1 /* 2 * Copyright 2016 Advanced Micro Devices, Inc. 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice shall be included in 12 * all copies or substantial portions of the Software. 13 * 14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 20 * OTHER DEALINGS IN THE SOFTWARE. 21 * 22 */ 23 24 #include <linux/module.h> 25 26 #ifdef CONFIG_X86 27 #include <asm/hypervisor.h> 28 #endif 29 30 #include <drm/drm_drv.h> 31 #include <xen/xen.h> 32 33 #include "amdgpu.h" 34 #include "amdgpu_ras.h" 35 #include "amdgpu_reset.h" 36 #include "amdgpu_dpm.h" 37 #include "vi.h" 38 #include "soc15.h" 39 #include "nv.h" 40 #include "amdgpu_virt_ras_cmd.h" 41 42 #define POPULATE_UCODE_INFO(vf2pf_info, ucode, ver) \ 43 do { \ 44 vf2pf_info->ucode_info[ucode].id = ucode; \ 45 vf2pf_info->ucode_info[ucode].version = ver; \ 46 } while (0) 47 48 #define mmRCC_CONFIG_MEMSIZE 0xde3 49 50 const char *amdgpu_virt_dynamic_crit_table_name[] = { 51 "IP DISCOVERY", 52 "VBIOS IMG", 53 "RAS TELEMETRY", 54 "DATA EXCHANGE", 55 "BAD PAGE INFO", 56 "INIT HEADER", 57 "LAST", 58 }; 59 60 bool amdgpu_virt_mmio_blocked(struct amdgpu_device *adev) 61 { 62 /* By now all MMIO pages except mailbox are blocked */ 63 /* if blocking is enabled in hypervisor. Choose the */ 64 /* SCRATCH_REG0 to test. */ 65 return RREG32_NO_KIQ(0xc040) == 0xffffffff; 66 } 67 68 void amdgpu_virt_init_setting(struct amdgpu_device *adev) 69 { 70 struct drm_device *ddev = adev_to_drm(adev); 71 72 /* enable virtual display */ 73 if (adev->asic_type != CHIP_ALDEBARAN && 74 adev->asic_type != CHIP_ARCTURUS && 75 ((adev->pdev->class >> 8) != PCI_CLASS_ACCELERATOR_PROCESSING)) { 76 if (adev->mode_info.num_crtc == 0) 77 adev->mode_info.num_crtc = 1; 78 adev->enable_virtual_display = true; 79 } 80 ddev->driver_features &= ~DRIVER_ATOMIC; 81 adev->cg_flags = 0; 82 adev->pg_flags = 0; 83 84 /* Reduce kcq number to 2 to reduce latency */ 85 if (amdgpu_num_kcq == -1) 86 amdgpu_num_kcq = 2; 87 } 88 89 /** 90 * amdgpu_virt_request_full_gpu() - request full gpu access 91 * @adev: amdgpu device. 92 * @init: is driver init time. 93 * When start to init/fini driver, first need to request full gpu access. 94 * Return: Zero if request success, otherwise will return error. 95 */ 96 int amdgpu_virt_request_full_gpu(struct amdgpu_device *adev, bool init) 97 { 98 struct amdgpu_virt *virt = &adev->virt; 99 int r; 100 101 if (virt->ops && virt->ops->req_full_gpu) { 102 r = virt->ops->req_full_gpu(adev, init); 103 if (r) { 104 adev->no_hw_access = true; 105 return r; 106 } 107 108 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME; 109 } 110 111 return 0; 112 } 113 114 /** 115 * amdgpu_virt_release_full_gpu() - release full gpu access 116 * @adev: amdgpu device. 117 * @init: is driver init time. 118 * When finishing driver init/fini, need to release full gpu access. 119 * Return: Zero if release success, otherwise will returen error. 120 */ 121 int amdgpu_virt_release_full_gpu(struct amdgpu_device *adev, bool init) 122 { 123 struct amdgpu_virt *virt = &adev->virt; 124 int r; 125 126 if (virt->ops && virt->ops->rel_full_gpu) { 127 r = virt->ops->rel_full_gpu(adev, init); 128 if (r) 129 return r; 130 131 adev->virt.caps |= AMDGPU_SRIOV_CAPS_RUNTIME; 132 } 133 return 0; 134 } 135 136 /** 137 * amdgpu_virt_reset_gpu() - reset gpu 138 * @adev: amdgpu device. 139 * Send reset command to GPU hypervisor to reset GPU that VM is using 140 * Return: Zero if reset success, otherwise will return error. 141 */ 142 int amdgpu_virt_reset_gpu(struct amdgpu_device *adev) 143 { 144 struct amdgpu_virt *virt = &adev->virt; 145 int r; 146 147 if (virt->ops && virt->ops->reset_gpu) { 148 r = virt->ops->reset_gpu(adev); 149 if (r) 150 return r; 151 152 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME; 153 } 154 155 return 0; 156 } 157 158 void amdgpu_virt_request_init_data(struct amdgpu_device *adev) 159 { 160 struct amdgpu_virt *virt = &adev->virt; 161 162 if (virt->ops && virt->ops->req_init_data) 163 virt->ops->req_init_data(adev); 164 165 if (adev->virt.req_init_data_ver > 0) 166 dev_info(adev->dev, "host supports REQ_INIT_DATA handshake of critical_region_version %d\n", 167 adev->virt.req_init_data_ver); 168 else 169 dev_warn(adev->dev, "host doesn't support REQ_INIT_DATA handshake\n"); 170 } 171 172 /** 173 * amdgpu_virt_ready_to_reset() - send ready to reset to host 174 * @adev: amdgpu device. 175 * Send ready to reset message to GPU hypervisor to signal we have stopped GPU 176 * activity and is ready for host FLR 177 */ 178 void amdgpu_virt_ready_to_reset(struct amdgpu_device *adev) 179 { 180 struct amdgpu_virt *virt = &adev->virt; 181 182 if (virt->ops && virt->ops->reset_gpu) 183 virt->ops->ready_to_reset(adev); 184 } 185 186 /** 187 * amdgpu_virt_wait_reset() - wait for reset gpu completed 188 * @adev: amdgpu device. 189 * Wait for GPU reset completed. 190 * Return: Zero if reset success, otherwise will return error. 191 */ 192 int amdgpu_virt_wait_reset(struct amdgpu_device *adev) 193 { 194 struct amdgpu_virt *virt = &adev->virt; 195 196 if (!virt->ops || !virt->ops->wait_reset) 197 return -EINVAL; 198 199 return virt->ops->wait_reset(adev); 200 } 201 202 /** 203 * amdgpu_virt_alloc_mm_table() - alloc memory for mm table 204 * @adev: amdgpu device. 205 * MM table is used by UVD and VCE for its initialization 206 * Return: Zero if allocate success. 207 */ 208 int amdgpu_virt_alloc_mm_table(struct amdgpu_device *adev) 209 { 210 int r; 211 212 if (!amdgpu_sriov_vf(adev) || adev->virt.mm_table.gpu_addr) 213 return 0; 214 215 r = amdgpu_bo_create_kernel(adev, PAGE_SIZE, PAGE_SIZE, 216 AMDGPU_GEM_DOMAIN_VRAM | 217 AMDGPU_GEM_DOMAIN_GTT, 218 &adev->virt.mm_table.bo, 219 &adev->virt.mm_table.gpu_addr, 220 (void *)&adev->virt.mm_table.cpu_addr); 221 if (r) { 222 dev_err(adev->dev, "failed to alloc mm table and error = %d.\n", r); 223 return r; 224 } 225 226 memset((void *)adev->virt.mm_table.cpu_addr, 0, PAGE_SIZE); 227 dev_info(adev->dev, "MM table gpu addr = 0x%llx, cpu addr = %p.\n", 228 adev->virt.mm_table.gpu_addr, 229 adev->virt.mm_table.cpu_addr); 230 return 0; 231 } 232 233 /** 234 * amdgpu_virt_free_mm_table() - free mm table memory 235 * @adev: amdgpu device. 236 * Free MM table memory 237 */ 238 void amdgpu_virt_free_mm_table(struct amdgpu_device *adev) 239 { 240 if (!amdgpu_sriov_vf(adev) || !adev->virt.mm_table.gpu_addr) 241 return; 242 243 amdgpu_bo_free_kernel(&adev->virt.mm_table.bo, 244 &adev->virt.mm_table.gpu_addr, 245 (void *)&adev->virt.mm_table.cpu_addr); 246 adev->virt.mm_table.gpu_addr = 0; 247 } 248 249 /** 250 * amdgpu_virt_rcvd_ras_interrupt() - receive ras interrupt 251 * @adev: amdgpu device. 252 * Check whether host sent RAS error message 253 * Return: true if found, otherwise false 254 */ 255 bool amdgpu_virt_rcvd_ras_interrupt(struct amdgpu_device *adev) 256 { 257 struct amdgpu_virt *virt = &adev->virt; 258 259 if (!virt->ops || !virt->ops->rcvd_ras_intr) 260 return false; 261 262 return virt->ops->rcvd_ras_intr(adev); 263 } 264 265 266 unsigned int amd_sriov_msg_checksum(void *obj, 267 unsigned long obj_size, 268 unsigned int key, 269 unsigned int checksum) 270 { 271 unsigned int ret = key; 272 unsigned long i = 0; 273 unsigned char *pos; 274 275 pos = (char *)obj; 276 /* calculate checksum */ 277 for (i = 0; i < obj_size; ++i) 278 ret += *(pos + i); 279 /* minus the checksum itself */ 280 pos = (char *)&checksum; 281 for (i = 0; i < sizeof(checksum); ++i) 282 ret -= *(pos + i); 283 return ret; 284 } 285 286 static int amdgpu_virt_init_ras_err_handler_data(struct amdgpu_device *adev) 287 { 288 struct amdgpu_virt *virt = &adev->virt; 289 struct amdgpu_virt_ras_err_handler_data **data = &virt->virt_eh_data; 290 /* GPU will be marked bad on host if bp count more then 10, 291 * so alloc 512 is enough. 292 */ 293 unsigned int align_space = 512; 294 void *bps = NULL; 295 struct amdgpu_bo **bps_bo = NULL; 296 297 *data = kmalloc_obj(struct amdgpu_virt_ras_err_handler_data); 298 if (!*data) 299 goto data_failure; 300 301 bps = kmalloc_objs(*(*data)->bps, align_space); 302 if (!bps) 303 goto bps_failure; 304 305 bps_bo = kmalloc_objs(*(*data)->bps_bo, align_space); 306 if (!bps_bo) 307 goto bps_bo_failure; 308 309 (*data)->bps = bps; 310 (*data)->bps_bo = bps_bo; 311 (*data)->count = 0; 312 (*data)->last_reserved = 0; 313 314 virt->ras_init_done = true; 315 316 return 0; 317 318 bps_bo_failure: 319 kfree(bps); 320 bps_failure: 321 kfree(*data); 322 data_failure: 323 return -ENOMEM; 324 } 325 326 static void amdgpu_virt_ras_release_bp(struct amdgpu_device *adev) 327 { 328 struct amdgpu_virt *virt = &adev->virt; 329 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data; 330 struct amdgpu_bo *bo; 331 int i; 332 333 if (!data) 334 return; 335 336 for (i = data->last_reserved - 1; i >= 0; i--) { 337 bo = data->bps_bo[i]; 338 if (bo) { 339 amdgpu_bo_free_kernel(&bo, NULL, NULL); 340 data->bps_bo[i] = bo; 341 } 342 data->last_reserved = i; 343 } 344 } 345 346 void amdgpu_virt_release_ras_err_handler_data(struct amdgpu_device *adev) 347 { 348 struct amdgpu_virt *virt = &adev->virt; 349 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data; 350 351 virt->ras_init_done = false; 352 353 if (!data) 354 return; 355 356 amdgpu_virt_ras_release_bp(adev); 357 358 kfree(data->bps); 359 kfree(data->bps_bo); 360 kfree(data); 361 virt->virt_eh_data = NULL; 362 } 363 364 static void amdgpu_virt_ras_add_bps(struct amdgpu_device *adev, 365 struct eeprom_table_record *bps, int pages) 366 { 367 struct amdgpu_virt *virt = &adev->virt; 368 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data; 369 370 if (!data) 371 return; 372 373 memcpy(&data->bps[data->count], bps, pages * sizeof(*data->bps)); 374 data->count += pages; 375 } 376 377 static void amdgpu_virt_ras_reserve_bps(struct amdgpu_device *adev) 378 { 379 struct amdgpu_virt *virt = &adev->virt; 380 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data; 381 struct amdgpu_vram_mgr *mgr = &adev->mman.vram_mgr; 382 struct ttm_resource_manager *man = &mgr->manager; 383 struct amdgpu_bo *bo = NULL; 384 uint64_t bp; 385 int i; 386 387 if (!data) 388 return; 389 390 for (i = data->last_reserved; i < data->count; i++) { 391 bp = data->bps[i].retired_page; 392 393 /* There are two cases of reserve error should be ignored: 394 * 1) a ras bad page has been allocated (used by someone); 395 * 2) a ras bad page has been reserved (duplicate error injection 396 * for one page); 397 */ 398 if (ttm_resource_manager_used(man)) { 399 amdgpu_vram_mgr_reserve_range(&adev->mman.vram_mgr, 400 bp << AMDGPU_GPU_PAGE_SHIFT, 401 AMDGPU_GPU_PAGE_SIZE); 402 data->bps_bo[i] = NULL; 403 } else { 404 if (amdgpu_bo_create_kernel_at(adev, bp << AMDGPU_GPU_PAGE_SHIFT, 405 AMDGPU_GPU_PAGE_SIZE, 406 &bo, NULL)) 407 dev_dbg(adev->dev, 408 "RAS WARN: reserve vram for retired page %llx fail\n", 409 bp); 410 data->bps_bo[i] = bo; 411 } 412 data->last_reserved = i + 1; 413 bo = NULL; 414 } 415 } 416 417 static bool amdgpu_virt_ras_check_bad_page(struct amdgpu_device *adev, 418 uint64_t retired_page) 419 { 420 struct amdgpu_virt *virt = &adev->virt; 421 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data; 422 int i; 423 424 if (!data) 425 return true; 426 427 for (i = 0; i < data->count; i++) 428 if (retired_page == data->bps[i].retired_page) 429 return true; 430 431 return false; 432 } 433 434 static void amdgpu_virt_add_bad_page(struct amdgpu_device *adev, 435 uint64_t bp_block_offset, uint32_t bp_block_size) 436 { 437 struct eeprom_table_record bp; 438 uint64_t retired_page; 439 uint32_t bp_idx, bp_cnt; 440 void *vram_usage_va = NULL; 441 442 if (adev->mman.fw_vram_usage_va) 443 vram_usage_va = adev->mman.fw_vram_usage_va; 444 else 445 vram_usage_va = adev->mman.drv_vram_usage_va; 446 447 memset(&bp, 0, sizeof(bp)); 448 449 if (bp_block_size) { 450 bp_cnt = bp_block_size / sizeof(uint64_t); 451 for (bp_idx = 0; bp_idx < bp_cnt; bp_idx++) { 452 retired_page = *(uint64_t *)(vram_usage_va + 453 bp_block_offset + bp_idx * sizeof(uint64_t)); 454 bp.retired_page = retired_page; 455 456 if (amdgpu_virt_ras_check_bad_page(adev, retired_page)) 457 continue; 458 459 amdgpu_virt_ras_add_bps(adev, &bp, 1); 460 461 amdgpu_virt_ras_reserve_bps(adev); 462 } 463 } 464 } 465 466 static int amdgpu_virt_read_pf2vf_data(struct amdgpu_device *adev) 467 { 468 struct amd_sriov_msg_pf2vf_info_header *pf2vf_info = adev->virt.fw_reserve.p_pf2vf; 469 uint32_t checksum; 470 uint32_t checkval; 471 472 uint32_t i; 473 uint32_t tmp; 474 475 if (adev->virt.fw_reserve.p_pf2vf == NULL) 476 return -EINVAL; 477 478 if (pf2vf_info->size > 1024) { 479 dev_err(adev->dev, "invalid pf2vf message size: 0x%x\n", pf2vf_info->size); 480 return -EINVAL; 481 } 482 483 switch (pf2vf_info->version) { 484 case 1: 485 checksum = ((struct amdgim_pf2vf_info_v1 *)pf2vf_info)->checksum; 486 checkval = amd_sriov_msg_checksum( 487 adev->virt.fw_reserve.p_pf2vf, pf2vf_info->size, 488 adev->virt.fw_reserve.checksum_key, checksum); 489 if (checksum != checkval) { 490 dev_err(adev->dev, 491 "invalid pf2vf message: header checksum=0x%x calculated checksum=0x%x\n", 492 checksum, checkval); 493 return -EINVAL; 494 } 495 496 adev->virt.gim_feature = 497 ((struct amdgim_pf2vf_info_v1 *)pf2vf_info)->feature_flags; 498 break; 499 case 2: 500 /* TODO: missing key, need to add it later */ 501 checksum = ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->checksum; 502 checkval = amd_sriov_msg_checksum( 503 adev->virt.fw_reserve.p_pf2vf, pf2vf_info->size, 504 0, checksum); 505 if (checksum != checkval) { 506 dev_err(adev->dev, 507 "invalid pf2vf message: header checksum=0x%x calculated checksum=0x%x\n", 508 checksum, checkval); 509 return -EINVAL; 510 } 511 512 adev->virt.vf2pf_update_interval_ms = 513 ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->vf2pf_update_interval_ms; 514 adev->virt.gim_feature = 515 ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->feature_flags.all; 516 adev->virt.reg_access = 517 ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->reg_access_flags.all; 518 519 adev->virt.decode_max_dimension_pixels = 0; 520 adev->virt.decode_max_frame_pixels = 0; 521 adev->virt.encode_max_dimension_pixels = 0; 522 adev->virt.encode_max_frame_pixels = 0; 523 adev->virt.is_mm_bw_enabled = false; 524 for (i = 0; i < AMD_SRIOV_MSG_RESERVE_VCN_INST; i++) { 525 tmp = ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->mm_bw_management[i].decode_max_dimension_pixels; 526 adev->virt.decode_max_dimension_pixels = max(tmp, adev->virt.decode_max_dimension_pixels); 527 528 tmp = ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->mm_bw_management[i].decode_max_frame_pixels; 529 adev->virt.decode_max_frame_pixels = max(tmp, adev->virt.decode_max_frame_pixels); 530 531 tmp = ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->mm_bw_management[i].encode_max_dimension_pixels; 532 adev->virt.encode_max_dimension_pixels = max(tmp, adev->virt.encode_max_dimension_pixels); 533 534 tmp = ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->mm_bw_management[i].encode_max_frame_pixels; 535 adev->virt.encode_max_frame_pixels = max(tmp, adev->virt.encode_max_frame_pixels); 536 } 537 if ((adev->virt.decode_max_dimension_pixels > 0) || (adev->virt.encode_max_dimension_pixels > 0)) 538 adev->virt.is_mm_bw_enabled = true; 539 540 adev->unique_id = 541 ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->uuid; 542 adev->virt.ras_en_caps.all = ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->ras_en_caps.all; 543 adev->virt.ras_telemetry_en_caps.all = 544 ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->ras_telemetry_en_caps.all; 545 break; 546 default: 547 dev_err(adev->dev, "invalid pf2vf version: 0x%x\n", pf2vf_info->version); 548 return -EINVAL; 549 } 550 551 /* correct too large or too little interval value */ 552 if (adev->virt.vf2pf_update_interval_ms < 200 || adev->virt.vf2pf_update_interval_ms > 10000) 553 adev->virt.vf2pf_update_interval_ms = 2000; 554 555 return 0; 556 } 557 558 static void amdgpu_virt_populate_vf2pf_ucode_info(struct amdgpu_device *adev) 559 { 560 struct amd_sriov_msg_vf2pf_info *vf2pf_info; 561 vf2pf_info = (struct amd_sriov_msg_vf2pf_info *) adev->virt.fw_reserve.p_vf2pf; 562 563 if (adev->virt.fw_reserve.p_vf2pf == NULL) 564 return; 565 566 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_VCE, adev->vce.fw_version); 567 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_UVD, adev->uvd.fw_version); 568 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_MC, adev->gmc.fw_version); 569 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_ME, adev->gfx.me_fw_version); 570 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_PFP, adev->gfx.pfp_fw_version); 571 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_CE, adev->gfx.ce_fw_version); 572 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_RLC, adev->gfx.rlc_fw_version); 573 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_RLC_SRLC, adev->gfx.rlc_srlc_fw_version); 574 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_RLC_SRLG, adev->gfx.rlc_srlg_fw_version); 575 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_RLC_SRLS, adev->gfx.rlc_srls_fw_version); 576 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_MEC, adev->gfx.mec_fw_version); 577 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_MEC2, adev->gfx.mec2_fw_version); 578 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_SOS, adev->psp.sos.fw_version); 579 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_ASD, 580 adev->psp.asd_context.bin_desc.fw_version); 581 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_TA_RAS, 582 adev->psp.ras_context.context.bin_desc.fw_version); 583 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_TA_XGMI, 584 adev->psp.xgmi_context.context.bin_desc.fw_version); 585 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_SMC, adev->pm.fw_version); 586 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_SDMA, adev->sdma.instance[0].fw_version); 587 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_SDMA2, adev->sdma.instance[1].fw_version); 588 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_VCN, adev->vcn.fw_version); 589 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_DMCU, adev->dm.dmcu_fw_version); 590 } 591 592 static int amdgpu_virt_write_vf2pf_data(struct amdgpu_device *adev) 593 { 594 struct amd_sriov_msg_vf2pf_info *vf2pf_info; 595 596 vf2pf_info = (struct amd_sriov_msg_vf2pf_info *) adev->virt.fw_reserve.p_vf2pf; 597 598 if (adev->virt.fw_reserve.p_vf2pf == NULL) 599 return -EINVAL; 600 601 memset(vf2pf_info, 0, sizeof(struct amd_sriov_msg_vf2pf_info)); 602 603 vf2pf_info->header.size = sizeof(struct amd_sriov_msg_vf2pf_info); 604 vf2pf_info->header.version = AMD_SRIOV_MSG_FW_VRAM_VF2PF_VER; 605 606 #ifdef MODULE 607 if (THIS_MODULE->version != NULL) 608 strscpy(vf2pf_info->driver_version, THIS_MODULE->version); 609 else 610 #endif 611 strscpy(vf2pf_info->driver_version, "N/A"); 612 613 vf2pf_info->pf2vf_version_required = 0; // no requirement, guest understands all 614 vf2pf_info->driver_cert = 0; 615 vf2pf_info->os_info.all = 0; 616 617 vf2pf_info->fb_usage = ttm_resource_manager_used(&adev->mman.vram_mgr.manager) ? 618 ttm_resource_manager_usage(&adev->mman.vram_mgr.manager) >> 20 : 0; 619 vf2pf_info->fb_vis_usage = 620 amdgpu_vram_mgr_vis_usage(&adev->mman.vram_mgr) >> 20; 621 vf2pf_info->fb_size = adev->gmc.real_vram_size >> 20; 622 vf2pf_info->fb_vis_size = adev->gmc.visible_vram_size >> 20; 623 624 amdgpu_virt_populate_vf2pf_ucode_info(adev); 625 626 /* TODO: read dynamic info */ 627 vf2pf_info->gfx_usage = 0; 628 vf2pf_info->compute_usage = 0; 629 vf2pf_info->encode_usage = 0; 630 vf2pf_info->decode_usage = 0; 631 632 vf2pf_info->dummy_page_addr = (uint64_t)adev->dummy_page_addr; 633 if (amdgpu_sriov_is_mes_info_enable(adev)) { 634 vf2pf_info->mes_info_addr = 635 (uint64_t)(adev->mes.resource_1_gpu_addr[0] + AMDGPU_GPU_PAGE_SIZE); 636 vf2pf_info->mes_info_size = 637 adev->mes.resource_1[0]->tbo.base.size - AMDGPU_GPU_PAGE_SIZE; 638 } 639 vf2pf_info->checksum = 640 amd_sriov_msg_checksum( 641 vf2pf_info, sizeof(*vf2pf_info), 0, 0); 642 643 return 0; 644 } 645 646 static void amdgpu_virt_update_vf2pf_work_item(struct work_struct *work) 647 { 648 struct amdgpu_device *adev = container_of(work, struct amdgpu_device, virt.vf2pf_work.work); 649 int ret; 650 651 ret = amdgpu_virt_read_pf2vf_data(adev); 652 if (ret) { 653 adev->virt.vf2pf_update_retry_cnt++; 654 655 if ((amdgpu_virt_rcvd_ras_interrupt(adev) || 656 adev->virt.vf2pf_update_retry_cnt >= AMDGPU_VF2PF_UPDATE_MAX_RETRY_LIMIT) && 657 amdgpu_sriov_runtime(adev)) { 658 659 amdgpu_ras_set_fed(adev, true); 660 if (amdgpu_reset_domain_schedule(adev->reset_domain, 661 &adev->kfd.reset_work)) 662 return; 663 else 664 dev_err(adev->dev, "Failed to queue work! at %s", __func__); 665 } 666 667 goto out; 668 } 669 670 adev->virt.vf2pf_update_retry_cnt = 0; 671 amdgpu_virt_write_vf2pf_data(adev); 672 673 out: 674 schedule_delayed_work(&(adev->virt.vf2pf_work), adev->virt.vf2pf_update_interval_ms); 675 } 676 677 static int amdgpu_virt_read_exchange_data_from_mem(struct amdgpu_device *adev, uint32_t *pfvf_data) 678 { 679 uint32_t dataexchange_offset = 680 adev->virt.crit_regn_tbl[AMD_SRIOV_MSG_DATAEXCHANGE_TABLE_ID].offset; 681 uint32_t dataexchange_size = 682 adev->virt.crit_regn_tbl[AMD_SRIOV_MSG_DATAEXCHANGE_TABLE_ID].size_kb << 10; 683 uint64_t pos = 0; 684 685 dev_info(adev->dev, 686 "Got data exchange info from dynamic crit_region_table at offset 0x%x with size of 0x%x bytes.\n", 687 dataexchange_offset, dataexchange_size); 688 689 if (!IS_ALIGNED(dataexchange_offset, 4) || !IS_ALIGNED(dataexchange_size, 4)) { 690 dev_err(adev->dev, "Data exchange data not aligned to 4 bytes\n"); 691 return -EINVAL; 692 } 693 694 pos = (uint64_t)dataexchange_offset; 695 amdgpu_device_vram_access(adev, pos, pfvf_data, 696 dataexchange_size, false); 697 698 return 0; 699 } 700 701 void amdgpu_virt_fini_data_exchange(struct amdgpu_device *adev) 702 { 703 if (adev->virt.vf2pf_update_interval_ms != 0) { 704 dev_info(adev->dev, "clean up the vf2pf work item\n"); 705 cancel_delayed_work_sync(&adev->virt.vf2pf_work); 706 adev->virt.vf2pf_update_interval_ms = 0; 707 } 708 } 709 710 void amdgpu_virt_init_data_exchange(struct amdgpu_device *adev) 711 { 712 uint32_t *pfvf_data = NULL; 713 714 adev->virt.fw_reserve.p_pf2vf = NULL; 715 adev->virt.fw_reserve.p_vf2pf = NULL; 716 adev->virt.vf2pf_update_interval_ms = 0; 717 adev->virt.vf2pf_update_retry_cnt = 0; 718 719 if (adev->mman.fw_vram_usage_va && adev->mman.drv_vram_usage_va) { 720 dev_warn(adev->dev, "Currently fw_vram and drv_vram should not have values at the same time!"); 721 } else if (adev->mman.fw_vram_usage_va || adev->mman.drv_vram_usage_va) { 722 /* go through this logic in ip_init and reset to init workqueue*/ 723 amdgpu_virt_exchange_data(adev); 724 725 INIT_DELAYED_WORK(&adev->virt.vf2pf_work, amdgpu_virt_update_vf2pf_work_item); 726 schedule_delayed_work(&(adev->virt.vf2pf_work), msecs_to_jiffies(adev->virt.vf2pf_update_interval_ms)); 727 } else if (adev->bios != NULL) { 728 /* got through this logic in early init stage to get necessary flags, e.g. rlcg_acc related*/ 729 if (adev->virt.req_init_data_ver == GPU_CRIT_REGION_V2) { 730 pfvf_data = 731 kzalloc(adev->virt.crit_regn_tbl[AMD_SRIOV_MSG_DATAEXCHANGE_TABLE_ID].size_kb << 10, 732 GFP_KERNEL); 733 if (!pfvf_data) { 734 dev_err(adev->dev, "Failed to allocate memory for pfvf_data\n"); 735 return; 736 } 737 738 if (amdgpu_virt_read_exchange_data_from_mem(adev, pfvf_data)) 739 goto free_pfvf_data; 740 741 adev->virt.fw_reserve.p_pf2vf = 742 (struct amd_sriov_msg_pf2vf_info_header *)pfvf_data; 743 744 amdgpu_virt_read_pf2vf_data(adev); 745 746 free_pfvf_data: 747 kfree(pfvf_data); 748 pfvf_data = NULL; 749 adev->virt.fw_reserve.p_pf2vf = NULL; 750 } else { 751 adev->virt.fw_reserve.p_pf2vf = 752 (struct amd_sriov_msg_pf2vf_info_header *) 753 (adev->bios + (AMD_SRIOV_MSG_PF2VF_OFFSET_KB_V1 << 10)); 754 755 amdgpu_virt_read_pf2vf_data(adev); 756 } 757 } 758 } 759 760 761 void amdgpu_virt_exchange_data(struct amdgpu_device *adev) 762 { 763 uint64_t bp_block_offset = 0; 764 uint32_t bp_block_size = 0; 765 struct amd_sriov_msg_pf2vf_info *pf2vf_v2 = NULL; 766 767 if (adev->mman.fw_vram_usage_va || adev->mman.drv_vram_usage_va) { 768 if (adev->mman.fw_vram_usage_va) { 769 if (adev->virt.req_init_data_ver == GPU_CRIT_REGION_V2) { 770 adev->virt.fw_reserve.p_pf2vf = 771 (struct amd_sriov_msg_pf2vf_info_header *) 772 (adev->mman.fw_vram_usage_va + 773 adev->virt.crit_regn_tbl[AMD_SRIOV_MSG_DATAEXCHANGE_TABLE_ID].offset); 774 adev->virt.fw_reserve.p_vf2pf = 775 (struct amd_sriov_msg_vf2pf_info_header *) 776 (adev->mman.fw_vram_usage_va + 777 adev->virt.crit_regn_tbl[AMD_SRIOV_MSG_DATAEXCHANGE_TABLE_ID].offset + 778 (AMD_SRIOV_MSG_SIZE_KB << 10)); 779 adev->virt.fw_reserve.ras_telemetry = 780 (adev->mman.fw_vram_usage_va + 781 adev->virt.crit_regn_tbl[AMD_SRIOV_MSG_RAS_TELEMETRY_TABLE_ID].offset); 782 } else { 783 adev->virt.fw_reserve.p_pf2vf = 784 (struct amd_sriov_msg_pf2vf_info_header *) 785 (adev->mman.fw_vram_usage_va + (AMD_SRIOV_MSG_PF2VF_OFFSET_KB_V1 << 10)); 786 adev->virt.fw_reserve.p_vf2pf = 787 (struct amd_sriov_msg_vf2pf_info_header *) 788 (adev->mman.fw_vram_usage_va + (AMD_SRIOV_MSG_VF2PF_OFFSET_KB_V1 << 10)); 789 adev->virt.fw_reserve.ras_telemetry = 790 (adev->mman.fw_vram_usage_va + (AMD_SRIOV_MSG_RAS_TELEMETRY_OFFSET_KB_V1 << 10)); 791 } 792 } else if (adev->mman.drv_vram_usage_va) { 793 adev->virt.fw_reserve.p_pf2vf = 794 (struct amd_sriov_msg_pf2vf_info_header *) 795 (adev->mman.drv_vram_usage_va + (AMD_SRIOV_MSG_PF2VF_OFFSET_KB_V1 << 10)); 796 adev->virt.fw_reserve.p_vf2pf = 797 (struct amd_sriov_msg_vf2pf_info_header *) 798 (adev->mman.drv_vram_usage_va + (AMD_SRIOV_MSG_VF2PF_OFFSET_KB_V1 << 10)); 799 adev->virt.fw_reserve.ras_telemetry = 800 (adev->mman.drv_vram_usage_va + (AMD_SRIOV_MSG_RAS_TELEMETRY_OFFSET_KB_V1 << 10)); 801 } 802 803 amdgpu_virt_read_pf2vf_data(adev); 804 amdgpu_virt_write_vf2pf_data(adev); 805 806 /* bad page handling for version 2 */ 807 if (adev->virt.fw_reserve.p_pf2vf->version == 2) { 808 pf2vf_v2 = (struct amd_sriov_msg_pf2vf_info *)adev->virt.fw_reserve.p_pf2vf; 809 810 bp_block_offset = ((uint64_t)pf2vf_v2->bp_block_offset_low & 0xFFFFFFFF) | 811 ((((uint64_t)pf2vf_v2->bp_block_offset_high) << 32) & 0xFFFFFFFF00000000); 812 bp_block_size = pf2vf_v2->bp_block_size; 813 814 if (bp_block_size && !adev->virt.ras_init_done) 815 amdgpu_virt_init_ras_err_handler_data(adev); 816 817 if (adev->virt.ras_init_done) 818 amdgpu_virt_add_bad_page(adev, bp_block_offset, bp_block_size); 819 } 820 } 821 } 822 823 static u32 amdgpu_virt_init_detect_asic(struct amdgpu_device *adev) 824 { 825 uint32_t reg; 826 827 switch (adev->asic_type) { 828 case CHIP_TONGA: 829 case CHIP_FIJI: 830 reg = RREG32(mmBIF_IOV_FUNC_IDENTIFIER); 831 break; 832 case CHIP_VEGA10: 833 case CHIP_VEGA20: 834 case CHIP_NAVI10: 835 case CHIP_NAVI12: 836 case CHIP_SIENNA_CICHLID: 837 case CHIP_ARCTURUS: 838 case CHIP_ALDEBARAN: 839 case CHIP_IP_DISCOVERY: 840 reg = RREG32(mmRCC_IOV_FUNC_IDENTIFIER); 841 break; 842 default: /* other chip doesn't support SRIOV */ 843 reg = 0; 844 break; 845 } 846 847 if (reg & 1) 848 adev->virt.caps |= AMDGPU_SRIOV_CAPS_IS_VF; 849 850 if (reg & 0x80000000) 851 adev->virt.caps |= AMDGPU_SRIOV_CAPS_ENABLE_IOV; 852 853 if (!reg) { 854 /* passthrough mode exclus sriov mod */ 855 if (is_virtual_machine() && !xen_initial_domain()) 856 adev->virt.caps |= AMDGPU_PASSTHROUGH_MODE; 857 } 858 859 return reg; 860 } 861 862 static bool amdgpu_virt_init_req_data(struct amdgpu_device *adev, u32 reg) 863 { 864 bool is_sriov = false; 865 866 /* we have the ability to check now */ 867 if (amdgpu_sriov_vf(adev)) { 868 is_sriov = true; 869 870 switch (adev->asic_type) { 871 case CHIP_TONGA: 872 case CHIP_FIJI: 873 vi_set_virt_ops(adev); 874 break; 875 case CHIP_VEGA10: 876 soc15_set_virt_ops(adev); 877 #ifdef CONFIG_X86 878 /* not send GPU_INIT_DATA with MS_HYPERV*/ 879 if (!hypervisor_is_type(X86_HYPER_MS_HYPERV)) 880 #endif 881 /* send a dummy GPU_INIT_DATA request to host on vega10 */ 882 amdgpu_virt_request_init_data(adev); 883 break; 884 case CHIP_VEGA20: 885 case CHIP_ARCTURUS: 886 case CHIP_ALDEBARAN: 887 soc15_set_virt_ops(adev); 888 break; 889 case CHIP_NAVI10: 890 case CHIP_NAVI12: 891 case CHIP_SIENNA_CICHLID: 892 case CHIP_IP_DISCOVERY: 893 nv_set_virt_ops(adev); 894 /* try send GPU_INIT_DATA request to host */ 895 amdgpu_virt_request_init_data(adev); 896 break; 897 default: /* other chip doesn't support SRIOV */ 898 is_sriov = false; 899 dev_err(adev->dev, "Unknown asic type: %d!\n", adev->asic_type); 900 break; 901 } 902 } 903 904 return is_sriov; 905 } 906 907 static void amdgpu_virt_init_ras(struct amdgpu_device *adev) 908 { 909 ratelimit_state_init(&adev->virt.ras.ras_error_cnt_rs, 5 * HZ, 1); 910 ratelimit_state_init(&adev->virt.ras.ras_cper_dump_rs, 5 * HZ, 1); 911 ratelimit_state_init(&adev->virt.ras.ras_chk_criti_rs, 5 * HZ, 1); 912 913 ratelimit_set_flags(&adev->virt.ras.ras_error_cnt_rs, 914 RATELIMIT_MSG_ON_RELEASE); 915 ratelimit_set_flags(&adev->virt.ras.ras_cper_dump_rs, 916 RATELIMIT_MSG_ON_RELEASE); 917 ratelimit_set_flags(&adev->virt.ras.ras_chk_criti_rs, 918 RATELIMIT_MSG_ON_RELEASE); 919 920 mutex_init(&adev->virt.ras.ras_telemetry_mutex); 921 mutex_init(&adev->virt.access_req_mutex); 922 923 adev->virt.ras.cper_rptr = 0; 924 } 925 926 static uint8_t amdgpu_virt_crit_region_calc_checksum(uint8_t *buf_start, uint8_t *buf_end) 927 { 928 uint32_t sum = 0; 929 930 if (buf_start >= buf_end) 931 return 0; 932 933 for (; buf_start < buf_end; buf_start++) 934 sum += buf_start[0]; 935 936 return 0xffffffff - sum; 937 } 938 939 int amdgpu_virt_init_critical_region(struct amdgpu_device *adev) 940 { 941 struct amd_sriov_msg_init_data_header *init_data_hdr = NULL; 942 u64 init_hdr_offset = adev->virt.init_data_header.offset; 943 u64 init_hdr_size = (u64)adev->virt.init_data_header.size_kb << 10; /* KB → bytes */ 944 u64 vram_size; 945 u64 end; 946 int r = 0; 947 uint8_t checksum = 0; 948 949 /* Skip below init if critical region version != v2 */ 950 if (adev->virt.req_init_data_ver != GPU_CRIT_REGION_V2) 951 return 0; 952 953 vram_size = RREG32(mmRCC_CONFIG_MEMSIZE); 954 if (!vram_size || vram_size == U32_MAX) 955 return -EINVAL; 956 vram_size <<= 20; 957 958 if (check_add_overflow(init_hdr_offset, init_hdr_size, &end) || end > vram_size) { 959 dev_err(adev->dev, "init_data_header exceeds VRAM size, exiting\n"); 960 return -EINVAL; 961 } 962 963 /* Allocate for init_data_hdr */ 964 init_data_hdr = kzalloc_obj(struct amd_sriov_msg_init_data_header); 965 if (!init_data_hdr) 966 return -ENOMEM; 967 968 amdgpu_device_vram_access(adev, (uint64_t)init_hdr_offset, (uint32_t *)init_data_hdr, 969 sizeof(struct amd_sriov_msg_init_data_header), false); 970 971 /* Table validation */ 972 if (strncmp(init_data_hdr->signature, 973 AMDGPU_SRIOV_CRIT_DATA_SIGNATURE, 974 AMDGPU_SRIOV_CRIT_DATA_SIG_LEN) != 0) { 975 dev_err(adev->dev, "Invalid init data signature: %.4s\n", 976 init_data_hdr->signature); 977 r = -EINVAL; 978 goto out; 979 } 980 981 checksum = amdgpu_virt_crit_region_calc_checksum( 982 (uint8_t *)&init_data_hdr->initdata_offset, 983 (uint8_t *)init_data_hdr + 984 sizeof(struct amd_sriov_msg_init_data_header)); 985 if (checksum != init_data_hdr->checksum) { 986 dev_err(adev->dev, "Found unmatching checksum from calculation 0x%x and init_data 0x%x\n", 987 checksum, init_data_hdr->checksum); 988 r = -EINVAL; 989 goto out; 990 } 991 992 memset(&adev->virt.crit_regn, 0, sizeof(adev->virt.crit_regn)); 993 memset(adev->virt.crit_regn_tbl, 0, sizeof(adev->virt.crit_regn_tbl)); 994 995 adev->virt.crit_regn.offset = init_data_hdr->initdata_offset; 996 adev->virt.crit_regn.size_kb = init_data_hdr->initdata_size_in_kb; 997 998 /* Validation and initialization for each table entry */ 999 if (IS_SRIOV_CRIT_REGN_ENTRY_VALID(init_data_hdr, AMD_SRIOV_MSG_IPD_TABLE_ID)) { 1000 if (!init_data_hdr->ip_discovery_size_in_kb || 1001 init_data_hdr->ip_discovery_size_in_kb > DISCOVERY_TMR_SIZE) { 1002 dev_err(adev->dev, "Invalid %s size: 0x%x\n", 1003 amdgpu_virt_dynamic_crit_table_name[AMD_SRIOV_MSG_IPD_TABLE_ID], 1004 init_data_hdr->ip_discovery_size_in_kb); 1005 r = -EINVAL; 1006 goto out; 1007 } 1008 1009 adev->virt.crit_regn_tbl[AMD_SRIOV_MSG_IPD_TABLE_ID].offset = 1010 init_data_hdr->ip_discovery_offset; 1011 adev->virt.crit_regn_tbl[AMD_SRIOV_MSG_IPD_TABLE_ID].size_kb = 1012 init_data_hdr->ip_discovery_size_in_kb; 1013 } 1014 1015 if (IS_SRIOV_CRIT_REGN_ENTRY_VALID(init_data_hdr, AMD_SRIOV_MSG_VBIOS_IMG_TABLE_ID)) { 1016 if (!init_data_hdr->vbios_img_size_in_kb) { 1017 dev_err(adev->dev, "Invalid %s size: 0x%x\n", 1018 amdgpu_virt_dynamic_crit_table_name[AMD_SRIOV_MSG_VBIOS_IMG_TABLE_ID], 1019 init_data_hdr->vbios_img_size_in_kb); 1020 r = -EINVAL; 1021 goto out; 1022 } 1023 1024 adev->virt.crit_regn_tbl[AMD_SRIOV_MSG_VBIOS_IMG_TABLE_ID].offset = 1025 init_data_hdr->vbios_img_offset; 1026 adev->virt.crit_regn_tbl[AMD_SRIOV_MSG_VBIOS_IMG_TABLE_ID].size_kb = 1027 init_data_hdr->vbios_img_size_in_kb; 1028 } 1029 1030 if (IS_SRIOV_CRIT_REGN_ENTRY_VALID(init_data_hdr, AMD_SRIOV_MSG_RAS_TELEMETRY_TABLE_ID)) { 1031 if (!init_data_hdr->ras_tele_info_size_in_kb) { 1032 dev_err(adev->dev, "Invalid %s size: 0x%x\n", 1033 amdgpu_virt_dynamic_crit_table_name[AMD_SRIOV_MSG_RAS_TELEMETRY_TABLE_ID], 1034 init_data_hdr->ras_tele_info_size_in_kb); 1035 r = -EINVAL; 1036 goto out; 1037 } 1038 1039 adev->virt.crit_regn_tbl[AMD_SRIOV_MSG_RAS_TELEMETRY_TABLE_ID].offset = 1040 init_data_hdr->ras_tele_info_offset; 1041 adev->virt.crit_regn_tbl[AMD_SRIOV_MSG_RAS_TELEMETRY_TABLE_ID].size_kb = 1042 init_data_hdr->ras_tele_info_size_in_kb; 1043 } 1044 1045 if (IS_SRIOV_CRIT_REGN_ENTRY_VALID(init_data_hdr, AMD_SRIOV_MSG_DATAEXCHANGE_TABLE_ID)) { 1046 if (!init_data_hdr->dataexchange_size_in_kb) { 1047 dev_err(adev->dev, "Invalid %s size: 0x%x\n", 1048 amdgpu_virt_dynamic_crit_table_name[AMD_SRIOV_MSG_DATAEXCHANGE_TABLE_ID], 1049 init_data_hdr->dataexchange_size_in_kb); 1050 r = -EINVAL; 1051 goto out; 1052 } 1053 1054 adev->virt.crit_regn_tbl[AMD_SRIOV_MSG_DATAEXCHANGE_TABLE_ID].offset = 1055 init_data_hdr->dataexchange_offset; 1056 adev->virt.crit_regn_tbl[AMD_SRIOV_MSG_DATAEXCHANGE_TABLE_ID].size_kb = 1057 init_data_hdr->dataexchange_size_in_kb; 1058 } 1059 1060 if (IS_SRIOV_CRIT_REGN_ENTRY_VALID(init_data_hdr, AMD_SRIOV_MSG_BAD_PAGE_INFO_TABLE_ID)) { 1061 if (!init_data_hdr->bad_page_size_in_kb) { 1062 dev_err(adev->dev, "Invalid %s size: 0x%x\n", 1063 amdgpu_virt_dynamic_crit_table_name[AMD_SRIOV_MSG_BAD_PAGE_INFO_TABLE_ID], 1064 init_data_hdr->bad_page_size_in_kb); 1065 r = -EINVAL; 1066 goto out; 1067 } 1068 1069 adev->virt.crit_regn_tbl[AMD_SRIOV_MSG_BAD_PAGE_INFO_TABLE_ID].offset = 1070 init_data_hdr->bad_page_info_offset; 1071 adev->virt.crit_regn_tbl[AMD_SRIOV_MSG_BAD_PAGE_INFO_TABLE_ID].size_kb = 1072 init_data_hdr->bad_page_size_in_kb; 1073 } 1074 1075 /* Validation for critical region info */ 1076 if (adev->virt.crit_regn_tbl[AMD_SRIOV_MSG_IPD_TABLE_ID].size_kb > DISCOVERY_TMR_SIZE) { 1077 dev_err(adev->dev, "Invalid IP discovery size: 0x%x\n", 1078 adev->virt.crit_regn_tbl[AMD_SRIOV_MSG_IPD_TABLE_ID].size_kb); 1079 r = -EINVAL; 1080 goto out; 1081 } 1082 1083 /* reserved memory starts from crit region base offset with the size of 5MB */ 1084 adev->mman.fw_vram_usage_start_offset = adev->virt.crit_regn.offset; 1085 adev->mman.fw_vram_usage_size = adev->virt.crit_regn.size_kb << 10; 1086 dev_info(adev->dev, 1087 "critical region v%d requested to reserve memory start at %08llx with %llu KB.\n", 1088 init_data_hdr->version, 1089 adev->mman.fw_vram_usage_start_offset, 1090 adev->mman.fw_vram_usage_size >> 10); 1091 1092 adev->virt.is_dynamic_crit_regn_enabled = true; 1093 1094 out: 1095 kfree(init_data_hdr); 1096 init_data_hdr = NULL; 1097 1098 return r; 1099 } 1100 1101 int amdgpu_virt_get_dynamic_data_info(struct amdgpu_device *adev, 1102 int data_id, uint8_t *binary, u32 *size) 1103 { 1104 uint32_t data_offset = 0; 1105 uint32_t data_size = 0; 1106 enum amd_sriov_msg_table_id_enum data_table_id = data_id; 1107 1108 if (data_table_id >= AMD_SRIOV_MSG_MAX_TABLE_ID) 1109 return -EINVAL; 1110 1111 data_offset = adev->virt.crit_regn_tbl[data_table_id].offset; 1112 data_size = adev->virt.crit_regn_tbl[data_table_id].size_kb << 10; 1113 1114 /* Validate on input params */ 1115 if (!binary || !size || *size < (uint64_t)data_size) 1116 return -EINVAL; 1117 1118 /* Proceed to copy the dynamic content */ 1119 amdgpu_device_vram_access(adev, 1120 (uint64_t)data_offset, (uint32_t *)binary, data_size, false); 1121 *size = (uint64_t)data_size; 1122 1123 dev_dbg(adev->dev, 1124 "Got %s info from dynamic crit_region_table at offset 0x%x with size of 0x%x bytes.\n", 1125 amdgpu_virt_dynamic_crit_table_name[data_id], data_offset, data_size); 1126 1127 return 0; 1128 } 1129 1130 void amdgpu_virt_init(struct amdgpu_device *adev) 1131 { 1132 bool is_sriov = false; 1133 uint32_t reg = amdgpu_virt_init_detect_asic(adev); 1134 1135 is_sriov = amdgpu_virt_init_req_data(adev, reg); 1136 1137 if (is_sriov) 1138 amdgpu_virt_init_ras(adev); 1139 } 1140 1141 static bool amdgpu_virt_access_debugfs_is_mmio(struct amdgpu_device *adev) 1142 { 1143 return amdgpu_sriov_is_debug(adev) ? true : false; 1144 } 1145 1146 static bool amdgpu_virt_access_debugfs_is_kiq(struct amdgpu_device *adev) 1147 { 1148 return amdgpu_sriov_is_normal(adev) ? true : false; 1149 } 1150 1151 int amdgpu_virt_enable_access_debugfs(struct amdgpu_device *adev) 1152 { 1153 if (!amdgpu_sriov_vf(adev) || 1154 amdgpu_virt_access_debugfs_is_kiq(adev)) 1155 return 0; 1156 1157 if (amdgpu_virt_access_debugfs_is_mmio(adev)) 1158 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME; 1159 else 1160 return -EPERM; 1161 1162 return 0; 1163 } 1164 1165 void amdgpu_virt_disable_access_debugfs(struct amdgpu_device *adev) 1166 { 1167 if (amdgpu_sriov_vf(adev)) 1168 adev->virt.caps |= AMDGPU_SRIOV_CAPS_RUNTIME; 1169 } 1170 1171 enum amdgpu_sriov_vf_mode amdgpu_virt_get_sriov_vf_mode(struct amdgpu_device *adev) 1172 { 1173 enum amdgpu_sriov_vf_mode mode; 1174 1175 if (amdgpu_sriov_vf(adev)) { 1176 if (amdgpu_sriov_is_pp_one_vf(adev)) 1177 mode = SRIOV_VF_MODE_ONE_VF; 1178 else 1179 mode = SRIOV_VF_MODE_MULTI_VF; 1180 } else { 1181 mode = SRIOV_VF_MODE_BARE_METAL; 1182 } 1183 1184 return mode; 1185 } 1186 1187 void amdgpu_virt_pre_reset(struct amdgpu_device *adev) 1188 { 1189 /* stop the data exchange thread */ 1190 amdgpu_virt_fini_data_exchange(adev); 1191 amdgpu_dpm_set_mp1_state(adev, PP_MP1_STATE_FLR); 1192 } 1193 1194 void amdgpu_virt_post_reset(struct amdgpu_device *adev) 1195 { 1196 if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(11, 0, 3)) { 1197 /* force set to GFXOFF state after reset, 1198 * to avoid some invalid operation before GC enable 1199 */ 1200 adev->gfx.is_poweron = false; 1201 } 1202 1203 adev->mes.ring[0].sched.ready = false; 1204 } 1205 1206 bool amdgpu_virt_fw_load_skip_check(struct amdgpu_device *adev, uint32_t ucode_id) 1207 { 1208 switch (amdgpu_ip_version(adev, MP0_HWIP, 0)) { 1209 case IP_VERSION(13, 0, 0): 1210 /* no vf autoload, white list */ 1211 if (ucode_id == AMDGPU_UCODE_ID_VCN1 || 1212 ucode_id == AMDGPU_UCODE_ID_VCN) 1213 return false; 1214 else 1215 return true; 1216 case IP_VERSION(11, 0, 9): 1217 case IP_VERSION(11, 0, 7): 1218 /* black list for CHIP_NAVI12 and CHIP_SIENNA_CICHLID */ 1219 if (ucode_id == AMDGPU_UCODE_ID_RLC_G 1220 || ucode_id == AMDGPU_UCODE_ID_RLC_RESTORE_LIST_CNTL 1221 || ucode_id == AMDGPU_UCODE_ID_RLC_RESTORE_LIST_GPM_MEM 1222 || ucode_id == AMDGPU_UCODE_ID_RLC_RESTORE_LIST_SRM_MEM 1223 || ucode_id == AMDGPU_UCODE_ID_SMC) 1224 return true; 1225 else 1226 return false; 1227 case IP_VERSION(13, 0, 10): 1228 /* white list */ 1229 if (ucode_id == AMDGPU_UCODE_ID_CAP 1230 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_PFP 1231 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_ME 1232 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_MEC 1233 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_PFP_P0_STACK 1234 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_PFP_P1_STACK 1235 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_ME_P0_STACK 1236 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_ME_P1_STACK 1237 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_MEC_P0_STACK 1238 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_MEC_P1_STACK 1239 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_MEC_P2_STACK 1240 || ucode_id == AMDGPU_UCODE_ID_CP_RS64_MEC_P3_STACK 1241 || ucode_id == AMDGPU_UCODE_ID_CP_MES 1242 || ucode_id == AMDGPU_UCODE_ID_CP_MES_DATA 1243 || ucode_id == AMDGPU_UCODE_ID_CP_MES1 1244 || ucode_id == AMDGPU_UCODE_ID_CP_MES1_DATA 1245 || ucode_id == AMDGPU_UCODE_ID_VCN1 1246 || ucode_id == AMDGPU_UCODE_ID_VCN) 1247 return false; 1248 else 1249 return true; 1250 default: 1251 /* lagacy black list */ 1252 if (ucode_id == AMDGPU_UCODE_ID_SDMA0 1253 || ucode_id == AMDGPU_UCODE_ID_SDMA1 1254 || ucode_id == AMDGPU_UCODE_ID_SDMA2 1255 || ucode_id == AMDGPU_UCODE_ID_SDMA3 1256 || ucode_id == AMDGPU_UCODE_ID_SDMA4 1257 || ucode_id == AMDGPU_UCODE_ID_SDMA5 1258 || ucode_id == AMDGPU_UCODE_ID_SDMA6 1259 || ucode_id == AMDGPU_UCODE_ID_SDMA7 1260 || ucode_id == AMDGPU_UCODE_ID_SDMA_RS64 1261 || ucode_id == AMDGPU_UCODE_ID_RLC_G 1262 || ucode_id == AMDGPU_UCODE_ID_RLC_RESTORE_LIST_CNTL 1263 || ucode_id == AMDGPU_UCODE_ID_RLC_RESTORE_LIST_GPM_MEM 1264 || ucode_id == AMDGPU_UCODE_ID_RLC_RESTORE_LIST_SRM_MEM 1265 || ucode_id == AMDGPU_UCODE_ID_SMC) 1266 return true; 1267 else 1268 return false; 1269 } 1270 } 1271 1272 void amdgpu_virt_update_sriov_video_codec(struct amdgpu_device *adev, 1273 struct amdgpu_video_codec_info *encode, uint32_t encode_array_size, 1274 struct amdgpu_video_codec_info *decode, uint32_t decode_array_size) 1275 { 1276 uint32_t i; 1277 1278 if (!adev->virt.is_mm_bw_enabled) 1279 return; 1280 1281 if (encode) { 1282 for (i = 0; i < encode_array_size; i++) { 1283 encode[i].max_width = adev->virt.encode_max_dimension_pixels; 1284 encode[i].max_pixels_per_frame = adev->virt.encode_max_frame_pixels; 1285 if (encode[i].max_width > 0) 1286 encode[i].max_height = encode[i].max_pixels_per_frame / encode[i].max_width; 1287 else 1288 encode[i].max_height = 0; 1289 } 1290 } 1291 1292 if (decode) { 1293 for (i = 0; i < decode_array_size; i++) { 1294 decode[i].max_width = adev->virt.decode_max_dimension_pixels; 1295 decode[i].max_pixels_per_frame = adev->virt.decode_max_frame_pixels; 1296 if (decode[i].max_width > 0) 1297 decode[i].max_height = decode[i].max_pixels_per_frame / decode[i].max_width; 1298 else 1299 decode[i].max_height = 0; 1300 } 1301 } 1302 } 1303 1304 bool amdgpu_virt_get_rlcg_reg_access_flag(struct amdgpu_device *adev, 1305 u32 acc_flags, u32 hwip, 1306 bool write, u32 *rlcg_flag) 1307 { 1308 bool ret = false; 1309 1310 switch (hwip) { 1311 case GC_HWIP: 1312 if (amdgpu_sriov_reg_indirect_gc(adev)) { 1313 *rlcg_flag = 1314 write ? AMDGPU_RLCG_GC_WRITE : AMDGPU_RLCG_GC_READ; 1315 ret = true; 1316 /* only in new version, AMDGPU_REGS_NO_KIQ and 1317 * AMDGPU_REGS_RLC are enabled simultaneously */ 1318 } else if ((acc_flags & AMDGPU_REGS_RLC) && 1319 !(acc_flags & AMDGPU_REGS_NO_KIQ) && write) { 1320 *rlcg_flag = AMDGPU_RLCG_GC_WRITE_LEGACY; 1321 ret = true; 1322 } 1323 break; 1324 case MMHUB_HWIP: 1325 if (amdgpu_sriov_reg_indirect_mmhub(adev) && 1326 (acc_flags & AMDGPU_REGS_RLC) && write) { 1327 *rlcg_flag = AMDGPU_RLCG_MMHUB_WRITE; 1328 ret = true; 1329 } 1330 break; 1331 default: 1332 break; 1333 } 1334 return ret; 1335 } 1336 1337 static u32 amdgpu_virt_rlcg_vfi_reg_rw(struct amdgpu_device *adev, u32 offset, u32 v, u32 flag, u32 xcc_id) 1338 { 1339 uint32_t timeout = 100; 1340 uint32_t i; 1341 1342 struct amdgpu_rlcg_reg_access_ctrl *reg_access_ctrl; 1343 void *vfi_cmd; 1344 void *vfi_stat; 1345 void *vfi_addr; 1346 void *vfi_data; 1347 void *vfi_grbm_cntl; 1348 void *vfi_grbm_idx; 1349 uint32_t cmd; 1350 uint32_t stat; 1351 uint32_t addr = offset; 1352 uint32_t data; 1353 uint32_t grbm_cntl_data; 1354 uint32_t grbm_idx_data; 1355 1356 unsigned long flags; 1357 bool is_err = true; 1358 1359 if (!adev->gfx.rlc.rlcg_reg_access_supported) { 1360 dev_err(adev->dev, "VFi interface is not available\n"); 1361 return 0; 1362 } 1363 1364 if (adev->gfx.xcc_mask && (((1 << xcc_id) & adev->gfx.xcc_mask) == 0)) { 1365 dev_err(adev->dev, "VFi invalid XCC, xcc_id=0x%x\n", xcc_id); 1366 return 0; 1367 } 1368 1369 if (amdgpu_device_skip_hw_access(adev)) 1370 return 0; 1371 1372 reg_access_ctrl = &adev->gfx.rlc.reg_access_ctrl[xcc_id]; 1373 vfi_cmd = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->vfi_cmd; 1374 vfi_stat = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->vfi_stat; 1375 vfi_addr = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->vfi_addr; 1376 vfi_data = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->vfi_data; 1377 vfi_grbm_cntl = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->vfi_grbm_cntl; 1378 vfi_grbm_idx = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->vfi_grbm_idx; 1379 grbm_cntl_data = reg_access_ctrl->vfi_grbm_cntl_data; 1380 grbm_idx_data = reg_access_ctrl->vfi_grbm_idx_data; 1381 1382 if (flag == AMDGPU_RLCG_GC_WRITE) { 1383 data = v; 1384 cmd = AMDGPU_RLCG_VFI_CMD__WR; 1385 1386 // the GRBM_GFX_CNTL and GRBM_GFX_INDEX are protected by mutex outside this call 1387 if (addr == reg_access_ctrl->grbm_cntl) { 1388 reg_access_ctrl->vfi_grbm_cntl_data = data; 1389 return 0; 1390 } else if (addr == reg_access_ctrl->grbm_idx) { 1391 reg_access_ctrl->vfi_grbm_idx_data = data; 1392 return 0; 1393 } 1394 1395 } else if (flag == AMDGPU_RLCG_GC_READ) { 1396 data = 0; 1397 cmd = AMDGPU_RLCG_VFI_CMD__RD; 1398 1399 // the GRBM_GFX_CNTL and GRBM_GFX_INDEX are protected by mutex outside this call 1400 if (addr == reg_access_ctrl->grbm_cntl) 1401 return grbm_cntl_data; 1402 else if (addr == reg_access_ctrl->grbm_idx) 1403 return grbm_idx_data; 1404 1405 } else { 1406 dev_err(adev->dev, "VFi invalid access, flag=0x%x\n", flag); 1407 return 0; 1408 } 1409 1410 spin_lock_irqsave(&adev->virt.rlcg_reg_lock, flags); 1411 1412 writel(addr, vfi_addr); 1413 writel(data, vfi_data); 1414 writel(grbm_cntl_data, vfi_grbm_cntl); 1415 writel(grbm_idx_data, vfi_grbm_idx); 1416 1417 writel(AMDGPU_RLCG_VFI_STAT__BUSY, vfi_stat); 1418 writel(cmd, vfi_cmd); 1419 1420 for (i = 0; i < timeout; i++) { 1421 stat = readl(vfi_stat); 1422 if (stat != AMDGPU_RLCG_VFI_STAT__BUSY) 1423 break; 1424 udelay(10); 1425 } 1426 1427 switch (stat) { 1428 case AMDGPU_RLCG_VFI_STAT__DONE: 1429 is_err = false; 1430 if (cmd == AMDGPU_RLCG_VFI_CMD__RD) 1431 data = readl(vfi_data); 1432 break; 1433 case AMDGPU_RLCG_VFI_STAT__BUSY: 1434 dev_err(adev->dev, "VFi access timeout\n"); 1435 break; 1436 case AMDGPU_RLCG_VFI_STAT__INV_CMD: 1437 dev_err(adev->dev, "VFi invalid command\n"); 1438 break; 1439 case AMDGPU_RLCG_VFI_STAT__INV_ADDR: 1440 dev_err(adev->dev, "VFi invalid address\n"); 1441 break; 1442 case AMDGPU_RLCG_VFI_STAT__ERR: 1443 dev_err(adev->dev, "VFi unknown error\n"); 1444 break; 1445 default: 1446 dev_err(adev->dev, "VFi unknown status code\n"); 1447 break; 1448 } 1449 1450 spin_unlock_irqrestore(&adev->virt.rlcg_reg_lock, flags); 1451 1452 if (is_err) 1453 dev_err(adev->dev, "VFi: [grbm_cntl=0x%x grbm_idx=0x%x] addr=0x%x (byte addr 0x%x), data=0x%x, cmd=0x%x\n", 1454 grbm_cntl_data, grbm_idx_data, 1455 addr, addr * 4, data, cmd); 1456 else 1457 dev_dbg(adev->dev, "VFi: [grbm_cntl=0x%x grbm_idx=0x%x] addr=0x%x (byte addr 0x%x), data=0x%x, cmd=0x%x\n", 1458 grbm_cntl_data, grbm_idx_data, 1459 addr, addr * 4, data, cmd); 1460 1461 return data; 1462 } 1463 1464 u32 amdgpu_virt_rlcg_reg_rw(struct amdgpu_device *adev, u32 offset, u32 v, u32 flag, u32 xcc_id) 1465 { 1466 struct amdgpu_rlcg_reg_access_ctrl *reg_access_ctrl; 1467 uint32_t timeout = 50000; 1468 uint32_t i, tmp; 1469 uint32_t ret = 0; 1470 void *scratch_reg0; 1471 void *scratch_reg1; 1472 void *scratch_reg2; 1473 void *scratch_reg3; 1474 void *spare_int; 1475 unsigned long flags; 1476 1477 if (amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(12, 1, 0)) 1478 return amdgpu_virt_rlcg_vfi_reg_rw(adev, offset, v, flag, xcc_id); 1479 1480 if (!adev->gfx.rlc.rlcg_reg_access_supported) { 1481 dev_err(adev->dev, 1482 "indirect registers access through rlcg is not available\n"); 1483 return 0; 1484 } 1485 1486 if (adev->gfx.xcc_mask && (((1 << xcc_id) & adev->gfx.xcc_mask) == 0)) { 1487 dev_err(adev->dev, "invalid xcc\n"); 1488 return 0; 1489 } 1490 1491 if (amdgpu_device_skip_hw_access(adev)) 1492 return 0; 1493 1494 reg_access_ctrl = &adev->gfx.rlc.reg_access_ctrl[xcc_id]; 1495 scratch_reg0 = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->scratch_reg0; 1496 scratch_reg1 = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->scratch_reg1; 1497 scratch_reg2 = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->scratch_reg2; 1498 scratch_reg3 = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->scratch_reg3; 1499 1500 spin_lock_irqsave(&adev->virt.rlcg_reg_lock, flags); 1501 1502 if (reg_access_ctrl->spare_int) 1503 spare_int = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->spare_int; 1504 1505 if (offset == reg_access_ctrl->grbm_cntl) { 1506 /* if the target reg offset is grbm_cntl, write to scratch_reg2 */ 1507 writel(v, scratch_reg2); 1508 if (flag == AMDGPU_RLCG_GC_WRITE_LEGACY) 1509 writel(v, ((void __iomem *)adev->rmmio) + (offset * 4)); 1510 } else if (offset == reg_access_ctrl->grbm_idx) { 1511 /* if the target reg offset is grbm_idx, write to scratch_reg3 */ 1512 writel(v, scratch_reg3); 1513 if (flag == AMDGPU_RLCG_GC_WRITE_LEGACY) 1514 writel(v, ((void __iomem *)adev->rmmio) + (offset * 4)); 1515 } else { 1516 /* 1517 * SCRATCH_REG0 = read/write value 1518 * SCRATCH_REG1[30:28] = command 1519 * SCRATCH_REG1[19:0] = address in dword 1520 * SCRATCH_REG1[27:24] = Error reporting 1521 */ 1522 writel(v, scratch_reg0); 1523 writel((offset | flag), scratch_reg1); 1524 if (reg_access_ctrl->spare_int) 1525 writel(1, spare_int); 1526 1527 for (i = 0; i < timeout; i++) { 1528 tmp = readl(scratch_reg1); 1529 if (!(tmp & AMDGPU_RLCG_SCRATCH1_ADDRESS_MASK)) 1530 break; 1531 udelay(10); 1532 } 1533 1534 tmp = readl(scratch_reg1); 1535 if (i >= timeout || (tmp & AMDGPU_RLCG_SCRATCH1_ERROR_MASK) != 0) { 1536 if (amdgpu_sriov_rlcg_error_report_enabled(adev)) { 1537 if (tmp & AMDGPU_RLCG_VFGATE_DISABLED) { 1538 dev_err(adev->dev, 1539 "vfgate is disabled, rlcg failed to program reg: 0x%05x\n", offset); 1540 } else if (tmp & AMDGPU_RLCG_WRONG_OPERATION_TYPE) { 1541 dev_err(adev->dev, 1542 "wrong operation type, rlcg failed to program reg: 0x%05x\n", offset); 1543 } else if (tmp & AMDGPU_RLCG_REG_NOT_IN_RANGE) { 1544 dev_err(adev->dev, 1545 "register is not in range, rlcg failed to program reg: 0x%05x\n", offset); 1546 } else { 1547 dev_err(adev->dev, 1548 "unknown error type, rlcg failed to program reg: 0x%05x\n", offset); 1549 } 1550 } else { 1551 dev_err(adev->dev, 1552 "timeout: rlcg faled to program reg: 0x%05x\n", offset); 1553 } 1554 } 1555 } 1556 1557 ret = readl(scratch_reg0); 1558 1559 spin_unlock_irqrestore(&adev->virt.rlcg_reg_lock, flags); 1560 1561 return ret; 1562 } 1563 1564 void amdgpu_sriov_wreg(struct amdgpu_device *adev, 1565 u32 offset, u32 value, 1566 u32 acc_flags, u32 hwip, u32 xcc_id) 1567 { 1568 u32 rlcg_flag; 1569 1570 if (amdgpu_device_skip_hw_access(adev)) 1571 return; 1572 1573 if (!amdgpu_sriov_runtime(adev) && 1574 amdgpu_virt_get_rlcg_reg_access_flag(adev, acc_flags, hwip, true, &rlcg_flag)) { 1575 amdgpu_virt_rlcg_reg_rw(adev, offset, value, rlcg_flag, xcc_id); 1576 return; 1577 } 1578 1579 if (acc_flags & AMDGPU_REGS_NO_KIQ) 1580 WREG32_NO_KIQ(offset, value); 1581 else 1582 WREG32(offset, value); 1583 } 1584 1585 u32 amdgpu_sriov_rreg(struct amdgpu_device *adev, 1586 u32 offset, u32 acc_flags, u32 hwip, u32 xcc_id) 1587 { 1588 u32 rlcg_flag; 1589 1590 if (amdgpu_device_skip_hw_access(adev)) 1591 return 0; 1592 1593 if (!amdgpu_sriov_runtime(adev) && 1594 amdgpu_virt_get_rlcg_reg_access_flag(adev, acc_flags, hwip, false, &rlcg_flag)) 1595 return amdgpu_virt_rlcg_reg_rw(adev, offset, 0, rlcg_flag, xcc_id); 1596 1597 if (acc_flags & AMDGPU_REGS_NO_KIQ) 1598 return RREG32_NO_KIQ(offset); 1599 else 1600 return RREG32(offset); 1601 } 1602 1603 bool amdgpu_sriov_xnack_support(struct amdgpu_device *adev) 1604 { 1605 bool xnack_mode = true; 1606 1607 if (amdgpu_sriov_vf(adev) && 1608 amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 2)) 1609 xnack_mode = false; 1610 1611 return xnack_mode; 1612 } 1613 1614 bool amdgpu_virt_get_ras_capability(struct amdgpu_device *adev) 1615 { 1616 struct amdgpu_ras *con = amdgpu_ras_get_context(adev); 1617 1618 if (!amdgpu_sriov_ras_caps_en(adev)) 1619 return false; 1620 1621 if (adev->virt.ras_en_caps.bits.block_umc) 1622 adev->ras_hw_enabled |= BIT(AMDGPU_RAS_BLOCK__UMC); 1623 if (adev->virt.ras_en_caps.bits.block_sdma) 1624 adev->ras_hw_enabled |= BIT(AMDGPU_RAS_BLOCK__SDMA); 1625 if (adev->virt.ras_en_caps.bits.block_gfx) 1626 adev->ras_hw_enabled |= BIT(AMDGPU_RAS_BLOCK__GFX); 1627 if (adev->virt.ras_en_caps.bits.block_mmhub) 1628 adev->ras_hw_enabled |= BIT(AMDGPU_RAS_BLOCK__MMHUB); 1629 if (adev->virt.ras_en_caps.bits.block_athub) 1630 adev->ras_hw_enabled |= BIT(AMDGPU_RAS_BLOCK__ATHUB); 1631 if (adev->virt.ras_en_caps.bits.block_pcie_bif) 1632 adev->ras_hw_enabled |= BIT(AMDGPU_RAS_BLOCK__PCIE_BIF); 1633 if (adev->virt.ras_en_caps.bits.block_hdp) 1634 adev->ras_hw_enabled |= BIT(AMDGPU_RAS_BLOCK__HDP); 1635 if (adev->virt.ras_en_caps.bits.block_xgmi_wafl) 1636 adev->ras_hw_enabled |= BIT(AMDGPU_RAS_BLOCK__XGMI_WAFL); 1637 if (adev->virt.ras_en_caps.bits.block_df) 1638 adev->ras_hw_enabled |= BIT(AMDGPU_RAS_BLOCK__DF); 1639 if (adev->virt.ras_en_caps.bits.block_smn) 1640 adev->ras_hw_enabled |= BIT(AMDGPU_RAS_BLOCK__SMN); 1641 if (adev->virt.ras_en_caps.bits.block_sem) 1642 adev->ras_hw_enabled |= BIT(AMDGPU_RAS_BLOCK__SEM); 1643 if (adev->virt.ras_en_caps.bits.block_mp0) 1644 adev->ras_hw_enabled |= BIT(AMDGPU_RAS_BLOCK__MP0); 1645 if (adev->virt.ras_en_caps.bits.block_mp1) 1646 adev->ras_hw_enabled |= BIT(AMDGPU_RAS_BLOCK__MP1); 1647 if (adev->virt.ras_en_caps.bits.block_fuse) 1648 adev->ras_hw_enabled |= BIT(AMDGPU_RAS_BLOCK__FUSE); 1649 if (adev->virt.ras_en_caps.bits.block_mca) 1650 adev->ras_hw_enabled |= BIT(AMDGPU_RAS_BLOCK__MCA); 1651 if (adev->virt.ras_en_caps.bits.block_vcn) 1652 adev->ras_hw_enabled |= BIT(AMDGPU_RAS_BLOCK__VCN); 1653 if (adev->virt.ras_en_caps.bits.block_jpeg) 1654 adev->ras_hw_enabled |= BIT(AMDGPU_RAS_BLOCK__JPEG); 1655 if (adev->virt.ras_en_caps.bits.block_ih) 1656 adev->ras_hw_enabled |= BIT(AMDGPU_RAS_BLOCK__IH); 1657 if (adev->virt.ras_en_caps.bits.block_mpio) 1658 adev->ras_hw_enabled |= BIT(AMDGPU_RAS_BLOCK__MPIO); 1659 1660 if (adev->virt.ras_en_caps.bits.poison_propogation_mode) 1661 con->poison_supported = true; /* Poison is handled by host */ 1662 1663 if (adev->virt.ras_en_caps.bits.uniras_supported) 1664 amdgpu_virt_ras_set_remote_uniras(adev, true); 1665 1666 return true; 1667 } 1668 1669 static inline enum amd_sriov_ras_telemetry_gpu_block 1670 amdgpu_ras_block_to_sriov(struct amdgpu_device *adev, enum amdgpu_ras_block block) { 1671 switch (block) { 1672 case AMDGPU_RAS_BLOCK__UMC: 1673 return RAS_TELEMETRY_GPU_BLOCK_UMC; 1674 case AMDGPU_RAS_BLOCK__SDMA: 1675 return RAS_TELEMETRY_GPU_BLOCK_SDMA; 1676 case AMDGPU_RAS_BLOCK__GFX: 1677 return RAS_TELEMETRY_GPU_BLOCK_GFX; 1678 case AMDGPU_RAS_BLOCK__MMHUB: 1679 return RAS_TELEMETRY_GPU_BLOCK_MMHUB; 1680 case AMDGPU_RAS_BLOCK__ATHUB: 1681 return RAS_TELEMETRY_GPU_BLOCK_ATHUB; 1682 case AMDGPU_RAS_BLOCK__PCIE_BIF: 1683 return RAS_TELEMETRY_GPU_BLOCK_PCIE_BIF; 1684 case AMDGPU_RAS_BLOCK__HDP: 1685 return RAS_TELEMETRY_GPU_BLOCK_HDP; 1686 case AMDGPU_RAS_BLOCK__XGMI_WAFL: 1687 return RAS_TELEMETRY_GPU_BLOCK_XGMI_WAFL; 1688 case AMDGPU_RAS_BLOCK__DF: 1689 return RAS_TELEMETRY_GPU_BLOCK_DF; 1690 case AMDGPU_RAS_BLOCK__SMN: 1691 return RAS_TELEMETRY_GPU_BLOCK_SMN; 1692 case AMDGPU_RAS_BLOCK__SEM: 1693 return RAS_TELEMETRY_GPU_BLOCK_SEM; 1694 case AMDGPU_RAS_BLOCK__MP0: 1695 return RAS_TELEMETRY_GPU_BLOCK_MP0; 1696 case AMDGPU_RAS_BLOCK__MP1: 1697 return RAS_TELEMETRY_GPU_BLOCK_MP1; 1698 case AMDGPU_RAS_BLOCK__FUSE: 1699 return RAS_TELEMETRY_GPU_BLOCK_FUSE; 1700 case AMDGPU_RAS_BLOCK__MCA: 1701 return RAS_TELEMETRY_GPU_BLOCK_MCA; 1702 case AMDGPU_RAS_BLOCK__VCN: 1703 return RAS_TELEMETRY_GPU_BLOCK_VCN; 1704 case AMDGPU_RAS_BLOCK__JPEG: 1705 return RAS_TELEMETRY_GPU_BLOCK_JPEG; 1706 case AMDGPU_RAS_BLOCK__IH: 1707 return RAS_TELEMETRY_GPU_BLOCK_IH; 1708 case AMDGPU_RAS_BLOCK__MPIO: 1709 return RAS_TELEMETRY_GPU_BLOCK_MPIO; 1710 default: 1711 dev_warn(adev->dev, "Unsupported SRIOV RAS telemetry block 0x%x\n", 1712 block); 1713 return RAS_TELEMETRY_GPU_BLOCK_COUNT; 1714 } 1715 } 1716 1717 static int amdgpu_virt_cache_host_error_counts(struct amdgpu_device *adev, 1718 struct amdsriov_ras_telemetry *host_telemetry) 1719 { 1720 struct amd_sriov_ras_telemetry_error_count *tmp = NULL; 1721 uint32_t checksum, used_size; 1722 1723 checksum = host_telemetry->header.checksum; 1724 used_size = host_telemetry->header.used_size; 1725 1726 if (used_size > (AMD_SRIOV_MSG_RAS_TELEMETRY_SIZE_KB_V1 << 10)) 1727 return 0; 1728 1729 tmp = kmemdup(&host_telemetry->body.error_count, used_size, GFP_KERNEL); 1730 if (!tmp) 1731 return -ENOMEM; 1732 1733 if (checksum != amd_sriov_msg_checksum(tmp, used_size, 0, 0)) 1734 goto out; 1735 1736 memcpy(&adev->virt.count_cache, tmp, 1737 min(used_size, sizeof(adev->virt.count_cache))); 1738 out: 1739 kfree(tmp); 1740 1741 return 0; 1742 } 1743 1744 static int amdgpu_virt_req_ras_err_count_internal(struct amdgpu_device *adev, bool force_update) 1745 { 1746 struct amdgpu_virt *virt = &adev->virt; 1747 1748 if (!virt->ops || !virt->ops->req_ras_err_count) 1749 return -EOPNOTSUPP; 1750 1751 /* Host allows 15 ras telemetry requests per 60 seconds. Afterwhich, the Host 1752 * will ignore incoming guest messages. Ratelimit the guest messages to 1753 * prevent guest self DOS. 1754 */ 1755 if (__ratelimit(&virt->ras.ras_error_cnt_rs) || force_update) { 1756 mutex_lock(&virt->ras.ras_telemetry_mutex); 1757 if (!virt->ops->req_ras_err_count(adev)) 1758 amdgpu_virt_cache_host_error_counts(adev, 1759 virt->fw_reserve.ras_telemetry); 1760 mutex_unlock(&virt->ras.ras_telemetry_mutex); 1761 } 1762 1763 return 0; 1764 } 1765 1766 /* Bypass ACA interface and query ECC counts directly from host */ 1767 int amdgpu_virt_req_ras_err_count(struct amdgpu_device *adev, enum amdgpu_ras_block block, 1768 struct ras_err_data *err_data) 1769 { 1770 enum amd_sriov_ras_telemetry_gpu_block sriov_block; 1771 1772 sriov_block = amdgpu_ras_block_to_sriov(adev, block); 1773 1774 if (sriov_block >= RAS_TELEMETRY_GPU_BLOCK_COUNT || 1775 !amdgpu_sriov_ras_telemetry_block_en(adev, sriov_block)) 1776 return -EOPNOTSUPP; 1777 1778 /* Host Access may be lost during reset, just return last cached data. */ 1779 if (down_read_trylock(&adev->reset_domain->sem)) { 1780 amdgpu_virt_req_ras_err_count_internal(adev, false); 1781 up_read(&adev->reset_domain->sem); 1782 } 1783 1784 err_data->ue_count = adev->virt.count_cache.block[sriov_block].ue_count; 1785 err_data->ce_count = adev->virt.count_cache.block[sriov_block].ce_count; 1786 err_data->de_count = adev->virt.count_cache.block[sriov_block].de_count; 1787 1788 return 0; 1789 } 1790 1791 static int 1792 amdgpu_virt_write_cpers_to_ring(struct amdgpu_device *adev, 1793 struct amdsriov_ras_telemetry *host_telemetry, 1794 u32 *more) 1795 { 1796 struct amd_sriov_ras_cper_dump *cper_dump = NULL; 1797 struct cper_hdr *entry = NULL; 1798 struct amdgpu_ring *ring = &adev->cper.ring_buf; 1799 uint32_t checksum, used_size, i; 1800 int ret = 0; 1801 1802 checksum = host_telemetry->header.checksum; 1803 used_size = host_telemetry->header.used_size; 1804 1805 if (used_size > (AMD_SRIOV_MSG_RAS_TELEMETRY_SIZE_KB_V1 << 10)) 1806 return -EINVAL; 1807 1808 cper_dump = kmemdup(&host_telemetry->body.cper_dump, used_size, GFP_KERNEL); 1809 if (!cper_dump) 1810 return -ENOMEM; 1811 1812 if (checksum != amd_sriov_msg_checksum(cper_dump, used_size, 0, 0)) { 1813 ret = -EINVAL; 1814 goto out; 1815 } 1816 1817 *more = cper_dump->more; 1818 1819 if (cper_dump->wptr < adev->virt.ras.cper_rptr) { 1820 dev_warn( 1821 adev->dev, 1822 "guest specified rptr that was too high! guest rptr: 0x%llx, host rptr: 0x%llx\n", 1823 adev->virt.ras.cper_rptr, cper_dump->wptr); 1824 1825 adev->virt.ras.cper_rptr = cper_dump->wptr; 1826 goto out; 1827 } 1828 1829 entry = (struct cper_hdr *)&cper_dump->buf[0]; 1830 1831 for (i = 0; i < cper_dump->count; i++) { 1832 amdgpu_cper_ring_write(ring, entry, entry->record_length); 1833 entry = (struct cper_hdr *)((char *)entry + 1834 entry->record_length); 1835 } 1836 1837 if (cper_dump->overflow_count) 1838 dev_warn(adev->dev, 1839 "host reported CPER overflow of 0x%llx entries!\n", 1840 cper_dump->overflow_count); 1841 1842 adev->virt.ras.cper_rptr = cper_dump->wptr; 1843 out: 1844 kfree(cper_dump); 1845 1846 return ret; 1847 } 1848 1849 static int amdgpu_virt_req_ras_cper_dump_internal(struct amdgpu_device *adev) 1850 { 1851 struct amdgpu_virt *virt = &adev->virt; 1852 int ret = 0; 1853 uint32_t more = 0; 1854 1855 if (!virt->ops || !virt->ops->req_ras_cper_dump) 1856 return -EOPNOTSUPP; 1857 1858 do { 1859 if (!virt->ops->req_ras_cper_dump(adev, virt->ras.cper_rptr)) 1860 ret = amdgpu_virt_write_cpers_to_ring( 1861 adev, virt->fw_reserve.ras_telemetry, &more); 1862 else 1863 ret = 0; 1864 } while (more && !ret); 1865 1866 return ret; 1867 } 1868 1869 int amdgpu_virt_req_ras_cper_dump(struct amdgpu_device *adev, bool force_update) 1870 { 1871 struct amdgpu_virt *virt = &adev->virt; 1872 int ret = 0; 1873 1874 if (!amdgpu_sriov_ras_cper_en(adev)) 1875 return -EOPNOTSUPP; 1876 1877 if ((__ratelimit(&virt->ras.ras_cper_dump_rs) || force_update) && 1878 down_read_trylock(&adev->reset_domain->sem)) { 1879 mutex_lock(&virt->ras.ras_telemetry_mutex); 1880 ret = amdgpu_virt_req_ras_cper_dump_internal(adev); 1881 mutex_unlock(&virt->ras.ras_telemetry_mutex); 1882 up_read(&adev->reset_domain->sem); 1883 } 1884 1885 return ret; 1886 } 1887 1888 int amdgpu_virt_ras_telemetry_post_reset(struct amdgpu_device *adev) 1889 { 1890 unsigned long ue_count, ce_count; 1891 1892 if (amdgpu_sriov_ras_telemetry_en(adev)) { 1893 amdgpu_virt_req_ras_err_count_internal(adev, true); 1894 amdgpu_ras_query_error_count(adev, &ce_count, &ue_count, NULL); 1895 } 1896 1897 return 0; 1898 } 1899 1900 bool amdgpu_virt_ras_telemetry_block_en(struct amdgpu_device *adev, 1901 enum amdgpu_ras_block block) 1902 { 1903 enum amd_sriov_ras_telemetry_gpu_block sriov_block; 1904 1905 sriov_block = amdgpu_ras_block_to_sriov(adev, block); 1906 1907 if (sriov_block >= RAS_TELEMETRY_GPU_BLOCK_COUNT || 1908 !amdgpu_sriov_ras_telemetry_block_en(adev, sriov_block)) 1909 return false; 1910 1911 return true; 1912 } 1913 1914 /* 1915 * amdgpu_virt_request_bad_pages() - request bad pages 1916 * @adev: amdgpu device. 1917 * Send command to GPU hypervisor to write new bad pages into the shared PF2VF region 1918 */ 1919 void amdgpu_virt_request_bad_pages(struct amdgpu_device *adev) 1920 { 1921 struct amdgpu_virt *virt = &adev->virt; 1922 1923 if (virt->ops && virt->ops->req_bad_pages) 1924 virt->ops->req_bad_pages(adev); 1925 } 1926 1927 static int amdgpu_virt_cache_chk_criti_hit(struct amdgpu_device *adev, 1928 struct amdsriov_ras_telemetry *host_telemetry, 1929 bool *hit) 1930 { 1931 struct amd_sriov_ras_chk_criti *tmp = NULL; 1932 uint32_t checksum, used_size; 1933 1934 checksum = host_telemetry->header.checksum; 1935 used_size = host_telemetry->header.used_size; 1936 1937 if (used_size > (AMD_SRIOV_MSG_RAS_TELEMETRY_SIZE_KB_V1 << 10)) 1938 return 0; 1939 1940 tmp = kmemdup(&host_telemetry->body.chk_criti, used_size, GFP_KERNEL); 1941 if (!tmp) 1942 return -ENOMEM; 1943 1944 if (checksum != amd_sriov_msg_checksum(tmp, used_size, 0, 0)) 1945 goto out; 1946 1947 if (hit) 1948 *hit = tmp->hit ? true : false; 1949 1950 out: 1951 kfree(tmp); 1952 1953 return 0; 1954 } 1955 1956 int amdgpu_virt_check_vf_critical_region(struct amdgpu_device *adev, u64 addr, bool *hit) 1957 { 1958 struct amdgpu_virt *virt = &adev->virt; 1959 int r = -EPERM; 1960 1961 if (!virt->ops || !virt->ops->req_ras_chk_criti) 1962 return -EOPNOTSUPP; 1963 1964 /* Host allows 15 ras telemetry requests per 60 seconds. Afterwhich, the Host 1965 * will ignore incoming guest messages. Ratelimit the guest messages to 1966 * prevent guest self DOS. 1967 */ 1968 if (__ratelimit(&virt->ras.ras_chk_criti_rs)) { 1969 mutex_lock(&virt->ras.ras_telemetry_mutex); 1970 if (!virt->ops->req_ras_chk_criti(adev, addr)) 1971 r = amdgpu_virt_cache_chk_criti_hit( 1972 adev, virt->fw_reserve.ras_telemetry, hit); 1973 mutex_unlock(&virt->ras.ras_telemetry_mutex); 1974 } 1975 1976 return r; 1977 } 1978 1979 static int req_remote_ras_cmd(struct amdgpu_device *adev, 1980 u32 param1, u32 param2, u32 param3) 1981 { 1982 struct amdgpu_virt *virt = &adev->virt; 1983 1984 if (virt->ops && virt->ops->req_remote_ras_cmd) 1985 return virt->ops->req_remote_ras_cmd(adev, param1, param2, param3); 1986 return -ENOENT; 1987 } 1988 1989 int amdgpu_virt_send_remote_ras_cmd(struct amdgpu_device *adev, 1990 uint64_t buf, uint32_t buf_len) 1991 { 1992 uint64_t gpa = buf; 1993 int ret = -EIO; 1994 1995 if (down_read_trylock(&adev->reset_domain->sem)) { 1996 ret = req_remote_ras_cmd(adev, 1997 lower_32_bits(gpa), upper_32_bits(gpa), buf_len); 1998 up_read(&adev->reset_domain->sem); 1999 } 2000 2001 return ret; 2002 } 2003