Linux Audio

Check our new training course

Loading...
v4.6
   1/*
   2 * Copyright 2011 Advanced Micro Devices, Inc.
   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
   7 * "Software"), to deal in the Software without restriction, including
   8 * without limitation the rights to use, copy, modify, merge, publish,
   9 * distribute, sub license, and/or sell copies of the Software, and to
  10 * permit persons to whom the Software is furnished to do so, subject to
  11 * the following conditions:
  12 *
  13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  15 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  16 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
  17 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  18 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  19 * USE OR OTHER DEALINGS IN THE SOFTWARE.
  20 *
  21 * The above copyright notice and this permission notice (including the
  22 * next paragraph) shall be included in all copies or substantial portions
  23 * of the Software.
  24 *
  25 */
  26/*
  27 * Authors:
  28 *    Christian König <deathsimple@vodafone.de>
  29 */
  30
  31#include <linux/firmware.h>
  32#include <linux/module.h>
  33#include <drm/drmP.h>
  34#include <drm/drm.h>
  35
  36#include "amdgpu.h"
  37#include "amdgpu_pm.h"
  38#include "amdgpu_uvd.h"
  39#include "cikd.h"
  40#include "uvd/uvd_4_2_d.h"
  41
  42/* 1 second timeout */
  43#define UVD_IDLE_TIMEOUT_MS	1000
 
 
 
 
 
 
 
 
 
  44
  45/* Firmware Names */
  46#ifdef CONFIG_DRM_AMDGPU_CIK
  47#define FIRMWARE_BONAIRE	"radeon/bonaire_uvd.bin"
  48#define FIRMWARE_KABINI 	"radeon/kabini_uvd.bin"
  49#define FIRMWARE_KAVERI 	"radeon/kaveri_uvd.bin"
  50#define FIRMWARE_HAWAII 	"radeon/hawaii_uvd.bin"
  51#define FIRMWARE_MULLINS	"radeon/mullins_uvd.bin"
  52#endif
  53#define FIRMWARE_TONGA		"amdgpu/tonga_uvd.bin"
  54#define FIRMWARE_CARRIZO	"amdgpu/carrizo_uvd.bin"
  55#define FIRMWARE_FIJI		"amdgpu/fiji_uvd.bin"
  56#define FIRMWARE_STONEY		"amdgpu/stoney_uvd.bin"
 
 
 
  57
  58/**
  59 * amdgpu_uvd_cs_ctx - Command submission parser context
  60 *
  61 * Used for emulating virtual memory support on UVD 4.2.
  62 */
  63struct amdgpu_uvd_cs_ctx {
  64	struct amdgpu_cs_parser *parser;
  65	unsigned reg, count;
  66	unsigned data0, data1;
  67	unsigned idx;
  68	unsigned ib_idx;
  69
  70	/* does the IB has a msg command */
  71	bool has_msg_cmd;
  72
  73	/* minimum buffer sizes */
  74	unsigned *buf_sizes;
  75};
  76
  77#ifdef CONFIG_DRM_AMDGPU_CIK
  78MODULE_FIRMWARE(FIRMWARE_BONAIRE);
  79MODULE_FIRMWARE(FIRMWARE_KABINI);
  80MODULE_FIRMWARE(FIRMWARE_KAVERI);
  81MODULE_FIRMWARE(FIRMWARE_HAWAII);
  82MODULE_FIRMWARE(FIRMWARE_MULLINS);
  83#endif
  84MODULE_FIRMWARE(FIRMWARE_TONGA);
  85MODULE_FIRMWARE(FIRMWARE_CARRIZO);
  86MODULE_FIRMWARE(FIRMWARE_FIJI);
  87MODULE_FIRMWARE(FIRMWARE_STONEY);
 
 
 
  88
  89static void amdgpu_uvd_note_usage(struct amdgpu_device *adev);
  90static void amdgpu_uvd_idle_work_handler(struct work_struct *work);
  91
  92int amdgpu_uvd_sw_init(struct amdgpu_device *adev)
  93{
  94	struct amdgpu_ring *ring;
  95	struct amd_sched_rq *rq;
  96	unsigned long bo_size;
  97	const char *fw_name;
  98	const struct common_firmware_header *hdr;
  99	unsigned version_major, version_minor, family_id;
 100	int i, r;
 101
 102	INIT_DELAYED_WORK(&adev->uvd.idle_work, amdgpu_uvd_idle_work_handler);
 103
 104	switch (adev->asic_type) {
 105#ifdef CONFIG_DRM_AMDGPU_CIK
 106	case CHIP_BONAIRE:
 107		fw_name = FIRMWARE_BONAIRE;
 108		break;
 109	case CHIP_KABINI:
 110		fw_name = FIRMWARE_KABINI;
 111		break;
 112	case CHIP_KAVERI:
 113		fw_name = FIRMWARE_KAVERI;
 114		break;
 115	case CHIP_HAWAII:
 116		fw_name = FIRMWARE_HAWAII;
 117		break;
 118	case CHIP_MULLINS:
 119		fw_name = FIRMWARE_MULLINS;
 120		break;
 121#endif
 122	case CHIP_TONGA:
 123		fw_name = FIRMWARE_TONGA;
 124		break;
 125	case CHIP_FIJI:
 126		fw_name = FIRMWARE_FIJI;
 127		break;
 128	case CHIP_CARRIZO:
 129		fw_name = FIRMWARE_CARRIZO;
 130		break;
 131	case CHIP_STONEY:
 132		fw_name = FIRMWARE_STONEY;
 133		break;
 
 
 
 
 
 
 
 
 
 134	default:
 135		return -EINVAL;
 136	}
 137
 138	r = request_firmware(&adev->uvd.fw, fw_name, adev->dev);
 139	if (r) {
 140		dev_err(adev->dev, "amdgpu_uvd: Can't load firmware \"%s\"\n",
 141			fw_name);
 142		return r;
 143	}
 144
 145	r = amdgpu_ucode_validate(adev->uvd.fw);
 146	if (r) {
 147		dev_err(adev->dev, "amdgpu_uvd: Can't validate firmware \"%s\"\n",
 148			fw_name);
 149		release_firmware(adev->uvd.fw);
 150		adev->uvd.fw = NULL;
 151		return r;
 152	}
 153
 
 
 
 154	hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
 155	family_id = le32_to_cpu(hdr->ucode_version) & 0xff;
 156	version_major = (le32_to_cpu(hdr->ucode_version) >> 24) & 0xff;
 157	version_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
 158	DRM_INFO("Found UVD firmware Version: %hu.%hu Family ID: %hu\n",
 159		version_major, version_minor, family_id);
 160
 
 
 
 
 
 
 
 
 
 
 161	adev->uvd.fw_version = ((version_major << 24) | (version_minor << 16) |
 162				(family_id << 8));
 163
 
 
 
 
 
 
 164	bo_size = AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr->ucode_size_bytes) + 8)
 165		 +  AMDGPU_UVD_STACK_SIZE + AMDGPU_UVD_HEAP_SIZE;
 166	r = amdgpu_bo_create(adev, bo_size, PAGE_SIZE, true,
 167			     AMDGPU_GEM_DOMAIN_VRAM,
 168			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
 169			     NULL, NULL, &adev->uvd.vcpu_bo);
 170	if (r) {
 171		dev_err(adev->dev, "(%d) failed to allocate UVD bo\n", r);
 172		return r;
 173	}
 174
 175	r = amdgpu_bo_reserve(adev->uvd.vcpu_bo, false);
 176	if (r) {
 177		amdgpu_bo_unref(&adev->uvd.vcpu_bo);
 178		dev_err(adev->dev, "(%d) failed to reserve UVD bo\n", r);
 179		return r;
 180	}
 181
 182	r = amdgpu_bo_pin(adev->uvd.vcpu_bo, AMDGPU_GEM_DOMAIN_VRAM,
 183			  &adev->uvd.gpu_addr);
 184	if (r) {
 185		amdgpu_bo_unreserve(adev->uvd.vcpu_bo);
 186		amdgpu_bo_unref(&adev->uvd.vcpu_bo);
 187		dev_err(adev->dev, "(%d) UVD bo pin failed\n", r);
 188		return r;
 189	}
 190
 191	r = amdgpu_bo_kmap(adev->uvd.vcpu_bo, &adev->uvd.cpu_addr);
 192	if (r) {
 193		dev_err(adev->dev, "(%d) UVD map failed\n", r);
 194		return r;
 195	}
 196
 197	amdgpu_bo_unreserve(adev->uvd.vcpu_bo);
 198
 199	ring = &adev->uvd.ring;
 200	rq = &ring->sched.sched_rq[AMD_SCHED_PRIORITY_NORMAL];
 201	r = amd_sched_entity_init(&ring->sched, &adev->uvd.entity,
 202				  rq, amdgpu_sched_jobs);
 203	if (r != 0) {
 204		DRM_ERROR("Failed setting up UVD run queue.\n");
 205		return r;
 206	}
 207
 208	for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
 209		atomic_set(&adev->uvd.handles[i], 0);
 210		adev->uvd.filp[i] = NULL;
 211	}
 212
 213	/* from uvd v5.0 HW addressing capacity increased to 64 bits */
 214	if (!amdgpu_ip_block_version_cmp(adev, AMD_IP_BLOCK_TYPE_UVD, 5, 0))
 215		adev->uvd.address_64_bit = true;
 216
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 217	return 0;
 218}
 219
 220int amdgpu_uvd_sw_fini(struct amdgpu_device *adev)
 221{
 222	int r;
 223
 224	if (adev->uvd.vcpu_bo == NULL)
 225		return 0;
 226
 227	amd_sched_entity_fini(&adev->uvd.ring.sched, &adev->uvd.entity);
 228
 229	r = amdgpu_bo_reserve(adev->uvd.vcpu_bo, false);
 230	if (!r) {
 231		amdgpu_bo_kunmap(adev->uvd.vcpu_bo);
 232		amdgpu_bo_unpin(adev->uvd.vcpu_bo);
 233		amdgpu_bo_unreserve(adev->uvd.vcpu_bo);
 234	}
 235
 236	amdgpu_bo_unref(&adev->uvd.vcpu_bo);
 237
 238	amdgpu_ring_fini(&adev->uvd.ring);
 239
 240	release_firmware(adev->uvd.fw);
 241
 242	return 0;
 243}
 244
 245int amdgpu_uvd_suspend(struct amdgpu_device *adev)
 246{
 247	unsigned size;
 248	void *ptr;
 249	int i;
 250
 251	if (adev->uvd.vcpu_bo == NULL)
 252		return 0;
 253
 254	for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i)
 255		if (atomic_read(&adev->uvd.handles[i]))
 256			break;
 257
 258	if (i == AMDGPU_MAX_UVD_HANDLES)
 259		return 0;
 260
 261	cancel_delayed_work_sync(&adev->uvd.idle_work);
 262
 263	size = amdgpu_bo_size(adev->uvd.vcpu_bo);
 264	ptr = adev->uvd.cpu_addr;
 265
 266	adev->uvd.saved_bo = kmalloc(size, GFP_KERNEL);
 267	if (!adev->uvd.saved_bo)
 268		return -ENOMEM;
 269
 270	memcpy(adev->uvd.saved_bo, ptr, size);
 271
 272	return 0;
 273}
 274
 275int amdgpu_uvd_resume(struct amdgpu_device *adev)
 276{
 277	unsigned size;
 278	void *ptr;
 279
 280	if (adev->uvd.vcpu_bo == NULL)
 281		return -EINVAL;
 282
 283	size = amdgpu_bo_size(adev->uvd.vcpu_bo);
 284	ptr = adev->uvd.cpu_addr;
 285
 286	if (adev->uvd.saved_bo != NULL) {
 287		memcpy(ptr, adev->uvd.saved_bo, size);
 288		kfree(adev->uvd.saved_bo);
 289		adev->uvd.saved_bo = NULL;
 290	} else {
 291		const struct common_firmware_header *hdr;
 292		unsigned offset;
 293
 294		hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
 295		offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
 296		memcpy(adev->uvd.cpu_addr, (adev->uvd.fw->data) + offset,
 297			(adev->uvd.fw->size) - offset);
 298		size -= le32_to_cpu(hdr->ucode_size_bytes);
 299		ptr += le32_to_cpu(hdr->ucode_size_bytes);
 300		memset(ptr, 0, size);
 301	}
 302
 303	return 0;
 304}
 305
 306void amdgpu_uvd_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
 307{
 308	struct amdgpu_ring *ring = &adev->uvd.ring;
 309	int i, r;
 310
 311	for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
 312		uint32_t handle = atomic_read(&adev->uvd.handles[i]);
 313		if (handle != 0 && adev->uvd.filp[i] == filp) {
 314			struct fence *fence;
 315
 316			amdgpu_uvd_note_usage(adev);
 317
 318			r = amdgpu_uvd_get_destroy_msg(ring, handle,
 319						       false, &fence);
 320			if (r) {
 321				DRM_ERROR("Error destroying UVD (%d)!\n", r);
 322				continue;
 323			}
 324
 325			fence_wait(fence, false);
 326			fence_put(fence);
 327
 328			adev->uvd.filp[i] = NULL;
 329			atomic_set(&adev->uvd.handles[i], 0);
 330		}
 331	}
 332}
 333
 334static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *rbo)
 335{
 336	int i;
 337	for (i = 0; i < rbo->placement.num_placement; ++i) {
 338		rbo->placements[i].fpfn = 0 >> PAGE_SHIFT;
 339		rbo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
 340	}
 341}
 342
 
 
 
 
 
 
 
 
 
 
 
 
 343/**
 344 * amdgpu_uvd_cs_pass1 - first parsing round
 345 *
 346 * @ctx: UVD parser context
 347 *
 348 * Make sure UVD message and feedback buffers are in VRAM and
 349 * nobody is violating an 256MB boundary.
 350 */
 351static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx *ctx)
 352{
 353	struct amdgpu_bo_va_mapping *mapping;
 354	struct amdgpu_bo *bo;
 355	uint32_t cmd, lo, hi;
 356	uint64_t addr;
 357	int r = 0;
 358
 359	lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
 360	hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
 361	addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
 362
 363	mapping = amdgpu_cs_find_mapping(ctx->parser, addr, &bo);
 364	if (mapping == NULL) {
 365		DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
 366		return -EINVAL;
 367	}
 368
 369	if (!ctx->parser->adev->uvd.address_64_bit) {
 370		/* check if it's a message or feedback command */
 371		cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
 372		if (cmd == 0x0 || cmd == 0x3) {
 373			/* yes, force it into VRAM */
 374			uint32_t domain = AMDGPU_GEM_DOMAIN_VRAM;
 375			amdgpu_ttm_placement_from_domain(bo, domain);
 376		}
 377		amdgpu_uvd_force_into_uvd_segment(bo);
 378
 379		r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
 380	}
 381
 382	return r;
 383}
 384
 385/**
 386 * amdgpu_uvd_cs_msg_decode - handle UVD decode message
 387 *
 388 * @msg: pointer to message structure
 389 * @buf_sizes: returned buffer sizes
 390 *
 391 * Peek into the decode message and calculate the necessary buffer sizes.
 392 */
 393static int amdgpu_uvd_cs_msg_decode(uint32_t *msg, unsigned buf_sizes[])
 
 394{
 395	unsigned stream_type = msg[4];
 396	unsigned width = msg[6];
 397	unsigned height = msg[7];
 398	unsigned dpb_size = msg[9];
 399	unsigned pitch = msg[28];
 400	unsigned level = msg[57];
 401
 402	unsigned width_in_mb = width / 16;
 403	unsigned height_in_mb = ALIGN(height / 16, 2);
 404	unsigned fs_in_mb = width_in_mb * height_in_mb;
 405
 406	unsigned image_size, tmp, min_dpb_size, num_dpb_buffer;
 407	unsigned min_ctx_size = 0;
 408
 409	image_size = width * height;
 410	image_size += image_size / 2;
 411	image_size = ALIGN(image_size, 1024);
 412
 413	switch (stream_type) {
 414	case 0: /* H264 */
 415	case 7: /* H264 Perf */
 416		switch(level) {
 417		case 30:
 418			num_dpb_buffer = 8100 / fs_in_mb;
 419			break;
 420		case 31:
 421			num_dpb_buffer = 18000 / fs_in_mb;
 422			break;
 423		case 32:
 424			num_dpb_buffer = 20480 / fs_in_mb;
 425			break;
 426		case 41:
 427			num_dpb_buffer = 32768 / fs_in_mb;
 428			break;
 429		case 42:
 430			num_dpb_buffer = 34816 / fs_in_mb;
 431			break;
 432		case 50:
 433			num_dpb_buffer = 110400 / fs_in_mb;
 434			break;
 435		case 51:
 436			num_dpb_buffer = 184320 / fs_in_mb;
 437			break;
 438		default:
 439			num_dpb_buffer = 184320 / fs_in_mb;
 440			break;
 441		}
 442		num_dpb_buffer++;
 443		if (num_dpb_buffer > 17)
 444			num_dpb_buffer = 17;
 445
 446		/* reference picture buffer */
 447		min_dpb_size = image_size * num_dpb_buffer;
 448
 449		/* macroblock context buffer */
 450		min_dpb_size += width_in_mb * height_in_mb * num_dpb_buffer * 192;
 451
 452		/* IT surface buffer */
 453		min_dpb_size += width_in_mb * height_in_mb * 32;
 454		break;
 455
 456	case 1: /* VC1 */
 457
 458		/* reference picture buffer */
 459		min_dpb_size = image_size * 3;
 460
 461		/* CONTEXT_BUFFER */
 462		min_dpb_size += width_in_mb * height_in_mb * 128;
 463
 464		/* IT surface buffer */
 465		min_dpb_size += width_in_mb * 64;
 466
 467		/* DB surface buffer */
 468		min_dpb_size += width_in_mb * 128;
 469
 470		/* BP */
 471		tmp = max(width_in_mb, height_in_mb);
 472		min_dpb_size += ALIGN(tmp * 7 * 16, 64);
 473		break;
 474
 475	case 3: /* MPEG2 */
 476
 477		/* reference picture buffer */
 478		min_dpb_size = image_size * 3;
 479		break;
 480
 481	case 4: /* MPEG4 */
 482
 483		/* reference picture buffer */
 484		min_dpb_size = image_size * 3;
 485
 486		/* CM */
 487		min_dpb_size += width_in_mb * height_in_mb * 64;
 488
 489		/* IT surface buffer */
 490		min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
 491		break;
 492
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 493	case 16: /* H265 */
 494		image_size = (ALIGN(width, 16) * ALIGN(height, 16) * 3) / 2;
 495		image_size = ALIGN(image_size, 256);
 496
 497		num_dpb_buffer = (le32_to_cpu(msg[59]) & 0xff) + 2;
 498		min_dpb_size = image_size * num_dpb_buffer;
 499		min_ctx_size = ((width + 255) / 16) * ((height + 255) / 16)
 500					   * 16 * num_dpb_buffer + 52 * 1024;
 501		break;
 502
 503	default:
 504		DRM_ERROR("UVD codec not handled %d!\n", stream_type);
 505		return -EINVAL;
 506	}
 507
 508	if (width > pitch) {
 509		DRM_ERROR("Invalid UVD decoding target pitch!\n");
 510		return -EINVAL;
 511	}
 512
 513	if (dpb_size < min_dpb_size) {
 514		DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
 515			  dpb_size, min_dpb_size);
 516		return -EINVAL;
 517	}
 518
 519	buf_sizes[0x1] = dpb_size;
 520	buf_sizes[0x2] = image_size;
 521	buf_sizes[0x4] = min_ctx_size;
 522	return 0;
 523}
 524
 525/**
 526 * amdgpu_uvd_cs_msg - handle UVD message
 527 *
 528 * @ctx: UVD parser context
 529 * @bo: buffer object containing the message
 530 * @offset: offset into the buffer object
 531 *
 532 * Peek into the UVD message and extract the session id.
 533 * Make sure that we don't open up to many sessions.
 534 */
 535static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx *ctx,
 536			     struct amdgpu_bo *bo, unsigned offset)
 537{
 538	struct amdgpu_device *adev = ctx->parser->adev;
 539	int32_t *msg, msg_type, handle;
 540	void *ptr;
 541	long r;
 542	int i;
 543
 544	if (offset & 0x3F) {
 545		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
 546		return -EINVAL;
 547	}
 548
 549	r = amdgpu_bo_kmap(bo, &ptr);
 550	if (r) {
 551		DRM_ERROR("Failed mapping the UVD message (%ld)!\n", r);
 552		return r;
 553	}
 554
 555	msg = ptr + offset;
 556
 557	msg_type = msg[1];
 558	handle = msg[2];
 559
 560	if (handle == 0) {
 561		DRM_ERROR("Invalid UVD handle!\n");
 562		return -EINVAL;
 563	}
 564
 565	switch (msg_type) {
 566	case 0:
 567		/* it's a create msg, calc image size (width * height) */
 568		amdgpu_bo_kunmap(bo);
 569
 570		/* try to alloc a new handle */
 571		for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
 572			if (atomic_read(&adev->uvd.handles[i]) == handle) {
 573				DRM_ERROR("Handle 0x%x already in use!\n", handle);
 574				return -EINVAL;
 575			}
 576
 577			if (!atomic_cmpxchg(&adev->uvd.handles[i], 0, handle)) {
 578				adev->uvd.filp[i] = ctx->parser->filp;
 579				return 0;
 580			}
 581		}
 582
 583		DRM_ERROR("No more free UVD handles!\n");
 584		return -EINVAL;
 585
 586	case 1:
 587		/* it's a decode msg, calc buffer sizes */
 588		r = amdgpu_uvd_cs_msg_decode(msg, ctx->buf_sizes);
 589		amdgpu_bo_kunmap(bo);
 590		if (r)
 591			return r;
 592
 593		/* validate the handle */
 594		for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i) {
 595			if (atomic_read(&adev->uvd.handles[i]) == handle) {
 596				if (adev->uvd.filp[i] != ctx->parser->filp) {
 597					DRM_ERROR("UVD handle collision detected!\n");
 598					return -EINVAL;
 599				}
 600				return 0;
 601			}
 602		}
 603
 604		DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
 605		return -ENOENT;
 606
 607	case 2:
 608		/* it's a destroy msg, free the handle */
 609		for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i)
 610			atomic_cmpxchg(&adev->uvd.handles[i], handle, 0);
 611		amdgpu_bo_kunmap(bo);
 612		return 0;
 613
 614	default:
 615		DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
 616		return -EINVAL;
 617	}
 618	BUG();
 619	return -EINVAL;
 620}
 621
 622/**
 623 * amdgpu_uvd_cs_pass2 - second parsing round
 624 *
 625 * @ctx: UVD parser context
 626 *
 627 * Patch buffer addresses, make sure buffer sizes are correct.
 628 */
 629static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx *ctx)
 630{
 631	struct amdgpu_bo_va_mapping *mapping;
 632	struct amdgpu_bo *bo;
 633	uint32_t cmd, lo, hi;
 634	uint64_t start, end;
 635	uint64_t addr;
 636	int r;
 637
 638	lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
 639	hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
 640	addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
 641
 642	mapping = amdgpu_cs_find_mapping(ctx->parser, addr, &bo);
 643	if (mapping == NULL)
 
 644		return -EINVAL;
 
 645
 646	start = amdgpu_bo_gpu_offset(bo);
 647
 648	end = (mapping->it.last + 1 - mapping->it.start);
 649	end = end * AMDGPU_GPU_PAGE_SIZE + start;
 650
 651	addr -= ((uint64_t)mapping->it.start) * AMDGPU_GPU_PAGE_SIZE;
 652	start += addr;
 653
 654	amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data0,
 655			    lower_32_bits(start));
 656	amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data1,
 657			    upper_32_bits(start));
 658
 659	cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
 660	if (cmd < 0x4) {
 661		if ((end - start) < ctx->buf_sizes[cmd]) {
 662			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
 663				  (unsigned)(end - start),
 664				  ctx->buf_sizes[cmd]);
 665			return -EINVAL;
 666		}
 667
 668	} else if (cmd == 0x206) {
 669		if ((end - start) < ctx->buf_sizes[4]) {
 670			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
 671					  (unsigned)(end - start),
 672					  ctx->buf_sizes[4]);
 673			return -EINVAL;
 674		}
 675	} else if ((cmd != 0x100) && (cmd != 0x204)) {
 676		DRM_ERROR("invalid UVD command %X!\n", cmd);
 677		return -EINVAL;
 678	}
 679
 680	if (!ctx->parser->adev->uvd.address_64_bit) {
 681		if ((start >> 28) != ((end - 1) >> 28)) {
 682			DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
 683				  start, end);
 684			return -EINVAL;
 685		}
 686
 687		if ((cmd == 0 || cmd == 0x3) &&
 688		    (start >> 28) != (ctx->parser->adev->uvd.gpu_addr >> 28)) {
 689			DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
 690				  start, end);
 691			return -EINVAL;
 692		}
 693	}
 694
 695	if (cmd == 0) {
 696		ctx->has_msg_cmd = true;
 697		r = amdgpu_uvd_cs_msg(ctx, bo, addr);
 698		if (r)
 699			return r;
 700	} else if (!ctx->has_msg_cmd) {
 701		DRM_ERROR("Message needed before other commands are send!\n");
 702		return -EINVAL;
 703	}
 704
 705	return 0;
 706}
 707
 708/**
 709 * amdgpu_uvd_cs_reg - parse register writes
 710 *
 711 * @ctx: UVD parser context
 712 * @cb: callback function
 713 *
 714 * Parse the register writes, call cb on each complete command.
 715 */
 716static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx *ctx,
 717			     int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
 718{
 719	struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
 720	int i, r;
 721
 722	ctx->idx++;
 723	for (i = 0; i <= ctx->count; ++i) {
 724		unsigned reg = ctx->reg + i;
 725
 726		if (ctx->idx >= ib->length_dw) {
 727			DRM_ERROR("Register command after end of CS!\n");
 728			return -EINVAL;
 729		}
 730
 731		switch (reg) {
 732		case mmUVD_GPCOM_VCPU_DATA0:
 733			ctx->data0 = ctx->idx;
 734			break;
 735		case mmUVD_GPCOM_VCPU_DATA1:
 736			ctx->data1 = ctx->idx;
 737			break;
 738		case mmUVD_GPCOM_VCPU_CMD:
 739			r = cb(ctx);
 740			if (r)
 741				return r;
 742			break;
 743		case mmUVD_ENGINE_CNTL:
 
 744			break;
 745		default:
 746			DRM_ERROR("Invalid reg 0x%X!\n", reg);
 747			return -EINVAL;
 748		}
 749		ctx->idx++;
 750	}
 751	return 0;
 752}
 753
 754/**
 755 * amdgpu_uvd_cs_packets - parse UVD packets
 756 *
 757 * @ctx: UVD parser context
 758 * @cb: callback function
 759 *
 760 * Parse the command stream packets.
 761 */
 762static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx *ctx,
 763				 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
 764{
 765	struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
 766	int r;
 767
 768	for (ctx->idx = 0 ; ctx->idx < ib->length_dw; ) {
 769		uint32_t cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx);
 770		unsigned type = CP_PACKET_GET_TYPE(cmd);
 771		switch (type) {
 772		case PACKET_TYPE0:
 773			ctx->reg = CP_PACKET0_GET_REG(cmd);
 774			ctx->count = CP_PACKET_GET_COUNT(cmd);
 775			r = amdgpu_uvd_cs_reg(ctx, cb);
 776			if (r)
 777				return r;
 778			break;
 779		case PACKET_TYPE2:
 780			++ctx->idx;
 781			break;
 782		default:
 783			DRM_ERROR("Unknown packet type %d !\n", type);
 784			return -EINVAL;
 785		}
 786	}
 787	return 0;
 788}
 789
 790/**
 791 * amdgpu_uvd_ring_parse_cs - UVD command submission parser
 792 *
 793 * @parser: Command submission parser context
 794 *
 795 * Parse the command stream, patch in addresses as necessary.
 796 */
 797int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser, uint32_t ib_idx)
 798{
 799	struct amdgpu_uvd_cs_ctx ctx = {};
 800	unsigned buf_sizes[] = {
 801		[0x00000000]	=	2048,
 802		[0x00000001]	=	0xFFFFFFFF,
 803		[0x00000002]	=	0xFFFFFFFF,
 804		[0x00000003]	=	2048,
 805		[0x00000004]	=	0xFFFFFFFF,
 806	};
 807	struct amdgpu_ib *ib = &parser->job->ibs[ib_idx];
 808	int r;
 809
 
 
 
 810	if (ib->length_dw % 16) {
 811		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
 812			  ib->length_dw);
 813		return -EINVAL;
 814	}
 815
 
 
 
 
 816	ctx.parser = parser;
 817	ctx.buf_sizes = buf_sizes;
 818	ctx.ib_idx = ib_idx;
 819
 820	/* first round, make sure the buffers are actually in the UVD segment */
 821	r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass1);
 822	if (r)
 823		return r;
 
 
 
 824
 825	/* second round, patch buffer addresses into the command stream */
 826	r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass2);
 827	if (r)
 828		return r;
 829
 830	if (!ctx.has_msg_cmd) {
 831		DRM_ERROR("UVD-IBs need a msg command!\n");
 832		return -EINVAL;
 833	}
 834
 835	amdgpu_uvd_note_usage(ctx.parser->adev);
 836
 837	return 0;
 838}
 839
 840static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
 841			       bool direct, struct fence **fence)
 842{
 843	struct ttm_validate_buffer tv;
 844	struct ww_acquire_ctx ticket;
 845	struct list_head head;
 846	struct amdgpu_job *job;
 847	struct amdgpu_ib *ib;
 848	struct fence *f = NULL;
 849	struct amdgpu_device *adev = ring->adev;
 850	uint64_t addr;
 851	int i, r;
 852
 853	memset(&tv, 0, sizeof(tv));
 854	tv.bo = &bo->tbo;
 855
 856	INIT_LIST_HEAD(&head);
 857	list_add(&tv.head, &head);
 858
 859	r = ttm_eu_reserve_buffers(&ticket, &head, true, NULL);
 860	if (r)
 861		return r;
 862
 863	if (!bo->adev->uvd.address_64_bit) {
 864		amdgpu_ttm_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
 865		amdgpu_uvd_force_into_uvd_segment(bo);
 866	}
 867
 868	r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
 869	if (r)
 870		goto err;
 871
 872	r = amdgpu_job_alloc_with_ib(adev, 64, &job);
 873	if (r)
 874		goto err;
 875
 876	ib = &job->ibs[0];
 877	addr = amdgpu_bo_gpu_offset(bo);
 878	ib->ptr[0] = PACKET0(mmUVD_GPCOM_VCPU_DATA0, 0);
 879	ib->ptr[1] = addr;
 880	ib->ptr[2] = PACKET0(mmUVD_GPCOM_VCPU_DATA1, 0);
 881	ib->ptr[3] = addr >> 32;
 882	ib->ptr[4] = PACKET0(mmUVD_GPCOM_VCPU_CMD, 0);
 883	ib->ptr[5] = 0;
 884	for (i = 6; i < 16; ++i)
 885		ib->ptr[i] = PACKET2(0);
 
 
 886	ib->length_dw = 16;
 887
 888	if (direct) {
 889		r = amdgpu_ib_schedule(ring, 1, ib, NULL, &f);
 890		job->fence = f;
 891		if (r)
 892			goto err_free;
 893
 894		amdgpu_job_free(job);
 895	} else {
 896		r = amdgpu_job_submit(job, ring, &adev->uvd.entity,
 897				      AMDGPU_FENCE_OWNER_UNDEFINED, &f);
 898		if (r)
 899			goto err_free;
 900	}
 901
 902	ttm_eu_fence_buffer_objects(&ticket, &head, f);
 903
 904	if (fence)
 905		*fence = fence_get(f);
 906	amdgpu_bo_unref(&bo);
 907	fence_put(f);
 908
 909	return 0;
 910
 911err_free:
 912	amdgpu_job_free(job);
 913
 914err:
 915	ttm_eu_backoff_reservation(&ticket, &head);
 916	return r;
 917}
 918
 919/* multiple fence commands without any stream commands in between can
 920   crash the vcpu so just try to emmit a dummy create/destroy msg to
 921   avoid this */
 922int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
 923			      struct fence **fence)
 924{
 925	struct amdgpu_device *adev = ring->adev;
 926	struct amdgpu_bo *bo;
 927	uint32_t *msg;
 928	int r, i;
 929
 930	r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
 931			     AMDGPU_GEM_DOMAIN_VRAM,
 932			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
 
 933			     NULL, NULL, &bo);
 934	if (r)
 935		return r;
 936
 937	r = amdgpu_bo_reserve(bo, false);
 938	if (r) {
 939		amdgpu_bo_unref(&bo);
 940		return r;
 941	}
 942
 943	r = amdgpu_bo_kmap(bo, (void **)&msg);
 944	if (r) {
 945		amdgpu_bo_unreserve(bo);
 946		amdgpu_bo_unref(&bo);
 947		return r;
 948	}
 949
 950	/* stitch together an UVD create msg */
 951	msg[0] = cpu_to_le32(0x00000de4);
 952	msg[1] = cpu_to_le32(0x00000000);
 953	msg[2] = cpu_to_le32(handle);
 954	msg[3] = cpu_to_le32(0x00000000);
 955	msg[4] = cpu_to_le32(0x00000000);
 956	msg[5] = cpu_to_le32(0x00000000);
 957	msg[6] = cpu_to_le32(0x00000000);
 958	msg[7] = cpu_to_le32(0x00000780);
 959	msg[8] = cpu_to_le32(0x00000440);
 960	msg[9] = cpu_to_le32(0x00000000);
 961	msg[10] = cpu_to_le32(0x01b37000);
 962	for (i = 11; i < 1024; ++i)
 963		msg[i] = cpu_to_le32(0x0);
 964
 965	amdgpu_bo_kunmap(bo);
 966	amdgpu_bo_unreserve(bo);
 967
 968	return amdgpu_uvd_send_msg(ring, bo, true, fence);
 969}
 970
 971int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
 972			       bool direct, struct fence **fence)
 973{
 974	struct amdgpu_device *adev = ring->adev;
 975	struct amdgpu_bo *bo;
 976	uint32_t *msg;
 977	int r, i;
 978
 979	r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
 980			     AMDGPU_GEM_DOMAIN_VRAM,
 981			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
 
 982			     NULL, NULL, &bo);
 983	if (r)
 984		return r;
 985
 986	r = amdgpu_bo_reserve(bo, false);
 987	if (r) {
 988		amdgpu_bo_unref(&bo);
 989		return r;
 990	}
 991
 992	r = amdgpu_bo_kmap(bo, (void **)&msg);
 993	if (r) {
 994		amdgpu_bo_unreserve(bo);
 995		amdgpu_bo_unref(&bo);
 996		return r;
 997	}
 998
 999	/* stitch together an UVD destroy msg */
1000	msg[0] = cpu_to_le32(0x00000de4);
1001	msg[1] = cpu_to_le32(0x00000002);
1002	msg[2] = cpu_to_le32(handle);
1003	msg[3] = cpu_to_le32(0x00000000);
1004	for (i = 4; i < 1024; ++i)
1005		msg[i] = cpu_to_le32(0x0);
1006
1007	amdgpu_bo_kunmap(bo);
1008	amdgpu_bo_unreserve(bo);
1009
1010	return amdgpu_uvd_send_msg(ring, bo, direct, fence);
1011}
1012
1013static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
1014{
1015	struct amdgpu_device *adev =
1016		container_of(work, struct amdgpu_device, uvd.idle_work.work);
1017	unsigned i, fences, handles = 0;
1018
1019	fences = amdgpu_fence_count_emitted(&adev->uvd.ring);
1020
1021	for (i = 0; i < AMDGPU_MAX_UVD_HANDLES; ++i)
1022		if (atomic_read(&adev->uvd.handles[i]))
1023			++handles;
1024
1025	if (fences == 0 && handles == 0) {
1026		if (adev->pm.dpm_enabled) {
1027			amdgpu_dpm_enable_uvd(adev, false);
1028		} else {
1029			amdgpu_asic_set_uvd_clocks(adev, 0, 0);
1030		}
1031	} else {
1032		schedule_delayed_work(&adev->uvd.idle_work,
1033				      msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
1034	}
1035}
1036
1037static void amdgpu_uvd_note_usage(struct amdgpu_device *adev)
1038{
 
1039	bool set_clocks = !cancel_delayed_work_sync(&adev->uvd.idle_work);
1040	set_clocks &= schedule_delayed_work(&adev->uvd.idle_work,
1041					    msecs_to_jiffies(UVD_IDLE_TIMEOUT_MS));
1042
1043	if (set_clocks) {
1044		if (adev->pm.dpm_enabled) {
1045			amdgpu_dpm_enable_uvd(adev, true);
1046		} else {
1047			amdgpu_asic_set_uvd_clocks(adev, 53300, 40000);
1048		}
1049	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1050}
v4.10.11
   1/*
   2 * Copyright 2011 Advanced Micro Devices, Inc.
   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
   7 * "Software"), to deal in the Software without restriction, including
   8 * without limitation the rights to use, copy, modify, merge, publish,
   9 * distribute, sub license, and/or sell copies of the Software, and to
  10 * permit persons to whom the Software is furnished to do so, subject to
  11 * the following conditions:
  12 *
  13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  15 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  16 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
  17 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  18 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  19 * USE OR OTHER DEALINGS IN THE SOFTWARE.
  20 *
  21 * The above copyright notice and this permission notice (including the
  22 * next paragraph) shall be included in all copies or substantial portions
  23 * of the Software.
  24 *
  25 */
  26/*
  27 * Authors:
  28 *    Christian König <deathsimple@vodafone.de>
  29 */
  30
  31#include <linux/firmware.h>
  32#include <linux/module.h>
  33#include <drm/drmP.h>
  34#include <drm/drm.h>
  35
  36#include "amdgpu.h"
  37#include "amdgpu_pm.h"
  38#include "amdgpu_uvd.h"
  39#include "cikd.h"
  40#include "uvd/uvd_4_2_d.h"
  41
  42/* 1 second timeout */
  43#define UVD_IDLE_TIMEOUT	msecs_to_jiffies(1000)
  44
  45/* Firmware versions for VI */
  46#define FW_1_65_10	((1 << 24) | (65 << 16) | (10 << 8))
  47#define FW_1_87_11	((1 << 24) | (87 << 16) | (11 << 8))
  48#define FW_1_87_12	((1 << 24) | (87 << 16) | (12 << 8))
  49#define FW_1_37_15	((1 << 24) | (37 << 16) | (15 << 8))
  50
  51/* Polaris10/11 firmware version */
  52#define FW_1_66_16	((1 << 24) | (66 << 16) | (16 << 8))
  53
  54/* Firmware Names */
  55#ifdef CONFIG_DRM_AMDGPU_CIK
  56#define FIRMWARE_BONAIRE	"radeon/bonaire_uvd.bin"
  57#define FIRMWARE_KABINI	"radeon/kabini_uvd.bin"
  58#define FIRMWARE_KAVERI	"radeon/kaveri_uvd.bin"
  59#define FIRMWARE_HAWAII	"radeon/hawaii_uvd.bin"
  60#define FIRMWARE_MULLINS	"radeon/mullins_uvd.bin"
  61#endif
  62#define FIRMWARE_TONGA		"amdgpu/tonga_uvd.bin"
  63#define FIRMWARE_CARRIZO	"amdgpu/carrizo_uvd.bin"
  64#define FIRMWARE_FIJI		"amdgpu/fiji_uvd.bin"
  65#define FIRMWARE_STONEY		"amdgpu/stoney_uvd.bin"
  66#define FIRMWARE_POLARIS10	"amdgpu/polaris10_uvd.bin"
  67#define FIRMWARE_POLARIS11	"amdgpu/polaris11_uvd.bin"
  68#define FIRMWARE_POLARIS12	"amdgpu/polaris12_uvd.bin"
  69
  70/**
  71 * amdgpu_uvd_cs_ctx - Command submission parser context
  72 *
  73 * Used for emulating virtual memory support on UVD 4.2.
  74 */
  75struct amdgpu_uvd_cs_ctx {
  76	struct amdgpu_cs_parser *parser;
  77	unsigned reg, count;
  78	unsigned data0, data1;
  79	unsigned idx;
  80	unsigned ib_idx;
  81
  82	/* does the IB has a msg command */
  83	bool has_msg_cmd;
  84
  85	/* minimum buffer sizes */
  86	unsigned *buf_sizes;
  87};
  88
  89#ifdef CONFIG_DRM_AMDGPU_CIK
  90MODULE_FIRMWARE(FIRMWARE_BONAIRE);
  91MODULE_FIRMWARE(FIRMWARE_KABINI);
  92MODULE_FIRMWARE(FIRMWARE_KAVERI);
  93MODULE_FIRMWARE(FIRMWARE_HAWAII);
  94MODULE_FIRMWARE(FIRMWARE_MULLINS);
  95#endif
  96MODULE_FIRMWARE(FIRMWARE_TONGA);
  97MODULE_FIRMWARE(FIRMWARE_CARRIZO);
  98MODULE_FIRMWARE(FIRMWARE_FIJI);
  99MODULE_FIRMWARE(FIRMWARE_STONEY);
 100MODULE_FIRMWARE(FIRMWARE_POLARIS10);
 101MODULE_FIRMWARE(FIRMWARE_POLARIS11);
 102MODULE_FIRMWARE(FIRMWARE_POLARIS12);
 103
 
 104static void amdgpu_uvd_idle_work_handler(struct work_struct *work);
 105
 106int amdgpu_uvd_sw_init(struct amdgpu_device *adev)
 107{
 108	struct amdgpu_ring *ring;
 109	struct amd_sched_rq *rq;
 110	unsigned long bo_size;
 111	const char *fw_name;
 112	const struct common_firmware_header *hdr;
 113	unsigned version_major, version_minor, family_id;
 114	int i, r;
 115
 116	INIT_DELAYED_WORK(&adev->uvd.idle_work, amdgpu_uvd_idle_work_handler);
 117
 118	switch (adev->asic_type) {
 119#ifdef CONFIG_DRM_AMDGPU_CIK
 120	case CHIP_BONAIRE:
 121		fw_name = FIRMWARE_BONAIRE;
 122		break;
 123	case CHIP_KABINI:
 124		fw_name = FIRMWARE_KABINI;
 125		break;
 126	case CHIP_KAVERI:
 127		fw_name = FIRMWARE_KAVERI;
 128		break;
 129	case CHIP_HAWAII:
 130		fw_name = FIRMWARE_HAWAII;
 131		break;
 132	case CHIP_MULLINS:
 133		fw_name = FIRMWARE_MULLINS;
 134		break;
 135#endif
 136	case CHIP_TONGA:
 137		fw_name = FIRMWARE_TONGA;
 138		break;
 139	case CHIP_FIJI:
 140		fw_name = FIRMWARE_FIJI;
 141		break;
 142	case CHIP_CARRIZO:
 143		fw_name = FIRMWARE_CARRIZO;
 144		break;
 145	case CHIP_STONEY:
 146		fw_name = FIRMWARE_STONEY;
 147		break;
 148	case CHIP_POLARIS10:
 149		fw_name = FIRMWARE_POLARIS10;
 150		break;
 151	case CHIP_POLARIS11:
 152		fw_name = FIRMWARE_POLARIS11;
 153		break;
 154	case CHIP_POLARIS12:
 155		fw_name = FIRMWARE_POLARIS12;
 156		break;
 157	default:
 158		return -EINVAL;
 159	}
 160
 161	r = request_firmware(&adev->uvd.fw, fw_name, adev->dev);
 162	if (r) {
 163		dev_err(adev->dev, "amdgpu_uvd: Can't load firmware \"%s\"\n",
 164			fw_name);
 165		return r;
 166	}
 167
 168	r = amdgpu_ucode_validate(adev->uvd.fw);
 169	if (r) {
 170		dev_err(adev->dev, "amdgpu_uvd: Can't validate firmware \"%s\"\n",
 171			fw_name);
 172		release_firmware(adev->uvd.fw);
 173		adev->uvd.fw = NULL;
 174		return r;
 175	}
 176
 177	/* Set the default UVD handles that the firmware can handle */
 178	adev->uvd.max_handles = AMDGPU_DEFAULT_UVD_HANDLES;
 179
 180	hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
 181	family_id = le32_to_cpu(hdr->ucode_version) & 0xff;
 182	version_major = (le32_to_cpu(hdr->ucode_version) >> 24) & 0xff;
 183	version_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
 184	DRM_INFO("Found UVD firmware Version: %hu.%hu Family ID: %hu\n",
 185		version_major, version_minor, family_id);
 186
 187	/*
 188	 * Limit the number of UVD handles depending on microcode major
 189	 * and minor versions. The firmware version which has 40 UVD
 190	 * instances support is 1.80. So all subsequent versions should
 191	 * also have the same support.
 192	 */
 193	if ((version_major > 0x01) ||
 194	    ((version_major == 0x01) && (version_minor >= 0x50)))
 195		adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
 196
 197	adev->uvd.fw_version = ((version_major << 24) | (version_minor << 16) |
 198				(family_id << 8));
 199
 200	if ((adev->asic_type == CHIP_POLARIS10 ||
 201	     adev->asic_type == CHIP_POLARIS11) &&
 202	    (adev->uvd.fw_version < FW_1_66_16))
 203		DRM_ERROR("POLARIS10/11 UVD firmware version %hu.%hu is too old.\n",
 204			  version_major, version_minor);
 205
 206	bo_size = AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr->ucode_size_bytes) + 8)
 207		  +  AMDGPU_UVD_STACK_SIZE + AMDGPU_UVD_HEAP_SIZE
 208		  +  AMDGPU_UVD_SESSION_SIZE * adev->uvd.max_handles;
 209	r = amdgpu_bo_create_kernel(adev, bo_size, PAGE_SIZE,
 210				    AMDGPU_GEM_DOMAIN_VRAM, &adev->uvd.vcpu_bo,
 211				    &adev->uvd.gpu_addr, &adev->uvd.cpu_addr);
 212	if (r) {
 213		dev_err(adev->dev, "(%d) failed to allocate UVD bo\n", r);
 214		return r;
 215	}
 216
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 217	ring = &adev->uvd.ring;
 218	rq = &ring->sched.sched_rq[AMD_SCHED_PRIORITY_NORMAL];
 219	r = amd_sched_entity_init(&ring->sched, &adev->uvd.entity,
 220				  rq, amdgpu_sched_jobs);
 221	if (r != 0) {
 222		DRM_ERROR("Failed setting up UVD run queue.\n");
 223		return r;
 224	}
 225
 226	for (i = 0; i < adev->uvd.max_handles; ++i) {
 227		atomic_set(&adev->uvd.handles[i], 0);
 228		adev->uvd.filp[i] = NULL;
 229	}
 230
 231	/* from uvd v5.0 HW addressing capacity increased to 64 bits */
 232	if (!amdgpu_ip_block_version_cmp(adev, AMD_IP_BLOCK_TYPE_UVD, 5, 0))
 233		adev->uvd.address_64_bit = true;
 234
 235	switch (adev->asic_type) {
 236	case CHIP_TONGA:
 237		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_65_10;
 238		break;
 239	case CHIP_CARRIZO:
 240		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_11;
 241		break;
 242	case CHIP_FIJI:
 243		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_12;
 244		break;
 245	case CHIP_STONEY:
 246		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_37_15;
 247		break;
 248	default:
 249		adev->uvd.use_ctx_buf = adev->asic_type >= CHIP_POLARIS10;
 250	}
 251
 252	return 0;
 253}
 254
 255int amdgpu_uvd_sw_fini(struct amdgpu_device *adev)
 256{
 257	kfree(adev->uvd.saved_bo);
 
 
 
 258
 259	amd_sched_entity_fini(&adev->uvd.ring.sched, &adev->uvd.entity);
 260
 261	amdgpu_bo_free_kernel(&adev->uvd.vcpu_bo,
 262			      &adev->uvd.gpu_addr,
 263			      (void **)&adev->uvd.cpu_addr);
 
 
 
 
 
 264
 265	amdgpu_ring_fini(&adev->uvd.ring);
 266
 267	release_firmware(adev->uvd.fw);
 268
 269	return 0;
 270}
 271
 272int amdgpu_uvd_suspend(struct amdgpu_device *adev)
 273{
 274	unsigned size;
 275	void *ptr;
 276	int i;
 277
 278	if (adev->uvd.vcpu_bo == NULL)
 279		return 0;
 280
 281	for (i = 0; i < adev->uvd.max_handles; ++i)
 282		if (atomic_read(&adev->uvd.handles[i]))
 283			break;
 284
 285	if (i == AMDGPU_MAX_UVD_HANDLES)
 286		return 0;
 287
 288	cancel_delayed_work_sync(&adev->uvd.idle_work);
 289
 290	size = amdgpu_bo_size(adev->uvd.vcpu_bo);
 291	ptr = adev->uvd.cpu_addr;
 292
 293	adev->uvd.saved_bo = kmalloc(size, GFP_KERNEL);
 294	if (!adev->uvd.saved_bo)
 295		return -ENOMEM;
 296
 297	memcpy_fromio(adev->uvd.saved_bo, ptr, size);
 298
 299	return 0;
 300}
 301
 302int amdgpu_uvd_resume(struct amdgpu_device *adev)
 303{
 304	unsigned size;
 305	void *ptr;
 306
 307	if (adev->uvd.vcpu_bo == NULL)
 308		return -EINVAL;
 309
 310	size = amdgpu_bo_size(adev->uvd.vcpu_bo);
 311	ptr = adev->uvd.cpu_addr;
 312
 313	if (adev->uvd.saved_bo != NULL) {
 314		memcpy_toio(ptr, adev->uvd.saved_bo, size);
 315		kfree(adev->uvd.saved_bo);
 316		adev->uvd.saved_bo = NULL;
 317	} else {
 318		const struct common_firmware_header *hdr;
 319		unsigned offset;
 320
 321		hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
 322		offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
 323		memcpy_toio(adev->uvd.cpu_addr, adev->uvd.fw->data + offset,
 324			    le32_to_cpu(hdr->ucode_size_bytes));
 325		size -= le32_to_cpu(hdr->ucode_size_bytes);
 326		ptr += le32_to_cpu(hdr->ucode_size_bytes);
 327		memset_io(ptr, 0, size);
 328	}
 329
 330	return 0;
 331}
 332
 333void amdgpu_uvd_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
 334{
 335	struct amdgpu_ring *ring = &adev->uvd.ring;
 336	int i, r;
 337
 338	for (i = 0; i < adev->uvd.max_handles; ++i) {
 339		uint32_t handle = atomic_read(&adev->uvd.handles[i]);
 340		if (handle != 0 && adev->uvd.filp[i] == filp) {
 341			struct dma_fence *fence;
 
 
 342
 343			r = amdgpu_uvd_get_destroy_msg(ring, handle,
 344						       false, &fence);
 345			if (r) {
 346				DRM_ERROR("Error destroying UVD (%d)!\n", r);
 347				continue;
 348			}
 349
 350			dma_fence_wait(fence, false);
 351			dma_fence_put(fence);
 352
 353			adev->uvd.filp[i] = NULL;
 354			atomic_set(&adev->uvd.handles[i], 0);
 355		}
 356	}
 357}
 358
 359static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *abo)
 360{
 361	int i;
 362	for (i = 0; i < abo->placement.num_placement; ++i) {
 363		abo->placements[i].fpfn = 0 >> PAGE_SHIFT;
 364		abo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
 365	}
 366}
 367
 368static u64 amdgpu_uvd_get_addr_from_ctx(struct amdgpu_uvd_cs_ctx *ctx)
 369{
 370	uint32_t lo, hi;
 371	uint64_t addr;
 372
 373	lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
 374	hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
 375	addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
 376
 377	return addr;
 378}
 379
 380/**
 381 * amdgpu_uvd_cs_pass1 - first parsing round
 382 *
 383 * @ctx: UVD parser context
 384 *
 385 * Make sure UVD message and feedback buffers are in VRAM and
 386 * nobody is violating an 256MB boundary.
 387 */
 388static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx *ctx)
 389{
 390	struct amdgpu_bo_va_mapping *mapping;
 391	struct amdgpu_bo *bo;
 392	uint32_t cmd;
 393	uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
 394	int r = 0;
 395
 
 
 
 
 396	mapping = amdgpu_cs_find_mapping(ctx->parser, addr, &bo);
 397	if (mapping == NULL) {
 398		DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
 399		return -EINVAL;
 400	}
 401
 402	if (!ctx->parser->adev->uvd.address_64_bit) {
 403		/* check if it's a message or feedback command */
 404		cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
 405		if (cmd == 0x0 || cmd == 0x3) {
 406			/* yes, force it into VRAM */
 407			uint32_t domain = AMDGPU_GEM_DOMAIN_VRAM;
 408			amdgpu_ttm_placement_from_domain(bo, domain);
 409		}
 410		amdgpu_uvd_force_into_uvd_segment(bo);
 411
 412		r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
 413	}
 414
 415	return r;
 416}
 417
 418/**
 419 * amdgpu_uvd_cs_msg_decode - handle UVD decode message
 420 *
 421 * @msg: pointer to message structure
 422 * @buf_sizes: returned buffer sizes
 423 *
 424 * Peek into the decode message and calculate the necessary buffer sizes.
 425 */
 426static int amdgpu_uvd_cs_msg_decode(struct amdgpu_device *adev, uint32_t *msg,
 427	unsigned buf_sizes[])
 428{
 429	unsigned stream_type = msg[4];
 430	unsigned width = msg[6];
 431	unsigned height = msg[7];
 432	unsigned dpb_size = msg[9];
 433	unsigned pitch = msg[28];
 434	unsigned level = msg[57];
 435
 436	unsigned width_in_mb = width / 16;
 437	unsigned height_in_mb = ALIGN(height / 16, 2);
 438	unsigned fs_in_mb = width_in_mb * height_in_mb;
 439
 440	unsigned image_size, tmp, min_dpb_size, num_dpb_buffer;
 441	unsigned min_ctx_size = ~0;
 442
 443	image_size = width * height;
 444	image_size += image_size / 2;
 445	image_size = ALIGN(image_size, 1024);
 446
 447	switch (stream_type) {
 448	case 0: /* H264 */
 
 449		switch(level) {
 450		case 30:
 451			num_dpb_buffer = 8100 / fs_in_mb;
 452			break;
 453		case 31:
 454			num_dpb_buffer = 18000 / fs_in_mb;
 455			break;
 456		case 32:
 457			num_dpb_buffer = 20480 / fs_in_mb;
 458			break;
 459		case 41:
 460			num_dpb_buffer = 32768 / fs_in_mb;
 461			break;
 462		case 42:
 463			num_dpb_buffer = 34816 / fs_in_mb;
 464			break;
 465		case 50:
 466			num_dpb_buffer = 110400 / fs_in_mb;
 467			break;
 468		case 51:
 469			num_dpb_buffer = 184320 / fs_in_mb;
 470			break;
 471		default:
 472			num_dpb_buffer = 184320 / fs_in_mb;
 473			break;
 474		}
 475		num_dpb_buffer++;
 476		if (num_dpb_buffer > 17)
 477			num_dpb_buffer = 17;
 478
 479		/* reference picture buffer */
 480		min_dpb_size = image_size * num_dpb_buffer;
 481
 482		/* macroblock context buffer */
 483		min_dpb_size += width_in_mb * height_in_mb * num_dpb_buffer * 192;
 484
 485		/* IT surface buffer */
 486		min_dpb_size += width_in_mb * height_in_mb * 32;
 487		break;
 488
 489	case 1: /* VC1 */
 490
 491		/* reference picture buffer */
 492		min_dpb_size = image_size * 3;
 493
 494		/* CONTEXT_BUFFER */
 495		min_dpb_size += width_in_mb * height_in_mb * 128;
 496
 497		/* IT surface buffer */
 498		min_dpb_size += width_in_mb * 64;
 499
 500		/* DB surface buffer */
 501		min_dpb_size += width_in_mb * 128;
 502
 503		/* BP */
 504		tmp = max(width_in_mb, height_in_mb);
 505		min_dpb_size += ALIGN(tmp * 7 * 16, 64);
 506		break;
 507
 508	case 3: /* MPEG2 */
 509
 510		/* reference picture buffer */
 511		min_dpb_size = image_size * 3;
 512		break;
 513
 514	case 4: /* MPEG4 */
 515
 516		/* reference picture buffer */
 517		min_dpb_size = image_size * 3;
 518
 519		/* CM */
 520		min_dpb_size += width_in_mb * height_in_mb * 64;
 521
 522		/* IT surface buffer */
 523		min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
 524		break;
 525
 526	case 7: /* H264 Perf */
 527		switch(level) {
 528		case 30:
 529			num_dpb_buffer = 8100 / fs_in_mb;
 530			break;
 531		case 31:
 532			num_dpb_buffer = 18000 / fs_in_mb;
 533			break;
 534		case 32:
 535			num_dpb_buffer = 20480 / fs_in_mb;
 536			break;
 537		case 41:
 538			num_dpb_buffer = 32768 / fs_in_mb;
 539			break;
 540		case 42:
 541			num_dpb_buffer = 34816 / fs_in_mb;
 542			break;
 543		case 50:
 544			num_dpb_buffer = 110400 / fs_in_mb;
 545			break;
 546		case 51:
 547			num_dpb_buffer = 184320 / fs_in_mb;
 548			break;
 549		default:
 550			num_dpb_buffer = 184320 / fs_in_mb;
 551			break;
 552		}
 553		num_dpb_buffer++;
 554		if (num_dpb_buffer > 17)
 555			num_dpb_buffer = 17;
 556
 557		/* reference picture buffer */
 558		min_dpb_size = image_size * num_dpb_buffer;
 559
 560		if (!adev->uvd.use_ctx_buf){
 561			/* macroblock context buffer */
 562			min_dpb_size +=
 563				width_in_mb * height_in_mb * num_dpb_buffer * 192;
 564
 565			/* IT surface buffer */
 566			min_dpb_size += width_in_mb * height_in_mb * 32;
 567		} else {
 568			/* macroblock context buffer */
 569			min_ctx_size =
 570				width_in_mb * height_in_mb * num_dpb_buffer * 192;
 571		}
 572		break;
 573
 574	case 16: /* H265 */
 575		image_size = (ALIGN(width, 16) * ALIGN(height, 16) * 3) / 2;
 576		image_size = ALIGN(image_size, 256);
 577
 578		num_dpb_buffer = (le32_to_cpu(msg[59]) & 0xff) + 2;
 579		min_dpb_size = image_size * num_dpb_buffer;
 580		min_ctx_size = ((width + 255) / 16) * ((height + 255) / 16)
 581					   * 16 * num_dpb_buffer + 52 * 1024;
 582		break;
 583
 584	default:
 585		DRM_ERROR("UVD codec not handled %d!\n", stream_type);
 586		return -EINVAL;
 587	}
 588
 589	if (width > pitch) {
 590		DRM_ERROR("Invalid UVD decoding target pitch!\n");
 591		return -EINVAL;
 592	}
 593
 594	if (dpb_size < min_dpb_size) {
 595		DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
 596			  dpb_size, min_dpb_size);
 597		return -EINVAL;
 598	}
 599
 600	buf_sizes[0x1] = dpb_size;
 601	buf_sizes[0x2] = image_size;
 602	buf_sizes[0x4] = min_ctx_size;
 603	return 0;
 604}
 605
 606/**
 607 * amdgpu_uvd_cs_msg - handle UVD message
 608 *
 609 * @ctx: UVD parser context
 610 * @bo: buffer object containing the message
 611 * @offset: offset into the buffer object
 612 *
 613 * Peek into the UVD message and extract the session id.
 614 * Make sure that we don't open up to many sessions.
 615 */
 616static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx *ctx,
 617			     struct amdgpu_bo *bo, unsigned offset)
 618{
 619	struct amdgpu_device *adev = ctx->parser->adev;
 620	int32_t *msg, msg_type, handle;
 621	void *ptr;
 622	long r;
 623	int i;
 624
 625	if (offset & 0x3F) {
 626		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
 627		return -EINVAL;
 628	}
 629
 630	r = amdgpu_bo_kmap(bo, &ptr);
 631	if (r) {
 632		DRM_ERROR("Failed mapping the UVD message (%ld)!\n", r);
 633		return r;
 634	}
 635
 636	msg = ptr + offset;
 637
 638	msg_type = msg[1];
 639	handle = msg[2];
 640
 641	if (handle == 0) {
 642		DRM_ERROR("Invalid UVD handle!\n");
 643		return -EINVAL;
 644	}
 645
 646	switch (msg_type) {
 647	case 0:
 648		/* it's a create msg, calc image size (width * height) */
 649		amdgpu_bo_kunmap(bo);
 650
 651		/* try to alloc a new handle */
 652		for (i = 0; i < adev->uvd.max_handles; ++i) {
 653			if (atomic_read(&adev->uvd.handles[i]) == handle) {
 654				DRM_ERROR("Handle 0x%x already in use!\n", handle);
 655				return -EINVAL;
 656			}
 657
 658			if (!atomic_cmpxchg(&adev->uvd.handles[i], 0, handle)) {
 659				adev->uvd.filp[i] = ctx->parser->filp;
 660				return 0;
 661			}
 662		}
 663
 664		DRM_ERROR("No more free UVD handles!\n");
 665		return -ENOSPC;
 666
 667	case 1:
 668		/* it's a decode msg, calc buffer sizes */
 669		r = amdgpu_uvd_cs_msg_decode(adev, msg, ctx->buf_sizes);
 670		amdgpu_bo_kunmap(bo);
 671		if (r)
 672			return r;
 673
 674		/* validate the handle */
 675		for (i = 0; i < adev->uvd.max_handles; ++i) {
 676			if (atomic_read(&adev->uvd.handles[i]) == handle) {
 677				if (adev->uvd.filp[i] != ctx->parser->filp) {
 678					DRM_ERROR("UVD handle collision detected!\n");
 679					return -EINVAL;
 680				}
 681				return 0;
 682			}
 683		}
 684
 685		DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
 686		return -ENOENT;
 687
 688	case 2:
 689		/* it's a destroy msg, free the handle */
 690		for (i = 0; i < adev->uvd.max_handles; ++i)
 691			atomic_cmpxchg(&adev->uvd.handles[i], handle, 0);
 692		amdgpu_bo_kunmap(bo);
 693		return 0;
 694
 695	default:
 696		DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
 697		return -EINVAL;
 698	}
 699	BUG();
 700	return -EINVAL;
 701}
 702
 703/**
 704 * amdgpu_uvd_cs_pass2 - second parsing round
 705 *
 706 * @ctx: UVD parser context
 707 *
 708 * Patch buffer addresses, make sure buffer sizes are correct.
 709 */
 710static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx *ctx)
 711{
 712	struct amdgpu_bo_va_mapping *mapping;
 713	struct amdgpu_bo *bo;
 714	uint32_t cmd;
 715	uint64_t start, end;
 716	uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
 717	int r;
 718
 
 
 
 
 719	mapping = amdgpu_cs_find_mapping(ctx->parser, addr, &bo);
 720	if (mapping == NULL) {
 721		DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
 722		return -EINVAL;
 723	}
 724
 725	start = amdgpu_bo_gpu_offset(bo);
 726
 727	end = (mapping->it.last + 1 - mapping->it.start);
 728	end = end * AMDGPU_GPU_PAGE_SIZE + start;
 729
 730	addr -= ((uint64_t)mapping->it.start) * AMDGPU_GPU_PAGE_SIZE;
 731	start += addr;
 732
 733	amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data0,
 734			    lower_32_bits(start));
 735	amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data1,
 736			    upper_32_bits(start));
 737
 738	cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
 739	if (cmd < 0x4) {
 740		if ((end - start) < ctx->buf_sizes[cmd]) {
 741			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
 742				  (unsigned)(end - start),
 743				  ctx->buf_sizes[cmd]);
 744			return -EINVAL;
 745		}
 746
 747	} else if (cmd == 0x206) {
 748		if ((end - start) < ctx->buf_sizes[4]) {
 749			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
 750					  (unsigned)(end - start),
 751					  ctx->buf_sizes[4]);
 752			return -EINVAL;
 753		}
 754	} else if ((cmd != 0x100) && (cmd != 0x204)) {
 755		DRM_ERROR("invalid UVD command %X!\n", cmd);
 756		return -EINVAL;
 757	}
 758
 759	if (!ctx->parser->adev->uvd.address_64_bit) {
 760		if ((start >> 28) != ((end - 1) >> 28)) {
 761			DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
 762				  start, end);
 763			return -EINVAL;
 764		}
 765
 766		if ((cmd == 0 || cmd == 0x3) &&
 767		    (start >> 28) != (ctx->parser->adev->uvd.gpu_addr >> 28)) {
 768			DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
 769				  start, end);
 770			return -EINVAL;
 771		}
 772	}
 773
 774	if (cmd == 0) {
 775		ctx->has_msg_cmd = true;
 776		r = amdgpu_uvd_cs_msg(ctx, bo, addr);
 777		if (r)
 778			return r;
 779	} else if (!ctx->has_msg_cmd) {
 780		DRM_ERROR("Message needed before other commands are send!\n");
 781		return -EINVAL;
 782	}
 783
 784	return 0;
 785}
 786
 787/**
 788 * amdgpu_uvd_cs_reg - parse register writes
 789 *
 790 * @ctx: UVD parser context
 791 * @cb: callback function
 792 *
 793 * Parse the register writes, call cb on each complete command.
 794 */
 795static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx *ctx,
 796			     int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
 797{
 798	struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
 799	int i, r;
 800
 801	ctx->idx++;
 802	for (i = 0; i <= ctx->count; ++i) {
 803		unsigned reg = ctx->reg + i;
 804
 805		if (ctx->idx >= ib->length_dw) {
 806			DRM_ERROR("Register command after end of CS!\n");
 807			return -EINVAL;
 808		}
 809
 810		switch (reg) {
 811		case mmUVD_GPCOM_VCPU_DATA0:
 812			ctx->data0 = ctx->idx;
 813			break;
 814		case mmUVD_GPCOM_VCPU_DATA1:
 815			ctx->data1 = ctx->idx;
 816			break;
 817		case mmUVD_GPCOM_VCPU_CMD:
 818			r = cb(ctx);
 819			if (r)
 820				return r;
 821			break;
 822		case mmUVD_ENGINE_CNTL:
 823		case mmUVD_NO_OP:
 824			break;
 825		default:
 826			DRM_ERROR("Invalid reg 0x%X!\n", reg);
 827			return -EINVAL;
 828		}
 829		ctx->idx++;
 830	}
 831	return 0;
 832}
 833
 834/**
 835 * amdgpu_uvd_cs_packets - parse UVD packets
 836 *
 837 * @ctx: UVD parser context
 838 * @cb: callback function
 839 *
 840 * Parse the command stream packets.
 841 */
 842static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx *ctx,
 843				 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
 844{
 845	struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
 846	int r;
 847
 848	for (ctx->idx = 0 ; ctx->idx < ib->length_dw; ) {
 849		uint32_t cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx);
 850		unsigned type = CP_PACKET_GET_TYPE(cmd);
 851		switch (type) {
 852		case PACKET_TYPE0:
 853			ctx->reg = CP_PACKET0_GET_REG(cmd);
 854			ctx->count = CP_PACKET_GET_COUNT(cmd);
 855			r = amdgpu_uvd_cs_reg(ctx, cb);
 856			if (r)
 857				return r;
 858			break;
 859		case PACKET_TYPE2:
 860			++ctx->idx;
 861			break;
 862		default:
 863			DRM_ERROR("Unknown packet type %d !\n", type);
 864			return -EINVAL;
 865		}
 866	}
 867	return 0;
 868}
 869
 870/**
 871 * amdgpu_uvd_ring_parse_cs - UVD command submission parser
 872 *
 873 * @parser: Command submission parser context
 874 *
 875 * Parse the command stream, patch in addresses as necessary.
 876 */
 877int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser, uint32_t ib_idx)
 878{
 879	struct amdgpu_uvd_cs_ctx ctx = {};
 880	unsigned buf_sizes[] = {
 881		[0x00000000]	=	2048,
 882		[0x00000001]	=	0xFFFFFFFF,
 883		[0x00000002]	=	0xFFFFFFFF,
 884		[0x00000003]	=	2048,
 885		[0x00000004]	=	0xFFFFFFFF,
 886	};
 887	struct amdgpu_ib *ib = &parser->job->ibs[ib_idx];
 888	int r;
 889
 890	parser->job->vm = NULL;
 891	ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
 892
 893	if (ib->length_dw % 16) {
 894		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
 895			  ib->length_dw);
 896		return -EINVAL;
 897	}
 898
 899	r = amdgpu_cs_sysvm_access_required(parser);
 900	if (r)
 901		return r;
 902
 903	ctx.parser = parser;
 904	ctx.buf_sizes = buf_sizes;
 905	ctx.ib_idx = ib_idx;
 906
 907	/* first round only required on chips without UVD 64 bit address support */
 908	if (!parser->adev->uvd.address_64_bit) {
 909		/* first round, make sure the buffers are actually in the UVD segment */
 910		r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass1);
 911		if (r)
 912			return r;
 913	}
 914
 915	/* second round, patch buffer addresses into the command stream */
 916	r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass2);
 917	if (r)
 918		return r;
 919
 920	if (!ctx.has_msg_cmd) {
 921		DRM_ERROR("UVD-IBs need a msg command!\n");
 922		return -EINVAL;
 923	}
 924
 
 
 925	return 0;
 926}
 927
 928static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
 929			       bool direct, struct dma_fence **fence)
 930{
 931	struct ttm_validate_buffer tv;
 932	struct ww_acquire_ctx ticket;
 933	struct list_head head;
 934	struct amdgpu_job *job;
 935	struct amdgpu_ib *ib;
 936	struct dma_fence *f = NULL;
 937	struct amdgpu_device *adev = ring->adev;
 938	uint64_t addr;
 939	int i, r;
 940
 941	memset(&tv, 0, sizeof(tv));
 942	tv.bo = &bo->tbo;
 943
 944	INIT_LIST_HEAD(&head);
 945	list_add(&tv.head, &head);
 946
 947	r = ttm_eu_reserve_buffers(&ticket, &head, true, NULL);
 948	if (r)
 949		return r;
 950
 951	if (!ring->adev->uvd.address_64_bit) {
 952		amdgpu_ttm_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
 953		amdgpu_uvd_force_into_uvd_segment(bo);
 954	}
 955
 956	r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
 957	if (r)
 958		goto err;
 959
 960	r = amdgpu_job_alloc_with_ib(adev, 64, &job);
 961	if (r)
 962		goto err;
 963
 964	ib = &job->ibs[0];
 965	addr = amdgpu_bo_gpu_offset(bo);
 966	ib->ptr[0] = PACKET0(mmUVD_GPCOM_VCPU_DATA0, 0);
 967	ib->ptr[1] = addr;
 968	ib->ptr[2] = PACKET0(mmUVD_GPCOM_VCPU_DATA1, 0);
 969	ib->ptr[3] = addr >> 32;
 970	ib->ptr[4] = PACKET0(mmUVD_GPCOM_VCPU_CMD, 0);
 971	ib->ptr[5] = 0;
 972	for (i = 6; i < 16; i += 2) {
 973		ib->ptr[i] = PACKET0(mmUVD_NO_OP, 0);
 974		ib->ptr[i+1] = 0;
 975	}
 976	ib->length_dw = 16;
 977
 978	if (direct) {
 979		r = amdgpu_ib_schedule(ring, 1, ib, NULL, NULL, &f);
 980		job->fence = dma_fence_get(f);
 981		if (r)
 982			goto err_free;
 983
 984		amdgpu_job_free(job);
 985	} else {
 986		r = amdgpu_job_submit(job, ring, &adev->uvd.entity,
 987				      AMDGPU_FENCE_OWNER_UNDEFINED, &f);
 988		if (r)
 989			goto err_free;
 990	}
 991
 992	ttm_eu_fence_buffer_objects(&ticket, &head, f);
 993
 994	if (fence)
 995		*fence = dma_fence_get(f);
 996	amdgpu_bo_unref(&bo);
 997	dma_fence_put(f);
 998
 999	return 0;
1000
1001err_free:
1002	amdgpu_job_free(job);
1003
1004err:
1005	ttm_eu_backoff_reservation(&ticket, &head);
1006	return r;
1007}
1008
1009/* multiple fence commands without any stream commands in between can
1010   crash the vcpu so just try to emmit a dummy create/destroy msg to
1011   avoid this */
1012int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
1013			      struct dma_fence **fence)
1014{
1015	struct amdgpu_device *adev = ring->adev;
1016	struct amdgpu_bo *bo;
1017	uint32_t *msg;
1018	int r, i;
1019
1020	r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
1021			     AMDGPU_GEM_DOMAIN_VRAM,
1022			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
1023			     AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS,
1024			     NULL, NULL, &bo);
1025	if (r)
1026		return r;
1027
1028	r = amdgpu_bo_reserve(bo, false);
1029	if (r) {
1030		amdgpu_bo_unref(&bo);
1031		return r;
1032	}
1033
1034	r = amdgpu_bo_kmap(bo, (void **)&msg);
1035	if (r) {
1036		amdgpu_bo_unreserve(bo);
1037		amdgpu_bo_unref(&bo);
1038		return r;
1039	}
1040
1041	/* stitch together an UVD create msg */
1042	msg[0] = cpu_to_le32(0x00000de4);
1043	msg[1] = cpu_to_le32(0x00000000);
1044	msg[2] = cpu_to_le32(handle);
1045	msg[3] = cpu_to_le32(0x00000000);
1046	msg[4] = cpu_to_le32(0x00000000);
1047	msg[5] = cpu_to_le32(0x00000000);
1048	msg[6] = cpu_to_le32(0x00000000);
1049	msg[7] = cpu_to_le32(0x00000780);
1050	msg[8] = cpu_to_le32(0x00000440);
1051	msg[9] = cpu_to_le32(0x00000000);
1052	msg[10] = cpu_to_le32(0x01b37000);
1053	for (i = 11; i < 1024; ++i)
1054		msg[i] = cpu_to_le32(0x0);
1055
1056	amdgpu_bo_kunmap(bo);
1057	amdgpu_bo_unreserve(bo);
1058
1059	return amdgpu_uvd_send_msg(ring, bo, true, fence);
1060}
1061
1062int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
1063			       bool direct, struct dma_fence **fence)
1064{
1065	struct amdgpu_device *adev = ring->adev;
1066	struct amdgpu_bo *bo;
1067	uint32_t *msg;
1068	int r, i;
1069
1070	r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
1071			     AMDGPU_GEM_DOMAIN_VRAM,
1072			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
1073			     AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS,
1074			     NULL, NULL, &bo);
1075	if (r)
1076		return r;
1077
1078	r = amdgpu_bo_reserve(bo, false);
1079	if (r) {
1080		amdgpu_bo_unref(&bo);
1081		return r;
1082	}
1083
1084	r = amdgpu_bo_kmap(bo, (void **)&msg);
1085	if (r) {
1086		amdgpu_bo_unreserve(bo);
1087		amdgpu_bo_unref(&bo);
1088		return r;
1089	}
1090
1091	/* stitch together an UVD destroy msg */
1092	msg[0] = cpu_to_le32(0x00000de4);
1093	msg[1] = cpu_to_le32(0x00000002);
1094	msg[2] = cpu_to_le32(handle);
1095	msg[3] = cpu_to_le32(0x00000000);
1096	for (i = 4; i < 1024; ++i)
1097		msg[i] = cpu_to_le32(0x0);
1098
1099	amdgpu_bo_kunmap(bo);
1100	amdgpu_bo_unreserve(bo);
1101
1102	return amdgpu_uvd_send_msg(ring, bo, direct, fence);
1103}
1104
1105static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
1106{
1107	struct amdgpu_device *adev =
1108		container_of(work, struct amdgpu_device, uvd.idle_work.work);
1109	unsigned fences = amdgpu_fence_count_emitted(&adev->uvd.ring);
 
 
1110
1111	if (fences == 0) {
 
 
 
 
1112		if (adev->pm.dpm_enabled) {
1113			amdgpu_dpm_enable_uvd(adev, false);
1114		} else {
1115			amdgpu_asic_set_uvd_clocks(adev, 0, 0);
1116		}
1117	} else {
1118		schedule_delayed_work(&adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
 
1119	}
1120}
1121
1122void amdgpu_uvd_ring_begin_use(struct amdgpu_ring *ring)
1123{
1124	struct amdgpu_device *adev = ring->adev;
1125	bool set_clocks = !cancel_delayed_work_sync(&adev->uvd.idle_work);
 
 
1126
1127	if (set_clocks) {
1128		if (adev->pm.dpm_enabled) {
1129			amdgpu_dpm_enable_uvd(adev, true);
1130		} else {
1131			amdgpu_asic_set_uvd_clocks(adev, 53300, 40000);
1132		}
1133	}
1134}
1135
1136void amdgpu_uvd_ring_end_use(struct amdgpu_ring *ring)
1137{
1138	schedule_delayed_work(&ring->adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1139}
1140
1141/**
1142 * amdgpu_uvd_ring_test_ib - test ib execution
1143 *
1144 * @ring: amdgpu_ring pointer
1145 *
1146 * Test if we can successfully execute an IB
1147 */
1148int amdgpu_uvd_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1149{
1150	struct dma_fence *fence;
1151	long r;
1152
1153	r = amdgpu_uvd_get_create_msg(ring, 1, NULL);
1154	if (r) {
1155		DRM_ERROR("amdgpu: failed to get create msg (%ld).\n", r);
1156		goto error;
1157	}
1158
1159	r = amdgpu_uvd_get_destroy_msg(ring, 1, true, &fence);
1160	if (r) {
1161		DRM_ERROR("amdgpu: failed to get destroy ib (%ld).\n", r);
1162		goto error;
1163	}
1164
1165	r = dma_fence_wait_timeout(fence, false, timeout);
1166	if (r == 0) {
1167		DRM_ERROR("amdgpu: IB test timed out.\n");
1168		r = -ETIMEDOUT;
1169	} else if (r < 0) {
1170		DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
1171	} else {
1172		DRM_INFO("ib test on ring %d succeeded\n",  ring->idx);
1173		r = 0;
1174	}
1175
1176	dma_fence_put(fence);
1177
1178error:
1179	return r;
1180}