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 #include <drm/drm_drv.h> 27 28 #include "amdgpu.h" 29 #include "amdgpu_ras.h" 30 #include "vi.h" 31 #include "soc15.h" 32 #include "nv.h" 33 34 #define POPULATE_UCODE_INFO(vf2pf_info, ucode, ver) \ 35 do { \ 36 vf2pf_info->ucode_info[ucode].id = ucode; \ 37 vf2pf_info->ucode_info[ucode].version = ver; \ 38 } while (0) 39 40 bool amdgpu_virt_mmio_blocked(struct amdgpu_device *adev) 41 { 42 /* By now all MMIO pages except mailbox are blocked */ 43 /* if blocking is enabled in hypervisor. Choose the */ 44 /* SCRATCH_REG0 to test. */ 45 return RREG32_NO_KIQ(0xc040) == 0xffffffff; 46 } 47 48 void amdgpu_virt_init_setting(struct amdgpu_device *adev) 49 { 50 struct drm_device *ddev = adev_to_drm(adev); 51 52 /* enable virtual display */ 53 if (adev->asic_type != CHIP_ALDEBARAN && 54 adev->asic_type != CHIP_ARCTURUS) { 55 if (adev->mode_info.num_crtc == 0) 56 adev->mode_info.num_crtc = 1; 57 adev->enable_virtual_display = true; 58 } 59 ddev->driver_features &= ~DRIVER_ATOMIC; 60 adev->cg_flags = 0; 61 adev->pg_flags = 0; 62 } 63 64 void amdgpu_virt_kiq_reg_write_reg_wait(struct amdgpu_device *adev, 65 uint32_t reg0, uint32_t reg1, 66 uint32_t ref, uint32_t mask) 67 { 68 struct amdgpu_kiq *kiq = &adev->gfx.kiq; 69 struct amdgpu_ring *ring = &kiq->ring; 70 signed long r, cnt = 0; 71 unsigned long flags; 72 uint32_t seq; 73 74 spin_lock_irqsave(&kiq->ring_lock, flags); 75 amdgpu_ring_alloc(ring, 32); 76 amdgpu_ring_emit_reg_write_reg_wait(ring, reg0, reg1, 77 ref, mask); 78 r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT); 79 if (r) 80 goto failed_undo; 81 82 amdgpu_ring_commit(ring); 83 spin_unlock_irqrestore(&kiq->ring_lock, flags); 84 85 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT); 86 87 /* don't wait anymore for IRQ context */ 88 if (r < 1 && in_interrupt()) 89 goto failed_kiq; 90 91 might_sleep(); 92 while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) { 93 94 msleep(MAX_KIQ_REG_BAILOUT_INTERVAL); 95 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT); 96 } 97 98 if (cnt > MAX_KIQ_REG_TRY) 99 goto failed_kiq; 100 101 return; 102 103 failed_undo: 104 amdgpu_ring_undo(ring); 105 spin_unlock_irqrestore(&kiq->ring_lock, flags); 106 failed_kiq: 107 dev_err(adev->dev, "failed to write reg %x wait reg %x\n", reg0, reg1); 108 } 109 110 /** 111 * amdgpu_virt_request_full_gpu() - request full gpu access 112 * @adev: amdgpu device. 113 * @init: is driver init time. 114 * When start to init/fini driver, first need to request full gpu access. 115 * Return: Zero if request success, otherwise will return error. 116 */ 117 int amdgpu_virt_request_full_gpu(struct amdgpu_device *adev, bool init) 118 { 119 struct amdgpu_virt *virt = &adev->virt; 120 int r; 121 122 if (virt->ops && virt->ops->req_full_gpu) { 123 r = virt->ops->req_full_gpu(adev, init); 124 if (r) 125 return r; 126 127 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME; 128 } 129 130 return 0; 131 } 132 133 /** 134 * amdgpu_virt_release_full_gpu() - release full gpu access 135 * @adev: amdgpu device. 136 * @init: is driver init time. 137 * When finishing driver init/fini, need to release full gpu access. 138 * Return: Zero if release success, otherwise will returen error. 139 */ 140 int amdgpu_virt_release_full_gpu(struct amdgpu_device *adev, bool init) 141 { 142 struct amdgpu_virt *virt = &adev->virt; 143 int r; 144 145 if (virt->ops && virt->ops->rel_full_gpu) { 146 r = virt->ops->rel_full_gpu(adev, init); 147 if (r) 148 return r; 149 150 adev->virt.caps |= AMDGPU_SRIOV_CAPS_RUNTIME; 151 } 152 return 0; 153 } 154 155 /** 156 * amdgpu_virt_reset_gpu() - reset gpu 157 * @adev: amdgpu device. 158 * Send reset command to GPU hypervisor to reset GPU that VM is using 159 * Return: Zero if reset success, otherwise will return error. 160 */ 161 int amdgpu_virt_reset_gpu(struct amdgpu_device *adev) 162 { 163 struct amdgpu_virt *virt = &adev->virt; 164 int r; 165 166 if (virt->ops && virt->ops->reset_gpu) { 167 r = virt->ops->reset_gpu(adev); 168 if (r) 169 return r; 170 171 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME; 172 } 173 174 return 0; 175 } 176 177 void amdgpu_virt_request_init_data(struct amdgpu_device *adev) 178 { 179 struct amdgpu_virt *virt = &adev->virt; 180 181 if (virt->ops && virt->ops->req_init_data) 182 virt->ops->req_init_data(adev); 183 184 if (adev->virt.req_init_data_ver > 0) 185 DRM_INFO("host supports REQ_INIT_DATA handshake\n"); 186 else 187 DRM_WARN("host doesn't support REQ_INIT_DATA handshake\n"); 188 } 189 190 /** 191 * amdgpu_virt_wait_reset() - wait for reset gpu completed 192 * @adev: amdgpu device. 193 * Wait for GPU reset completed. 194 * Return: Zero if reset success, otherwise will return error. 195 */ 196 int amdgpu_virt_wait_reset(struct amdgpu_device *adev) 197 { 198 struct amdgpu_virt *virt = &adev->virt; 199 200 if (!virt->ops || !virt->ops->wait_reset) 201 return -EINVAL; 202 203 return virt->ops->wait_reset(adev); 204 } 205 206 /** 207 * amdgpu_virt_alloc_mm_table() - alloc memory for mm table 208 * @adev: amdgpu device. 209 * MM table is used by UVD and VCE for its initialization 210 * Return: Zero if allocate success. 211 */ 212 int amdgpu_virt_alloc_mm_table(struct amdgpu_device *adev) 213 { 214 int r; 215 216 if (!amdgpu_sriov_vf(adev) || adev->virt.mm_table.gpu_addr) 217 return 0; 218 219 r = amdgpu_bo_create_kernel(adev, PAGE_SIZE, PAGE_SIZE, 220 AMDGPU_GEM_DOMAIN_VRAM, 221 &adev->virt.mm_table.bo, 222 &adev->virt.mm_table.gpu_addr, 223 (void *)&adev->virt.mm_table.cpu_addr); 224 if (r) { 225 DRM_ERROR("failed to alloc mm table and error = %d.\n", r); 226 return r; 227 } 228 229 memset((void *)adev->virt.mm_table.cpu_addr, 0, PAGE_SIZE); 230 DRM_INFO("MM table gpu addr = 0x%llx, cpu addr = %p.\n", 231 adev->virt.mm_table.gpu_addr, 232 adev->virt.mm_table.cpu_addr); 233 return 0; 234 } 235 236 /** 237 * amdgpu_virt_free_mm_table() - free mm table memory 238 * @adev: amdgpu device. 239 * Free MM table memory 240 */ 241 void amdgpu_virt_free_mm_table(struct amdgpu_device *adev) 242 { 243 if (!amdgpu_sriov_vf(adev) || !adev->virt.mm_table.gpu_addr) 244 return; 245 246 amdgpu_bo_free_kernel(&adev->virt.mm_table.bo, 247 &adev->virt.mm_table.gpu_addr, 248 (void *)&adev->virt.mm_table.cpu_addr); 249 adev->virt.mm_table.gpu_addr = 0; 250 } 251 252 253 unsigned int amd_sriov_msg_checksum(void *obj, 254 unsigned long obj_size, 255 unsigned int key, 256 unsigned int checksum) 257 { 258 unsigned int ret = key; 259 unsigned long i = 0; 260 unsigned char *pos; 261 262 pos = (char *)obj; 263 /* calculate checksum */ 264 for (i = 0; i < obj_size; ++i) 265 ret += *(pos + i); 266 /* minus the checksum itself */ 267 pos = (char *)&checksum; 268 for (i = 0; i < sizeof(checksum); ++i) 269 ret -= *(pos + i); 270 return ret; 271 } 272 273 static int amdgpu_virt_init_ras_err_handler_data(struct amdgpu_device *adev) 274 { 275 struct amdgpu_virt *virt = &adev->virt; 276 struct amdgpu_virt_ras_err_handler_data **data = &virt->virt_eh_data; 277 /* GPU will be marked bad on host if bp count more then 10, 278 * so alloc 512 is enough. 279 */ 280 unsigned int align_space = 512; 281 void *bps = NULL; 282 struct amdgpu_bo **bps_bo = NULL; 283 284 *data = kmalloc(sizeof(struct amdgpu_virt_ras_err_handler_data), GFP_KERNEL); 285 if (!*data) 286 goto data_failure; 287 288 bps = kmalloc_array(align_space, sizeof((*data)->bps), GFP_KERNEL); 289 if (!bps) 290 goto bps_failure; 291 292 bps_bo = kmalloc_array(align_space, sizeof((*data)->bps_bo), GFP_KERNEL); 293 if (!bps_bo) 294 goto bps_bo_failure; 295 296 (*data)->bps = bps; 297 (*data)->bps_bo = bps_bo; 298 (*data)->count = 0; 299 (*data)->last_reserved = 0; 300 301 virt->ras_init_done = true; 302 303 return 0; 304 305 bps_bo_failure: 306 kfree(bps); 307 bps_failure: 308 kfree(*data); 309 data_failure: 310 return -ENOMEM; 311 } 312 313 static void amdgpu_virt_ras_release_bp(struct amdgpu_device *adev) 314 { 315 struct amdgpu_virt *virt = &adev->virt; 316 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data; 317 struct amdgpu_bo *bo; 318 int i; 319 320 if (!data) 321 return; 322 323 for (i = data->last_reserved - 1; i >= 0; i--) { 324 bo = data->bps_bo[i]; 325 amdgpu_bo_free_kernel(&bo, NULL, NULL); 326 data->bps_bo[i] = bo; 327 data->last_reserved = i; 328 } 329 } 330 331 void amdgpu_virt_release_ras_err_handler_data(struct amdgpu_device *adev) 332 { 333 struct amdgpu_virt *virt = &adev->virt; 334 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data; 335 336 virt->ras_init_done = false; 337 338 if (!data) 339 return; 340 341 amdgpu_virt_ras_release_bp(adev); 342 343 kfree(data->bps); 344 kfree(data->bps_bo); 345 kfree(data); 346 virt->virt_eh_data = NULL; 347 } 348 349 static void amdgpu_virt_ras_add_bps(struct amdgpu_device *adev, 350 struct eeprom_table_record *bps, int pages) 351 { 352 struct amdgpu_virt *virt = &adev->virt; 353 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data; 354 355 if (!data) 356 return; 357 358 memcpy(&data->bps[data->count], bps, pages * sizeof(*data->bps)); 359 data->count += pages; 360 } 361 362 static void amdgpu_virt_ras_reserve_bps(struct amdgpu_device *adev) 363 { 364 struct amdgpu_virt *virt = &adev->virt; 365 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data; 366 struct amdgpu_bo *bo = NULL; 367 uint64_t bp; 368 int i; 369 370 if (!data) 371 return; 372 373 for (i = data->last_reserved; i < data->count; i++) { 374 bp = data->bps[i].retired_page; 375 376 /* There are two cases of reserve error should be ignored: 377 * 1) a ras bad page has been allocated (used by someone); 378 * 2) a ras bad page has been reserved (duplicate error injection 379 * for one page); 380 */ 381 if (amdgpu_bo_create_kernel_at(adev, bp << AMDGPU_GPU_PAGE_SHIFT, 382 AMDGPU_GPU_PAGE_SIZE, 383 AMDGPU_GEM_DOMAIN_VRAM, 384 &bo, NULL)) 385 DRM_DEBUG("RAS WARN: reserve vram for retired page %llx fail\n", bp); 386 387 data->bps_bo[i] = bo; 388 data->last_reserved = i + 1; 389 bo = NULL; 390 } 391 } 392 393 static bool amdgpu_virt_ras_check_bad_page(struct amdgpu_device *adev, 394 uint64_t retired_page) 395 { 396 struct amdgpu_virt *virt = &adev->virt; 397 struct amdgpu_virt_ras_err_handler_data *data = virt->virt_eh_data; 398 int i; 399 400 if (!data) 401 return true; 402 403 for (i = 0; i < data->count; i++) 404 if (retired_page == data->bps[i].retired_page) 405 return true; 406 407 return false; 408 } 409 410 static void amdgpu_virt_add_bad_page(struct amdgpu_device *adev, 411 uint64_t bp_block_offset, uint32_t bp_block_size) 412 { 413 struct eeprom_table_record bp; 414 uint64_t retired_page; 415 uint32_t bp_idx, bp_cnt; 416 417 if (bp_block_size) { 418 bp_cnt = bp_block_size / sizeof(uint64_t); 419 for (bp_idx = 0; bp_idx < bp_cnt; bp_idx++) { 420 retired_page = *(uint64_t *)(adev->mman.fw_vram_usage_va + 421 bp_block_offset + bp_idx * sizeof(uint64_t)); 422 bp.retired_page = retired_page; 423 424 if (amdgpu_virt_ras_check_bad_page(adev, retired_page)) 425 continue; 426 427 amdgpu_virt_ras_add_bps(adev, &bp, 1); 428 429 amdgpu_virt_ras_reserve_bps(adev); 430 } 431 } 432 } 433 434 static int amdgpu_virt_read_pf2vf_data(struct amdgpu_device *adev) 435 { 436 struct amd_sriov_msg_pf2vf_info_header *pf2vf_info = adev->virt.fw_reserve.p_pf2vf; 437 uint32_t checksum; 438 uint32_t checkval; 439 440 uint32_t i; 441 uint32_t tmp; 442 443 if (adev->virt.fw_reserve.p_pf2vf == NULL) 444 return -EINVAL; 445 446 if (pf2vf_info->size > 1024) { 447 DRM_ERROR("invalid pf2vf message size\n"); 448 return -EINVAL; 449 } 450 451 switch (pf2vf_info->version) { 452 case 1: 453 checksum = ((struct amdgim_pf2vf_info_v1 *)pf2vf_info)->checksum; 454 checkval = amd_sriov_msg_checksum( 455 adev->virt.fw_reserve.p_pf2vf, pf2vf_info->size, 456 adev->virt.fw_reserve.checksum_key, checksum); 457 if (checksum != checkval) { 458 DRM_ERROR("invalid pf2vf message\n"); 459 return -EINVAL; 460 } 461 462 adev->virt.gim_feature = 463 ((struct amdgim_pf2vf_info_v1 *)pf2vf_info)->feature_flags; 464 break; 465 case 2: 466 /* TODO: missing key, need to add it later */ 467 checksum = ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->checksum; 468 checkval = amd_sriov_msg_checksum( 469 adev->virt.fw_reserve.p_pf2vf, pf2vf_info->size, 470 0, checksum); 471 if (checksum != checkval) { 472 DRM_ERROR("invalid pf2vf message\n"); 473 return -EINVAL; 474 } 475 476 adev->virt.vf2pf_update_interval_ms = 477 ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->vf2pf_update_interval_ms; 478 adev->virt.gim_feature = 479 ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->feature_flags.all; 480 adev->virt.reg_access = 481 ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->reg_access_flags.all; 482 483 adev->virt.decode_max_dimension_pixels = 0; 484 adev->virt.decode_max_frame_pixels = 0; 485 adev->virt.encode_max_dimension_pixels = 0; 486 adev->virt.encode_max_frame_pixels = 0; 487 adev->virt.is_mm_bw_enabled = false; 488 for (i = 0; i < AMD_SRIOV_MSG_RESERVE_VCN_INST; i++) { 489 tmp = ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->mm_bw_management[i].decode_max_dimension_pixels; 490 adev->virt.decode_max_dimension_pixels = max(tmp, adev->virt.decode_max_dimension_pixels); 491 492 tmp = ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->mm_bw_management[i].decode_max_frame_pixels; 493 adev->virt.decode_max_frame_pixels = max(tmp, adev->virt.decode_max_frame_pixels); 494 495 tmp = ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->mm_bw_management[i].encode_max_dimension_pixels; 496 adev->virt.encode_max_dimension_pixels = max(tmp, adev->virt.encode_max_dimension_pixels); 497 498 tmp = ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->mm_bw_management[i].encode_max_frame_pixels; 499 adev->virt.encode_max_frame_pixels = max(tmp, adev->virt.encode_max_frame_pixels); 500 } 501 if((adev->virt.decode_max_dimension_pixels > 0) || (adev->virt.encode_max_dimension_pixels > 0)) 502 adev->virt.is_mm_bw_enabled = true; 503 504 adev->unique_id = 505 ((struct amd_sriov_msg_pf2vf_info *)pf2vf_info)->uuid; 506 break; 507 default: 508 DRM_ERROR("invalid pf2vf version\n"); 509 return -EINVAL; 510 } 511 512 /* correct too large or too little interval value */ 513 if (adev->virt.vf2pf_update_interval_ms < 200 || adev->virt.vf2pf_update_interval_ms > 10000) 514 adev->virt.vf2pf_update_interval_ms = 2000; 515 516 return 0; 517 } 518 519 static void amdgpu_virt_populate_vf2pf_ucode_info(struct amdgpu_device *adev) 520 { 521 struct amd_sriov_msg_vf2pf_info *vf2pf_info; 522 vf2pf_info = (struct amd_sriov_msg_vf2pf_info *) adev->virt.fw_reserve.p_vf2pf; 523 524 if (adev->virt.fw_reserve.p_vf2pf == NULL) 525 return; 526 527 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_VCE, adev->vce.fw_version); 528 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_UVD, adev->uvd.fw_version); 529 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_MC, adev->gmc.fw_version); 530 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_ME, adev->gfx.me_fw_version); 531 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_PFP, adev->gfx.pfp_fw_version); 532 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_CE, adev->gfx.ce_fw_version); 533 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_RLC, adev->gfx.rlc_fw_version); 534 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_RLC_SRLC, adev->gfx.rlc_srlc_fw_version); 535 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_RLC_SRLG, adev->gfx.rlc_srlg_fw_version); 536 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_RLC_SRLS, adev->gfx.rlc_srls_fw_version); 537 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_MEC, adev->gfx.mec_fw_version); 538 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_MEC2, adev->gfx.mec2_fw_version); 539 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_SOS, adev->psp.sos.fw_version); 540 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_ASD, 541 adev->psp.asd_context.bin_desc.fw_version); 542 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_TA_RAS, 543 adev->psp.ras_context.context.bin_desc.fw_version); 544 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_TA_XGMI, 545 adev->psp.xgmi_context.context.bin_desc.fw_version); 546 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_SMC, adev->pm.fw_version); 547 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_SDMA, adev->sdma.instance[0].fw_version); 548 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_SDMA2, adev->sdma.instance[1].fw_version); 549 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_VCN, adev->vcn.fw_version); 550 POPULATE_UCODE_INFO(vf2pf_info, AMD_SRIOV_UCODE_ID_DMCU, adev->dm.dmcu_fw_version); 551 } 552 553 static int amdgpu_virt_write_vf2pf_data(struct amdgpu_device *adev) 554 { 555 struct amd_sriov_msg_vf2pf_info *vf2pf_info; 556 557 vf2pf_info = (struct amd_sriov_msg_vf2pf_info *) adev->virt.fw_reserve.p_vf2pf; 558 559 if (adev->virt.fw_reserve.p_vf2pf == NULL) 560 return -EINVAL; 561 562 memset(vf2pf_info, 0, sizeof(struct amd_sriov_msg_vf2pf_info)); 563 564 vf2pf_info->header.size = sizeof(struct amd_sriov_msg_vf2pf_info); 565 vf2pf_info->header.version = AMD_SRIOV_MSG_FW_VRAM_VF2PF_VER; 566 567 #ifdef MODULE 568 if (THIS_MODULE->version != NULL) 569 strcpy(vf2pf_info->driver_version, THIS_MODULE->version); 570 else 571 #endif 572 strcpy(vf2pf_info->driver_version, "N/A"); 573 574 vf2pf_info->pf2vf_version_required = 0; // no requirement, guest understands all 575 vf2pf_info->driver_cert = 0; 576 vf2pf_info->os_info.all = 0; 577 578 vf2pf_info->fb_usage = 579 ttm_resource_manager_usage(&adev->mman.vram_mgr.manager) >> 20; 580 vf2pf_info->fb_vis_usage = 581 amdgpu_vram_mgr_vis_usage(&adev->mman.vram_mgr) >> 20; 582 vf2pf_info->fb_size = adev->gmc.real_vram_size >> 20; 583 vf2pf_info->fb_vis_size = adev->gmc.visible_vram_size >> 20; 584 585 amdgpu_virt_populate_vf2pf_ucode_info(adev); 586 587 /* TODO: read dynamic info */ 588 vf2pf_info->gfx_usage = 0; 589 vf2pf_info->compute_usage = 0; 590 vf2pf_info->encode_usage = 0; 591 vf2pf_info->decode_usage = 0; 592 593 vf2pf_info->dummy_page_addr = (uint64_t)adev->dummy_page_addr; 594 vf2pf_info->checksum = 595 amd_sriov_msg_checksum( 596 vf2pf_info, vf2pf_info->header.size, 0, 0); 597 598 return 0; 599 } 600 601 static void amdgpu_virt_update_vf2pf_work_item(struct work_struct *work) 602 { 603 struct amdgpu_device *adev = container_of(work, struct amdgpu_device, virt.vf2pf_work.work); 604 int ret; 605 606 ret = amdgpu_virt_read_pf2vf_data(adev); 607 if (ret) 608 goto out; 609 amdgpu_virt_write_vf2pf_data(adev); 610 611 out: 612 schedule_delayed_work(&(adev->virt.vf2pf_work), adev->virt.vf2pf_update_interval_ms); 613 } 614 615 void amdgpu_virt_fini_data_exchange(struct amdgpu_device *adev) 616 { 617 if (adev->virt.vf2pf_update_interval_ms != 0) { 618 DRM_INFO("clean up the vf2pf work item\n"); 619 cancel_delayed_work_sync(&adev->virt.vf2pf_work); 620 adev->virt.vf2pf_update_interval_ms = 0; 621 } 622 } 623 624 void amdgpu_virt_init_data_exchange(struct amdgpu_device *adev) 625 { 626 adev->virt.fw_reserve.p_pf2vf = NULL; 627 adev->virt.fw_reserve.p_vf2pf = NULL; 628 adev->virt.vf2pf_update_interval_ms = 0; 629 630 if (adev->mman.fw_vram_usage_va != NULL) { 631 /* go through this logic in ip_init and reset to init workqueue*/ 632 amdgpu_virt_exchange_data(adev); 633 634 INIT_DELAYED_WORK(&adev->virt.vf2pf_work, amdgpu_virt_update_vf2pf_work_item); 635 schedule_delayed_work(&(adev->virt.vf2pf_work), msecs_to_jiffies(adev->virt.vf2pf_update_interval_ms)); 636 } else if (adev->bios != NULL) { 637 /* got through this logic in early init stage to get necessary flags, e.g. rlcg_acc related*/ 638 adev->virt.fw_reserve.p_pf2vf = 639 (struct amd_sriov_msg_pf2vf_info_header *) 640 (adev->bios + (AMD_SRIOV_MSG_PF2VF_OFFSET_KB << 10)); 641 642 amdgpu_virt_read_pf2vf_data(adev); 643 } 644 } 645 646 647 void amdgpu_virt_exchange_data(struct amdgpu_device *adev) 648 { 649 uint64_t bp_block_offset = 0; 650 uint32_t bp_block_size = 0; 651 struct amd_sriov_msg_pf2vf_info *pf2vf_v2 = NULL; 652 653 if (adev->mman.fw_vram_usage_va != NULL) { 654 655 adev->virt.fw_reserve.p_pf2vf = 656 (struct amd_sriov_msg_pf2vf_info_header *) 657 (adev->mman.fw_vram_usage_va + (AMD_SRIOV_MSG_PF2VF_OFFSET_KB << 10)); 658 adev->virt.fw_reserve.p_vf2pf = 659 (struct amd_sriov_msg_vf2pf_info_header *) 660 (adev->mman.fw_vram_usage_va + (AMD_SRIOV_MSG_VF2PF_OFFSET_KB << 10)); 661 662 amdgpu_virt_read_pf2vf_data(adev); 663 amdgpu_virt_write_vf2pf_data(adev); 664 665 /* bad page handling for version 2 */ 666 if (adev->virt.fw_reserve.p_pf2vf->version == 2) { 667 pf2vf_v2 = (struct amd_sriov_msg_pf2vf_info *)adev->virt.fw_reserve.p_pf2vf; 668 669 bp_block_offset = ((uint64_t)pf2vf_v2->bp_block_offset_low & 0xFFFFFFFF) | 670 ((((uint64_t)pf2vf_v2->bp_block_offset_high) << 32) & 0xFFFFFFFF00000000); 671 bp_block_size = pf2vf_v2->bp_block_size; 672 673 if (bp_block_size && !adev->virt.ras_init_done) 674 amdgpu_virt_init_ras_err_handler_data(adev); 675 676 if (adev->virt.ras_init_done) 677 amdgpu_virt_add_bad_page(adev, bp_block_offset, bp_block_size); 678 } 679 } 680 } 681 682 683 void amdgpu_detect_virtualization(struct amdgpu_device *adev) 684 { 685 uint32_t reg; 686 687 switch (adev->asic_type) { 688 case CHIP_TONGA: 689 case CHIP_FIJI: 690 reg = RREG32(mmBIF_IOV_FUNC_IDENTIFIER); 691 break; 692 case CHIP_VEGA10: 693 case CHIP_VEGA20: 694 case CHIP_NAVI10: 695 case CHIP_NAVI12: 696 case CHIP_SIENNA_CICHLID: 697 case CHIP_ARCTURUS: 698 case CHIP_ALDEBARAN: 699 reg = RREG32(mmRCC_IOV_FUNC_IDENTIFIER); 700 break; 701 default: /* other chip doesn't support SRIOV */ 702 reg = 0; 703 break; 704 } 705 706 if (reg & 1) 707 adev->virt.caps |= AMDGPU_SRIOV_CAPS_IS_VF; 708 709 if (reg & 0x80000000) 710 adev->virt.caps |= AMDGPU_SRIOV_CAPS_ENABLE_IOV; 711 712 if (!reg) { 713 if (is_virtual_machine()) /* passthrough mode exclus sriov mod */ 714 adev->virt.caps |= AMDGPU_PASSTHROUGH_MODE; 715 } 716 717 /* we have the ability to check now */ 718 if (amdgpu_sriov_vf(adev)) { 719 switch (adev->asic_type) { 720 case CHIP_TONGA: 721 case CHIP_FIJI: 722 vi_set_virt_ops(adev); 723 break; 724 case CHIP_VEGA10: 725 soc15_set_virt_ops(adev); 726 /* send a dummy GPU_INIT_DATA request to host on vega10 */ 727 amdgpu_virt_request_init_data(adev); 728 break; 729 case CHIP_VEGA20: 730 case CHIP_ARCTURUS: 731 case CHIP_ALDEBARAN: 732 soc15_set_virt_ops(adev); 733 break; 734 case CHIP_NAVI10: 735 case CHIP_NAVI12: 736 case CHIP_SIENNA_CICHLID: 737 nv_set_virt_ops(adev); 738 /* try send GPU_INIT_DATA request to host */ 739 amdgpu_virt_request_init_data(adev); 740 break; 741 default: /* other chip doesn't support SRIOV */ 742 DRM_ERROR("Unknown asic type: %d!\n", adev->asic_type); 743 break; 744 } 745 } 746 } 747 748 static bool amdgpu_virt_access_debugfs_is_mmio(struct amdgpu_device *adev) 749 { 750 return amdgpu_sriov_is_debug(adev) ? true : false; 751 } 752 753 static bool amdgpu_virt_access_debugfs_is_kiq(struct amdgpu_device *adev) 754 { 755 return amdgpu_sriov_is_normal(adev) ? true : false; 756 } 757 758 int amdgpu_virt_enable_access_debugfs(struct amdgpu_device *adev) 759 { 760 if (!amdgpu_sriov_vf(adev) || 761 amdgpu_virt_access_debugfs_is_kiq(adev)) 762 return 0; 763 764 if (amdgpu_virt_access_debugfs_is_mmio(adev)) 765 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME; 766 else 767 return -EPERM; 768 769 return 0; 770 } 771 772 void amdgpu_virt_disable_access_debugfs(struct amdgpu_device *adev) 773 { 774 if (amdgpu_sriov_vf(adev)) 775 adev->virt.caps |= AMDGPU_SRIOV_CAPS_RUNTIME; 776 } 777 778 enum amdgpu_sriov_vf_mode amdgpu_virt_get_sriov_vf_mode(struct amdgpu_device *adev) 779 { 780 enum amdgpu_sriov_vf_mode mode; 781 782 if (amdgpu_sriov_vf(adev)) { 783 if (amdgpu_sriov_is_pp_one_vf(adev)) 784 mode = SRIOV_VF_MODE_ONE_VF; 785 else 786 mode = SRIOV_VF_MODE_MULTI_VF; 787 } else { 788 mode = SRIOV_VF_MODE_BARE_METAL; 789 } 790 791 return mode; 792 } 793 794 void amdgpu_virt_update_sriov_video_codec(struct amdgpu_device *adev, 795 struct amdgpu_video_codec_info *encode, uint32_t encode_array_size, 796 struct amdgpu_video_codec_info *decode, uint32_t decode_array_size) 797 { 798 uint32_t i; 799 800 if (!adev->virt.is_mm_bw_enabled) 801 return; 802 803 if (encode) { 804 for (i = 0; i < encode_array_size; i++) { 805 encode[i].max_width = adev->virt.encode_max_dimension_pixels; 806 encode[i].max_pixels_per_frame = adev->virt.encode_max_frame_pixels; 807 if (encode[i].max_width > 0) 808 encode[i].max_height = encode[i].max_pixels_per_frame / encode[i].max_width; 809 else 810 encode[i].max_height = 0; 811 } 812 } 813 814 if (decode) { 815 for (i = 0; i < decode_array_size; i++) { 816 decode[i].max_width = adev->virt.decode_max_dimension_pixels; 817 decode[i].max_pixels_per_frame = adev->virt.decode_max_frame_pixels; 818 if (decode[i].max_width > 0) 819 decode[i].max_height = decode[i].max_pixels_per_frame / decode[i].max_width; 820 else 821 decode[i].max_height = 0; 822 } 823 } 824 } 825 826 static bool amdgpu_virt_get_rlcg_reg_access_flag(struct amdgpu_device *adev, 827 u32 acc_flags, u32 hwip, 828 bool write, u32 *rlcg_flag) 829 { 830 bool ret = false; 831 832 switch (hwip) { 833 case GC_HWIP: 834 if (amdgpu_sriov_reg_indirect_gc(adev)) { 835 *rlcg_flag = 836 write ? AMDGPU_RLCG_GC_WRITE : AMDGPU_RLCG_GC_READ; 837 ret = true; 838 /* only in new version, AMDGPU_REGS_NO_KIQ and 839 * AMDGPU_REGS_RLC are enabled simultaneously */ 840 } else if ((acc_flags & AMDGPU_REGS_RLC) && 841 !(acc_flags & AMDGPU_REGS_NO_KIQ) && write) { 842 *rlcg_flag = AMDGPU_RLCG_GC_WRITE_LEGACY; 843 ret = true; 844 } 845 break; 846 case MMHUB_HWIP: 847 if (amdgpu_sriov_reg_indirect_mmhub(adev) && 848 (acc_flags & AMDGPU_REGS_RLC) && write) { 849 *rlcg_flag = AMDGPU_RLCG_MMHUB_WRITE; 850 ret = true; 851 } 852 break; 853 default: 854 break; 855 } 856 return ret; 857 } 858 859 static u32 amdgpu_virt_rlcg_reg_rw(struct amdgpu_device *adev, u32 offset, u32 v, u32 flag) 860 { 861 struct amdgpu_rlcg_reg_access_ctrl *reg_access_ctrl; 862 uint32_t timeout = 50000; 863 uint32_t i, tmp; 864 uint32_t ret = 0; 865 static void *scratch_reg0; 866 static void *scratch_reg1; 867 static void *scratch_reg2; 868 static void *scratch_reg3; 869 static void *spare_int; 870 871 if (!adev->gfx.rlc.rlcg_reg_access_supported) { 872 dev_err(adev->dev, 873 "indirect registers access through rlcg is not available\n"); 874 return 0; 875 } 876 877 reg_access_ctrl = &adev->gfx.rlc.reg_access_ctrl; 878 scratch_reg0 = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->scratch_reg0; 879 scratch_reg1 = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->scratch_reg1; 880 scratch_reg2 = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->scratch_reg2; 881 scratch_reg3 = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->scratch_reg3; 882 if (reg_access_ctrl->spare_int) 883 spare_int = (void __iomem *)adev->rmmio + 4 * reg_access_ctrl->spare_int; 884 885 if (offset == reg_access_ctrl->grbm_cntl) { 886 /* if the target reg offset is grbm_cntl, write to scratch_reg2 */ 887 writel(v, scratch_reg2); 888 writel(v, ((void __iomem *)adev->rmmio) + (offset * 4)); 889 } else if (offset == reg_access_ctrl->grbm_idx) { 890 /* if the target reg offset is grbm_idx, write to scratch_reg3 */ 891 writel(v, scratch_reg3); 892 writel(v, ((void __iomem *)adev->rmmio) + (offset * 4)); 893 } else { 894 /* 895 * SCRATCH_REG0 = read/write value 896 * SCRATCH_REG1[30:28] = command 897 * SCRATCH_REG1[19:0] = address in dword 898 * SCRATCH_REG1[26:24] = Error reporting 899 */ 900 writel(v, scratch_reg0); 901 writel((offset | flag), scratch_reg1); 902 if (reg_access_ctrl->spare_int) 903 writel(1, spare_int); 904 905 for (i = 0; i < timeout; i++) { 906 tmp = readl(scratch_reg1); 907 if (!(tmp & AMDGPU_RLCG_SCRATCH1_ADDRESS_MASK)) 908 break; 909 udelay(10); 910 } 911 912 if (i >= timeout) { 913 if (amdgpu_sriov_rlcg_error_report_enabled(adev)) { 914 if (tmp & AMDGPU_RLCG_VFGATE_DISABLED) { 915 dev_err(adev->dev, 916 "vfgate is disabled, rlcg failed to program reg: 0x%05x\n", offset); 917 } else if (tmp & AMDGPU_RLCG_WRONG_OPERATION_TYPE) { 918 dev_err(adev->dev, 919 "wrong operation type, rlcg failed to program reg: 0x%05x\n", offset); 920 } else if (tmp & AMDGPU_RLCG_REG_NOT_IN_RANGE) { 921 dev_err(adev->dev, 922 "regiser is not in range, rlcg failed to program reg: 0x%05x\n", offset); 923 } else { 924 dev_err(adev->dev, 925 "unknown error type, rlcg failed to program reg: 0x%05x\n", offset); 926 } 927 } else { 928 dev_err(adev->dev, 929 "timeout: rlcg faled to program reg: 0x%05x\n", offset); 930 } 931 } 932 } 933 934 ret = readl(scratch_reg0); 935 return ret; 936 } 937 938 void amdgpu_sriov_wreg(struct amdgpu_device *adev, 939 u32 offset, u32 value, 940 u32 acc_flags, u32 hwip) 941 { 942 u32 rlcg_flag; 943 944 if (!amdgpu_sriov_runtime(adev) && 945 amdgpu_virt_get_rlcg_reg_access_flag(adev, acc_flags, hwip, true, &rlcg_flag)) { 946 amdgpu_virt_rlcg_reg_rw(adev, offset, value, rlcg_flag); 947 return; 948 } 949 950 if (acc_flags & AMDGPU_REGS_NO_KIQ) 951 WREG32_NO_KIQ(offset, value); 952 else 953 WREG32(offset, value); 954 } 955 956 u32 amdgpu_sriov_rreg(struct amdgpu_device *adev, 957 u32 offset, u32 acc_flags, u32 hwip) 958 { 959 u32 rlcg_flag; 960 961 if (!amdgpu_sriov_runtime(adev) && 962 amdgpu_virt_get_rlcg_reg_access_flag(adev, acc_flags, hwip, false, &rlcg_flag)) 963 return amdgpu_virt_rlcg_reg_rw(adev, offset, 0, rlcg_flag); 964 965 if (acc_flags & AMDGPU_REGS_NO_KIQ) 966 return RREG32_NO_KIQ(offset); 967 else 968 return RREG32(offset); 969 } 970