drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c

   1 /*
   2  * Copyright 2014 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 #include <linux/firmware.h>
  24 #include "drmP.h"
  25 #include "amdgpu.h"
  26 #include "gmc_v8_0.h"
  27 #include "amdgpu_ucode.h"
  28
  29 #include "gmc/gmc_8_1_d.h"
  30 #include "gmc/gmc_8_1_sh_mask.h"
  31
  32 #include "bif/bif_5_0_d.h"
  33 #include "bif/bif_5_0_sh_mask.h"
  34
  35 #include "oss/oss_3_0_d.h"
  36 #include "oss/oss_3_0_sh_mask.h"
  37
  38 #include "vid.h"
  39 #include "vi.h"
  40
  41
  42 static void gmc_v8_0_set_gart_funcs(struct amdgpu_device *adev);
  43 static void gmc_v8_0_set_irq_funcs(struct amdgpu_device *adev);
  44
  45 MODULE_FIRMWARE("amdgpu/topaz_mc.bin");
  46 MODULE_FIRMWARE("amdgpu/tonga_mc.bin");
  47 MODULE_FIRMWARE("amdgpu/fiji_mc.bin");
  48
  49 static const u32 golden_settings_tonga_a11[] =
  50 {
  51         mmMC_ARB_WTM_GRPWT_RD, 0x00000003, 0x00000000,
  52         mmMC_HUB_RDREQ_DMIF_LIMIT, 0x0000007f, 0x00000028,
  53         mmMC_HUB_WDP_UMC, 0x00007fb6, 0x00000991,
  54         mmVM_PRT_APERTURE0_LOW_ADDR, 0x0fffffff, 0x0fffffff,
  55         mmVM_PRT_APERTURE1_LOW_ADDR, 0x0fffffff, 0x0fffffff,
  56         mmVM_PRT_APERTURE2_LOW_ADDR, 0x0fffffff, 0x0fffffff,
  57         mmVM_PRT_APERTURE3_LOW_ADDR, 0x0fffffff, 0x0fffffff,
  58 };
  59
  60 static const u32 tonga_mgcg_cgcg_init[] =
  61 {
  62         mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104
  63 };
  64
  65 static const u32 golden_settings_fiji_a10[] =
  66 {
  67         mmVM_PRT_APERTURE0_LOW_ADDR, 0x0fffffff, 0x0fffffff,
  68         mmVM_PRT_APERTURE1_LOW_ADDR, 0x0fffffff, 0x0fffffff,
  69         mmVM_PRT_APERTURE2_LOW_ADDR, 0x0fffffff, 0x0fffffff,
  70         mmVM_PRT_APERTURE3_LOW_ADDR, 0x0fffffff, 0x0fffffff,
  71 };
  72
  73 static const u32 fiji_mgcg_cgcg_init[] =
  74 {
  75         mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104
  76 };
  77
  78 static const u32 golden_settings_iceland_a11[] =
  79 {
  80         mmVM_PRT_APERTURE0_LOW_ADDR, 0x0fffffff, 0x0fffffff,
  81         mmVM_PRT_APERTURE1_LOW_ADDR, 0x0fffffff, 0x0fffffff,
  82         mmVM_PRT_APERTURE2_LOW_ADDR, 0x0fffffff, 0x0fffffff,
  83         mmVM_PRT_APERTURE3_LOW_ADDR, 0x0fffffff, 0x0fffffff
  84 };
  85
  86 static const u32 iceland_mgcg_cgcg_init[] =
  87 {
  88         mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104
  89 };
  90
  91 static const u32 cz_mgcg_cgcg_init[] =
  92 {
  93         mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104
  94 };
  95
  96 static const u32 stoney_mgcg_cgcg_init[] =
  97 {
  98         mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104
  99 };
 100
 101
 102 static void gmc_v8_0_init_golden_registers(struct amdgpu_device *adev)
 103 {
 104         switch (adev->asic_type) {
 105         case CHIP_TOPAZ:
 106                 amdgpu_program_register_sequence(adev,
 107                                                  iceland_mgcg_cgcg_init,
 108                                                  (const u32)ARRAY_SIZE(iceland_mgcg_cgcg_init));
 109                 amdgpu_program_register_sequence(adev,
 110                                                  golden_settings_iceland_a11,
 111                                                  (const u32)ARRAY_SIZE(golden_settings_iceland_a11));
 112                 break;
 113         case CHIP_FIJI:
 114                 amdgpu_program_register_sequence(adev,
 115                                                  fiji_mgcg_cgcg_init,
 116                                                  (const u32)ARRAY_SIZE(fiji_mgcg_cgcg_init));
 117                 amdgpu_program_register_sequence(adev,
 118                                                  golden_settings_fiji_a10,
 119                                                  (const u32)ARRAY_SIZE(golden_settings_fiji_a10));
 120                 break;
 121         case CHIP_TONGA:
 122                 amdgpu_program_register_sequence(adev,
 123                                                  tonga_mgcg_cgcg_init,
 124                                                  (const u32)ARRAY_SIZE(tonga_mgcg_cgcg_init));
 125                 amdgpu_program_register_sequence(adev,
 126                                                  golden_settings_tonga_a11,
 127                                                  (const u32)ARRAY_SIZE(golden_settings_tonga_a11));
 128                 break;
 129         case CHIP_CARRIZO:
 130                 amdgpu_program_register_sequence(adev,
 131                                                  cz_mgcg_cgcg_init,
 132                                                  (const u32)ARRAY_SIZE(cz_mgcg_cgcg_init));
 133                 break;
 134         case CHIP_STONEY:
 135                 amdgpu_program_register_sequence(adev,
 136                                                  stoney_mgcg_cgcg_init,
 137                                                  (const u32)ARRAY_SIZE(stoney_mgcg_cgcg_init));
 138                 break;
 139         default:
 140                 break;
 141         }
 142 }
 143
 144 /**
 145  * gmc8_mc_wait_for_idle - wait for MC idle callback.
 146  *
 147  * @adev: amdgpu_device pointer
 148  *
 149  * Wait for the MC (memory controller) to be idle.
 150  * (evergreen+).
 151  * Returns 0 if the MC is idle, -1 if not.
 152  */
 153 int gmc_v8_0_mc_wait_for_idle(struct amdgpu_device *adev)
 154 {
 155         unsigned i;
 156         u32 tmp;
 157
 158         for (i = 0; i < adev->usec_timeout; i++) {
 159                 /* read MC_STATUS */
 160                 tmp = RREG32(mmSRBM_STATUS) & (SRBM_STATUS__VMC_BUSY_MASK |
 161                                                SRBM_STATUS__MCB_BUSY_MASK |
 162                                                SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK |
 163                                                SRBM_STATUS__MCC_BUSY_MASK |
 164                                                SRBM_STATUS__MCD_BUSY_MASK |
 165                                                SRBM_STATUS__VMC1_BUSY_MASK);
 166                 if (!tmp)
 167                         return 0;
 168                 udelay(1);
 169         }
 170         return -1;
 171 }
 172
 173 void gmc_v8_0_mc_stop(struct amdgpu_device *adev,
 174                       struct amdgpu_mode_mc_save *save)
 175 {
 176         u32 blackout;
 177
 178         if (adev->mode_info.num_crtc)
 179                 amdgpu_display_stop_mc_access(adev, save);
 180
 181         amdgpu_asic_wait_for_mc_idle(adev);
 182
 183         blackout = RREG32(mmMC_SHARED_BLACKOUT_CNTL);
 184         if (REG_GET_FIELD(blackout, MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE) != 1) {
 185                 /* Block CPU access */
 186                 WREG32(mmBIF_FB_EN, 0);
 187                 /* blackout the MC */
 188                 blackout = REG_SET_FIELD(blackout,
 189                                          MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE, 1);
 190                 WREG32(mmMC_SHARED_BLACKOUT_CNTL, blackout);
 191         }
 192         /* wait for the MC to settle */
 193         udelay(100);
 194 }
 195
 196 void gmc_v8_0_mc_resume(struct amdgpu_device *adev,
 197                         struct amdgpu_mode_mc_save *save)
 198 {
 199         u32 tmp;
 200
 201         /* unblackout the MC */
 202         tmp = RREG32(mmMC_SHARED_BLACKOUT_CNTL);
 203         tmp = REG_SET_FIELD(tmp, MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE, 0);
 204         WREG32(mmMC_SHARED_BLACKOUT_CNTL, tmp);
 205         /* allow CPU access */
 206         tmp = REG_SET_FIELD(0, BIF_FB_EN, FB_READ_EN, 1);
 207         tmp = REG_SET_FIELD(tmp, BIF_FB_EN, FB_WRITE_EN, 1);
 208         WREG32(mmBIF_FB_EN, tmp);
 209
 210         if (adev->mode_info.num_crtc)
 211                 amdgpu_display_resume_mc_access(adev, save);
 212 }
 213
 214 /**
 215  * gmc_v8_0_init_microcode - load ucode images from disk
 216  *
 217  * @adev: amdgpu_device pointer
 218  *
 219  * Use the firmware interface to load the ucode images into
 220  * the driver (not loaded into hw).
 221  * Returns 0 on success, error on failure.
 222  */
 223 static int gmc_v8_0_init_microcode(struct amdgpu_device *adev)
 224 {
 225         const char *chip_name;
 226         char fw_name[30];
 227         int err;
 228
 229         DRM_DEBUG("\n");
 230
 231         switch (adev->asic_type) {
 232         case CHIP_TOPAZ:
 233                 chip_name = "topaz";
 234                 break;
 235         case CHIP_TONGA:
 236                 chip_name = "tonga";
 237                 break;
 238         case CHIP_FIJI:
 239                 chip_name = "fiji";
 240                 break;
 241         case CHIP_CARRIZO:
 242         case CHIP_STONEY:
 243                 return 0;
 244         default: BUG();
 245         }
 246
 247         snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mc.bin", chip_name);
 248         err = request_firmware(&adev->mc.fw, fw_name, adev->dev);
 249         if (err)
 250                 goto out;
 251         err = amdgpu_ucode_validate(adev->mc.fw);
 252
 253 out:
 254         if (err) {
 255                 printk(KERN_ERR
 256                        "mc: Failed to load firmware \"%s\"\n",
 257                        fw_name);
 258                 release_firmware(adev->mc.fw);
 259                 adev->mc.fw = NULL;
 260         }
 261         return err;
 262 }
 263
 264 /**
 265  * gmc_v8_0_mc_load_microcode - load MC ucode into the hw
 266  *
 267  * @adev: amdgpu_device pointer
 268  *
 269  * Load the GDDR MC ucode into the hw (CIK).
 270  * Returns 0 on success, error on failure.
 271  */
 272 static int gmc_v8_0_mc_load_microcode(struct amdgpu_device *adev)
 273 {
 274         const struct mc_firmware_header_v1_0 *hdr;
 275         const __le32 *fw_data = NULL;
 276         const __le32 *io_mc_regs = NULL;
 277         u32 running, blackout = 0;
 278         int i, ucode_size, regs_size;
 279
 280         if (!adev->mc.fw)
 281                 return -EINVAL;
 282
 283         hdr = (const struct mc_firmware_header_v1_0 *)adev->mc.fw->data;
 284         amdgpu_ucode_print_mc_hdr(&hdr->header);
 285
 286         adev->mc.fw_version = le32_to_cpu(hdr->header.ucode_version);
 287         regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
 288         io_mc_regs = (const __le32 *)
 289                 (adev->mc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
 290         ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
 291         fw_data = (const __le32 *)
 292                 (adev->mc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
 293
 294         running = REG_GET_FIELD(RREG32(mmMC_SEQ_SUP_CNTL), MC_SEQ_SUP_CNTL, RUN);
 295
 296         if (running == 0) {
 297                 if (running) {
 298                         blackout = RREG32(mmMC_SHARED_BLACKOUT_CNTL);
 299                         WREG32(mmMC_SHARED_BLACKOUT_CNTL, blackout | 1);
 300                 }
 301
 302                 /* reset the engine and set to writable */
 303                 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008);
 304                 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000010);
 305
 306                 /* load mc io regs */
 307                 for (i = 0; i < regs_size; i++) {
 308                         WREG32(mmMC_SEQ_IO_DEBUG_INDEX, le32_to_cpup(io_mc_regs++));
 309                         WREG32(mmMC_SEQ_IO_DEBUG_DATA, le32_to_cpup(io_mc_regs++));
 310                 }
 311                 /* load the MC ucode */
 312                 for (i = 0; i < ucode_size; i++)
 313                         WREG32(mmMC_SEQ_SUP_PGM, le32_to_cpup(fw_data++));
 314
 315                 /* put the engine back into the active state */
 316                 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008);
 317                 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000004);
 318                 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000001);
 319
 320                 /* wait for training to complete */
 321                 for (i = 0; i < adev->usec_timeout; i++) {
 322                         if (REG_GET_FIELD(RREG32(mmMC_SEQ_TRAIN_WAKEUP_CNTL),
 323                                           MC_SEQ_TRAIN_WAKEUP_CNTL, TRAIN_DONE_D0))
 324                                 break;
 325                         udelay(1);
 326                 }
 327                 for (i = 0; i < adev->usec_timeout; i++) {
 328                         if (REG_GET_FIELD(RREG32(mmMC_SEQ_TRAIN_WAKEUP_CNTL),
 329                                           MC_SEQ_TRAIN_WAKEUP_CNTL, TRAIN_DONE_D1))
 330                                 break;
 331                         udelay(1);
 332                 }
 333
 334                 if (running)
 335                         WREG32(mmMC_SHARED_BLACKOUT_CNTL, blackout);
 336         }
 337
 338         return 0;
 339 }
 340
 341 static void gmc_v8_0_vram_gtt_location(struct amdgpu_device *adev,
 342                                        struct amdgpu_mc *mc)
 343 {
 344         if (mc->mc_vram_size > 0xFFC0000000ULL) {
 345                 /* leave room for at least 1024M GTT */
 346                 dev_warn(adev->dev, "limiting VRAM\n");
 347                 mc->real_vram_size = 0xFFC0000000ULL;
 348                 mc->mc_vram_size = 0xFFC0000000ULL;
 349         }
 350         amdgpu_vram_location(adev, &adev->mc, 0);
 351         adev->mc.gtt_base_align = 0;
 352         amdgpu_gtt_location(adev, mc);
 353 }
 354
 355 /**
 356  * gmc_v8_0_mc_program - program the GPU memory controller
 357  *
 358  * @adev: amdgpu_device pointer
 359  *
 360  * Set the location of vram, gart, and AGP in the GPU's
 361  * physical address space (CIK).
 362  */
 363 static void gmc_v8_0_mc_program(struct amdgpu_device *adev)
 364 {
 365         struct amdgpu_mode_mc_save save;
 366         u32 tmp;
 367         int i, j;
 368
 369         /* Initialize HDP */
 370         for (i = 0, j = 0; i < 32; i++, j += 0x6) {
 371                 WREG32((0xb05 + j), 0x00000000);
 372                 WREG32((0xb06 + j), 0x00000000);
 373                 WREG32((0xb07 + j), 0x00000000);
 374                 WREG32((0xb08 + j), 0x00000000);
 375                 WREG32((0xb09 + j), 0x00000000);
 376         }
 377         WREG32(mmHDP_REG_COHERENCY_FLUSH_CNTL, 0);
 378
 379         if (adev->mode_info.num_crtc)
 380                 amdgpu_display_set_vga_render_state(adev, false);
 381
 382         gmc_v8_0_mc_stop(adev, &save);
 383         if (amdgpu_asic_wait_for_mc_idle(adev)) {
 384                 dev_warn(adev->dev, "Wait for MC idle timedout !\n");
 385         }
 386         /* Update configuration */
 387         WREG32(mmMC_VM_SYSTEM_APERTURE_LOW_ADDR,
 388                adev->mc.vram_start >> 12);
 389         WREG32(mmMC_VM_SYSTEM_APERTURE_HIGH_ADDR,
 390                adev->mc.vram_end >> 12);
 391         WREG32(mmMC_VM_SYSTEM_APERTURE_DEFAULT_ADDR,
 392                adev->vram_scratch.gpu_addr >> 12);
 393         tmp = ((adev->mc.vram_end >> 24) & 0xFFFF) << 16;
 394         tmp |= ((adev->mc.vram_start >> 24) & 0xFFFF);
 395         WREG32(mmMC_VM_FB_LOCATION, tmp);
 396         /* XXX double check these! */
 397         WREG32(mmHDP_NONSURFACE_BASE, (adev->mc.vram_start >> 8));
 398         WREG32(mmHDP_NONSURFACE_INFO, (2 << 7) | (1 << 30));
 399         WREG32(mmHDP_NONSURFACE_SIZE, 0x3FFFFFFF);
 400         WREG32(mmMC_VM_AGP_BASE, 0);
 401         WREG32(mmMC_VM_AGP_TOP, 0x0FFFFFFF);
 402         WREG32(mmMC_VM_AGP_BOT, 0x0FFFFFFF);
 403         if (amdgpu_asic_wait_for_mc_idle(adev)) {
 404                 dev_warn(adev->dev, "Wait for MC idle timedout !\n");
 405         }
 406         gmc_v8_0_mc_resume(adev, &save);
 407
 408         WREG32(mmBIF_FB_EN, BIF_FB_EN__FB_READ_EN_MASK | BIF_FB_EN__FB_WRITE_EN_MASK);
 409
 410         tmp = RREG32(mmHDP_MISC_CNTL);
 411         tmp = REG_SET_FIELD(tmp, HDP_MISC_CNTL, FLUSH_INVALIDATE_CACHE, 1);
 412         WREG32(mmHDP_MISC_CNTL, tmp);
 413
 414         tmp = RREG32(mmHDP_HOST_PATH_CNTL);
 415         WREG32(mmHDP_HOST_PATH_CNTL, tmp);
 416 }
 417
 418 /**
 419  * gmc_v8_0_mc_init - initialize the memory controller driver params
 420  *
 421  * @adev: amdgpu_device pointer
 422  *
 423  * Look up the amount of vram, vram width, and decide how to place
 424  * vram and gart within the GPU's physical address space (CIK).
 425  * Returns 0 for success.
 426  */
 427 static int gmc_v8_0_mc_init(struct amdgpu_device *adev)
 428 {
 429         u32 tmp;
 430         int chansize, numchan;
 431
 432         /* Get VRAM informations */
 433         tmp = RREG32(mmMC_ARB_RAMCFG);
 434         if (REG_GET_FIELD(tmp, MC_ARB_RAMCFG, CHANSIZE)) {
 435                 chansize = 64;
 436         } else {
 437                 chansize = 32;
 438         }
 439         tmp = RREG32(mmMC_SHARED_CHMAP);
 440         switch (REG_GET_FIELD(tmp, MC_SHARED_CHMAP, NOOFCHAN)) {
 441         case 0:
 442         default:
 443                 numchan = 1;
 444                 break;
 445         case 1:
 446                 numchan = 2;
 447                 break;
 448         case 2:
 449                 numchan = 4;
 450                 break;
 451         case 3:
 452                 numchan = 8;
 453                 break;
 454         case 4:
 455                 numchan = 3;
 456                 break;
 457         case 5:
 458                 numchan = 6;
 459                 break;
 460         case 6:
 461                 numchan = 10;
 462                 break;
 463         case 7:
 464                 numchan = 12;
 465                 break;
 466         case 8:
 467                 numchan = 16;
 468                 break;
 469         }
 470         adev->mc.vram_width = numchan * chansize;
 471         /* Could aper size report 0 ? */
 472         adev->mc.aper_base = pci_resource_start(adev->pdev, 0);
 473         adev->mc.aper_size = pci_resource_len(adev->pdev, 0);
 474         /* size in MB on si */
 475         adev->mc.mc_vram_size = RREG32(mmCONFIG_MEMSIZE) * 1024ULL * 1024ULL;
 476         adev->mc.real_vram_size = RREG32(mmCONFIG_MEMSIZE) * 1024ULL * 1024ULL;
 477         adev->mc.visible_vram_size = adev->mc.aper_size;
 478
 479         /* unless the user had overridden it, set the gart
 480          * size equal to the 1024 or vram, whichever is larger.
 481          */
 482         if (amdgpu_gart_size == -1)
 483                 adev->mc.gtt_size = max((1024ULL << 20), adev->mc.mc_vram_size);
 484         else
 485                 adev->mc.gtt_size = (uint64_t)amdgpu_gart_size << 20;
 486
 487         gmc_v8_0_vram_gtt_location(adev, &adev->mc);
 488
 489         return 0;
 490 }
 491
 492 /*
 493  * GART
 494  * VMID 0 is the physical GPU addresses as used by the kernel.
 495  * VMIDs 1-15 are used for userspace clients and are handled
 496  * by the amdgpu vm/hsa code.
 497  */
 498
 499 /**
 500  * gmc_v8_0_gart_flush_gpu_tlb - gart tlb flush callback
 501  *
 502  * @adev: amdgpu_device pointer
 503  * @vmid: vm instance to flush
 504  *
 505  * Flush the TLB for the requested page table (CIK).
 506  */
 507 static void gmc_v8_0_gart_flush_gpu_tlb(struct amdgpu_device *adev,
 508                                         uint32_t vmid)
 509 {
 510         /* flush hdp cache */
 511         WREG32(mmHDP_MEM_COHERENCY_FLUSH_CNTL, 0);
 512
 513         /* bits 0-15 are the VM contexts0-15 */
 514         WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
 515 }
 516
 517 /**
 518  * gmc_v8_0_gart_set_pte_pde - update the page tables using MMIO
 519  *
 520  * @adev: amdgpu_device pointer
 521  * @cpu_pt_addr: cpu address of the page table
 522  * @gpu_page_idx: entry in the page table to update
 523  * @addr: dst addr to write into pte/pde
 524  * @flags: access flags
 525  *
 526  * Update the page tables using the CPU.
 527  */
 528 static int gmc_v8_0_gart_set_pte_pde(struct amdgpu_device *adev,
 529                                      void *cpu_pt_addr,
 530                                      uint32_t gpu_page_idx,
 531                                      uint64_t addr,
 532                                      uint32_t flags)
 533 {
 534         void __iomem *ptr = (void *)cpu_pt_addr;
 535         uint64_t value;
 536
 537         /*
 538          * PTE format on VI:
 539          * 63:40 reserved
 540          * 39:12 4k physical page base address
 541          * 11:7 fragment
 542          * 6 write
 543          * 5 read
 544          * 4 exe
 545          * 3 reserved
 546          * 2 snooped
 547          * 1 system
 548          * 0 valid
 549          *
 550          * PDE format on VI:
 551          * 63:59 block fragment size
 552          * 58:40 reserved
 553          * 39:1 physical base address of PTE
 554          * bits 5:1 must be 0.
 555          * 0 valid
 556          */
 557         value = addr & 0x000000FFFFFFF000ULL;
 558         value |= flags;
 559         writeq(value, ptr + (gpu_page_idx * 8));
 560
 561         return 0;
 562 }
 563
 564 /**
 565  * gmc_v8_0_set_fault_enable_default - update VM fault handling
 566  *
 567  * @adev: amdgpu_device pointer
 568  * @value: true redirects VM faults to the default page
 569  */
 570 static void gmc_v8_0_set_fault_enable_default(struct amdgpu_device *adev,
 571                                               bool value)
 572 {
 573         u32 tmp;
 574
 575         tmp = RREG32(mmVM_CONTEXT1_CNTL);
 576         tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
 577                             RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, value);
 578         tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
 579                             DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT, value);
 580         tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
 581                             PDE0_PROTECTION_FAULT_ENABLE_DEFAULT, value);
 582         tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
 583                             VALID_PROTECTION_FAULT_ENABLE_DEFAULT, value);
 584         tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
 585                             READ_PROTECTION_FAULT_ENABLE_DEFAULT, value);
 586         tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
 587                             WRITE_PROTECTION_FAULT_ENABLE_DEFAULT, value);
 588         tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
 589                             EXECUTE_PROTECTION_FAULT_ENABLE_DEFAULT, value);
 590         WREG32(mmVM_CONTEXT1_CNTL, tmp);
 591 }
 592
 593 /**
 594  * gmc_v8_0_gart_enable - gart enable
 595  *
 596  * @adev: amdgpu_device pointer
 597  *
 598  * This sets up the TLBs, programs the page tables for VMID0,
 599  * sets up the hw for VMIDs 1-15 which are allocated on
 600  * demand, and sets up the global locations for the LDS, GDS,
 601  * and GPUVM for FSA64 clients (CIK).
 602  * Returns 0 for success, errors for failure.
 603  */
 604 static int gmc_v8_0_gart_enable(struct amdgpu_device *adev)
 605 {
 606         int r, i;
 607         u32 tmp;
 608
 609         if (adev->gart.robj == NULL) {
 610                 dev_err(adev->dev, "No VRAM object for PCIE GART.\n");
 611                 return -EINVAL;
 612         }
 613         r = amdgpu_gart_table_vram_pin(adev);
 614         if (r)
 615                 return r;
 616         /* Setup TLB control */
 617         tmp = RREG32(mmMC_VM_MX_L1_TLB_CNTL);
 618         tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_TLB, 1);
 619         tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_FRAGMENT_PROCESSING, 1);
 620         tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, SYSTEM_ACCESS_MODE, 3);
 621         tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_ADVANCED_DRIVER_MODEL, 1);
 622         tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, SYSTEM_APERTURE_UNMAPPED_ACCESS, 0);
 623         WREG32(mmMC_VM_MX_L1_TLB_CNTL, tmp);
 624         /* Setup L2 cache */
 625         tmp = RREG32(mmVM_L2_CNTL);
 626         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 1);
 627         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING, 1);
 628         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE, 1);
 629         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE, 1);
 630         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, EFFECTIVE_L2_QUEUE_SIZE, 7);
 631         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, CONTEXT1_IDENTITY_ACCESS_MODE, 1);
 632         WREG32(mmVM_L2_CNTL, tmp);
 633         tmp = RREG32(mmVM_L2_CNTL2);
 634         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL2, INVALIDATE_ALL_L1_TLBS, 1);
 635         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL2, INVALIDATE_L2_CACHE, 1);
 636         WREG32(mmVM_L2_CNTL2, tmp);
 637         tmp = RREG32(mmVM_L2_CNTL3);
 638         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY, 1);
 639         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, BANK_SELECT, 4);
 640         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_FRAGMENT_SIZE, 4);
 641         WREG32(mmVM_L2_CNTL3, tmp);
 642         /* XXX: set to enable PTE/PDE in system memory */
 643         tmp = RREG32(mmVM_L2_CNTL4);
 644         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PDE_REQUEST_PHYSICAL, 0);
 645         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PDE_REQUEST_SHARED, 0);
 646         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PDE_REQUEST_SNOOP, 0);
 647         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PTE_REQUEST_PHYSICAL, 0);
 648         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PTE_REQUEST_SHARED, 0);
 649         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PTE_REQUEST_SNOOP, 0);
 650         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PDE_REQUEST_PHYSICAL, 0);
 651         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PDE_REQUEST_SHARED, 0);
 652         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PDE_REQUEST_SNOOP, 0);
 653         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PTE_REQUEST_PHYSICAL, 0);
 654         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PTE_REQUEST_SHARED, 0);
 655         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PTE_REQUEST_SNOOP, 0);
 656         WREG32(mmVM_L2_CNTL4, tmp);
 657         /* setup context0 */
 658         WREG32(mmVM_CONTEXT0_PAGE_TABLE_START_ADDR, adev->mc.gtt_start >> 12);
 659         WREG32(mmVM_CONTEXT0_PAGE_TABLE_END_ADDR, (adev->mc.gtt_end >> 12) - 1);
 660         WREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR, adev->gart.table_addr >> 12);
 661         WREG32(mmVM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR,
 662                         (u32)(adev->dummy_page.addr >> 12));
 663         WREG32(mmVM_CONTEXT0_CNTL2, 0);
 664         tmp = RREG32(mmVM_CONTEXT0_CNTL);
 665         tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, ENABLE_CONTEXT, 1);
 666         tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, PAGE_TABLE_DEPTH, 0);
 667         tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
 668         WREG32(mmVM_CONTEXT0_CNTL, tmp);
 669
 670         WREG32(mmVM_L2_CONTEXT1_IDENTITY_APERTURE_LOW_ADDR, 0);
 671         WREG32(mmVM_L2_CONTEXT1_IDENTITY_APERTURE_HIGH_ADDR, 0);
 672         WREG32(mmVM_L2_CONTEXT_IDENTITY_PHYSICAL_OFFSET, 0);
 673
 674         /* empty context1-15 */
 675         /* FIXME start with 4G, once using 2 level pt switch to full
 676          * vm size space
 677          */
 678         /* set vm size, must be a multiple of 4 */
 679         WREG32(mmVM_CONTEXT1_PAGE_TABLE_START_ADDR, 0);
 680         WREG32(mmVM_CONTEXT1_PAGE_TABLE_END_ADDR, adev->vm_manager.max_pfn - 1);
 681         for (i = 1; i < 16; i++) {
 682                 if (i < 8)
 683                         WREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + i,
 684                                adev->gart.table_addr >> 12);
 685                 else
 686                         WREG32(mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + i - 8,
 687                                adev->gart.table_addr >> 12);
 688         }
 689
 690         /* enable context1-15 */
 691         WREG32(mmVM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR,
 692                (u32)(adev->dummy_page.addr >> 12));
 693         WREG32(mmVM_CONTEXT1_CNTL2, 4);
 694         tmp = RREG32(mmVM_CONTEXT1_CNTL);
 695         tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, ENABLE_CONTEXT, 1);
 696         tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_DEPTH, 1);
 697         tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
 698         tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
 699         tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PDE0_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
 700         tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, VALID_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
 701         tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, READ_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
 702         tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, WRITE_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
 703         tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, EXECUTE_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
 704         tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_BLOCK_SIZE,
 705                             amdgpu_vm_block_size - 9);
 706         WREG32(mmVM_CONTEXT1_CNTL, tmp);
 707         if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS)
 708                 gmc_v8_0_set_fault_enable_default(adev, false);
 709         else
 710                 gmc_v8_0_set_fault_enable_default(adev, true);
 711
 712         gmc_v8_0_gart_flush_gpu_tlb(adev, 0);
 713         DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
 714                  (unsigned)(adev->mc.gtt_size >> 20),
 715                  (unsigned long long)adev->gart.table_addr);
 716         adev->gart.ready = true;
 717         return 0;
 718 }
 719
 720 static int gmc_v8_0_gart_init(struct amdgpu_device *adev)
 721 {
 722         int r;
 723
 724         if (adev->gart.robj) {
 725                 WARN(1, "R600 PCIE GART already initialized\n");
 726                 return 0;
 727         }
 728         /* Initialize common gart structure */
 729         r = amdgpu_gart_init(adev);
 730         if (r)
 731                 return r;
 732         adev->gart.table_size = adev->gart.num_gpu_pages * 8;
 733         return amdgpu_gart_table_vram_alloc(adev);
 734 }
 735
 736 /**
 737  * gmc_v8_0_gart_disable - gart disable
 738  *
 739  * @adev: amdgpu_device pointer
 740  *
 741  * This disables all VM page table (CIK).
 742  */
 743 static void gmc_v8_0_gart_disable(struct amdgpu_device *adev)
 744 {
 745         u32 tmp;
 746
 747         /* Disable all tables */
 748         WREG32(mmVM_CONTEXT0_CNTL, 0);
 749         WREG32(mmVM_CONTEXT1_CNTL, 0);
 750         /* Setup TLB control */
 751         tmp = RREG32(mmMC_VM_MX_L1_TLB_CNTL);
 752         tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_TLB, 0);
 753         tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_FRAGMENT_PROCESSING, 0);
 754         tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_ADVANCED_DRIVER_MODEL, 0);
 755         WREG32(mmMC_VM_MX_L1_TLB_CNTL, tmp);
 756         /* Setup L2 cache */
 757         tmp = RREG32(mmVM_L2_CNTL);
 758         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 0);
 759         WREG32(mmVM_L2_CNTL, tmp);
 760         WREG32(mmVM_L2_CNTL2, 0);
 761         amdgpu_gart_table_vram_unpin(adev);
 762 }
 763
 764 /**
 765  * gmc_v8_0_gart_fini - vm fini callback
 766  *
 767  * @adev: amdgpu_device pointer
 768  *
 769  * Tears down the driver GART/VM setup (CIK).
 770  */
 771 static void gmc_v8_0_gart_fini(struct amdgpu_device *adev)
 772 {
 773         amdgpu_gart_table_vram_free(adev);
 774         amdgpu_gart_fini(adev);
 775 }
 776
 777 /*
 778  * vm
 779  * VMID 0 is the physical GPU addresses as used by the kernel.
 780  * VMIDs 1-15 are used for userspace clients and are handled
 781  * by the amdgpu vm/hsa code.
 782  */
 783 /**
 784  * gmc_v8_0_vm_init - cik vm init callback
 785  *
 786  * @adev: amdgpu_device pointer
 787  *
 788  * Inits cik specific vm parameters (number of VMs, base of vram for
 789  * VMIDs 1-15) (CIK).
 790  * Returns 0 for success.
 791  */
 792 static int gmc_v8_0_vm_init(struct amdgpu_device *adev)
 793 {
 794         /*
 795          * number of VMs
 796          * VMID 0 is reserved for System
 797          * amdgpu graphics/compute will use VMIDs 1-7
 798          * amdkfd will use VMIDs 8-15
 799          */
 800         adev->vm_manager.nvm = AMDGPU_NUM_OF_VMIDS;
 801
 802         /* base offset of vram pages */
 803         if (adev->flags & AMD_IS_APU) {
 804                 u64 tmp = RREG32(mmMC_VM_FB_OFFSET);
 805                 tmp <<= 22;
 806                 adev->vm_manager.vram_base_offset = tmp;
 807         } else
 808                 adev->vm_manager.vram_base_offset = 0;
 809
 810         return 0;
 811 }
 812
 813 /**
 814  * gmc_v8_0_vm_fini - cik vm fini callback
 815  *
 816  * @adev: amdgpu_device pointer
 817  *
 818  * Tear down any asic specific VM setup (CIK).
 819  */
 820 static void gmc_v8_0_vm_fini(struct amdgpu_device *adev)
 821 {
 822 }
 823
 824 /**
 825  * gmc_v8_0_vm_decode_fault - print human readable fault info
 826  *
 827  * @adev: amdgpu_device pointer
 828  * @status: VM_CONTEXT1_PROTECTION_FAULT_STATUS register value
 829  * @addr: VM_CONTEXT1_PROTECTION_FAULT_ADDR register value
 830  *
 831  * Print human readable fault information (CIK).
 832  */
 833 static void gmc_v8_0_vm_decode_fault(struct amdgpu_device *adev,
 834                                      u32 status, u32 addr, u32 mc_client)
 835 {
 836         u32 mc_id;
 837         u32 vmid = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, VMID);
 838         u32 protections = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS,
 839                                         PROTECTIONS);
 840         char block[5] = { mc_client >> 24, (mc_client >> 16) & 0xff,
 841                 (mc_client >> 8) & 0xff, mc_client & 0xff, 0 };
 842
 843         mc_id = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS,
 844                               MEMORY_CLIENT_ID);
 845
 846         printk("VM fault (0x%02x, vmid %d) at page %u, %s from '%s' (0x%08x) (%d)\n",
 847                protections, vmid, addr,
 848                REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS,
 849                              MEMORY_CLIENT_RW) ?
 850                "write" : "read", block, mc_client, mc_id);
 851 }
 852
 853 static int gmc_v8_0_convert_vram_type(int mc_seq_vram_type)
 854 {
 855         switch (mc_seq_vram_type) {
 856         case MC_SEQ_MISC0__MT__GDDR1:
 857                 return AMDGPU_VRAM_TYPE_GDDR1;
 858         case MC_SEQ_MISC0__MT__DDR2:
 859                 return AMDGPU_VRAM_TYPE_DDR2;
 860         case MC_SEQ_MISC0__MT__GDDR3:
 861                 return AMDGPU_VRAM_TYPE_GDDR3;
 862         case MC_SEQ_MISC0__MT__GDDR4:
 863                 return AMDGPU_VRAM_TYPE_GDDR4;
 864         case MC_SEQ_MISC0__MT__GDDR5:
 865                 return AMDGPU_VRAM_TYPE_GDDR5;
 866         case MC_SEQ_MISC0__MT__HBM:
 867                 return AMDGPU_VRAM_TYPE_HBM;
 868         case MC_SEQ_MISC0__MT__DDR3:
 869                 return AMDGPU_VRAM_TYPE_DDR3;
 870         default:
 871                 return AMDGPU_VRAM_TYPE_UNKNOWN;
 872         }
 873 }
 874
 875 static int gmc_v8_0_early_init(void *handle)
 876 {
 877         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 878
 879         gmc_v8_0_set_gart_funcs(adev);
 880         gmc_v8_0_set_irq_funcs(adev);
 881
 882         if (adev->flags & AMD_IS_APU) {
 883                 adev->mc.vram_type = AMDGPU_VRAM_TYPE_UNKNOWN;
 884         } else {
 885                 u32 tmp = RREG32(mmMC_SEQ_MISC0);
 886                 tmp &= MC_SEQ_MISC0__MT__MASK;
 887                 adev->mc.vram_type = gmc_v8_0_convert_vram_type(tmp);
 888         }
 889
 890         return 0;
 891 }
 892
 893 static int gmc_v8_0_late_init(void *handle)
 894 {
 895         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 896
 897         return amdgpu_irq_get(adev, &adev->mc.vm_fault, 0);
 898 }
 899
 900 static int gmc_v8_0_sw_init(void *handle)
 901 {
 902         int r;
 903         int dma_bits;
 904         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 905
 906         r = amdgpu_gem_init(adev);
 907         if (r)
 908                 return r;
 909
 910         r = amdgpu_irq_add_id(adev, 146, &adev->mc.vm_fault);
 911         if (r)
 912                 return r;
 913
 914         r = amdgpu_irq_add_id(adev, 147, &adev->mc.vm_fault);
 915         if (r)
 916                 return r;
 917
 918         /* Adjust VM size here.
 919          * Currently set to 4GB ((1 << 20) 4k pages).
 920          * Max GPUVM size for cayman and SI is 40 bits.
 921          */
 922         adev->vm_manager.max_pfn = amdgpu_vm_size << 18;
 923
 924         /* Set the internal MC address mask
 925          * This is the max address of the GPU's
 926          * internal address space.
 927          */
 928         adev->mc.mc_mask = 0xffffffffffULL; /* 40 bit MC */
 929
 930         /* set DMA mask + need_dma32 flags.
 931          * PCIE - can handle 40-bits.
 932          * IGP - can handle 40-bits
 933          * PCI - dma32 for legacy pci gart, 40 bits on newer asics
 934          */
 935         adev->need_dma32 = false;
 936         dma_bits = adev->need_dma32 ? 32 : 40;
 937         r = pci_set_dma_mask(adev->pdev, DMA_BIT_MASK(dma_bits));
 938         if (r) {
 939                 adev->need_dma32 = true;
 940                 dma_bits = 32;
 941                 printk(KERN_WARNING "amdgpu: No suitable DMA available.\n");
 942         }
 943         r = pci_set_consistent_dma_mask(adev->pdev, DMA_BIT_MASK(dma_bits));
 944         if (r) {
 945                 pci_set_consistent_dma_mask(adev->pdev, DMA_BIT_MASK(32));
 946                 printk(KERN_WARNING "amdgpu: No coherent DMA available.\n");
 947         }
 948
 949         r = gmc_v8_0_init_microcode(adev);
 950         if (r) {
 951                 DRM_ERROR("Failed to load mc firmware!\n");
 952                 return r;
 953         }
 954
 955         r = gmc_v8_0_mc_init(adev);
 956         if (r)
 957                 return r;
 958
 959         /* Memory manager */
 960         r = amdgpu_bo_init(adev);
 961         if (r)
 962                 return r;
 963
 964         r = gmc_v8_0_gart_init(adev);
 965         if (r)
 966                 return r;
 967
 968         if (!adev->vm_manager.enabled) {
 969                 r = gmc_v8_0_vm_init(adev);
 970                 if (r) {
 971                         dev_err(adev->dev, "vm manager initialization failed (%d).\n", r);
 972                         return r;
 973                 }
 974                 adev->vm_manager.enabled = true;
 975         }
 976
 977         return r;
 978 }
 979
 980 static int gmc_v8_0_sw_fini(void *handle)
 981 {
 982         int i;
 983         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 984
 985         if (adev->vm_manager.enabled) {
 986                 for (i = 0; i < AMDGPU_NUM_VM; ++i)
 987                         fence_put(adev->vm_manager.active[i]);
 988                 gmc_v8_0_vm_fini(adev);
 989                 adev->vm_manager.enabled = false;
 990         }
 991         gmc_v8_0_gart_fini(adev);
 992         amdgpu_gem_fini(adev);
 993         amdgpu_bo_fini(adev);
 994
 995         return 0;
 996 }
 997
 998 static int gmc_v8_0_hw_init(void *handle)
 999 {
1000         int r;
1001         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1002
1003         gmc_v8_0_init_golden_registers(adev);
1004
1005         gmc_v8_0_mc_program(adev);
1006
1007         if (!(adev->flags & AMD_IS_APU)) {
1008                 r = gmc_v8_0_mc_load_microcode(adev);
1009                 if (r) {
1010                         DRM_ERROR("Failed to load MC firmware!\n");
1011                         return r;
1012                 }
1013         }
1014
1015         r = gmc_v8_0_gart_enable(adev);
1016         if (r)
1017                 return r;
1018
1019         return r;
1020 }
1021
1022 static int gmc_v8_0_hw_fini(void *handle)
1023 {
1024         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1025
1026         amdgpu_irq_put(adev, &adev->mc.vm_fault, 0);
1027         gmc_v8_0_gart_disable(adev);
1028
1029         return 0;
1030 }
1031
1032 static int gmc_v8_0_suspend(void *handle)
1033 {
1034         int i;
1035         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1036
1037         if (adev->vm_manager.enabled) {
1038                 for (i = 0; i < AMDGPU_NUM_VM; ++i)
1039                         fence_put(adev->vm_manager.active[i]);
1040                 gmc_v8_0_vm_fini(adev);
1041                 adev->vm_manager.enabled = false;
1042         }
1043         gmc_v8_0_hw_fini(adev);
1044
1045         return 0;
1046 }
1047
1048 static int gmc_v8_0_resume(void *handle)
1049 {
1050         int r;
1051         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1052
1053         r = gmc_v8_0_hw_init(adev);
1054         if (r)
1055                 return r;
1056
1057         if (!adev->vm_manager.enabled) {
1058                 r = gmc_v8_0_vm_init(adev);
1059                 if (r) {
1060                         dev_err(adev->dev, "vm manager initialization failed (%d).\n", r);
1061                         return r;
1062                 }
1063                 adev->vm_manager.enabled = true;
1064         }
1065
1066         return r;
1067 }
1068
1069 static bool gmc_v8_0_is_idle(void *handle)
1070 {
1071         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1072         u32 tmp = RREG32(mmSRBM_STATUS);
1073
1074         if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK |
1075                    SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK | SRBM_STATUS__VMC_BUSY_MASK))
1076                 return false;
1077
1078         return true;
1079 }
1080
1081 static int gmc_v8_0_wait_for_idle(void *handle)
1082 {
1083         unsigned i;
1084         u32 tmp;
1085         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1086
1087         for (i = 0; i < adev->usec_timeout; i++) {
1088                 /* read MC_STATUS */
1089                 tmp = RREG32(mmSRBM_STATUS) & (SRBM_STATUS__MCB_BUSY_MASK |
1090                                                SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK |
1091                                                SRBM_STATUS__MCC_BUSY_MASK |
1092                                                SRBM_STATUS__MCD_BUSY_MASK |
1093                                                SRBM_STATUS__VMC_BUSY_MASK |
1094                                                SRBM_STATUS__VMC1_BUSY_MASK);
1095                 if (!tmp)
1096                         return 0;
1097                 udelay(1);
1098         }
1099         return -ETIMEDOUT;
1100
1101 }
1102
1103 static void gmc_v8_0_print_status(void *handle)
1104 {
1105         int i, j;
1106         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1107
1108         dev_info(adev->dev, "GMC 8.x registers\n");
1109         dev_info(adev->dev, "  SRBM_STATUS=0x%08X\n",
1110                 RREG32(mmSRBM_STATUS));
1111         dev_info(adev->dev, "  SRBM_STATUS2=0x%08X\n",
1112                 RREG32(mmSRBM_STATUS2));
1113
1114         dev_info(adev->dev, "  VM_CONTEXT1_PROTECTION_FAULT_ADDR   0x%08X\n",
1115                  RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_ADDR));
1116         dev_info(adev->dev, "  VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n",
1117                  RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_STATUS));
1118         dev_info(adev->dev, "  MC_VM_MX_L1_TLB_CNTL=0x%08X\n",
1119                  RREG32(mmMC_VM_MX_L1_TLB_CNTL));
1120         dev_info(adev->dev, "  VM_L2_CNTL=0x%08X\n",
1121                  RREG32(mmVM_L2_CNTL));
1122         dev_info(adev->dev, "  VM_L2_CNTL2=0x%08X\n",
1123                  RREG32(mmVM_L2_CNTL2));
1124         dev_info(adev->dev, "  VM_L2_CNTL3=0x%08X\n",
1125                  RREG32(mmVM_L2_CNTL3));
1126         dev_info(adev->dev, "  VM_L2_CNTL4=0x%08X\n",
1127                  RREG32(mmVM_L2_CNTL4));
1128         dev_info(adev->dev, "  VM_CONTEXT0_PAGE_TABLE_START_ADDR=0x%08X\n",
1129                  RREG32(mmVM_CONTEXT0_PAGE_TABLE_START_ADDR));
1130         dev_info(adev->dev, "  VM_CONTEXT0_PAGE_TABLE_END_ADDR=0x%08X\n",
1131                  RREG32(mmVM_CONTEXT0_PAGE_TABLE_END_ADDR));
1132         dev_info(adev->dev, "  VM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR=0x%08X\n",
1133                  RREG32(mmVM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR));
1134         dev_info(adev->dev, "  VM_CONTEXT0_CNTL2=0x%08X\n",
1135                  RREG32(mmVM_CONTEXT0_CNTL2));
1136         dev_info(adev->dev, "  VM_CONTEXT0_CNTL=0x%08X\n",
1137                  RREG32(mmVM_CONTEXT0_CNTL));
1138         dev_info(adev->dev, "  VM_L2_CONTEXT1_IDENTITY_APERTURE_LOW_ADDR=0x%08X\n",
1139                  RREG32(mmVM_L2_CONTEXT1_IDENTITY_APERTURE_LOW_ADDR));
1140         dev_info(adev->dev, "  VM_L2_CONTEXT1_IDENTITY_APERTURE_HIGH_ADDR=0x%08X\n",
1141                  RREG32(mmVM_L2_CONTEXT1_IDENTITY_APERTURE_HIGH_ADDR));
1142         dev_info(adev->dev, "  mmVM_L2_CONTEXT_IDENTITY_PHYSICAL_OFFSET=0x%08X\n",
1143                  RREG32(mmVM_L2_CONTEXT_IDENTITY_PHYSICAL_OFFSET));
1144         dev_info(adev->dev, "  VM_CONTEXT1_PAGE_TABLE_START_ADDR=0x%08X\n",
1145                  RREG32(mmVM_CONTEXT1_PAGE_TABLE_START_ADDR));
1146         dev_info(adev->dev, "  VM_CONTEXT1_PAGE_TABLE_END_ADDR=0x%08X\n",
1147                  RREG32(mmVM_CONTEXT1_PAGE_TABLE_END_ADDR));
1148         dev_info(adev->dev, "  VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR=0x%08X\n",
1149                  RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR));
1150         dev_info(adev->dev, "  VM_CONTEXT1_CNTL2=0x%08X\n",
1151                  RREG32(mmVM_CONTEXT1_CNTL2));
1152         dev_info(adev->dev, "  VM_CONTEXT1_CNTL=0x%08X\n",
1153                  RREG32(mmVM_CONTEXT1_CNTL));
1154         for (i = 0; i < 16; i++) {
1155                 if (i < 8)
1156                         dev_info(adev->dev, "  VM_CONTEXT%d_PAGE_TABLE_BASE_ADDR=0x%08X\n",
1157                                  i, RREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + i));
1158                 else
1159                         dev_info(adev->dev, "  VM_CONTEXT%d_PAGE_TABLE_BASE_ADDR=0x%08X\n",
1160                                  i, RREG32(mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + i - 8));
1161         }
1162         dev_info(adev->dev, "  MC_VM_SYSTEM_APERTURE_LOW_ADDR=0x%08X\n",
1163                  RREG32(mmMC_VM_SYSTEM_APERTURE_LOW_ADDR));
1164         dev_info(adev->dev, "  MC_VM_SYSTEM_APERTURE_HIGH_ADDR=0x%08X\n",
1165                  RREG32(mmMC_VM_SYSTEM_APERTURE_HIGH_ADDR));
1166         dev_info(adev->dev, "  MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR=0x%08X\n",
1167                  RREG32(mmMC_VM_SYSTEM_APERTURE_DEFAULT_ADDR));
1168         dev_info(adev->dev, "  MC_VM_FB_LOCATION=0x%08X\n",
1169                  RREG32(mmMC_VM_FB_LOCATION));
1170         dev_info(adev->dev, "  MC_VM_AGP_BASE=0x%08X\n",
1171                  RREG32(mmMC_VM_AGP_BASE));
1172         dev_info(adev->dev, "  MC_VM_AGP_TOP=0x%08X\n",
1173                  RREG32(mmMC_VM_AGP_TOP));
1174         dev_info(adev->dev, "  MC_VM_AGP_BOT=0x%08X\n",
1175                  RREG32(mmMC_VM_AGP_BOT));
1176
1177         dev_info(adev->dev, "  HDP_REG_COHERENCY_FLUSH_CNTL=0x%08X\n",
1178                  RREG32(mmHDP_REG_COHERENCY_FLUSH_CNTL));
1179         dev_info(adev->dev, "  HDP_NONSURFACE_BASE=0x%08X\n",
1180                  RREG32(mmHDP_NONSURFACE_BASE));
1181         dev_info(adev->dev, "  HDP_NONSURFACE_INFO=0x%08X\n",
1182                  RREG32(mmHDP_NONSURFACE_INFO));
1183         dev_info(adev->dev, "  HDP_NONSURFACE_SIZE=0x%08X\n",
1184                  RREG32(mmHDP_NONSURFACE_SIZE));
1185         dev_info(adev->dev, "  HDP_MISC_CNTL=0x%08X\n",
1186                  RREG32(mmHDP_MISC_CNTL));
1187         dev_info(adev->dev, "  HDP_HOST_PATH_CNTL=0x%08X\n",
1188                  RREG32(mmHDP_HOST_PATH_CNTL));
1189
1190         for (i = 0, j = 0; i < 32; i++, j += 0x6) {
1191                 dev_info(adev->dev, "  %d:\n", i);
1192                 dev_info(adev->dev, "  0x%04X=0x%08X\n",
1193                          0xb05 + j, RREG32(0xb05 + j));
1194                 dev_info(adev->dev, "  0x%04X=0x%08X\n",
1195                          0xb06 + j, RREG32(0xb06 + j));
1196                 dev_info(adev->dev, "  0x%04X=0x%08X\n",
1197                          0xb07 + j, RREG32(0xb07 + j));
1198                 dev_info(adev->dev, "  0x%04X=0x%08X\n",
1199                          0xb08 + j, RREG32(0xb08 + j));
1200                 dev_info(adev->dev, "  0x%04X=0x%08X\n",
1201                          0xb09 + j, RREG32(0xb09 + j));
1202         }
1203
1204         dev_info(adev->dev, "  BIF_FB_EN=0x%08X\n",
1205                  RREG32(mmBIF_FB_EN));
1206 }
1207
1208 static int gmc_v8_0_soft_reset(void *handle)
1209 {
1210         struct amdgpu_mode_mc_save save;
1211         u32 srbm_soft_reset = 0;
1212         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1213         u32 tmp = RREG32(mmSRBM_STATUS);
1214
1215         if (tmp & SRBM_STATUS__VMC_BUSY_MASK)
1216                 srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset,
1217                                                 SRBM_SOFT_RESET, SOFT_RESET_VMC, 1);
1218
1219         if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK |
1220                    SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK)) {
1221                 if (!(adev->flags & AMD_IS_APU))
1222                         srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset,
1223                                                         SRBM_SOFT_RESET, SOFT_RESET_MC, 1);
1224         }
1225
1226         if (srbm_soft_reset) {
1227                 gmc_v8_0_print_status((void *)adev);
1228
1229                 gmc_v8_0_mc_stop(adev, &save);
1230                 if (gmc_v8_0_wait_for_idle(adev)) {
1231                         dev_warn(adev->dev, "Wait for GMC idle timed out !\n");
1232                 }
1233
1234
1235                 tmp = RREG32(mmSRBM_SOFT_RESET);
1236                 tmp |= srbm_soft_reset;
1237                 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1238                 WREG32(mmSRBM_SOFT_RESET, tmp);
1239                 tmp = RREG32(mmSRBM_SOFT_RESET);
1240
1241                 udelay(50);
1242
1243                 tmp &= ~srbm_soft_reset;
1244                 WREG32(mmSRBM_SOFT_RESET, tmp);
1245                 tmp = RREG32(mmSRBM_SOFT_RESET);
1246
1247                 /* Wait a little for things to settle down */
1248                 udelay(50);
1249
1250                 gmc_v8_0_mc_resume(adev, &save);
1251                 udelay(50);
1252
1253                 gmc_v8_0_print_status((void *)adev);
1254         }
1255
1256         return 0;
1257 }
1258
1259 static int gmc_v8_0_vm_fault_interrupt_state(struct amdgpu_device *adev,
1260                                              struct amdgpu_irq_src *src,
1261                                              unsigned type,
1262                                              enum amdgpu_interrupt_state state)
1263 {
1264         u32 tmp;
1265         u32 bits = (VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1266                     VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1267                     VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1268                     VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1269                     VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1270                     VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1271                     VM_CONTEXT1_CNTL__EXECUTE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK);
1272
1273         switch (state) {
1274         case AMDGPU_IRQ_STATE_DISABLE:
1275                 /* system context */
1276                 tmp = RREG32(mmVM_CONTEXT0_CNTL);
1277                 tmp &= ~bits;
1278                 WREG32(mmVM_CONTEXT0_CNTL, tmp);
1279                 /* VMs */
1280                 tmp = RREG32(mmVM_CONTEXT1_CNTL);
1281                 tmp &= ~bits;
1282                 WREG32(mmVM_CONTEXT1_CNTL, tmp);
1283                 break;
1284         case AMDGPU_IRQ_STATE_ENABLE:
1285                 /* system context */
1286                 tmp = RREG32(mmVM_CONTEXT0_CNTL);
1287                 tmp |= bits;
1288                 WREG32(mmVM_CONTEXT0_CNTL, tmp);
1289                 /* VMs */
1290                 tmp = RREG32(mmVM_CONTEXT1_CNTL);
1291                 tmp |= bits;
1292                 WREG32(mmVM_CONTEXT1_CNTL, tmp);
1293                 break;
1294         default:
1295                 break;
1296         }
1297
1298         return 0;
1299 }
1300
1301 static int gmc_v8_0_process_interrupt(struct amdgpu_device *adev,
1302                                       struct amdgpu_irq_src *source,
1303                                       struct amdgpu_iv_entry *entry)
1304 {
1305         u32 addr, status, mc_client;
1306
1307         addr = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_ADDR);
1308         status = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_STATUS);
1309         mc_client = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_MCCLIENT);
1310         /* reset addr and status */
1311         WREG32_P(mmVM_CONTEXT1_CNTL2, 1, ~1);
1312
1313         if (!addr && !status)
1314                 return 0;
1315
1316         if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_FIRST)
1317                 gmc_v8_0_set_fault_enable_default(adev, false);
1318
1319         dev_err(adev->dev, "GPU fault detected: %d 0x%08x\n",
1320                 entry->src_id, entry->src_data);
1321         dev_err(adev->dev, "  VM_CONTEXT1_PROTECTION_FAULT_ADDR   0x%08X\n",
1322                 addr);
1323         dev_err(adev->dev, "  VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n",
1324                 status);
1325         gmc_v8_0_vm_decode_fault(adev, status, addr, mc_client);
1326
1327         return 0;
1328 }
1329
1330 static int gmc_v8_0_set_clockgating_state(void *handle,
1331                                           enum amd_clockgating_state state)
1332 {
1333         return 0;
1334 }
1335
1336 static int gmc_v8_0_set_powergating_state(void *handle,
1337                                           enum amd_powergating_state state)
1338 {
1339         return 0;
1340 }
1341
1342 const struct amd_ip_funcs gmc_v8_0_ip_funcs = {
1343         .early_init = gmc_v8_0_early_init,
1344         .late_init = gmc_v8_0_late_init,
1345         .sw_init = gmc_v8_0_sw_init,
1346         .sw_fini = gmc_v8_0_sw_fini,
1347         .hw_init = gmc_v8_0_hw_init,
1348         .hw_fini = gmc_v8_0_hw_fini,
1349         .suspend = gmc_v8_0_suspend,
1350         .resume = gmc_v8_0_resume,
1351         .is_idle = gmc_v8_0_is_idle,
1352         .wait_for_idle = gmc_v8_0_wait_for_idle,
1353         .soft_reset = gmc_v8_0_soft_reset,
1354         .print_status = gmc_v8_0_print_status,
1355         .set_clockgating_state = gmc_v8_0_set_clockgating_state,
1356         .set_powergating_state = gmc_v8_0_set_powergating_state,
1357 };
1358
1359 static const struct amdgpu_gart_funcs gmc_v8_0_gart_funcs = {
1360         .flush_gpu_tlb = gmc_v8_0_gart_flush_gpu_tlb,
1361         .set_pte_pde = gmc_v8_0_gart_set_pte_pde,
1362 };
1363
1364 static const struct amdgpu_irq_src_funcs gmc_v8_0_irq_funcs = {
1365         .set = gmc_v8_0_vm_fault_interrupt_state,
1366         .process = gmc_v8_0_process_interrupt,
1367 };
1368
1369 static void gmc_v8_0_set_gart_funcs(struct amdgpu_device *adev)
1370 {
1371         if (adev->gart.gart_funcs == NULL)
1372                 adev->gart.gart_funcs = &gmc_v8_0_gart_funcs;
1373 }
1374
1375 static void gmc_v8_0_set_irq_funcs(struct amdgpu_device *adev)
1376 {
1377         adev->mc.vm_fault.num_types = 1;
1378         adev->mc.vm_fault.funcs = &gmc_v8_0_irq_funcs;
1379 }