Linux Audio

Check our new training course

Loading...
v4.17
   1/*
   2 * Copyright 2008 Advanced Micro Devices, Inc.
   3 * Copyright 2008 Red Hat Inc.
   4 * Copyright 2009 Jerome Glisse.
   5 *
   6 * Permission is hereby granted, free of charge, to any person obtaining a
   7 * copy of this software and associated documentation files (the "Software"),
   8 * to deal in the Software without restriction, including without limitation
   9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10 * and/or sell copies of the Software, and to permit persons to whom the
  11 * Software is furnished to do so, subject to the following conditions:
  12 *
  13 * The above copyright notice and this permission notice shall be included in
  14 * all copies or substantial portions of the Software.
  15 *
  16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  22 * OTHER DEALINGS IN THE SOFTWARE.
  23 *
  24 * Authors: Dave Airlie
  25 *          Alex Deucher
  26 *          Jerome Glisse
  27 */
 
  28#include <linux/kthread.h>
 
  29#include <linux/console.h>
  30#include <linux/slab.h>
  31#include <drm/drmP.h>
  32#include <drm/drm_crtc_helper.h>
  33#include <drm/drm_atomic_helper.h>
 
  34#include <drm/amdgpu_drm.h>
  35#include <linux/vgaarb.h>
  36#include <linux/vga_switcheroo.h>
  37#include <linux/efi.h>
  38#include "amdgpu.h"
  39#include "amdgpu_trace.h"
  40#include "amdgpu_i2c.h"
  41#include "atom.h"
  42#include "amdgpu_atombios.h"
  43#include "amdgpu_atomfirmware.h"
  44#include "amd_pcie.h"
  45#ifdef CONFIG_DRM_AMDGPU_SI
  46#include "si.h"
  47#endif
  48#ifdef CONFIG_DRM_AMDGPU_CIK
  49#include "cik.h"
  50#endif
  51#include "vi.h"
  52#include "soc15.h"
 
  53#include "bif/bif_4_1_d.h"
  54#include <linux/pci.h>
  55#include <linux/firmware.h>
  56#include "amdgpu_vf_error.h"
  57
  58#include "amdgpu_amdkfd.h"
  59#include "amdgpu_pm.h"
  60
 
 
 
 
 
 
 
 
 
 
 
 
  61MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin");
  62MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin");
  63MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin");
 
 
 
 
 
 
 
 
 
  64
  65#define AMDGPU_RESUME_MS		2000
  66
  67static const char *amdgpu_asic_name[] = {
  68	"TAHITI",
  69	"PITCAIRN",
  70	"VERDE",
  71	"OLAND",
  72	"HAINAN",
  73	"BONAIRE",
  74	"KAVERI",
  75	"KABINI",
  76	"HAWAII",
  77	"MULLINS",
  78	"TOPAZ",
  79	"TONGA",
  80	"FIJI",
  81	"CARRIZO",
  82	"STONEY",
  83	"POLARIS10",
  84	"POLARIS11",
  85	"POLARIS12",
 
  86	"VEGA10",
  87	"VEGA12",
 
  88	"RAVEN",
 
 
 
 
 
 
 
 
 
 
 
 
  89	"LAST",
  90};
  91
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  92static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev);
  93
  94/**
  95 * amdgpu_device_is_px - Is the device is a dGPU with HG/PX power control
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  96 *
  97 * @dev: drm_device pointer
  98 *
  99 * Returns true if the device is a dGPU with HG/PX power control,
 100 * otherwise return false.
 101 */
 102bool amdgpu_device_is_px(struct drm_device *dev)
 103{
 104	struct amdgpu_device *adev = dev->dev_private;
 105
 106	if (adev->flags & AMD_IS_PX)
 
 107		return true;
 108	return false;
 109}
 110
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 111/*
 112 * MMIO register access helper functions.
 113 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 114/**
 115 * amdgpu_mm_rreg - read a memory mapped IO register
 116 *
 117 * @adev: amdgpu_device pointer
 118 * @reg: dword aligned register offset
 119 * @acc_flags: access flags which require special behavior
 120 *
 121 * Returns the 32 bit value from the offset specified.
 122 */
 123uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg,
 124			uint32_t acc_flags)
 125{
 126	uint32_t ret;
 127
 128	if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
 129		return amdgpu_virt_kiq_rreg(adev, reg);
 130
 131	if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
 132		ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
 133	else {
 134		unsigned long flags;
 135
 136		spin_lock_irqsave(&adev->mmio_idx_lock, flags);
 137		writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
 138		ret = readl(((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
 139		spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
 
 
 140	}
 141	trace_amdgpu_mm_rreg(adev->pdev->device, reg, ret);
 
 
 142	return ret;
 143}
 144
 145/*
 146 * MMIO register read with bytes helper functions
 147 * @offset:bytes offset from MMIO start
 148 *
 149*/
 150
 151/**
 152 * amdgpu_mm_rreg8 - read a memory mapped IO register
 153 *
 154 * @adev: amdgpu_device pointer
 155 * @offset: byte aligned register offset
 156 *
 157 * Returns the 8 bit value from the offset specified.
 158 */
 159uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset) {
 
 
 
 
 160	if (offset < adev->rmmio_size)
 161		return (readb(adev->rmmio + offset));
 162	BUG();
 163}
 164
 165/*
 166 * MMIO register write with bytes helper functions
 167 * @offset:bytes offset from MMIO start
 168 * @value: the value want to be written to the register
 169 *
 170*/
 171/**
 172 * amdgpu_mm_wreg8 - read a memory mapped IO register
 173 *
 174 * @adev: amdgpu_device pointer
 175 * @offset: byte aligned register offset
 176 * @value: 8 bit value to write
 177 *
 178 * Writes the value specified to the offset specified.
 179 */
 180void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value) {
 
 
 
 
 181	if (offset < adev->rmmio_size)
 182		writeb(value, adev->rmmio + offset);
 183	else
 184		BUG();
 185}
 186
 187/**
 188 * amdgpu_mm_wreg - write to a memory mapped IO register
 189 *
 190 * @adev: amdgpu_device pointer
 191 * @reg: dword aligned register offset
 192 * @v: 32 bit value to write to the register
 193 * @acc_flags: access flags which require special behavior
 194 *
 195 * Writes the value specified to the offset specified.
 196 */
 197void amdgpu_mm_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v,
 198		    uint32_t acc_flags)
 
 199{
 200	trace_amdgpu_mm_wreg(adev->pdev->device, reg, v);
 201
 202	if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
 203		adev->last_mm_index = v;
 204	}
 205
 206	if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
 207		return amdgpu_virt_kiq_wreg(adev, reg, v);
 208
 209	if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
 210		writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
 211	else {
 212		unsigned long flags;
 213
 214		spin_lock_irqsave(&adev->mmio_idx_lock, flags);
 215		writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
 216		writel(v, ((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
 217		spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
 
 
 
 
 
 
 
 218	}
 219
 220	if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
 221		udelay(500);
 222	}
 223}
 224
 225/**
 226 * amdgpu_io_rreg - read an IO register
 227 *
 228 * @adev: amdgpu_device pointer
 229 * @reg: dword aligned register offset
 230 *
 231 * Returns the 32 bit value from the offset specified.
 232 */
 233u32 amdgpu_io_rreg(struct amdgpu_device *adev, u32 reg)
 234{
 235	if ((reg * 4) < adev->rio_mem_size)
 236		return ioread32(adev->rio_mem + (reg * 4));
 237	else {
 238		iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
 239		return ioread32(adev->rio_mem + (mmMM_DATA * 4));
 240	}
 241}
 242
 243/**
 244 * amdgpu_io_wreg - write to an IO register
 245 *
 246 * @adev: amdgpu_device pointer
 247 * @reg: dword aligned register offset
 248 * @v: 32 bit value to write to the register
 249 *
 250 * Writes the value specified to the offset specified.
 251 */
 252void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32 v)
 
 253{
 254	if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
 255		adev->last_mm_index = v;
 256	}
 257
 258	if ((reg * 4) < adev->rio_mem_size)
 259		iowrite32(v, adev->rio_mem + (reg * 4));
 260	else {
 261		iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
 262		iowrite32(v, adev->rio_mem + (mmMM_DATA * 4));
 263	}
 264
 265	if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
 266		udelay(500);
 
 
 
 
 
 267	}
 268}
 269
 270/**
 271 * amdgpu_mm_rdoorbell - read a doorbell dword
 272 *
 273 * @adev: amdgpu_device pointer
 274 * @index: doorbell index
 275 *
 276 * Returns the value in the doorbell aperture at the
 277 * requested doorbell index (CIK).
 278 */
 279u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
 280{
 
 
 
 281	if (index < adev->doorbell.num_doorbells) {
 282		return readl(adev->doorbell.ptr + index);
 283	} else {
 284		DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
 285		return 0;
 286	}
 287}
 288
 289/**
 290 * amdgpu_mm_wdoorbell - write a doorbell dword
 291 *
 292 * @adev: amdgpu_device pointer
 293 * @index: doorbell index
 294 * @v: value to write
 295 *
 296 * Writes @v to the doorbell aperture at the
 297 * requested doorbell index (CIK).
 298 */
 299void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v)
 300{
 
 
 
 301	if (index < adev->doorbell.num_doorbells) {
 302		writel(v, adev->doorbell.ptr + index);
 303	} else {
 304		DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
 305	}
 306}
 307
 308/**
 309 * amdgpu_mm_rdoorbell64 - read a doorbell Qword
 310 *
 311 * @adev: amdgpu_device pointer
 312 * @index: doorbell index
 313 *
 314 * Returns the value in the doorbell aperture at the
 315 * requested doorbell index (VEGA10+).
 316 */
 317u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index)
 318{
 
 
 
 319	if (index < adev->doorbell.num_doorbells) {
 320		return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index));
 321	} else {
 322		DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
 323		return 0;
 324	}
 325}
 326
 327/**
 328 * amdgpu_mm_wdoorbell64 - write a doorbell Qword
 329 *
 330 * @adev: amdgpu_device pointer
 331 * @index: doorbell index
 332 * @v: value to write
 333 *
 334 * Writes @v to the doorbell aperture at the
 335 * requested doorbell index (VEGA10+).
 336 */
 337void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v)
 338{
 
 
 
 339	if (index < adev->doorbell.num_doorbells) {
 340		atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v);
 341	} else {
 342		DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
 343	}
 344}
 345
 346/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 347 * amdgpu_invalid_rreg - dummy reg read function
 348 *
 349 * @adev: amdgpu device pointer
 350 * @reg: offset of register
 351 *
 352 * Dummy register read function.  Used for register blocks
 353 * that certain asics don't have (all asics).
 354 * Returns the value in the register.
 355 */
 356static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
 357{
 358	DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
 359	BUG();
 360	return 0;
 361}
 362
 363/**
 364 * amdgpu_invalid_wreg - dummy reg write function
 365 *
 366 * @adev: amdgpu device pointer
 367 * @reg: offset of register
 368 * @v: value to write to the register
 369 *
 370 * Dummy register read function.  Used for register blocks
 371 * that certain asics don't have (all asics).
 372 */
 373static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
 374{
 375	DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
 376		  reg, v);
 377	BUG();
 378}
 379
 380/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 381 * amdgpu_block_invalid_rreg - dummy reg read function
 382 *
 383 * @adev: amdgpu device pointer
 384 * @block: offset of instance
 385 * @reg: offset of register
 386 *
 387 * Dummy register read function.  Used for register blocks
 388 * that certain asics don't have (all asics).
 389 * Returns the value in the register.
 390 */
 391static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
 392					  uint32_t block, uint32_t reg)
 393{
 394	DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
 395		  reg, block);
 396	BUG();
 397	return 0;
 398}
 399
 400/**
 401 * amdgpu_block_invalid_wreg - dummy reg write function
 402 *
 403 * @adev: amdgpu device pointer
 404 * @block: offset of instance
 405 * @reg: offset of register
 406 * @v: value to write to the register
 407 *
 408 * Dummy register read function.  Used for register blocks
 409 * that certain asics don't have (all asics).
 410 */
 411static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
 412				      uint32_t block,
 413				      uint32_t reg, uint32_t v)
 414{
 415	DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
 416		  reg, block, v);
 417	BUG();
 418}
 419
 420/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 421 * amdgpu_device_vram_scratch_init - allocate the VRAM scratch page
 422 *
 423 * @adev: amdgpu device pointer
 424 *
 425 * Allocates a scratch page of VRAM for use by various things in the
 426 * driver.
 427 */
 428static int amdgpu_device_vram_scratch_init(struct amdgpu_device *adev)
 429{
 430	return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE,
 431				       PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
 432				       &adev->vram_scratch.robj,
 433				       &adev->vram_scratch.gpu_addr,
 434				       (void **)&adev->vram_scratch.ptr);
 435}
 436
 437/**
 438 * amdgpu_device_vram_scratch_fini - Free the VRAM scratch page
 439 *
 440 * @adev: amdgpu device pointer
 441 *
 442 * Frees the VRAM scratch page.
 443 */
 444static void amdgpu_device_vram_scratch_fini(struct amdgpu_device *adev)
 445{
 446	amdgpu_bo_free_kernel(&adev->vram_scratch.robj, NULL, NULL);
 447}
 448
 449/**
 450 * amdgpu_device_program_register_sequence - program an array of registers.
 451 *
 452 * @adev: amdgpu_device pointer
 453 * @registers: pointer to the register array
 454 * @array_size: size of the register array
 455 *
 456 * Programs an array or registers with and and or masks.
 457 * This is a helper for setting golden registers.
 458 */
 459void amdgpu_device_program_register_sequence(struct amdgpu_device *adev,
 460					     const u32 *registers,
 461					     const u32 array_size)
 462{
 463	u32 tmp, reg, and_mask, or_mask;
 464	int i;
 465
 466	if (array_size % 3)
 467		return;
 468
 469	for (i = 0; i < array_size; i +=3) {
 470		reg = registers[i + 0];
 471		and_mask = registers[i + 1];
 472		or_mask = registers[i + 2];
 473
 474		if (and_mask == 0xffffffff) {
 475			tmp = or_mask;
 476		} else {
 477			tmp = RREG32(reg);
 478			tmp &= ~and_mask;
 479			tmp |= or_mask;
 
 
 
 480		}
 481		WREG32(reg, tmp);
 482	}
 483}
 484
 485/**
 486 * amdgpu_device_pci_config_reset - reset the GPU
 487 *
 488 * @adev: amdgpu_device pointer
 489 *
 490 * Resets the GPU using the pci config reset sequence.
 491 * Only applicable to asics prior to vega10.
 492 */
 493void amdgpu_device_pci_config_reset(struct amdgpu_device *adev)
 494{
 495	pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
 496}
 497
 
 
 
 
 
 
 
 
 
 
 
 
 498/*
 499 * GPU doorbell aperture helpers function.
 500 */
 501/**
 502 * amdgpu_device_doorbell_init - Init doorbell driver information.
 503 *
 504 * @adev: amdgpu_device pointer
 505 *
 506 * Init doorbell driver information (CIK)
 507 * Returns 0 on success, error on failure.
 508 */
 509static int amdgpu_device_doorbell_init(struct amdgpu_device *adev)
 510{
 
 511	/* No doorbell on SI hardware generation */
 512	if (adev->asic_type < CHIP_BONAIRE) {
 513		adev->doorbell.base = 0;
 514		adev->doorbell.size = 0;
 515		adev->doorbell.num_doorbells = 0;
 516		adev->doorbell.ptr = NULL;
 517		return 0;
 518	}
 519
 520	if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET)
 521		return -EINVAL;
 522
 
 
 523	/* doorbell bar mapping */
 524	adev->doorbell.base = pci_resource_start(adev->pdev, 2);
 525	adev->doorbell.size = pci_resource_len(adev->pdev, 2);
 526
 527	adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32),
 528					     AMDGPU_DOORBELL_MAX_ASSIGNMENT+1);
 529	if (adev->doorbell.num_doorbells == 0)
 530		return -EINVAL;
 531
 
 
 
 
 
 
 
 
 
 532	adev->doorbell.ptr = ioremap(adev->doorbell.base,
 533				     adev->doorbell.num_doorbells *
 534				     sizeof(u32));
 535	if (adev->doorbell.ptr == NULL)
 536		return -ENOMEM;
 537
 538	return 0;
 539}
 540
 541/**
 542 * amdgpu_device_doorbell_fini - Tear down doorbell driver information.
 543 *
 544 * @adev: amdgpu_device pointer
 545 *
 546 * Tear down doorbell driver information (CIK)
 547 */
 548static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev)
 549{
 550	iounmap(adev->doorbell.ptr);
 551	adev->doorbell.ptr = NULL;
 552}
 553
 554
 555
 556/*
 557 * amdgpu_device_wb_*()
 558 * Writeback is the method by which the GPU updates special pages in memory
 559 * with the status of certain GPU events (fences, ring pointers,etc.).
 560 */
 561
 562/**
 563 * amdgpu_device_wb_fini - Disable Writeback and free memory
 564 *
 565 * @adev: amdgpu_device pointer
 566 *
 567 * Disables Writeback and frees the Writeback memory (all asics).
 568 * Used at driver shutdown.
 569 */
 570static void amdgpu_device_wb_fini(struct amdgpu_device *adev)
 571{
 572	if (adev->wb.wb_obj) {
 573		amdgpu_bo_free_kernel(&adev->wb.wb_obj,
 574				      &adev->wb.gpu_addr,
 575				      (void **)&adev->wb.wb);
 576		adev->wb.wb_obj = NULL;
 577	}
 578}
 579
 580/**
 581 * amdgpu_device_wb_init- Init Writeback driver info and allocate memory
 582 *
 583 * @adev: amdgpu_device pointer
 584 *
 585 * Initializes writeback and allocates writeback memory (all asics).
 586 * Used at driver startup.
 587 * Returns 0 on success or an -error on failure.
 588 */
 589static int amdgpu_device_wb_init(struct amdgpu_device *adev)
 590{
 591	int r;
 592
 593	if (adev->wb.wb_obj == NULL) {
 594		/* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */
 595		r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8,
 596					    PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
 597					    &adev->wb.wb_obj, &adev->wb.gpu_addr,
 598					    (void **)&adev->wb.wb);
 599		if (r) {
 600			dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
 601			return r;
 602		}
 603
 604		adev->wb.num_wb = AMDGPU_MAX_WB;
 605		memset(&adev->wb.used, 0, sizeof(adev->wb.used));
 606
 607		/* clear wb memory */
 608		memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8);
 609	}
 610
 611	return 0;
 612}
 613
 614/**
 615 * amdgpu_device_wb_get - Allocate a wb entry
 616 *
 617 * @adev: amdgpu_device pointer
 618 * @wb: wb index
 619 *
 620 * Allocate a wb slot for use by the driver (all asics).
 621 * Returns 0 on success or -EINVAL on failure.
 622 */
 623int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb)
 624{
 625	unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
 626
 627	if (offset < adev->wb.num_wb) {
 628		__set_bit(offset, adev->wb.used);
 629		*wb = offset << 3; /* convert to dw offset */
 630		return 0;
 631	} else {
 632		return -EINVAL;
 633	}
 634}
 635
 636/**
 637 * amdgpu_device_wb_free - Free a wb entry
 638 *
 639 * @adev: amdgpu_device pointer
 640 * @wb: wb index
 641 *
 642 * Free a wb slot allocated for use by the driver (all asics)
 643 */
 644void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb)
 645{
 646	wb >>= 3;
 647	if (wb < adev->wb.num_wb)
 648		__clear_bit(wb, adev->wb.used);
 649}
 650
 651/**
 652 * amdgpu_device_vram_location - try to find VRAM location
 653 *
 654 * @adev: amdgpu device structure holding all necessary informations
 655 * @mc: memory controller structure holding memory informations
 656 * @base: base address at which to put VRAM
 657 *
 658 * Function will try to place VRAM at base address provided
 659 * as parameter.
 660 */
 661void amdgpu_device_vram_location(struct amdgpu_device *adev,
 662				 struct amdgpu_gmc *mc, u64 base)
 663{
 664	uint64_t limit = (uint64_t)amdgpu_vram_limit << 20;
 665
 666	mc->vram_start = base;
 667	mc->vram_end = mc->vram_start + mc->mc_vram_size - 1;
 668	if (limit && limit < mc->real_vram_size)
 669		mc->real_vram_size = limit;
 670	dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
 671			mc->mc_vram_size >> 20, mc->vram_start,
 672			mc->vram_end, mc->real_vram_size >> 20);
 673}
 674
 675/**
 676 * amdgpu_device_gart_location - try to find GTT location
 677 *
 678 * @adev: amdgpu device structure holding all necessary informations
 679 * @mc: memory controller structure holding memory informations
 680 *
 681 * Function will place try to place GTT before or after VRAM.
 682 *
 683 * If GTT size is bigger than space left then we ajust GTT size.
 684 * Thus function will never fails.
 685 *
 686 * FIXME: when reducing GTT size align new size on power of 2.
 687 */
 688void amdgpu_device_gart_location(struct amdgpu_device *adev,
 689				 struct amdgpu_gmc *mc)
 690{
 691	u64 size_af, size_bf;
 692
 693	size_af = adev->gmc.mc_mask - mc->vram_end;
 694	size_bf = mc->vram_start;
 695	if (size_bf > size_af) {
 696		if (mc->gart_size > size_bf) {
 697			dev_warn(adev->dev, "limiting GTT\n");
 698			mc->gart_size = size_bf;
 699		}
 700		mc->gart_start = 0;
 701	} else {
 702		if (mc->gart_size > size_af) {
 703			dev_warn(adev->dev, "limiting GTT\n");
 704			mc->gart_size = size_af;
 705		}
 706		/* VCE doesn't like it when BOs cross a 4GB segment, so align
 707		 * the GART base on a 4GB boundary as well.
 708		 */
 709		mc->gart_start = ALIGN(mc->vram_end + 1, 0x100000000ULL);
 710	}
 711	mc->gart_end = mc->gart_start + mc->gart_size - 1;
 712	dev_info(adev->dev, "GTT: %lluM 0x%016llX - 0x%016llX\n",
 713			mc->gart_size >> 20, mc->gart_start, mc->gart_end);
 714}
 715
 716/**
 717 * amdgpu_device_resize_fb_bar - try to resize FB BAR
 718 *
 719 * @adev: amdgpu_device pointer
 720 *
 721 * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not
 722 * to fail, but if any of the BARs is not accessible after the size we abort
 723 * driver loading by returning -ENODEV.
 724 */
 725int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev)
 726{
 727	u64 space_needed = roundup_pow_of_two(adev->gmc.real_vram_size);
 728	u32 rbar_size = order_base_2(((space_needed >> 20) | 1)) - 1;
 729	struct pci_bus *root;
 730	struct resource *res;
 731	unsigned i;
 732	u16 cmd;
 733	int r;
 734
 735	/* Bypass for VF */
 736	if (amdgpu_sriov_vf(adev))
 737		return 0;
 738
 
 
 
 
 
 739	/* Check if the root BUS has 64bit memory resources */
 740	root = adev->pdev->bus;
 741	while (root->parent)
 742		root = root->parent;
 743
 744	pci_bus_for_each_resource(root, res, i) {
 745		if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) &&
 746		    res->start > 0x100000000ull)
 747			break;
 748	}
 749
 750	/* Trying to resize is pointless without a root hub window above 4GB */
 751	if (!res)
 752		return 0;
 753
 
 
 
 
 754	/* Disable memory decoding while we change the BAR addresses and size */
 755	pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd);
 756	pci_write_config_word(adev->pdev, PCI_COMMAND,
 757			      cmd & ~PCI_COMMAND_MEMORY);
 758
 759	/* Free the VRAM and doorbell BAR, we most likely need to move both. */
 760	amdgpu_device_doorbell_fini(adev);
 761	if (adev->asic_type >= CHIP_BONAIRE)
 762		pci_release_resource(adev->pdev, 2);
 763
 764	pci_release_resource(adev->pdev, 0);
 765
 766	r = pci_resize_resource(adev->pdev, 0, rbar_size);
 767	if (r == -ENOSPC)
 768		DRM_INFO("Not enough PCI address space for a large BAR.");
 769	else if (r && r != -ENOTSUPP)
 770		DRM_ERROR("Problem resizing BAR0 (%d).", r);
 771
 772	pci_assign_unassigned_bus_resources(adev->pdev->bus);
 773
 774	/* When the doorbell or fb BAR isn't available we have no chance of
 775	 * using the device.
 776	 */
 777	r = amdgpu_device_doorbell_init(adev);
 778	if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET))
 779		return -ENODEV;
 780
 781	pci_write_config_word(adev->pdev, PCI_COMMAND, cmd);
 782
 783	return 0;
 784}
 785
 786/*
 787 * GPU helpers function.
 788 */
 789/**
 790 * amdgpu_device_need_post - check if the hw need post or not
 791 *
 792 * @adev: amdgpu_device pointer
 793 *
 794 * Check if the asic has been initialized (all asics) at driver startup
 795 * or post is needed if  hw reset is performed.
 796 * Returns true if need or false if not.
 797 */
 798bool amdgpu_device_need_post(struct amdgpu_device *adev)
 799{
 800	uint32_t reg;
 801
 802	if (amdgpu_sriov_vf(adev))
 803		return false;
 804
 805	if (amdgpu_passthrough(adev)) {
 806		/* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
 807		 * some old smc fw still need driver do vPost otherwise gpu hang, while
 808		 * those smc fw version above 22.15 doesn't have this flaw, so we force
 809		 * vpost executed for smc version below 22.15
 810		 */
 811		if (adev->asic_type == CHIP_FIJI) {
 812			int err;
 813			uint32_t fw_ver;
 814			err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
 815			/* force vPost if error occured */
 816			if (err)
 817				return true;
 818
 819			fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
 820			if (fw_ver < 0x00160e00)
 821				return true;
 822		}
 823	}
 824
 
 
 
 
 825	if (adev->has_hw_reset) {
 826		adev->has_hw_reset = false;
 827		return true;
 828	}
 829
 830	/* bios scratch used on CIK+ */
 831	if (adev->asic_type >= CHIP_BONAIRE)
 832		return amdgpu_atombios_scratch_need_asic_init(adev);
 833
 834	/* check MEM_SIZE for older asics */
 835	reg = amdgpu_asic_get_config_memsize(adev);
 836
 837	if ((reg != 0) && (reg != 0xffffffff))
 838		return false;
 839
 840	return true;
 841}
 842
 843/* if we get transitioned to only one device, take VGA back */
 844/**
 845 * amdgpu_device_vga_set_decode - enable/disable vga decode
 846 *
 847 * @cookie: amdgpu_device pointer
 848 * @state: enable/disable vga decode
 849 *
 850 * Enable/disable vga decode (all asics).
 851 * Returns VGA resource flags.
 852 */
 853static unsigned int amdgpu_device_vga_set_decode(void *cookie, bool state)
 854{
 855	struct amdgpu_device *adev = cookie;
 856	amdgpu_asic_set_vga_state(adev, state);
 857	if (state)
 858		return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
 859		       VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
 860	else
 861		return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
 862}
 863
 864/**
 865 * amdgpu_device_check_block_size - validate the vm block size
 866 *
 867 * @adev: amdgpu_device pointer
 868 *
 869 * Validates the vm block size specified via module parameter.
 870 * The vm block size defines number of bits in page table versus page directory,
 871 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
 872 * page table and the remaining bits are in the page directory.
 873 */
 874static void amdgpu_device_check_block_size(struct amdgpu_device *adev)
 875{
 876	/* defines number of bits in page table versus page directory,
 877	 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
 878	 * page table and the remaining bits are in the page directory */
 879	if (amdgpu_vm_block_size == -1)
 880		return;
 881
 882	if (amdgpu_vm_block_size < 9) {
 883		dev_warn(adev->dev, "VM page table size (%d) too small\n",
 884			 amdgpu_vm_block_size);
 885		amdgpu_vm_block_size = -1;
 886	}
 887}
 888
 889/**
 890 * amdgpu_device_check_vm_size - validate the vm size
 891 *
 892 * @adev: amdgpu_device pointer
 893 *
 894 * Validates the vm size in GB specified via module parameter.
 895 * The VM size is the size of the GPU virtual memory space in GB.
 896 */
 897static void amdgpu_device_check_vm_size(struct amdgpu_device *adev)
 898{
 899	/* no need to check the default value */
 900	if (amdgpu_vm_size == -1)
 901		return;
 902
 903	if (amdgpu_vm_size < 1) {
 904		dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
 905			 amdgpu_vm_size);
 906		amdgpu_vm_size = -1;
 907	}
 908}
 909
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 910/**
 911 * amdgpu_device_check_arguments - validate module params
 912 *
 913 * @adev: amdgpu_device pointer
 914 *
 915 * Validates certain module parameters and updates
 916 * the associated values used by the driver (all asics).
 917 */
 918static void amdgpu_device_check_arguments(struct amdgpu_device *adev)
 919{
 920	if (amdgpu_sched_jobs < 4) {
 921		dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
 922			 amdgpu_sched_jobs);
 923		amdgpu_sched_jobs = 4;
 924	} else if (!is_power_of_2(amdgpu_sched_jobs)){
 925		dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
 926			 amdgpu_sched_jobs);
 927		amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
 928	}
 929
 930	if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) {
 931		/* gart size must be greater or equal to 32M */
 932		dev_warn(adev->dev, "gart size (%d) too small\n",
 933			 amdgpu_gart_size);
 934		amdgpu_gart_size = -1;
 935	}
 936
 937	if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) {
 938		/* gtt size must be greater or equal to 32M */
 939		dev_warn(adev->dev, "gtt size (%d) too small\n",
 940				 amdgpu_gtt_size);
 941		amdgpu_gtt_size = -1;
 942	}
 943
 944	/* valid range is between 4 and 9 inclusive */
 945	if (amdgpu_vm_fragment_size != -1 &&
 946	    (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) {
 947		dev_warn(adev->dev, "valid range is between 4 and 9\n");
 948		amdgpu_vm_fragment_size = -1;
 949	}
 950
 
 
 
 
 
 
 
 
 
 
 
 
 951	amdgpu_device_check_vm_size(adev);
 952
 953	amdgpu_device_check_block_size(adev);
 954
 955	if (amdgpu_vram_page_split != -1 && (amdgpu_vram_page_split < 16 ||
 956	    !is_power_of_2(amdgpu_vram_page_split))) {
 957		dev_warn(adev->dev, "invalid VRAM page split (%d)\n",
 958			 amdgpu_vram_page_split);
 959		amdgpu_vram_page_split = 1024;
 960	}
 961
 962	if (amdgpu_lockup_timeout == 0) {
 963		dev_warn(adev->dev, "lockup_timeout msut be > 0, adjusting to 10000\n");
 964		amdgpu_lockup_timeout = 10000;
 965	}
 966
 967	adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type);
 
 
 968}
 969
 970/**
 971 * amdgpu_switcheroo_set_state - set switcheroo state
 972 *
 973 * @pdev: pci dev pointer
 974 * @state: vga_switcheroo state
 975 *
 976 * Callback for the switcheroo driver.  Suspends or resumes the
 977 * the asics before or after it is powered up using ACPI methods.
 978 */
 979static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, enum vga_switcheroo_state state)
 
 980{
 981	struct drm_device *dev = pci_get_drvdata(pdev);
 
 982
 983	if (amdgpu_device_is_px(dev) && state == VGA_SWITCHEROO_OFF)
 984		return;
 985
 986	if (state == VGA_SWITCHEROO_ON) {
 987		pr_info("amdgpu: switched on\n");
 988		/* don't suspend or resume card normally */
 989		dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
 990
 991		amdgpu_device_resume(dev, true, true);
 
 
 
 
 
 992
 993		dev->switch_power_state = DRM_SWITCH_POWER_ON;
 994		drm_kms_helper_poll_enable(dev);
 995	} else {
 996		pr_info("amdgpu: switched off\n");
 997		drm_kms_helper_poll_disable(dev);
 998		dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
 999		amdgpu_device_suspend(dev, true, true);
 
 
 
 
1000		dev->switch_power_state = DRM_SWITCH_POWER_OFF;
1001	}
1002}
1003
1004/**
1005 * amdgpu_switcheroo_can_switch - see if switcheroo state can change
1006 *
1007 * @pdev: pci dev pointer
1008 *
1009 * Callback for the switcheroo driver.  Check of the switcheroo
1010 * state can be changed.
1011 * Returns true if the state can be changed, false if not.
1012 */
1013static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
1014{
1015	struct drm_device *dev = pci_get_drvdata(pdev);
1016
1017	/*
1018	* FIXME: open_count is protected by drm_global_mutex but that would lead to
1019	* locking inversion with the driver load path. And the access here is
1020	* completely racy anyway. So don't bother with locking for now.
1021	*/
1022	return dev->open_count == 0;
1023}
1024
1025static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
1026	.set_gpu_state = amdgpu_switcheroo_set_state,
1027	.reprobe = NULL,
1028	.can_switch = amdgpu_switcheroo_can_switch,
1029};
1030
1031/**
1032 * amdgpu_device_ip_set_clockgating_state - set the CG state
1033 *
1034 * @adev: amdgpu_device pointer
1035 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1036 * @state: clockgating state (gate or ungate)
1037 *
1038 * Sets the requested clockgating state for all instances of
1039 * the hardware IP specified.
1040 * Returns the error code from the last instance.
1041 */
1042int amdgpu_device_ip_set_clockgating_state(struct amdgpu_device *adev,
1043					   enum amd_ip_block_type block_type,
1044					   enum amd_clockgating_state state)
1045{
 
1046	int i, r = 0;
1047
1048	for (i = 0; i < adev->num_ip_blocks; i++) {
1049		if (!adev->ip_blocks[i].status.valid)
1050			continue;
1051		if (adev->ip_blocks[i].version->type != block_type)
1052			continue;
1053		if (!adev->ip_blocks[i].version->funcs->set_clockgating_state)
1054			continue;
1055		r = adev->ip_blocks[i].version->funcs->set_clockgating_state(
1056			(void *)adev, state);
1057		if (r)
1058			DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",
1059				  adev->ip_blocks[i].version->funcs->name, r);
1060	}
1061	return r;
1062}
1063
1064/**
1065 * amdgpu_device_ip_set_powergating_state - set the PG state
1066 *
1067 * @adev: amdgpu_device pointer
1068 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1069 * @state: powergating state (gate or ungate)
1070 *
1071 * Sets the requested powergating state for all instances of
1072 * the hardware IP specified.
1073 * Returns the error code from the last instance.
1074 */
1075int amdgpu_device_ip_set_powergating_state(struct amdgpu_device *adev,
1076					   enum amd_ip_block_type block_type,
1077					   enum amd_powergating_state state)
1078{
 
1079	int i, r = 0;
1080
1081	for (i = 0; i < adev->num_ip_blocks; i++) {
1082		if (!adev->ip_blocks[i].status.valid)
1083			continue;
1084		if (adev->ip_blocks[i].version->type != block_type)
1085			continue;
1086		if (!adev->ip_blocks[i].version->funcs->set_powergating_state)
1087			continue;
1088		r = adev->ip_blocks[i].version->funcs->set_powergating_state(
1089			(void *)adev, state);
1090		if (r)
1091			DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",
1092				  adev->ip_blocks[i].version->funcs->name, r);
1093	}
1094	return r;
1095}
1096
1097/**
1098 * amdgpu_device_ip_get_clockgating_state - get the CG state
1099 *
1100 * @adev: amdgpu_device pointer
1101 * @flags: clockgating feature flags
1102 *
1103 * Walks the list of IPs on the device and updates the clockgating
1104 * flags for each IP.
1105 * Updates @flags with the feature flags for each hardware IP where
1106 * clockgating is enabled.
1107 */
1108void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev,
1109					    u32 *flags)
1110{
1111	int i;
1112
1113	for (i = 0; i < adev->num_ip_blocks; i++) {
1114		if (!adev->ip_blocks[i].status.valid)
1115			continue;
1116		if (adev->ip_blocks[i].version->funcs->get_clockgating_state)
1117			adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags);
1118	}
1119}
1120
1121/**
1122 * amdgpu_device_ip_wait_for_idle - wait for idle
1123 *
1124 * @adev: amdgpu_device pointer
1125 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1126 *
1127 * Waits for the request hardware IP to be idle.
1128 * Returns 0 for success or a negative error code on failure.
1129 */
1130int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev,
1131				   enum amd_ip_block_type block_type)
1132{
1133	int i, r;
1134
1135	for (i = 0; i < adev->num_ip_blocks; i++) {
1136		if (!adev->ip_blocks[i].status.valid)
1137			continue;
1138		if (adev->ip_blocks[i].version->type == block_type) {
1139			r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
1140			if (r)
1141				return r;
1142			break;
1143		}
1144	}
1145	return 0;
1146
1147}
1148
1149/**
1150 * amdgpu_device_ip_is_idle - is the hardware IP idle
1151 *
1152 * @adev: amdgpu_device pointer
1153 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1154 *
1155 * Check if the hardware IP is idle or not.
1156 * Returns true if it the IP is idle, false if not.
1157 */
1158bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev,
1159			      enum amd_ip_block_type block_type)
1160{
1161	int i;
1162
1163	for (i = 0; i < adev->num_ip_blocks; i++) {
1164		if (!adev->ip_blocks[i].status.valid)
1165			continue;
1166		if (adev->ip_blocks[i].version->type == block_type)
1167			return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
1168	}
1169	return true;
1170
1171}
1172
1173/**
1174 * amdgpu_device_ip_get_ip_block - get a hw IP pointer
1175 *
1176 * @adev: amdgpu_device pointer
1177 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1178 *
1179 * Returns a pointer to the hardware IP block structure
1180 * if it exists for the asic, otherwise NULL.
1181 */
1182struct amdgpu_ip_block *
1183amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,
1184			      enum amd_ip_block_type type)
1185{
1186	int i;
1187
1188	for (i = 0; i < adev->num_ip_blocks; i++)
1189		if (adev->ip_blocks[i].version->type == type)
1190			return &adev->ip_blocks[i];
1191
1192	return NULL;
1193}
1194
1195/**
1196 * amdgpu_device_ip_block_version_cmp
1197 *
1198 * @adev: amdgpu_device pointer
1199 * @type: enum amd_ip_block_type
1200 * @major: major version
1201 * @minor: minor version
1202 *
1203 * return 0 if equal or greater
1204 * return 1 if smaller or the ip_block doesn't exist
1205 */
1206int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev,
1207				       enum amd_ip_block_type type,
1208				       u32 major, u32 minor)
1209{
1210	struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type);
1211
1212	if (ip_block && ((ip_block->version->major > major) ||
1213			((ip_block->version->major == major) &&
1214			(ip_block->version->minor >= minor))))
1215		return 0;
1216
1217	return 1;
1218}
1219
1220/**
1221 * amdgpu_device_ip_block_add
1222 *
1223 * @adev: amdgpu_device pointer
1224 * @ip_block_version: pointer to the IP to add
1225 *
1226 * Adds the IP block driver information to the collection of IPs
1227 * on the asic.
1228 */
1229int amdgpu_device_ip_block_add(struct amdgpu_device *adev,
1230			       const struct amdgpu_ip_block_version *ip_block_version)
1231{
1232	if (!ip_block_version)
1233		return -EINVAL;
1234
 
 
 
 
 
 
 
 
 
 
 
 
 
1235	DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks,
1236		  ip_block_version->funcs->name);
1237
1238	adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
1239
1240	return 0;
1241}
1242
1243/**
1244 * amdgpu_device_enable_virtual_display - enable virtual display feature
1245 *
1246 * @adev: amdgpu_device pointer
1247 *
1248 * Enabled the virtual display feature if the user has enabled it via
1249 * the module parameter virtual_display.  This feature provides a virtual
1250 * display hardware on headless boards or in virtualized environments.
1251 * This function parses and validates the configuration string specified by
1252 * the user and configues the virtual display configuration (number of
1253 * virtual connectors, crtcs, etc.) specified.
1254 */
1255static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
1256{
1257	adev->enable_virtual_display = false;
1258
1259	if (amdgpu_virtual_display) {
1260		struct drm_device *ddev = adev->ddev;
1261		const char *pci_address_name = pci_name(ddev->pdev);
1262		char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
1263
1264		pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
1265		pciaddstr_tmp = pciaddstr;
1266		while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
1267			pciaddname = strsep(&pciaddname_tmp, ",");
1268			if (!strcmp("all", pciaddname)
1269			    || !strcmp(pci_address_name, pciaddname)) {
1270				long num_crtc;
1271				int res = -1;
1272
1273				adev->enable_virtual_display = true;
1274
1275				if (pciaddname_tmp)
1276					res = kstrtol(pciaddname_tmp, 10,
1277						      &num_crtc);
1278
1279				if (!res) {
1280					if (num_crtc < 1)
1281						num_crtc = 1;
1282					if (num_crtc > 6)
1283						num_crtc = 6;
1284					adev->mode_info.num_crtc = num_crtc;
1285				} else {
1286					adev->mode_info.num_crtc = 1;
1287				}
1288				break;
1289			}
1290		}
1291
1292		DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
1293			 amdgpu_virtual_display, pci_address_name,
1294			 adev->enable_virtual_display, adev->mode_info.num_crtc);
1295
1296		kfree(pciaddstr);
1297	}
1298}
1299
1300/**
1301 * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware
1302 *
1303 * @adev: amdgpu_device pointer
1304 *
1305 * Parses the asic configuration parameters specified in the gpu info
1306 * firmware and makes them availale to the driver for use in configuring
1307 * the asic.
1308 * Returns 0 on success, -EINVAL on failure.
1309 */
1310static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev)
1311{
1312	const char *chip_name;
1313	char fw_name[30];
1314	int err;
1315	const struct gpu_info_firmware_header_v1_0 *hdr;
1316
1317	adev->firmware.gpu_info_fw = NULL;
1318
 
 
 
 
 
 
 
 
 
 
 
 
1319	switch (adev->asic_type) {
1320	case CHIP_TOPAZ:
1321	case CHIP_TONGA:
1322	case CHIP_FIJI:
1323	case CHIP_POLARIS11:
1324	case CHIP_POLARIS10:
1325	case CHIP_POLARIS12:
1326	case CHIP_CARRIZO:
1327	case CHIP_STONEY:
1328#ifdef CONFIG_DRM_AMDGPU_SI
1329	case CHIP_VERDE:
1330	case CHIP_TAHITI:
1331	case CHIP_PITCAIRN:
1332	case CHIP_OLAND:
1333	case CHIP_HAINAN:
1334#endif
1335#ifdef CONFIG_DRM_AMDGPU_CIK
1336	case CHIP_BONAIRE:
1337	case CHIP_HAWAII:
1338	case CHIP_KAVERI:
1339	case CHIP_KABINI:
1340	case CHIP_MULLINS:
1341#endif
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1342	default:
1343		return 0;
1344	case CHIP_VEGA10:
1345		chip_name = "vega10";
1346		break;
1347	case CHIP_VEGA12:
1348		chip_name = "vega12";
1349		break;
1350	case CHIP_RAVEN:
1351		chip_name = "raven";
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1352		break;
1353	}
1354
1355	snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name);
1356	err = request_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev);
1357	if (err) {
1358		dev_err(adev->dev,
1359			"Failed to load gpu_info firmware \"%s\"\n",
1360			fw_name);
1361		goto out;
1362	}
1363	err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw);
1364	if (err) {
1365		dev_err(adev->dev,
1366			"Failed to validate gpu_info firmware \"%s\"\n",
1367			fw_name);
1368		goto out;
1369	}
1370
1371	hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data;
1372	amdgpu_ucode_print_gpu_info_hdr(&hdr->header);
1373
1374	switch (hdr->version_major) {
1375	case 1:
1376	{
1377		const struct gpu_info_firmware_v1_0 *gpu_info_fw =
1378			(const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data +
1379								le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1380
 
 
 
 
 
 
1381		adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se);
1382		adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh);
1383		adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se);
1384		adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se);
1385		adev->gfx.config.max_texture_channel_caches =
1386			le32_to_cpu(gpu_info_fw->gc_num_tccs);
1387		adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs);
1388		adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds);
1389		adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth);
1390		adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth);
1391		adev->gfx.config.double_offchip_lds_buf =
1392			le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer);
1393		adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size);
1394		adev->gfx.cu_info.max_waves_per_simd =
1395			le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd);
1396		adev->gfx.cu_info.max_scratch_slots_per_cu =
1397			le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu);
1398		adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1399		break;
1400	}
1401	default:
1402		dev_err(adev->dev,
1403			"Unsupported gpu_info table %d\n", hdr->header.ucode_version);
1404		err = -EINVAL;
1405		goto out;
1406	}
1407out:
1408	return err;
1409}
1410
1411/**
1412 * amdgpu_device_ip_early_init - run early init for hardware IPs
1413 *
1414 * @adev: amdgpu_device pointer
1415 *
1416 * Early initialization pass for hardware IPs.  The hardware IPs that make
1417 * up each asic are discovered each IP's early_init callback is run.  This
1418 * is the first stage in initializing the asic.
1419 * Returns 0 on success, negative error code on failure.
1420 */
1421static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
1422{
1423	int i, r;
1424
1425	amdgpu_device_enable_virtual_display(adev);
1426
1427	switch (adev->asic_type) {
1428	case CHIP_TOPAZ:
1429	case CHIP_TONGA:
1430	case CHIP_FIJI:
1431	case CHIP_POLARIS11:
1432	case CHIP_POLARIS10:
1433	case CHIP_POLARIS12:
1434	case CHIP_CARRIZO:
1435	case CHIP_STONEY:
1436		if (adev->asic_type == CHIP_CARRIZO || adev->asic_type == CHIP_STONEY)
1437			adev->family = AMDGPU_FAMILY_CZ;
1438		else
1439			adev->family = AMDGPU_FAMILY_VI;
1440
1441		r = vi_set_ip_blocks(adev);
1442		if (r)
1443			return r;
1444		break;
 
 
1445#ifdef CONFIG_DRM_AMDGPU_SI
1446	case CHIP_VERDE:
1447	case CHIP_TAHITI:
1448	case CHIP_PITCAIRN:
1449	case CHIP_OLAND:
1450	case CHIP_HAINAN:
1451		adev->family = AMDGPU_FAMILY_SI;
1452		r = si_set_ip_blocks(adev);
1453		if (r)
1454			return r;
1455		break;
1456#endif
1457#ifdef CONFIG_DRM_AMDGPU_CIK
1458	case CHIP_BONAIRE:
1459	case CHIP_HAWAII:
1460	case CHIP_KAVERI:
1461	case CHIP_KABINI:
1462	case CHIP_MULLINS:
1463		if ((adev->asic_type == CHIP_BONAIRE) || (adev->asic_type == CHIP_HAWAII))
1464			adev->family = AMDGPU_FAMILY_CI;
1465		else
1466			adev->family = AMDGPU_FAMILY_KV;
 
 
1467
1468		r = cik_set_ip_blocks(adev);
1469		if (r)
1470			return r;
1471		break;
1472#endif
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1473	case CHIP_VEGA10:
1474	case CHIP_VEGA12:
 
1475	case CHIP_RAVEN:
1476		if (adev->asic_type == CHIP_RAVEN)
 
 
 
1477			adev->family = AMDGPU_FAMILY_RV;
1478		else
1479			adev->family = AMDGPU_FAMILY_AI;
1480
1481		r = soc15_set_ip_blocks(adev);
1482		if (r)
1483			return r;
1484		break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1485	default:
1486		/* FIXME: not supported yet */
1487		return -EINVAL;
1488	}
1489
1490	r = amdgpu_device_parse_gpu_info_fw(adev);
1491	if (r)
1492		return r;
1493
1494	amdgpu_amdkfd_device_probe(adev);
1495
1496	if (amdgpu_sriov_vf(adev)) {
1497		r = amdgpu_virt_request_full_gpu(adev, true);
1498		if (r)
1499			return -EAGAIN;
1500	}
1501
1502	for (i = 0; i < adev->num_ip_blocks; i++) {
1503		if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
1504			DRM_ERROR("disabled ip block: %d <%s>\n",
1505				  i, adev->ip_blocks[i].version->funcs->name);
1506			adev->ip_blocks[i].status.valid = false;
1507		} else {
1508			if (adev->ip_blocks[i].version->funcs->early_init) {
1509				r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
1510				if (r == -ENOENT) {
1511					adev->ip_blocks[i].status.valid = false;
1512				} else if (r) {
1513					DRM_ERROR("early_init of IP block <%s> failed %d\n",
1514						  adev->ip_blocks[i].version->funcs->name, r);
1515					return r;
1516				} else {
1517					adev->ip_blocks[i].status.valid = true;
1518				}
1519			} else {
1520				adev->ip_blocks[i].status.valid = true;
1521			}
1522		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1523	}
1524
1525	adev->cg_flags &= amdgpu_cg_mask;
1526	adev->pg_flags &= amdgpu_pg_mask;
1527
1528	return 0;
1529}
1530
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1531/**
1532 * amdgpu_device_ip_init - run init for hardware IPs
1533 *
1534 * @adev: amdgpu_device pointer
1535 *
1536 * Main initialization pass for hardware IPs.  The list of all the hardware
1537 * IPs that make up the asic is walked and the sw_init and hw_init callbacks
1538 * are run.  sw_init initializes the software state associated with each IP
1539 * and hw_init initializes the hardware associated with each IP.
1540 * Returns 0 on success, negative error code on failure.
1541 */
1542static int amdgpu_device_ip_init(struct amdgpu_device *adev)
1543{
1544	int i, r;
1545
 
 
 
 
1546	for (i = 0; i < adev->num_ip_blocks; i++) {
1547		if (!adev->ip_blocks[i].status.valid)
1548			continue;
1549		r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
1550		if (r) {
1551			DRM_ERROR("sw_init of IP block <%s> failed %d\n",
1552				  adev->ip_blocks[i].version->funcs->name, r);
1553			return r;
1554		}
1555		adev->ip_blocks[i].status.sw = true;
1556
1557		/* need to do gmc hw init early so we can allocate gpu mem */
1558		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1559			r = amdgpu_device_vram_scratch_init(adev);
1560			if (r) {
1561				DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r);
1562				return r;
1563			}
1564			r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1565			if (r) {
1566				DRM_ERROR("hw_init %d failed %d\n", i, r);
1567				return r;
1568			}
1569			r = amdgpu_device_wb_init(adev);
1570			if (r) {
1571				DRM_ERROR("amdgpu_device_wb_init failed %d\n", r);
1572				return r;
1573			}
1574			adev->ip_blocks[i].status.hw = true;
1575
1576			/* right after GMC hw init, we create CSA */
1577			if (amdgpu_sriov_vf(adev)) {
1578				r = amdgpu_allocate_static_csa(adev);
 
 
1579				if (r) {
1580					DRM_ERROR("allocate CSA failed %d\n", r);
1581					return r;
1582				}
1583			}
1584		}
1585	}
1586
1587	for (i = 0; i < adev->num_ip_blocks; i++) {
1588		if (!adev->ip_blocks[i].status.sw)
1589			continue;
1590		if (adev->ip_blocks[i].status.hw)
1591			continue;
1592		r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1593		if (r) {
1594			DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1595				  adev->ip_blocks[i].version->funcs->name, r);
1596			return r;
1597		}
1598		adev->ip_blocks[i].status.hw = true;
1599	}
1600
1601	amdgpu_amdkfd_device_init(adev);
 
 
 
 
 
 
 
 
 
 
1602
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1603	if (amdgpu_sriov_vf(adev))
1604		amdgpu_virt_release_full_gpu(adev, true);
1605
1606	return 0;
1607}
1608
1609/**
1610 * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer
1611 *
1612 * @adev: amdgpu_device pointer
1613 *
1614 * Writes a reset magic value to the gart pointer in VRAM.  The driver calls
1615 * this function before a GPU reset.  If the value is retained after a
1616 * GPU reset, VRAM has not been lost.  Some GPU resets may destry VRAM contents.
1617 */
1618static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev)
1619{
1620	memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
1621}
1622
1623/**
1624 * amdgpu_device_check_vram_lost - check if vram is valid
1625 *
1626 * @adev: amdgpu_device pointer
1627 *
1628 * Checks the reset magic value written to the gart pointer in VRAM.
1629 * The driver calls this after a GPU reset to see if the contents of
1630 * VRAM is lost or now.
1631 * returns true if vram is lost, false if not.
1632 */
1633static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev)
1634{
1635	return !!memcmp(adev->gart.ptr, adev->reset_magic,
1636			AMDGPU_RESET_MAGIC_NUM);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1637}
1638
1639/**
1640 * amdgpu_device_ip_late_set_cg_state - late init for clockgating
1641 *
1642 * @adev: amdgpu_device pointer
 
1643 *
1644 * Late initialization pass enabling clockgating for hardware IPs.
1645 * The list of all the hardware IPs that make up the asic is walked and the
1646 * set_clockgating_state callbacks are run.  This stage is run late
1647 * in the init process.
 
1648 * Returns 0 on success, negative error code on failure.
1649 */
1650static int amdgpu_device_ip_late_set_cg_state(struct amdgpu_device *adev)
 
 
1651{
1652	int i = 0, r;
1653
1654	if (amdgpu_emu_mode == 1)
1655		return 0;
1656
1657	for (i = 0; i < adev->num_ip_blocks; i++) {
1658		if (!adev->ip_blocks[i].status.valid)
 
 
 
 
 
1659			continue;
1660		/* skip CG for VCE/UVD, it's handled specially */
1661		if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1662		    adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
 
 
1663		    adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1664			/* enable clockgating to save power */
1665			r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1666										     AMD_CG_STATE_GATE);
1667			if (r) {
1668				DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
1669					  adev->ip_blocks[i].version->funcs->name, r);
1670				return r;
1671			}
1672		}
1673	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1674	return 0;
1675}
1676
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1677/**
1678 * amdgpu_device_ip_late_init - run late init for hardware IPs
1679 *
1680 * @adev: amdgpu_device pointer
1681 *
1682 * Late initialization pass for hardware IPs.  The list of all the hardware
1683 * IPs that make up the asic is walked and the late_init callbacks are run.
1684 * late_init covers any special initialization that an IP requires
1685 * after all of the have been initialized or something that needs to happen
1686 * late in the init process.
1687 * Returns 0 on success, negative error code on failure.
1688 */
1689static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
1690{
 
1691	int i = 0, r;
1692
1693	for (i = 0; i < adev->num_ip_blocks; i++) {
1694		if (!adev->ip_blocks[i].status.valid)
1695			continue;
1696		if (adev->ip_blocks[i].version->funcs->late_init) {
1697			r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
1698			if (r) {
1699				DRM_ERROR("late_init of IP block <%s> failed %d\n",
1700					  adev->ip_blocks[i].version->funcs->name, r);
1701				return r;
1702			}
1703			adev->ip_blocks[i].status.late_initialized = true;
1704		}
 
1705	}
1706
1707	mod_delayed_work(system_wq, &adev->late_init_work,
1708			msecs_to_jiffies(AMDGPU_RESUME_MS));
 
 
1709
1710	amdgpu_device_fill_reset_magic(adev);
1711
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1712	return 0;
1713}
1714
1715/**
1716 * amdgpu_device_ip_fini - run fini for hardware IPs
1717 *
1718 * @adev: amdgpu_device pointer
1719 *
1720 * Main teardown pass for hardware IPs.  The list of all the hardware
1721 * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks
1722 * are run.  hw_fini tears down the hardware associated with each IP
1723 * and sw_fini tears down any software state associated with each IP.
1724 * Returns 0 on success, negative error code on failure.
1725 */
1726static int amdgpu_device_ip_fini(struct amdgpu_device *adev)
1727{
1728	int i, r;
1729
1730	amdgpu_amdkfd_device_fini(adev);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1731	/* need to disable SMC first */
1732	for (i = 0; i < adev->num_ip_blocks; i++) {
1733		if (!adev->ip_blocks[i].status.hw)
1734			continue;
1735		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC &&
1736			adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1737			/* ungate blocks before hw fini so that we can shutdown the blocks safely */
1738			r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1739										     AMD_CG_STATE_UNGATE);
1740			if (r) {
1741				DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1742					  adev->ip_blocks[i].version->funcs->name, r);
1743				return r;
1744			}
1745			r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
1746			/* XXX handle errors */
1747			if (r) {
1748				DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
1749					  adev->ip_blocks[i].version->funcs->name, r);
1750			}
1751			adev->ip_blocks[i].status.hw = false;
1752			break;
1753		}
1754	}
1755
1756	for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1757		if (!adev->ip_blocks[i].status.hw)
1758			continue;
1759
1760		if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1761			adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1762			adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1763			/* ungate blocks before hw fini so that we can shutdown the blocks safely */
1764			r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1765										     AMD_CG_STATE_UNGATE);
1766			if (r) {
1767				DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1768					  adev->ip_blocks[i].version->funcs->name, r);
1769				return r;
1770			}
1771		}
1772
1773		r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
1774		/* XXX handle errors */
1775		if (r) {
1776			DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
1777				  adev->ip_blocks[i].version->funcs->name, r);
1778		}
1779
1780		adev->ip_blocks[i].status.hw = false;
1781	}
1782
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1783
1784	for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1785		if (!adev->ip_blocks[i].status.sw)
1786			continue;
1787
1788		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1789			amdgpu_free_static_csa(adev);
 
1790			amdgpu_device_wb_fini(adev);
1791			amdgpu_device_vram_scratch_fini(adev);
 
1792		}
1793
1794		r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
1795		/* XXX handle errors */
1796		if (r) {
1797			DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
1798				  adev->ip_blocks[i].version->funcs->name, r);
1799		}
1800		adev->ip_blocks[i].status.sw = false;
1801		adev->ip_blocks[i].status.valid = false;
1802	}
1803
1804	for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1805		if (!adev->ip_blocks[i].status.late_initialized)
1806			continue;
1807		if (adev->ip_blocks[i].version->funcs->late_fini)
1808			adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
1809		adev->ip_blocks[i].status.late_initialized = false;
1810	}
1811
 
 
1812	if (amdgpu_sriov_vf(adev))
1813		if (amdgpu_virt_release_full_gpu(adev, false))
1814			DRM_ERROR("failed to release exclusive mode on fini\n");
1815
1816	return 0;
1817}
1818
1819/**
1820 * amdgpu_device_ip_late_init_func_handler - work handler for clockgating
1821 *
1822 * @work: work_struct
1823 *
1824 * Work handler for amdgpu_device_ip_late_set_cg_state.  We put the
1825 * clockgating setup into a worker thread to speed up driver init and
1826 * resume from suspend.
1827 */
1828static void amdgpu_device_ip_late_init_func_handler(struct work_struct *work)
1829{
1830	struct amdgpu_device *adev =
1831		container_of(work, struct amdgpu_device, late_init_work.work);
1832	amdgpu_device_ip_late_set_cg_state(adev);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1833}
1834
1835/**
1836 * amdgpu_device_ip_suspend - run suspend for hardware IPs
1837 *
1838 * @adev: amdgpu_device pointer
1839 *
1840 * Main suspend function for hardware IPs.  The list of all the hardware
1841 * IPs that make up the asic is walked, clockgating is disabled and the
1842 * suspend callbacks are run.  suspend puts the hardware and software state
1843 * in each IP into a state suitable for suspend.
1844 * Returns 0 on success, negative error code on failure.
1845 */
1846int amdgpu_device_ip_suspend(struct amdgpu_device *adev)
1847{
1848	int i, r;
1849
1850	if (amdgpu_sriov_vf(adev))
1851		amdgpu_virt_request_full_gpu(adev, false);
1852
1853	/* ungate SMC block first */
1854	r = amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_SMC,
1855						   AMD_CG_STATE_UNGATE);
1856	if (r) {
1857		DRM_ERROR("set_clockgating_state(ungate) SMC failed %d\n", r);
 
 
 
 
 
 
 
 
 
 
 
 
 
1858	}
1859
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1860	for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1861		if (!adev->ip_blocks[i].status.valid)
1862			continue;
1863		/* ungate blocks so that suspend can properly shut them down */
1864		if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_SMC &&
1865			adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1866			r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1867										     AMD_CG_STATE_UNGATE);
1868			if (r) {
1869				DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1870					  adev->ip_blocks[i].version->funcs->name, r);
1871			}
 
 
 
 
 
 
 
 
 
1872		}
 
 
 
 
 
 
 
 
 
 
 
1873		/* XXX handle errors */
1874		r = adev->ip_blocks[i].version->funcs->suspend(adev);
1875		/* XXX handle errors */
1876		if (r) {
1877			DRM_ERROR("suspend of IP block <%s> failed %d\n",
1878				  adev->ip_blocks[i].version->funcs->name, r);
1879		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1880	}
1881
 
 
 
 
 
1882	if (amdgpu_sriov_vf(adev))
1883		amdgpu_virt_release_full_gpu(adev, false);
1884
1885	return 0;
1886}
1887
1888static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
1889{
1890	int i, r;
1891
1892	static enum amd_ip_block_type ip_order[] = {
1893		AMD_IP_BLOCK_TYPE_GMC,
1894		AMD_IP_BLOCK_TYPE_COMMON,
 
1895		AMD_IP_BLOCK_TYPE_IH,
1896	};
1897
1898	for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
1899		int j;
1900		struct amdgpu_ip_block *block;
1901
1902		for (j = 0; j < adev->num_ip_blocks; j++) {
1903			block = &adev->ip_blocks[j];
1904
1905			if (block->version->type != ip_order[i] ||
 
 
1906				!block->status.valid)
1907				continue;
1908
1909			r = block->version->funcs->hw_init(adev);
1910			DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"successed");
1911			if (r)
1912				return r;
 
1913		}
1914	}
1915
1916	return 0;
1917}
1918
1919static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
1920{
1921	int i, r;
1922
1923	static enum amd_ip_block_type ip_order[] = {
1924		AMD_IP_BLOCK_TYPE_SMC,
1925		AMD_IP_BLOCK_TYPE_PSP,
1926		AMD_IP_BLOCK_TYPE_DCE,
1927		AMD_IP_BLOCK_TYPE_GFX,
1928		AMD_IP_BLOCK_TYPE_SDMA,
1929		AMD_IP_BLOCK_TYPE_UVD,
1930		AMD_IP_BLOCK_TYPE_VCE
 
1931	};
1932
1933	for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
1934		int j;
1935		struct amdgpu_ip_block *block;
1936
1937		for (j = 0; j < adev->num_ip_blocks; j++) {
1938			block = &adev->ip_blocks[j];
1939
1940			if (block->version->type != ip_order[i] ||
1941				!block->status.valid)
 
1942				continue;
1943
1944			r = block->version->funcs->hw_init(adev);
1945			DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"successed");
 
 
 
 
1946			if (r)
1947				return r;
 
1948		}
1949	}
1950
1951	return 0;
1952}
1953
1954/**
1955 * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs
1956 *
1957 * @adev: amdgpu_device pointer
1958 *
1959 * First resume function for hardware IPs.  The list of all the hardware
1960 * IPs that make up the asic is walked and the resume callbacks are run for
1961 * COMMON, GMC, and IH.  resume puts the hardware into a functional state
1962 * after a suspend and updates the software state as necessary.  This
1963 * function is also used for restoring the GPU after a GPU reset.
1964 * Returns 0 on success, negative error code on failure.
1965 */
1966static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)
1967{
1968	int i, r;
1969
1970	for (i = 0; i < adev->num_ip_blocks; i++) {
1971		if (!adev->ip_blocks[i].status.valid)
1972			continue;
1973		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
1974		    adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
1975		    adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
 
1976			r = adev->ip_blocks[i].version->funcs->resume(adev);
1977			if (r) {
1978				DRM_ERROR("resume of IP block <%s> failed %d\n",
1979					  adev->ip_blocks[i].version->funcs->name, r);
1980				return r;
1981			}
 
1982		}
1983	}
1984
1985	return 0;
1986}
1987
1988/**
1989 * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs
1990 *
1991 * @adev: amdgpu_device pointer
1992 *
1993 * First resume function for hardware IPs.  The list of all the hardware
1994 * IPs that make up the asic is walked and the resume callbacks are run for
1995 * all blocks except COMMON, GMC, and IH.  resume puts the hardware into a
1996 * functional state after a suspend and updates the software state as
1997 * necessary.  This function is also used for restoring the GPU after a GPU
1998 * reset.
1999 * Returns 0 on success, negative error code on failure.
2000 */
2001static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)
2002{
2003	int i, r;
2004
2005	for (i = 0; i < adev->num_ip_blocks; i++) {
2006		if (!adev->ip_blocks[i].status.valid)
2007			continue;
2008		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2009		    adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2010		    adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH)
 
2011			continue;
2012		r = adev->ip_blocks[i].version->funcs->resume(adev);
2013		if (r) {
2014			DRM_ERROR("resume of IP block <%s> failed %d\n",
2015				  adev->ip_blocks[i].version->funcs->name, r);
2016			return r;
2017		}
 
2018	}
2019
2020	return 0;
2021}
2022
2023/**
2024 * amdgpu_device_ip_resume - run resume for hardware IPs
2025 *
2026 * @adev: amdgpu_device pointer
2027 *
2028 * Main resume function for hardware IPs.  The hardware IPs
2029 * are split into two resume functions because they are
2030 * are also used in in recovering from a GPU reset and some additional
2031 * steps need to be take between them.  In this case (S3/S4) they are
2032 * run sequentially.
2033 * Returns 0 on success, negative error code on failure.
2034 */
2035static int amdgpu_device_ip_resume(struct amdgpu_device *adev)
2036{
2037	int r;
2038
 
 
 
 
2039	r = amdgpu_device_ip_resume_phase1(adev);
2040	if (r)
2041		return r;
 
 
 
 
 
2042	r = amdgpu_device_ip_resume_phase2(adev);
2043
2044	return r;
2045}
2046
2047/**
2048 * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV
2049 *
2050 * @adev: amdgpu_device pointer
2051 *
2052 * Query the VBIOS data tables to determine if the board supports SR-IOV.
2053 */
2054static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
2055{
2056	if (amdgpu_sriov_vf(adev)) {
2057		if (adev->is_atom_fw) {
2058			if (amdgpu_atomfirmware_gpu_supports_virtualization(adev))
2059				adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2060		} else {
2061			if (amdgpu_atombios_has_gpu_virtualization_table(adev))
2062				adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2063		}
2064
2065		if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
2066			amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
2067	}
2068}
2069
2070/**
2071 * amdgpu_device_asic_has_dc_support - determine if DC supports the asic
2072 *
2073 * @asic_type: AMD asic type
2074 *
2075 * Check if there is DC (new modesetting infrastructre) support for an asic.
2076 * returns true if DC has support, false if not.
2077 */
2078bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
2079{
2080	switch (asic_type) {
2081#if defined(CONFIG_DRM_AMD_DC)
 
 
 
 
 
 
2082	case CHIP_BONAIRE:
2083	case CHIP_HAWAII:
2084	case CHIP_KAVERI:
2085	case CHIP_KABINI:
2086	case CHIP_MULLINS:
 
 
 
 
 
 
 
 
 
2087	case CHIP_CARRIZO:
2088	case CHIP_STONEY:
2089	case CHIP_POLARIS11:
2090	case CHIP_POLARIS10:
 
2091	case CHIP_POLARIS12:
 
2092	case CHIP_TONGA:
2093	case CHIP_FIJI:
2094#if defined(CONFIG_DRM_AMD_DC_PRE_VEGA)
2095		return amdgpu_dc != 0;
2096#endif
2097	case CHIP_VEGA10:
2098	case CHIP_VEGA12:
2099#if defined(CONFIG_DRM_AMD_DC_DCN1_0)
 
2100	case CHIP_RAVEN:
 
 
 
 
 
 
 
 
 
 
2101#endif
2102		return amdgpu_dc != 0;
2103#endif
2104	default:
 
 
 
2105		return false;
2106	}
2107}
2108
2109/**
2110 * amdgpu_device_has_dc_support - check if dc is supported
2111 *
2112 * @adev: amdgpu_device_pointer
2113 *
2114 * Returns true for supported, false for not supported
2115 */
2116bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
2117{
2118	if (amdgpu_sriov_vf(adev))
 
 
2119		return false;
2120
2121	return amdgpu_device_asic_has_dc_support(adev->asic_type);
2122}
2123
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2124/**
2125 * amdgpu_device_init - initialize the driver
2126 *
2127 * @adev: amdgpu_device pointer
2128 * @pdev: drm dev pointer
2129 * @pdev: pci dev pointer
2130 * @flags: driver flags
2131 *
2132 * Initializes the driver info and hw (all asics).
2133 * Returns 0 for success or an error on failure.
2134 * Called at driver startup.
2135 */
2136int amdgpu_device_init(struct amdgpu_device *adev,
2137		       struct drm_device *ddev,
2138		       struct pci_dev *pdev,
2139		       uint32_t flags)
2140{
 
 
2141	int r, i;
2142	bool runtime = false;
2143	u32 max_MBps;
2144
2145	adev->shutdown = false;
2146	adev->dev = &pdev->dev;
2147	adev->ddev = ddev;
2148	adev->pdev = pdev;
2149	adev->flags = flags;
2150	adev->asic_type = flags & AMD_ASIC_MASK;
 
 
 
 
 
2151	adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
2152	if (amdgpu_emu_mode == 1)
2153		adev->usec_timeout *= 2;
2154	adev->gmc.gart_size = 512 * 1024 * 1024;
2155	adev->accel_working = false;
2156	adev->num_rings = 0;
2157	adev->mman.buffer_funcs = NULL;
2158	adev->mman.buffer_funcs_ring = NULL;
2159	adev->vm_manager.vm_pte_funcs = NULL;
2160	adev->vm_manager.vm_pte_num_rings = 0;
2161	adev->gmc.gmc_funcs = NULL;
 
2162	adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
2163	bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
2164
2165	adev->smc_rreg = &amdgpu_invalid_rreg;
2166	adev->smc_wreg = &amdgpu_invalid_wreg;
2167	adev->pcie_rreg = &amdgpu_invalid_rreg;
2168	adev->pcie_wreg = &amdgpu_invalid_wreg;
2169	adev->pciep_rreg = &amdgpu_invalid_rreg;
2170	adev->pciep_wreg = &amdgpu_invalid_wreg;
 
 
2171	adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
2172	adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
2173	adev->didt_rreg = &amdgpu_invalid_rreg;
2174	adev->didt_wreg = &amdgpu_invalid_wreg;
2175	adev->gc_cac_rreg = &amdgpu_invalid_rreg;
2176	adev->gc_cac_wreg = &amdgpu_invalid_wreg;
2177	adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
2178	adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
2179
2180	DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
2181		 amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
2182		 pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
2183
2184	/* mutex initialization are all done here so we
2185	 * can recall function without having locking issues */
2186	atomic_set(&adev->irq.ih.lock, 0);
2187	mutex_init(&adev->firmware.mutex);
2188	mutex_init(&adev->pm.mutex);
2189	mutex_init(&adev->gfx.gpu_clock_mutex);
2190	mutex_init(&adev->srbm_mutex);
2191	mutex_init(&adev->gfx.pipe_reserve_mutex);
 
2192	mutex_init(&adev->grbm_idx_mutex);
2193	mutex_init(&adev->mn_lock);
2194	mutex_init(&adev->virt.vf_errors.lock);
2195	hash_init(adev->mn_hash);
2196	mutex_init(&adev->lock_reset);
 
 
 
 
 
 
 
2197
2198	amdgpu_device_check_arguments(adev);
 
 
2199
2200	spin_lock_init(&adev->mmio_idx_lock);
2201	spin_lock_init(&adev->smc_idx_lock);
2202	spin_lock_init(&adev->pcie_idx_lock);
2203	spin_lock_init(&adev->uvd_ctx_idx_lock);
2204	spin_lock_init(&adev->didt_idx_lock);
2205	spin_lock_init(&adev->gc_cac_idx_lock);
2206	spin_lock_init(&adev->se_cac_idx_lock);
2207	spin_lock_init(&adev->audio_endpt_idx_lock);
2208	spin_lock_init(&adev->mm_stats.lock);
2209
2210	INIT_LIST_HEAD(&adev->shadow_list);
2211	mutex_init(&adev->shadow_list_lock);
2212
2213	INIT_LIST_HEAD(&adev->ring_lru_list);
2214	spin_lock_init(&adev->ring_lru_list_lock);
2215
2216	INIT_DELAYED_WORK(&adev->late_init_work,
2217			  amdgpu_device_ip_late_init_func_handler);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2218
2219	/* Registers mapping */
2220	/* TODO: block userspace mapping of io register */
2221	if (adev->asic_type >= CHIP_BONAIRE) {
2222		adev->rmmio_base = pci_resource_start(adev->pdev, 5);
2223		adev->rmmio_size = pci_resource_len(adev->pdev, 5);
2224	} else {
2225		adev->rmmio_base = pci_resource_start(adev->pdev, 2);
2226		adev->rmmio_size = pci_resource_len(adev->pdev, 2);
2227	}
2228
2229	adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
2230	if (adev->rmmio == NULL) {
2231		return -ENOMEM;
2232	}
2233	DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
2234	DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
2235
2236	/* doorbell bar mapping */
2237	amdgpu_device_doorbell_init(adev);
2238
2239	/* io port mapping */
2240	for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
2241		if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) {
2242			adev->rio_mem_size = pci_resource_len(adev->pdev, i);
2243			adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size);
2244			break;
2245		}
2246	}
2247	if (adev->rio_mem == NULL)
2248		DRM_INFO("PCI I/O BAR is not found.\n");
2249
2250	amdgpu_device_get_pcie_info(adev);
2251
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2252	/* early init functions */
2253	r = amdgpu_device_ip_early_init(adev);
2254	if (r)
2255		return r;
2256
2257	/* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
2258	/* this will fail for cards that aren't VGA class devices, just
2259	 * ignore it */
2260	vga_client_register(adev->pdev, adev, NULL, amdgpu_device_vga_set_decode);
2261
2262	if (amdgpu_device_is_px(ddev))
2263		runtime = true;
2264	if (!pci_is_thunderbolt_attached(adev->pdev))
2265		vga_switcheroo_register_client(adev->pdev,
2266					       &amdgpu_switcheroo_ops, runtime);
2267	if (runtime)
2268		vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
2269
2270	if (amdgpu_emu_mode == 1) {
2271		/* post the asic on emulation mode */
2272		emu_soc_asic_init(adev);
2273		goto fence_driver_init;
2274	}
2275
2276	/* Read BIOS */
2277	if (!amdgpu_get_bios(adev)) {
2278		r = -EINVAL;
2279		goto failed;
2280	}
2281
2282	r = amdgpu_atombios_init(adev);
2283	if (r) {
2284		dev_err(adev->dev, "amdgpu_atombios_init failed\n");
2285		amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0);
2286		goto failed;
2287	}
2288
2289	/* detect if we are with an SRIOV vbios */
2290	amdgpu_device_detect_sriov_bios(adev);
2291
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2292	/* Post card if necessary */
2293	if (amdgpu_device_need_post(adev)) {
2294		if (!adev->bios) {
2295			dev_err(adev->dev, "no vBIOS found\n");
2296			r = -EINVAL;
2297			goto failed;
2298		}
2299		DRM_INFO("GPU posting now...\n");
2300		r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2301		if (r) {
2302			dev_err(adev->dev, "gpu post error!\n");
2303			goto failed;
2304		}
2305	}
2306
2307	if (adev->is_atom_fw) {
2308		/* Initialize clocks */
2309		r = amdgpu_atomfirmware_get_clock_info(adev);
2310		if (r) {
2311			dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
2312			amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
2313			goto failed;
2314		}
2315	} else {
2316		/* Initialize clocks */
2317		r = amdgpu_atombios_get_clock_info(adev);
2318		if (r) {
2319			dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
2320			amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
2321			goto failed;
2322		}
2323		/* init i2c buses */
2324		if (!amdgpu_device_has_dc_support(adev))
2325			amdgpu_atombios_i2c_init(adev);
2326	}
2327
2328fence_driver_init:
2329	/* Fence driver */
2330	r = amdgpu_fence_driver_init(adev);
2331	if (r) {
2332		dev_err(adev->dev, "amdgpu_fence_driver_init failed\n");
2333		amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
2334		goto failed;
2335	}
2336
2337	/* init the mode config */
2338	drm_mode_config_init(adev->ddev);
2339
2340	r = amdgpu_device_ip_init(adev);
2341	if (r) {
2342		/* failed in exclusive mode due to timeout */
2343		if (amdgpu_sriov_vf(adev) &&
2344		    !amdgpu_sriov_runtime(adev) &&
2345		    amdgpu_virt_mmio_blocked(adev) &&
2346		    !amdgpu_virt_wait_reset(adev)) {
2347			dev_err(adev->dev, "VF exclusive mode timeout\n");
2348			/* Don't send request since VF is inactive. */
2349			adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
2350			adev->virt.ops = NULL;
2351			r = -EAGAIN;
2352			goto failed;
2353		}
2354		dev_err(adev->dev, "amdgpu_device_ip_init failed\n");
2355		amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
2356		goto failed;
2357	}
2358
 
 
 
 
 
 
 
 
 
2359	adev->accel_working = true;
2360
2361	amdgpu_vm_check_compute_bug(adev);
2362
2363	/* Initialize the buffer migration limit. */
2364	if (amdgpu_moverate >= 0)
2365		max_MBps = amdgpu_moverate;
2366	else
2367		max_MBps = 8; /* Allow 8 MB/s. */
2368	/* Get a log2 for easy divisions. */
2369	adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
2370
2371	r = amdgpu_ib_pool_init(adev);
2372	if (r) {
2373		dev_err(adev->dev, "IB initialization failed (%d).\n", r);
2374		amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r);
2375		goto failed;
2376	}
2377
2378	r = amdgpu_ib_ring_tests(adev);
2379	if (r)
2380		DRM_ERROR("ib ring test failed (%d).\n", r);
2381
2382	if (amdgpu_sriov_vf(adev))
2383		amdgpu_virt_init_data_exchange(adev);
2384
2385	amdgpu_fbdev_init(adev);
2386
2387	r = amdgpu_pm_sysfs_init(adev);
2388	if (r)
 
2389		DRM_ERROR("registering pm debugfs failed (%d).\n", r);
 
 
2390
2391	r = amdgpu_debugfs_gem_init(adev);
2392	if (r)
2393		DRM_ERROR("registering gem debugfs failed (%d).\n", r);
2394
2395	r = amdgpu_debugfs_regs_init(adev);
2396	if (r)
2397		DRM_ERROR("registering register debugfs failed (%d).\n", r);
2398
2399	r = amdgpu_debugfs_firmware_init(adev);
2400	if (r)
2401		DRM_ERROR("registering firmware debugfs failed (%d).\n", r);
2402
2403	r = amdgpu_debugfs_init(adev);
2404	if (r)
2405		DRM_ERROR("Creating debugfs files failed (%d).\n", r);
2406
2407	if ((amdgpu_testing & 1)) {
2408		if (adev->accel_working)
2409			amdgpu_test_moves(adev);
2410		else
2411			DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n");
2412	}
2413	if (amdgpu_benchmarking) {
2414		if (adev->accel_working)
2415			amdgpu_benchmark(adev, amdgpu_benchmarking);
2416		else
2417			DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n");
2418	}
2419
 
 
 
 
 
 
 
2420	/* enable clockgating, etc. after ib tests, etc. since some blocks require
2421	 * explicit gating rather than handling it automatically.
2422	 */
2423	r = amdgpu_device_ip_late_init(adev);
2424	if (r) {
2425		dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n");
2426		amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
2427		goto failed;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2428	}
2429
 
 
 
 
2430	return 0;
2431
 
 
 
2432failed:
2433	amdgpu_vf_error_trans_all(adev);
2434	if (runtime)
2435		vga_switcheroo_fini_domain_pm_ops(adev->dev);
2436
2437	return r;
2438}
2439
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2440/**
2441 * amdgpu_device_fini - tear down the driver
2442 *
2443 * @adev: amdgpu_device pointer
2444 *
2445 * Tear down the driver info (all asics).
2446 * Called at driver shutdown.
2447 */
2448void amdgpu_device_fini(struct amdgpu_device *adev)
2449{
2450	int r;
2451
2452	DRM_INFO("amdgpu: finishing device.\n");
2453	adev->shutdown = true;
 
 
 
 
 
 
 
 
 
2454	/* disable all interrupts */
2455	amdgpu_irq_disable_all(adev);
2456	if (adev->mode_info.mode_config_initialized){
2457		if (!amdgpu_device_has_dc_support(adev))
2458			drm_crtc_force_disable_all(adev->ddev);
2459		else
2460			drm_atomic_helper_shutdown(adev->ddev);
2461	}
2462	amdgpu_ib_pool_fini(adev);
2463	amdgpu_fence_driver_fini(adev);
2464	amdgpu_pm_sysfs_fini(adev);
 
 
 
 
 
2465	amdgpu_fbdev_fini(adev);
2466	r = amdgpu_device_ip_fini(adev);
2467	if (adev->firmware.gpu_info_fw) {
2468		release_firmware(adev->firmware.gpu_info_fw);
2469		adev->firmware.gpu_info_fw = NULL;
2470	}
 
 
 
 
 
 
 
 
 
 
 
2471	adev->accel_working = false;
2472	cancel_delayed_work_sync(&adev->late_init_work);
 
 
2473	/* free i2c buses */
2474	if (!amdgpu_device_has_dc_support(adev))
2475		amdgpu_i2c_fini(adev);
2476
2477	if (amdgpu_emu_mode != 1)
2478		amdgpu_atombios_fini(adev);
2479
2480	kfree(adev->bios);
2481	adev->bios = NULL;
2482	if (!pci_is_thunderbolt_attached(adev->pdev))
2483		vga_switcheroo_unregister_client(adev->pdev);
2484	if (adev->flags & AMD_IS_PX)
2485		vga_switcheroo_fini_domain_pm_ops(adev->dev);
2486	vga_client_register(adev->pdev, NULL, NULL, NULL);
2487	if (adev->rio_mem)
2488		pci_iounmap(adev->pdev, adev->rio_mem);
2489	adev->rio_mem = NULL;
2490	iounmap(adev->rmmio);
2491	adev->rmmio = NULL;
2492	amdgpu_device_doorbell_fini(adev);
2493	amdgpu_debugfs_regs_cleanup(adev);
 
 
 
2494}
2495
2496
2497/*
2498 * Suspend & resume.
2499 */
2500/**
2501 * amdgpu_device_suspend - initiate device suspend
2502 *
2503 * @pdev: drm dev pointer
2504 * @state: suspend state
2505 *
2506 * Puts the hw in the suspend state (all asics).
2507 * Returns 0 for success or an error on failure.
2508 * Called at driver suspend.
2509 */
2510int amdgpu_device_suspend(struct drm_device *dev, bool suspend, bool fbcon)
2511{
2512	struct amdgpu_device *adev;
2513	struct drm_crtc *crtc;
2514	struct drm_connector *connector;
2515	int r;
2516
2517	if (dev == NULL || dev->dev_private == NULL) {
2518		return -ENODEV;
2519	}
2520
2521	adev = dev->dev_private;
2522
2523	if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
2524		return 0;
2525
 
 
 
 
 
2526	drm_kms_helper_poll_disable(dev);
2527
2528	if (!amdgpu_device_has_dc_support(adev)) {
2529		/* turn off display hw */
2530		drm_modeset_lock_all(dev);
2531		list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2532			drm_helper_connector_dpms(connector, DRM_MODE_DPMS_OFF);
2533		}
2534		drm_modeset_unlock_all(dev);
2535	}
2536
2537	amdgpu_amdkfd_suspend(adev);
2538
2539	/* unpin the front buffers and cursors */
2540	list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2541		struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2542		struct amdgpu_framebuffer *rfb = to_amdgpu_framebuffer(crtc->primary->fb);
2543		struct amdgpu_bo *robj;
2544
2545		if (amdgpu_crtc->cursor_bo) {
2546			struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2547			r = amdgpu_bo_reserve(aobj, true);
2548			if (r == 0) {
2549				amdgpu_bo_unpin(aobj);
2550				amdgpu_bo_unreserve(aobj);
2551			}
2552		}
2553
2554		if (rfb == NULL || rfb->obj == NULL) {
2555			continue;
2556		}
2557		robj = gem_to_amdgpu_bo(rfb->obj);
2558		/* don't unpin kernel fb objects */
2559		if (!amdgpu_fbdev_robj_is_fb(adev, robj)) {
2560			r = amdgpu_bo_reserve(robj, true);
2561			if (r == 0) {
2562				amdgpu_bo_unpin(robj);
2563				amdgpu_bo_unreserve(robj);
2564			}
2565		}
2566	}
2567	/* evict vram memory */
2568	amdgpu_bo_evict_vram(adev);
2569
2570	amdgpu_fence_driver_suspend(adev);
2571
2572	r = amdgpu_device_ip_suspend(adev);
2573
 
2574	/* evict remaining vram memory
2575	 * This second call to evict vram is to evict the gart page table
2576	 * using the CPU.
2577	 */
2578	amdgpu_bo_evict_vram(adev);
2579
2580	pci_save_state(dev->pdev);
2581	if (suspend) {
2582		/* Shut down the device */
2583		pci_disable_device(dev->pdev);
2584		pci_set_power_state(dev->pdev, PCI_D3hot);
2585	} else {
2586		r = amdgpu_asic_reset(adev);
2587		if (r)
2588			DRM_ERROR("amdgpu asic reset failed\n");
2589	}
2590
2591	if (fbcon) {
2592		console_lock();
2593		amdgpu_fbdev_set_suspend(adev, 1);
2594		console_unlock();
2595	}
2596	return 0;
2597}
2598
2599/**
2600 * amdgpu_device_resume - initiate device resume
2601 *
2602 * @pdev: drm dev pointer
 
2603 *
2604 * Bring the hw back to operating state (all asics).
2605 * Returns 0 for success or an error on failure.
2606 * Called at driver resume.
2607 */
2608int amdgpu_device_resume(struct drm_device *dev, bool resume, bool fbcon)
2609{
2610	struct drm_connector *connector;
2611	struct amdgpu_device *adev = dev->dev_private;
2612	struct drm_crtc *crtc;
2613	int r = 0;
2614
2615	if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
2616		return 0;
2617
2618	if (fbcon)
2619		console_lock();
2620
2621	if (resume) {
2622		pci_set_power_state(dev->pdev, PCI_D0);
2623		pci_restore_state(dev->pdev);
2624		r = pci_enable_device(dev->pdev);
2625		if (r)
2626			goto unlock;
2627	}
2628
2629	/* post card */
2630	if (amdgpu_device_need_post(adev)) {
2631		r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2632		if (r)
2633			DRM_ERROR("amdgpu asic init failed\n");
2634	}
2635
2636	r = amdgpu_device_ip_resume(adev);
2637	if (r) {
2638		DRM_ERROR("amdgpu_device_ip_resume failed (%d).\n", r);
2639		goto unlock;
2640	}
2641	amdgpu_fence_driver_resume(adev);
2642
2643	if (resume) {
2644		r = amdgpu_ib_ring_tests(adev);
2645		if (r)
2646			DRM_ERROR("ib ring test failed (%d).\n", r);
2647	}
 
2648
2649	r = amdgpu_device_ip_late_init(adev);
2650	if (r)
2651		goto unlock;
2652
2653	/* pin cursors */
2654	list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2655		struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2656
2657		if (amdgpu_crtc->cursor_bo) {
2658			struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2659			r = amdgpu_bo_reserve(aobj, true);
2660			if (r == 0) {
2661				r = amdgpu_bo_pin(aobj,
2662						  AMDGPU_GEM_DOMAIN_VRAM,
2663						  &amdgpu_crtc->cursor_addr);
2664				if (r != 0)
2665					DRM_ERROR("Failed to pin cursor BO (%d)\n", r);
2666				amdgpu_bo_unreserve(aobj);
2667			}
2668		}
2669	}
2670	r = amdgpu_amdkfd_resume(adev);
2671	if (r)
2672		return r;
2673
2674	/* blat the mode back in */
2675	if (fbcon) {
2676		if (!amdgpu_device_has_dc_support(adev)) {
2677			/* pre DCE11 */
2678			drm_helper_resume_force_mode(dev);
2679
2680			/* turn on display hw */
2681			drm_modeset_lock_all(dev);
2682			list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2683				drm_helper_connector_dpms(connector, DRM_MODE_DPMS_ON);
2684			}
2685			drm_modeset_unlock_all(dev);
2686		}
2687	}
2688
 
 
 
 
 
 
2689	drm_kms_helper_poll_enable(dev);
2690
 
 
2691	/*
2692	 * Most of the connector probing functions try to acquire runtime pm
2693	 * refs to ensure that the GPU is powered on when connector polling is
2694	 * performed. Since we're calling this from a runtime PM callback,
2695	 * trying to acquire rpm refs will cause us to deadlock.
2696	 *
2697	 * Since we're guaranteed to be holding the rpm lock, it's safe to
2698	 * temporarily disable the rpm helpers so this doesn't deadlock us.
2699	 */
2700#ifdef CONFIG_PM
2701	dev->dev->power.disable_depth++;
2702#endif
2703	if (!amdgpu_device_has_dc_support(adev))
2704		drm_helper_hpd_irq_event(dev);
2705	else
2706		drm_kms_helper_hotplug_event(dev);
2707#ifdef CONFIG_PM
2708	dev->dev->power.disable_depth--;
2709#endif
 
2710
2711	if (fbcon)
2712		amdgpu_fbdev_set_suspend(adev, 0);
2713
2714unlock:
2715	if (fbcon)
2716		console_unlock();
2717
2718	return r;
2719}
2720
2721/**
2722 * amdgpu_device_ip_check_soft_reset - did soft reset succeed
2723 *
2724 * @adev: amdgpu_device pointer
2725 *
2726 * The list of all the hardware IPs that make up the asic is walked and
2727 * the check_soft_reset callbacks are run.  check_soft_reset determines
2728 * if the asic is still hung or not.
2729 * Returns true if any of the IPs are still in a hung state, false if not.
2730 */
2731static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)
2732{
2733	int i;
2734	bool asic_hang = false;
2735
2736	if (amdgpu_sriov_vf(adev))
2737		return true;
2738
 
 
 
2739	for (i = 0; i < adev->num_ip_blocks; i++) {
2740		if (!adev->ip_blocks[i].status.valid)
2741			continue;
2742		if (adev->ip_blocks[i].version->funcs->check_soft_reset)
2743			adev->ip_blocks[i].status.hang =
2744				adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
2745		if (adev->ip_blocks[i].status.hang) {
2746			DRM_INFO("IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
2747			asic_hang = true;
2748		}
2749	}
2750	return asic_hang;
2751}
2752
2753/**
2754 * amdgpu_device_ip_pre_soft_reset - prepare for soft reset
2755 *
2756 * @adev: amdgpu_device pointer
2757 *
2758 * The list of all the hardware IPs that make up the asic is walked and the
2759 * pre_soft_reset callbacks are run if the block is hung.  pre_soft_reset
2760 * handles any IP specific hardware or software state changes that are
2761 * necessary for a soft reset to succeed.
2762 * Returns 0 on success, negative error code on failure.
2763 */
2764static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)
2765{
2766	int i, r = 0;
2767
2768	for (i = 0; i < adev->num_ip_blocks; i++) {
2769		if (!adev->ip_blocks[i].status.valid)
2770			continue;
2771		if (adev->ip_blocks[i].status.hang &&
2772		    adev->ip_blocks[i].version->funcs->pre_soft_reset) {
2773			r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
2774			if (r)
2775				return r;
2776		}
2777	}
2778
2779	return 0;
2780}
2781
2782/**
2783 * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed
2784 *
2785 * @adev: amdgpu_device pointer
2786 *
2787 * Some hardware IPs cannot be soft reset.  If they are hung, a full gpu
2788 * reset is necessary to recover.
2789 * Returns true if a full asic reset is required, false if not.
2790 */
2791static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)
2792{
2793	int i;
2794
 
 
 
2795	for (i = 0; i < adev->num_ip_blocks; i++) {
2796		if (!adev->ip_blocks[i].status.valid)
2797			continue;
2798		if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
2799		    (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
2800		    (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
2801		    (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
2802		     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
2803			if (adev->ip_blocks[i].status.hang) {
2804				DRM_INFO("Some block need full reset!\n");
2805				return true;
2806			}
2807		}
2808	}
2809	return false;
2810}
2811
2812/**
2813 * amdgpu_device_ip_soft_reset - do a soft reset
2814 *
2815 * @adev: amdgpu_device pointer
2816 *
2817 * The list of all the hardware IPs that make up the asic is walked and the
2818 * soft_reset callbacks are run if the block is hung.  soft_reset handles any
2819 * IP specific hardware or software state changes that are necessary to soft
2820 * reset the IP.
2821 * Returns 0 on success, negative error code on failure.
2822 */
2823static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)
2824{
2825	int i, r = 0;
2826
2827	for (i = 0; i < adev->num_ip_blocks; i++) {
2828		if (!adev->ip_blocks[i].status.valid)
2829			continue;
2830		if (adev->ip_blocks[i].status.hang &&
2831		    adev->ip_blocks[i].version->funcs->soft_reset) {
2832			r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
2833			if (r)
2834				return r;
2835		}
2836	}
2837
2838	return 0;
2839}
2840
2841/**
2842 * amdgpu_device_ip_post_soft_reset - clean up from soft reset
2843 *
2844 * @adev: amdgpu_device pointer
2845 *
2846 * The list of all the hardware IPs that make up the asic is walked and the
2847 * post_soft_reset callbacks are run if the asic was hung.  post_soft_reset
2848 * handles any IP specific hardware or software state changes that are
2849 * necessary after the IP has been soft reset.
2850 * Returns 0 on success, negative error code on failure.
2851 */
2852static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)
2853{
2854	int i, r = 0;
2855
2856	for (i = 0; i < adev->num_ip_blocks; i++) {
2857		if (!adev->ip_blocks[i].status.valid)
2858			continue;
2859		if (adev->ip_blocks[i].status.hang &&
2860		    adev->ip_blocks[i].version->funcs->post_soft_reset)
2861			r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
2862		if (r)
2863			return r;
2864	}
2865
2866	return 0;
2867}
2868
2869/**
2870 * amdgpu_device_recover_vram_from_shadow - restore shadowed VRAM buffers
2871 *
2872 * @adev: amdgpu_device pointer
2873 * @ring: amdgpu_ring for the engine handling the buffer operations
2874 * @bo: amdgpu_bo buffer whose shadow is being restored
2875 * @fence: dma_fence associated with the operation
2876 *
2877 * Restores the VRAM buffer contents from the shadow in GTT.  Used to
2878 * restore things like GPUVM page tables after a GPU reset where
2879 * the contents of VRAM might be lost.
2880 * Returns 0 on success, negative error code on failure.
2881 */
2882static int amdgpu_device_recover_vram_from_shadow(struct amdgpu_device *adev,
2883						  struct amdgpu_ring *ring,
2884						  struct amdgpu_bo *bo,
2885						  struct dma_fence **fence)
2886{
2887	uint32_t domain;
2888	int r;
2889
2890	if (!bo->shadow)
2891		return 0;
2892
2893	r = amdgpu_bo_reserve(bo, true);
2894	if (r)
2895		return r;
2896	domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
2897	/* if bo has been evicted, then no need to recover */
2898	if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
2899		r = amdgpu_bo_validate(bo->shadow);
2900		if (r) {
2901			DRM_ERROR("bo validate failed!\n");
2902			goto err;
2903		}
2904
2905		r = amdgpu_bo_restore_from_shadow(adev, ring, bo,
2906						 NULL, fence, true);
2907		if (r) {
2908			DRM_ERROR("recover page table failed!\n");
2909			goto err;
2910		}
2911	}
2912err:
2913	amdgpu_bo_unreserve(bo);
2914	return r;
2915}
2916
2917/**
2918 * amdgpu_device_handle_vram_lost - Handle the loss of VRAM contents
2919 *
2920 * @adev: amdgpu_device pointer
2921 *
2922 * Restores the contents of VRAM buffers from the shadows in GTT.  Used to
2923 * restore things like GPUVM page tables after a GPU reset where
2924 * the contents of VRAM might be lost.
2925 * Returns 0 on success, 1 on failure.
 
 
2926 */
2927static int amdgpu_device_handle_vram_lost(struct amdgpu_device *adev)
2928{
2929	struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
2930	struct amdgpu_bo *bo, *tmp;
2931	struct dma_fence *fence = NULL, *next = NULL;
2932	long r = 1;
2933	int i = 0;
2934	long tmo;
2935
2936	if (amdgpu_sriov_runtime(adev))
2937		tmo = msecs_to_jiffies(amdgpu_lockup_timeout);
2938	else
2939		tmo = msecs_to_jiffies(100);
2940
2941	DRM_INFO("recover vram bo from shadow start\n");
2942	mutex_lock(&adev->shadow_list_lock);
2943	list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) {
2944		next = NULL;
2945		amdgpu_device_recover_vram_from_shadow(adev, ring, bo, &next);
 
 
 
 
 
 
 
 
 
2946		if (fence) {
2947			r = dma_fence_wait_timeout(fence, false, tmo);
2948			if (r == 0)
2949				pr_err("wait fence %p[%d] timeout\n", fence, i);
2950			else if (r < 0)
2951				pr_err("wait fence %p[%d] interrupted\n", fence, i);
2952			if (r < 1) {
2953				dma_fence_put(fence);
2954				fence = next;
2955				break;
2956			}
2957			i++;
 
2958		}
2959
2960		dma_fence_put(fence);
2961		fence = next;
2962	}
2963	mutex_unlock(&adev->shadow_list_lock);
2964
2965	if (fence) {
2966		r = dma_fence_wait_timeout(fence, false, tmo);
2967		if (r == 0)
2968			pr_err("wait fence %p[%d] timeout\n", fence, i);
2969		else if (r < 0)
2970			pr_err("wait fence %p[%d] interrupted\n", fence, i);
2971
2972	}
2973	dma_fence_put(fence);
2974
2975	if (r > 0)
2976		DRM_INFO("recover vram bo from shadow done\n");
2977	else
2978		DRM_ERROR("recover vram bo from shadow failed\n");
2979
2980	return (r > 0) ? 0 : 1;
2981}
2982
2983/**
2984 * amdgpu_device_reset - reset ASIC/GPU for bare-metal or passthrough
2985 *
2986 * @adev: amdgpu device pointer
2987 *
2988 * attempt to do soft-reset or full-reset and reinitialize Asic
2989 * return 0 means successed otherwise failed
2990 */
2991static int amdgpu_device_reset(struct amdgpu_device *adev)
2992{
2993	bool need_full_reset, vram_lost = 0;
2994	int r;
2995
2996	need_full_reset = amdgpu_device_ip_need_full_reset(adev);
2997
2998	if (!need_full_reset) {
2999		amdgpu_device_ip_pre_soft_reset(adev);
3000		r = amdgpu_device_ip_soft_reset(adev);
3001		amdgpu_device_ip_post_soft_reset(adev);
3002		if (r || amdgpu_device_ip_check_soft_reset(adev)) {
3003			DRM_INFO("soft reset failed, will fallback to full reset!\n");
3004			need_full_reset = true;
3005		}
3006	}
3007
3008	if (need_full_reset) {
3009		r = amdgpu_device_ip_suspend(adev);
3010
3011retry:
3012		r = amdgpu_asic_reset(adev);
3013		/* post card */
3014		amdgpu_atom_asic_init(adev->mode_info.atom_context);
3015
3016		if (!r) {
3017			dev_info(adev->dev, "GPU reset succeeded, trying to resume\n");
3018			r = amdgpu_device_ip_resume_phase1(adev);
3019			if (r)
3020				goto out;
3021
3022			vram_lost = amdgpu_device_check_vram_lost(adev);
3023			if (vram_lost) {
3024				DRM_ERROR("VRAM is lost!\n");
3025				atomic_inc(&adev->vram_lost_counter);
3026			}
3027
3028			r = amdgpu_gtt_mgr_recover(
3029				&adev->mman.bdev.man[TTM_PL_TT]);
3030			if (r)
3031				goto out;
3032
3033			r = amdgpu_device_ip_resume_phase2(adev);
3034			if (r)
3035				goto out;
3036
3037			if (vram_lost)
3038				amdgpu_device_fill_reset_magic(adev);
3039		}
3040	}
3041
3042out:
3043	if (!r) {
3044		amdgpu_irq_gpu_reset_resume_helper(adev);
3045		r = amdgpu_ib_ring_tests(adev);
3046		if (r) {
3047			dev_err(adev->dev, "ib ring test failed (%d).\n", r);
3048			r = amdgpu_device_ip_suspend(adev);
3049			need_full_reset = true;
3050			goto retry;
3051		}
3052	}
3053
3054	if (!r && ((need_full_reset && !(adev->flags & AMD_IS_APU)) || vram_lost))
3055		r = amdgpu_device_handle_vram_lost(adev);
3056
3057	return r;
3058}
3059
 
3060/**
3061 * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
3062 *
3063 * @adev: amdgpu device pointer
 
3064 *
3065 * do VF FLR and reinitialize Asic
3066 * return 0 means successed otherwise failed
3067 */
3068static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
3069				     bool from_hypervisor)
3070{
3071	int r;
3072
3073	if (from_hypervisor)
3074		r = amdgpu_virt_request_full_gpu(adev, true);
3075	else
3076		r = amdgpu_virt_reset_gpu(adev);
3077	if (r)
3078		return r;
3079
 
 
3080	/* Resume IP prior to SMC */
3081	r = amdgpu_device_ip_reinit_early_sriov(adev);
3082	if (r)
3083		goto error;
3084
 
3085	/* we need recover gart prior to run SMC/CP/SDMA resume */
3086	amdgpu_gtt_mgr_recover(&adev->mman.bdev.man[TTM_PL_TT]);
 
 
 
 
3087
3088	/* now we are okay to resume SMC/CP/SDMA */
3089	r = amdgpu_device_ip_reinit_late_sriov(adev);
3090	amdgpu_virt_release_full_gpu(adev, true);
3091	if (r)
3092		goto error;
3093
3094	amdgpu_irq_gpu_reset_resume_helper(adev);
3095	r = amdgpu_ib_ring_tests(adev);
 
3096
 
3097	if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
3098		atomic_inc(&adev->vram_lost_counter);
3099		r = amdgpu_device_handle_vram_lost(adev);
3100	}
3101
3102error:
3103
3104	return r;
3105}
3106
3107/**
3108 * amdgpu_device_gpu_recover - reset the asic and recover scheduler
3109 *
3110 * @adev: amdgpu device pointer
3111 * @job: which job trigger hang
3112 * @force forces reset regardless of amdgpu_gpu_recovery
3113 *
3114 * Attempt to reset the GPU if it has hung (all asics).
3115 * Returns 0 for success or an error on failure.
3116 */
3117int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
3118			      struct amdgpu_job *job, bool force)
3119{
3120	struct drm_atomic_state *state = NULL;
3121	int i, r, resched;
3122
3123	if (!force && !amdgpu_device_ip_check_soft_reset(adev)) {
3124		DRM_INFO("No hardware hang detected. Did some blocks stall?\n");
3125		return 0;
 
 
 
 
 
 
 
 
 
3126	}
 
 
3127
3128	if (!force && (amdgpu_gpu_recovery == 0 ||
3129			(amdgpu_gpu_recovery == -1  && !amdgpu_sriov_vf(adev)))) {
3130		DRM_INFO("GPU recovery disabled.\n");
3131		return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3132	}
3133
3134	dev_info(adev->dev, "GPU reset begin!\n");
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3135
3136	mutex_lock(&adev->lock_reset);
3137	atomic_inc(&adev->gpu_reset_counter);
3138	adev->in_gpu_reset = 1;
3139
3140	/* block TTM */
3141	resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
3142
3143	/* store modesetting */
3144	if (amdgpu_device_has_dc_support(adev))
3145		state = drm_atomic_helper_suspend(adev->ddev);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3146
3147	/* block all schedulers and reset given job's ring */
3148	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3149		struct amdgpu_ring *ring = adev->rings[i];
3150
3151		if (!ring || !ring->sched.thread)
3152			continue;
3153
3154		kthread_park(ring->sched.thread);
 
 
3155
3156		if (job && job->ring->idx != i)
3157			continue;
3158
3159		drm_sched_hw_job_reset(&ring->sched, &job->base);
 
 
 
 
 
3160
3161		/* after all hw jobs are reset, hw fence is meaningless, so force_completion */
3162		amdgpu_fence_driver_force_completion(ring);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3163	}
3164
3165	if (amdgpu_sriov_vf(adev))
3166		r = amdgpu_device_reset_sriov(adev, job ? false : true);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3167	else
3168		r = amdgpu_device_reset(adev);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3169
3170	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3171		struct amdgpu_ring *ring = adev->rings[i];
 
 
3172
3173		if (!ring || !ring->sched.thread)
3174			continue;
3175
3176		/* only need recovery sched of the given job's ring
3177		 * or all rings (in the case @job is NULL)
3178		 * after above amdgpu_reset accomplished
3179		 */
3180		if ((!job || job->ring->idx == i) && !r)
3181			drm_sched_job_recovery(&ring->sched);
 
 
 
 
 
 
 
3182
3183		kthread_unpark(ring->sched.thread);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3184	}
 
3185
3186	if (amdgpu_device_has_dc_support(adev)) {
3187		if (drm_atomic_helper_resume(adev->ddev, state))
3188			dev_info(adev->dev, "drm resume failed:%d\n", r);
3189	} else {
3190		drm_helper_resume_force_mode(adev->ddev);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3191	}
3192
3193	ttm_bo_unlock_delayed_workqueue(&adev->mman.bdev, resched);
 
3194
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3195	if (r) {
3196		/* bad news, how to tell it to userspace ? */
3197		dev_info(adev->dev, "GPU reset(%d) failed\n", atomic_read(&adev->gpu_reset_counter));
3198		amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3199	} else {
3200		dev_info(adev->dev, "GPU reset(%d) successed!\n",atomic_read(&adev->gpu_reset_counter));
 
3201	}
3202
3203	amdgpu_vf_error_trans_all(adev);
3204	adev->in_gpu_reset = 0;
3205	mutex_unlock(&adev->lock_reset);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3206	return r;
3207}
3208
3209/**
3210 * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot
3211 *
3212 * @adev: amdgpu_device pointer
3213 *
3214 * Fetchs and stores in the driver the PCIE capabilities (gen speed
3215 * and lanes) of the slot the device is in. Handles APUs and
3216 * virtualized environments where PCIE config space may not be available.
3217 */
3218static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev)
3219{
3220	u32 mask;
3221	int ret;
 
3222
3223	if (amdgpu_pcie_gen_cap)
3224		adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
3225
3226	if (amdgpu_pcie_lane_cap)
3227		adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
3228
3229	/* covers APUs as well */
3230	if (pci_is_root_bus(adev->pdev->bus)) {
3231		if (adev->pm.pcie_gen_mask == 0)
3232			adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
3233		if (adev->pm.pcie_mlw_mask == 0)
3234			adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
3235		return;
3236	}
3237
 
 
 
 
 
 
3238	if (adev->pm.pcie_gen_mask == 0) {
3239		ret = drm_pcie_get_speed_cap_mask(adev->ddev, &mask);
3240		if (!ret) {
3241			adev->pm.pcie_gen_mask = (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
 
 
3242						  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3243						  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
3244
3245			if (mask & DRM_PCIE_SPEED_25)
3246				adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
3247			if (mask & DRM_PCIE_SPEED_50)
3248				adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2;
3249			if (mask & DRM_PCIE_SPEED_80)
3250				adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3;
3251		} else {
3252			adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3253		}
3254	}
3255	if (adev->pm.pcie_mlw_mask == 0) {
3256		ret = drm_pcie_get_max_link_width(adev->ddev, &mask);
3257		if (!ret) {
3258			switch (mask) {
3259			case 32:
 
3260				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
3261							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
3262							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3263							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3264							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3265							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3266							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3267				break;
3268			case 16:
3269				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
3270							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3271							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3272							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3273							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3274							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3275				break;
3276			case 12:
3277				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3278							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3279							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3280							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3281							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3282				break;
3283			case 8:
3284				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3285							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3286							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3287							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3288				break;
3289			case 4:
3290				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3291							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3292							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3293				break;
3294			case 2:
3295				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3296							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3297				break;
3298			case 1:
3299				adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;
3300				break;
3301			default:
3302				break;
3303			}
3304		} else {
3305			adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
3306		}
3307	}
3308}
3309
v5.14.15
   1/*
   2 * Copyright 2008 Advanced Micro Devices, Inc.
   3 * Copyright 2008 Red Hat Inc.
   4 * Copyright 2009 Jerome Glisse.
   5 *
   6 * Permission is hereby granted, free of charge, to any person obtaining a
   7 * copy of this software and associated documentation files (the "Software"),
   8 * to deal in the Software without restriction, including without limitation
   9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10 * and/or sell copies of the Software, and to permit persons to whom the
  11 * Software is furnished to do so, subject to the following conditions:
  12 *
  13 * The above copyright notice and this permission notice shall be included in
  14 * all copies or substantial portions of the Software.
  15 *
  16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  22 * OTHER DEALINGS IN THE SOFTWARE.
  23 *
  24 * Authors: Dave Airlie
  25 *          Alex Deucher
  26 *          Jerome Glisse
  27 */
  28#include <linux/power_supply.h>
  29#include <linux/kthread.h>
  30#include <linux/module.h>
  31#include <linux/console.h>
  32#include <linux/slab.h>
  33
 
  34#include <drm/drm_atomic_helper.h>
  35#include <drm/drm_probe_helper.h>
  36#include <drm/amdgpu_drm.h>
  37#include <linux/vgaarb.h>
  38#include <linux/vga_switcheroo.h>
  39#include <linux/efi.h>
  40#include "amdgpu.h"
  41#include "amdgpu_trace.h"
  42#include "amdgpu_i2c.h"
  43#include "atom.h"
  44#include "amdgpu_atombios.h"
  45#include "amdgpu_atomfirmware.h"
  46#include "amd_pcie.h"
  47#ifdef CONFIG_DRM_AMDGPU_SI
  48#include "si.h"
  49#endif
  50#ifdef CONFIG_DRM_AMDGPU_CIK
  51#include "cik.h"
  52#endif
  53#include "vi.h"
  54#include "soc15.h"
  55#include "nv.h"
  56#include "bif/bif_4_1_d.h"
  57#include <linux/pci.h>
  58#include <linux/firmware.h>
  59#include "amdgpu_vf_error.h"
  60
  61#include "amdgpu_amdkfd.h"
  62#include "amdgpu_pm.h"
  63
  64#include "amdgpu_xgmi.h"
  65#include "amdgpu_ras.h"
  66#include "amdgpu_pmu.h"
  67#include "amdgpu_fru_eeprom.h"
  68#include "amdgpu_reset.h"
  69
  70#include <linux/suspend.h>
  71#include <drm/task_barrier.h>
  72#include <linux/pm_runtime.h>
  73
  74#include <drm/drm_drv.h>
  75
  76MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin");
  77MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin");
  78MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin");
  79MODULE_FIRMWARE("amdgpu/picasso_gpu_info.bin");
  80MODULE_FIRMWARE("amdgpu/raven2_gpu_info.bin");
  81MODULE_FIRMWARE("amdgpu/arcturus_gpu_info.bin");
  82MODULE_FIRMWARE("amdgpu/renoir_gpu_info.bin");
  83MODULE_FIRMWARE("amdgpu/navi10_gpu_info.bin");
  84MODULE_FIRMWARE("amdgpu/navi14_gpu_info.bin");
  85MODULE_FIRMWARE("amdgpu/navi12_gpu_info.bin");
  86MODULE_FIRMWARE("amdgpu/vangogh_gpu_info.bin");
  87MODULE_FIRMWARE("amdgpu/yellow_carp_gpu_info.bin");
  88
  89#define AMDGPU_RESUME_MS		2000
  90
  91const char *amdgpu_asic_name[] = {
  92	"TAHITI",
  93	"PITCAIRN",
  94	"VERDE",
  95	"OLAND",
  96	"HAINAN",
  97	"BONAIRE",
  98	"KAVERI",
  99	"KABINI",
 100	"HAWAII",
 101	"MULLINS",
 102	"TOPAZ",
 103	"TONGA",
 104	"FIJI",
 105	"CARRIZO",
 106	"STONEY",
 107	"POLARIS10",
 108	"POLARIS11",
 109	"POLARIS12",
 110	"VEGAM",
 111	"VEGA10",
 112	"VEGA12",
 113	"VEGA20",
 114	"RAVEN",
 115	"ARCTURUS",
 116	"RENOIR",
 117	"ALDEBARAN",
 118	"NAVI10",
 119	"NAVI14",
 120	"NAVI12",
 121	"SIENNA_CICHLID",
 122	"NAVY_FLOUNDER",
 123	"VANGOGH",
 124	"DIMGREY_CAVEFISH",
 125	"BEIGE_GOBY",
 126	"YELLOW_CARP",
 127	"LAST",
 128};
 129
 130/**
 131 * DOC: pcie_replay_count
 132 *
 133 * The amdgpu driver provides a sysfs API for reporting the total number
 134 * of PCIe replays (NAKs)
 135 * The file pcie_replay_count is used for this and returns the total
 136 * number of replays as a sum of the NAKs generated and NAKs received
 137 */
 138
 139static ssize_t amdgpu_device_get_pcie_replay_count(struct device *dev,
 140		struct device_attribute *attr, char *buf)
 141{
 142	struct drm_device *ddev = dev_get_drvdata(dev);
 143	struct amdgpu_device *adev = drm_to_adev(ddev);
 144	uint64_t cnt = amdgpu_asic_get_pcie_replay_count(adev);
 145
 146	return sysfs_emit(buf, "%llu\n", cnt);
 147}
 148
 149static DEVICE_ATTR(pcie_replay_count, S_IRUGO,
 150		amdgpu_device_get_pcie_replay_count, NULL);
 151
 152static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev);
 153
 154/**
 155 * DOC: product_name
 156 *
 157 * The amdgpu driver provides a sysfs API for reporting the product name
 158 * for the device
 159 * The file serial_number is used for this and returns the product name
 160 * as returned from the FRU.
 161 * NOTE: This is only available for certain server cards
 162 */
 163
 164static ssize_t amdgpu_device_get_product_name(struct device *dev,
 165		struct device_attribute *attr, char *buf)
 166{
 167	struct drm_device *ddev = dev_get_drvdata(dev);
 168	struct amdgpu_device *adev = drm_to_adev(ddev);
 169
 170	return sysfs_emit(buf, "%s\n", adev->product_name);
 171}
 172
 173static DEVICE_ATTR(product_name, S_IRUGO,
 174		amdgpu_device_get_product_name, NULL);
 175
 176/**
 177 * DOC: product_number
 178 *
 179 * The amdgpu driver provides a sysfs API for reporting the part number
 180 * for the device
 181 * The file serial_number is used for this and returns the part number
 182 * as returned from the FRU.
 183 * NOTE: This is only available for certain server cards
 184 */
 185
 186static ssize_t amdgpu_device_get_product_number(struct device *dev,
 187		struct device_attribute *attr, char *buf)
 188{
 189	struct drm_device *ddev = dev_get_drvdata(dev);
 190	struct amdgpu_device *adev = drm_to_adev(ddev);
 191
 192	return sysfs_emit(buf, "%s\n", adev->product_number);
 193}
 194
 195static DEVICE_ATTR(product_number, S_IRUGO,
 196		amdgpu_device_get_product_number, NULL);
 197
 198/**
 199 * DOC: serial_number
 200 *
 201 * The amdgpu driver provides a sysfs API for reporting the serial number
 202 * for the device
 203 * The file serial_number is used for this and returns the serial number
 204 * as returned from the FRU.
 205 * NOTE: This is only available for certain server cards
 206 */
 207
 208static ssize_t amdgpu_device_get_serial_number(struct device *dev,
 209		struct device_attribute *attr, char *buf)
 210{
 211	struct drm_device *ddev = dev_get_drvdata(dev);
 212	struct amdgpu_device *adev = drm_to_adev(ddev);
 213
 214	return sysfs_emit(buf, "%s\n", adev->serial);
 215}
 216
 217static DEVICE_ATTR(serial_number, S_IRUGO,
 218		amdgpu_device_get_serial_number, NULL);
 219
 220/**
 221 * amdgpu_device_supports_px - Is the device a dGPU with ATPX power control
 222 *
 223 * @dev: drm_device pointer
 224 *
 225 * Returns true if the device is a dGPU with ATPX power control,
 226 * otherwise return false.
 227 */
 228bool amdgpu_device_supports_px(struct drm_device *dev)
 229{
 230	struct amdgpu_device *adev = drm_to_adev(dev);
 231
 232	if ((adev->flags & AMD_IS_PX) && !amdgpu_is_atpx_hybrid())
 233		return true;
 234	return false;
 235}
 236
 237/**
 238 * amdgpu_device_supports_boco - Is the device a dGPU with ACPI power resources
 239 *
 240 * @dev: drm_device pointer
 241 *
 242 * Returns true if the device is a dGPU with ACPI power control,
 243 * otherwise return false.
 244 */
 245bool amdgpu_device_supports_boco(struct drm_device *dev)
 246{
 247	struct amdgpu_device *adev = drm_to_adev(dev);
 248
 249	if (adev->has_pr3 ||
 250	    ((adev->flags & AMD_IS_PX) && amdgpu_is_atpx_hybrid()))
 251		return true;
 252	return false;
 253}
 254
 255/**
 256 * amdgpu_device_supports_baco - Does the device support BACO
 257 *
 258 * @dev: drm_device pointer
 259 *
 260 * Returns true if the device supporte BACO,
 261 * otherwise return false.
 262 */
 263bool amdgpu_device_supports_baco(struct drm_device *dev)
 264{
 265	struct amdgpu_device *adev = drm_to_adev(dev);
 266
 267	return amdgpu_asic_supports_baco(adev);
 268}
 269
 270/**
 271 * amdgpu_device_supports_smart_shift - Is the device dGPU with
 272 * smart shift support
 273 *
 274 * @dev: drm_device pointer
 275 *
 276 * Returns true if the device is a dGPU with Smart Shift support,
 277 * otherwise returns false.
 278 */
 279bool amdgpu_device_supports_smart_shift(struct drm_device *dev)
 280{
 281	return (amdgpu_device_supports_boco(dev) &&
 282		amdgpu_acpi_is_power_shift_control_supported());
 283}
 284
 285/*
 286 * VRAM access helper functions
 287 */
 288
 289/**
 290 * amdgpu_device_vram_access - read/write a buffer in vram
 291 *
 292 * @adev: amdgpu_device pointer
 293 * @pos: offset of the buffer in vram
 294 * @buf: virtual address of the buffer in system memory
 295 * @size: read/write size, sizeof(@buf) must > @size
 296 * @write: true - write to vram, otherwise - read from vram
 297 */
 298void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos,
 299			       uint32_t *buf, size_t size, bool write)
 300{
 301	unsigned long flags;
 302	uint32_t hi = ~0;
 303	uint64_t last;
 304	int idx;
 305
 306	if (!drm_dev_enter(&adev->ddev, &idx))
 307		return;
 308
 309#ifdef CONFIG_64BIT
 310	last = min(pos + size, adev->gmc.visible_vram_size);
 311	if (last > pos) {
 312		void __iomem *addr = adev->mman.aper_base_kaddr + pos;
 313		size_t count = last - pos;
 314
 315		if (write) {
 316			memcpy_toio(addr, buf, count);
 317			mb();
 318			amdgpu_device_flush_hdp(adev, NULL);
 319		} else {
 320			amdgpu_device_invalidate_hdp(adev, NULL);
 321			mb();
 322			memcpy_fromio(buf, addr, count);
 323		}
 324
 325		if (count == size)
 326			goto exit;
 327
 328		pos += count;
 329		buf += count / 4;
 330		size -= count;
 331	}
 332#endif
 333
 334	spin_lock_irqsave(&adev->mmio_idx_lock, flags);
 335	for (last = pos + size; pos < last; pos += 4) {
 336		uint32_t tmp = pos >> 31;
 337
 338		WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)pos) | 0x80000000);
 339		if (tmp != hi) {
 340			WREG32_NO_KIQ(mmMM_INDEX_HI, tmp);
 341			hi = tmp;
 342		}
 343		if (write)
 344			WREG32_NO_KIQ(mmMM_DATA, *buf++);
 345		else
 346			*buf++ = RREG32_NO_KIQ(mmMM_DATA);
 347	}
 348	spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
 349
 350#ifdef CONFIG_64BIT
 351exit:
 352#endif
 353	drm_dev_exit(idx);
 354}
 355
 356/*
 357 * register access helper functions.
 358 */
 359
 360/* Check if hw access should be skipped because of hotplug or device error */
 361bool amdgpu_device_skip_hw_access(struct amdgpu_device *adev)
 362{
 363	if (adev->no_hw_access)
 364		return true;
 365
 366#ifdef CONFIG_LOCKDEP
 367	/*
 368	 * This is a bit complicated to understand, so worth a comment. What we assert
 369	 * here is that the GPU reset is not running on another thread in parallel.
 370	 *
 371	 * For this we trylock the read side of the reset semaphore, if that succeeds
 372	 * we know that the reset is not running in paralell.
 373	 *
 374	 * If the trylock fails we assert that we are either already holding the read
 375	 * side of the lock or are the reset thread itself and hold the write side of
 376	 * the lock.
 377	 */
 378	if (in_task()) {
 379		if (down_read_trylock(&adev->reset_sem))
 380			up_read(&adev->reset_sem);
 381		else
 382			lockdep_assert_held(&adev->reset_sem);
 383	}
 384#endif
 385	return false;
 386}
 387
 388/**
 389 * amdgpu_device_rreg - read a memory mapped IO or indirect register
 390 *
 391 * @adev: amdgpu_device pointer
 392 * @reg: dword aligned register offset
 393 * @acc_flags: access flags which require special behavior
 394 *
 395 * Returns the 32 bit value from the offset specified.
 396 */
 397uint32_t amdgpu_device_rreg(struct amdgpu_device *adev,
 398			    uint32_t reg, uint32_t acc_flags)
 399{
 400	uint32_t ret;
 401
 402	if (amdgpu_device_skip_hw_access(adev))
 403		return 0;
 404
 405	if ((reg * 4) < adev->rmmio_size) {
 406		if (!(acc_flags & AMDGPU_REGS_NO_KIQ) &&
 407		    amdgpu_sriov_runtime(adev) &&
 408		    down_read_trylock(&adev->reset_sem)) {
 409			ret = amdgpu_kiq_rreg(adev, reg);
 410			up_read(&adev->reset_sem);
 411		} else {
 412			ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
 413		}
 414	} else {
 415		ret = adev->pcie_rreg(adev, reg * 4);
 416	}
 417
 418	trace_amdgpu_device_rreg(adev->pdev->device, reg, ret);
 419
 420	return ret;
 421}
 422
 423/*
 424 * MMIO register read with bytes helper functions
 425 * @offset:bytes offset from MMIO start
 426 *
 427*/
 428
 429/**
 430 * amdgpu_mm_rreg8 - read a memory mapped IO register
 431 *
 432 * @adev: amdgpu_device pointer
 433 * @offset: byte aligned register offset
 434 *
 435 * Returns the 8 bit value from the offset specified.
 436 */
 437uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset)
 438{
 439	if (amdgpu_device_skip_hw_access(adev))
 440		return 0;
 441
 442	if (offset < adev->rmmio_size)
 443		return (readb(adev->rmmio + offset));
 444	BUG();
 445}
 446
 447/*
 448 * MMIO register write with bytes helper functions
 449 * @offset:bytes offset from MMIO start
 450 * @value: the value want to be written to the register
 451 *
 452*/
 453/**
 454 * amdgpu_mm_wreg8 - read a memory mapped IO register
 455 *
 456 * @adev: amdgpu_device pointer
 457 * @offset: byte aligned register offset
 458 * @value: 8 bit value to write
 459 *
 460 * Writes the value specified to the offset specified.
 461 */
 462void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value)
 463{
 464	if (amdgpu_device_skip_hw_access(adev))
 465		return;
 466
 467	if (offset < adev->rmmio_size)
 468		writeb(value, adev->rmmio + offset);
 469	else
 470		BUG();
 471}
 472
 473/**
 474 * amdgpu_device_wreg - write to a memory mapped IO or indirect register
 475 *
 476 * @adev: amdgpu_device pointer
 477 * @reg: dword aligned register offset
 478 * @v: 32 bit value to write to the register
 479 * @acc_flags: access flags which require special behavior
 480 *
 481 * Writes the value specified to the offset specified.
 482 */
 483void amdgpu_device_wreg(struct amdgpu_device *adev,
 484			uint32_t reg, uint32_t v,
 485			uint32_t acc_flags)
 486{
 487	if (amdgpu_device_skip_hw_access(adev))
 488		return;
 
 
 
 
 
 
 
 
 
 
 
 489
 490	if ((reg * 4) < adev->rmmio_size) {
 491		if (!(acc_flags & AMDGPU_REGS_NO_KIQ) &&
 492		    amdgpu_sriov_runtime(adev) &&
 493		    down_read_trylock(&adev->reset_sem)) {
 494			amdgpu_kiq_wreg(adev, reg, v);
 495			up_read(&adev->reset_sem);
 496		} else {
 497			writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
 498		}
 499	} else {
 500		adev->pcie_wreg(adev, reg * 4, v);
 501	}
 502
 503	trace_amdgpu_device_wreg(adev->pdev->device, reg, v);
 
 
 504}
 505
 506/*
 507 * amdgpu_mm_wreg_mmio_rlc -  write register either with mmio or with RLC path if in range
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 508 *
 509 * this function is invoked only the debugfs register access
 510 * */
 511void amdgpu_mm_wreg_mmio_rlc(struct amdgpu_device *adev,
 512			     uint32_t reg, uint32_t v)
 513{
 514	if (amdgpu_device_skip_hw_access(adev))
 515		return;
 
 
 
 
 
 
 
 
 516
 517	if (amdgpu_sriov_fullaccess(adev) &&
 518	    adev->gfx.rlc.funcs &&
 519	    adev->gfx.rlc.funcs->is_rlcg_access_range) {
 520		if (adev->gfx.rlc.funcs->is_rlcg_access_range(adev, reg))
 521			return adev->gfx.rlc.funcs->rlcg_wreg(adev, reg, v, 0, 0);
 522	} else {
 523		writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
 524	}
 525}
 526
 527/**
 528 * amdgpu_mm_rdoorbell - read a doorbell dword
 529 *
 530 * @adev: amdgpu_device pointer
 531 * @index: doorbell index
 532 *
 533 * Returns the value in the doorbell aperture at the
 534 * requested doorbell index (CIK).
 535 */
 536u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
 537{
 538	if (amdgpu_device_skip_hw_access(adev))
 539		return 0;
 540
 541	if (index < adev->doorbell.num_doorbells) {
 542		return readl(adev->doorbell.ptr + index);
 543	} else {
 544		DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
 545		return 0;
 546	}
 547}
 548
 549/**
 550 * amdgpu_mm_wdoorbell - write a doorbell dword
 551 *
 552 * @adev: amdgpu_device pointer
 553 * @index: doorbell index
 554 * @v: value to write
 555 *
 556 * Writes @v to the doorbell aperture at the
 557 * requested doorbell index (CIK).
 558 */
 559void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v)
 560{
 561	if (amdgpu_device_skip_hw_access(adev))
 562		return;
 563
 564	if (index < adev->doorbell.num_doorbells) {
 565		writel(v, adev->doorbell.ptr + index);
 566	} else {
 567		DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
 568	}
 569}
 570
 571/**
 572 * amdgpu_mm_rdoorbell64 - read a doorbell Qword
 573 *
 574 * @adev: amdgpu_device pointer
 575 * @index: doorbell index
 576 *
 577 * Returns the value in the doorbell aperture at the
 578 * requested doorbell index (VEGA10+).
 579 */
 580u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index)
 581{
 582	if (amdgpu_device_skip_hw_access(adev))
 583		return 0;
 584
 585	if (index < adev->doorbell.num_doorbells) {
 586		return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index));
 587	} else {
 588		DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
 589		return 0;
 590	}
 591}
 592
 593/**
 594 * amdgpu_mm_wdoorbell64 - write a doorbell Qword
 595 *
 596 * @adev: amdgpu_device pointer
 597 * @index: doorbell index
 598 * @v: value to write
 599 *
 600 * Writes @v to the doorbell aperture at the
 601 * requested doorbell index (VEGA10+).
 602 */
 603void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v)
 604{
 605	if (amdgpu_device_skip_hw_access(adev))
 606		return;
 607
 608	if (index < adev->doorbell.num_doorbells) {
 609		atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v);
 610	} else {
 611		DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
 612	}
 613}
 614
 615/**
 616 * amdgpu_device_indirect_rreg - read an indirect register
 617 *
 618 * @adev: amdgpu_device pointer
 619 * @pcie_index: mmio register offset
 620 * @pcie_data: mmio register offset
 621 * @reg_addr: indirect register address to read from
 622 *
 623 * Returns the value of indirect register @reg_addr
 624 */
 625u32 amdgpu_device_indirect_rreg(struct amdgpu_device *adev,
 626				u32 pcie_index, u32 pcie_data,
 627				u32 reg_addr)
 628{
 629	unsigned long flags;
 630	u32 r;
 631	void __iomem *pcie_index_offset;
 632	void __iomem *pcie_data_offset;
 633
 634	spin_lock_irqsave(&adev->pcie_idx_lock, flags);
 635	pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
 636	pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
 637
 638	writel(reg_addr, pcie_index_offset);
 639	readl(pcie_index_offset);
 640	r = readl(pcie_data_offset);
 641	spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
 642
 643	return r;
 644}
 645
 646/**
 647 * amdgpu_device_indirect_rreg64 - read a 64bits indirect register
 648 *
 649 * @adev: amdgpu_device pointer
 650 * @pcie_index: mmio register offset
 651 * @pcie_data: mmio register offset
 652 * @reg_addr: indirect register address to read from
 653 *
 654 * Returns the value of indirect register @reg_addr
 655 */
 656u64 amdgpu_device_indirect_rreg64(struct amdgpu_device *adev,
 657				  u32 pcie_index, u32 pcie_data,
 658				  u32 reg_addr)
 659{
 660	unsigned long flags;
 661	u64 r;
 662	void __iomem *pcie_index_offset;
 663	void __iomem *pcie_data_offset;
 664
 665	spin_lock_irqsave(&adev->pcie_idx_lock, flags);
 666	pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
 667	pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
 668
 669	/* read low 32 bits */
 670	writel(reg_addr, pcie_index_offset);
 671	readl(pcie_index_offset);
 672	r = readl(pcie_data_offset);
 673	/* read high 32 bits */
 674	writel(reg_addr + 4, pcie_index_offset);
 675	readl(pcie_index_offset);
 676	r |= ((u64)readl(pcie_data_offset) << 32);
 677	spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
 678
 679	return r;
 680}
 681
 682/**
 683 * amdgpu_device_indirect_wreg - write an indirect register address
 684 *
 685 * @adev: amdgpu_device pointer
 686 * @pcie_index: mmio register offset
 687 * @pcie_data: mmio register offset
 688 * @reg_addr: indirect register offset
 689 * @reg_data: indirect register data
 690 *
 691 */
 692void amdgpu_device_indirect_wreg(struct amdgpu_device *adev,
 693				 u32 pcie_index, u32 pcie_data,
 694				 u32 reg_addr, u32 reg_data)
 695{
 696	unsigned long flags;
 697	void __iomem *pcie_index_offset;
 698	void __iomem *pcie_data_offset;
 699
 700	spin_lock_irqsave(&adev->pcie_idx_lock, flags);
 701	pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
 702	pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
 703
 704	writel(reg_addr, pcie_index_offset);
 705	readl(pcie_index_offset);
 706	writel(reg_data, pcie_data_offset);
 707	readl(pcie_data_offset);
 708	spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
 709}
 710
 711/**
 712 * amdgpu_device_indirect_wreg64 - write a 64bits indirect register address
 713 *
 714 * @adev: amdgpu_device pointer
 715 * @pcie_index: mmio register offset
 716 * @pcie_data: mmio register offset
 717 * @reg_addr: indirect register offset
 718 * @reg_data: indirect register data
 719 *
 720 */
 721void amdgpu_device_indirect_wreg64(struct amdgpu_device *adev,
 722				   u32 pcie_index, u32 pcie_data,
 723				   u32 reg_addr, u64 reg_data)
 724{
 725	unsigned long flags;
 726	void __iomem *pcie_index_offset;
 727	void __iomem *pcie_data_offset;
 728
 729	spin_lock_irqsave(&adev->pcie_idx_lock, flags);
 730	pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
 731	pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
 732
 733	/* write low 32 bits */
 734	writel(reg_addr, pcie_index_offset);
 735	readl(pcie_index_offset);
 736	writel((u32)(reg_data & 0xffffffffULL), pcie_data_offset);
 737	readl(pcie_data_offset);
 738	/* write high 32 bits */
 739	writel(reg_addr + 4, pcie_index_offset);
 740	readl(pcie_index_offset);
 741	writel((u32)(reg_data >> 32), pcie_data_offset);
 742	readl(pcie_data_offset);
 743	spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
 744}
 745
 746/**
 747 * amdgpu_invalid_rreg - dummy reg read function
 748 *
 749 * @adev: amdgpu_device pointer
 750 * @reg: offset of register
 751 *
 752 * Dummy register read function.  Used for register blocks
 753 * that certain asics don't have (all asics).
 754 * Returns the value in the register.
 755 */
 756static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
 757{
 758	DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
 759	BUG();
 760	return 0;
 761}
 762
 763/**
 764 * amdgpu_invalid_wreg - dummy reg write function
 765 *
 766 * @adev: amdgpu_device pointer
 767 * @reg: offset of register
 768 * @v: value to write to the register
 769 *
 770 * Dummy register read function.  Used for register blocks
 771 * that certain asics don't have (all asics).
 772 */
 773static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
 774{
 775	DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
 776		  reg, v);
 777	BUG();
 778}
 779
 780/**
 781 * amdgpu_invalid_rreg64 - dummy 64 bit reg read function
 782 *
 783 * @adev: amdgpu_device pointer
 784 * @reg: offset of register
 785 *
 786 * Dummy register read function.  Used for register blocks
 787 * that certain asics don't have (all asics).
 788 * Returns the value in the register.
 789 */
 790static uint64_t amdgpu_invalid_rreg64(struct amdgpu_device *adev, uint32_t reg)
 791{
 792	DRM_ERROR("Invalid callback to read 64 bit register 0x%04X\n", reg);
 793	BUG();
 794	return 0;
 795}
 796
 797/**
 798 * amdgpu_invalid_wreg64 - dummy reg write function
 799 *
 800 * @adev: amdgpu_device pointer
 801 * @reg: offset of register
 802 * @v: value to write to the register
 803 *
 804 * Dummy register read function.  Used for register blocks
 805 * that certain asics don't have (all asics).
 806 */
 807static void amdgpu_invalid_wreg64(struct amdgpu_device *adev, uint32_t reg, uint64_t v)
 808{
 809	DRM_ERROR("Invalid callback to write 64 bit register 0x%04X with 0x%08llX\n",
 810		  reg, v);
 811	BUG();
 812}
 813
 814/**
 815 * amdgpu_block_invalid_rreg - dummy reg read function
 816 *
 817 * @adev: amdgpu_device pointer
 818 * @block: offset of instance
 819 * @reg: offset of register
 820 *
 821 * Dummy register read function.  Used for register blocks
 822 * that certain asics don't have (all asics).
 823 * Returns the value in the register.
 824 */
 825static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
 826					  uint32_t block, uint32_t reg)
 827{
 828	DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
 829		  reg, block);
 830	BUG();
 831	return 0;
 832}
 833
 834/**
 835 * amdgpu_block_invalid_wreg - dummy reg write function
 836 *
 837 * @adev: amdgpu_device pointer
 838 * @block: offset of instance
 839 * @reg: offset of register
 840 * @v: value to write to the register
 841 *
 842 * Dummy register read function.  Used for register blocks
 843 * that certain asics don't have (all asics).
 844 */
 845static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
 846				      uint32_t block,
 847				      uint32_t reg, uint32_t v)
 848{
 849	DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
 850		  reg, block, v);
 851	BUG();
 852}
 853
 854/**
 855 * amdgpu_device_asic_init - Wrapper for atom asic_init
 856 *
 857 * @adev: amdgpu_device pointer
 858 *
 859 * Does any asic specific work and then calls atom asic init.
 860 */
 861static int amdgpu_device_asic_init(struct amdgpu_device *adev)
 862{
 863	amdgpu_asic_pre_asic_init(adev);
 864
 865	return amdgpu_atom_asic_init(adev->mode_info.atom_context);
 866}
 867
 868/**
 869 * amdgpu_device_vram_scratch_init - allocate the VRAM scratch page
 870 *
 871 * @adev: amdgpu_device pointer
 872 *
 873 * Allocates a scratch page of VRAM for use by various things in the
 874 * driver.
 875 */
 876static int amdgpu_device_vram_scratch_init(struct amdgpu_device *adev)
 877{
 878	return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE,
 879				       PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
 880				       &adev->vram_scratch.robj,
 881				       &adev->vram_scratch.gpu_addr,
 882				       (void **)&adev->vram_scratch.ptr);
 883}
 884
 885/**
 886 * amdgpu_device_vram_scratch_fini - Free the VRAM scratch page
 887 *
 888 * @adev: amdgpu_device pointer
 889 *
 890 * Frees the VRAM scratch page.
 891 */
 892static void amdgpu_device_vram_scratch_fini(struct amdgpu_device *adev)
 893{
 894	amdgpu_bo_free_kernel(&adev->vram_scratch.robj, NULL, NULL);
 895}
 896
 897/**
 898 * amdgpu_device_program_register_sequence - program an array of registers.
 899 *
 900 * @adev: amdgpu_device pointer
 901 * @registers: pointer to the register array
 902 * @array_size: size of the register array
 903 *
 904 * Programs an array or registers with and and or masks.
 905 * This is a helper for setting golden registers.
 906 */
 907void amdgpu_device_program_register_sequence(struct amdgpu_device *adev,
 908					     const u32 *registers,
 909					     const u32 array_size)
 910{
 911	u32 tmp, reg, and_mask, or_mask;
 912	int i;
 913
 914	if (array_size % 3)
 915		return;
 916
 917	for (i = 0; i < array_size; i +=3) {
 918		reg = registers[i + 0];
 919		and_mask = registers[i + 1];
 920		or_mask = registers[i + 2];
 921
 922		if (and_mask == 0xffffffff) {
 923			tmp = or_mask;
 924		} else {
 925			tmp = RREG32(reg);
 926			tmp &= ~and_mask;
 927			if (adev->family >= AMDGPU_FAMILY_AI)
 928				tmp |= (or_mask & and_mask);
 929			else
 930				tmp |= or_mask;
 931		}
 932		WREG32(reg, tmp);
 933	}
 934}
 935
 936/**
 937 * amdgpu_device_pci_config_reset - reset the GPU
 938 *
 939 * @adev: amdgpu_device pointer
 940 *
 941 * Resets the GPU using the pci config reset sequence.
 942 * Only applicable to asics prior to vega10.
 943 */
 944void amdgpu_device_pci_config_reset(struct amdgpu_device *adev)
 945{
 946	pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
 947}
 948
 949/**
 950 * amdgpu_device_pci_reset - reset the GPU using generic PCI means
 951 *
 952 * @adev: amdgpu_device pointer
 953 *
 954 * Resets the GPU using generic pci reset interfaces (FLR, SBR, etc.).
 955 */
 956int amdgpu_device_pci_reset(struct amdgpu_device *adev)
 957{
 958	return pci_reset_function(adev->pdev);
 959}
 960
 961/*
 962 * GPU doorbell aperture helpers function.
 963 */
 964/**
 965 * amdgpu_device_doorbell_init - Init doorbell driver information.
 966 *
 967 * @adev: amdgpu_device pointer
 968 *
 969 * Init doorbell driver information (CIK)
 970 * Returns 0 on success, error on failure.
 971 */
 972static int amdgpu_device_doorbell_init(struct amdgpu_device *adev)
 973{
 974
 975	/* No doorbell on SI hardware generation */
 976	if (adev->asic_type < CHIP_BONAIRE) {
 977		adev->doorbell.base = 0;
 978		adev->doorbell.size = 0;
 979		adev->doorbell.num_doorbells = 0;
 980		adev->doorbell.ptr = NULL;
 981		return 0;
 982	}
 983
 984	if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET)
 985		return -EINVAL;
 986
 987	amdgpu_asic_init_doorbell_index(adev);
 988
 989	/* doorbell bar mapping */
 990	adev->doorbell.base = pci_resource_start(adev->pdev, 2);
 991	adev->doorbell.size = pci_resource_len(adev->pdev, 2);
 992
 993	adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32),
 994					     adev->doorbell_index.max_assignment+1);
 995	if (adev->doorbell.num_doorbells == 0)
 996		return -EINVAL;
 997
 998	/* For Vega, reserve and map two pages on doorbell BAR since SDMA
 999	 * paging queue doorbell use the second page. The
1000	 * AMDGPU_DOORBELL64_MAX_ASSIGNMENT definition assumes all the
1001	 * doorbells are in the first page. So with paging queue enabled,
1002	 * the max num_doorbells should + 1 page (0x400 in dword)
1003	 */
1004	if (adev->asic_type >= CHIP_VEGA10)
1005		adev->doorbell.num_doorbells += 0x400;
1006
1007	adev->doorbell.ptr = ioremap(adev->doorbell.base,
1008				     adev->doorbell.num_doorbells *
1009				     sizeof(u32));
1010	if (adev->doorbell.ptr == NULL)
1011		return -ENOMEM;
1012
1013	return 0;
1014}
1015
1016/**
1017 * amdgpu_device_doorbell_fini - Tear down doorbell driver information.
1018 *
1019 * @adev: amdgpu_device pointer
1020 *
1021 * Tear down doorbell driver information (CIK)
1022 */
1023static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev)
1024{
1025	iounmap(adev->doorbell.ptr);
1026	adev->doorbell.ptr = NULL;
1027}
1028
1029
1030
1031/*
1032 * amdgpu_device_wb_*()
1033 * Writeback is the method by which the GPU updates special pages in memory
1034 * with the status of certain GPU events (fences, ring pointers,etc.).
1035 */
1036
1037/**
1038 * amdgpu_device_wb_fini - Disable Writeback and free memory
1039 *
1040 * @adev: amdgpu_device pointer
1041 *
1042 * Disables Writeback and frees the Writeback memory (all asics).
1043 * Used at driver shutdown.
1044 */
1045static void amdgpu_device_wb_fini(struct amdgpu_device *adev)
1046{
1047	if (adev->wb.wb_obj) {
1048		amdgpu_bo_free_kernel(&adev->wb.wb_obj,
1049				      &adev->wb.gpu_addr,
1050				      (void **)&adev->wb.wb);
1051		adev->wb.wb_obj = NULL;
1052	}
1053}
1054
1055/**
1056 * amdgpu_device_wb_init- Init Writeback driver info and allocate memory
1057 *
1058 * @adev: amdgpu_device pointer
1059 *
1060 * Initializes writeback and allocates writeback memory (all asics).
1061 * Used at driver startup.
1062 * Returns 0 on success or an -error on failure.
1063 */
1064static int amdgpu_device_wb_init(struct amdgpu_device *adev)
1065{
1066	int r;
1067
1068	if (adev->wb.wb_obj == NULL) {
1069		/* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */
1070		r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8,
1071					    PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
1072					    &adev->wb.wb_obj, &adev->wb.gpu_addr,
1073					    (void **)&adev->wb.wb);
1074		if (r) {
1075			dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
1076			return r;
1077		}
1078
1079		adev->wb.num_wb = AMDGPU_MAX_WB;
1080		memset(&adev->wb.used, 0, sizeof(adev->wb.used));
1081
1082		/* clear wb memory */
1083		memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8);
1084	}
1085
1086	return 0;
1087}
1088
1089/**
1090 * amdgpu_device_wb_get - Allocate a wb entry
1091 *
1092 * @adev: amdgpu_device pointer
1093 * @wb: wb index
1094 *
1095 * Allocate a wb slot for use by the driver (all asics).
1096 * Returns 0 on success or -EINVAL on failure.
1097 */
1098int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb)
1099{
1100	unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
1101
1102	if (offset < adev->wb.num_wb) {
1103		__set_bit(offset, adev->wb.used);
1104		*wb = offset << 3; /* convert to dw offset */
1105		return 0;
1106	} else {
1107		return -EINVAL;
1108	}
1109}
1110
1111/**
1112 * amdgpu_device_wb_free - Free a wb entry
1113 *
1114 * @adev: amdgpu_device pointer
1115 * @wb: wb index
1116 *
1117 * Free a wb slot allocated for use by the driver (all asics)
1118 */
1119void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb)
1120{
1121	wb >>= 3;
1122	if (wb < adev->wb.num_wb)
1123		__clear_bit(wb, adev->wb.used);
1124}
1125
1126/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1127 * amdgpu_device_resize_fb_bar - try to resize FB BAR
1128 *
1129 * @adev: amdgpu_device pointer
1130 *
1131 * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not
1132 * to fail, but if any of the BARs is not accessible after the size we abort
1133 * driver loading by returning -ENODEV.
1134 */
1135int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev)
1136{
1137	int rbar_size = pci_rebar_bytes_to_size(adev->gmc.real_vram_size);
 
1138	struct pci_bus *root;
1139	struct resource *res;
1140	unsigned i;
1141	u16 cmd;
1142	int r;
1143
1144	/* Bypass for VF */
1145	if (amdgpu_sriov_vf(adev))
1146		return 0;
1147
1148	/* skip if the bios has already enabled large BAR */
1149	if (adev->gmc.real_vram_size &&
1150	    (pci_resource_len(adev->pdev, 0) >= adev->gmc.real_vram_size))
1151		return 0;
1152
1153	/* Check if the root BUS has 64bit memory resources */
1154	root = adev->pdev->bus;
1155	while (root->parent)
1156		root = root->parent;
1157
1158	pci_bus_for_each_resource(root, res, i) {
1159		if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) &&
1160		    res->start > 0x100000000ull)
1161			break;
1162	}
1163
1164	/* Trying to resize is pointless without a root hub window above 4GB */
1165	if (!res)
1166		return 0;
1167
1168	/* Limit the BAR size to what is available */
1169	rbar_size = min(fls(pci_rebar_get_possible_sizes(adev->pdev, 0)) - 1,
1170			rbar_size);
1171
1172	/* Disable memory decoding while we change the BAR addresses and size */
1173	pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd);
1174	pci_write_config_word(adev->pdev, PCI_COMMAND,
1175			      cmd & ~PCI_COMMAND_MEMORY);
1176
1177	/* Free the VRAM and doorbell BAR, we most likely need to move both. */
1178	amdgpu_device_doorbell_fini(adev);
1179	if (adev->asic_type >= CHIP_BONAIRE)
1180		pci_release_resource(adev->pdev, 2);
1181
1182	pci_release_resource(adev->pdev, 0);
1183
1184	r = pci_resize_resource(adev->pdev, 0, rbar_size);
1185	if (r == -ENOSPC)
1186		DRM_INFO("Not enough PCI address space for a large BAR.");
1187	else if (r && r != -ENOTSUPP)
1188		DRM_ERROR("Problem resizing BAR0 (%d).", r);
1189
1190	pci_assign_unassigned_bus_resources(adev->pdev->bus);
1191
1192	/* When the doorbell or fb BAR isn't available we have no chance of
1193	 * using the device.
1194	 */
1195	r = amdgpu_device_doorbell_init(adev);
1196	if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET))
1197		return -ENODEV;
1198
1199	pci_write_config_word(adev->pdev, PCI_COMMAND, cmd);
1200
1201	return 0;
1202}
1203
1204/*
1205 * GPU helpers function.
1206 */
1207/**
1208 * amdgpu_device_need_post - check if the hw need post or not
1209 *
1210 * @adev: amdgpu_device pointer
1211 *
1212 * Check if the asic has been initialized (all asics) at driver startup
1213 * or post is needed if  hw reset is performed.
1214 * Returns true if need or false if not.
1215 */
1216bool amdgpu_device_need_post(struct amdgpu_device *adev)
1217{
1218	uint32_t reg;
1219
1220	if (amdgpu_sriov_vf(adev))
1221		return false;
1222
1223	if (amdgpu_passthrough(adev)) {
1224		/* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
1225		 * some old smc fw still need driver do vPost otherwise gpu hang, while
1226		 * those smc fw version above 22.15 doesn't have this flaw, so we force
1227		 * vpost executed for smc version below 22.15
1228		 */
1229		if (adev->asic_type == CHIP_FIJI) {
1230			int err;
1231			uint32_t fw_ver;
1232			err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
1233			/* force vPost if error occured */
1234			if (err)
1235				return true;
1236
1237			fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
1238			if (fw_ver < 0x00160e00)
1239				return true;
1240		}
1241	}
1242
1243	/* Don't post if we need to reset whole hive on init */
1244	if (adev->gmc.xgmi.pending_reset)
1245		return false;
1246
1247	if (adev->has_hw_reset) {
1248		adev->has_hw_reset = false;
1249		return true;
1250	}
1251
1252	/* bios scratch used on CIK+ */
1253	if (adev->asic_type >= CHIP_BONAIRE)
1254		return amdgpu_atombios_scratch_need_asic_init(adev);
1255
1256	/* check MEM_SIZE for older asics */
1257	reg = amdgpu_asic_get_config_memsize(adev);
1258
1259	if ((reg != 0) && (reg != 0xffffffff))
1260		return false;
1261
1262	return true;
1263}
1264
1265/* if we get transitioned to only one device, take VGA back */
1266/**
1267 * amdgpu_device_vga_set_decode - enable/disable vga decode
1268 *
1269 * @cookie: amdgpu_device pointer
1270 * @state: enable/disable vga decode
1271 *
1272 * Enable/disable vga decode (all asics).
1273 * Returns VGA resource flags.
1274 */
1275static unsigned int amdgpu_device_vga_set_decode(void *cookie, bool state)
1276{
1277	struct amdgpu_device *adev = cookie;
1278	amdgpu_asic_set_vga_state(adev, state);
1279	if (state)
1280		return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
1281		       VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
1282	else
1283		return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
1284}
1285
1286/**
1287 * amdgpu_device_check_block_size - validate the vm block size
1288 *
1289 * @adev: amdgpu_device pointer
1290 *
1291 * Validates the vm block size specified via module parameter.
1292 * The vm block size defines number of bits in page table versus page directory,
1293 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
1294 * page table and the remaining bits are in the page directory.
1295 */
1296static void amdgpu_device_check_block_size(struct amdgpu_device *adev)
1297{
1298	/* defines number of bits in page table versus page directory,
1299	 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
1300	 * page table and the remaining bits are in the page directory */
1301	if (amdgpu_vm_block_size == -1)
1302		return;
1303
1304	if (amdgpu_vm_block_size < 9) {
1305		dev_warn(adev->dev, "VM page table size (%d) too small\n",
1306			 amdgpu_vm_block_size);
1307		amdgpu_vm_block_size = -1;
1308	}
1309}
1310
1311/**
1312 * amdgpu_device_check_vm_size - validate the vm size
1313 *
1314 * @adev: amdgpu_device pointer
1315 *
1316 * Validates the vm size in GB specified via module parameter.
1317 * The VM size is the size of the GPU virtual memory space in GB.
1318 */
1319static void amdgpu_device_check_vm_size(struct amdgpu_device *adev)
1320{
1321	/* no need to check the default value */
1322	if (amdgpu_vm_size == -1)
1323		return;
1324
1325	if (amdgpu_vm_size < 1) {
1326		dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
1327			 amdgpu_vm_size);
1328		amdgpu_vm_size = -1;
1329	}
1330}
1331
1332static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev)
1333{
1334	struct sysinfo si;
1335	bool is_os_64 = (sizeof(void *) == 8);
1336	uint64_t total_memory;
1337	uint64_t dram_size_seven_GB = 0x1B8000000;
1338	uint64_t dram_size_three_GB = 0xB8000000;
1339
1340	if (amdgpu_smu_memory_pool_size == 0)
1341		return;
1342
1343	if (!is_os_64) {
1344		DRM_WARN("Not 64-bit OS, feature not supported\n");
1345		goto def_value;
1346	}
1347	si_meminfo(&si);
1348	total_memory = (uint64_t)si.totalram * si.mem_unit;
1349
1350	if ((amdgpu_smu_memory_pool_size == 1) ||
1351		(amdgpu_smu_memory_pool_size == 2)) {
1352		if (total_memory < dram_size_three_GB)
1353			goto def_value1;
1354	} else if ((amdgpu_smu_memory_pool_size == 4) ||
1355		(amdgpu_smu_memory_pool_size == 8)) {
1356		if (total_memory < dram_size_seven_GB)
1357			goto def_value1;
1358	} else {
1359		DRM_WARN("Smu memory pool size not supported\n");
1360		goto def_value;
1361	}
1362	adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28;
1363
1364	return;
1365
1366def_value1:
1367	DRM_WARN("No enough system memory\n");
1368def_value:
1369	adev->pm.smu_prv_buffer_size = 0;
1370}
1371
1372static int amdgpu_device_init_apu_flags(struct amdgpu_device *adev)
1373{
1374	if (!(adev->flags & AMD_IS_APU) ||
1375	    adev->asic_type < CHIP_RAVEN)
1376		return 0;
1377
1378	switch (adev->asic_type) {
1379	case CHIP_RAVEN:
1380		if (adev->pdev->device == 0x15dd)
1381			adev->apu_flags |= AMD_APU_IS_RAVEN;
1382		if (adev->pdev->device == 0x15d8)
1383			adev->apu_flags |= AMD_APU_IS_PICASSO;
1384		break;
1385	case CHIP_RENOIR:
1386		if ((adev->pdev->device == 0x1636) ||
1387		    (adev->pdev->device == 0x164c))
1388			adev->apu_flags |= AMD_APU_IS_RENOIR;
1389		else
1390			adev->apu_flags |= AMD_APU_IS_GREEN_SARDINE;
1391		break;
1392	case CHIP_VANGOGH:
1393		adev->apu_flags |= AMD_APU_IS_VANGOGH;
1394		break;
1395	case CHIP_YELLOW_CARP:
1396		break;
1397	default:
1398		return -EINVAL;
1399	}
1400
1401	return 0;
1402}
1403
1404/**
1405 * amdgpu_device_check_arguments - validate module params
1406 *
1407 * @adev: amdgpu_device pointer
1408 *
1409 * Validates certain module parameters and updates
1410 * the associated values used by the driver (all asics).
1411 */
1412static int amdgpu_device_check_arguments(struct amdgpu_device *adev)
1413{
1414	if (amdgpu_sched_jobs < 4) {
1415		dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
1416			 amdgpu_sched_jobs);
1417		amdgpu_sched_jobs = 4;
1418	} else if (!is_power_of_2(amdgpu_sched_jobs)){
1419		dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
1420			 amdgpu_sched_jobs);
1421		amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
1422	}
1423
1424	if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) {
1425		/* gart size must be greater or equal to 32M */
1426		dev_warn(adev->dev, "gart size (%d) too small\n",
1427			 amdgpu_gart_size);
1428		amdgpu_gart_size = -1;
1429	}
1430
1431	if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) {
1432		/* gtt size must be greater or equal to 32M */
1433		dev_warn(adev->dev, "gtt size (%d) too small\n",
1434				 amdgpu_gtt_size);
1435		amdgpu_gtt_size = -1;
1436	}
1437
1438	/* valid range is between 4 and 9 inclusive */
1439	if (amdgpu_vm_fragment_size != -1 &&
1440	    (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) {
1441		dev_warn(adev->dev, "valid range is between 4 and 9\n");
1442		amdgpu_vm_fragment_size = -1;
1443	}
1444
1445	if (amdgpu_sched_hw_submission < 2) {
1446		dev_warn(adev->dev, "sched hw submission jobs (%d) must be at least 2\n",
1447			 amdgpu_sched_hw_submission);
1448		amdgpu_sched_hw_submission = 2;
1449	} else if (!is_power_of_2(amdgpu_sched_hw_submission)) {
1450		dev_warn(adev->dev, "sched hw submission jobs (%d) must be a power of 2\n",
1451			 amdgpu_sched_hw_submission);
1452		amdgpu_sched_hw_submission = roundup_pow_of_two(amdgpu_sched_hw_submission);
1453	}
1454
1455	amdgpu_device_check_smu_prv_buffer_size(adev);
1456
1457	amdgpu_device_check_vm_size(adev);
1458
1459	amdgpu_device_check_block_size(adev);
1460
1461	adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type);
 
 
 
 
 
1462
1463	amdgpu_gmc_tmz_set(adev);
 
 
 
1464
1465	amdgpu_gmc_noretry_set(adev);
1466
1467	return 0;
1468}
1469
1470/**
1471 * amdgpu_switcheroo_set_state - set switcheroo state
1472 *
1473 * @pdev: pci dev pointer
1474 * @state: vga_switcheroo state
1475 *
1476 * Callback for the switcheroo driver.  Suspends or resumes the
1477 * the asics before or after it is powered up using ACPI methods.
1478 */
1479static void amdgpu_switcheroo_set_state(struct pci_dev *pdev,
1480					enum vga_switcheroo_state state)
1481{
1482	struct drm_device *dev = pci_get_drvdata(pdev);
1483	int r;
1484
1485	if (amdgpu_device_supports_px(dev) && state == VGA_SWITCHEROO_OFF)
1486		return;
1487
1488	if (state == VGA_SWITCHEROO_ON) {
1489		pr_info("switched on\n");
1490		/* don't suspend or resume card normally */
1491		dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1492
1493		pci_set_power_state(pdev, PCI_D0);
1494		amdgpu_device_load_pci_state(pdev);
1495		r = pci_enable_device(pdev);
1496		if (r)
1497			DRM_WARN("pci_enable_device failed (%d)\n", r);
1498		amdgpu_device_resume(dev, true);
1499
1500		dev->switch_power_state = DRM_SWITCH_POWER_ON;
 
1501	} else {
1502		pr_info("switched off\n");
 
1503		dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1504		amdgpu_device_suspend(dev, true);
1505		amdgpu_device_cache_pci_state(pdev);
1506		/* Shut down the device */
1507		pci_disable_device(pdev);
1508		pci_set_power_state(pdev, PCI_D3cold);
1509		dev->switch_power_state = DRM_SWITCH_POWER_OFF;
1510	}
1511}
1512
1513/**
1514 * amdgpu_switcheroo_can_switch - see if switcheroo state can change
1515 *
1516 * @pdev: pci dev pointer
1517 *
1518 * Callback for the switcheroo driver.  Check of the switcheroo
1519 * state can be changed.
1520 * Returns true if the state can be changed, false if not.
1521 */
1522static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
1523{
1524	struct drm_device *dev = pci_get_drvdata(pdev);
1525
1526	/*
1527	* FIXME: open_count is protected by drm_global_mutex but that would lead to
1528	* locking inversion with the driver load path. And the access here is
1529	* completely racy anyway. So don't bother with locking for now.
1530	*/
1531	return atomic_read(&dev->open_count) == 0;
1532}
1533
1534static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
1535	.set_gpu_state = amdgpu_switcheroo_set_state,
1536	.reprobe = NULL,
1537	.can_switch = amdgpu_switcheroo_can_switch,
1538};
1539
1540/**
1541 * amdgpu_device_ip_set_clockgating_state - set the CG state
1542 *
1543 * @dev: amdgpu_device pointer
1544 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1545 * @state: clockgating state (gate or ungate)
1546 *
1547 * Sets the requested clockgating state for all instances of
1548 * the hardware IP specified.
1549 * Returns the error code from the last instance.
1550 */
1551int amdgpu_device_ip_set_clockgating_state(void *dev,
1552					   enum amd_ip_block_type block_type,
1553					   enum amd_clockgating_state state)
1554{
1555	struct amdgpu_device *adev = dev;
1556	int i, r = 0;
1557
1558	for (i = 0; i < adev->num_ip_blocks; i++) {
1559		if (!adev->ip_blocks[i].status.valid)
1560			continue;
1561		if (adev->ip_blocks[i].version->type != block_type)
1562			continue;
1563		if (!adev->ip_blocks[i].version->funcs->set_clockgating_state)
1564			continue;
1565		r = adev->ip_blocks[i].version->funcs->set_clockgating_state(
1566			(void *)adev, state);
1567		if (r)
1568			DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",
1569				  adev->ip_blocks[i].version->funcs->name, r);
1570	}
1571	return r;
1572}
1573
1574/**
1575 * amdgpu_device_ip_set_powergating_state - set the PG state
1576 *
1577 * @dev: amdgpu_device pointer
1578 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1579 * @state: powergating state (gate or ungate)
1580 *
1581 * Sets the requested powergating state for all instances of
1582 * the hardware IP specified.
1583 * Returns the error code from the last instance.
1584 */
1585int amdgpu_device_ip_set_powergating_state(void *dev,
1586					   enum amd_ip_block_type block_type,
1587					   enum amd_powergating_state state)
1588{
1589	struct amdgpu_device *adev = dev;
1590	int i, r = 0;
1591
1592	for (i = 0; i < adev->num_ip_blocks; i++) {
1593		if (!adev->ip_blocks[i].status.valid)
1594			continue;
1595		if (adev->ip_blocks[i].version->type != block_type)
1596			continue;
1597		if (!adev->ip_blocks[i].version->funcs->set_powergating_state)
1598			continue;
1599		r = adev->ip_blocks[i].version->funcs->set_powergating_state(
1600			(void *)adev, state);
1601		if (r)
1602			DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",
1603				  adev->ip_blocks[i].version->funcs->name, r);
1604	}
1605	return r;
1606}
1607
1608/**
1609 * amdgpu_device_ip_get_clockgating_state - get the CG state
1610 *
1611 * @adev: amdgpu_device pointer
1612 * @flags: clockgating feature flags
1613 *
1614 * Walks the list of IPs on the device and updates the clockgating
1615 * flags for each IP.
1616 * Updates @flags with the feature flags for each hardware IP where
1617 * clockgating is enabled.
1618 */
1619void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev,
1620					    u32 *flags)
1621{
1622	int i;
1623
1624	for (i = 0; i < adev->num_ip_blocks; i++) {
1625		if (!adev->ip_blocks[i].status.valid)
1626			continue;
1627		if (adev->ip_blocks[i].version->funcs->get_clockgating_state)
1628			adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags);
1629	}
1630}
1631
1632/**
1633 * amdgpu_device_ip_wait_for_idle - wait for idle
1634 *
1635 * @adev: amdgpu_device pointer
1636 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1637 *
1638 * Waits for the request hardware IP to be idle.
1639 * Returns 0 for success or a negative error code on failure.
1640 */
1641int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev,
1642				   enum amd_ip_block_type block_type)
1643{
1644	int i, r;
1645
1646	for (i = 0; i < adev->num_ip_blocks; i++) {
1647		if (!adev->ip_blocks[i].status.valid)
1648			continue;
1649		if (adev->ip_blocks[i].version->type == block_type) {
1650			r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
1651			if (r)
1652				return r;
1653			break;
1654		}
1655	}
1656	return 0;
1657
1658}
1659
1660/**
1661 * amdgpu_device_ip_is_idle - is the hardware IP idle
1662 *
1663 * @adev: amdgpu_device pointer
1664 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1665 *
1666 * Check if the hardware IP is idle or not.
1667 * Returns true if it the IP is idle, false if not.
1668 */
1669bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev,
1670			      enum amd_ip_block_type block_type)
1671{
1672	int i;
1673
1674	for (i = 0; i < adev->num_ip_blocks; i++) {
1675		if (!adev->ip_blocks[i].status.valid)
1676			continue;
1677		if (adev->ip_blocks[i].version->type == block_type)
1678			return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
1679	}
1680	return true;
1681
1682}
1683
1684/**
1685 * amdgpu_device_ip_get_ip_block - get a hw IP pointer
1686 *
1687 * @adev: amdgpu_device pointer
1688 * @type: Type of hardware IP (SMU, GFX, UVD, etc.)
1689 *
1690 * Returns a pointer to the hardware IP block structure
1691 * if it exists for the asic, otherwise NULL.
1692 */
1693struct amdgpu_ip_block *
1694amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,
1695			      enum amd_ip_block_type type)
1696{
1697	int i;
1698
1699	for (i = 0; i < adev->num_ip_blocks; i++)
1700		if (adev->ip_blocks[i].version->type == type)
1701			return &adev->ip_blocks[i];
1702
1703	return NULL;
1704}
1705
1706/**
1707 * amdgpu_device_ip_block_version_cmp
1708 *
1709 * @adev: amdgpu_device pointer
1710 * @type: enum amd_ip_block_type
1711 * @major: major version
1712 * @minor: minor version
1713 *
1714 * return 0 if equal or greater
1715 * return 1 if smaller or the ip_block doesn't exist
1716 */
1717int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev,
1718				       enum amd_ip_block_type type,
1719				       u32 major, u32 minor)
1720{
1721	struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type);
1722
1723	if (ip_block && ((ip_block->version->major > major) ||
1724			((ip_block->version->major == major) &&
1725			(ip_block->version->minor >= minor))))
1726		return 0;
1727
1728	return 1;
1729}
1730
1731/**
1732 * amdgpu_device_ip_block_add
1733 *
1734 * @adev: amdgpu_device pointer
1735 * @ip_block_version: pointer to the IP to add
1736 *
1737 * Adds the IP block driver information to the collection of IPs
1738 * on the asic.
1739 */
1740int amdgpu_device_ip_block_add(struct amdgpu_device *adev,
1741			       const struct amdgpu_ip_block_version *ip_block_version)
1742{
1743	if (!ip_block_version)
1744		return -EINVAL;
1745
1746	switch (ip_block_version->type) {
1747	case AMD_IP_BLOCK_TYPE_VCN:
1748		if (adev->harvest_ip_mask & AMD_HARVEST_IP_VCN_MASK)
1749			return 0;
1750		break;
1751	case AMD_IP_BLOCK_TYPE_JPEG:
1752		if (adev->harvest_ip_mask & AMD_HARVEST_IP_JPEG_MASK)
1753			return 0;
1754		break;
1755	default:
1756		break;
1757	}
1758
1759	DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks,
1760		  ip_block_version->funcs->name);
1761
1762	adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
1763
1764	return 0;
1765}
1766
1767/**
1768 * amdgpu_device_enable_virtual_display - enable virtual display feature
1769 *
1770 * @adev: amdgpu_device pointer
1771 *
1772 * Enabled the virtual display feature if the user has enabled it via
1773 * the module parameter virtual_display.  This feature provides a virtual
1774 * display hardware on headless boards or in virtualized environments.
1775 * This function parses and validates the configuration string specified by
1776 * the user and configues the virtual display configuration (number of
1777 * virtual connectors, crtcs, etc.) specified.
1778 */
1779static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
1780{
1781	adev->enable_virtual_display = false;
1782
1783	if (amdgpu_virtual_display) {
1784		const char *pci_address_name = pci_name(adev->pdev);
 
1785		char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
1786
1787		pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
1788		pciaddstr_tmp = pciaddstr;
1789		while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
1790			pciaddname = strsep(&pciaddname_tmp, ",");
1791			if (!strcmp("all", pciaddname)
1792			    || !strcmp(pci_address_name, pciaddname)) {
1793				long num_crtc;
1794				int res = -1;
1795
1796				adev->enable_virtual_display = true;
1797
1798				if (pciaddname_tmp)
1799					res = kstrtol(pciaddname_tmp, 10,
1800						      &num_crtc);
1801
1802				if (!res) {
1803					if (num_crtc < 1)
1804						num_crtc = 1;
1805					if (num_crtc > 6)
1806						num_crtc = 6;
1807					adev->mode_info.num_crtc = num_crtc;
1808				} else {
1809					adev->mode_info.num_crtc = 1;
1810				}
1811				break;
1812			}
1813		}
1814
1815		DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
1816			 amdgpu_virtual_display, pci_address_name,
1817			 adev->enable_virtual_display, adev->mode_info.num_crtc);
1818
1819		kfree(pciaddstr);
1820	}
1821}
1822
1823/**
1824 * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware
1825 *
1826 * @adev: amdgpu_device pointer
1827 *
1828 * Parses the asic configuration parameters specified in the gpu info
1829 * firmware and makes them availale to the driver for use in configuring
1830 * the asic.
1831 * Returns 0 on success, -EINVAL on failure.
1832 */
1833static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev)
1834{
1835	const char *chip_name;
1836	char fw_name[40];
1837	int err;
1838	const struct gpu_info_firmware_header_v1_0 *hdr;
1839
1840	adev->firmware.gpu_info_fw = NULL;
1841
1842	if (adev->mman.discovery_bin) {
1843		amdgpu_discovery_get_gfx_info(adev);
1844
1845		/*
1846		 * FIXME: The bounding box is still needed by Navi12, so
1847		 * temporarily read it from gpu_info firmware. Should be droped
1848		 * when DAL no longer needs it.
1849		 */
1850		if (adev->asic_type != CHIP_NAVI12)
1851			return 0;
1852	}
1853
1854	switch (adev->asic_type) {
 
 
 
 
 
 
 
 
1855#ifdef CONFIG_DRM_AMDGPU_SI
1856	case CHIP_VERDE:
1857	case CHIP_TAHITI:
1858	case CHIP_PITCAIRN:
1859	case CHIP_OLAND:
1860	case CHIP_HAINAN:
1861#endif
1862#ifdef CONFIG_DRM_AMDGPU_CIK
1863	case CHIP_BONAIRE:
1864	case CHIP_HAWAII:
1865	case CHIP_KAVERI:
1866	case CHIP_KABINI:
1867	case CHIP_MULLINS:
1868#endif
1869	case CHIP_TOPAZ:
1870	case CHIP_TONGA:
1871	case CHIP_FIJI:
1872	case CHIP_POLARIS10:
1873	case CHIP_POLARIS11:
1874	case CHIP_POLARIS12:
1875	case CHIP_VEGAM:
1876	case CHIP_CARRIZO:
1877	case CHIP_STONEY:
1878	case CHIP_VEGA20:
1879	case CHIP_ALDEBARAN:
1880	case CHIP_SIENNA_CICHLID:
1881	case CHIP_NAVY_FLOUNDER:
1882	case CHIP_DIMGREY_CAVEFISH:
1883	case CHIP_BEIGE_GOBY:
1884	default:
1885		return 0;
1886	case CHIP_VEGA10:
1887		chip_name = "vega10";
1888		break;
1889	case CHIP_VEGA12:
1890		chip_name = "vega12";
1891		break;
1892	case CHIP_RAVEN:
1893		if (adev->apu_flags & AMD_APU_IS_RAVEN2)
1894			chip_name = "raven2";
1895		else if (adev->apu_flags & AMD_APU_IS_PICASSO)
1896			chip_name = "picasso";
1897		else
1898			chip_name = "raven";
1899		break;
1900	case CHIP_ARCTURUS:
1901		chip_name = "arcturus";
1902		break;
1903	case CHIP_RENOIR:
1904		if (adev->apu_flags & AMD_APU_IS_RENOIR)
1905			chip_name = "renoir";
1906		else
1907			chip_name = "green_sardine";
1908		break;
1909	case CHIP_NAVI10:
1910		chip_name = "navi10";
1911		break;
1912	case CHIP_NAVI14:
1913		chip_name = "navi14";
1914		break;
1915	case CHIP_NAVI12:
1916		chip_name = "navi12";
1917		break;
1918	case CHIP_VANGOGH:
1919		chip_name = "vangogh";
1920		break;
1921	case CHIP_YELLOW_CARP:
1922		chip_name = "yellow_carp";
1923		break;
1924	}
1925
1926	snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name);
1927	err = request_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev);
1928	if (err) {
1929		dev_err(adev->dev,
1930			"Failed to load gpu_info firmware \"%s\"\n",
1931			fw_name);
1932		goto out;
1933	}
1934	err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw);
1935	if (err) {
1936		dev_err(adev->dev,
1937			"Failed to validate gpu_info firmware \"%s\"\n",
1938			fw_name);
1939		goto out;
1940	}
1941
1942	hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data;
1943	amdgpu_ucode_print_gpu_info_hdr(&hdr->header);
1944
1945	switch (hdr->version_major) {
1946	case 1:
1947	{
1948		const struct gpu_info_firmware_v1_0 *gpu_info_fw =
1949			(const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data +
1950								le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1951
1952		/*
1953		 * Should be droped when DAL no longer needs it.
1954		 */
1955		if (adev->asic_type == CHIP_NAVI12)
1956			goto parse_soc_bounding_box;
1957
1958		adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se);
1959		adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh);
1960		adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se);
1961		adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se);
1962		adev->gfx.config.max_texture_channel_caches =
1963			le32_to_cpu(gpu_info_fw->gc_num_tccs);
1964		adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs);
1965		adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds);
1966		adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth);
1967		adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth);
1968		adev->gfx.config.double_offchip_lds_buf =
1969			le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer);
1970		adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size);
1971		adev->gfx.cu_info.max_waves_per_simd =
1972			le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd);
1973		adev->gfx.cu_info.max_scratch_slots_per_cu =
1974			le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu);
1975		adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size);
1976		if (hdr->version_minor >= 1) {
1977			const struct gpu_info_firmware_v1_1 *gpu_info_fw =
1978				(const struct gpu_info_firmware_v1_1 *)(adev->firmware.gpu_info_fw->data +
1979									le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1980			adev->gfx.config.num_sc_per_sh =
1981				le32_to_cpu(gpu_info_fw->num_sc_per_sh);
1982			adev->gfx.config.num_packer_per_sc =
1983				le32_to_cpu(gpu_info_fw->num_packer_per_sc);
1984		}
1985
1986parse_soc_bounding_box:
1987		/*
1988		 * soc bounding box info is not integrated in disocovery table,
1989		 * we always need to parse it from gpu info firmware if needed.
1990		 */
1991		if (hdr->version_minor == 2) {
1992			const struct gpu_info_firmware_v1_2 *gpu_info_fw =
1993				(const struct gpu_info_firmware_v1_2 *)(adev->firmware.gpu_info_fw->data +
1994									le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1995			adev->dm.soc_bounding_box = &gpu_info_fw->soc_bounding_box;
1996		}
1997		break;
1998	}
1999	default:
2000		dev_err(adev->dev,
2001			"Unsupported gpu_info table %d\n", hdr->header.ucode_version);
2002		err = -EINVAL;
2003		goto out;
2004	}
2005out:
2006	return err;
2007}
2008
2009/**
2010 * amdgpu_device_ip_early_init - run early init for hardware IPs
2011 *
2012 * @adev: amdgpu_device pointer
2013 *
2014 * Early initialization pass for hardware IPs.  The hardware IPs that make
2015 * up each asic are discovered each IP's early_init callback is run.  This
2016 * is the first stage in initializing the asic.
2017 * Returns 0 on success, negative error code on failure.
2018 */
2019static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
2020{
2021	int i, r;
2022
2023	amdgpu_device_enable_virtual_display(adev);
2024
2025	if (amdgpu_sriov_vf(adev)) {
2026		r = amdgpu_virt_request_full_gpu(adev, true);
 
 
 
 
 
 
 
 
 
 
 
 
 
2027		if (r)
2028			return r;
2029	}
2030
2031	switch (adev->asic_type) {
2032#ifdef CONFIG_DRM_AMDGPU_SI
2033	case CHIP_VERDE:
2034	case CHIP_TAHITI:
2035	case CHIP_PITCAIRN:
2036	case CHIP_OLAND:
2037	case CHIP_HAINAN:
2038		adev->family = AMDGPU_FAMILY_SI;
2039		r = si_set_ip_blocks(adev);
2040		if (r)
2041			return r;
2042		break;
2043#endif
2044#ifdef CONFIG_DRM_AMDGPU_CIK
2045	case CHIP_BONAIRE:
2046	case CHIP_HAWAII:
2047	case CHIP_KAVERI:
2048	case CHIP_KABINI:
2049	case CHIP_MULLINS:
2050		if (adev->flags & AMD_IS_APU)
 
 
2051			adev->family = AMDGPU_FAMILY_KV;
2052		else
2053			adev->family = AMDGPU_FAMILY_CI;
2054
2055		r = cik_set_ip_blocks(adev);
2056		if (r)
2057			return r;
2058		break;
2059#endif
2060	case CHIP_TOPAZ:
2061	case CHIP_TONGA:
2062	case CHIP_FIJI:
2063	case CHIP_POLARIS10:
2064	case CHIP_POLARIS11:
2065	case CHIP_POLARIS12:
2066	case CHIP_VEGAM:
2067	case CHIP_CARRIZO:
2068	case CHIP_STONEY:
2069		if (adev->flags & AMD_IS_APU)
2070			adev->family = AMDGPU_FAMILY_CZ;
2071		else
2072			adev->family = AMDGPU_FAMILY_VI;
2073
2074		r = vi_set_ip_blocks(adev);
2075		if (r)
2076			return r;
2077		break;
2078	case CHIP_VEGA10:
2079	case CHIP_VEGA12:
2080	case CHIP_VEGA20:
2081	case CHIP_RAVEN:
2082	case CHIP_ARCTURUS:
2083	case CHIP_RENOIR:
2084	case CHIP_ALDEBARAN:
2085		if (adev->flags & AMD_IS_APU)
2086			adev->family = AMDGPU_FAMILY_RV;
2087		else
2088			adev->family = AMDGPU_FAMILY_AI;
2089
2090		r = soc15_set_ip_blocks(adev);
2091		if (r)
2092			return r;
2093		break;
2094	case  CHIP_NAVI10:
2095	case  CHIP_NAVI14:
2096	case  CHIP_NAVI12:
2097	case  CHIP_SIENNA_CICHLID:
2098	case  CHIP_NAVY_FLOUNDER:
2099	case  CHIP_DIMGREY_CAVEFISH:
2100	case  CHIP_BEIGE_GOBY:
2101	case CHIP_VANGOGH:
2102	case CHIP_YELLOW_CARP:
2103		if (adev->asic_type == CHIP_VANGOGH)
2104			adev->family = AMDGPU_FAMILY_VGH;
2105		else if (adev->asic_type == CHIP_YELLOW_CARP)
2106			adev->family = AMDGPU_FAMILY_YC;
2107		else
2108			adev->family = AMDGPU_FAMILY_NV;
2109
2110		r = nv_set_ip_blocks(adev);
2111		if (r)
2112			return r;
2113		break;
2114	default:
2115		/* FIXME: not supported yet */
2116		return -EINVAL;
2117	}
2118
 
 
 
 
2119	amdgpu_amdkfd_device_probe(adev);
2120
2121	adev->pm.pp_feature = amdgpu_pp_feature_mask;
2122	if (amdgpu_sriov_vf(adev) || sched_policy == KFD_SCHED_POLICY_NO_HWS)
2123		adev->pm.pp_feature &= ~PP_GFXOFF_MASK;
2124	if (amdgpu_sriov_vf(adev) && adev->asic_type == CHIP_SIENNA_CICHLID)
2125		adev->pm.pp_feature &= ~PP_OVERDRIVE_MASK;
2126
2127	for (i = 0; i < adev->num_ip_blocks; i++) {
2128		if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
2129			DRM_ERROR("disabled ip block: %d <%s>\n",
2130				  i, adev->ip_blocks[i].version->funcs->name);
2131			adev->ip_blocks[i].status.valid = false;
2132		} else {
2133			if (adev->ip_blocks[i].version->funcs->early_init) {
2134				r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
2135				if (r == -ENOENT) {
2136					adev->ip_blocks[i].status.valid = false;
2137				} else if (r) {
2138					DRM_ERROR("early_init of IP block <%s> failed %d\n",
2139						  adev->ip_blocks[i].version->funcs->name, r);
2140					return r;
2141				} else {
2142					adev->ip_blocks[i].status.valid = true;
2143				}
2144			} else {
2145				adev->ip_blocks[i].status.valid = true;
2146			}
2147		}
2148		/* get the vbios after the asic_funcs are set up */
2149		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) {
2150			r = amdgpu_device_parse_gpu_info_fw(adev);
2151			if (r)
2152				return r;
2153
2154			/* Read BIOS */
2155			if (!amdgpu_get_bios(adev))
2156				return -EINVAL;
2157
2158			r = amdgpu_atombios_init(adev);
2159			if (r) {
2160				dev_err(adev->dev, "amdgpu_atombios_init failed\n");
2161				amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0);
2162				return r;
2163			}
2164
2165			/*get pf2vf msg info at it's earliest time*/
2166			if (amdgpu_sriov_vf(adev))
2167				amdgpu_virt_init_data_exchange(adev);
2168
2169		}
2170	}
2171
2172	adev->cg_flags &= amdgpu_cg_mask;
2173	adev->pg_flags &= amdgpu_pg_mask;
2174
2175	return 0;
2176}
2177
2178static int amdgpu_device_ip_hw_init_phase1(struct amdgpu_device *adev)
2179{
2180	int i, r;
2181
2182	for (i = 0; i < adev->num_ip_blocks; i++) {
2183		if (!adev->ip_blocks[i].status.sw)
2184			continue;
2185		if (adev->ip_blocks[i].status.hw)
2186			continue;
2187		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2188		    (amdgpu_sriov_vf(adev) && (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)) ||
2189		    adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
2190			r = adev->ip_blocks[i].version->funcs->hw_init(adev);
2191			if (r) {
2192				DRM_ERROR("hw_init of IP block <%s> failed %d\n",
2193					  adev->ip_blocks[i].version->funcs->name, r);
2194				return r;
2195			}
2196			adev->ip_blocks[i].status.hw = true;
2197		}
2198	}
2199
2200	return 0;
2201}
2202
2203static int amdgpu_device_ip_hw_init_phase2(struct amdgpu_device *adev)
2204{
2205	int i, r;
2206
2207	for (i = 0; i < adev->num_ip_blocks; i++) {
2208		if (!adev->ip_blocks[i].status.sw)
2209			continue;
2210		if (adev->ip_blocks[i].status.hw)
2211			continue;
2212		r = adev->ip_blocks[i].version->funcs->hw_init(adev);
2213		if (r) {
2214			DRM_ERROR("hw_init of IP block <%s> failed %d\n",
2215				  adev->ip_blocks[i].version->funcs->name, r);
2216			return r;
2217		}
2218		adev->ip_blocks[i].status.hw = true;
2219	}
2220
2221	return 0;
2222}
2223
2224static int amdgpu_device_fw_loading(struct amdgpu_device *adev)
2225{
2226	int r = 0;
2227	int i;
2228	uint32_t smu_version;
2229
2230	if (adev->asic_type >= CHIP_VEGA10) {
2231		for (i = 0; i < adev->num_ip_blocks; i++) {
2232			if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_PSP)
2233				continue;
2234
2235			if (!adev->ip_blocks[i].status.sw)
2236				continue;
2237
2238			/* no need to do the fw loading again if already done*/
2239			if (adev->ip_blocks[i].status.hw == true)
2240				break;
2241
2242			if (amdgpu_in_reset(adev) || adev->in_suspend) {
2243				r = adev->ip_blocks[i].version->funcs->resume(adev);
2244				if (r) {
2245					DRM_ERROR("resume of IP block <%s> failed %d\n",
2246							  adev->ip_blocks[i].version->funcs->name, r);
2247					return r;
2248				}
2249			} else {
2250				r = adev->ip_blocks[i].version->funcs->hw_init(adev);
2251				if (r) {
2252					DRM_ERROR("hw_init of IP block <%s> failed %d\n",
2253							  adev->ip_blocks[i].version->funcs->name, r);
2254					return r;
2255				}
2256			}
2257
2258			adev->ip_blocks[i].status.hw = true;
2259			break;
2260		}
2261	}
2262
2263	if (!amdgpu_sriov_vf(adev) || adev->asic_type == CHIP_TONGA)
2264		r = amdgpu_pm_load_smu_firmware(adev, &smu_version);
2265
2266	return r;
2267}
2268
2269/**
2270 * amdgpu_device_ip_init - run init for hardware IPs
2271 *
2272 * @adev: amdgpu_device pointer
2273 *
2274 * Main initialization pass for hardware IPs.  The list of all the hardware
2275 * IPs that make up the asic is walked and the sw_init and hw_init callbacks
2276 * are run.  sw_init initializes the software state associated with each IP
2277 * and hw_init initializes the hardware associated with each IP.
2278 * Returns 0 on success, negative error code on failure.
2279 */
2280static int amdgpu_device_ip_init(struct amdgpu_device *adev)
2281{
2282	int i, r;
2283
2284	r = amdgpu_ras_init(adev);
2285	if (r)
2286		return r;
2287
2288	for (i = 0; i < adev->num_ip_blocks; i++) {
2289		if (!adev->ip_blocks[i].status.valid)
2290			continue;
2291		r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
2292		if (r) {
2293			DRM_ERROR("sw_init of IP block <%s> failed %d\n",
2294				  adev->ip_blocks[i].version->funcs->name, r);
2295			goto init_failed;
2296		}
2297		adev->ip_blocks[i].status.sw = true;
2298
2299		/* need to do gmc hw init early so we can allocate gpu mem */
2300		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
2301			r = amdgpu_device_vram_scratch_init(adev);
2302			if (r) {
2303				DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r);
2304				goto init_failed;
2305			}
2306			r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
2307			if (r) {
2308				DRM_ERROR("hw_init %d failed %d\n", i, r);
2309				goto init_failed;
2310			}
2311			r = amdgpu_device_wb_init(adev);
2312			if (r) {
2313				DRM_ERROR("amdgpu_device_wb_init failed %d\n", r);
2314				goto init_failed;
2315			}
2316			adev->ip_blocks[i].status.hw = true;
2317
2318			/* right after GMC hw init, we create CSA */
2319			if (amdgpu_mcbp || amdgpu_sriov_vf(adev)) {
2320				r = amdgpu_allocate_static_csa(adev, &adev->virt.csa_obj,
2321								AMDGPU_GEM_DOMAIN_VRAM,
2322								AMDGPU_CSA_SIZE);
2323				if (r) {
2324					DRM_ERROR("allocate CSA failed %d\n", r);
2325					goto init_failed;
2326				}
2327			}
2328		}
2329	}
2330
2331	if (amdgpu_sriov_vf(adev))
2332		amdgpu_virt_init_data_exchange(adev);
2333
2334	r = amdgpu_ib_pool_init(adev);
2335	if (r) {
2336		dev_err(adev->dev, "IB initialization failed (%d).\n", r);
2337		amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r);
2338		goto init_failed;
 
 
 
 
2339	}
2340
2341	r = amdgpu_ucode_create_bo(adev); /* create ucode bo when sw_init complete*/
2342	if (r)
2343		goto init_failed;
2344
2345	r = amdgpu_device_ip_hw_init_phase1(adev);
2346	if (r)
2347		goto init_failed;
2348
2349	r = amdgpu_device_fw_loading(adev);
2350	if (r)
2351		goto init_failed;
2352
2353	r = amdgpu_device_ip_hw_init_phase2(adev);
2354	if (r)
2355		goto init_failed;
2356
2357	/*
2358	 * retired pages will be loaded from eeprom and reserved here,
2359	 * it should be called after amdgpu_device_ip_hw_init_phase2  since
2360	 * for some ASICs the RAS EEPROM code relies on SMU fully functioning
2361	 * for I2C communication which only true at this point.
2362	 *
2363	 * amdgpu_ras_recovery_init may fail, but the upper only cares the
2364	 * failure from bad gpu situation and stop amdgpu init process
2365	 * accordingly. For other failed cases, it will still release all
2366	 * the resource and print error message, rather than returning one
2367	 * negative value to upper level.
2368	 *
2369	 * Note: theoretically, this should be called before all vram allocations
2370	 * to protect retired page from abusing
2371	 */
2372	r = amdgpu_ras_recovery_init(adev);
2373	if (r)
2374		goto init_failed;
2375
2376	if (adev->gmc.xgmi.num_physical_nodes > 1)
2377		amdgpu_xgmi_add_device(adev);
2378
2379	/* Don't init kfd if whole hive need to be reset during init */
2380	if (!adev->gmc.xgmi.pending_reset)
2381		amdgpu_amdkfd_device_init(adev);
2382
2383	r = amdgpu_amdkfd_resume_iommu(adev);
2384	if (r)
2385		goto init_failed;
2386
2387	amdgpu_fru_get_product_info(adev);
2388
2389init_failed:
2390	if (amdgpu_sriov_vf(adev))
2391		amdgpu_virt_release_full_gpu(adev, true);
2392
2393	return r;
2394}
2395
2396/**
2397 * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer
2398 *
2399 * @adev: amdgpu_device pointer
2400 *
2401 * Writes a reset magic value to the gart pointer in VRAM.  The driver calls
2402 * this function before a GPU reset.  If the value is retained after a
2403 * GPU reset, VRAM has not been lost.  Some GPU resets may destry VRAM contents.
2404 */
2405static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev)
2406{
2407	memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
2408}
2409
2410/**
2411 * amdgpu_device_check_vram_lost - check if vram is valid
2412 *
2413 * @adev: amdgpu_device pointer
2414 *
2415 * Checks the reset magic value written to the gart pointer in VRAM.
2416 * The driver calls this after a GPU reset to see if the contents of
2417 * VRAM is lost or now.
2418 * returns true if vram is lost, false if not.
2419 */
2420static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev)
2421{
2422	if (memcmp(adev->gart.ptr, adev->reset_magic,
2423			AMDGPU_RESET_MAGIC_NUM))
2424		return true;
2425
2426	if (!amdgpu_in_reset(adev))
2427		return false;
2428
2429	/*
2430	 * For all ASICs with baco/mode1 reset, the VRAM is
2431	 * always assumed to be lost.
2432	 */
2433	switch (amdgpu_asic_reset_method(adev)) {
2434	case AMD_RESET_METHOD_BACO:
2435	case AMD_RESET_METHOD_MODE1:
2436		return true;
2437	default:
2438		return false;
2439	}
2440}
2441
2442/**
2443 * amdgpu_device_set_cg_state - set clockgating for amdgpu device
2444 *
2445 * @adev: amdgpu_device pointer
2446 * @state: clockgating state (gate or ungate)
2447 *
 
2448 * The list of all the hardware IPs that make up the asic is walked and the
2449 * set_clockgating_state callbacks are run.
2450 * Late initialization pass enabling clockgating for hardware IPs.
2451 * Fini or suspend, pass disabling clockgating for hardware IPs.
2452 * Returns 0 on success, negative error code on failure.
2453 */
2454
2455int amdgpu_device_set_cg_state(struct amdgpu_device *adev,
2456			       enum amd_clockgating_state state)
2457{
2458	int i, j, r;
2459
2460	if (amdgpu_emu_mode == 1)
2461		return 0;
2462
2463	for (j = 0; j < adev->num_ip_blocks; j++) {
2464		i = state == AMD_CG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
2465		if (!adev->ip_blocks[i].status.late_initialized)
2466			continue;
2467		/* skip CG for GFX on S0ix */
2468		if (adev->in_s0ix &&
2469		    adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX)
2470			continue;
2471		/* skip CG for VCE/UVD, it's handled specially */
2472		if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
2473		    adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
2474		    adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
2475		    adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG &&
2476		    adev->ip_blocks[i].version->funcs->set_clockgating_state) {
2477			/* enable clockgating to save power */
2478			r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
2479										     state);
2480			if (r) {
2481				DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
2482					  adev->ip_blocks[i].version->funcs->name, r);
2483				return r;
2484			}
2485		}
2486	}
2487
2488	return 0;
2489}
2490
2491int amdgpu_device_set_pg_state(struct amdgpu_device *adev,
2492			       enum amd_powergating_state state)
2493{
2494	int i, j, r;
2495
2496	if (amdgpu_emu_mode == 1)
2497		return 0;
2498
2499	for (j = 0; j < adev->num_ip_blocks; j++) {
2500		i = state == AMD_PG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
2501		if (!adev->ip_blocks[i].status.late_initialized)
2502			continue;
2503		/* skip PG for GFX on S0ix */
2504		if (adev->in_s0ix &&
2505		    adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX)
2506			continue;
2507		/* skip CG for VCE/UVD, it's handled specially */
2508		if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
2509		    adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
2510		    adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
2511		    adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG &&
2512		    adev->ip_blocks[i].version->funcs->set_powergating_state) {
2513			/* enable powergating to save power */
2514			r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev,
2515											state);
2516			if (r) {
2517				DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n",
2518					  adev->ip_blocks[i].version->funcs->name, r);
2519				return r;
2520			}
2521		}
2522	}
2523	return 0;
2524}
2525
2526static int amdgpu_device_enable_mgpu_fan_boost(void)
2527{
2528	struct amdgpu_gpu_instance *gpu_ins;
2529	struct amdgpu_device *adev;
2530	int i, ret = 0;
2531
2532	mutex_lock(&mgpu_info.mutex);
2533
2534	/*
2535	 * MGPU fan boost feature should be enabled
2536	 * only when there are two or more dGPUs in
2537	 * the system
2538	 */
2539	if (mgpu_info.num_dgpu < 2)
2540		goto out;
2541
2542	for (i = 0; i < mgpu_info.num_dgpu; i++) {
2543		gpu_ins = &(mgpu_info.gpu_ins[i]);
2544		adev = gpu_ins->adev;
2545		if (!(adev->flags & AMD_IS_APU) &&
2546		    !gpu_ins->mgpu_fan_enabled) {
2547			ret = amdgpu_dpm_enable_mgpu_fan_boost(adev);
2548			if (ret)
2549				break;
2550
2551			gpu_ins->mgpu_fan_enabled = 1;
2552		}
2553	}
2554
2555out:
2556	mutex_unlock(&mgpu_info.mutex);
2557
2558	return ret;
2559}
2560
2561/**
2562 * amdgpu_device_ip_late_init - run late init for hardware IPs
2563 *
2564 * @adev: amdgpu_device pointer
2565 *
2566 * Late initialization pass for hardware IPs.  The list of all the hardware
2567 * IPs that make up the asic is walked and the late_init callbacks are run.
2568 * late_init covers any special initialization that an IP requires
2569 * after all of the have been initialized or something that needs to happen
2570 * late in the init process.
2571 * Returns 0 on success, negative error code on failure.
2572 */
2573static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
2574{
2575	struct amdgpu_gpu_instance *gpu_instance;
2576	int i = 0, r;
2577
2578	for (i = 0; i < adev->num_ip_blocks; i++) {
2579		if (!adev->ip_blocks[i].status.hw)
2580			continue;
2581		if (adev->ip_blocks[i].version->funcs->late_init) {
2582			r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
2583			if (r) {
2584				DRM_ERROR("late_init of IP block <%s> failed %d\n",
2585					  adev->ip_blocks[i].version->funcs->name, r);
2586				return r;
2587			}
 
2588		}
2589		adev->ip_blocks[i].status.late_initialized = true;
2590	}
2591
2592	amdgpu_ras_set_error_query_ready(adev, true);
2593
2594	amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE);
2595	amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE);
2596
2597	amdgpu_device_fill_reset_magic(adev);
2598
2599	r = amdgpu_device_enable_mgpu_fan_boost();
2600	if (r)
2601		DRM_ERROR("enable mgpu fan boost failed (%d).\n", r);
2602
2603	/* For XGMI + passthrough configuration on arcturus, enable light SBR */
2604	if (adev->asic_type == CHIP_ARCTURUS &&
2605	    amdgpu_passthrough(adev) &&
2606	    adev->gmc.xgmi.num_physical_nodes > 1)
2607		smu_set_light_sbr(&adev->smu, true);
2608
2609	if (adev->gmc.xgmi.num_physical_nodes > 1) {
2610		mutex_lock(&mgpu_info.mutex);
2611
2612		/*
2613		 * Reset device p-state to low as this was booted with high.
2614		 *
2615		 * This should be performed only after all devices from the same
2616		 * hive get initialized.
2617		 *
2618		 * However, it's unknown how many device in the hive in advance.
2619		 * As this is counted one by one during devices initializations.
2620		 *
2621		 * So, we wait for all XGMI interlinked devices initialized.
2622		 * This may bring some delays as those devices may come from
2623		 * different hives. But that should be OK.
2624		 */
2625		if (mgpu_info.num_dgpu == adev->gmc.xgmi.num_physical_nodes) {
2626			for (i = 0; i < mgpu_info.num_gpu; i++) {
2627				gpu_instance = &(mgpu_info.gpu_ins[i]);
2628				if (gpu_instance->adev->flags & AMD_IS_APU)
2629					continue;
2630
2631				r = amdgpu_xgmi_set_pstate(gpu_instance->adev,
2632						AMDGPU_XGMI_PSTATE_MIN);
2633				if (r) {
2634					DRM_ERROR("pstate setting failed (%d).\n", r);
2635					break;
2636				}
2637			}
2638		}
2639
2640		mutex_unlock(&mgpu_info.mutex);
2641	}
2642
2643	return 0;
2644}
2645
2646static int amdgpu_device_ip_fini_early(struct amdgpu_device *adev)
 
 
 
 
 
 
 
 
 
 
 
2647{
2648	int i, r;
2649
2650	for (i = 0; i < adev->num_ip_blocks; i++) {
2651		if (!adev->ip_blocks[i].version->funcs->early_fini)
2652			continue;
2653
2654		r = adev->ip_blocks[i].version->funcs->early_fini((void *)adev);
2655		if (r) {
2656			DRM_DEBUG("early_fini of IP block <%s> failed %d\n",
2657				  adev->ip_blocks[i].version->funcs->name, r);
2658		}
2659	}
2660
2661	amdgpu_amdkfd_suspend(adev, false);
2662
2663	amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
2664	amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
2665
2666	/* need to disable SMC first */
2667	for (i = 0; i < adev->num_ip_blocks; i++) {
2668		if (!adev->ip_blocks[i].status.hw)
2669			continue;
2670		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
 
 
 
 
 
 
 
 
 
2671			r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
2672			/* XXX handle errors */
2673			if (r) {
2674				DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
2675					  adev->ip_blocks[i].version->funcs->name, r);
2676			}
2677			adev->ip_blocks[i].status.hw = false;
2678			break;
2679		}
2680	}
2681
2682	for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2683		if (!adev->ip_blocks[i].status.hw)
2684			continue;
2685
 
 
 
 
 
 
 
 
 
 
 
 
 
2686		r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
2687		/* XXX handle errors */
2688		if (r) {
2689			DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
2690				  adev->ip_blocks[i].version->funcs->name, r);
2691		}
2692
2693		adev->ip_blocks[i].status.hw = false;
2694	}
2695
2696	return 0;
2697}
2698
2699/**
2700 * amdgpu_device_ip_fini - run fini for hardware IPs
2701 *
2702 * @adev: amdgpu_device pointer
2703 *
2704 * Main teardown pass for hardware IPs.  The list of all the hardware
2705 * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks
2706 * are run.  hw_fini tears down the hardware associated with each IP
2707 * and sw_fini tears down any software state associated with each IP.
2708 * Returns 0 on success, negative error code on failure.
2709 */
2710static int amdgpu_device_ip_fini(struct amdgpu_device *adev)
2711{
2712	int i, r;
2713
2714	if (amdgpu_sriov_vf(adev) && adev->virt.ras_init_done)
2715		amdgpu_virt_release_ras_err_handler_data(adev);
2716
2717	amdgpu_ras_pre_fini(adev);
2718
2719	if (adev->gmc.xgmi.num_physical_nodes > 1)
2720		amdgpu_xgmi_remove_device(adev);
2721
2722	amdgpu_amdkfd_device_fini_sw(adev);
2723
2724	for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2725		if (!adev->ip_blocks[i].status.sw)
2726			continue;
2727
2728		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
2729			amdgpu_ucode_free_bo(adev);
2730			amdgpu_free_static_csa(&adev->virt.csa_obj);
2731			amdgpu_device_wb_fini(adev);
2732			amdgpu_device_vram_scratch_fini(adev);
2733			amdgpu_ib_pool_fini(adev);
2734		}
2735
2736		r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
2737		/* XXX handle errors */
2738		if (r) {
2739			DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
2740				  adev->ip_blocks[i].version->funcs->name, r);
2741		}
2742		adev->ip_blocks[i].status.sw = false;
2743		adev->ip_blocks[i].status.valid = false;
2744	}
2745
2746	for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2747		if (!adev->ip_blocks[i].status.late_initialized)
2748			continue;
2749		if (adev->ip_blocks[i].version->funcs->late_fini)
2750			adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
2751		adev->ip_blocks[i].status.late_initialized = false;
2752	}
2753
2754	amdgpu_ras_fini(adev);
2755
2756	if (amdgpu_sriov_vf(adev))
2757		if (amdgpu_virt_release_full_gpu(adev, false))
2758			DRM_ERROR("failed to release exclusive mode on fini\n");
2759
2760	return 0;
2761}
2762
2763/**
2764 * amdgpu_device_delayed_init_work_handler - work handler for IB tests
 
 
2765 *
2766 * @work: work_struct.
 
 
2767 */
2768static void amdgpu_device_delayed_init_work_handler(struct work_struct *work)
2769{
2770	struct amdgpu_device *adev =
2771		container_of(work, struct amdgpu_device, delayed_init_work.work);
2772	int r;
2773
2774	r = amdgpu_ib_ring_tests(adev);
2775	if (r)
2776		DRM_ERROR("ib ring test failed (%d).\n", r);
2777}
2778
2779static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work)
2780{
2781	struct amdgpu_device *adev =
2782		container_of(work, struct amdgpu_device, gfx.gfx_off_delay_work.work);
2783
2784	WARN_ON_ONCE(adev->gfx.gfx_off_state);
2785	WARN_ON_ONCE(adev->gfx.gfx_off_req_count);
2786
2787	if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, true))
2788		adev->gfx.gfx_off_state = true;
2789}
2790
2791/**
2792 * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1)
2793 *
2794 * @adev: amdgpu_device pointer
2795 *
2796 * Main suspend function for hardware IPs.  The list of all the hardware
2797 * IPs that make up the asic is walked, clockgating is disabled and the
2798 * suspend callbacks are run.  suspend puts the hardware and software state
2799 * in each IP into a state suitable for suspend.
2800 * Returns 0 on success, negative error code on failure.
2801 */
2802static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev)
2803{
2804	int i, r;
2805
2806	amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
2807	amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
2808
2809	for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2810		if (!adev->ip_blocks[i].status.valid)
2811			continue;
2812
2813		/* displays are handled separately */
2814		if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_DCE)
2815			continue;
2816
2817		/* XXX handle errors */
2818		r = adev->ip_blocks[i].version->funcs->suspend(adev);
2819		/* XXX handle errors */
2820		if (r) {
2821			DRM_ERROR("suspend of IP block <%s> failed %d\n",
2822				  adev->ip_blocks[i].version->funcs->name, r);
2823			return r;
2824		}
2825
2826		adev->ip_blocks[i].status.hw = false;
2827	}
2828
2829	return 0;
2830}
2831
2832/**
2833 * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2)
2834 *
2835 * @adev: amdgpu_device pointer
2836 *
2837 * Main suspend function for hardware IPs.  The list of all the hardware
2838 * IPs that make up the asic is walked, clockgating is disabled and the
2839 * suspend callbacks are run.  suspend puts the hardware and software state
2840 * in each IP into a state suitable for suspend.
2841 * Returns 0 on success, negative error code on failure.
2842 */
2843static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev)
2844{
2845	int i, r;
2846
2847	if (adev->in_s0ix)
2848		amdgpu_gfx_state_change_set(adev, sGpuChangeState_D3Entry);
2849
2850	for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2851		if (!adev->ip_blocks[i].status.valid)
2852			continue;
2853		/* displays are handled in phase1 */
2854		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE)
2855			continue;
2856		/* PSP lost connection when err_event_athub occurs */
2857		if (amdgpu_ras_intr_triggered() &&
2858		    adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
2859			adev->ip_blocks[i].status.hw = false;
2860			continue;
2861		}
2862
2863		/* skip unnecessary suspend if we do not initialize them yet */
2864		if (adev->gmc.xgmi.pending_reset &&
2865		    !(adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2866		      adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC ||
2867		      adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2868		      adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH)) {
2869			adev->ip_blocks[i].status.hw = false;
2870			continue;
2871		}
2872
2873		/* skip suspend of gfx and psp for S0ix
2874		 * gfx is in gfxoff state, so on resume it will exit gfxoff just
2875		 * like at runtime. PSP is also part of the always on hardware
2876		 * so no need to suspend it.
2877		 */
2878		if (adev->in_s0ix &&
2879		    (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP ||
2880		     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX))
2881			continue;
2882
2883		/* XXX handle errors */
2884		r = adev->ip_blocks[i].version->funcs->suspend(adev);
2885		/* XXX handle errors */
2886		if (r) {
2887			DRM_ERROR("suspend of IP block <%s> failed %d\n",
2888				  adev->ip_blocks[i].version->funcs->name, r);
2889		}
2890		adev->ip_blocks[i].status.hw = false;
2891		/* handle putting the SMC in the appropriate state */
2892		if(!amdgpu_sriov_vf(adev)){
2893			if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
2894				r = amdgpu_dpm_set_mp1_state(adev, adev->mp1_state);
2895				if (r) {
2896					DRM_ERROR("SMC failed to set mp1 state %d, %d\n",
2897							adev->mp1_state, r);
2898					return r;
2899				}
2900			}
2901		}
2902	}
2903
2904	return 0;
2905}
2906
2907/**
2908 * amdgpu_device_ip_suspend - run suspend for hardware IPs
2909 *
2910 * @adev: amdgpu_device pointer
2911 *
2912 * Main suspend function for hardware IPs.  The list of all the hardware
2913 * IPs that make up the asic is walked, clockgating is disabled and the
2914 * suspend callbacks are run.  suspend puts the hardware and software state
2915 * in each IP into a state suitable for suspend.
2916 * Returns 0 on success, negative error code on failure.
2917 */
2918int amdgpu_device_ip_suspend(struct amdgpu_device *adev)
2919{
2920	int r;
2921
2922	if (amdgpu_sriov_vf(adev)) {
2923		amdgpu_virt_fini_data_exchange(adev);
2924		amdgpu_virt_request_full_gpu(adev, false);
2925	}
2926
2927	r = amdgpu_device_ip_suspend_phase1(adev);
2928	if (r)
2929		return r;
2930	r = amdgpu_device_ip_suspend_phase2(adev);
2931
2932	if (amdgpu_sriov_vf(adev))
2933		amdgpu_virt_release_full_gpu(adev, false);
2934
2935	return r;
2936}
2937
2938static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
2939{
2940	int i, r;
2941
2942	static enum amd_ip_block_type ip_order[] = {
2943		AMD_IP_BLOCK_TYPE_GMC,
2944		AMD_IP_BLOCK_TYPE_COMMON,
2945		AMD_IP_BLOCK_TYPE_PSP,
2946		AMD_IP_BLOCK_TYPE_IH,
2947	};
2948
2949	for (i = 0; i < adev->num_ip_blocks; i++) {
2950		int j;
2951		struct amdgpu_ip_block *block;
2952
2953		block = &adev->ip_blocks[i];
2954		block->status.hw = false;
2955
2956		for (j = 0; j < ARRAY_SIZE(ip_order); j++) {
2957
2958			if (block->version->type != ip_order[j] ||
2959				!block->status.valid)
2960				continue;
2961
2962			r = block->version->funcs->hw_init(adev);
2963			DRM_INFO("RE-INIT-early: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
2964			if (r)
2965				return r;
2966			block->status.hw = true;
2967		}
2968	}
2969
2970	return 0;
2971}
2972
2973static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
2974{
2975	int i, r;
2976
2977	static enum amd_ip_block_type ip_order[] = {
2978		AMD_IP_BLOCK_TYPE_SMC,
 
2979		AMD_IP_BLOCK_TYPE_DCE,
2980		AMD_IP_BLOCK_TYPE_GFX,
2981		AMD_IP_BLOCK_TYPE_SDMA,
2982		AMD_IP_BLOCK_TYPE_UVD,
2983		AMD_IP_BLOCK_TYPE_VCE,
2984		AMD_IP_BLOCK_TYPE_VCN
2985	};
2986
2987	for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
2988		int j;
2989		struct amdgpu_ip_block *block;
2990
2991		for (j = 0; j < adev->num_ip_blocks; j++) {
2992			block = &adev->ip_blocks[j];
2993
2994			if (block->version->type != ip_order[i] ||
2995				!block->status.valid ||
2996				block->status.hw)
2997				continue;
2998
2999			if (block->version->type == AMD_IP_BLOCK_TYPE_SMC)
3000				r = block->version->funcs->resume(adev);
3001			else
3002				r = block->version->funcs->hw_init(adev);
3003
3004			DRM_INFO("RE-INIT-late: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
3005			if (r)
3006				return r;
3007			block->status.hw = true;
3008		}
3009	}
3010
3011	return 0;
3012}
3013
3014/**
3015 * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs
3016 *
3017 * @adev: amdgpu_device pointer
3018 *
3019 * First resume function for hardware IPs.  The list of all the hardware
3020 * IPs that make up the asic is walked and the resume callbacks are run for
3021 * COMMON, GMC, and IH.  resume puts the hardware into a functional state
3022 * after a suspend and updates the software state as necessary.  This
3023 * function is also used for restoring the GPU after a GPU reset.
3024 * Returns 0 on success, negative error code on failure.
3025 */
3026static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)
3027{
3028	int i, r;
3029
3030	for (i = 0; i < adev->num_ip_blocks; i++) {
3031		if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
3032			continue;
3033		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
3034		    adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
3035		    adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
3036
3037			r = adev->ip_blocks[i].version->funcs->resume(adev);
3038			if (r) {
3039				DRM_ERROR("resume of IP block <%s> failed %d\n",
3040					  adev->ip_blocks[i].version->funcs->name, r);
3041				return r;
3042			}
3043			adev->ip_blocks[i].status.hw = true;
3044		}
3045	}
3046
3047	return 0;
3048}
3049
3050/**
3051 * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs
3052 *
3053 * @adev: amdgpu_device pointer
3054 *
3055 * First resume function for hardware IPs.  The list of all the hardware
3056 * IPs that make up the asic is walked and the resume callbacks are run for
3057 * all blocks except COMMON, GMC, and IH.  resume puts the hardware into a
3058 * functional state after a suspend and updates the software state as
3059 * necessary.  This function is also used for restoring the GPU after a GPU
3060 * reset.
3061 * Returns 0 on success, negative error code on failure.
3062 */
3063static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)
3064{
3065	int i, r;
3066
3067	for (i = 0; i < adev->num_ip_blocks; i++) {
3068		if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
3069			continue;
3070		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
3071		    adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
3072		    adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH ||
3073		    adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)
3074			continue;
3075		r = adev->ip_blocks[i].version->funcs->resume(adev);
3076		if (r) {
3077			DRM_ERROR("resume of IP block <%s> failed %d\n",
3078				  adev->ip_blocks[i].version->funcs->name, r);
3079			return r;
3080		}
3081		adev->ip_blocks[i].status.hw = true;
3082	}
3083
3084	return 0;
3085}
3086
3087/**
3088 * amdgpu_device_ip_resume - run resume for hardware IPs
3089 *
3090 * @adev: amdgpu_device pointer
3091 *
3092 * Main resume function for hardware IPs.  The hardware IPs
3093 * are split into two resume functions because they are
3094 * are also used in in recovering from a GPU reset and some additional
3095 * steps need to be take between them.  In this case (S3/S4) they are
3096 * run sequentially.
3097 * Returns 0 on success, negative error code on failure.
3098 */
3099static int amdgpu_device_ip_resume(struct amdgpu_device *adev)
3100{
3101	int r;
3102
3103	r = amdgpu_amdkfd_resume_iommu(adev);
3104	if (r)
3105		return r;
3106
3107	r = amdgpu_device_ip_resume_phase1(adev);
3108	if (r)
3109		return r;
3110
3111	r = amdgpu_device_fw_loading(adev);
3112	if (r)
3113		return r;
3114
3115	r = amdgpu_device_ip_resume_phase2(adev);
3116
3117	return r;
3118}
3119
3120/**
3121 * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV
3122 *
3123 * @adev: amdgpu_device pointer
3124 *
3125 * Query the VBIOS data tables to determine if the board supports SR-IOV.
3126 */
3127static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
3128{
3129	if (amdgpu_sriov_vf(adev)) {
3130		if (adev->is_atom_fw) {
3131			if (amdgpu_atomfirmware_gpu_virtualization_supported(adev))
3132				adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
3133		} else {
3134			if (amdgpu_atombios_has_gpu_virtualization_table(adev))
3135				adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
3136		}
3137
3138		if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
3139			amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
3140	}
3141}
3142
3143/**
3144 * amdgpu_device_asic_has_dc_support - determine if DC supports the asic
3145 *
3146 * @asic_type: AMD asic type
3147 *
3148 * Check if there is DC (new modesetting infrastructre) support for an asic.
3149 * returns true if DC has support, false if not.
3150 */
3151bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
3152{
3153	switch (asic_type) {
3154#if defined(CONFIG_DRM_AMD_DC)
3155#if defined(CONFIG_DRM_AMD_DC_SI)
3156	case CHIP_TAHITI:
3157	case CHIP_PITCAIRN:
3158	case CHIP_VERDE:
3159	case CHIP_OLAND:
3160#endif
3161	case CHIP_BONAIRE:
 
3162	case CHIP_KAVERI:
3163	case CHIP_KABINI:
3164	case CHIP_MULLINS:
3165		/*
3166		 * We have systems in the wild with these ASICs that require
3167		 * LVDS and VGA support which is not supported with DC.
3168		 *
3169		 * Fallback to the non-DC driver here by default so as not to
3170		 * cause regressions.
3171		 */
3172		return amdgpu_dc > 0;
3173	case CHIP_HAWAII:
3174	case CHIP_CARRIZO:
3175	case CHIP_STONEY:
 
3176	case CHIP_POLARIS10:
3177	case CHIP_POLARIS11:
3178	case CHIP_POLARIS12:
3179	case CHIP_VEGAM:
3180	case CHIP_TONGA:
3181	case CHIP_FIJI:
 
 
 
3182	case CHIP_VEGA10:
3183	case CHIP_VEGA12:
3184	case CHIP_VEGA20:
3185#if defined(CONFIG_DRM_AMD_DC_DCN)
3186	case CHIP_RAVEN:
3187	case CHIP_NAVI10:
3188	case CHIP_NAVI14:
3189	case CHIP_NAVI12:
3190	case CHIP_RENOIR:
3191	case CHIP_SIENNA_CICHLID:
3192	case CHIP_NAVY_FLOUNDER:
3193	case CHIP_DIMGREY_CAVEFISH:
3194	case CHIP_BEIGE_GOBY:
3195	case CHIP_VANGOGH:
3196	case CHIP_YELLOW_CARP:
3197#endif
3198		return amdgpu_dc != 0;
3199#endif
3200	default:
3201		if (amdgpu_dc > 0)
3202			DRM_INFO_ONCE("Display Core has been requested via kernel parameter "
3203					 "but isn't supported by ASIC, ignoring\n");
3204		return false;
3205	}
3206}
3207
3208/**
3209 * amdgpu_device_has_dc_support - check if dc is supported
3210 *
3211 * @adev: amdgpu_device pointer
3212 *
3213 * Returns true for supported, false for not supported
3214 */
3215bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
3216{
3217	if (amdgpu_sriov_vf(adev) || 
3218	    adev->enable_virtual_display ||
3219	    (adev->harvest_ip_mask & AMD_HARVEST_IP_DMU_MASK))
3220		return false;
3221
3222	return amdgpu_device_asic_has_dc_support(adev->asic_type);
3223}
3224
3225static void amdgpu_device_xgmi_reset_func(struct work_struct *__work)
3226{
3227	struct amdgpu_device *adev =
3228		container_of(__work, struct amdgpu_device, xgmi_reset_work);
3229	struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev);
3230
3231	/* It's a bug to not have a hive within this function */
3232	if (WARN_ON(!hive))
3233		return;
3234
3235	/*
3236	 * Use task barrier to synchronize all xgmi reset works across the
3237	 * hive. task_barrier_enter and task_barrier_exit will block
3238	 * until all the threads running the xgmi reset works reach
3239	 * those points. task_barrier_full will do both blocks.
3240	 */
3241	if (amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) {
3242
3243		task_barrier_enter(&hive->tb);
3244		adev->asic_reset_res = amdgpu_device_baco_enter(adev_to_drm(adev));
3245
3246		if (adev->asic_reset_res)
3247			goto fail;
3248
3249		task_barrier_exit(&hive->tb);
3250		adev->asic_reset_res = amdgpu_device_baco_exit(adev_to_drm(adev));
3251
3252		if (adev->asic_reset_res)
3253			goto fail;
3254
3255		if (adev->mmhub.ras_funcs &&
3256		    adev->mmhub.ras_funcs->reset_ras_error_count)
3257			adev->mmhub.ras_funcs->reset_ras_error_count(adev);
3258	} else {
3259
3260		task_barrier_full(&hive->tb);
3261		adev->asic_reset_res =  amdgpu_asic_reset(adev);
3262	}
3263
3264fail:
3265	if (adev->asic_reset_res)
3266		DRM_WARN("ASIC reset failed with error, %d for drm dev, %s",
3267			 adev->asic_reset_res, adev_to_drm(adev)->unique);
3268	amdgpu_put_xgmi_hive(hive);
3269}
3270
3271static int amdgpu_device_get_job_timeout_settings(struct amdgpu_device *adev)
3272{
3273	char *input = amdgpu_lockup_timeout;
3274	char *timeout_setting = NULL;
3275	int index = 0;
3276	long timeout;
3277	int ret = 0;
3278
3279	/*
3280	 * By default timeout for non compute jobs is 10000
3281	 * and 60000 for compute jobs.
3282	 * In SR-IOV or passthrough mode, timeout for compute
3283	 * jobs are 60000 by default.
3284	 */
3285	adev->gfx_timeout = msecs_to_jiffies(10000);
3286	adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout;
3287	if (amdgpu_sriov_vf(adev))
3288		adev->compute_timeout = amdgpu_sriov_is_pp_one_vf(adev) ?
3289					msecs_to_jiffies(60000) : msecs_to_jiffies(10000);
3290	else
3291		adev->compute_timeout =  msecs_to_jiffies(60000);
3292
3293	if (strnlen(input, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) {
3294		while ((timeout_setting = strsep(&input, ",")) &&
3295				strnlen(timeout_setting, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) {
3296			ret = kstrtol(timeout_setting, 0, &timeout);
3297			if (ret)
3298				return ret;
3299
3300			if (timeout == 0) {
3301				index++;
3302				continue;
3303			} else if (timeout < 0) {
3304				timeout = MAX_SCHEDULE_TIMEOUT;
3305			} else {
3306				timeout = msecs_to_jiffies(timeout);
3307			}
3308
3309			switch (index++) {
3310			case 0:
3311				adev->gfx_timeout = timeout;
3312				break;
3313			case 1:
3314				adev->compute_timeout = timeout;
3315				break;
3316			case 2:
3317				adev->sdma_timeout = timeout;
3318				break;
3319			case 3:
3320				adev->video_timeout = timeout;
3321				break;
3322			default:
3323				break;
3324			}
3325		}
3326		/*
3327		 * There is only one value specified and
3328		 * it should apply to all non-compute jobs.
3329		 */
3330		if (index == 1) {
3331			adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout;
3332			if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev))
3333				adev->compute_timeout = adev->gfx_timeout;
3334		}
3335	}
3336
3337	return ret;
3338}
3339
3340static const struct attribute *amdgpu_dev_attributes[] = {
3341	&dev_attr_product_name.attr,
3342	&dev_attr_product_number.attr,
3343	&dev_attr_serial_number.attr,
3344	&dev_attr_pcie_replay_count.attr,
3345	NULL
3346};
3347
3348/**
3349 * amdgpu_device_init - initialize the driver
3350 *
3351 * @adev: amdgpu_device pointer
 
 
3352 * @flags: driver flags
3353 *
3354 * Initializes the driver info and hw (all asics).
3355 * Returns 0 for success or an error on failure.
3356 * Called at driver startup.
3357 */
3358int amdgpu_device_init(struct amdgpu_device *adev,
 
 
3359		       uint32_t flags)
3360{
3361	struct drm_device *ddev = adev_to_drm(adev);
3362	struct pci_dev *pdev = adev->pdev;
3363	int r, i;
3364	bool px = false;
3365	u32 max_MBps;
3366
3367	adev->shutdown = false;
 
 
 
3368	adev->flags = flags;
3369
3370	if (amdgpu_force_asic_type >= 0 && amdgpu_force_asic_type < CHIP_LAST)
3371		adev->asic_type = amdgpu_force_asic_type;
3372	else
3373		adev->asic_type = flags & AMD_ASIC_MASK;
3374
3375	adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
3376	if (amdgpu_emu_mode == 1)
3377		adev->usec_timeout *= 10;
3378	adev->gmc.gart_size = 512 * 1024 * 1024;
3379	adev->accel_working = false;
3380	adev->num_rings = 0;
3381	adev->mman.buffer_funcs = NULL;
3382	adev->mman.buffer_funcs_ring = NULL;
3383	adev->vm_manager.vm_pte_funcs = NULL;
3384	adev->vm_manager.vm_pte_num_scheds = 0;
3385	adev->gmc.gmc_funcs = NULL;
3386	adev->harvest_ip_mask = 0x0;
3387	adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
3388	bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
3389
3390	adev->smc_rreg = &amdgpu_invalid_rreg;
3391	adev->smc_wreg = &amdgpu_invalid_wreg;
3392	adev->pcie_rreg = &amdgpu_invalid_rreg;
3393	adev->pcie_wreg = &amdgpu_invalid_wreg;
3394	adev->pciep_rreg = &amdgpu_invalid_rreg;
3395	adev->pciep_wreg = &amdgpu_invalid_wreg;
3396	adev->pcie_rreg64 = &amdgpu_invalid_rreg64;
3397	adev->pcie_wreg64 = &amdgpu_invalid_wreg64;
3398	adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
3399	adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
3400	adev->didt_rreg = &amdgpu_invalid_rreg;
3401	adev->didt_wreg = &amdgpu_invalid_wreg;
3402	adev->gc_cac_rreg = &amdgpu_invalid_rreg;
3403	adev->gc_cac_wreg = &amdgpu_invalid_wreg;
3404	adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
3405	adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
3406
3407	DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
3408		 amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
3409		 pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
3410
3411	/* mutex initialization are all done here so we
3412	 * can recall function without having locking issues */
 
3413	mutex_init(&adev->firmware.mutex);
3414	mutex_init(&adev->pm.mutex);
3415	mutex_init(&adev->gfx.gpu_clock_mutex);
3416	mutex_init(&adev->srbm_mutex);
3417	mutex_init(&adev->gfx.pipe_reserve_mutex);
3418	mutex_init(&adev->gfx.gfx_off_mutex);
3419	mutex_init(&adev->grbm_idx_mutex);
3420	mutex_init(&adev->mn_lock);
3421	mutex_init(&adev->virt.vf_errors.lock);
3422	hash_init(adev->mn_hash);
3423	atomic_set(&adev->in_gpu_reset, 0);
3424	init_rwsem(&adev->reset_sem);
3425	mutex_init(&adev->psp.mutex);
3426	mutex_init(&adev->notifier_lock);
3427
3428	r = amdgpu_device_init_apu_flags(adev);
3429	if (r)
3430		return r;
3431
3432	r = amdgpu_device_check_arguments(adev);
3433	if (r)
3434		return r;
3435
3436	spin_lock_init(&adev->mmio_idx_lock);
3437	spin_lock_init(&adev->smc_idx_lock);
3438	spin_lock_init(&adev->pcie_idx_lock);
3439	spin_lock_init(&adev->uvd_ctx_idx_lock);
3440	spin_lock_init(&adev->didt_idx_lock);
3441	spin_lock_init(&adev->gc_cac_idx_lock);
3442	spin_lock_init(&adev->se_cac_idx_lock);
3443	spin_lock_init(&adev->audio_endpt_idx_lock);
3444	spin_lock_init(&adev->mm_stats.lock);
3445
3446	INIT_LIST_HEAD(&adev->shadow_list);
3447	mutex_init(&adev->shadow_list_lock);
3448
3449	INIT_LIST_HEAD(&adev->reset_list);
 
3450
3451	INIT_DELAYED_WORK(&adev->delayed_init_work,
3452			  amdgpu_device_delayed_init_work_handler);
3453	INIT_DELAYED_WORK(&adev->gfx.gfx_off_delay_work,
3454			  amdgpu_device_delay_enable_gfx_off);
3455
3456	INIT_WORK(&adev->xgmi_reset_work, amdgpu_device_xgmi_reset_func);
3457
3458	adev->gfx.gfx_off_req_count = 1;
3459	adev->pm.ac_power = power_supply_is_system_supplied() > 0;
3460
3461	atomic_set(&adev->throttling_logging_enabled, 1);
3462	/*
3463	 * If throttling continues, logging will be performed every minute
3464	 * to avoid log flooding. "-1" is subtracted since the thermal
3465	 * throttling interrupt comes every second. Thus, the total logging
3466	 * interval is 59 seconds(retelimited printk interval) + 1(waiting
3467	 * for throttling interrupt) = 60 seconds.
3468	 */
3469	ratelimit_state_init(&adev->throttling_logging_rs, (60 - 1) * HZ, 1);
3470	ratelimit_set_flags(&adev->throttling_logging_rs, RATELIMIT_MSG_ON_RELEASE);
3471
3472	/* Registers mapping */
3473	/* TODO: block userspace mapping of io register */
3474	if (adev->asic_type >= CHIP_BONAIRE) {
3475		adev->rmmio_base = pci_resource_start(adev->pdev, 5);
3476		adev->rmmio_size = pci_resource_len(adev->pdev, 5);
3477	} else {
3478		adev->rmmio_base = pci_resource_start(adev->pdev, 2);
3479		adev->rmmio_size = pci_resource_len(adev->pdev, 2);
3480	}
3481
3482	adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
3483	if (adev->rmmio == NULL) {
3484		return -ENOMEM;
3485	}
3486	DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
3487	DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
3488
3489	/* enable PCIE atomic ops */
3490	r = pci_enable_atomic_ops_to_root(adev->pdev,
3491					  PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
3492					  PCI_EXP_DEVCAP2_ATOMIC_COMP64);
3493	if (r) {
3494		adev->have_atomics_support = false;
3495		DRM_INFO("PCIE atomic ops is not supported\n");
3496	} else {
3497		adev->have_atomics_support = true;
 
3498	}
 
 
3499
3500	amdgpu_device_get_pcie_info(adev);
3501
3502	if (amdgpu_mcbp)
3503		DRM_INFO("MCBP is enabled\n");
3504
3505	if (amdgpu_mes && adev->asic_type >= CHIP_NAVI10)
3506		adev->enable_mes = true;
3507
3508	/* detect hw virtualization here */
3509	amdgpu_detect_virtualization(adev);
3510
3511	r = amdgpu_device_get_job_timeout_settings(adev);
3512	if (r) {
3513		dev_err(adev->dev, "invalid lockup_timeout parameter syntax\n");
3514		return r;
3515	}
3516
3517	/* early init functions */
3518	r = amdgpu_device_ip_early_init(adev);
3519	if (r)
3520		return r;
3521
3522	/* doorbell bar mapping and doorbell index init*/
3523	amdgpu_device_doorbell_init(adev);
 
 
 
 
 
 
 
 
 
 
3524
3525	if (amdgpu_emu_mode == 1) {
3526		/* post the asic on emulation mode */
3527		emu_soc_asic_init(adev);
3528		goto fence_driver_init;
3529	}
3530
3531	amdgpu_reset_init(adev);
 
 
 
 
 
 
 
 
 
 
 
3532
3533	/* detect if we are with an SRIOV vbios */
3534	amdgpu_device_detect_sriov_bios(adev);
3535
3536	/* check if we need to reset the asic
3537	 *  E.g., driver was not cleanly unloaded previously, etc.
3538	 */
3539	if (!amdgpu_sriov_vf(adev) && amdgpu_asic_need_reset_on_init(adev)) {
3540		if (adev->gmc.xgmi.num_physical_nodes) {
3541			dev_info(adev->dev, "Pending hive reset.\n");
3542			adev->gmc.xgmi.pending_reset = true;
3543			/* Only need to init necessary block for SMU to handle the reset */
3544			for (i = 0; i < adev->num_ip_blocks; i++) {
3545				if (!adev->ip_blocks[i].status.valid)
3546					continue;
3547				if (!(adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
3548				      adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
3549				      adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH ||
3550				      adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC)) {
3551					DRM_DEBUG("IP %s disabled for hw_init.\n",
3552						adev->ip_blocks[i].version->funcs->name);
3553					adev->ip_blocks[i].status.hw = true;
3554				}
3555			}
3556		} else {
3557			r = amdgpu_asic_reset(adev);
3558			if (r) {
3559				dev_err(adev->dev, "asic reset on init failed\n");
3560				goto failed;
3561			}
3562		}
3563	}
3564
3565	pci_enable_pcie_error_reporting(adev->pdev);
3566
3567	/* Post card if necessary */
3568	if (amdgpu_device_need_post(adev)) {
3569		if (!adev->bios) {
3570			dev_err(adev->dev, "no vBIOS found\n");
3571			r = -EINVAL;
3572			goto failed;
3573		}
3574		DRM_INFO("GPU posting now...\n");
3575		r = amdgpu_device_asic_init(adev);
3576		if (r) {
3577			dev_err(adev->dev, "gpu post error!\n");
3578			goto failed;
3579		}
3580	}
3581
3582	if (adev->is_atom_fw) {
3583		/* Initialize clocks */
3584		r = amdgpu_atomfirmware_get_clock_info(adev);
3585		if (r) {
3586			dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
3587			amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
3588			goto failed;
3589		}
3590	} else {
3591		/* Initialize clocks */
3592		r = amdgpu_atombios_get_clock_info(adev);
3593		if (r) {
3594			dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
3595			amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
3596			goto failed;
3597		}
3598		/* init i2c buses */
3599		if (!amdgpu_device_has_dc_support(adev))
3600			amdgpu_atombios_i2c_init(adev);
3601	}
3602
3603fence_driver_init:
3604	/* Fence driver */
3605	r = amdgpu_fence_driver_sw_init(adev);
3606	if (r) {
3607		dev_err(adev->dev, "amdgpu_fence_driver_sw_init failed\n");
3608		amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
3609		goto failed;
3610	}
3611
3612	/* init the mode config */
3613	drm_mode_config_init(adev_to_drm(adev));
3614
3615	r = amdgpu_device_ip_init(adev);
3616	if (r) {
3617		/* failed in exclusive mode due to timeout */
3618		if (amdgpu_sriov_vf(adev) &&
3619		    !amdgpu_sriov_runtime(adev) &&
3620		    amdgpu_virt_mmio_blocked(adev) &&
3621		    !amdgpu_virt_wait_reset(adev)) {
3622			dev_err(adev->dev, "VF exclusive mode timeout\n");
3623			/* Don't send request since VF is inactive. */
3624			adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
3625			adev->virt.ops = NULL;
3626			r = -EAGAIN;
3627			goto release_ras_con;
3628		}
3629		dev_err(adev->dev, "amdgpu_device_ip_init failed\n");
3630		amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
3631		goto release_ras_con;
3632	}
3633
3634	amdgpu_fence_driver_hw_init(adev);
3635
3636	dev_info(adev->dev,
3637		"SE %d, SH per SE %d, CU per SH %d, active_cu_number %d\n",
3638			adev->gfx.config.max_shader_engines,
3639			adev->gfx.config.max_sh_per_se,
3640			adev->gfx.config.max_cu_per_sh,
3641			adev->gfx.cu_info.number);
3642
3643	adev->accel_working = true;
3644
3645	amdgpu_vm_check_compute_bug(adev);
3646
3647	/* Initialize the buffer migration limit. */
3648	if (amdgpu_moverate >= 0)
3649		max_MBps = amdgpu_moverate;
3650	else
3651		max_MBps = 8; /* Allow 8 MB/s. */
3652	/* Get a log2 for easy divisions. */
3653	adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
3654
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3655	amdgpu_fbdev_init(adev);
3656
3657	r = amdgpu_pm_sysfs_init(adev);
3658	if (r) {
3659		adev->pm_sysfs_en = false;
3660		DRM_ERROR("registering pm debugfs failed (%d).\n", r);
3661	} else
3662		adev->pm_sysfs_en = true;
3663
3664	r = amdgpu_ucode_sysfs_init(adev);
3665	if (r) {
3666		adev->ucode_sysfs_en = false;
3667		DRM_ERROR("Creating firmware sysfs failed (%d).\n", r);
3668	} else
3669		adev->ucode_sysfs_en = true;
 
 
 
 
 
 
 
 
 
3670
3671	if ((amdgpu_testing & 1)) {
3672		if (adev->accel_working)
3673			amdgpu_test_moves(adev);
3674		else
3675			DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n");
3676	}
3677	if (amdgpu_benchmarking) {
3678		if (adev->accel_working)
3679			amdgpu_benchmark(adev, amdgpu_benchmarking);
3680		else
3681			DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n");
3682	}
3683
3684	/*
3685	 * Register gpu instance before amdgpu_device_enable_mgpu_fan_boost.
3686	 * Otherwise the mgpu fan boost feature will be skipped due to the
3687	 * gpu instance is counted less.
3688	 */
3689	amdgpu_register_gpu_instance(adev);
3690
3691	/* enable clockgating, etc. after ib tests, etc. since some blocks require
3692	 * explicit gating rather than handling it automatically.
3693	 */
3694	if (!adev->gmc.xgmi.pending_reset) {
3695		r = amdgpu_device_ip_late_init(adev);
3696		if (r) {
3697			dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n");
3698			amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
3699			goto release_ras_con;
3700		}
3701		/* must succeed. */
3702		amdgpu_ras_resume(adev);
3703		queue_delayed_work(system_wq, &adev->delayed_init_work,
3704				   msecs_to_jiffies(AMDGPU_RESUME_MS));
3705	}
3706
3707	if (amdgpu_sriov_vf(adev))
3708		flush_delayed_work(&adev->delayed_init_work);
3709
3710	r = sysfs_create_files(&adev->dev->kobj, amdgpu_dev_attributes);
3711	if (r)
3712		dev_err(adev->dev, "Could not create amdgpu device attr\n");
3713
3714	if (IS_ENABLED(CONFIG_PERF_EVENTS))
3715		r = amdgpu_pmu_init(adev);
3716	if (r)
3717		dev_err(adev->dev, "amdgpu_pmu_init failed\n");
3718
3719	/* Have stored pci confspace at hand for restore in sudden PCI error */
3720	if (amdgpu_device_cache_pci_state(adev->pdev))
3721		pci_restore_state(pdev);
3722
3723	/* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
3724	/* this will fail for cards that aren't VGA class devices, just
3725	 * ignore it */
3726	if ((adev->pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA)
3727		vga_client_register(adev->pdev, adev, NULL, amdgpu_device_vga_set_decode);
3728
3729	if (amdgpu_device_supports_px(ddev)) {
3730		px = true;
3731		vga_switcheroo_register_client(adev->pdev,
3732					       &amdgpu_switcheroo_ops, px);
3733		vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
3734	}
3735
3736	if (adev->gmc.xgmi.pending_reset)
3737		queue_delayed_work(system_wq, &mgpu_info.delayed_reset_work,
3738				   msecs_to_jiffies(AMDGPU_RESUME_MS));
3739
3740	return 0;
3741
3742release_ras_con:
3743	amdgpu_release_ras_context(adev);
3744
3745failed:
3746	amdgpu_vf_error_trans_all(adev);
 
 
3747
3748	return r;
3749}
3750
3751static void amdgpu_device_unmap_mmio(struct amdgpu_device *adev)
3752{
3753	/* Clear all CPU mappings pointing to this device */
3754	unmap_mapping_range(adev->ddev.anon_inode->i_mapping, 0, 0, 1);
3755
3756	/* Unmap all mapped bars - Doorbell, registers and VRAM */
3757	amdgpu_device_doorbell_fini(adev);
3758
3759	iounmap(adev->rmmio);
3760	adev->rmmio = NULL;
3761	if (adev->mman.aper_base_kaddr)
3762		iounmap(adev->mman.aper_base_kaddr);
3763	adev->mman.aper_base_kaddr = NULL;
3764
3765	/* Memory manager related */
3766	if (!adev->gmc.xgmi.connected_to_cpu) {
3767		arch_phys_wc_del(adev->gmc.vram_mtrr);
3768		arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
3769	}
3770}
3771
3772/**
3773 * amdgpu_device_fini - tear down the driver
3774 *
3775 * @adev: amdgpu_device pointer
3776 *
3777 * Tear down the driver info (all asics).
3778 * Called at driver shutdown.
3779 */
3780void amdgpu_device_fini_hw(struct amdgpu_device *adev)
3781{
3782	dev_info(adev->dev, "amdgpu: finishing device.\n");
3783	flush_delayed_work(&adev->delayed_init_work);
3784	ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
3785	adev->shutdown = true;
3786
3787	/* make sure IB test finished before entering exclusive mode
3788	 * to avoid preemption on IB test
3789	 * */
3790	if (amdgpu_sriov_vf(adev)) {
3791		amdgpu_virt_request_full_gpu(adev, false);
3792		amdgpu_virt_fini_data_exchange(adev);
3793	}
3794
3795	/* disable all interrupts */
3796	amdgpu_irq_disable_all(adev);
3797	if (adev->mode_info.mode_config_initialized){
3798		if (!amdgpu_device_has_dc_support(adev))
3799			drm_helper_force_disable_all(adev_to_drm(adev));
3800		else
3801			drm_atomic_helper_shutdown(adev_to_drm(adev));
3802	}
3803	amdgpu_fence_driver_hw_fini(adev);
3804
3805	if (adev->pm_sysfs_en)
3806		amdgpu_pm_sysfs_fini(adev);
3807	if (adev->ucode_sysfs_en)
3808		amdgpu_ucode_sysfs_fini(adev);
3809	sysfs_remove_files(&adev->dev->kobj, amdgpu_dev_attributes);
3810
3811	amdgpu_fbdev_fini(adev);
3812
3813	amdgpu_irq_fini_hw(adev);
3814
3815	amdgpu_device_ip_fini_early(adev);
3816
3817	amdgpu_gart_dummy_page_fini(adev);
3818
3819	amdgpu_device_unmap_mmio(adev);
3820}
3821
3822void amdgpu_device_fini_sw(struct amdgpu_device *adev)
3823{
3824	amdgpu_device_ip_fini(adev);
3825	amdgpu_fence_driver_sw_fini(adev);
3826	release_firmware(adev->firmware.gpu_info_fw);
3827	adev->firmware.gpu_info_fw = NULL;
3828	adev->accel_working = false;
3829
3830	amdgpu_reset_fini(adev);
3831
3832	/* free i2c buses */
3833	if (!amdgpu_device_has_dc_support(adev))
3834		amdgpu_i2c_fini(adev);
3835
3836	if (amdgpu_emu_mode != 1)
3837		amdgpu_atombios_fini(adev);
3838
3839	kfree(adev->bios);
3840	adev->bios = NULL;
3841	if (amdgpu_device_supports_px(adev_to_drm(adev))) {
3842		vga_switcheroo_unregister_client(adev->pdev);
 
3843		vga_switcheroo_fini_domain_pm_ops(adev->dev);
3844	}
3845	if ((adev->pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA)
3846		vga_client_register(adev->pdev, NULL, NULL, NULL);
3847
3848	if (IS_ENABLED(CONFIG_PERF_EVENTS))
3849		amdgpu_pmu_fini(adev);
3850	if (adev->mman.discovery_bin)
3851		amdgpu_discovery_fini(adev);
3852
3853	kfree(adev->pci_state);
3854
3855}
3856
3857
3858/*
3859 * Suspend & resume.
3860 */
3861/**
3862 * amdgpu_device_suspend - initiate device suspend
3863 *
3864 * @dev: drm dev pointer
3865 * @fbcon : notify the fbdev of suspend
3866 *
3867 * Puts the hw in the suspend state (all asics).
3868 * Returns 0 for success or an error on failure.
3869 * Called at driver suspend.
3870 */
3871int amdgpu_device_suspend(struct drm_device *dev, bool fbcon)
3872{
3873	struct amdgpu_device *adev = drm_to_adev(dev);
 
 
 
 
 
 
 
 
 
3874
3875	if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
3876		return 0;
3877
3878	adev->in_suspend = true;
3879
3880	if (amdgpu_acpi_smart_shift_update(dev, AMDGPU_SS_DEV_D3))
3881		DRM_WARN("smart shift update failed\n");
3882
3883	drm_kms_helper_poll_disable(dev);
3884
3885	if (fbcon)
3886		amdgpu_fbdev_set_suspend(adev, 1);
 
 
 
 
 
 
3887
3888	cancel_delayed_work_sync(&adev->delayed_init_work);
3889
3890	amdgpu_ras_suspend(adev);
 
 
 
 
3891
3892	amdgpu_device_ip_suspend_phase1(adev);
3893
3894	if (!adev->in_s0ix)
3895		amdgpu_amdkfd_suspend(adev, adev->in_runpm);
 
 
 
 
3896
 
 
 
 
 
 
 
 
 
 
 
 
 
3897	/* evict vram memory */
3898	amdgpu_bo_evict_vram(adev);
3899
3900	amdgpu_fence_driver_hw_fini(adev);
 
 
3901
3902	amdgpu_device_ip_suspend_phase2(adev);
3903	/* evict remaining vram memory
3904	 * This second call to evict vram is to evict the gart page table
3905	 * using the CPU.
3906	 */
3907	amdgpu_bo_evict_vram(adev);
3908
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3909	return 0;
3910}
3911
3912/**
3913 * amdgpu_device_resume - initiate device resume
3914 *
3915 * @dev: drm dev pointer
3916 * @fbcon : notify the fbdev of resume
3917 *
3918 * Bring the hw back to operating state (all asics).
3919 * Returns 0 for success or an error on failure.
3920 * Called at driver resume.
3921 */
3922int amdgpu_device_resume(struct drm_device *dev, bool fbcon)
3923{
3924	struct amdgpu_device *adev = drm_to_adev(dev);
 
 
3925	int r = 0;
3926
3927	if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
3928		return 0;
3929
3930	if (adev->in_s0ix)
3931		amdgpu_gfx_state_change_set(adev, sGpuChangeState_D0Entry);
 
 
 
 
 
 
 
 
3932
3933	/* post card */
3934	if (amdgpu_device_need_post(adev)) {
3935		r = amdgpu_device_asic_init(adev);
3936		if (r)
3937			dev_err(adev->dev, "amdgpu asic init failed\n");
3938	}
3939
3940	r = amdgpu_device_ip_resume(adev);
3941	if (r) {
3942		dev_err(adev->dev, "amdgpu_device_ip_resume failed (%d).\n", r);
3943		return r;
 
 
 
 
 
 
 
3944	}
3945	amdgpu_fence_driver_hw_init(adev);
3946
3947	r = amdgpu_device_ip_late_init(adev);
3948	if (r)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3949		return r;
3950
3951	queue_delayed_work(system_wq, &adev->delayed_init_work,
3952			   msecs_to_jiffies(AMDGPU_RESUME_MS));
3953
3954	if (!adev->in_s0ix) {
3955		r = amdgpu_amdkfd_resume(adev, adev->in_runpm);
3956		if (r)
3957			return r;
 
 
 
 
 
 
3958	}
3959
3960	/* Make sure IB tests flushed */
3961	flush_delayed_work(&adev->delayed_init_work);
3962
3963	if (fbcon)
3964		amdgpu_fbdev_set_suspend(adev, 0);
3965
3966	drm_kms_helper_poll_enable(dev);
3967
3968	amdgpu_ras_resume(adev);
3969
3970	/*
3971	 * Most of the connector probing functions try to acquire runtime pm
3972	 * refs to ensure that the GPU is powered on when connector polling is
3973	 * performed. Since we're calling this from a runtime PM callback,
3974	 * trying to acquire rpm refs will cause us to deadlock.
3975	 *
3976	 * Since we're guaranteed to be holding the rpm lock, it's safe to
3977	 * temporarily disable the rpm helpers so this doesn't deadlock us.
3978	 */
3979#ifdef CONFIG_PM
3980	dev->dev->power.disable_depth++;
3981#endif
3982	if (!amdgpu_device_has_dc_support(adev))
3983		drm_helper_hpd_irq_event(dev);
3984	else
3985		drm_kms_helper_hotplug_event(dev);
3986#ifdef CONFIG_PM
3987	dev->dev->power.disable_depth--;
3988#endif
3989	adev->in_suspend = false;
3990
3991	if (amdgpu_acpi_smart_shift_update(dev, AMDGPU_SS_DEV_D0))
3992		DRM_WARN("smart shift update failed\n");
3993
3994	return 0;
 
 
 
 
3995}
3996
3997/**
3998 * amdgpu_device_ip_check_soft_reset - did soft reset succeed
3999 *
4000 * @adev: amdgpu_device pointer
4001 *
4002 * The list of all the hardware IPs that make up the asic is walked and
4003 * the check_soft_reset callbacks are run.  check_soft_reset determines
4004 * if the asic is still hung or not.
4005 * Returns true if any of the IPs are still in a hung state, false if not.
4006 */
4007static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)
4008{
4009	int i;
4010	bool asic_hang = false;
4011
4012	if (amdgpu_sriov_vf(adev))
4013		return true;
4014
4015	if (amdgpu_asic_need_full_reset(adev))
4016		return true;
4017
4018	for (i = 0; i < adev->num_ip_blocks; i++) {
4019		if (!adev->ip_blocks[i].status.valid)
4020			continue;
4021		if (adev->ip_blocks[i].version->funcs->check_soft_reset)
4022			adev->ip_blocks[i].status.hang =
4023				adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
4024		if (adev->ip_blocks[i].status.hang) {
4025			dev_info(adev->dev, "IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
4026			asic_hang = true;
4027		}
4028	}
4029	return asic_hang;
4030}
4031
4032/**
4033 * amdgpu_device_ip_pre_soft_reset - prepare for soft reset
4034 *
4035 * @adev: amdgpu_device pointer
4036 *
4037 * The list of all the hardware IPs that make up the asic is walked and the
4038 * pre_soft_reset callbacks are run if the block is hung.  pre_soft_reset
4039 * handles any IP specific hardware or software state changes that are
4040 * necessary for a soft reset to succeed.
4041 * Returns 0 on success, negative error code on failure.
4042 */
4043static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)
4044{
4045	int i, r = 0;
4046
4047	for (i = 0; i < adev->num_ip_blocks; i++) {
4048		if (!adev->ip_blocks[i].status.valid)
4049			continue;
4050		if (adev->ip_blocks[i].status.hang &&
4051		    adev->ip_blocks[i].version->funcs->pre_soft_reset) {
4052			r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
4053			if (r)
4054				return r;
4055		}
4056	}
4057
4058	return 0;
4059}
4060
4061/**
4062 * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed
4063 *
4064 * @adev: amdgpu_device pointer
4065 *
4066 * Some hardware IPs cannot be soft reset.  If they are hung, a full gpu
4067 * reset is necessary to recover.
4068 * Returns true if a full asic reset is required, false if not.
4069 */
4070static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)
4071{
4072	int i;
4073
4074	if (amdgpu_asic_need_full_reset(adev))
4075		return true;
4076
4077	for (i = 0; i < adev->num_ip_blocks; i++) {
4078		if (!adev->ip_blocks[i].status.valid)
4079			continue;
4080		if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
4081		    (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
4082		    (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
4083		    (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
4084		     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
4085			if (adev->ip_blocks[i].status.hang) {
4086				dev_info(adev->dev, "Some block need full reset!\n");
4087				return true;
4088			}
4089		}
4090	}
4091	return false;
4092}
4093
4094/**
4095 * amdgpu_device_ip_soft_reset - do a soft reset
4096 *
4097 * @adev: amdgpu_device pointer
4098 *
4099 * The list of all the hardware IPs that make up the asic is walked and the
4100 * soft_reset callbacks are run if the block is hung.  soft_reset handles any
4101 * IP specific hardware or software state changes that are necessary to soft
4102 * reset the IP.
4103 * Returns 0 on success, negative error code on failure.
4104 */
4105static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)
4106{
4107	int i, r = 0;
4108
4109	for (i = 0; i < adev->num_ip_blocks; i++) {
4110		if (!adev->ip_blocks[i].status.valid)
4111			continue;
4112		if (adev->ip_blocks[i].status.hang &&
4113		    adev->ip_blocks[i].version->funcs->soft_reset) {
4114			r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
4115			if (r)
4116				return r;
4117		}
4118	}
4119
4120	return 0;
4121}
4122
4123/**
4124 * amdgpu_device_ip_post_soft_reset - clean up from soft reset
4125 *
4126 * @adev: amdgpu_device pointer
4127 *
4128 * The list of all the hardware IPs that make up the asic is walked and the
4129 * post_soft_reset callbacks are run if the asic was hung.  post_soft_reset
4130 * handles any IP specific hardware or software state changes that are
4131 * necessary after the IP has been soft reset.
4132 * Returns 0 on success, negative error code on failure.
4133 */
4134static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)
4135{
4136	int i, r = 0;
4137
4138	for (i = 0; i < adev->num_ip_blocks; i++) {
4139		if (!adev->ip_blocks[i].status.valid)
4140			continue;
4141		if (adev->ip_blocks[i].status.hang &&
4142		    adev->ip_blocks[i].version->funcs->post_soft_reset)
4143			r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
4144		if (r)
4145			return r;
4146	}
4147
4148	return 0;
4149}
4150
4151/**
4152 * amdgpu_device_recover_vram - Recover some VRAM contents
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4153 *
4154 * @adev: amdgpu_device pointer
4155 *
4156 * Restores the contents of VRAM buffers from the shadows in GTT.  Used to
4157 * restore things like GPUVM page tables after a GPU reset where
4158 * the contents of VRAM might be lost.
4159 *
4160 * Returns:
4161 * 0 on success, negative error code on failure.
4162 */
4163static int amdgpu_device_recover_vram(struct amdgpu_device *adev)
4164{
 
 
4165	struct dma_fence *fence = NULL, *next = NULL;
4166	struct amdgpu_bo *shadow;
4167	struct amdgpu_bo_vm *vmbo;
4168	long r = 1, tmo;
4169
4170	if (amdgpu_sriov_runtime(adev))
4171		tmo = msecs_to_jiffies(8000);
4172	else
4173		tmo = msecs_to_jiffies(100);
4174
4175	dev_info(adev->dev, "recover vram bo from shadow start\n");
4176	mutex_lock(&adev->shadow_list_lock);
4177	list_for_each_entry(vmbo, &adev->shadow_list, shadow_list) {
4178		shadow = &vmbo->bo;
4179		/* No need to recover an evicted BO */
4180		if (shadow->tbo.resource->mem_type != TTM_PL_TT ||
4181		    shadow->tbo.resource->start == AMDGPU_BO_INVALID_OFFSET ||
4182		    shadow->parent->tbo.resource->mem_type != TTM_PL_VRAM)
4183			continue;
4184
4185		r = amdgpu_bo_restore_shadow(shadow, &next);
4186		if (r)
4187			break;
4188
4189		if (fence) {
4190			tmo = dma_fence_wait_timeout(fence, false, tmo);
4191			dma_fence_put(fence);
4192			fence = next;
4193			if (tmo == 0) {
4194				r = -ETIMEDOUT;
4195				break;
4196			} else if (tmo < 0) {
4197				r = tmo;
4198				break;
4199			}
4200		} else {
4201			fence = next;
4202		}
 
 
 
4203	}
4204	mutex_unlock(&adev->shadow_list_lock);
4205
4206	if (fence)
4207		tmo = dma_fence_wait_timeout(fence, false, tmo);
 
 
 
 
 
 
4208	dma_fence_put(fence);
4209
4210	if (r < 0 || tmo <= 0) {
4211		dev_err(adev->dev, "recover vram bo from shadow failed, r is %ld, tmo is %ld\n", r, tmo);
4212		return -EIO;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4213	}
4214
4215	dev_info(adev->dev, "recover vram bo from shadow done\n");
4216	return 0;
 
 
4217}
4218
4219
4220/**
4221 * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
4222 *
4223 * @adev: amdgpu_device pointer
4224 * @from_hypervisor: request from hypervisor
4225 *
4226 * do VF FLR and reinitialize Asic
4227 * return 0 means succeeded otherwise failed
4228 */
4229static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
4230				     bool from_hypervisor)
4231{
4232	int r;
4233
4234	if (from_hypervisor)
4235		r = amdgpu_virt_request_full_gpu(adev, true);
4236	else
4237		r = amdgpu_virt_reset_gpu(adev);
4238	if (r)
4239		return r;
4240
4241	amdgpu_amdkfd_pre_reset(adev);
4242
4243	/* Resume IP prior to SMC */
4244	r = amdgpu_device_ip_reinit_early_sriov(adev);
4245	if (r)
4246		goto error;
4247
4248	amdgpu_virt_init_data_exchange(adev);
4249	/* we need recover gart prior to run SMC/CP/SDMA resume */
4250	amdgpu_gtt_mgr_recover(ttm_manager_type(&adev->mman.bdev, TTM_PL_TT));
4251
4252	r = amdgpu_device_fw_loading(adev);
4253	if (r)
4254		return r;
4255
4256	/* now we are okay to resume SMC/CP/SDMA */
4257	r = amdgpu_device_ip_reinit_late_sriov(adev);
 
4258	if (r)
4259		goto error;
4260
4261	amdgpu_irq_gpu_reset_resume_helper(adev);
4262	r = amdgpu_ib_ring_tests(adev);
4263	amdgpu_amdkfd_post_reset(adev);
4264
4265error:
4266	if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
4267		amdgpu_inc_vram_lost(adev);
4268		r = amdgpu_device_recover_vram(adev);
4269	}
4270	amdgpu_virt_release_full_gpu(adev, true);
 
4271
4272	return r;
4273}
4274
4275/**
4276 * amdgpu_device_has_job_running - check if there is any job in mirror list
4277 *
4278 * @adev: amdgpu_device pointer
 
 
4279 *
4280 * check if there is any job in mirror list
 
4281 */
4282bool amdgpu_device_has_job_running(struct amdgpu_device *adev)
 
4283{
4284	int i;
4285	struct drm_sched_job *job;
4286
4287	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
4288		struct amdgpu_ring *ring = adev->rings[i];
4289
4290		if (!ring || !ring->sched.thread)
4291			continue;
4292
4293		spin_lock(&ring->sched.job_list_lock);
4294		job = list_first_entry_or_null(&ring->sched.pending_list,
4295					       struct drm_sched_job, list);
4296		spin_unlock(&ring->sched.job_list_lock);
4297		if (job)
4298			return true;
4299	}
4300	return false;
4301}
4302
4303/**
4304 * amdgpu_device_should_recover_gpu - check if we should try GPU recovery
4305 *
4306 * @adev: amdgpu_device pointer
4307 *
4308 * Check amdgpu_gpu_recovery and SRIOV status to see if we should try to recover
4309 * a hung GPU.
4310 */
4311bool amdgpu_device_should_recover_gpu(struct amdgpu_device *adev)
4312{
4313	if (!amdgpu_device_ip_check_soft_reset(adev)) {
4314		dev_info(adev->dev, "Timeout, but no hardware hang detected.\n");
4315		return false;
4316	}
4317
4318	if (amdgpu_gpu_recovery == 0)
4319		goto disabled;
4320
4321	if (amdgpu_sriov_vf(adev))
4322		return true;
4323
4324	if (amdgpu_gpu_recovery == -1) {
4325		switch (adev->asic_type) {
4326		case CHIP_BONAIRE:
4327		case CHIP_HAWAII:
4328		case CHIP_TOPAZ:
4329		case CHIP_TONGA:
4330		case CHIP_FIJI:
4331		case CHIP_POLARIS10:
4332		case CHIP_POLARIS11:
4333		case CHIP_POLARIS12:
4334		case CHIP_VEGAM:
4335		case CHIP_VEGA20:
4336		case CHIP_VEGA10:
4337		case CHIP_VEGA12:
4338		case CHIP_RAVEN:
4339		case CHIP_ARCTURUS:
4340		case CHIP_RENOIR:
4341		case CHIP_NAVI10:
4342		case CHIP_NAVI14:
4343		case CHIP_NAVI12:
4344		case CHIP_SIENNA_CICHLID:
4345		case CHIP_NAVY_FLOUNDER:
4346		case CHIP_DIMGREY_CAVEFISH:
4347		case CHIP_BEIGE_GOBY:
4348		case CHIP_VANGOGH:
4349		case CHIP_ALDEBARAN:
4350			break;
4351		default:
4352			goto disabled;
4353		}
4354	}
4355
4356	return true;
4357
4358disabled:
4359		dev_info(adev->dev, "GPU recovery disabled.\n");
4360		return false;
4361}
4362
4363int amdgpu_device_mode1_reset(struct amdgpu_device *adev)
4364{
4365        u32 i;
4366        int ret = 0;
4367
4368        amdgpu_atombios_scratch_regs_engine_hung(adev, true);
4369
4370        dev_info(adev->dev, "GPU mode1 reset\n");
4371
4372        /* disable BM */
4373        pci_clear_master(adev->pdev);
4374
4375        amdgpu_device_cache_pci_state(adev->pdev);
4376
4377        if (amdgpu_dpm_is_mode1_reset_supported(adev)) {
4378                dev_info(adev->dev, "GPU smu mode1 reset\n");
4379                ret = amdgpu_dpm_mode1_reset(adev);
4380        } else {
4381                dev_info(adev->dev, "GPU psp mode1 reset\n");
4382                ret = psp_gpu_reset(adev);
4383        }
4384
4385        if (ret)
4386                dev_err(adev->dev, "GPU mode1 reset failed\n");
4387
4388        amdgpu_device_load_pci_state(adev->pdev);
4389
4390        /* wait for asic to come out of reset */
4391        for (i = 0; i < adev->usec_timeout; i++) {
4392                u32 memsize = adev->nbio.funcs->get_memsize(adev);
4393
4394                if (memsize != 0xffffffff)
4395                        break;
4396                udelay(1);
4397        }
4398
4399        amdgpu_atombios_scratch_regs_engine_hung(adev, false);
4400        return ret;
4401}
4402
4403int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,
4404				 struct amdgpu_reset_context *reset_context)
4405{
4406	int i, r = 0;
4407	struct amdgpu_job *job = NULL;
4408	bool need_full_reset =
4409		test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
4410
4411	if (reset_context->reset_req_dev == adev)
4412		job = reset_context->job;
4413
4414	/* no need to dump if device is not in good state during probe period */
4415	if (!adev->gmc.xgmi.pending_reset)
4416		amdgpu_debugfs_wait_dump(adev);
4417
4418	if (amdgpu_sriov_vf(adev)) {
4419		/* stop the data exchange thread */
4420		amdgpu_virt_fini_data_exchange(adev);
4421	}
4422
4423	/* block all schedulers and reset given job's ring */
4424	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
4425		struct amdgpu_ring *ring = adev->rings[i];
4426
4427		if (!ring || !ring->sched.thread)
4428			continue;
4429
4430		/* after all hw jobs are reset, hw fence is meaningless, so force_completion */
4431		amdgpu_fence_driver_force_completion(ring);
4432	}
4433
4434	if(job)
4435		drm_sched_increase_karma(&job->base);
4436
4437	r = amdgpu_reset_prepare_hwcontext(adev, reset_context);
4438	/* If reset handler not implemented, continue; otherwise return */
4439	if (r == -ENOSYS)
4440		r = 0;
4441	else
4442		return r;
4443
4444	/* Don't suspend on bare metal if we are not going to HW reset the ASIC */
4445	if (!amdgpu_sriov_vf(adev)) {
4446
4447		if (!need_full_reset)
4448			need_full_reset = amdgpu_device_ip_need_full_reset(adev);
4449
4450		if (!need_full_reset) {
4451			amdgpu_device_ip_pre_soft_reset(adev);
4452			r = amdgpu_device_ip_soft_reset(adev);
4453			amdgpu_device_ip_post_soft_reset(adev);
4454			if (r || amdgpu_device_ip_check_soft_reset(adev)) {
4455				dev_info(adev->dev, "soft reset failed, will fallback to full reset!\n");
4456				need_full_reset = true;
4457			}
4458		}
4459
4460		if (need_full_reset)
4461			r = amdgpu_device_ip_suspend(adev);
4462		if (need_full_reset)
4463			set_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
4464		else
4465			clear_bit(AMDGPU_NEED_FULL_RESET,
4466				  &reset_context->flags);
4467	}
4468
4469	return r;
4470}
4471
4472int amdgpu_do_asic_reset(struct list_head *device_list_handle,
4473			 struct amdgpu_reset_context *reset_context)
4474{
4475	struct amdgpu_device *tmp_adev = NULL;
4476	bool need_full_reset, skip_hw_reset, vram_lost = false;
4477	int r = 0;
4478
4479	/* Try reset handler method first */
4480	tmp_adev = list_first_entry(device_list_handle, struct amdgpu_device,
4481				    reset_list);
4482	r = amdgpu_reset_perform_reset(tmp_adev, reset_context);
4483	/* If reset handler not implemented, continue; otherwise return */
4484	if (r == -ENOSYS)
4485		r = 0;
4486	else
4487		return r;
4488
4489	/* Reset handler not implemented, use the default method */
4490	need_full_reset =
4491		test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
4492	skip_hw_reset = test_bit(AMDGPU_SKIP_HW_RESET, &reset_context->flags);
4493
4494	/*
4495	 * ASIC reset has to be done on all XGMI hive nodes ASAP
4496	 * to allow proper links negotiation in FW (within 1 sec)
4497	 */
4498	if (!skip_hw_reset && need_full_reset) {
4499		list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
4500			/* For XGMI run all resets in parallel to speed up the process */
4501			if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
4502				tmp_adev->gmc.xgmi.pending_reset = false;
4503				if (!queue_work(system_unbound_wq, &tmp_adev->xgmi_reset_work))
4504					r = -EALREADY;
4505			} else
4506				r = amdgpu_asic_reset(tmp_adev);
4507
4508			if (r) {
4509				dev_err(tmp_adev->dev, "ASIC reset failed with error, %d for drm dev, %s",
4510					 r, adev_to_drm(tmp_adev)->unique);
4511				break;
4512			}
4513		}
4514
4515		/* For XGMI wait for all resets to complete before proceed */
4516		if (!r) {
4517			list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
4518				if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
4519					flush_work(&tmp_adev->xgmi_reset_work);
4520					r = tmp_adev->asic_reset_res;
4521					if (r)
4522						break;
4523				}
4524			}
4525		}
4526	}
4527
4528	if (!r && amdgpu_ras_intr_triggered()) {
4529		list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
4530			if (tmp_adev->mmhub.ras_funcs &&
4531			    tmp_adev->mmhub.ras_funcs->reset_ras_error_count)
4532				tmp_adev->mmhub.ras_funcs->reset_ras_error_count(tmp_adev);
4533		}
4534
4535		amdgpu_ras_intr_cleared();
4536	}
4537
4538	list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
4539		if (need_full_reset) {
4540			/* post card */
4541			r = amdgpu_device_asic_init(tmp_adev);
4542			if (r) {
4543				dev_warn(tmp_adev->dev, "asic atom init failed!");
4544			} else {
4545				dev_info(tmp_adev->dev, "GPU reset succeeded, trying to resume\n");
4546				r = amdgpu_amdkfd_resume_iommu(tmp_adev);
4547				if (r)
4548					goto out;
4549
4550				r = amdgpu_device_ip_resume_phase1(tmp_adev);
4551				if (r)
4552					goto out;
4553
4554				vram_lost = amdgpu_device_check_vram_lost(tmp_adev);
4555				if (vram_lost) {
4556					DRM_INFO("VRAM is lost due to GPU reset!\n");
4557					amdgpu_inc_vram_lost(tmp_adev);
4558				}
4559
4560				r = amdgpu_gtt_mgr_recover(ttm_manager_type(&tmp_adev->mman.bdev, TTM_PL_TT));
4561				if (r)
4562					goto out;
4563
4564				r = amdgpu_device_fw_loading(tmp_adev);
4565				if (r)
4566					return r;
4567
4568				r = amdgpu_device_ip_resume_phase2(tmp_adev);
4569				if (r)
4570					goto out;
4571
4572				if (vram_lost)
4573					amdgpu_device_fill_reset_magic(tmp_adev);
4574
4575				/*
4576				 * Add this ASIC as tracked as reset was already
4577				 * complete successfully.
4578				 */
4579				amdgpu_register_gpu_instance(tmp_adev);
4580
4581				if (!reset_context->hive &&
4582				    tmp_adev->gmc.xgmi.num_physical_nodes > 1)
4583					amdgpu_xgmi_add_device(tmp_adev);
4584
4585				r = amdgpu_device_ip_late_init(tmp_adev);
4586				if (r)
4587					goto out;
4588
4589				amdgpu_fbdev_set_suspend(tmp_adev, 0);
4590
4591				/*
4592				 * The GPU enters bad state once faulty pages
4593				 * by ECC has reached the threshold, and ras
4594				 * recovery is scheduled next. So add one check
4595				 * here to break recovery if it indeed exceeds
4596				 * bad page threshold, and remind user to
4597				 * retire this GPU or setting one bigger
4598				 * bad_page_threshold value to fix this once
4599				 * probing driver again.
4600				 */
4601				if (!amdgpu_ras_eeprom_check_err_threshold(tmp_adev)) {
4602					/* must succeed. */
4603					amdgpu_ras_resume(tmp_adev);
4604				} else {
4605					r = -EINVAL;
4606					goto out;
4607				}
4608
4609				/* Update PSP FW topology after reset */
4610				if (reset_context->hive &&
4611				    tmp_adev->gmc.xgmi.num_physical_nodes > 1)
4612					r = amdgpu_xgmi_update_topology(
4613						reset_context->hive, tmp_adev);
4614			}
4615		}
4616
4617out:
4618		if (!r) {
4619			amdgpu_irq_gpu_reset_resume_helper(tmp_adev);
4620			r = amdgpu_ib_ring_tests(tmp_adev);
4621			if (r) {
4622				dev_err(tmp_adev->dev, "ib ring test failed (%d).\n", r);
4623				need_full_reset = true;
4624				r = -EAGAIN;
4625				goto end;
4626			}
4627		}
4628
4629		if (!r)
4630			r = amdgpu_device_recover_vram(tmp_adev);
4631		else
4632			tmp_adev->asic_reset_res = r;
4633	}
4634
4635end:
4636	if (need_full_reset)
4637		set_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
4638	else
4639		clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
4640	return r;
4641}
4642
4643static bool amdgpu_device_lock_adev(struct amdgpu_device *adev,
4644				struct amdgpu_hive_info *hive)
4645{
4646	if (atomic_cmpxchg(&adev->in_gpu_reset, 0, 1) != 0)
4647		return false;
4648
4649	if (hive) {
4650		down_write_nest_lock(&adev->reset_sem, &hive->hive_lock);
4651	} else {
4652		down_write(&adev->reset_sem);
4653	}
4654
4655	switch (amdgpu_asic_reset_method(adev)) {
4656	case AMD_RESET_METHOD_MODE1:
4657		adev->mp1_state = PP_MP1_STATE_SHUTDOWN;
4658		break;
4659	case AMD_RESET_METHOD_MODE2:
4660		adev->mp1_state = PP_MP1_STATE_RESET;
4661		break;
4662	default:
4663		adev->mp1_state = PP_MP1_STATE_NONE;
4664		break;
4665	}
4666
4667	return true;
4668}
4669
4670static void amdgpu_device_unlock_adev(struct amdgpu_device *adev)
4671{
4672	amdgpu_vf_error_trans_all(adev);
4673	adev->mp1_state = PP_MP1_STATE_NONE;
4674	atomic_set(&adev->in_gpu_reset, 0);
4675	up_write(&adev->reset_sem);
4676}
4677
4678/*
4679 * to lockup a list of amdgpu devices in a hive safely, if not a hive
4680 * with multiple nodes, it will be similar as amdgpu_device_lock_adev.
4681 *
4682 * unlock won't require roll back.
4683 */
4684static int amdgpu_device_lock_hive_adev(struct amdgpu_device *adev, struct amdgpu_hive_info *hive)
4685{
4686	struct amdgpu_device *tmp_adev = NULL;
4687
4688	if (adev->gmc.xgmi.num_physical_nodes > 1) {
4689		if (!hive) {
4690			dev_err(adev->dev, "Hive is NULL while device has multiple xgmi nodes");
4691			return -ENODEV;
4692		}
4693		list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
4694			if (!amdgpu_device_lock_adev(tmp_adev, hive))
4695				goto roll_back;
4696		}
4697	} else if (!amdgpu_device_lock_adev(adev, hive))
4698		return -EAGAIN;
4699
4700	return 0;
4701roll_back:
4702	if (!list_is_first(&tmp_adev->gmc.xgmi.head, &hive->device_list)) {
4703		/*
4704		 * if the lockup iteration break in the middle of a hive,
4705		 * it may means there may has a race issue,
4706		 * or a hive device locked up independently.
4707		 * we may be in trouble and may not, so will try to roll back
4708		 * the lock and give out a warnning.
4709		 */
4710		dev_warn(tmp_adev->dev, "Hive lock iteration broke in the middle. Rolling back to unlock");
4711		list_for_each_entry_continue_reverse(tmp_adev, &hive->device_list, gmc.xgmi.head) {
4712			amdgpu_device_unlock_adev(tmp_adev);
4713		}
4714	}
4715	return -EAGAIN;
4716}
4717
4718static void amdgpu_device_resume_display_audio(struct amdgpu_device *adev)
4719{
4720	struct pci_dev *p = NULL;
4721
4722	p = pci_get_domain_bus_and_slot(pci_domain_nr(adev->pdev->bus),
4723			adev->pdev->bus->number, 1);
4724	if (p) {
4725		pm_runtime_enable(&(p->dev));
4726		pm_runtime_resume(&(p->dev));
4727	}
4728}
4729
4730static int amdgpu_device_suspend_display_audio(struct amdgpu_device *adev)
4731{
4732	enum amd_reset_method reset_method;
4733	struct pci_dev *p = NULL;
4734	u64 expires;
4735
4736	/*
4737	 * For now, only BACO and mode1 reset are confirmed
4738	 * to suffer the audio issue without proper suspended.
4739	 */
4740	reset_method = amdgpu_asic_reset_method(adev);
4741	if ((reset_method != AMD_RESET_METHOD_BACO) &&
4742	     (reset_method != AMD_RESET_METHOD_MODE1))
4743		return -EINVAL;
4744
4745	p = pci_get_domain_bus_and_slot(pci_domain_nr(adev->pdev->bus),
4746			adev->pdev->bus->number, 1);
4747	if (!p)
4748		return -ENODEV;
4749
4750	expires = pm_runtime_autosuspend_expiration(&(p->dev));
4751	if (!expires)
4752		/*
4753		 * If we cannot get the audio device autosuspend delay,
4754		 * a fixed 4S interval will be used. Considering 3S is
4755		 * the audio controller default autosuspend delay setting.
4756		 * 4S used here is guaranteed to cover that.
4757		 */
4758		expires = ktime_get_mono_fast_ns() + NSEC_PER_SEC * 4ULL;
4759
4760	while (!pm_runtime_status_suspended(&(p->dev))) {
4761		if (!pm_runtime_suspend(&(p->dev)))
4762			break;
4763
4764		if (expires < ktime_get_mono_fast_ns()) {
4765			dev_warn(adev->dev, "failed to suspend display audio\n");
4766			/* TODO: abort the succeeding gpu reset? */
4767			return -ETIMEDOUT;
4768		}
4769	}
4770
4771	pm_runtime_disable(&(p->dev));
4772
4773	return 0;
4774}
4775
4776static void amdgpu_device_recheck_guilty_jobs(
4777	struct amdgpu_device *adev, struct list_head *device_list_handle,
4778	struct amdgpu_reset_context *reset_context)
4779{
4780	int i, r = 0;
4781
4782	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
4783		struct amdgpu_ring *ring = adev->rings[i];
4784		int ret = 0;
4785		struct drm_sched_job *s_job;
4786
4787		if (!ring || !ring->sched.thread)
4788			continue;
4789
4790		s_job = list_first_entry_or_null(&ring->sched.pending_list,
4791				struct drm_sched_job, list);
4792		if (s_job == NULL)
4793			continue;
4794
4795		/* clear job's guilty and depend the folowing step to decide the real one */
4796		drm_sched_reset_karma(s_job);
4797		drm_sched_resubmit_jobs_ext(&ring->sched, 1);
4798
4799		ret = dma_fence_wait_timeout(s_job->s_fence->parent, false, ring->sched.timeout);
4800		if (ret == 0) { /* timeout */
4801			DRM_ERROR("Found the real bad job! ring:%s, job_id:%llx\n",
4802						ring->sched.name, s_job->id);
4803
4804			/* set guilty */
4805			drm_sched_increase_karma(s_job);
4806retry:
4807			/* do hw reset */
4808			if (amdgpu_sriov_vf(adev)) {
4809				amdgpu_virt_fini_data_exchange(adev);
4810				r = amdgpu_device_reset_sriov(adev, false);
4811				if (r)
4812					adev->asic_reset_res = r;
4813			} else {
4814				clear_bit(AMDGPU_SKIP_HW_RESET,
4815					  &reset_context->flags);
4816				r = amdgpu_do_asic_reset(device_list_handle,
4817							 reset_context);
4818				if (r && r == -EAGAIN)
4819					goto retry;
4820			}
4821
4822			/*
4823			 * add reset counter so that the following
4824			 * resubmitted job could flush vmid
4825			 */
4826			atomic_inc(&adev->gpu_reset_counter);
4827			continue;
4828		}
4829
4830		/* got the hw fence, signal finished fence */
4831		atomic_dec(ring->sched.score);
4832		dma_fence_get(&s_job->s_fence->finished);
4833		dma_fence_signal(&s_job->s_fence->finished);
4834		dma_fence_put(&s_job->s_fence->finished);
4835
4836		/* remove node from list and free the job */
4837		spin_lock(&ring->sched.job_list_lock);
4838		list_del_init(&s_job->list);
4839		spin_unlock(&ring->sched.job_list_lock);
4840		ring->sched.ops->free_job(s_job);
4841	}
4842}
4843
4844/**
4845 * amdgpu_device_gpu_recover - reset the asic and recover scheduler
4846 *
4847 * @adev: amdgpu_device pointer
4848 * @job: which job trigger hang
4849 *
4850 * Attempt to reset the GPU if it has hung (all asics).
4851 * Attempt to do soft-reset or full-reset and reinitialize Asic
4852 * Returns 0 for success or an error on failure.
4853 */
4854
4855int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
4856			      struct amdgpu_job *job)
4857{
4858	struct list_head device_list, *device_list_handle =  NULL;
4859	bool job_signaled = false;
4860	struct amdgpu_hive_info *hive = NULL;
4861	struct amdgpu_device *tmp_adev = NULL;
4862	int i, r = 0;
4863	bool need_emergency_restart = false;
4864	bool audio_suspended = false;
4865	int tmp_vram_lost_counter;
4866	struct amdgpu_reset_context reset_context;
4867
4868	memset(&reset_context, 0, sizeof(reset_context));
4869
4870	/*
4871	 * Special case: RAS triggered and full reset isn't supported
4872	 */
4873	need_emergency_restart = amdgpu_ras_need_emergency_restart(adev);
4874
4875	/*
4876	 * Flush RAM to disk so that after reboot
4877	 * the user can read log and see why the system rebooted.
4878	 */
4879	if (need_emergency_restart && amdgpu_ras_get_context(adev)->reboot) {
4880		DRM_WARN("Emergency reboot.");
4881
4882		ksys_sync_helper();
4883		emergency_restart();
4884	}
4885
4886	dev_info(adev->dev, "GPU %s begin!\n",
4887		need_emergency_restart ? "jobs stop":"reset");
4888
4889	/*
4890	 * Here we trylock to avoid chain of resets executing from
4891	 * either trigger by jobs on different adevs in XGMI hive or jobs on
4892	 * different schedulers for same device while this TO handler is running.
4893	 * We always reset all schedulers for device and all devices for XGMI
4894	 * hive so that should take care of them too.
4895	 */
4896	hive = amdgpu_get_xgmi_hive(adev);
4897	if (hive) {
4898		if (atomic_cmpxchg(&hive->in_reset, 0, 1) != 0) {
4899			DRM_INFO("Bailing on TDR for s_job:%llx, hive: %llx as another already in progress",
4900				job ? job->base.id : -1, hive->hive_id);
4901			amdgpu_put_xgmi_hive(hive);
4902			if (job)
4903				drm_sched_increase_karma(&job->base);
4904			return 0;
4905		}
4906		mutex_lock(&hive->hive_lock);
4907	}
4908
4909	reset_context.method = AMD_RESET_METHOD_NONE;
4910	reset_context.reset_req_dev = adev;
4911	reset_context.job = job;
4912	reset_context.hive = hive;
4913	clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
4914
4915	/*
4916	 * lock the device before we try to operate the linked list
4917	 * if didn't get the device lock, don't touch the linked list since
4918	 * others may iterating it.
4919	 */
4920	r = amdgpu_device_lock_hive_adev(adev, hive);
4921	if (r) {
4922		dev_info(adev->dev, "Bailing on TDR for s_job:%llx, as another already in progress",
4923					job ? job->base.id : -1);
4924
4925		/* even we skipped this reset, still need to set the job to guilty */
4926		if (job)
4927			drm_sched_increase_karma(&job->base);
4928		goto skip_recovery;
4929	}
4930
4931	/*
4932	 * Build list of devices to reset.
4933	 * In case we are in XGMI hive mode, resort the device list
4934	 * to put adev in the 1st position.
4935	 */
4936	INIT_LIST_HEAD(&device_list);
4937	if (adev->gmc.xgmi.num_physical_nodes > 1) {
4938		list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head)
4939			list_add_tail(&tmp_adev->reset_list, &device_list);
4940		if (!list_is_first(&adev->reset_list, &device_list))
4941			list_rotate_to_front(&adev->reset_list, &device_list);
4942		device_list_handle = &device_list;
4943	} else {
4944		list_add_tail(&adev->reset_list, &device_list);
4945		device_list_handle = &device_list;
4946	}
4947
4948	/* block all schedulers and reset given job's ring */
4949	list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
4950		/*
4951		 * Try to put the audio codec into suspend state
4952		 * before gpu reset started.
4953		 *
4954		 * Due to the power domain of the graphics device
4955		 * is shared with AZ power domain. Without this,
4956		 * we may change the audio hardware from behind
4957		 * the audio driver's back. That will trigger
4958		 * some audio codec errors.
4959		 */
4960		if (!amdgpu_device_suspend_display_audio(tmp_adev))
4961			audio_suspended = true;
4962
4963		amdgpu_ras_set_error_query_ready(tmp_adev, false);
4964
4965		cancel_delayed_work_sync(&tmp_adev->delayed_init_work);
4966
4967		if (!amdgpu_sriov_vf(tmp_adev))
4968			amdgpu_amdkfd_pre_reset(tmp_adev);
4969
4970		/*
4971		 * Mark these ASICs to be reseted as untracked first
4972		 * And add them back after reset completed
4973		 */
4974		amdgpu_unregister_gpu_instance(tmp_adev);
4975
4976		amdgpu_fbdev_set_suspend(tmp_adev, 1);
4977
4978		/* disable ras on ALL IPs */
4979		if (!need_emergency_restart &&
4980		      amdgpu_device_ip_need_full_reset(tmp_adev))
4981			amdgpu_ras_suspend(tmp_adev);
4982
4983		for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
4984			struct amdgpu_ring *ring = tmp_adev->rings[i];
4985
4986			if (!ring || !ring->sched.thread)
4987				continue;
4988
4989			drm_sched_stop(&ring->sched, job ? &job->base : NULL);
4990
4991			if (need_emergency_restart)
4992				amdgpu_job_stop_all_jobs_on_sched(&ring->sched);
4993		}
4994		atomic_inc(&tmp_adev->gpu_reset_counter);
4995	}
4996
4997	if (need_emergency_restart)
4998		goto skip_sched_resume;
4999
5000	/*
5001	 * Must check guilty signal here since after this point all old
5002	 * HW fences are force signaled.
5003	 *
5004	 * job->base holds a reference to parent fence
5005	 */
5006	if (job && job->base.s_fence->parent &&
5007	    dma_fence_is_signaled(job->base.s_fence->parent)) {
5008		job_signaled = true;
5009		dev_info(adev->dev, "Guilty job already signaled, skipping HW reset");
5010		goto skip_hw_reset;
5011	}
5012
5013retry:	/* Rest of adevs pre asic reset from XGMI hive. */
5014	list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5015		r = amdgpu_device_pre_asic_reset(tmp_adev, &reset_context);
5016		/*TODO Should we stop ?*/
5017		if (r) {
5018			dev_err(tmp_adev->dev, "GPU pre asic reset failed with err, %d for drm dev, %s ",
5019				  r, adev_to_drm(tmp_adev)->unique);
5020			tmp_adev->asic_reset_res = r;
5021		}
5022	}
5023
5024	tmp_vram_lost_counter = atomic_read(&((adev)->vram_lost_counter));
5025	/* Actual ASIC resets if needed.*/
5026	/* TODO Implement XGMI hive reset logic for SRIOV */
5027	if (amdgpu_sriov_vf(adev)) {
5028		r = amdgpu_device_reset_sriov(adev, job ? false : true);
5029		if (r)
5030			adev->asic_reset_res = r;
5031	} else {
5032		r = amdgpu_do_asic_reset(device_list_handle, &reset_context);
5033		if (r && r == -EAGAIN)
5034			goto retry;
5035	}
5036
5037skip_hw_reset:
5038
5039	/* Post ASIC reset for all devs .*/
5040	list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5041
5042		/*
5043		 * Sometimes a later bad compute job can block a good gfx job as gfx
5044		 * and compute ring share internal GC HW mutually. We add an additional
5045		 * guilty jobs recheck step to find the real guilty job, it synchronously
5046		 * submits and pends for the first job being signaled. If it gets timeout,
5047		 * we identify it as a real guilty job.
5048		 */
5049		if (amdgpu_gpu_recovery == 2 &&
5050			!(tmp_vram_lost_counter < atomic_read(&adev->vram_lost_counter)))
5051			amdgpu_device_recheck_guilty_jobs(
5052				tmp_adev, device_list_handle, &reset_context);
5053
5054		for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
5055			struct amdgpu_ring *ring = tmp_adev->rings[i];
5056
5057			if (!ring || !ring->sched.thread)
5058				continue;
5059
5060			/* No point to resubmit jobs if we didn't HW reset*/
5061			if (!tmp_adev->asic_reset_res && !job_signaled)
5062				drm_sched_resubmit_jobs(&ring->sched);
5063
5064			drm_sched_start(&ring->sched, !tmp_adev->asic_reset_res);
5065		}
5066
5067		if (!amdgpu_device_has_dc_support(tmp_adev) && !job_signaled) {
5068			drm_helper_resume_force_mode(adev_to_drm(tmp_adev));
5069		}
5070
5071		tmp_adev->asic_reset_res = 0;
5072
5073		if (r) {
5074			/* bad news, how to tell it to userspace ? */
5075			dev_info(tmp_adev->dev, "GPU reset(%d) failed\n", atomic_read(&tmp_adev->gpu_reset_counter));
5076			amdgpu_vf_error_put(tmp_adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r);
5077		} else {
5078			dev_info(tmp_adev->dev, "GPU reset(%d) succeeded!\n", atomic_read(&tmp_adev->gpu_reset_counter));
5079			if (amdgpu_acpi_smart_shift_update(adev_to_drm(tmp_adev), AMDGPU_SS_DEV_D0))
5080				DRM_WARN("smart shift update failed\n");
5081		}
5082	}
5083
5084skip_sched_resume:
5085	list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5086		/* unlock kfd: SRIOV would do it separately */
5087		if (!need_emergency_restart && !amdgpu_sriov_vf(tmp_adev))
5088	                amdgpu_amdkfd_post_reset(tmp_adev);
5089
5090		/* kfd_post_reset will do nothing if kfd device is not initialized,
5091		 * need to bring up kfd here if it's not be initialized before
5092		 */
5093		if (!adev->kfd.init_complete)
5094			amdgpu_amdkfd_device_init(adev);
5095
5096		if (audio_suspended)
5097			amdgpu_device_resume_display_audio(tmp_adev);
5098		amdgpu_device_unlock_adev(tmp_adev);
5099	}
5100
5101skip_recovery:
5102	if (hive) {
5103		atomic_set(&hive->in_reset, 0);
5104		mutex_unlock(&hive->hive_lock);
5105		amdgpu_put_xgmi_hive(hive);
5106	}
5107
5108	if (r && r != -EAGAIN)
5109		dev_info(adev->dev, "GPU reset end with ret = %d\n", r);
5110	return r;
5111}
5112
5113/**
5114 * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot
5115 *
5116 * @adev: amdgpu_device pointer
5117 *
5118 * Fetchs and stores in the driver the PCIE capabilities (gen speed
5119 * and lanes) of the slot the device is in. Handles APUs and
5120 * virtualized environments where PCIE config space may not be available.
5121 */
5122static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev)
5123{
5124	struct pci_dev *pdev;
5125	enum pci_bus_speed speed_cap, platform_speed_cap;
5126	enum pcie_link_width platform_link_width;
5127
5128	if (amdgpu_pcie_gen_cap)
5129		adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
5130
5131	if (amdgpu_pcie_lane_cap)
5132		adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
5133
5134	/* covers APUs as well */
5135	if (pci_is_root_bus(adev->pdev->bus)) {
5136		if (adev->pm.pcie_gen_mask == 0)
5137			adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
5138		if (adev->pm.pcie_mlw_mask == 0)
5139			adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
5140		return;
5141	}
5142
5143	if (adev->pm.pcie_gen_mask && adev->pm.pcie_mlw_mask)
5144		return;
5145
5146	pcie_bandwidth_available(adev->pdev, NULL,
5147				 &platform_speed_cap, &platform_link_width);
5148
5149	if (adev->pm.pcie_gen_mask == 0) {
5150		/* asic caps */
5151		pdev = adev->pdev;
5152		speed_cap = pcie_get_speed_cap(pdev);
5153		if (speed_cap == PCI_SPEED_UNKNOWN) {
5154			adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5155						  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5156						  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
 
 
 
 
 
 
 
5157		} else {
5158			if (speed_cap == PCIE_SPEED_32_0GT)
5159				adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5160							  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5161							  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
5162							  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4 |
5163							  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN5);
5164			else if (speed_cap == PCIE_SPEED_16_0GT)
5165				adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5166							  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5167							  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
5168							  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4);
5169			else if (speed_cap == PCIE_SPEED_8_0GT)
5170				adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5171							  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5172							  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
5173			else if (speed_cap == PCIE_SPEED_5_0GT)
5174				adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5175							  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2);
5176			else
5177				adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1;
5178		}
5179		/* platform caps */
5180		if (platform_speed_cap == PCI_SPEED_UNKNOWN) {
5181			adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5182						   CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
5183		} else {
5184			if (platform_speed_cap == PCIE_SPEED_32_0GT)
5185				adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5186							   CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5187							   CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
5188							   CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4 |
5189							   CAIL_PCIE_LINK_SPEED_SUPPORT_GEN5);
5190			else if (platform_speed_cap == PCIE_SPEED_16_0GT)
5191				adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5192							   CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5193							   CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
5194							   CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4);
5195			else if (platform_speed_cap == PCIE_SPEED_8_0GT)
5196				adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5197							   CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5198							   CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3);
5199			else if (platform_speed_cap == PCIE_SPEED_5_0GT)
5200				adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5201							   CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
5202			else
5203				adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
5204
5205		}
5206	}
5207	if (adev->pm.pcie_mlw_mask == 0) {
5208		if (platform_link_width == PCIE_LNK_WIDTH_UNKNOWN) {
5209			adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK;
5210		} else {
5211			switch (platform_link_width) {
5212			case PCIE_LNK_X32:
5213				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
5214							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
5215							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
5216							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
5217							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5218							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5219							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5220				break;
5221			case PCIE_LNK_X16:
5222				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
5223							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
5224							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
5225							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5226							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5227							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5228				break;
5229			case PCIE_LNK_X12:
5230				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
5231							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
5232							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5233							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5234							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5235				break;
5236			case PCIE_LNK_X8:
5237				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
5238							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5239							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5240							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5241				break;
5242			case PCIE_LNK_X4:
5243				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5244							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5245							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5246				break;
5247			case PCIE_LNK_X2:
5248				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5249							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5250				break;
5251			case PCIE_LNK_X1:
5252				adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;
5253				break;
5254			default:
5255				break;
5256			}
 
 
5257		}
5258	}
5259}
5260
5261int amdgpu_device_baco_enter(struct drm_device *dev)
5262{
5263	struct amdgpu_device *adev = drm_to_adev(dev);
5264	struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
5265
5266	if (!amdgpu_device_supports_baco(adev_to_drm(adev)))
5267		return -ENOTSUPP;
5268
5269	if (ras && adev->ras_enabled &&
5270	    adev->nbio.funcs->enable_doorbell_interrupt)
5271		adev->nbio.funcs->enable_doorbell_interrupt(adev, false);
5272
5273	return amdgpu_dpm_baco_enter(adev);
5274}
5275
5276int amdgpu_device_baco_exit(struct drm_device *dev)
5277{
5278	struct amdgpu_device *adev = drm_to_adev(dev);
5279	struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
5280	int ret = 0;
5281
5282	if (!amdgpu_device_supports_baco(adev_to_drm(adev)))
5283		return -ENOTSUPP;
5284
5285	ret = amdgpu_dpm_baco_exit(adev);
5286	if (ret)
5287		return ret;
5288
5289	if (ras && adev->ras_enabled &&
5290	    adev->nbio.funcs->enable_doorbell_interrupt)
5291		adev->nbio.funcs->enable_doorbell_interrupt(adev, true);
5292
5293	return 0;
5294}
5295
5296static void amdgpu_cancel_all_tdr(struct amdgpu_device *adev)
5297{
5298	int i;
5299
5300	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
5301		struct amdgpu_ring *ring = adev->rings[i];
5302
5303		if (!ring || !ring->sched.thread)
5304			continue;
5305
5306		cancel_delayed_work_sync(&ring->sched.work_tdr);
5307	}
5308}
5309
5310/**
5311 * amdgpu_pci_error_detected - Called when a PCI error is detected.
5312 * @pdev: PCI device struct
5313 * @state: PCI channel state
5314 *
5315 * Description: Called when a PCI error is detected.
5316 *
5317 * Return: PCI_ERS_RESULT_NEED_RESET or PCI_ERS_RESULT_DISCONNECT.
5318 */
5319pci_ers_result_t amdgpu_pci_error_detected(struct pci_dev *pdev, pci_channel_state_t state)
5320{
5321	struct drm_device *dev = pci_get_drvdata(pdev);
5322	struct amdgpu_device *adev = drm_to_adev(dev);
5323	int i;
5324
5325	DRM_INFO("PCI error: detected callback, state(%d)!!\n", state);
5326
5327	if (adev->gmc.xgmi.num_physical_nodes > 1) {
5328		DRM_WARN("No support for XGMI hive yet...");
5329		return PCI_ERS_RESULT_DISCONNECT;
5330	}
5331
5332	adev->pci_channel_state = state;
5333
5334	switch (state) {
5335	case pci_channel_io_normal:
5336		return PCI_ERS_RESULT_CAN_RECOVER;
5337	/* Fatal error, prepare for slot reset */
5338	case pci_channel_io_frozen:
5339		/*
5340		 * Cancel and wait for all TDRs in progress if failing to
5341		 * set  adev->in_gpu_reset in amdgpu_device_lock_adev
5342		 *
5343		 * Locking adev->reset_sem will prevent any external access
5344		 * to GPU during PCI error recovery
5345		 */
5346		while (!amdgpu_device_lock_adev(adev, NULL))
5347			amdgpu_cancel_all_tdr(adev);
5348
5349		/*
5350		 * Block any work scheduling as we do for regular GPU reset
5351		 * for the duration of the recovery
5352		 */
5353		for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
5354			struct amdgpu_ring *ring = adev->rings[i];
5355
5356			if (!ring || !ring->sched.thread)
5357				continue;
5358
5359			drm_sched_stop(&ring->sched, NULL);
5360		}
5361		atomic_inc(&adev->gpu_reset_counter);
5362		return PCI_ERS_RESULT_NEED_RESET;
5363	case pci_channel_io_perm_failure:
5364		/* Permanent error, prepare for device removal */
5365		return PCI_ERS_RESULT_DISCONNECT;
5366	}
5367
5368	return PCI_ERS_RESULT_NEED_RESET;
5369}
5370
5371/**
5372 * amdgpu_pci_mmio_enabled - Enable MMIO and dump debug registers
5373 * @pdev: pointer to PCI device
5374 */
5375pci_ers_result_t amdgpu_pci_mmio_enabled(struct pci_dev *pdev)
5376{
5377
5378	DRM_INFO("PCI error: mmio enabled callback!!\n");
5379
5380	/* TODO - dump whatever for debugging purposes */
5381
5382	/* This called only if amdgpu_pci_error_detected returns
5383	 * PCI_ERS_RESULT_CAN_RECOVER. Read/write to the device still
5384	 * works, no need to reset slot.
5385	 */
5386
5387	return PCI_ERS_RESULT_RECOVERED;
5388}
5389
5390/**
5391 * amdgpu_pci_slot_reset - Called when PCI slot has been reset.
5392 * @pdev: PCI device struct
5393 *
5394 * Description: This routine is called by the pci error recovery
5395 * code after the PCI slot has been reset, just before we
5396 * should resume normal operations.
5397 */
5398pci_ers_result_t amdgpu_pci_slot_reset(struct pci_dev *pdev)
5399{
5400	struct drm_device *dev = pci_get_drvdata(pdev);
5401	struct amdgpu_device *adev = drm_to_adev(dev);
5402	int r, i;
5403	struct amdgpu_reset_context reset_context;
5404	u32 memsize;
5405	struct list_head device_list;
5406
5407	DRM_INFO("PCI error: slot reset callback!!\n");
5408
5409	memset(&reset_context, 0, sizeof(reset_context));
5410
5411	INIT_LIST_HEAD(&device_list);
5412	list_add_tail(&adev->reset_list, &device_list);
5413
5414	/* wait for asic to come out of reset */
5415	msleep(500);
5416
5417	/* Restore PCI confspace */
5418	amdgpu_device_load_pci_state(pdev);
5419
5420	/* confirm  ASIC came out of reset */
5421	for (i = 0; i < adev->usec_timeout; i++) {
5422		memsize = amdgpu_asic_get_config_memsize(adev);
5423
5424		if (memsize != 0xffffffff)
5425			break;
5426		udelay(1);
5427	}
5428	if (memsize == 0xffffffff) {
5429		r = -ETIME;
5430		goto out;
5431	}
5432
5433	reset_context.method = AMD_RESET_METHOD_NONE;
5434	reset_context.reset_req_dev = adev;
5435	set_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
5436	set_bit(AMDGPU_SKIP_HW_RESET, &reset_context.flags);
5437
5438	adev->no_hw_access = true;
5439	r = amdgpu_device_pre_asic_reset(adev, &reset_context);
5440	adev->no_hw_access = false;
5441	if (r)
5442		goto out;
5443
5444	r = amdgpu_do_asic_reset(&device_list, &reset_context);
5445
5446out:
5447	if (!r) {
5448		if (amdgpu_device_cache_pci_state(adev->pdev))
5449			pci_restore_state(adev->pdev);
5450
5451		DRM_INFO("PCIe error recovery succeeded\n");
5452	} else {
5453		DRM_ERROR("PCIe error recovery failed, err:%d", r);
5454		amdgpu_device_unlock_adev(adev);
5455	}
5456
5457	return r ? PCI_ERS_RESULT_DISCONNECT : PCI_ERS_RESULT_RECOVERED;
5458}
5459
5460/**
5461 * amdgpu_pci_resume() - resume normal ops after PCI reset
5462 * @pdev: pointer to PCI device
5463 *
5464 * Called when the error recovery driver tells us that its
5465 * OK to resume normal operation.
5466 */
5467void amdgpu_pci_resume(struct pci_dev *pdev)
5468{
5469	struct drm_device *dev = pci_get_drvdata(pdev);
5470	struct amdgpu_device *adev = drm_to_adev(dev);
5471	int i;
5472
5473
5474	DRM_INFO("PCI error: resume callback!!\n");
5475
5476	/* Only continue execution for the case of pci_channel_io_frozen */
5477	if (adev->pci_channel_state != pci_channel_io_frozen)
5478		return;
5479
5480	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
5481		struct amdgpu_ring *ring = adev->rings[i];
5482
5483		if (!ring || !ring->sched.thread)
5484			continue;
5485
5486
5487		drm_sched_resubmit_jobs(&ring->sched);
5488		drm_sched_start(&ring->sched, true);
5489	}
5490
5491	amdgpu_device_unlock_adev(adev);
5492}
5493
5494bool amdgpu_device_cache_pci_state(struct pci_dev *pdev)
5495{
5496	struct drm_device *dev = pci_get_drvdata(pdev);
5497	struct amdgpu_device *adev = drm_to_adev(dev);
5498	int r;
5499
5500	r = pci_save_state(pdev);
5501	if (!r) {
5502		kfree(adev->pci_state);
5503
5504		adev->pci_state = pci_store_saved_state(pdev);
5505
5506		if (!adev->pci_state) {
5507			DRM_ERROR("Failed to store PCI saved state");
5508			return false;
5509		}
5510	} else {
5511		DRM_WARN("Failed to save PCI state, err:%d\n", r);
5512		return false;
5513	}
5514
5515	return true;
5516}
5517
5518bool amdgpu_device_load_pci_state(struct pci_dev *pdev)
5519{
5520	struct drm_device *dev = pci_get_drvdata(pdev);
5521	struct amdgpu_device *adev = drm_to_adev(dev);
5522	int r;
5523
5524	if (!adev->pci_state)
5525		return false;
5526
5527	r = pci_load_saved_state(pdev, adev->pci_state);
5528
5529	if (!r) {
5530		pci_restore_state(pdev);
5531	} else {
5532		DRM_WARN("Failed to load PCI state, err:%d\n", r);
5533		return false;
5534	}
5535
5536	return true;
5537}
5538
5539void amdgpu_device_flush_hdp(struct amdgpu_device *adev,
5540		struct amdgpu_ring *ring)
5541{
5542#ifdef CONFIG_X86_64
5543	if (adev->flags & AMD_IS_APU)
5544		return;
5545#endif
5546	if (adev->gmc.xgmi.connected_to_cpu)
5547		return;
5548
5549	if (ring && ring->funcs->emit_hdp_flush)
5550		amdgpu_ring_emit_hdp_flush(ring);
5551	else
5552		amdgpu_asic_flush_hdp(adev, ring);
5553}
5554
5555void amdgpu_device_invalidate_hdp(struct amdgpu_device *adev,
5556		struct amdgpu_ring *ring)
5557{
5558#ifdef CONFIG_X86_64
5559	if (adev->flags & AMD_IS_APU)
5560		return;
5561#endif
5562	if (adev->gmc.xgmi.connected_to_cpu)
5563		return;
5564
5565	amdgpu_asic_invalidate_hdp(adev, ring);
5566}