttm_bo_move_ttm(bo, true, no_wait_reserve, no_wait_gpu, &tmp_mem);
 789	if (ret)
 790		goto out;
 791
 792	ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_reserve, no_wait_gpu, new_mem);
 793	if (ret)
 794		goto out;
 795
 796out:
 797	ttm_bo_mem_put(bo, &tmp_mem);
 798	return ret;
 799}
 800
 801static void
 802nouveau_bo_move_ntfy(struct ttm_buffer_object *bo, struct ttm_mem_reg *new_mem)
 803{
 804	struct nouveau_mem *node = new_mem->mm_node;
 805	struct nouveau_bo *nvbo = nouveau_bo(bo);
 806	struct nouveau_vma *vma;
 
 
 
 
 807
 808	list_for_each_entry(vma, &nvbo->vma_list, head) {
 809		if (new_mem->mem_type == TTM_PL_VRAM) {
 810			nouveau_vm_map(vma, new_mem->mm_node);
 811		} else
 812		if (new_mem->mem_type == TTM_PL_TT &&
 813		    nvbo->page_shift == vma->vm->spg_shift) {
 814			nouveau_vm_map_sg(vma, 0, new_mem->
 815					  num_pages << PAGE_SHIFT,
 816					  node, node->pages);
 817		} else {
 818			nouveau_vm_unmap(vma);
 819		}
 820	}
 821}
 822
 823static int
 824nouveau_bo_vm_bind(struct ttm_buffer_object *bo, struct ttm_mem_reg *new_mem,
 825		   struct nouveau_tile_reg **new_tile)
 826{
 827	struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
 828	struct drm_device *dev = dev_priv->dev;
 829	struct nouveau_bo *nvbo = nouveau_bo(bo);
 830	u64 offset = new_mem->start << PAGE_SHIFT;
 831
 832	*new_tile = NULL;
 833	if (new_mem->mem_type != TTM_PL_VRAM)
 834		return 0;
 835
 836	if (dev_priv->card_type >= NV_10) {
 837		*new_tile = nv10_mem_set_tiling(dev, offset, new_mem->size,
 838						nvbo->tile_mode,
 839						nvbo->tile_flags);
 840	}
 841
 842	return 0;
 843}
 844
 845static void
 846nouveau_bo_vm_cleanup(struct ttm_buffer_object *bo,
 847		      struct nouveau_tile_reg *new_tile,
 848		      struct nouveau_tile_reg **old_tile)
 849{
 850	struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
 851	struct drm_device *dev = dev_priv->dev;
 
 852
 853	nv10_mem_put_tile_region(dev, *old_tile, bo->sync_obj);
 854	*old_tile = new_tile;
 855}
 856
 857static int
 858nouveau_bo_move(struct ttm_buffer_object *bo, bool evict, bool intr,
 859		bool no_wait_reserve, bool no_wait_gpu,
 860		struct ttm_mem_reg *new_mem)
 861{
 862	struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
 863	struct nouveau_bo *nvbo = nouveau_bo(bo);
 864	struct ttm_mem_reg *old_mem = &bo->mem;
 865	struct nouveau_tile_reg *new_tile = NULL;
 866	int ret = 0;
 867
 868	if (dev_priv->card_type < NV_50) {
 
 
 
 869		ret = nouveau_bo_vm_bind(bo, new_mem, &new_tile);
 870		if (ret)
 871			return ret;
 872	}
 873
 874	/* Fake bo copy. */
 875	if (old_mem->mem_type == TTM_PL_SYSTEM && !bo->ttm) {
 876		BUG_ON(bo->mem.mm_node != NULL);
 877		bo->mem = *new_mem;
 878		new_mem->mm_node = NULL;
 879		goto out;
 880	}
 881
 882	/* Software copy if the card isn't up and running yet. */
 883	if (!dev_priv->channel) {
 884		ret = ttm_bo_move_memcpy(bo, evict, no_wait_reserve, no_wait_gpu, new_mem);
 885		goto out;
 886	}
 887
 888	/* Hardware assisted copy. */
 889	if (new_mem->mem_type == TTM_PL_SYSTEM)
 890		ret = nouveau_bo_move_flipd(bo, evict, intr, no_wait_reserve, no_wait_gpu, new_mem);
 891	else if (old_mem->mem_type == TTM_PL_SYSTEM)
 892		ret = nouveau_bo_move_flips(bo, evict, intr, no_wait_reserve, no_wait_gpu, new_mem);
 893	else
 894		ret = nouveau_bo_move_m2mf(bo, evict, intr, no_wait_reserve, no_wait_gpu, new_mem);
 895
 896	if (!ret)
 897		goto out;
 
 
 
 
 898
 899	/* Fallback to software copy. */
 900	ret = ttm_bo_move_memcpy(bo, evict, no_wait_reserve, no_wait_gpu, new_mem);
 
 
 901
 902out:
 903	if (dev_priv->card_type < NV_50) {
 904		if (ret)
 905			nouveau_bo_vm_cleanup(bo, NULL, &new_tile);
 906		else
 907			nouveau_bo_vm_cleanup(bo, new_tile, &nvbo->tile);
 908	}
 909
 910	return ret;
 911}
 912
 913static int
 914nouveau_bo_verify_access(struct ttm_buffer_object *bo, struct file *filp)
 915{
 916	return 0;
 
 
 917}
 918
 919static int
 920nouveau_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
 921{
 922	struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
 923	struct drm_nouveau_private *dev_priv = nouveau_bdev(bdev);
 924	struct drm_device *dev = dev_priv->dev;
 
 925	int ret;
 926
 927	mem->bus.addr = NULL;
 928	mem->bus.offset = 0;
 929	mem->bus.size = mem->num_pages << PAGE_SHIFT;
 930	mem->bus.base = 0;
 931	mem->bus.is_iomem = false;
 932	if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
 933		return -EINVAL;
 934	switch (mem->mem_type) {
 935	case TTM_PL_SYSTEM:
 936		/* System memory */
 937		return 0;
 938	case TTM_PL_TT:
 939#if __OS_HAS_AGP
 940		if (dev_priv->gart_info.type == NOUVEAU_GART_AGP) {
 941			mem->bus.offset = mem->start << PAGE_SHIFT;
 942			mem->bus.base = dev_priv->gart_info.aper_base;
 943			mem->bus.is_iomem = true;
 944		}
 945#endif
 946		break;
 947	case TTM_PL_VRAM:
 948	{
 949		struct nouveau_mem *node = mem->mm_node;
 950		u8 page_shift;
 951
 952		if (!dev_priv->bar1_vm) {
 953			mem->bus.offset = mem->start << PAGE_SHIFT;
 954			mem->bus.base = pci_resource_start(dev->pdev, 1);
 955			mem->bus.is_iomem = true;
 956			break;
 957		}
 958
 959		if (dev_priv->card_type == NV_C0)
 960			page_shift = node->page_shift;
 961		else
 962			page_shift = 12;
 
 
 
 
 963
 964		ret = nouveau_vm_get(dev_priv->bar1_vm, mem->bus.size,
 965				     page_shift, NV_MEM_ACCESS_RW,
 966				     &node->bar_vma);
 967		if (ret)
 968			return ret;
 969
 970		nouveau_vm_map(&node->bar_vma, node);
 971		if (ret) {
 972			nouveau_vm_put(&node->bar_vma);
 973			return ret;
 974		}
 975
 976		mem->bus.offset = node->bar_vma.offset;
 977		if (dev_priv->card_type == NV_50) /*XXX*/
 978			mem->bus.offset -= 0x0020000000ULL;
 979		mem->bus.base = pci_resource_start(dev->pdev, 1);
 980		mem->bus.is_iomem = true;
 981	}
 982		break;
 983	default:
 984		return -EINVAL;
 985	}
 986	return 0;
 987}
 988
 989static void
 990nouveau_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
 991{
 992	struct drm_nouveau_private *dev_priv = nouveau_bdev(bdev);
 993	struct nouveau_mem *node = mem->mm_node;
 994
 995	if (!dev_priv->bar1_vm || mem->mem_type != TTM_PL_VRAM)
 996		return;
 997
 998	if (!node->bar_vma.node)
 999		return;
1000
1001	nouveau_vm_unmap(&node->bar_vma);
1002	nouveau_vm_put(&node->bar_vma);
1003}
1004
1005static int
1006nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
1007{
1008	struct drm_nouveau_private *dev_priv = nouveau_bdev(bo->bdev);
1009	struct nouveau_bo *nvbo = nouveau_bo(bo);
 
 
 
1010
1011	/* as long as the bo isn't in vram, and isn't tiled, we've got
1012	 * nothing to do here.
1013	 */
1014	if (bo->mem.mem_type != TTM_PL_VRAM) {
1015		if (dev_priv->card_type < NV_50 ||
1016		    !nouveau_bo_tile_layout(nvbo))
1017			return 0;
 
 
 
 
 
 
 
 
 
1018	}
1019
1020	/* make sure bo is in mappable vram */
1021	if (bo->mem.start + bo->mem.num_pages < dev_priv->fb_mappable_pages)
 
1022		return 0;
1023
 
 
 
 
 
 
 
 
 
1024
1025	nvbo->placement.fpfn = 0;
1026	nvbo->placement.lpfn = dev_priv->fb_mappable_pages;
1027	nouveau_bo_placement_set(nvbo, TTM_PL_VRAM, 0);
1028	return nouveau_bo_validate(nvbo, false, true, false);
1029}
1030
1031void
1032nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1033{
1034	struct nouveau_fence *old_fence;
 
 
 
 
 
 
1035
1036	if (likely(fence))
1037		nouveau_fence_ref(fence);
1038
1039	spin_lock(&nvbo->bo.bdev->fence_lock);
1040	old_fence = nvbo->bo.sync_obj;
1041	nvbo->bo.sync_obj = fence;
1042	spin_unlock(&nvbo->bo.bdev->fence_lock);
 
 
 
 
 
 
 
 
 
 
1043
1044	nouveau_fence_unref(&old_fence);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1045}
1046
1047struct ttm_bo_driver nouveau_bo_driver = {
1048	.create_ttm_backend_entry = nouveau_bo_create_ttm_backend_entry,
 
 
1049	.invalidate_caches = nouveau_bo_invalidate_caches,
1050	.init_mem_type = nouveau_bo_init_mem_type,
1051	.evict_flags = nouveau_bo_evict_flags,
1052	.move_notify = nouveau_bo_move_ntfy,
1053	.move = nouveau_bo_move,
1054	.verify_access = nouveau_bo_verify_access,
1055	.sync_obj_signaled = __nouveau_fence_signalled,
1056	.sync_obj_wait = __nouveau_fence_wait,
1057	.sync_obj_flush = __nouveau_fence_flush,
1058	.sync_obj_unref = __nouveau_fence_unref,
1059	.sync_obj_ref = __nouveau_fence_ref,
1060	.fault_reserve_notify = &nouveau_ttm_fault_reserve_notify,
1061	.io_mem_reserve = &nouveau_ttm_io_mem_reserve,
1062	.io_mem_free = &nouveau_ttm_io_mem_free,
1063};
1064
1065struct nouveau_vma *
1066nouveau_bo_vma_find(struct nouveau_bo *nvbo, struct nouveau_vm *vm)
1067{
1068	struct nouveau_vma *vma;
1069	list_for_each_entry(vma, &nvbo->vma_list, head) {
1070		if (vma->vm == vm)
1071			return vma;
1072	}
1073
1074	return NULL;
1075}
1076
1077int
1078nouveau_bo_vma_add(struct nouveau_bo *nvbo, struct nouveau_vm *vm,
1079		   struct nouveau_vma *vma)
1080{
1081	const u32 size = nvbo->bo.mem.num_pages << PAGE_SHIFT;
1082	struct nouveau_mem *node = nvbo->bo.mem.mm_node;
1083	int ret;
1084
1085	ret = nouveau_vm_get(vm, size, nvbo->page_shift,
1086			     NV_MEM_ACCESS_RW, vma);
1087	if (ret)
1088		return ret;
1089
1090	if (nvbo->bo.mem.mem_type == TTM_PL_VRAM)
1091		nouveau_vm_map(vma, nvbo->bo.mem.mm_node);
1092	else
1093	if (nvbo->bo.mem.mem_type == TTM_PL_TT)
1094		nouveau_vm_map_sg(vma, 0, size, node, node->pages);
1095
1096	list_add_tail(&vma->head, &nvbo->vma_list);
1097	vma->refcount = 1;
1098	return 0;
1099}
1100
1101void
1102nouveau_bo_vma_del(struct nouveau_bo *nvbo, struct nouveau_vma *vma)
1103{
1104	if (vma->node) {
1105		if (nvbo->bo.mem.mem_type != TTM_PL_SYSTEM) {
1106			spin_lock(&nvbo->bo.bdev->fence_lock);
1107			ttm_bo_wait(&nvbo->bo, false, false, false);
1108			spin_unlock(&nvbo->bo.bdev->fence_lock);
1109			nouveau_vm_unmap(vma);
1110		}
1111
1112		nouveau_vm_put(vma);
1113		list_del(&vma->head);
1114	}
1115}
v4.6
   1/*
   2 * Copyright 2007 Dave Airlied
   3 * All Rights Reserved.
   4 *
   5 * Permission is hereby granted, free of charge, to any person obtaining a
   6 * copy of this software and associated documentation files (the "Software"),
   7 * to deal in the Software without restriction, including without limitation
   8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9 * and/or sell copies of the Software, and to permit persons to whom the
  10 * Software is furnished to do so, subject to the following conditions:
  11 *
  12 * The above copyright notice and this permission notice (including the next
  13 * paragraph) shall be included in all copies or substantial portions of the
  14 * Software.
  15 *
  16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19 * VA LINUX SYSTEMS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
  20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  22 * OTHER DEALINGS IN THE SOFTWARE.
  23 */
  24/*
  25 * Authors: Dave Airlied <airlied@linux.ie>
  26 *	    Ben Skeggs   <darktama@iinet.net.au>
  27 *	    Jeremy Kolb  <jkolb@brandeis.edu>
  28 */
  29
  30#include <linux/dma-mapping.h>
  31#include <linux/swiotlb.h>
  32
  33#include "nouveau_drm.h"
 
  34#include "nouveau_dma.h"
  35#include "nouveau_fence.h"
 
  36
  37#include "nouveau_bo.h"
  38#include "nouveau_ttm.h"
  39#include "nouveau_gem.h"
  40
  41/*
  42 * NV10-NV40 tiling helpers
  43 */
  44
  45static void
  46nv10_bo_update_tile_region(struct drm_device *dev, struct nouveau_drm_tile *reg,
  47			   u32 addr, u32 size, u32 pitch, u32 flags)
  48{
  49	struct nouveau_drm *drm = nouveau_drm(dev);
  50	int i = reg - drm->tile.reg;
  51	struct nvkm_device *device = nvxx_device(&drm->device);
  52	struct nvkm_fb *fb = device->fb;
  53	struct nvkm_fb_tile *tile = &fb->tile.region[i];
  54
  55	nouveau_fence_unref(&reg->fence);
  56
  57	if (tile->pitch)
  58		nvkm_fb_tile_fini(fb, i, tile);
  59
  60	if (pitch)
  61		nvkm_fb_tile_init(fb, i, addr, size, pitch, flags, tile);
  62
  63	nvkm_fb_tile_prog(fb, i, tile);
  64}
  65
  66static struct nouveau_drm_tile *
  67nv10_bo_get_tile_region(struct drm_device *dev, int i)
  68{
  69	struct nouveau_drm *drm = nouveau_drm(dev);
  70	struct nouveau_drm_tile *tile = &drm->tile.reg[i];
  71
  72	spin_lock(&drm->tile.lock);
  73
  74	if (!tile->used &&
  75	    (!tile->fence || nouveau_fence_done(tile->fence)))
  76		tile->used = true;
  77	else
  78		tile = NULL;
  79
  80	spin_unlock(&drm->tile.lock);
  81	return tile;
  82}
  83
  84static void
  85nv10_bo_put_tile_region(struct drm_device *dev, struct nouveau_drm_tile *tile,
  86			struct fence *fence)
  87{
  88	struct nouveau_drm *drm = nouveau_drm(dev);
  89
  90	if (tile) {
  91		spin_lock(&drm->tile.lock);
  92		tile->fence = (struct nouveau_fence *)fence_get(fence);
  93		tile->used = false;
  94		spin_unlock(&drm->tile.lock);
  95	}
  96}
  97
  98static struct nouveau_drm_tile *
  99nv10_bo_set_tiling(struct drm_device *dev, u32 addr,
 100		   u32 size, u32 pitch, u32 flags)
 101{
 102	struct nouveau_drm *drm = nouveau_drm(dev);
 103	struct nvkm_fb *fb = nvxx_fb(&drm->device);
 104	struct nouveau_drm_tile *tile, *found = NULL;
 105	int i;
 106
 107	for (i = 0; i < fb->tile.regions; i++) {
 108		tile = nv10_bo_get_tile_region(dev, i);
 109
 110		if (pitch && !found) {
 111			found = tile;
 112			continue;
 113
 114		} else if (tile && fb->tile.region[i].pitch) {
 115			/* Kill an unused tile region. */
 116			nv10_bo_update_tile_region(dev, tile, 0, 0, 0, 0);
 117		}
 118
 119		nv10_bo_put_tile_region(dev, tile, NULL);
 120	}
 121
 122	if (found)
 123		nv10_bo_update_tile_region(dev, found, addr, size,
 124					    pitch, flags);
 125	return found;
 126}
 127
 128static void
 129nouveau_bo_del_ttm(struct ttm_buffer_object *bo)
 130{
 131	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
 132	struct drm_device *dev = drm->dev;
 133	struct nouveau_bo *nvbo = nouveau_bo(bo);
 134
 135	if (unlikely(nvbo->gem.filp))
 136		DRM_ERROR("bo %p still attached to GEM object\n", bo);
 137	WARN_ON(nvbo->pin_refcnt > 0);
 138	nv10_bo_put_tile_region(dev, nvbo->tile, NULL);
 139	kfree(nvbo);
 140}
 141
 142static void
 143nouveau_bo_fixup_align(struct nouveau_bo *nvbo, u32 flags,
 144		       int *align, int *size)
 145{
 146	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
 147	struct nvif_device *device = &drm->device;
 148
 149	if (device->info.family < NV_DEVICE_INFO_V0_TESLA) {
 150		if (nvbo->tile_mode) {
 151			if (device->info.chipset >= 0x40) {
 152				*align = 65536;
 153				*size = roundup(*size, 64 * nvbo->tile_mode);
 154
 155			} else if (device->info.chipset >= 0x30) {
 156				*align = 32768;
 157				*size = roundup(*size, 64 * nvbo->tile_mode);
 158
 159			} else if (device->info.chipset >= 0x20) {
 160				*align = 16384;
 161				*size = roundup(*size, 64 * nvbo->tile_mode);
 162
 163			} else if (device->info.chipset >= 0x10) {
 164				*align = 16384;
 165				*size = roundup(*size, 32 * nvbo->tile_mode);
 166			}
 167		}
 168	} else {
 169		*size = roundup(*size, (1 << nvbo->page_shift));
 170		*align = max((1 <<  nvbo->page_shift), *align);
 171	}
 172
 173	*size = roundup(*size, PAGE_SIZE);
 174}
 175
 176int
 177nouveau_bo_new(struct drm_device *dev, int size, int align,
 178	       uint32_t flags, uint32_t tile_mode, uint32_t tile_flags,
 179	       struct sg_table *sg, struct reservation_object *robj,
 180	       struct nouveau_bo **pnvbo)
 181{
 182	struct nouveau_drm *drm = nouveau_drm(dev);
 183	struct nouveau_bo *nvbo;
 184	size_t acc_size;
 185	int ret;
 186	int type = ttm_bo_type_device;
 187	int lpg_shift = 12;
 188	int max_size;
 189
 190	if (drm->client.vm)
 191		lpg_shift = drm->client.vm->mmu->lpg_shift;
 192	max_size = INT_MAX & ~((1 << lpg_shift) - 1);
 193
 194	if (size <= 0 || size > max_size) {
 195		NV_WARN(drm, "skipped size %x\n", (u32)size);
 196		return -EINVAL;
 197	}
 198
 199	if (sg)
 200		type = ttm_bo_type_sg;
 201
 202	nvbo = kzalloc(sizeof(struct nouveau_bo), GFP_KERNEL);
 203	if (!nvbo)
 204		return -ENOMEM;
 205	INIT_LIST_HEAD(&nvbo->head);
 206	INIT_LIST_HEAD(&nvbo->entry);
 207	INIT_LIST_HEAD(&nvbo->vma_list);
 208	nvbo->tile_mode = tile_mode;
 209	nvbo->tile_flags = tile_flags;
 210	nvbo->bo.bdev = &drm->ttm.bdev;
 211
 212	if (!nvxx_device(&drm->device)->func->cpu_coherent)
 213		nvbo->force_coherent = flags & TTM_PL_FLAG_UNCACHED;
 214
 215	nvbo->page_shift = 12;
 216	if (drm->client.vm) {
 217		if (!(flags & TTM_PL_FLAG_TT) && size > 256 * 1024)
 218			nvbo->page_shift = drm->client.vm->mmu->lpg_shift;
 219	}
 220
 221	nouveau_bo_fixup_align(nvbo, flags, &align, &size);
 222	nvbo->bo.mem.num_pages = size >> PAGE_SHIFT;
 223	nouveau_bo_placement_set(nvbo, flags, 0);
 224
 225	acc_size = ttm_bo_dma_acc_size(&drm->ttm.bdev, size,
 226				       sizeof(struct nouveau_bo));
 227
 228	ret = ttm_bo_init(&drm->ttm.bdev, &nvbo->bo, size,
 229			  type, &nvbo->placement,
 230			  align >> PAGE_SHIFT, false, NULL, acc_size, sg,
 231			  robj, nouveau_bo_del_ttm);
 232	if (ret) {
 233		/* ttm will call nouveau_bo_del_ttm if it fails.. */
 234		return ret;
 235	}
 236
 237	*pnvbo = nvbo;
 238	return 0;
 239}
 240
 241static void
 242set_placement_list(struct ttm_place *pl, unsigned *n, uint32_t type, uint32_t flags)
 243{
 244	*n = 0;
 245
 246	if (type & TTM_PL_FLAG_VRAM)
 247		pl[(*n)++].flags = TTM_PL_FLAG_VRAM | flags;
 248	if (type & TTM_PL_FLAG_TT)
 249		pl[(*n)++].flags = TTM_PL_FLAG_TT | flags;
 250	if (type & TTM_PL_FLAG_SYSTEM)
 251		pl[(*n)++].flags = TTM_PL_FLAG_SYSTEM | flags;
 252}
 253
 254static void
 255set_placement_range(struct nouveau_bo *nvbo, uint32_t type)
 256{
 257	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
 258	u32 vram_pages = drm->device.info.ram_size >> PAGE_SHIFT;
 259	unsigned i, fpfn, lpfn;
 260
 261	if (drm->device.info.family == NV_DEVICE_INFO_V0_CELSIUS &&
 262	    nvbo->tile_mode && (type & TTM_PL_FLAG_VRAM) &&
 263	    nvbo->bo.mem.num_pages < vram_pages / 4) {
 264		/*
 265		 * Make sure that the color and depth buffers are handled
 266		 * by independent memory controller units. Up to a 9x
 267		 * speed up when alpha-blending and depth-test are enabled
 268		 * at the same time.
 269		 */
 270		if (nvbo->tile_flags & NOUVEAU_GEM_TILE_ZETA) {
 271			fpfn = vram_pages / 2;
 272			lpfn = ~0;
 273		} else {
 274			fpfn = 0;
 275			lpfn = vram_pages / 2;
 276		}
 277		for (i = 0; i < nvbo->placement.num_placement; ++i) {
 278			nvbo->placements[i].fpfn = fpfn;
 279			nvbo->placements[i].lpfn = lpfn;
 280		}
 281		for (i = 0; i < nvbo->placement.num_busy_placement; ++i) {
 282			nvbo->busy_placements[i].fpfn = fpfn;
 283			nvbo->busy_placements[i].lpfn = lpfn;
 284		}
 285	}
 286}
 287
 288void
 289nouveau_bo_placement_set(struct nouveau_bo *nvbo, uint32_t type, uint32_t busy)
 290{
 291	struct ttm_placement *pl = &nvbo->placement;
 292	uint32_t flags = (nvbo->force_coherent ? TTM_PL_FLAG_UNCACHED :
 293						 TTM_PL_MASK_CACHING) |
 294			 (nvbo->pin_refcnt ? TTM_PL_FLAG_NO_EVICT : 0);
 295
 296	pl->placement = nvbo->placements;
 297	set_placement_list(nvbo->placements, &pl->num_placement,
 298			   type, flags);
 299
 300	pl->busy_placement = nvbo->busy_placements;
 301	set_placement_list(nvbo->busy_placements, &pl->num_busy_placement,
 302			   type | busy, flags);
 303
 304	set_placement_range(nvbo, type);
 305}
 306
 307int
 308nouveau_bo_pin(struct nouveau_bo *nvbo, uint32_t memtype, bool contig)
 309{
 310	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
 311	struct ttm_buffer_object *bo = &nvbo->bo;
 312	bool force = false, evict = false;
 313	int ret;
 314
 315	ret = ttm_bo_reserve(bo, false, false, false, NULL);
 316	if (ret)
 317		return ret;
 
 
 
 318
 319	if (drm->device.info.family >= NV_DEVICE_INFO_V0_TESLA &&
 320	    memtype == TTM_PL_FLAG_VRAM && contig) {
 321		if (nvbo->tile_flags & NOUVEAU_GEM_TILE_NONCONTIG) {
 322			if (bo->mem.mem_type == TTM_PL_VRAM) {
 323				struct nvkm_mem *mem = bo->mem.mm_node;
 324				if (!list_is_singular(&mem->regions))
 325					evict = true;
 326			}
 327			nvbo->tile_flags &= ~NOUVEAU_GEM_TILE_NONCONTIG;
 328			force = true;
 329		}
 330	}
 331
 332	if (nvbo->pin_refcnt) {
 333		if (!(memtype & (1 << bo->mem.mem_type)) || evict) {
 334			NV_ERROR(drm, "bo %p pinned elsewhere: "
 335				      "0x%08x vs 0x%08x\n", bo,
 336				 1 << bo->mem.mem_type, memtype);
 337			ret = -EBUSY;
 338		}
 339		nvbo->pin_refcnt++;
 340		goto out;
 341	}
 342
 343	if (evict) {
 344		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_TT, 0);
 345		ret = nouveau_bo_validate(nvbo, false, false);
 346		if (ret)
 347			goto out;
 348	}
 349
 350	nvbo->pin_refcnt++;
 351	nouveau_bo_placement_set(nvbo, memtype, 0);
 352
 353	/* drop pin_refcnt temporarily, so we don't trip the assertion
 354	 * in nouveau_bo_move() that makes sure we're not trying to
 355	 * move a pinned buffer
 356	 */
 357	nvbo->pin_refcnt--;
 358	ret = nouveau_bo_validate(nvbo, false, false);
 359	if (ret)
 360		goto out;
 361	nvbo->pin_refcnt++;
 362
 363	switch (bo->mem.mem_type) {
 364	case TTM_PL_VRAM:
 365		drm->gem.vram_available -= bo->mem.size;
 366		break;
 367	case TTM_PL_TT:
 368		drm->gem.gart_available -= bo->mem.size;
 369		break;
 370	default:
 371		break;
 372	}
 373
 374out:
 375	if (force && ret)
 376		nvbo->tile_flags |= NOUVEAU_GEM_TILE_NONCONTIG;
 377	ttm_bo_unreserve(bo);
 378	return ret;
 379}
 380
 381int
 382nouveau_bo_unpin(struct nouveau_bo *nvbo)
 383{
 384	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
 385	struct ttm_buffer_object *bo = &nvbo->bo;
 386	int ret, ref;
 387
 388	ret = ttm_bo_reserve(bo, false, false, false, NULL);
 
 
 
 389	if (ret)
 390		return ret;
 391
 392	ref = --nvbo->pin_refcnt;
 393	WARN_ON_ONCE(ref < 0);
 394	if (ref)
 395		goto out;
 396
 397	nouveau_bo_placement_set(nvbo, bo->mem.placement, 0);
 398
 399	ret = nouveau_bo_validate(nvbo, false, false);
 400	if (ret == 0) {
 401		switch (bo->mem.mem_type) {
 402		case TTM_PL_VRAM:
 403			drm->gem.vram_available += bo->mem.size;
 404			break;
 405		case TTM_PL_TT:
 406			drm->gem.gart_available += bo->mem.size;
 407			break;
 408		default:
 409			break;
 410		}
 411	}
 412
 413out:
 414	ttm_bo_unreserve(bo);
 415	return ret;
 416}
 417
 418int
 419nouveau_bo_map(struct nouveau_bo *nvbo)
 420{
 421	int ret;
 422
 423	ret = ttm_bo_reserve(&nvbo->bo, false, false, false, NULL);
 424	if (ret)
 425		return ret;
 426
 427	/*
 428	 * TTM buffers allocated using the DMA API already have a mapping, let's
 429	 * use it instead.
 430	 */
 431	if (!nvbo->force_coherent)
 432		ret = ttm_bo_kmap(&nvbo->bo, 0, nvbo->bo.mem.num_pages,
 433				  &nvbo->kmap);
 434
 435	ttm_bo_unreserve(&nvbo->bo);
 436	return ret;
 437}
 438
 439void
 440nouveau_bo_unmap(struct nouveau_bo *nvbo)
 441{
 442	if (!nvbo)
 443		return;
 444
 445	/*
 446	 * TTM buffers allocated using the DMA API already had a coherent
 447	 * mapping which we used, no need to unmap.
 448	 */
 449	if (!nvbo->force_coherent)
 450		ttm_bo_kunmap(&nvbo->kmap);
 451}
 452
 453void
 454nouveau_bo_sync_for_device(struct nouveau_bo *nvbo)
 455{
 456	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
 457	struct nvkm_device *device = nvxx_device(&drm->device);
 458	struct ttm_dma_tt *ttm_dma = (struct ttm_dma_tt *)nvbo->bo.ttm;
 459	int i;
 460
 461	if (!ttm_dma)
 462		return;
 463
 464	/* Don't waste time looping if the object is coherent */
 465	if (nvbo->force_coherent)
 466		return;
 467
 468	for (i = 0; i < ttm_dma->ttm.num_pages; i++)
 469		dma_sync_single_for_device(device->dev, ttm_dma->dma_address[i],
 470					   PAGE_SIZE, DMA_TO_DEVICE);
 471}
 472
 473void
 474nouveau_bo_sync_for_cpu(struct nouveau_bo *nvbo)
 475{
 476	struct nouveau_drm *drm = nouveau_bdev(nvbo->bo.bdev);
 477	struct nvkm_device *device = nvxx_device(&drm->device);
 478	struct ttm_dma_tt *ttm_dma = (struct ttm_dma_tt *)nvbo->bo.ttm;
 479	int i;
 480
 481	if (!ttm_dma)
 482		return;
 483
 484	/* Don't waste time looping if the object is coherent */
 485	if (nvbo->force_coherent)
 486		return;
 487
 488	for (i = 0; i < ttm_dma->ttm.num_pages; i++)
 489		dma_sync_single_for_cpu(device->dev, ttm_dma->dma_address[i],
 490					PAGE_SIZE, DMA_FROM_DEVICE);
 491}
 492
 493int
 494nouveau_bo_validate(struct nouveau_bo *nvbo, bool interruptible,
 495		    bool no_wait_gpu)
 496{
 497	int ret;
 498
 499	ret = ttm_bo_validate(&nvbo->bo, &nvbo->placement,
 500			      interruptible, no_wait_gpu);
 501	if (ret)
 502		return ret;
 503
 504	nouveau_bo_sync_for_device(nvbo);
 505
 506	return 0;
 507}
 508
 509static inline void *
 510_nouveau_bo_mem_index(struct nouveau_bo *nvbo, unsigned index, void *mem, u8 sz)
 511{
 512	struct ttm_dma_tt *dma_tt;
 513	u8 *m = mem;
 514
 515	index *= sz;
 516
 517	if (m) {
 518		/* kmap'd address, return the corresponding offset */
 519		m += index;
 520	} else {
 521		/* DMA-API mapping, lookup the right address */
 522		dma_tt = (struct ttm_dma_tt *)nvbo->bo.ttm;
 523		m = dma_tt->cpu_address[index / PAGE_SIZE];
 524		m += index % PAGE_SIZE;
 525	}
 526
 527	return m;
 528}
 529#define nouveau_bo_mem_index(o, i, m) _nouveau_bo_mem_index(o, i, m, sizeof(*m))
 530
 531void
 532nouveau_bo_wr16(struct nouveau_bo *nvbo, unsigned index, u16 val)
 533{
 534	bool is_iomem;
 535	u16 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
 536
 537	mem = nouveau_bo_mem_index(nvbo, index, mem);
 538
 539	if (is_iomem)
 540		iowrite16_native(val, (void __force __iomem *)mem);
 541	else
 542		*mem = val;
 543}
 544
 545u32
 546nouveau_bo_rd32(struct nouveau_bo *nvbo, unsigned index)
 547{
 548	bool is_iomem;
 549	u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
 550
 551	mem = nouveau_bo_mem_index(nvbo, index, mem);
 552
 553	if (is_iomem)
 554		return ioread32_native((void __force __iomem *)mem);
 555	else
 556		return *mem;
 557}
 558
 559void
 560nouveau_bo_wr32(struct nouveau_bo *nvbo, unsigned index, u32 val)
 561{
 562	bool is_iomem;
 563	u32 *mem = ttm_kmap_obj_virtual(&nvbo->kmap, &is_iomem);
 564
 565	mem = nouveau_bo_mem_index(nvbo, index, mem);
 566
 567	if (is_iomem)
 568		iowrite32_native(val, (void __force __iomem *)mem);
 569	else
 570		*mem = val;
 571}
 572
 573static struct ttm_tt *
 574nouveau_ttm_tt_create(struct ttm_bo_device *bdev, unsigned long size,
 575		      uint32_t page_flags, struct page *dummy_read)
 576{
 577#if IS_ENABLED(CONFIG_AGP)
 578	struct nouveau_drm *drm = nouveau_bdev(bdev);
 579
 580	if (drm->agp.bridge) {
 581		return ttm_agp_tt_create(bdev, drm->agp.bridge, size,
 582					 page_flags, dummy_read);
 
 
 
 
 
 
 
 
 583	}
 584#endif
 585
 586	return nouveau_sgdma_create_ttm(bdev, size, page_flags, dummy_read);
 587}
 588
 589static int
 590nouveau_bo_invalidate_caches(struct ttm_bo_device *bdev, uint32_t flags)
 591{
 592	/* We'll do this from user space. */
 593	return 0;
 594}
 595
 596static int
 597nouveau_bo_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
 598			 struct ttm_mem_type_manager *man)
 599{
 600	struct nouveau_drm *drm = nouveau_bdev(bdev);
 
 601
 602	switch (type) {
 603	case TTM_PL_SYSTEM:
 604		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
 605		man->available_caching = TTM_PL_MASK_CACHING;
 606		man->default_caching = TTM_PL_FLAG_CACHED;
 607		break;
 608	case TTM_PL_VRAM:
 609		man->flags = TTM_MEMTYPE_FLAG_FIXED |
 610			     TTM_MEMTYPE_FLAG_MAPPABLE;
 611		man->available_caching = TTM_PL_FLAG_UNCACHED |
 612					 TTM_PL_FLAG_WC;
 613		man->default_caching = TTM_PL_FLAG_WC;
 614
 615		if (drm->device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
 616			/* Some BARs do not support being ioremapped WC */
 617			if (nvxx_bar(&drm->device)->iomap_uncached) {
 618				man->available_caching = TTM_PL_FLAG_UNCACHED;
 619				man->default_caching = TTM_PL_FLAG_UNCACHED;
 620			}
 621
 622			man->func = &nouveau_vram_manager;
 623			man->io_reserve_fastpath = false;
 624			man->use_io_reserve_lru = true;
 625		} else {
 626			man->func = &ttm_bo_manager_func;
 627		}
 
 
 
 
 
 628		break;
 629	case TTM_PL_TT:
 630		if (drm->device.info.family >= NV_DEVICE_INFO_V0_TESLA)
 631			man->func = &nouveau_gart_manager;
 632		else
 633		if (!drm->agp.bridge)
 634			man->func = &nv04_gart_manager;
 635		else
 636			man->func = &ttm_bo_manager_func;
 637
 638		if (drm->agp.bridge) {
 639			man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
 640			man->available_caching = TTM_PL_FLAG_UNCACHED |
 641				TTM_PL_FLAG_WC;
 642			man->default_caching = TTM_PL_FLAG_WC;
 643		} else {
 
 
 644			man->flags = TTM_MEMTYPE_FLAG_MAPPABLE |
 645				     TTM_MEMTYPE_FLAG_CMA;
 646			man->available_caching = TTM_PL_MASK_CACHING;
 647			man->default_caching = TTM_PL_FLAG_CACHED;
 
 
 
 
 
 648		}
 649
 650		break;
 651	default:
 
 652		return -EINVAL;
 653	}
 654	return 0;
 655}
 656
 657static void
 658nouveau_bo_evict_flags(struct ttm_buffer_object *bo, struct ttm_placement *pl)
 659{
 660	struct nouveau_bo *nvbo = nouveau_bo(bo);
 661
 662	switch (bo->mem.mem_type) {
 663	case TTM_PL_VRAM:
 664		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_TT,
 665					 TTM_PL_FLAG_SYSTEM);
 666		break;
 667	default:
 668		nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_SYSTEM, 0);
 669		break;
 670	}
 671
 672	*pl = nvbo->placement;
 673}
 674
 675
 676static int
 677nve0_bo_move_init(struct nouveau_channel *chan, u32 handle)
 678{
 679	int ret = RING_SPACE(chan, 2);
 680	if (ret == 0) {
 681		BEGIN_NVC0(chan, NvSubCopy, 0x0000, 1);
 682		OUT_RING  (chan, handle & 0x0000ffff);
 683		FIRE_RING (chan);
 684	}
 685	return ret;
 686}
 687
 688static int
 689nve0_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
 690		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
 691{
 692	struct nvkm_mem *node = old_mem->mm_node;
 693	int ret = RING_SPACE(chan, 10);
 694	if (ret == 0) {
 695		BEGIN_NVC0(chan, NvSubCopy, 0x0400, 8);
 696		OUT_RING  (chan, upper_32_bits(node->vma[0].offset));
 697		OUT_RING  (chan, lower_32_bits(node->vma[0].offset));
 698		OUT_RING  (chan, upper_32_bits(node->vma[1].offset));
 699		OUT_RING  (chan, lower_32_bits(node->vma[1].offset));
 700		OUT_RING  (chan, PAGE_SIZE);
 701		OUT_RING  (chan, PAGE_SIZE);
 702		OUT_RING  (chan, PAGE_SIZE);
 703		OUT_RING  (chan, new_mem->num_pages);
 704		BEGIN_IMC0(chan, NvSubCopy, 0x0300, 0x0386);
 705	}
 706	return ret;
 707}
 708
 709static int
 710nvc0_bo_move_init(struct nouveau_channel *chan, u32 handle)
 
 
 
 711{
 712	int ret = RING_SPACE(chan, 2);
 713	if (ret == 0) {
 714		BEGIN_NVC0(chan, NvSubCopy, 0x0000, 1);
 715		OUT_RING  (chan, handle);
 716	}
 717	return ret;
 718}
 719
 720static int
 721nvc0_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
 722		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
 723{
 724	struct nvkm_mem *node = old_mem->mm_node;
 725	u64 src_offset = node->vma[0].offset;
 726	u64 dst_offset = node->vma[1].offset;
 727	u32 page_count = new_mem->num_pages;
 728	int ret;
 729
 730	page_count = new_mem->num_pages;
 731	while (page_count) {
 732		int line_count = (page_count > 8191) ? 8191 : page_count;
 733
 734		ret = RING_SPACE(chan, 11);
 735		if (ret)
 736			return ret;
 737
 738		BEGIN_NVC0(chan, NvSubCopy, 0x030c, 8);
 739		OUT_RING  (chan, upper_32_bits(src_offset));
 740		OUT_RING  (chan, lower_32_bits(src_offset));
 741		OUT_RING  (chan, upper_32_bits(dst_offset));
 742		OUT_RING  (chan, lower_32_bits(dst_offset));
 743		OUT_RING  (chan, PAGE_SIZE);
 744		OUT_RING  (chan, PAGE_SIZE);
 745		OUT_RING  (chan, PAGE_SIZE);
 746		OUT_RING  (chan, line_count);
 747		BEGIN_NVC0(chan, NvSubCopy, 0x0300, 1);
 748		OUT_RING  (chan, 0x00000110);
 749
 750		page_count -= line_count;
 751		src_offset += (PAGE_SIZE * line_count);
 752		dst_offset += (PAGE_SIZE * line_count);
 753	}
 754
 755	return 0;
 756}
 757
 758static int
 759nvc0_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
 760		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
 761{
 762	struct nvkm_mem *node = old_mem->mm_node;
 763	u64 src_offset = node->vma[0].offset;
 764	u64 dst_offset = node->vma[1].offset;
 765	u32 page_count = new_mem->num_pages;
 766	int ret;
 767
 768	page_count = new_mem->num_pages;
 769	while (page_count) {
 770		int line_count = (page_count > 2047) ? 2047 : page_count;
 771
 772		ret = RING_SPACE(chan, 12);
 773		if (ret)
 774			return ret;
 775
 776		BEGIN_NVC0(chan, NvSubCopy, 0x0238, 2);
 777		OUT_RING  (chan, upper_32_bits(dst_offset));
 778		OUT_RING  (chan, lower_32_bits(dst_offset));
 779		BEGIN_NVC0(chan, NvSubCopy, 0x030c, 6);
 780		OUT_RING  (chan, upper_32_bits(src_offset));
 781		OUT_RING  (chan, lower_32_bits(src_offset));
 782		OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
 783		OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
 784		OUT_RING  (chan, PAGE_SIZE); /* line_length */
 785		OUT_RING  (chan, line_count);
 786		BEGIN_NVC0(chan, NvSubCopy, 0x0300, 1);
 787		OUT_RING  (chan, 0x00100110);
 788
 789		page_count -= line_count;
 790		src_offset += (PAGE_SIZE * line_count);
 791		dst_offset += (PAGE_SIZE * line_count);
 792	}
 793
 794	return 0;
 795}
 796
 797static int
 798nva3_bo_move_copy(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
 799		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
 800{
 801	struct nvkm_mem *node = old_mem->mm_node;
 802	u64 src_offset = node->vma[0].offset;
 803	u64 dst_offset = node->vma[1].offset;
 804	u32 page_count = new_mem->num_pages;
 805	int ret;
 806
 807	page_count = new_mem->num_pages;
 808	while (page_count) {
 809		int line_count = (page_count > 8191) ? 8191 : page_count;
 810
 811		ret = RING_SPACE(chan, 11);
 812		if (ret)
 813			return ret;
 814
 815		BEGIN_NV04(chan, NvSubCopy, 0x030c, 8);
 816		OUT_RING  (chan, upper_32_bits(src_offset));
 817		OUT_RING  (chan, lower_32_bits(src_offset));
 818		OUT_RING  (chan, upper_32_bits(dst_offset));
 819		OUT_RING  (chan, lower_32_bits(dst_offset));
 820		OUT_RING  (chan, PAGE_SIZE);
 821		OUT_RING  (chan, PAGE_SIZE);
 822		OUT_RING  (chan, PAGE_SIZE);
 823		OUT_RING  (chan, line_count);
 824		BEGIN_NV04(chan, NvSubCopy, 0x0300, 1);
 825		OUT_RING  (chan, 0x00000110);
 826
 827		page_count -= line_count;
 828		src_offset += (PAGE_SIZE * line_count);
 829		dst_offset += (PAGE_SIZE * line_count);
 830	}
 831
 832	return 0;
 833}
 834
 835static int
 836nv98_bo_move_exec(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
 837		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
 838{
 839	struct nvkm_mem *node = old_mem->mm_node;
 840	int ret = RING_SPACE(chan, 7);
 841	if (ret == 0) {
 842		BEGIN_NV04(chan, NvSubCopy, 0x0320, 6);
 843		OUT_RING  (chan, upper_32_bits(node->vma[0].offset));
 844		OUT_RING  (chan, lower_32_bits(node->vma[0].offset));
 845		OUT_RING  (chan, upper_32_bits(node->vma[1].offset));
 846		OUT_RING  (chan, lower_32_bits(node->vma[1].offset));
 847		OUT_RING  (chan, 0x00000000 /* COPY */);
 848		OUT_RING  (chan, new_mem->num_pages << PAGE_SHIFT);
 849	}
 850	return ret;
 851}
 852
 853static int
 854nv84_bo_move_exec(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
 855		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
 856{
 857	struct nvkm_mem *node = old_mem->mm_node;
 858	int ret = RING_SPACE(chan, 7);
 859	if (ret == 0) {
 860		BEGIN_NV04(chan, NvSubCopy, 0x0304, 6);
 861		OUT_RING  (chan, new_mem->num_pages << PAGE_SHIFT);
 862		OUT_RING  (chan, upper_32_bits(node->vma[0].offset));
 863		OUT_RING  (chan, lower_32_bits(node->vma[0].offset));
 864		OUT_RING  (chan, upper_32_bits(node->vma[1].offset));
 865		OUT_RING  (chan, lower_32_bits(node->vma[1].offset));
 866		OUT_RING  (chan, 0x00000000 /* MODE_COPY, QUERY_NONE */);
 867	}
 868	return ret;
 869}
 870
 871static int
 872nv50_bo_move_init(struct nouveau_channel *chan, u32 handle)
 873{
 874	int ret = RING_SPACE(chan, 6);
 875	if (ret == 0) {
 876		BEGIN_NV04(chan, NvSubCopy, 0x0000, 1);
 877		OUT_RING  (chan, handle);
 878		BEGIN_NV04(chan, NvSubCopy, 0x0180, 3);
 879		OUT_RING  (chan, chan->drm->ntfy.handle);
 880		OUT_RING  (chan, chan->vram.handle);
 881		OUT_RING  (chan, chan->vram.handle);
 882	}
 883
 884	return ret;
 885}
 886
 887static int
 888nv50_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
 889		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
 890{
 891	struct nvkm_mem *node = old_mem->mm_node;
 
 892	u64 length = (new_mem->num_pages << PAGE_SHIFT);
 893	u64 src_offset = node->vma[0].offset;
 894	u64 dst_offset = node->vma[1].offset;
 895	int src_tiled = !!node->memtype;
 896	int dst_tiled = !!((struct nvkm_mem *)new_mem->mm_node)->memtype;
 897	int ret;
 898
 899	while (length) {
 900		u32 amount, stride, height;
 901
 902		ret = RING_SPACE(chan, 18 + 6 * (src_tiled + dst_tiled));
 903		if (ret)
 904			return ret;
 905
 906		amount  = min(length, (u64)(4 * 1024 * 1024));
 907		stride  = 16 * 4;
 908		height  = amount / stride;
 909
 910		if (src_tiled) {
 911			BEGIN_NV04(chan, NvSubCopy, 0x0200, 7);
 
 
 
 
 
 912			OUT_RING  (chan, 0);
 913			OUT_RING  (chan, 0);
 914			OUT_RING  (chan, stride);
 915			OUT_RING  (chan, height);
 916			OUT_RING  (chan, 1);
 917			OUT_RING  (chan, 0);
 918			OUT_RING  (chan, 0);
 919		} else {
 920			BEGIN_NV04(chan, NvSubCopy, 0x0200, 1);
 
 
 
 
 921			OUT_RING  (chan, 1);
 922		}
 923		if (dst_tiled) {
 924			BEGIN_NV04(chan, NvSubCopy, 0x021c, 7);
 
 
 
 
 
 925			OUT_RING  (chan, 0);
 926			OUT_RING  (chan, 0);
 927			OUT_RING  (chan, stride);
 928			OUT_RING  (chan, height);
 929			OUT_RING  (chan, 1);
 930			OUT_RING  (chan, 0);
 931			OUT_RING  (chan, 0);
 932		} else {
 933			BEGIN_NV04(chan, NvSubCopy, 0x021c, 1);
 
 
 
 
 934			OUT_RING  (chan, 1);
 935		}
 936
 937		BEGIN_NV04(chan, NvSubCopy, 0x0238, 2);
 
 
 
 
 938		OUT_RING  (chan, upper_32_bits(src_offset));
 939		OUT_RING  (chan, upper_32_bits(dst_offset));
 940		BEGIN_NV04(chan, NvSubCopy, 0x030c, 8);
 941		OUT_RING  (chan, lower_32_bits(src_offset));
 942		OUT_RING  (chan, lower_32_bits(dst_offset));
 943		OUT_RING  (chan, stride);
 944		OUT_RING  (chan, stride);
 945		OUT_RING  (chan, stride);
 946		OUT_RING  (chan, height);
 947		OUT_RING  (chan, 0x00000101);
 948		OUT_RING  (chan, 0x00000000);
 949		BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
 950		OUT_RING  (chan, 0);
 951
 952		length -= amount;
 953		src_offset += amount;
 954		dst_offset += amount;
 955	}
 956
 957	return 0;
 958}
 959
 960static int
 961nv04_bo_move_init(struct nouveau_channel *chan, u32 handle)
 962{
 963	int ret = RING_SPACE(chan, 4);
 964	if (ret == 0) {
 965		BEGIN_NV04(chan, NvSubCopy, 0x0000, 1);
 966		OUT_RING  (chan, handle);
 967		BEGIN_NV04(chan, NvSubCopy, 0x0180, 1);
 968		OUT_RING  (chan, chan->drm->ntfy.handle);
 969	}
 970
 971	return ret;
 972}
 973
 974static inline uint32_t
 975nouveau_bo_mem_ctxdma(struct ttm_buffer_object *bo,
 976		      struct nouveau_channel *chan, struct ttm_mem_reg *mem)
 977{
 978	if (mem->mem_type == TTM_PL_TT)
 979		return NvDmaTT;
 980	return chan->vram.handle;
 981}
 982
 983static int
 984nv04_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
 985		  struct ttm_mem_reg *old_mem, struct ttm_mem_reg *new_mem)
 986{
 987	u32 src_offset = old_mem->start << PAGE_SHIFT;
 988	u32 dst_offset = new_mem->start << PAGE_SHIFT;
 989	u32 page_count = new_mem->num_pages;
 990	int ret;
 991
 992	ret = RING_SPACE(chan, 3);
 993	if (ret)
 994		return ret;
 995
 996	BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_DMA_SOURCE, 2);
 997	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, old_mem));
 998	OUT_RING  (chan, nouveau_bo_mem_ctxdma(bo, chan, new_mem));
 999
1000	page_count = new_mem->num_pages;
1001	while (page_count) {
1002		int line_count = (page_count > 2047) ? 2047 : page_count;
1003
1004		ret = RING_SPACE(chan, 11);
1005		if (ret)
1006			return ret;
1007
1008		BEGIN_NV04(chan, NvSubCopy,
1009				 NV_MEMORY_TO_MEMORY_FORMAT_OFFSET_IN, 8);
1010		OUT_RING  (chan, src_offset);
1011		OUT_RING  (chan, dst_offset);
1012		OUT_RING  (chan, PAGE_SIZE); /* src_pitch */
1013		OUT_RING  (chan, PAGE_SIZE); /* dst_pitch */
1014		OUT_RING  (chan, PAGE_SIZE); /* line_length */
1015		OUT_RING  (chan, line_count);
1016		OUT_RING  (chan, 0x00000101);
1017		OUT_RING  (chan, 0x00000000);
1018		BEGIN_NV04(chan, NvSubCopy, NV_MEMORY_TO_MEMORY_FORMAT_NOP, 1);
1019		OUT_RING  (chan, 0);
1020
1021		page_count -= line_count;
1022		src_offset += (PAGE_SIZE * line_count);
1023		dst_offset += (PAGE_SIZE * line_count);
1024	}
1025
1026	return 0;
1027}
1028
1029static int
1030nouveau_bo_move_prep(struct nouveau_drm *drm, struct ttm_buffer_object *bo,
1031		     struct ttm_mem_reg *mem)
1032{
1033	struct nvkm_mem *old_node = bo->mem.mm_node;
1034	struct nvkm_mem *new_node = mem->mm_node;
1035	u64 size = (u64)mem->num_pages << PAGE_SHIFT;
1036	int ret;
1037
1038	ret = nvkm_vm_get(drm->client.vm, size, old_node->page_shift,
1039			  NV_MEM_ACCESS_RW, &old_node->vma[0]);
1040	if (ret)
1041		return ret;
1042
1043	ret = nvkm_vm_get(drm->client.vm, size, new_node->page_shift,
1044			  NV_MEM_ACCESS_RW, &old_node->vma[1]);
1045	if (ret) {
1046		nvkm_vm_put(&old_node->vma[0]);
1047		return ret;
1048	}
1049
1050	nvkm_vm_map(&old_node->vma[0], old_node);
1051	nvkm_vm_map(&old_node->vma[1], new_node);
1052	return 0;
1053}
1054
1055static int
1056nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict, bool intr,
1057		     bool no_wait_gpu, struct ttm_mem_reg *new_mem)
 
1058{
1059	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1060	struct nouveau_channel *chan = drm->ttm.chan;
1061	struct nouveau_cli *cli = (void *)chan->user.client;
1062	struct nouveau_fence *fence;
1063	int ret;
1064
 
 
 
 
 
 
1065	/* create temporary vmas for the transfer and attach them to the
1066	 * old nvkm_mem node, these will get cleaned up after ttm has
1067	 * destroyed the ttm_mem_reg
1068	 */
1069	if (drm->device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
1070		ret = nouveau_bo_move_prep(drm, bo, new_mem);
 
 
 
 
 
 
1071		if (ret)
1072			return ret;
1073	}
1074
1075	mutex_lock_nested(&cli->mutex, SINGLE_DEPTH_NESTING);
1076	ret = nouveau_fence_sync(nouveau_bo(bo), chan, true, intr);
 
 
 
 
 
1077	if (ret == 0) {
1078		ret = drm->ttm.move(chan, bo, &bo->mem, new_mem);
1079		if (ret == 0) {
1080			ret = nouveau_fence_new(chan, false, &fence);
1081			if (ret == 0) {
1082				ret = ttm_bo_move_accel_cleanup(bo,
1083								&fence->base,
1084								evict,
1085								no_wait_gpu,
1086								new_mem);
1087				nouveau_fence_unref(&fence);
1088			}
1089		}
1090	}
1091	mutex_unlock(&cli->mutex);
 
 
 
1092	return ret;
1093}
1094
1095void
1096nouveau_bo_move_init(struct nouveau_drm *drm)
1097{
1098	static const struct {
1099		const char *name;
1100		int engine;
1101		s32 oclass;
1102		int (*exec)(struct nouveau_channel *,
1103			    struct ttm_buffer_object *,
1104			    struct ttm_mem_reg *, struct ttm_mem_reg *);
1105		int (*init)(struct nouveau_channel *, u32 handle);
1106	} _methods[] = {
1107		{  "COPY", 4, 0xb0b5, nve0_bo_move_copy, nve0_bo_move_init },
1108		{  "GRCE", 0, 0xb0b5, nve0_bo_move_copy, nvc0_bo_move_init },
1109		{  "COPY", 4, 0xa0b5, nve0_bo_move_copy, nve0_bo_move_init },
1110		{  "GRCE", 0, 0xa0b5, nve0_bo_move_copy, nvc0_bo_move_init },
1111		{ "COPY1", 5, 0x90b8, nvc0_bo_move_copy, nvc0_bo_move_init },
1112		{ "COPY0", 4, 0x90b5, nvc0_bo_move_copy, nvc0_bo_move_init },
1113		{  "COPY", 0, 0x85b5, nva3_bo_move_copy, nv50_bo_move_init },
1114		{ "CRYPT", 0, 0x74c1, nv84_bo_move_exec, nv50_bo_move_init },
1115		{  "M2MF", 0, 0x9039, nvc0_bo_move_m2mf, nvc0_bo_move_init },
1116		{  "M2MF", 0, 0x5039, nv50_bo_move_m2mf, nv50_bo_move_init },
1117		{  "M2MF", 0, 0x0039, nv04_bo_move_m2mf, nv04_bo_move_init },
1118		{},
1119		{ "CRYPT", 0, 0x88b4, nv98_bo_move_exec, nv50_bo_move_init },
1120	}, *mthd = _methods;
1121	const char *name = "CPU";
1122	int ret;
1123
1124	do {
1125		struct nouveau_channel *chan;
1126
1127		if (mthd->engine)
1128			chan = drm->cechan;
1129		else
1130			chan = drm->channel;
1131		if (chan == NULL)
1132			continue;
1133
1134		ret = nvif_object_init(&chan->user,
1135				       mthd->oclass | (mthd->engine << 16),
1136				       mthd->oclass, NULL, 0,
1137				       &drm->ttm.copy);
1138		if (ret == 0) {
1139			ret = mthd->init(chan, drm->ttm.copy.handle);
1140			if (ret) {
1141				nvif_object_fini(&drm->ttm.copy);
1142				continue;
1143			}
1144
1145			drm->ttm.move = mthd->exec;
1146			drm->ttm.chan = chan;
1147			name = mthd->name;
1148			break;
1149		}
1150	} while ((++mthd)->exec);
1151
1152	NV_INFO(drm, "MM: using %s for buffer copies\n", name);
1153}
1154
1155static int
1156nouveau_bo_move_flipd(struct ttm_buffer_object *bo, bool evict, bool intr,
1157		      bool no_wait_gpu, struct ttm_mem_reg *new_mem)
 
1158{
1159	struct ttm_place placement_memtype = {
1160		.fpfn = 0,
1161		.lpfn = 0,
1162		.flags = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING
1163	};
1164	struct ttm_placement placement;
1165	struct ttm_mem_reg tmp_mem;
1166	int ret;
1167
 
1168	placement.num_placement = placement.num_busy_placement = 1;
1169	placement.placement = placement.busy_placement = &placement_memtype;
1170
1171	tmp_mem = *new_mem;
1172	tmp_mem.mm_node = NULL;
1173	ret = ttm_bo_mem_space(bo, &placement, &tmp_mem, intr, no_wait_gpu);
1174	if (ret)
1175		return ret;
1176
1177	ret = ttm_tt_bind(bo->ttm, &tmp_mem);
1178	if (ret)
1179		goto out;
1180
1181	ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_gpu, &tmp_mem);
1182	if (ret)
1183		goto out;
1184
1185	ret = ttm_bo_move_ttm(bo, true, no_wait_gpu, new_mem);
1186out:
1187	ttm_bo_mem_put(bo, &tmp_mem);
1188	return ret;
1189}
1190
1191static int
1192nouveau_bo_move_flips(struct ttm_buffer_object *bo, bool evict, bool intr,
1193		      bool no_wait_gpu, struct ttm_mem_reg *new_mem)
 
1194{
1195	struct ttm_place placement_memtype = {
1196		.fpfn = 0,
1197		.lpfn = 0,
1198		.flags = TTM_PL_FLAG_TT | TTM_PL_MASK_CACHING
1199	};
1200	struct ttm_placement placement;
1201	struct ttm_mem_reg tmp_mem;
1202	int ret;
1203
 
1204	placement.num_placement = placement.num_busy_placement = 1;
1205	placement.placement = placement.busy_placement = &placement_memtype;
1206
1207	tmp_mem = *new_mem;
1208	tmp_mem.mm_node = NULL;
1209	ret = ttm_bo_mem_space(bo, &placement, &tmp_mem, intr, no_wait_gpu);
1210	if (ret)
1211		return ret;
1212
1213	ret = ttm_bo_move_ttm(bo, true, no_wait_gpu, &tmp_mem);
1214	if (ret)
1215		goto out;
1216
1217	ret = nouveau_bo_move_m2mf(bo, true, intr, no_wait_gpu, new_mem);
1218	if (ret)
1219		goto out;
1220
1221out:
1222	ttm_bo_mem_put(bo, &tmp_mem);
1223	return ret;
1224}
1225
1226static void
1227nouveau_bo_move_ntfy(struct ttm_buffer_object *bo, struct ttm_mem_reg *new_mem)
1228{
 
1229	struct nouveau_bo *nvbo = nouveau_bo(bo);
1230	struct nvkm_vma *vma;
1231
1232	/* ttm can now (stupidly) pass the driver bos it didn't create... */
1233	if (bo->destroy != nouveau_bo_del_ttm)
1234		return;
1235
1236	list_for_each_entry(vma, &nvbo->vma_list, head) {
1237		if (new_mem && new_mem->mem_type != TTM_PL_SYSTEM &&
1238			      (new_mem->mem_type == TTM_PL_VRAM ||
1239			       nvbo->page_shift != vma->vm->mmu->lpg_shift)) {
1240			nvkm_vm_map(vma, new_mem->mm_node);
 
 
 
 
1241		} else {
1242			nvkm_vm_unmap(vma);
1243		}
1244	}
1245}
1246
1247static int
1248nouveau_bo_vm_bind(struct ttm_buffer_object *bo, struct ttm_mem_reg *new_mem,
1249		   struct nouveau_drm_tile **new_tile)
1250{
1251	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1252	struct drm_device *dev = drm->dev;
1253	struct nouveau_bo *nvbo = nouveau_bo(bo);
1254	u64 offset = new_mem->start << PAGE_SHIFT;
1255
1256	*new_tile = NULL;
1257	if (new_mem->mem_type != TTM_PL_VRAM)
1258		return 0;
1259
1260	if (drm->device.info.family >= NV_DEVICE_INFO_V0_CELSIUS) {
1261		*new_tile = nv10_bo_set_tiling(dev, offset, new_mem->size,
1262						nvbo->tile_mode,
1263						nvbo->tile_flags);
1264	}
1265
1266	return 0;
1267}
1268
1269static void
1270nouveau_bo_vm_cleanup(struct ttm_buffer_object *bo,
1271		      struct nouveau_drm_tile *new_tile,
1272		      struct nouveau_drm_tile **old_tile)
1273{
1274	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1275	struct drm_device *dev = drm->dev;
1276	struct fence *fence = reservation_object_get_excl(bo->resv);
1277
1278	nv10_bo_put_tile_region(dev, *old_tile, fence);
1279	*old_tile = new_tile;
1280}
1281
1282static int
1283nouveau_bo_move(struct ttm_buffer_object *bo, bool evict, bool intr,
1284		bool no_wait_gpu, struct ttm_mem_reg *new_mem)
 
1285{
1286	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1287	struct nouveau_bo *nvbo = nouveau_bo(bo);
1288	struct ttm_mem_reg *old_mem = &bo->mem;
1289	struct nouveau_drm_tile *new_tile = NULL;
1290	int ret = 0;
1291
1292	if (nvbo->pin_refcnt)
1293		NV_WARN(drm, "Moving pinned object %p!\n", nvbo);
1294
1295	if (drm->device.info.family < NV_DEVICE_INFO_V0_TESLA) {
1296		ret = nouveau_bo_vm_bind(bo, new_mem, &new_tile);
1297		if (ret)
1298			return ret;
1299	}
1300
1301	/* Fake bo copy. */
1302	if (old_mem->mem_type == TTM_PL_SYSTEM && !bo->ttm) {
1303		BUG_ON(bo->mem.mm_node != NULL);
1304		bo->mem = *new_mem;
1305		new_mem->mm_node = NULL;
1306		goto out;
1307	}
1308
 
 
 
 
 
 
1309	/* Hardware assisted copy. */
1310	if (drm->ttm.move) {
1311		if (new_mem->mem_type == TTM_PL_SYSTEM)
1312			ret = nouveau_bo_move_flipd(bo, evict, intr,
1313						    no_wait_gpu, new_mem);
1314		else if (old_mem->mem_type == TTM_PL_SYSTEM)
1315			ret = nouveau_bo_move_flips(bo, evict, intr,
1316						    no_wait_gpu, new_mem);
1317		else
1318			ret = nouveau_bo_move_m2mf(bo, evict, intr,
1319						   no_wait_gpu, new_mem);
1320		if (!ret)
1321			goto out;
1322	}
1323
1324	/* Fallback to software copy. */
1325	ret = ttm_bo_wait(bo, true, intr, no_wait_gpu);
1326	if (ret == 0)
1327		ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, new_mem);
1328
1329out:
1330	if (drm->device.info.family < NV_DEVICE_INFO_V0_TESLA) {
1331		if (ret)
1332			nouveau_bo_vm_cleanup(bo, NULL, &new_tile);
1333		else
1334			nouveau_bo_vm_cleanup(bo, new_tile, &nvbo->tile);
1335	}
1336
1337	return ret;
1338}
1339
1340static int
1341nouveau_bo_verify_access(struct ttm_buffer_object *bo, struct file *filp)
1342{
1343	struct nouveau_bo *nvbo = nouveau_bo(bo);
1344
1345	return drm_vma_node_verify_access(&nvbo->gem.vma_node, filp);
1346}
1347
1348static int
1349nouveau_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1350{
1351	struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1352	struct nouveau_drm *drm = nouveau_bdev(bdev);
1353	struct nvkm_device *device = nvxx_device(&drm->device);
1354	struct nvkm_mem *node = mem->mm_node;
1355	int ret;
1356
1357	mem->bus.addr = NULL;
1358	mem->bus.offset = 0;
1359	mem->bus.size = mem->num_pages << PAGE_SHIFT;
1360	mem->bus.base = 0;
1361	mem->bus.is_iomem = false;
1362	if (!(man->flags & TTM_MEMTYPE_FLAG_MAPPABLE))
1363		return -EINVAL;
1364	switch (mem->mem_type) {
1365	case TTM_PL_SYSTEM:
1366		/* System memory */
1367		return 0;
1368	case TTM_PL_TT:
1369#if IS_ENABLED(CONFIG_AGP)
1370		if (drm->agp.bridge) {
1371			mem->bus.offset = mem->start << PAGE_SHIFT;
1372			mem->bus.base = drm->agp.base;
1373			mem->bus.is_iomem = !drm->agp.cma;
1374		}
1375#endif
1376		if (drm->device.info.family < NV_DEVICE_INFO_V0_TESLA || !node->memtype)
1377			/* untiled */
 
 
 
 
 
 
 
 
1378			break;
1379		/* fallthrough, tiled memory */
1380	case TTM_PL_VRAM:
1381		mem->bus.offset = mem->start << PAGE_SHIFT;
1382		mem->bus.base = device->func->resource_addr(device, 1);
1383		mem->bus.is_iomem = true;
1384		if (drm->device.info.family >= NV_DEVICE_INFO_V0_TESLA) {
1385			struct nvkm_bar *bar = nvxx_bar(&drm->device);
1386			int page_shift = 12;
1387			if (drm->device.info.family >= NV_DEVICE_INFO_V0_FERMI)
1388				page_shift = node->page_shift;
1389
1390			ret = nvkm_bar_umap(bar, node->size << 12, page_shift,
1391					    &node->bar_vma);
1392			if (ret)
1393				return ret;
 
1394
1395			nvkm_vm_map(&node->bar_vma, node);
1396			mem->bus.offset = node->bar_vma.offset;
 
 
1397		}
 
 
 
 
 
 
 
1398		break;
1399	default:
1400		return -EINVAL;
1401	}
1402	return 0;
1403}
1404
1405static void
1406nouveau_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1407{
1408	struct nvkm_mem *node = mem->mm_node;
 
 
 
 
1409
1410	if (!node->bar_vma.node)
1411		return;
1412
1413	nvkm_vm_unmap(&node->bar_vma);
1414	nvkm_vm_put(&node->bar_vma);
1415}
1416
1417static int
1418nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
1419{
1420	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
1421	struct nouveau_bo *nvbo = nouveau_bo(bo);
1422	struct nvkm_device *device = nvxx_device(&drm->device);
1423	u32 mappable = device->func->resource_size(device, 1) >> PAGE_SHIFT;
1424	int i, ret;
1425
1426	/* as long as the bo isn't in vram, and isn't tiled, we've got
1427	 * nothing to do here.
1428	 */
1429	if (bo->mem.mem_type != TTM_PL_VRAM) {
1430		if (drm->device.info.family < NV_DEVICE_INFO_V0_TESLA ||
1431		    !nouveau_bo_tile_layout(nvbo))
1432			return 0;
1433
1434		if (bo->mem.mem_type == TTM_PL_SYSTEM) {
1435			nouveau_bo_placement_set(nvbo, TTM_PL_TT, 0);
1436
1437			ret = nouveau_bo_validate(nvbo, false, false);
1438			if (ret)
1439				return ret;
1440		}
1441		return 0;
1442	}
1443
1444	/* make sure bo is in mappable vram */
1445	if (drm->device.info.family >= NV_DEVICE_INFO_V0_TESLA ||
1446	    bo->mem.start + bo->mem.num_pages < mappable)
1447		return 0;
1448
1449	for (i = 0; i < nvbo->placement.num_placement; ++i) {
1450		nvbo->placements[i].fpfn = 0;
1451		nvbo->placements[i].lpfn = mappable;
1452	}
1453
1454	for (i = 0; i < nvbo->placement.num_busy_placement; ++i) {
1455		nvbo->busy_placements[i].fpfn = 0;
1456		nvbo->busy_placements[i].lpfn = mappable;
1457	}
1458
1459	nouveau_bo_placement_set(nvbo, TTM_PL_FLAG_VRAM, 0);
1460	return nouveau_bo_validate(nvbo, false, false);
 
 
1461}
1462
1463static int
1464nouveau_ttm_tt_populate(struct ttm_tt *ttm)
1465{
1466	struct ttm_dma_tt *ttm_dma = (void *)ttm;
1467	struct nouveau_drm *drm;
1468	struct nvkm_device *device;
1469	struct drm_device *dev;
1470	struct device *pdev;
1471	unsigned i;
1472	int r;
1473	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1474
1475	if (ttm->state != tt_unpopulated)
1476		return 0;
1477
1478	if (slave && ttm->sg) {
1479		/* make userspace faulting work */
1480		drm_prime_sg_to_page_addr_arrays(ttm->sg, ttm->pages,
1481						 ttm_dma->dma_address, ttm->num_pages);
1482		ttm->state = tt_unbound;
1483		return 0;
1484	}
1485
1486	drm = nouveau_bdev(ttm->bdev);
1487	device = nvxx_device(&drm->device);
1488	dev = drm->dev;
1489	pdev = device->dev;
1490
1491	/*
1492	 * Objects matching this condition have been marked as force_coherent,
1493	 * so use the DMA API for them.
1494	 */
1495	if (!nvxx_device(&drm->device)->func->cpu_coherent &&
1496	    ttm->caching_state == tt_uncached)
1497		return ttm_dma_populate(ttm_dma, dev->dev);
1498
1499#if IS_ENABLED(CONFIG_AGP)
1500	if (drm->agp.bridge) {
1501		return ttm_agp_tt_populate(ttm);
1502	}
1503#endif
1504
1505#if IS_ENABLED(CONFIG_SWIOTLB) && IS_ENABLED(CONFIG_X86)
1506	if (swiotlb_nr_tbl()) {
1507		return ttm_dma_populate((void *)ttm, dev->dev);
1508	}
1509#endif
1510
1511	r = ttm_pool_populate(ttm);
1512	if (r) {
1513		return r;
1514	}
1515
1516	for (i = 0; i < ttm->num_pages; i++) {
1517		dma_addr_t addr;
1518
1519		addr = dma_map_page(pdev, ttm->pages[i], 0, PAGE_SIZE,
1520				    DMA_BIDIRECTIONAL);
1521
1522		if (dma_mapping_error(pdev, addr)) {
1523			while (i--) {
1524				dma_unmap_page(pdev, ttm_dma->dma_address[i],
1525					       PAGE_SIZE, DMA_BIDIRECTIONAL);
1526				ttm_dma->dma_address[i] = 0;
1527			}
1528			ttm_pool_unpopulate(ttm);
1529			return -EFAULT;
1530		}
1531
1532		ttm_dma->dma_address[i] = addr;
1533	}
1534	return 0;
1535}
1536
1537static void
1538nouveau_ttm_tt_unpopulate(struct ttm_tt *ttm)
1539{
1540	struct ttm_dma_tt *ttm_dma = (void *)ttm;
1541	struct nouveau_drm *drm;
1542	struct nvkm_device *device;
1543	struct drm_device *dev;
1544	struct device *pdev;
1545	unsigned i;
1546	bool slave = !!(ttm->page_flags & TTM_PAGE_FLAG_SG);
1547
1548	if (slave)
1549		return;
1550
1551	drm = nouveau_bdev(ttm->bdev);
1552	device = nvxx_device(&drm->device);
1553	dev = drm->dev;
1554	pdev = device->dev;
1555
1556	/*
1557	 * Objects matching this condition have been marked as force_coherent,
1558	 * so use the DMA API for them.
1559	 */
1560	if (!nvxx_device(&drm->device)->func->cpu_coherent &&
1561	    ttm->caching_state == tt_uncached) {
1562		ttm_dma_unpopulate(ttm_dma, dev->dev);
1563		return;
1564	}
1565
1566#if IS_ENABLED(CONFIG_AGP)
1567	if (drm->agp.bridge) {
1568		ttm_agp_tt_unpopulate(ttm);
1569		return;
1570	}
1571#endif
1572
1573#if IS_ENABLED(CONFIG_SWIOTLB) && IS_ENABLED(CONFIG_X86)
1574	if (swiotlb_nr_tbl()) {
1575		ttm_dma_unpopulate((void *)ttm, dev->dev);
1576		return;
1577	}
1578#endif
1579
1580	for (i = 0; i < ttm->num_pages; i++) {
1581		if (ttm_dma->dma_address[i]) {
1582			dma_unmap_page(pdev, ttm_dma->dma_address[i], PAGE_SIZE,
1583				       DMA_BIDIRECTIONAL);
1584		}
1585	}
1586
1587	ttm_pool_unpopulate(ttm);
1588}
1589
1590void
1591nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence, bool exclusive)
1592{
1593	struct reservation_object *resv = nvbo->bo.resv;
1594
1595	if (exclusive)
1596		reservation_object_add_excl_fence(resv, &fence->base);
1597	else if (fence)
1598		reservation_object_add_shared_fence(resv, &fence->base);
1599}
1600
1601struct ttm_bo_driver nouveau_bo_driver = {
1602	.ttm_tt_create = &nouveau_ttm_tt_create,
1603	.ttm_tt_populate = &nouveau_ttm_tt_populate,
1604	.ttm_tt_unpopulate = &nouveau_ttm_tt_unpopulate,
1605	.invalidate_caches = nouveau_bo_invalidate_caches,
1606	.init_mem_type = nouveau_bo_init_mem_type,
1607	.evict_flags = nouveau_bo_evict_flags,
1608	.move_notify = nouveau_bo_move_ntfy,
1609	.move = nouveau_bo_move,
1610	.verify_access = nouveau_bo_verify_access,
 
 
 
 
 
1611	.fault_reserve_notify = &nouveau_ttm_fault_reserve_notify,
1612	.io_mem_reserve = &nouveau_ttm_io_mem_reserve,
1613	.io_mem_free = &nouveau_ttm_io_mem_free,
1614};
1615
1616struct nvkm_vma *
1617nouveau_bo_vma_find(struct nouveau_bo *nvbo, struct nvkm_vm *vm)
1618{
1619	struct nvkm_vma *vma;
1620	list_for_each_entry(vma, &nvbo->vma_list, head) {
1621		if (vma->vm == vm)
1622			return vma;
1623	}
1624
1625	return NULL;
1626}
1627
1628int
1629nouveau_bo_vma_add(struct nouveau_bo *nvbo, struct nvkm_vm *vm,
1630		   struct nvkm_vma *vma)
1631{
1632	const u32 size = nvbo->bo.mem.num_pages << PAGE_SHIFT;
 
1633	int ret;
1634
1635	ret = nvkm_vm_get(vm, size, nvbo->page_shift,
1636			     NV_MEM_ACCESS_RW, vma);
1637	if (ret)
1638		return ret;
1639
1640	if ( nvbo->bo.mem.mem_type != TTM_PL_SYSTEM &&
1641	    (nvbo->bo.mem.mem_type == TTM_PL_VRAM ||
1642	     nvbo->page_shift != vma->vm->mmu->lpg_shift))
1643		nvkm_vm_map(vma, nvbo->bo.mem.mm_node);
 
1644
1645	list_add_tail(&vma->head, &nvbo->vma_list);
1646	vma->refcount = 1;
1647	return 0;
1648}
1649
1650void
1651nouveau_bo_vma_del(struct nouveau_bo *nvbo, struct nvkm_vma *vma)
1652{
1653	if (vma->node) {
1654		if (nvbo->bo.mem.mem_type != TTM_PL_SYSTEM)
1655			nvkm_vm_unmap(vma);
1656		nvkm_vm_put(vma);
 
 
 
 
 
1657		list_del(&vma->head);
1658	}
1659}