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 24 #include <linux/firmware.h> 25 #include <linux/module.h> 26 #include <linux/pci.h> 27 28 #include <drm/drm_cache.h> 29 #include "amdgpu.h" 30 #include "gmc_v8_0.h" 31 #include "amdgpu_ucode.h" 32 #include "amdgpu_amdkfd.h" 33 #include "amdgpu_gem.h" 34 35 #include "gmc/gmc_8_1_d.h" 36 #include "gmc/gmc_8_1_sh_mask.h" 37 38 #include "bif/bif_5_0_d.h" 39 #include "bif/bif_5_0_sh_mask.h" 40 41 #include "oss/oss_3_0_d.h" 42 #include "oss/oss_3_0_sh_mask.h" 43 44 #include "dce/dce_10_0_d.h" 45 #include "dce/dce_10_0_sh_mask.h" 46 47 #include "vid.h" 48 #include "vi.h" 49 50 #include "amdgpu_atombios.h" 51 52 #include "ivsrcid/ivsrcid_vislands30.h" 53 54 static void gmc_v8_0_set_gmc_funcs(struct amdgpu_device *adev); 55 static void gmc_v8_0_set_irq_funcs(struct amdgpu_device *adev); 56 static int gmc_v8_0_wait_for_idle(void *handle); 57 58 MODULE_FIRMWARE("amdgpu/tonga_mc.bin"); 59 MODULE_FIRMWARE("amdgpu/polaris11_mc.bin"); 60 MODULE_FIRMWARE("amdgpu/polaris10_mc.bin"); 61 MODULE_FIRMWARE("amdgpu/polaris12_mc.bin"); 62 MODULE_FIRMWARE("amdgpu/polaris11_k_mc.bin"); 63 MODULE_FIRMWARE("amdgpu/polaris10_k_mc.bin"); 64 MODULE_FIRMWARE("amdgpu/polaris12_k_mc.bin"); 65 66 static const u32 golden_settings_tonga_a11[] = 67 { 68 mmMC_ARB_WTM_GRPWT_RD, 0x00000003, 0x00000000, 69 mmMC_HUB_RDREQ_DMIF_LIMIT, 0x0000007f, 0x00000028, 70 mmMC_HUB_WDP_UMC, 0x00007fb6, 0x00000991, 71 mmVM_PRT_APERTURE0_LOW_ADDR, 0x0fffffff, 0x0fffffff, 72 mmVM_PRT_APERTURE1_LOW_ADDR, 0x0fffffff, 0x0fffffff, 73 mmVM_PRT_APERTURE2_LOW_ADDR, 0x0fffffff, 0x0fffffff, 74 mmVM_PRT_APERTURE3_LOW_ADDR, 0x0fffffff, 0x0fffffff, 75 }; 76 77 static const u32 tonga_mgcg_cgcg_init[] = 78 { 79 mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104 80 }; 81 82 static const u32 golden_settings_fiji_a10[] = 83 { 84 mmVM_PRT_APERTURE0_LOW_ADDR, 0x0fffffff, 0x0fffffff, 85 mmVM_PRT_APERTURE1_LOW_ADDR, 0x0fffffff, 0x0fffffff, 86 mmVM_PRT_APERTURE2_LOW_ADDR, 0x0fffffff, 0x0fffffff, 87 mmVM_PRT_APERTURE3_LOW_ADDR, 0x0fffffff, 0x0fffffff, 88 }; 89 90 static const u32 fiji_mgcg_cgcg_init[] = 91 { 92 mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104 93 }; 94 95 static const u32 golden_settings_polaris11_a11[] = 96 { 97 mmVM_PRT_APERTURE0_LOW_ADDR, 0x0fffffff, 0x0fffffff, 98 mmVM_PRT_APERTURE1_LOW_ADDR, 0x0fffffff, 0x0fffffff, 99 mmVM_PRT_APERTURE2_LOW_ADDR, 0x0fffffff, 0x0fffffff, 100 mmVM_PRT_APERTURE3_LOW_ADDR, 0x0fffffff, 0x0fffffff 101 }; 102 103 static const u32 golden_settings_polaris10_a11[] = 104 { 105 mmMC_ARB_WTM_GRPWT_RD, 0x00000003, 0x00000000, 106 mmVM_PRT_APERTURE0_LOW_ADDR, 0x0fffffff, 0x0fffffff, 107 mmVM_PRT_APERTURE1_LOW_ADDR, 0x0fffffff, 0x0fffffff, 108 mmVM_PRT_APERTURE2_LOW_ADDR, 0x0fffffff, 0x0fffffff, 109 mmVM_PRT_APERTURE3_LOW_ADDR, 0x0fffffff, 0x0fffffff 110 }; 111 112 static const u32 cz_mgcg_cgcg_init[] = 113 { 114 mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104 115 }; 116 117 static const u32 stoney_mgcg_cgcg_init[] = 118 { 119 mmATC_MISC_CG, 0xffffffff, 0x000c0200, 120 mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104 121 }; 122 123 static const u32 golden_settings_stoney_common[] = 124 { 125 mmMC_HUB_RDREQ_UVD, MC_HUB_RDREQ_UVD__PRESCALE_MASK, 0x00000004, 126 mmMC_RD_GRP_OTH, MC_RD_GRP_OTH__UVD_MASK, 0x00600000 127 }; 128 129 static void gmc_v8_0_init_golden_registers(struct amdgpu_device *adev) 130 { 131 switch (adev->asic_type) { 132 case CHIP_FIJI: 133 amdgpu_device_program_register_sequence(adev, 134 fiji_mgcg_cgcg_init, 135 ARRAY_SIZE(fiji_mgcg_cgcg_init)); 136 amdgpu_device_program_register_sequence(adev, 137 golden_settings_fiji_a10, 138 ARRAY_SIZE(golden_settings_fiji_a10)); 139 break; 140 case CHIP_TONGA: 141 amdgpu_device_program_register_sequence(adev, 142 tonga_mgcg_cgcg_init, 143 ARRAY_SIZE(tonga_mgcg_cgcg_init)); 144 amdgpu_device_program_register_sequence(adev, 145 golden_settings_tonga_a11, 146 ARRAY_SIZE(golden_settings_tonga_a11)); 147 break; 148 case CHIP_POLARIS11: 149 case CHIP_POLARIS12: 150 case CHIP_VEGAM: 151 amdgpu_device_program_register_sequence(adev, 152 golden_settings_polaris11_a11, 153 ARRAY_SIZE(golden_settings_polaris11_a11)); 154 break; 155 case CHIP_POLARIS10: 156 amdgpu_device_program_register_sequence(adev, 157 golden_settings_polaris10_a11, 158 ARRAY_SIZE(golden_settings_polaris10_a11)); 159 break; 160 case CHIP_CARRIZO: 161 amdgpu_device_program_register_sequence(adev, 162 cz_mgcg_cgcg_init, 163 ARRAY_SIZE(cz_mgcg_cgcg_init)); 164 break; 165 case CHIP_STONEY: 166 amdgpu_device_program_register_sequence(adev, 167 stoney_mgcg_cgcg_init, 168 ARRAY_SIZE(stoney_mgcg_cgcg_init)); 169 amdgpu_device_program_register_sequence(adev, 170 golden_settings_stoney_common, 171 ARRAY_SIZE(golden_settings_stoney_common)); 172 break; 173 default: 174 break; 175 } 176 } 177 178 static void gmc_v8_0_mc_stop(struct amdgpu_device *adev) 179 { 180 u32 blackout; 181 182 gmc_v8_0_wait_for_idle(adev); 183 184 blackout = RREG32(mmMC_SHARED_BLACKOUT_CNTL); 185 if (REG_GET_FIELD(blackout, MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE) != 1) { 186 /* Block CPU access */ 187 WREG32(mmBIF_FB_EN, 0); 188 /* blackout the MC */ 189 blackout = REG_SET_FIELD(blackout, 190 MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE, 1); 191 WREG32(mmMC_SHARED_BLACKOUT_CNTL, blackout); 192 } 193 /* wait for the MC to settle */ 194 udelay(100); 195 } 196 197 static void gmc_v8_0_mc_resume(struct amdgpu_device *adev) 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 } 210 211 /** 212 * gmc_v8_0_init_microcode - load ucode images from disk 213 * 214 * @adev: amdgpu_device pointer 215 * 216 * Use the firmware interface to load the ucode images into 217 * the driver (not loaded into hw). 218 * Returns 0 on success, error on failure. 219 */ 220 static int gmc_v8_0_init_microcode(struct amdgpu_device *adev) 221 { 222 const char *chip_name; 223 char fw_name[30]; 224 int err; 225 226 DRM_DEBUG("\n"); 227 228 switch (adev->asic_type) { 229 case CHIP_TONGA: 230 chip_name = "tonga"; 231 break; 232 case CHIP_POLARIS11: 233 if (((adev->pdev->device == 0x67ef) && 234 ((adev->pdev->revision == 0xe0) || 235 (adev->pdev->revision == 0xe5))) || 236 ((adev->pdev->device == 0x67ff) && 237 ((adev->pdev->revision == 0xcf) || 238 (adev->pdev->revision == 0xef) || 239 (adev->pdev->revision == 0xff)))) 240 chip_name = "polaris11_k"; 241 else if ((adev->pdev->device == 0x67ef) && 242 (adev->pdev->revision == 0xe2)) 243 chip_name = "polaris11_k"; 244 else 245 chip_name = "polaris11"; 246 break; 247 case CHIP_POLARIS10: 248 if ((adev->pdev->device == 0x67df) && 249 ((adev->pdev->revision == 0xe1) || 250 (adev->pdev->revision == 0xf7))) 251 chip_name = "polaris10_k"; 252 else 253 chip_name = "polaris10"; 254 break; 255 case CHIP_POLARIS12: 256 if (((adev->pdev->device == 0x6987) && 257 ((adev->pdev->revision == 0xc0) || 258 (adev->pdev->revision == 0xc3))) || 259 ((adev->pdev->device == 0x6981) && 260 ((adev->pdev->revision == 0x00) || 261 (adev->pdev->revision == 0x01) || 262 (adev->pdev->revision == 0x10)))) 263 chip_name = "polaris12_k"; 264 else 265 chip_name = "polaris12"; 266 break; 267 case CHIP_FIJI: 268 case CHIP_CARRIZO: 269 case CHIP_STONEY: 270 case CHIP_VEGAM: 271 return 0; 272 default: BUG(); 273 } 274 275 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mc.bin", chip_name); 276 err = request_firmware(&adev->gmc.fw, fw_name, adev->dev); 277 if (err) 278 goto out; 279 err = amdgpu_ucode_validate(adev->gmc.fw); 280 281 out: 282 if (err) { 283 pr_err("mc: Failed to load firmware \"%s\"\n", fw_name); 284 release_firmware(adev->gmc.fw); 285 adev->gmc.fw = NULL; 286 } 287 return err; 288 } 289 290 /** 291 * gmc_v8_0_tonga_mc_load_microcode - load tonga MC ucode into the hw 292 * 293 * @adev: amdgpu_device pointer 294 * 295 * Load the GDDR MC ucode into the hw (VI). 296 * Returns 0 on success, error on failure. 297 */ 298 static int gmc_v8_0_tonga_mc_load_microcode(struct amdgpu_device *adev) 299 { 300 const struct mc_firmware_header_v1_0 *hdr; 301 const __le32 *fw_data = NULL; 302 const __le32 *io_mc_regs = NULL; 303 u32 running; 304 int i, ucode_size, regs_size; 305 306 /* Skip MC ucode loading on SR-IOV capable boards. 307 * vbios does this for us in asic_init in that case. 308 * Skip MC ucode loading on VF, because hypervisor will do that 309 * for this adaptor. 310 */ 311 if (amdgpu_sriov_bios(adev)) 312 return 0; 313 314 if (!adev->gmc.fw) 315 return -EINVAL; 316 317 hdr = (const struct mc_firmware_header_v1_0 *)adev->gmc.fw->data; 318 amdgpu_ucode_print_mc_hdr(&hdr->header); 319 320 adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version); 321 regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2); 322 io_mc_regs = (const __le32 *) 323 (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes)); 324 ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4; 325 fw_data = (const __le32 *) 326 (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes)); 327 328 running = REG_GET_FIELD(RREG32(mmMC_SEQ_SUP_CNTL), MC_SEQ_SUP_CNTL, RUN); 329 330 if (running == 0) { 331 /* reset the engine and set to writable */ 332 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008); 333 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000010); 334 335 /* load mc io regs */ 336 for (i = 0; i < regs_size; i++) { 337 WREG32(mmMC_SEQ_IO_DEBUG_INDEX, le32_to_cpup(io_mc_regs++)); 338 WREG32(mmMC_SEQ_IO_DEBUG_DATA, le32_to_cpup(io_mc_regs++)); 339 } 340 /* load the MC ucode */ 341 for (i = 0; i < ucode_size; i++) 342 WREG32(mmMC_SEQ_SUP_PGM, le32_to_cpup(fw_data++)); 343 344 /* put the engine back into the active state */ 345 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008); 346 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000004); 347 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000001); 348 349 /* wait for training to complete */ 350 for (i = 0; i < adev->usec_timeout; i++) { 351 if (REG_GET_FIELD(RREG32(mmMC_SEQ_TRAIN_WAKEUP_CNTL), 352 MC_SEQ_TRAIN_WAKEUP_CNTL, TRAIN_DONE_D0)) 353 break; 354 udelay(1); 355 } 356 for (i = 0; i < adev->usec_timeout; i++) { 357 if (REG_GET_FIELD(RREG32(mmMC_SEQ_TRAIN_WAKEUP_CNTL), 358 MC_SEQ_TRAIN_WAKEUP_CNTL, TRAIN_DONE_D1)) 359 break; 360 udelay(1); 361 } 362 } 363 364 return 0; 365 } 366 367 static int gmc_v8_0_polaris_mc_load_microcode(struct amdgpu_device *adev) 368 { 369 const struct mc_firmware_header_v1_0 *hdr; 370 const __le32 *fw_data = NULL; 371 const __le32 *io_mc_regs = NULL; 372 u32 data; 373 int i, ucode_size, regs_size; 374 375 /* Skip MC ucode loading on SR-IOV capable boards. 376 * vbios does this for us in asic_init in that case. 377 * Skip MC ucode loading on VF, because hypervisor will do that 378 * for this adaptor. 379 */ 380 if (amdgpu_sriov_bios(adev)) 381 return 0; 382 383 if (!adev->gmc.fw) 384 return -EINVAL; 385 386 hdr = (const struct mc_firmware_header_v1_0 *)adev->gmc.fw->data; 387 amdgpu_ucode_print_mc_hdr(&hdr->header); 388 389 adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version); 390 regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2); 391 io_mc_regs = (const __le32 *) 392 (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes)); 393 ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4; 394 fw_data = (const __le32 *) 395 (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes)); 396 397 data = RREG32(mmMC_SEQ_MISC0); 398 data &= ~(0x40); 399 WREG32(mmMC_SEQ_MISC0, data); 400 401 /* load mc io regs */ 402 for (i = 0; i < regs_size; i++) { 403 WREG32(mmMC_SEQ_IO_DEBUG_INDEX, le32_to_cpup(io_mc_regs++)); 404 WREG32(mmMC_SEQ_IO_DEBUG_DATA, le32_to_cpup(io_mc_regs++)); 405 } 406 407 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008); 408 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000010); 409 410 /* load the MC ucode */ 411 for (i = 0; i < ucode_size; i++) 412 WREG32(mmMC_SEQ_SUP_PGM, le32_to_cpup(fw_data++)); 413 414 /* put the engine back into the active state */ 415 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008); 416 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000004); 417 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000001); 418 419 /* wait for training to complete */ 420 for (i = 0; i < adev->usec_timeout; i++) { 421 data = RREG32(mmMC_SEQ_MISC0); 422 if (data & 0x80) 423 break; 424 udelay(1); 425 } 426 427 return 0; 428 } 429 430 static void gmc_v8_0_vram_gtt_location(struct amdgpu_device *adev, 431 struct amdgpu_gmc *mc) 432 { 433 u64 base = 0; 434 435 if (!amdgpu_sriov_vf(adev)) 436 base = RREG32(mmMC_VM_FB_LOCATION) & 0xFFFF; 437 base <<= 24; 438 439 amdgpu_gmc_vram_location(adev, mc, base); 440 amdgpu_gmc_gart_location(adev, mc); 441 } 442 443 /** 444 * gmc_v8_0_mc_program - program the GPU memory controller 445 * 446 * @adev: amdgpu_device pointer 447 * 448 * Set the location of vram, gart, and AGP in the GPU's 449 * physical address space (VI). 450 */ 451 static void gmc_v8_0_mc_program(struct amdgpu_device *adev) 452 { 453 u32 tmp; 454 int i, j; 455 456 /* Initialize HDP */ 457 for (i = 0, j = 0; i < 32; i++, j += 0x6) { 458 WREG32((0xb05 + j), 0x00000000); 459 WREG32((0xb06 + j), 0x00000000); 460 WREG32((0xb07 + j), 0x00000000); 461 WREG32((0xb08 + j), 0x00000000); 462 WREG32((0xb09 + j), 0x00000000); 463 } 464 WREG32(mmHDP_REG_COHERENCY_FLUSH_CNTL, 0); 465 466 if (gmc_v8_0_wait_for_idle((void *)adev)) { 467 dev_warn(adev->dev, "Wait for MC idle timedout !\n"); 468 } 469 if (adev->mode_info.num_crtc) { 470 /* Lockout access through VGA aperture*/ 471 tmp = RREG32(mmVGA_HDP_CONTROL); 472 tmp = REG_SET_FIELD(tmp, VGA_HDP_CONTROL, VGA_MEMORY_DISABLE, 1); 473 WREG32(mmVGA_HDP_CONTROL, tmp); 474 475 /* disable VGA render */ 476 tmp = RREG32(mmVGA_RENDER_CONTROL); 477 tmp = REG_SET_FIELD(tmp, VGA_RENDER_CONTROL, VGA_VSTATUS_CNTL, 0); 478 WREG32(mmVGA_RENDER_CONTROL, tmp); 479 } 480 /* Update configuration */ 481 WREG32(mmMC_VM_SYSTEM_APERTURE_LOW_ADDR, 482 adev->gmc.vram_start >> 12); 483 WREG32(mmMC_VM_SYSTEM_APERTURE_HIGH_ADDR, 484 adev->gmc.vram_end >> 12); 485 WREG32(mmMC_VM_SYSTEM_APERTURE_DEFAULT_ADDR, 486 adev->vram_scratch.gpu_addr >> 12); 487 488 if (amdgpu_sriov_vf(adev)) { 489 tmp = ((adev->gmc.vram_end >> 24) & 0xFFFF) << 16; 490 tmp |= ((adev->gmc.vram_start >> 24) & 0xFFFF); 491 WREG32(mmMC_VM_FB_LOCATION, tmp); 492 /* XXX double check these! */ 493 WREG32(mmHDP_NONSURFACE_BASE, (adev->gmc.vram_start >> 8)); 494 WREG32(mmHDP_NONSURFACE_INFO, (2 << 7) | (1 << 30)); 495 WREG32(mmHDP_NONSURFACE_SIZE, 0x3FFFFFFF); 496 } 497 498 WREG32(mmMC_VM_AGP_BASE, 0); 499 WREG32(mmMC_VM_AGP_TOP, 0x0FFFFFFF); 500 WREG32(mmMC_VM_AGP_BOT, 0x0FFFFFFF); 501 if (gmc_v8_0_wait_for_idle((void *)adev)) { 502 dev_warn(adev->dev, "Wait for MC idle timedout !\n"); 503 } 504 505 WREG32(mmBIF_FB_EN, BIF_FB_EN__FB_READ_EN_MASK | BIF_FB_EN__FB_WRITE_EN_MASK); 506 507 tmp = RREG32(mmHDP_MISC_CNTL); 508 tmp = REG_SET_FIELD(tmp, HDP_MISC_CNTL, FLUSH_INVALIDATE_CACHE, 0); 509 WREG32(mmHDP_MISC_CNTL, tmp); 510 511 tmp = RREG32(mmHDP_HOST_PATH_CNTL); 512 WREG32(mmHDP_HOST_PATH_CNTL, tmp); 513 } 514 515 /** 516 * gmc_v8_0_mc_init - initialize the memory controller driver params 517 * 518 * @adev: amdgpu_device pointer 519 * 520 * Look up the amount of vram, vram width, and decide how to place 521 * vram and gart within the GPU's physical address space (VI). 522 * Returns 0 for success. 523 */ 524 static int gmc_v8_0_mc_init(struct amdgpu_device *adev) 525 { 526 int r; 527 528 adev->gmc.vram_width = amdgpu_atombios_get_vram_width(adev); 529 if (!adev->gmc.vram_width) { 530 u32 tmp; 531 int chansize, numchan; 532 533 /* Get VRAM informations */ 534 tmp = RREG32(mmMC_ARB_RAMCFG); 535 if (REG_GET_FIELD(tmp, MC_ARB_RAMCFG, CHANSIZE)) { 536 chansize = 64; 537 } else { 538 chansize = 32; 539 } 540 tmp = RREG32(mmMC_SHARED_CHMAP); 541 switch (REG_GET_FIELD(tmp, MC_SHARED_CHMAP, NOOFCHAN)) { 542 case 0: 543 default: 544 numchan = 1; 545 break; 546 case 1: 547 numchan = 2; 548 break; 549 case 2: 550 numchan = 4; 551 break; 552 case 3: 553 numchan = 8; 554 break; 555 case 4: 556 numchan = 3; 557 break; 558 case 5: 559 numchan = 6; 560 break; 561 case 6: 562 numchan = 10; 563 break; 564 case 7: 565 numchan = 12; 566 break; 567 case 8: 568 numchan = 16; 569 break; 570 } 571 adev->gmc.vram_width = numchan * chansize; 572 } 573 /* size in MB on si */ 574 adev->gmc.mc_vram_size = RREG32(mmCONFIG_MEMSIZE) * 1024ULL * 1024ULL; 575 adev->gmc.real_vram_size = RREG32(mmCONFIG_MEMSIZE) * 1024ULL * 1024ULL; 576 577 if (!(adev->flags & AMD_IS_APU)) { 578 r = amdgpu_device_resize_fb_bar(adev); 579 if (r) 580 return r; 581 } 582 adev->gmc.aper_base = pci_resource_start(adev->pdev, 0); 583 adev->gmc.aper_size = pci_resource_len(adev->pdev, 0); 584 585 #ifdef CONFIG_X86_64 586 if (adev->flags & AMD_IS_APU) { 587 adev->gmc.aper_base = ((u64)RREG32(mmMC_VM_FB_OFFSET)) << 22; 588 adev->gmc.aper_size = adev->gmc.real_vram_size; 589 } 590 #endif 591 592 /* In case the PCI BAR is larger than the actual amount of vram */ 593 adev->gmc.visible_vram_size = adev->gmc.aper_size; 594 if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size) 595 adev->gmc.visible_vram_size = adev->gmc.real_vram_size; 596 597 /* set the gart size */ 598 if (amdgpu_gart_size == -1) { 599 switch (adev->asic_type) { 600 case CHIP_POLARIS10: /* all engines support GPUVM */ 601 case CHIP_POLARIS11: /* all engines support GPUVM */ 602 case CHIP_POLARIS12: /* all engines support GPUVM */ 603 case CHIP_VEGAM: /* all engines support GPUVM */ 604 default: 605 adev->gmc.gart_size = 256ULL << 20; 606 break; 607 case CHIP_TONGA: /* UVD, VCE do not support GPUVM */ 608 case CHIP_FIJI: /* UVD, VCE do not support GPUVM */ 609 case CHIP_CARRIZO: /* UVD, VCE do not support GPUVM, DCE SG support */ 610 case CHIP_STONEY: /* UVD does not support GPUVM, DCE SG support */ 611 adev->gmc.gart_size = 1024ULL << 20; 612 break; 613 } 614 } else { 615 adev->gmc.gart_size = (u64)amdgpu_gart_size << 20; 616 } 617 618 gmc_v8_0_vram_gtt_location(adev, &adev->gmc); 619 620 return 0; 621 } 622 623 /* 624 * GART 625 * VMID 0 is the physical GPU addresses as used by the kernel. 626 * VMIDs 1-15 are used for userspace clients and are handled 627 * by the amdgpu vm/hsa code. 628 */ 629 630 /** 631 * gmc_v8_0_flush_gpu_tlb - gart tlb flush callback 632 * 633 * @adev: amdgpu_device pointer 634 * @vmid: vm instance to flush 635 * 636 * Flush the TLB for the requested page table (VI). 637 */ 638 static void gmc_v8_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid, 639 uint32_t vmhub, uint32_t flush_type) 640 { 641 /* bits 0-15 are the VM contexts0-15 */ 642 WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid); 643 } 644 645 static uint64_t gmc_v8_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring, 646 unsigned vmid, uint64_t pd_addr) 647 { 648 uint32_t reg; 649 650 if (vmid < 8) 651 reg = mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + vmid; 652 else 653 reg = mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vmid - 8; 654 amdgpu_ring_emit_wreg(ring, reg, pd_addr >> 12); 655 656 /* bits 0-15 are the VM contexts0-15 */ 657 amdgpu_ring_emit_wreg(ring, mmVM_INVALIDATE_REQUEST, 1 << vmid); 658 659 return pd_addr; 660 } 661 662 static void gmc_v8_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid, 663 unsigned pasid) 664 { 665 amdgpu_ring_emit_wreg(ring, mmIH_VMID_0_LUT + vmid, pasid); 666 } 667 668 /* 669 * PTE format on VI: 670 * 63:40 reserved 671 * 39:12 4k physical page base address 672 * 11:7 fragment 673 * 6 write 674 * 5 read 675 * 4 exe 676 * 3 reserved 677 * 2 snooped 678 * 1 system 679 * 0 valid 680 * 681 * PDE format on VI: 682 * 63:59 block fragment size 683 * 58:40 reserved 684 * 39:1 physical base address of PTE 685 * bits 5:1 must be 0. 686 * 0 valid 687 */ 688 689 static void gmc_v8_0_get_vm_pde(struct amdgpu_device *adev, int level, 690 uint64_t *addr, uint64_t *flags) 691 { 692 BUG_ON(*addr & 0xFFFFFF0000000FFFULL); 693 } 694 695 static void gmc_v8_0_get_vm_pte(struct amdgpu_device *adev, 696 struct amdgpu_bo_va_mapping *mapping, 697 uint64_t *flags) 698 { 699 *flags &= ~AMDGPU_PTE_EXECUTABLE; 700 *flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE; 701 *flags &= ~AMDGPU_PTE_PRT; 702 } 703 704 /** 705 * gmc_v8_0_set_fault_enable_default - update VM fault handling 706 * 707 * @adev: amdgpu_device pointer 708 * @value: true redirects VM faults to the default page 709 */ 710 static void gmc_v8_0_set_fault_enable_default(struct amdgpu_device *adev, 711 bool value) 712 { 713 u32 tmp; 714 715 tmp = RREG32(mmVM_CONTEXT1_CNTL); 716 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 717 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, value); 718 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 719 DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT, value); 720 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 721 PDE0_PROTECTION_FAULT_ENABLE_DEFAULT, value); 722 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 723 VALID_PROTECTION_FAULT_ENABLE_DEFAULT, value); 724 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 725 READ_PROTECTION_FAULT_ENABLE_DEFAULT, value); 726 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 727 WRITE_PROTECTION_FAULT_ENABLE_DEFAULT, value); 728 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 729 EXECUTE_PROTECTION_FAULT_ENABLE_DEFAULT, value); 730 WREG32(mmVM_CONTEXT1_CNTL, tmp); 731 } 732 733 /** 734 * gmc_v8_0_set_prt - set PRT VM fault 735 * 736 * @adev: amdgpu_device pointer 737 * @enable: enable/disable VM fault handling for PRT 738 */ 739 static void gmc_v8_0_set_prt(struct amdgpu_device *adev, bool enable) 740 { 741 u32 tmp; 742 743 if (enable && !adev->gmc.prt_warning) { 744 dev_warn(adev->dev, "Disabling VM faults because of PRT request!\n"); 745 adev->gmc.prt_warning = true; 746 } 747 748 tmp = RREG32(mmVM_PRT_CNTL); 749 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 750 CB_DISABLE_READ_FAULT_ON_UNMAPPED_ACCESS, enable); 751 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 752 CB_DISABLE_WRITE_FAULT_ON_UNMAPPED_ACCESS, enable); 753 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 754 TC_DISABLE_READ_FAULT_ON_UNMAPPED_ACCESS, enable); 755 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 756 TC_DISABLE_WRITE_FAULT_ON_UNMAPPED_ACCESS, enable); 757 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 758 L2_CACHE_STORE_INVALID_ENTRIES, enable); 759 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 760 L1_TLB_STORE_INVALID_ENTRIES, enable); 761 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 762 MASK_PDE0_FAULT, enable); 763 WREG32(mmVM_PRT_CNTL, tmp); 764 765 if (enable) { 766 uint32_t low = AMDGPU_VA_RESERVED_SIZE >> AMDGPU_GPU_PAGE_SHIFT; 767 uint32_t high = adev->vm_manager.max_pfn - 768 (AMDGPU_VA_RESERVED_SIZE >> AMDGPU_GPU_PAGE_SHIFT); 769 770 WREG32(mmVM_PRT_APERTURE0_LOW_ADDR, low); 771 WREG32(mmVM_PRT_APERTURE1_LOW_ADDR, low); 772 WREG32(mmVM_PRT_APERTURE2_LOW_ADDR, low); 773 WREG32(mmVM_PRT_APERTURE3_LOW_ADDR, low); 774 WREG32(mmVM_PRT_APERTURE0_HIGH_ADDR, high); 775 WREG32(mmVM_PRT_APERTURE1_HIGH_ADDR, high); 776 WREG32(mmVM_PRT_APERTURE2_HIGH_ADDR, high); 777 WREG32(mmVM_PRT_APERTURE3_HIGH_ADDR, high); 778 } else { 779 WREG32(mmVM_PRT_APERTURE0_LOW_ADDR, 0xfffffff); 780 WREG32(mmVM_PRT_APERTURE1_LOW_ADDR, 0xfffffff); 781 WREG32(mmVM_PRT_APERTURE2_LOW_ADDR, 0xfffffff); 782 WREG32(mmVM_PRT_APERTURE3_LOW_ADDR, 0xfffffff); 783 WREG32(mmVM_PRT_APERTURE0_HIGH_ADDR, 0x0); 784 WREG32(mmVM_PRT_APERTURE1_HIGH_ADDR, 0x0); 785 WREG32(mmVM_PRT_APERTURE2_HIGH_ADDR, 0x0); 786 WREG32(mmVM_PRT_APERTURE3_HIGH_ADDR, 0x0); 787 } 788 } 789 790 /** 791 * gmc_v8_0_gart_enable - gart enable 792 * 793 * @adev: amdgpu_device pointer 794 * 795 * This sets up the TLBs, programs the page tables for VMID0, 796 * sets up the hw for VMIDs 1-15 which are allocated on 797 * demand, and sets up the global locations for the LDS, GDS, 798 * and GPUVM for FSA64 clients (VI). 799 * Returns 0 for success, errors for failure. 800 */ 801 static int gmc_v8_0_gart_enable(struct amdgpu_device *adev) 802 { 803 uint64_t table_addr; 804 int r, i; 805 u32 tmp, field; 806 807 if (adev->gart.bo == NULL) { 808 dev_err(adev->dev, "No VRAM object for PCIE GART.\n"); 809 return -EINVAL; 810 } 811 r = amdgpu_gart_table_vram_pin(adev); 812 if (r) 813 return r; 814 815 table_addr = amdgpu_bo_gpu_offset(adev->gart.bo); 816 817 /* Setup TLB control */ 818 tmp = RREG32(mmMC_VM_MX_L1_TLB_CNTL); 819 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_TLB, 1); 820 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_FRAGMENT_PROCESSING, 1); 821 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, SYSTEM_ACCESS_MODE, 3); 822 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_ADVANCED_DRIVER_MODEL, 1); 823 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, SYSTEM_APERTURE_UNMAPPED_ACCESS, 0); 824 WREG32(mmMC_VM_MX_L1_TLB_CNTL, tmp); 825 /* Setup L2 cache */ 826 tmp = RREG32(mmVM_L2_CNTL); 827 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 1); 828 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING, 1); 829 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE, 1); 830 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE, 1); 831 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, EFFECTIVE_L2_QUEUE_SIZE, 7); 832 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, CONTEXT1_IDENTITY_ACCESS_MODE, 1); 833 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_DEFAULT_PAGE_OUT_TO_SYSTEM_MEMORY, 1); 834 WREG32(mmVM_L2_CNTL, tmp); 835 tmp = RREG32(mmVM_L2_CNTL2); 836 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL2, INVALIDATE_ALL_L1_TLBS, 1); 837 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL2, INVALIDATE_L2_CACHE, 1); 838 WREG32(mmVM_L2_CNTL2, tmp); 839 840 field = adev->vm_manager.fragment_size; 841 tmp = RREG32(mmVM_L2_CNTL3); 842 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY, 1); 843 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, BANK_SELECT, field); 844 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_FRAGMENT_SIZE, field); 845 WREG32(mmVM_L2_CNTL3, tmp); 846 /* XXX: set to enable PTE/PDE in system memory */ 847 tmp = RREG32(mmVM_L2_CNTL4); 848 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PDE_REQUEST_PHYSICAL, 0); 849 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PDE_REQUEST_SHARED, 0); 850 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PDE_REQUEST_SNOOP, 0); 851 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PTE_REQUEST_PHYSICAL, 0); 852 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PTE_REQUEST_SHARED, 0); 853 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PTE_REQUEST_SNOOP, 0); 854 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PDE_REQUEST_PHYSICAL, 0); 855 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PDE_REQUEST_SHARED, 0); 856 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PDE_REQUEST_SNOOP, 0); 857 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PTE_REQUEST_PHYSICAL, 0); 858 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PTE_REQUEST_SHARED, 0); 859 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PTE_REQUEST_SNOOP, 0); 860 WREG32(mmVM_L2_CNTL4, tmp); 861 /* setup context0 */ 862 WREG32(mmVM_CONTEXT0_PAGE_TABLE_START_ADDR, adev->gmc.gart_start >> 12); 863 WREG32(mmVM_CONTEXT0_PAGE_TABLE_END_ADDR, adev->gmc.gart_end >> 12); 864 WREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR, table_addr >> 12); 865 WREG32(mmVM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR, 866 (u32)(adev->dummy_page_addr >> 12)); 867 WREG32(mmVM_CONTEXT0_CNTL2, 0); 868 tmp = RREG32(mmVM_CONTEXT0_CNTL); 869 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, ENABLE_CONTEXT, 1); 870 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, PAGE_TABLE_DEPTH, 0); 871 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1); 872 WREG32(mmVM_CONTEXT0_CNTL, tmp); 873 874 WREG32(mmVM_L2_CONTEXT1_IDENTITY_APERTURE_LOW_ADDR, 0); 875 WREG32(mmVM_L2_CONTEXT1_IDENTITY_APERTURE_HIGH_ADDR, 0); 876 WREG32(mmVM_L2_CONTEXT_IDENTITY_PHYSICAL_OFFSET, 0); 877 878 /* empty context1-15 */ 879 /* FIXME start with 4G, once using 2 level pt switch to full 880 * vm size space 881 */ 882 /* set vm size, must be a multiple of 4 */ 883 WREG32(mmVM_CONTEXT1_PAGE_TABLE_START_ADDR, 0); 884 WREG32(mmVM_CONTEXT1_PAGE_TABLE_END_ADDR, adev->vm_manager.max_pfn - 1); 885 for (i = 1; i < 16; i++) { 886 if (i < 8) 887 WREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + i, 888 table_addr >> 12); 889 else 890 WREG32(mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + i - 8, 891 table_addr >> 12); 892 } 893 894 /* enable context1-15 */ 895 WREG32(mmVM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR, 896 (u32)(adev->dummy_page_addr >> 12)); 897 WREG32(mmVM_CONTEXT1_CNTL2, 4); 898 tmp = RREG32(mmVM_CONTEXT1_CNTL); 899 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, ENABLE_CONTEXT, 1); 900 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_DEPTH, 1); 901 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1); 902 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1); 903 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PDE0_PROTECTION_FAULT_ENABLE_DEFAULT, 1); 904 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, VALID_PROTECTION_FAULT_ENABLE_DEFAULT, 1); 905 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, READ_PROTECTION_FAULT_ENABLE_DEFAULT, 1); 906 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, WRITE_PROTECTION_FAULT_ENABLE_DEFAULT, 1); 907 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, EXECUTE_PROTECTION_FAULT_ENABLE_DEFAULT, 1); 908 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_BLOCK_SIZE, 909 adev->vm_manager.block_size - 9); 910 WREG32(mmVM_CONTEXT1_CNTL, tmp); 911 if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS) 912 gmc_v8_0_set_fault_enable_default(adev, false); 913 else 914 gmc_v8_0_set_fault_enable_default(adev, true); 915 916 gmc_v8_0_flush_gpu_tlb(adev, 0, 0, 0); 917 DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n", 918 (unsigned)(adev->gmc.gart_size >> 20), 919 (unsigned long long)table_addr); 920 adev->gart.ready = true; 921 return 0; 922 } 923 924 static int gmc_v8_0_gart_init(struct amdgpu_device *adev) 925 { 926 int r; 927 928 if (adev->gart.bo) { 929 WARN(1, "R600 PCIE GART already initialized\n"); 930 return 0; 931 } 932 /* Initialize common gart structure */ 933 r = amdgpu_gart_init(adev); 934 if (r) 935 return r; 936 adev->gart.table_size = adev->gart.num_gpu_pages * 8; 937 adev->gart.gart_pte_flags = AMDGPU_PTE_EXECUTABLE; 938 return amdgpu_gart_table_vram_alloc(adev); 939 } 940 941 /** 942 * gmc_v8_0_gart_disable - gart disable 943 * 944 * @adev: amdgpu_device pointer 945 * 946 * This disables all VM page table (VI). 947 */ 948 static void gmc_v8_0_gart_disable(struct amdgpu_device *adev) 949 { 950 u32 tmp; 951 952 /* Disable all tables */ 953 WREG32(mmVM_CONTEXT0_CNTL, 0); 954 WREG32(mmVM_CONTEXT1_CNTL, 0); 955 /* Setup TLB control */ 956 tmp = RREG32(mmMC_VM_MX_L1_TLB_CNTL); 957 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_TLB, 0); 958 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_FRAGMENT_PROCESSING, 0); 959 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_ADVANCED_DRIVER_MODEL, 0); 960 WREG32(mmMC_VM_MX_L1_TLB_CNTL, tmp); 961 /* Setup L2 cache */ 962 tmp = RREG32(mmVM_L2_CNTL); 963 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 0); 964 WREG32(mmVM_L2_CNTL, tmp); 965 WREG32(mmVM_L2_CNTL2, 0); 966 amdgpu_gart_table_vram_unpin(adev); 967 } 968 969 /** 970 * gmc_v8_0_vm_decode_fault - print human readable fault info 971 * 972 * @adev: amdgpu_device pointer 973 * @status: VM_CONTEXT1_PROTECTION_FAULT_STATUS register value 974 * @addr: VM_CONTEXT1_PROTECTION_FAULT_ADDR register value 975 * 976 * Print human readable fault information (VI). 977 */ 978 static void gmc_v8_0_vm_decode_fault(struct amdgpu_device *adev, u32 status, 979 u32 addr, u32 mc_client, unsigned pasid) 980 { 981 u32 vmid = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, VMID); 982 u32 protections = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, 983 PROTECTIONS); 984 char block[5] = { mc_client >> 24, (mc_client >> 16) & 0xff, 985 (mc_client >> 8) & 0xff, mc_client & 0xff, 0 }; 986 u32 mc_id; 987 988 mc_id = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, 989 MEMORY_CLIENT_ID); 990 991 dev_err(adev->dev, "VM fault (0x%02x, vmid %d, pasid %d) at page %u, %s from '%s' (0x%08x) (%d)\n", 992 protections, vmid, pasid, addr, 993 REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, 994 MEMORY_CLIENT_RW) ? 995 "write" : "read", block, mc_client, mc_id); 996 } 997 998 static int gmc_v8_0_convert_vram_type(int mc_seq_vram_type) 999 { 1000 switch (mc_seq_vram_type) { 1001 case MC_SEQ_MISC0__MT__GDDR1: 1002 return AMDGPU_VRAM_TYPE_GDDR1; 1003 case MC_SEQ_MISC0__MT__DDR2: 1004 return AMDGPU_VRAM_TYPE_DDR2; 1005 case MC_SEQ_MISC0__MT__GDDR3: 1006 return AMDGPU_VRAM_TYPE_GDDR3; 1007 case MC_SEQ_MISC0__MT__GDDR4: 1008 return AMDGPU_VRAM_TYPE_GDDR4; 1009 case MC_SEQ_MISC0__MT__GDDR5: 1010 return AMDGPU_VRAM_TYPE_GDDR5; 1011 case MC_SEQ_MISC0__MT__HBM: 1012 return AMDGPU_VRAM_TYPE_HBM; 1013 case MC_SEQ_MISC0__MT__DDR3: 1014 return AMDGPU_VRAM_TYPE_DDR3; 1015 default: 1016 return AMDGPU_VRAM_TYPE_UNKNOWN; 1017 } 1018 } 1019 1020 static int gmc_v8_0_early_init(void *handle) 1021 { 1022 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1023 1024 gmc_v8_0_set_gmc_funcs(adev); 1025 gmc_v8_0_set_irq_funcs(adev); 1026 1027 adev->gmc.shared_aperture_start = 0x2000000000000000ULL; 1028 adev->gmc.shared_aperture_end = 1029 adev->gmc.shared_aperture_start + (4ULL << 30) - 1; 1030 adev->gmc.private_aperture_start = 1031 adev->gmc.shared_aperture_end + 1; 1032 adev->gmc.private_aperture_end = 1033 adev->gmc.private_aperture_start + (4ULL << 30) - 1; 1034 1035 return 0; 1036 } 1037 1038 static int gmc_v8_0_late_init(void *handle) 1039 { 1040 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1041 1042 amdgpu_bo_late_init(adev); 1043 1044 if (amdgpu_vm_fault_stop != AMDGPU_VM_FAULT_STOP_ALWAYS) 1045 return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0); 1046 else 1047 return 0; 1048 } 1049 1050 static unsigned gmc_v8_0_get_vbios_fb_size(struct amdgpu_device *adev) 1051 { 1052 u32 d1vga_control = RREG32(mmD1VGA_CONTROL); 1053 unsigned size; 1054 1055 if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) { 1056 size = 9 * 1024 * 1024; /* reserve 8MB for vga emulator and 1 MB for FB */ 1057 } else { 1058 u32 viewport = RREG32(mmVIEWPORT_SIZE); 1059 size = (REG_GET_FIELD(viewport, VIEWPORT_SIZE, VIEWPORT_HEIGHT) * 1060 REG_GET_FIELD(viewport, VIEWPORT_SIZE, VIEWPORT_WIDTH) * 1061 4); 1062 } 1063 /* return 0 if the pre-OS buffer uses up most of vram */ 1064 if ((adev->gmc.real_vram_size - size) < (8 * 1024 * 1024)) 1065 return 0; 1066 return size; 1067 } 1068 1069 #define mmMC_SEQ_MISC0_FIJI 0xA71 1070 1071 static int gmc_v8_0_sw_init(void *handle) 1072 { 1073 int r; 1074 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1075 1076 adev->num_vmhubs = 1; 1077 1078 if (adev->flags & AMD_IS_APU) { 1079 adev->gmc.vram_type = AMDGPU_VRAM_TYPE_UNKNOWN; 1080 } else { 1081 u32 tmp; 1082 1083 if ((adev->asic_type == CHIP_FIJI) || 1084 (adev->asic_type == CHIP_VEGAM)) 1085 tmp = RREG32(mmMC_SEQ_MISC0_FIJI); 1086 else 1087 tmp = RREG32(mmMC_SEQ_MISC0); 1088 tmp &= MC_SEQ_MISC0__MT__MASK; 1089 adev->gmc.vram_type = gmc_v8_0_convert_vram_type(tmp); 1090 } 1091 1092 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_GFX_PAGE_INV_FAULT, &adev->gmc.vm_fault); 1093 if (r) 1094 return r; 1095 1096 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_GFX_MEM_PROT_FAULT, &adev->gmc.vm_fault); 1097 if (r) 1098 return r; 1099 1100 /* Adjust VM size here. 1101 * Currently set to 4GB ((1 << 20) 4k pages). 1102 * Max GPUVM size for cayman and SI is 40 bits. 1103 */ 1104 amdgpu_vm_adjust_size(adev, 64, 9, 1, 40); 1105 1106 /* Set the internal MC address mask 1107 * This is the max address of the GPU's 1108 * internal address space. 1109 */ 1110 adev->gmc.mc_mask = 0xffffffffffULL; /* 40 bit MC */ 1111 1112 r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(40)); 1113 if (r) { 1114 pr_warn("amdgpu: No suitable DMA available\n"); 1115 return r; 1116 } 1117 adev->need_swiotlb = drm_need_swiotlb(40); 1118 1119 r = gmc_v8_0_init_microcode(adev); 1120 if (r) { 1121 DRM_ERROR("Failed to load mc firmware!\n"); 1122 return r; 1123 } 1124 1125 r = gmc_v8_0_mc_init(adev); 1126 if (r) 1127 return r; 1128 1129 adev->gmc.stolen_size = gmc_v8_0_get_vbios_fb_size(adev); 1130 1131 /* Memory manager */ 1132 r = amdgpu_bo_init(adev); 1133 if (r) 1134 return r; 1135 1136 r = gmc_v8_0_gart_init(adev); 1137 if (r) 1138 return r; 1139 1140 /* 1141 * number of VMs 1142 * VMID 0 is reserved for System 1143 * amdgpu graphics/compute will use VMIDs 1-7 1144 * amdkfd will use VMIDs 8-15 1145 */ 1146 adev->vm_manager.id_mgr[0].num_ids = AMDGPU_NUM_OF_VMIDS; 1147 amdgpu_vm_manager_init(adev); 1148 1149 /* base offset of vram pages */ 1150 if (adev->flags & AMD_IS_APU) { 1151 u64 tmp = RREG32(mmMC_VM_FB_OFFSET); 1152 1153 tmp <<= 22; 1154 adev->vm_manager.vram_base_offset = tmp; 1155 } else { 1156 adev->vm_manager.vram_base_offset = 0; 1157 } 1158 1159 adev->gmc.vm_fault_info = kmalloc(sizeof(struct kfd_vm_fault_info), 1160 GFP_KERNEL); 1161 if (!adev->gmc.vm_fault_info) 1162 return -ENOMEM; 1163 atomic_set(&adev->gmc.vm_fault_info_updated, 0); 1164 1165 return 0; 1166 } 1167 1168 static int gmc_v8_0_sw_fini(void *handle) 1169 { 1170 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1171 1172 amdgpu_gem_force_release(adev); 1173 amdgpu_vm_manager_fini(adev); 1174 kfree(adev->gmc.vm_fault_info); 1175 amdgpu_gart_table_vram_free(adev); 1176 amdgpu_bo_fini(adev); 1177 amdgpu_gart_fini(adev); 1178 release_firmware(adev->gmc.fw); 1179 adev->gmc.fw = NULL; 1180 1181 return 0; 1182 } 1183 1184 static int gmc_v8_0_hw_init(void *handle) 1185 { 1186 int r; 1187 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1188 1189 gmc_v8_0_init_golden_registers(adev); 1190 1191 gmc_v8_0_mc_program(adev); 1192 1193 if (adev->asic_type == CHIP_TONGA) { 1194 r = gmc_v8_0_tonga_mc_load_microcode(adev); 1195 if (r) { 1196 DRM_ERROR("Failed to load MC firmware!\n"); 1197 return r; 1198 } 1199 } else if (adev->asic_type == CHIP_POLARIS11 || 1200 adev->asic_type == CHIP_POLARIS10 || 1201 adev->asic_type == CHIP_POLARIS12) { 1202 r = gmc_v8_0_polaris_mc_load_microcode(adev); 1203 if (r) { 1204 DRM_ERROR("Failed to load MC firmware!\n"); 1205 return r; 1206 } 1207 } 1208 1209 r = gmc_v8_0_gart_enable(adev); 1210 if (r) 1211 return r; 1212 1213 return r; 1214 } 1215 1216 static int gmc_v8_0_hw_fini(void *handle) 1217 { 1218 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1219 1220 amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0); 1221 gmc_v8_0_gart_disable(adev); 1222 1223 return 0; 1224 } 1225 1226 static int gmc_v8_0_suspend(void *handle) 1227 { 1228 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1229 1230 gmc_v8_0_hw_fini(adev); 1231 1232 return 0; 1233 } 1234 1235 static int gmc_v8_0_resume(void *handle) 1236 { 1237 int r; 1238 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1239 1240 r = gmc_v8_0_hw_init(adev); 1241 if (r) 1242 return r; 1243 1244 amdgpu_vmid_reset_all(adev); 1245 1246 return 0; 1247 } 1248 1249 static bool gmc_v8_0_is_idle(void *handle) 1250 { 1251 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1252 u32 tmp = RREG32(mmSRBM_STATUS); 1253 1254 if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK | 1255 SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK | SRBM_STATUS__VMC_BUSY_MASK)) 1256 return false; 1257 1258 return true; 1259 } 1260 1261 static int gmc_v8_0_wait_for_idle(void *handle) 1262 { 1263 unsigned i; 1264 u32 tmp; 1265 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1266 1267 for (i = 0; i < adev->usec_timeout; i++) { 1268 /* read MC_STATUS */ 1269 tmp = RREG32(mmSRBM_STATUS) & (SRBM_STATUS__MCB_BUSY_MASK | 1270 SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK | 1271 SRBM_STATUS__MCC_BUSY_MASK | 1272 SRBM_STATUS__MCD_BUSY_MASK | 1273 SRBM_STATUS__VMC_BUSY_MASK | 1274 SRBM_STATUS__VMC1_BUSY_MASK); 1275 if (!tmp) 1276 return 0; 1277 udelay(1); 1278 } 1279 return -ETIMEDOUT; 1280 1281 } 1282 1283 static bool gmc_v8_0_check_soft_reset(void *handle) 1284 { 1285 u32 srbm_soft_reset = 0; 1286 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1287 u32 tmp = RREG32(mmSRBM_STATUS); 1288 1289 if (tmp & SRBM_STATUS__VMC_BUSY_MASK) 1290 srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset, 1291 SRBM_SOFT_RESET, SOFT_RESET_VMC, 1); 1292 1293 if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK | 1294 SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK)) { 1295 if (!(adev->flags & AMD_IS_APU)) 1296 srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset, 1297 SRBM_SOFT_RESET, SOFT_RESET_MC, 1); 1298 } 1299 if (srbm_soft_reset) { 1300 adev->gmc.srbm_soft_reset = srbm_soft_reset; 1301 return true; 1302 } else { 1303 adev->gmc.srbm_soft_reset = 0; 1304 return false; 1305 } 1306 } 1307 1308 static int gmc_v8_0_pre_soft_reset(void *handle) 1309 { 1310 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1311 1312 if (!adev->gmc.srbm_soft_reset) 1313 return 0; 1314 1315 gmc_v8_0_mc_stop(adev); 1316 if (gmc_v8_0_wait_for_idle(adev)) { 1317 dev_warn(adev->dev, "Wait for GMC idle timed out !\n"); 1318 } 1319 1320 return 0; 1321 } 1322 1323 static int gmc_v8_0_soft_reset(void *handle) 1324 { 1325 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1326 u32 srbm_soft_reset; 1327 1328 if (!adev->gmc.srbm_soft_reset) 1329 return 0; 1330 srbm_soft_reset = adev->gmc.srbm_soft_reset; 1331 1332 if (srbm_soft_reset) { 1333 u32 tmp; 1334 1335 tmp = RREG32(mmSRBM_SOFT_RESET); 1336 tmp |= srbm_soft_reset; 1337 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp); 1338 WREG32(mmSRBM_SOFT_RESET, tmp); 1339 tmp = RREG32(mmSRBM_SOFT_RESET); 1340 1341 udelay(50); 1342 1343 tmp &= ~srbm_soft_reset; 1344 WREG32(mmSRBM_SOFT_RESET, tmp); 1345 tmp = RREG32(mmSRBM_SOFT_RESET); 1346 1347 /* Wait a little for things to settle down */ 1348 udelay(50); 1349 } 1350 1351 return 0; 1352 } 1353 1354 static int gmc_v8_0_post_soft_reset(void *handle) 1355 { 1356 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1357 1358 if (!adev->gmc.srbm_soft_reset) 1359 return 0; 1360 1361 gmc_v8_0_mc_resume(adev); 1362 return 0; 1363 } 1364 1365 static int gmc_v8_0_vm_fault_interrupt_state(struct amdgpu_device *adev, 1366 struct amdgpu_irq_src *src, 1367 unsigned type, 1368 enum amdgpu_interrupt_state state) 1369 { 1370 u32 tmp; 1371 u32 bits = (VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1372 VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1373 VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1374 VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1375 VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1376 VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1377 VM_CONTEXT1_CNTL__EXECUTE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK); 1378 1379 switch (state) { 1380 case AMDGPU_IRQ_STATE_DISABLE: 1381 /* system context */ 1382 tmp = RREG32(mmVM_CONTEXT0_CNTL); 1383 tmp &= ~bits; 1384 WREG32(mmVM_CONTEXT0_CNTL, tmp); 1385 /* VMs */ 1386 tmp = RREG32(mmVM_CONTEXT1_CNTL); 1387 tmp &= ~bits; 1388 WREG32(mmVM_CONTEXT1_CNTL, tmp); 1389 break; 1390 case AMDGPU_IRQ_STATE_ENABLE: 1391 /* system context */ 1392 tmp = RREG32(mmVM_CONTEXT0_CNTL); 1393 tmp |= bits; 1394 WREG32(mmVM_CONTEXT0_CNTL, tmp); 1395 /* VMs */ 1396 tmp = RREG32(mmVM_CONTEXT1_CNTL); 1397 tmp |= bits; 1398 WREG32(mmVM_CONTEXT1_CNTL, tmp); 1399 break; 1400 default: 1401 break; 1402 } 1403 1404 return 0; 1405 } 1406 1407 static int gmc_v8_0_process_interrupt(struct amdgpu_device *adev, 1408 struct amdgpu_irq_src *source, 1409 struct amdgpu_iv_entry *entry) 1410 { 1411 u32 addr, status, mc_client, vmid; 1412 1413 if (amdgpu_sriov_vf(adev)) { 1414 dev_err(adev->dev, "GPU fault detected: %d 0x%08x\n", 1415 entry->src_id, entry->src_data[0]); 1416 dev_err(adev->dev, " Can't decode VM fault info here on SRIOV VF\n"); 1417 return 0; 1418 } 1419 1420 addr = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_ADDR); 1421 status = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_STATUS); 1422 mc_client = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_MCCLIENT); 1423 /* reset addr and status */ 1424 WREG32_P(mmVM_CONTEXT1_CNTL2, 1, ~1); 1425 1426 if (!addr && !status) 1427 return 0; 1428 1429 if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_FIRST) 1430 gmc_v8_0_set_fault_enable_default(adev, false); 1431 1432 if (printk_ratelimit()) { 1433 struct amdgpu_task_info task_info; 1434 1435 memset(&task_info, 0, sizeof(struct amdgpu_task_info)); 1436 amdgpu_vm_get_task_info(adev, entry->pasid, &task_info); 1437 1438 dev_err(adev->dev, "GPU fault detected: %d 0x%08x for process %s pid %d thread %s pid %d\n", 1439 entry->src_id, entry->src_data[0], task_info.process_name, 1440 task_info.tgid, task_info.task_name, task_info.pid); 1441 dev_err(adev->dev, " VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x%08X\n", 1442 addr); 1443 dev_err(adev->dev, " VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n", 1444 status); 1445 gmc_v8_0_vm_decode_fault(adev, status, addr, mc_client, 1446 entry->pasid); 1447 } 1448 1449 vmid = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, 1450 VMID); 1451 if (amdgpu_amdkfd_is_kfd_vmid(adev, vmid) 1452 && !atomic_read(&adev->gmc.vm_fault_info_updated)) { 1453 struct kfd_vm_fault_info *info = adev->gmc.vm_fault_info; 1454 u32 protections = REG_GET_FIELD(status, 1455 VM_CONTEXT1_PROTECTION_FAULT_STATUS, 1456 PROTECTIONS); 1457 1458 info->vmid = vmid; 1459 info->mc_id = REG_GET_FIELD(status, 1460 VM_CONTEXT1_PROTECTION_FAULT_STATUS, 1461 MEMORY_CLIENT_ID); 1462 info->status = status; 1463 info->page_addr = addr; 1464 info->prot_valid = protections & 0x7 ? true : false; 1465 info->prot_read = protections & 0x8 ? true : false; 1466 info->prot_write = protections & 0x10 ? true : false; 1467 info->prot_exec = protections & 0x20 ? true : false; 1468 mb(); 1469 atomic_set(&adev->gmc.vm_fault_info_updated, 1); 1470 } 1471 1472 return 0; 1473 } 1474 1475 static void fiji_update_mc_medium_grain_clock_gating(struct amdgpu_device *adev, 1476 bool enable) 1477 { 1478 uint32_t data; 1479 1480 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_MC_MGCG)) { 1481 data = RREG32(mmMC_HUB_MISC_HUB_CG); 1482 data |= MC_HUB_MISC_HUB_CG__ENABLE_MASK; 1483 WREG32(mmMC_HUB_MISC_HUB_CG, data); 1484 1485 data = RREG32(mmMC_HUB_MISC_SIP_CG); 1486 data |= MC_HUB_MISC_SIP_CG__ENABLE_MASK; 1487 WREG32(mmMC_HUB_MISC_SIP_CG, data); 1488 1489 data = RREG32(mmMC_HUB_MISC_VM_CG); 1490 data |= MC_HUB_MISC_VM_CG__ENABLE_MASK; 1491 WREG32(mmMC_HUB_MISC_VM_CG, data); 1492 1493 data = RREG32(mmMC_XPB_CLK_GAT); 1494 data |= MC_XPB_CLK_GAT__ENABLE_MASK; 1495 WREG32(mmMC_XPB_CLK_GAT, data); 1496 1497 data = RREG32(mmATC_MISC_CG); 1498 data |= ATC_MISC_CG__ENABLE_MASK; 1499 WREG32(mmATC_MISC_CG, data); 1500 1501 data = RREG32(mmMC_CITF_MISC_WR_CG); 1502 data |= MC_CITF_MISC_WR_CG__ENABLE_MASK; 1503 WREG32(mmMC_CITF_MISC_WR_CG, data); 1504 1505 data = RREG32(mmMC_CITF_MISC_RD_CG); 1506 data |= MC_CITF_MISC_RD_CG__ENABLE_MASK; 1507 WREG32(mmMC_CITF_MISC_RD_CG, data); 1508 1509 data = RREG32(mmMC_CITF_MISC_VM_CG); 1510 data |= MC_CITF_MISC_VM_CG__ENABLE_MASK; 1511 WREG32(mmMC_CITF_MISC_VM_CG, data); 1512 1513 data = RREG32(mmVM_L2_CG); 1514 data |= VM_L2_CG__ENABLE_MASK; 1515 WREG32(mmVM_L2_CG, data); 1516 } else { 1517 data = RREG32(mmMC_HUB_MISC_HUB_CG); 1518 data &= ~MC_HUB_MISC_HUB_CG__ENABLE_MASK; 1519 WREG32(mmMC_HUB_MISC_HUB_CG, data); 1520 1521 data = RREG32(mmMC_HUB_MISC_SIP_CG); 1522 data &= ~MC_HUB_MISC_SIP_CG__ENABLE_MASK; 1523 WREG32(mmMC_HUB_MISC_SIP_CG, data); 1524 1525 data = RREG32(mmMC_HUB_MISC_VM_CG); 1526 data &= ~MC_HUB_MISC_VM_CG__ENABLE_MASK; 1527 WREG32(mmMC_HUB_MISC_VM_CG, data); 1528 1529 data = RREG32(mmMC_XPB_CLK_GAT); 1530 data &= ~MC_XPB_CLK_GAT__ENABLE_MASK; 1531 WREG32(mmMC_XPB_CLK_GAT, data); 1532 1533 data = RREG32(mmATC_MISC_CG); 1534 data &= ~ATC_MISC_CG__ENABLE_MASK; 1535 WREG32(mmATC_MISC_CG, data); 1536 1537 data = RREG32(mmMC_CITF_MISC_WR_CG); 1538 data &= ~MC_CITF_MISC_WR_CG__ENABLE_MASK; 1539 WREG32(mmMC_CITF_MISC_WR_CG, data); 1540 1541 data = RREG32(mmMC_CITF_MISC_RD_CG); 1542 data &= ~MC_CITF_MISC_RD_CG__ENABLE_MASK; 1543 WREG32(mmMC_CITF_MISC_RD_CG, data); 1544 1545 data = RREG32(mmMC_CITF_MISC_VM_CG); 1546 data &= ~MC_CITF_MISC_VM_CG__ENABLE_MASK; 1547 WREG32(mmMC_CITF_MISC_VM_CG, data); 1548 1549 data = RREG32(mmVM_L2_CG); 1550 data &= ~VM_L2_CG__ENABLE_MASK; 1551 WREG32(mmVM_L2_CG, data); 1552 } 1553 } 1554 1555 static void fiji_update_mc_light_sleep(struct amdgpu_device *adev, 1556 bool enable) 1557 { 1558 uint32_t data; 1559 1560 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_MC_LS)) { 1561 data = RREG32(mmMC_HUB_MISC_HUB_CG); 1562 data |= MC_HUB_MISC_HUB_CG__MEM_LS_ENABLE_MASK; 1563 WREG32(mmMC_HUB_MISC_HUB_CG, data); 1564 1565 data = RREG32(mmMC_HUB_MISC_SIP_CG); 1566 data |= MC_HUB_MISC_SIP_CG__MEM_LS_ENABLE_MASK; 1567 WREG32(mmMC_HUB_MISC_SIP_CG, data); 1568 1569 data = RREG32(mmMC_HUB_MISC_VM_CG); 1570 data |= MC_HUB_MISC_VM_CG__MEM_LS_ENABLE_MASK; 1571 WREG32(mmMC_HUB_MISC_VM_CG, data); 1572 1573 data = RREG32(mmMC_XPB_CLK_GAT); 1574 data |= MC_XPB_CLK_GAT__MEM_LS_ENABLE_MASK; 1575 WREG32(mmMC_XPB_CLK_GAT, data); 1576 1577 data = RREG32(mmATC_MISC_CG); 1578 data |= ATC_MISC_CG__MEM_LS_ENABLE_MASK; 1579 WREG32(mmATC_MISC_CG, data); 1580 1581 data = RREG32(mmMC_CITF_MISC_WR_CG); 1582 data |= MC_CITF_MISC_WR_CG__MEM_LS_ENABLE_MASK; 1583 WREG32(mmMC_CITF_MISC_WR_CG, data); 1584 1585 data = RREG32(mmMC_CITF_MISC_RD_CG); 1586 data |= MC_CITF_MISC_RD_CG__MEM_LS_ENABLE_MASK; 1587 WREG32(mmMC_CITF_MISC_RD_CG, data); 1588 1589 data = RREG32(mmMC_CITF_MISC_VM_CG); 1590 data |= MC_CITF_MISC_VM_CG__MEM_LS_ENABLE_MASK; 1591 WREG32(mmMC_CITF_MISC_VM_CG, data); 1592 1593 data = RREG32(mmVM_L2_CG); 1594 data |= VM_L2_CG__MEM_LS_ENABLE_MASK; 1595 WREG32(mmVM_L2_CG, data); 1596 } else { 1597 data = RREG32(mmMC_HUB_MISC_HUB_CG); 1598 data &= ~MC_HUB_MISC_HUB_CG__MEM_LS_ENABLE_MASK; 1599 WREG32(mmMC_HUB_MISC_HUB_CG, data); 1600 1601 data = RREG32(mmMC_HUB_MISC_SIP_CG); 1602 data &= ~MC_HUB_MISC_SIP_CG__MEM_LS_ENABLE_MASK; 1603 WREG32(mmMC_HUB_MISC_SIP_CG, data); 1604 1605 data = RREG32(mmMC_HUB_MISC_VM_CG); 1606 data &= ~MC_HUB_MISC_VM_CG__MEM_LS_ENABLE_MASK; 1607 WREG32(mmMC_HUB_MISC_VM_CG, data); 1608 1609 data = RREG32(mmMC_XPB_CLK_GAT); 1610 data &= ~MC_XPB_CLK_GAT__MEM_LS_ENABLE_MASK; 1611 WREG32(mmMC_XPB_CLK_GAT, data); 1612 1613 data = RREG32(mmATC_MISC_CG); 1614 data &= ~ATC_MISC_CG__MEM_LS_ENABLE_MASK; 1615 WREG32(mmATC_MISC_CG, data); 1616 1617 data = RREG32(mmMC_CITF_MISC_WR_CG); 1618 data &= ~MC_CITF_MISC_WR_CG__MEM_LS_ENABLE_MASK; 1619 WREG32(mmMC_CITF_MISC_WR_CG, data); 1620 1621 data = RREG32(mmMC_CITF_MISC_RD_CG); 1622 data &= ~MC_CITF_MISC_RD_CG__MEM_LS_ENABLE_MASK; 1623 WREG32(mmMC_CITF_MISC_RD_CG, data); 1624 1625 data = RREG32(mmMC_CITF_MISC_VM_CG); 1626 data &= ~MC_CITF_MISC_VM_CG__MEM_LS_ENABLE_MASK; 1627 WREG32(mmMC_CITF_MISC_VM_CG, data); 1628 1629 data = RREG32(mmVM_L2_CG); 1630 data &= ~VM_L2_CG__MEM_LS_ENABLE_MASK; 1631 WREG32(mmVM_L2_CG, data); 1632 } 1633 } 1634 1635 static int gmc_v8_0_set_clockgating_state(void *handle, 1636 enum amd_clockgating_state state) 1637 { 1638 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1639 1640 if (amdgpu_sriov_vf(adev)) 1641 return 0; 1642 1643 switch (adev->asic_type) { 1644 case CHIP_FIJI: 1645 fiji_update_mc_medium_grain_clock_gating(adev, 1646 state == AMD_CG_STATE_GATE); 1647 fiji_update_mc_light_sleep(adev, 1648 state == AMD_CG_STATE_GATE); 1649 break; 1650 default: 1651 break; 1652 } 1653 return 0; 1654 } 1655 1656 static int gmc_v8_0_set_powergating_state(void *handle, 1657 enum amd_powergating_state state) 1658 { 1659 return 0; 1660 } 1661 1662 static void gmc_v8_0_get_clockgating_state(void *handle, u32 *flags) 1663 { 1664 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1665 int data; 1666 1667 if (amdgpu_sriov_vf(adev)) 1668 *flags = 0; 1669 1670 /* AMD_CG_SUPPORT_MC_MGCG */ 1671 data = RREG32(mmMC_HUB_MISC_HUB_CG); 1672 if (data & MC_HUB_MISC_HUB_CG__ENABLE_MASK) 1673 *flags |= AMD_CG_SUPPORT_MC_MGCG; 1674 1675 /* AMD_CG_SUPPORT_MC_LS */ 1676 if (data & MC_HUB_MISC_HUB_CG__MEM_LS_ENABLE_MASK) 1677 *flags |= AMD_CG_SUPPORT_MC_LS; 1678 } 1679 1680 static const struct amd_ip_funcs gmc_v8_0_ip_funcs = { 1681 .name = "gmc_v8_0", 1682 .early_init = gmc_v8_0_early_init, 1683 .late_init = gmc_v8_0_late_init, 1684 .sw_init = gmc_v8_0_sw_init, 1685 .sw_fini = gmc_v8_0_sw_fini, 1686 .hw_init = gmc_v8_0_hw_init, 1687 .hw_fini = gmc_v8_0_hw_fini, 1688 .suspend = gmc_v8_0_suspend, 1689 .resume = gmc_v8_0_resume, 1690 .is_idle = gmc_v8_0_is_idle, 1691 .wait_for_idle = gmc_v8_0_wait_for_idle, 1692 .check_soft_reset = gmc_v8_0_check_soft_reset, 1693 .pre_soft_reset = gmc_v8_0_pre_soft_reset, 1694 .soft_reset = gmc_v8_0_soft_reset, 1695 .post_soft_reset = gmc_v8_0_post_soft_reset, 1696 .set_clockgating_state = gmc_v8_0_set_clockgating_state, 1697 .set_powergating_state = gmc_v8_0_set_powergating_state, 1698 .get_clockgating_state = gmc_v8_0_get_clockgating_state, 1699 }; 1700 1701 static const struct amdgpu_gmc_funcs gmc_v8_0_gmc_funcs = { 1702 .flush_gpu_tlb = gmc_v8_0_flush_gpu_tlb, 1703 .emit_flush_gpu_tlb = gmc_v8_0_emit_flush_gpu_tlb, 1704 .emit_pasid_mapping = gmc_v8_0_emit_pasid_mapping, 1705 .set_prt = gmc_v8_0_set_prt, 1706 .get_vm_pde = gmc_v8_0_get_vm_pde, 1707 .get_vm_pte = gmc_v8_0_get_vm_pte 1708 }; 1709 1710 static const struct amdgpu_irq_src_funcs gmc_v8_0_irq_funcs = { 1711 .set = gmc_v8_0_vm_fault_interrupt_state, 1712 .process = gmc_v8_0_process_interrupt, 1713 }; 1714 1715 static void gmc_v8_0_set_gmc_funcs(struct amdgpu_device *adev) 1716 { 1717 adev->gmc.gmc_funcs = &gmc_v8_0_gmc_funcs; 1718 } 1719 1720 static void gmc_v8_0_set_irq_funcs(struct amdgpu_device *adev) 1721 { 1722 adev->gmc.vm_fault.num_types = 1; 1723 adev->gmc.vm_fault.funcs = &gmc_v8_0_irq_funcs; 1724 } 1725 1726 const struct amdgpu_ip_block_version gmc_v8_0_ip_block = 1727 { 1728 .type = AMD_IP_BLOCK_TYPE_GMC, 1729 .major = 8, 1730 .minor = 0, 1731 .rev = 0, 1732 .funcs = &gmc_v8_0_ip_funcs, 1733 }; 1734 1735 const struct amdgpu_ip_block_version gmc_v8_1_ip_block = 1736 { 1737 .type = AMD_IP_BLOCK_TYPE_GMC, 1738 .major = 8, 1739 .minor = 1, 1740 .rev = 0, 1741 .funcs = &gmc_v8_0_ip_funcs, 1742 }; 1743 1744 const struct amdgpu_ip_block_version gmc_v8_5_ip_block = 1745 { 1746 .type = AMD_IP_BLOCK_TYPE_GMC, 1747 .major = 8, 1748 .minor = 5, 1749 .rev = 0, 1750 .funcs = &gmc_v8_0_ip_funcs, 1751 }; 1752