1 /* 2 * Copyright 2022 Advanced Micro Devices, Inc. 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice shall be included in 12 * all copies or substantial portions of the Software. 13 * 14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 20 * OTHER DEALINGS IN THE SOFTWARE. 21 * 22 */ 23 #include "amdgpu.h" 24 #include "soc15.h" 25 26 #include "soc15_common.h" 27 #include "amdgpu_reg_state.h" 28 #include "amdgpu_xcp.h" 29 #include "gfx_v9_4_3.h" 30 #include "gfxhub_v1_2.h" 31 #include "sdma_v4_4_2.h" 32 33 #define XCP_INST_MASK(num_inst, xcp_id) \ 34 (num_inst ? GENMASK(num_inst - 1, 0) << (xcp_id * num_inst) : 0) 35 36 #define AMDGPU_XCP_OPS_KFD (1 << 0) 37 38 void aqua_vanjaram_doorbell_index_init(struct amdgpu_device *adev) 39 { 40 int i; 41 42 adev->doorbell_index.kiq = AMDGPU_DOORBELL_LAYOUT1_KIQ_START; 43 44 adev->doorbell_index.mec_ring0 = AMDGPU_DOORBELL_LAYOUT1_MEC_RING_START; 45 46 adev->doorbell_index.userqueue_start = AMDGPU_DOORBELL_LAYOUT1_USERQUEUE_START; 47 adev->doorbell_index.userqueue_end = AMDGPU_DOORBELL_LAYOUT1_USERQUEUE_END; 48 adev->doorbell_index.xcc_doorbell_range = AMDGPU_DOORBELL_LAYOUT1_XCC_RANGE; 49 50 adev->doorbell_index.sdma_doorbell_range = 20; 51 for (i = 0; i < adev->sdma.num_instances; i++) 52 adev->doorbell_index.sdma_engine[i] = 53 AMDGPU_DOORBELL_LAYOUT1_sDMA_ENGINE_START + 54 i * (adev->doorbell_index.sdma_doorbell_range >> 1); 55 56 adev->doorbell_index.ih = AMDGPU_DOORBELL_LAYOUT1_IH; 57 adev->doorbell_index.vcn.vcn_ring0_1 = AMDGPU_DOORBELL_LAYOUT1_VCN_START; 58 59 adev->doorbell_index.first_non_cp = AMDGPU_DOORBELL_LAYOUT1_FIRST_NON_CP; 60 adev->doorbell_index.last_non_cp = AMDGPU_DOORBELL_LAYOUT1_LAST_NON_CP; 61 62 adev->doorbell_index.max_assignment = AMDGPU_DOORBELL_LAYOUT1_MAX_ASSIGNMENT << 1; 63 } 64 65 static bool aqua_vanjaram_xcp_vcn_shared(struct amdgpu_device *adev) 66 { 67 return (adev->xcp_mgr->num_xcps > adev->vcn.num_vcn_inst); 68 } 69 70 static void aqua_vanjaram_set_xcp_id(struct amdgpu_device *adev, 71 uint32_t inst_idx, struct amdgpu_ring *ring) 72 { 73 int xcp_id; 74 enum AMDGPU_XCP_IP_BLOCK ip_blk; 75 uint32_t inst_mask; 76 77 ring->xcp_id = AMDGPU_XCP_NO_PARTITION; 78 if (adev->xcp_mgr->mode == AMDGPU_XCP_MODE_NONE) 79 return; 80 81 inst_mask = 1 << inst_idx; 82 83 switch (ring->funcs->type) { 84 case AMDGPU_HW_IP_GFX: 85 case AMDGPU_RING_TYPE_COMPUTE: 86 case AMDGPU_RING_TYPE_KIQ: 87 ip_blk = AMDGPU_XCP_GFX; 88 break; 89 case AMDGPU_RING_TYPE_SDMA: 90 ip_blk = AMDGPU_XCP_SDMA; 91 break; 92 case AMDGPU_RING_TYPE_VCN_ENC: 93 case AMDGPU_RING_TYPE_VCN_JPEG: 94 ip_blk = AMDGPU_XCP_VCN; 95 if (aqua_vanjaram_xcp_vcn_shared(adev)) 96 inst_mask = 1 << (inst_idx * 2); 97 break; 98 default: 99 DRM_ERROR("Not support ring type %d!", ring->funcs->type); 100 return; 101 } 102 103 for (xcp_id = 0; xcp_id < adev->xcp_mgr->num_xcps; xcp_id++) { 104 if (adev->xcp_mgr->xcp[xcp_id].ip[ip_blk].inst_mask & inst_mask) { 105 ring->xcp_id = xcp_id; 106 break; 107 } 108 } 109 } 110 111 static void aqua_vanjaram_xcp_gpu_sched_update( 112 struct amdgpu_device *adev, 113 struct amdgpu_ring *ring, 114 unsigned int sel_xcp_id) 115 { 116 unsigned int *num_gpu_sched; 117 118 num_gpu_sched = &adev->xcp_mgr->xcp[sel_xcp_id] 119 .gpu_sched[ring->funcs->type][ring->hw_prio].num_scheds; 120 adev->xcp_mgr->xcp[sel_xcp_id].gpu_sched[ring->funcs->type][ring->hw_prio] 121 .sched[(*num_gpu_sched)++] = &ring->sched; 122 DRM_DEBUG("%s :[%d] gpu_sched[%d][%d] = %d", ring->name, 123 sel_xcp_id, ring->funcs->type, 124 ring->hw_prio, *num_gpu_sched); 125 } 126 127 static int aqua_vanjaram_xcp_sched_list_update( 128 struct amdgpu_device *adev) 129 { 130 struct amdgpu_ring *ring; 131 int i; 132 133 for (i = 0; i < MAX_XCP; i++) { 134 atomic_set(&adev->xcp_mgr->xcp[i].ref_cnt, 0); 135 memset(adev->xcp_mgr->xcp[i].gpu_sched, 0, sizeof(adev->xcp_mgr->xcp->gpu_sched)); 136 } 137 138 if (adev->xcp_mgr->mode == AMDGPU_XCP_MODE_NONE) 139 return 0; 140 141 for (i = 0; i < AMDGPU_MAX_RINGS; i++) { 142 ring = adev->rings[i]; 143 if (!ring || !ring->sched.ready || ring->no_scheduler) 144 continue; 145 146 aqua_vanjaram_xcp_gpu_sched_update(adev, ring, ring->xcp_id); 147 148 /* VCN may be shared by two partitions under CPX MODE in certain 149 * configs. 150 */ 151 if ((ring->funcs->type == AMDGPU_RING_TYPE_VCN_ENC || 152 ring->funcs->type == AMDGPU_RING_TYPE_VCN_JPEG) && 153 aqua_vanjaram_xcp_vcn_shared(adev)) 154 aqua_vanjaram_xcp_gpu_sched_update(adev, ring, ring->xcp_id + 1); 155 } 156 157 return 0; 158 } 159 160 static int aqua_vanjaram_update_partition_sched_list(struct amdgpu_device *adev) 161 { 162 int i; 163 164 for (i = 0; i < adev->num_rings; i++) { 165 struct amdgpu_ring *ring = adev->rings[i]; 166 167 if (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE || 168 ring->funcs->type == AMDGPU_RING_TYPE_KIQ) 169 aqua_vanjaram_set_xcp_id(adev, ring->xcc_id, ring); 170 else 171 aqua_vanjaram_set_xcp_id(adev, ring->me, ring); 172 } 173 174 return aqua_vanjaram_xcp_sched_list_update(adev); 175 } 176 177 static int aqua_vanjaram_select_scheds( 178 struct amdgpu_device *adev, 179 u32 hw_ip, 180 u32 hw_prio, 181 struct amdgpu_fpriv *fpriv, 182 unsigned int *num_scheds, 183 struct drm_gpu_scheduler ***scheds) 184 { 185 u32 sel_xcp_id; 186 int i; 187 188 if (fpriv->xcp_id == AMDGPU_XCP_NO_PARTITION) { 189 u32 least_ref_cnt = ~0; 190 191 fpriv->xcp_id = 0; 192 for (i = 0; i < adev->xcp_mgr->num_xcps; i++) { 193 u32 total_ref_cnt; 194 195 total_ref_cnt = atomic_read(&adev->xcp_mgr->xcp[i].ref_cnt); 196 if (total_ref_cnt < least_ref_cnt) { 197 fpriv->xcp_id = i; 198 least_ref_cnt = total_ref_cnt; 199 } 200 } 201 } 202 sel_xcp_id = fpriv->xcp_id; 203 204 if (adev->xcp_mgr->xcp[sel_xcp_id].gpu_sched[hw_ip][hw_prio].num_scheds) { 205 *num_scheds = adev->xcp_mgr->xcp[fpriv->xcp_id].gpu_sched[hw_ip][hw_prio].num_scheds; 206 *scheds = adev->xcp_mgr->xcp[fpriv->xcp_id].gpu_sched[hw_ip][hw_prio].sched; 207 atomic_inc(&adev->xcp_mgr->xcp[sel_xcp_id].ref_cnt); 208 DRM_DEBUG("Selected partition #%d", sel_xcp_id); 209 } else { 210 DRM_ERROR("Failed to schedule partition #%d.", sel_xcp_id); 211 return -ENOENT; 212 } 213 214 return 0; 215 } 216 217 static int8_t aqua_vanjaram_logical_to_dev_inst(struct amdgpu_device *adev, 218 enum amd_hw_ip_block_type block, 219 int8_t inst) 220 { 221 int8_t dev_inst; 222 223 switch (block) { 224 case GC_HWIP: 225 case SDMA0_HWIP: 226 /* Both JPEG and VCN as JPEG is only alias of VCN */ 227 case VCN_HWIP: 228 dev_inst = adev->ip_map.dev_inst[block][inst]; 229 break; 230 default: 231 /* For rest of the IPs, no look up required. 232 * Assume 'logical instance == physical instance' for all configs. */ 233 dev_inst = inst; 234 break; 235 } 236 237 return dev_inst; 238 } 239 240 static uint32_t aqua_vanjaram_logical_to_dev_mask(struct amdgpu_device *adev, 241 enum amd_hw_ip_block_type block, 242 uint32_t mask) 243 { 244 uint32_t dev_mask = 0; 245 int8_t log_inst, dev_inst; 246 247 while (mask) { 248 log_inst = ffs(mask) - 1; 249 dev_inst = aqua_vanjaram_logical_to_dev_inst(adev, block, log_inst); 250 dev_mask |= (1 << dev_inst); 251 mask &= ~(1 << log_inst); 252 } 253 254 return dev_mask; 255 } 256 257 static void aqua_vanjaram_populate_ip_map(struct amdgpu_device *adev, 258 enum amd_hw_ip_block_type ip_block, 259 uint32_t inst_mask) 260 { 261 int l = 0, i; 262 263 while (inst_mask) { 264 i = ffs(inst_mask) - 1; 265 adev->ip_map.dev_inst[ip_block][l++] = i; 266 inst_mask &= ~(1 << i); 267 } 268 for (; l < HWIP_MAX_INSTANCE; l++) 269 adev->ip_map.dev_inst[ip_block][l] = -1; 270 } 271 272 void aqua_vanjaram_ip_map_init(struct amdgpu_device *adev) 273 { 274 u32 ip_map[][2] = { 275 { GC_HWIP, adev->gfx.xcc_mask }, 276 { SDMA0_HWIP, adev->sdma.sdma_mask }, 277 { VCN_HWIP, adev->vcn.inst_mask }, 278 }; 279 int i; 280 281 for (i = 0; i < ARRAY_SIZE(ip_map); ++i) 282 aqua_vanjaram_populate_ip_map(adev, ip_map[i][0], ip_map[i][1]); 283 284 adev->ip_map.logical_to_dev_inst = aqua_vanjaram_logical_to_dev_inst; 285 adev->ip_map.logical_to_dev_mask = aqua_vanjaram_logical_to_dev_mask; 286 } 287 288 /* Fixed pattern for smn addressing on different AIDs: 289 * bit[34]: indicate cross AID access 290 * bit[33:32]: indicate target AID id 291 * AID id range is 0 ~ 3 as maximum AID number is 4. 292 */ 293 u64 aqua_vanjaram_encode_ext_smn_addressing(int ext_id) 294 { 295 u64 ext_offset; 296 297 /* local routing and bit[34:32] will be zeros */ 298 if (ext_id == 0) 299 return 0; 300 301 /* Initiated from host, accessing to all non-zero aids are cross traffic */ 302 ext_offset = ((u64)(ext_id & 0x3) << 32) | (1ULL << 34); 303 304 return ext_offset; 305 } 306 307 static int aqua_vanjaram_query_partition_mode(struct amdgpu_xcp_mgr *xcp_mgr) 308 { 309 enum amdgpu_gfx_partition mode = AMDGPU_UNKNOWN_COMPUTE_PARTITION_MODE; 310 struct amdgpu_device *adev = xcp_mgr->adev; 311 312 if (adev->nbio.funcs->get_compute_partition_mode) 313 mode = adev->nbio.funcs->get_compute_partition_mode(adev); 314 315 return mode; 316 } 317 318 static int __aqua_vanjaram_get_xcc_per_xcp(struct amdgpu_xcp_mgr *xcp_mgr, int mode) 319 { 320 int num_xcc, num_xcc_per_xcp = 0; 321 322 num_xcc = NUM_XCC(xcp_mgr->adev->gfx.xcc_mask); 323 324 switch (mode) { 325 case AMDGPU_SPX_PARTITION_MODE: 326 num_xcc_per_xcp = num_xcc; 327 break; 328 case AMDGPU_DPX_PARTITION_MODE: 329 num_xcc_per_xcp = num_xcc / 2; 330 break; 331 case AMDGPU_TPX_PARTITION_MODE: 332 num_xcc_per_xcp = num_xcc / 3; 333 break; 334 case AMDGPU_QPX_PARTITION_MODE: 335 num_xcc_per_xcp = num_xcc / 4; 336 break; 337 case AMDGPU_CPX_PARTITION_MODE: 338 num_xcc_per_xcp = 1; 339 break; 340 } 341 342 return num_xcc_per_xcp; 343 } 344 345 static int __aqua_vanjaram_get_xcp_ip_info(struct amdgpu_xcp_mgr *xcp_mgr, int xcp_id, 346 enum AMDGPU_XCP_IP_BLOCK ip_id, 347 struct amdgpu_xcp_ip *ip) 348 { 349 struct amdgpu_device *adev = xcp_mgr->adev; 350 int num_xcc_xcp, num_sdma_xcp, num_vcn_xcp; 351 int num_sdma, num_vcn; 352 353 num_sdma = adev->sdma.num_instances; 354 num_vcn = adev->vcn.num_vcn_inst; 355 356 switch (xcp_mgr->mode) { 357 case AMDGPU_SPX_PARTITION_MODE: 358 num_sdma_xcp = num_sdma; 359 num_vcn_xcp = num_vcn; 360 break; 361 case AMDGPU_DPX_PARTITION_MODE: 362 num_sdma_xcp = num_sdma / 2; 363 num_vcn_xcp = num_vcn / 2; 364 break; 365 case AMDGPU_TPX_PARTITION_MODE: 366 num_sdma_xcp = num_sdma / 3; 367 num_vcn_xcp = num_vcn / 3; 368 break; 369 case AMDGPU_QPX_PARTITION_MODE: 370 num_sdma_xcp = num_sdma / 4; 371 num_vcn_xcp = num_vcn / 4; 372 break; 373 case AMDGPU_CPX_PARTITION_MODE: 374 num_sdma_xcp = 2; 375 num_vcn_xcp = num_vcn ? 1 : 0; 376 break; 377 default: 378 return -EINVAL; 379 } 380 381 num_xcc_xcp = adev->gfx.num_xcc_per_xcp; 382 383 switch (ip_id) { 384 case AMDGPU_XCP_GFXHUB: 385 ip->inst_mask = XCP_INST_MASK(num_xcc_xcp, xcp_id); 386 ip->ip_funcs = &gfxhub_v1_2_xcp_funcs; 387 break; 388 case AMDGPU_XCP_GFX: 389 ip->inst_mask = XCP_INST_MASK(num_xcc_xcp, xcp_id); 390 ip->ip_funcs = &gfx_v9_4_3_xcp_funcs; 391 break; 392 case AMDGPU_XCP_SDMA: 393 ip->inst_mask = XCP_INST_MASK(num_sdma_xcp, xcp_id); 394 ip->ip_funcs = &sdma_v4_4_2_xcp_funcs; 395 break; 396 case AMDGPU_XCP_VCN: 397 ip->inst_mask = XCP_INST_MASK(num_vcn_xcp, xcp_id); 398 /* TODO : Assign IP funcs */ 399 break; 400 default: 401 return -EINVAL; 402 } 403 404 ip->ip_id = ip_id; 405 406 return 0; 407 } 408 409 static enum amdgpu_gfx_partition 410 __aqua_vanjaram_get_auto_mode(struct amdgpu_xcp_mgr *xcp_mgr) 411 { 412 struct amdgpu_device *adev = xcp_mgr->adev; 413 int num_xcc; 414 415 num_xcc = NUM_XCC(xcp_mgr->adev->gfx.xcc_mask); 416 417 if (adev->gmc.num_mem_partitions == 1) 418 return AMDGPU_SPX_PARTITION_MODE; 419 420 if (adev->gmc.num_mem_partitions == num_xcc) 421 return AMDGPU_CPX_PARTITION_MODE; 422 423 if (adev->gmc.num_mem_partitions == num_xcc / 2) 424 return (adev->flags & AMD_IS_APU) ? AMDGPU_TPX_PARTITION_MODE : 425 AMDGPU_QPX_PARTITION_MODE; 426 427 if (adev->gmc.num_mem_partitions == 2 && !(adev->flags & AMD_IS_APU)) 428 return AMDGPU_DPX_PARTITION_MODE; 429 430 return AMDGPU_UNKNOWN_COMPUTE_PARTITION_MODE; 431 } 432 433 static bool __aqua_vanjaram_is_valid_mode(struct amdgpu_xcp_mgr *xcp_mgr, 434 enum amdgpu_gfx_partition mode) 435 { 436 struct amdgpu_device *adev = xcp_mgr->adev; 437 int num_xcc, num_xccs_per_xcp; 438 439 num_xcc = NUM_XCC(adev->gfx.xcc_mask); 440 switch (mode) { 441 case AMDGPU_SPX_PARTITION_MODE: 442 return adev->gmc.num_mem_partitions == 1 && num_xcc > 0; 443 case AMDGPU_DPX_PARTITION_MODE: 444 return adev->gmc.num_mem_partitions != 8 && (num_xcc % 4) == 0; 445 case AMDGPU_TPX_PARTITION_MODE: 446 return (adev->gmc.num_mem_partitions == 1 || 447 adev->gmc.num_mem_partitions == 3) && 448 ((num_xcc % 3) == 0); 449 case AMDGPU_QPX_PARTITION_MODE: 450 num_xccs_per_xcp = num_xcc / 4; 451 return (adev->gmc.num_mem_partitions == 1 || 452 adev->gmc.num_mem_partitions == 4) && 453 (num_xccs_per_xcp >= 2); 454 case AMDGPU_CPX_PARTITION_MODE: 455 return ((num_xcc > 1) && 456 (adev->gmc.num_mem_partitions == 1 || adev->gmc.num_mem_partitions == 4) && 457 (num_xcc % adev->gmc.num_mem_partitions) == 0); 458 default: 459 return false; 460 } 461 462 return false; 463 } 464 465 static int __aqua_vanjaram_pre_partition_switch(struct amdgpu_xcp_mgr *xcp_mgr, u32 flags) 466 { 467 /* TODO: 468 * Stop user queues and threads, and make sure GPU is empty of work. 469 */ 470 471 if (flags & AMDGPU_XCP_OPS_KFD) 472 amdgpu_amdkfd_device_fini_sw(xcp_mgr->adev); 473 474 return 0; 475 } 476 477 static int __aqua_vanjaram_post_partition_switch(struct amdgpu_xcp_mgr *xcp_mgr, u32 flags) 478 { 479 int ret = 0; 480 481 if (flags & AMDGPU_XCP_OPS_KFD) { 482 amdgpu_amdkfd_device_probe(xcp_mgr->adev); 483 amdgpu_amdkfd_device_init(xcp_mgr->adev); 484 /* If KFD init failed, return failure */ 485 if (!xcp_mgr->adev->kfd.init_complete) 486 ret = -EIO; 487 } 488 489 return ret; 490 } 491 492 static int aqua_vanjaram_switch_partition_mode(struct amdgpu_xcp_mgr *xcp_mgr, 493 int mode, int *num_xcps) 494 { 495 int num_xcc_per_xcp, num_xcc, ret; 496 struct amdgpu_device *adev; 497 u32 flags = 0; 498 499 adev = xcp_mgr->adev; 500 num_xcc = NUM_XCC(adev->gfx.xcc_mask); 501 502 if (mode == AMDGPU_AUTO_COMPUTE_PARTITION_MODE) { 503 mode = __aqua_vanjaram_get_auto_mode(xcp_mgr); 504 } else if (!__aqua_vanjaram_is_valid_mode(xcp_mgr, mode)) { 505 dev_err(adev->dev, 506 "Invalid compute partition mode requested, requested: %s, available memory partitions: %d", 507 amdgpu_gfx_compute_mode_desc(mode), adev->gmc.num_mem_partitions); 508 return -EINVAL; 509 } 510 511 if (adev->kfd.init_complete && !amdgpu_in_reset(adev)) 512 flags |= AMDGPU_XCP_OPS_KFD; 513 514 if (flags & AMDGPU_XCP_OPS_KFD) { 515 ret = amdgpu_amdkfd_check_and_lock_kfd(adev); 516 if (ret) 517 goto out; 518 } 519 520 ret = __aqua_vanjaram_pre_partition_switch(xcp_mgr, flags); 521 if (ret) 522 goto unlock; 523 524 num_xcc_per_xcp = __aqua_vanjaram_get_xcc_per_xcp(xcp_mgr, mode); 525 if (adev->gfx.funcs->switch_partition_mode) 526 adev->gfx.funcs->switch_partition_mode(xcp_mgr->adev, 527 num_xcc_per_xcp); 528 529 /* Init info about new xcps */ 530 *num_xcps = num_xcc / num_xcc_per_xcp; 531 amdgpu_xcp_init(xcp_mgr, *num_xcps, mode); 532 533 ret = __aqua_vanjaram_post_partition_switch(xcp_mgr, flags); 534 unlock: 535 if (flags & AMDGPU_XCP_OPS_KFD) 536 amdgpu_amdkfd_unlock_kfd(adev); 537 out: 538 return ret; 539 } 540 541 static int __aqua_vanjaram_get_xcp_mem_id(struct amdgpu_device *adev, 542 int xcc_id, uint8_t *mem_id) 543 { 544 /* memory/spatial modes validation check is already done */ 545 *mem_id = xcc_id / adev->gfx.num_xcc_per_xcp; 546 *mem_id /= adev->xcp_mgr->num_xcp_per_mem_partition; 547 548 return 0; 549 } 550 551 static int aqua_vanjaram_get_xcp_mem_id(struct amdgpu_xcp_mgr *xcp_mgr, 552 struct amdgpu_xcp *xcp, uint8_t *mem_id) 553 { 554 struct amdgpu_numa_info numa_info; 555 struct amdgpu_device *adev; 556 uint32_t xcc_mask; 557 int r, i, xcc_id; 558 559 adev = xcp_mgr->adev; 560 /* TODO: BIOS is not returning the right info now 561 * Check on this later 562 */ 563 /* 564 if (adev->gmc.gmc_funcs->query_mem_partition_mode) 565 mode = adev->gmc.gmc_funcs->query_mem_partition_mode(adev); 566 */ 567 if (adev->gmc.num_mem_partitions == 1) { 568 /* Only one range */ 569 *mem_id = 0; 570 return 0; 571 } 572 573 r = amdgpu_xcp_get_inst_details(xcp, AMDGPU_XCP_GFX, &xcc_mask); 574 if (r || !xcc_mask) 575 return -EINVAL; 576 577 xcc_id = ffs(xcc_mask) - 1; 578 if (!adev->gmc.is_app_apu) 579 return __aqua_vanjaram_get_xcp_mem_id(adev, xcc_id, mem_id); 580 581 r = amdgpu_acpi_get_mem_info(adev, xcc_id, &numa_info); 582 583 if (r) 584 return r; 585 586 r = -EINVAL; 587 for (i = 0; i < adev->gmc.num_mem_partitions; ++i) { 588 if (adev->gmc.mem_partitions[i].numa.node == numa_info.nid) { 589 *mem_id = i; 590 r = 0; 591 break; 592 } 593 } 594 595 return r; 596 } 597 598 static int aqua_vanjaram_get_xcp_ip_details(struct amdgpu_xcp_mgr *xcp_mgr, int xcp_id, 599 enum AMDGPU_XCP_IP_BLOCK ip_id, 600 struct amdgpu_xcp_ip *ip) 601 { 602 if (!ip) 603 return -EINVAL; 604 605 return __aqua_vanjaram_get_xcp_ip_info(xcp_mgr, xcp_id, ip_id, ip); 606 } 607 608 struct amdgpu_xcp_mgr_funcs aqua_vanjaram_xcp_funcs = { 609 .switch_partition_mode = &aqua_vanjaram_switch_partition_mode, 610 .query_partition_mode = &aqua_vanjaram_query_partition_mode, 611 .get_ip_details = &aqua_vanjaram_get_xcp_ip_details, 612 .get_xcp_mem_id = &aqua_vanjaram_get_xcp_mem_id, 613 .select_scheds = &aqua_vanjaram_select_scheds, 614 .update_partition_sched_list = &aqua_vanjaram_update_partition_sched_list 615 }; 616 617 static int aqua_vanjaram_xcp_mgr_init(struct amdgpu_device *adev) 618 { 619 int ret; 620 621 ret = amdgpu_xcp_mgr_init(adev, AMDGPU_UNKNOWN_COMPUTE_PARTITION_MODE, 1, 622 &aqua_vanjaram_xcp_funcs); 623 if (ret) 624 return ret; 625 626 /* TODO: Default memory node affinity init */ 627 628 return ret; 629 } 630 631 int aqua_vanjaram_init_soc_config(struct amdgpu_device *adev) 632 { 633 u32 mask, avail_inst, inst_mask = adev->sdma.sdma_mask; 634 int ret, i; 635 636 /* generally 1 AID supports 4 instances */ 637 adev->sdma.num_inst_per_aid = 4; 638 adev->sdma.num_instances = NUM_SDMA(adev->sdma.sdma_mask); 639 640 adev->aid_mask = i = 1; 641 inst_mask >>= adev->sdma.num_inst_per_aid; 642 643 for (mask = (1 << adev->sdma.num_inst_per_aid) - 1; inst_mask; 644 inst_mask >>= adev->sdma.num_inst_per_aid, ++i) { 645 avail_inst = inst_mask & mask; 646 if (avail_inst == mask || avail_inst == 0x3 || 647 avail_inst == 0xc) 648 adev->aid_mask |= (1 << i); 649 } 650 651 /* Harvest config is not used for aqua vanjaram. VCN and JPEGs will be 652 * addressed based on logical instance ids. 653 */ 654 adev->vcn.harvest_config = 0; 655 adev->vcn.num_inst_per_aid = 1; 656 adev->vcn.num_vcn_inst = hweight32(adev->vcn.inst_mask); 657 adev->jpeg.harvest_config = 0; 658 adev->jpeg.num_inst_per_aid = 1; 659 adev->jpeg.num_jpeg_inst = hweight32(adev->jpeg.inst_mask); 660 661 ret = aqua_vanjaram_xcp_mgr_init(adev); 662 if (ret) 663 return ret; 664 665 aqua_vanjaram_ip_map_init(adev); 666 667 return 0; 668 } 669 670 static void aqua_read_smn(struct amdgpu_device *adev, 671 struct amdgpu_smn_reg_data *regdata, 672 uint64_t smn_addr) 673 { 674 regdata->addr = smn_addr; 675 regdata->value = RREG32_PCIE(smn_addr); 676 } 677 678 struct aqua_reg_list { 679 uint64_t start_addr; 680 uint32_t num_regs; 681 uint32_t incrx; 682 }; 683 684 #define DW_ADDR_INCR 4 685 686 static void aqua_read_smn_ext(struct amdgpu_device *adev, 687 struct amdgpu_smn_reg_data *regdata, 688 uint64_t smn_addr, int i) 689 { 690 regdata->addr = 691 smn_addr + adev->asic_funcs->encode_ext_smn_addressing(i); 692 regdata->value = RREG32_PCIE_EXT(regdata->addr); 693 } 694 695 #define smnreg_0x1A340218 0x1A340218 696 #define smnreg_0x1A3402E4 0x1A3402E4 697 #define smnreg_0x1A340294 0x1A340294 698 #define smreg_0x1A380088 0x1A380088 699 700 #define NUM_PCIE_SMN_REGS 14 701 702 static struct aqua_reg_list pcie_reg_addrs[] = { 703 { smnreg_0x1A340218, 1, 0 }, 704 { smnreg_0x1A3402E4, 1, 0 }, 705 { smnreg_0x1A340294, 6, DW_ADDR_INCR }, 706 { smreg_0x1A380088, 6, DW_ADDR_INCR }, 707 }; 708 709 static ssize_t aqua_vanjaram_read_pcie_state(struct amdgpu_device *adev, 710 void *buf, size_t max_size) 711 { 712 struct amdgpu_reg_state_pcie_v1_0 *pcie_reg_state; 713 uint32_t start_addr, incrx, num_regs, szbuf; 714 struct amdgpu_regs_pcie_v1_0 *pcie_regs; 715 struct amdgpu_smn_reg_data *reg_data; 716 struct pci_dev *us_pdev, *ds_pdev; 717 int aer_cap, r, n; 718 719 if (!buf || !max_size) 720 return -EINVAL; 721 722 pcie_reg_state = (struct amdgpu_reg_state_pcie_v1_0 *)buf; 723 724 szbuf = sizeof(*pcie_reg_state) + 725 amdgpu_reginst_size(1, sizeof(*pcie_regs), NUM_PCIE_SMN_REGS); 726 /* Only one instance of pcie regs */ 727 if (max_size < szbuf) 728 return -EOVERFLOW; 729 730 pcie_regs = (struct amdgpu_regs_pcie_v1_0 *)((uint8_t *)buf + 731 sizeof(*pcie_reg_state)); 732 pcie_regs->inst_header.instance = 0; 733 pcie_regs->inst_header.state = AMDGPU_INST_S_OK; 734 pcie_regs->inst_header.num_smn_regs = NUM_PCIE_SMN_REGS; 735 736 reg_data = pcie_regs->smn_reg_values; 737 738 for (r = 0; r < ARRAY_SIZE(pcie_reg_addrs); r++) { 739 start_addr = pcie_reg_addrs[r].start_addr; 740 incrx = pcie_reg_addrs[r].incrx; 741 num_regs = pcie_reg_addrs[r].num_regs; 742 for (n = 0; n < num_regs; n++) { 743 aqua_read_smn(adev, reg_data, start_addr + n * incrx); 744 ++reg_data; 745 } 746 } 747 748 ds_pdev = pci_upstream_bridge(adev->pdev); 749 us_pdev = pci_upstream_bridge(ds_pdev); 750 751 pcie_capability_read_word(us_pdev, PCI_EXP_DEVSTA, 752 &pcie_regs->device_status); 753 pcie_capability_read_word(us_pdev, PCI_EXP_LNKSTA, 754 &pcie_regs->link_status); 755 756 aer_cap = pci_find_ext_capability(us_pdev, PCI_EXT_CAP_ID_ERR); 757 if (aer_cap) { 758 pci_read_config_dword(us_pdev, aer_cap + PCI_ERR_COR_STATUS, 759 &pcie_regs->pcie_corr_err_status); 760 pci_read_config_dword(us_pdev, aer_cap + PCI_ERR_UNCOR_STATUS, 761 &pcie_regs->pcie_uncorr_err_status); 762 } 763 764 pci_read_config_dword(us_pdev, PCI_PRIMARY_BUS, 765 &pcie_regs->sub_bus_number_latency); 766 767 pcie_reg_state->common_header.structure_size = szbuf; 768 pcie_reg_state->common_header.format_revision = 1; 769 pcie_reg_state->common_header.content_revision = 0; 770 pcie_reg_state->common_header.state_type = AMDGPU_REG_STATE_TYPE_PCIE; 771 pcie_reg_state->common_header.num_instances = 1; 772 773 return pcie_reg_state->common_header.structure_size; 774 } 775 776 #define smnreg_0x11A00050 0x11A00050 777 #define smnreg_0x11A00180 0x11A00180 778 #define smnreg_0x11A00070 0x11A00070 779 #define smnreg_0x11A00200 0x11A00200 780 #define smnreg_0x11A0020C 0x11A0020C 781 #define smnreg_0x11A00210 0x11A00210 782 #define smnreg_0x11A00108 0x11A00108 783 784 #define XGMI_LINK_REG(smnreg, l) ((smnreg) | (l << 20)) 785 786 #define NUM_XGMI_SMN_REGS 25 787 788 static struct aqua_reg_list xgmi_reg_addrs[] = { 789 { smnreg_0x11A00050, 1, 0 }, 790 { smnreg_0x11A00180, 16, DW_ADDR_INCR }, 791 { smnreg_0x11A00070, 4, DW_ADDR_INCR }, 792 { smnreg_0x11A00200, 1, 0 }, 793 { smnreg_0x11A0020C, 1, 0 }, 794 { smnreg_0x11A00210, 1, 0 }, 795 { smnreg_0x11A00108, 1, 0 }, 796 }; 797 798 static ssize_t aqua_vanjaram_read_xgmi_state(struct amdgpu_device *adev, 799 void *buf, size_t max_size) 800 { 801 struct amdgpu_reg_state_xgmi_v1_0 *xgmi_reg_state; 802 uint32_t start_addr, incrx, num_regs, szbuf; 803 struct amdgpu_regs_xgmi_v1_0 *xgmi_regs; 804 struct amdgpu_smn_reg_data *reg_data; 805 const int max_xgmi_instances = 8; 806 int inst = 0, i, j, r, n; 807 const int xgmi_inst = 2; 808 void *p; 809 810 if (!buf || !max_size) 811 return -EINVAL; 812 813 xgmi_reg_state = (struct amdgpu_reg_state_xgmi_v1_0 *)buf; 814 815 szbuf = sizeof(*xgmi_reg_state) + 816 amdgpu_reginst_size(max_xgmi_instances, sizeof(*xgmi_regs), 817 NUM_XGMI_SMN_REGS); 818 /* Only one instance of pcie regs */ 819 if (max_size < szbuf) 820 return -EOVERFLOW; 821 822 p = &xgmi_reg_state->xgmi_state_regs[0]; 823 for_each_inst(i, adev->aid_mask) { 824 for (j = 0; j < xgmi_inst; ++j) { 825 xgmi_regs = (struct amdgpu_regs_xgmi_v1_0 *)p; 826 xgmi_regs->inst_header.instance = inst++; 827 828 xgmi_regs->inst_header.state = AMDGPU_INST_S_OK; 829 xgmi_regs->inst_header.num_smn_regs = NUM_XGMI_SMN_REGS; 830 831 reg_data = xgmi_regs->smn_reg_values; 832 833 for (r = 0; r < ARRAY_SIZE(xgmi_reg_addrs); r++) { 834 start_addr = xgmi_reg_addrs[r].start_addr; 835 incrx = xgmi_reg_addrs[r].incrx; 836 num_regs = xgmi_reg_addrs[r].num_regs; 837 838 for (n = 0; n < num_regs; n++) { 839 aqua_read_smn_ext( 840 adev, reg_data, 841 XGMI_LINK_REG(start_addr, j) + 842 n * incrx, 843 i); 844 ++reg_data; 845 } 846 } 847 p = reg_data; 848 } 849 } 850 851 xgmi_reg_state->common_header.structure_size = szbuf; 852 xgmi_reg_state->common_header.format_revision = 1; 853 xgmi_reg_state->common_header.content_revision = 0; 854 xgmi_reg_state->common_header.state_type = AMDGPU_REG_STATE_TYPE_XGMI; 855 xgmi_reg_state->common_header.num_instances = max_xgmi_instances; 856 857 return xgmi_reg_state->common_header.structure_size; 858 } 859 860 #define smnreg_0x11C00070 0x11C00070 861 #define smnreg_0x11C00210 0x11C00210 862 863 static struct aqua_reg_list wafl_reg_addrs[] = { 864 { smnreg_0x11C00070, 4, DW_ADDR_INCR }, 865 { smnreg_0x11C00210, 1, 0 }, 866 }; 867 868 #define WAFL_LINK_REG(smnreg, l) ((smnreg) | (l << 20)) 869 870 #define NUM_WAFL_SMN_REGS 5 871 872 static ssize_t aqua_vanjaram_read_wafl_state(struct amdgpu_device *adev, 873 void *buf, size_t max_size) 874 { 875 struct amdgpu_reg_state_wafl_v1_0 *wafl_reg_state; 876 uint32_t start_addr, incrx, num_regs, szbuf; 877 struct amdgpu_regs_wafl_v1_0 *wafl_regs; 878 struct amdgpu_smn_reg_data *reg_data; 879 const int max_wafl_instances = 8; 880 int inst = 0, i, j, r, n; 881 const int wafl_inst = 2; 882 void *p; 883 884 if (!buf || !max_size) 885 return -EINVAL; 886 887 wafl_reg_state = (struct amdgpu_reg_state_wafl_v1_0 *)buf; 888 889 szbuf = sizeof(*wafl_reg_state) + 890 amdgpu_reginst_size(max_wafl_instances, sizeof(*wafl_regs), 891 NUM_WAFL_SMN_REGS); 892 893 if (max_size < szbuf) 894 return -EOVERFLOW; 895 896 p = &wafl_reg_state->wafl_state_regs[0]; 897 for_each_inst(i, adev->aid_mask) { 898 for (j = 0; j < wafl_inst; ++j) { 899 wafl_regs = (struct amdgpu_regs_wafl_v1_0 *)p; 900 wafl_regs->inst_header.instance = inst++; 901 902 wafl_regs->inst_header.state = AMDGPU_INST_S_OK; 903 wafl_regs->inst_header.num_smn_regs = NUM_WAFL_SMN_REGS; 904 905 reg_data = wafl_regs->smn_reg_values; 906 907 for (r = 0; r < ARRAY_SIZE(wafl_reg_addrs); r++) { 908 start_addr = wafl_reg_addrs[r].start_addr; 909 incrx = wafl_reg_addrs[r].incrx; 910 num_regs = wafl_reg_addrs[r].num_regs; 911 for (n = 0; n < num_regs; n++) { 912 aqua_read_smn_ext( 913 adev, reg_data, 914 WAFL_LINK_REG(start_addr, j) + 915 n * incrx, 916 i); 917 ++reg_data; 918 } 919 } 920 p = reg_data; 921 } 922 } 923 924 wafl_reg_state->common_header.structure_size = szbuf; 925 wafl_reg_state->common_header.format_revision = 1; 926 wafl_reg_state->common_header.content_revision = 0; 927 wafl_reg_state->common_header.state_type = AMDGPU_REG_STATE_TYPE_WAFL; 928 wafl_reg_state->common_header.num_instances = max_wafl_instances; 929 930 return wafl_reg_state->common_header.structure_size; 931 } 932 933 #define smnreg_0x1B311060 0x1B311060 934 #define smnreg_0x1B411060 0x1B411060 935 #define smnreg_0x1B511060 0x1B511060 936 #define smnreg_0x1B611060 0x1B611060 937 938 #define smnreg_0x1C307120 0x1C307120 939 #define smnreg_0x1C317120 0x1C317120 940 941 #define smnreg_0x1C320830 0x1C320830 942 #define smnreg_0x1C380830 0x1C380830 943 #define smnreg_0x1C3D0830 0x1C3D0830 944 #define smnreg_0x1C420830 0x1C420830 945 946 #define smnreg_0x1C320100 0x1C320100 947 #define smnreg_0x1C380100 0x1C380100 948 #define smnreg_0x1C3D0100 0x1C3D0100 949 #define smnreg_0x1C420100 0x1C420100 950 951 #define smnreg_0x1B310500 0x1B310500 952 #define smnreg_0x1C300400 0x1C300400 953 954 #define USR_CAKE_INCR 0x11000 955 #define USR_LINK_INCR 0x100000 956 #define USR_CP_INCR 0x10000 957 958 #define NUM_USR_SMN_REGS 20 959 960 struct aqua_reg_list usr_reg_addrs[] = { 961 { smnreg_0x1B311060, 4, DW_ADDR_INCR }, 962 { smnreg_0x1B411060, 4, DW_ADDR_INCR }, 963 { smnreg_0x1B511060, 4, DW_ADDR_INCR }, 964 { smnreg_0x1B611060, 4, DW_ADDR_INCR }, 965 { smnreg_0x1C307120, 2, DW_ADDR_INCR }, 966 { smnreg_0x1C317120, 2, DW_ADDR_INCR }, 967 }; 968 969 #define NUM_USR1_SMN_REGS 46 970 struct aqua_reg_list usr1_reg_addrs[] = { 971 { smnreg_0x1C320830, 6, USR_CAKE_INCR }, 972 { smnreg_0x1C380830, 5, USR_CAKE_INCR }, 973 { smnreg_0x1C3D0830, 5, USR_CAKE_INCR }, 974 { smnreg_0x1C420830, 4, USR_CAKE_INCR }, 975 { smnreg_0x1C320100, 6, USR_CAKE_INCR }, 976 { smnreg_0x1C380100, 5, USR_CAKE_INCR }, 977 { smnreg_0x1C3D0100, 5, USR_CAKE_INCR }, 978 { smnreg_0x1C420100, 4, USR_CAKE_INCR }, 979 { smnreg_0x1B310500, 4, USR_LINK_INCR }, 980 { smnreg_0x1C300400, 2, USR_CP_INCR }, 981 }; 982 983 static ssize_t aqua_vanjaram_read_usr_state(struct amdgpu_device *adev, 984 void *buf, size_t max_size, 985 int reg_state) 986 { 987 uint32_t start_addr, incrx, num_regs, szbuf, num_smn; 988 struct amdgpu_reg_state_usr_v1_0 *usr_reg_state; 989 struct amdgpu_regs_usr_v1_0 *usr_regs; 990 struct amdgpu_smn_reg_data *reg_data; 991 const int max_usr_instances = 4; 992 struct aqua_reg_list *reg_addrs; 993 int inst = 0, i, n, r, arr_size; 994 void *p; 995 996 if (!buf || !max_size) 997 return -EINVAL; 998 999 switch (reg_state) { 1000 case AMDGPU_REG_STATE_TYPE_USR: 1001 arr_size = ARRAY_SIZE(usr_reg_addrs); 1002 reg_addrs = usr_reg_addrs; 1003 num_smn = NUM_USR_SMN_REGS; 1004 break; 1005 case AMDGPU_REG_STATE_TYPE_USR_1: 1006 arr_size = ARRAY_SIZE(usr1_reg_addrs); 1007 reg_addrs = usr1_reg_addrs; 1008 num_smn = NUM_USR1_SMN_REGS; 1009 break; 1010 default: 1011 return -EINVAL; 1012 } 1013 1014 usr_reg_state = (struct amdgpu_reg_state_usr_v1_0 *)buf; 1015 1016 szbuf = sizeof(*usr_reg_state) + amdgpu_reginst_size(max_usr_instances, 1017 sizeof(*usr_regs), 1018 num_smn); 1019 if (max_size < szbuf) 1020 return -EOVERFLOW; 1021 1022 p = &usr_reg_state->usr_state_regs[0]; 1023 for_each_inst(i, adev->aid_mask) { 1024 usr_regs = (struct amdgpu_regs_usr_v1_0 *)p; 1025 usr_regs->inst_header.instance = inst++; 1026 usr_regs->inst_header.state = AMDGPU_INST_S_OK; 1027 usr_regs->inst_header.num_smn_regs = num_smn; 1028 reg_data = usr_regs->smn_reg_values; 1029 1030 for (r = 0; r < arr_size; r++) { 1031 start_addr = reg_addrs[r].start_addr; 1032 incrx = reg_addrs[r].incrx; 1033 num_regs = reg_addrs[r].num_regs; 1034 for (n = 0; n < num_regs; n++) { 1035 aqua_read_smn_ext(adev, reg_data, 1036 start_addr + n * incrx, i); 1037 reg_data++; 1038 } 1039 } 1040 p = reg_data; 1041 } 1042 1043 usr_reg_state->common_header.structure_size = szbuf; 1044 usr_reg_state->common_header.format_revision = 1; 1045 usr_reg_state->common_header.content_revision = 0; 1046 usr_reg_state->common_header.state_type = AMDGPU_REG_STATE_TYPE_USR; 1047 usr_reg_state->common_header.num_instances = max_usr_instances; 1048 1049 return usr_reg_state->common_header.structure_size; 1050 } 1051 1052 ssize_t aqua_vanjaram_get_reg_state(struct amdgpu_device *adev, 1053 enum amdgpu_reg_state reg_state, void *buf, 1054 size_t max_size) 1055 { 1056 ssize_t size; 1057 1058 switch (reg_state) { 1059 case AMDGPU_REG_STATE_TYPE_PCIE: 1060 size = aqua_vanjaram_read_pcie_state(adev, buf, max_size); 1061 break; 1062 case AMDGPU_REG_STATE_TYPE_XGMI: 1063 size = aqua_vanjaram_read_xgmi_state(adev, buf, max_size); 1064 break; 1065 case AMDGPU_REG_STATE_TYPE_WAFL: 1066 size = aqua_vanjaram_read_wafl_state(adev, buf, max_size); 1067 break; 1068 case AMDGPU_REG_STATE_TYPE_USR: 1069 size = aqua_vanjaram_read_usr_state(adev, buf, max_size, 1070 AMDGPU_REG_STATE_TYPE_USR); 1071 break; 1072 case AMDGPU_REG_STATE_TYPE_USR_1: 1073 size = aqua_vanjaram_read_usr_state( 1074 adev, buf, max_size, AMDGPU_REG_STATE_TYPE_USR_1); 1075 break; 1076 default: 1077 return -EINVAL; 1078 } 1079 1080 return size; 1081 } 1082