Linux Audio

Check our new training course

Loading...
v6.9.4
   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#include <linux/iommu.h>
  34#include <linux/pci.h>
  35#include <linux/pci-p2pdma.h>
  36#include <linux/apple-gmux.h>
  37
  38#include <drm/drm_aperture.h>
  39#include <drm/drm_atomic_helper.h>
  40#include <drm/drm_crtc_helper.h>
  41#include <drm/drm_fb_helper.h>
  42#include <drm/drm_probe_helper.h>
  43#include <drm/amdgpu_drm.h>
  44#include <linux/device.h>
  45#include <linux/vgaarb.h>
  46#include <linux/vga_switcheroo.h>
  47#include <linux/efi.h>
  48#include "amdgpu.h"
  49#include "amdgpu_trace.h"
  50#include "amdgpu_i2c.h"
  51#include "atom.h"
  52#include "amdgpu_atombios.h"
  53#include "amdgpu_atomfirmware.h"
  54#include "amd_pcie.h"
  55#ifdef CONFIG_DRM_AMDGPU_SI
  56#include "si.h"
  57#endif
  58#ifdef CONFIG_DRM_AMDGPU_CIK
  59#include "cik.h"
  60#endif
  61#include "vi.h"
  62#include "soc15.h"
  63#include "nv.h"
  64#include "bif/bif_4_1_d.h"
 
  65#include <linux/firmware.h>
  66#include "amdgpu_vf_error.h"
  67
  68#include "amdgpu_amdkfd.h"
  69#include "amdgpu_pm.h"
  70
  71#include "amdgpu_xgmi.h"
  72#include "amdgpu_ras.h"
  73#include "amdgpu_pmu.h"
  74#include "amdgpu_fru_eeprom.h"
  75#include "amdgpu_reset.h"
  76#include "amdgpu_virt.h"
  77
  78#include <linux/suspend.h>
  79#include <drm/task_barrier.h>
  80#include <linux/pm_runtime.h>
  81
  82#include <drm/drm_drv.h>
  83
  84#if IS_ENABLED(CONFIG_X86)
  85#include <asm/intel-family.h>
  86#endif
  87
  88MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin");
  89MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin");
  90MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin");
  91MODULE_FIRMWARE("amdgpu/picasso_gpu_info.bin");
  92MODULE_FIRMWARE("amdgpu/raven2_gpu_info.bin");
  93MODULE_FIRMWARE("amdgpu/arcturus_gpu_info.bin");
 
 
 
  94MODULE_FIRMWARE("amdgpu/navi12_gpu_info.bin");
 
 
  95
  96#define AMDGPU_RESUME_MS		2000
  97#define AMDGPU_MAX_RETRY_LIMIT		2
  98#define AMDGPU_RETRY_SRIOV_RESET(r) ((r) == -EBUSY || (r) == -ETIMEDOUT || (r) == -EINVAL)
  99#define AMDGPU_PCIE_INDEX_FALLBACK (0x38 >> 2)
 100#define AMDGPU_PCIE_INDEX_HI_FALLBACK (0x44 >> 2)
 101#define AMDGPU_PCIE_DATA_FALLBACK (0x3C >> 2)
 102
 103static const struct drm_driver amdgpu_kms_driver;
 104
 105const char *amdgpu_asic_name[] = {
 106	"TAHITI",
 107	"PITCAIRN",
 108	"VERDE",
 109	"OLAND",
 110	"HAINAN",
 111	"BONAIRE",
 112	"KAVERI",
 113	"KABINI",
 114	"HAWAII",
 115	"MULLINS",
 116	"TOPAZ",
 117	"TONGA",
 118	"FIJI",
 119	"CARRIZO",
 120	"STONEY",
 121	"POLARIS10",
 122	"POLARIS11",
 123	"POLARIS12",
 124	"VEGAM",
 125	"VEGA10",
 126	"VEGA12",
 127	"VEGA20",
 128	"RAVEN",
 129	"ARCTURUS",
 130	"RENOIR",
 131	"ALDEBARAN",
 132	"NAVI10",
 133	"CYAN_SKILLFISH",
 134	"NAVI14",
 135	"NAVI12",
 136	"SIENNA_CICHLID",
 137	"NAVY_FLOUNDER",
 138	"VANGOGH",
 139	"DIMGREY_CAVEFISH",
 140	"BEIGE_GOBY",
 141	"YELLOW_CARP",
 142	"IP DISCOVERY",
 143	"LAST",
 144};
 145
 146/**
 147 * DOC: pcie_replay_count
 148 *
 149 * The amdgpu driver provides a sysfs API for reporting the total number
 150 * of PCIe replays (NAKs)
 151 * The file pcie_replay_count is used for this and returns the total
 152 * number of replays as a sum of the NAKs generated and NAKs received
 153 */
 154
 155static ssize_t amdgpu_device_get_pcie_replay_count(struct device *dev,
 156		struct device_attribute *attr, char *buf)
 157{
 158	struct drm_device *ddev = dev_get_drvdata(dev);
 159	struct amdgpu_device *adev = drm_to_adev(ddev);
 160	uint64_t cnt = amdgpu_asic_get_pcie_replay_count(adev);
 161
 162	return sysfs_emit(buf, "%llu\n", cnt);
 163}
 164
 165static DEVICE_ATTR(pcie_replay_count, 0444,
 166		amdgpu_device_get_pcie_replay_count, NULL);
 167
 168static ssize_t amdgpu_sysfs_reg_state_get(struct file *f, struct kobject *kobj,
 169					  struct bin_attribute *attr, char *buf,
 170					  loff_t ppos, size_t count)
 171{
 172	struct device *dev = kobj_to_dev(kobj);
 173	struct drm_device *ddev = dev_get_drvdata(dev);
 174	struct amdgpu_device *adev = drm_to_adev(ddev);
 175	ssize_t bytes_read;
 176
 177	switch (ppos) {
 178	case AMDGPU_SYS_REG_STATE_XGMI:
 179		bytes_read = amdgpu_asic_get_reg_state(
 180			adev, AMDGPU_REG_STATE_TYPE_XGMI, buf, count);
 181		break;
 182	case AMDGPU_SYS_REG_STATE_WAFL:
 183		bytes_read = amdgpu_asic_get_reg_state(
 184			adev, AMDGPU_REG_STATE_TYPE_WAFL, buf, count);
 185		break;
 186	case AMDGPU_SYS_REG_STATE_PCIE:
 187		bytes_read = amdgpu_asic_get_reg_state(
 188			adev, AMDGPU_REG_STATE_TYPE_PCIE, buf, count);
 189		break;
 190	case AMDGPU_SYS_REG_STATE_USR:
 191		bytes_read = amdgpu_asic_get_reg_state(
 192			adev, AMDGPU_REG_STATE_TYPE_USR, buf, count);
 193		break;
 194	case AMDGPU_SYS_REG_STATE_USR_1:
 195		bytes_read = amdgpu_asic_get_reg_state(
 196			adev, AMDGPU_REG_STATE_TYPE_USR_1, buf, count);
 197		break;
 198	default:
 199		return -EINVAL;
 200	}
 201
 202	return bytes_read;
 203}
 204
 205BIN_ATTR(reg_state, 0444, amdgpu_sysfs_reg_state_get, NULL,
 206	 AMDGPU_SYS_REG_STATE_END);
 
 
 
 
 
 
 
 207
 208int amdgpu_reg_state_sysfs_init(struct amdgpu_device *adev)
 
 209{
 210	int ret;
 211
 212	if (!amdgpu_asic_get_reg_state_supported(adev))
 213		return 0;
 214
 215	ret = sysfs_create_bin_file(&adev->dev->kobj, &bin_attr_reg_state);
 216
 217	return ret;
 218}
 219
 220void amdgpu_reg_state_sysfs_fini(struct amdgpu_device *adev)
 221{
 222	if (!amdgpu_asic_get_reg_state_supported(adev))
 223		return;
 224	sysfs_remove_bin_file(&adev->dev->kobj, &bin_attr_reg_state);
 225}
 226
 227/**
 228 * DOC: board_info
 229 *
 230 * The amdgpu driver provides a sysfs API for giving board related information.
 231 * It provides the form factor information in the format
 232 *
 233 *   type : form factor
 234 *
 235 * Possible form factor values
 236 *
 237 * - "cem"		- PCIE CEM card
 238 * - "oam"		- Open Compute Accelerator Module
 239 * - "unknown"	- Not known
 240 *
 
 
 
 
 
 241 */
 242
 243static ssize_t amdgpu_device_get_board_info(struct device *dev,
 244					    struct device_attribute *attr,
 245					    char *buf)
 246{
 247	struct drm_device *ddev = dev_get_drvdata(dev);
 248	struct amdgpu_device *adev = drm_to_adev(ddev);
 249	enum amdgpu_pkg_type pkg_type = AMDGPU_PKG_TYPE_CEM;
 250	const char *pkg;
 251
 252	if (adev->smuio.funcs && adev->smuio.funcs->get_pkg_type)
 253		pkg_type = adev->smuio.funcs->get_pkg_type(adev);
 254
 255	switch (pkg_type) {
 256	case AMDGPU_PKG_TYPE_CEM:
 257		pkg = "cem";
 258		break;
 259	case AMDGPU_PKG_TYPE_OAM:
 260		pkg = "oam";
 261		break;
 262	default:
 263		pkg = "unknown";
 264		break;
 265	}
 266
 267	return sysfs_emit(buf, "%s : %s\n", "type", pkg);
 268}
 269
 270static DEVICE_ATTR(board_info, 0444, amdgpu_device_get_board_info, NULL);
 
 271
 272static struct attribute *amdgpu_board_attrs[] = {
 273	&dev_attr_board_info.attr,
 274	NULL,
 275};
 
 
 
 
 
 276
 277static umode_t amdgpu_board_attrs_is_visible(struct kobject *kobj,
 278					     struct attribute *attr, int n)
 279{
 280	struct device *dev = kobj_to_dev(kobj);
 281	struct drm_device *ddev = dev_get_drvdata(dev);
 282	struct amdgpu_device *adev = drm_to_adev(ddev);
 283
 284	if (adev->flags & AMD_IS_APU)
 285		return 0;
 286
 287	return attr->mode;
 288}
 289
 290static const struct attribute_group amdgpu_board_attrs_group = {
 291	.attrs = amdgpu_board_attrs,
 292	.is_visible = amdgpu_board_attrs_is_visible
 293};
 294
 295static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev);
 296
 297
 298/**
 299 * amdgpu_device_supports_px - Is the device a dGPU with ATPX power control
 300 *
 301 * @dev: drm_device pointer
 302 *
 303 * Returns true if the device is a dGPU with ATPX power control,
 304 * otherwise return false.
 305 */
 306bool amdgpu_device_supports_px(struct drm_device *dev)
 307{
 308	struct amdgpu_device *adev = drm_to_adev(dev);
 309
 310	if ((adev->flags & AMD_IS_PX) && !amdgpu_is_atpx_hybrid())
 311		return true;
 312	return false;
 313}
 314
 315/**
 316 * amdgpu_device_supports_boco - Is the device a dGPU with ACPI power resources
 317 *
 318 * @dev: drm_device pointer
 319 *
 320 * Returns true if the device is a dGPU with ACPI power control,
 321 * otherwise return false.
 322 */
 323bool amdgpu_device_supports_boco(struct drm_device *dev)
 324{
 325	struct amdgpu_device *adev = drm_to_adev(dev);
 326
 327	if (adev->has_pr3 ||
 328	    ((adev->flags & AMD_IS_PX) && amdgpu_is_atpx_hybrid()))
 329		return true;
 330	return false;
 331}
 332
 333/**
 334 * amdgpu_device_supports_baco - Does the device support BACO
 335 *
 336 * @dev: drm_device pointer
 337 *
 338 * Returns true if the device supporte BACO,
 339 * otherwise return false.
 340 */
 341bool amdgpu_device_supports_baco(struct drm_device *dev)
 342{
 343	struct amdgpu_device *adev = drm_to_adev(dev);
 344
 345	return amdgpu_asic_supports_baco(adev);
 346}
 347
 348/**
 349 * amdgpu_device_supports_smart_shift - Is the device dGPU with
 350 * smart shift support
 351 *
 352 * @dev: drm_device pointer
 353 *
 354 * Returns true if the device is a dGPU with Smart Shift support,
 355 * otherwise returns false.
 356 */
 357bool amdgpu_device_supports_smart_shift(struct drm_device *dev)
 358{
 359	return (amdgpu_device_supports_boco(dev) &&
 360		amdgpu_acpi_is_power_shift_control_supported());
 361}
 362
 363/*
 364 * VRAM access helper functions
 365 */
 366
 367/**
 368 * amdgpu_device_mm_access - access vram by MM_INDEX/MM_DATA
 369 *
 370 * @adev: amdgpu_device pointer
 371 * @pos: offset of the buffer in vram
 372 * @buf: virtual address of the buffer in system memory
 373 * @size: read/write size, sizeof(@buf) must > @size
 374 * @write: true - write to vram, otherwise - read from vram
 375 */
 376void amdgpu_device_mm_access(struct amdgpu_device *adev, loff_t pos,
 377			     void *buf, size_t size, bool write)
 378{
 379	unsigned long flags;
 380	uint32_t hi = ~0, tmp = 0;
 381	uint32_t *data = buf;
 382	uint64_t last;
 383	int idx;
 384
 385	if (!drm_dev_enter(adev_to_drm(adev), &idx))
 386		return;
 387
 388	BUG_ON(!IS_ALIGNED(pos, 4) || !IS_ALIGNED(size, 4));
 389
 390	spin_lock_irqsave(&adev->mmio_idx_lock, flags);
 391	for (last = pos + size; pos < last; pos += 4) {
 392		tmp = pos >> 31;
 393
 394		WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)pos) | 0x80000000);
 395		if (tmp != hi) {
 396			WREG32_NO_KIQ(mmMM_INDEX_HI, tmp);
 397			hi = tmp;
 398		}
 399		if (write)
 400			WREG32_NO_KIQ(mmMM_DATA, *data++);
 401		else
 402			*data++ = RREG32_NO_KIQ(mmMM_DATA);
 403	}
 404
 405	spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
 406	drm_dev_exit(idx);
 407}
 408
 409/**
 410 * amdgpu_device_aper_access - access vram by vram aperature
 411 *
 412 * @adev: amdgpu_device pointer
 413 * @pos: offset of the buffer in vram
 414 * @buf: virtual address of the buffer in system memory
 415 * @size: read/write size, sizeof(@buf) must > @size
 416 * @write: true - write to vram, otherwise - read from vram
 417 *
 418 * The return value means how many bytes have been transferred.
 419 */
 420size_t amdgpu_device_aper_access(struct amdgpu_device *adev, loff_t pos,
 421				 void *buf, size_t size, bool write)
 422{
 423#ifdef CONFIG_64BIT
 424	void __iomem *addr;
 425	size_t count = 0;
 426	uint64_t last;
 427
 428	if (!adev->mman.aper_base_kaddr)
 429		return 0;
 430
 431	last = min(pos + size, adev->gmc.visible_vram_size);
 432	if (last > pos) {
 433		addr = adev->mman.aper_base_kaddr + pos;
 434		count = last - pos;
 435
 436		if (write) {
 437			memcpy_toio(addr, buf, count);
 438			/* Make sure HDP write cache flush happens without any reordering
 439			 * after the system memory contents are sent over PCIe device
 440			 */
 441			mb();
 442			amdgpu_device_flush_hdp(adev, NULL);
 443		} else {
 444			amdgpu_device_invalidate_hdp(adev, NULL);
 445			/* Make sure HDP read cache is invalidated before issuing a read
 446			 * to the PCIe device
 447			 */
 448			mb();
 449			memcpy_fromio(buf, addr, count);
 450		}
 451
 452	}
 
 453
 454	return count;
 455#else
 456	return 0;
 
 457#endif
 458}
 459
 460/**
 461 * amdgpu_device_vram_access - read/write a buffer in vram
 462 *
 463 * @adev: amdgpu_device pointer
 464 * @pos: offset of the buffer in vram
 465 * @buf: virtual address of the buffer in system memory
 466 * @size: read/write size, sizeof(@buf) must > @size
 467 * @write: true - write to vram, otherwise - read from vram
 468 */
 469void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos,
 470			       void *buf, size_t size, bool write)
 471{
 472	size_t count;
 473
 474	/* try to using vram apreature to access vram first */
 475	count = amdgpu_device_aper_access(adev, pos, buf, size, write);
 476	size -= count;
 477	if (size) {
 478		/* using MM to access rest vram */
 479		pos += count;
 480		buf += count;
 481		amdgpu_device_mm_access(adev, pos, buf, size, write);
 
 482	}
 
 
 
 
 
 
 483}
 484
 485/*
 486 * register access helper functions.
 487 */
 488
 489/* Check if hw access should be skipped because of hotplug or device error */
 490bool amdgpu_device_skip_hw_access(struct amdgpu_device *adev)
 491{
 492	if (adev->no_hw_access)
 493		return true;
 494
 495#ifdef CONFIG_LOCKDEP
 496	/*
 497	 * This is a bit complicated to understand, so worth a comment. What we assert
 498	 * here is that the GPU reset is not running on another thread in parallel.
 499	 *
 500	 * For this we trylock the read side of the reset semaphore, if that succeeds
 501	 * we know that the reset is not running in paralell.
 502	 *
 503	 * If the trylock fails we assert that we are either already holding the read
 504	 * side of the lock or are the reset thread itself and hold the write side of
 505	 * the lock.
 506	 */
 507	if (in_task()) {
 508		if (down_read_trylock(&adev->reset_domain->sem))
 509			up_read(&adev->reset_domain->sem);
 510		else
 511			lockdep_assert_held(&adev->reset_domain->sem);
 512	}
 513#endif
 514	return false;
 515}
 516
 517/**
 518 * amdgpu_device_rreg - read a memory mapped IO or indirect register
 519 *
 520 * @adev: amdgpu_device pointer
 521 * @reg: dword aligned register offset
 522 * @acc_flags: access flags which require special behavior
 523 *
 524 * Returns the 32 bit value from the offset specified.
 525 */
 526uint32_t amdgpu_device_rreg(struct amdgpu_device *adev,
 527			    uint32_t reg, uint32_t acc_flags)
 528{
 529	uint32_t ret;
 530
 531	if (amdgpu_device_skip_hw_access(adev))
 532		return 0;
 533
 534	if ((reg * 4) < adev->rmmio_size) {
 535		if (!(acc_flags & AMDGPU_REGS_NO_KIQ) &&
 536		    amdgpu_sriov_runtime(adev) &&
 537		    down_read_trylock(&adev->reset_domain->sem)) {
 538			ret = amdgpu_kiq_rreg(adev, reg, 0);
 539			up_read(&adev->reset_domain->sem);
 540		} else {
 541			ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
 542		}
 543	} else {
 544		ret = adev->pcie_rreg(adev, reg * 4);
 545	}
 546
 547	trace_amdgpu_device_rreg(adev->pdev->device, reg, ret);
 548
 549	return ret;
 550}
 551
 552/*
 553 * MMIO register read with bytes helper functions
 554 * @offset:bytes offset from MMIO start
 555 */
 
 556
 557/**
 558 * amdgpu_mm_rreg8 - read a memory mapped IO register
 559 *
 560 * @adev: amdgpu_device pointer
 561 * @offset: byte aligned register offset
 562 *
 563 * Returns the 8 bit value from the offset specified.
 564 */
 565uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset)
 566{
 567	if (amdgpu_device_skip_hw_access(adev))
 568		return 0;
 569
 570	if (offset < adev->rmmio_size)
 571		return (readb(adev->rmmio + offset));
 572	BUG();
 573}
 574
 575
 576/**
 577 * amdgpu_device_xcc_rreg - read a memory mapped IO or indirect register with specific XCC
 578 *
 579 * @adev: amdgpu_device pointer
 580 * @reg: dword aligned register offset
 581 * @acc_flags: access flags which require special behavior
 582 * @xcc_id: xcc accelerated compute core id
 583 *
 584 * Returns the 32 bit value from the offset specified.
 585 */
 586uint32_t amdgpu_device_xcc_rreg(struct amdgpu_device *adev,
 587				uint32_t reg, uint32_t acc_flags,
 588				uint32_t xcc_id)
 589{
 590	uint32_t ret, rlcg_flag;
 591
 592	if (amdgpu_device_skip_hw_access(adev))
 593		return 0;
 594
 595	if ((reg * 4) < adev->rmmio_size) {
 596		if (amdgpu_sriov_vf(adev) &&
 597		    !amdgpu_sriov_runtime(adev) &&
 598		    adev->gfx.rlc.rlcg_reg_access_supported &&
 599		    amdgpu_virt_get_rlcg_reg_access_flag(adev, acc_flags,
 600							 GC_HWIP, false,
 601							 &rlcg_flag)) {
 602			ret = amdgpu_virt_rlcg_reg_rw(adev, reg, 0, rlcg_flag, xcc_id);
 603		} else if (!(acc_flags & AMDGPU_REGS_NO_KIQ) &&
 604		    amdgpu_sriov_runtime(adev) &&
 605		    down_read_trylock(&adev->reset_domain->sem)) {
 606			ret = amdgpu_kiq_rreg(adev, reg, xcc_id);
 607			up_read(&adev->reset_domain->sem);
 608		} else {
 609			ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
 610		}
 611	} else {
 612		ret = adev->pcie_rreg(adev, reg * 4);
 613	}
 614
 615	return ret;
 616}
 617
 618/*
 619 * MMIO register write with bytes helper functions
 620 * @offset:bytes offset from MMIO start
 621 * @value: the value want to be written to the register
 622 */
 623
 624/**
 625 * amdgpu_mm_wreg8 - read a memory mapped IO register
 626 *
 627 * @adev: amdgpu_device pointer
 628 * @offset: byte aligned register offset
 629 * @value: 8 bit value to write
 630 *
 631 * Writes the value specified to the offset specified.
 632 */
 633void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value)
 634{
 635	if (amdgpu_device_skip_hw_access(adev))
 636		return;
 637
 638	if (offset < adev->rmmio_size)
 639		writeb(value, adev->rmmio + offset);
 640	else
 641		BUG();
 642}
 643
 644/**
 645 * amdgpu_device_wreg - write to a memory mapped IO or indirect register
 646 *
 647 * @adev: amdgpu_device pointer
 648 * @reg: dword aligned register offset
 649 * @v: 32 bit value to write to the register
 650 * @acc_flags: access flags which require special behavior
 651 *
 652 * Writes the value specified to the offset specified.
 653 */
 654void amdgpu_device_wreg(struct amdgpu_device *adev,
 655			uint32_t reg, uint32_t v,
 656			uint32_t acc_flags)
 657{
 658	if (amdgpu_device_skip_hw_access(adev))
 659		return;
 660
 661	if ((reg * 4) < adev->rmmio_size) {
 662		if (!(acc_flags & AMDGPU_REGS_NO_KIQ) &&
 663		    amdgpu_sriov_runtime(adev) &&
 664		    down_read_trylock(&adev->reset_domain->sem)) {
 665			amdgpu_kiq_wreg(adev, reg, v, 0);
 666			up_read(&adev->reset_domain->sem);
 667		} else {
 668			writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
 669		}
 670	} else {
 671		adev->pcie_wreg(adev, reg * 4, v);
 672	}
 673
 674	trace_amdgpu_device_wreg(adev->pdev->device, reg, v);
 675}
 676
 677/**
 678 * amdgpu_mm_wreg_mmio_rlc -  write register either with direct/indirect mmio or with RLC path if in range
 679 *
 680 * @adev: amdgpu_device pointer
 681 * @reg: mmio/rlc register
 682 * @v: value to write
 683 * @xcc_id: xcc accelerated compute core id
 684 *
 685 * this function is invoked only for the debugfs register access
 686 */
 687void amdgpu_mm_wreg_mmio_rlc(struct amdgpu_device *adev,
 688			     uint32_t reg, uint32_t v,
 689			     uint32_t xcc_id)
 690{
 691	if (amdgpu_device_skip_hw_access(adev))
 692		return;
 693
 694	if (amdgpu_sriov_fullaccess(adev) &&
 695	    adev->gfx.rlc.funcs &&
 696	    adev->gfx.rlc.funcs->is_rlcg_access_range) {
 697		if (adev->gfx.rlc.funcs->is_rlcg_access_range(adev, reg))
 698			return amdgpu_sriov_wreg(adev, reg, v, 0, 0, xcc_id);
 699	} else if ((reg * 4) >= adev->rmmio_size) {
 700		adev->pcie_wreg(adev, reg * 4, v);
 701	} else {
 702		writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
 703	}
 704}
 705
 706/**
 707 * amdgpu_device_xcc_wreg - write to a memory mapped IO or indirect register with specific XCC
 708 *
 709 * @adev: amdgpu_device pointer
 710 * @reg: dword aligned register offset
 711 * @v: 32 bit value to write to the register
 712 * @acc_flags: access flags which require special behavior
 713 * @xcc_id: xcc accelerated compute core id
 714 *
 715 * Writes the value specified to the offset specified.
 
 716 */
 717void amdgpu_device_xcc_wreg(struct amdgpu_device *adev,
 718			uint32_t reg, uint32_t v,
 719			uint32_t acc_flags, uint32_t xcc_id)
 720{
 721	uint32_t rlcg_flag;
 722
 723	if (amdgpu_device_skip_hw_access(adev))
 724		return;
 725
 726	if ((reg * 4) < adev->rmmio_size) {
 727		if (amdgpu_sriov_vf(adev) &&
 728		    !amdgpu_sriov_runtime(adev) &&
 729		    adev->gfx.rlc.rlcg_reg_access_supported &&
 730		    amdgpu_virt_get_rlcg_reg_access_flag(adev, acc_flags,
 731							 GC_HWIP, true,
 732							 &rlcg_flag)) {
 733			amdgpu_virt_rlcg_reg_rw(adev, reg, v, rlcg_flag, xcc_id);
 734		} else if (!(acc_flags & AMDGPU_REGS_NO_KIQ) &&
 735		    amdgpu_sriov_runtime(adev) &&
 736		    down_read_trylock(&adev->reset_domain->sem)) {
 737			amdgpu_kiq_wreg(adev, reg, v, xcc_id);
 738			up_read(&adev->reset_domain->sem);
 739		} else {
 740			writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
 741		}
 742	} else {
 743		adev->pcie_wreg(adev, reg * 4, v);
 
 744	}
 745}
 746
 747/**
 748 * amdgpu_device_indirect_rreg - read an indirect register
 749 *
 750 * @adev: amdgpu_device pointer
 751 * @reg_addr: indirect register address to read from
 
 752 *
 753 * Returns the value of indirect register @reg_addr
 
 754 */
 755u32 amdgpu_device_indirect_rreg(struct amdgpu_device *adev,
 756				u32 reg_addr)
 757{
 758	unsigned long flags, pcie_index, pcie_data;
 759	void __iomem *pcie_index_offset;
 760	void __iomem *pcie_data_offset;
 761	u32 r;
 762
 763	pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev);
 764	pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev);
 765
 766	spin_lock_irqsave(&adev->pcie_idx_lock, flags);
 767	pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
 768	pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
 769
 770	writel(reg_addr, pcie_index_offset);
 771	readl(pcie_index_offset);
 772	r = readl(pcie_data_offset);
 773	spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
 774
 775	return r;
 
 
 
 
 776}
 777
 778u32 amdgpu_device_indirect_rreg_ext(struct amdgpu_device *adev,
 779				    u64 reg_addr)
 
 
 
 
 
 
 
 
 780{
 781	unsigned long flags, pcie_index, pcie_index_hi, pcie_data;
 782	u32 r;
 783	void __iomem *pcie_index_offset;
 784	void __iomem *pcie_index_hi_offset;
 785	void __iomem *pcie_data_offset;
 786
 787	if (unlikely(!adev->nbio.funcs)) {
 788		pcie_index = AMDGPU_PCIE_INDEX_FALLBACK;
 789		pcie_data = AMDGPU_PCIE_DATA_FALLBACK;
 790	} else {
 791		pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev);
 792		pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev);
 793	}
 794
 795	if (reg_addr >> 32) {
 796		if (unlikely(!adev->nbio.funcs))
 797			pcie_index_hi = AMDGPU_PCIE_INDEX_HI_FALLBACK;
 798		else
 799			pcie_index_hi = adev->nbio.funcs->get_pcie_index_hi_offset(adev);
 800	} else {
 801		pcie_index_hi = 0;
 
 802	}
 
 803
 804	spin_lock_irqsave(&adev->pcie_idx_lock, flags);
 805	pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
 806	pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
 807	if (pcie_index_hi != 0)
 808		pcie_index_hi_offset = (void __iomem *)adev->rmmio +
 809				pcie_index_hi * 4;
 810
 811	writel(reg_addr, pcie_index_offset);
 812	readl(pcie_index_offset);
 813	if (pcie_index_hi != 0) {
 814		writel((reg_addr >> 32) & 0xff, pcie_index_hi_offset);
 815		readl(pcie_index_hi_offset);
 816	}
 817	r = readl(pcie_data_offset);
 818
 819	/* clear the high bits */
 820	if (pcie_index_hi != 0) {
 821		writel(0, pcie_index_hi_offset);
 822		readl(pcie_index_hi_offset);
 823	}
 824
 825	spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
 826
 827	return r;
 828}
 829
 830/**
 831 * amdgpu_device_indirect_rreg64 - read a 64bits indirect register
 832 *
 833 * @adev: amdgpu_device pointer
 
 
 834 * @reg_addr: indirect register address to read from
 835 *
 836 * Returns the value of indirect register @reg_addr
 837 */
 838u64 amdgpu_device_indirect_rreg64(struct amdgpu_device *adev,
 839				  u32 reg_addr)
 
 840{
 841	unsigned long flags, pcie_index, pcie_data;
 
 842	void __iomem *pcie_index_offset;
 843	void __iomem *pcie_data_offset;
 844	u64 r;
 845
 846	pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev);
 847	pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev);
 848
 849	spin_lock_irqsave(&adev->pcie_idx_lock, flags);
 850	pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
 851	pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
 852
 853	/* read low 32 bits */
 854	writel(reg_addr, pcie_index_offset);
 855	readl(pcie_index_offset);
 856	r = readl(pcie_data_offset);
 857	/* read high 32 bits */
 858	writel(reg_addr + 4, pcie_index_offset);
 859	readl(pcie_index_offset);
 860	r |= ((u64)readl(pcie_data_offset) << 32);
 861	spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
 862
 863	return r;
 864}
 865
 866u64 amdgpu_device_indirect_rreg64_ext(struct amdgpu_device *adev,
 867				  u64 reg_addr)
 
 
 
 
 
 
 
 
 
 
 
 868{
 869	unsigned long flags, pcie_index, pcie_data;
 870	unsigned long pcie_index_hi = 0;
 871	void __iomem *pcie_index_offset;
 872	void __iomem *pcie_index_hi_offset;
 873	void __iomem *pcie_data_offset;
 874	u64 r;
 875
 876	pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev);
 877	pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev);
 878	if ((reg_addr >> 32) && (adev->nbio.funcs->get_pcie_index_hi_offset))
 879		pcie_index_hi = adev->nbio.funcs->get_pcie_index_hi_offset(adev);
 880
 881	spin_lock_irqsave(&adev->pcie_idx_lock, flags);
 882	pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
 883	pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
 884	if (pcie_index_hi != 0)
 885		pcie_index_hi_offset = (void __iomem *)adev->rmmio +
 886			pcie_index_hi * 4;
 887
 888	/* read low 32 bits */
 889	writel(reg_addr, pcie_index_offset);
 890	readl(pcie_index_offset);
 891	if (pcie_index_hi != 0) {
 892		writel((reg_addr >> 32) & 0xff, pcie_index_hi_offset);
 893		readl(pcie_index_hi_offset);
 894	}
 895	r = readl(pcie_data_offset);
 896	/* read high 32 bits */
 897	writel(reg_addr + 4, pcie_index_offset);
 898	readl(pcie_index_offset);
 899	if (pcie_index_hi != 0) {
 900		writel((reg_addr >> 32) & 0xff, pcie_index_hi_offset);
 901		readl(pcie_index_hi_offset);
 902	}
 903	r |= ((u64)readl(pcie_data_offset) << 32);
 904
 905	/* clear the high bits */
 906	if (pcie_index_hi != 0) {
 907		writel(0, pcie_index_hi_offset);
 908		readl(pcie_index_hi_offset);
 909	}
 910
 911	spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
 912
 913	return r;
 914}
 915
 916/**
 917 * amdgpu_device_indirect_wreg - write an indirect register address
 918 *
 919 * @adev: amdgpu_device pointer
 
 
 920 * @reg_addr: indirect register offset
 921 * @reg_data: indirect register data
 922 *
 923 */
 924void amdgpu_device_indirect_wreg(struct amdgpu_device *adev,
 
 925				 u32 reg_addr, u32 reg_data)
 926{
 927	unsigned long flags, pcie_index, pcie_data;
 928	void __iomem *pcie_index_offset;
 929	void __iomem *pcie_data_offset;
 930
 931	pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev);
 932	pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev);
 933
 934	spin_lock_irqsave(&adev->pcie_idx_lock, flags);
 935	pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
 936	pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
 937
 938	writel(reg_addr, pcie_index_offset);
 939	readl(pcie_index_offset);
 940	writel(reg_data, pcie_data_offset);
 941	readl(pcie_data_offset);
 942	spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
 943}
 944
 945void amdgpu_device_indirect_wreg_ext(struct amdgpu_device *adev,
 946				     u64 reg_addr, u32 reg_data)
 947{
 948	unsigned long flags, pcie_index, pcie_index_hi, pcie_data;
 949	void __iomem *pcie_index_offset;
 950	void __iomem *pcie_index_hi_offset;
 951	void __iomem *pcie_data_offset;
 952
 953	pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev);
 954	pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev);
 955	if ((reg_addr >> 32) && (adev->nbio.funcs->get_pcie_index_hi_offset))
 956		pcie_index_hi = adev->nbio.funcs->get_pcie_index_hi_offset(adev);
 957	else
 958		pcie_index_hi = 0;
 959
 960	spin_lock_irqsave(&adev->pcie_idx_lock, flags);
 961	pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
 962	pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
 963	if (pcie_index_hi != 0)
 964		pcie_index_hi_offset = (void __iomem *)adev->rmmio +
 965				pcie_index_hi * 4;
 966
 967	writel(reg_addr, pcie_index_offset);
 968	readl(pcie_index_offset);
 969	if (pcie_index_hi != 0) {
 970		writel((reg_addr >> 32) & 0xff, pcie_index_hi_offset);
 971		readl(pcie_index_hi_offset);
 972	}
 973	writel(reg_data, pcie_data_offset);
 974	readl(pcie_data_offset);
 975
 976	/* clear the high bits */
 977	if (pcie_index_hi != 0) {
 978		writel(0, pcie_index_hi_offset);
 979		readl(pcie_index_hi_offset);
 980	}
 981
 982	spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
 983}
 984
 985/**
 986 * amdgpu_device_indirect_wreg64 - write a 64bits indirect register address
 987 *
 988 * @adev: amdgpu_device pointer
 
 
 989 * @reg_addr: indirect register offset
 990 * @reg_data: indirect register data
 991 *
 992 */
 993void amdgpu_device_indirect_wreg64(struct amdgpu_device *adev,
 
 994				   u32 reg_addr, u64 reg_data)
 995{
 996	unsigned long flags, pcie_index, pcie_data;
 997	void __iomem *pcie_index_offset;
 998	void __iomem *pcie_data_offset;
 999
1000	pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev);
1001	pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev);
1002
1003	spin_lock_irqsave(&adev->pcie_idx_lock, flags);
1004	pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
1005	pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
1006
1007	/* write low 32 bits */
1008	writel(reg_addr, pcie_index_offset);
1009	readl(pcie_index_offset);
1010	writel((u32)(reg_data & 0xffffffffULL), pcie_data_offset);
1011	readl(pcie_data_offset);
1012	/* write high 32 bits */
1013	writel(reg_addr + 4, pcie_index_offset);
1014	readl(pcie_index_offset);
1015	writel((u32)(reg_data >> 32), pcie_data_offset);
1016	readl(pcie_data_offset);
1017	spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
1018}
1019
1020void amdgpu_device_indirect_wreg64_ext(struct amdgpu_device *adev,
1021				   u64 reg_addr, u64 reg_data)
1022{
1023	unsigned long flags, pcie_index, pcie_data;
1024	unsigned long pcie_index_hi = 0;
1025	void __iomem *pcie_index_offset;
1026	void __iomem *pcie_index_hi_offset;
1027	void __iomem *pcie_data_offset;
1028
1029	pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev);
1030	pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev);
1031	if ((reg_addr >> 32) && (adev->nbio.funcs->get_pcie_index_hi_offset))
1032		pcie_index_hi = adev->nbio.funcs->get_pcie_index_hi_offset(adev);
1033
1034	spin_lock_irqsave(&adev->pcie_idx_lock, flags);
1035	pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
1036	pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
1037	if (pcie_index_hi != 0)
1038		pcie_index_hi_offset = (void __iomem *)adev->rmmio +
1039				pcie_index_hi * 4;
1040
1041	/* write low 32 bits */
1042	writel(reg_addr, pcie_index_offset);
1043	readl(pcie_index_offset);
1044	if (pcie_index_hi != 0) {
1045		writel((reg_addr >> 32) & 0xff, pcie_index_hi_offset);
1046		readl(pcie_index_hi_offset);
1047	}
1048	writel((u32)(reg_data & 0xffffffffULL), pcie_data_offset);
1049	readl(pcie_data_offset);
1050	/* write high 32 bits */
1051	writel(reg_addr + 4, pcie_index_offset);
1052	readl(pcie_index_offset);
1053	if (pcie_index_hi != 0) {
1054		writel((reg_addr >> 32) & 0xff, pcie_index_hi_offset);
1055		readl(pcie_index_hi_offset);
1056	}
1057	writel((u32)(reg_data >> 32), pcie_data_offset);
1058	readl(pcie_data_offset);
1059
1060	/* clear the high bits */
1061	if (pcie_index_hi != 0) {
1062		writel(0, pcie_index_hi_offset);
1063		readl(pcie_index_hi_offset);
1064	}
1065
1066	spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
1067}
1068
1069/**
1070 * amdgpu_device_get_rev_id - query device rev_id
1071 *
1072 * @adev: amdgpu_device pointer
1073 *
1074 * Return device rev_id
1075 */
1076u32 amdgpu_device_get_rev_id(struct amdgpu_device *adev)
1077{
1078	return adev->nbio.funcs->get_rev_id(adev);
1079}
1080
1081/**
1082 * amdgpu_invalid_rreg - dummy reg read function
1083 *
1084 * @adev: amdgpu_device pointer
1085 * @reg: offset of register
1086 *
1087 * Dummy register read function.  Used for register blocks
1088 * that certain asics don't have (all asics).
1089 * Returns the value in the register.
1090 */
1091static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
1092{
1093	DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
1094	BUG();
1095	return 0;
1096}
1097
1098static uint32_t amdgpu_invalid_rreg_ext(struct amdgpu_device *adev, uint64_t reg)
1099{
1100	DRM_ERROR("Invalid callback to read register 0x%llX\n", reg);
1101	BUG();
1102	return 0;
1103}
1104
1105/**
1106 * amdgpu_invalid_wreg - dummy reg write function
1107 *
1108 * @adev: amdgpu_device pointer
1109 * @reg: offset of register
1110 * @v: value to write to the register
1111 *
1112 * Dummy register read function.  Used for register blocks
1113 * that certain asics don't have (all asics).
1114 */
1115static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
1116{
1117	DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
1118		  reg, v);
1119	BUG();
1120}
1121
1122static void amdgpu_invalid_wreg_ext(struct amdgpu_device *adev, uint64_t reg, uint32_t v)
1123{
1124	DRM_ERROR("Invalid callback to write register 0x%llX with 0x%08X\n",
1125		  reg, v);
1126	BUG();
1127}
1128
1129/**
1130 * amdgpu_invalid_rreg64 - dummy 64 bit reg read function
1131 *
1132 * @adev: amdgpu_device pointer
1133 * @reg: offset of register
1134 *
1135 * Dummy register read function.  Used for register blocks
1136 * that certain asics don't have (all asics).
1137 * Returns the value in the register.
1138 */
1139static uint64_t amdgpu_invalid_rreg64(struct amdgpu_device *adev, uint32_t reg)
1140{
1141	DRM_ERROR("Invalid callback to read 64 bit register 0x%04X\n", reg);
1142	BUG();
1143	return 0;
1144}
1145
1146static uint64_t amdgpu_invalid_rreg64_ext(struct amdgpu_device *adev, uint64_t reg)
1147{
1148	DRM_ERROR("Invalid callback to read register 0x%llX\n", reg);
1149	BUG();
1150	return 0;
1151}
1152
1153/**
1154 * amdgpu_invalid_wreg64 - dummy reg write function
1155 *
1156 * @adev: amdgpu_device pointer
1157 * @reg: offset of register
1158 * @v: value to write to the register
1159 *
1160 * Dummy register read function.  Used for register blocks
1161 * that certain asics don't have (all asics).
1162 */
1163static void amdgpu_invalid_wreg64(struct amdgpu_device *adev, uint32_t reg, uint64_t v)
1164{
1165	DRM_ERROR("Invalid callback to write 64 bit register 0x%04X with 0x%08llX\n",
1166		  reg, v);
1167	BUG();
1168}
1169
1170static void amdgpu_invalid_wreg64_ext(struct amdgpu_device *adev, uint64_t reg, uint64_t v)
1171{
1172	DRM_ERROR("Invalid callback to write 64 bit register 0x%llX with 0x%08llX\n",
1173		  reg, v);
1174	BUG();
1175}
1176
1177/**
1178 * amdgpu_block_invalid_rreg - dummy reg read function
1179 *
1180 * @adev: amdgpu_device pointer
1181 * @block: offset of instance
1182 * @reg: offset of register
1183 *
1184 * Dummy register read function.  Used for register blocks
1185 * that certain asics don't have (all asics).
1186 * Returns the value in the register.
1187 */
1188static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
1189					  uint32_t block, uint32_t reg)
1190{
1191	DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
1192		  reg, block);
1193	BUG();
1194	return 0;
1195}
1196
1197/**
1198 * amdgpu_block_invalid_wreg - dummy reg write function
1199 *
1200 * @adev: amdgpu_device pointer
1201 * @block: offset of instance
1202 * @reg: offset of register
1203 * @v: value to write to the register
1204 *
1205 * Dummy register read function.  Used for register blocks
1206 * that certain asics don't have (all asics).
1207 */
1208static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
1209				      uint32_t block,
1210				      uint32_t reg, uint32_t v)
1211{
1212	DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
1213		  reg, block, v);
1214	BUG();
1215}
1216
1217/**
1218 * amdgpu_device_asic_init - Wrapper for atom asic_init
1219 *
1220 * @adev: amdgpu_device pointer
1221 *
1222 * Does any asic specific work and then calls atom asic init.
1223 */
1224static int amdgpu_device_asic_init(struct amdgpu_device *adev)
1225{
1226	int ret;
1227
1228	amdgpu_asic_pre_asic_init(adev);
1229
1230	if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3) ||
1231	    amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(11, 0, 0)) {
1232		amdgpu_psp_wait_for_bootloader(adev);
1233		ret = amdgpu_atomfirmware_asic_init(adev, true);
1234		return ret;
1235	} else {
1236		return amdgpu_atom_asic_init(adev->mode_info.atom_context);
1237	}
1238
1239	return 0;
1240}
1241
1242/**
1243 * amdgpu_device_mem_scratch_init - allocate the VRAM scratch page
1244 *
1245 * @adev: amdgpu_device pointer
1246 *
1247 * Allocates a scratch page of VRAM for use by various things in the
1248 * driver.
1249 */
1250static int amdgpu_device_mem_scratch_init(struct amdgpu_device *adev)
1251{
1252	return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE, PAGE_SIZE,
1253				       AMDGPU_GEM_DOMAIN_VRAM |
1254				       AMDGPU_GEM_DOMAIN_GTT,
1255				       &adev->mem_scratch.robj,
1256				       &adev->mem_scratch.gpu_addr,
1257				       (void **)&adev->mem_scratch.ptr);
1258}
1259
1260/**
1261 * amdgpu_device_mem_scratch_fini - Free the VRAM scratch page
1262 *
1263 * @adev: amdgpu_device pointer
1264 *
1265 * Frees the VRAM scratch page.
1266 */
1267static void amdgpu_device_mem_scratch_fini(struct amdgpu_device *adev)
1268{
1269	amdgpu_bo_free_kernel(&adev->mem_scratch.robj, NULL, NULL);
1270}
1271
1272/**
1273 * amdgpu_device_program_register_sequence - program an array of registers.
1274 *
1275 * @adev: amdgpu_device pointer
1276 * @registers: pointer to the register array
1277 * @array_size: size of the register array
1278 *
1279 * Programs an array or registers with and or masks.
1280 * This is a helper for setting golden registers.
1281 */
1282void amdgpu_device_program_register_sequence(struct amdgpu_device *adev,
1283					     const u32 *registers,
1284					     const u32 array_size)
1285{
1286	u32 tmp, reg, and_mask, or_mask;
1287	int i;
1288
1289	if (array_size % 3)
1290		return;
1291
1292	for (i = 0; i < array_size; i += 3) {
1293		reg = registers[i + 0];
1294		and_mask = registers[i + 1];
1295		or_mask = registers[i + 2];
1296
1297		if (and_mask == 0xffffffff) {
1298			tmp = or_mask;
1299		} else {
1300			tmp = RREG32(reg);
1301			tmp &= ~and_mask;
1302			if (adev->family >= AMDGPU_FAMILY_AI)
1303				tmp |= (or_mask & and_mask);
1304			else
1305				tmp |= or_mask;
1306		}
1307		WREG32(reg, tmp);
1308	}
1309}
1310
1311/**
1312 * amdgpu_device_pci_config_reset - reset the GPU
1313 *
1314 * @adev: amdgpu_device pointer
1315 *
1316 * Resets the GPU using the pci config reset sequence.
1317 * Only applicable to asics prior to vega10.
1318 */
1319void amdgpu_device_pci_config_reset(struct amdgpu_device *adev)
1320{
1321	pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
1322}
1323
1324/**
1325 * amdgpu_device_pci_reset - reset the GPU using generic PCI means
1326 *
1327 * @adev: amdgpu_device pointer
1328 *
1329 * Resets the GPU using generic pci reset interfaces (FLR, SBR, etc.).
1330 */
1331int amdgpu_device_pci_reset(struct amdgpu_device *adev)
1332{
1333	return pci_reset_function(adev->pdev);
1334}
1335
1336/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1337 * amdgpu_device_wb_*()
1338 * Writeback is the method by which the GPU updates special pages in memory
1339 * with the status of certain GPU events (fences, ring pointers,etc.).
1340 */
1341
1342/**
1343 * amdgpu_device_wb_fini - Disable Writeback and free memory
1344 *
1345 * @adev: amdgpu_device pointer
1346 *
1347 * Disables Writeback and frees the Writeback memory (all asics).
1348 * Used at driver shutdown.
1349 */
1350static void amdgpu_device_wb_fini(struct amdgpu_device *adev)
1351{
1352	if (adev->wb.wb_obj) {
1353		amdgpu_bo_free_kernel(&adev->wb.wb_obj,
1354				      &adev->wb.gpu_addr,
1355				      (void **)&adev->wb.wb);
1356		adev->wb.wb_obj = NULL;
1357	}
1358}
1359
1360/**
1361 * amdgpu_device_wb_init - Init Writeback driver info and allocate memory
1362 *
1363 * @adev: amdgpu_device pointer
1364 *
1365 * Initializes writeback and allocates writeback memory (all asics).
1366 * Used at driver startup.
1367 * Returns 0 on success or an -error on failure.
1368 */
1369static int amdgpu_device_wb_init(struct amdgpu_device *adev)
1370{
1371	int r;
1372
1373	if (adev->wb.wb_obj == NULL) {
1374		/* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */
1375		r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8,
1376					    PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
1377					    &adev->wb.wb_obj, &adev->wb.gpu_addr,
1378					    (void **)&adev->wb.wb);
1379		if (r) {
1380			dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
1381			return r;
1382		}
1383
1384		adev->wb.num_wb = AMDGPU_MAX_WB;
1385		memset(&adev->wb.used, 0, sizeof(adev->wb.used));
1386
1387		/* clear wb memory */
1388		memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8);
1389	}
1390
1391	return 0;
1392}
1393
1394/**
1395 * amdgpu_device_wb_get - Allocate a wb entry
1396 *
1397 * @adev: amdgpu_device pointer
1398 * @wb: wb index
1399 *
1400 * Allocate a wb slot for use by the driver (all asics).
1401 * Returns 0 on success or -EINVAL on failure.
1402 */
1403int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb)
1404{
1405	unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
1406
1407	if (offset < adev->wb.num_wb) {
1408		__set_bit(offset, adev->wb.used);
1409		*wb = offset << 3; /* convert to dw offset */
1410		return 0;
1411	} else {
1412		return -EINVAL;
1413	}
1414}
1415
1416/**
1417 * amdgpu_device_wb_free - Free a wb entry
1418 *
1419 * @adev: amdgpu_device pointer
1420 * @wb: wb index
1421 *
1422 * Free a wb slot allocated for use by the driver (all asics)
1423 */
1424void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb)
1425{
1426	wb >>= 3;
1427	if (wb < adev->wb.num_wb)
1428		__clear_bit(wb, adev->wb.used);
1429}
1430
1431/**
1432 * amdgpu_device_resize_fb_bar - try to resize FB BAR
1433 *
1434 * @adev: amdgpu_device pointer
1435 *
1436 * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not
1437 * to fail, but if any of the BARs is not accessible after the size we abort
1438 * driver loading by returning -ENODEV.
1439 */
1440int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev)
1441{
1442	int rbar_size = pci_rebar_bytes_to_size(adev->gmc.real_vram_size);
1443	struct pci_bus *root;
1444	struct resource *res;
1445	unsigned int i;
1446	u16 cmd;
1447	int r;
1448
1449	if (!IS_ENABLED(CONFIG_PHYS_ADDR_T_64BIT))
1450		return 0;
1451
1452	/* Bypass for VF */
1453	if (amdgpu_sriov_vf(adev))
1454		return 0;
1455
1456	/* PCI_EXT_CAP_ID_VNDR extended capability is located at 0x100 */
1457	if (!pci_find_ext_capability(adev->pdev, PCI_EXT_CAP_ID_VNDR))
1458		DRM_WARN("System can't access extended configuration space,please check!!\n");
1459
1460	/* skip if the bios has already enabled large BAR */
1461	if (adev->gmc.real_vram_size &&
1462	    (pci_resource_len(adev->pdev, 0) >= adev->gmc.real_vram_size))
1463		return 0;
1464
1465	/* Check if the root BUS has 64bit memory resources */
1466	root = adev->pdev->bus;
1467	while (root->parent)
1468		root = root->parent;
1469
1470	pci_bus_for_each_resource(root, res, i) {
1471		if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) &&
1472		    res->start > 0x100000000ull)
1473			break;
1474	}
1475
1476	/* Trying to resize is pointless without a root hub window above 4GB */
1477	if (!res)
1478		return 0;
1479
1480	/* Limit the BAR size to what is available */
1481	rbar_size = min(fls(pci_rebar_get_possible_sizes(adev->pdev, 0)) - 1,
1482			rbar_size);
1483
1484	/* Disable memory decoding while we change the BAR addresses and size */
1485	pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd);
1486	pci_write_config_word(adev->pdev, PCI_COMMAND,
1487			      cmd & ~PCI_COMMAND_MEMORY);
1488
1489	/* Free the VRAM and doorbell BAR, we most likely need to move both. */
1490	amdgpu_doorbell_fini(adev);
1491	if (adev->asic_type >= CHIP_BONAIRE)
1492		pci_release_resource(adev->pdev, 2);
1493
1494	pci_release_resource(adev->pdev, 0);
1495
1496	r = pci_resize_resource(adev->pdev, 0, rbar_size);
1497	if (r == -ENOSPC)
1498		DRM_INFO("Not enough PCI address space for a large BAR.");
1499	else if (r && r != -ENOTSUPP)
1500		DRM_ERROR("Problem resizing BAR0 (%d).", r);
1501
1502	pci_assign_unassigned_bus_resources(adev->pdev->bus);
1503
1504	/* When the doorbell or fb BAR isn't available we have no chance of
1505	 * using the device.
1506	 */
1507	r = amdgpu_doorbell_init(adev);
1508	if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET))
1509		return -ENODEV;
1510
1511	pci_write_config_word(adev->pdev, PCI_COMMAND, cmd);
1512
1513	return 0;
1514}
1515
1516static bool amdgpu_device_read_bios(struct amdgpu_device *adev)
1517{
1518	if (hweight32(adev->aid_mask) && (adev->flags & AMD_IS_APU))
1519		return false;
1520
1521	return true;
1522}
1523
1524/*
1525 * GPU helpers function.
1526 */
1527/**
1528 * amdgpu_device_need_post - check if the hw need post or not
1529 *
1530 * @adev: amdgpu_device pointer
1531 *
1532 * Check if the asic has been initialized (all asics) at driver startup
1533 * or post is needed if  hw reset is performed.
1534 * Returns true if need or false if not.
1535 */
1536bool amdgpu_device_need_post(struct amdgpu_device *adev)
1537{
1538	uint32_t reg;
1539
1540	if (amdgpu_sriov_vf(adev))
1541		return false;
1542
1543	if (!amdgpu_device_read_bios(adev))
1544		return false;
1545
1546	if (amdgpu_passthrough(adev)) {
1547		/* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
1548		 * some old smc fw still need driver do vPost otherwise gpu hang, while
1549		 * those smc fw version above 22.15 doesn't have this flaw, so we force
1550		 * vpost executed for smc version below 22.15
1551		 */
1552		if (adev->asic_type == CHIP_FIJI) {
1553			int err;
1554			uint32_t fw_ver;
1555
1556			err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
1557			/* force vPost if error occured */
1558			if (err)
1559				return true;
1560
1561			fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
1562			release_firmware(adev->pm.fw);
1563			if (fw_ver < 0x00160e00)
1564				return true;
1565		}
1566	}
1567
1568	/* Don't post if we need to reset whole hive on init */
1569	if (adev->gmc.xgmi.pending_reset)
1570		return false;
1571
1572	if (adev->has_hw_reset) {
1573		adev->has_hw_reset = false;
1574		return true;
1575	}
1576
1577	/* bios scratch used on CIK+ */
1578	if (adev->asic_type >= CHIP_BONAIRE)
1579		return amdgpu_atombios_scratch_need_asic_init(adev);
1580
1581	/* check MEM_SIZE for older asics */
1582	reg = amdgpu_asic_get_config_memsize(adev);
1583
1584	if ((reg != 0) && (reg != 0xffffffff))
1585		return false;
1586
1587	return true;
1588}
1589
1590/*
1591 * Check whether seamless boot is supported.
1592 *
1593 * So far we only support seamless boot on DCE 3.0 or later.
1594 * If users report that it works on older ASICS as well, we may
1595 * loosen this.
1596 */
1597bool amdgpu_device_seamless_boot_supported(struct amdgpu_device *adev)
1598{
1599	switch (amdgpu_seamless) {
1600	case -1:
1601		break;
1602	case 1:
1603		return true;
1604	case 0:
1605		return false;
1606	default:
1607		DRM_ERROR("Invalid value for amdgpu.seamless: %d\n",
1608			  amdgpu_seamless);
1609		return false;
1610	}
1611
1612	if (!(adev->flags & AMD_IS_APU))
1613		return false;
1614
1615	if (adev->mman.keep_stolen_vga_memory)
1616		return false;
1617
1618	return amdgpu_ip_version(adev, DCE_HWIP, 0) >= IP_VERSION(3, 0, 0);
1619}
1620
1621/*
1622 * Intel hosts such as Rocket Lake, Alder Lake, Raptor Lake and Sapphire Rapids
1623 * don't support dynamic speed switching. Until we have confirmation from Intel
1624 * that a specific host supports it, it's safer that we keep it disabled for all.
1625 *
1626 * https://edc.intel.com/content/www/us/en/design/products/platforms/details/raptor-lake-s/13th-generation-core-processors-datasheet-volume-1-of-2/005/pci-express-support/
1627 * https://gitlab.freedesktop.org/drm/amd/-/issues/2663
1628 */
1629static bool amdgpu_device_pcie_dynamic_switching_supported(struct amdgpu_device *adev)
1630{
1631#if IS_ENABLED(CONFIG_X86)
1632	struct cpuinfo_x86 *c = &cpu_data(0);
1633
1634	/* eGPU change speeds based on USB4 fabric conditions */
1635	if (dev_is_removable(adev->dev))
1636		return true;
1637
1638	if (c->x86_vendor == X86_VENDOR_INTEL)
1639		return false;
1640#endif
1641	return true;
1642}
1643
1644/**
1645 * amdgpu_device_should_use_aspm - check if the device should program ASPM
1646 *
1647 * @adev: amdgpu_device pointer
1648 *
1649 * Confirm whether the module parameter and pcie bridge agree that ASPM should
1650 * be set for this device.
1651 *
1652 * Returns true if it should be used or false if not.
1653 */
1654bool amdgpu_device_should_use_aspm(struct amdgpu_device *adev)
1655{
1656	switch (amdgpu_aspm) {
1657	case -1:
1658		break;
1659	case 0:
1660		return false;
1661	case 1:
1662		return true;
1663	default:
1664		return false;
1665	}
1666	if (adev->flags & AMD_IS_APU)
1667		return false;
1668	if (!(adev->pm.pp_feature & PP_PCIE_DPM_MASK))
1669		return false;
1670	return pcie_aspm_enabled(adev->pdev);
1671}
1672
1673/* if we get transitioned to only one device, take VGA back */
1674/**
1675 * amdgpu_device_vga_set_decode - enable/disable vga decode
1676 *
1677 * @pdev: PCI device pointer
1678 * @state: enable/disable vga decode
1679 *
1680 * Enable/disable vga decode (all asics).
1681 * Returns VGA resource flags.
1682 */
1683static unsigned int amdgpu_device_vga_set_decode(struct pci_dev *pdev,
1684		bool state)
1685{
1686	struct amdgpu_device *adev = drm_to_adev(pci_get_drvdata(pdev));
1687
1688	amdgpu_asic_set_vga_state(adev, state);
1689	if (state)
1690		return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
1691		       VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
1692	else
1693		return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
1694}
1695
1696/**
1697 * amdgpu_device_check_block_size - validate the vm block size
1698 *
1699 * @adev: amdgpu_device pointer
1700 *
1701 * Validates the vm block size specified via module parameter.
1702 * The vm block size defines number of bits in page table versus page directory,
1703 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
1704 * page table and the remaining bits are in the page directory.
1705 */
1706static void amdgpu_device_check_block_size(struct amdgpu_device *adev)
1707{
1708	/* defines number of bits in page table versus page directory,
1709	 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
1710	 * page table and the remaining bits are in the page directory
1711	 */
1712	if (amdgpu_vm_block_size == -1)
1713		return;
1714
1715	if (amdgpu_vm_block_size < 9) {
1716		dev_warn(adev->dev, "VM page table size (%d) too small\n",
1717			 amdgpu_vm_block_size);
1718		amdgpu_vm_block_size = -1;
1719	}
1720}
1721
1722/**
1723 * amdgpu_device_check_vm_size - validate the vm size
1724 *
1725 * @adev: amdgpu_device pointer
1726 *
1727 * Validates the vm size in GB specified via module parameter.
1728 * The VM size is the size of the GPU virtual memory space in GB.
1729 */
1730static void amdgpu_device_check_vm_size(struct amdgpu_device *adev)
1731{
1732	/* no need to check the default value */
1733	if (amdgpu_vm_size == -1)
1734		return;
1735
1736	if (amdgpu_vm_size < 1) {
1737		dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
1738			 amdgpu_vm_size);
1739		amdgpu_vm_size = -1;
1740	}
1741}
1742
1743static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev)
1744{
1745	struct sysinfo si;
1746	bool is_os_64 = (sizeof(void *) == 8);
1747	uint64_t total_memory;
1748	uint64_t dram_size_seven_GB = 0x1B8000000;
1749	uint64_t dram_size_three_GB = 0xB8000000;
1750
1751	if (amdgpu_smu_memory_pool_size == 0)
1752		return;
1753
1754	if (!is_os_64) {
1755		DRM_WARN("Not 64-bit OS, feature not supported\n");
1756		goto def_value;
1757	}
1758	si_meminfo(&si);
1759	total_memory = (uint64_t)si.totalram * si.mem_unit;
1760
1761	if ((amdgpu_smu_memory_pool_size == 1) ||
1762		(amdgpu_smu_memory_pool_size == 2)) {
1763		if (total_memory < dram_size_three_GB)
1764			goto def_value1;
1765	} else if ((amdgpu_smu_memory_pool_size == 4) ||
1766		(amdgpu_smu_memory_pool_size == 8)) {
1767		if (total_memory < dram_size_seven_GB)
1768			goto def_value1;
1769	} else {
1770		DRM_WARN("Smu memory pool size not supported\n");
1771		goto def_value;
1772	}
1773	adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28;
1774
1775	return;
1776
1777def_value1:
1778	DRM_WARN("No enough system memory\n");
1779def_value:
1780	adev->pm.smu_prv_buffer_size = 0;
1781}
1782
1783static int amdgpu_device_init_apu_flags(struct amdgpu_device *adev)
1784{
1785	if (!(adev->flags & AMD_IS_APU) ||
1786	    adev->asic_type < CHIP_RAVEN)
1787		return 0;
1788
1789	switch (adev->asic_type) {
1790	case CHIP_RAVEN:
1791		if (adev->pdev->device == 0x15dd)
1792			adev->apu_flags |= AMD_APU_IS_RAVEN;
1793		if (adev->pdev->device == 0x15d8)
1794			adev->apu_flags |= AMD_APU_IS_PICASSO;
1795		break;
1796	case CHIP_RENOIR:
1797		if ((adev->pdev->device == 0x1636) ||
1798		    (adev->pdev->device == 0x164c))
1799			adev->apu_flags |= AMD_APU_IS_RENOIR;
1800		else
1801			adev->apu_flags |= AMD_APU_IS_GREEN_SARDINE;
1802		break;
1803	case CHIP_VANGOGH:
1804		adev->apu_flags |= AMD_APU_IS_VANGOGH;
1805		break;
1806	case CHIP_YELLOW_CARP:
1807		break;
1808	case CHIP_CYAN_SKILLFISH:
1809		if ((adev->pdev->device == 0x13FE) ||
1810		    (adev->pdev->device == 0x143F))
1811			adev->apu_flags |= AMD_APU_IS_CYAN_SKILLFISH2;
1812		break;
1813	default:
1814		break;
1815	}
1816
1817	return 0;
1818}
1819
1820/**
1821 * amdgpu_device_check_arguments - validate module params
1822 *
1823 * @adev: amdgpu_device pointer
1824 *
1825 * Validates certain module parameters and updates
1826 * the associated values used by the driver (all asics).
1827 */
1828static int amdgpu_device_check_arguments(struct amdgpu_device *adev)
1829{
1830	if (amdgpu_sched_jobs < 4) {
1831		dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
1832			 amdgpu_sched_jobs);
1833		amdgpu_sched_jobs = 4;
1834	} else if (!is_power_of_2(amdgpu_sched_jobs)) {
1835		dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
1836			 amdgpu_sched_jobs);
1837		amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
1838	}
1839
1840	if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) {
1841		/* gart size must be greater or equal to 32M */
1842		dev_warn(adev->dev, "gart size (%d) too small\n",
1843			 amdgpu_gart_size);
1844		amdgpu_gart_size = -1;
1845	}
1846
1847	if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) {
1848		/* gtt size must be greater or equal to 32M */
1849		dev_warn(adev->dev, "gtt size (%d) too small\n",
1850				 amdgpu_gtt_size);
1851		amdgpu_gtt_size = -1;
1852	}
1853
1854	/* valid range is between 4 and 9 inclusive */
1855	if (amdgpu_vm_fragment_size != -1 &&
1856	    (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) {
1857		dev_warn(adev->dev, "valid range is between 4 and 9\n");
1858		amdgpu_vm_fragment_size = -1;
1859	}
1860
1861	if (amdgpu_sched_hw_submission < 2) {
1862		dev_warn(adev->dev, "sched hw submission jobs (%d) must be at least 2\n",
1863			 amdgpu_sched_hw_submission);
1864		amdgpu_sched_hw_submission = 2;
1865	} else if (!is_power_of_2(amdgpu_sched_hw_submission)) {
1866		dev_warn(adev->dev, "sched hw submission jobs (%d) must be a power of 2\n",
1867			 amdgpu_sched_hw_submission);
1868		amdgpu_sched_hw_submission = roundup_pow_of_two(amdgpu_sched_hw_submission);
1869	}
1870
1871	if (amdgpu_reset_method < -1 || amdgpu_reset_method > 4) {
1872		dev_warn(adev->dev, "invalid option for reset method, reverting to default\n");
1873		amdgpu_reset_method = -1;
1874	}
1875
1876	amdgpu_device_check_smu_prv_buffer_size(adev);
1877
1878	amdgpu_device_check_vm_size(adev);
1879
1880	amdgpu_device_check_block_size(adev);
1881
1882	adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type);
1883
 
 
 
 
1884	return 0;
1885}
1886
1887/**
1888 * amdgpu_switcheroo_set_state - set switcheroo state
1889 *
1890 * @pdev: pci dev pointer
1891 * @state: vga_switcheroo state
1892 *
1893 * Callback for the switcheroo driver.  Suspends or resumes
1894 * the asics before or after it is powered up using ACPI methods.
1895 */
1896static void amdgpu_switcheroo_set_state(struct pci_dev *pdev,
1897					enum vga_switcheroo_state state)
1898{
1899	struct drm_device *dev = pci_get_drvdata(pdev);
1900	int r;
1901
1902	if (amdgpu_device_supports_px(dev) && state == VGA_SWITCHEROO_OFF)
1903		return;
1904
1905	if (state == VGA_SWITCHEROO_ON) {
1906		pr_info("switched on\n");
1907		/* don't suspend or resume card normally */
1908		dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1909
1910		pci_set_power_state(pdev, PCI_D0);
1911		amdgpu_device_load_pci_state(pdev);
1912		r = pci_enable_device(pdev);
1913		if (r)
1914			DRM_WARN("pci_enable_device failed (%d)\n", r);
1915		amdgpu_device_resume(dev, true);
1916
1917		dev->switch_power_state = DRM_SWITCH_POWER_ON;
1918	} else {
1919		pr_info("switched off\n");
1920		dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1921		amdgpu_device_prepare(dev);
1922		amdgpu_device_suspend(dev, true);
1923		amdgpu_device_cache_pci_state(pdev);
1924		/* Shut down the device */
1925		pci_disable_device(pdev);
1926		pci_set_power_state(pdev, PCI_D3cold);
1927		dev->switch_power_state = DRM_SWITCH_POWER_OFF;
1928	}
1929}
1930
1931/**
1932 * amdgpu_switcheroo_can_switch - see if switcheroo state can change
1933 *
1934 * @pdev: pci dev pointer
1935 *
1936 * Callback for the switcheroo driver.  Check of the switcheroo
1937 * state can be changed.
1938 * Returns true if the state can be changed, false if not.
1939 */
1940static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
1941{
1942	struct drm_device *dev = pci_get_drvdata(pdev);
1943
1944       /*
1945	* FIXME: open_count is protected by drm_global_mutex but that would lead to
1946	* locking inversion with the driver load path. And the access here is
1947	* completely racy anyway. So don't bother with locking for now.
1948	*/
1949	return atomic_read(&dev->open_count) == 0;
1950}
1951
1952static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
1953	.set_gpu_state = amdgpu_switcheroo_set_state,
1954	.reprobe = NULL,
1955	.can_switch = amdgpu_switcheroo_can_switch,
1956};
1957
1958/**
1959 * amdgpu_device_ip_set_clockgating_state - set the CG state
1960 *
1961 * @dev: amdgpu_device pointer
1962 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1963 * @state: clockgating state (gate or ungate)
1964 *
1965 * Sets the requested clockgating state for all instances of
1966 * the hardware IP specified.
1967 * Returns the error code from the last instance.
1968 */
1969int amdgpu_device_ip_set_clockgating_state(void *dev,
1970					   enum amd_ip_block_type block_type,
1971					   enum amd_clockgating_state state)
1972{
1973	struct amdgpu_device *adev = dev;
1974	int i, r = 0;
1975
1976	for (i = 0; i < adev->num_ip_blocks; i++) {
1977		if (!adev->ip_blocks[i].status.valid)
1978			continue;
1979		if (adev->ip_blocks[i].version->type != block_type)
1980			continue;
1981		if (!adev->ip_blocks[i].version->funcs->set_clockgating_state)
1982			continue;
1983		r = adev->ip_blocks[i].version->funcs->set_clockgating_state(
1984			(void *)adev, state);
1985		if (r)
1986			DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",
1987				  adev->ip_blocks[i].version->funcs->name, r);
1988	}
1989	return r;
1990}
1991
1992/**
1993 * amdgpu_device_ip_set_powergating_state - set the PG state
1994 *
1995 * @dev: amdgpu_device pointer
1996 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1997 * @state: powergating state (gate or ungate)
1998 *
1999 * Sets the requested powergating state for all instances of
2000 * the hardware IP specified.
2001 * Returns the error code from the last instance.
2002 */
2003int amdgpu_device_ip_set_powergating_state(void *dev,
2004					   enum amd_ip_block_type block_type,
2005					   enum amd_powergating_state state)
2006{
2007	struct amdgpu_device *adev = dev;
2008	int i, r = 0;
2009
2010	for (i = 0; i < adev->num_ip_blocks; i++) {
2011		if (!adev->ip_blocks[i].status.valid)
2012			continue;
2013		if (adev->ip_blocks[i].version->type != block_type)
2014			continue;
2015		if (!adev->ip_blocks[i].version->funcs->set_powergating_state)
2016			continue;
2017		r = adev->ip_blocks[i].version->funcs->set_powergating_state(
2018			(void *)adev, state);
2019		if (r)
2020			DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",
2021				  adev->ip_blocks[i].version->funcs->name, r);
2022	}
2023	return r;
2024}
2025
2026/**
2027 * amdgpu_device_ip_get_clockgating_state - get the CG state
2028 *
2029 * @adev: amdgpu_device pointer
2030 * @flags: clockgating feature flags
2031 *
2032 * Walks the list of IPs on the device and updates the clockgating
2033 * flags for each IP.
2034 * Updates @flags with the feature flags for each hardware IP where
2035 * clockgating is enabled.
2036 */
2037void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev,
2038					    u64 *flags)
2039{
2040	int i;
2041
2042	for (i = 0; i < adev->num_ip_blocks; i++) {
2043		if (!adev->ip_blocks[i].status.valid)
2044			continue;
2045		if (adev->ip_blocks[i].version->funcs->get_clockgating_state)
2046			adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags);
2047	}
2048}
2049
2050/**
2051 * amdgpu_device_ip_wait_for_idle - wait for idle
2052 *
2053 * @adev: amdgpu_device pointer
2054 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
2055 *
2056 * Waits for the request hardware IP to be idle.
2057 * Returns 0 for success or a negative error code on failure.
2058 */
2059int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev,
2060				   enum amd_ip_block_type block_type)
2061{
2062	int i, r;
2063
2064	for (i = 0; i < adev->num_ip_blocks; i++) {
2065		if (!adev->ip_blocks[i].status.valid)
2066			continue;
2067		if (adev->ip_blocks[i].version->type == block_type) {
2068			r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
2069			if (r)
2070				return r;
2071			break;
2072		}
2073	}
2074	return 0;
2075
2076}
2077
2078/**
2079 * amdgpu_device_ip_is_idle - is the hardware IP idle
2080 *
2081 * @adev: amdgpu_device pointer
2082 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
2083 *
2084 * Check if the hardware IP is idle or not.
2085 * Returns true if it the IP is idle, false if not.
2086 */
2087bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev,
2088			      enum amd_ip_block_type block_type)
2089{
2090	int i;
2091
2092	for (i = 0; i < adev->num_ip_blocks; i++) {
2093		if (!adev->ip_blocks[i].status.valid)
2094			continue;
2095		if (adev->ip_blocks[i].version->type == block_type)
2096			return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
2097	}
2098	return true;
2099
2100}
2101
2102/**
2103 * amdgpu_device_ip_get_ip_block - get a hw IP pointer
2104 *
2105 * @adev: amdgpu_device pointer
2106 * @type: Type of hardware IP (SMU, GFX, UVD, etc.)
2107 *
2108 * Returns a pointer to the hardware IP block structure
2109 * if it exists for the asic, otherwise NULL.
2110 */
2111struct amdgpu_ip_block *
2112amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,
2113			      enum amd_ip_block_type type)
2114{
2115	int i;
2116
2117	for (i = 0; i < adev->num_ip_blocks; i++)
2118		if (adev->ip_blocks[i].version->type == type)
2119			return &adev->ip_blocks[i];
2120
2121	return NULL;
2122}
2123
2124/**
2125 * amdgpu_device_ip_block_version_cmp
2126 *
2127 * @adev: amdgpu_device pointer
2128 * @type: enum amd_ip_block_type
2129 * @major: major version
2130 * @minor: minor version
2131 *
2132 * return 0 if equal or greater
2133 * return 1 if smaller or the ip_block doesn't exist
2134 */
2135int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev,
2136				       enum amd_ip_block_type type,
2137				       u32 major, u32 minor)
2138{
2139	struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type);
2140
2141	if (ip_block && ((ip_block->version->major > major) ||
2142			((ip_block->version->major == major) &&
2143			(ip_block->version->minor >= minor))))
2144		return 0;
2145
2146	return 1;
2147}
2148
2149/**
2150 * amdgpu_device_ip_block_add
2151 *
2152 * @adev: amdgpu_device pointer
2153 * @ip_block_version: pointer to the IP to add
2154 *
2155 * Adds the IP block driver information to the collection of IPs
2156 * on the asic.
2157 */
2158int amdgpu_device_ip_block_add(struct amdgpu_device *adev,
2159			       const struct amdgpu_ip_block_version *ip_block_version)
2160{
2161	if (!ip_block_version)
2162		return -EINVAL;
2163
2164	switch (ip_block_version->type) {
2165	case AMD_IP_BLOCK_TYPE_VCN:
2166		if (adev->harvest_ip_mask & AMD_HARVEST_IP_VCN_MASK)
2167			return 0;
2168		break;
2169	case AMD_IP_BLOCK_TYPE_JPEG:
2170		if (adev->harvest_ip_mask & AMD_HARVEST_IP_JPEG_MASK)
2171			return 0;
2172		break;
2173	default:
2174		break;
2175	}
2176
2177	DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks,
2178		  ip_block_version->funcs->name);
2179
2180	adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
2181
2182	return 0;
2183}
2184
2185/**
2186 * amdgpu_device_enable_virtual_display - enable virtual display feature
2187 *
2188 * @adev: amdgpu_device pointer
2189 *
2190 * Enabled the virtual display feature if the user has enabled it via
2191 * the module parameter virtual_display.  This feature provides a virtual
2192 * display hardware on headless boards or in virtualized environments.
2193 * This function parses and validates the configuration string specified by
2194 * the user and configues the virtual display configuration (number of
2195 * virtual connectors, crtcs, etc.) specified.
2196 */
2197static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
2198{
2199	adev->enable_virtual_display = false;
2200
2201	if (amdgpu_virtual_display) {
2202		const char *pci_address_name = pci_name(adev->pdev);
2203		char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
2204
2205		pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
2206		pciaddstr_tmp = pciaddstr;
2207		while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
2208			pciaddname = strsep(&pciaddname_tmp, ",");
2209			if (!strcmp("all", pciaddname)
2210			    || !strcmp(pci_address_name, pciaddname)) {
2211				long num_crtc;
2212				int res = -1;
2213
2214				adev->enable_virtual_display = true;
2215
2216				if (pciaddname_tmp)
2217					res = kstrtol(pciaddname_tmp, 10,
2218						      &num_crtc);
2219
2220				if (!res) {
2221					if (num_crtc < 1)
2222						num_crtc = 1;
2223					if (num_crtc > 6)
2224						num_crtc = 6;
2225					adev->mode_info.num_crtc = num_crtc;
2226				} else {
2227					adev->mode_info.num_crtc = 1;
2228				}
2229				break;
2230			}
2231		}
2232
2233		DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
2234			 amdgpu_virtual_display, pci_address_name,
2235			 adev->enable_virtual_display, adev->mode_info.num_crtc);
2236
2237		kfree(pciaddstr);
2238	}
2239}
2240
2241void amdgpu_device_set_sriov_virtual_display(struct amdgpu_device *adev)
2242{
2243	if (amdgpu_sriov_vf(adev) && !adev->enable_virtual_display) {
2244		adev->mode_info.num_crtc = 1;
2245		adev->enable_virtual_display = true;
2246		DRM_INFO("virtual_display:%d, num_crtc:%d\n",
2247			 adev->enable_virtual_display, adev->mode_info.num_crtc);
2248	}
2249}
2250
2251/**
2252 * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware
2253 *
2254 * @adev: amdgpu_device pointer
2255 *
2256 * Parses the asic configuration parameters specified in the gpu info
2257 * firmware and makes them availale to the driver for use in configuring
2258 * the asic.
2259 * Returns 0 on success, -EINVAL on failure.
2260 */
2261static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev)
2262{
2263	const char *chip_name;
2264	char fw_name[40];
2265	int err;
2266	const struct gpu_info_firmware_header_v1_0 *hdr;
2267
2268	adev->firmware.gpu_info_fw = NULL;
2269
2270	if (adev->mman.discovery_bin)
2271		return 0;
 
 
 
 
 
 
 
 
 
2272
2273	switch (adev->asic_type) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2274	default:
2275		return 0;
2276	case CHIP_VEGA10:
2277		chip_name = "vega10";
2278		break;
2279	case CHIP_VEGA12:
2280		chip_name = "vega12";
2281		break;
2282	case CHIP_RAVEN:
2283		if (adev->apu_flags & AMD_APU_IS_RAVEN2)
2284			chip_name = "raven2";
2285		else if (adev->apu_flags & AMD_APU_IS_PICASSO)
2286			chip_name = "picasso";
2287		else
2288			chip_name = "raven";
2289		break;
2290	case CHIP_ARCTURUS:
2291		chip_name = "arcturus";
2292		break;
 
 
 
 
 
 
 
 
 
 
 
 
2293	case CHIP_NAVI12:
2294		chip_name = "navi12";
2295		break;
 
 
 
 
 
 
2296	}
2297
2298	snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name);
2299	err = amdgpu_ucode_request(adev, &adev->firmware.gpu_info_fw, fw_name);
 
 
 
 
 
 
 
2300	if (err) {
2301		dev_err(adev->dev,
2302			"Failed to get gpu_info firmware \"%s\"\n",
2303			fw_name);
2304		goto out;
2305	}
2306
2307	hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data;
2308	amdgpu_ucode_print_gpu_info_hdr(&hdr->header);
2309
2310	switch (hdr->version_major) {
2311	case 1:
2312	{
2313		const struct gpu_info_firmware_v1_0 *gpu_info_fw =
2314			(const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data +
2315								le32_to_cpu(hdr->header.ucode_array_offset_bytes));
2316
2317		/*
2318		 * Should be droped when DAL no longer needs it.
2319		 */
2320		if (adev->asic_type == CHIP_NAVI12)
2321			goto parse_soc_bounding_box;
2322
2323		adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se);
2324		adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh);
2325		adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se);
2326		adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se);
2327		adev->gfx.config.max_texture_channel_caches =
2328			le32_to_cpu(gpu_info_fw->gc_num_tccs);
2329		adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs);
2330		adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds);
2331		adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth);
2332		adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth);
2333		adev->gfx.config.double_offchip_lds_buf =
2334			le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer);
2335		adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size);
2336		adev->gfx.cu_info.max_waves_per_simd =
2337			le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd);
2338		adev->gfx.cu_info.max_scratch_slots_per_cu =
2339			le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu);
2340		adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size);
2341		if (hdr->version_minor >= 1) {
2342			const struct gpu_info_firmware_v1_1 *gpu_info_fw =
2343				(const struct gpu_info_firmware_v1_1 *)(adev->firmware.gpu_info_fw->data +
2344									le32_to_cpu(hdr->header.ucode_array_offset_bytes));
2345			adev->gfx.config.num_sc_per_sh =
2346				le32_to_cpu(gpu_info_fw->num_sc_per_sh);
2347			adev->gfx.config.num_packer_per_sc =
2348				le32_to_cpu(gpu_info_fw->num_packer_per_sc);
2349		}
2350
2351parse_soc_bounding_box:
2352		/*
2353		 * soc bounding box info is not integrated in disocovery table,
2354		 * we always need to parse it from gpu info firmware if needed.
2355		 */
2356		if (hdr->version_minor == 2) {
2357			const struct gpu_info_firmware_v1_2 *gpu_info_fw =
2358				(const struct gpu_info_firmware_v1_2 *)(adev->firmware.gpu_info_fw->data +
2359									le32_to_cpu(hdr->header.ucode_array_offset_bytes));
2360			adev->dm.soc_bounding_box = &gpu_info_fw->soc_bounding_box;
2361		}
2362		break;
2363	}
2364	default:
2365		dev_err(adev->dev,
2366			"Unsupported gpu_info table %d\n", hdr->header.ucode_version);
2367		err = -EINVAL;
2368		goto out;
2369	}
2370out:
2371	return err;
2372}
2373
2374/**
2375 * amdgpu_device_ip_early_init - run early init for hardware IPs
2376 *
2377 * @adev: amdgpu_device pointer
2378 *
2379 * Early initialization pass for hardware IPs.  The hardware IPs that make
2380 * up each asic are discovered each IP's early_init callback is run.  This
2381 * is the first stage in initializing the asic.
2382 * Returns 0 on success, negative error code on failure.
2383 */
2384static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
2385{
2386	struct pci_dev *parent;
2387	int i, r;
2388	bool total;
2389
2390	amdgpu_device_enable_virtual_display(adev);
2391
2392	if (amdgpu_sriov_vf(adev)) {
2393		r = amdgpu_virt_request_full_gpu(adev, true);
2394		if (r)
2395			return r;
2396	}
2397
2398	switch (adev->asic_type) {
2399#ifdef CONFIG_DRM_AMDGPU_SI
2400	case CHIP_VERDE:
2401	case CHIP_TAHITI:
2402	case CHIP_PITCAIRN:
2403	case CHIP_OLAND:
2404	case CHIP_HAINAN:
2405		adev->family = AMDGPU_FAMILY_SI;
2406		r = si_set_ip_blocks(adev);
2407		if (r)
2408			return r;
2409		break;
2410#endif
2411#ifdef CONFIG_DRM_AMDGPU_CIK
2412	case CHIP_BONAIRE:
2413	case CHIP_HAWAII:
2414	case CHIP_KAVERI:
2415	case CHIP_KABINI:
2416	case CHIP_MULLINS:
2417		if (adev->flags & AMD_IS_APU)
2418			adev->family = AMDGPU_FAMILY_KV;
2419		else
2420			adev->family = AMDGPU_FAMILY_CI;
2421
2422		r = cik_set_ip_blocks(adev);
2423		if (r)
2424			return r;
2425		break;
2426#endif
2427	case CHIP_TOPAZ:
2428	case CHIP_TONGA:
2429	case CHIP_FIJI:
2430	case CHIP_POLARIS10:
2431	case CHIP_POLARIS11:
2432	case CHIP_POLARIS12:
2433	case CHIP_VEGAM:
2434	case CHIP_CARRIZO:
2435	case CHIP_STONEY:
2436		if (adev->flags & AMD_IS_APU)
2437			adev->family = AMDGPU_FAMILY_CZ;
2438		else
2439			adev->family = AMDGPU_FAMILY_VI;
2440
2441		r = vi_set_ip_blocks(adev);
2442		if (r)
2443			return r;
2444		break;
2445	default:
2446		r = amdgpu_discovery_set_ip_blocks(adev);
 
 
 
 
 
 
 
 
 
 
 
2447		if (r)
2448			return r;
2449		break;
2450	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2451
2452	if (amdgpu_has_atpx() &&
2453	    (amdgpu_is_atpx_hybrid() ||
2454	     amdgpu_has_atpx_dgpu_power_cntl()) &&
2455	    ((adev->flags & AMD_IS_APU) == 0) &&
2456	    !dev_is_removable(&adev->pdev->dev))
2457		adev->flags |= AMD_IS_PX;
2458
2459	if (!(adev->flags & AMD_IS_APU)) {
2460		parent = pcie_find_root_port(adev->pdev);
2461		adev->has_pr3 = parent ? pci_pr3_present(parent) : false;
2462	}
2463
 
2464
2465	adev->pm.pp_feature = amdgpu_pp_feature_mask;
2466	if (amdgpu_sriov_vf(adev) || sched_policy == KFD_SCHED_POLICY_NO_HWS)
2467		adev->pm.pp_feature &= ~PP_GFXOFF_MASK;
2468	if (amdgpu_sriov_vf(adev) && adev->asic_type == CHIP_SIENNA_CICHLID)
2469		adev->pm.pp_feature &= ~PP_OVERDRIVE_MASK;
2470	if (!amdgpu_device_pcie_dynamic_switching_supported(adev))
2471		adev->pm.pp_feature &= ~PP_PCIE_DPM_MASK;
2472
2473	total = true;
2474	for (i = 0; i < adev->num_ip_blocks; i++) {
2475		if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
2476			DRM_WARN("disabled ip block: %d <%s>\n",
2477				  i, adev->ip_blocks[i].version->funcs->name);
2478			adev->ip_blocks[i].status.valid = false;
2479		} else {
2480			if (adev->ip_blocks[i].version->funcs->early_init) {
2481				r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
2482				if (r == -ENOENT) {
2483					adev->ip_blocks[i].status.valid = false;
2484				} else if (r) {
2485					DRM_ERROR("early_init of IP block <%s> failed %d\n",
2486						  adev->ip_blocks[i].version->funcs->name, r);
2487					total = false;
2488				} else {
2489					adev->ip_blocks[i].status.valid = true;
2490				}
2491			} else {
2492				adev->ip_blocks[i].status.valid = true;
2493			}
2494		}
2495		/* get the vbios after the asic_funcs are set up */
2496		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) {
2497			r = amdgpu_device_parse_gpu_info_fw(adev);
2498			if (r)
2499				return r;
2500
2501			/* Read BIOS */
2502			if (amdgpu_device_read_bios(adev)) {
2503				if (!amdgpu_get_bios(adev))
2504					return -EINVAL;
2505
2506				r = amdgpu_atombios_init(adev);
2507				if (r) {
2508					dev_err(adev->dev, "amdgpu_atombios_init failed\n");
2509					amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0);
2510					return r;
2511				}
2512			}
2513
2514			/*get pf2vf msg info at it's earliest time*/
2515			if (amdgpu_sriov_vf(adev))
2516				amdgpu_virt_init_data_exchange(adev);
2517
2518		}
2519	}
2520	if (!total)
2521		return -ENODEV;
2522
2523	amdgpu_amdkfd_device_probe(adev);
2524	adev->cg_flags &= amdgpu_cg_mask;
2525	adev->pg_flags &= amdgpu_pg_mask;
2526
2527	return 0;
2528}
2529
2530static int amdgpu_device_ip_hw_init_phase1(struct amdgpu_device *adev)
2531{
2532	int i, r;
2533
2534	for (i = 0; i < adev->num_ip_blocks; i++) {
2535		if (!adev->ip_blocks[i].status.sw)
2536			continue;
2537		if (adev->ip_blocks[i].status.hw)
2538			continue;
2539		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2540		    (amdgpu_sriov_vf(adev) && (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)) ||
2541		    adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
2542			r = adev->ip_blocks[i].version->funcs->hw_init(adev);
2543			if (r) {
2544				DRM_ERROR("hw_init of IP block <%s> failed %d\n",
2545					  adev->ip_blocks[i].version->funcs->name, r);
2546				return r;
2547			}
2548			adev->ip_blocks[i].status.hw = true;
2549		}
2550	}
2551
2552	return 0;
2553}
2554
2555static int amdgpu_device_ip_hw_init_phase2(struct amdgpu_device *adev)
2556{
2557	int i, r;
2558
2559	for (i = 0; i < adev->num_ip_blocks; i++) {
2560		if (!adev->ip_blocks[i].status.sw)
2561			continue;
2562		if (adev->ip_blocks[i].status.hw)
2563			continue;
2564		r = adev->ip_blocks[i].version->funcs->hw_init(adev);
2565		if (r) {
2566			DRM_ERROR("hw_init of IP block <%s> failed %d\n",
2567				  adev->ip_blocks[i].version->funcs->name, r);
2568			return r;
2569		}
2570		adev->ip_blocks[i].status.hw = true;
2571	}
2572
2573	return 0;
2574}
2575
2576static int amdgpu_device_fw_loading(struct amdgpu_device *adev)
2577{
2578	int r = 0;
2579	int i;
2580	uint32_t smu_version;
2581
2582	if (adev->asic_type >= CHIP_VEGA10) {
2583		for (i = 0; i < adev->num_ip_blocks; i++) {
2584			if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_PSP)
2585				continue;
2586
2587			if (!adev->ip_blocks[i].status.sw)
2588				continue;
2589
2590			/* no need to do the fw loading again if already done*/
2591			if (adev->ip_blocks[i].status.hw == true)
2592				break;
2593
2594			if (amdgpu_in_reset(adev) || adev->in_suspend) {
2595				r = adev->ip_blocks[i].version->funcs->resume(adev);
2596				if (r) {
2597					DRM_ERROR("resume of IP block <%s> failed %d\n",
2598							  adev->ip_blocks[i].version->funcs->name, r);
2599					return r;
2600				}
2601			} else {
2602				r = adev->ip_blocks[i].version->funcs->hw_init(adev);
2603				if (r) {
2604					DRM_ERROR("hw_init of IP block <%s> failed %d\n",
2605							  adev->ip_blocks[i].version->funcs->name, r);
2606					return r;
2607				}
2608			}
2609
2610			adev->ip_blocks[i].status.hw = true;
2611			break;
2612		}
2613	}
2614
2615	if (!amdgpu_sriov_vf(adev) || adev->asic_type == CHIP_TONGA)
2616		r = amdgpu_pm_load_smu_firmware(adev, &smu_version);
2617
2618	return r;
2619}
2620
2621static int amdgpu_device_init_schedulers(struct amdgpu_device *adev)
2622{
2623	long timeout;
2624	int r, i;
2625
2626	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
2627		struct amdgpu_ring *ring = adev->rings[i];
2628
2629		/* No need to setup the GPU scheduler for rings that don't need it */
2630		if (!ring || ring->no_scheduler)
2631			continue;
2632
2633		switch (ring->funcs->type) {
2634		case AMDGPU_RING_TYPE_GFX:
2635			timeout = adev->gfx_timeout;
2636			break;
2637		case AMDGPU_RING_TYPE_COMPUTE:
2638			timeout = adev->compute_timeout;
2639			break;
2640		case AMDGPU_RING_TYPE_SDMA:
2641			timeout = adev->sdma_timeout;
2642			break;
2643		default:
2644			timeout = adev->video_timeout;
2645			break;
2646		}
2647
2648		r = drm_sched_init(&ring->sched, &amdgpu_sched_ops, NULL,
2649				   DRM_SCHED_PRIORITY_COUNT,
2650				   ring->num_hw_submission, 0,
2651				   timeout, adev->reset_domain->wq,
2652				   ring->sched_score, ring->name,
2653				   adev->dev);
2654		if (r) {
2655			DRM_ERROR("Failed to create scheduler on ring %s.\n",
2656				  ring->name);
2657			return r;
2658		}
2659		r = amdgpu_uvd_entity_init(adev, ring);
2660		if (r) {
2661			DRM_ERROR("Failed to create UVD scheduling entity on ring %s.\n",
2662				  ring->name);
2663			return r;
2664		}
2665		r = amdgpu_vce_entity_init(adev, ring);
2666		if (r) {
2667			DRM_ERROR("Failed to create VCE scheduling entity on ring %s.\n",
2668				  ring->name);
2669			return r;
2670		}
2671	}
2672
2673	amdgpu_xcp_update_partition_sched_list(adev);
2674
2675	return 0;
2676}
2677
2678
2679/**
2680 * amdgpu_device_ip_init - run init for hardware IPs
2681 *
2682 * @adev: amdgpu_device pointer
2683 *
2684 * Main initialization pass for hardware IPs.  The list of all the hardware
2685 * IPs that make up the asic is walked and the sw_init and hw_init callbacks
2686 * are run.  sw_init initializes the software state associated with each IP
2687 * and hw_init initializes the hardware associated with each IP.
2688 * Returns 0 on success, negative error code on failure.
2689 */
2690static int amdgpu_device_ip_init(struct amdgpu_device *adev)
2691{
2692	int i, r;
2693
2694	r = amdgpu_ras_init(adev);
2695	if (r)
2696		return r;
2697
2698	for (i = 0; i < adev->num_ip_blocks; i++) {
2699		if (!adev->ip_blocks[i].status.valid)
2700			continue;
2701		r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
2702		if (r) {
2703			DRM_ERROR("sw_init of IP block <%s> failed %d\n",
2704				  adev->ip_blocks[i].version->funcs->name, r);
2705			goto init_failed;
2706		}
2707		adev->ip_blocks[i].status.sw = true;
2708
2709		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) {
2710			/* need to do common hw init early so everything is set up for gmc */
2711			r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
2712			if (r) {
2713				DRM_ERROR("hw_init %d failed %d\n", i, r);
2714				goto init_failed;
2715			}
2716			adev->ip_blocks[i].status.hw = true;
2717		} else if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
2718			/* need to do gmc hw init early so we can allocate gpu mem */
2719			/* Try to reserve bad pages early */
2720			if (amdgpu_sriov_vf(adev))
2721				amdgpu_virt_exchange_data(adev);
2722
2723			r = amdgpu_device_mem_scratch_init(adev);
2724			if (r) {
2725				DRM_ERROR("amdgpu_mem_scratch_init failed %d\n", r);
2726				goto init_failed;
2727			}
2728			r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
2729			if (r) {
2730				DRM_ERROR("hw_init %d failed %d\n", i, r);
2731				goto init_failed;
2732			}
2733			r = amdgpu_device_wb_init(adev);
2734			if (r) {
2735				DRM_ERROR("amdgpu_device_wb_init failed %d\n", r);
2736				goto init_failed;
2737			}
2738			adev->ip_blocks[i].status.hw = true;
2739
2740			/* right after GMC hw init, we create CSA */
2741			if (adev->gfx.mcbp) {
2742				r = amdgpu_allocate_static_csa(adev, &adev->virt.csa_obj,
2743							       AMDGPU_GEM_DOMAIN_VRAM |
2744							       AMDGPU_GEM_DOMAIN_GTT,
2745							       AMDGPU_CSA_SIZE);
2746				if (r) {
2747					DRM_ERROR("allocate CSA failed %d\n", r);
2748					goto init_failed;
2749				}
2750			}
2751
2752			r = amdgpu_seq64_init(adev);
2753			if (r) {
2754				DRM_ERROR("allocate seq64 failed %d\n", r);
2755				goto init_failed;
2756			}
2757		}
2758	}
2759
2760	if (amdgpu_sriov_vf(adev))
2761		amdgpu_virt_init_data_exchange(adev);
2762
2763	r = amdgpu_ib_pool_init(adev);
2764	if (r) {
2765		dev_err(adev->dev, "IB initialization failed (%d).\n", r);
2766		amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r);
2767		goto init_failed;
2768	}
2769
2770	r = amdgpu_ucode_create_bo(adev); /* create ucode bo when sw_init complete*/
2771	if (r)
2772		goto init_failed;
2773
2774	r = amdgpu_device_ip_hw_init_phase1(adev);
2775	if (r)
2776		goto init_failed;
2777
2778	r = amdgpu_device_fw_loading(adev);
2779	if (r)
2780		goto init_failed;
2781
2782	r = amdgpu_device_ip_hw_init_phase2(adev);
2783	if (r)
2784		goto init_failed;
2785
2786	/*
2787	 * retired pages will be loaded from eeprom and reserved here,
2788	 * it should be called after amdgpu_device_ip_hw_init_phase2  since
2789	 * for some ASICs the RAS EEPROM code relies on SMU fully functioning
2790	 * for I2C communication which only true at this point.
2791	 *
2792	 * amdgpu_ras_recovery_init may fail, but the upper only cares the
2793	 * failure from bad gpu situation and stop amdgpu init process
2794	 * accordingly. For other failed cases, it will still release all
2795	 * the resource and print error message, rather than returning one
2796	 * negative value to upper level.
2797	 *
2798	 * Note: theoretically, this should be called before all vram allocations
2799	 * to protect retired page from abusing
2800	 */
2801	r = amdgpu_ras_recovery_init(adev);
2802	if (r)
2803		goto init_failed;
2804
2805	/**
2806	 * In case of XGMI grab extra reference for reset domain for this device
2807	 */
2808	if (adev->gmc.xgmi.num_physical_nodes > 1) {
2809		if (amdgpu_xgmi_add_device(adev) == 0) {
2810			if (!amdgpu_sriov_vf(adev)) {
2811				struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev);
2812
2813				if (WARN_ON(!hive)) {
2814					r = -ENOENT;
2815					goto init_failed;
2816				}
2817
2818				if (!hive->reset_domain ||
2819				    !amdgpu_reset_get_reset_domain(hive->reset_domain)) {
2820					r = -ENOENT;
2821					amdgpu_put_xgmi_hive(hive);
2822					goto init_failed;
2823				}
2824
2825				/* Drop the early temporary reset domain we created for device */
2826				amdgpu_reset_put_reset_domain(adev->reset_domain);
2827				adev->reset_domain = hive->reset_domain;
2828				amdgpu_put_xgmi_hive(hive);
2829			}
2830		}
2831	}
2832
2833	r = amdgpu_device_init_schedulers(adev);
2834	if (r)
2835		goto init_failed;
2836
2837	if (adev->mman.buffer_funcs_ring->sched.ready)
2838		amdgpu_ttm_set_buffer_funcs_status(adev, true);
2839
2840	/* Don't init kfd if whole hive need to be reset during init */
2841	if (!adev->gmc.xgmi.pending_reset) {
2842		kgd2kfd_init_zone_device(adev);
2843		amdgpu_amdkfd_device_init(adev);
2844	}
2845
2846	amdgpu_fru_get_product_info(adev);
2847
2848init_failed:
 
 
2849
2850	return r;
2851}
2852
2853/**
2854 * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer
2855 *
2856 * @adev: amdgpu_device pointer
2857 *
2858 * Writes a reset magic value to the gart pointer in VRAM.  The driver calls
2859 * this function before a GPU reset.  If the value is retained after a
2860 * GPU reset, VRAM has not been lost.  Some GPU resets may destry VRAM contents.
2861 */
2862static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev)
2863{
2864	memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
2865}
2866
2867/**
2868 * amdgpu_device_check_vram_lost - check if vram is valid
2869 *
2870 * @adev: amdgpu_device pointer
2871 *
2872 * Checks the reset magic value written to the gart pointer in VRAM.
2873 * The driver calls this after a GPU reset to see if the contents of
2874 * VRAM is lost or now.
2875 * returns true if vram is lost, false if not.
2876 */
2877static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev)
2878{
2879	if (memcmp(adev->gart.ptr, adev->reset_magic,
2880			AMDGPU_RESET_MAGIC_NUM))
2881		return true;
2882
2883	if (!amdgpu_in_reset(adev))
2884		return false;
2885
2886	/*
2887	 * For all ASICs with baco/mode1 reset, the VRAM is
2888	 * always assumed to be lost.
2889	 */
2890	switch (amdgpu_asic_reset_method(adev)) {
2891	case AMD_RESET_METHOD_BACO:
2892	case AMD_RESET_METHOD_MODE1:
2893		return true;
2894	default:
2895		return false;
2896	}
2897}
2898
2899/**
2900 * amdgpu_device_set_cg_state - set clockgating for amdgpu device
2901 *
2902 * @adev: amdgpu_device pointer
2903 * @state: clockgating state (gate or ungate)
2904 *
2905 * The list of all the hardware IPs that make up the asic is walked and the
2906 * set_clockgating_state callbacks are run.
2907 * Late initialization pass enabling clockgating for hardware IPs.
2908 * Fini or suspend, pass disabling clockgating for hardware IPs.
2909 * Returns 0 on success, negative error code on failure.
2910 */
2911
2912int amdgpu_device_set_cg_state(struct amdgpu_device *adev,
2913			       enum amd_clockgating_state state)
2914{
2915	int i, j, r;
2916
2917	if (amdgpu_emu_mode == 1)
2918		return 0;
2919
2920	for (j = 0; j < adev->num_ip_blocks; j++) {
2921		i = state == AMD_CG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
2922		if (!adev->ip_blocks[i].status.late_initialized)
2923			continue;
2924		/* skip CG for GFX, SDMA on S0ix */
2925		if (adev->in_s0ix &&
2926		    (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX ||
2927		     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SDMA))
2928			continue;
2929		/* skip CG for VCE/UVD, it's handled specially */
2930		if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
2931		    adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
2932		    adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
2933		    adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG &&
2934		    adev->ip_blocks[i].version->funcs->set_clockgating_state) {
2935			/* enable clockgating to save power */
2936			r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
2937										     state);
2938			if (r) {
2939				DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
2940					  adev->ip_blocks[i].version->funcs->name, r);
2941				return r;
2942			}
2943		}
2944	}
2945
2946	return 0;
2947}
2948
2949int amdgpu_device_set_pg_state(struct amdgpu_device *adev,
2950			       enum amd_powergating_state state)
2951{
2952	int i, j, r;
2953
2954	if (amdgpu_emu_mode == 1)
2955		return 0;
2956
2957	for (j = 0; j < adev->num_ip_blocks; j++) {
2958		i = state == AMD_PG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
2959		if (!adev->ip_blocks[i].status.late_initialized)
2960			continue;
2961		/* skip PG for GFX, SDMA on S0ix */
2962		if (adev->in_s0ix &&
2963		    (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX ||
2964		     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SDMA))
2965			continue;
2966		/* skip CG for VCE/UVD, it's handled specially */
2967		if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
2968		    adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
2969		    adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
2970		    adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG &&
2971		    adev->ip_blocks[i].version->funcs->set_powergating_state) {
2972			/* enable powergating to save power */
2973			r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev,
2974											state);
2975			if (r) {
2976				DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n",
2977					  adev->ip_blocks[i].version->funcs->name, r);
2978				return r;
2979			}
2980		}
2981	}
2982	return 0;
2983}
2984
2985static int amdgpu_device_enable_mgpu_fan_boost(void)
2986{
2987	struct amdgpu_gpu_instance *gpu_ins;
2988	struct amdgpu_device *adev;
2989	int i, ret = 0;
2990
2991	mutex_lock(&mgpu_info.mutex);
2992
2993	/*
2994	 * MGPU fan boost feature should be enabled
2995	 * only when there are two or more dGPUs in
2996	 * the system
2997	 */
2998	if (mgpu_info.num_dgpu < 2)
2999		goto out;
3000
3001	for (i = 0; i < mgpu_info.num_dgpu; i++) {
3002		gpu_ins = &(mgpu_info.gpu_ins[i]);
3003		adev = gpu_ins->adev;
3004		if (!(adev->flags & AMD_IS_APU) &&
3005		    !gpu_ins->mgpu_fan_enabled) {
3006			ret = amdgpu_dpm_enable_mgpu_fan_boost(adev);
3007			if (ret)
3008				break;
3009
3010			gpu_ins->mgpu_fan_enabled = 1;
3011		}
3012	}
3013
3014out:
3015	mutex_unlock(&mgpu_info.mutex);
3016
3017	return ret;
3018}
3019
3020/**
3021 * amdgpu_device_ip_late_init - run late init for hardware IPs
3022 *
3023 * @adev: amdgpu_device pointer
3024 *
3025 * Late initialization pass for hardware IPs.  The list of all the hardware
3026 * IPs that make up the asic is walked and the late_init callbacks are run.
3027 * late_init covers any special initialization that an IP requires
3028 * after all of the have been initialized or something that needs to happen
3029 * late in the init process.
3030 * Returns 0 on success, negative error code on failure.
3031 */
3032static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
3033{
3034	struct amdgpu_gpu_instance *gpu_instance;
3035	int i = 0, r;
3036
3037	for (i = 0; i < adev->num_ip_blocks; i++) {
3038		if (!adev->ip_blocks[i].status.hw)
3039			continue;
3040		if (adev->ip_blocks[i].version->funcs->late_init) {
3041			r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
3042			if (r) {
3043				DRM_ERROR("late_init of IP block <%s> failed %d\n",
3044					  adev->ip_blocks[i].version->funcs->name, r);
3045				return r;
3046			}
3047		}
3048		adev->ip_blocks[i].status.late_initialized = true;
3049	}
3050
3051	r = amdgpu_ras_late_init(adev);
3052	if (r) {
3053		DRM_ERROR("amdgpu_ras_late_init failed %d", r);
3054		return r;
3055	}
3056
3057	amdgpu_ras_set_error_query_ready(adev, true);
3058
3059	amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE);
3060	amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE);
3061
3062	amdgpu_device_fill_reset_magic(adev);
3063
3064	r = amdgpu_device_enable_mgpu_fan_boost();
3065	if (r)
3066		DRM_ERROR("enable mgpu fan boost failed (%d).\n", r);
3067
3068	/* For passthrough configuration on arcturus and aldebaran, enable special handling SBR */
3069	if (amdgpu_passthrough(adev) &&
3070	    ((adev->asic_type == CHIP_ARCTURUS && adev->gmc.xgmi.num_physical_nodes > 1) ||
3071	     adev->asic_type == CHIP_ALDEBARAN))
3072		amdgpu_dpm_handle_passthrough_sbr(adev, true);
3073
3074	if (adev->gmc.xgmi.num_physical_nodes > 1) {
3075		mutex_lock(&mgpu_info.mutex);
3076
3077		/*
3078		 * Reset device p-state to low as this was booted with high.
3079		 *
3080		 * This should be performed only after all devices from the same
3081		 * hive get initialized.
3082		 *
3083		 * However, it's unknown how many device in the hive in advance.
3084		 * As this is counted one by one during devices initializations.
3085		 *
3086		 * So, we wait for all XGMI interlinked devices initialized.
3087		 * This may bring some delays as those devices may come from
3088		 * different hives. But that should be OK.
3089		 */
3090		if (mgpu_info.num_dgpu == adev->gmc.xgmi.num_physical_nodes) {
3091			for (i = 0; i < mgpu_info.num_gpu; i++) {
3092				gpu_instance = &(mgpu_info.gpu_ins[i]);
3093				if (gpu_instance->adev->flags & AMD_IS_APU)
3094					continue;
3095
3096				r = amdgpu_xgmi_set_pstate(gpu_instance->adev,
3097						AMDGPU_XGMI_PSTATE_MIN);
3098				if (r) {
3099					DRM_ERROR("pstate setting failed (%d).\n", r);
3100					break;
3101				}
3102			}
3103		}
3104
3105		mutex_unlock(&mgpu_info.mutex);
3106	}
3107
3108	return 0;
3109}
3110
3111/**
3112 * amdgpu_device_smu_fini_early - smu hw_fini wrapper
3113 *
3114 * @adev: amdgpu_device pointer
3115 *
3116 * For ASICs need to disable SMC first
3117 */
3118static void amdgpu_device_smu_fini_early(struct amdgpu_device *adev)
3119{
3120	int i, r;
3121
3122	if (amdgpu_ip_version(adev, GC_HWIP, 0) > IP_VERSION(9, 0, 0))
3123		return;
3124
3125	for (i = 0; i < adev->num_ip_blocks; i++) {
3126		if (!adev->ip_blocks[i].status.hw)
3127			continue;
3128		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
3129			r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
3130			/* XXX handle errors */
3131			if (r) {
3132				DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
3133					  adev->ip_blocks[i].version->funcs->name, r);
3134			}
3135			adev->ip_blocks[i].status.hw = false;
3136			break;
3137		}
3138	}
3139}
3140
3141static int amdgpu_device_ip_fini_early(struct amdgpu_device *adev)
3142{
3143	int i, r;
3144
3145	for (i = 0; i < adev->num_ip_blocks; i++) {
3146		if (!adev->ip_blocks[i].version->funcs->early_fini)
3147			continue;
3148
3149		r = adev->ip_blocks[i].version->funcs->early_fini((void *)adev);
3150		if (r) {
3151			DRM_DEBUG("early_fini of IP block <%s> failed %d\n",
3152				  adev->ip_blocks[i].version->funcs->name, r);
3153		}
3154	}
3155
 
 
3156	amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
3157	amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
3158
3159	amdgpu_amdkfd_suspend(adev, false);
3160
3161	/* Workaroud for ASICs need to disable SMC first */
3162	amdgpu_device_smu_fini_early(adev);
 
 
 
 
 
 
 
 
 
 
 
3163
3164	for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
3165		if (!adev->ip_blocks[i].status.hw)
3166			continue;
3167
3168		r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
3169		/* XXX handle errors */
3170		if (r) {
3171			DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
3172				  adev->ip_blocks[i].version->funcs->name, r);
3173		}
3174
3175		adev->ip_blocks[i].status.hw = false;
3176	}
3177
3178	if (amdgpu_sriov_vf(adev)) {
3179		if (amdgpu_virt_release_full_gpu(adev, false))
3180			DRM_ERROR("failed to release exclusive mode on fini\n");
3181	}
3182
3183	return 0;
3184}
3185
3186/**
3187 * amdgpu_device_ip_fini - run fini for hardware IPs
3188 *
3189 * @adev: amdgpu_device pointer
3190 *
3191 * Main teardown pass for hardware IPs.  The list of all the hardware
3192 * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks
3193 * are run.  hw_fini tears down the hardware associated with each IP
3194 * and sw_fini tears down any software state associated with each IP.
3195 * Returns 0 on success, negative error code on failure.
3196 */
3197static int amdgpu_device_ip_fini(struct amdgpu_device *adev)
3198{
3199	int i, r;
3200
3201	if (amdgpu_sriov_vf(adev) && adev->virt.ras_init_done)
3202		amdgpu_virt_release_ras_err_handler_data(adev);
3203
 
 
3204	if (adev->gmc.xgmi.num_physical_nodes > 1)
3205		amdgpu_xgmi_remove_device(adev);
3206
3207	amdgpu_amdkfd_device_fini_sw(adev);
3208
3209	for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
3210		if (!adev->ip_blocks[i].status.sw)
3211			continue;
3212
3213		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
3214			amdgpu_ucode_free_bo(adev);
3215			amdgpu_free_static_csa(&adev->virt.csa_obj);
3216			amdgpu_device_wb_fini(adev);
3217			amdgpu_device_mem_scratch_fini(adev);
3218			amdgpu_ib_pool_fini(adev);
3219			amdgpu_seq64_fini(adev);
3220		}
3221
3222		r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
3223		/* XXX handle errors */
3224		if (r) {
3225			DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
3226				  adev->ip_blocks[i].version->funcs->name, r);
3227		}
3228		adev->ip_blocks[i].status.sw = false;
3229		adev->ip_blocks[i].status.valid = false;
3230	}
3231
3232	for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
3233		if (!adev->ip_blocks[i].status.late_initialized)
3234			continue;
3235		if (adev->ip_blocks[i].version->funcs->late_fini)
3236			adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
3237		adev->ip_blocks[i].status.late_initialized = false;
3238	}
3239
3240	amdgpu_ras_fini(adev);
3241
 
 
 
 
3242	return 0;
3243}
3244
3245/**
3246 * amdgpu_device_delayed_init_work_handler - work handler for IB tests
3247 *
3248 * @work: work_struct.
3249 */
3250static void amdgpu_device_delayed_init_work_handler(struct work_struct *work)
3251{
3252	struct amdgpu_device *adev =
3253		container_of(work, struct amdgpu_device, delayed_init_work.work);
3254	int r;
3255
3256	r = amdgpu_ib_ring_tests(adev);
3257	if (r)
3258		DRM_ERROR("ib ring test failed (%d).\n", r);
3259}
3260
3261static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work)
3262{
3263	struct amdgpu_device *adev =
3264		container_of(work, struct amdgpu_device, gfx.gfx_off_delay_work.work);
3265
3266	WARN_ON_ONCE(adev->gfx.gfx_off_state);
3267	WARN_ON_ONCE(adev->gfx.gfx_off_req_count);
3268
3269	if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, true))
3270		adev->gfx.gfx_off_state = true;
3271}
3272
3273/**
3274 * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1)
3275 *
3276 * @adev: amdgpu_device pointer
3277 *
3278 * Main suspend function for hardware IPs.  The list of all the hardware
3279 * IPs that make up the asic is walked, clockgating is disabled and the
3280 * suspend callbacks are run.  suspend puts the hardware and software state
3281 * in each IP into a state suitable for suspend.
3282 * Returns 0 on success, negative error code on failure.
3283 */
3284static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev)
3285{
3286	int i, r;
3287
3288	amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
3289	amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
3290
3291	/*
3292	 * Per PMFW team's suggestion, driver needs to handle gfxoff
3293	 * and df cstate features disablement for gpu reset(e.g. Mode1Reset)
3294	 * scenario. Add the missing df cstate disablement here.
3295	 */
3296	if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_DISALLOW))
3297		dev_warn(adev->dev, "Failed to disallow df cstate");
3298
3299	for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
3300		if (!adev->ip_blocks[i].status.valid)
3301			continue;
3302
3303		/* displays are handled separately */
3304		if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_DCE)
3305			continue;
3306
3307		/* XXX handle errors */
3308		r = adev->ip_blocks[i].version->funcs->suspend(adev);
3309		/* XXX handle errors */
3310		if (r) {
3311			DRM_ERROR("suspend of IP block <%s> failed %d\n",
3312				  adev->ip_blocks[i].version->funcs->name, r);
3313			return r;
3314		}
3315
3316		adev->ip_blocks[i].status.hw = false;
3317	}
3318
3319	return 0;
3320}
3321
3322/**
3323 * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2)
3324 *
3325 * @adev: amdgpu_device pointer
3326 *
3327 * Main suspend function for hardware IPs.  The list of all the hardware
3328 * IPs that make up the asic is walked, clockgating is disabled and the
3329 * suspend callbacks are run.  suspend puts the hardware and software state
3330 * in each IP into a state suitable for suspend.
3331 * Returns 0 on success, negative error code on failure.
3332 */
3333static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev)
3334{
3335	int i, r;
3336
3337	if (adev->in_s0ix)
3338		amdgpu_dpm_gfx_state_change(adev, sGpuChangeState_D3Entry);
3339
3340	for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
3341		if (!adev->ip_blocks[i].status.valid)
3342			continue;
3343		/* displays are handled in phase1 */
3344		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE)
3345			continue;
3346		/* PSP lost connection when err_event_athub occurs */
3347		if (amdgpu_ras_intr_triggered() &&
3348		    adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
3349			adev->ip_blocks[i].status.hw = false;
3350			continue;
3351		}
3352
3353		/* skip unnecessary suspend if we do not initialize them yet */
3354		if (adev->gmc.xgmi.pending_reset &&
3355		    !(adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
3356		      adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC ||
3357		      adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
3358		      adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH)) {
3359			adev->ip_blocks[i].status.hw = false;
3360			continue;
3361		}
3362
3363		/* skip suspend of gfx/mes and psp for S0ix
3364		 * gfx is in gfxoff state, so on resume it will exit gfxoff just
3365		 * like at runtime. PSP is also part of the always on hardware
3366		 * so no need to suspend it.
3367		 */
3368		if (adev->in_s0ix &&
3369		    (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP ||
3370		     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX ||
3371		     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_MES))
3372			continue;
3373
3374		/* SDMA 5.x+ is part of GFX power domain so it's covered by GFXOFF */
3375		if (adev->in_s0ix &&
3376		    (amdgpu_ip_version(adev, SDMA0_HWIP, 0) >=
3377		     IP_VERSION(5, 0, 0)) &&
3378		    (adev->ip_blocks[i].version->type ==
3379		     AMD_IP_BLOCK_TYPE_SDMA))
3380			continue;
3381
3382		/* Once swPSP provides the IMU, RLC FW binaries to TOS during cold-boot.
3383		 * These are in TMR, hence are expected to be reused by PSP-TOS to reload
3384		 * from this location and RLC Autoload automatically also gets loaded
3385		 * from here based on PMFW -> PSP message during re-init sequence.
3386		 * Therefore, the psp suspend & resume should be skipped to avoid destroy
3387		 * the TMR and reload FWs again for IMU enabled APU ASICs.
3388		 */
3389		if (amdgpu_in_reset(adev) &&
3390		    (adev->flags & AMD_IS_APU) && adev->gfx.imu.funcs &&
3391		    adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)
3392			continue;
3393
3394		/* XXX handle errors */
3395		r = adev->ip_blocks[i].version->funcs->suspend(adev);
3396		/* XXX handle errors */
3397		if (r) {
3398			DRM_ERROR("suspend of IP block <%s> failed %d\n",
3399				  adev->ip_blocks[i].version->funcs->name, r);
3400		}
3401		adev->ip_blocks[i].status.hw = false;
3402		/* handle putting the SMC in the appropriate state */
3403		if (!amdgpu_sriov_vf(adev)) {
3404			if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
3405				r = amdgpu_dpm_set_mp1_state(adev, adev->mp1_state);
3406				if (r) {
3407					DRM_ERROR("SMC failed to set mp1 state %d, %d\n",
3408							adev->mp1_state, r);
3409					return r;
3410				}
3411			}
3412		}
3413	}
3414
3415	return 0;
3416}
3417
3418/**
3419 * amdgpu_device_ip_suspend - run suspend for hardware IPs
3420 *
3421 * @adev: amdgpu_device pointer
3422 *
3423 * Main suspend function for hardware IPs.  The list of all the hardware
3424 * IPs that make up the asic is walked, clockgating is disabled and the
3425 * suspend callbacks are run.  suspend puts the hardware and software state
3426 * in each IP into a state suitable for suspend.
3427 * Returns 0 on success, negative error code on failure.
3428 */
3429int amdgpu_device_ip_suspend(struct amdgpu_device *adev)
3430{
3431	int r;
3432
3433	if (amdgpu_sriov_vf(adev)) {
3434		amdgpu_virt_fini_data_exchange(adev);
3435		amdgpu_virt_request_full_gpu(adev, false);
3436	}
3437
3438	amdgpu_ttm_set_buffer_funcs_status(adev, false);
3439
3440	r = amdgpu_device_ip_suspend_phase1(adev);
3441	if (r)
3442		return r;
3443	r = amdgpu_device_ip_suspend_phase2(adev);
3444
3445	if (amdgpu_sriov_vf(adev))
3446		amdgpu_virt_release_full_gpu(adev, false);
3447
3448	return r;
3449}
3450
3451static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
3452{
3453	int i, r;
3454
3455	static enum amd_ip_block_type ip_order[] = {
3456		AMD_IP_BLOCK_TYPE_COMMON,
3457		AMD_IP_BLOCK_TYPE_GMC,
 
3458		AMD_IP_BLOCK_TYPE_PSP,
3459		AMD_IP_BLOCK_TYPE_IH,
3460	};
3461
3462	for (i = 0; i < adev->num_ip_blocks; i++) {
3463		int j;
3464		struct amdgpu_ip_block *block;
3465
3466		block = &adev->ip_blocks[i];
3467		block->status.hw = false;
3468
3469		for (j = 0; j < ARRAY_SIZE(ip_order); j++) {
3470
3471			if (block->version->type != ip_order[j] ||
3472				!block->status.valid)
3473				continue;
3474
3475			r = block->version->funcs->hw_init(adev);
3476			DRM_INFO("RE-INIT-early: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
3477			if (r)
3478				return r;
3479			block->status.hw = true;
3480		}
3481	}
3482
3483	return 0;
3484}
3485
3486static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
3487{
3488	int i, r;
3489
3490	static enum amd_ip_block_type ip_order[] = {
3491		AMD_IP_BLOCK_TYPE_SMC,
3492		AMD_IP_BLOCK_TYPE_DCE,
3493		AMD_IP_BLOCK_TYPE_GFX,
3494		AMD_IP_BLOCK_TYPE_SDMA,
3495		AMD_IP_BLOCK_TYPE_MES,
3496		AMD_IP_BLOCK_TYPE_UVD,
3497		AMD_IP_BLOCK_TYPE_VCE,
3498		AMD_IP_BLOCK_TYPE_VCN,
3499		AMD_IP_BLOCK_TYPE_JPEG
3500	};
3501
3502	for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
3503		int j;
3504		struct amdgpu_ip_block *block;
3505
3506		for (j = 0; j < adev->num_ip_blocks; j++) {
3507			block = &adev->ip_blocks[j];
3508
3509			if (block->version->type != ip_order[i] ||
3510				!block->status.valid ||
3511				block->status.hw)
3512				continue;
3513
3514			if (block->version->type == AMD_IP_BLOCK_TYPE_SMC)
3515				r = block->version->funcs->resume(adev);
3516			else
3517				r = block->version->funcs->hw_init(adev);
3518
3519			DRM_INFO("RE-INIT-late: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
3520			if (r)
3521				return r;
3522			block->status.hw = true;
3523		}
3524	}
3525
3526	return 0;
3527}
3528
3529/**
3530 * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs
3531 *
3532 * @adev: amdgpu_device pointer
3533 *
3534 * First resume function for hardware IPs.  The list of all the hardware
3535 * IPs that make up the asic is walked and the resume callbacks are run for
3536 * COMMON, GMC, and IH.  resume puts the hardware into a functional state
3537 * after a suspend and updates the software state as necessary.  This
3538 * function is also used for restoring the GPU after a GPU reset.
3539 * Returns 0 on success, negative error code on failure.
3540 */
3541static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)
3542{
3543	int i, r;
3544
3545	for (i = 0; i < adev->num_ip_blocks; i++) {
3546		if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
3547			continue;
3548		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
3549		    adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
3550		    adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH ||
3551		    (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP && amdgpu_sriov_vf(adev))) {
3552
3553			r = adev->ip_blocks[i].version->funcs->resume(adev);
3554			if (r) {
3555				DRM_ERROR("resume of IP block <%s> failed %d\n",
3556					  adev->ip_blocks[i].version->funcs->name, r);
3557				return r;
3558			}
3559			adev->ip_blocks[i].status.hw = true;
3560		}
3561	}
3562
3563	return 0;
3564}
3565
3566/**
3567 * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs
3568 *
3569 * @adev: amdgpu_device pointer
3570 *
3571 * First resume function for hardware IPs.  The list of all the hardware
3572 * IPs that make up the asic is walked and the resume callbacks are run for
3573 * all blocks except COMMON, GMC, and IH.  resume puts the hardware into a
3574 * functional state after a suspend and updates the software state as
3575 * necessary.  This function is also used for restoring the GPU after a GPU
3576 * reset.
3577 * Returns 0 on success, negative error code on failure.
3578 */
3579static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)
3580{
3581	int i, r;
3582
3583	for (i = 0; i < adev->num_ip_blocks; i++) {
3584		if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
3585			continue;
3586		if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
3587		    adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
3588		    adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH ||
3589		    adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)
3590			continue;
3591		r = adev->ip_blocks[i].version->funcs->resume(adev);
3592		if (r) {
3593			DRM_ERROR("resume of IP block <%s> failed %d\n",
3594				  adev->ip_blocks[i].version->funcs->name, r);
3595			return r;
3596		}
3597		adev->ip_blocks[i].status.hw = true;
3598	}
3599
3600	return 0;
3601}
3602
3603/**
3604 * amdgpu_device_ip_resume - run resume for hardware IPs
3605 *
3606 * @adev: amdgpu_device pointer
3607 *
3608 * Main resume function for hardware IPs.  The hardware IPs
3609 * are split into two resume functions because they are
3610 * also used in recovering from a GPU reset and some additional
3611 * steps need to be take between them.  In this case (S3/S4) they are
3612 * run sequentially.
3613 * Returns 0 on success, negative error code on failure.
3614 */
3615static int amdgpu_device_ip_resume(struct amdgpu_device *adev)
3616{
3617	int r;
3618
 
 
 
 
3619	r = amdgpu_device_ip_resume_phase1(adev);
3620	if (r)
3621		return r;
3622
3623	r = amdgpu_device_fw_loading(adev);
3624	if (r)
3625		return r;
3626
3627	r = amdgpu_device_ip_resume_phase2(adev);
3628
3629	if (adev->mman.buffer_funcs_ring->sched.ready)
3630		amdgpu_ttm_set_buffer_funcs_status(adev, true);
3631
3632	return r;
3633}
3634
3635/**
3636 * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV
3637 *
3638 * @adev: amdgpu_device pointer
3639 *
3640 * Query the VBIOS data tables to determine if the board supports SR-IOV.
3641 */
3642static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
3643{
3644	if (amdgpu_sriov_vf(adev)) {
3645		if (adev->is_atom_fw) {
3646			if (amdgpu_atomfirmware_gpu_virtualization_supported(adev))
3647				adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
3648		} else {
3649			if (amdgpu_atombios_has_gpu_virtualization_table(adev))
3650				adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
3651		}
3652
3653		if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
3654			amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
3655	}
3656}
3657
3658/**
3659 * amdgpu_device_asic_has_dc_support - determine if DC supports the asic
3660 *
3661 * @asic_type: AMD asic type
3662 *
3663 * Check if there is DC (new modesetting infrastructre) support for an asic.
3664 * returns true if DC has support, false if not.
3665 */
3666bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
3667{
3668	switch (asic_type) {
3669#ifdef CONFIG_DRM_AMDGPU_SI
3670	case CHIP_HAINAN:
3671#endif
3672	case CHIP_TOPAZ:
3673		/* chips with no display hardware */
3674		return false;
3675#if defined(CONFIG_DRM_AMD_DC)
 
3676	case CHIP_TAHITI:
3677	case CHIP_PITCAIRN:
3678	case CHIP_VERDE:
3679	case CHIP_OLAND:
3680		/*
3681		 * We have systems in the wild with these ASICs that require
3682		 * LVDS and VGA support which is not supported with DC.
3683		 *
3684		 * Fallback to the non-DC driver here by default so as not to
3685		 * cause regressions.
3686		 */
3687#if defined(CONFIG_DRM_AMD_DC_SI)
3688		return amdgpu_dc > 0;
3689#else
3690		return false;
3691#endif
3692	case CHIP_BONAIRE:
3693	case CHIP_KAVERI:
3694	case CHIP_KABINI:
3695	case CHIP_MULLINS:
3696		/*
3697		 * We have systems in the wild with these ASICs that require
3698		 * VGA support which is not supported with DC.
3699		 *
3700		 * Fallback to the non-DC driver here by default so as not to
3701		 * cause regressions.
3702		 */
3703		return amdgpu_dc > 0;
3704	default:
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3705		return amdgpu_dc != 0;
3706#else
3707	default:
3708		if (amdgpu_dc > 0)
3709			DRM_INFO_ONCE("Display Core has been requested via kernel parameter but isn't supported by ASIC, ignoring\n");
 
3710		return false;
3711#endif
3712	}
3713}
3714
3715/**
3716 * amdgpu_device_has_dc_support - check if dc is supported
3717 *
3718 * @adev: amdgpu_device pointer
3719 *
3720 * Returns true for supported, false for not supported
3721 */
3722bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
3723{
3724	if (adev->enable_virtual_display ||
 
3725	    (adev->harvest_ip_mask & AMD_HARVEST_IP_DMU_MASK))
3726		return false;
3727
3728	return amdgpu_device_asic_has_dc_support(adev->asic_type);
3729}
3730
3731static void amdgpu_device_xgmi_reset_func(struct work_struct *__work)
3732{
3733	struct amdgpu_device *adev =
3734		container_of(__work, struct amdgpu_device, xgmi_reset_work);
3735	struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev);
3736
3737	/* It's a bug to not have a hive within this function */
3738	if (WARN_ON(!hive))
3739		return;
3740
3741	/*
3742	 * Use task barrier to synchronize all xgmi reset works across the
3743	 * hive. task_barrier_enter and task_barrier_exit will block
3744	 * until all the threads running the xgmi reset works reach
3745	 * those points. task_barrier_full will do both blocks.
3746	 */
3747	if (amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) {
3748
3749		task_barrier_enter(&hive->tb);
3750		adev->asic_reset_res = amdgpu_device_baco_enter(adev_to_drm(adev));
3751
3752		if (adev->asic_reset_res)
3753			goto fail;
3754
3755		task_barrier_exit(&hive->tb);
3756		adev->asic_reset_res = amdgpu_device_baco_exit(adev_to_drm(adev));
3757
3758		if (adev->asic_reset_res)
3759			goto fail;
3760
3761		amdgpu_ras_reset_error_count(adev, AMDGPU_RAS_BLOCK__MMHUB);
 
 
3762	} else {
3763
3764		task_barrier_full(&hive->tb);
3765		adev->asic_reset_res =  amdgpu_asic_reset(adev);
3766	}
3767
3768fail:
3769	if (adev->asic_reset_res)
3770		DRM_WARN("ASIC reset failed with error, %d for drm dev, %s",
3771			 adev->asic_reset_res, adev_to_drm(adev)->unique);
3772	amdgpu_put_xgmi_hive(hive);
3773}
3774
3775static int amdgpu_device_get_job_timeout_settings(struct amdgpu_device *adev)
3776{
3777	char *input = amdgpu_lockup_timeout;
3778	char *timeout_setting = NULL;
3779	int index = 0;
3780	long timeout;
3781	int ret = 0;
3782
3783	/*
3784	 * By default timeout for non compute jobs is 10000
3785	 * and 60000 for compute jobs.
3786	 * In SR-IOV or passthrough mode, timeout for compute
3787	 * jobs are 60000 by default.
3788	 */
3789	adev->gfx_timeout = msecs_to_jiffies(10000);
3790	adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout;
3791	if (amdgpu_sriov_vf(adev))
3792		adev->compute_timeout = amdgpu_sriov_is_pp_one_vf(adev) ?
3793					msecs_to_jiffies(60000) : msecs_to_jiffies(10000);
3794	else
3795		adev->compute_timeout =  msecs_to_jiffies(60000);
3796
3797	if (strnlen(input, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) {
3798		while ((timeout_setting = strsep(&input, ",")) &&
3799				strnlen(timeout_setting, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) {
3800			ret = kstrtol(timeout_setting, 0, &timeout);
3801			if (ret)
3802				return ret;
3803
3804			if (timeout == 0) {
3805				index++;
3806				continue;
3807			} else if (timeout < 0) {
3808				timeout = MAX_SCHEDULE_TIMEOUT;
3809				dev_warn(adev->dev, "lockup timeout disabled");
3810				add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
3811			} else {
3812				timeout = msecs_to_jiffies(timeout);
3813			}
3814
3815			switch (index++) {
3816			case 0:
3817				adev->gfx_timeout = timeout;
3818				break;
3819			case 1:
3820				adev->compute_timeout = timeout;
3821				break;
3822			case 2:
3823				adev->sdma_timeout = timeout;
3824				break;
3825			case 3:
3826				adev->video_timeout = timeout;
3827				break;
3828			default:
3829				break;
3830			}
3831		}
3832		/*
3833		 * There is only one value specified and
3834		 * it should apply to all non-compute jobs.
3835		 */
3836		if (index == 1) {
3837			adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout;
3838			if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev))
3839				adev->compute_timeout = adev->gfx_timeout;
3840		}
3841	}
3842
3843	return ret;
3844}
3845
3846/**
3847 * amdgpu_device_check_iommu_direct_map - check if RAM direct mapped to GPU
3848 *
3849 * @adev: amdgpu_device pointer
3850 *
3851 * RAM direct mapped to GPU if IOMMU is not enabled or is pass through mode
3852 */
3853static void amdgpu_device_check_iommu_direct_map(struct amdgpu_device *adev)
3854{
3855	struct iommu_domain *domain;
3856
3857	domain = iommu_get_domain_for_dev(adev->dev);
3858	if (!domain || domain->type == IOMMU_DOMAIN_IDENTITY)
3859		adev->ram_is_direct_mapped = true;
3860}
3861
3862static const struct attribute *amdgpu_dev_attributes[] = {
 
 
 
3863	&dev_attr_pcie_replay_count.attr,
3864	NULL
3865};
3866
3867static void amdgpu_device_set_mcbp(struct amdgpu_device *adev)
3868{
3869	if (amdgpu_mcbp == 1)
3870		adev->gfx.mcbp = true;
3871	else if (amdgpu_mcbp == 0)
3872		adev->gfx.mcbp = false;
3873
3874	if (amdgpu_sriov_vf(adev))
3875		adev->gfx.mcbp = true;
3876
3877	if (adev->gfx.mcbp)
3878		DRM_INFO("MCBP is enabled\n");
3879}
3880
3881/**
3882 * amdgpu_device_init - initialize the driver
3883 *
3884 * @adev: amdgpu_device pointer
3885 * @flags: driver flags
3886 *
3887 * Initializes the driver info and hw (all asics).
3888 * Returns 0 for success or an error on failure.
3889 * Called at driver startup.
3890 */
3891int amdgpu_device_init(struct amdgpu_device *adev,
3892		       uint32_t flags)
3893{
3894	struct drm_device *ddev = adev_to_drm(adev);
3895	struct pci_dev *pdev = adev->pdev;
3896	int r, i;
3897	bool px = false;
3898	u32 max_MBps;
3899	int tmp;
3900
3901	adev->shutdown = false;
3902	adev->flags = flags;
3903
3904	if (amdgpu_force_asic_type >= 0 && amdgpu_force_asic_type < CHIP_LAST)
3905		adev->asic_type = amdgpu_force_asic_type;
3906	else
3907		adev->asic_type = flags & AMD_ASIC_MASK;
3908
3909	adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
3910	if (amdgpu_emu_mode == 1)
3911		adev->usec_timeout *= 10;
3912	adev->gmc.gart_size = 512 * 1024 * 1024;
3913	adev->accel_working = false;
3914	adev->num_rings = 0;
3915	RCU_INIT_POINTER(adev->gang_submit, dma_fence_get_stub());
3916	adev->mman.buffer_funcs = NULL;
3917	adev->mman.buffer_funcs_ring = NULL;
3918	adev->vm_manager.vm_pte_funcs = NULL;
3919	adev->vm_manager.vm_pte_num_scheds = 0;
3920	adev->gmc.gmc_funcs = NULL;
3921	adev->harvest_ip_mask = 0x0;
3922	adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
3923	bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
3924
3925	adev->smc_rreg = &amdgpu_invalid_rreg;
3926	adev->smc_wreg = &amdgpu_invalid_wreg;
3927	adev->pcie_rreg = &amdgpu_invalid_rreg;
3928	adev->pcie_wreg = &amdgpu_invalid_wreg;
3929	adev->pcie_rreg_ext = &amdgpu_invalid_rreg_ext;
3930	adev->pcie_wreg_ext = &amdgpu_invalid_wreg_ext;
3931	adev->pciep_rreg = &amdgpu_invalid_rreg;
3932	adev->pciep_wreg = &amdgpu_invalid_wreg;
3933	adev->pcie_rreg64 = &amdgpu_invalid_rreg64;
3934	adev->pcie_wreg64 = &amdgpu_invalid_wreg64;
3935	adev->pcie_rreg64_ext = &amdgpu_invalid_rreg64_ext;
3936	adev->pcie_wreg64_ext = &amdgpu_invalid_wreg64_ext;
3937	adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
3938	adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
3939	adev->didt_rreg = &amdgpu_invalid_rreg;
3940	adev->didt_wreg = &amdgpu_invalid_wreg;
3941	adev->gc_cac_rreg = &amdgpu_invalid_rreg;
3942	adev->gc_cac_wreg = &amdgpu_invalid_wreg;
3943	adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
3944	adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
3945
3946	DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
3947		 amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
3948		 pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
3949
3950	/* mutex initialization are all done here so we
3951	 * can recall function without having locking issues
3952	 */
3953	mutex_init(&adev->firmware.mutex);
3954	mutex_init(&adev->pm.mutex);
3955	mutex_init(&adev->gfx.gpu_clock_mutex);
3956	mutex_init(&adev->srbm_mutex);
3957	mutex_init(&adev->gfx.pipe_reserve_mutex);
3958	mutex_init(&adev->gfx.gfx_off_mutex);
3959	mutex_init(&adev->gfx.partition_mutex);
3960	mutex_init(&adev->grbm_idx_mutex);
3961	mutex_init(&adev->mn_lock);
3962	mutex_init(&adev->virt.vf_errors.lock);
3963	hash_init(adev->mn_hash);
 
 
3964	mutex_init(&adev->psp.mutex);
3965	mutex_init(&adev->notifier_lock);
3966	mutex_init(&adev->pm.stable_pstate_ctx_lock);
3967	mutex_init(&adev->benchmark_mutex);
3968
3969	amdgpu_device_init_apu_flags(adev);
 
 
3970
3971	r = amdgpu_device_check_arguments(adev);
3972	if (r)
3973		return r;
3974
3975	spin_lock_init(&adev->mmio_idx_lock);
3976	spin_lock_init(&adev->smc_idx_lock);
3977	spin_lock_init(&adev->pcie_idx_lock);
3978	spin_lock_init(&adev->uvd_ctx_idx_lock);
3979	spin_lock_init(&adev->didt_idx_lock);
3980	spin_lock_init(&adev->gc_cac_idx_lock);
3981	spin_lock_init(&adev->se_cac_idx_lock);
3982	spin_lock_init(&adev->audio_endpt_idx_lock);
3983	spin_lock_init(&adev->mm_stats.lock);
3984
3985	INIT_LIST_HEAD(&adev->shadow_list);
3986	mutex_init(&adev->shadow_list_lock);
3987
3988	INIT_LIST_HEAD(&adev->reset_list);
3989
3990	INIT_LIST_HEAD(&adev->ras_list);
3991
3992	INIT_LIST_HEAD(&adev->pm.od_kobj_list);
3993
3994	INIT_DELAYED_WORK(&adev->delayed_init_work,
3995			  amdgpu_device_delayed_init_work_handler);
3996	INIT_DELAYED_WORK(&adev->gfx.gfx_off_delay_work,
3997			  amdgpu_device_delay_enable_gfx_off);
3998
3999	INIT_WORK(&adev->xgmi_reset_work, amdgpu_device_xgmi_reset_func);
4000
4001	adev->gfx.gfx_off_req_count = 1;
4002	adev->gfx.gfx_off_residency = 0;
4003	adev->gfx.gfx_off_entrycount = 0;
4004	adev->pm.ac_power = power_supply_is_system_supplied() > 0;
4005
4006	atomic_set(&adev->throttling_logging_enabled, 1);
4007	/*
4008	 * If throttling continues, logging will be performed every minute
4009	 * to avoid log flooding. "-1" is subtracted since the thermal
4010	 * throttling interrupt comes every second. Thus, the total logging
4011	 * interval is 59 seconds(retelimited printk interval) + 1(waiting
4012	 * for throttling interrupt) = 60 seconds.
4013	 */
4014	ratelimit_state_init(&adev->throttling_logging_rs, (60 - 1) * HZ, 1);
4015	ratelimit_set_flags(&adev->throttling_logging_rs, RATELIMIT_MSG_ON_RELEASE);
4016
4017	/* Registers mapping */
4018	/* TODO: block userspace mapping of io register */
4019	if (adev->asic_type >= CHIP_BONAIRE) {
4020		adev->rmmio_base = pci_resource_start(adev->pdev, 5);
4021		adev->rmmio_size = pci_resource_len(adev->pdev, 5);
4022	} else {
4023		adev->rmmio_base = pci_resource_start(adev->pdev, 2);
4024		adev->rmmio_size = pci_resource_len(adev->pdev, 2);
4025	}
4026
4027	for (i = 0; i < AMD_IP_BLOCK_TYPE_NUM; i++)
4028		atomic_set(&adev->pm.pwr_state[i], POWER_STATE_UNKNOWN);
4029
4030	adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
4031	if (!adev->rmmio)
4032		return -ENOMEM;
4033
4034	DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
4035	DRM_INFO("register mmio size: %u\n", (unsigned int)adev->rmmio_size);
4036
4037	/*
4038	 * Reset domain needs to be present early, before XGMI hive discovered
4039	 * (if any) and intitialized to use reset sem and in_gpu reset flag
4040	 * early on during init and before calling to RREG32.
4041	 */
4042	adev->reset_domain = amdgpu_reset_create_reset_domain(SINGLE_DEVICE, "amdgpu-reset-dev");
4043	if (!adev->reset_domain)
4044		return -ENOMEM;
 
 
 
 
 
 
 
 
 
 
4045
4046	/* detect hw virtualization here */
4047	amdgpu_detect_virtualization(adev);
4048
4049	amdgpu_device_get_pcie_info(adev);
4050
4051	r = amdgpu_device_get_job_timeout_settings(adev);
4052	if (r) {
4053		dev_err(adev->dev, "invalid lockup_timeout parameter syntax\n");
4054		return r;
4055	}
4056
4057	amdgpu_device_set_mcbp(adev);
4058
4059	/* early init functions */
4060	r = amdgpu_device_ip_early_init(adev);
4061	if (r)
4062		return r;
4063
4064	/* Get rid of things like offb */
4065	r = drm_aperture_remove_conflicting_pci_framebuffers(adev->pdev, &amdgpu_kms_driver);
4066	if (r)
4067		return r;
4068
4069	/* Enable TMZ based on IP_VERSION */
4070	amdgpu_gmc_tmz_set(adev);
4071
4072	amdgpu_gmc_noretry_set(adev);
4073	/* Need to get xgmi info early to decide the reset behavior*/
4074	if (adev->gmc.xgmi.supported) {
4075		r = adev->gfxhub.funcs->get_xgmi_info(adev);
4076		if (r)
4077			return r;
4078	}
4079
4080	/* enable PCIE atomic ops */
4081	if (amdgpu_sriov_vf(adev)) {
4082		if (adev->virt.fw_reserve.p_pf2vf)
4083			adev->have_atomics_support = ((struct amd_sriov_msg_pf2vf_info *)
4084						      adev->virt.fw_reserve.p_pf2vf)->pcie_atomic_ops_support_flags ==
4085				(PCI_EXP_DEVCAP2_ATOMIC_COMP32 | PCI_EXP_DEVCAP2_ATOMIC_COMP64);
4086	/* APUs w/ gfx9 onwards doesn't reply on PCIe atomics, rather it is a
4087	 * internal path natively support atomics, set have_atomics_support to true.
4088	 */
4089	} else if ((adev->flags & AMD_IS_APU) &&
4090		   (amdgpu_ip_version(adev, GC_HWIP, 0) >
4091		    IP_VERSION(9, 0, 0))) {
4092		adev->have_atomics_support = true;
4093	} else {
4094		adev->have_atomics_support =
4095			!pci_enable_atomic_ops_to_root(adev->pdev,
4096					  PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
4097					  PCI_EXP_DEVCAP2_ATOMIC_COMP64);
4098	}
4099
4100	if (!adev->have_atomics_support)
4101		dev_info(adev->dev, "PCIE atomic ops is not supported\n");
4102
4103	/* doorbell bar mapping and doorbell index init*/
4104	amdgpu_doorbell_init(adev);
4105
4106	if (amdgpu_emu_mode == 1) {
4107		/* post the asic on emulation mode */
4108		emu_soc_asic_init(adev);
4109		goto fence_driver_init;
4110	}
4111
4112	amdgpu_reset_init(adev);
4113
4114	/* detect if we are with an SRIOV vbios */
4115	if (adev->bios)
4116		amdgpu_device_detect_sriov_bios(adev);
4117
4118	/* check if we need to reset the asic
4119	 *  E.g., driver was not cleanly unloaded previously, etc.
4120	 */
4121	if (!amdgpu_sriov_vf(adev) && amdgpu_asic_need_reset_on_init(adev)) {
4122		if (adev->gmc.xgmi.num_physical_nodes) {
4123			dev_info(adev->dev, "Pending hive reset.\n");
4124			adev->gmc.xgmi.pending_reset = true;
4125			/* Only need to init necessary block for SMU to handle the reset */
4126			for (i = 0; i < adev->num_ip_blocks; i++) {
4127				if (!adev->ip_blocks[i].status.valid)
4128					continue;
4129				if (!(adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
4130				      adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
4131				      adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH ||
4132				      adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC)) {
4133					DRM_DEBUG("IP %s disabled for hw_init.\n",
4134						adev->ip_blocks[i].version->funcs->name);
4135					adev->ip_blocks[i].status.hw = true;
4136				}
4137			}
4138		} else if (amdgpu_ip_version(adev, MP1_HWIP, 0) == IP_VERSION(13, 0, 10) &&
4139				   !amdgpu_device_has_display_hardware(adev)) {
4140					r = psp_gpu_reset(adev);
4141		} else {
4142				tmp = amdgpu_reset_method;
4143				/* It should do a default reset when loading or reloading the driver,
4144				 * regardless of the module parameter reset_method.
4145				 */
4146				amdgpu_reset_method = AMD_RESET_METHOD_NONE;
4147				r = amdgpu_asic_reset(adev);
4148				amdgpu_reset_method = tmp;
4149		}
4150
4151		if (r) {
4152		  dev_err(adev->dev, "asic reset on init failed\n");
4153		  goto failed;
4154		}
4155	}
4156
 
 
4157	/* Post card if necessary */
4158	if (amdgpu_device_need_post(adev)) {
4159		if (!adev->bios) {
4160			dev_err(adev->dev, "no vBIOS found\n");
4161			r = -EINVAL;
4162			goto failed;
4163		}
4164		DRM_INFO("GPU posting now...\n");
4165		r = amdgpu_device_asic_init(adev);
4166		if (r) {
4167			dev_err(adev->dev, "gpu post error!\n");
4168			goto failed;
4169		}
4170	}
4171
4172	if (adev->bios) {
4173		if (adev->is_atom_fw) {
4174			/* Initialize clocks */
4175			r = amdgpu_atomfirmware_get_clock_info(adev);
4176			if (r) {
4177				dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
4178				amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
4179				goto failed;
4180			}
4181		} else {
4182			/* Initialize clocks */
4183			r = amdgpu_atombios_get_clock_info(adev);
4184			if (r) {
4185				dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
4186				amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
4187				goto failed;
4188			}
4189			/* init i2c buses */
4190			if (!amdgpu_device_has_dc_support(adev))
4191				amdgpu_atombios_i2c_init(adev);
4192		}
 
 
 
 
 
 
 
 
 
 
 
4193	}
4194
4195fence_driver_init:
4196	/* Fence driver */
4197	r = amdgpu_fence_driver_sw_init(adev);
4198	if (r) {
4199		dev_err(adev->dev, "amdgpu_fence_driver_sw_init failed\n");
4200		amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
4201		goto failed;
4202	}
4203
4204	/* init the mode config */
4205	drm_mode_config_init(adev_to_drm(adev));
4206
4207	r = amdgpu_device_ip_init(adev);
4208	if (r) {
 
 
 
 
 
 
 
 
 
 
 
 
4209		dev_err(adev->dev, "amdgpu_device_ip_init failed\n");
4210		amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
4211		goto release_ras_con;
4212	}
4213
4214	amdgpu_fence_driver_hw_init(adev);
4215
4216	dev_info(adev->dev,
4217		"SE %d, SH per SE %d, CU per SH %d, active_cu_number %d\n",
4218			adev->gfx.config.max_shader_engines,
4219			adev->gfx.config.max_sh_per_se,
4220			adev->gfx.config.max_cu_per_sh,
4221			adev->gfx.cu_info.number);
4222
4223	adev->accel_working = true;
4224
4225	amdgpu_vm_check_compute_bug(adev);
4226
4227	/* Initialize the buffer migration limit. */
4228	if (amdgpu_moverate >= 0)
4229		max_MBps = amdgpu_moverate;
4230	else
4231		max_MBps = 8; /* Allow 8 MB/s. */
4232	/* Get a log2 for easy divisions. */
4233	adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
4234
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4235	/*
4236	 * Register gpu instance before amdgpu_device_enable_mgpu_fan_boost.
4237	 * Otherwise the mgpu fan boost feature will be skipped due to the
4238	 * gpu instance is counted less.
4239	 */
4240	amdgpu_register_gpu_instance(adev);
4241
4242	/* enable clockgating, etc. after ib tests, etc. since some blocks require
4243	 * explicit gating rather than handling it automatically.
4244	 */
4245	if (!adev->gmc.xgmi.pending_reset) {
4246		r = amdgpu_device_ip_late_init(adev);
4247		if (r) {
4248			dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n");
4249			amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
4250			goto release_ras_con;
4251		}
4252		/* must succeed. */
4253		amdgpu_ras_resume(adev);
4254		queue_delayed_work(system_wq, &adev->delayed_init_work,
4255				   msecs_to_jiffies(AMDGPU_RESUME_MS));
4256	}
4257
4258	if (amdgpu_sriov_vf(adev)) {
4259		amdgpu_virt_release_full_gpu(adev, true);
4260		flush_delayed_work(&adev->delayed_init_work);
4261	}
4262
4263	/*
4264	 * Place those sysfs registering after `late_init`. As some of those
4265	 * operations performed in `late_init` might affect the sysfs
4266	 * interfaces creating.
4267	 */
4268	r = amdgpu_atombios_sysfs_init(adev);
4269	if (r)
4270		drm_err(&adev->ddev,
4271			"registering atombios sysfs failed (%d).\n", r);
4272
4273	r = amdgpu_pm_sysfs_init(adev);
4274	if (r)
4275		DRM_ERROR("registering pm sysfs failed (%d).\n", r);
4276
4277	r = amdgpu_ucode_sysfs_init(adev);
4278	if (r) {
4279		adev->ucode_sysfs_en = false;
4280		DRM_ERROR("Creating firmware sysfs failed (%d).\n", r);
4281	} else
4282		adev->ucode_sysfs_en = true;
4283
4284	r = sysfs_create_files(&adev->dev->kobj, amdgpu_dev_attributes);
4285	if (r)
4286		dev_err(adev->dev, "Could not create amdgpu device attr\n");
4287
4288	r = devm_device_add_group(adev->dev, &amdgpu_board_attrs_group);
4289	if (r)
4290		dev_err(adev->dev,
4291			"Could not create amdgpu board attributes\n");
4292
4293	amdgpu_fru_sysfs_init(adev);
4294	amdgpu_reg_state_sysfs_init(adev);
4295
4296	if (IS_ENABLED(CONFIG_PERF_EVENTS))
4297		r = amdgpu_pmu_init(adev);
4298	if (r)
4299		dev_err(adev->dev, "amdgpu_pmu_init failed\n");
4300
4301	/* Have stored pci confspace at hand for restore in sudden PCI error */
4302	if (amdgpu_device_cache_pci_state(adev->pdev))
4303		pci_restore_state(pdev);
4304
4305	/* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
4306	/* this will fail for cards that aren't VGA class devices, just
4307	 * ignore it
4308	 */
4309	if ((adev->pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA)
4310		vga_client_register(adev->pdev, amdgpu_device_vga_set_decode);
4311
4312	px = amdgpu_device_supports_px(ddev);
4313
4314	if (px || (!dev_is_removable(&adev->pdev->dev) &&
4315				apple_gmux_detect(NULL, NULL)))
4316		vga_switcheroo_register_client(adev->pdev,
4317					       &amdgpu_switcheroo_ops, px);
4318
4319	if (px)
4320		vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
 
4321
4322	if (adev->gmc.xgmi.pending_reset)
4323		queue_delayed_work(system_wq, &mgpu_info.delayed_reset_work,
4324				   msecs_to_jiffies(AMDGPU_RESUME_MS));
4325
4326	amdgpu_device_check_iommu_direct_map(adev);
4327
4328	return 0;
4329
4330release_ras_con:
4331	if (amdgpu_sriov_vf(adev))
4332		amdgpu_virt_release_full_gpu(adev, true);
4333
4334	/* failed in exclusive mode due to timeout */
4335	if (amdgpu_sriov_vf(adev) &&
4336		!amdgpu_sriov_runtime(adev) &&
4337		amdgpu_virt_mmio_blocked(adev) &&
4338		!amdgpu_virt_wait_reset(adev)) {
4339		dev_err(adev->dev, "VF exclusive mode timeout\n");
4340		/* Don't send request since VF is inactive. */
4341		adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
4342		adev->virt.ops = NULL;
4343		r = -EAGAIN;
4344	}
4345	amdgpu_release_ras_context(adev);
4346
4347failed:
4348	amdgpu_vf_error_trans_all(adev);
4349
4350	return r;
4351}
4352
4353static void amdgpu_device_unmap_mmio(struct amdgpu_device *adev)
4354{
4355
4356	/* Clear all CPU mappings pointing to this device */
4357	unmap_mapping_range(adev->ddev.anon_inode->i_mapping, 0, 0, 1);
4358
4359	/* Unmap all mapped bars - Doorbell, registers and VRAM */
4360	amdgpu_doorbell_fini(adev);
4361
4362	iounmap(adev->rmmio);
4363	adev->rmmio = NULL;
4364	if (adev->mman.aper_base_kaddr)
4365		iounmap(adev->mman.aper_base_kaddr);
4366	adev->mman.aper_base_kaddr = NULL;
4367
4368	/* Memory manager related */
4369	if (!adev->gmc.xgmi.connected_to_cpu && !adev->gmc.is_app_apu) {
4370		arch_phys_wc_del(adev->gmc.vram_mtrr);
4371		arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
4372	}
4373}
4374
4375/**
4376 * amdgpu_device_fini_hw - tear down the driver
4377 *
4378 * @adev: amdgpu_device pointer
4379 *
4380 * Tear down the driver info (all asics).
4381 * Called at driver shutdown.
4382 */
4383void amdgpu_device_fini_hw(struct amdgpu_device *adev)
4384{
4385	dev_info(adev->dev, "amdgpu: finishing device.\n");
4386	flush_delayed_work(&adev->delayed_init_work);
 
4387	adev->shutdown = true;
4388
4389	/* make sure IB test finished before entering exclusive mode
4390	 * to avoid preemption on IB test
4391	 */
4392	if (amdgpu_sriov_vf(adev)) {
4393		amdgpu_virt_request_full_gpu(adev, false);
4394		amdgpu_virt_fini_data_exchange(adev);
4395	}
4396
4397	/* disable all interrupts */
4398	amdgpu_irq_disable_all(adev);
4399	if (adev->mode_info.mode_config_initialized) {
4400		if (!drm_drv_uses_atomic_modeset(adev_to_drm(adev)))
4401			drm_helper_force_disable_all(adev_to_drm(adev));
4402		else
4403			drm_atomic_helper_shutdown(adev_to_drm(adev));
4404	}
4405	amdgpu_fence_driver_hw_fini(adev);
4406
4407	if (adev->mman.initialized)
4408		drain_workqueue(adev->mman.bdev.wq);
4409
4410	if (adev->pm.sysfs_initialized)
4411		amdgpu_pm_sysfs_fini(adev);
4412	if (adev->ucode_sysfs_en)
4413		amdgpu_ucode_sysfs_fini(adev);
4414	sysfs_remove_files(&adev->dev->kobj, amdgpu_dev_attributes);
4415	amdgpu_fru_sysfs_fini(adev);
4416
4417	amdgpu_reg_state_sysfs_fini(adev);
4418
4419	/* disable ras feature must before hw fini */
4420	amdgpu_ras_pre_fini(adev);
4421
4422	amdgpu_ttm_set_buffer_funcs_status(adev, false);
4423
4424	amdgpu_device_ip_fini_early(adev);
4425
4426	amdgpu_irq_fini_hw(adev);
4427
4428	if (adev->mman.initialized)
4429		ttm_device_clear_dma_mappings(&adev->mman.bdev);
4430
4431	amdgpu_gart_dummy_page_fini(adev);
4432
4433	if (drm_dev_is_unplugged(adev_to_drm(adev)))
4434		amdgpu_device_unmap_mmio(adev);
4435
4436}
4437
4438void amdgpu_device_fini_sw(struct amdgpu_device *adev)
4439{
4440	int idx;
4441	bool px;
4442
4443	amdgpu_fence_driver_sw_fini(adev);
4444	amdgpu_device_ip_fini(adev);
4445	amdgpu_ucode_release(&adev->firmware.gpu_info_fw);
 
 
4446	adev->accel_working = false;
4447	dma_fence_put(rcu_dereference_protected(adev->gang_submit, true));
4448
4449	amdgpu_reset_fini(adev);
4450
4451	/* free i2c buses */
4452	if (!amdgpu_device_has_dc_support(adev))
4453		amdgpu_i2c_fini(adev);
4454
4455	if (amdgpu_emu_mode != 1)
4456		amdgpu_atombios_fini(adev);
4457
4458	kfree(adev->bios);
4459	adev->bios = NULL;
4460
4461	kfree(adev->fru_info);
4462	adev->fru_info = NULL;
4463
4464	px = amdgpu_device_supports_px(adev_to_drm(adev));
4465
4466	if (px || (!dev_is_removable(&adev->pdev->dev) &&
4467				apple_gmux_detect(NULL, NULL)))
4468		vga_switcheroo_unregister_client(adev->pdev);
4469
4470	if (px)
4471		vga_switcheroo_fini_domain_pm_ops(adev->dev);
4472
4473	if ((adev->pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA)
4474		vga_client_unregister(adev->pdev);
4475
4476	if (drm_dev_enter(adev_to_drm(adev), &idx)) {
4477
4478		iounmap(adev->rmmio);
4479		adev->rmmio = NULL;
4480		amdgpu_doorbell_fini(adev);
4481		drm_dev_exit(idx);
4482	}
 
 
4483
4484	if (IS_ENABLED(CONFIG_PERF_EVENTS))
4485		amdgpu_pmu_fini(adev);
4486	if (adev->mman.discovery_bin)
4487		amdgpu_discovery_fini(adev);
4488
4489	amdgpu_reset_put_reset_domain(adev->reset_domain);
4490	adev->reset_domain = NULL;
4491
4492	kfree(adev->pci_state);
4493
4494}
4495
4496/**
4497 * amdgpu_device_evict_resources - evict device resources
4498 * @adev: amdgpu device object
4499 *
4500 * Evicts all ttm device resources(vram BOs, gart table) from the lru list
4501 * of the vram memory type. Mainly used for evicting device resources
4502 * at suspend time.
4503 *
4504 */
4505static int amdgpu_device_evict_resources(struct amdgpu_device *adev)
4506{
4507	int ret;
4508
4509	/* No need to evict vram on APUs for suspend to ram or s2idle */
4510	if ((adev->in_s3 || adev->in_s0ix) && (adev->flags & AMD_IS_APU))
4511		return 0;
4512
4513	ret = amdgpu_ttm_evict_resources(adev, TTM_PL_VRAM);
4514	if (ret)
4515		DRM_WARN("evicting device resources failed\n");
4516	return ret;
4517}
4518
4519/*
4520 * Suspend & resume.
4521 */
4522/**
4523 * amdgpu_device_prepare - prepare for device suspend
4524 *
4525 * @dev: drm dev pointer
4526 *
4527 * Prepare to put the hw in the suspend state (all asics).
4528 * Returns 0 for success or an error on failure.
4529 * Called at driver suspend.
4530 */
4531int amdgpu_device_prepare(struct drm_device *dev)
4532{
4533	struct amdgpu_device *adev = drm_to_adev(dev);
4534	int i, r;
4535
4536	amdgpu_choose_low_power_state(adev);
4537
4538	if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
4539		return 0;
4540
4541	/* Evict the majority of BOs before starting suspend sequence */
4542	r = amdgpu_device_evict_resources(adev);
4543	if (r)
4544		goto unprepare;
4545
4546	flush_delayed_work(&adev->gfx.gfx_off_delay_work);
4547
4548	for (i = 0; i < adev->num_ip_blocks; i++) {
4549		if (!adev->ip_blocks[i].status.valid)
4550			continue;
4551		if (!adev->ip_blocks[i].version->funcs->prepare_suspend)
4552			continue;
4553		r = adev->ip_blocks[i].version->funcs->prepare_suspend((void *)adev);
4554		if (r)
4555			goto unprepare;
4556	}
4557
4558	return 0;
4559
4560unprepare:
4561	adev->in_s0ix = adev->in_s3 = false;
4562
4563	return r;
4564}
4565
4566/**
4567 * amdgpu_device_suspend - initiate device suspend
4568 *
4569 * @dev: drm dev pointer
4570 * @fbcon : notify the fbdev of suspend
4571 *
4572 * Puts the hw in the suspend state (all asics).
4573 * Returns 0 for success or an error on failure.
4574 * Called at driver suspend.
4575 */
4576int amdgpu_device_suspend(struct drm_device *dev, bool fbcon)
4577{
4578	struct amdgpu_device *adev = drm_to_adev(dev);
4579	int r = 0;
4580
4581	if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
4582		return 0;
4583
4584	adev->in_suspend = true;
4585
4586	if (amdgpu_sriov_vf(adev)) {
4587		amdgpu_virt_fini_data_exchange(adev);
4588		r = amdgpu_virt_request_full_gpu(adev, false);
4589		if (r)
4590			return r;
4591	}
4592
4593	if (amdgpu_acpi_smart_shift_update(dev, AMDGPU_SS_DEV_D3))
4594		DRM_WARN("smart shift update failed\n");
4595
 
 
4596	if (fbcon)
4597		drm_fb_helper_set_suspend_unlocked(adev_to_drm(adev)->fb_helper, true);
4598
4599	cancel_delayed_work_sync(&adev->delayed_init_work);
4600
4601	amdgpu_ras_suspend(adev);
4602
4603	amdgpu_device_ip_suspend_phase1(adev);
4604
4605	if (!adev->in_s0ix)
4606		amdgpu_amdkfd_suspend(adev, adev->in_runpm);
4607
4608	r = amdgpu_device_evict_resources(adev);
4609	if (r)
4610		return r;
4611
4612	amdgpu_ttm_set_buffer_funcs_status(adev, false);
4613
4614	amdgpu_fence_driver_hw_fini(adev);
4615
4616	amdgpu_device_ip_suspend_phase2(adev);
4617
4618	if (amdgpu_sriov_vf(adev))
4619		amdgpu_virt_release_full_gpu(adev, false);
4620
4621	r = amdgpu_dpm_notify_rlc_state(adev, false);
4622	if (r)
4623		return r;
4624
4625	return 0;
4626}
4627
4628/**
4629 * amdgpu_device_resume - initiate device resume
4630 *
4631 * @dev: drm dev pointer
4632 * @fbcon : notify the fbdev of resume
4633 *
4634 * Bring the hw back to operating state (all asics).
4635 * Returns 0 for success or an error on failure.
4636 * Called at driver resume.
4637 */
4638int amdgpu_device_resume(struct drm_device *dev, bool fbcon)
4639{
4640	struct amdgpu_device *adev = drm_to_adev(dev);
4641	int r = 0;
4642
4643	if (amdgpu_sriov_vf(adev)) {
4644		r = amdgpu_virt_request_full_gpu(adev, true);
4645		if (r)
4646			return r;
4647	}
4648
4649	if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
4650		return 0;
4651
4652	if (adev->in_s0ix)
4653		amdgpu_dpm_gfx_state_change(adev, sGpuChangeState_D0Entry);
4654
4655	/* post card */
4656	if (amdgpu_device_need_post(adev)) {
4657		r = amdgpu_device_asic_init(adev);
4658		if (r)
4659			dev_err(adev->dev, "amdgpu asic init failed\n");
4660	}
4661
4662	r = amdgpu_device_ip_resume(adev);
4663
4664	if (r) {
4665		dev_err(adev->dev, "amdgpu_device_ip_resume failed (%d).\n", r);
4666		goto exit;
4667	}
4668	amdgpu_fence_driver_hw_init(adev);
4669
4670	if (!adev->in_s0ix) {
4671		r = amdgpu_amdkfd_resume(adev, adev->in_runpm);
4672		if (r)
4673			goto exit;
4674	}
4675
4676	r = amdgpu_device_ip_late_init(adev);
4677	if (r)
4678		goto exit;
4679
4680	queue_delayed_work(system_wq, &adev->delayed_init_work,
4681			   msecs_to_jiffies(AMDGPU_RESUME_MS));
4682exit:
4683	if (amdgpu_sriov_vf(adev)) {
4684		amdgpu_virt_init_data_exchange(adev);
4685		amdgpu_virt_release_full_gpu(adev, true);
4686	}
4687
4688	if (r)
4689		return r;
 
 
 
4690
4691	/* Make sure IB tests flushed */
4692	flush_delayed_work(&adev->delayed_init_work);
4693
4694	if (fbcon)
4695		drm_fb_helper_set_suspend_unlocked(adev_to_drm(adev)->fb_helper, false);
 
 
4696
4697	amdgpu_ras_resume(adev);
4698
4699	if (adev->mode_info.num_crtc) {
4700		/*
4701		 * Most of the connector probing functions try to acquire runtime pm
4702		 * refs to ensure that the GPU is powered on when connector polling is
4703		 * performed. Since we're calling this from a runtime PM callback,
4704		 * trying to acquire rpm refs will cause us to deadlock.
4705		 *
4706		 * Since we're guaranteed to be holding the rpm lock, it's safe to
4707		 * temporarily disable the rpm helpers so this doesn't deadlock us.
4708		 */
4709#ifdef CONFIG_PM
4710		dev->dev->power.disable_depth++;
4711#endif
4712		if (!adev->dc_enabled)
4713			drm_helper_hpd_irq_event(dev);
4714		else
4715			drm_kms_helper_hotplug_event(dev);
4716#ifdef CONFIG_PM
4717		dev->dev->power.disable_depth--;
4718#endif
4719	}
4720	adev->in_suspend = false;
4721
4722	if (adev->enable_mes)
4723		amdgpu_mes_self_test(adev);
4724
4725	if (amdgpu_acpi_smart_shift_update(dev, AMDGPU_SS_DEV_D0))
4726		DRM_WARN("smart shift update failed\n");
4727
4728	return 0;
4729}
4730
4731/**
4732 * amdgpu_device_ip_check_soft_reset - did soft reset succeed
4733 *
4734 * @adev: amdgpu_device pointer
4735 *
4736 * The list of all the hardware IPs that make up the asic is walked and
4737 * the check_soft_reset callbacks are run.  check_soft_reset determines
4738 * if the asic is still hung or not.
4739 * Returns true if any of the IPs are still in a hung state, false if not.
4740 */
4741static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)
4742{
4743	int i;
4744	bool asic_hang = false;
4745
4746	if (amdgpu_sriov_vf(adev))
4747		return true;
4748
4749	if (amdgpu_asic_need_full_reset(adev))
4750		return true;
4751
4752	for (i = 0; i < adev->num_ip_blocks; i++) {
4753		if (!adev->ip_blocks[i].status.valid)
4754			continue;
4755		if (adev->ip_blocks[i].version->funcs->check_soft_reset)
4756			adev->ip_blocks[i].status.hang =
4757				adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
4758		if (adev->ip_blocks[i].status.hang) {
4759			dev_info(adev->dev, "IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
4760			asic_hang = true;
4761		}
4762	}
4763	return asic_hang;
4764}
4765
4766/**
4767 * amdgpu_device_ip_pre_soft_reset - prepare for soft reset
4768 *
4769 * @adev: amdgpu_device pointer
4770 *
4771 * The list of all the hardware IPs that make up the asic is walked and the
4772 * pre_soft_reset callbacks are run if the block is hung.  pre_soft_reset
4773 * handles any IP specific hardware or software state changes that are
4774 * necessary for a soft reset to succeed.
4775 * Returns 0 on success, negative error code on failure.
4776 */
4777static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)
4778{
4779	int i, r = 0;
4780
4781	for (i = 0; i < adev->num_ip_blocks; i++) {
4782		if (!adev->ip_blocks[i].status.valid)
4783			continue;
4784		if (adev->ip_blocks[i].status.hang &&
4785		    adev->ip_blocks[i].version->funcs->pre_soft_reset) {
4786			r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
4787			if (r)
4788				return r;
4789		}
4790	}
4791
4792	return 0;
4793}
4794
4795/**
4796 * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed
4797 *
4798 * @adev: amdgpu_device pointer
4799 *
4800 * Some hardware IPs cannot be soft reset.  If they are hung, a full gpu
4801 * reset is necessary to recover.
4802 * Returns true if a full asic reset is required, false if not.
4803 */
4804static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)
4805{
4806	int i;
4807
4808	if (amdgpu_asic_need_full_reset(adev))
4809		return true;
4810
4811	for (i = 0; i < adev->num_ip_blocks; i++) {
4812		if (!adev->ip_blocks[i].status.valid)
4813			continue;
4814		if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
4815		    (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
4816		    (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
4817		    (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
4818		     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
4819			if (adev->ip_blocks[i].status.hang) {
4820				dev_info(adev->dev, "Some block need full reset!\n");
4821				return true;
4822			}
4823		}
4824	}
4825	return false;
4826}
4827
4828/**
4829 * amdgpu_device_ip_soft_reset - do a soft reset
4830 *
4831 * @adev: amdgpu_device pointer
4832 *
4833 * The list of all the hardware IPs that make up the asic is walked and the
4834 * soft_reset callbacks are run if the block is hung.  soft_reset handles any
4835 * IP specific hardware or software state changes that are necessary to soft
4836 * reset the IP.
4837 * Returns 0 on success, negative error code on failure.
4838 */
4839static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)
4840{
4841	int i, r = 0;
4842
4843	for (i = 0; i < adev->num_ip_blocks; i++) {
4844		if (!adev->ip_blocks[i].status.valid)
4845			continue;
4846		if (adev->ip_blocks[i].status.hang &&
4847		    adev->ip_blocks[i].version->funcs->soft_reset) {
4848			r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
4849			if (r)
4850				return r;
4851		}
4852	}
4853
4854	return 0;
4855}
4856
4857/**
4858 * amdgpu_device_ip_post_soft_reset - clean up from soft reset
4859 *
4860 * @adev: amdgpu_device pointer
4861 *
4862 * The list of all the hardware IPs that make up the asic is walked and the
4863 * post_soft_reset callbacks are run if the asic was hung.  post_soft_reset
4864 * handles any IP specific hardware or software state changes that are
4865 * necessary after the IP has been soft reset.
4866 * Returns 0 on success, negative error code on failure.
4867 */
4868static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)
4869{
4870	int i, r = 0;
4871
4872	for (i = 0; i < adev->num_ip_blocks; i++) {
4873		if (!adev->ip_blocks[i].status.valid)
4874			continue;
4875		if (adev->ip_blocks[i].status.hang &&
4876		    adev->ip_blocks[i].version->funcs->post_soft_reset)
4877			r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
4878		if (r)
4879			return r;
4880	}
4881
4882	return 0;
4883}
4884
4885/**
4886 * amdgpu_device_recover_vram - Recover some VRAM contents
4887 *
4888 * @adev: amdgpu_device pointer
4889 *
4890 * Restores the contents of VRAM buffers from the shadows in GTT.  Used to
4891 * restore things like GPUVM page tables after a GPU reset where
4892 * the contents of VRAM might be lost.
4893 *
4894 * Returns:
4895 * 0 on success, negative error code on failure.
4896 */
4897static int amdgpu_device_recover_vram(struct amdgpu_device *adev)
4898{
4899	struct dma_fence *fence = NULL, *next = NULL;
4900	struct amdgpu_bo *shadow;
4901	struct amdgpu_bo_vm *vmbo;
4902	long r = 1, tmo;
4903
4904	if (amdgpu_sriov_runtime(adev))
4905		tmo = msecs_to_jiffies(8000);
4906	else
4907		tmo = msecs_to_jiffies(100);
4908
4909	dev_info(adev->dev, "recover vram bo from shadow start\n");
4910	mutex_lock(&adev->shadow_list_lock);
4911	list_for_each_entry(vmbo, &adev->shadow_list, shadow_list) {
4912		/* If vm is compute context or adev is APU, shadow will be NULL */
4913		if (!vmbo->shadow)
4914			continue;
4915		shadow = vmbo->shadow;
4916
4917		/* No need to recover an evicted BO */
4918		if (shadow->tbo.resource->mem_type != TTM_PL_TT ||
4919		    shadow->tbo.resource->start == AMDGPU_BO_INVALID_OFFSET ||
4920		    shadow->parent->tbo.resource->mem_type != TTM_PL_VRAM)
4921			continue;
4922
4923		r = amdgpu_bo_restore_shadow(shadow, &next);
4924		if (r)
4925			break;
4926
4927		if (fence) {
4928			tmo = dma_fence_wait_timeout(fence, false, tmo);
4929			dma_fence_put(fence);
4930			fence = next;
4931			if (tmo == 0) {
4932				r = -ETIMEDOUT;
4933				break;
4934			} else if (tmo < 0) {
4935				r = tmo;
4936				break;
4937			}
4938		} else {
4939			fence = next;
4940		}
4941	}
4942	mutex_unlock(&adev->shadow_list_lock);
4943
4944	if (fence)
4945		tmo = dma_fence_wait_timeout(fence, false, tmo);
4946	dma_fence_put(fence);
4947
4948	if (r < 0 || tmo <= 0) {
4949		dev_err(adev->dev, "recover vram bo from shadow failed, r is %ld, tmo is %ld\n", r, tmo);
4950		return -EIO;
4951	}
4952
4953	dev_info(adev->dev, "recover vram bo from shadow done\n");
4954	return 0;
4955}
4956
4957
4958/**
4959 * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
4960 *
4961 * @adev: amdgpu_device pointer
4962 * @from_hypervisor: request from hypervisor
4963 *
4964 * do VF FLR and reinitialize Asic
4965 * return 0 means succeeded otherwise failed
4966 */
4967static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
4968				     bool from_hypervisor)
4969{
4970	int r;
4971	struct amdgpu_hive_info *hive = NULL;
4972	int retry_limit = 0;
4973
4974retry:
4975	amdgpu_amdkfd_pre_reset(adev);
4976
4977	if (from_hypervisor)
4978		r = amdgpu_virt_request_full_gpu(adev, true);
4979	else
4980		r = amdgpu_virt_reset_gpu(adev);
4981	if (r)
4982		return r;
4983	amdgpu_irq_gpu_reset_resume_helper(adev);
4984
4985	/* some sw clean up VF needs to do before recover */
4986	amdgpu_virt_post_reset(adev);
4987
4988	/* Resume IP prior to SMC */
4989	r = amdgpu_device_ip_reinit_early_sriov(adev);
4990	if (r)
4991		goto error;
4992
4993	amdgpu_virt_init_data_exchange(adev);
 
 
4994
4995	r = amdgpu_device_fw_loading(adev);
4996	if (r)
4997		return r;
4998
4999	/* now we are okay to resume SMC/CP/SDMA */
5000	r = amdgpu_device_ip_reinit_late_sriov(adev);
5001	if (r)
5002		goto error;
5003
5004	hive = amdgpu_get_xgmi_hive(adev);
5005	/* Update PSP FW topology after reset */
5006	if (hive && adev->gmc.xgmi.num_physical_nodes > 1)
5007		r = amdgpu_xgmi_update_topology(hive, adev);
5008
5009	if (hive)
5010		amdgpu_put_xgmi_hive(hive);
5011
5012	if (!r) {
5013		r = amdgpu_ib_ring_tests(adev);
5014
5015		amdgpu_amdkfd_post_reset(adev);
5016	}
5017
5018error:
5019	if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
5020		amdgpu_inc_vram_lost(adev);
5021		r = amdgpu_device_recover_vram(adev);
5022	}
5023	amdgpu_virt_release_full_gpu(adev, true);
5024
5025	if (AMDGPU_RETRY_SRIOV_RESET(r)) {
5026		if (retry_limit < AMDGPU_MAX_RETRY_LIMIT) {
5027			retry_limit++;
5028			goto retry;
5029		} else
5030			DRM_ERROR("GPU reset retry is beyond the retry limit\n");
5031	}
5032
5033	return r;
5034}
5035
5036/**
5037 * amdgpu_device_has_job_running - check if there is any job in mirror list
5038 *
5039 * @adev: amdgpu_device pointer
5040 *
5041 * check if there is any job in mirror list
5042 */
5043bool amdgpu_device_has_job_running(struct amdgpu_device *adev)
5044{
5045	int i;
5046	struct drm_sched_job *job;
5047
5048	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
5049		struct amdgpu_ring *ring = adev->rings[i];
5050
5051		if (!amdgpu_ring_sched_ready(ring))
5052			continue;
5053
5054		spin_lock(&ring->sched.job_list_lock);
5055		job = list_first_entry_or_null(&ring->sched.pending_list,
5056					       struct drm_sched_job, list);
5057		spin_unlock(&ring->sched.job_list_lock);
5058		if (job)
5059			return true;
5060	}
5061	return false;
5062}
5063
5064/**
5065 * amdgpu_device_should_recover_gpu - check if we should try GPU recovery
5066 *
5067 * @adev: amdgpu_device pointer
5068 *
5069 * Check amdgpu_gpu_recovery and SRIOV status to see if we should try to recover
5070 * a hung GPU.
5071 */
5072bool amdgpu_device_should_recover_gpu(struct amdgpu_device *adev)
5073{
 
 
 
 
5074
5075	if (amdgpu_gpu_recovery == 0)
5076		goto disabled;
5077
5078	/* Skip soft reset check in fatal error mode */
5079	if (!amdgpu_ras_is_poison_mode_supported(adev))
5080		return true;
5081
5082	if (amdgpu_sriov_vf(adev))
5083		return true;
5084
5085	if (amdgpu_gpu_recovery == -1) {
5086		switch (adev->asic_type) {
5087#ifdef CONFIG_DRM_AMDGPU_SI
5088		case CHIP_VERDE:
5089		case CHIP_TAHITI:
5090		case CHIP_PITCAIRN:
5091		case CHIP_OLAND:
5092		case CHIP_HAINAN:
5093#endif
5094#ifdef CONFIG_DRM_AMDGPU_CIK
5095		case CHIP_KAVERI:
5096		case CHIP_KABINI:
5097		case CHIP_MULLINS:
5098#endif
5099		case CHIP_CARRIZO:
5100		case CHIP_STONEY:
5101		case CHIP_CYAN_SKILLFISH:
5102			goto disabled;
5103		default:
 
 
 
 
 
 
 
5104			break;
 
 
5105		}
5106	}
5107
5108	return true;
5109
5110disabled:
5111		dev_info(adev->dev, "GPU recovery disabled.\n");
5112		return false;
5113}
5114
5115int amdgpu_device_mode1_reset(struct amdgpu_device *adev)
5116{
5117	u32 i;
5118	int ret = 0;
5119
5120	amdgpu_atombios_scratch_regs_engine_hung(adev, true);
5121
5122	dev_info(adev->dev, "GPU mode1 reset\n");
5123
5124	/* disable BM */
5125	pci_clear_master(adev->pdev);
5126
5127	amdgpu_device_cache_pci_state(adev->pdev);
5128
5129	if (amdgpu_dpm_is_mode1_reset_supported(adev)) {
5130		dev_info(adev->dev, "GPU smu mode1 reset\n");
5131		ret = amdgpu_dpm_mode1_reset(adev);
5132	} else {
5133		dev_info(adev->dev, "GPU psp mode1 reset\n");
5134		ret = psp_gpu_reset(adev);
5135	}
5136
5137	if (ret)
5138		goto mode1_reset_failed;
5139
5140	amdgpu_device_load_pci_state(adev->pdev);
5141	ret = amdgpu_psp_wait_for_bootloader(adev);
5142	if (ret)
5143		goto mode1_reset_failed;
5144
5145	/* wait for asic to come out of reset */
5146	for (i = 0; i < adev->usec_timeout; i++) {
5147		u32 memsize = adev->nbio.funcs->get_memsize(adev);
5148
5149		if (memsize != 0xffffffff)
5150			break;
5151		udelay(1);
5152	}
5153
5154	if (i >= adev->usec_timeout) {
5155		ret = -ETIMEDOUT;
5156		goto mode1_reset_failed;
5157	}
5158
5159	amdgpu_atombios_scratch_regs_engine_hung(adev, false);
 
 
 
5160
5161	return 0;
5162
5163mode1_reset_failed:
5164	dev_err(adev->dev, "GPU mode1 reset failed\n");
5165	return ret;
5166}
5167
5168int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,
5169				 struct amdgpu_reset_context *reset_context)
5170{
5171	int i, r = 0;
5172	struct amdgpu_job *job = NULL;
5173	bool need_full_reset =
5174		test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
5175
5176	if (reset_context->reset_req_dev == adev)
5177		job = reset_context->job;
5178
 
 
 
 
5179	if (amdgpu_sriov_vf(adev)) {
5180		/* stop the data exchange thread */
5181		amdgpu_virt_fini_data_exchange(adev);
5182	}
5183
5184	amdgpu_fence_driver_isr_toggle(adev, true);
5185
5186	/* block all schedulers and reset given job's ring */
5187	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
5188		struct amdgpu_ring *ring = adev->rings[i];
5189
5190		if (!amdgpu_ring_sched_ready(ring))
5191			continue;
5192
5193		/* Clear job fence from fence drv to avoid force_completion
5194		 * leave NULL and vm flush fence in fence drv
5195		 */
5196		amdgpu_fence_driver_clear_job_fences(ring);
5197
5198		/* after all hw jobs are reset, hw fence is meaningless, so force_completion */
5199		amdgpu_fence_driver_force_completion(ring);
5200	}
5201
5202	amdgpu_fence_driver_isr_toggle(adev, false);
5203
5204	if (job && job->vm)
5205		drm_sched_increase_karma(&job->base);
5206
5207	r = amdgpu_reset_prepare_hwcontext(adev, reset_context);
5208	/* If reset handler not implemented, continue; otherwise return */
5209	if (r == -EOPNOTSUPP)
5210		r = 0;
5211	else
5212		return r;
5213
5214	/* Don't suspend on bare metal if we are not going to HW reset the ASIC */
5215	if (!amdgpu_sriov_vf(adev)) {
5216
5217		if (!need_full_reset)
5218			need_full_reset = amdgpu_device_ip_need_full_reset(adev);
5219
5220		if (!need_full_reset && amdgpu_gpu_recovery &&
5221		    amdgpu_device_ip_check_soft_reset(adev)) {
5222			amdgpu_device_ip_pre_soft_reset(adev);
5223			r = amdgpu_device_ip_soft_reset(adev);
5224			amdgpu_device_ip_post_soft_reset(adev);
5225			if (r || amdgpu_device_ip_check_soft_reset(adev)) {
5226				dev_info(adev->dev, "soft reset failed, will fallback to full reset!\n");
5227				need_full_reset = true;
5228			}
5229		}
5230
5231		if (need_full_reset)
5232			r = amdgpu_device_ip_suspend(adev);
5233		if (need_full_reset)
5234			set_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
5235		else
5236			clear_bit(AMDGPU_NEED_FULL_RESET,
5237				  &reset_context->flags);
5238	}
5239
5240	return r;
5241}
5242
5243static int amdgpu_reset_reg_dumps(struct amdgpu_device *adev)
5244{
5245	int i;
5246
5247	lockdep_assert_held(&adev->reset_domain->sem);
5248
5249	for (i = 0; i < adev->reset_info.num_regs; i++) {
5250		adev->reset_info.reset_dump_reg_value[i] =
5251			RREG32(adev->reset_info.reset_dump_reg_list[i]);
5252
5253		trace_amdgpu_reset_reg_dumps(adev->reset_info.reset_dump_reg_list[i],
5254					     adev->reset_info.reset_dump_reg_value[i]);
5255	}
5256
5257	return 0;
5258}
5259
5260int amdgpu_do_asic_reset(struct list_head *device_list_handle,
5261			 struct amdgpu_reset_context *reset_context)
5262{
5263	struct amdgpu_device *tmp_adev = NULL;
5264	bool need_full_reset, skip_hw_reset, vram_lost = false;
5265	int r = 0;
5266
5267	/* Try reset handler method first */
5268	tmp_adev = list_first_entry(device_list_handle, struct amdgpu_device,
5269				    reset_list);
5270	amdgpu_reset_reg_dumps(tmp_adev);
5271
5272	reset_context->reset_device_list = device_list_handle;
5273	r = amdgpu_reset_perform_reset(tmp_adev, reset_context);
5274	/* If reset handler not implemented, continue; otherwise return */
5275	if (r == -EOPNOTSUPP)
5276		r = 0;
5277	else
5278		return r;
5279
5280	/* Reset handler not implemented, use the default method */
5281	need_full_reset =
5282		test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
5283	skip_hw_reset = test_bit(AMDGPU_SKIP_HW_RESET, &reset_context->flags);
5284
5285	/*
5286	 * ASIC reset has to be done on all XGMI hive nodes ASAP
5287	 * to allow proper links negotiation in FW (within 1 sec)
5288	 */
5289	if (!skip_hw_reset && need_full_reset) {
5290		list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5291			/* For XGMI run all resets in parallel to speed up the process */
5292			if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
5293				tmp_adev->gmc.xgmi.pending_reset = false;
5294				if (!queue_work(system_unbound_wq, &tmp_adev->xgmi_reset_work))
5295					r = -EALREADY;
5296			} else
5297				r = amdgpu_asic_reset(tmp_adev);
5298
5299			if (r) {
5300				dev_err(tmp_adev->dev, "ASIC reset failed with error, %d for drm dev, %s",
5301					 r, adev_to_drm(tmp_adev)->unique);
5302				goto out;
5303			}
5304		}
5305
5306		/* For XGMI wait for all resets to complete before proceed */
5307		if (!r) {
5308			list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5309				if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
5310					flush_work(&tmp_adev->xgmi_reset_work);
5311					r = tmp_adev->asic_reset_res;
5312					if (r)
5313						break;
5314				}
5315			}
5316		}
5317	}
5318
5319	if (!r && amdgpu_ras_intr_triggered()) {
5320		list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5321			amdgpu_ras_reset_error_count(tmp_adev, AMDGPU_RAS_BLOCK__MMHUB);
 
 
5322		}
5323
5324		amdgpu_ras_intr_cleared();
5325	}
5326
5327	list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5328		if (need_full_reset) {
5329			/* post card */
5330			amdgpu_ras_set_fed(tmp_adev, false);
5331			r = amdgpu_device_asic_init(tmp_adev);
5332			if (r) {
5333				dev_warn(tmp_adev->dev, "asic atom init failed!");
5334			} else {
5335				dev_info(tmp_adev->dev, "GPU reset succeeded, trying to resume\n");
 
 
 
5336
5337				r = amdgpu_device_ip_resume_phase1(tmp_adev);
5338				if (r)
5339					goto out;
5340
5341				vram_lost = amdgpu_device_check_vram_lost(tmp_adev);
5342
5343				amdgpu_coredump(tmp_adev, vram_lost, reset_context);
5344
5345				if (vram_lost) {
5346					DRM_INFO("VRAM is lost due to GPU reset!\n");
5347					amdgpu_inc_vram_lost(tmp_adev);
5348				}
5349
5350				r = amdgpu_device_fw_loading(tmp_adev);
5351				if (r)
5352					return r;
5353
5354				r = amdgpu_xcp_restore_partition_mode(
5355					tmp_adev->xcp_mgr);
5356				if (r)
5357					goto out;
5358
5359				r = amdgpu_device_ip_resume_phase2(tmp_adev);
5360				if (r)
5361					goto out;
5362
5363				if (tmp_adev->mman.buffer_funcs_ring->sched.ready)
5364					amdgpu_ttm_set_buffer_funcs_status(tmp_adev, true);
5365
5366				if (vram_lost)
5367					amdgpu_device_fill_reset_magic(tmp_adev);
5368
5369				/*
5370				 * Add this ASIC as tracked as reset was already
5371				 * complete successfully.
5372				 */
5373				amdgpu_register_gpu_instance(tmp_adev);
5374
5375				if (!reset_context->hive &&
5376				    tmp_adev->gmc.xgmi.num_physical_nodes > 1)
5377					amdgpu_xgmi_add_device(tmp_adev);
5378
5379				r = amdgpu_device_ip_late_init(tmp_adev);
5380				if (r)
5381					goto out;
5382
5383				drm_fb_helper_set_suspend_unlocked(adev_to_drm(tmp_adev)->fb_helper, false);
5384
5385				/*
5386				 * The GPU enters bad state once faulty pages
5387				 * by ECC has reached the threshold, and ras
5388				 * recovery is scheduled next. So add one check
5389				 * here to break recovery if it indeed exceeds
5390				 * bad page threshold, and remind user to
5391				 * retire this GPU or setting one bigger
5392				 * bad_page_threshold value to fix this once
5393				 * probing driver again.
5394				 */
5395				if (!amdgpu_ras_eeprom_check_err_threshold(tmp_adev)) {
5396					/* must succeed. */
5397					amdgpu_ras_resume(tmp_adev);
5398				} else {
5399					r = -EINVAL;
5400					goto out;
5401				}
5402
5403				/* Update PSP FW topology after reset */
5404				if (reset_context->hive &&
5405				    tmp_adev->gmc.xgmi.num_physical_nodes > 1)
5406					r = amdgpu_xgmi_update_topology(
5407						reset_context->hive, tmp_adev);
5408			}
5409		}
5410
5411out:
5412		if (!r) {
5413			amdgpu_irq_gpu_reset_resume_helper(tmp_adev);
5414			r = amdgpu_ib_ring_tests(tmp_adev);
5415			if (r) {
5416				dev_err(tmp_adev->dev, "ib ring test failed (%d).\n", r);
5417				need_full_reset = true;
5418				r = -EAGAIN;
5419				goto end;
5420			}
5421		}
5422
5423		if (!r)
5424			r = amdgpu_device_recover_vram(tmp_adev);
5425		else
5426			tmp_adev->asic_reset_res = r;
5427	}
5428
5429end:
5430	if (need_full_reset)
5431		set_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
5432	else
5433		clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
5434	return r;
5435}
5436
5437static void amdgpu_device_set_mp1_state(struct amdgpu_device *adev)
 
5438{
 
 
 
 
 
 
 
 
5439
5440	switch (amdgpu_asic_reset_method(adev)) {
5441	case AMD_RESET_METHOD_MODE1:
5442		adev->mp1_state = PP_MP1_STATE_SHUTDOWN;
5443		break;
5444	case AMD_RESET_METHOD_MODE2:
5445		adev->mp1_state = PP_MP1_STATE_RESET;
5446		break;
5447	default:
5448		adev->mp1_state = PP_MP1_STATE_NONE;
5449		break;
5450	}
 
 
5451}
5452
5453static void amdgpu_device_unset_mp1_state(struct amdgpu_device *adev)
5454{
5455	amdgpu_vf_error_trans_all(adev);
5456	adev->mp1_state = PP_MP1_STATE_NONE;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5457}
5458
5459static void amdgpu_device_resume_display_audio(struct amdgpu_device *adev)
5460{
5461	struct pci_dev *p = NULL;
5462
5463	p = pci_get_domain_bus_and_slot(pci_domain_nr(adev->pdev->bus),
5464			adev->pdev->bus->number, 1);
5465	if (p) {
5466		pm_runtime_enable(&(p->dev));
5467		pm_runtime_resume(&(p->dev));
5468	}
5469
5470	pci_dev_put(p);
5471}
5472
5473static int amdgpu_device_suspend_display_audio(struct amdgpu_device *adev)
5474{
5475	enum amd_reset_method reset_method;
5476	struct pci_dev *p = NULL;
5477	u64 expires;
5478
5479	/*
5480	 * For now, only BACO and mode1 reset are confirmed
5481	 * to suffer the audio issue without proper suspended.
5482	 */
5483	reset_method = amdgpu_asic_reset_method(adev);
5484	if ((reset_method != AMD_RESET_METHOD_BACO) &&
5485	     (reset_method != AMD_RESET_METHOD_MODE1))
5486		return -EINVAL;
5487
5488	p = pci_get_domain_bus_and_slot(pci_domain_nr(adev->pdev->bus),
5489			adev->pdev->bus->number, 1);
5490	if (!p)
5491		return -ENODEV;
5492
5493	expires = pm_runtime_autosuspend_expiration(&(p->dev));
5494	if (!expires)
5495		/*
5496		 * If we cannot get the audio device autosuspend delay,
5497		 * a fixed 4S interval will be used. Considering 3S is
5498		 * the audio controller default autosuspend delay setting.
5499		 * 4S used here is guaranteed to cover that.
5500		 */
5501		expires = ktime_get_mono_fast_ns() + NSEC_PER_SEC * 4ULL;
5502
5503	while (!pm_runtime_status_suspended(&(p->dev))) {
5504		if (!pm_runtime_suspend(&(p->dev)))
5505			break;
5506
5507		if (expires < ktime_get_mono_fast_ns()) {
5508			dev_warn(adev->dev, "failed to suspend display audio\n");
5509			pci_dev_put(p);
5510			/* TODO: abort the succeeding gpu reset? */
5511			return -ETIMEDOUT;
5512		}
5513	}
5514
5515	pm_runtime_disable(&(p->dev));
5516
5517	pci_dev_put(p);
5518	return 0;
5519}
5520
5521static inline void amdgpu_device_stop_pending_resets(struct amdgpu_device *adev)
 
 
5522{
5523	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
5524
5525#if defined(CONFIG_DEBUG_FS)
5526	if (!amdgpu_sriov_vf(adev))
5527		cancel_work(&adev->reset_work);
5528#endif
5529
5530	if (adev->kfd.dev)
5531		cancel_work(&adev->kfd.reset_work);
5532
5533	if (amdgpu_sriov_vf(adev))
5534		cancel_work(&adev->virt.flr_work);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5535
5536	if (con && adev->ras_enabled)
5537		cancel_work(&con->recovery_work);
 
 
 
 
 
5538
 
 
 
 
 
 
 
 
 
 
 
 
5539}
5540
5541/**
5542 * amdgpu_device_gpu_recover - reset the asic and recover scheduler
5543 *
5544 * @adev: amdgpu_device pointer
5545 * @job: which job trigger hang
5546 * @reset_context: amdgpu reset context pointer
5547 *
5548 * Attempt to reset the GPU if it has hung (all asics).
5549 * Attempt to do soft-reset or full-reset and reinitialize Asic
5550 * Returns 0 for success or an error on failure.
5551 */
5552
5553int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
5554			      struct amdgpu_job *job,
5555			      struct amdgpu_reset_context *reset_context)
5556{
5557	struct list_head device_list, *device_list_handle =  NULL;
5558	bool job_signaled = false;
5559	struct amdgpu_hive_info *hive = NULL;
5560	struct amdgpu_device *tmp_adev = NULL;
5561	int i, r = 0;
5562	bool need_emergency_restart = false;
5563	bool audio_suspended = false;
 
 
 
 
5564
5565	/*
5566	 * Special case: RAS triggered and full reset isn't supported
5567	 */
5568	need_emergency_restart = amdgpu_ras_need_emergency_restart(adev);
5569
5570	/*
5571	 * Flush RAM to disk so that after reboot
5572	 * the user can read log and see why the system rebooted.
5573	 */
5574	if (need_emergency_restart && amdgpu_ras_get_context(adev) &&
5575		amdgpu_ras_get_context(adev)->reboot) {
5576		DRM_WARN("Emergency reboot.");
5577
5578		ksys_sync_helper();
5579		emergency_restart();
5580	}
5581
5582	dev_info(adev->dev, "GPU %s begin!\n",
5583		need_emergency_restart ? "jobs stop":"reset");
5584
5585	if (!amdgpu_sriov_vf(adev))
5586		hive = amdgpu_get_xgmi_hive(adev);
5587	if (hive)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5588		mutex_lock(&hive->hive_lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5589
5590	reset_context->job = job;
5591	reset_context->hive = hive;
5592	/*
5593	 * Build list of devices to reset.
5594	 * In case we are in XGMI hive mode, resort the device list
5595	 * to put adev in the 1st position.
5596	 */
5597	INIT_LIST_HEAD(&device_list);
5598	if (!amdgpu_sriov_vf(adev) && (adev->gmc.xgmi.num_physical_nodes > 1)) {
5599		list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
5600			list_add_tail(&tmp_adev->reset_list, &device_list);
5601			if (adev->shutdown)
5602				tmp_adev->shutdown = true;
5603		}
5604		if (!list_is_first(&adev->reset_list, &device_list))
5605			list_rotate_to_front(&adev->reset_list, &device_list);
5606		device_list_handle = &device_list;
5607	} else {
5608		list_add_tail(&adev->reset_list, &device_list);
5609		device_list_handle = &device_list;
5610	}
5611
5612	/* We need to lock reset domain only once both for XGMI and single device */
5613	tmp_adev = list_first_entry(device_list_handle, struct amdgpu_device,
5614				    reset_list);
5615	amdgpu_device_lock_reset_domain(tmp_adev->reset_domain);
5616
5617	/* block all schedulers and reset given job's ring */
5618	list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5619
5620		amdgpu_device_set_mp1_state(tmp_adev);
5621
5622		/*
5623		 * Try to put the audio codec into suspend state
5624		 * before gpu reset started.
5625		 *
5626		 * Due to the power domain of the graphics device
5627		 * is shared with AZ power domain. Without this,
5628		 * we may change the audio hardware from behind
5629		 * the audio driver's back. That will trigger
5630		 * some audio codec errors.
5631		 */
5632		if (!amdgpu_device_suspend_display_audio(tmp_adev))
5633			audio_suspended = true;
5634
5635		amdgpu_ras_set_error_query_ready(tmp_adev, false);
5636
5637		cancel_delayed_work_sync(&tmp_adev->delayed_init_work);
5638
5639		if (!amdgpu_sriov_vf(tmp_adev))
5640			amdgpu_amdkfd_pre_reset(tmp_adev);
5641
5642		/*
5643		 * Mark these ASICs to be reseted as untracked first
5644		 * And add them back after reset completed
5645		 */
5646		amdgpu_unregister_gpu_instance(tmp_adev);
5647
5648		drm_fb_helper_set_suspend_unlocked(adev_to_drm(tmp_adev)->fb_helper, true);
5649
5650		/* disable ras on ALL IPs */
5651		if (!need_emergency_restart &&
5652		      amdgpu_device_ip_need_full_reset(tmp_adev))
5653			amdgpu_ras_suspend(tmp_adev);
5654
5655		for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
5656			struct amdgpu_ring *ring = tmp_adev->rings[i];
5657
5658			if (!amdgpu_ring_sched_ready(ring))
5659				continue;
5660
5661			drm_sched_stop(&ring->sched, job ? &job->base : NULL);
5662
5663			if (need_emergency_restart)
5664				amdgpu_job_stop_all_jobs_on_sched(&ring->sched);
5665		}
5666		atomic_inc(&tmp_adev->gpu_reset_counter);
5667	}
5668
5669	if (need_emergency_restart)
5670		goto skip_sched_resume;
5671
5672	/*
5673	 * Must check guilty signal here since after this point all old
5674	 * HW fences are force signaled.
5675	 *
5676	 * job->base holds a reference to parent fence
5677	 */
5678	if (job && dma_fence_is_signaled(&job->hw_fence)) {
 
5679		job_signaled = true;
5680		dev_info(adev->dev, "Guilty job already signaled, skipping HW reset");
5681		goto skip_hw_reset;
5682	}
5683
5684retry:	/* Rest of adevs pre asic reset from XGMI hive. */
5685	list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5686		r = amdgpu_device_pre_asic_reset(tmp_adev, reset_context);
5687		/*TODO Should we stop ?*/
5688		if (r) {
5689			dev_err(tmp_adev->dev, "GPU pre asic reset failed with err, %d for drm dev, %s ",
5690				  r, adev_to_drm(tmp_adev)->unique);
5691			tmp_adev->asic_reset_res = r;
5692		}
5693
5694		/*
5695		 * Drop all pending non scheduler resets. Scheduler resets
5696		 * were already dropped during drm_sched_stop
5697		 */
5698		amdgpu_device_stop_pending_resets(tmp_adev);
5699	}
5700
 
5701	/* Actual ASIC resets if needed.*/
5702	/* Host driver will handle XGMI hive reset for SRIOV */
5703	if (amdgpu_sriov_vf(adev)) {
5704		r = amdgpu_device_reset_sriov(adev, job ? false : true);
5705		if (r)
5706			adev->asic_reset_res = r;
5707
5708		/* Aldebaran and gfx_11_0_3 support ras in SRIOV, so need resume ras during reset */
5709		if (amdgpu_ip_version(adev, GC_HWIP, 0) ==
5710			    IP_VERSION(9, 4, 2) ||
5711		    amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3) ||
5712		    amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(11, 0, 3))
5713			amdgpu_ras_resume(adev);
5714	} else {
5715		r = amdgpu_do_asic_reset(device_list_handle, reset_context);
5716		if (r && r == -EAGAIN)
5717			goto retry;
5718	}
5719
5720skip_hw_reset:
5721
5722	/* Post ASIC reset for all devs .*/
5723	list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5724
 
 
 
 
 
 
 
 
 
 
 
 
5725		for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
5726			struct amdgpu_ring *ring = tmp_adev->rings[i];
5727
5728			if (!amdgpu_ring_sched_ready(ring))
5729				continue;
5730
5731			drm_sched_start(&ring->sched, true);
 
 
 
 
5732		}
5733
5734		if (!drm_drv_uses_atomic_modeset(adev_to_drm(tmp_adev)) && !job_signaled)
5735			drm_helper_resume_force_mode(adev_to_drm(tmp_adev));
5736
5737		if (tmp_adev->asic_reset_res)
5738			r = tmp_adev->asic_reset_res;
5739
5740		tmp_adev->asic_reset_res = 0;
5741
5742		if (r) {
5743			/* bad news, how to tell it to userspace ? */
5744			dev_info(tmp_adev->dev, "GPU reset(%d) failed\n", atomic_read(&tmp_adev->gpu_reset_counter));
5745			amdgpu_vf_error_put(tmp_adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r);
5746		} else {
5747			dev_info(tmp_adev->dev, "GPU reset(%d) succeeded!\n", atomic_read(&tmp_adev->gpu_reset_counter));
5748			if (amdgpu_acpi_smart_shift_update(adev_to_drm(tmp_adev), AMDGPU_SS_DEV_D0))
5749				DRM_WARN("smart shift update failed\n");
5750		}
5751	}
5752
5753skip_sched_resume:
5754	list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5755		/* unlock kfd: SRIOV would do it separately */
5756		if (!need_emergency_restart && !amdgpu_sriov_vf(tmp_adev))
5757			amdgpu_amdkfd_post_reset(tmp_adev);
5758
5759		/* kfd_post_reset will do nothing if kfd device is not initialized,
5760		 * need to bring up kfd here if it's not be initialized before
5761		 */
5762		if (!adev->kfd.init_complete)
5763			amdgpu_amdkfd_device_init(adev);
5764
5765		if (audio_suspended)
5766			amdgpu_device_resume_display_audio(tmp_adev);
5767
5768		amdgpu_device_unset_mp1_state(tmp_adev);
5769
5770		amdgpu_ras_set_error_query_ready(tmp_adev, true);
5771	}
5772
5773	tmp_adev = list_first_entry(device_list_handle, struct amdgpu_device,
5774					    reset_list);
5775	amdgpu_device_unlock_reset_domain(tmp_adev->reset_domain);
5776
5777	if (hive) {
 
5778		mutex_unlock(&hive->hive_lock);
5779		amdgpu_put_xgmi_hive(hive);
5780	}
5781
5782	if (r)
5783		dev_info(adev->dev, "GPU reset end with ret = %d\n", r);
5784
5785	atomic_set(&adev->reset_domain->reset_res, r);
5786	return r;
5787}
5788
5789/**
5790 * amdgpu_device_partner_bandwidth - find the bandwidth of appropriate partner
5791 *
5792 * @adev: amdgpu_device pointer
5793 * @speed: pointer to the speed of the link
5794 * @width: pointer to the width of the link
5795 *
5796 * Evaluate the hierarchy to find the speed and bandwidth capabilities of the
5797 * first physical partner to an AMD dGPU.
5798 * This will exclude any virtual switches and links.
5799 */
5800static void amdgpu_device_partner_bandwidth(struct amdgpu_device *adev,
5801					    enum pci_bus_speed *speed,
5802					    enum pcie_link_width *width)
5803{
5804	struct pci_dev *parent = adev->pdev;
5805
5806	if (!speed || !width)
5807		return;
5808
5809	*speed = PCI_SPEED_UNKNOWN;
5810	*width = PCIE_LNK_WIDTH_UNKNOWN;
5811
5812	if (amdgpu_device_pcie_dynamic_switching_supported(adev)) {
5813		while ((parent = pci_upstream_bridge(parent))) {
5814			/* skip upstream/downstream switches internal to dGPU*/
5815			if (parent->vendor == PCI_VENDOR_ID_ATI)
5816				continue;
5817			*speed = pcie_get_speed_cap(parent);
5818			*width = pcie_get_width_cap(parent);
5819			break;
5820		}
5821	} else {
5822		/* use the current speeds rather than max if switching is not supported */
5823		pcie_bandwidth_available(adev->pdev, NULL, speed, width);
5824	}
5825}
5826
5827/**
5828 * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot
5829 *
5830 * @adev: amdgpu_device pointer
5831 *
5832 * Fetchs and stores in the driver the PCIE capabilities (gen speed
5833 * and lanes) of the slot the device is in. Handles APUs and
5834 * virtualized environments where PCIE config space may not be available.
5835 */
5836static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev)
5837{
5838	struct pci_dev *pdev;
5839	enum pci_bus_speed speed_cap, platform_speed_cap;
5840	enum pcie_link_width platform_link_width;
5841
5842	if (amdgpu_pcie_gen_cap)
5843		adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
5844
5845	if (amdgpu_pcie_lane_cap)
5846		adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
5847
5848	/* covers APUs as well */
5849	if (pci_is_root_bus(adev->pdev->bus) && !amdgpu_passthrough(adev)) {
5850		if (adev->pm.pcie_gen_mask == 0)
5851			adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
5852		if (adev->pm.pcie_mlw_mask == 0)
5853			adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
5854		return;
5855	}
5856
5857	if (adev->pm.pcie_gen_mask && adev->pm.pcie_mlw_mask)
5858		return;
5859
5860	amdgpu_device_partner_bandwidth(adev, &platform_speed_cap,
5861					&platform_link_width);
5862
5863	if (adev->pm.pcie_gen_mask == 0) {
5864		/* asic caps */
5865		pdev = adev->pdev;
5866		speed_cap = pcie_get_speed_cap(pdev);
5867		if (speed_cap == PCI_SPEED_UNKNOWN) {
5868			adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5869						  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5870						  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
5871		} else {
5872			if (speed_cap == PCIE_SPEED_32_0GT)
5873				adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5874							  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5875							  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
5876							  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4 |
5877							  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN5);
5878			else if (speed_cap == PCIE_SPEED_16_0GT)
5879				adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5880							  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5881							  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
5882							  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4);
5883			else if (speed_cap == PCIE_SPEED_8_0GT)
5884				adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5885							  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5886							  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
5887			else if (speed_cap == PCIE_SPEED_5_0GT)
5888				adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5889							  CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2);
5890			else
5891				adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1;
5892		}
5893		/* platform caps */
5894		if (platform_speed_cap == PCI_SPEED_UNKNOWN) {
5895			adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5896						   CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
5897		} else {
5898			if (platform_speed_cap == PCIE_SPEED_32_0GT)
5899				adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5900							   CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5901							   CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
5902							   CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4 |
5903							   CAIL_PCIE_LINK_SPEED_SUPPORT_GEN5);
5904			else if (platform_speed_cap == PCIE_SPEED_16_0GT)
5905				adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5906							   CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5907							   CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
5908							   CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4);
5909			else if (platform_speed_cap == PCIE_SPEED_8_0GT)
5910				adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5911							   CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5912							   CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3);
5913			else if (platform_speed_cap == PCIE_SPEED_5_0GT)
5914				adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5915							   CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
5916			else
5917				adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
5918
5919		}
5920	}
5921	if (adev->pm.pcie_mlw_mask == 0) {
5922		if (platform_link_width == PCIE_LNK_WIDTH_UNKNOWN) {
5923			adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK;
5924		} else {
5925			switch (platform_link_width) {
5926			case PCIE_LNK_X32:
5927				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
5928							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
5929							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
5930							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
5931							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5932							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5933							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5934				break;
5935			case PCIE_LNK_X16:
5936				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
5937							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
5938							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
5939							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5940							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5941							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5942				break;
5943			case PCIE_LNK_X12:
5944				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
5945							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
5946							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5947							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5948							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5949				break;
5950			case PCIE_LNK_X8:
5951				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
5952							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5953							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5954							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5955				break;
5956			case PCIE_LNK_X4:
5957				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5958							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5959							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5960				break;
5961			case PCIE_LNK_X2:
5962				adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5963							  CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5964				break;
5965			case PCIE_LNK_X1:
5966				adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;
5967				break;
5968			default:
5969				break;
5970			}
5971		}
5972	}
5973}
5974
5975/**
5976 * amdgpu_device_is_peer_accessible - Check peer access through PCIe BAR
5977 *
5978 * @adev: amdgpu_device pointer
5979 * @peer_adev: amdgpu_device pointer for peer device trying to access @adev
5980 *
5981 * Return true if @peer_adev can access (DMA) @adev through the PCIe
5982 * BAR, i.e. @adev is "large BAR" and the BAR matches the DMA mask of
5983 * @peer_adev.
5984 */
5985bool amdgpu_device_is_peer_accessible(struct amdgpu_device *adev,
5986				      struct amdgpu_device *peer_adev)
5987{
5988#ifdef CONFIG_HSA_AMD_P2P
5989	uint64_t address_mask = peer_adev->dev->dma_mask ?
5990		~*peer_adev->dev->dma_mask : ~((1ULL << 32) - 1);
5991	resource_size_t aper_limit =
5992		adev->gmc.aper_base + adev->gmc.aper_size - 1;
5993	bool p2p_access =
5994		!adev->gmc.xgmi.connected_to_cpu &&
5995		!(pci_p2pdma_distance(adev->pdev, peer_adev->dev, false) < 0);
5996
5997	return pcie_p2p && p2p_access && (adev->gmc.visible_vram_size &&
5998		adev->gmc.real_vram_size == adev->gmc.visible_vram_size &&
5999		!(adev->gmc.aper_base & address_mask ||
6000		  aper_limit & address_mask));
6001#else
6002	return false;
6003#endif
6004}
6005
6006int amdgpu_device_baco_enter(struct drm_device *dev)
6007{
6008	struct amdgpu_device *adev = drm_to_adev(dev);
6009	struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
6010
6011	if (!amdgpu_device_supports_baco(dev))
6012		return -ENOTSUPP;
6013
6014	if (ras && adev->ras_enabled &&
6015	    adev->nbio.funcs->enable_doorbell_interrupt)
6016		adev->nbio.funcs->enable_doorbell_interrupt(adev, false);
6017
6018	return amdgpu_dpm_baco_enter(adev);
6019}
6020
6021int amdgpu_device_baco_exit(struct drm_device *dev)
6022{
6023	struct amdgpu_device *adev = drm_to_adev(dev);
6024	struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
6025	int ret = 0;
6026
6027	if (!amdgpu_device_supports_baco(dev))
6028		return -ENOTSUPP;
6029
6030	ret = amdgpu_dpm_baco_exit(adev);
6031	if (ret)
6032		return ret;
6033
6034	if (ras && adev->ras_enabled &&
6035	    adev->nbio.funcs->enable_doorbell_interrupt)
6036		adev->nbio.funcs->enable_doorbell_interrupt(adev, true);
6037
6038	if (amdgpu_passthrough(adev) &&
6039	    adev->nbio.funcs->clear_doorbell_interrupt)
6040		adev->nbio.funcs->clear_doorbell_interrupt(adev);
6041
6042	return 0;
6043}
6044
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6045/**
6046 * amdgpu_pci_error_detected - Called when a PCI error is detected.
6047 * @pdev: PCI device struct
6048 * @state: PCI channel state
6049 *
6050 * Description: Called when a PCI error is detected.
6051 *
6052 * Return: PCI_ERS_RESULT_NEED_RESET or PCI_ERS_RESULT_DISCONNECT.
6053 */
6054pci_ers_result_t amdgpu_pci_error_detected(struct pci_dev *pdev, pci_channel_state_t state)
6055{
6056	struct drm_device *dev = pci_get_drvdata(pdev);
6057	struct amdgpu_device *adev = drm_to_adev(dev);
6058	int i;
6059
6060	DRM_INFO("PCI error: detected callback, state(%d)!!\n", state);
6061
6062	if (adev->gmc.xgmi.num_physical_nodes > 1) {
6063		DRM_WARN("No support for XGMI hive yet...");
6064		return PCI_ERS_RESULT_DISCONNECT;
6065	}
6066
6067	adev->pci_channel_state = state;
6068
6069	switch (state) {
6070	case pci_channel_io_normal:
6071		return PCI_ERS_RESULT_CAN_RECOVER;
6072	/* Fatal error, prepare for slot reset */
6073	case pci_channel_io_frozen:
6074		/*
6075		 * Locking adev->reset_domain->sem will prevent any external access
 
 
 
6076		 * to GPU during PCI error recovery
6077		 */
6078		amdgpu_device_lock_reset_domain(adev->reset_domain);
6079		amdgpu_device_set_mp1_state(adev);
6080
6081		/*
6082		 * Block any work scheduling as we do for regular GPU reset
6083		 * for the duration of the recovery
6084		 */
6085		for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
6086			struct amdgpu_ring *ring = adev->rings[i];
6087
6088			if (!amdgpu_ring_sched_ready(ring))
6089				continue;
6090
6091			drm_sched_stop(&ring->sched, NULL);
6092		}
6093		atomic_inc(&adev->gpu_reset_counter);
6094		return PCI_ERS_RESULT_NEED_RESET;
6095	case pci_channel_io_perm_failure:
6096		/* Permanent error, prepare for device removal */
6097		return PCI_ERS_RESULT_DISCONNECT;
6098	}
6099
6100	return PCI_ERS_RESULT_NEED_RESET;
6101}
6102
6103/**
6104 * amdgpu_pci_mmio_enabled - Enable MMIO and dump debug registers
6105 * @pdev: pointer to PCI device
6106 */
6107pci_ers_result_t amdgpu_pci_mmio_enabled(struct pci_dev *pdev)
6108{
6109
6110	DRM_INFO("PCI error: mmio enabled callback!!\n");
6111
6112	/* TODO - dump whatever for debugging purposes */
6113
6114	/* This called only if amdgpu_pci_error_detected returns
6115	 * PCI_ERS_RESULT_CAN_RECOVER. Read/write to the device still
6116	 * works, no need to reset slot.
6117	 */
6118
6119	return PCI_ERS_RESULT_RECOVERED;
6120}
6121
6122/**
6123 * amdgpu_pci_slot_reset - Called when PCI slot has been reset.
6124 * @pdev: PCI device struct
6125 *
6126 * Description: This routine is called by the pci error recovery
6127 * code after the PCI slot has been reset, just before we
6128 * should resume normal operations.
6129 */
6130pci_ers_result_t amdgpu_pci_slot_reset(struct pci_dev *pdev)
6131{
6132	struct drm_device *dev = pci_get_drvdata(pdev);
6133	struct amdgpu_device *adev = drm_to_adev(dev);
6134	int r, i;
6135	struct amdgpu_reset_context reset_context;
6136	u32 memsize;
6137	struct list_head device_list;
6138	struct amdgpu_hive_info *hive;
6139	int hive_ras_recovery = 0;
6140	struct amdgpu_ras *ras;
6141
6142	/* PCI error slot reset should be skipped During RAS recovery */
6143	hive = amdgpu_get_xgmi_hive(adev);
6144	if (hive) {
6145		hive_ras_recovery = atomic_read(&hive->ras_recovery);
6146		amdgpu_put_xgmi_hive(hive);
6147	}
6148	ras = amdgpu_ras_get_context(adev);
6149	if ((amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3)) &&
6150		 ras && (atomic_read(&ras->in_recovery) || hive_ras_recovery))
6151		return PCI_ERS_RESULT_RECOVERED;
6152
6153	DRM_INFO("PCI error: slot reset callback!!\n");
6154
6155	memset(&reset_context, 0, sizeof(reset_context));
6156
6157	INIT_LIST_HEAD(&device_list);
6158	list_add_tail(&adev->reset_list, &device_list);
6159
6160	/* wait for asic to come out of reset */
6161	msleep(500);
6162
6163	/* Restore PCI confspace */
6164	amdgpu_device_load_pci_state(pdev);
6165
6166	/* confirm  ASIC came out of reset */
6167	for (i = 0; i < adev->usec_timeout; i++) {
6168		memsize = amdgpu_asic_get_config_memsize(adev);
6169
6170		if (memsize != 0xffffffff)
6171			break;
6172		udelay(1);
6173	}
6174	if (memsize == 0xffffffff) {
6175		r = -ETIME;
6176		goto out;
6177	}
6178
6179	reset_context.method = AMD_RESET_METHOD_NONE;
6180	reset_context.reset_req_dev = adev;
6181	set_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
6182	set_bit(AMDGPU_SKIP_HW_RESET, &reset_context.flags);
6183
6184	adev->no_hw_access = true;
6185	r = amdgpu_device_pre_asic_reset(adev, &reset_context);
6186	adev->no_hw_access = false;
6187	if (r)
6188		goto out;
6189
6190	r = amdgpu_do_asic_reset(&device_list, &reset_context);
6191
6192out:
6193	if (!r) {
6194		if (amdgpu_device_cache_pci_state(adev->pdev))
6195			pci_restore_state(adev->pdev);
6196
6197		DRM_INFO("PCIe error recovery succeeded\n");
6198	} else {
6199		DRM_ERROR("PCIe error recovery failed, err:%d", r);
6200		amdgpu_device_unset_mp1_state(adev);
6201		amdgpu_device_unlock_reset_domain(adev->reset_domain);
6202	}
6203
6204	return r ? PCI_ERS_RESULT_DISCONNECT : PCI_ERS_RESULT_RECOVERED;
6205}
6206
6207/**
6208 * amdgpu_pci_resume() - resume normal ops after PCI reset
6209 * @pdev: pointer to PCI device
6210 *
6211 * Called when the error recovery driver tells us that its
6212 * OK to resume normal operation.
6213 */
6214void amdgpu_pci_resume(struct pci_dev *pdev)
6215{
6216	struct drm_device *dev = pci_get_drvdata(pdev);
6217	struct amdgpu_device *adev = drm_to_adev(dev);
6218	int i;
6219
6220
6221	DRM_INFO("PCI error: resume callback!!\n");
6222
6223	/* Only continue execution for the case of pci_channel_io_frozen */
6224	if (adev->pci_channel_state != pci_channel_io_frozen)
6225		return;
6226
6227	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
6228		struct amdgpu_ring *ring = adev->rings[i];
6229
6230		if (!amdgpu_ring_sched_ready(ring))
6231			continue;
6232
 
 
6233		drm_sched_start(&ring->sched, true);
6234	}
6235
6236	amdgpu_device_unset_mp1_state(adev);
6237	amdgpu_device_unlock_reset_domain(adev->reset_domain);
6238}
6239
6240bool amdgpu_device_cache_pci_state(struct pci_dev *pdev)
6241{
6242	struct drm_device *dev = pci_get_drvdata(pdev);
6243	struct amdgpu_device *adev = drm_to_adev(dev);
6244	int r;
6245
6246	r = pci_save_state(pdev);
6247	if (!r) {
6248		kfree(adev->pci_state);
6249
6250		adev->pci_state = pci_store_saved_state(pdev);
6251
6252		if (!adev->pci_state) {
6253			DRM_ERROR("Failed to store PCI saved state");
6254			return false;
6255		}
6256	} else {
6257		DRM_WARN("Failed to save PCI state, err:%d\n", r);
6258		return false;
6259	}
6260
6261	return true;
6262}
6263
6264bool amdgpu_device_load_pci_state(struct pci_dev *pdev)
6265{
6266	struct drm_device *dev = pci_get_drvdata(pdev);
6267	struct amdgpu_device *adev = drm_to_adev(dev);
6268	int r;
6269
6270	if (!adev->pci_state)
6271		return false;
6272
6273	r = pci_load_saved_state(pdev, adev->pci_state);
6274
6275	if (!r) {
6276		pci_restore_state(pdev);
6277	} else {
6278		DRM_WARN("Failed to load PCI state, err:%d\n", r);
6279		return false;
6280	}
6281
6282	return true;
6283}
6284
6285void amdgpu_device_flush_hdp(struct amdgpu_device *adev,
6286		struct amdgpu_ring *ring)
6287{
6288#ifdef CONFIG_X86_64
6289	if ((adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev))
6290		return;
6291#endif
6292	if (adev->gmc.xgmi.connected_to_cpu)
6293		return;
6294
6295	if (ring && ring->funcs->emit_hdp_flush)
6296		amdgpu_ring_emit_hdp_flush(ring);
6297	else
6298		amdgpu_asic_flush_hdp(adev, ring);
6299}
6300
6301void amdgpu_device_invalidate_hdp(struct amdgpu_device *adev,
6302		struct amdgpu_ring *ring)
6303{
6304#ifdef CONFIG_X86_64
6305	if ((adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev))
6306		return;
6307#endif
6308	if (adev->gmc.xgmi.connected_to_cpu)
6309		return;
6310
6311	amdgpu_asic_invalidate_hdp(adev, ring);
6312}
6313
6314int amdgpu_in_reset(struct amdgpu_device *adev)
6315{
6316	return atomic_read(&adev->reset_domain->in_gpu_reset);
6317}
6318
6319/**
6320 * amdgpu_device_halt() - bring hardware to some kind of halt state
6321 *
6322 * @adev: amdgpu_device pointer
6323 *
6324 * Bring hardware to some kind of halt state so that no one can touch it
6325 * any more. It will help to maintain error context when error occurred.
6326 * Compare to a simple hang, the system will keep stable at least for SSH
6327 * access. Then it should be trivial to inspect the hardware state and
6328 * see what's going on. Implemented as following:
6329 *
6330 * 1. drm_dev_unplug() makes device inaccessible to user space(IOCTLs, etc),
6331 *    clears all CPU mappings to device, disallows remappings through page faults
6332 * 2. amdgpu_irq_disable_all() disables all interrupts
6333 * 3. amdgpu_fence_driver_hw_fini() signals all HW fences
6334 * 4. set adev->no_hw_access to avoid potential crashes after setp 5
6335 * 5. amdgpu_device_unmap_mmio() clears all MMIO mappings
6336 * 6. pci_disable_device() and pci_wait_for_pending_transaction()
6337 *    flush any in flight DMA operations
6338 */
6339void amdgpu_device_halt(struct amdgpu_device *adev)
6340{
6341	struct pci_dev *pdev = adev->pdev;
6342	struct drm_device *ddev = adev_to_drm(adev);
6343
6344	amdgpu_xcp_dev_unplug(adev);
6345	drm_dev_unplug(ddev);
6346
6347	amdgpu_irq_disable_all(adev);
6348
6349	amdgpu_fence_driver_hw_fini(adev);
6350
6351	adev->no_hw_access = true;
6352
6353	amdgpu_device_unmap_mmio(adev);
6354
6355	pci_disable_device(pdev);
6356	pci_wait_for_pending_transaction(pdev);
6357}
6358
6359u32 amdgpu_device_pcie_port_rreg(struct amdgpu_device *adev,
6360				u32 reg)
6361{
6362	unsigned long flags, address, data;
6363	u32 r;
6364
6365	address = adev->nbio.funcs->get_pcie_port_index_offset(adev);
6366	data = adev->nbio.funcs->get_pcie_port_data_offset(adev);
6367
6368	spin_lock_irqsave(&adev->pcie_idx_lock, flags);
6369	WREG32(address, reg * 4);
6370	(void)RREG32(address);
6371	r = RREG32(data);
6372	spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
6373	return r;
6374}
6375
6376void amdgpu_device_pcie_port_wreg(struct amdgpu_device *adev,
6377				u32 reg, u32 v)
6378{
6379	unsigned long flags, address, data;
6380
6381	address = adev->nbio.funcs->get_pcie_port_index_offset(adev);
6382	data = adev->nbio.funcs->get_pcie_port_data_offset(adev);
6383
6384	spin_lock_irqsave(&adev->pcie_idx_lock, flags);
6385	WREG32(address, reg * 4);
6386	(void)RREG32(address);
6387	WREG32(data, v);
6388	(void)RREG32(data);
6389	spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
6390}
6391
6392/**
6393 * amdgpu_device_switch_gang - switch to a new gang
6394 * @adev: amdgpu_device pointer
6395 * @gang: the gang to switch to
6396 *
6397 * Try to switch to a new gang.
6398 * Returns: NULL if we switched to the new gang or a reference to the current
6399 * gang leader.
6400 */
6401struct dma_fence *amdgpu_device_switch_gang(struct amdgpu_device *adev,
6402					    struct dma_fence *gang)
6403{
6404	struct dma_fence *old = NULL;
6405
6406	do {
6407		dma_fence_put(old);
6408		rcu_read_lock();
6409		old = dma_fence_get_rcu_safe(&adev->gang_submit);
6410		rcu_read_unlock();
6411
6412		if (old == gang)
6413			break;
6414
6415		if (!dma_fence_is_signaled(old))
6416			return old;
6417
6418	} while (cmpxchg((struct dma_fence __force **)&adev->gang_submit,
6419			 old, gang) != old);
6420
6421	dma_fence_put(old);
6422	return NULL;
6423}
6424
6425bool amdgpu_device_has_display_hardware(struct amdgpu_device *adev)
6426{
6427	switch (adev->asic_type) {
6428#ifdef CONFIG_DRM_AMDGPU_SI
6429	case CHIP_HAINAN:
6430#endif
6431	case CHIP_TOPAZ:
6432		/* chips with no display hardware */
6433		return false;
6434#ifdef CONFIG_DRM_AMDGPU_SI
6435	case CHIP_TAHITI:
6436	case CHIP_PITCAIRN:
6437	case CHIP_VERDE:
6438	case CHIP_OLAND:
6439#endif
6440#ifdef CONFIG_DRM_AMDGPU_CIK
6441	case CHIP_BONAIRE:
6442	case CHIP_HAWAII:
6443	case CHIP_KAVERI:
6444	case CHIP_KABINI:
6445	case CHIP_MULLINS:
6446#endif
6447	case CHIP_TONGA:
6448	case CHIP_FIJI:
6449	case CHIP_POLARIS10:
6450	case CHIP_POLARIS11:
6451	case CHIP_POLARIS12:
6452	case CHIP_VEGAM:
6453	case CHIP_CARRIZO:
6454	case CHIP_STONEY:
6455		/* chips with display hardware */
6456		return true;
6457	default:
6458		/* IP discovery */
6459		if (!amdgpu_ip_version(adev, DCE_HWIP, 0) ||
6460		    (adev->harvest_ip_mask & AMD_HARVEST_IP_DMU_MASK))
6461			return false;
6462		return true;
6463	}
6464}
6465
6466uint32_t amdgpu_device_wait_on_rreg(struct amdgpu_device *adev,
6467		uint32_t inst, uint32_t reg_addr, char reg_name[],
6468		uint32_t expected_value, uint32_t mask)
6469{
6470	uint32_t ret = 0;
6471	uint32_t old_ = 0;
6472	uint32_t tmp_ = RREG32(reg_addr);
6473	uint32_t loop = adev->usec_timeout;
6474
6475	while ((tmp_ & (mask)) != (expected_value)) {
6476		if (old_ != tmp_) {
6477			loop = adev->usec_timeout;
6478			old_ = tmp_;
6479		} else
6480			udelay(1);
6481		tmp_ = RREG32(reg_addr);
6482		loop--;
6483		if (!loop) {
6484			DRM_WARN("Register(%d) [%s] failed to reach value 0x%08x != 0x%08xn",
6485				  inst, reg_name, (uint32_t)expected_value,
6486				  (uint32_t)(tmp_ & (mask)));
6487			ret = -ETIMEDOUT;
6488			break;
6489		}
6490	}
6491	return ret;
6492}
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}