drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c

   1 /*
   2  * Copyright 2016 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  *
  22  */
  23
  24 #include <linux/firmware.h>
  25 #include <linux/pci.h>
  26
  27 #include <drm/drm_cache.h>
  28
  29 #include "amdgpu.h"
  30 #include "gmc_v9_0.h"
  31 #include "amdgpu_atomfirmware.h"
  32 #include "amdgpu_gem.h"
  33
  34 #include "gc/gc_9_0_sh_mask.h"
  35 #include "dce/dce_12_0_offset.h"
  36 #include "dce/dce_12_0_sh_mask.h"
  37 #include "vega10_enum.h"
  38 #include "mmhub/mmhub_1_0_offset.h"
  39 #include "athub/athub_1_0_sh_mask.h"
  40 #include "athub/athub_1_0_offset.h"
  41 #include "oss/osssys_4_0_offset.h"
  42
  43 #include "soc15.h"
  44 #include "soc15d.h"
  45 #include "soc15_common.h"
  46 #include "umc/umc_6_0_sh_mask.h"
  47
  48 #include "gfxhub_v1_0.h"
  49 #include "mmhub_v1_0.h"
  50 #include "athub_v1_0.h"
  51 #include "gfxhub_v1_1.h"
  52 #include "mmhub_v9_4.h"
  53 #include "mmhub_v1_7.h"
  54 #include "umc_v6_1.h"
  55 #include "umc_v6_0.h"
  56
  57 #include "ivsrcid/vmc/irqsrcs_vmc_1_0.h"
  58
  59 #include "amdgpu_ras.h"
  60 #include "amdgpu_xgmi.h"
  61
  62 /* add these here since we already include dce12 headers and these are for DCN */
  63 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION                                                          0x055d
  64 #define mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION_BASE_IDX                                                 2
  65 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_WIDTH__SHIFT                                        0x0
  66 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_HEIGHT__SHIFT                                       0x10
  67 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_WIDTH_MASK                                          0x00003FFFL
  68 #define HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION__PRI_VIEWPORT_HEIGHT_MASK                                         0x3FFF0000L
  69 #define mmDCHUBBUB_SDPIF_MMIO_CNTRL_0                                                                  0x049d
  70 #define mmDCHUBBUB_SDPIF_MMIO_CNTRL_0_BASE_IDX                                                         2
  71
  72
  73 static const char *gfxhub_client_ids[] = {
  74         "CB",
  75         "DB",
  76         "IA",
  77         "WD",
  78         "CPF",
  79         "CPC",
  80         "CPG",
  81         "RLC",
  82         "TCP",
  83         "SQC (inst)",
  84         "SQC (data)",
  85         "SQG",
  86         "PA",
  87 };
  88
  89 static const char *mmhub_client_ids_raven[][2] = {
  90         [0][0] = "MP1",
  91         [1][0] = "MP0",
  92         [2][0] = "VCN",
  93         [3][0] = "VCNU",
  94         [4][0] = "HDP",
  95         [5][0] = "DCE",
  96         [13][0] = "UTCL2",
  97         [19][0] = "TLS",
  98         [26][0] = "OSS",
  99         [27][0] = "SDMA0",
 100         [0][1] = "MP1",
 101         [1][1] = "MP0",
 102         [2][1] = "VCN",
 103         [3][1] = "VCNU",
 104         [4][1] = "HDP",
 105         [5][1] = "XDP",
 106         [6][1] = "DBGU0",
 107         [7][1] = "DCE",
 108         [8][1] = "DCEDWB0",
 109         [9][1] = "DCEDWB1",
 110         [26][1] = "OSS",
 111         [27][1] = "SDMA0",
 112 };
 113
 114 static const char *mmhub_client_ids_renoir[][2] = {
 115         [0][0] = "MP1",
 116         [1][0] = "MP0",
 117         [2][0] = "HDP",
 118         [4][0] = "DCEDMC",
 119         [5][0] = "DCEVGA",
 120         [13][0] = "UTCL2",
 121         [19][0] = "TLS",
 122         [26][0] = "OSS",
 123         [27][0] = "SDMA0",
 124         [28][0] = "VCN",
 125         [29][0] = "VCNU",
 126         [30][0] = "JPEG",
 127         [0][1] = "MP1",
 128         [1][1] = "MP0",
 129         [2][1] = "HDP",
 130         [3][1] = "XDP",
 131         [6][1] = "DBGU0",
 132         [7][1] = "DCEDMC",
 133         [8][1] = "DCEVGA",
 134         [9][1] = "DCEDWB",
 135         [26][1] = "OSS",
 136         [27][1] = "SDMA0",
 137         [28][1] = "VCN",
 138         [29][1] = "VCNU",
 139         [30][1] = "JPEG",
 140 };
 141
 142 static const char *mmhub_client_ids_vega10[][2] = {
 143         [0][0] = "MP0",
 144         [1][0] = "UVD",
 145         [2][0] = "UVDU",
 146         [3][0] = "HDP",
 147         [13][0] = "UTCL2",
 148         [14][0] = "OSS",
 149         [15][0] = "SDMA1",
 150         [32+0][0] = "VCE0",
 151         [32+1][0] = "VCE0U",
 152         [32+2][0] = "XDMA",
 153         [32+3][0] = "DCE",
 154         [32+4][0] = "MP1",
 155         [32+14][0] = "SDMA0",
 156         [0][1] = "MP0",
 157         [1][1] = "UVD",
 158         [2][1] = "UVDU",
 159         [3][1] = "DBGU0",
 160         [4][1] = "HDP",
 161         [5][1] = "XDP",
 162         [14][1] = "OSS",
 163         [15][1] = "SDMA0",
 164         [32+0][1] = "VCE0",
 165         [32+1][1] = "VCE0U",
 166         [32+2][1] = "XDMA",
 167         [32+3][1] = "DCE",
 168         [32+4][1] = "DCEDWB",
 169         [32+5][1] = "MP1",
 170         [32+6][1] = "DBGU1",
 171         [32+14][1] = "SDMA1",
 172 };
 173
 174 static const char *mmhub_client_ids_vega12[][2] = {
 175         [0][0] = "MP0",
 176         [1][0] = "VCE0",
 177         [2][0] = "VCE0U",
 178         [3][0] = "HDP",
 179         [13][0] = "UTCL2",
 180         [14][0] = "OSS",
 181         [15][0] = "SDMA1",
 182         [32+0][0] = "DCE",
 183         [32+1][0] = "XDMA",
 184         [32+2][0] = "UVD",
 185         [32+3][0] = "UVDU",
 186         [32+4][0] = "MP1",
 187         [32+15][0] = "SDMA0",
 188         [0][1] = "MP0",
 189         [1][1] = "VCE0",
 190         [2][1] = "VCE0U",
 191         [3][1] = "DBGU0",
 192         [4][1] = "HDP",
 193         [5][1] = "XDP",
 194         [14][1] = "OSS",
 195         [15][1] = "SDMA0",
 196         [32+0][1] = "DCE",
 197         [32+1][1] = "DCEDWB",
 198         [32+2][1] = "XDMA",
 199         [32+3][1] = "UVD",
 200         [32+4][1] = "UVDU",
 201         [32+5][1] = "MP1",
 202         [32+6][1] = "DBGU1",
 203         [32+15][1] = "SDMA1",
 204 };
 205
 206 static const char *mmhub_client_ids_vega20[][2] = {
 207         [0][0] = "XDMA",
 208         [1][0] = "DCE",
 209         [2][0] = "VCE0",
 210         [3][0] = "VCE0U",
 211         [4][0] = "UVD",
 212         [5][0] = "UVD1U",
 213         [13][0] = "OSS",
 214         [14][0] = "HDP",
 215         [15][0] = "SDMA0",
 216         [32+0][0] = "UVD",
 217         [32+1][0] = "UVDU",
 218         [32+2][0] = "MP1",
 219         [32+3][0] = "MP0",
 220         [32+12][0] = "UTCL2",
 221         [32+14][0] = "SDMA1",
 222         [0][1] = "XDMA",
 223         [1][1] = "DCE",
 224         [2][1] = "DCEDWB",
 225         [3][1] = "VCE0",
 226         [4][1] = "VCE0U",
 227         [5][1] = "UVD1",
 228         [6][1] = "UVD1U",
 229         [7][1] = "DBGU0",
 230         [8][1] = "XDP",
 231         [13][1] = "OSS",
 232         [14][1] = "HDP",
 233         [15][1] = "SDMA0",
 234         [32+0][1] = "UVD",
 235         [32+1][1] = "UVDU",
 236         [32+2][1] = "DBGU1",
 237         [32+3][1] = "MP1",
 238         [32+4][1] = "MP0",
 239         [32+14][1] = "SDMA1",
 240 };
 241
 242 static const char *mmhub_client_ids_arcturus[][2] = {
 243         [0][0] = "DBGU1",
 244         [1][0] = "XDP",
 245         [2][0] = "MP1",
 246         [14][0] = "HDP",
 247         [171][0] = "JPEG",
 248         [172][0] = "VCN",
 249         [173][0] = "VCNU",
 250         [203][0] = "JPEG1",
 251         [204][0] = "VCN1",
 252         [205][0] = "VCN1U",
 253         [256][0] = "SDMA0",
 254         [257][0] = "SDMA1",
 255         [258][0] = "SDMA2",
 256         [259][0] = "SDMA3",
 257         [260][0] = "SDMA4",
 258         [261][0] = "SDMA5",
 259         [262][0] = "SDMA6",
 260         [263][0] = "SDMA7",
 261         [384][0] = "OSS",
 262         [0][1] = "DBGU1",
 263         [1][1] = "XDP",
 264         [2][1] = "MP1",
 265         [14][1] = "HDP",
 266         [171][1] = "JPEG",
 267         [172][1] = "VCN",
 268         [173][1] = "VCNU",
 269         [203][1] = "JPEG1",
 270         [204][1] = "VCN1",
 271         [205][1] = "VCN1U",
 272         [256][1] = "SDMA0",
 273         [257][1] = "SDMA1",
 274         [258][1] = "SDMA2",
 275         [259][1] = "SDMA3",
 276         [260][1] = "SDMA4",
 277         [261][1] = "SDMA5",
 278         [262][1] = "SDMA6",
 279         [263][1] = "SDMA7",
 280         [384][1] = "OSS",
 281 };
 282
 283 static const char *mmhub_client_ids_aldebaran[][2] = {
 284         [2][0] = "MP1",
 285         [3][0] = "MP0",
 286         [32+1][0] = "DBGU_IO0",
 287         [32+2][0] = "DBGU_IO2",
 288         [32+4][0] = "MPIO",
 289         [96+11][0] = "JPEG0",
 290         [96+12][0] = "VCN0",
 291         [96+13][0] = "VCNU0",
 292         [128+11][0] = "JPEG1",
 293         [128+12][0] = "VCN1",
 294         [128+13][0] = "VCNU1",
 295         [160+1][0] = "XDP",
 296         [160+14][0] = "HDP",
 297         [256+0][0] = "SDMA0",
 298         [256+1][0] = "SDMA1",
 299         [256+2][0] = "SDMA2",
 300         [256+3][0] = "SDMA3",
 301         [256+4][0] = "SDMA4",
 302         [384+0][0] = "OSS",
 303         [2][1] = "MP1",
 304         [3][1] = "MP0",
 305         [32+1][1] = "DBGU_IO0",
 306         [32+2][1] = "DBGU_IO2",
 307         [32+4][1] = "MPIO",
 308         [96+11][1] = "JPEG0",
 309         [96+12][1] = "VCN0",
 310         [96+13][1] = "VCNU0",
 311         [128+11][1] = "JPEG1",
 312         [128+12][1] = "VCN1",
 313         [128+13][1] = "VCNU1",
 314         [160+1][1] = "XDP",
 315         [160+14][1] = "HDP",
 316         [256+0][1] = "SDMA0",
 317         [256+1][1] = "SDMA1",
 318         [256+2][1] = "SDMA2",
 319         [256+3][1] = "SDMA3",
 320         [256+4][1] = "SDMA4",
 321         [384+0][1] = "OSS",
 322 };
 323
 324 static const struct soc15_reg_golden golden_settings_mmhub_1_0_0[] =
 325 {
 326         SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmDAGB1_WRCLI2, 0x00000007, 0xfe5fe0fa),
 327         SOC15_REG_GOLDEN_VALUE(MMHUB, 0, mmMMEA1_DRAM_WR_CLI2GRP_MAP0, 0x00000030, 0x55555565)
 328 };
 329
 330 static const struct soc15_reg_golden golden_settings_athub_1_0_0[] =
 331 {
 332         SOC15_REG_GOLDEN_VALUE(ATHUB, 0, mmRPB_ARB_CNTL, 0x0000ff00, 0x00000800),
 333         SOC15_REG_GOLDEN_VALUE(ATHUB, 0, mmRPB_ARB_CNTL2, 0x00ff00ff, 0x00080008)
 334 };
 335
 336 static const uint32_t ecc_umc_mcumc_ctrl_addrs[] = {
 337         (0x000143c0 + 0x00000000),
 338         (0x000143c0 + 0x00000800),
 339         (0x000143c0 + 0x00001000),
 340         (0x000143c0 + 0x00001800),
 341         (0x000543c0 + 0x00000000),
 342         (0x000543c0 + 0x00000800),
 343         (0x000543c0 + 0x00001000),
 344         (0x000543c0 + 0x00001800),
 345         (0x000943c0 + 0x00000000),
 346         (0x000943c0 + 0x00000800),
 347         (0x000943c0 + 0x00001000),
 348         (0x000943c0 + 0x00001800),
 349         (0x000d43c0 + 0x00000000),
 350         (0x000d43c0 + 0x00000800),
 351         (0x000d43c0 + 0x00001000),
 352         (0x000d43c0 + 0x00001800),
 353         (0x001143c0 + 0x00000000),
 354         (0x001143c0 + 0x00000800),
 355         (0x001143c0 + 0x00001000),
 356         (0x001143c0 + 0x00001800),
 357         (0x001543c0 + 0x00000000),
 358         (0x001543c0 + 0x00000800),
 359         (0x001543c0 + 0x00001000),
 360         (0x001543c0 + 0x00001800),
 361         (0x001943c0 + 0x00000000),
 362         (0x001943c0 + 0x00000800),
 363         (0x001943c0 + 0x00001000),
 364         (0x001943c0 + 0x00001800),
 365         (0x001d43c0 + 0x00000000),
 366         (0x001d43c0 + 0x00000800),
 367         (0x001d43c0 + 0x00001000),
 368         (0x001d43c0 + 0x00001800),
 369 };
 370
 371 static const uint32_t ecc_umc_mcumc_ctrl_mask_addrs[] = {
 372         (0x000143e0 + 0x00000000),
 373         (0x000143e0 + 0x00000800),
 374         (0x000143e0 + 0x00001000),
 375         (0x000143e0 + 0x00001800),
 376         (0x000543e0 + 0x00000000),
 377         (0x000543e0 + 0x00000800),
 378         (0x000543e0 + 0x00001000),
 379         (0x000543e0 + 0x00001800),
 380         (0x000943e0 + 0x00000000),
 381         (0x000943e0 + 0x00000800),
 382         (0x000943e0 + 0x00001000),
 383         (0x000943e0 + 0x00001800),
 384         (0x000d43e0 + 0x00000000),
 385         (0x000d43e0 + 0x00000800),
 386         (0x000d43e0 + 0x00001000),
 387         (0x000d43e0 + 0x00001800),
 388         (0x001143e0 + 0x00000000),
 389         (0x001143e0 + 0x00000800),
 390         (0x001143e0 + 0x00001000),
 391         (0x001143e0 + 0x00001800),
 392         (0x001543e0 + 0x00000000),
 393         (0x001543e0 + 0x00000800),
 394         (0x001543e0 + 0x00001000),
 395         (0x001543e0 + 0x00001800),
 396         (0x001943e0 + 0x00000000),
 397         (0x001943e0 + 0x00000800),
 398         (0x001943e0 + 0x00001000),
 399         (0x001943e0 + 0x00001800),
 400         (0x001d43e0 + 0x00000000),
 401         (0x001d43e0 + 0x00000800),
 402         (0x001d43e0 + 0x00001000),
 403         (0x001d43e0 + 0x00001800),
 404 };
 405
 406 static int gmc_v9_0_ecc_interrupt_state(struct amdgpu_device *adev,
 407                 struct amdgpu_irq_src *src,
 408                 unsigned type,
 409                 enum amdgpu_interrupt_state state)
 410 {
 411         u32 bits, i, tmp, reg;
 412
 413         /* Devices newer then VEGA10/12 shall have these programming
 414              sequences performed by PSP BL */
 415         if (adev->asic_type >= CHIP_VEGA20)
 416                 return 0;
 417
 418         bits = 0x7f;
 419
 420         switch (state) {
 421         case AMDGPU_IRQ_STATE_DISABLE:
 422                 for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_addrs); i++) {
 423                         reg = ecc_umc_mcumc_ctrl_addrs[i];
 424                         tmp = RREG32(reg);
 425                         tmp &= ~bits;
 426                         WREG32(reg, tmp);
 427                 }
 428                 for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_mask_addrs); i++) {
 429                         reg = ecc_umc_mcumc_ctrl_mask_addrs[i];
 430                         tmp = RREG32(reg);
 431                         tmp &= ~bits;
 432                         WREG32(reg, tmp);
 433                 }
 434                 break;
 435         case AMDGPU_IRQ_STATE_ENABLE:
 436                 for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_addrs); i++) {
 437                         reg = ecc_umc_mcumc_ctrl_addrs[i];
 438                         tmp = RREG32(reg);
 439                         tmp |= bits;
 440                         WREG32(reg, tmp);
 441                 }
 442                 for (i = 0; i < ARRAY_SIZE(ecc_umc_mcumc_ctrl_mask_addrs); i++) {
 443                         reg = ecc_umc_mcumc_ctrl_mask_addrs[i];
 444                         tmp = RREG32(reg);
 445                         tmp |= bits;
 446                         WREG32(reg, tmp);
 447                 }
 448                 break;
 449         default:
 450                 break;
 451         }
 452
 453         return 0;
 454 }
 455
 456 static int gmc_v9_0_vm_fault_interrupt_state(struct amdgpu_device *adev,
 457                                         struct amdgpu_irq_src *src,
 458                                         unsigned type,
 459                                         enum amdgpu_interrupt_state state)
 460 {
 461         struct amdgpu_vmhub *hub;
 462         u32 tmp, reg, bits, i, j;
 463
 464         bits = VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
 465                 VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
 466                 VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
 467                 VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
 468                 VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
 469                 VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
 470                 VM_CONTEXT1_CNTL__EXECUTE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK;
 471
 472         switch (state) {
 473         case AMDGPU_IRQ_STATE_DISABLE:
 474                 for (j = 0; j < adev->num_vmhubs; j++) {
 475                         hub = &adev->vmhub[j];
 476                         for (i = 0; i < 16; i++) {
 477                                 reg = hub->vm_context0_cntl + i;
 478                                 tmp = RREG32(reg);
 479                                 tmp &= ~bits;
 480                                 WREG32(reg, tmp);
 481                         }
 482                 }
 483                 break;
 484         case AMDGPU_IRQ_STATE_ENABLE:
 485                 for (j = 0; j < adev->num_vmhubs; j++) {
 486                         hub = &adev->vmhub[j];
 487                         for (i = 0; i < 16; i++) {
 488                                 reg = hub->vm_context0_cntl + i;
 489                                 tmp = RREG32(reg);
 490                                 tmp |= bits;
 491                                 WREG32(reg, tmp);
 492                         }
 493                 }
 494                 break;
 495         default:
 496                 break;
 497         }
 498
 499         return 0;
 500 }
 501
 502 static int gmc_v9_0_process_interrupt(struct amdgpu_device *adev,
 503                                       struct amdgpu_irq_src *source,
 504                                       struct amdgpu_iv_entry *entry)
 505 {
 506         bool retry_fault = !!(entry->src_data[1] & 0x80);
 507         uint32_t status = 0, cid = 0, rw = 0;
 508         struct amdgpu_task_info task_info;
 509         struct amdgpu_vmhub *hub;
 510         const char *mmhub_cid;
 511         const char *hub_name;
 512         u64 addr;
 513
 514         addr = (u64)entry->src_data[0] << 12;
 515         addr |= ((u64)entry->src_data[1] & 0xf) << 44;
 516
 517         if (retry_fault) {
 518                 /* Returning 1 here also prevents sending the IV to the KFD */
 519
 520                 /* Process it onyl if it's the first fault for this address */
 521                 if (entry->ih != &adev->irq.ih_soft &&
 522                     amdgpu_gmc_filter_faults(adev, addr, entry->pasid,
 523                                              entry->timestamp))
 524                         return 1;
 525
 526                 /* Delegate it to a different ring if the hardware hasn't
 527                  * already done it.
 528                  */
 529                 if (entry->ih == &adev->irq.ih) {
 530                         amdgpu_irq_delegate(adev, entry, 8);
 531                         return 1;
 532                 }
 533
 534                 /* Try to handle the recoverable page faults by filling page
 535                  * tables
 536                  */
 537                 if (amdgpu_vm_handle_fault(adev, entry->pasid, addr))
 538                         return 1;
 539         }
 540
 541         if (!printk_ratelimit())
 542                 return 0;
 543
 544         if (entry->client_id == SOC15_IH_CLIENTID_VMC) {
 545                 hub_name = "mmhub0";
 546                 hub = &adev->vmhub[AMDGPU_MMHUB_0];
 547         } else if (entry->client_id == SOC15_IH_CLIENTID_VMC1) {
 548                 hub_name = "mmhub1";
 549                 hub = &adev->vmhub[AMDGPU_MMHUB_1];
 550         } else {
 551                 hub_name = "gfxhub0";
 552                 hub = &adev->vmhub[AMDGPU_GFXHUB_0];
 553         }
 554
 555         memset(&task_info, 0, sizeof(struct amdgpu_task_info));
 556         amdgpu_vm_get_task_info(adev, entry->pasid, &task_info);
 557
 558         dev_err(adev->dev,
 559                 "[%s] %s page fault (src_id:%u ring:%u vmid:%u "
 560                 "pasid:%u, for process %s pid %d thread %s pid %d)\n",
 561                 hub_name, retry_fault ? "retry" : "no-retry",
 562                 entry->src_id, entry->ring_id, entry->vmid,
 563                 entry->pasid, task_info.process_name, task_info.tgid,
 564                 task_info.task_name, task_info.pid);
 565         dev_err(adev->dev, "  in page starting at address 0x%016llx from IH client 0x%x (%s)\n",
 566                 addr, entry->client_id,
 567                 soc15_ih_clientid_name[entry->client_id]);
 568
 569         if (amdgpu_sriov_vf(adev))
 570                 return 0;
 571
 572         /*
 573          * Issue a dummy read to wait for the status register to
 574          * be updated to avoid reading an incorrect value due to
 575          * the new fast GRBM interface.
 576          */
 577         if (entry->vmid_src == AMDGPU_GFXHUB_0)
 578                 RREG32(hub->vm_l2_pro_fault_status);
 579
 580         status = RREG32(hub->vm_l2_pro_fault_status);
 581         cid = REG_GET_FIELD(status, VM_L2_PROTECTION_FAULT_STATUS, CID);
 582         rw = REG_GET_FIELD(status, VM_L2_PROTECTION_FAULT_STATUS, RW);
 583         WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1);
 584
 585
 586         dev_err(adev->dev,
 587                 "VM_L2_PROTECTION_FAULT_STATUS:0x%08X\n",
 588                 status);
 589         if (hub == &adev->vmhub[AMDGPU_GFXHUB_0]) {
 590                 dev_err(adev->dev, "\t Faulty UTCL2 client ID: %s (0x%x)\n",
 591                         cid >= ARRAY_SIZE(gfxhub_client_ids) ? "unknown" :
 592                         gfxhub_client_ids[cid],
 593                         cid);
 594         } else {
 595                 switch (adev->asic_type) {
 596                 case CHIP_VEGA10:
 597                         mmhub_cid = mmhub_client_ids_vega10[cid][rw];
 598                         break;
 599                 case CHIP_VEGA12:
 600                         mmhub_cid = mmhub_client_ids_vega12[cid][rw];
 601                         break;
 602                 case CHIP_VEGA20:
 603                         mmhub_cid = mmhub_client_ids_vega20[cid][rw];
 604                         break;
 605                 case CHIP_ARCTURUS:
 606                         mmhub_cid = mmhub_client_ids_arcturus[cid][rw];
 607                         break;
 608                 case CHIP_RAVEN:
 609                         mmhub_cid = mmhub_client_ids_raven[cid][rw];
 610                         break;
 611                 case CHIP_RENOIR:
 612                         mmhub_cid = mmhub_client_ids_renoir[cid][rw];
 613                         break;
 614                 case CHIP_ALDEBARAN:
 615                         mmhub_cid = mmhub_client_ids_aldebaran[cid][rw];
 616                         break;
 617                 default:
 618                         mmhub_cid = NULL;
 619                         break;
 620                 }
 621                 dev_err(adev->dev, "\t Faulty UTCL2 client ID: %s (0x%x)\n",
 622                         mmhub_cid ? mmhub_cid : "unknown", cid);
 623         }
 624         dev_err(adev->dev, "\t MORE_FAULTS: 0x%lx\n",
 625                 REG_GET_FIELD(status,
 626                 VM_L2_PROTECTION_FAULT_STATUS, MORE_FAULTS));
 627         dev_err(adev->dev, "\t WALKER_ERROR: 0x%lx\n",
 628                 REG_GET_FIELD(status,
 629                 VM_L2_PROTECTION_FAULT_STATUS, WALKER_ERROR));
 630         dev_err(adev->dev, "\t PERMISSION_FAULTS: 0x%lx\n",
 631                 REG_GET_FIELD(status,
 632                 VM_L2_PROTECTION_FAULT_STATUS, PERMISSION_FAULTS));
 633         dev_err(adev->dev, "\t MAPPING_ERROR: 0x%lx\n",
 634                 REG_GET_FIELD(status,
 635                 VM_L2_PROTECTION_FAULT_STATUS, MAPPING_ERROR));
 636         dev_err(adev->dev, "\t RW: 0x%x\n", rw);
 637         return 0;
 638 }
 639
 640 static const struct amdgpu_irq_src_funcs gmc_v9_0_irq_funcs = {
 641         .set = gmc_v9_0_vm_fault_interrupt_state,
 642         .process = gmc_v9_0_process_interrupt,
 643 };
 644
 645
 646 static const struct amdgpu_irq_src_funcs gmc_v9_0_ecc_funcs = {
 647         .set = gmc_v9_0_ecc_interrupt_state,
 648         .process = amdgpu_umc_process_ecc_irq,
 649 };
 650
 651 static void gmc_v9_0_set_irq_funcs(struct amdgpu_device *adev)
 652 {
 653         adev->gmc.vm_fault.num_types = 1;
 654         adev->gmc.vm_fault.funcs = &gmc_v9_0_irq_funcs;
 655
 656         if (!amdgpu_sriov_vf(adev)) {
 657                 adev->gmc.ecc_irq.num_types = 1;
 658                 adev->gmc.ecc_irq.funcs = &gmc_v9_0_ecc_funcs;
 659         }
 660 }
 661
 662 static uint32_t gmc_v9_0_get_invalidate_req(unsigned int vmid,
 663                                         uint32_t flush_type)
 664 {
 665         u32 req = 0;
 666
 667         req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ,
 668                             PER_VMID_INVALIDATE_REQ, 1 << vmid);
 669         req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, FLUSH_TYPE, flush_type);
 670         req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PTES, 1);
 671         req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE0, 1);
 672         req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE1, 1);
 673         req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE2, 1);
 674         req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ, INVALIDATE_L1_PTES, 1);
 675         req = REG_SET_FIELD(req, VM_INVALIDATE_ENG0_REQ,
 676                             CLEAR_PROTECTION_FAULT_STATUS_ADDR, 0);
 677
 678         return req;
 679 }
 680
 681 /**
 682  * gmc_v9_0_use_invalidate_semaphore - judge whether to use semaphore
 683  *
 684  * @adev: amdgpu_device pointer
 685  * @vmhub: vmhub type
 686  *
 687  */
 688 static bool gmc_v9_0_use_invalidate_semaphore(struct amdgpu_device *adev,
 689                                        uint32_t vmhub)
 690 {
 691         if (adev->asic_type == CHIP_ALDEBARAN)
 692                 return false;
 693
 694         return ((vmhub == AMDGPU_MMHUB_0 ||
 695                  vmhub == AMDGPU_MMHUB_1) &&
 696                 (!amdgpu_sriov_vf(adev)) &&
 697                 (!(!(adev->apu_flags & AMD_APU_IS_RAVEN2) &&
 698                    (adev->apu_flags & AMD_APU_IS_PICASSO))));
 699 }
 700
 701 static bool gmc_v9_0_get_atc_vmid_pasid_mapping_info(struct amdgpu_device *adev,
 702                                         uint8_t vmid, uint16_t *p_pasid)
 703 {
 704         uint32_t value;
 705
 706         value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
 707                      + vmid);
 708         *p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK;
 709
 710         return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK);
 711 }
 712
 713 /*
 714  * GART
 715  * VMID 0 is the physical GPU addresses as used by the kernel.
 716  * VMIDs 1-15 are used for userspace clients and are handled
 717  * by the amdgpu vm/hsa code.
 718  */
 719
 720 /**
 721  * gmc_v9_0_flush_gpu_tlb - tlb flush with certain type
 722  *
 723  * @adev: amdgpu_device pointer
 724  * @vmid: vm instance to flush
 725  * @vmhub: which hub to flush
 726  * @flush_type: the flush type
 727  *
 728  * Flush the TLB for the requested page table using certain type.
 729  */
 730 static void gmc_v9_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
 731                                         uint32_t vmhub, uint32_t flush_type)
 732 {
 733         bool use_semaphore = gmc_v9_0_use_invalidate_semaphore(adev, vmhub);
 734         const unsigned eng = 17;
 735         u32 j, inv_req, inv_req2, tmp;
 736         struct amdgpu_vmhub *hub;
 737
 738         BUG_ON(vmhub >= adev->num_vmhubs);
 739
 740         hub = &adev->vmhub[vmhub];
 741         if (adev->gmc.xgmi.num_physical_nodes &&
 742             adev->asic_type == CHIP_VEGA20) {
 743                 /* Vega20+XGMI caches PTEs in TC and TLB. Add a
 744                  * heavy-weight TLB flush (type 2), which flushes
 745                  * both. Due to a race condition with concurrent
 746                  * memory accesses using the same TLB cache line, we
 747                  * still need a second TLB flush after this.
 748                  */
 749                 inv_req = gmc_v9_0_get_invalidate_req(vmid, 2);
 750                 inv_req2 = gmc_v9_0_get_invalidate_req(vmid, flush_type);
 751         } else {
 752                 inv_req = gmc_v9_0_get_invalidate_req(vmid, flush_type);
 753                 inv_req2 = 0;
 754         }
 755
 756         /* This is necessary for a HW workaround under SRIOV as well
 757          * as GFXOFF under bare metal
 758          */
 759         if (adev->gfx.kiq.ring.sched.ready &&
 760             (amdgpu_sriov_runtime(adev) || !amdgpu_sriov_vf(adev)) &&
 761             down_read_trylock(&adev->reset_sem)) {
 762                 uint32_t req = hub->vm_inv_eng0_req + hub->eng_distance * eng;
 763                 uint32_t ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng;
 764
 765                 amdgpu_virt_kiq_reg_write_reg_wait(adev, req, ack, inv_req,
 766                                                    1 << vmid);
 767                 up_read(&adev->reset_sem);
 768                 return;
 769         }
 770
 771         spin_lock(&adev->gmc.invalidate_lock);
 772
 773         /*
 774          * It may lose gpuvm invalidate acknowldege state across power-gating
 775          * off cycle, add semaphore acquire before invalidation and semaphore
 776          * release after invalidation to avoid entering power gated state
 777          * to WA the Issue
 778          */
 779
 780         /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
 781         if (use_semaphore) {
 782                 for (j = 0; j < adev->usec_timeout; j++) {
 783                         /* a read return value of 1 means semaphore acuqire */
 784                         tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_sem +
 785                                             hub->eng_distance * eng);
 786                         if (tmp & 0x1)
 787                                 break;
 788                         udelay(1);
 789                 }
 790
 791                 if (j >= adev->usec_timeout)
 792                         DRM_ERROR("Timeout waiting for sem acquire in VM flush!\n");
 793         }
 794
 795         do {
 796                 WREG32_NO_KIQ(hub->vm_inv_eng0_req +
 797                               hub->eng_distance * eng, inv_req);
 798
 799                 /*
 800                  * Issue a dummy read to wait for the ACK register to
 801                  * be cleared to avoid a false ACK due to the new fast
 802                  * GRBM interface.
 803                  */
 804                 if (vmhub == AMDGPU_GFXHUB_0)
 805                         RREG32_NO_KIQ(hub->vm_inv_eng0_req +
 806                                       hub->eng_distance * eng);
 807
 808                 for (j = 0; j < adev->usec_timeout; j++) {
 809                         tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_ack +
 810                                             hub->eng_distance * eng);
 811                         if (tmp & (1 << vmid))
 812                                 break;
 813                         udelay(1);
 814                 }
 815
 816                 inv_req = inv_req2;
 817                 inv_req2 = 0;
 818         } while (inv_req);
 819
 820         /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
 821         if (use_semaphore)
 822                 /*
 823                  * add semaphore release after invalidation,
 824                  * write with 0 means semaphore release
 825                  */
 826                 WREG32_NO_KIQ(hub->vm_inv_eng0_sem +
 827                               hub->eng_distance * eng, 0);
 828
 829         spin_unlock(&adev->gmc.invalidate_lock);
 830
 831         if (j < adev->usec_timeout)
 832                 return;
 833
 834         DRM_ERROR("Timeout waiting for VM flush ACK!\n");
 835 }
 836
 837 /**
 838  * gmc_v9_0_flush_gpu_tlb_pasid - tlb flush via pasid
 839  *
 840  * @adev: amdgpu_device pointer
 841  * @pasid: pasid to be flush
 842  * @flush_type: the flush type
 843  * @all_hub: flush all hubs
 844  *
 845  * Flush the TLB for the requested pasid.
 846  */
 847 static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
 848                                         uint16_t pasid, uint32_t flush_type,
 849                                         bool all_hub)
 850 {
 851         int vmid, i;
 852         signed long r;
 853         uint32_t seq;
 854         uint16_t queried_pasid;
 855         bool ret;
 856         struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
 857         struct amdgpu_kiq *kiq = &adev->gfx.kiq;
 858
 859         if (amdgpu_in_reset(adev))
 860                 return -EIO;
 861
 862         if (ring->sched.ready && down_read_trylock(&adev->reset_sem)) {
 863                 /* Vega20+XGMI caches PTEs in TC and TLB. Add a
 864                  * heavy-weight TLB flush (type 2), which flushes
 865                  * both. Due to a race condition with concurrent
 866                  * memory accesses using the same TLB cache line, we
 867                  * still need a second TLB flush after this.
 868                  */
 869                 bool vega20_xgmi_wa = (adev->gmc.xgmi.num_physical_nodes &&
 870                                        adev->asic_type == CHIP_VEGA20);
 871                 /* 2 dwords flush + 8 dwords fence */
 872                 unsigned int ndw = kiq->pmf->invalidate_tlbs_size + 8;
 873
 874                 if (vega20_xgmi_wa)
 875                         ndw += kiq->pmf->invalidate_tlbs_size;
 876
 877                 spin_lock(&adev->gfx.kiq.ring_lock);
 878                 /* 2 dwords flush + 8 dwords fence */
 879                 amdgpu_ring_alloc(ring, ndw);
 880                 if (vega20_xgmi_wa)
 881                         kiq->pmf->kiq_invalidate_tlbs(ring,
 882                                                       pasid, 2, all_hub);
 883                 kiq->pmf->kiq_invalidate_tlbs(ring,
 884                                         pasid, flush_type, all_hub);
 885                 r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
 886                 if (r) {
 887                         amdgpu_ring_undo(ring);
 888                         spin_unlock(&adev->gfx.kiq.ring_lock);
 889                         up_read(&adev->reset_sem);
 890                         return -ETIME;
 891                 }
 892
 893                 amdgpu_ring_commit(ring);
 894                 spin_unlock(&adev->gfx.kiq.ring_lock);
 895                 r = amdgpu_fence_wait_polling(ring, seq, adev->usec_timeout);
 896                 if (r < 1) {
 897                         dev_err(adev->dev, "wait for kiq fence error: %ld.\n", r);
 898                         up_read(&adev->reset_sem);
 899                         return -ETIME;
 900                 }
 901                 up_read(&adev->reset_sem);
 902                 return 0;
 903         }
 904
 905         for (vmid = 1; vmid < 16; vmid++) {
 906
 907                 ret = gmc_v9_0_get_atc_vmid_pasid_mapping_info(adev, vmid,
 908                                 &queried_pasid);
 909                 if (ret && queried_pasid == pasid) {
 910                         if (all_hub) {
 911                                 for (i = 0; i < adev->num_vmhubs; i++)
 912                                         gmc_v9_0_flush_gpu_tlb(adev, vmid,
 913                                                         i, flush_type);
 914                         } else {
 915                                 gmc_v9_0_flush_gpu_tlb(adev, vmid,
 916                                                 AMDGPU_GFXHUB_0, flush_type);
 917                         }
 918                         break;
 919                 }
 920         }
 921
 922         return 0;
 923
 924 }
 925
 926 static uint64_t gmc_v9_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
 927                                             unsigned vmid, uint64_t pd_addr)
 928 {
 929         bool use_semaphore = gmc_v9_0_use_invalidate_semaphore(ring->adev, ring->funcs->vmhub);
 930         struct amdgpu_device *adev = ring->adev;
 931         struct amdgpu_vmhub *hub = &adev->vmhub[ring->funcs->vmhub];
 932         uint32_t req = gmc_v9_0_get_invalidate_req(vmid, 0);
 933         unsigned eng = ring->vm_inv_eng;
 934
 935         /*
 936          * It may lose gpuvm invalidate acknowldege state across power-gating
 937          * off cycle, add semaphore acquire before invalidation and semaphore
 938          * release after invalidation to avoid entering power gated state
 939          * to WA the Issue
 940          */
 941
 942         /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
 943         if (use_semaphore)
 944                 /* a read return value of 1 means semaphore acuqire */
 945                 amdgpu_ring_emit_reg_wait(ring,
 946                                           hub->vm_inv_eng0_sem +
 947                                           hub->eng_distance * eng, 0x1, 0x1);
 948
 949         amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 +
 950                               (hub->ctx_addr_distance * vmid),
 951                               lower_32_bits(pd_addr));
 952
 953         amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 +
 954                               (hub->ctx_addr_distance * vmid),
 955                               upper_32_bits(pd_addr));
 956
 957         amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req +
 958                                             hub->eng_distance * eng,
 959                                             hub->vm_inv_eng0_ack +
 960                                             hub->eng_distance * eng,
 961                                             req, 1 << vmid);
 962
 963         /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
 964         if (use_semaphore)
 965                 /*
 966                  * add semaphore release after invalidation,
 967                  * write with 0 means semaphore release
 968                  */
 969                 amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_sem +
 970                                       hub->eng_distance * eng, 0);
 971
 972         return pd_addr;
 973 }
 974
 975 static void gmc_v9_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid,
 976                                         unsigned pasid)
 977 {
 978         struct amdgpu_device *adev = ring->adev;
 979         uint32_t reg;
 980
 981         /* Do nothing because there's no lut register for mmhub1. */
 982         if (ring->funcs->vmhub == AMDGPU_MMHUB_1)
 983                 return;
 984
 985         if (ring->funcs->vmhub == AMDGPU_GFXHUB_0)
 986                 reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid;
 987         else
 988                 reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid;
 989
 990         amdgpu_ring_emit_wreg(ring, reg, pasid);
 991 }
 992
 993 /*
 994  * PTE format on VEGA 10:
 995  * 63:59 reserved
 996  * 58:57 mtype
 997  * 56 F
 998  * 55 L
 999  * 54 P
1000  * 53 SW
1001  * 52 T
1002  * 50:48 reserved
1003  * 47:12 4k physical page base address
1004  * 11:7 fragment
1005  * 6 write
1006  * 5 read
1007  * 4 exe
1008  * 3 Z
1009  * 2 snooped
1010  * 1 system
1011  * 0 valid
1012  *
1013  * PDE format on VEGA 10:
1014  * 63:59 block fragment size
1015  * 58:55 reserved
1016  * 54 P
1017  * 53:48 reserved
1018  * 47:6 physical base address of PD or PTE
1019  * 5:3 reserved
1020  * 2 C
1021  * 1 system
1022  * 0 valid
1023  */
1024
1025 static uint64_t gmc_v9_0_map_mtype(struct amdgpu_device *adev, uint32_t flags)
1026
1027 {
1028         switch (flags) {
1029         case AMDGPU_VM_MTYPE_DEFAULT:
1030                 return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC);
1031         case AMDGPU_VM_MTYPE_NC:
1032                 return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC);
1033         case AMDGPU_VM_MTYPE_WC:
1034                 return AMDGPU_PTE_MTYPE_VG10(MTYPE_WC);
1035         case AMDGPU_VM_MTYPE_RW:
1036                 return AMDGPU_PTE_MTYPE_VG10(MTYPE_RW);
1037         case AMDGPU_VM_MTYPE_CC:
1038                 return AMDGPU_PTE_MTYPE_VG10(MTYPE_CC);
1039         case AMDGPU_VM_MTYPE_UC:
1040                 return AMDGPU_PTE_MTYPE_VG10(MTYPE_UC);
1041         default:
1042                 return AMDGPU_PTE_MTYPE_VG10(MTYPE_NC);
1043         }
1044 }
1045
1046 static void gmc_v9_0_get_vm_pde(struct amdgpu_device *adev, int level,
1047                                 uint64_t *addr, uint64_t *flags)
1048 {
1049         if (!(*flags & AMDGPU_PDE_PTE) && !(*flags & AMDGPU_PTE_SYSTEM))
1050                 *addr = adev->vm_manager.vram_base_offset + *addr -
1051                         adev->gmc.vram_start;
1052         BUG_ON(*addr & 0xFFFF00000000003FULL);
1053
1054         if (!adev->gmc.translate_further)
1055                 return;
1056
1057         if (level == AMDGPU_VM_PDB1) {
1058                 /* Set the block fragment size */
1059                 if (!(*flags & AMDGPU_PDE_PTE))
1060                         *flags |= AMDGPU_PDE_BFS(0x9);
1061
1062         } else if (level == AMDGPU_VM_PDB0) {
1063                 if (*flags & AMDGPU_PDE_PTE)
1064                         *flags &= ~AMDGPU_PDE_PTE;
1065                 else
1066                         *flags |= AMDGPU_PTE_TF;
1067         }
1068 }
1069
1070 static void gmc_v9_0_get_vm_pte(struct amdgpu_device *adev,
1071                                 struct amdgpu_bo_va_mapping *mapping,
1072                                 uint64_t *flags)
1073 {
1074         *flags &= ~AMDGPU_PTE_EXECUTABLE;
1075         *flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE;
1076
1077         *flags &= ~AMDGPU_PTE_MTYPE_VG10_MASK;
1078         *flags |= mapping->flags & AMDGPU_PTE_MTYPE_VG10_MASK;
1079
1080         if (mapping->flags & AMDGPU_PTE_PRT) {
1081                 *flags |= AMDGPU_PTE_PRT;
1082                 *flags &= ~AMDGPU_PTE_VALID;
1083         }
1084
1085         if ((adev->asic_type == CHIP_ARCTURUS ||
1086             adev->asic_type == CHIP_ALDEBARAN) &&
1087             !(*flags & AMDGPU_PTE_SYSTEM) &&
1088             mapping->bo_va->is_xgmi)
1089                 *flags |= AMDGPU_PTE_SNOOPED;
1090
1091         if (adev->asic_type == CHIP_ALDEBARAN)
1092                 *flags |= mapping->flags & AMDGPU_PTE_SNOOPED;
1093 }
1094
1095 static unsigned gmc_v9_0_get_vbios_fb_size(struct amdgpu_device *adev)
1096 {
1097         u32 d1vga_control = RREG32_SOC15(DCE, 0, mmD1VGA_CONTROL);
1098         unsigned size;
1099
1100         if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) {
1101                 size = AMDGPU_VBIOS_VGA_ALLOCATION;
1102         } else {
1103                 u32 viewport;
1104
1105                 switch (adev->asic_type) {
1106                 case CHIP_RAVEN:
1107                 case CHIP_RENOIR:
1108                         viewport = RREG32_SOC15(DCE, 0, mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION);
1109                         size = (REG_GET_FIELD(viewport,
1110                                               HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) *
1111                                 REG_GET_FIELD(viewport,
1112                                               HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_WIDTH) *
1113                                 4);
1114                         break;
1115                 case CHIP_VEGA10:
1116                 case CHIP_VEGA12:
1117                 case CHIP_VEGA20:
1118                 default:
1119                         viewport = RREG32_SOC15(DCE, 0, mmSCL0_VIEWPORT_SIZE);
1120                         size = (REG_GET_FIELD(viewport, SCL0_VIEWPORT_SIZE, VIEWPORT_HEIGHT) *
1121                                 REG_GET_FIELD(viewport, SCL0_VIEWPORT_SIZE, VIEWPORT_WIDTH) *
1122                                 4);
1123                         break;
1124                 }
1125         }
1126
1127         return size;
1128 }
1129
1130 static const struct amdgpu_gmc_funcs gmc_v9_0_gmc_funcs = {
1131         .flush_gpu_tlb = gmc_v9_0_flush_gpu_tlb,
1132         .flush_gpu_tlb_pasid = gmc_v9_0_flush_gpu_tlb_pasid,
1133         .emit_flush_gpu_tlb = gmc_v9_0_emit_flush_gpu_tlb,
1134         .emit_pasid_mapping = gmc_v9_0_emit_pasid_mapping,
1135         .map_mtype = gmc_v9_0_map_mtype,
1136         .get_vm_pde = gmc_v9_0_get_vm_pde,
1137         .get_vm_pte = gmc_v9_0_get_vm_pte,
1138         .get_vbios_fb_size = gmc_v9_0_get_vbios_fb_size,
1139 };
1140
1141 static void gmc_v9_0_set_gmc_funcs(struct amdgpu_device *adev)
1142 {
1143         adev->gmc.gmc_funcs = &gmc_v9_0_gmc_funcs;
1144 }
1145
1146 static void gmc_v9_0_set_umc_funcs(struct amdgpu_device *adev)
1147 {
1148         switch (adev->asic_type) {
1149         case CHIP_VEGA10:
1150                 adev->umc.funcs = &umc_v6_0_funcs;
1151                 break;
1152         case CHIP_VEGA20:
1153                 adev->umc.max_ras_err_cnt_per_query = UMC_V6_1_TOTAL_CHANNEL_NUM;
1154                 adev->umc.channel_inst_num = UMC_V6_1_CHANNEL_INSTANCE_NUM;
1155                 adev->umc.umc_inst_num = UMC_V6_1_UMC_INSTANCE_NUM;
1156                 adev->umc.channel_offs = UMC_V6_1_PER_CHANNEL_OFFSET_VG20;
1157                 adev->umc.channel_idx_tbl = &umc_v6_1_channel_idx_tbl[0][0];
1158                 adev->umc.funcs = &umc_v6_1_funcs;
1159                 break;
1160         case CHIP_ARCTURUS:
1161                 adev->umc.max_ras_err_cnt_per_query = UMC_V6_1_TOTAL_CHANNEL_NUM;
1162                 adev->umc.channel_inst_num = UMC_V6_1_CHANNEL_INSTANCE_NUM;
1163                 adev->umc.umc_inst_num = UMC_V6_1_UMC_INSTANCE_NUM;
1164                 adev->umc.channel_offs = UMC_V6_1_PER_CHANNEL_OFFSET_ARCT;
1165                 adev->umc.channel_idx_tbl = &umc_v6_1_channel_idx_tbl[0][0];
1166                 adev->umc.funcs = &umc_v6_1_funcs;
1167                 break;
1168         default:
1169                 break;
1170         }
1171 }
1172
1173 static void gmc_v9_0_set_mmhub_funcs(struct amdgpu_device *adev)
1174 {
1175         switch (adev->asic_type) {
1176         case CHIP_ARCTURUS:
1177                 adev->mmhub.funcs = &mmhub_v9_4_funcs;
1178                 break;
1179         case CHIP_ALDEBARAN:
1180                 adev->mmhub.funcs = &mmhub_v1_7_funcs;
1181                 break;
1182         default:
1183                 adev->mmhub.funcs = &mmhub_v1_0_funcs;
1184                 break;
1185         }
1186 }
1187
1188 static void gmc_v9_0_set_gfxhub_funcs(struct amdgpu_device *adev)
1189 {
1190         adev->gfxhub.funcs = &gfxhub_v1_0_funcs;
1191 }
1192
1193 static int gmc_v9_0_early_init(void *handle)
1194 {
1195         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1196
1197         gmc_v9_0_set_gmc_funcs(adev);
1198         gmc_v9_0_set_irq_funcs(adev);
1199         gmc_v9_0_set_umc_funcs(adev);
1200         gmc_v9_0_set_mmhub_funcs(adev);
1201         gmc_v9_0_set_gfxhub_funcs(adev);
1202
1203         if (adev->asic_type == CHIP_VEGA20 ||
1204             adev->asic_type == CHIP_ARCTURUS)
1205                 adev->gmc.xgmi.supported = true;
1206
1207         if (adev->asic_type == CHIP_ALDEBARAN) {
1208                 adev->gmc.xgmi.supported = true;
1209                 adev->gmc.xgmi.connected_to_cpu =
1210                         adev->smuio.funcs->is_host_gpu_xgmi_supported(adev);
1211         }
1212
1213         adev->gmc.shared_aperture_start = 0x2000000000000000ULL;
1214         adev->gmc.shared_aperture_end =
1215                 adev->gmc.shared_aperture_start + (4ULL << 30) - 1;
1216         adev->gmc.private_aperture_start = 0x1000000000000000ULL;
1217         adev->gmc.private_aperture_end =
1218                 adev->gmc.private_aperture_start + (4ULL << 30) - 1;
1219
1220         return 0;
1221 }
1222
1223 static int gmc_v9_0_late_init(void *handle)
1224 {
1225         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1226         int r;
1227
1228         r = amdgpu_gmc_allocate_vm_inv_eng(adev);
1229         if (r)
1230                 return r;
1231
1232         /*
1233          * Workaround performance drop issue with VBIOS enables partial
1234          * writes, while disables HBM ECC for vega10.
1235          */
1236         if (!amdgpu_sriov_vf(adev) && (adev->asic_type == CHIP_VEGA10)) {
1237                 if (!(adev->ras_features & (1 << AMDGPU_RAS_BLOCK__UMC))) {
1238                         if (adev->df.funcs->enable_ecc_force_par_wr_rmw)
1239                                 adev->df.funcs->enable_ecc_force_par_wr_rmw(adev, false);
1240                 }
1241         }
1242
1243         if (adev->mmhub.funcs && adev->mmhub.funcs->reset_ras_error_count)
1244                 adev->mmhub.funcs->reset_ras_error_count(adev);
1245
1246         r = amdgpu_gmc_ras_late_init(adev);
1247         if (r)
1248                 return r;
1249
1250         return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0);
1251 }
1252
1253 static void gmc_v9_0_vram_gtt_location(struct amdgpu_device *adev,
1254                                         struct amdgpu_gmc *mc)
1255 {
1256         u64 base = 0;
1257
1258         if (!amdgpu_sriov_vf(adev))
1259                 base = adev->mmhub.funcs->get_fb_location(adev);
1260
1261         /* add the xgmi offset of the physical node */
1262         base += adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
1263         if (adev->gmc.xgmi.connected_to_cpu) {
1264                 amdgpu_gmc_sysvm_location(adev, mc);
1265         } else {
1266                 amdgpu_gmc_vram_location(adev, mc, base);
1267                 amdgpu_gmc_gart_location(adev, mc);
1268                 amdgpu_gmc_agp_location(adev, mc);
1269         }
1270         /* base offset of vram pages */
1271         adev->vm_manager.vram_base_offset = adev->gfxhub.funcs->get_mc_fb_offset(adev);
1272
1273         /* XXX: add the xgmi offset of the physical node? */
1274         adev->vm_manager.vram_base_offset +=
1275                 adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
1276 }
1277
1278 /**
1279  * gmc_v9_0_mc_init - initialize the memory controller driver params
1280  *
1281  * @adev: amdgpu_device pointer
1282  *
1283  * Look up the amount of vram, vram width, and decide how to place
1284  * vram and gart within the GPU's physical address space.
1285  * Returns 0 for success.
1286  */
1287 static int gmc_v9_0_mc_init(struct amdgpu_device *adev)
1288 {
1289         int r;
1290
1291         /* size in MB on si */
1292         adev->gmc.mc_vram_size =
1293                 adev->nbio.funcs->get_memsize(adev) * 1024ULL * 1024ULL;
1294         adev->gmc.real_vram_size = adev->gmc.mc_vram_size;
1295
1296         if (!(adev->flags & AMD_IS_APU) &&
1297             !adev->gmc.xgmi.connected_to_cpu) {
1298                 r = amdgpu_device_resize_fb_bar(adev);
1299                 if (r)
1300                         return r;
1301         }
1302         adev->gmc.aper_base = pci_resource_start(adev->pdev, 0);
1303         adev->gmc.aper_size = pci_resource_len(adev->pdev, 0);
1304
1305 #ifdef CONFIG_X86_64
1306         /*
1307          * AMD Accelerated Processing Platform (APP) supporting GPU-HOST xgmi
1308          * interface can use VRAM through here as it appears system reserved
1309          * memory in host address space.
1310          *
1311          * For APUs, VRAM is just the stolen system memory and can be accessed
1312          * directly.
1313          *
1314          * Otherwise, use the legacy Host Data Path (HDP) through PCIe BAR.
1315          */
1316
1317         /* check whether both host-gpu and gpu-gpu xgmi links exist */
1318         if ((adev->flags & AMD_IS_APU) ||
1319             (adev->gmc.xgmi.supported &&
1320              adev->gmc.xgmi.connected_to_cpu)) {
1321                 adev->gmc.aper_base =
1322                         adev->gfxhub.funcs->get_mc_fb_offset(adev) +
1323                         adev->gmc.xgmi.physical_node_id *
1324                         adev->gmc.xgmi.node_segment_size;
1325                 adev->gmc.aper_size = adev->gmc.real_vram_size;
1326         }
1327
1328 #endif
1329         /* In case the PCI BAR is larger than the actual amount of vram */
1330         adev->gmc.visible_vram_size = adev->gmc.aper_size;
1331         if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size)
1332                 adev->gmc.visible_vram_size = adev->gmc.real_vram_size;
1333
1334         /* set the gart size */
1335         if (amdgpu_gart_size == -1) {
1336                 switch (adev->asic_type) {
1337                 case CHIP_VEGA10:  /* all engines support GPUVM */
1338                 case CHIP_VEGA12:  /* all engines support GPUVM */
1339                 case CHIP_VEGA20:
1340                 case CHIP_ARCTURUS:
1341                 case CHIP_ALDEBARAN:
1342                 default:
1343                         adev->gmc.gart_size = 512ULL << 20;
1344                         break;
1345                 case CHIP_RAVEN:   /* DCE SG support */
1346                 case CHIP_RENOIR:
1347                         adev->gmc.gart_size = 1024ULL << 20;
1348                         break;
1349                 }
1350         } else {
1351                 adev->gmc.gart_size = (u64)amdgpu_gart_size << 20;
1352         }
1353
1354         adev->gmc.gart_size += adev->pm.smu_prv_buffer_size;
1355
1356         gmc_v9_0_vram_gtt_location(adev, &adev->gmc);
1357
1358         return 0;
1359 }
1360
1361 static int gmc_v9_0_gart_init(struct amdgpu_device *adev)
1362 {
1363         int r;
1364
1365         if (adev->gart.bo) {
1366                 WARN(1, "VEGA10 PCIE GART already initialized\n");
1367                 return 0;
1368         }
1369
1370         if (adev->gmc.xgmi.connected_to_cpu) {
1371                 adev->gmc.vmid0_page_table_depth = 1;
1372                 adev->gmc.vmid0_page_table_block_size = 12;
1373         } else {
1374                 adev->gmc.vmid0_page_table_depth = 0;
1375                 adev->gmc.vmid0_page_table_block_size = 0;
1376         }
1377
1378         /* Initialize common gart structure */
1379         r = amdgpu_gart_init(adev);
1380         if (r)
1381                 return r;
1382         adev->gart.table_size = adev->gart.num_gpu_pages * 8;
1383         adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_VG10(MTYPE_UC) |
1384                                  AMDGPU_PTE_EXECUTABLE;
1385
1386         r = amdgpu_gart_table_vram_alloc(adev);
1387         if (r)
1388                 return r;
1389
1390         if (adev->gmc.xgmi.connected_to_cpu) {
1391                 r = amdgpu_gmc_pdb0_alloc(adev);
1392         }
1393
1394         return r;
1395 }
1396
1397 /**
1398  * gmc_v9_0_save_registers - saves regs
1399  *
1400  * @adev: amdgpu_device pointer
1401  *
1402  * This saves potential register values that should be
1403  * restored upon resume
1404  */
1405 static void gmc_v9_0_save_registers(struct amdgpu_device *adev)
1406 {
1407         if (adev->asic_type == CHIP_RAVEN)
1408                 adev->gmc.sdpif_register = RREG32_SOC15(DCE, 0, mmDCHUBBUB_SDPIF_MMIO_CNTRL_0);
1409 }
1410
1411 static int gmc_v9_0_sw_init(void *handle)
1412 {
1413         int r, vram_width = 0, vram_type = 0, vram_vendor = 0;
1414         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1415
1416         adev->gfxhub.funcs->init(adev);
1417
1418         adev->mmhub.funcs->init(adev);
1419
1420         spin_lock_init(&adev->gmc.invalidate_lock);
1421
1422         r = amdgpu_atomfirmware_get_vram_info(adev,
1423                 &vram_width, &vram_type, &vram_vendor);
1424         if (amdgpu_sriov_vf(adev))
1425                 /* For Vega10 SR-IOV, vram_width can't be read from ATOM as RAVEN,
1426                  * and DF related registers is not readable, seems hardcord is the
1427                  * only way to set the correct vram_width
1428                  */
1429                 adev->gmc.vram_width = 2048;
1430         else if (amdgpu_emu_mode != 1)
1431                 adev->gmc.vram_width = vram_width;
1432
1433         if (!adev->gmc.vram_width) {
1434                 int chansize, numchan;
1435
1436                 /* hbm memory channel size */
1437                 if (adev->flags & AMD_IS_APU)
1438                         chansize = 64;
1439                 else
1440                         chansize = 128;
1441
1442                 numchan = adev->df.funcs->get_hbm_channel_number(adev);
1443                 adev->gmc.vram_width = numchan * chansize;
1444         }
1445
1446         adev->gmc.vram_type = vram_type;
1447         adev->gmc.vram_vendor = vram_vendor;
1448         switch (adev->asic_type) {
1449         case CHIP_RAVEN:
1450                 adev->num_vmhubs = 2;
1451
1452                 if (adev->rev_id == 0x0 || adev->rev_id == 0x1) {
1453                         amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
1454                 } else {
1455                         /* vm_size is 128TB + 512GB for legacy 3-level page support */
1456                         amdgpu_vm_adjust_size(adev, 128 * 1024 + 512, 9, 2, 48);
1457                         adev->gmc.translate_further =
1458                                 adev->vm_manager.num_level > 1;
1459                 }
1460                 break;
1461         case CHIP_VEGA10:
1462         case CHIP_VEGA12:
1463         case CHIP_VEGA20:
1464         case CHIP_RENOIR:
1465         case CHIP_ALDEBARAN:
1466                 adev->num_vmhubs = 2;
1467
1468
1469                 /*
1470                  * To fulfill 4-level page support,
1471                  * vm size is 256TB (48bit), maximum size of Vega10,
1472                  * block size 512 (9bit)
1473                  */
1474                 /* sriov restrict max_pfn below AMDGPU_GMC_HOLE */
1475                 if (amdgpu_sriov_vf(adev))
1476                         amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 47);
1477                 else
1478                         amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
1479                 break;
1480         case CHIP_ARCTURUS:
1481                 adev->num_vmhubs = 3;
1482
1483                 /* Keep the vm size same with Vega20 */
1484                 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
1485                 break;
1486         default:
1487                 break;
1488         }
1489
1490         /* This interrupt is VMC page fault.*/
1491         r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC, VMC_1_0__SRCID__VM_FAULT,
1492                                 &adev->gmc.vm_fault);
1493         if (r)
1494                 return r;
1495
1496         if (adev->asic_type == CHIP_ARCTURUS) {
1497                 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC1, VMC_1_0__SRCID__VM_FAULT,
1498                                         &adev->gmc.vm_fault);
1499                 if (r)
1500                         return r;
1501         }
1502
1503         r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_UTCL2, UTCL2_1_0__SRCID__FAULT,
1504                                 &adev->gmc.vm_fault);
1505
1506         if (r)
1507                 return r;
1508
1509         if (!amdgpu_sriov_vf(adev)) {
1510                 /* interrupt sent to DF. */
1511                 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_DF, 0,
1512                                       &adev->gmc.ecc_irq);
1513                 if (r)
1514                         return r;
1515         }
1516
1517         /* Set the internal MC address mask
1518          * This is the max address of the GPU's
1519          * internal address space.
1520          */
1521         adev->gmc.mc_mask = 0xffffffffffffULL; /* 48 bit MC */
1522
1523         r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(44));
1524         if (r) {
1525                 printk(KERN_WARNING "amdgpu: No suitable DMA available.\n");
1526                 return r;
1527         }
1528         adev->need_swiotlb = drm_need_swiotlb(44);
1529
1530         if (adev->gmc.xgmi.supported) {
1531                 r = adev->gfxhub.funcs->get_xgmi_info(adev);
1532                 if (r)
1533                         return r;
1534         }
1535
1536         r = gmc_v9_0_mc_init(adev);
1537         if (r)
1538                 return r;
1539
1540         amdgpu_gmc_get_vbios_allocations(adev);
1541
1542         /* Memory manager */
1543         r = amdgpu_bo_init(adev);
1544         if (r)
1545                 return r;
1546
1547         r = gmc_v9_0_gart_init(adev);
1548         if (r)
1549                 return r;
1550
1551         /*
1552          * number of VMs
1553          * VMID 0 is reserved for System
1554          * amdgpu graphics/compute will use VMIDs 1..n-1
1555          * amdkfd will use VMIDs n..15
1556          *
1557          * The first KFD VMID is 8 for GPUs with graphics, 3 for
1558          * compute-only GPUs. On compute-only GPUs that leaves 2 VMIDs
1559          * for video processing.
1560          */
1561         adev->vm_manager.first_kfd_vmid =
1562                 (adev->asic_type == CHIP_ARCTURUS ||
1563                  adev->asic_type == CHIP_ALDEBARAN) ? 3 : 8;
1564
1565         amdgpu_vm_manager_init(adev);
1566
1567         gmc_v9_0_save_registers(adev);
1568
1569         return 0;
1570 }
1571
1572 static int gmc_v9_0_sw_fini(void *handle)
1573 {
1574         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1575
1576         amdgpu_gmc_ras_fini(adev);
1577         amdgpu_gem_force_release(adev);
1578         amdgpu_vm_manager_fini(adev);
1579         amdgpu_gart_table_vram_free(adev);
1580         amdgpu_bo_unref(&adev->gmc.pdb0_bo);
1581         amdgpu_bo_fini(adev);
1582         amdgpu_gart_fini(adev);
1583
1584         return 0;
1585 }
1586
1587 static void gmc_v9_0_init_golden_registers(struct amdgpu_device *adev)
1588 {
1589
1590         switch (adev->asic_type) {
1591         case CHIP_VEGA10:
1592                 if (amdgpu_sriov_vf(adev))
1593                         break;
1594                 fallthrough;
1595         case CHIP_VEGA20:
1596                 soc15_program_register_sequence(adev,
1597                                                 golden_settings_mmhub_1_0_0,
1598                                                 ARRAY_SIZE(golden_settings_mmhub_1_0_0));
1599                 soc15_program_register_sequence(adev,
1600                                                 golden_settings_athub_1_0_0,
1601                                                 ARRAY_SIZE(golden_settings_athub_1_0_0));
1602                 break;
1603         case CHIP_VEGA12:
1604                 break;
1605         case CHIP_RAVEN:
1606                 /* TODO for renoir */
1607                 soc15_program_register_sequence(adev,
1608                                                 golden_settings_athub_1_0_0,
1609                                                 ARRAY_SIZE(golden_settings_athub_1_0_0));
1610                 break;
1611         default:
1612                 break;
1613         }
1614 }
1615
1616 /**
1617  * gmc_v9_0_restore_registers - restores regs
1618  *
1619  * @adev: amdgpu_device pointer
1620  *
1621  * This restores register values, saved at suspend.
1622  */
1623 void gmc_v9_0_restore_registers(struct amdgpu_device *adev)
1624 {
1625         if (adev->asic_type == CHIP_RAVEN) {
1626                 WREG32_SOC15(DCE, 0, mmDCHUBBUB_SDPIF_MMIO_CNTRL_0, adev->gmc.sdpif_register);
1627                 WARN_ON(adev->gmc.sdpif_register !=
1628                         RREG32_SOC15(DCE, 0, mmDCHUBBUB_SDPIF_MMIO_CNTRL_0));
1629         }
1630 }
1631
1632 /**
1633  * gmc_v9_0_gart_enable - gart enable
1634  *
1635  * @adev: amdgpu_device pointer
1636  */
1637 static int gmc_v9_0_gart_enable(struct amdgpu_device *adev)
1638 {
1639         int r;
1640
1641         if (adev->gmc.xgmi.connected_to_cpu)
1642                 amdgpu_gmc_init_pdb0(adev);
1643
1644         if (adev->gart.bo == NULL) {
1645                 dev_err(adev->dev, "No VRAM object for PCIE GART.\n");
1646                 return -EINVAL;
1647         }
1648
1649         r = amdgpu_gart_table_vram_pin(adev);
1650         if (r)
1651                 return r;
1652
1653         r = adev->gfxhub.funcs->gart_enable(adev);
1654         if (r)
1655                 return r;
1656
1657         r = adev->mmhub.funcs->gart_enable(adev);
1658         if (r)
1659                 return r;
1660
1661         DRM_INFO("PCIE GART of %uM enabled.\n",
1662                  (unsigned)(adev->gmc.gart_size >> 20));
1663         if (adev->gmc.pdb0_bo)
1664                 DRM_INFO("PDB0 located at 0x%016llX\n",
1665                                 (unsigned long long)amdgpu_bo_gpu_offset(adev->gmc.pdb0_bo));
1666         DRM_INFO("PTB located at 0x%016llX\n",
1667                         (unsigned long long)amdgpu_bo_gpu_offset(adev->gart.bo));
1668
1669         adev->gart.ready = true;
1670         return 0;
1671 }
1672
1673 static int gmc_v9_0_hw_init(void *handle)
1674 {
1675         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1676         bool value;
1677         int r, i;
1678
1679         /* The sequence of these two function calls matters.*/
1680         gmc_v9_0_init_golden_registers(adev);
1681
1682         if (adev->mode_info.num_crtc) {
1683                 /* Lockout access through VGA aperture*/
1684                 WREG32_FIELD15(DCE, 0, VGA_HDP_CONTROL, VGA_MEMORY_DISABLE, 1);
1685                 /* disable VGA render */
1686                 WREG32_FIELD15(DCE, 0, VGA_RENDER_CONTROL, VGA_VSTATUS_CNTL, 0);
1687         }
1688
1689         if (adev->mmhub.funcs->update_power_gating)
1690                 adev->mmhub.funcs->update_power_gating(adev, true);
1691
1692         adev->hdp.funcs->init_registers(adev);
1693
1694         /* After HDP is initialized, flush HDP.*/
1695         adev->hdp.funcs->flush_hdp(adev, NULL);
1696
1697         if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS)
1698                 value = false;
1699         else
1700                 value = true;
1701
1702         if (!amdgpu_sriov_vf(adev)) {
1703                 adev->gfxhub.funcs->set_fault_enable_default(adev, value);
1704                 adev->mmhub.funcs->set_fault_enable_default(adev, value);
1705         }
1706         for (i = 0; i < adev->num_vmhubs; ++i)
1707                 gmc_v9_0_flush_gpu_tlb(adev, 0, i, 0);
1708
1709         if (adev->umc.funcs && adev->umc.funcs->init_registers)
1710                 adev->umc.funcs->init_registers(adev);
1711
1712         r = gmc_v9_0_gart_enable(adev);
1713
1714         return r;
1715 }
1716
1717 /**
1718  * gmc_v9_0_gart_disable - gart disable
1719  *
1720  * @adev: amdgpu_device pointer
1721  *
1722  * This disables all VM page table.
1723  */
1724 static void gmc_v9_0_gart_disable(struct amdgpu_device *adev)
1725 {
1726         adev->gfxhub.funcs->gart_disable(adev);
1727         adev->mmhub.funcs->gart_disable(adev);
1728         amdgpu_gart_table_vram_unpin(adev);
1729 }
1730
1731 static int gmc_v9_0_hw_fini(void *handle)
1732 {
1733         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1734
1735         if (amdgpu_sriov_vf(adev)) {
1736                 /* full access mode, so don't touch any GMC register */
1737                 DRM_DEBUG("For SRIOV client, shouldn't do anything.\n");
1738                 return 0;
1739         }
1740
1741         amdgpu_irq_put(adev, &adev->gmc.ecc_irq, 0);
1742         amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0);
1743         gmc_v9_0_gart_disable(adev);
1744
1745         return 0;
1746 }
1747
1748 static int gmc_v9_0_suspend(void *handle)
1749 {
1750         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1751
1752         return gmc_v9_0_hw_fini(adev);
1753 }
1754
1755 static int gmc_v9_0_resume(void *handle)
1756 {
1757         int r;
1758         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1759
1760         r = gmc_v9_0_hw_init(adev);
1761         if (r)
1762                 return r;
1763
1764         amdgpu_vmid_reset_all(adev);
1765
1766         return 0;
1767 }
1768
1769 static bool gmc_v9_0_is_idle(void *handle)
1770 {
1771         /* MC is always ready in GMC v9.*/
1772         return true;
1773 }
1774
1775 static int gmc_v9_0_wait_for_idle(void *handle)
1776 {
1777         /* There is no need to wait for MC idle in GMC v9.*/
1778         return 0;
1779 }
1780
1781 static int gmc_v9_0_soft_reset(void *handle)
1782 {
1783         /* XXX for emulation.*/
1784         return 0;
1785 }
1786
1787 static int gmc_v9_0_set_clockgating_state(void *handle,
1788                                         enum amd_clockgating_state state)
1789 {
1790         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1791
1792         adev->mmhub.funcs->set_clockgating(adev, state);
1793
1794         athub_v1_0_set_clockgating(adev, state);
1795
1796         return 0;
1797 }
1798
1799 static void gmc_v9_0_get_clockgating_state(void *handle, u32 *flags)
1800 {
1801         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1802
1803         adev->mmhub.funcs->get_clockgating(adev, flags);
1804
1805         athub_v1_0_get_clockgating(adev, flags);
1806 }
1807
1808 static int gmc_v9_0_set_powergating_state(void *handle,
1809                                         enum amd_powergating_state state)
1810 {
1811         return 0;
1812 }
1813
1814 const struct amd_ip_funcs gmc_v9_0_ip_funcs = {
1815         .name = "gmc_v9_0",
1816         .early_init = gmc_v9_0_early_init,
1817         .late_init = gmc_v9_0_late_init,
1818         .sw_init = gmc_v9_0_sw_init,
1819         .sw_fini = gmc_v9_0_sw_fini,
1820         .hw_init = gmc_v9_0_hw_init,
1821         .hw_fini = gmc_v9_0_hw_fini,
1822         .suspend = gmc_v9_0_suspend,
1823         .resume = gmc_v9_0_resume,
1824         .is_idle = gmc_v9_0_is_idle,
1825         .wait_for_idle = gmc_v9_0_wait_for_idle,
1826         .soft_reset = gmc_v9_0_soft_reset,
1827         .set_clockgating_state = gmc_v9_0_set_clockgating_state,
1828         .set_powergating_state = gmc_v9_0_set_powergating_state,
1829         .get_clockgating_state = gmc_v9_0_get_clockgating_state,
1830 };
1831
1832 const struct amdgpu_ip_block_version gmc_v9_0_ip_block =
1833 {
1834         .type = AMD_IP_BLOCK_TYPE_GMC,
1835         .major = 9,
1836         .minor = 0,
1837         .rev = 0,
1838         .funcs = &gmc_v9_0_ip_funcs,
1839 };