1 // SPDX-License-Identifier: GPL-2.0 OR MIT
3 * Copyright 2015-2022 Advanced Micro Devices, Inc.
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
12 * The above copyright notice and this permission notice shall be included in
13 * all copies or substantial portions of the Software.
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21 * OTHER DEALINGS IN THE SOFTWARE.
24 #include <linux/pci.h>
25 #include <linux/acpi.h>
28 #include "kfd_topology.h"
29 #include "kfd_iommu.h"
31 #include "amdgpu_amdkfd.h"
33 /* GPU Processor ID base for dGPUs for which VCRAT needs to be created.
34 * GPU processor ID are expressed with Bit[31]=1.
35 * The base is set to 0x8000_0000 + 0x1000 to avoid collision with GPU IDs
38 static uint32_t gpu_processor_id_low = 0x80001000;
40 /* Return the next available gpu_processor_id and increment it for next GPU
41 * @total_cu_count - Total CUs present in the GPU including ones
44 static inline unsigned int get_and_inc_gpu_processor_id(
45 unsigned int total_cu_count)
47 int current_id = gpu_processor_id_low;
49 gpu_processor_id_low += total_cu_count;
54 static struct kfd_gpu_cache_info kaveri_cache_info[] = {
56 /* TCP L1 Cache per CU */
59 .flags = (CRAT_CACHE_FLAGS_ENABLED |
60 CRAT_CACHE_FLAGS_DATA_CACHE |
61 CRAT_CACHE_FLAGS_SIMD_CACHE),
65 /* Scalar L1 Instruction Cache (in SQC module) per bank */
68 .flags = (CRAT_CACHE_FLAGS_ENABLED |
69 CRAT_CACHE_FLAGS_INST_CACHE |
70 CRAT_CACHE_FLAGS_SIMD_CACHE),
74 /* Scalar L1 Data Cache (in SQC module) per bank */
77 .flags = (CRAT_CACHE_FLAGS_ENABLED |
78 CRAT_CACHE_FLAGS_DATA_CACHE |
79 CRAT_CACHE_FLAGS_SIMD_CACHE),
83 /* TODO: Add L2 Cache information */
87 static struct kfd_gpu_cache_info carrizo_cache_info[] = {
89 /* TCP L1 Cache per CU */
92 .flags = (CRAT_CACHE_FLAGS_ENABLED |
93 CRAT_CACHE_FLAGS_DATA_CACHE |
94 CRAT_CACHE_FLAGS_SIMD_CACHE),
98 /* Scalar L1 Instruction Cache (in SQC module) per bank */
101 .flags = (CRAT_CACHE_FLAGS_ENABLED |
102 CRAT_CACHE_FLAGS_INST_CACHE |
103 CRAT_CACHE_FLAGS_SIMD_CACHE),
107 /* Scalar L1 Data Cache (in SQC module) per bank. */
110 .flags = (CRAT_CACHE_FLAGS_ENABLED |
111 CRAT_CACHE_FLAGS_DATA_CACHE |
112 CRAT_CACHE_FLAGS_SIMD_CACHE),
116 /* TODO: Add L2 Cache information */
119 #define hawaii_cache_info kaveri_cache_info
120 #define tonga_cache_info carrizo_cache_info
121 #define fiji_cache_info carrizo_cache_info
122 #define polaris10_cache_info carrizo_cache_info
123 #define polaris11_cache_info carrizo_cache_info
124 #define polaris12_cache_info carrizo_cache_info
125 #define vegam_cache_info carrizo_cache_info
127 /* NOTE: L1 cache information has been updated and L2/L3
128 * cache information has been added for Vega10 and
129 * newer ASICs. The unit for cache_size is KiB.
130 * In future, check & update cache details
131 * for every new ASIC is required.
134 static struct kfd_gpu_cache_info vega10_cache_info[] = {
136 /* TCP L1 Cache per CU */
139 .flags = (CRAT_CACHE_FLAGS_ENABLED |
140 CRAT_CACHE_FLAGS_DATA_CACHE |
141 CRAT_CACHE_FLAGS_SIMD_CACHE),
145 /* Scalar L1 Instruction Cache per SQC */
148 .flags = (CRAT_CACHE_FLAGS_ENABLED |
149 CRAT_CACHE_FLAGS_INST_CACHE |
150 CRAT_CACHE_FLAGS_SIMD_CACHE),
154 /* Scalar L1 Data Cache per SQC */
157 .flags = (CRAT_CACHE_FLAGS_ENABLED |
158 CRAT_CACHE_FLAGS_DATA_CACHE |
159 CRAT_CACHE_FLAGS_SIMD_CACHE),
163 /* L2 Data Cache per GPU (Total Tex Cache) */
166 .flags = (CRAT_CACHE_FLAGS_ENABLED |
167 CRAT_CACHE_FLAGS_DATA_CACHE |
168 CRAT_CACHE_FLAGS_SIMD_CACHE),
173 static struct kfd_gpu_cache_info raven_cache_info[] = {
175 /* TCP L1 Cache per CU */
178 .flags = (CRAT_CACHE_FLAGS_ENABLED |
179 CRAT_CACHE_FLAGS_DATA_CACHE |
180 CRAT_CACHE_FLAGS_SIMD_CACHE),
184 /* Scalar L1 Instruction Cache per SQC */
187 .flags = (CRAT_CACHE_FLAGS_ENABLED |
188 CRAT_CACHE_FLAGS_INST_CACHE |
189 CRAT_CACHE_FLAGS_SIMD_CACHE),
193 /* Scalar L1 Data Cache per SQC */
196 .flags = (CRAT_CACHE_FLAGS_ENABLED |
197 CRAT_CACHE_FLAGS_DATA_CACHE |
198 CRAT_CACHE_FLAGS_SIMD_CACHE),
202 /* L2 Data Cache per GPU (Total Tex Cache) */
205 .flags = (CRAT_CACHE_FLAGS_ENABLED |
206 CRAT_CACHE_FLAGS_DATA_CACHE |
207 CRAT_CACHE_FLAGS_SIMD_CACHE),
212 static struct kfd_gpu_cache_info renoir_cache_info[] = {
214 /* TCP L1 Cache per CU */
217 .flags = (CRAT_CACHE_FLAGS_ENABLED |
218 CRAT_CACHE_FLAGS_DATA_CACHE |
219 CRAT_CACHE_FLAGS_SIMD_CACHE),
223 /* Scalar L1 Instruction Cache per SQC */
226 .flags = (CRAT_CACHE_FLAGS_ENABLED |
227 CRAT_CACHE_FLAGS_INST_CACHE |
228 CRAT_CACHE_FLAGS_SIMD_CACHE),
232 /* Scalar L1 Data Cache per SQC */
235 .flags = (CRAT_CACHE_FLAGS_ENABLED |
236 CRAT_CACHE_FLAGS_DATA_CACHE |
237 CRAT_CACHE_FLAGS_SIMD_CACHE),
241 /* L2 Data Cache per GPU (Total Tex Cache) */
244 .flags = (CRAT_CACHE_FLAGS_ENABLED |
245 CRAT_CACHE_FLAGS_DATA_CACHE |
246 CRAT_CACHE_FLAGS_SIMD_CACHE),
251 static struct kfd_gpu_cache_info vega12_cache_info[] = {
253 /* TCP L1 Cache per CU */
256 .flags = (CRAT_CACHE_FLAGS_ENABLED |
257 CRAT_CACHE_FLAGS_DATA_CACHE |
258 CRAT_CACHE_FLAGS_SIMD_CACHE),
262 /* Scalar L1 Instruction Cache per SQC */
265 .flags = (CRAT_CACHE_FLAGS_ENABLED |
266 CRAT_CACHE_FLAGS_INST_CACHE |
267 CRAT_CACHE_FLAGS_SIMD_CACHE),
271 /* Scalar L1 Data Cache per SQC */
274 .flags = (CRAT_CACHE_FLAGS_ENABLED |
275 CRAT_CACHE_FLAGS_DATA_CACHE |
276 CRAT_CACHE_FLAGS_SIMD_CACHE),
280 /* L2 Data Cache per GPU (Total Tex Cache) */
283 .flags = (CRAT_CACHE_FLAGS_ENABLED |
284 CRAT_CACHE_FLAGS_DATA_CACHE |
285 CRAT_CACHE_FLAGS_SIMD_CACHE),
290 static struct kfd_gpu_cache_info vega20_cache_info[] = {
292 /* TCP L1 Cache per CU */
295 .flags = (CRAT_CACHE_FLAGS_ENABLED |
296 CRAT_CACHE_FLAGS_DATA_CACHE |
297 CRAT_CACHE_FLAGS_SIMD_CACHE),
301 /* Scalar L1 Instruction Cache per SQC */
304 .flags = (CRAT_CACHE_FLAGS_ENABLED |
305 CRAT_CACHE_FLAGS_INST_CACHE |
306 CRAT_CACHE_FLAGS_SIMD_CACHE),
310 /* Scalar L1 Data Cache per SQC */
313 .flags = (CRAT_CACHE_FLAGS_ENABLED |
314 CRAT_CACHE_FLAGS_DATA_CACHE |
315 CRAT_CACHE_FLAGS_SIMD_CACHE),
319 /* L2 Data Cache per GPU (Total Tex Cache) */
322 .flags = (CRAT_CACHE_FLAGS_ENABLED |
323 CRAT_CACHE_FLAGS_DATA_CACHE |
324 CRAT_CACHE_FLAGS_SIMD_CACHE),
329 static struct kfd_gpu_cache_info aldebaran_cache_info[] = {
331 /* TCP L1 Cache per CU */
334 .flags = (CRAT_CACHE_FLAGS_ENABLED |
335 CRAT_CACHE_FLAGS_DATA_CACHE |
336 CRAT_CACHE_FLAGS_SIMD_CACHE),
340 /* Scalar L1 Instruction Cache per SQC */
343 .flags = (CRAT_CACHE_FLAGS_ENABLED |
344 CRAT_CACHE_FLAGS_INST_CACHE |
345 CRAT_CACHE_FLAGS_SIMD_CACHE),
349 /* Scalar L1 Data Cache per SQC */
352 .flags = (CRAT_CACHE_FLAGS_ENABLED |
353 CRAT_CACHE_FLAGS_DATA_CACHE |
354 CRAT_CACHE_FLAGS_SIMD_CACHE),
358 /* L2 Data Cache per GPU (Total Tex Cache) */
361 .flags = (CRAT_CACHE_FLAGS_ENABLED |
362 CRAT_CACHE_FLAGS_DATA_CACHE |
363 CRAT_CACHE_FLAGS_SIMD_CACHE),
368 static struct kfd_gpu_cache_info navi10_cache_info[] = {
370 /* TCP L1 Cache per CU */
373 .flags = (CRAT_CACHE_FLAGS_ENABLED |
374 CRAT_CACHE_FLAGS_DATA_CACHE |
375 CRAT_CACHE_FLAGS_SIMD_CACHE),
379 /* Scalar L1 Instruction Cache per SQC */
382 .flags = (CRAT_CACHE_FLAGS_ENABLED |
383 CRAT_CACHE_FLAGS_INST_CACHE |
384 CRAT_CACHE_FLAGS_SIMD_CACHE),
388 /* Scalar L1 Data Cache per SQC */
391 .flags = (CRAT_CACHE_FLAGS_ENABLED |
392 CRAT_CACHE_FLAGS_DATA_CACHE |
393 CRAT_CACHE_FLAGS_SIMD_CACHE),
397 /* GL1 Data Cache per SA */
400 .flags = (CRAT_CACHE_FLAGS_ENABLED |
401 CRAT_CACHE_FLAGS_DATA_CACHE |
402 CRAT_CACHE_FLAGS_SIMD_CACHE),
406 /* L2 Data Cache per GPU (Total Tex Cache) */
409 .flags = (CRAT_CACHE_FLAGS_ENABLED |
410 CRAT_CACHE_FLAGS_DATA_CACHE |
411 CRAT_CACHE_FLAGS_SIMD_CACHE),
416 static struct kfd_gpu_cache_info vangogh_cache_info[] = {
418 /* TCP L1 Cache per CU */
421 .flags = (CRAT_CACHE_FLAGS_ENABLED |
422 CRAT_CACHE_FLAGS_DATA_CACHE |
423 CRAT_CACHE_FLAGS_SIMD_CACHE),
427 /* Scalar L1 Instruction Cache per SQC */
430 .flags = (CRAT_CACHE_FLAGS_ENABLED |
431 CRAT_CACHE_FLAGS_INST_CACHE |
432 CRAT_CACHE_FLAGS_SIMD_CACHE),
436 /* Scalar L1 Data Cache per SQC */
439 .flags = (CRAT_CACHE_FLAGS_ENABLED |
440 CRAT_CACHE_FLAGS_DATA_CACHE |
441 CRAT_CACHE_FLAGS_SIMD_CACHE),
445 /* GL1 Data Cache per SA */
448 .flags = (CRAT_CACHE_FLAGS_ENABLED |
449 CRAT_CACHE_FLAGS_DATA_CACHE |
450 CRAT_CACHE_FLAGS_SIMD_CACHE),
454 /* L2 Data Cache per GPU (Total Tex Cache) */
457 .flags = (CRAT_CACHE_FLAGS_ENABLED |
458 CRAT_CACHE_FLAGS_DATA_CACHE |
459 CRAT_CACHE_FLAGS_SIMD_CACHE),
464 static struct kfd_gpu_cache_info navi14_cache_info[] = {
466 /* TCP L1 Cache per CU */
469 .flags = (CRAT_CACHE_FLAGS_ENABLED |
470 CRAT_CACHE_FLAGS_DATA_CACHE |
471 CRAT_CACHE_FLAGS_SIMD_CACHE),
475 /* Scalar L1 Instruction Cache per SQC */
478 .flags = (CRAT_CACHE_FLAGS_ENABLED |
479 CRAT_CACHE_FLAGS_INST_CACHE |
480 CRAT_CACHE_FLAGS_SIMD_CACHE),
484 /* Scalar L1 Data Cache per SQC */
487 .flags = (CRAT_CACHE_FLAGS_ENABLED |
488 CRAT_CACHE_FLAGS_DATA_CACHE |
489 CRAT_CACHE_FLAGS_SIMD_CACHE),
493 /* GL1 Data Cache per SA */
496 .flags = (CRAT_CACHE_FLAGS_ENABLED |
497 CRAT_CACHE_FLAGS_DATA_CACHE |
498 CRAT_CACHE_FLAGS_SIMD_CACHE),
502 /* L2 Data Cache per GPU (Total Tex Cache) */
505 .flags = (CRAT_CACHE_FLAGS_ENABLED |
506 CRAT_CACHE_FLAGS_DATA_CACHE |
507 CRAT_CACHE_FLAGS_SIMD_CACHE),
512 static struct kfd_gpu_cache_info sienna_cichlid_cache_info[] = {
514 /* TCP L1 Cache per CU */
517 .flags = (CRAT_CACHE_FLAGS_ENABLED |
518 CRAT_CACHE_FLAGS_DATA_CACHE |
519 CRAT_CACHE_FLAGS_SIMD_CACHE),
523 /* Scalar L1 Instruction Cache per SQC */
526 .flags = (CRAT_CACHE_FLAGS_ENABLED |
527 CRAT_CACHE_FLAGS_INST_CACHE |
528 CRAT_CACHE_FLAGS_SIMD_CACHE),
532 /* Scalar L1 Data Cache per SQC */
535 .flags = (CRAT_CACHE_FLAGS_ENABLED |
536 CRAT_CACHE_FLAGS_DATA_CACHE |
537 CRAT_CACHE_FLAGS_SIMD_CACHE),
541 /* GL1 Data Cache per SA */
544 .flags = (CRAT_CACHE_FLAGS_ENABLED |
545 CRAT_CACHE_FLAGS_DATA_CACHE |
546 CRAT_CACHE_FLAGS_SIMD_CACHE),
550 /* L2 Data Cache per GPU (Total Tex Cache) */
553 .flags = (CRAT_CACHE_FLAGS_ENABLED |
554 CRAT_CACHE_FLAGS_DATA_CACHE |
555 CRAT_CACHE_FLAGS_SIMD_CACHE),
559 /* L3 Data Cache per GPU */
560 .cache_size = 128*1024,
562 .flags = (CRAT_CACHE_FLAGS_ENABLED |
563 CRAT_CACHE_FLAGS_DATA_CACHE |
564 CRAT_CACHE_FLAGS_SIMD_CACHE),
569 static struct kfd_gpu_cache_info navy_flounder_cache_info[] = {
571 /* TCP L1 Cache per CU */
574 .flags = (CRAT_CACHE_FLAGS_ENABLED |
575 CRAT_CACHE_FLAGS_DATA_CACHE |
576 CRAT_CACHE_FLAGS_SIMD_CACHE),
580 /* Scalar L1 Instruction Cache per SQC */
583 .flags = (CRAT_CACHE_FLAGS_ENABLED |
584 CRAT_CACHE_FLAGS_INST_CACHE |
585 CRAT_CACHE_FLAGS_SIMD_CACHE),
589 /* Scalar L1 Data Cache per SQC */
592 .flags = (CRAT_CACHE_FLAGS_ENABLED |
593 CRAT_CACHE_FLAGS_DATA_CACHE |
594 CRAT_CACHE_FLAGS_SIMD_CACHE),
598 /* GL1 Data Cache per SA */
601 .flags = (CRAT_CACHE_FLAGS_ENABLED |
602 CRAT_CACHE_FLAGS_DATA_CACHE |
603 CRAT_CACHE_FLAGS_SIMD_CACHE),
607 /* L2 Data Cache per GPU (Total Tex Cache) */
610 .flags = (CRAT_CACHE_FLAGS_ENABLED |
611 CRAT_CACHE_FLAGS_DATA_CACHE |
612 CRAT_CACHE_FLAGS_SIMD_CACHE),
616 /* L3 Data Cache per GPU */
617 .cache_size = 96*1024,
619 .flags = (CRAT_CACHE_FLAGS_ENABLED |
620 CRAT_CACHE_FLAGS_DATA_CACHE |
621 CRAT_CACHE_FLAGS_SIMD_CACHE),
626 static struct kfd_gpu_cache_info dimgrey_cavefish_cache_info[] = {
628 /* TCP L1 Cache per CU */
631 .flags = (CRAT_CACHE_FLAGS_ENABLED |
632 CRAT_CACHE_FLAGS_DATA_CACHE |
633 CRAT_CACHE_FLAGS_SIMD_CACHE),
637 /* Scalar L1 Instruction Cache per SQC */
640 .flags = (CRAT_CACHE_FLAGS_ENABLED |
641 CRAT_CACHE_FLAGS_INST_CACHE |
642 CRAT_CACHE_FLAGS_SIMD_CACHE),
646 /* Scalar L1 Data Cache per SQC */
649 .flags = (CRAT_CACHE_FLAGS_ENABLED |
650 CRAT_CACHE_FLAGS_DATA_CACHE |
651 CRAT_CACHE_FLAGS_SIMD_CACHE),
655 /* GL1 Data Cache per SA */
658 .flags = (CRAT_CACHE_FLAGS_ENABLED |
659 CRAT_CACHE_FLAGS_DATA_CACHE |
660 CRAT_CACHE_FLAGS_SIMD_CACHE),
664 /* L2 Data Cache per GPU (Total Tex Cache) */
667 .flags = (CRAT_CACHE_FLAGS_ENABLED |
668 CRAT_CACHE_FLAGS_DATA_CACHE |
669 CRAT_CACHE_FLAGS_SIMD_CACHE),
673 /* L3 Data Cache per GPU */
674 .cache_size = 32*1024,
676 .flags = (CRAT_CACHE_FLAGS_ENABLED |
677 CRAT_CACHE_FLAGS_DATA_CACHE |
678 CRAT_CACHE_FLAGS_SIMD_CACHE),
683 static struct kfd_gpu_cache_info beige_goby_cache_info[] = {
685 /* TCP L1 Cache per CU */
688 .flags = (CRAT_CACHE_FLAGS_ENABLED |
689 CRAT_CACHE_FLAGS_DATA_CACHE |
690 CRAT_CACHE_FLAGS_SIMD_CACHE),
694 /* Scalar L1 Instruction Cache per SQC */
697 .flags = (CRAT_CACHE_FLAGS_ENABLED |
698 CRAT_CACHE_FLAGS_INST_CACHE |
699 CRAT_CACHE_FLAGS_SIMD_CACHE),
703 /* Scalar L1 Data Cache per SQC */
706 .flags = (CRAT_CACHE_FLAGS_ENABLED |
707 CRAT_CACHE_FLAGS_DATA_CACHE |
708 CRAT_CACHE_FLAGS_SIMD_CACHE),
712 /* GL1 Data Cache per SA */
715 .flags = (CRAT_CACHE_FLAGS_ENABLED |
716 CRAT_CACHE_FLAGS_DATA_CACHE |
717 CRAT_CACHE_FLAGS_SIMD_CACHE),
721 /* L2 Data Cache per GPU (Total Tex Cache) */
724 .flags = (CRAT_CACHE_FLAGS_ENABLED |
725 CRAT_CACHE_FLAGS_DATA_CACHE |
726 CRAT_CACHE_FLAGS_SIMD_CACHE),
730 /* L3 Data Cache per GPU */
731 .cache_size = 16*1024,
733 .flags = (CRAT_CACHE_FLAGS_ENABLED |
734 CRAT_CACHE_FLAGS_DATA_CACHE |
735 CRAT_CACHE_FLAGS_SIMD_CACHE),
740 static struct kfd_gpu_cache_info yellow_carp_cache_info[] = {
742 /* TCP L1 Cache per CU */
745 .flags = (CRAT_CACHE_FLAGS_ENABLED |
746 CRAT_CACHE_FLAGS_DATA_CACHE |
747 CRAT_CACHE_FLAGS_SIMD_CACHE),
751 /* Scalar L1 Instruction Cache per SQC */
754 .flags = (CRAT_CACHE_FLAGS_ENABLED |
755 CRAT_CACHE_FLAGS_INST_CACHE |
756 CRAT_CACHE_FLAGS_SIMD_CACHE),
760 /* Scalar L1 Data Cache per SQC */
763 .flags = (CRAT_CACHE_FLAGS_ENABLED |
764 CRAT_CACHE_FLAGS_DATA_CACHE |
765 CRAT_CACHE_FLAGS_SIMD_CACHE),
769 /* GL1 Data Cache per SA */
772 .flags = (CRAT_CACHE_FLAGS_ENABLED |
773 CRAT_CACHE_FLAGS_DATA_CACHE |
774 CRAT_CACHE_FLAGS_SIMD_CACHE),
778 /* L2 Data Cache per GPU (Total Tex Cache) */
781 .flags = (CRAT_CACHE_FLAGS_ENABLED |
782 CRAT_CACHE_FLAGS_DATA_CACHE |
783 CRAT_CACHE_FLAGS_SIMD_CACHE),
788 static struct kfd_gpu_cache_info gfx1037_cache_info[] = {
790 /* TCP L1 Cache per CU */
793 .flags = (CRAT_CACHE_FLAGS_ENABLED |
794 CRAT_CACHE_FLAGS_DATA_CACHE |
795 CRAT_CACHE_FLAGS_SIMD_CACHE),
799 /* Scalar L1 Instruction Cache per SQC */
802 .flags = (CRAT_CACHE_FLAGS_ENABLED |
803 CRAT_CACHE_FLAGS_INST_CACHE |
804 CRAT_CACHE_FLAGS_SIMD_CACHE),
808 /* Scalar L1 Data Cache per SQC */
811 .flags = (CRAT_CACHE_FLAGS_ENABLED |
812 CRAT_CACHE_FLAGS_DATA_CACHE |
813 CRAT_CACHE_FLAGS_SIMD_CACHE),
817 /* GL1 Data Cache per SA */
820 .flags = (CRAT_CACHE_FLAGS_ENABLED |
821 CRAT_CACHE_FLAGS_DATA_CACHE |
822 CRAT_CACHE_FLAGS_SIMD_CACHE),
826 /* L2 Data Cache per GPU (Total Tex Cache) */
829 .flags = (CRAT_CACHE_FLAGS_ENABLED |
830 CRAT_CACHE_FLAGS_DATA_CACHE |
831 CRAT_CACHE_FLAGS_SIMD_CACHE),
836 static struct kfd_gpu_cache_info gc_10_3_6_cache_info[] = {
838 /* TCP L1 Cache per CU */
841 .flags = (CRAT_CACHE_FLAGS_ENABLED |
842 CRAT_CACHE_FLAGS_DATA_CACHE |
843 CRAT_CACHE_FLAGS_SIMD_CACHE),
847 /* Scalar L1 Instruction Cache per SQC */
850 .flags = (CRAT_CACHE_FLAGS_ENABLED |
851 CRAT_CACHE_FLAGS_INST_CACHE |
852 CRAT_CACHE_FLAGS_SIMD_CACHE),
856 /* Scalar L1 Data Cache per SQC */
859 .flags = (CRAT_CACHE_FLAGS_ENABLED |
860 CRAT_CACHE_FLAGS_DATA_CACHE |
861 CRAT_CACHE_FLAGS_SIMD_CACHE),
865 /* GL1 Data Cache per SA */
868 .flags = (CRAT_CACHE_FLAGS_ENABLED |
869 CRAT_CACHE_FLAGS_DATA_CACHE |
870 CRAT_CACHE_FLAGS_SIMD_CACHE),
874 /* L2 Data Cache per GPU (Total Tex Cache) */
877 .flags = (CRAT_CACHE_FLAGS_ENABLED |
878 CRAT_CACHE_FLAGS_DATA_CACHE |
879 CRAT_CACHE_FLAGS_SIMD_CACHE),
884 static struct kfd_gpu_cache_info dummy_cache_info[] = {
886 /* TCP L1 Cache per CU */
889 .flags = (CRAT_CACHE_FLAGS_ENABLED |
890 CRAT_CACHE_FLAGS_DATA_CACHE |
891 CRAT_CACHE_FLAGS_SIMD_CACHE),
895 /* Scalar L1 Instruction Cache per SQC */
898 .flags = (CRAT_CACHE_FLAGS_ENABLED |
899 CRAT_CACHE_FLAGS_INST_CACHE |
900 CRAT_CACHE_FLAGS_SIMD_CACHE),
904 /* Scalar L1 Data Cache per SQC */
907 .flags = (CRAT_CACHE_FLAGS_ENABLED |
908 CRAT_CACHE_FLAGS_DATA_CACHE |
909 CRAT_CACHE_FLAGS_SIMD_CACHE),
913 /* GL1 Data Cache per SA */
916 .flags = (CRAT_CACHE_FLAGS_ENABLED |
917 CRAT_CACHE_FLAGS_DATA_CACHE |
918 CRAT_CACHE_FLAGS_SIMD_CACHE),
922 /* L2 Data Cache per GPU (Total Tex Cache) */
925 .flags = (CRAT_CACHE_FLAGS_ENABLED |
926 CRAT_CACHE_FLAGS_DATA_CACHE |
927 CRAT_CACHE_FLAGS_SIMD_CACHE),
932 static void kfd_populated_cu_info_cpu(struct kfd_topology_device *dev,
933 struct crat_subtype_computeunit *cu)
935 dev->node_props.cpu_cores_count = cu->num_cpu_cores;
936 dev->node_props.cpu_core_id_base = cu->processor_id_low;
937 if (cu->hsa_capability & CRAT_CU_FLAGS_IOMMU_PRESENT)
938 dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
940 pr_debug("CU CPU: cores=%d id_base=%d\n", cu->num_cpu_cores,
941 cu->processor_id_low);
944 static void kfd_populated_cu_info_gpu(struct kfd_topology_device *dev,
945 struct crat_subtype_computeunit *cu)
947 dev->node_props.simd_id_base = cu->processor_id_low;
948 dev->node_props.simd_count = cu->num_simd_cores;
949 dev->node_props.lds_size_in_kb = cu->lds_size_in_kb;
950 dev->node_props.max_waves_per_simd = cu->max_waves_simd;
951 dev->node_props.wave_front_size = cu->wave_front_size;
952 dev->node_props.array_count = cu->array_count;
953 dev->node_props.cu_per_simd_array = cu->num_cu_per_array;
954 dev->node_props.simd_per_cu = cu->num_simd_per_cu;
955 dev->node_props.max_slots_scratch_cu = cu->max_slots_scatch_cu;
956 if (cu->hsa_capability & CRAT_CU_FLAGS_HOT_PLUGGABLE)
957 dev->node_props.capability |= HSA_CAP_HOT_PLUGGABLE;
958 pr_debug("CU GPU: id_base=%d\n", cu->processor_id_low);
961 /* kfd_parse_subtype_cu - parse compute unit subtypes and attach it to correct
962 * topology device present in the device_list
964 static int kfd_parse_subtype_cu(struct crat_subtype_computeunit *cu,
965 struct list_head *device_list)
967 struct kfd_topology_device *dev;
969 pr_debug("Found CU entry in CRAT table with proximity_domain=%d caps=%x\n",
970 cu->proximity_domain, cu->hsa_capability);
971 list_for_each_entry(dev, device_list, list) {
972 if (cu->proximity_domain == dev->proximity_domain) {
973 if (cu->flags & CRAT_CU_FLAGS_CPU_PRESENT)
974 kfd_populated_cu_info_cpu(dev, cu);
976 if (cu->flags & CRAT_CU_FLAGS_GPU_PRESENT)
977 kfd_populated_cu_info_gpu(dev, cu);
985 static struct kfd_mem_properties *
986 find_subtype_mem(uint32_t heap_type, uint32_t flags, uint32_t width,
987 struct kfd_topology_device *dev)
989 struct kfd_mem_properties *props;
991 list_for_each_entry(props, &dev->mem_props, list) {
992 if (props->heap_type == heap_type
993 && props->flags == flags
994 && props->width == width)
1000 /* kfd_parse_subtype_mem - parse memory subtypes and attach it to correct
1001 * topology device present in the device_list
1003 static int kfd_parse_subtype_mem(struct crat_subtype_memory *mem,
1004 struct list_head *device_list)
1006 struct kfd_mem_properties *props;
1007 struct kfd_topology_device *dev;
1009 uint64_t size_in_bytes;
1013 pr_debug("Found memory entry in CRAT table with proximity_domain=%d\n",
1014 mem->proximity_domain);
1015 list_for_each_entry(dev, device_list, list) {
1016 if (mem->proximity_domain == dev->proximity_domain) {
1017 /* We're on GPU node */
1018 if (dev->node_props.cpu_cores_count == 0) {
1020 if (mem->visibility_type == 0)
1022 HSA_MEM_HEAP_TYPE_FB_PRIVATE;
1025 heap_type = mem->visibility_type;
1027 heap_type = HSA_MEM_HEAP_TYPE_SYSTEM;
1029 if (mem->flags & CRAT_MEM_FLAGS_HOT_PLUGGABLE)
1030 flags |= HSA_MEM_FLAGS_HOT_PLUGGABLE;
1031 if (mem->flags & CRAT_MEM_FLAGS_NON_VOLATILE)
1032 flags |= HSA_MEM_FLAGS_NON_VOLATILE;
1035 ((uint64_t)mem->length_high << 32) +
1039 /* Multiple banks of the same type are aggregated into
1040 * one. User mode doesn't care about multiple physical
1041 * memory segments. It's managed as a single virtual
1042 * heap for user mode.
1044 props = find_subtype_mem(heap_type, flags, width, dev);
1046 props->size_in_bytes += size_in_bytes;
1050 props = kfd_alloc_struct(props);
1054 props->heap_type = heap_type;
1055 props->flags = flags;
1056 props->size_in_bytes = size_in_bytes;
1057 props->width = width;
1059 dev->node_props.mem_banks_count++;
1060 list_add_tail(&props->list, &dev->mem_props);
1069 /* kfd_parse_subtype_cache - parse cache subtypes and attach it to correct
1070 * topology device present in the device_list
1072 static int kfd_parse_subtype_cache(struct crat_subtype_cache *cache,
1073 struct list_head *device_list)
1075 struct kfd_cache_properties *props;
1076 struct kfd_topology_device *dev;
1078 uint32_t total_num_of_cu;
1080 id = cache->processor_id_low;
1082 pr_debug("Found cache entry in CRAT table with processor_id=%d\n", id);
1083 list_for_each_entry(dev, device_list, list) {
1084 total_num_of_cu = (dev->node_props.array_count *
1085 dev->node_props.cu_per_simd_array);
1087 /* Cache infomration in CRAT doesn't have proximity_domain
1088 * information as it is associated with a CPU core or GPU
1089 * Compute Unit. So map the cache using CPU core Id or SIMD
1091 * TODO: This works because currently we can safely assume that
1092 * Compute Units are parsed before caches are parsed. In
1093 * future, remove this dependency
1095 if ((id >= dev->node_props.cpu_core_id_base &&
1096 id <= dev->node_props.cpu_core_id_base +
1097 dev->node_props.cpu_cores_count) ||
1098 (id >= dev->node_props.simd_id_base &&
1099 id < dev->node_props.simd_id_base +
1101 props = kfd_alloc_struct(props);
1105 props->processor_id_low = id;
1106 props->cache_level = cache->cache_level;
1107 props->cache_size = cache->cache_size;
1108 props->cacheline_size = cache->cache_line_size;
1109 props->cachelines_per_tag = cache->lines_per_tag;
1110 props->cache_assoc = cache->associativity;
1111 props->cache_latency = cache->cache_latency;
1113 memcpy(props->sibling_map, cache->sibling_map,
1114 CRAT_SIBLINGMAP_SIZE);
1116 /* set the sibling_map_size as 32 for CRAT from ACPI */
1117 props->sibling_map_size = CRAT_SIBLINGMAP_SIZE;
1119 if (cache->flags & CRAT_CACHE_FLAGS_DATA_CACHE)
1120 props->cache_type |= HSA_CACHE_TYPE_DATA;
1121 if (cache->flags & CRAT_CACHE_FLAGS_INST_CACHE)
1122 props->cache_type |= HSA_CACHE_TYPE_INSTRUCTION;
1123 if (cache->flags & CRAT_CACHE_FLAGS_CPU_CACHE)
1124 props->cache_type |= HSA_CACHE_TYPE_CPU;
1125 if (cache->flags & CRAT_CACHE_FLAGS_SIMD_CACHE)
1126 props->cache_type |= HSA_CACHE_TYPE_HSACU;
1128 dev->node_props.caches_count++;
1129 list_add_tail(&props->list, &dev->cache_props);
1138 /* kfd_parse_subtype_iolink - parse iolink subtypes and attach it to correct
1139 * topology device present in the device_list
1141 static int kfd_parse_subtype_iolink(struct crat_subtype_iolink *iolink,
1142 struct list_head *device_list)
1144 struct kfd_iolink_properties *props = NULL, *props2;
1145 struct kfd_topology_device *dev, *to_dev;
1149 id_from = iolink->proximity_domain_from;
1150 id_to = iolink->proximity_domain_to;
1152 pr_debug("Found IO link entry in CRAT table with id_from=%d, id_to %d\n",
1154 list_for_each_entry(dev, device_list, list) {
1155 if (id_from == dev->proximity_domain) {
1156 props = kfd_alloc_struct(props);
1160 props->node_from = id_from;
1161 props->node_to = id_to;
1162 props->ver_maj = iolink->version_major;
1163 props->ver_min = iolink->version_minor;
1164 props->iolink_type = iolink->io_interface_type;
1166 if (props->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS)
1168 else if (props->iolink_type == CRAT_IOLINK_TYPE_XGMI)
1169 props->weight = iolink->weight_xgmi;
1171 props->weight = node_distance(id_from, id_to);
1173 props->min_latency = iolink->minimum_latency;
1174 props->max_latency = iolink->maximum_latency;
1175 props->min_bandwidth = iolink->minimum_bandwidth_mbs;
1176 props->max_bandwidth = iolink->maximum_bandwidth_mbs;
1177 props->rec_transfer_size =
1178 iolink->recommended_transfer_size;
1180 dev->node_props.io_links_count++;
1181 list_add_tail(&props->list, &dev->io_link_props);
1186 /* CPU topology is created before GPUs are detected, so CPU->GPU
1187 * links are not built at that time. If a PCIe type is discovered, it
1188 * means a GPU is detected and we are adding GPU->CPU to the topology.
1189 * At this time, also add the corresponded CPU->GPU link if GPU
1191 * For xGMI, we only added the link with one direction in the crat
1192 * table, add corresponded reversed direction link now.
1194 if (props && (iolink->flags & CRAT_IOLINK_FLAGS_BI_DIRECTIONAL)) {
1195 to_dev = kfd_topology_device_by_proximity_domain_no_lock(id_to);
1198 /* same everything but the other direction */
1199 props2 = kmemdup(props, sizeof(*props2), GFP_KERNEL);
1203 props2->node_from = id_to;
1204 props2->node_to = id_from;
1205 props2->kobj = NULL;
1206 to_dev->node_props.io_links_count++;
1207 list_add_tail(&props2->list, &to_dev->io_link_props);
1213 /* kfd_parse_subtype - parse subtypes and attach it to correct topology device
1214 * present in the device_list
1215 * @sub_type_hdr - subtype section of crat_image
1216 * @device_list - list of topology devices present in this crat_image
1218 static int kfd_parse_subtype(struct crat_subtype_generic *sub_type_hdr,
1219 struct list_head *device_list)
1221 struct crat_subtype_computeunit *cu;
1222 struct crat_subtype_memory *mem;
1223 struct crat_subtype_cache *cache;
1224 struct crat_subtype_iolink *iolink;
1227 switch (sub_type_hdr->type) {
1228 case CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY:
1229 cu = (struct crat_subtype_computeunit *)sub_type_hdr;
1230 ret = kfd_parse_subtype_cu(cu, device_list);
1232 case CRAT_SUBTYPE_MEMORY_AFFINITY:
1233 mem = (struct crat_subtype_memory *)sub_type_hdr;
1234 ret = kfd_parse_subtype_mem(mem, device_list);
1236 case CRAT_SUBTYPE_CACHE_AFFINITY:
1237 cache = (struct crat_subtype_cache *)sub_type_hdr;
1238 ret = kfd_parse_subtype_cache(cache, device_list);
1240 case CRAT_SUBTYPE_TLB_AFFINITY:
1242 * For now, nothing to do here
1244 pr_debug("Found TLB entry in CRAT table (not processing)\n");
1246 case CRAT_SUBTYPE_CCOMPUTE_AFFINITY:
1248 * For now, nothing to do here
1250 pr_debug("Found CCOMPUTE entry in CRAT table (not processing)\n");
1252 case CRAT_SUBTYPE_IOLINK_AFFINITY:
1253 iolink = (struct crat_subtype_iolink *)sub_type_hdr;
1254 ret = kfd_parse_subtype_iolink(iolink, device_list);
1257 pr_warn("Unknown subtype %d in CRAT\n",
1258 sub_type_hdr->type);
1264 /* kfd_parse_crat_table - parse CRAT table. For each node present in CRAT
1265 * create a kfd_topology_device and add in to device_list. Also parse
1266 * CRAT subtypes and attach it to appropriate kfd_topology_device
1267 * @crat_image - input image containing CRAT
1268 * @device_list - [OUT] list of kfd_topology_device generated after
1269 * parsing crat_image
1270 * @proximity_domain - Proximity domain of the first device in the table
1272 * Return - 0 if successful else -ve value
1274 int kfd_parse_crat_table(void *crat_image, struct list_head *device_list,
1275 uint32_t proximity_domain)
1277 struct kfd_topology_device *top_dev = NULL;
1278 struct crat_subtype_generic *sub_type_hdr;
1281 struct crat_header *crat_table = (struct crat_header *)crat_image;
1288 if (!list_empty(device_list)) {
1289 pr_warn("Error device list should be empty\n");
1293 num_nodes = crat_table->num_domains;
1294 image_len = crat_table->length;
1296 pr_debug("Parsing CRAT table with %d nodes\n", num_nodes);
1298 for (node_id = 0; node_id < num_nodes; node_id++) {
1299 top_dev = kfd_create_topology_device(device_list);
1302 top_dev->proximity_domain = proximity_domain++;
1310 memcpy(top_dev->oem_id, crat_table->oem_id, CRAT_OEMID_LENGTH);
1311 memcpy(top_dev->oem_table_id, crat_table->oem_table_id,
1312 CRAT_OEMTABLEID_LENGTH);
1313 top_dev->oem_revision = crat_table->oem_revision;
1315 sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1);
1316 while ((char *)sub_type_hdr + sizeof(struct crat_subtype_generic) <
1317 ((char *)crat_image) + image_len) {
1318 if (sub_type_hdr->flags & CRAT_SUBTYPE_FLAGS_ENABLED) {
1319 ret = kfd_parse_subtype(sub_type_hdr, device_list);
1324 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1325 sub_type_hdr->length);
1330 kfd_release_topology_device_list(device_list);
1336 static int kfd_fill_gpu_cache_info_from_gfx_config(struct kfd_dev *kdev,
1337 struct kfd_gpu_cache_info *pcache_info)
1339 struct amdgpu_device *adev = kdev->adev;
1342 /* TCP L1 Cache per CU */
1343 if (adev->gfx.config.gc_tcp_l1_size) {
1344 pcache_info[i].cache_size = adev->gfx.config.gc_tcp_l1_size;
1345 pcache_info[i].cache_level = 1;
1346 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1347 CRAT_CACHE_FLAGS_DATA_CACHE |
1348 CRAT_CACHE_FLAGS_SIMD_CACHE);
1349 pcache_info[0].num_cu_shared = adev->gfx.config.gc_num_tcp_per_wpg / 2;
1352 /* Scalar L1 Instruction Cache per SQC */
1353 if (adev->gfx.config.gc_l1_instruction_cache_size_per_sqc) {
1354 pcache_info[i].cache_size =
1355 adev->gfx.config.gc_l1_instruction_cache_size_per_sqc;
1356 pcache_info[i].cache_level = 1;
1357 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1358 CRAT_CACHE_FLAGS_INST_CACHE |
1359 CRAT_CACHE_FLAGS_SIMD_CACHE);
1360 pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_sqc_per_wgp * 2;
1363 /* Scalar L1 Data Cache per SQC */
1364 if (adev->gfx.config.gc_l1_data_cache_size_per_sqc) {
1365 pcache_info[i].cache_size = adev->gfx.config.gc_l1_data_cache_size_per_sqc;
1366 pcache_info[i].cache_level = 1;
1367 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1368 CRAT_CACHE_FLAGS_DATA_CACHE |
1369 CRAT_CACHE_FLAGS_SIMD_CACHE);
1370 pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_sqc_per_wgp * 2;
1373 /* GL1 Data Cache per SA */
1374 if (adev->gfx.config.gc_gl1c_per_sa &&
1375 adev->gfx.config.gc_gl1c_size_per_instance) {
1376 pcache_info[i].cache_size = adev->gfx.config.gc_gl1c_per_sa *
1377 adev->gfx.config.gc_gl1c_size_per_instance;
1378 pcache_info[i].cache_level = 1;
1379 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1380 CRAT_CACHE_FLAGS_DATA_CACHE |
1381 CRAT_CACHE_FLAGS_SIMD_CACHE);
1382 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
1385 /* L2 Data Cache per GPU (Total Tex Cache) */
1386 if (adev->gfx.config.gc_gl2c_per_gpu) {
1387 pcache_info[i].cache_size = adev->gfx.config.gc_gl2c_per_gpu;
1388 pcache_info[i].cache_level = 2;
1389 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1390 CRAT_CACHE_FLAGS_DATA_CACHE |
1391 CRAT_CACHE_FLAGS_SIMD_CACHE);
1392 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
1395 /* L3 Data Cache per GPU */
1396 if (adev->gmc.mall_size) {
1397 pcache_info[i].cache_size = adev->gmc.mall_size / 1024;
1398 pcache_info[i].cache_level = 3;
1399 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1400 CRAT_CACHE_FLAGS_DATA_CACHE |
1401 CRAT_CACHE_FLAGS_SIMD_CACHE);
1402 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
1408 int kfd_get_gpu_cache_info(struct kfd_node *kdev, struct kfd_gpu_cache_info **pcache_info)
1410 int num_of_cache_types = 0;
1412 switch (kdev->adev->asic_type) {
1414 *pcache_info = kaveri_cache_info;
1415 num_of_cache_types = ARRAY_SIZE(kaveri_cache_info);
1418 *pcache_info = hawaii_cache_info;
1419 num_of_cache_types = ARRAY_SIZE(hawaii_cache_info);
1422 *pcache_info = carrizo_cache_info;
1423 num_of_cache_types = ARRAY_SIZE(carrizo_cache_info);
1426 *pcache_info = tonga_cache_info;
1427 num_of_cache_types = ARRAY_SIZE(tonga_cache_info);
1430 *pcache_info = fiji_cache_info;
1431 num_of_cache_types = ARRAY_SIZE(fiji_cache_info);
1433 case CHIP_POLARIS10:
1434 *pcache_info = polaris10_cache_info;
1435 num_of_cache_types = ARRAY_SIZE(polaris10_cache_info);
1437 case CHIP_POLARIS11:
1438 *pcache_info = polaris11_cache_info;
1439 num_of_cache_types = ARRAY_SIZE(polaris11_cache_info);
1441 case CHIP_POLARIS12:
1442 *pcache_info = polaris12_cache_info;
1443 num_of_cache_types = ARRAY_SIZE(polaris12_cache_info);
1446 *pcache_info = vegam_cache_info;
1447 num_of_cache_types = ARRAY_SIZE(vegam_cache_info);
1450 switch (KFD_GC_VERSION(kdev)) {
1451 case IP_VERSION(9, 0, 1):
1452 *pcache_info = vega10_cache_info;
1453 num_of_cache_types = ARRAY_SIZE(vega10_cache_info);
1455 case IP_VERSION(9, 2, 1):
1456 *pcache_info = vega12_cache_info;
1457 num_of_cache_types = ARRAY_SIZE(vega12_cache_info);
1459 case IP_VERSION(9, 4, 0):
1460 case IP_VERSION(9, 4, 1):
1461 *pcache_info = vega20_cache_info;
1462 num_of_cache_types = ARRAY_SIZE(vega20_cache_info);
1464 case IP_VERSION(9, 4, 2):
1465 case IP_VERSION(9, 4, 3):
1466 *pcache_info = aldebaran_cache_info;
1467 num_of_cache_types = ARRAY_SIZE(aldebaran_cache_info);
1469 case IP_VERSION(9, 1, 0):
1470 case IP_VERSION(9, 2, 2):
1471 *pcache_info = raven_cache_info;
1472 num_of_cache_types = ARRAY_SIZE(raven_cache_info);
1474 case IP_VERSION(9, 3, 0):
1475 *pcache_info = renoir_cache_info;
1476 num_of_cache_types = ARRAY_SIZE(renoir_cache_info);
1478 case IP_VERSION(10, 1, 10):
1479 case IP_VERSION(10, 1, 2):
1480 case IP_VERSION(10, 1, 3):
1481 case IP_VERSION(10, 1, 4):
1482 *pcache_info = navi10_cache_info;
1483 num_of_cache_types = ARRAY_SIZE(navi10_cache_info);
1485 case IP_VERSION(10, 1, 1):
1486 *pcache_info = navi14_cache_info;
1487 num_of_cache_types = ARRAY_SIZE(navi14_cache_info);
1489 case IP_VERSION(10, 3, 0):
1490 *pcache_info = sienna_cichlid_cache_info;
1491 num_of_cache_types = ARRAY_SIZE(sienna_cichlid_cache_info);
1493 case IP_VERSION(10, 3, 2):
1494 *pcache_info = navy_flounder_cache_info;
1495 num_of_cache_types = ARRAY_SIZE(navy_flounder_cache_info);
1497 case IP_VERSION(10, 3, 4):
1498 *pcache_info = dimgrey_cavefish_cache_info;
1499 num_of_cache_types = ARRAY_SIZE(dimgrey_cavefish_cache_info);
1501 case IP_VERSION(10, 3, 1):
1502 *pcache_info = vangogh_cache_info;
1503 num_of_cache_types = ARRAY_SIZE(vangogh_cache_info);
1505 case IP_VERSION(10, 3, 5):
1506 *pcache_info = beige_goby_cache_info;
1507 num_of_cache_types = ARRAY_SIZE(beige_goby_cache_info);
1509 case IP_VERSION(10, 3, 3):
1510 *pcache_info = yellow_carp_cache_info;
1511 num_of_cache_types = ARRAY_SIZE(yellow_carp_cache_info);
1513 case IP_VERSION(10, 3, 6):
1514 *pcache_info = gc_10_3_6_cache_info;
1515 num_of_cache_types = ARRAY_SIZE(gc_10_3_6_cache_info);
1517 case IP_VERSION(10, 3, 7):
1518 *pcache_info = gfx1037_cache_info;
1519 num_of_cache_types = ARRAY_SIZE(gfx1037_cache_info);
1521 case IP_VERSION(11, 0, 0):
1522 case IP_VERSION(11, 0, 1):
1523 case IP_VERSION(11, 0, 2):
1524 case IP_VERSION(11, 0, 3):
1525 case IP_VERSION(11, 0, 4):
1526 num_of_cache_types =
1527 kfd_fill_gpu_cache_info_from_gfx_config(kdev->kfd, *pcache_info);
1530 *pcache_info = dummy_cache_info;
1531 num_of_cache_types = ARRAY_SIZE(dummy_cache_info);
1532 pr_warn("dummy cache info is used temporarily and real cache info need update later.\n");
1536 return num_of_cache_types;
1539 static bool kfd_ignore_crat(void)
1546 #ifndef KFD_SUPPORT_IOMMU_V2
1556 * kfd_create_crat_image_acpi - Allocates memory for CRAT image and
1557 * copies CRAT from ACPI (if available).
1558 * NOTE: Call kfd_destroy_crat_image to free CRAT image memory
1560 * @crat_image: CRAT read from ACPI. If no CRAT in ACPI then
1561 * crat_image will be NULL
1562 * @size: [OUT] size of crat_image
1564 * Return 0 if successful else return error code
1566 int kfd_create_crat_image_acpi(void **crat_image, size_t *size)
1568 struct acpi_table_header *crat_table;
1578 if (kfd_ignore_crat()) {
1579 pr_info("CRAT table disabled by module option\n");
1583 /* Fetch the CRAT table from ACPI */
1584 status = acpi_get_table(CRAT_SIGNATURE, 0, &crat_table);
1585 if (status == AE_NOT_FOUND) {
1586 pr_info("CRAT table not found\n");
1588 } else if (ACPI_FAILURE(status)) {
1589 const char *err = acpi_format_exception(status);
1591 pr_err("CRAT table error: %s\n", err);
1595 pcrat_image = kvmalloc(crat_table->length, GFP_KERNEL);
1601 memcpy(pcrat_image, crat_table, crat_table->length);
1602 *crat_image = pcrat_image;
1603 *size = crat_table->length;
1605 acpi_put_table(crat_table);
1609 /* Memory required to create Virtual CRAT.
1610 * Since there is no easy way to predict the amount of memory required, the
1611 * following amount is allocated for GPU Virtual CRAT. This is
1612 * expected to cover all known conditions. But to be safe additional check
1613 * is put in the code to ensure we don't overwrite.
1615 #define VCRAT_SIZE_FOR_GPU (4 * PAGE_SIZE)
1617 /* kfd_fill_cu_for_cpu - Fill in Compute info for the given CPU NUMA node
1619 * @numa_node_id: CPU NUMA node id
1620 * @avail_size: Available size in the memory
1621 * @sub_type_hdr: Memory into which compute info will be filled in
1623 * Return 0 if successful else return -ve value
1625 static int kfd_fill_cu_for_cpu(int numa_node_id, int *avail_size,
1626 int proximity_domain,
1627 struct crat_subtype_computeunit *sub_type_hdr)
1629 const struct cpumask *cpumask;
1631 *avail_size -= sizeof(struct crat_subtype_computeunit);
1632 if (*avail_size < 0)
1635 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit));
1637 /* Fill in subtype header data */
1638 sub_type_hdr->type = CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY;
1639 sub_type_hdr->length = sizeof(struct crat_subtype_computeunit);
1640 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
1642 cpumask = cpumask_of_node(numa_node_id);
1644 /* Fill in CU data */
1645 sub_type_hdr->flags |= CRAT_CU_FLAGS_CPU_PRESENT;
1646 sub_type_hdr->proximity_domain = proximity_domain;
1647 sub_type_hdr->processor_id_low = kfd_numa_node_to_apic_id(numa_node_id);
1648 if (sub_type_hdr->processor_id_low == -1)
1651 sub_type_hdr->num_cpu_cores = cpumask_weight(cpumask);
1656 /* kfd_fill_mem_info_for_cpu - Fill in Memory info for the given CPU NUMA node
1658 * @numa_node_id: CPU NUMA node id
1659 * @avail_size: Available size in the memory
1660 * @sub_type_hdr: Memory into which compute info will be filled in
1662 * Return 0 if successful else return -ve value
1664 static int kfd_fill_mem_info_for_cpu(int numa_node_id, int *avail_size,
1665 int proximity_domain,
1666 struct crat_subtype_memory *sub_type_hdr)
1668 uint64_t mem_in_bytes = 0;
1672 *avail_size -= sizeof(struct crat_subtype_memory);
1673 if (*avail_size < 0)
1676 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_memory));
1678 /* Fill in subtype header data */
1679 sub_type_hdr->type = CRAT_SUBTYPE_MEMORY_AFFINITY;
1680 sub_type_hdr->length = sizeof(struct crat_subtype_memory);
1681 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
1683 /* Fill in Memory Subunit data */
1685 /* Unlike si_meminfo, si_meminfo_node is not exported. So
1686 * the following lines are duplicated from si_meminfo_node
1689 pgdat = NODE_DATA(numa_node_id);
1690 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
1691 mem_in_bytes += zone_managed_pages(&pgdat->node_zones[zone_type]);
1692 mem_in_bytes <<= PAGE_SHIFT;
1694 sub_type_hdr->length_low = lower_32_bits(mem_in_bytes);
1695 sub_type_hdr->length_high = upper_32_bits(mem_in_bytes);
1696 sub_type_hdr->proximity_domain = proximity_domain;
1701 #ifdef CONFIG_X86_64
1702 static int kfd_fill_iolink_info_for_cpu(int numa_node_id, int *avail_size,
1703 uint32_t *num_entries,
1704 struct crat_subtype_iolink *sub_type_hdr)
1707 struct cpuinfo_x86 *c = &cpu_data(0);
1710 if (c->x86_vendor == X86_VENDOR_AMD)
1711 link_type = CRAT_IOLINK_TYPE_HYPERTRANSPORT;
1713 link_type = CRAT_IOLINK_TYPE_QPI_1_1;
1717 /* Create IO links from this node to other CPU nodes */
1718 for_each_online_node(nid) {
1719 if (nid == numa_node_id) /* node itself */
1722 *avail_size -= sizeof(struct crat_subtype_iolink);
1723 if (*avail_size < 0)
1726 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
1728 /* Fill in subtype header data */
1729 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
1730 sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
1731 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
1733 /* Fill in IO link data */
1734 sub_type_hdr->proximity_domain_from = numa_node_id;
1735 sub_type_hdr->proximity_domain_to = nid;
1736 sub_type_hdr->io_interface_type = link_type;
1746 /* kfd_create_vcrat_image_cpu - Create Virtual CRAT for CPU
1748 * @pcrat_image: Fill in VCRAT for CPU
1749 * @size: [IN] allocated size of crat_image.
1750 * [OUT] actual size of data filled in crat_image
1752 static int kfd_create_vcrat_image_cpu(void *pcrat_image, size_t *size)
1754 struct crat_header *crat_table = (struct crat_header *)pcrat_image;
1755 struct acpi_table_header *acpi_table;
1757 struct crat_subtype_generic *sub_type_hdr;
1758 int avail_size = *size;
1760 #ifdef CONFIG_X86_64
1761 uint32_t entries = 0;
1768 /* Fill in CRAT Header.
1769 * Modify length and total_entries as subunits are added.
1771 avail_size -= sizeof(struct crat_header);
1775 memset(crat_table, 0, sizeof(struct crat_header));
1776 memcpy(&crat_table->signature, CRAT_SIGNATURE,
1777 sizeof(crat_table->signature));
1778 crat_table->length = sizeof(struct crat_header);
1780 status = acpi_get_table("DSDT", 0, &acpi_table);
1781 if (status != AE_OK)
1782 pr_warn("DSDT table not found for OEM information\n");
1784 crat_table->oem_revision = acpi_table->revision;
1785 memcpy(crat_table->oem_id, acpi_table->oem_id,
1787 memcpy(crat_table->oem_table_id, acpi_table->oem_table_id,
1788 CRAT_OEMTABLEID_LENGTH);
1789 acpi_put_table(acpi_table);
1791 crat_table->total_entries = 0;
1792 crat_table->num_domains = 0;
1794 sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1);
1796 for_each_online_node(numa_node_id) {
1797 if (kfd_numa_node_to_apic_id(numa_node_id) == -1)
1800 /* Fill in Subtype: Compute Unit */
1801 ret = kfd_fill_cu_for_cpu(numa_node_id, &avail_size,
1802 crat_table->num_domains,
1803 (struct crat_subtype_computeunit *)sub_type_hdr);
1806 crat_table->length += sub_type_hdr->length;
1807 crat_table->total_entries++;
1809 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1810 sub_type_hdr->length);
1812 /* Fill in Subtype: Memory */
1813 ret = kfd_fill_mem_info_for_cpu(numa_node_id, &avail_size,
1814 crat_table->num_domains,
1815 (struct crat_subtype_memory *)sub_type_hdr);
1818 crat_table->length += sub_type_hdr->length;
1819 crat_table->total_entries++;
1821 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1822 sub_type_hdr->length);
1824 /* Fill in Subtype: IO Link */
1825 #ifdef CONFIG_X86_64
1826 ret = kfd_fill_iolink_info_for_cpu(numa_node_id, &avail_size,
1828 (struct crat_subtype_iolink *)sub_type_hdr);
1833 crat_table->length += (sub_type_hdr->length * entries);
1834 crat_table->total_entries += entries;
1836 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1837 sub_type_hdr->length * entries);
1840 pr_info("IO link not available for non x86 platforms\n");
1843 crat_table->num_domains++;
1846 /* TODO: Add cache Subtype for CPU.
1847 * Currently, CPU cache information is available in function
1848 * detect_cache_attributes(cpu) defined in the file
1849 * ./arch/x86/kernel/cpu/intel_cacheinfo.c. This function is not
1850 * exported and to get the same information the code needs to be
1854 *size = crat_table->length;
1855 pr_info("Virtual CRAT table created for CPU\n");
1860 static int kfd_fill_gpu_memory_affinity(int *avail_size,
1861 struct kfd_node *kdev, uint8_t type, uint64_t size,
1862 struct crat_subtype_memory *sub_type_hdr,
1863 uint32_t proximity_domain,
1864 const struct kfd_local_mem_info *local_mem_info)
1866 *avail_size -= sizeof(struct crat_subtype_memory);
1867 if (*avail_size < 0)
1870 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_memory));
1871 sub_type_hdr->type = CRAT_SUBTYPE_MEMORY_AFFINITY;
1872 sub_type_hdr->length = sizeof(struct crat_subtype_memory);
1873 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED;
1875 sub_type_hdr->proximity_domain = proximity_domain;
1877 pr_debug("Fill gpu memory affinity - type 0x%x size 0x%llx\n",
1880 sub_type_hdr->length_low = lower_32_bits(size);
1881 sub_type_hdr->length_high = upper_32_bits(size);
1883 sub_type_hdr->width = local_mem_info->vram_width;
1884 sub_type_hdr->visibility_type = type;
1889 #ifdef CONFIG_ACPI_NUMA
1890 static void kfd_find_numa_node_in_srat(struct kfd_node *kdev)
1892 struct acpi_table_header *table_header = NULL;
1893 struct acpi_subtable_header *sub_header = NULL;
1894 unsigned long table_end, subtable_len;
1895 u32 pci_id = pci_domain_nr(kdev->adev->pdev->bus) << 16 |
1896 pci_dev_id(kdev->adev->pdev);
1899 struct acpi_srat_cpu_affinity *cpu;
1900 struct acpi_srat_generic_affinity *gpu;
1901 int pxm = 0, max_pxm = 0;
1902 int numa_node = NUMA_NO_NODE;
1905 /* Fetch the SRAT table from ACPI */
1906 status = acpi_get_table(ACPI_SIG_SRAT, 0, &table_header);
1907 if (status == AE_NOT_FOUND) {
1908 pr_warn("SRAT table not found\n");
1910 } else if (ACPI_FAILURE(status)) {
1911 const char *err = acpi_format_exception(status);
1912 pr_err("SRAT table error: %s\n", err);
1916 table_end = (unsigned long)table_header + table_header->length;
1918 /* Parse all entries looking for a match. */
1919 sub_header = (struct acpi_subtable_header *)
1920 ((unsigned long)table_header +
1921 sizeof(struct acpi_table_srat));
1922 subtable_len = sub_header->length;
1924 while (((unsigned long)sub_header) + subtable_len < table_end) {
1926 * If length is 0, break from this loop to avoid
1929 if (subtable_len == 0) {
1930 pr_err("SRAT invalid zero length\n");
1934 switch (sub_header->type) {
1935 case ACPI_SRAT_TYPE_CPU_AFFINITY:
1936 cpu = (struct acpi_srat_cpu_affinity *)sub_header;
1937 pxm = *((u32 *)cpu->proximity_domain_hi) << 8 |
1938 cpu->proximity_domain_lo;
1942 case ACPI_SRAT_TYPE_GENERIC_AFFINITY:
1943 gpu = (struct acpi_srat_generic_affinity *)sub_header;
1944 bdf = *((u16 *)(&gpu->device_handle[0])) << 16 |
1945 *((u16 *)(&gpu->device_handle[2]));
1946 if (bdf == pci_id) {
1948 numa_node = pxm_to_node(gpu->proximity_domain);
1958 sub_header = (struct acpi_subtable_header *)
1959 ((unsigned long)sub_header + subtable_len);
1960 subtable_len = sub_header->length;
1963 acpi_put_table(table_header);
1965 /* Workaround bad cpu-gpu binding case */
1966 if (found && (numa_node < 0 ||
1967 numa_node > pxm_to_node(max_pxm)))
1970 if (numa_node != NUMA_NO_NODE)
1971 set_dev_node(&kdev->adev->pdev->dev, numa_node);
1975 #define KFD_CRAT_INTRA_SOCKET_WEIGHT 13
1976 #define KFD_CRAT_XGMI_WEIGHT 15
1978 /* kfd_fill_gpu_direct_io_link - Fill in direct io link from GPU
1980 * @avail_size: Available size in the memory
1981 * @kdev - [IN] GPU device
1982 * @sub_type_hdr: Memory into which io link info will be filled in
1983 * @proximity_domain - proximity domain of the GPU node
1985 * Return 0 if successful else return -ve value
1987 static int kfd_fill_gpu_direct_io_link_to_cpu(int *avail_size,
1988 struct kfd_node *kdev,
1989 struct crat_subtype_iolink *sub_type_hdr,
1990 uint32_t proximity_domain)
1992 *avail_size -= sizeof(struct crat_subtype_iolink);
1993 if (*avail_size < 0)
1996 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
1998 /* Fill in subtype header data */
1999 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
2000 sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
2001 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED;
2002 if (kfd_dev_is_large_bar(kdev))
2003 sub_type_hdr->flags |= CRAT_IOLINK_FLAGS_BI_DIRECTIONAL;
2005 /* Fill in IOLINK subtype.
2006 * TODO: Fill-in other fields of iolink subtype
2008 if (kdev->adev->gmc.xgmi.connected_to_cpu ||
2009 (KFD_GC_VERSION(kdev) == IP_VERSION(9, 4, 3) &&
2010 kdev->adev->smuio.funcs->get_pkg_type(kdev->adev) ==
2011 AMDGPU_PKG_TYPE_APU)) {
2012 bool ext_cpu = KFD_GC_VERSION(kdev) != IP_VERSION(9, 4, 3);
2013 int mem_bw = 819200, weight = ext_cpu ? KFD_CRAT_XGMI_WEIGHT :
2014 KFD_CRAT_INTRA_SOCKET_WEIGHT;
2015 uint32_t bandwidth = ext_cpu ? amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(
2016 kdev->adev, NULL, true) : mem_bw;
2019 * with host gpu xgmi link, host can access gpu memory whether
2020 * or not pcie bar type is large, so always create bidirectional
2023 sub_type_hdr->flags |= CRAT_IOLINK_FLAGS_BI_DIRECTIONAL;
2024 sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_XGMI;
2025 sub_type_hdr->weight_xgmi = weight;
2026 sub_type_hdr->minimum_bandwidth_mbs = bandwidth;
2027 sub_type_hdr->maximum_bandwidth_mbs = bandwidth;
2029 sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_PCIEXPRESS;
2030 sub_type_hdr->minimum_bandwidth_mbs =
2031 amdgpu_amdkfd_get_pcie_bandwidth_mbytes(kdev->adev, true);
2032 sub_type_hdr->maximum_bandwidth_mbs =
2033 amdgpu_amdkfd_get_pcie_bandwidth_mbytes(kdev->adev, false);
2036 sub_type_hdr->proximity_domain_from = proximity_domain;
2038 #ifdef CONFIG_ACPI_NUMA
2039 if (kdev->adev->pdev->dev.numa_node == NUMA_NO_NODE &&
2040 num_possible_nodes() > 1)
2041 kfd_find_numa_node_in_srat(kdev);
2044 if (kdev->adev->pdev->dev.numa_node == NUMA_NO_NODE)
2045 sub_type_hdr->proximity_domain_to = 0;
2047 sub_type_hdr->proximity_domain_to = kdev->adev->pdev->dev.numa_node;
2049 sub_type_hdr->proximity_domain_to = 0;
2054 static int kfd_fill_gpu_xgmi_link_to_gpu(int *avail_size,
2055 struct kfd_node *kdev,
2056 struct kfd_node *peer_kdev,
2057 struct crat_subtype_iolink *sub_type_hdr,
2058 uint32_t proximity_domain_from,
2059 uint32_t proximity_domain_to)
2061 bool use_ta_info = kdev->kfd->num_nodes == 1;
2063 *avail_size -= sizeof(struct crat_subtype_iolink);
2064 if (*avail_size < 0)
2067 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
2069 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
2070 sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
2071 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED |
2072 CRAT_IOLINK_FLAGS_BI_DIRECTIONAL;
2074 sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_XGMI;
2075 sub_type_hdr->proximity_domain_from = proximity_domain_from;
2076 sub_type_hdr->proximity_domain_to = proximity_domain_to;
2079 sub_type_hdr->weight_xgmi = KFD_CRAT_XGMI_WEIGHT *
2080 amdgpu_amdkfd_get_xgmi_hops_count(kdev->adev, peer_kdev->adev);
2081 sub_type_hdr->maximum_bandwidth_mbs =
2082 amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(kdev->adev,
2083 peer_kdev->adev, false);
2084 sub_type_hdr->minimum_bandwidth_mbs = sub_type_hdr->maximum_bandwidth_mbs ?
2085 amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(kdev->adev, NULL, true) : 0;
2087 bool is_single_hop = kdev->kfd == peer_kdev->kfd;
2088 int weight = is_single_hop ? KFD_CRAT_INTRA_SOCKET_WEIGHT :
2089 (2 * KFD_CRAT_INTRA_SOCKET_WEIGHT) + KFD_CRAT_XGMI_WEIGHT;
2090 int mem_bw = 819200;
2092 sub_type_hdr->weight_xgmi = weight;
2093 sub_type_hdr->maximum_bandwidth_mbs = is_single_hop ? mem_bw : 0;
2094 sub_type_hdr->minimum_bandwidth_mbs = is_single_hop ? mem_bw : 0;
2100 /* kfd_create_vcrat_image_gpu - Create Virtual CRAT for CPU
2102 * @pcrat_image: Fill in VCRAT for GPU
2103 * @size: [IN] allocated size of crat_image.
2104 * [OUT] actual size of data filled in crat_image
2106 static int kfd_create_vcrat_image_gpu(void *pcrat_image,
2107 size_t *size, struct kfd_node *kdev,
2108 uint32_t proximity_domain)
2110 struct crat_header *crat_table = (struct crat_header *)pcrat_image;
2111 struct crat_subtype_generic *sub_type_hdr;
2112 struct kfd_local_mem_info local_mem_info;
2113 struct kfd_topology_device *peer_dev;
2114 struct crat_subtype_computeunit *cu;
2115 struct kfd_cu_info cu_info;
2116 int avail_size = *size;
2117 uint32_t total_num_of_cu;
2121 if (!pcrat_image || avail_size < VCRAT_SIZE_FOR_GPU)
2124 /* Fill the CRAT Header.
2125 * Modify length and total_entries as subunits are added.
2127 avail_size -= sizeof(struct crat_header);
2131 memset(crat_table, 0, sizeof(struct crat_header));
2133 memcpy(&crat_table->signature, CRAT_SIGNATURE,
2134 sizeof(crat_table->signature));
2135 /* Change length as we add more subtypes*/
2136 crat_table->length = sizeof(struct crat_header);
2137 crat_table->num_domains = 1;
2138 crat_table->total_entries = 0;
2140 /* Fill in Subtype: Compute Unit
2141 * First fill in the sub type header and then sub type data
2143 avail_size -= sizeof(struct crat_subtype_computeunit);
2147 sub_type_hdr = (struct crat_subtype_generic *)(crat_table + 1);
2148 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit));
2150 sub_type_hdr->type = CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY;
2151 sub_type_hdr->length = sizeof(struct crat_subtype_computeunit);
2152 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
2154 /* Fill CU subtype data */
2155 cu = (struct crat_subtype_computeunit *)sub_type_hdr;
2156 cu->flags |= CRAT_CU_FLAGS_GPU_PRESENT;
2157 cu->proximity_domain = proximity_domain;
2159 amdgpu_amdkfd_get_cu_info(kdev->adev, &cu_info);
2160 cu->num_simd_per_cu = cu_info.simd_per_cu;
2161 cu->num_simd_cores = cu_info.simd_per_cu * cu_info.cu_active_number;
2162 cu->max_waves_simd = cu_info.max_waves_per_simd;
2164 cu->wave_front_size = cu_info.wave_front_size;
2165 cu->array_count = cu_info.num_shader_arrays_per_engine *
2166 cu_info.num_shader_engines;
2167 total_num_of_cu = (cu->array_count * cu_info.num_cu_per_sh);
2168 cu->processor_id_low = get_and_inc_gpu_processor_id(total_num_of_cu);
2169 cu->num_cu_per_array = cu_info.num_cu_per_sh;
2170 cu->max_slots_scatch_cu = cu_info.max_scratch_slots_per_cu;
2171 cu->num_banks = cu_info.num_shader_engines;
2172 cu->lds_size_in_kb = cu_info.lds_size;
2174 cu->hsa_capability = 0;
2176 /* Check if this node supports IOMMU. During parsing this flag will
2177 * translate to HSA_CAP_ATS_PRESENT
2179 if (!kfd_iommu_check_device(kdev->kfd))
2180 cu->hsa_capability |= CRAT_CU_FLAGS_IOMMU_PRESENT;
2182 crat_table->length += sub_type_hdr->length;
2183 crat_table->total_entries++;
2185 /* Fill in Subtype: Memory. Only on systems with large BAR (no
2186 * private FB), report memory as public. On other systems
2187 * report the total FB size (public+private) as a single
2190 local_mem_info = kdev->local_mem_info;
2191 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
2192 sub_type_hdr->length);
2195 local_mem_info.local_mem_size_private = 0;
2197 if (local_mem_info.local_mem_size_private == 0)
2198 ret = kfd_fill_gpu_memory_affinity(&avail_size,
2199 kdev, HSA_MEM_HEAP_TYPE_FB_PUBLIC,
2200 local_mem_info.local_mem_size_public,
2201 (struct crat_subtype_memory *)sub_type_hdr,
2205 ret = kfd_fill_gpu_memory_affinity(&avail_size,
2206 kdev, HSA_MEM_HEAP_TYPE_FB_PRIVATE,
2207 local_mem_info.local_mem_size_public +
2208 local_mem_info.local_mem_size_private,
2209 (struct crat_subtype_memory *)sub_type_hdr,
2215 crat_table->length += sizeof(struct crat_subtype_memory);
2216 crat_table->total_entries++;
2218 /* Fill in Subtype: IO_LINKS
2219 * Only direct links are added here which is Link from GPU to
2220 * its NUMA node. Indirect links are added by userspace.
2222 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
2223 sub_type_hdr->length);
2224 ret = kfd_fill_gpu_direct_io_link_to_cpu(&avail_size, kdev,
2225 (struct crat_subtype_iolink *)sub_type_hdr, proximity_domain);
2230 crat_table->length += sub_type_hdr->length;
2231 crat_table->total_entries++;
2234 /* Fill in Subtype: IO_LINKS
2235 * Direct links from GPU to other GPUs through xGMI.
2236 * We will loop GPUs that already be processed (with lower value
2237 * of proximity_domain), add the link for the GPUs with same
2238 * hive id (from this GPU to other GPU) . The reversed iolink
2239 * (from other GPU to this GPU) will be added
2240 * in kfd_parse_subtype_iolink.
2242 if (kdev->kfd->hive_id) {
2243 for (nid = 0; nid < proximity_domain; ++nid) {
2244 peer_dev = kfd_topology_device_by_proximity_domain_no_lock(nid);
2247 if (peer_dev->gpu->kfd->hive_id != kdev->kfd->hive_id)
2249 sub_type_hdr = (typeof(sub_type_hdr))(
2250 (char *)sub_type_hdr +
2251 sizeof(struct crat_subtype_iolink));
2252 ret = kfd_fill_gpu_xgmi_link_to_gpu(
2253 &avail_size, kdev, peer_dev->gpu,
2254 (struct crat_subtype_iolink *)sub_type_hdr,
2255 proximity_domain, nid);
2258 crat_table->length += sub_type_hdr->length;
2259 crat_table->total_entries++;
2262 *size = crat_table->length;
2263 pr_info("Virtual CRAT table created for GPU\n");
2268 /* kfd_create_crat_image_virtual - Allocates memory for CRAT image and
2269 * creates a Virtual CRAT (VCRAT) image
2271 * NOTE: Call kfd_destroy_crat_image to free CRAT image memory
2273 * @crat_image: VCRAT image created because ACPI does not have a
2274 * CRAT for this device
2275 * @size: [OUT] size of virtual crat_image
2276 * @flags: COMPUTE_UNIT_CPU - Create VCRAT for CPU device
2277 * COMPUTE_UNIT_GPU - Create VCRAT for GPU
2278 * (COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU) - Create VCRAT for APU
2279 * -- this option is not currently implemented.
2280 * The assumption is that all AMD APUs will have CRAT
2281 * @kdev: Valid kfd_node required if flags contain COMPUTE_UNIT_GPU
2283 * Return 0 if successful else return -ve value
2285 int kfd_create_crat_image_virtual(void **crat_image, size_t *size,
2286 int flags, struct kfd_node *kdev,
2287 uint32_t proximity_domain)
2289 void *pcrat_image = NULL;
2290 int ret = 0, num_nodes;
2298 /* Allocate the CPU Virtual CRAT size based on the number of online
2299 * nodes. Allocate VCRAT_SIZE_FOR_GPU for GPU virtual CRAT image.
2300 * This should cover all the current conditions. A check is put not
2301 * to overwrite beyond allocated size for GPUs
2304 case COMPUTE_UNIT_CPU:
2305 num_nodes = num_online_nodes();
2306 dyn_size = sizeof(struct crat_header) +
2307 num_nodes * (sizeof(struct crat_subtype_computeunit) +
2308 sizeof(struct crat_subtype_memory) +
2309 (num_nodes - 1) * sizeof(struct crat_subtype_iolink));
2310 pcrat_image = kvmalloc(dyn_size, GFP_KERNEL);
2314 pr_debug("CRAT size is %ld", dyn_size);
2315 ret = kfd_create_vcrat_image_cpu(pcrat_image, size);
2317 case COMPUTE_UNIT_GPU:
2320 pcrat_image = kvmalloc(VCRAT_SIZE_FOR_GPU, GFP_KERNEL);
2323 *size = VCRAT_SIZE_FOR_GPU;
2324 ret = kfd_create_vcrat_image_gpu(pcrat_image, size, kdev,
2327 case (COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU):
2330 pr_err("VCRAT not implemented for APU\n");
2337 *crat_image = pcrat_image;
2339 kvfree(pcrat_image);
2345 /* kfd_destroy_crat_image
2347 * @crat_image: [IN] - crat_image from kfd_create_crat_image_xxx(..)
2350 void kfd_destroy_crat_image(void *crat_image)