1 /* 2 * Copyright 2018 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 "amdgpu.h" 24 #include "df_v3_6.h" 25 26 #include "df/df_3_6_default.h" 27 #include "df/df_3_6_offset.h" 28 #include "df/df_3_6_sh_mask.h" 29 30 #define DF_3_6_SMN_REG_INST_DIST 0x8 31 #define DF_3_6_INST_CNT 8 32 33 static u32 df_v3_6_channel_number[] = {1, 2, 0, 4, 0, 8, 0, 34 16, 32, 0, 0, 0, 2, 4, 8}; 35 36 /* init df format attrs */ 37 AMDGPU_PMU_ATTR(event, "config:0-7"); 38 AMDGPU_PMU_ATTR(instance, "config:8-15"); 39 AMDGPU_PMU_ATTR(umask, "config:16-23"); 40 41 /* df format attributes */ 42 static struct attribute *df_v3_6_format_attrs[] = { 43 &pmu_attr_event.attr, 44 &pmu_attr_instance.attr, 45 &pmu_attr_umask.attr, 46 NULL 47 }; 48 49 /* df format attribute group */ 50 static struct attribute_group df_v3_6_format_attr_group = { 51 .name = "format", 52 .attrs = df_v3_6_format_attrs, 53 }; 54 55 /* df event attrs */ 56 AMDGPU_PMU_ATTR(cake0_pcsout_txdata, 57 "event=0x7,instance=0x46,umask=0x2"); 58 AMDGPU_PMU_ATTR(cake1_pcsout_txdata, 59 "event=0x7,instance=0x47,umask=0x2"); 60 AMDGPU_PMU_ATTR(cake0_pcsout_txmeta, 61 "event=0x7,instance=0x46,umask=0x4"); 62 AMDGPU_PMU_ATTR(cake1_pcsout_txmeta, 63 "event=0x7,instance=0x47,umask=0x4"); 64 AMDGPU_PMU_ATTR(cake0_ftiinstat_reqalloc, 65 "event=0xb,instance=0x46,umask=0x4"); 66 AMDGPU_PMU_ATTR(cake1_ftiinstat_reqalloc, 67 "event=0xb,instance=0x47,umask=0x4"); 68 AMDGPU_PMU_ATTR(cake0_ftiinstat_rspalloc, 69 "event=0xb,instance=0x46,umask=0x8"); 70 AMDGPU_PMU_ATTR(cake1_ftiinstat_rspalloc, 71 "event=0xb,instance=0x47,umask=0x8"); 72 73 /* df event attributes */ 74 static struct attribute *df_v3_6_event_attrs[] = { 75 &pmu_attr_cake0_pcsout_txdata.attr, 76 &pmu_attr_cake1_pcsout_txdata.attr, 77 &pmu_attr_cake0_pcsout_txmeta.attr, 78 &pmu_attr_cake1_pcsout_txmeta.attr, 79 &pmu_attr_cake0_ftiinstat_reqalloc.attr, 80 &pmu_attr_cake1_ftiinstat_reqalloc.attr, 81 &pmu_attr_cake0_ftiinstat_rspalloc.attr, 82 &pmu_attr_cake1_ftiinstat_rspalloc.attr, 83 NULL 84 }; 85 86 /* df event attribute group */ 87 static struct attribute_group df_v3_6_event_attr_group = { 88 .name = "events", 89 .attrs = df_v3_6_event_attrs 90 }; 91 92 /* df event attr groups */ 93 const struct attribute_group *df_v3_6_attr_groups[] = { 94 &df_v3_6_format_attr_group, 95 &df_v3_6_event_attr_group, 96 NULL 97 }; 98 99 static uint64_t df_v3_6_get_fica(struct amdgpu_device *adev, 100 uint32_t ficaa_val) 101 { 102 unsigned long flags, address, data; 103 uint32_t ficadl_val, ficadh_val; 104 105 address = adev->nbio.funcs->get_pcie_index_offset(adev); 106 data = adev->nbio.funcs->get_pcie_data_offset(adev); 107 108 spin_lock_irqsave(&adev->pcie_idx_lock, flags); 109 WREG32(address, smnDF_PIE_AON_FabricIndirectConfigAccessAddress3); 110 WREG32(data, ficaa_val); 111 112 WREG32(address, smnDF_PIE_AON_FabricIndirectConfigAccessDataLo3); 113 ficadl_val = RREG32(data); 114 115 WREG32(address, smnDF_PIE_AON_FabricIndirectConfigAccessDataHi3); 116 ficadh_val = RREG32(data); 117 118 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags); 119 120 return (((ficadh_val & 0xFFFFFFFFFFFFFFFF) << 32) | ficadl_val); 121 } 122 123 static void df_v3_6_set_fica(struct amdgpu_device *adev, uint32_t ficaa_val, 124 uint32_t ficadl_val, uint32_t ficadh_val) 125 { 126 unsigned long flags, address, data; 127 128 address = adev->nbio.funcs->get_pcie_index_offset(adev); 129 data = adev->nbio.funcs->get_pcie_data_offset(adev); 130 131 spin_lock_irqsave(&adev->pcie_idx_lock, flags); 132 WREG32(address, smnDF_PIE_AON_FabricIndirectConfigAccessAddress3); 133 WREG32(data, ficaa_val); 134 135 WREG32(address, smnDF_PIE_AON_FabricIndirectConfigAccessDataLo3); 136 WREG32(data, ficadl_val); 137 138 WREG32(address, smnDF_PIE_AON_FabricIndirectConfigAccessDataHi3); 139 WREG32(data, ficadh_val); 140 141 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags); 142 } 143 144 /* 145 * df_v3_6_perfmon_rreg - read perfmon lo and hi 146 * 147 * required to be atomic. no mmio method provided so subsequent reads for lo 148 * and hi require to preserve df finite state machine 149 */ 150 static void df_v3_6_perfmon_rreg(struct amdgpu_device *adev, 151 uint32_t lo_addr, uint32_t *lo_val, 152 uint32_t hi_addr, uint32_t *hi_val) 153 { 154 unsigned long flags, address, data; 155 156 address = adev->nbio.funcs->get_pcie_index_offset(adev); 157 data = adev->nbio.funcs->get_pcie_data_offset(adev); 158 159 spin_lock_irqsave(&adev->pcie_idx_lock, flags); 160 WREG32(address, lo_addr); 161 *lo_val = RREG32(data); 162 WREG32(address, hi_addr); 163 *hi_val = RREG32(data); 164 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags); 165 } 166 167 /* 168 * df_v3_6_perfmon_wreg - write to perfmon lo and hi 169 * 170 * required to be atomic. no mmio method provided so subsequent reads after 171 * data writes cannot occur to preserve data fabrics finite state machine. 172 */ 173 static void df_v3_6_perfmon_wreg(struct amdgpu_device *adev, uint32_t lo_addr, 174 uint32_t lo_val, uint32_t hi_addr, uint32_t hi_val) 175 { 176 unsigned long flags, address, data; 177 178 address = adev->nbio.funcs->get_pcie_index_offset(adev); 179 data = adev->nbio.funcs->get_pcie_data_offset(adev); 180 181 spin_lock_irqsave(&adev->pcie_idx_lock, flags); 182 WREG32(address, lo_addr); 183 WREG32(data, lo_val); 184 WREG32(address, hi_addr); 185 WREG32(data, hi_val); 186 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags); 187 } 188 189 /* same as perfmon_wreg but return status on write value check */ 190 static int df_v3_6_perfmon_arm_with_status(struct amdgpu_device *adev, 191 uint32_t lo_addr, uint32_t lo_val, 192 uint32_t hi_addr, uint32_t hi_val) 193 { 194 unsigned long flags, address, data; 195 uint32_t lo_val_rb, hi_val_rb; 196 197 address = adev->nbio.funcs->get_pcie_index_offset(adev); 198 data = adev->nbio.funcs->get_pcie_data_offset(adev); 199 200 spin_lock_irqsave(&adev->pcie_idx_lock, flags); 201 WREG32(address, lo_addr); 202 WREG32(data, lo_val); 203 WREG32(address, hi_addr); 204 WREG32(data, hi_val); 205 206 WREG32(address, lo_addr); 207 lo_val_rb = RREG32(data); 208 WREG32(address, hi_addr); 209 hi_val_rb = RREG32(data); 210 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags); 211 212 if (!(lo_val == lo_val_rb && hi_val == hi_val_rb)) 213 return -EBUSY; 214 215 return 0; 216 } 217 218 219 /* 220 * retry arming counters every 100 usecs within 1 millisecond interval. 221 * if retry fails after time out, return error. 222 */ 223 #define ARM_RETRY_USEC_TIMEOUT 1000 224 #define ARM_RETRY_USEC_INTERVAL 100 225 static int df_v3_6_perfmon_arm_with_retry(struct amdgpu_device *adev, 226 uint32_t lo_addr, uint32_t lo_val, 227 uint32_t hi_addr, uint32_t hi_val) 228 { 229 int countdown = ARM_RETRY_USEC_TIMEOUT; 230 231 while (countdown) { 232 233 if (!df_v3_6_perfmon_arm_with_status(adev, lo_addr, lo_val, 234 hi_addr, hi_val)) 235 break; 236 237 countdown -= ARM_RETRY_USEC_INTERVAL; 238 udelay(ARM_RETRY_USEC_INTERVAL); 239 } 240 241 return countdown > 0 ? 0 : -ETIME; 242 } 243 244 /* get the number of df counters available */ 245 static ssize_t df_v3_6_get_df_cntr_avail(struct device *dev, 246 struct device_attribute *attr, 247 char *buf) 248 { 249 struct amdgpu_device *adev; 250 struct drm_device *ddev; 251 int i, count; 252 253 ddev = dev_get_drvdata(dev); 254 adev = ddev->dev_private; 255 count = 0; 256 257 for (i = 0; i < DF_V3_6_MAX_COUNTERS; i++) { 258 if (adev->df_perfmon_config_assign_mask[i] == 0) 259 count++; 260 } 261 262 return snprintf(buf, PAGE_SIZE, "%i\n", count); 263 } 264 265 /* device attr for available perfmon counters */ 266 static DEVICE_ATTR(df_cntr_avail, S_IRUGO, df_v3_6_get_df_cntr_avail, NULL); 267 268 static void df_v3_6_query_hashes(struct amdgpu_device *adev) 269 { 270 u32 tmp; 271 272 adev->df.hash_status.hash_64k = false; 273 adev->df.hash_status.hash_2m = false; 274 adev->df.hash_status.hash_1g = false; 275 276 if (adev->asic_type != CHIP_ARCTURUS) 277 return; 278 279 /* encoding for hash-enabled on Arcturus */ 280 if (adev->df.funcs->get_fb_channel_number(adev) == 0xe) { 281 tmp = RREG32_SOC15(DF, 0, mmDF_CS_UMC_AON0_DfGlobalCtrl); 282 adev->df.hash_status.hash_64k = REG_GET_FIELD(tmp, 283 DF_CS_UMC_AON0_DfGlobalCtrl, 284 GlbHashIntlvCtl64K); 285 adev->df.hash_status.hash_2m = REG_GET_FIELD(tmp, 286 DF_CS_UMC_AON0_DfGlobalCtrl, 287 GlbHashIntlvCtl2M); 288 adev->df.hash_status.hash_1g = REG_GET_FIELD(tmp, 289 DF_CS_UMC_AON0_DfGlobalCtrl, 290 GlbHashIntlvCtl1G); 291 } 292 } 293 294 /* init perfmons */ 295 static void df_v3_6_sw_init(struct amdgpu_device *adev) 296 { 297 int i, ret; 298 299 ret = device_create_file(adev->dev, &dev_attr_df_cntr_avail); 300 if (ret) 301 DRM_ERROR("failed to create file for available df counters\n"); 302 303 for (i = 0; i < AMDGPU_MAX_DF_PERFMONS; i++) 304 adev->df_perfmon_config_assign_mask[i] = 0; 305 306 df_v3_6_query_hashes(adev); 307 } 308 309 static void df_v3_6_sw_fini(struct amdgpu_device *adev) 310 { 311 312 device_remove_file(adev->dev, &dev_attr_df_cntr_avail); 313 314 } 315 316 static void df_v3_6_enable_broadcast_mode(struct amdgpu_device *adev, 317 bool enable) 318 { 319 u32 tmp; 320 321 if (enable) { 322 tmp = RREG32_SOC15(DF, 0, mmFabricConfigAccessControl); 323 tmp &= ~FabricConfigAccessControl__CfgRegInstAccEn_MASK; 324 WREG32_SOC15(DF, 0, mmFabricConfigAccessControl, tmp); 325 } else 326 WREG32_SOC15(DF, 0, mmFabricConfigAccessControl, 327 mmFabricConfigAccessControl_DEFAULT); 328 } 329 330 static u32 df_v3_6_get_fb_channel_number(struct amdgpu_device *adev) 331 { 332 u32 tmp; 333 334 tmp = RREG32_SOC15(DF, 0, mmDF_CS_UMC_AON0_DramBaseAddress0); 335 tmp &= DF_CS_UMC_AON0_DramBaseAddress0__IntLvNumChan_MASK; 336 tmp >>= DF_CS_UMC_AON0_DramBaseAddress0__IntLvNumChan__SHIFT; 337 338 return tmp; 339 } 340 341 static u32 df_v3_6_get_hbm_channel_number(struct amdgpu_device *adev) 342 { 343 int fb_channel_number; 344 345 fb_channel_number = adev->df.funcs->get_fb_channel_number(adev); 346 if (fb_channel_number >= ARRAY_SIZE(df_v3_6_channel_number)) 347 fb_channel_number = 0; 348 349 return df_v3_6_channel_number[fb_channel_number]; 350 } 351 352 static void df_v3_6_update_medium_grain_clock_gating(struct amdgpu_device *adev, 353 bool enable) 354 { 355 u32 tmp; 356 357 if (adev->cg_flags & AMD_CG_SUPPORT_DF_MGCG) { 358 /* Put DF on broadcast mode */ 359 adev->df.funcs->enable_broadcast_mode(adev, true); 360 361 if (enable) { 362 tmp = RREG32_SOC15(DF, 0, 363 mmDF_PIE_AON0_DfGlobalClkGater); 364 tmp &= ~DF_PIE_AON0_DfGlobalClkGater__MGCGMode_MASK; 365 tmp |= DF_V3_6_MGCG_ENABLE_15_CYCLE_DELAY; 366 WREG32_SOC15(DF, 0, 367 mmDF_PIE_AON0_DfGlobalClkGater, tmp); 368 } else { 369 tmp = RREG32_SOC15(DF, 0, 370 mmDF_PIE_AON0_DfGlobalClkGater); 371 tmp &= ~DF_PIE_AON0_DfGlobalClkGater__MGCGMode_MASK; 372 tmp |= DF_V3_6_MGCG_DISABLE; 373 WREG32_SOC15(DF, 0, 374 mmDF_PIE_AON0_DfGlobalClkGater, tmp); 375 } 376 377 /* Exit broadcast mode */ 378 adev->df.funcs->enable_broadcast_mode(adev, false); 379 } 380 } 381 382 static void df_v3_6_get_clockgating_state(struct amdgpu_device *adev, 383 u32 *flags) 384 { 385 u32 tmp; 386 387 /* AMD_CG_SUPPORT_DF_MGCG */ 388 tmp = RREG32_SOC15(DF, 0, mmDF_PIE_AON0_DfGlobalClkGater); 389 if (tmp & DF_V3_6_MGCG_ENABLE_15_CYCLE_DELAY) 390 *flags |= AMD_CG_SUPPORT_DF_MGCG; 391 } 392 393 /* get assigned df perfmon ctr as int */ 394 static int df_v3_6_pmc_config_2_cntr(struct amdgpu_device *adev, 395 uint64_t config) 396 { 397 int i; 398 399 for (i = 0; i < DF_V3_6_MAX_COUNTERS; i++) { 400 if ((config & 0x0FFFFFFUL) == 401 adev->df_perfmon_config_assign_mask[i]) 402 return i; 403 } 404 405 return -EINVAL; 406 } 407 408 /* get address based on counter assignment */ 409 static void df_v3_6_pmc_get_addr(struct amdgpu_device *adev, 410 uint64_t config, 411 int is_ctrl, 412 uint32_t *lo_base_addr, 413 uint32_t *hi_base_addr) 414 { 415 int target_cntr = df_v3_6_pmc_config_2_cntr(adev, config); 416 417 if (target_cntr < 0) 418 return; 419 420 switch (target_cntr) { 421 422 case 0: 423 *lo_base_addr = is_ctrl ? smnPerfMonCtlLo4 : smnPerfMonCtrLo4; 424 *hi_base_addr = is_ctrl ? smnPerfMonCtlHi4 : smnPerfMonCtrHi4; 425 break; 426 case 1: 427 *lo_base_addr = is_ctrl ? smnPerfMonCtlLo5 : smnPerfMonCtrLo5; 428 *hi_base_addr = is_ctrl ? smnPerfMonCtlHi5 : smnPerfMonCtrHi5; 429 break; 430 case 2: 431 *lo_base_addr = is_ctrl ? smnPerfMonCtlLo6 : smnPerfMonCtrLo6; 432 *hi_base_addr = is_ctrl ? smnPerfMonCtlHi6 : smnPerfMonCtrHi6; 433 break; 434 case 3: 435 *lo_base_addr = is_ctrl ? smnPerfMonCtlLo7 : smnPerfMonCtrLo7; 436 *hi_base_addr = is_ctrl ? smnPerfMonCtlHi7 : smnPerfMonCtrHi7; 437 break; 438 439 } 440 441 } 442 443 /* get read counter address */ 444 static void df_v3_6_pmc_get_read_settings(struct amdgpu_device *adev, 445 uint64_t config, 446 uint32_t *lo_base_addr, 447 uint32_t *hi_base_addr) 448 { 449 df_v3_6_pmc_get_addr(adev, config, 0, lo_base_addr, hi_base_addr); 450 } 451 452 /* get control counter settings i.e. address and values to set */ 453 static int df_v3_6_pmc_get_ctrl_settings(struct amdgpu_device *adev, 454 uint64_t config, 455 uint32_t *lo_base_addr, 456 uint32_t *hi_base_addr, 457 uint32_t *lo_val, 458 uint32_t *hi_val) 459 { 460 461 uint32_t eventsel, instance, unitmask; 462 uint32_t instance_10, instance_5432, instance_76; 463 464 df_v3_6_pmc_get_addr(adev, config, 1, lo_base_addr, hi_base_addr); 465 466 if ((*lo_base_addr == 0) || (*hi_base_addr == 0)) { 467 DRM_ERROR("[DF PMC] addressing not retrieved! Lo: %x, Hi: %x", 468 *lo_base_addr, *hi_base_addr); 469 return -ENXIO; 470 } 471 472 eventsel = DF_V3_6_GET_EVENT(config) & 0x3f; 473 unitmask = DF_V3_6_GET_UNITMASK(config) & 0xf; 474 instance = DF_V3_6_GET_INSTANCE(config); 475 476 instance_10 = instance & 0x3; 477 instance_5432 = (instance >> 2) & 0xf; 478 instance_76 = (instance >> 6) & 0x3; 479 480 *lo_val = (unitmask << 8) | (instance_10 << 6) | eventsel | (1 << 22); 481 *hi_val = (instance_76 << 29) | instance_5432; 482 483 DRM_DEBUG_DRIVER("config=%llx addr=%08x:%08x val=%08x:%08x", 484 config, *lo_base_addr, *hi_base_addr, *lo_val, *hi_val); 485 486 return 0; 487 } 488 489 /* add df performance counters for read */ 490 static int df_v3_6_pmc_add_cntr(struct amdgpu_device *adev, 491 uint64_t config) 492 { 493 int i, target_cntr; 494 495 target_cntr = df_v3_6_pmc_config_2_cntr(adev, config); 496 497 if (target_cntr >= 0) 498 return 0; 499 500 for (i = 0; i < DF_V3_6_MAX_COUNTERS; i++) { 501 if (adev->df_perfmon_config_assign_mask[i] == 0U) { 502 adev->df_perfmon_config_assign_mask[i] = 503 config & 0x0FFFFFFUL; 504 return 0; 505 } 506 } 507 508 return -ENOSPC; 509 } 510 511 #define DEFERRED_ARM_MASK (1 << 31) 512 static int df_v3_6_pmc_set_deferred(struct amdgpu_device *adev, 513 uint64_t config, bool is_deferred) 514 { 515 int target_cntr; 516 517 target_cntr = df_v3_6_pmc_config_2_cntr(adev, config); 518 519 if (target_cntr < 0) 520 return -EINVAL; 521 522 if (is_deferred) 523 adev->df_perfmon_config_assign_mask[target_cntr] |= 524 DEFERRED_ARM_MASK; 525 else 526 adev->df_perfmon_config_assign_mask[target_cntr] &= 527 ~DEFERRED_ARM_MASK; 528 529 return 0; 530 } 531 532 static bool df_v3_6_pmc_is_deferred(struct amdgpu_device *adev, 533 uint64_t config) 534 { 535 int target_cntr; 536 537 target_cntr = df_v3_6_pmc_config_2_cntr(adev, config); 538 539 /* 540 * we never get target_cntr < 0 since this funciton is only called in 541 * pmc_count for now but we should check anyways. 542 */ 543 return (target_cntr >= 0 && 544 (adev->df_perfmon_config_assign_mask[target_cntr] 545 & DEFERRED_ARM_MASK)); 546 547 } 548 549 /* release performance counter */ 550 static void df_v3_6_pmc_release_cntr(struct amdgpu_device *adev, 551 uint64_t config) 552 { 553 int target_cntr = df_v3_6_pmc_config_2_cntr(adev, config); 554 555 if (target_cntr >= 0) 556 adev->df_perfmon_config_assign_mask[target_cntr] = 0ULL; 557 } 558 559 560 static void df_v3_6_reset_perfmon_cntr(struct amdgpu_device *adev, 561 uint64_t config) 562 { 563 uint32_t lo_base_addr, hi_base_addr; 564 565 df_v3_6_pmc_get_read_settings(adev, config, &lo_base_addr, 566 &hi_base_addr); 567 568 if ((lo_base_addr == 0) || (hi_base_addr == 0)) 569 return; 570 571 df_v3_6_perfmon_wreg(adev, lo_base_addr, 0, hi_base_addr, 0); 572 } 573 574 static int df_v3_6_pmc_start(struct amdgpu_device *adev, uint64_t config, 575 int is_enable) 576 { 577 uint32_t lo_base_addr, hi_base_addr, lo_val, hi_val; 578 int err = 0, ret = 0; 579 580 switch (adev->asic_type) { 581 case CHIP_VEGA20: 582 if (is_enable) 583 return df_v3_6_pmc_add_cntr(adev, config); 584 585 df_v3_6_reset_perfmon_cntr(adev, config); 586 587 ret = df_v3_6_pmc_get_ctrl_settings(adev, 588 config, 589 &lo_base_addr, 590 &hi_base_addr, 591 &lo_val, 592 &hi_val); 593 594 if (ret) 595 return ret; 596 597 err = df_v3_6_perfmon_arm_with_retry(adev, 598 lo_base_addr, 599 lo_val, 600 hi_base_addr, 601 hi_val); 602 603 if (err) 604 ret = df_v3_6_pmc_set_deferred(adev, config, true); 605 606 break; 607 default: 608 break; 609 } 610 611 return ret; 612 } 613 614 static int df_v3_6_pmc_stop(struct amdgpu_device *adev, uint64_t config, 615 int is_disable) 616 { 617 uint32_t lo_base_addr, hi_base_addr, lo_val, hi_val; 618 int ret = 0; 619 620 switch (adev->asic_type) { 621 case CHIP_VEGA20: 622 ret = df_v3_6_pmc_get_ctrl_settings(adev, 623 config, 624 &lo_base_addr, 625 &hi_base_addr, 626 &lo_val, 627 &hi_val); 628 629 if (ret) 630 return ret; 631 632 df_v3_6_reset_perfmon_cntr(adev, config); 633 634 if (is_disable) 635 df_v3_6_pmc_release_cntr(adev, config); 636 637 break; 638 default: 639 break; 640 } 641 642 return ret; 643 } 644 645 static void df_v3_6_pmc_get_count(struct amdgpu_device *adev, 646 uint64_t config, 647 uint64_t *count) 648 { 649 uint32_t lo_base_addr, hi_base_addr, lo_val = 0, hi_val = 0; 650 *count = 0; 651 652 switch (adev->asic_type) { 653 case CHIP_VEGA20: 654 df_v3_6_pmc_get_read_settings(adev, config, &lo_base_addr, 655 &hi_base_addr); 656 657 if ((lo_base_addr == 0) || (hi_base_addr == 0)) 658 return; 659 660 /* rearm the counter or throw away count value on failure */ 661 if (df_v3_6_pmc_is_deferred(adev, config)) { 662 int rearm_err = df_v3_6_perfmon_arm_with_status(adev, 663 lo_base_addr, lo_val, 664 hi_base_addr, hi_val); 665 666 if (rearm_err) 667 return; 668 669 df_v3_6_pmc_set_deferred(adev, config, false); 670 } 671 672 df_v3_6_perfmon_rreg(adev, lo_base_addr, &lo_val, 673 hi_base_addr, &hi_val); 674 675 *count = ((hi_val | 0ULL) << 32) | (lo_val | 0ULL); 676 677 if (*count >= DF_V3_6_PERFMON_OVERFLOW) 678 *count = 0; 679 680 DRM_DEBUG_DRIVER("config=%llx addr=%08x:%08x val=%08x:%08x", 681 config, lo_base_addr, hi_base_addr, lo_val, hi_val); 682 683 break; 684 default: 685 break; 686 } 687 } 688 689 static uint64_t df_v3_6_get_dram_base_addr(struct amdgpu_device *adev, 690 uint32_t df_inst) 691 { 692 uint32_t base_addr_reg_val = 0; 693 uint64_t base_addr = 0; 694 695 base_addr_reg_val = RREG32_PCIE(smnDF_CS_UMC_AON0_DramBaseAddress0 + 696 df_inst * DF_3_6_SMN_REG_INST_DIST); 697 698 if (REG_GET_FIELD(base_addr_reg_val, 699 DF_CS_UMC_AON0_DramBaseAddress0, 700 AddrRngVal) == 0) { 701 DRM_WARN("address range not valid"); 702 return 0; 703 } 704 705 base_addr = REG_GET_FIELD(base_addr_reg_val, 706 DF_CS_UMC_AON0_DramBaseAddress0, 707 DramBaseAddr); 708 709 return base_addr << 28; 710 } 711 712 static uint32_t df_v3_6_get_df_inst_id(struct amdgpu_device *adev) 713 { 714 uint32_t xgmi_node_id = 0; 715 uint32_t df_inst_id = 0; 716 717 /* Walk through DF dst nodes to find current XGMI node */ 718 for (df_inst_id = 0; df_inst_id < DF_3_6_INST_CNT; df_inst_id++) { 719 720 xgmi_node_id = RREG32_PCIE(smnDF_CS_UMC_AON0_DramLimitAddress0 + 721 df_inst_id * DF_3_6_SMN_REG_INST_DIST); 722 xgmi_node_id = REG_GET_FIELD(xgmi_node_id, 723 DF_CS_UMC_AON0_DramLimitAddress0, 724 DstFabricID); 725 726 /* TODO: establish reason dest fabric id is offset by 7 */ 727 xgmi_node_id = xgmi_node_id >> 7; 728 729 if (adev->gmc.xgmi.physical_node_id == xgmi_node_id) 730 break; 731 } 732 733 if (df_inst_id == DF_3_6_INST_CNT) { 734 DRM_WARN("cant match df dst id with gpu node"); 735 return 0; 736 } 737 738 return df_inst_id; 739 } 740 741 const struct amdgpu_df_funcs df_v3_6_funcs = { 742 .sw_init = df_v3_6_sw_init, 743 .sw_fini = df_v3_6_sw_fini, 744 .enable_broadcast_mode = df_v3_6_enable_broadcast_mode, 745 .get_fb_channel_number = df_v3_6_get_fb_channel_number, 746 .get_hbm_channel_number = df_v3_6_get_hbm_channel_number, 747 .update_medium_grain_clock_gating = 748 df_v3_6_update_medium_grain_clock_gating, 749 .get_clockgating_state = df_v3_6_get_clockgating_state, 750 .pmc_start = df_v3_6_pmc_start, 751 .pmc_stop = df_v3_6_pmc_stop, 752 .pmc_get_count = df_v3_6_pmc_get_count, 753 .get_fica = df_v3_6_get_fica, 754 .set_fica = df_v3_6_set_fica, 755 .get_dram_base_addr = df_v3_6_get_dram_base_addr, 756 .get_df_inst_id = df_v3_6_get_df_inst_id 757 }; 758