Linux Audio

Check our new training course

Loading...
v6.2
   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
  34#include <drm/drm.h>
  35#include <drm/drm_drv.h>
  36
  37#include "amdgpu.h"
  38#include "amdgpu_pm.h"
  39#include "amdgpu_uvd.h"
  40#include "amdgpu_cs.h"
  41#include "cikd.h"
  42#include "uvd/uvd_4_2_d.h"
  43
  44#include "amdgpu_ras.h"
  45
  46/* 1 second timeout */
  47#define UVD_IDLE_TIMEOUT	msecs_to_jiffies(1000)
  48
  49/* Firmware versions for VI */
  50#define FW_1_65_10	((1 << 24) | (65 << 16) | (10 << 8))
  51#define FW_1_87_11	((1 << 24) | (87 << 16) | (11 << 8))
  52#define FW_1_87_12	((1 << 24) | (87 << 16) | (12 << 8))
  53#define FW_1_37_15	((1 << 24) | (37 << 16) | (15 << 8))
  54
  55/* Polaris10/11 firmware version */
  56#define FW_1_66_16	((1 << 24) | (66 << 16) | (16 << 8))
  57
  58/* Firmware Names */
  59#ifdef CONFIG_DRM_AMDGPU_SI
  60#define FIRMWARE_TAHITI		"amdgpu/tahiti_uvd.bin"
  61#define FIRMWARE_VERDE		"amdgpu/verde_uvd.bin"
  62#define FIRMWARE_PITCAIRN	"amdgpu/pitcairn_uvd.bin"
  63#define FIRMWARE_OLAND		"amdgpu/oland_uvd.bin"
  64#endif
  65#ifdef CONFIG_DRM_AMDGPU_CIK
  66#define FIRMWARE_BONAIRE	"amdgpu/bonaire_uvd.bin"
  67#define FIRMWARE_KABINI	"amdgpu/kabini_uvd.bin"
  68#define FIRMWARE_KAVERI	"amdgpu/kaveri_uvd.bin"
  69#define FIRMWARE_HAWAII	"amdgpu/hawaii_uvd.bin"
  70#define FIRMWARE_MULLINS	"amdgpu/mullins_uvd.bin"
  71#endif
  72#define FIRMWARE_TONGA		"amdgpu/tonga_uvd.bin"
  73#define FIRMWARE_CARRIZO	"amdgpu/carrizo_uvd.bin"
  74#define FIRMWARE_FIJI		"amdgpu/fiji_uvd.bin"
  75#define FIRMWARE_STONEY		"amdgpu/stoney_uvd.bin"
  76#define FIRMWARE_POLARIS10	"amdgpu/polaris10_uvd.bin"
  77#define FIRMWARE_POLARIS11	"amdgpu/polaris11_uvd.bin"
  78#define FIRMWARE_POLARIS12	"amdgpu/polaris12_uvd.bin"
  79#define FIRMWARE_VEGAM		"amdgpu/vegam_uvd.bin"
  80
  81#define FIRMWARE_VEGA10		"amdgpu/vega10_uvd.bin"
  82#define FIRMWARE_VEGA12		"amdgpu/vega12_uvd.bin"
  83#define FIRMWARE_VEGA20		"amdgpu/vega20_uvd.bin"
  84
  85/* These are common relative offsets for all asics, from uvd_7_0_offset.h,  */
  86#define UVD_GPCOM_VCPU_CMD		0x03c3
  87#define UVD_GPCOM_VCPU_DATA0	0x03c4
  88#define UVD_GPCOM_VCPU_DATA1	0x03c5
  89#define UVD_NO_OP				0x03ff
  90#define UVD_BASE_SI				0x3800
  91
  92/*
  93 * amdgpu_uvd_cs_ctx - Command submission parser context
  94 *
  95 * Used for emulating virtual memory support on UVD 4.2.
  96 */
  97struct amdgpu_uvd_cs_ctx {
  98	struct amdgpu_cs_parser *parser;
  99	unsigned reg, count;
 100	unsigned data0, data1;
 101	unsigned idx;
 102	struct amdgpu_ib *ib;
 103
 104	/* does the IB has a msg command */
 105	bool has_msg_cmd;
 106
 107	/* minimum buffer sizes */
 108	unsigned *buf_sizes;
 109};
 110
 111#ifdef CONFIG_DRM_AMDGPU_SI
 112MODULE_FIRMWARE(FIRMWARE_TAHITI);
 113MODULE_FIRMWARE(FIRMWARE_VERDE);
 114MODULE_FIRMWARE(FIRMWARE_PITCAIRN);
 115MODULE_FIRMWARE(FIRMWARE_OLAND);
 116#endif
 117#ifdef CONFIG_DRM_AMDGPU_CIK
 118MODULE_FIRMWARE(FIRMWARE_BONAIRE);
 119MODULE_FIRMWARE(FIRMWARE_KABINI);
 120MODULE_FIRMWARE(FIRMWARE_KAVERI);
 121MODULE_FIRMWARE(FIRMWARE_HAWAII);
 122MODULE_FIRMWARE(FIRMWARE_MULLINS);
 123#endif
 124MODULE_FIRMWARE(FIRMWARE_TONGA);
 125MODULE_FIRMWARE(FIRMWARE_CARRIZO);
 126MODULE_FIRMWARE(FIRMWARE_FIJI);
 127MODULE_FIRMWARE(FIRMWARE_STONEY);
 128MODULE_FIRMWARE(FIRMWARE_POLARIS10);
 129MODULE_FIRMWARE(FIRMWARE_POLARIS11);
 130MODULE_FIRMWARE(FIRMWARE_POLARIS12);
 131MODULE_FIRMWARE(FIRMWARE_VEGAM);
 132
 133MODULE_FIRMWARE(FIRMWARE_VEGA10);
 134MODULE_FIRMWARE(FIRMWARE_VEGA12);
 135MODULE_FIRMWARE(FIRMWARE_VEGA20);
 136
 137static void amdgpu_uvd_idle_work_handler(struct work_struct *work);
 138static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *abo);
 139
 140static int amdgpu_uvd_create_msg_bo_helper(struct amdgpu_device *adev,
 141					   uint32_t size,
 142					   struct amdgpu_bo **bo_ptr)
 143{
 144	struct ttm_operation_ctx ctx = { true, false };
 145	struct amdgpu_bo *bo = NULL;
 146	void *addr;
 147	int r;
 148
 149	r = amdgpu_bo_create_reserved(adev, size, PAGE_SIZE,
 150				      AMDGPU_GEM_DOMAIN_GTT,
 151				      &bo, NULL, &addr);
 152	if (r)
 153		return r;
 154
 155	if (adev->uvd.address_64_bit)
 156		goto succ;
 157
 158	amdgpu_bo_kunmap(bo);
 159	amdgpu_bo_unpin(bo);
 160	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
 161	amdgpu_uvd_force_into_uvd_segment(bo);
 162	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
 163	if (r)
 164		goto err;
 165	r = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_VRAM);
 166	if (r)
 167		goto err_pin;
 168	r = amdgpu_bo_kmap(bo, &addr);
 169	if (r)
 170		goto err_kmap;
 171succ:
 172	amdgpu_bo_unreserve(bo);
 173	*bo_ptr = bo;
 174	return 0;
 175err_kmap:
 176	amdgpu_bo_unpin(bo);
 177err_pin:
 178err:
 179	amdgpu_bo_unreserve(bo);
 180	amdgpu_bo_unref(&bo);
 181	return r;
 182}
 183
 184int amdgpu_uvd_sw_init(struct amdgpu_device *adev)
 185{
 186	unsigned long bo_size;
 187	const char *fw_name;
 188	const struct common_firmware_header *hdr;
 189	unsigned family_id;
 190	int i, j, r;
 191
 192	INIT_DELAYED_WORK(&adev->uvd.idle_work, amdgpu_uvd_idle_work_handler);
 193
 194	switch (adev->asic_type) {
 195#ifdef CONFIG_DRM_AMDGPU_SI
 196	case CHIP_TAHITI:
 197		fw_name = FIRMWARE_TAHITI;
 198		break;
 199	case CHIP_VERDE:
 200		fw_name = FIRMWARE_VERDE;
 201		break;
 202	case CHIP_PITCAIRN:
 203		fw_name = FIRMWARE_PITCAIRN;
 204		break;
 205	case CHIP_OLAND:
 206		fw_name = FIRMWARE_OLAND;
 207		break;
 208#endif
 209#ifdef CONFIG_DRM_AMDGPU_CIK
 210	case CHIP_BONAIRE:
 211		fw_name = FIRMWARE_BONAIRE;
 212		break;
 213	case CHIP_KABINI:
 214		fw_name = FIRMWARE_KABINI;
 215		break;
 216	case CHIP_KAVERI:
 217		fw_name = FIRMWARE_KAVERI;
 218		break;
 219	case CHIP_HAWAII:
 220		fw_name = FIRMWARE_HAWAII;
 221		break;
 222	case CHIP_MULLINS:
 223		fw_name = FIRMWARE_MULLINS;
 224		break;
 225#endif
 226	case CHIP_TONGA:
 227		fw_name = FIRMWARE_TONGA;
 228		break;
 229	case CHIP_FIJI:
 230		fw_name = FIRMWARE_FIJI;
 231		break;
 232	case CHIP_CARRIZO:
 233		fw_name = FIRMWARE_CARRIZO;
 234		break;
 235	case CHIP_STONEY:
 236		fw_name = FIRMWARE_STONEY;
 237		break;
 238	case CHIP_POLARIS10:
 239		fw_name = FIRMWARE_POLARIS10;
 240		break;
 241	case CHIP_POLARIS11:
 242		fw_name = FIRMWARE_POLARIS11;
 243		break;
 244	case CHIP_POLARIS12:
 245		fw_name = FIRMWARE_POLARIS12;
 246		break;
 247	case CHIP_VEGA10:
 248		fw_name = FIRMWARE_VEGA10;
 249		break;
 250	case CHIP_VEGA12:
 251		fw_name = FIRMWARE_VEGA12;
 252		break;
 253	case CHIP_VEGAM:
 254		fw_name = FIRMWARE_VEGAM;
 255		break;
 256	case CHIP_VEGA20:
 257		fw_name = FIRMWARE_VEGA20;
 258		break;
 259	default:
 260		return -EINVAL;
 261	}
 262
 263	r = request_firmware(&adev->uvd.fw, fw_name, adev->dev);
 264	if (r) {
 265		dev_err(adev->dev, "amdgpu_uvd: Can't load firmware \"%s\"\n",
 266			fw_name);
 267		return r;
 268	}
 269
 270	r = amdgpu_ucode_validate(adev->uvd.fw);
 271	if (r) {
 272		dev_err(adev->dev, "amdgpu_uvd: Can't validate firmware \"%s\"\n",
 273			fw_name);
 274		release_firmware(adev->uvd.fw);
 275		adev->uvd.fw = NULL;
 276		return r;
 277	}
 278
 279	/* Set the default UVD handles that the firmware can handle */
 280	adev->uvd.max_handles = AMDGPU_DEFAULT_UVD_HANDLES;
 281
 282	hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
 283	family_id = le32_to_cpu(hdr->ucode_version) & 0xff;
 284
 285	if (adev->asic_type < CHIP_VEGA20) {
 286		unsigned version_major, version_minor;
 287
 288		version_major = (le32_to_cpu(hdr->ucode_version) >> 24) & 0xff;
 289		version_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
 290		DRM_INFO("Found UVD firmware Version: %u.%u Family ID: %u\n",
 291			version_major, version_minor, family_id);
 292
 293		/*
 294		 * Limit the number of UVD handles depending on microcode major
 295		 * and minor versions. The firmware version which has 40 UVD
 296		 * instances support is 1.80. So all subsequent versions should
 297		 * also have the same support.
 298		 */
 299		if ((version_major > 0x01) ||
 300		    ((version_major == 0x01) && (version_minor >= 0x50)))
 301			adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
 302
 303		adev->uvd.fw_version = ((version_major << 24) | (version_minor << 16) |
 304					(family_id << 8));
 305
 306		if ((adev->asic_type == CHIP_POLARIS10 ||
 307		     adev->asic_type == CHIP_POLARIS11) &&
 308		    (adev->uvd.fw_version < FW_1_66_16))
 309			DRM_ERROR("POLARIS10/11 UVD firmware version %u.%u is too old.\n",
 310				  version_major, version_minor);
 311	} else {
 312		unsigned int enc_major, enc_minor, dec_minor;
 313
 314		dec_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
 315		enc_minor = (le32_to_cpu(hdr->ucode_version) >> 24) & 0x3f;
 316		enc_major = (le32_to_cpu(hdr->ucode_version) >> 30) & 0x3;
 317		DRM_INFO("Found UVD firmware ENC: %u.%u DEC: .%u Family ID: %u\n",
 318			enc_major, enc_minor, dec_minor, family_id);
 319
 320		adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
 321
 322		adev->uvd.fw_version = le32_to_cpu(hdr->ucode_version);
 323	}
 324
 325	bo_size = AMDGPU_UVD_STACK_SIZE + AMDGPU_UVD_HEAP_SIZE
 326		  +  AMDGPU_UVD_SESSION_SIZE * adev->uvd.max_handles;
 327	if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP)
 328		bo_size += AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr->ucode_size_bytes) + 8);
 329
 330	for (j = 0; j < adev->uvd.num_uvd_inst; j++) {
 331		if (adev->uvd.harvest_config & (1 << j))
 332			continue;
 333		r = amdgpu_bo_create_kernel(adev, bo_size, PAGE_SIZE,
 334					    AMDGPU_GEM_DOMAIN_VRAM, &adev->uvd.inst[j].vcpu_bo,
 335					    &adev->uvd.inst[j].gpu_addr, &adev->uvd.inst[j].cpu_addr);
 336		if (r) {
 337			dev_err(adev->dev, "(%d) failed to allocate UVD bo\n", r);
 338			return r;
 339		}
 340	}
 341
 342	for (i = 0; i < adev->uvd.max_handles; ++i) {
 343		atomic_set(&adev->uvd.handles[i], 0);
 344		adev->uvd.filp[i] = NULL;
 345	}
 346
 347	/* from uvd v5.0 HW addressing capacity increased to 64 bits */
 348	if (!amdgpu_device_ip_block_version_cmp(adev, AMD_IP_BLOCK_TYPE_UVD, 5, 0))
 349		adev->uvd.address_64_bit = true;
 350
 351	r = amdgpu_uvd_create_msg_bo_helper(adev, 128 << 10, &adev->uvd.ib_bo);
 352	if (r)
 353		return r;
 354
 355	switch (adev->asic_type) {
 356	case CHIP_TONGA:
 357		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_65_10;
 358		break;
 359	case CHIP_CARRIZO:
 360		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_11;
 361		break;
 362	case CHIP_FIJI:
 363		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_12;
 364		break;
 365	case CHIP_STONEY:
 366		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_37_15;
 367		break;
 368	default:
 369		adev->uvd.use_ctx_buf = adev->asic_type >= CHIP_POLARIS10;
 370	}
 371
 372	return 0;
 373}
 374
 375int amdgpu_uvd_sw_fini(struct amdgpu_device *adev)
 376{
 377	void *addr = amdgpu_bo_kptr(adev->uvd.ib_bo);
 378	int i, j;
 379
 380	drm_sched_entity_destroy(&adev->uvd.entity);
 381
 382	for (j = 0; j < adev->uvd.num_uvd_inst; ++j) {
 383		if (adev->uvd.harvest_config & (1 << j))
 384			continue;
 385		kvfree(adev->uvd.inst[j].saved_bo);
 386
 387		amdgpu_bo_free_kernel(&adev->uvd.inst[j].vcpu_bo,
 388				      &adev->uvd.inst[j].gpu_addr,
 389				      (void **)&adev->uvd.inst[j].cpu_addr);
 390
 391		amdgpu_ring_fini(&adev->uvd.inst[j].ring);
 392
 393		for (i = 0; i < AMDGPU_MAX_UVD_ENC_RINGS; ++i)
 394			amdgpu_ring_fini(&adev->uvd.inst[j].ring_enc[i]);
 395	}
 396	amdgpu_bo_free_kernel(&adev->uvd.ib_bo, NULL, &addr);
 397	release_firmware(adev->uvd.fw);
 398
 399	return 0;
 400}
 401
 402/**
 403 * amdgpu_uvd_entity_init - init entity
 404 *
 405 * @adev: amdgpu_device pointer
 406 *
 407 */
 408int amdgpu_uvd_entity_init(struct amdgpu_device *adev)
 409{
 410	struct amdgpu_ring *ring;
 411	struct drm_gpu_scheduler *sched;
 412	int r;
 413
 414	ring = &adev->uvd.inst[0].ring;
 415	sched = &ring->sched;
 416	r = drm_sched_entity_init(&adev->uvd.entity, DRM_SCHED_PRIORITY_NORMAL,
 417				  &sched, 1, NULL);
 418	if (r) {
 419		DRM_ERROR("Failed setting up UVD kernel entity.\n");
 420		return r;
 421	}
 422
 423	return 0;
 424}
 425
 426int amdgpu_uvd_suspend(struct amdgpu_device *adev)
 427{
 428	unsigned size;
 429	void *ptr;
 430	int i, j, idx;
 431	bool in_ras_intr = amdgpu_ras_intr_triggered();
 432
 433	cancel_delayed_work_sync(&adev->uvd.idle_work);
 434
 435	/* only valid for physical mode */
 436	if (adev->asic_type < CHIP_POLARIS10) {
 437		for (i = 0; i < adev->uvd.max_handles; ++i)
 438			if (atomic_read(&adev->uvd.handles[i]))
 439				break;
 440
 441		if (i == adev->uvd.max_handles)
 442			return 0;
 443	}
 444
 445	for (j = 0; j < adev->uvd.num_uvd_inst; ++j) {
 446		if (adev->uvd.harvest_config & (1 << j))
 447			continue;
 448		if (adev->uvd.inst[j].vcpu_bo == NULL)
 449			continue;
 450
 451		size = amdgpu_bo_size(adev->uvd.inst[j].vcpu_bo);
 452		ptr = adev->uvd.inst[j].cpu_addr;
 453
 454		adev->uvd.inst[j].saved_bo = kvmalloc(size, GFP_KERNEL);
 455		if (!adev->uvd.inst[j].saved_bo)
 456			return -ENOMEM;
 457
 458		if (drm_dev_enter(adev_to_drm(adev), &idx)) {
 459			/* re-write 0 since err_event_athub will corrupt VCPU buffer */
 460			if (in_ras_intr)
 461				memset(adev->uvd.inst[j].saved_bo, 0, size);
 462			else
 463				memcpy_fromio(adev->uvd.inst[j].saved_bo, ptr, size);
 464
 465			drm_dev_exit(idx);
 466		}
 467	}
 468
 469	if (in_ras_intr)
 470		DRM_WARN("UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT\n");
 471
 472	return 0;
 473}
 474
 475int amdgpu_uvd_resume(struct amdgpu_device *adev)
 476{
 477	unsigned size;
 478	void *ptr;
 479	int i, idx;
 480
 481	for (i = 0; i < adev->uvd.num_uvd_inst; i++) {
 482		if (adev->uvd.harvest_config & (1 << i))
 483			continue;
 484		if (adev->uvd.inst[i].vcpu_bo == NULL)
 485			return -EINVAL;
 486
 487		size = amdgpu_bo_size(adev->uvd.inst[i].vcpu_bo);
 488		ptr = adev->uvd.inst[i].cpu_addr;
 489
 490		if (adev->uvd.inst[i].saved_bo != NULL) {
 491			if (drm_dev_enter(adev_to_drm(adev), &idx)) {
 492				memcpy_toio(ptr, adev->uvd.inst[i].saved_bo, size);
 493				drm_dev_exit(idx);
 494			}
 495			kvfree(adev->uvd.inst[i].saved_bo);
 496			adev->uvd.inst[i].saved_bo = NULL;
 497		} else {
 498			const struct common_firmware_header *hdr;
 499			unsigned offset;
 500
 501			hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
 502			if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP) {
 503				offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
 504				if (drm_dev_enter(adev_to_drm(adev), &idx)) {
 505					memcpy_toio(adev->uvd.inst[i].cpu_addr, adev->uvd.fw->data + offset,
 506						    le32_to_cpu(hdr->ucode_size_bytes));
 507					drm_dev_exit(idx);
 508				}
 509				size -= le32_to_cpu(hdr->ucode_size_bytes);
 510				ptr += le32_to_cpu(hdr->ucode_size_bytes);
 511			}
 512			memset_io(ptr, 0, size);
 513			/* to restore uvd fence seq */
 514			amdgpu_fence_driver_force_completion(&adev->uvd.inst[i].ring);
 515		}
 516	}
 517	return 0;
 518}
 519
 520void amdgpu_uvd_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
 521{
 522	struct amdgpu_ring *ring = &adev->uvd.inst[0].ring;
 523	int i, r;
 524
 525	for (i = 0; i < adev->uvd.max_handles; ++i) {
 526		uint32_t handle = atomic_read(&adev->uvd.handles[i]);
 527
 528		if (handle != 0 && adev->uvd.filp[i] == filp) {
 529			struct dma_fence *fence;
 530
 531			r = amdgpu_uvd_get_destroy_msg(ring, handle, false,
 532						       &fence);
 533			if (r) {
 534				DRM_ERROR("Error destroying UVD %d!\n", r);
 535				continue;
 536			}
 537
 538			dma_fence_wait(fence, false);
 539			dma_fence_put(fence);
 540
 541			adev->uvd.filp[i] = NULL;
 542			atomic_set(&adev->uvd.handles[i], 0);
 543		}
 544	}
 545}
 546
 547static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *abo)
 548{
 549	int i;
 550	for (i = 0; i < abo->placement.num_placement; ++i) {
 551		abo->placements[i].fpfn = 0 >> PAGE_SHIFT;
 552		abo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
 553	}
 554}
 555
 556static u64 amdgpu_uvd_get_addr_from_ctx(struct amdgpu_uvd_cs_ctx *ctx)
 557{
 558	uint32_t lo, hi;
 559	uint64_t addr;
 560
 561	lo = amdgpu_ib_get_value(ctx->ib, ctx->data0);
 562	hi = amdgpu_ib_get_value(ctx->ib, ctx->data1);
 563	addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
 564
 565	return addr;
 566}
 567
 568/**
 569 * amdgpu_uvd_cs_pass1 - first parsing round
 570 *
 571 * @ctx: UVD parser context
 572 *
 573 * Make sure UVD message and feedback buffers are in VRAM and
 574 * nobody is violating an 256MB boundary.
 575 */
 576static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx *ctx)
 577{
 578	struct ttm_operation_ctx tctx = { false, false };
 579	struct amdgpu_bo_va_mapping *mapping;
 580	struct amdgpu_bo *bo;
 581	uint32_t cmd;
 582	uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
 583	int r = 0;
 584
 585	r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
 586	if (r) {
 587		DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
 588		return r;
 589	}
 590
 591	if (!ctx->parser->adev->uvd.address_64_bit) {
 592		/* check if it's a message or feedback command */
 593		cmd = amdgpu_ib_get_value(ctx->ib, ctx->idx) >> 1;
 594		if (cmd == 0x0 || cmd == 0x3) {
 595			/* yes, force it into VRAM */
 596			uint32_t domain = AMDGPU_GEM_DOMAIN_VRAM;
 597			amdgpu_bo_placement_from_domain(bo, domain);
 598		}
 599		amdgpu_uvd_force_into_uvd_segment(bo);
 600
 601		r = ttm_bo_validate(&bo->tbo, &bo->placement, &tctx);
 602	}
 603
 604	return r;
 605}
 606
 607/**
 608 * amdgpu_uvd_cs_msg_decode - handle UVD decode message
 609 *
 610 * @adev: amdgpu_device pointer
 611 * @msg: pointer to message structure
 612 * @buf_sizes: placeholder to put the different buffer lengths
 613 *
 614 * Peek into the decode message and calculate the necessary buffer sizes.
 615 */
 616static int amdgpu_uvd_cs_msg_decode(struct amdgpu_device *adev, uint32_t *msg,
 617	unsigned buf_sizes[])
 618{
 619	unsigned stream_type = msg[4];
 620	unsigned width = msg[6];
 621	unsigned height = msg[7];
 622	unsigned dpb_size = msg[9];
 623	unsigned pitch = msg[28];
 624	unsigned level = msg[57];
 625
 626	unsigned width_in_mb = width / 16;
 627	unsigned height_in_mb = ALIGN(height / 16, 2);
 628	unsigned fs_in_mb = width_in_mb * height_in_mb;
 629
 630	unsigned image_size, tmp, min_dpb_size, num_dpb_buffer;
 631	unsigned min_ctx_size = ~0;
 632
 633	image_size = width * height;
 634	image_size += image_size / 2;
 635	image_size = ALIGN(image_size, 1024);
 636
 637	switch (stream_type) {
 638	case 0: /* H264 */
 639		switch(level) {
 640		case 30:
 641			num_dpb_buffer = 8100 / fs_in_mb;
 642			break;
 643		case 31:
 644			num_dpb_buffer = 18000 / fs_in_mb;
 645			break;
 646		case 32:
 647			num_dpb_buffer = 20480 / fs_in_mb;
 648			break;
 649		case 41:
 650			num_dpb_buffer = 32768 / fs_in_mb;
 651			break;
 652		case 42:
 653			num_dpb_buffer = 34816 / fs_in_mb;
 654			break;
 655		case 50:
 656			num_dpb_buffer = 110400 / fs_in_mb;
 657			break;
 658		case 51:
 659			num_dpb_buffer = 184320 / fs_in_mb;
 660			break;
 661		default:
 662			num_dpb_buffer = 184320 / fs_in_mb;
 663			break;
 664		}
 665		num_dpb_buffer++;
 666		if (num_dpb_buffer > 17)
 667			num_dpb_buffer = 17;
 668
 669		/* reference picture buffer */
 670		min_dpb_size = image_size * num_dpb_buffer;
 671
 672		/* macroblock context buffer */
 673		min_dpb_size += width_in_mb * height_in_mb * num_dpb_buffer * 192;
 674
 675		/* IT surface buffer */
 676		min_dpb_size += width_in_mb * height_in_mb * 32;
 677		break;
 678
 679	case 1: /* VC1 */
 680
 681		/* reference picture buffer */
 682		min_dpb_size = image_size * 3;
 683
 684		/* CONTEXT_BUFFER */
 685		min_dpb_size += width_in_mb * height_in_mb * 128;
 686
 687		/* IT surface buffer */
 688		min_dpb_size += width_in_mb * 64;
 689
 690		/* DB surface buffer */
 691		min_dpb_size += width_in_mb * 128;
 692
 693		/* BP */
 694		tmp = max(width_in_mb, height_in_mb);
 695		min_dpb_size += ALIGN(tmp * 7 * 16, 64);
 696		break;
 697
 698	case 3: /* MPEG2 */
 699
 700		/* reference picture buffer */
 701		min_dpb_size = image_size * 3;
 702		break;
 703
 704	case 4: /* MPEG4 */
 705
 706		/* reference picture buffer */
 707		min_dpb_size = image_size * 3;
 708
 709		/* CM */
 710		min_dpb_size += width_in_mb * height_in_mb * 64;
 711
 712		/* IT surface buffer */
 713		min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
 714		break;
 715
 716	case 7: /* H264 Perf */
 717		switch(level) {
 718		case 30:
 719			num_dpb_buffer = 8100 / fs_in_mb;
 720			break;
 721		case 31:
 722			num_dpb_buffer = 18000 / fs_in_mb;
 723			break;
 724		case 32:
 725			num_dpb_buffer = 20480 / fs_in_mb;
 726			break;
 727		case 41:
 728			num_dpb_buffer = 32768 / fs_in_mb;
 729			break;
 730		case 42:
 731			num_dpb_buffer = 34816 / fs_in_mb;
 732			break;
 733		case 50:
 734			num_dpb_buffer = 110400 / fs_in_mb;
 735			break;
 736		case 51:
 737			num_dpb_buffer = 184320 / fs_in_mb;
 738			break;
 739		default:
 740			num_dpb_buffer = 184320 / fs_in_mb;
 741			break;
 742		}
 743		num_dpb_buffer++;
 744		if (num_dpb_buffer > 17)
 745			num_dpb_buffer = 17;
 746
 747		/* reference picture buffer */
 748		min_dpb_size = image_size * num_dpb_buffer;
 749
 750		if (!adev->uvd.use_ctx_buf){
 751			/* macroblock context buffer */
 752			min_dpb_size +=
 753				width_in_mb * height_in_mb * num_dpb_buffer * 192;
 754
 755			/* IT surface buffer */
 756			min_dpb_size += width_in_mb * height_in_mb * 32;
 757		} else {
 758			/* macroblock context buffer */
 759			min_ctx_size =
 760				width_in_mb * height_in_mb * num_dpb_buffer * 192;
 761		}
 762		break;
 763
 764	case 8: /* MJPEG */
 765		min_dpb_size = 0;
 766		break;
 767
 768	case 16: /* H265 */
 769		image_size = (ALIGN(width, 16) * ALIGN(height, 16) * 3) / 2;
 770		image_size = ALIGN(image_size, 256);
 771
 772		num_dpb_buffer = (le32_to_cpu(msg[59]) & 0xff) + 2;
 773		min_dpb_size = image_size * num_dpb_buffer;
 774		min_ctx_size = ((width + 255) / 16) * ((height + 255) / 16)
 775					   * 16 * num_dpb_buffer + 52 * 1024;
 776		break;
 777
 778	default:
 779		DRM_ERROR("UVD codec not handled %d!\n", stream_type);
 780		return -EINVAL;
 781	}
 782
 783	if (width > pitch) {
 784		DRM_ERROR("Invalid UVD decoding target pitch!\n");
 785		return -EINVAL;
 786	}
 787
 788	if (dpb_size < min_dpb_size) {
 789		DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
 790			  dpb_size, min_dpb_size);
 791		return -EINVAL;
 792	}
 793
 794	buf_sizes[0x1] = dpb_size;
 795	buf_sizes[0x2] = image_size;
 796	buf_sizes[0x4] = min_ctx_size;
 797	/* store image width to adjust nb memory pstate */
 798	adev->uvd.decode_image_width = width;
 799	return 0;
 800}
 801
 802/**
 803 * amdgpu_uvd_cs_msg - handle UVD message
 804 *
 805 * @ctx: UVD parser context
 806 * @bo: buffer object containing the message
 807 * @offset: offset into the buffer object
 808 *
 809 * Peek into the UVD message and extract the session id.
 810 * Make sure that we don't open up to many sessions.
 811 */
 812static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx *ctx,
 813			     struct amdgpu_bo *bo, unsigned offset)
 814{
 815	struct amdgpu_device *adev = ctx->parser->adev;
 816	int32_t *msg, msg_type, handle;
 817	void *ptr;
 818	long r;
 819	int i;
 820
 821	if (offset & 0x3F) {
 822		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
 823		return -EINVAL;
 824	}
 825
 826	r = amdgpu_bo_kmap(bo, &ptr);
 827	if (r) {
 828		DRM_ERROR("Failed mapping the UVD) message (%ld)!\n", r);
 829		return r;
 830	}
 831
 832	msg = ptr + offset;
 833
 834	msg_type = msg[1];
 835	handle = msg[2];
 836
 837	if (handle == 0) {
 838		amdgpu_bo_kunmap(bo);
 839		DRM_ERROR("Invalid UVD handle!\n");
 840		return -EINVAL;
 841	}
 842
 843	switch (msg_type) {
 844	case 0:
 845		/* it's a create msg, calc image size (width * height) */
 846		amdgpu_bo_kunmap(bo);
 847
 848		/* try to alloc a new handle */
 849		for (i = 0; i < adev->uvd.max_handles; ++i) {
 850			if (atomic_read(&adev->uvd.handles[i]) == handle) {
 851				DRM_ERROR(")Handle 0x%x already in use!\n",
 852					  handle);
 853				return -EINVAL;
 854			}
 855
 856			if (!atomic_cmpxchg(&adev->uvd.handles[i], 0, handle)) {
 857				adev->uvd.filp[i] = ctx->parser->filp;
 858				return 0;
 859			}
 860		}
 861
 862		DRM_ERROR("No more free UVD handles!\n");
 863		return -ENOSPC;
 864
 865	case 1:
 866		/* it's a decode msg, calc buffer sizes */
 867		r = amdgpu_uvd_cs_msg_decode(adev, msg, ctx->buf_sizes);
 868		amdgpu_bo_kunmap(bo);
 869		if (r)
 870			return r;
 871
 872		/* validate the handle */
 873		for (i = 0; i < adev->uvd.max_handles; ++i) {
 874			if (atomic_read(&adev->uvd.handles[i]) == handle) {
 875				if (adev->uvd.filp[i] != ctx->parser->filp) {
 876					DRM_ERROR("UVD handle collision detected!\n");
 877					return -EINVAL;
 878				}
 879				return 0;
 880			}
 881		}
 882
 883		DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
 884		return -ENOENT;
 885
 886	case 2:
 887		/* it's a destroy msg, free the handle */
 888		for (i = 0; i < adev->uvd.max_handles; ++i)
 889			atomic_cmpxchg(&adev->uvd.handles[i], handle, 0);
 890		amdgpu_bo_kunmap(bo);
 891		return 0;
 892
 893	default:
 894		DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
 
 895	}
 896
 897	amdgpu_bo_kunmap(bo);
 898	return -EINVAL;
 899}
 900
 901/**
 902 * amdgpu_uvd_cs_pass2 - second parsing round
 903 *
 904 * @ctx: UVD parser context
 905 *
 906 * Patch buffer addresses, make sure buffer sizes are correct.
 907 */
 908static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx *ctx)
 909{
 910	struct amdgpu_bo_va_mapping *mapping;
 911	struct amdgpu_bo *bo;
 912	uint32_t cmd;
 913	uint64_t start, end;
 914	uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
 915	int r;
 916
 917	r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
 918	if (r) {
 919		DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
 920		return r;
 921	}
 922
 923	start = amdgpu_bo_gpu_offset(bo);
 924
 925	end = (mapping->last + 1 - mapping->start);
 926	end = end * AMDGPU_GPU_PAGE_SIZE + start;
 927
 928	addr -= mapping->start * AMDGPU_GPU_PAGE_SIZE;
 929	start += addr;
 930
 931	amdgpu_ib_set_value(ctx->ib, ctx->data0, lower_32_bits(start));
 932	amdgpu_ib_set_value(ctx->ib, ctx->data1, upper_32_bits(start));
 
 
 933
 934	cmd = amdgpu_ib_get_value(ctx->ib, ctx->idx) >> 1;
 935	if (cmd < 0x4) {
 936		if ((end - start) < ctx->buf_sizes[cmd]) {
 937			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
 938				  (unsigned)(end - start),
 939				  ctx->buf_sizes[cmd]);
 940			return -EINVAL;
 941		}
 942
 943	} else if (cmd == 0x206) {
 944		if ((end - start) < ctx->buf_sizes[4]) {
 945			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
 946					  (unsigned)(end - start),
 947					  ctx->buf_sizes[4]);
 948			return -EINVAL;
 949		}
 950	} else if ((cmd != 0x100) && (cmd != 0x204)) {
 951		DRM_ERROR("invalid UVD command %X!\n", cmd);
 952		return -EINVAL;
 953	}
 954
 955	if (!ctx->parser->adev->uvd.address_64_bit) {
 956		if ((start >> 28) != ((end - 1) >> 28)) {
 957			DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
 958				  start, end);
 959			return -EINVAL;
 960		}
 961
 962		if ((cmd == 0 || cmd == 0x3) &&
 963		    (start >> 28) != (ctx->parser->adev->uvd.inst->gpu_addr >> 28)) {
 964			DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
 965				  start, end);
 966			return -EINVAL;
 967		}
 968	}
 969
 970	if (cmd == 0) {
 971		ctx->has_msg_cmd = true;
 972		r = amdgpu_uvd_cs_msg(ctx, bo, addr);
 973		if (r)
 974			return r;
 975	} else if (!ctx->has_msg_cmd) {
 976		DRM_ERROR("Message needed before other commands are send!\n");
 977		return -EINVAL;
 978	}
 979
 980	return 0;
 981}
 982
 983/**
 984 * amdgpu_uvd_cs_reg - parse register writes
 985 *
 986 * @ctx: UVD parser context
 987 * @cb: callback function
 988 *
 989 * Parse the register writes, call cb on each complete command.
 990 */
 991static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx *ctx,
 992			     int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
 993{
 
 994	int i, r;
 995
 996	ctx->idx++;
 997	for (i = 0; i <= ctx->count; ++i) {
 998		unsigned reg = ctx->reg + i;
 999
1000		if (ctx->idx >= ctx->ib->length_dw) {
1001			DRM_ERROR("Register command after end of CS!\n");
1002			return -EINVAL;
1003		}
1004
1005		switch (reg) {
1006		case mmUVD_GPCOM_VCPU_DATA0:
1007			ctx->data0 = ctx->idx;
1008			break;
1009		case mmUVD_GPCOM_VCPU_DATA1:
1010			ctx->data1 = ctx->idx;
1011			break;
1012		case mmUVD_GPCOM_VCPU_CMD:
1013			r = cb(ctx);
1014			if (r)
1015				return r;
1016			break;
1017		case mmUVD_ENGINE_CNTL:
1018		case mmUVD_NO_OP:
1019			break;
1020		default:
1021			DRM_ERROR("Invalid reg 0x%X!\n", reg);
1022			return -EINVAL;
1023		}
1024		ctx->idx++;
1025	}
1026	return 0;
1027}
1028
1029/**
1030 * amdgpu_uvd_cs_packets - parse UVD packets
1031 *
1032 * @ctx: UVD parser context
1033 * @cb: callback function
1034 *
1035 * Parse the command stream packets.
1036 */
1037static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx *ctx,
1038				 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
1039{
 
1040	int r;
1041
1042	for (ctx->idx = 0 ; ctx->idx < ctx->ib->length_dw; ) {
1043		uint32_t cmd = amdgpu_ib_get_value(ctx->ib, ctx->idx);
1044		unsigned type = CP_PACKET_GET_TYPE(cmd);
1045		switch (type) {
1046		case PACKET_TYPE0:
1047			ctx->reg = CP_PACKET0_GET_REG(cmd);
1048			ctx->count = CP_PACKET_GET_COUNT(cmd);
1049			r = amdgpu_uvd_cs_reg(ctx, cb);
1050			if (r)
1051				return r;
1052			break;
1053		case PACKET_TYPE2:
1054			++ctx->idx;
1055			break;
1056		default:
1057			DRM_ERROR("Unknown packet type %d !\n", type);
1058			return -EINVAL;
1059		}
1060	}
1061	return 0;
1062}
1063
1064/**
1065 * amdgpu_uvd_ring_parse_cs - UVD command submission parser
1066 *
1067 * @parser: Command submission parser context
1068 * @job: the job to parse
1069 * @ib: the IB to patch
1070 *
1071 * Parse the command stream, patch in addresses as necessary.
1072 */
1073int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser,
1074			     struct amdgpu_job *job,
1075			     struct amdgpu_ib *ib)
1076{
1077	struct amdgpu_uvd_cs_ctx ctx = {};
1078	unsigned buf_sizes[] = {
1079		[0x00000000]	=	2048,
1080		[0x00000001]	=	0xFFFFFFFF,
1081		[0x00000002]	=	0xFFFFFFFF,
1082		[0x00000003]	=	2048,
1083		[0x00000004]	=	0xFFFFFFFF,
1084	};
 
1085	int r;
1086
1087	job->vm = NULL;
1088	ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
1089
1090	if (ib->length_dw % 16) {
1091		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
1092			  ib->length_dw);
1093		return -EINVAL;
1094	}
1095
1096	ctx.parser = parser;
1097	ctx.buf_sizes = buf_sizes;
1098	ctx.ib = ib;
1099
1100	/* first round only required on chips without UVD 64 bit address support */
1101	if (!parser->adev->uvd.address_64_bit) {
1102		/* first round, make sure the buffers are actually in the UVD segment */
1103		r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass1);
1104		if (r)
1105			return r;
1106	}
1107
1108	/* second round, patch buffer addresses into the command stream */
1109	r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass2);
1110	if (r)
1111		return r;
1112
1113	if (!ctx.has_msg_cmd) {
1114		DRM_ERROR("UVD-IBs need a msg command!\n");
1115		return -EINVAL;
1116	}
1117
1118	return 0;
1119}
1120
1121static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
1122			       bool direct, struct dma_fence **fence)
1123{
1124	struct amdgpu_device *adev = ring->adev;
1125	struct dma_fence *f = NULL;
1126	struct amdgpu_job *job;
1127	struct amdgpu_ib *ib;
1128	uint32_t data[4];
1129	uint64_t addr;
1130	long r;
1131	int i;
1132	unsigned offset_idx = 0;
1133	unsigned offset[3] = { UVD_BASE_SI, 0, 0 };
1134
1135	r = amdgpu_job_alloc_with_ib(ring->adev, &adev->uvd.entity,
1136				     AMDGPU_FENCE_OWNER_UNDEFINED,
1137				     64, direct ? AMDGPU_IB_POOL_DIRECT :
1138				     AMDGPU_IB_POOL_DELAYED, &job);
 
 
 
 
 
 
 
 
 
 
1139	if (r)
1140		return r;
1141
1142	if (adev->asic_type >= CHIP_VEGA10) {
1143		offset_idx = 1 + ring->me;
1144		offset[1] = adev->reg_offset[UVD_HWIP][0][1];
1145		offset[2] = adev->reg_offset[UVD_HWIP][1][1];
1146	}
1147
1148	data[0] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_DATA0, 0);
1149	data[1] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_DATA1, 0);
1150	data[2] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_CMD, 0);
1151	data[3] = PACKET0(offset[offset_idx] + UVD_NO_OP, 0);
1152
1153	ib = &job->ibs[0];
1154	addr = amdgpu_bo_gpu_offset(bo);
1155	ib->ptr[0] = data[0];
1156	ib->ptr[1] = addr;
1157	ib->ptr[2] = data[1];
1158	ib->ptr[3] = addr >> 32;
1159	ib->ptr[4] = data[2];
1160	ib->ptr[5] = 0;
1161	for (i = 6; i < 16; i += 2) {
1162		ib->ptr[i] = data[3];
1163		ib->ptr[i+1] = 0;
1164	}
1165	ib->length_dw = 16;
1166
1167	if (direct) {
1168		r = dma_resv_wait_timeout(bo->tbo.base.resv,
1169					  DMA_RESV_USAGE_KERNEL, false,
1170					  msecs_to_jiffies(10));
1171		if (r == 0)
1172			r = -ETIMEDOUT;
1173		if (r < 0)
1174			goto err_free;
1175
1176		r = amdgpu_job_submit_direct(job, ring, &f);
1177		if (r)
1178			goto err_free;
1179	} else {
1180		r = drm_sched_job_add_resv_dependencies(&job->base,
1181							bo->tbo.base.resv,
1182							DMA_RESV_USAGE_KERNEL);
1183		if (r)
1184			goto err_free;
1185
1186		f = amdgpu_job_submit(job);
 
 
 
1187	}
1188
1189	amdgpu_bo_reserve(bo, true);
1190	amdgpu_bo_fence(bo, f, false);
1191	amdgpu_bo_unreserve(bo);
 
1192
1193	if (fence)
1194		*fence = dma_fence_get(f);
1195	dma_fence_put(f);
1196
1197	return 0;
1198
1199err_free:
1200	amdgpu_job_free(job);
 
 
 
 
1201	return r;
1202}
1203
1204/* multiple fence commands without any stream commands in between can
1205   crash the vcpu so just try to emmit a dummy create/destroy msg to
1206   avoid this */
1207int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
1208			      struct dma_fence **fence)
1209{
1210	struct amdgpu_device *adev = ring->adev;
1211	struct amdgpu_bo *bo = adev->uvd.ib_bo;
1212	uint32_t *msg;
1213	int i;
 
 
 
 
 
 
1214
1215	msg = amdgpu_bo_kptr(bo);
1216	/* stitch together an UVD create msg */
1217	msg[0] = cpu_to_le32(0x00000de4);
1218	msg[1] = cpu_to_le32(0x00000000);
1219	msg[2] = cpu_to_le32(handle);
1220	msg[3] = cpu_to_le32(0x00000000);
1221	msg[4] = cpu_to_le32(0x00000000);
1222	msg[5] = cpu_to_le32(0x00000000);
1223	msg[6] = cpu_to_le32(0x00000000);
1224	msg[7] = cpu_to_le32(0x00000780);
1225	msg[8] = cpu_to_le32(0x00000440);
1226	msg[9] = cpu_to_le32(0x00000000);
1227	msg[10] = cpu_to_le32(0x01b37000);
1228	for (i = 11; i < 1024; ++i)
1229		msg[i] = cpu_to_le32(0x0);
1230
1231	return amdgpu_uvd_send_msg(ring, bo, true, fence);
1232
1233}
1234
1235int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
1236			       bool direct, struct dma_fence **fence)
1237{
1238	struct amdgpu_device *adev = ring->adev;
1239	struct amdgpu_bo *bo = NULL;
1240	uint32_t *msg;
1241	int r, i;
1242
1243	if (direct) {
1244		bo = adev->uvd.ib_bo;
1245	} else {
1246		r = amdgpu_uvd_create_msg_bo_helper(adev, 4096, &bo);
1247		if (r)
1248			return r;
1249	}
1250
1251	msg = amdgpu_bo_kptr(bo);
1252	/* stitch together an UVD destroy msg */
1253	msg[0] = cpu_to_le32(0x00000de4);
1254	msg[1] = cpu_to_le32(0x00000002);
1255	msg[2] = cpu_to_le32(handle);
1256	msg[3] = cpu_to_le32(0x00000000);
1257	for (i = 4; i < 1024; ++i)
1258		msg[i] = cpu_to_le32(0x0);
1259
1260	r = amdgpu_uvd_send_msg(ring, bo, direct, fence);
1261
1262	if (!direct)
1263		amdgpu_bo_free_kernel(&bo, NULL, (void **)&msg);
1264
1265	return r;
1266}
1267
1268static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
1269{
1270	struct amdgpu_device *adev =
1271		container_of(work, struct amdgpu_device, uvd.idle_work.work);
1272	unsigned fences = 0, i, j;
1273
1274	for (i = 0; i < adev->uvd.num_uvd_inst; ++i) {
1275		if (adev->uvd.harvest_config & (1 << i))
1276			continue;
1277		fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring);
1278		for (j = 0; j < adev->uvd.num_enc_rings; ++j) {
1279			fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring_enc[j]);
1280		}
1281	}
1282
1283	if (fences == 0) {
1284		if (adev->pm.dpm_enabled) {
1285			amdgpu_dpm_enable_uvd(adev, false);
1286		} else {
1287			amdgpu_asic_set_uvd_clocks(adev, 0, 0);
1288			/* shutdown the UVD block */
1289			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1290							       AMD_PG_STATE_GATE);
1291			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1292							       AMD_CG_STATE_GATE);
1293		}
1294	} else {
1295		schedule_delayed_work(&adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1296	}
1297}
1298
1299void amdgpu_uvd_ring_begin_use(struct amdgpu_ring *ring)
1300{
1301	struct amdgpu_device *adev = ring->adev;
1302	bool set_clocks;
1303
1304	if (amdgpu_sriov_vf(adev))
1305		return;
1306
1307	set_clocks = !cancel_delayed_work_sync(&adev->uvd.idle_work);
1308	if (set_clocks) {
1309		if (adev->pm.dpm_enabled) {
1310			amdgpu_dpm_enable_uvd(adev, true);
1311		} else {
1312			amdgpu_asic_set_uvd_clocks(adev, 53300, 40000);
1313			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1314							       AMD_CG_STATE_UNGATE);
1315			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1316							       AMD_PG_STATE_UNGATE);
1317		}
1318	}
1319}
1320
1321void amdgpu_uvd_ring_end_use(struct amdgpu_ring *ring)
1322{
1323	if (!amdgpu_sriov_vf(ring->adev))
1324		schedule_delayed_work(&ring->adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1325}
1326
1327/**
1328 * amdgpu_uvd_ring_test_ib - test ib execution
1329 *
1330 * @ring: amdgpu_ring pointer
1331 * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
1332 *
1333 * Test if we can successfully execute an IB
1334 */
1335int amdgpu_uvd_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1336{
1337	struct dma_fence *fence;
1338	long r;
1339
1340	r = amdgpu_uvd_get_create_msg(ring, 1, &fence);
1341	if (r)
1342		goto error;
1343
1344	r = dma_fence_wait_timeout(fence, false, timeout);
1345	dma_fence_put(fence);
1346	if (r == 0)
1347		r = -ETIMEDOUT;
1348	if (r < 0)
1349		goto error;
1350
1351	r = amdgpu_uvd_get_destroy_msg(ring, 1, true, &fence);
1352	if (r)
1353		goto error;
1354
1355	r = dma_fence_wait_timeout(fence, false, timeout);
1356	if (r == 0)
1357		r = -ETIMEDOUT;
1358	else if (r > 0)
1359		r = 0;
1360
1361	dma_fence_put(fence);
1362
1363error:
1364	return r;
1365}
1366
1367/**
1368 * amdgpu_uvd_used_handles - returns used UVD handles
1369 *
1370 * @adev: amdgpu_device pointer
1371 *
1372 * Returns the number of UVD handles in use
1373 */
1374uint32_t amdgpu_uvd_used_handles(struct amdgpu_device *adev)
1375{
1376	unsigned i;
1377	uint32_t used_handles = 0;
1378
1379	for (i = 0; i < adev->uvd.max_handles; ++i) {
1380		/*
1381		 * Handles can be freed in any order, and not
1382		 * necessarily linear. So we need to count
1383		 * all non-zero handles.
1384		 */
1385		if (atomic_read(&adev->uvd.handles[i]))
1386			used_handles++;
1387	}
1388
1389	return used_handles;
1390}
v5.4
   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
  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	"amdgpu/bonaire_uvd.bin"
  57#define FIRMWARE_KABINI	"amdgpu/kabini_uvd.bin"
  58#define FIRMWARE_KAVERI	"amdgpu/kaveri_uvd.bin"
  59#define FIRMWARE_HAWAII	"amdgpu/hawaii_uvd.bin"
  60#define FIRMWARE_MULLINS	"amdgpu/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#define FIRMWARE_VEGAM		"amdgpu/vegam_uvd.bin"
  70
  71#define FIRMWARE_VEGA10		"amdgpu/vega10_uvd.bin"
  72#define FIRMWARE_VEGA12		"amdgpu/vega12_uvd.bin"
  73#define FIRMWARE_VEGA20		"amdgpu/vega20_uvd.bin"
  74
  75/* These are common relative offsets for all asics, from uvd_7_0_offset.h,  */
  76#define UVD_GPCOM_VCPU_CMD		0x03c3
  77#define UVD_GPCOM_VCPU_DATA0	0x03c4
  78#define UVD_GPCOM_VCPU_DATA1	0x03c5
  79#define UVD_NO_OP				0x03ff
  80#define UVD_BASE_SI				0x3800
  81
  82/**
  83 * amdgpu_uvd_cs_ctx - Command submission parser context
  84 *
  85 * Used for emulating virtual memory support on UVD 4.2.
  86 */
  87struct amdgpu_uvd_cs_ctx {
  88	struct amdgpu_cs_parser *parser;
  89	unsigned reg, count;
  90	unsigned data0, data1;
  91	unsigned idx;
  92	unsigned ib_idx;
  93
  94	/* does the IB has a msg command */
  95	bool has_msg_cmd;
  96
  97	/* minimum buffer sizes */
  98	unsigned *buf_sizes;
  99};
 100
 
 
 
 
 
 
 101#ifdef CONFIG_DRM_AMDGPU_CIK
 102MODULE_FIRMWARE(FIRMWARE_BONAIRE);
 103MODULE_FIRMWARE(FIRMWARE_KABINI);
 104MODULE_FIRMWARE(FIRMWARE_KAVERI);
 105MODULE_FIRMWARE(FIRMWARE_HAWAII);
 106MODULE_FIRMWARE(FIRMWARE_MULLINS);
 107#endif
 108MODULE_FIRMWARE(FIRMWARE_TONGA);
 109MODULE_FIRMWARE(FIRMWARE_CARRIZO);
 110MODULE_FIRMWARE(FIRMWARE_FIJI);
 111MODULE_FIRMWARE(FIRMWARE_STONEY);
 112MODULE_FIRMWARE(FIRMWARE_POLARIS10);
 113MODULE_FIRMWARE(FIRMWARE_POLARIS11);
 114MODULE_FIRMWARE(FIRMWARE_POLARIS12);
 115MODULE_FIRMWARE(FIRMWARE_VEGAM);
 116
 117MODULE_FIRMWARE(FIRMWARE_VEGA10);
 118MODULE_FIRMWARE(FIRMWARE_VEGA12);
 119MODULE_FIRMWARE(FIRMWARE_VEGA20);
 120
 121static void amdgpu_uvd_idle_work_handler(struct work_struct *work);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 122
 123int amdgpu_uvd_sw_init(struct amdgpu_device *adev)
 124{
 125	unsigned long bo_size;
 126	const char *fw_name;
 127	const struct common_firmware_header *hdr;
 128	unsigned family_id;
 129	int i, j, r;
 130
 131	INIT_DELAYED_WORK(&adev->uvd.idle_work, amdgpu_uvd_idle_work_handler);
 132
 133	switch (adev->asic_type) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 134#ifdef CONFIG_DRM_AMDGPU_CIK
 135	case CHIP_BONAIRE:
 136		fw_name = FIRMWARE_BONAIRE;
 137		break;
 138	case CHIP_KABINI:
 139		fw_name = FIRMWARE_KABINI;
 140		break;
 141	case CHIP_KAVERI:
 142		fw_name = FIRMWARE_KAVERI;
 143		break;
 144	case CHIP_HAWAII:
 145		fw_name = FIRMWARE_HAWAII;
 146		break;
 147	case CHIP_MULLINS:
 148		fw_name = FIRMWARE_MULLINS;
 149		break;
 150#endif
 151	case CHIP_TONGA:
 152		fw_name = FIRMWARE_TONGA;
 153		break;
 154	case CHIP_FIJI:
 155		fw_name = FIRMWARE_FIJI;
 156		break;
 157	case CHIP_CARRIZO:
 158		fw_name = FIRMWARE_CARRIZO;
 159		break;
 160	case CHIP_STONEY:
 161		fw_name = FIRMWARE_STONEY;
 162		break;
 163	case CHIP_POLARIS10:
 164		fw_name = FIRMWARE_POLARIS10;
 165		break;
 166	case CHIP_POLARIS11:
 167		fw_name = FIRMWARE_POLARIS11;
 168		break;
 169	case CHIP_POLARIS12:
 170		fw_name = FIRMWARE_POLARIS12;
 171		break;
 172	case CHIP_VEGA10:
 173		fw_name = FIRMWARE_VEGA10;
 174		break;
 175	case CHIP_VEGA12:
 176		fw_name = FIRMWARE_VEGA12;
 177		break;
 178	case CHIP_VEGAM:
 179		fw_name = FIRMWARE_VEGAM;
 180		break;
 181	case CHIP_VEGA20:
 182		fw_name = FIRMWARE_VEGA20;
 183		break;
 184	default:
 185		return -EINVAL;
 186	}
 187
 188	r = request_firmware(&adev->uvd.fw, fw_name, adev->dev);
 189	if (r) {
 190		dev_err(adev->dev, "amdgpu_uvd: Can't load firmware \"%s\"\n",
 191			fw_name);
 192		return r;
 193	}
 194
 195	r = amdgpu_ucode_validate(adev->uvd.fw);
 196	if (r) {
 197		dev_err(adev->dev, "amdgpu_uvd: Can't validate firmware \"%s\"\n",
 198			fw_name);
 199		release_firmware(adev->uvd.fw);
 200		adev->uvd.fw = NULL;
 201		return r;
 202	}
 203
 204	/* Set the default UVD handles that the firmware can handle */
 205	adev->uvd.max_handles = AMDGPU_DEFAULT_UVD_HANDLES;
 206
 207	hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
 208	family_id = le32_to_cpu(hdr->ucode_version) & 0xff;
 209
 210	if (adev->asic_type < CHIP_VEGA20) {
 211		unsigned version_major, version_minor;
 212
 213		version_major = (le32_to_cpu(hdr->ucode_version) >> 24) & 0xff;
 214		version_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
 215		DRM_INFO("Found UVD firmware Version: %hu.%hu Family ID: %hu\n",
 216			version_major, version_minor, family_id);
 217
 218		/*
 219		 * Limit the number of UVD handles depending on microcode major
 220		 * and minor versions. The firmware version which has 40 UVD
 221		 * instances support is 1.80. So all subsequent versions should
 222		 * also have the same support.
 223		 */
 224		if ((version_major > 0x01) ||
 225		    ((version_major == 0x01) && (version_minor >= 0x50)))
 226			adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
 227
 228		adev->uvd.fw_version = ((version_major << 24) | (version_minor << 16) |
 229					(family_id << 8));
 230
 231		if ((adev->asic_type == CHIP_POLARIS10 ||
 232		     adev->asic_type == CHIP_POLARIS11) &&
 233		    (adev->uvd.fw_version < FW_1_66_16))
 234			DRM_ERROR("POLARIS10/11 UVD firmware version %hu.%hu is too old.\n",
 235				  version_major, version_minor);
 236	} else {
 237		unsigned int enc_major, enc_minor, dec_minor;
 238
 239		dec_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff;
 240		enc_minor = (le32_to_cpu(hdr->ucode_version) >> 24) & 0x3f;
 241		enc_major = (le32_to_cpu(hdr->ucode_version) >> 30) & 0x3;
 242		DRM_INFO("Found UVD firmware ENC: %hu.%hu DEC: .%hu Family ID: %hu\n",
 243			enc_major, enc_minor, dec_minor, family_id);
 244
 245		adev->uvd.max_handles = AMDGPU_MAX_UVD_HANDLES;
 246
 247		adev->uvd.fw_version = le32_to_cpu(hdr->ucode_version);
 248	}
 249
 250	bo_size = AMDGPU_UVD_STACK_SIZE + AMDGPU_UVD_HEAP_SIZE
 251		  +  AMDGPU_UVD_SESSION_SIZE * adev->uvd.max_handles;
 252	if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP)
 253		bo_size += AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr->ucode_size_bytes) + 8);
 254
 255	for (j = 0; j < adev->uvd.num_uvd_inst; j++) {
 256		if (adev->uvd.harvest_config & (1 << j))
 257			continue;
 258		r = amdgpu_bo_create_kernel(adev, bo_size, PAGE_SIZE,
 259					    AMDGPU_GEM_DOMAIN_VRAM, &adev->uvd.inst[j].vcpu_bo,
 260					    &adev->uvd.inst[j].gpu_addr, &adev->uvd.inst[j].cpu_addr);
 261		if (r) {
 262			dev_err(adev->dev, "(%d) failed to allocate UVD bo\n", r);
 263			return r;
 264		}
 265	}
 266
 267	for (i = 0; i < adev->uvd.max_handles; ++i) {
 268		atomic_set(&adev->uvd.handles[i], 0);
 269		adev->uvd.filp[i] = NULL;
 270	}
 271
 272	/* from uvd v5.0 HW addressing capacity increased to 64 bits */
 273	if (!amdgpu_device_ip_block_version_cmp(adev, AMD_IP_BLOCK_TYPE_UVD, 5, 0))
 274		adev->uvd.address_64_bit = true;
 275
 
 
 
 
 276	switch (adev->asic_type) {
 277	case CHIP_TONGA:
 278		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_65_10;
 279		break;
 280	case CHIP_CARRIZO:
 281		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_11;
 282		break;
 283	case CHIP_FIJI:
 284		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_87_12;
 285		break;
 286	case CHIP_STONEY:
 287		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_37_15;
 288		break;
 289	default:
 290		adev->uvd.use_ctx_buf = adev->asic_type >= CHIP_POLARIS10;
 291	}
 292
 293	return 0;
 294}
 295
 296int amdgpu_uvd_sw_fini(struct amdgpu_device *adev)
 297{
 
 298	int i, j;
 299
 300	drm_sched_entity_destroy(&adev->uvd.entity);
 301
 302	for (j = 0; j < adev->uvd.num_uvd_inst; ++j) {
 303		if (adev->uvd.harvest_config & (1 << j))
 304			continue;
 305		kvfree(adev->uvd.inst[j].saved_bo);
 306
 307		amdgpu_bo_free_kernel(&adev->uvd.inst[j].vcpu_bo,
 308				      &adev->uvd.inst[j].gpu_addr,
 309				      (void **)&adev->uvd.inst[j].cpu_addr);
 310
 311		amdgpu_ring_fini(&adev->uvd.inst[j].ring);
 312
 313		for (i = 0; i < AMDGPU_MAX_UVD_ENC_RINGS; ++i)
 314			amdgpu_ring_fini(&adev->uvd.inst[j].ring_enc[i]);
 315	}
 
 316	release_firmware(adev->uvd.fw);
 317
 318	return 0;
 319}
 320
 321/**
 322 * amdgpu_uvd_entity_init - init entity
 323 *
 324 * @adev: amdgpu_device pointer
 325 *
 326 */
 327int amdgpu_uvd_entity_init(struct amdgpu_device *adev)
 328{
 329	struct amdgpu_ring *ring;
 330	struct drm_sched_rq *rq;
 331	int r;
 332
 333	ring = &adev->uvd.inst[0].ring;
 334	rq = &ring->sched.sched_rq[DRM_SCHED_PRIORITY_NORMAL];
 335	r = drm_sched_entity_init(&adev->uvd.entity, &rq, 1, NULL);
 
 336	if (r) {
 337		DRM_ERROR("Failed setting up UVD kernel entity.\n");
 338		return r;
 339	}
 340
 341	return 0;
 342}
 343
 344int amdgpu_uvd_suspend(struct amdgpu_device *adev)
 345{
 346	unsigned size;
 347	void *ptr;
 348	int i, j;
 
 349
 350	cancel_delayed_work_sync(&adev->uvd.idle_work);
 351
 352	/* only valid for physical mode */
 353	if (adev->asic_type < CHIP_POLARIS10) {
 354		for (i = 0; i < adev->uvd.max_handles; ++i)
 355			if (atomic_read(&adev->uvd.handles[i]))
 356				break;
 357
 358		if (i == adev->uvd.max_handles)
 359			return 0;
 360	}
 361
 362	for (j = 0; j < adev->uvd.num_uvd_inst; ++j) {
 363		if (adev->uvd.harvest_config & (1 << j))
 364			continue;
 365		if (adev->uvd.inst[j].vcpu_bo == NULL)
 366			continue;
 367
 368		size = amdgpu_bo_size(adev->uvd.inst[j].vcpu_bo);
 369		ptr = adev->uvd.inst[j].cpu_addr;
 370
 371		adev->uvd.inst[j].saved_bo = kvmalloc(size, GFP_KERNEL);
 372		if (!adev->uvd.inst[j].saved_bo)
 373			return -ENOMEM;
 374
 375		memcpy_fromio(adev->uvd.inst[j].saved_bo, ptr, size);
 
 
 
 
 
 
 
 
 376	}
 
 
 
 
 377	return 0;
 378}
 379
 380int amdgpu_uvd_resume(struct amdgpu_device *adev)
 381{
 382	unsigned size;
 383	void *ptr;
 384	int i;
 385
 386	for (i = 0; i < adev->uvd.num_uvd_inst; i++) {
 387		if (adev->uvd.harvest_config & (1 << i))
 388			continue;
 389		if (adev->uvd.inst[i].vcpu_bo == NULL)
 390			return -EINVAL;
 391
 392		size = amdgpu_bo_size(adev->uvd.inst[i].vcpu_bo);
 393		ptr = adev->uvd.inst[i].cpu_addr;
 394
 395		if (adev->uvd.inst[i].saved_bo != NULL) {
 396			memcpy_toio(ptr, adev->uvd.inst[i].saved_bo, size);
 
 
 
 397			kvfree(adev->uvd.inst[i].saved_bo);
 398			adev->uvd.inst[i].saved_bo = NULL;
 399		} else {
 400			const struct common_firmware_header *hdr;
 401			unsigned offset;
 402
 403			hdr = (const struct common_firmware_header *)adev->uvd.fw->data;
 404			if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP) {
 405				offset = le32_to_cpu(hdr->ucode_array_offset_bytes);
 406				memcpy_toio(adev->uvd.inst[i].cpu_addr, adev->uvd.fw->data + offset,
 407					    le32_to_cpu(hdr->ucode_size_bytes));
 
 
 
 408				size -= le32_to_cpu(hdr->ucode_size_bytes);
 409				ptr += le32_to_cpu(hdr->ucode_size_bytes);
 410			}
 411			memset_io(ptr, 0, size);
 412			/* to restore uvd fence seq */
 413			amdgpu_fence_driver_force_completion(&adev->uvd.inst[i].ring);
 414		}
 415	}
 416	return 0;
 417}
 418
 419void amdgpu_uvd_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
 420{
 421	struct amdgpu_ring *ring = &adev->uvd.inst[0].ring;
 422	int i, r;
 423
 424	for (i = 0; i < adev->uvd.max_handles; ++i) {
 425		uint32_t handle = atomic_read(&adev->uvd.handles[i]);
 426
 427		if (handle != 0 && adev->uvd.filp[i] == filp) {
 428			struct dma_fence *fence;
 429
 430			r = amdgpu_uvd_get_destroy_msg(ring, handle, false,
 431						       &fence);
 432			if (r) {
 433				DRM_ERROR("Error destroying UVD %d!\n", r);
 434				continue;
 435			}
 436
 437			dma_fence_wait(fence, false);
 438			dma_fence_put(fence);
 439
 440			adev->uvd.filp[i] = NULL;
 441			atomic_set(&adev->uvd.handles[i], 0);
 442		}
 443	}
 444}
 445
 446static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *abo)
 447{
 448	int i;
 449	for (i = 0; i < abo->placement.num_placement; ++i) {
 450		abo->placements[i].fpfn = 0 >> PAGE_SHIFT;
 451		abo->placements[i].lpfn = (256 * 1024 * 1024) >> PAGE_SHIFT;
 452	}
 453}
 454
 455static u64 amdgpu_uvd_get_addr_from_ctx(struct amdgpu_uvd_cs_ctx *ctx)
 456{
 457	uint32_t lo, hi;
 458	uint64_t addr;
 459
 460	lo = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data0);
 461	hi = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->data1);
 462	addr = ((uint64_t)lo) | (((uint64_t)hi) << 32);
 463
 464	return addr;
 465}
 466
 467/**
 468 * amdgpu_uvd_cs_pass1 - first parsing round
 469 *
 470 * @ctx: UVD parser context
 471 *
 472 * Make sure UVD message and feedback buffers are in VRAM and
 473 * nobody is violating an 256MB boundary.
 474 */
 475static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx *ctx)
 476{
 477	struct ttm_operation_ctx tctx = { false, false };
 478	struct amdgpu_bo_va_mapping *mapping;
 479	struct amdgpu_bo *bo;
 480	uint32_t cmd;
 481	uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
 482	int r = 0;
 483
 484	r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
 485	if (r) {
 486		DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
 487		return r;
 488	}
 489
 490	if (!ctx->parser->adev->uvd.address_64_bit) {
 491		/* check if it's a message or feedback command */
 492		cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
 493		if (cmd == 0x0 || cmd == 0x3) {
 494			/* yes, force it into VRAM */
 495			uint32_t domain = AMDGPU_GEM_DOMAIN_VRAM;
 496			amdgpu_bo_placement_from_domain(bo, domain);
 497		}
 498		amdgpu_uvd_force_into_uvd_segment(bo);
 499
 500		r = ttm_bo_validate(&bo->tbo, &bo->placement, &tctx);
 501	}
 502
 503	return r;
 504}
 505
 506/**
 507 * amdgpu_uvd_cs_msg_decode - handle UVD decode message
 508 *
 
 509 * @msg: pointer to message structure
 510 * @buf_sizes: returned buffer sizes
 511 *
 512 * Peek into the decode message and calculate the necessary buffer sizes.
 513 */
 514static int amdgpu_uvd_cs_msg_decode(struct amdgpu_device *adev, uint32_t *msg,
 515	unsigned buf_sizes[])
 516{
 517	unsigned stream_type = msg[4];
 518	unsigned width = msg[6];
 519	unsigned height = msg[7];
 520	unsigned dpb_size = msg[9];
 521	unsigned pitch = msg[28];
 522	unsigned level = msg[57];
 523
 524	unsigned width_in_mb = width / 16;
 525	unsigned height_in_mb = ALIGN(height / 16, 2);
 526	unsigned fs_in_mb = width_in_mb * height_in_mb;
 527
 528	unsigned image_size, tmp, min_dpb_size, num_dpb_buffer;
 529	unsigned min_ctx_size = ~0;
 530
 531	image_size = width * height;
 532	image_size += image_size / 2;
 533	image_size = ALIGN(image_size, 1024);
 534
 535	switch (stream_type) {
 536	case 0: /* H264 */
 537		switch(level) {
 538		case 30:
 539			num_dpb_buffer = 8100 / fs_in_mb;
 540			break;
 541		case 31:
 542			num_dpb_buffer = 18000 / fs_in_mb;
 543			break;
 544		case 32:
 545			num_dpb_buffer = 20480 / fs_in_mb;
 546			break;
 547		case 41:
 548			num_dpb_buffer = 32768 / fs_in_mb;
 549			break;
 550		case 42:
 551			num_dpb_buffer = 34816 / fs_in_mb;
 552			break;
 553		case 50:
 554			num_dpb_buffer = 110400 / fs_in_mb;
 555			break;
 556		case 51:
 557			num_dpb_buffer = 184320 / fs_in_mb;
 558			break;
 559		default:
 560			num_dpb_buffer = 184320 / fs_in_mb;
 561			break;
 562		}
 563		num_dpb_buffer++;
 564		if (num_dpb_buffer > 17)
 565			num_dpb_buffer = 17;
 566
 567		/* reference picture buffer */
 568		min_dpb_size = image_size * num_dpb_buffer;
 569
 570		/* macroblock context buffer */
 571		min_dpb_size += width_in_mb * height_in_mb * num_dpb_buffer * 192;
 572
 573		/* IT surface buffer */
 574		min_dpb_size += width_in_mb * height_in_mb * 32;
 575		break;
 576
 577	case 1: /* VC1 */
 578
 579		/* reference picture buffer */
 580		min_dpb_size = image_size * 3;
 581
 582		/* CONTEXT_BUFFER */
 583		min_dpb_size += width_in_mb * height_in_mb * 128;
 584
 585		/* IT surface buffer */
 586		min_dpb_size += width_in_mb * 64;
 587
 588		/* DB surface buffer */
 589		min_dpb_size += width_in_mb * 128;
 590
 591		/* BP */
 592		tmp = max(width_in_mb, height_in_mb);
 593		min_dpb_size += ALIGN(tmp * 7 * 16, 64);
 594		break;
 595
 596	case 3: /* MPEG2 */
 597
 598		/* reference picture buffer */
 599		min_dpb_size = image_size * 3;
 600		break;
 601
 602	case 4: /* MPEG4 */
 603
 604		/* reference picture buffer */
 605		min_dpb_size = image_size * 3;
 606
 607		/* CM */
 608		min_dpb_size += width_in_mb * height_in_mb * 64;
 609
 610		/* IT surface buffer */
 611		min_dpb_size += ALIGN(width_in_mb * height_in_mb * 32, 64);
 612		break;
 613
 614	case 7: /* H264 Perf */
 615		switch(level) {
 616		case 30:
 617			num_dpb_buffer = 8100 / fs_in_mb;
 618			break;
 619		case 31:
 620			num_dpb_buffer = 18000 / fs_in_mb;
 621			break;
 622		case 32:
 623			num_dpb_buffer = 20480 / fs_in_mb;
 624			break;
 625		case 41:
 626			num_dpb_buffer = 32768 / fs_in_mb;
 627			break;
 628		case 42:
 629			num_dpb_buffer = 34816 / fs_in_mb;
 630			break;
 631		case 50:
 632			num_dpb_buffer = 110400 / fs_in_mb;
 633			break;
 634		case 51:
 635			num_dpb_buffer = 184320 / fs_in_mb;
 636			break;
 637		default:
 638			num_dpb_buffer = 184320 / fs_in_mb;
 639			break;
 640		}
 641		num_dpb_buffer++;
 642		if (num_dpb_buffer > 17)
 643			num_dpb_buffer = 17;
 644
 645		/* reference picture buffer */
 646		min_dpb_size = image_size * num_dpb_buffer;
 647
 648		if (!adev->uvd.use_ctx_buf){
 649			/* macroblock context buffer */
 650			min_dpb_size +=
 651				width_in_mb * height_in_mb * num_dpb_buffer * 192;
 652
 653			/* IT surface buffer */
 654			min_dpb_size += width_in_mb * height_in_mb * 32;
 655		} else {
 656			/* macroblock context buffer */
 657			min_ctx_size =
 658				width_in_mb * height_in_mb * num_dpb_buffer * 192;
 659		}
 660		break;
 661
 662	case 8: /* MJPEG */
 663		min_dpb_size = 0;
 664		break;
 665
 666	case 16: /* H265 */
 667		image_size = (ALIGN(width, 16) * ALIGN(height, 16) * 3) / 2;
 668		image_size = ALIGN(image_size, 256);
 669
 670		num_dpb_buffer = (le32_to_cpu(msg[59]) & 0xff) + 2;
 671		min_dpb_size = image_size * num_dpb_buffer;
 672		min_ctx_size = ((width + 255) / 16) * ((height + 255) / 16)
 673					   * 16 * num_dpb_buffer + 52 * 1024;
 674		break;
 675
 676	default:
 677		DRM_ERROR("UVD codec not handled %d!\n", stream_type);
 678		return -EINVAL;
 679	}
 680
 681	if (width > pitch) {
 682		DRM_ERROR("Invalid UVD decoding target pitch!\n");
 683		return -EINVAL;
 684	}
 685
 686	if (dpb_size < min_dpb_size) {
 687		DRM_ERROR("Invalid dpb_size in UVD message (%d / %d)!\n",
 688			  dpb_size, min_dpb_size);
 689		return -EINVAL;
 690	}
 691
 692	buf_sizes[0x1] = dpb_size;
 693	buf_sizes[0x2] = image_size;
 694	buf_sizes[0x4] = min_ctx_size;
 695	/* store image width to adjust nb memory pstate */
 696	adev->uvd.decode_image_width = width;
 697	return 0;
 698}
 699
 700/**
 701 * amdgpu_uvd_cs_msg - handle UVD message
 702 *
 703 * @ctx: UVD parser context
 704 * @bo: buffer object containing the message
 705 * @offset: offset into the buffer object
 706 *
 707 * Peek into the UVD message and extract the session id.
 708 * Make sure that we don't open up to many sessions.
 709 */
 710static int amdgpu_uvd_cs_msg(struct amdgpu_uvd_cs_ctx *ctx,
 711			     struct amdgpu_bo *bo, unsigned offset)
 712{
 713	struct amdgpu_device *adev = ctx->parser->adev;
 714	int32_t *msg, msg_type, handle;
 715	void *ptr;
 716	long r;
 717	int i;
 718
 719	if (offset & 0x3F) {
 720		DRM_ERROR("UVD messages must be 64 byte aligned!\n");
 721		return -EINVAL;
 722	}
 723
 724	r = amdgpu_bo_kmap(bo, &ptr);
 725	if (r) {
 726		DRM_ERROR("Failed mapping the UVD) message (%ld)!\n", r);
 727		return r;
 728	}
 729
 730	msg = ptr + offset;
 731
 732	msg_type = msg[1];
 733	handle = msg[2];
 734
 735	if (handle == 0) {
 
 736		DRM_ERROR("Invalid UVD handle!\n");
 737		return -EINVAL;
 738	}
 739
 740	switch (msg_type) {
 741	case 0:
 742		/* it's a create msg, calc image size (width * height) */
 743		amdgpu_bo_kunmap(bo);
 744
 745		/* try to alloc a new handle */
 746		for (i = 0; i < adev->uvd.max_handles; ++i) {
 747			if (atomic_read(&adev->uvd.handles[i]) == handle) {
 748				DRM_ERROR(")Handle 0x%x already in use!\n",
 749					  handle);
 750				return -EINVAL;
 751			}
 752
 753			if (!atomic_cmpxchg(&adev->uvd.handles[i], 0, handle)) {
 754				adev->uvd.filp[i] = ctx->parser->filp;
 755				return 0;
 756			}
 757		}
 758
 759		DRM_ERROR("No more free UVD handles!\n");
 760		return -ENOSPC;
 761
 762	case 1:
 763		/* it's a decode msg, calc buffer sizes */
 764		r = amdgpu_uvd_cs_msg_decode(adev, msg, ctx->buf_sizes);
 765		amdgpu_bo_kunmap(bo);
 766		if (r)
 767			return r;
 768
 769		/* validate the handle */
 770		for (i = 0; i < adev->uvd.max_handles; ++i) {
 771			if (atomic_read(&adev->uvd.handles[i]) == handle) {
 772				if (adev->uvd.filp[i] != ctx->parser->filp) {
 773					DRM_ERROR("UVD handle collision detected!\n");
 774					return -EINVAL;
 775				}
 776				return 0;
 777			}
 778		}
 779
 780		DRM_ERROR("Invalid UVD handle 0x%x!\n", handle);
 781		return -ENOENT;
 782
 783	case 2:
 784		/* it's a destroy msg, free the handle */
 785		for (i = 0; i < adev->uvd.max_handles; ++i)
 786			atomic_cmpxchg(&adev->uvd.handles[i], handle, 0);
 787		amdgpu_bo_kunmap(bo);
 788		return 0;
 789
 790	default:
 791		DRM_ERROR("Illegal UVD message type (%d)!\n", msg_type);
 792		return -EINVAL;
 793	}
 794	BUG();
 
 795	return -EINVAL;
 796}
 797
 798/**
 799 * amdgpu_uvd_cs_pass2 - second parsing round
 800 *
 801 * @ctx: UVD parser context
 802 *
 803 * Patch buffer addresses, make sure buffer sizes are correct.
 804 */
 805static int amdgpu_uvd_cs_pass2(struct amdgpu_uvd_cs_ctx *ctx)
 806{
 807	struct amdgpu_bo_va_mapping *mapping;
 808	struct amdgpu_bo *bo;
 809	uint32_t cmd;
 810	uint64_t start, end;
 811	uint64_t addr = amdgpu_uvd_get_addr_from_ctx(ctx);
 812	int r;
 813
 814	r = amdgpu_cs_find_mapping(ctx->parser, addr, &bo, &mapping);
 815	if (r) {
 816		DRM_ERROR("Can't find BO for addr 0x%08Lx\n", addr);
 817		return r;
 818	}
 819
 820	start = amdgpu_bo_gpu_offset(bo);
 821
 822	end = (mapping->last + 1 - mapping->start);
 823	end = end * AMDGPU_GPU_PAGE_SIZE + start;
 824
 825	addr -= mapping->start * AMDGPU_GPU_PAGE_SIZE;
 826	start += addr;
 827
 828	amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data0,
 829			    lower_32_bits(start));
 830	amdgpu_set_ib_value(ctx->parser, ctx->ib_idx, ctx->data1,
 831			    upper_32_bits(start));
 832
 833	cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx) >> 1;
 834	if (cmd < 0x4) {
 835		if ((end - start) < ctx->buf_sizes[cmd]) {
 836			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
 837				  (unsigned)(end - start),
 838				  ctx->buf_sizes[cmd]);
 839			return -EINVAL;
 840		}
 841
 842	} else if (cmd == 0x206) {
 843		if ((end - start) < ctx->buf_sizes[4]) {
 844			DRM_ERROR("buffer (%d) to small (%d / %d)!\n", cmd,
 845					  (unsigned)(end - start),
 846					  ctx->buf_sizes[4]);
 847			return -EINVAL;
 848		}
 849	} else if ((cmd != 0x100) && (cmd != 0x204)) {
 850		DRM_ERROR("invalid UVD command %X!\n", cmd);
 851		return -EINVAL;
 852	}
 853
 854	if (!ctx->parser->adev->uvd.address_64_bit) {
 855		if ((start >> 28) != ((end - 1) >> 28)) {
 856			DRM_ERROR("reloc %LX-%LX crossing 256MB boundary!\n",
 857				  start, end);
 858			return -EINVAL;
 859		}
 860
 861		if ((cmd == 0 || cmd == 0x3) &&
 862		    (start >> 28) != (ctx->parser->adev->uvd.inst->gpu_addr >> 28)) {
 863			DRM_ERROR("msg/fb buffer %LX-%LX out of 256MB segment!\n",
 864				  start, end);
 865			return -EINVAL;
 866		}
 867	}
 868
 869	if (cmd == 0) {
 870		ctx->has_msg_cmd = true;
 871		r = amdgpu_uvd_cs_msg(ctx, bo, addr);
 872		if (r)
 873			return r;
 874	} else if (!ctx->has_msg_cmd) {
 875		DRM_ERROR("Message needed before other commands are send!\n");
 876		return -EINVAL;
 877	}
 878
 879	return 0;
 880}
 881
 882/**
 883 * amdgpu_uvd_cs_reg - parse register writes
 884 *
 885 * @ctx: UVD parser context
 886 * @cb: callback function
 887 *
 888 * Parse the register writes, call cb on each complete command.
 889 */
 890static int amdgpu_uvd_cs_reg(struct amdgpu_uvd_cs_ctx *ctx,
 891			     int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
 892{
 893	struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
 894	int i, r;
 895
 896	ctx->idx++;
 897	for (i = 0; i <= ctx->count; ++i) {
 898		unsigned reg = ctx->reg + i;
 899
 900		if (ctx->idx >= ib->length_dw) {
 901			DRM_ERROR("Register command after end of CS!\n");
 902			return -EINVAL;
 903		}
 904
 905		switch (reg) {
 906		case mmUVD_GPCOM_VCPU_DATA0:
 907			ctx->data0 = ctx->idx;
 908			break;
 909		case mmUVD_GPCOM_VCPU_DATA1:
 910			ctx->data1 = ctx->idx;
 911			break;
 912		case mmUVD_GPCOM_VCPU_CMD:
 913			r = cb(ctx);
 914			if (r)
 915				return r;
 916			break;
 917		case mmUVD_ENGINE_CNTL:
 918		case mmUVD_NO_OP:
 919			break;
 920		default:
 921			DRM_ERROR("Invalid reg 0x%X!\n", reg);
 922			return -EINVAL;
 923		}
 924		ctx->idx++;
 925	}
 926	return 0;
 927}
 928
 929/**
 930 * amdgpu_uvd_cs_packets - parse UVD packets
 931 *
 932 * @ctx: UVD parser context
 933 * @cb: callback function
 934 *
 935 * Parse the command stream packets.
 936 */
 937static int amdgpu_uvd_cs_packets(struct amdgpu_uvd_cs_ctx *ctx,
 938				 int (*cb)(struct amdgpu_uvd_cs_ctx *ctx))
 939{
 940	struct amdgpu_ib *ib = &ctx->parser->job->ibs[ctx->ib_idx];
 941	int r;
 942
 943	for (ctx->idx = 0 ; ctx->idx < ib->length_dw; ) {
 944		uint32_t cmd = amdgpu_get_ib_value(ctx->parser, ctx->ib_idx, ctx->idx);
 945		unsigned type = CP_PACKET_GET_TYPE(cmd);
 946		switch (type) {
 947		case PACKET_TYPE0:
 948			ctx->reg = CP_PACKET0_GET_REG(cmd);
 949			ctx->count = CP_PACKET_GET_COUNT(cmd);
 950			r = amdgpu_uvd_cs_reg(ctx, cb);
 951			if (r)
 952				return r;
 953			break;
 954		case PACKET_TYPE2:
 955			++ctx->idx;
 956			break;
 957		default:
 958			DRM_ERROR("Unknown packet type %d !\n", type);
 959			return -EINVAL;
 960		}
 961	}
 962	return 0;
 963}
 964
 965/**
 966 * amdgpu_uvd_ring_parse_cs - UVD command submission parser
 967 *
 968 * @parser: Command submission parser context
 
 
 969 *
 970 * Parse the command stream, patch in addresses as necessary.
 971 */
 972int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser, uint32_t ib_idx)
 
 
 973{
 974	struct amdgpu_uvd_cs_ctx ctx = {};
 975	unsigned buf_sizes[] = {
 976		[0x00000000]	=	2048,
 977		[0x00000001]	=	0xFFFFFFFF,
 978		[0x00000002]	=	0xFFFFFFFF,
 979		[0x00000003]	=	2048,
 980		[0x00000004]	=	0xFFFFFFFF,
 981	};
 982	struct amdgpu_ib *ib = &parser->job->ibs[ib_idx];
 983	int r;
 984
 985	parser->job->vm = NULL;
 986	ib->gpu_addr = amdgpu_sa_bo_gpu_addr(ib->sa_bo);
 987
 988	if (ib->length_dw % 16) {
 989		DRM_ERROR("UVD IB length (%d) not 16 dwords aligned!\n",
 990			  ib->length_dw);
 991		return -EINVAL;
 992	}
 993
 994	ctx.parser = parser;
 995	ctx.buf_sizes = buf_sizes;
 996	ctx.ib_idx = ib_idx;
 997
 998	/* first round only required on chips without UVD 64 bit address support */
 999	if (!parser->adev->uvd.address_64_bit) {
1000		/* first round, make sure the buffers are actually in the UVD segment */
1001		r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass1);
1002		if (r)
1003			return r;
1004	}
1005
1006	/* second round, patch buffer addresses into the command stream */
1007	r = amdgpu_uvd_cs_packets(&ctx, amdgpu_uvd_cs_pass2);
1008	if (r)
1009		return r;
1010
1011	if (!ctx.has_msg_cmd) {
1012		DRM_ERROR("UVD-IBs need a msg command!\n");
1013		return -EINVAL;
1014	}
1015
1016	return 0;
1017}
1018
1019static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
1020			       bool direct, struct dma_fence **fence)
1021{
1022	struct amdgpu_device *adev = ring->adev;
1023	struct dma_fence *f = NULL;
1024	struct amdgpu_job *job;
1025	struct amdgpu_ib *ib;
1026	uint32_t data[4];
1027	uint64_t addr;
1028	long r;
1029	int i;
1030	unsigned offset_idx = 0;
1031	unsigned offset[3] = { UVD_BASE_SI, 0, 0 };
1032
1033	amdgpu_bo_kunmap(bo);
1034	amdgpu_bo_unpin(bo);
1035
1036	if (!ring->adev->uvd.address_64_bit) {
1037		struct ttm_operation_ctx ctx = { true, false };
1038
1039		amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
1040		amdgpu_uvd_force_into_uvd_segment(bo);
1041		r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
1042		if (r)
1043			goto err;
1044	}
1045
1046	r = amdgpu_job_alloc_with_ib(adev, 64, &job);
1047	if (r)
1048		goto err;
1049
1050	if (adev->asic_type >= CHIP_VEGA10) {
1051		offset_idx = 1 + ring->me;
1052		offset[1] = adev->reg_offset[UVD_HWIP][0][1];
1053		offset[2] = adev->reg_offset[UVD_HWIP][1][1];
1054	}
1055
1056	data[0] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_DATA0, 0);
1057	data[1] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_DATA1, 0);
1058	data[2] = PACKET0(offset[offset_idx] + UVD_GPCOM_VCPU_CMD, 0);
1059	data[3] = PACKET0(offset[offset_idx] + UVD_NO_OP, 0);
1060
1061	ib = &job->ibs[0];
1062	addr = amdgpu_bo_gpu_offset(bo);
1063	ib->ptr[0] = data[0];
1064	ib->ptr[1] = addr;
1065	ib->ptr[2] = data[1];
1066	ib->ptr[3] = addr >> 32;
1067	ib->ptr[4] = data[2];
1068	ib->ptr[5] = 0;
1069	for (i = 6; i < 16; i += 2) {
1070		ib->ptr[i] = data[3];
1071		ib->ptr[i+1] = 0;
1072	}
1073	ib->length_dw = 16;
1074
1075	if (direct) {
1076		r = dma_resv_wait_timeout_rcu(bo->tbo.base.resv,
1077							true, false,
1078							msecs_to_jiffies(10));
1079		if (r == 0)
1080			r = -ETIMEDOUT;
1081		if (r < 0)
1082			goto err_free;
1083
1084		r = amdgpu_job_submit_direct(job, ring, &f);
1085		if (r)
1086			goto err_free;
1087	} else {
1088		r = amdgpu_sync_resv(adev, &job->sync, bo->tbo.base.resv,
1089				     AMDGPU_FENCE_OWNER_UNDEFINED, false);
 
1090		if (r)
1091			goto err_free;
1092
1093		r = amdgpu_job_submit(job, &adev->uvd.entity,
1094				      AMDGPU_FENCE_OWNER_UNDEFINED, &f);
1095		if (r)
1096			goto err_free;
1097	}
1098
 
1099	amdgpu_bo_fence(bo, f, false);
1100	amdgpu_bo_unreserve(bo);
1101	amdgpu_bo_unref(&bo);
1102
1103	if (fence)
1104		*fence = dma_fence_get(f);
1105	dma_fence_put(f);
1106
1107	return 0;
1108
1109err_free:
1110	amdgpu_job_free(job);
1111
1112err:
1113	amdgpu_bo_unreserve(bo);
1114	amdgpu_bo_unref(&bo);
1115	return r;
1116}
1117
1118/* multiple fence commands without any stream commands in between can
1119   crash the vcpu so just try to emmit a dummy create/destroy msg to
1120   avoid this */
1121int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
1122			      struct dma_fence **fence)
1123{
1124	struct amdgpu_device *adev = ring->adev;
1125	struct amdgpu_bo *bo = NULL;
1126	uint32_t *msg;
1127	int r, i;
1128
1129	r = amdgpu_bo_create_reserved(adev, 1024, PAGE_SIZE,
1130				      AMDGPU_GEM_DOMAIN_VRAM,
1131				      &bo, NULL, (void **)&msg);
1132	if (r)
1133		return r;
1134
 
1135	/* stitch together an UVD create msg */
1136	msg[0] = cpu_to_le32(0x00000de4);
1137	msg[1] = cpu_to_le32(0x00000000);
1138	msg[2] = cpu_to_le32(handle);
1139	msg[3] = cpu_to_le32(0x00000000);
1140	msg[4] = cpu_to_le32(0x00000000);
1141	msg[5] = cpu_to_le32(0x00000000);
1142	msg[6] = cpu_to_le32(0x00000000);
1143	msg[7] = cpu_to_le32(0x00000780);
1144	msg[8] = cpu_to_le32(0x00000440);
1145	msg[9] = cpu_to_le32(0x00000000);
1146	msg[10] = cpu_to_le32(0x01b37000);
1147	for (i = 11; i < 1024; ++i)
1148		msg[i] = cpu_to_le32(0x0);
1149
1150	return amdgpu_uvd_send_msg(ring, bo, true, fence);
 
1151}
1152
1153int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
1154			       bool direct, struct dma_fence **fence)
1155{
1156	struct amdgpu_device *adev = ring->adev;
1157	struct amdgpu_bo *bo = NULL;
1158	uint32_t *msg;
1159	int r, i;
1160
1161	r = amdgpu_bo_create_reserved(adev, 1024, PAGE_SIZE,
1162				      AMDGPU_GEM_DOMAIN_VRAM,
1163				      &bo, NULL, (void **)&msg);
1164	if (r)
1165		return r;
 
 
1166
 
1167	/* stitch together an UVD destroy msg */
1168	msg[0] = cpu_to_le32(0x00000de4);
1169	msg[1] = cpu_to_le32(0x00000002);
1170	msg[2] = cpu_to_le32(handle);
1171	msg[3] = cpu_to_le32(0x00000000);
1172	for (i = 4; i < 1024; ++i)
1173		msg[i] = cpu_to_le32(0x0);
1174
1175	return amdgpu_uvd_send_msg(ring, bo, direct, fence);
 
 
 
 
 
1176}
1177
1178static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
1179{
1180	struct amdgpu_device *adev =
1181		container_of(work, struct amdgpu_device, uvd.idle_work.work);
1182	unsigned fences = 0, i, j;
1183
1184	for (i = 0; i < adev->uvd.num_uvd_inst; ++i) {
1185		if (adev->uvd.harvest_config & (1 << i))
1186			continue;
1187		fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring);
1188		for (j = 0; j < adev->uvd.num_enc_rings; ++j) {
1189			fences += amdgpu_fence_count_emitted(&adev->uvd.inst[i].ring_enc[j]);
1190		}
1191	}
1192
1193	if (fences == 0) {
1194		if (adev->pm.dpm_enabled) {
1195			amdgpu_dpm_enable_uvd(adev, false);
1196		} else {
1197			amdgpu_asic_set_uvd_clocks(adev, 0, 0);
1198			/* shutdown the UVD block */
1199			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1200							       AMD_PG_STATE_GATE);
1201			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1202							       AMD_CG_STATE_GATE);
1203		}
1204	} else {
1205		schedule_delayed_work(&adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1206	}
1207}
1208
1209void amdgpu_uvd_ring_begin_use(struct amdgpu_ring *ring)
1210{
1211	struct amdgpu_device *adev = ring->adev;
1212	bool set_clocks;
1213
1214	if (amdgpu_sriov_vf(adev))
1215		return;
1216
1217	set_clocks = !cancel_delayed_work_sync(&adev->uvd.idle_work);
1218	if (set_clocks) {
1219		if (adev->pm.dpm_enabled) {
1220			amdgpu_dpm_enable_uvd(adev, true);
1221		} else {
1222			amdgpu_asic_set_uvd_clocks(adev, 53300, 40000);
1223			amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1224							       AMD_CG_STATE_UNGATE);
1225			amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_UVD,
1226							       AMD_PG_STATE_UNGATE);
1227		}
1228	}
1229}
1230
1231void amdgpu_uvd_ring_end_use(struct amdgpu_ring *ring)
1232{
1233	if (!amdgpu_sriov_vf(ring->adev))
1234		schedule_delayed_work(&ring->adev->uvd.idle_work, UVD_IDLE_TIMEOUT);
1235}
1236
1237/**
1238 * amdgpu_uvd_ring_test_ib - test ib execution
1239 *
1240 * @ring: amdgpu_ring pointer
 
1241 *
1242 * Test if we can successfully execute an IB
1243 */
1244int amdgpu_uvd_ring_test_ib(struct amdgpu_ring *ring, long timeout)
1245{
1246	struct dma_fence *fence;
1247	long r;
1248
1249	r = amdgpu_uvd_get_create_msg(ring, 1, NULL);
1250	if (r)
 
 
 
 
 
 
 
1251		goto error;
1252
1253	r = amdgpu_uvd_get_destroy_msg(ring, 1, true, &fence);
1254	if (r)
1255		goto error;
1256
1257	r = dma_fence_wait_timeout(fence, false, timeout);
1258	if (r == 0)
1259		r = -ETIMEDOUT;
1260	else if (r > 0)
1261		r = 0;
1262
1263	dma_fence_put(fence);
1264
1265error:
1266	return r;
1267}
1268
1269/**
1270 * amdgpu_uvd_used_handles - returns used UVD handles
1271 *
1272 * @adev: amdgpu_device pointer
1273 *
1274 * Returns the number of UVD handles in use
1275 */
1276uint32_t amdgpu_uvd_used_handles(struct amdgpu_device *adev)
1277{
1278	unsigned i;
1279	uint32_t used_handles = 0;
1280
1281	for (i = 0; i < adev->uvd.max_handles; ++i) {
1282		/*
1283		 * Handles can be freed in any order, and not
1284		 * necessarily linear. So we need to count
1285		 * all non-zero handles.
1286		 */
1287		if (atomic_read(&adev->uvd.handles[i]))
1288			used_handles++;
1289	}
1290
1291	return used_handles;
1292}