1 // SPDX-License-Identifier: GPL-2.0 OR MIT
2 /*
3 * Copyright 2015-2022 Advanced Micro Devices, Inc.
4 *
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:
11 *
12 * The above copyright notice and this permission notice shall be included in
13 * all copies or substantial portions of the Software.
14 *
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.
22 */
23
24 #include <linux/pci.h>
25 #include <linux/acpi.h>
26 #include "kfd_crat.h"
27 #include "kfd_priv.h"
28 #include "kfd_topology.h"
29 #include "amdgpu.h"
30 #include "amdgpu_amdkfd.h"
31 #include "amdgpu_xgmi.h"
32
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
36 * used in the CRAT.
37 */
38 static uint32_t gpu_processor_id_low = 0x80001000;
39
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
42 * masked off
43 */
get_and_inc_gpu_processor_id(unsigned int total_cu_count)44 static inline unsigned int get_and_inc_gpu_processor_id(
45 unsigned int total_cu_count)
46 {
47 int current_id = gpu_processor_id_low;
48
49 gpu_processor_id_low += total_cu_count;
50 return current_id;
51 }
52
53
54 static struct kfd_gpu_cache_info kaveri_cache_info[] = {
55 {
56 /* TCP L1 Cache per CU */
57 .cache_size = 16,
58 .cache_level = 1,
59 .cache_line_size = 64,
60 .flags = (CRAT_CACHE_FLAGS_ENABLED |
61 CRAT_CACHE_FLAGS_DATA_CACHE |
62 CRAT_CACHE_FLAGS_SIMD_CACHE),
63 .num_cu_shared = 1,
64 },
65 {
66 /* Scalar L1 Instruction Cache (in SQC module) per bank */
67 .cache_size = 16,
68 .cache_level = 1,
69 .cache_line_size = 64,
70 .flags = (CRAT_CACHE_FLAGS_ENABLED |
71 CRAT_CACHE_FLAGS_INST_CACHE |
72 CRAT_CACHE_FLAGS_SIMD_CACHE),
73 .num_cu_shared = 2,
74 },
75 {
76 /* Scalar L1 Data Cache (in SQC module) per bank */
77 .cache_size = 8,
78 .cache_level = 1,
79 .cache_line_size = 64,
80 .flags = (CRAT_CACHE_FLAGS_ENABLED |
81 CRAT_CACHE_FLAGS_DATA_CACHE |
82 CRAT_CACHE_FLAGS_SIMD_CACHE),
83 .num_cu_shared = 2,
84 },
85
86 /* TODO: Add L2 Cache information */
87 };
88
89
90 static struct kfd_gpu_cache_info carrizo_cache_info[] = {
91 {
92 /* TCP L1 Cache per CU */
93 .cache_size = 16,
94 .cache_level = 1,
95 .cache_line_size = 64,
96 .flags = (CRAT_CACHE_FLAGS_ENABLED |
97 CRAT_CACHE_FLAGS_DATA_CACHE |
98 CRAT_CACHE_FLAGS_SIMD_CACHE),
99 .num_cu_shared = 1,
100 },
101 {
102 /* Scalar L1 Instruction Cache (in SQC module) per bank */
103 .cache_size = 32,
104 .cache_level = 1,
105 .cache_line_size = 64,
106 .flags = (CRAT_CACHE_FLAGS_ENABLED |
107 CRAT_CACHE_FLAGS_INST_CACHE |
108 CRAT_CACHE_FLAGS_SIMD_CACHE),
109 .num_cu_shared = 4,
110 },
111 {
112 /* Scalar L1 Data Cache (in SQC module) per bank. */
113 .cache_size = 16,
114 .cache_level = 1,
115 .cache_line_size = 64,
116 .flags = (CRAT_CACHE_FLAGS_ENABLED |
117 CRAT_CACHE_FLAGS_DATA_CACHE |
118 CRAT_CACHE_FLAGS_SIMD_CACHE),
119 .num_cu_shared = 4,
120 },
121
122 /* TODO: Add L2 Cache information */
123 };
124
125 #define hawaii_cache_info kaveri_cache_info
126 #define tonga_cache_info carrizo_cache_info
127 #define fiji_cache_info carrizo_cache_info
128 #define polaris10_cache_info carrizo_cache_info
129 #define polaris11_cache_info carrizo_cache_info
130 #define polaris12_cache_info carrizo_cache_info
131 #define vegam_cache_info carrizo_cache_info
132
133 /* NOTE: L1 cache information has been updated and L2/L3
134 * cache information has been added for Vega10 and
135 * newer ASICs. The unit for cache_size is KiB.
136 * In future, check & update cache details
137 * for every new ASIC is required.
138 */
139
140 static struct kfd_gpu_cache_info vega10_cache_info[] = {
141 {
142 /* TCP L1 Cache per CU */
143 .cache_size = 16,
144 .cache_level = 1,
145 .cache_line_size = 64,
146 .flags = (CRAT_CACHE_FLAGS_ENABLED |
147 CRAT_CACHE_FLAGS_DATA_CACHE |
148 CRAT_CACHE_FLAGS_SIMD_CACHE),
149 .num_cu_shared = 1,
150 },
151 {
152 /* Scalar L1 Instruction Cache per SQC */
153 .cache_size = 32,
154 .cache_level = 1,
155 .cache_line_size = 64,
156 .flags = (CRAT_CACHE_FLAGS_ENABLED |
157 CRAT_CACHE_FLAGS_INST_CACHE |
158 CRAT_CACHE_FLAGS_SIMD_CACHE),
159 .num_cu_shared = 3,
160 },
161 {
162 /* Scalar L1 Data Cache per SQC */
163 .cache_size = 16,
164 .cache_level = 1,
165 .cache_line_size = 64,
166 .flags = (CRAT_CACHE_FLAGS_ENABLED |
167 CRAT_CACHE_FLAGS_DATA_CACHE |
168 CRAT_CACHE_FLAGS_SIMD_CACHE),
169 .num_cu_shared = 3,
170 },
171 {
172 /* L2 Data Cache per GPU (Total Tex Cache) */
173 .cache_size = 4096,
174 .cache_level = 2,
175 .cache_line_size = 64,
176 .flags = (CRAT_CACHE_FLAGS_ENABLED |
177 CRAT_CACHE_FLAGS_DATA_CACHE |
178 CRAT_CACHE_FLAGS_SIMD_CACHE),
179 .num_cu_shared = 16,
180 },
181 };
182
183 static struct kfd_gpu_cache_info raven_cache_info[] = {
184 {
185 /* TCP L1 Cache per CU */
186 .cache_size = 16,
187 .cache_level = 1,
188 .cache_line_size = 64,
189 .flags = (CRAT_CACHE_FLAGS_ENABLED |
190 CRAT_CACHE_FLAGS_DATA_CACHE |
191 CRAT_CACHE_FLAGS_SIMD_CACHE),
192 .num_cu_shared = 1,
193 },
194 {
195 /* Scalar L1 Instruction Cache per SQC */
196 .cache_size = 32,
197 .cache_level = 1,
198 .cache_line_size = 64,
199 .flags = (CRAT_CACHE_FLAGS_ENABLED |
200 CRAT_CACHE_FLAGS_INST_CACHE |
201 CRAT_CACHE_FLAGS_SIMD_CACHE),
202 .num_cu_shared = 3,
203 },
204 {
205 /* Scalar L1 Data Cache per SQC */
206 .cache_size = 16,
207 .cache_level = 1,
208 .cache_line_size = 64,
209 .flags = (CRAT_CACHE_FLAGS_ENABLED |
210 CRAT_CACHE_FLAGS_DATA_CACHE |
211 CRAT_CACHE_FLAGS_SIMD_CACHE),
212 .num_cu_shared = 3,
213 },
214 {
215 /* L2 Data Cache per GPU (Total Tex Cache) */
216 .cache_size = 1024,
217 .cache_level = 2,
218 .cache_line_size = 64,
219 .flags = (CRAT_CACHE_FLAGS_ENABLED |
220 CRAT_CACHE_FLAGS_DATA_CACHE |
221 CRAT_CACHE_FLAGS_SIMD_CACHE),
222 .num_cu_shared = 11,
223 },
224 };
225
226 static struct kfd_gpu_cache_info renoir_cache_info[] = {
227 {
228 /* TCP L1 Cache per CU */
229 .cache_size = 16,
230 .cache_level = 1,
231 .cache_line_size = 64,
232 .flags = (CRAT_CACHE_FLAGS_ENABLED |
233 CRAT_CACHE_FLAGS_DATA_CACHE |
234 CRAT_CACHE_FLAGS_SIMD_CACHE),
235 .num_cu_shared = 1,
236 },
237 {
238 /* Scalar L1 Instruction Cache per SQC */
239 .cache_size = 32,
240 .cache_level = 1,
241 .cache_line_size = 64,
242 .flags = (CRAT_CACHE_FLAGS_ENABLED |
243 CRAT_CACHE_FLAGS_INST_CACHE |
244 CRAT_CACHE_FLAGS_SIMD_CACHE),
245 .num_cu_shared = 3,
246 },
247 {
248 /* Scalar L1 Data Cache per SQC */
249 .cache_size = 16,
250 .cache_level = 1,
251 .cache_line_size = 64,
252 .flags = (CRAT_CACHE_FLAGS_ENABLED |
253 CRAT_CACHE_FLAGS_DATA_CACHE |
254 CRAT_CACHE_FLAGS_SIMD_CACHE),
255 .num_cu_shared = 3,
256 },
257 {
258 /* L2 Data Cache per GPU (Total Tex Cache) */
259 .cache_size = 1024,
260 .cache_level = 2,
261 .cache_line_size = 64,
262 .flags = (CRAT_CACHE_FLAGS_ENABLED |
263 CRAT_CACHE_FLAGS_DATA_CACHE |
264 CRAT_CACHE_FLAGS_SIMD_CACHE),
265 .num_cu_shared = 8,
266 },
267 };
268
269 static struct kfd_gpu_cache_info vega12_cache_info[] = {
270 {
271 /* TCP L1 Cache per CU */
272 .cache_size = 16,
273 .cache_level = 1,
274 .cache_line_size = 64,
275 .flags = (CRAT_CACHE_FLAGS_ENABLED |
276 CRAT_CACHE_FLAGS_DATA_CACHE |
277 CRAT_CACHE_FLAGS_SIMD_CACHE),
278 .num_cu_shared = 1,
279 },
280 {
281 /* Scalar L1 Instruction Cache per SQC */
282 .cache_size = 32,
283 .cache_level = 1,
284 .cache_line_size = 64,
285 .flags = (CRAT_CACHE_FLAGS_ENABLED |
286 CRAT_CACHE_FLAGS_INST_CACHE |
287 CRAT_CACHE_FLAGS_SIMD_CACHE),
288 .num_cu_shared = 3,
289 },
290 {
291 /* Scalar L1 Data Cache per SQC */
292 .cache_size = 16,
293 .cache_level = 1,
294 .cache_line_size = 64,
295 .flags = (CRAT_CACHE_FLAGS_ENABLED |
296 CRAT_CACHE_FLAGS_DATA_CACHE |
297 CRAT_CACHE_FLAGS_SIMD_CACHE),
298 .num_cu_shared = 3,
299 },
300 {
301 /* L2 Data Cache per GPU (Total Tex Cache) */
302 .cache_size = 2048,
303 .cache_level = 2,
304 .cache_line_size = 64,
305 .flags = (CRAT_CACHE_FLAGS_ENABLED |
306 CRAT_CACHE_FLAGS_DATA_CACHE |
307 CRAT_CACHE_FLAGS_SIMD_CACHE),
308 .num_cu_shared = 5,
309 },
310 };
311
312 static struct kfd_gpu_cache_info vega20_cache_info[] = {
313 {
314 /* TCP L1 Cache per CU */
315 .cache_size = 16,
316 .cache_level = 1,
317 .cache_line_size = 64,
318 .flags = (CRAT_CACHE_FLAGS_ENABLED |
319 CRAT_CACHE_FLAGS_DATA_CACHE |
320 CRAT_CACHE_FLAGS_SIMD_CACHE),
321 .num_cu_shared = 1,
322 },
323 {
324 /* Scalar L1 Instruction Cache per SQC */
325 .cache_size = 32,
326 .cache_level = 1,
327 .cache_line_size = 64,
328 .flags = (CRAT_CACHE_FLAGS_ENABLED |
329 CRAT_CACHE_FLAGS_INST_CACHE |
330 CRAT_CACHE_FLAGS_SIMD_CACHE),
331 .num_cu_shared = 3,
332 },
333 {
334 /* Scalar L1 Data Cache per SQC */
335 .cache_size = 16,
336 .cache_level = 1,
337 .cache_line_size = 64,
338 .flags = (CRAT_CACHE_FLAGS_ENABLED |
339 CRAT_CACHE_FLAGS_DATA_CACHE |
340 CRAT_CACHE_FLAGS_SIMD_CACHE),
341 .num_cu_shared = 3,
342 },
343 {
344 /* L2 Data Cache per GPU (Total Tex Cache) */
345 .cache_size = 8192,
346 .cache_level = 2,
347 .cache_line_size = 64,
348 .flags = (CRAT_CACHE_FLAGS_ENABLED |
349 CRAT_CACHE_FLAGS_DATA_CACHE |
350 CRAT_CACHE_FLAGS_SIMD_CACHE),
351 .num_cu_shared = 16,
352 },
353 };
354
355 static struct kfd_gpu_cache_info aldebaran_cache_info[] = {
356 {
357 /* TCP L1 Cache per CU */
358 .cache_size = 16,
359 .cache_level = 1,
360 .cache_line_size = 64,
361 .flags = (CRAT_CACHE_FLAGS_ENABLED |
362 CRAT_CACHE_FLAGS_DATA_CACHE |
363 CRAT_CACHE_FLAGS_SIMD_CACHE),
364 .num_cu_shared = 1,
365 },
366 {
367 /* Scalar L1 Instruction Cache per SQC */
368 .cache_size = 32,
369 .cache_level = 1,
370 .cache_line_size = 64,
371 .flags = (CRAT_CACHE_FLAGS_ENABLED |
372 CRAT_CACHE_FLAGS_INST_CACHE |
373 CRAT_CACHE_FLAGS_SIMD_CACHE),
374 .num_cu_shared = 2,
375 },
376 {
377 /* Scalar L1 Data Cache per SQC */
378 .cache_size = 16,
379 .cache_level = 1,
380 .cache_line_size = 64,
381 .flags = (CRAT_CACHE_FLAGS_ENABLED |
382 CRAT_CACHE_FLAGS_DATA_CACHE |
383 CRAT_CACHE_FLAGS_SIMD_CACHE),
384 .num_cu_shared = 2,
385 },
386 {
387 /* L2 Data Cache per GPU (Total Tex Cache) */
388 .cache_size = 8192,
389 .cache_level = 2,
390 .cache_line_size = 128,
391 .flags = (CRAT_CACHE_FLAGS_ENABLED |
392 CRAT_CACHE_FLAGS_DATA_CACHE |
393 CRAT_CACHE_FLAGS_SIMD_CACHE),
394 .num_cu_shared = 14,
395 },
396 };
397
398 static struct kfd_gpu_cache_info navi10_cache_info[] = {
399 {
400 /* TCP L1 Cache per CU */
401 .cache_size = 16,
402 .cache_level = 1,
403 .cache_line_size = 128,
404 .flags = (CRAT_CACHE_FLAGS_ENABLED |
405 CRAT_CACHE_FLAGS_DATA_CACHE |
406 CRAT_CACHE_FLAGS_SIMD_CACHE),
407 .num_cu_shared = 1,
408 },
409 {
410 /* Scalar L1 Instruction Cache per SQC */
411 .cache_size = 32,
412 .cache_level = 1,
413 .cache_line_size = 64,
414 .flags = (CRAT_CACHE_FLAGS_ENABLED |
415 CRAT_CACHE_FLAGS_INST_CACHE |
416 CRAT_CACHE_FLAGS_SIMD_CACHE),
417 .num_cu_shared = 2,
418 },
419 {
420 /* Scalar L1 Data Cache per SQC */
421 .cache_size = 16,
422 .cache_level = 1,
423 .cache_line_size = 64,
424 .flags = (CRAT_CACHE_FLAGS_ENABLED |
425 CRAT_CACHE_FLAGS_DATA_CACHE |
426 CRAT_CACHE_FLAGS_SIMD_CACHE),
427 .num_cu_shared = 2,
428 },
429 {
430 /* GL1 Data Cache per SA */
431 .cache_size = 128,
432 .cache_level = 1,
433 .cache_line_size = 128,
434 .flags = (CRAT_CACHE_FLAGS_ENABLED |
435 CRAT_CACHE_FLAGS_DATA_CACHE |
436 CRAT_CACHE_FLAGS_SIMD_CACHE),
437 .num_cu_shared = 10,
438 },
439 {
440 /* L2 Data Cache per GPU (Total Tex Cache) */
441 .cache_size = 4096,
442 .cache_level = 2,
443 .cache_line_size = 128,
444 .flags = (CRAT_CACHE_FLAGS_ENABLED |
445 CRAT_CACHE_FLAGS_DATA_CACHE |
446 CRAT_CACHE_FLAGS_SIMD_CACHE),
447 .num_cu_shared = 10,
448 },
449 };
450
451 static struct kfd_gpu_cache_info vangogh_cache_info[] = {
452 {
453 /* TCP L1 Cache per CU */
454 .cache_size = 16,
455 .cache_level = 1,
456 .cache_line_size = 128,
457 .flags = (CRAT_CACHE_FLAGS_ENABLED |
458 CRAT_CACHE_FLAGS_DATA_CACHE |
459 CRAT_CACHE_FLAGS_SIMD_CACHE),
460 .num_cu_shared = 1,
461 },
462 {
463 /* Scalar L1 Instruction Cache per SQC */
464 .cache_size = 32,
465 .cache_level = 1,
466 .cache_line_size = 64,
467 .flags = (CRAT_CACHE_FLAGS_ENABLED |
468 CRAT_CACHE_FLAGS_INST_CACHE |
469 CRAT_CACHE_FLAGS_SIMD_CACHE),
470 .num_cu_shared = 2,
471 },
472 {
473 /* Scalar L1 Data Cache per SQC */
474 .cache_size = 16,
475 .cache_level = 1,
476 .cache_line_size = 64,
477 .flags = (CRAT_CACHE_FLAGS_ENABLED |
478 CRAT_CACHE_FLAGS_DATA_CACHE |
479 CRAT_CACHE_FLAGS_SIMD_CACHE),
480 .num_cu_shared = 2,
481 },
482 {
483 /* GL1 Data Cache per SA */
484 .cache_size = 128,
485 .cache_level = 1,
486 .cache_line_size = 128,
487 .flags = (CRAT_CACHE_FLAGS_ENABLED |
488 CRAT_CACHE_FLAGS_DATA_CACHE |
489 CRAT_CACHE_FLAGS_SIMD_CACHE),
490 .num_cu_shared = 8,
491 },
492 {
493 /* L2 Data Cache per GPU (Total Tex Cache) */
494 .cache_size = 1024,
495 .cache_level = 2,
496 .cache_line_size = 128,
497 .flags = (CRAT_CACHE_FLAGS_ENABLED |
498 CRAT_CACHE_FLAGS_DATA_CACHE |
499 CRAT_CACHE_FLAGS_SIMD_CACHE),
500 .num_cu_shared = 8,
501 },
502 };
503
504 static struct kfd_gpu_cache_info navi14_cache_info[] = {
505 {
506 /* TCP L1 Cache per CU */
507 .cache_size = 16,
508 .cache_level = 1,
509 .cache_line_size = 128,
510 .flags = (CRAT_CACHE_FLAGS_ENABLED |
511 CRAT_CACHE_FLAGS_DATA_CACHE |
512 CRAT_CACHE_FLAGS_SIMD_CACHE),
513 .num_cu_shared = 1,
514 },
515 {
516 /* Scalar L1 Instruction Cache per SQC */
517 .cache_size = 32,
518 .cache_level = 1,
519 .cache_line_size = 64,
520 .flags = (CRAT_CACHE_FLAGS_ENABLED |
521 CRAT_CACHE_FLAGS_INST_CACHE |
522 CRAT_CACHE_FLAGS_SIMD_CACHE),
523 .num_cu_shared = 2,
524 },
525 {
526 /* Scalar L1 Data Cache per SQC */
527 .cache_size = 16,
528 .cache_level = 1,
529 .cache_line_size = 64,
530 .flags = (CRAT_CACHE_FLAGS_ENABLED |
531 CRAT_CACHE_FLAGS_DATA_CACHE |
532 CRAT_CACHE_FLAGS_SIMD_CACHE),
533 .num_cu_shared = 2,
534 },
535 {
536 /* GL1 Data Cache per SA */
537 .cache_size = 128,
538 .cache_level = 1,
539 .cache_line_size = 128,
540 .flags = (CRAT_CACHE_FLAGS_ENABLED |
541 CRAT_CACHE_FLAGS_DATA_CACHE |
542 CRAT_CACHE_FLAGS_SIMD_CACHE),
543 .num_cu_shared = 12,
544 },
545 {
546 /* L2 Data Cache per GPU (Total Tex Cache) */
547 .cache_size = 2048,
548 .cache_level = 2,
549 .cache_line_size = 128,
550 .flags = (CRAT_CACHE_FLAGS_ENABLED |
551 CRAT_CACHE_FLAGS_DATA_CACHE |
552 CRAT_CACHE_FLAGS_SIMD_CACHE),
553 .num_cu_shared = 12,
554 },
555 };
556
557 static struct kfd_gpu_cache_info sienna_cichlid_cache_info[] = {
558 {
559 /* TCP L1 Cache per CU */
560 .cache_size = 16,
561 .cache_level = 1,
562 .cache_line_size = 128,
563 .flags = (CRAT_CACHE_FLAGS_ENABLED |
564 CRAT_CACHE_FLAGS_DATA_CACHE |
565 CRAT_CACHE_FLAGS_SIMD_CACHE),
566 .num_cu_shared = 1,
567 },
568 {
569 /* Scalar L1 Instruction Cache per SQC */
570 .cache_size = 32,
571 .cache_level = 1,
572 .cache_line_size = 64,
573 .flags = (CRAT_CACHE_FLAGS_ENABLED |
574 CRAT_CACHE_FLAGS_INST_CACHE |
575 CRAT_CACHE_FLAGS_SIMD_CACHE),
576 .num_cu_shared = 2,
577 },
578 {
579 /* Scalar L1 Data Cache per SQC */
580 .cache_size = 16,
581 .cache_level = 1,
582 .cache_line_size = 64,
583 .flags = (CRAT_CACHE_FLAGS_ENABLED |
584 CRAT_CACHE_FLAGS_DATA_CACHE |
585 CRAT_CACHE_FLAGS_SIMD_CACHE),
586 .num_cu_shared = 2,
587 },
588 {
589 /* GL1 Data Cache per SA */
590 .cache_size = 128,
591 .cache_level = 1,
592 .cache_line_size = 128,
593 .flags = (CRAT_CACHE_FLAGS_ENABLED |
594 CRAT_CACHE_FLAGS_DATA_CACHE |
595 CRAT_CACHE_FLAGS_SIMD_CACHE),
596 .num_cu_shared = 10,
597 },
598 {
599 /* L2 Data Cache per GPU (Total Tex Cache) */
600 .cache_size = 4096,
601 .cache_level = 2,
602 .cache_line_size = 128,
603 .flags = (CRAT_CACHE_FLAGS_ENABLED |
604 CRAT_CACHE_FLAGS_DATA_CACHE |
605 CRAT_CACHE_FLAGS_SIMD_CACHE),
606 .num_cu_shared = 10,
607 },
608 {
609 /* L3 Data Cache per GPU */
610 .cache_size = 128*1024,
611 .cache_level = 3,
612 .cache_line_size = 64,
613 .flags = (CRAT_CACHE_FLAGS_ENABLED |
614 CRAT_CACHE_FLAGS_DATA_CACHE |
615 CRAT_CACHE_FLAGS_SIMD_CACHE),
616 .num_cu_shared = 10,
617 },
618 };
619
620 static struct kfd_gpu_cache_info navy_flounder_cache_info[] = {
621 {
622 /* TCP L1 Cache per CU */
623 .cache_size = 16,
624 .cache_level = 1,
625 .cache_line_size = 128,
626 .flags = (CRAT_CACHE_FLAGS_ENABLED |
627 CRAT_CACHE_FLAGS_DATA_CACHE |
628 CRAT_CACHE_FLAGS_SIMD_CACHE),
629 .num_cu_shared = 1,
630 },
631 {
632 /* Scalar L1 Instruction Cache per SQC */
633 .cache_size = 32,
634 .cache_level = 1,
635 .cache_line_size = 64,
636 .flags = (CRAT_CACHE_FLAGS_ENABLED |
637 CRAT_CACHE_FLAGS_INST_CACHE |
638 CRAT_CACHE_FLAGS_SIMD_CACHE),
639 .num_cu_shared = 2,
640 },
641 {
642 /* Scalar L1 Data Cache per SQC */
643 .cache_size = 16,
644 .cache_level = 1,
645 .cache_line_size = 64,
646 .flags = (CRAT_CACHE_FLAGS_ENABLED |
647 CRAT_CACHE_FLAGS_DATA_CACHE |
648 CRAT_CACHE_FLAGS_SIMD_CACHE),
649 .num_cu_shared = 2,
650 },
651 {
652 /* GL1 Data Cache per SA */
653 .cache_size = 128,
654 .cache_level = 1,
655 .cache_line_size = 128,
656 .flags = (CRAT_CACHE_FLAGS_ENABLED |
657 CRAT_CACHE_FLAGS_DATA_CACHE |
658 CRAT_CACHE_FLAGS_SIMD_CACHE),
659 .num_cu_shared = 10,
660 },
661 {
662 /* L2 Data Cache per GPU (Total Tex Cache) */
663 .cache_size = 3072,
664 .cache_level = 2,
665 .cache_line_size = 128,
666 .flags = (CRAT_CACHE_FLAGS_ENABLED |
667 CRAT_CACHE_FLAGS_DATA_CACHE |
668 CRAT_CACHE_FLAGS_SIMD_CACHE),
669 .num_cu_shared = 10,
670 },
671 {
672 /* L3 Data Cache per GPU */
673 .cache_size = 96*1024,
674 .cache_level = 3,
675 .cache_line_size = 64,
676 .flags = (CRAT_CACHE_FLAGS_ENABLED |
677 CRAT_CACHE_FLAGS_DATA_CACHE |
678 CRAT_CACHE_FLAGS_SIMD_CACHE),
679 .num_cu_shared = 10,
680 },
681 };
682
683 static struct kfd_gpu_cache_info dimgrey_cavefish_cache_info[] = {
684 {
685 /* TCP L1 Cache per CU */
686 .cache_size = 16,
687 .cache_level = 1,
688 .cache_line_size = 128,
689 .flags = (CRAT_CACHE_FLAGS_ENABLED |
690 CRAT_CACHE_FLAGS_DATA_CACHE |
691 CRAT_CACHE_FLAGS_SIMD_CACHE),
692 .num_cu_shared = 1,
693 },
694 {
695 /* Scalar L1 Instruction Cache per SQC */
696 .cache_size = 32,
697 .cache_level = 1,
698 .cache_line_size = 64,
699 .flags = (CRAT_CACHE_FLAGS_ENABLED |
700 CRAT_CACHE_FLAGS_INST_CACHE |
701 CRAT_CACHE_FLAGS_SIMD_CACHE),
702 .num_cu_shared = 2,
703 },
704 {
705 /* Scalar L1 Data Cache per SQC */
706 .cache_size = 16,
707 .cache_level = 1,
708 .cache_line_size = 64,
709 .flags = (CRAT_CACHE_FLAGS_ENABLED |
710 CRAT_CACHE_FLAGS_DATA_CACHE |
711 CRAT_CACHE_FLAGS_SIMD_CACHE),
712 .num_cu_shared = 2,
713 },
714 {
715 /* GL1 Data Cache per SA */
716 .cache_size = 128,
717 .cache_level = 1,
718 .cache_line_size = 128,
719 .flags = (CRAT_CACHE_FLAGS_ENABLED |
720 CRAT_CACHE_FLAGS_DATA_CACHE |
721 CRAT_CACHE_FLAGS_SIMD_CACHE),
722 .num_cu_shared = 8,
723 },
724 {
725 /* L2 Data Cache per GPU (Total Tex Cache) */
726 .cache_size = 2048,
727 .cache_level = 2,
728 .cache_line_size = 128,
729 .flags = (CRAT_CACHE_FLAGS_ENABLED |
730 CRAT_CACHE_FLAGS_DATA_CACHE |
731 CRAT_CACHE_FLAGS_SIMD_CACHE),
732 .num_cu_shared = 8,
733 },
734 {
735 /* L3 Data Cache per GPU */
736 .cache_size = 32*1024,
737 .cache_level = 3,
738 .cache_line_size = 64,
739 .flags = (CRAT_CACHE_FLAGS_ENABLED |
740 CRAT_CACHE_FLAGS_DATA_CACHE |
741 CRAT_CACHE_FLAGS_SIMD_CACHE),
742 .num_cu_shared = 8,
743 },
744 };
745
746 static struct kfd_gpu_cache_info beige_goby_cache_info[] = {
747 {
748 /* TCP L1 Cache per CU */
749 .cache_size = 16,
750 .cache_level = 1,
751 .cache_line_size = 128,
752 .flags = (CRAT_CACHE_FLAGS_ENABLED |
753 CRAT_CACHE_FLAGS_DATA_CACHE |
754 CRAT_CACHE_FLAGS_SIMD_CACHE),
755 .num_cu_shared = 1,
756 },
757 {
758 /* Scalar L1 Instruction Cache per SQC */
759 .cache_size = 32,
760 .cache_level = 1,
761 .cache_line_size = 64,
762 .flags = (CRAT_CACHE_FLAGS_ENABLED |
763 CRAT_CACHE_FLAGS_INST_CACHE |
764 CRAT_CACHE_FLAGS_SIMD_CACHE),
765 .num_cu_shared = 2,
766 },
767 {
768 /* Scalar L1 Data Cache per SQC */
769 .cache_size = 16,
770 .cache_level = 1,
771 .cache_line_size = 64,
772 .flags = (CRAT_CACHE_FLAGS_ENABLED |
773 CRAT_CACHE_FLAGS_DATA_CACHE |
774 CRAT_CACHE_FLAGS_SIMD_CACHE),
775 .num_cu_shared = 2,
776 },
777 {
778 /* GL1 Data Cache per SA */
779 .cache_size = 128,
780 .cache_level = 1,
781 .cache_line_size = 128,
782 .flags = (CRAT_CACHE_FLAGS_ENABLED |
783 CRAT_CACHE_FLAGS_DATA_CACHE |
784 CRAT_CACHE_FLAGS_SIMD_CACHE),
785 .num_cu_shared = 8,
786 },
787 {
788 /* L2 Data Cache per GPU (Total Tex Cache) */
789 .cache_size = 1024,
790 .cache_level = 2,
791 .cache_line_size = 128,
792 .flags = (CRAT_CACHE_FLAGS_ENABLED |
793 CRAT_CACHE_FLAGS_DATA_CACHE |
794 CRAT_CACHE_FLAGS_SIMD_CACHE),
795 .num_cu_shared = 8,
796 },
797 {
798 /* L3 Data Cache per GPU */
799 .cache_size = 16*1024,
800 .cache_level = 3,
801 .cache_line_size = 64,
802 .flags = (CRAT_CACHE_FLAGS_ENABLED |
803 CRAT_CACHE_FLAGS_DATA_CACHE |
804 CRAT_CACHE_FLAGS_SIMD_CACHE),
805 .num_cu_shared = 8,
806 },
807 };
808
809 static struct kfd_gpu_cache_info yellow_carp_cache_info[] = {
810 {
811 /* TCP L1 Cache per CU */
812 .cache_size = 16,
813 .cache_level = 1,
814 .cache_line_size = 128,
815 .flags = (CRAT_CACHE_FLAGS_ENABLED |
816 CRAT_CACHE_FLAGS_DATA_CACHE |
817 CRAT_CACHE_FLAGS_SIMD_CACHE),
818 .num_cu_shared = 1,
819 },
820 {
821 /* Scalar L1 Instruction Cache per SQC */
822 .cache_size = 32,
823 .cache_level = 1,
824 .cache_line_size = 64,
825 .flags = (CRAT_CACHE_FLAGS_ENABLED |
826 CRAT_CACHE_FLAGS_INST_CACHE |
827 CRAT_CACHE_FLAGS_SIMD_CACHE),
828 .num_cu_shared = 2,
829 },
830 {
831 /* Scalar L1 Data Cache per SQC */
832 .cache_size = 16,
833 .cache_level = 1,
834 .cache_line_size = 64,
835 .flags = (CRAT_CACHE_FLAGS_ENABLED |
836 CRAT_CACHE_FLAGS_DATA_CACHE |
837 CRAT_CACHE_FLAGS_SIMD_CACHE),
838 .num_cu_shared = 2,
839 },
840 {
841 /* GL1 Data Cache per SA */
842 .cache_size = 128,
843 .cache_level = 1,
844 .cache_line_size = 128,
845 .flags = (CRAT_CACHE_FLAGS_ENABLED |
846 CRAT_CACHE_FLAGS_DATA_CACHE |
847 CRAT_CACHE_FLAGS_SIMD_CACHE),
848 .num_cu_shared = 6,
849 },
850 {
851 /* L2 Data Cache per GPU (Total Tex Cache) */
852 .cache_size = 2048,
853 .cache_level = 2,
854 .cache_line_size = 128,
855 .flags = (CRAT_CACHE_FLAGS_ENABLED |
856 CRAT_CACHE_FLAGS_DATA_CACHE |
857 CRAT_CACHE_FLAGS_SIMD_CACHE),
858 .num_cu_shared = 6,
859 },
860 };
861
862 static struct kfd_gpu_cache_info gfx1037_cache_info[] = {
863 {
864 /* TCP L1 Cache per CU */
865 .cache_size = 16,
866 .cache_level = 1,
867 .cache_line_size = 128,
868 .flags = (CRAT_CACHE_FLAGS_ENABLED |
869 CRAT_CACHE_FLAGS_DATA_CACHE |
870 CRAT_CACHE_FLAGS_SIMD_CACHE),
871 .num_cu_shared = 1,
872 },
873 {
874 /* Scalar L1 Instruction Cache per SQC */
875 .cache_size = 32,
876 .cache_level = 1,
877 .cache_line_size = 64,
878 .flags = (CRAT_CACHE_FLAGS_ENABLED |
879 CRAT_CACHE_FLAGS_INST_CACHE |
880 CRAT_CACHE_FLAGS_SIMD_CACHE),
881 .num_cu_shared = 2,
882 },
883 {
884 /* Scalar L1 Data Cache per SQC */
885 .cache_size = 16,
886 .cache_level = 1,
887 .cache_line_size = 64,
888 .flags = (CRAT_CACHE_FLAGS_ENABLED |
889 CRAT_CACHE_FLAGS_DATA_CACHE |
890 CRAT_CACHE_FLAGS_SIMD_CACHE),
891 .num_cu_shared = 2,
892 },
893 {
894 /* GL1 Data Cache per SA */
895 .cache_size = 128,
896 .cache_level = 1,
897 .cache_line_size = 128,
898 .flags = (CRAT_CACHE_FLAGS_ENABLED |
899 CRAT_CACHE_FLAGS_DATA_CACHE |
900 CRAT_CACHE_FLAGS_SIMD_CACHE),
901 .num_cu_shared = 2,
902 },
903 {
904 /* L2 Data Cache per GPU (Total Tex Cache) */
905 .cache_size = 256,
906 .cache_level = 2,
907 .cache_line_size = 128,
908 .flags = (CRAT_CACHE_FLAGS_ENABLED |
909 CRAT_CACHE_FLAGS_DATA_CACHE |
910 CRAT_CACHE_FLAGS_SIMD_CACHE),
911 .num_cu_shared = 2,
912 },
913 };
914
915 static struct kfd_gpu_cache_info gc_10_3_6_cache_info[] = {
916 {
917 /* TCP L1 Cache per CU */
918 .cache_size = 16,
919 .cache_level = 1,
920 .cache_line_size = 128,
921 .flags = (CRAT_CACHE_FLAGS_ENABLED |
922 CRAT_CACHE_FLAGS_DATA_CACHE |
923 CRAT_CACHE_FLAGS_SIMD_CACHE),
924 .num_cu_shared = 1,
925 },
926 {
927 /* Scalar L1 Instruction Cache per SQC */
928 .cache_size = 32,
929 .cache_level = 1,
930 .cache_line_size = 64,
931 .flags = (CRAT_CACHE_FLAGS_ENABLED |
932 CRAT_CACHE_FLAGS_INST_CACHE |
933 CRAT_CACHE_FLAGS_SIMD_CACHE),
934 .num_cu_shared = 2,
935 },
936 {
937 /* Scalar L1 Data Cache per SQC */
938 .cache_size = 16,
939 .cache_level = 1,
940 .cache_line_size = 64,
941 .flags = (CRAT_CACHE_FLAGS_ENABLED |
942 CRAT_CACHE_FLAGS_DATA_CACHE |
943 CRAT_CACHE_FLAGS_SIMD_CACHE),
944 .num_cu_shared = 2,
945 },
946 {
947 /* GL1 Data Cache per SA */
948 .cache_size = 128,
949 .cache_level = 1,
950 .cache_line_size = 128,
951 .flags = (CRAT_CACHE_FLAGS_ENABLED |
952 CRAT_CACHE_FLAGS_DATA_CACHE |
953 CRAT_CACHE_FLAGS_SIMD_CACHE),
954 .num_cu_shared = 2,
955 },
956 {
957 /* L2 Data Cache per GPU (Total Tex Cache) */
958 .cache_size = 256,
959 .cache_level = 2,
960 .cache_line_size = 128,
961 .flags = (CRAT_CACHE_FLAGS_ENABLED |
962 CRAT_CACHE_FLAGS_DATA_CACHE |
963 CRAT_CACHE_FLAGS_SIMD_CACHE),
964 .num_cu_shared = 2,
965 },
966 };
967
968 static struct kfd_gpu_cache_info dummy_cache_info[] = {
969 {
970 /* TCP L1 Cache per CU */
971 .cache_size = 16,
972 .cache_level = 1,
973 .cache_line_size = 64,
974 .flags = (CRAT_CACHE_FLAGS_ENABLED |
975 CRAT_CACHE_FLAGS_DATA_CACHE |
976 CRAT_CACHE_FLAGS_SIMD_CACHE),
977 .num_cu_shared = 1,
978 },
979 {
980 /* Scalar L1 Instruction Cache per SQC */
981 .cache_size = 32,
982 .cache_level = 1,
983 .cache_line_size = 64,
984 .flags = (CRAT_CACHE_FLAGS_ENABLED |
985 CRAT_CACHE_FLAGS_INST_CACHE |
986 CRAT_CACHE_FLAGS_SIMD_CACHE),
987 .num_cu_shared = 2,
988 },
989 {
990 /* Scalar L1 Data Cache per SQC */
991 .cache_size = 16,
992 .cache_level = 1,
993 .cache_line_size = 64,
994 .flags = (CRAT_CACHE_FLAGS_ENABLED |
995 CRAT_CACHE_FLAGS_DATA_CACHE |
996 CRAT_CACHE_FLAGS_SIMD_CACHE),
997 .num_cu_shared = 2,
998 },
999 {
1000 /* GL1 Data Cache per SA */
1001 .cache_size = 128,
1002 .cache_level = 1,
1003 .cache_line_size = 64,
1004 .flags = (CRAT_CACHE_FLAGS_ENABLED |
1005 CRAT_CACHE_FLAGS_DATA_CACHE |
1006 CRAT_CACHE_FLAGS_SIMD_CACHE),
1007 .num_cu_shared = 6,
1008 },
1009 {
1010 /* L2 Data Cache per GPU (Total Tex Cache) */
1011 .cache_size = 2048,
1012 .cache_level = 2,
1013 .cache_line_size = 64,
1014 .flags = (CRAT_CACHE_FLAGS_ENABLED |
1015 CRAT_CACHE_FLAGS_DATA_CACHE |
1016 CRAT_CACHE_FLAGS_SIMD_CACHE),
1017 .num_cu_shared = 6,
1018 },
1019 };
1020
kfd_populated_cu_info_cpu(struct kfd_topology_device * dev,struct crat_subtype_computeunit * cu)1021 static void kfd_populated_cu_info_cpu(struct kfd_topology_device *dev,
1022 struct crat_subtype_computeunit *cu)
1023 {
1024 dev->node_props.cpu_cores_count = cu->num_cpu_cores;
1025 dev->node_props.cpu_core_id_base = cu->processor_id_low;
1026 if (cu->hsa_capability & CRAT_CU_FLAGS_IOMMU_PRESENT)
1027 dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
1028
1029 pr_debug("CU CPU: cores=%d id_base=%d\n", cu->num_cpu_cores,
1030 cu->processor_id_low);
1031 }
1032
kfd_populated_cu_info_gpu(struct kfd_topology_device * dev,struct crat_subtype_computeunit * cu)1033 static void kfd_populated_cu_info_gpu(struct kfd_topology_device *dev,
1034 struct crat_subtype_computeunit *cu)
1035 {
1036 dev->node_props.simd_id_base = cu->processor_id_low;
1037 dev->node_props.simd_count = cu->num_simd_cores;
1038 dev->node_props.lds_size_in_kb = cu->lds_size_in_kb;
1039 dev->node_props.max_waves_per_simd = cu->max_waves_simd;
1040 dev->node_props.wave_front_size = cu->wave_front_size;
1041 dev->node_props.array_count = cu->array_count;
1042 dev->node_props.cu_per_simd_array = cu->num_cu_per_array;
1043 dev->node_props.simd_per_cu = cu->num_simd_per_cu;
1044 dev->node_props.max_slots_scratch_cu = cu->max_slots_scatch_cu;
1045 if (cu->hsa_capability & CRAT_CU_FLAGS_HOT_PLUGGABLE)
1046 dev->node_props.capability |= HSA_CAP_HOT_PLUGGABLE;
1047 pr_debug("CU GPU: id_base=%d\n", cu->processor_id_low);
1048 }
1049
1050 /* kfd_parse_subtype_cu - parse compute unit subtypes and attach it to correct
1051 * topology device present in the device_list
1052 */
kfd_parse_subtype_cu(struct crat_subtype_computeunit * cu,struct list_head * device_list)1053 static int kfd_parse_subtype_cu(struct crat_subtype_computeunit *cu,
1054 struct list_head *device_list)
1055 {
1056 struct kfd_topology_device *dev;
1057
1058 pr_debug("Found CU entry in CRAT table with proximity_domain=%d caps=%x\n",
1059 cu->proximity_domain, cu->hsa_capability);
1060 list_for_each_entry(dev, device_list, list) {
1061 if (cu->proximity_domain == dev->proximity_domain) {
1062 if (cu->flags & CRAT_CU_FLAGS_CPU_PRESENT)
1063 kfd_populated_cu_info_cpu(dev, cu);
1064
1065 if (cu->flags & CRAT_CU_FLAGS_GPU_PRESENT)
1066 kfd_populated_cu_info_gpu(dev, cu);
1067 break;
1068 }
1069 }
1070
1071 return 0;
1072 }
1073
1074 static struct kfd_mem_properties *
find_subtype_mem(uint32_t heap_type,uint32_t flags,uint32_t width,struct kfd_topology_device * dev)1075 find_subtype_mem(uint32_t heap_type, uint32_t flags, uint32_t width,
1076 struct kfd_topology_device *dev)
1077 {
1078 struct kfd_mem_properties *props;
1079
1080 list_for_each_entry(props, &dev->mem_props, list) {
1081 if (props->heap_type == heap_type
1082 && props->flags == flags
1083 && props->width == width)
1084 return props;
1085 }
1086
1087 return NULL;
1088 }
1089 /* kfd_parse_subtype_mem - parse memory subtypes and attach it to correct
1090 * topology device present in the device_list
1091 */
kfd_parse_subtype_mem(struct crat_subtype_memory * mem,struct list_head * device_list)1092 static int kfd_parse_subtype_mem(struct crat_subtype_memory *mem,
1093 struct list_head *device_list)
1094 {
1095 struct kfd_mem_properties *props;
1096 struct kfd_topology_device *dev;
1097 uint32_t heap_type;
1098 uint64_t size_in_bytes;
1099 uint32_t flags = 0;
1100 uint32_t width;
1101
1102 pr_debug("Found memory entry in CRAT table with proximity_domain=%d\n",
1103 mem->proximity_domain);
1104 list_for_each_entry(dev, device_list, list) {
1105 if (mem->proximity_domain == dev->proximity_domain) {
1106 /* We're on GPU node */
1107 if (dev->node_props.cpu_cores_count == 0) {
1108 /* APU */
1109 if (mem->visibility_type == 0)
1110 heap_type =
1111 HSA_MEM_HEAP_TYPE_FB_PRIVATE;
1112 /* dGPU */
1113 else
1114 heap_type = mem->visibility_type;
1115 } else
1116 heap_type = HSA_MEM_HEAP_TYPE_SYSTEM;
1117
1118 if (mem->flags & CRAT_MEM_FLAGS_HOT_PLUGGABLE)
1119 flags |= HSA_MEM_FLAGS_HOT_PLUGGABLE;
1120 if (mem->flags & CRAT_MEM_FLAGS_NON_VOLATILE)
1121 flags |= HSA_MEM_FLAGS_NON_VOLATILE;
1122
1123 size_in_bytes =
1124 ((uint64_t)mem->length_high << 32) +
1125 mem->length_low;
1126 width = mem->width;
1127
1128 /* Multiple banks of the same type are aggregated into
1129 * one. User mode doesn't care about multiple physical
1130 * memory segments. It's managed as a single virtual
1131 * heap for user mode.
1132 */
1133 props = find_subtype_mem(heap_type, flags, width, dev);
1134 if (props) {
1135 props->size_in_bytes += size_in_bytes;
1136 break;
1137 }
1138
1139 props = kfd_alloc_struct(props);
1140 if (!props)
1141 return -ENOMEM;
1142
1143 props->heap_type = heap_type;
1144 props->flags = flags;
1145 props->size_in_bytes = size_in_bytes;
1146 props->width = width;
1147
1148 dev->node_props.mem_banks_count++;
1149 list_add_tail(&props->list, &dev->mem_props);
1150
1151 break;
1152 }
1153 }
1154
1155 return 0;
1156 }
1157
1158 /* kfd_parse_subtype_cache - parse cache subtypes and attach it to correct
1159 * topology device present in the device_list
1160 */
kfd_parse_subtype_cache(struct crat_subtype_cache * cache,struct list_head * device_list)1161 static int kfd_parse_subtype_cache(struct crat_subtype_cache *cache,
1162 struct list_head *device_list)
1163 {
1164 struct kfd_cache_properties *props;
1165 struct kfd_topology_device *dev;
1166 uint32_t id;
1167 uint32_t total_num_of_cu;
1168
1169 id = cache->processor_id_low;
1170
1171 pr_debug("Found cache entry in CRAT table with processor_id=%d\n", id);
1172 list_for_each_entry(dev, device_list, list) {
1173 total_num_of_cu = (dev->node_props.array_count *
1174 dev->node_props.cu_per_simd_array);
1175
1176 /* Cache infomration in CRAT doesn't have proximity_domain
1177 * information as it is associated with a CPU core or GPU
1178 * Compute Unit. So map the cache using CPU core Id or SIMD
1179 * (GPU) ID.
1180 * TODO: This works because currently we can safely assume that
1181 * Compute Units are parsed before caches are parsed. In
1182 * future, remove this dependency
1183 */
1184 if ((id >= dev->node_props.cpu_core_id_base &&
1185 id <= dev->node_props.cpu_core_id_base +
1186 dev->node_props.cpu_cores_count) ||
1187 (id >= dev->node_props.simd_id_base &&
1188 id < dev->node_props.simd_id_base +
1189 total_num_of_cu)) {
1190 props = kfd_alloc_struct(props);
1191 if (!props)
1192 return -ENOMEM;
1193
1194 props->processor_id_low = id;
1195 props->cache_level = cache->cache_level;
1196 props->cache_size = cache->cache_size;
1197 props->cacheline_size = cache->cache_line_size;
1198 props->cachelines_per_tag = cache->lines_per_tag;
1199 props->cache_assoc = cache->associativity;
1200 props->cache_latency = cache->cache_latency;
1201
1202 memcpy(props->sibling_map, cache->sibling_map,
1203 CRAT_SIBLINGMAP_SIZE);
1204
1205 /* set the sibling_map_size as 32 for CRAT from ACPI */
1206 props->sibling_map_size = CRAT_SIBLINGMAP_SIZE;
1207
1208 if (cache->flags & CRAT_CACHE_FLAGS_DATA_CACHE)
1209 props->cache_type |= HSA_CACHE_TYPE_DATA;
1210 if (cache->flags & CRAT_CACHE_FLAGS_INST_CACHE)
1211 props->cache_type |= HSA_CACHE_TYPE_INSTRUCTION;
1212 if (cache->flags & CRAT_CACHE_FLAGS_CPU_CACHE)
1213 props->cache_type |= HSA_CACHE_TYPE_CPU;
1214 if (cache->flags & CRAT_CACHE_FLAGS_SIMD_CACHE)
1215 props->cache_type |= HSA_CACHE_TYPE_HSACU;
1216
1217 dev->node_props.caches_count++;
1218 list_add_tail(&props->list, &dev->cache_props);
1219
1220 break;
1221 }
1222 }
1223
1224 return 0;
1225 }
1226
1227 /* kfd_parse_subtype_iolink - parse iolink subtypes and attach it to correct
1228 * topology device present in the device_list
1229 */
kfd_parse_subtype_iolink(struct crat_subtype_iolink * iolink,struct list_head * device_list)1230 static int kfd_parse_subtype_iolink(struct crat_subtype_iolink *iolink,
1231 struct list_head *device_list)
1232 {
1233 struct kfd_iolink_properties *props = NULL, *props2;
1234 struct kfd_topology_device *dev, *to_dev;
1235 uint32_t id_from;
1236 uint32_t id_to;
1237
1238 id_from = iolink->proximity_domain_from;
1239 id_to = iolink->proximity_domain_to;
1240
1241 pr_debug("Found IO link entry in CRAT table with id_from=%d, id_to %d\n",
1242 id_from, id_to);
1243 list_for_each_entry(dev, device_list, list) {
1244 if (id_from == dev->proximity_domain) {
1245 props = kfd_alloc_struct(props);
1246 if (!props)
1247 return -ENOMEM;
1248
1249 props->node_from = id_from;
1250 props->node_to = id_to;
1251 props->ver_maj = iolink->version_major;
1252 props->ver_min = iolink->version_minor;
1253 props->iolink_type = iolink->io_interface_type;
1254
1255 if (props->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS)
1256 props->weight = 20;
1257 else if (props->iolink_type == CRAT_IOLINK_TYPE_XGMI)
1258 props->weight = iolink->weight_xgmi;
1259 else
1260 props->weight = node_distance(id_from, id_to);
1261
1262 props->min_latency = iolink->minimum_latency;
1263 props->max_latency = iolink->maximum_latency;
1264 props->min_bandwidth = iolink->minimum_bandwidth_mbs;
1265 props->max_bandwidth = iolink->maximum_bandwidth_mbs;
1266 props->rec_transfer_size =
1267 iolink->recommended_transfer_size;
1268
1269 dev->node_props.io_links_count++;
1270 list_add_tail(&props->list, &dev->io_link_props);
1271 break;
1272 }
1273 }
1274
1275 /* CPU topology is created before GPUs are detected, so CPU->GPU
1276 * links are not built at that time. If a PCIe type is discovered, it
1277 * means a GPU is detected and we are adding GPU->CPU to the topology.
1278 * At this time, also add the corresponded CPU->GPU link if GPU
1279 * is large bar.
1280 * For xGMI, we only added the link with one direction in the crat
1281 * table, add corresponded reversed direction link now.
1282 */
1283 if (props && (iolink->flags & CRAT_IOLINK_FLAGS_BI_DIRECTIONAL)) {
1284 to_dev = kfd_topology_device_by_proximity_domain_no_lock(id_to);
1285 if (!to_dev)
1286 return -ENODEV;
1287 /* same everything but the other direction */
1288 props2 = kmemdup(props, sizeof(*props2), GFP_KERNEL);
1289 if (!props2)
1290 return -ENOMEM;
1291
1292 props2->node_from = id_to;
1293 props2->node_to = id_from;
1294 props2->kobj = NULL;
1295 to_dev->node_props.io_links_count++;
1296 list_add_tail(&props2->list, &to_dev->io_link_props);
1297 }
1298
1299 return 0;
1300 }
1301
1302 /* kfd_parse_subtype - parse subtypes and attach it to correct topology device
1303 * present in the device_list
1304 * @sub_type_hdr - subtype section of crat_image
1305 * @device_list - list of topology devices present in this crat_image
1306 */
kfd_parse_subtype(struct crat_subtype_generic * sub_type_hdr,struct list_head * device_list)1307 static int kfd_parse_subtype(struct crat_subtype_generic *sub_type_hdr,
1308 struct list_head *device_list)
1309 {
1310 struct crat_subtype_computeunit *cu;
1311 struct crat_subtype_memory *mem;
1312 struct crat_subtype_cache *cache;
1313 struct crat_subtype_iolink *iolink;
1314 int ret = 0;
1315
1316 switch (sub_type_hdr->type) {
1317 case CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY:
1318 cu = (struct crat_subtype_computeunit *)sub_type_hdr;
1319 ret = kfd_parse_subtype_cu(cu, device_list);
1320 break;
1321 case CRAT_SUBTYPE_MEMORY_AFFINITY:
1322 mem = (struct crat_subtype_memory *)sub_type_hdr;
1323 ret = kfd_parse_subtype_mem(mem, device_list);
1324 break;
1325 case CRAT_SUBTYPE_CACHE_AFFINITY:
1326 cache = (struct crat_subtype_cache *)sub_type_hdr;
1327 ret = kfd_parse_subtype_cache(cache, device_list);
1328 break;
1329 case CRAT_SUBTYPE_TLB_AFFINITY:
1330 /*
1331 * For now, nothing to do here
1332 */
1333 pr_debug("Found TLB entry in CRAT table (not processing)\n");
1334 break;
1335 case CRAT_SUBTYPE_CCOMPUTE_AFFINITY:
1336 /*
1337 * For now, nothing to do here
1338 */
1339 pr_debug("Found CCOMPUTE entry in CRAT table (not processing)\n");
1340 break;
1341 case CRAT_SUBTYPE_IOLINK_AFFINITY:
1342 iolink = (struct crat_subtype_iolink *)sub_type_hdr;
1343 ret = kfd_parse_subtype_iolink(iolink, device_list);
1344 break;
1345 default:
1346 pr_warn("Unknown subtype %d in CRAT\n",
1347 sub_type_hdr->type);
1348 }
1349
1350 return ret;
1351 }
1352
1353 /* kfd_parse_crat_table - parse CRAT table. For each node present in CRAT
1354 * create a kfd_topology_device and add in to device_list. Also parse
1355 * CRAT subtypes and attach it to appropriate kfd_topology_device
1356 * @crat_image - input image containing CRAT
1357 * @device_list - [OUT] list of kfd_topology_device generated after
1358 * parsing crat_image
1359 * @proximity_domain - Proximity domain of the first device in the table
1360 *
1361 * Return - 0 if successful else -ve value
1362 */
kfd_parse_crat_table(void * crat_image,struct list_head * device_list,uint32_t proximity_domain)1363 int kfd_parse_crat_table(void *crat_image, struct list_head *device_list,
1364 uint32_t proximity_domain)
1365 {
1366 struct kfd_topology_device *top_dev = NULL;
1367 struct crat_subtype_generic *sub_type_hdr;
1368 uint16_t node_id;
1369 int ret = 0;
1370 struct crat_header *crat_table = (struct crat_header *)crat_image;
1371 uint16_t num_nodes;
1372 uint32_t image_len;
1373
1374 if (!crat_image)
1375 return -EINVAL;
1376
1377 if (!list_empty(device_list)) {
1378 pr_warn("Error device list should be empty\n");
1379 return -EINVAL;
1380 }
1381
1382 num_nodes = crat_table->num_domains;
1383 image_len = crat_table->length;
1384
1385 pr_debug("Parsing CRAT table with %d nodes\n", num_nodes);
1386
1387 for (node_id = 0; node_id < num_nodes; node_id++) {
1388 top_dev = kfd_create_topology_device(device_list);
1389 if (!top_dev)
1390 break;
1391 top_dev->proximity_domain = proximity_domain++;
1392 }
1393
1394 if (!top_dev) {
1395 ret = -ENOMEM;
1396 goto err;
1397 }
1398
1399 memcpy(top_dev->oem_id, crat_table->oem_id, CRAT_OEMID_LENGTH);
1400 memcpy(top_dev->oem_table_id, crat_table->oem_table_id,
1401 CRAT_OEMTABLEID_LENGTH);
1402 top_dev->oem_revision = crat_table->oem_revision;
1403
1404 sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1);
1405 while ((char *)sub_type_hdr + sizeof(struct crat_subtype_generic) <
1406 ((char *)crat_image) + image_len) {
1407 if (!sub_type_hdr->length ||
1408 sub_type_hdr->length < sizeof(struct crat_subtype_generic)) {
1409 pr_warn("Invalid CRAT subtype length %u\n",
1410 sub_type_hdr->length);
1411 ret = -EINVAL;
1412 break;
1413 }
1414
1415 /* Validate subtype fits within remaining image */
1416 if ((char *)sub_type_hdr + sub_type_hdr->length >
1417 (char *)crat_image + image_len) {
1418 pr_warn("CRAT subtype length %u exceeds image bounds\n",
1419 sub_type_hdr->length);
1420 ret = -EINVAL;
1421 break;
1422 }
1423
1424 if (sub_type_hdr->flags & CRAT_SUBTYPE_FLAGS_ENABLED) {
1425 ret = kfd_parse_subtype(sub_type_hdr, device_list);
1426 if (ret)
1427 break;
1428 }
1429
1430 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1431 sub_type_hdr->length);
1432 }
1433
1434 err:
1435 if (ret)
1436 kfd_release_topology_device_list(device_list);
1437
1438 return ret;
1439 }
1440
1441
kfd_fill_gpu_cache_info_from_gfx_config(struct kfd_dev * kdev,bool cache_line_size_missing,struct kfd_gpu_cache_info * pcache_info)1442 static int kfd_fill_gpu_cache_info_from_gfx_config(struct kfd_dev *kdev,
1443 bool cache_line_size_missing,
1444 struct kfd_gpu_cache_info *pcache_info)
1445 {
1446 struct amdgpu_device *adev = kdev->adev;
1447 int i = 0;
1448
1449 /* TCP L1 Cache per CU */
1450 if (adev->gfx.config.gc_tcp_l1_size) {
1451 pcache_info[i].cache_size = adev->gfx.config.gc_tcp_l1_size;
1452 pcache_info[i].cache_level = 1;
1453 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1454 CRAT_CACHE_FLAGS_DATA_CACHE |
1455 CRAT_CACHE_FLAGS_SIMD_CACHE);
1456 pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_tcp_per_wpg / 2;
1457 pcache_info[i].cache_line_size = adev->gfx.config.gc_tcp_cache_line_size;
1458 if (cache_line_size_missing && !pcache_info[i].cache_line_size)
1459 pcache_info[i].cache_line_size = 128;
1460 i++;
1461 }
1462 /* Scalar L1 Instruction Cache per SQC */
1463 if (adev->gfx.config.gc_l1_instruction_cache_size_per_sqc) {
1464 pcache_info[i].cache_size =
1465 adev->gfx.config.gc_l1_instruction_cache_size_per_sqc;
1466 pcache_info[i].cache_level = 1;
1467 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1468 CRAT_CACHE_FLAGS_INST_CACHE |
1469 CRAT_CACHE_FLAGS_SIMD_CACHE);
1470 pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_sqc_per_wgp * 2;
1471 pcache_info[i].cache_line_size = adev->gfx.config.gc_instruction_cache_line_size;
1472 if (cache_line_size_missing && !pcache_info[i].cache_line_size)
1473 pcache_info[i].cache_line_size = 128;
1474 i++;
1475 }
1476 /* Scalar L1 Data Cache per SQC */
1477 if (adev->gfx.config.gc_l1_data_cache_size_per_sqc) {
1478 pcache_info[i].cache_size = adev->gfx.config.gc_l1_data_cache_size_per_sqc;
1479 pcache_info[i].cache_level = 1;
1480 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1481 CRAT_CACHE_FLAGS_DATA_CACHE |
1482 CRAT_CACHE_FLAGS_SIMD_CACHE);
1483 pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_sqc_per_wgp * 2;
1484 pcache_info[i].cache_line_size = adev->gfx.config.gc_scalar_data_cache_line_size;
1485 if (cache_line_size_missing && !pcache_info[i].cache_line_size)
1486 pcache_info[i].cache_line_size = 64;
1487 i++;
1488 }
1489 /* GL1 Data Cache per SA */
1490 if (adev->gfx.config.gc_gl1c_per_sa &&
1491 adev->gfx.config.gc_gl1c_size_per_instance) {
1492 pcache_info[i].cache_size = adev->gfx.config.gc_gl1c_per_sa *
1493 adev->gfx.config.gc_gl1c_size_per_instance;
1494 pcache_info[i].cache_level = 1;
1495 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1496 CRAT_CACHE_FLAGS_DATA_CACHE |
1497 CRAT_CACHE_FLAGS_SIMD_CACHE);
1498 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
1499 if (cache_line_size_missing)
1500 pcache_info[i].cache_line_size = 128;
1501 i++;
1502 }
1503 /* L2 Data Cache per GPU (Total Tex Cache) */
1504 if (adev->gfx.config.gc_gl2c_per_gpu) {
1505 pcache_info[i].cache_size = adev->gfx.config.gc_gl2c_per_gpu;
1506 pcache_info[i].cache_level = 2;
1507 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1508 CRAT_CACHE_FLAGS_DATA_CACHE |
1509 CRAT_CACHE_FLAGS_SIMD_CACHE);
1510 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
1511 pcache_info[i].cache_line_size = adev->gfx.config.gc_tcc_cache_line_size;
1512 if (cache_line_size_missing && !pcache_info[i].cache_line_size)
1513 pcache_info[i].cache_line_size = 128;
1514 i++;
1515 }
1516 /* L3 Data Cache per GPU */
1517 if (adev->gmc.mall_size) {
1518 pcache_info[i].cache_size = adev->gmc.mall_size / 1024;
1519 pcache_info[i].cache_level = 3;
1520 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1521 CRAT_CACHE_FLAGS_DATA_CACHE |
1522 CRAT_CACHE_FLAGS_SIMD_CACHE);
1523 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
1524 pcache_info[i].cache_line_size = 64;
1525 i++;
1526 }
1527 return i;
1528 }
1529
kfd_fill_gpu_cache_info_from_gfx_config_v2(struct kfd_dev * kdev,struct kfd_gpu_cache_info * pcache_info)1530 static int kfd_fill_gpu_cache_info_from_gfx_config_v2(struct kfd_dev *kdev,
1531 struct kfd_gpu_cache_info *pcache_info)
1532 {
1533 struct amdgpu_device *adev = kdev->adev;
1534 int i = 0;
1535
1536 /* TCP L1 Cache per CU */
1537 if (adev->gfx.config.gc_tcp_size_per_cu) {
1538 pcache_info[i].cache_size = adev->gfx.config.gc_tcp_size_per_cu;
1539 pcache_info[i].cache_level = 1;
1540 /* Cacheline size not available in IP discovery for gc943,gc944 */
1541 pcache_info[i].cache_line_size = 128;
1542 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1543 CRAT_CACHE_FLAGS_DATA_CACHE |
1544 CRAT_CACHE_FLAGS_SIMD_CACHE);
1545 pcache_info[i].num_cu_shared = 1;
1546 i++;
1547 }
1548 /* Scalar L1 Instruction Cache per SQC */
1549 if (adev->gfx.config.gc_l1_instruction_cache_size_per_sqc) {
1550 pcache_info[i].cache_size =
1551 adev->gfx.config.gc_l1_instruction_cache_size_per_sqc;
1552 pcache_info[i].cache_level = 1;
1553 pcache_info[i].cache_line_size = 64;
1554 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1555 CRAT_CACHE_FLAGS_INST_CACHE |
1556 CRAT_CACHE_FLAGS_SIMD_CACHE);
1557 pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_cu_per_sqc;
1558 i++;
1559 }
1560 /* Scalar L1 Data Cache per SQC */
1561 if (adev->gfx.config.gc_l1_data_cache_size_per_sqc) {
1562 pcache_info[i].cache_size = adev->gfx.config.gc_l1_data_cache_size_per_sqc;
1563 pcache_info[i].cache_level = 1;
1564 pcache_info[i].cache_line_size = 64;
1565 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1566 CRAT_CACHE_FLAGS_DATA_CACHE |
1567 CRAT_CACHE_FLAGS_SIMD_CACHE);
1568 pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_cu_per_sqc;
1569 i++;
1570 }
1571 /* L2 Data Cache per GPU (Total Tex Cache) */
1572 if (adev->gfx.config.gc_tcc_size) {
1573 pcache_info[i].cache_size = adev->gfx.config.gc_tcc_size;
1574 pcache_info[i].cache_level = 2;
1575 pcache_info[i].cache_line_size = 128;
1576 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1577 CRAT_CACHE_FLAGS_DATA_CACHE |
1578 CRAT_CACHE_FLAGS_SIMD_CACHE);
1579 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
1580 i++;
1581 }
1582 /* L3 Data Cache per GPU */
1583 if (adev->gmc.mall_size) {
1584 pcache_info[i].cache_size = adev->gmc.mall_size / 1024;
1585 pcache_info[i].cache_level = 3;
1586 pcache_info[i].cache_line_size = 64;
1587 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1588 CRAT_CACHE_FLAGS_DATA_CACHE |
1589 CRAT_CACHE_FLAGS_SIMD_CACHE);
1590 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
1591 i++;
1592 }
1593 return i;
1594 }
1595
kfd_get_gpu_cache_info(struct kfd_node * kdev,struct kfd_gpu_cache_info ** pcache_info)1596 int kfd_get_gpu_cache_info(struct kfd_node *kdev, struct kfd_gpu_cache_info **pcache_info)
1597 {
1598 int num_of_cache_types = 0;
1599 bool cache_line_size_missing = false;
1600
1601 switch (kdev->adev->asic_type) {
1602 case CHIP_KAVERI:
1603 *pcache_info = kaveri_cache_info;
1604 num_of_cache_types = ARRAY_SIZE(kaveri_cache_info);
1605 break;
1606 case CHIP_HAWAII:
1607 *pcache_info = hawaii_cache_info;
1608 num_of_cache_types = ARRAY_SIZE(hawaii_cache_info);
1609 break;
1610 case CHIP_CARRIZO:
1611 *pcache_info = carrizo_cache_info;
1612 num_of_cache_types = ARRAY_SIZE(carrizo_cache_info);
1613 break;
1614 case CHIP_TONGA:
1615 *pcache_info = tonga_cache_info;
1616 num_of_cache_types = ARRAY_SIZE(tonga_cache_info);
1617 break;
1618 case CHIP_FIJI:
1619 *pcache_info = fiji_cache_info;
1620 num_of_cache_types = ARRAY_SIZE(fiji_cache_info);
1621 break;
1622 case CHIP_POLARIS10:
1623 *pcache_info = polaris10_cache_info;
1624 num_of_cache_types = ARRAY_SIZE(polaris10_cache_info);
1625 break;
1626 case CHIP_POLARIS11:
1627 *pcache_info = polaris11_cache_info;
1628 num_of_cache_types = ARRAY_SIZE(polaris11_cache_info);
1629 break;
1630 case CHIP_POLARIS12:
1631 *pcache_info = polaris12_cache_info;
1632 num_of_cache_types = ARRAY_SIZE(polaris12_cache_info);
1633 break;
1634 case CHIP_VEGAM:
1635 *pcache_info = vegam_cache_info;
1636 num_of_cache_types = ARRAY_SIZE(vegam_cache_info);
1637 break;
1638 default:
1639 switch (KFD_GC_VERSION(kdev)) {
1640 case IP_VERSION(9, 0, 1):
1641 *pcache_info = vega10_cache_info;
1642 num_of_cache_types = ARRAY_SIZE(vega10_cache_info);
1643 break;
1644 case IP_VERSION(9, 2, 1):
1645 *pcache_info = vega12_cache_info;
1646 num_of_cache_types = ARRAY_SIZE(vega12_cache_info);
1647 break;
1648 case IP_VERSION(9, 4, 0):
1649 case IP_VERSION(9, 4, 1):
1650 *pcache_info = vega20_cache_info;
1651 num_of_cache_types = ARRAY_SIZE(vega20_cache_info);
1652 break;
1653 case IP_VERSION(9, 4, 2):
1654 *pcache_info = aldebaran_cache_info;
1655 num_of_cache_types = ARRAY_SIZE(aldebaran_cache_info);
1656 break;
1657 case IP_VERSION(9, 4, 3):
1658 case IP_VERSION(9, 4, 4):
1659 case IP_VERSION(9, 5, 0):
1660 num_of_cache_types =
1661 kfd_fill_gpu_cache_info_from_gfx_config_v2(kdev->kfd,
1662 *pcache_info);
1663 break;
1664 case IP_VERSION(9, 1, 0):
1665 case IP_VERSION(9, 2, 2):
1666 *pcache_info = raven_cache_info;
1667 num_of_cache_types = ARRAY_SIZE(raven_cache_info);
1668 break;
1669 case IP_VERSION(9, 3, 0):
1670 *pcache_info = renoir_cache_info;
1671 num_of_cache_types = ARRAY_SIZE(renoir_cache_info);
1672 break;
1673 case IP_VERSION(10, 1, 10):
1674 case IP_VERSION(10, 1, 2):
1675 case IP_VERSION(10, 1, 3):
1676 case IP_VERSION(10, 1, 4):
1677 *pcache_info = navi10_cache_info;
1678 num_of_cache_types = ARRAY_SIZE(navi10_cache_info);
1679 break;
1680 case IP_VERSION(10, 1, 1):
1681 *pcache_info = navi14_cache_info;
1682 num_of_cache_types = ARRAY_SIZE(navi14_cache_info);
1683 break;
1684 case IP_VERSION(10, 3, 0):
1685 *pcache_info = sienna_cichlid_cache_info;
1686 num_of_cache_types = ARRAY_SIZE(sienna_cichlid_cache_info);
1687 break;
1688 case IP_VERSION(10, 3, 2):
1689 *pcache_info = navy_flounder_cache_info;
1690 num_of_cache_types = ARRAY_SIZE(navy_flounder_cache_info);
1691 break;
1692 case IP_VERSION(10, 3, 4):
1693 *pcache_info = dimgrey_cavefish_cache_info;
1694 num_of_cache_types = ARRAY_SIZE(dimgrey_cavefish_cache_info);
1695 break;
1696 case IP_VERSION(10, 3, 1):
1697 *pcache_info = vangogh_cache_info;
1698 num_of_cache_types = ARRAY_SIZE(vangogh_cache_info);
1699 break;
1700 case IP_VERSION(10, 3, 5):
1701 *pcache_info = beige_goby_cache_info;
1702 num_of_cache_types = ARRAY_SIZE(beige_goby_cache_info);
1703 break;
1704 case IP_VERSION(10, 3, 3):
1705 *pcache_info = yellow_carp_cache_info;
1706 num_of_cache_types = ARRAY_SIZE(yellow_carp_cache_info);
1707 break;
1708 case IP_VERSION(10, 3, 6):
1709 *pcache_info = gc_10_3_6_cache_info;
1710 num_of_cache_types = ARRAY_SIZE(gc_10_3_6_cache_info);
1711 break;
1712 case IP_VERSION(10, 3, 7):
1713 *pcache_info = gfx1037_cache_info;
1714 num_of_cache_types = ARRAY_SIZE(gfx1037_cache_info);
1715 break;
1716 case IP_VERSION(11, 0, 0):
1717 case IP_VERSION(11, 0, 1):
1718 case IP_VERSION(11, 0, 2):
1719 case IP_VERSION(11, 0, 3):
1720 case IP_VERSION(11, 0, 4):
1721 case IP_VERSION(11, 5, 0):
1722 case IP_VERSION(11, 5, 1):
1723 case IP_VERSION(11, 5, 2):
1724 case IP_VERSION(11, 5, 3):
1725 case IP_VERSION(11, 5, 4):
1726 case IP_VERSION(11, 5, 6):
1727 case IP_VERSION(11, 7, 0):
1728 case IP_VERSION(11, 7, 1):
1729 /* Cacheline size not available in IP discovery for gc11.
1730 * kfd_fill_gpu_cache_info_from_gfx_config to hard code it
1731 */
1732 cache_line_size_missing = true;
1733 fallthrough;
1734 case IP_VERSION(12, 0, 0):
1735 case IP_VERSION(12, 0, 1):
1736 case IP_VERSION(12, 1, 0):
1737 num_of_cache_types =
1738 kfd_fill_gpu_cache_info_from_gfx_config(kdev->kfd,
1739 cache_line_size_missing,
1740 *pcache_info);
1741 break;
1742 default:
1743 *pcache_info = dummy_cache_info;
1744 num_of_cache_types = ARRAY_SIZE(dummy_cache_info);
1745 pr_warn("dummy cache info is used temporarily and real cache info need update later.\n");
1746 break;
1747 }
1748 }
1749 return num_of_cache_types;
1750 }
1751
1752 /* Memory required to create Virtual CRAT.
1753 * Since there is no easy way to predict the amount of memory required, the
1754 * following amount is allocated for GPU Virtual CRAT. This is
1755 * expected to cover all known conditions. But to be safe additional check
1756 * is put in the code to ensure we don't overwrite.
1757 */
1758 #define VCRAT_SIZE_FOR_GPU (4 * PAGE_SIZE)
1759
1760 /* kfd_fill_cu_for_cpu - Fill in Compute info for the given CPU NUMA node
1761 *
1762 * @numa_node_id: CPU NUMA node id
1763 * @avail_size: Available size in the memory
1764 * @sub_type_hdr: Memory into which compute info will be filled in
1765 *
1766 * Return 0 if successful else return -ve value
1767 */
kfd_fill_cu_for_cpu(int numa_node_id,int * avail_size,int proximity_domain,struct crat_subtype_computeunit * sub_type_hdr)1768 static int kfd_fill_cu_for_cpu(int numa_node_id, int *avail_size,
1769 int proximity_domain,
1770 struct crat_subtype_computeunit *sub_type_hdr)
1771 {
1772 const struct cpumask *cpumask;
1773
1774 *avail_size -= sizeof(struct crat_subtype_computeunit);
1775 if (*avail_size < 0)
1776 return -ENOMEM;
1777
1778 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit));
1779
1780 /* Fill in subtype header data */
1781 sub_type_hdr->type = CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY;
1782 sub_type_hdr->length = sizeof(struct crat_subtype_computeunit);
1783 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
1784
1785 cpumask = cpumask_of_node(numa_node_id);
1786
1787 /* Fill in CU data */
1788 sub_type_hdr->flags |= CRAT_CU_FLAGS_CPU_PRESENT;
1789 sub_type_hdr->proximity_domain = proximity_domain;
1790 sub_type_hdr->processor_id_low = kfd_numa_node_to_apic_id(numa_node_id);
1791 if (sub_type_hdr->processor_id_low == -1)
1792 return -EINVAL;
1793
1794 sub_type_hdr->num_cpu_cores = cpumask_weight(cpumask);
1795
1796 return 0;
1797 }
1798
1799 /* kfd_fill_mem_info_for_cpu - Fill in Memory info for the given CPU NUMA node
1800 *
1801 * @numa_node_id: CPU NUMA node id
1802 * @avail_size: Available size in the memory
1803 * @sub_type_hdr: Memory into which compute info will be filled in
1804 *
1805 * Return 0 if successful else return -ve value
1806 */
kfd_fill_mem_info_for_cpu(int numa_node_id,int * avail_size,int proximity_domain,struct crat_subtype_memory * sub_type_hdr)1807 static int kfd_fill_mem_info_for_cpu(int numa_node_id, int *avail_size,
1808 int proximity_domain,
1809 struct crat_subtype_memory *sub_type_hdr)
1810 {
1811 uint64_t mem_in_bytes = 0;
1812 pg_data_t *pgdat;
1813 int zone_type;
1814
1815 *avail_size -= sizeof(struct crat_subtype_memory);
1816 if (*avail_size < 0)
1817 return -ENOMEM;
1818
1819 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_memory));
1820
1821 /* Fill in subtype header data */
1822 sub_type_hdr->type = CRAT_SUBTYPE_MEMORY_AFFINITY;
1823 sub_type_hdr->length = sizeof(struct crat_subtype_memory);
1824 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
1825
1826 /* Fill in Memory Subunit data */
1827
1828 /* Unlike si_meminfo, si_meminfo_node is not exported. So
1829 * the following lines are duplicated from si_meminfo_node
1830 * function
1831 */
1832 pgdat = NODE_DATA(numa_node_id);
1833 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
1834 mem_in_bytes += zone_managed_pages(&pgdat->node_zones[zone_type]);
1835 mem_in_bytes <<= PAGE_SHIFT;
1836
1837 sub_type_hdr->length_low = lower_32_bits(mem_in_bytes);
1838 sub_type_hdr->length_high = upper_32_bits(mem_in_bytes);
1839 sub_type_hdr->proximity_domain = proximity_domain;
1840
1841 return 0;
1842 }
1843
1844 #if defined(CONFIG_X86_64) && !defined(CONFIG_UML)
kfd_fill_iolink_info_for_cpu(int numa_node_id,int * avail_size,uint32_t * num_entries,struct crat_subtype_iolink * sub_type_hdr)1845 static int kfd_fill_iolink_info_for_cpu(int numa_node_id, int *avail_size,
1846 uint32_t *num_entries,
1847 struct crat_subtype_iolink *sub_type_hdr)
1848 {
1849 int nid;
1850 struct cpuinfo_x86 *c = &cpu_data(0);
1851 uint8_t link_type;
1852
1853 if (c->x86_vendor == X86_VENDOR_AMD)
1854 link_type = CRAT_IOLINK_TYPE_HYPERTRANSPORT;
1855 else
1856 link_type = CRAT_IOLINK_TYPE_QPI_1_1;
1857
1858 *num_entries = 0;
1859
1860 /* Create IO links from this node to other CPU nodes */
1861 for_each_online_node(nid) {
1862 if (nid == numa_node_id) /* node itself */
1863 continue;
1864
1865 *avail_size -= sizeof(struct crat_subtype_iolink);
1866 if (*avail_size < 0)
1867 return -ENOMEM;
1868
1869 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
1870
1871 /* Fill in subtype header data */
1872 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
1873 sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
1874 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
1875
1876 /* Fill in IO link data */
1877 sub_type_hdr->proximity_domain_from = numa_node_id;
1878 sub_type_hdr->proximity_domain_to = nid;
1879 sub_type_hdr->io_interface_type = link_type;
1880
1881 (*num_entries)++;
1882 sub_type_hdr++;
1883 }
1884
1885 return 0;
1886 }
1887 #endif
1888
1889 /* kfd_create_vcrat_image_cpu - Create Virtual CRAT for CPU
1890 *
1891 * @pcrat_image: Fill in VCRAT for CPU
1892 * @size: [IN] allocated size of crat_image.
1893 * [OUT] actual size of data filled in crat_image
1894 */
kfd_create_vcrat_image_cpu(void * pcrat_image,size_t * size)1895 static int kfd_create_vcrat_image_cpu(void *pcrat_image, size_t *size)
1896 {
1897 struct crat_header *crat_table = (struct crat_header *)pcrat_image;
1898 struct acpi_table_header *acpi_table;
1899 acpi_status status;
1900 struct crat_subtype_generic *sub_type_hdr;
1901 int avail_size = *size;
1902 int numa_node_id;
1903 #if defined(CONFIG_X86_64) && !defined(CONFIG_UML)
1904 uint32_t entries = 0;
1905 #endif
1906 int ret = 0;
1907
1908 if (!pcrat_image)
1909 return -EINVAL;
1910
1911 /* Fill in CRAT Header.
1912 * Modify length and total_entries as subunits are added.
1913 */
1914 avail_size -= sizeof(struct crat_header);
1915 if (avail_size < 0)
1916 return -ENOMEM;
1917
1918 memset(crat_table, 0, sizeof(struct crat_header));
1919 memcpy(&crat_table->signature, CRAT_SIGNATURE,
1920 sizeof(crat_table->signature));
1921 crat_table->length = sizeof(struct crat_header);
1922
1923 status = acpi_get_table("DSDT", 0, &acpi_table);
1924 if (status != AE_OK)
1925 pr_warn("DSDT table not found for OEM information\n");
1926 else {
1927 crat_table->oem_revision = acpi_table->revision;
1928 memcpy(crat_table->oem_id, acpi_table->oem_id,
1929 CRAT_OEMID_LENGTH);
1930 memcpy(crat_table->oem_table_id, acpi_table->oem_table_id,
1931 CRAT_OEMTABLEID_LENGTH);
1932 acpi_put_table(acpi_table);
1933 }
1934 crat_table->total_entries = 0;
1935 crat_table->num_domains = 0;
1936
1937 sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1);
1938
1939 for_each_online_node(numa_node_id) {
1940 if (kfd_numa_node_to_apic_id(numa_node_id) == -1)
1941 continue;
1942
1943 /* Fill in Subtype: Compute Unit */
1944 ret = kfd_fill_cu_for_cpu(numa_node_id, &avail_size,
1945 crat_table->num_domains,
1946 (struct crat_subtype_computeunit *)sub_type_hdr);
1947 if (ret < 0)
1948 return ret;
1949 crat_table->length += sub_type_hdr->length;
1950 crat_table->total_entries++;
1951
1952 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1953 sub_type_hdr->length);
1954
1955 /* Fill in Subtype: Memory */
1956 ret = kfd_fill_mem_info_for_cpu(numa_node_id, &avail_size,
1957 crat_table->num_domains,
1958 (struct crat_subtype_memory *)sub_type_hdr);
1959 if (ret < 0)
1960 return ret;
1961 crat_table->length += sub_type_hdr->length;
1962 crat_table->total_entries++;
1963
1964 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1965 sub_type_hdr->length);
1966
1967 /* Fill in Subtype: IO Link */
1968 #if defined(CONFIG_X86_64) && !defined(CONFIG_UML)
1969 ret = kfd_fill_iolink_info_for_cpu(numa_node_id, &avail_size,
1970 &entries,
1971 (struct crat_subtype_iolink *)sub_type_hdr);
1972 if (ret < 0)
1973 return ret;
1974
1975 if (entries) {
1976 crat_table->length += (sub_type_hdr->length * entries);
1977 crat_table->total_entries += entries;
1978
1979 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1980 sub_type_hdr->length * entries);
1981 }
1982 #else
1983 pr_info("IO link not available for non x86 platforms\n");
1984 #endif
1985
1986 crat_table->num_domains++;
1987 }
1988
1989 /* TODO: Add cache Subtype for CPU.
1990 * Currently, CPU cache information is available in function
1991 * detect_cache_attributes(cpu) defined in the file
1992 * ./arch/x86/kernel/cpu/intel_cacheinfo.c. This function is not
1993 * exported and to get the same information the code needs to be
1994 * duplicated.
1995 */
1996
1997 *size = crat_table->length;
1998 pr_info("Virtual CRAT table created for CPU\n");
1999
2000 return 0;
2001 }
2002
kfd_fill_gpu_memory_affinity(int * avail_size,struct kfd_node * kdev,uint8_t type,uint64_t size,struct crat_subtype_memory * sub_type_hdr,uint32_t proximity_domain,const struct kfd_local_mem_info * local_mem_info)2003 static int kfd_fill_gpu_memory_affinity(int *avail_size,
2004 struct kfd_node *kdev, uint8_t type, uint64_t size,
2005 struct crat_subtype_memory *sub_type_hdr,
2006 uint32_t proximity_domain,
2007 const struct kfd_local_mem_info *local_mem_info)
2008 {
2009 *avail_size -= sizeof(struct crat_subtype_memory);
2010 if (*avail_size < 0)
2011 return -ENOMEM;
2012
2013 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_memory));
2014 sub_type_hdr->type = CRAT_SUBTYPE_MEMORY_AFFINITY;
2015 sub_type_hdr->length = sizeof(struct crat_subtype_memory);
2016 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED;
2017
2018 sub_type_hdr->proximity_domain = proximity_domain;
2019
2020 pr_debug("Fill gpu memory affinity - type 0x%x size 0x%llx\n",
2021 type, size);
2022
2023 sub_type_hdr->length_low = lower_32_bits(size);
2024 sub_type_hdr->length_high = upper_32_bits(size);
2025
2026 sub_type_hdr->width = local_mem_info->vram_width;
2027 sub_type_hdr->visibility_type = type;
2028
2029 return 0;
2030 }
2031
2032 #ifdef CONFIG_ACPI_NUMA
kfd_find_numa_node_in_srat(struct kfd_node * kdev)2033 static void kfd_find_numa_node_in_srat(struct kfd_node *kdev)
2034 {
2035 struct acpi_table_header *table_header = NULL;
2036 struct acpi_subtable_header *sub_header = NULL;
2037 unsigned long table_end, subtable_len;
2038 u32 pci_id = pci_domain_nr(kdev->adev->pdev->bus) << 16 |
2039 pci_dev_id(kdev->adev->pdev);
2040 u32 bdf;
2041 acpi_status status;
2042 struct acpi_srat_cpu_affinity *cpu;
2043 struct acpi_srat_generic_affinity *gpu;
2044 int pxm = 0, max_pxm = 0;
2045 int numa_node = NUMA_NO_NODE;
2046 bool found = false;
2047
2048 /* Fetch the SRAT table from ACPI */
2049 status = acpi_get_table(ACPI_SIG_SRAT, 0, &table_header);
2050 if (status == AE_NOT_FOUND) {
2051 pr_warn("SRAT table not found\n");
2052 return;
2053 } else if (ACPI_FAILURE(status)) {
2054 const char *err = acpi_format_exception(status);
2055 pr_err("SRAT table error: %s\n", err);
2056 return;
2057 }
2058
2059 table_end = (unsigned long)table_header + table_header->length;
2060
2061 /* Parse all entries looking for a match. */
2062 sub_header = (struct acpi_subtable_header *)
2063 ((unsigned long)table_header +
2064 sizeof(struct acpi_table_srat));
2065 subtable_len = sub_header->length;
2066
2067 while (((unsigned long)sub_header) + subtable_len < table_end) {
2068 /*
2069 * If length is 0, break from this loop to avoid
2070 * infinite loop.
2071 */
2072 if (subtable_len == 0) {
2073 pr_err("SRAT invalid zero length\n");
2074 break;
2075 }
2076
2077 switch (sub_header->type) {
2078 case ACPI_SRAT_TYPE_CPU_AFFINITY:
2079 cpu = (struct acpi_srat_cpu_affinity *)sub_header;
2080 pxm = *((u32 *)cpu->proximity_domain_hi) << 8 |
2081 cpu->proximity_domain_lo;
2082 if (pxm > max_pxm)
2083 max_pxm = pxm;
2084 break;
2085 case ACPI_SRAT_TYPE_GENERIC_AFFINITY:
2086 gpu = (struct acpi_srat_generic_affinity *)sub_header;
2087 bdf = *((u16 *)(&gpu->device_handle[0])) << 16 |
2088 *((u16 *)(&gpu->device_handle[2]));
2089 if (bdf == pci_id) {
2090 found = true;
2091 numa_node = pxm_to_node(gpu->proximity_domain);
2092 }
2093 break;
2094 default:
2095 break;
2096 }
2097
2098 if (found)
2099 break;
2100
2101 sub_header = (struct acpi_subtable_header *)
2102 ((unsigned long)sub_header + subtable_len);
2103 subtable_len = sub_header->length;
2104 }
2105
2106 acpi_put_table(table_header);
2107
2108 /* Workaround bad cpu-gpu binding case */
2109 if (found && (numa_node < 0 ||
2110 numa_node > pxm_to_node(max_pxm)))
2111 numa_node = 0;
2112
2113 if (numa_node != NUMA_NO_NODE)
2114 set_dev_node(&kdev->adev->pdev->dev, numa_node);
2115 }
2116 #endif
2117
2118 #define KFD_CRAT_INTRA_SOCKET_WEIGHT 13
2119 #define KFD_CRAT_XGMI_WEIGHT 15
2120
2121 /* kfd_fill_gpu_direct_io_link - Fill in direct io link from GPU
2122 * to its NUMA node
2123 * @avail_size: Available size in the memory
2124 * @kdev - [IN] GPU device
2125 * @sub_type_hdr: Memory into which io link info will be filled in
2126 * @proximity_domain - proximity domain of the GPU node
2127 *
2128 * Return 0 if successful else return -ve value
2129 */
kfd_fill_gpu_direct_io_link_to_cpu(int * avail_size,struct kfd_node * kdev,struct crat_subtype_iolink * sub_type_hdr,uint32_t proximity_domain)2130 static int kfd_fill_gpu_direct_io_link_to_cpu(int *avail_size,
2131 struct kfd_node *kdev,
2132 struct crat_subtype_iolink *sub_type_hdr,
2133 uint32_t proximity_domain)
2134 {
2135 *avail_size -= sizeof(struct crat_subtype_iolink);
2136 if (*avail_size < 0)
2137 return -ENOMEM;
2138
2139 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
2140
2141 /* Fill in subtype header data */
2142 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
2143 sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
2144 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED;
2145 if (kfd_dev_is_large_bar(kdev))
2146 sub_type_hdr->flags |= CRAT_IOLINK_FLAGS_BI_DIRECTIONAL;
2147
2148 /* Fill in IOLINK subtype.
2149 * TODO: Fill-in other fields of iolink subtype
2150 */
2151 if (kdev->adev->gmc.xgmi.connected_to_cpu ||
2152 (KFD_GC_VERSION(kdev) == IP_VERSION(9, 4, 3) &&
2153 kdev->adev->smuio.funcs->get_pkg_type(kdev->adev) ==
2154 AMDGPU_PKG_TYPE_APU)) {
2155 bool ext_cpu = KFD_GC_VERSION(kdev) != IP_VERSION(9, 4, 3);
2156 int mem_bw = 819200, weight = ext_cpu ? KFD_CRAT_XGMI_WEIGHT :
2157 KFD_CRAT_INTRA_SOCKET_WEIGHT;
2158 /*
2159 * with host gpu xgmi link, host can access gpu memory whether
2160 * or not pcie bar type is large, so always create bidirectional
2161 * io link.
2162 */
2163 sub_type_hdr->flags |= CRAT_IOLINK_FLAGS_BI_DIRECTIONAL;
2164 sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_XGMI;
2165 sub_type_hdr->weight_xgmi = weight;
2166 if (ext_cpu) {
2167 amdgpu_xgmi_get_bandwidth(kdev->adev, NULL,
2168 AMDGPU_XGMI_BW_MODE_PER_LINK,
2169 AMDGPU_XGMI_BW_UNIT_MBYTES,
2170 &sub_type_hdr->minimum_bandwidth_mbs,
2171 &sub_type_hdr->maximum_bandwidth_mbs);
2172 } else {
2173 sub_type_hdr->minimum_bandwidth_mbs = mem_bw;
2174 sub_type_hdr->maximum_bandwidth_mbs = mem_bw;
2175 }
2176 } else {
2177 sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_PCIEXPRESS;
2178 sub_type_hdr->minimum_bandwidth_mbs =
2179 amdgpu_amdkfd_get_pcie_bandwidth_mbytes(kdev->adev, true);
2180 sub_type_hdr->maximum_bandwidth_mbs =
2181 amdgpu_amdkfd_get_pcie_bandwidth_mbytes(kdev->adev, false);
2182 }
2183
2184 sub_type_hdr->proximity_domain_from = proximity_domain;
2185
2186 #ifdef CONFIG_ACPI_NUMA
2187 if (kdev->adev->pdev->dev.numa_node == NUMA_NO_NODE &&
2188 num_possible_nodes() > 1)
2189 kfd_find_numa_node_in_srat(kdev);
2190 #endif
2191 #ifdef CONFIG_NUMA
2192 if (kdev->adev->pdev->dev.numa_node == NUMA_NO_NODE)
2193 sub_type_hdr->proximity_domain_to = 0;
2194 else
2195 sub_type_hdr->proximity_domain_to = kdev->adev->pdev->dev.numa_node;
2196 #else
2197 sub_type_hdr->proximity_domain_to = 0;
2198 #endif
2199 return 0;
2200 }
2201
kfd_fill_gpu_xgmi_link_to_gpu(int * avail_size,struct kfd_node * kdev,struct kfd_node * peer_kdev,struct crat_subtype_iolink * sub_type_hdr,uint32_t proximity_domain_from,uint32_t proximity_domain_to)2202 static int kfd_fill_gpu_xgmi_link_to_gpu(int *avail_size,
2203 struct kfd_node *kdev,
2204 struct kfd_node *peer_kdev,
2205 struct crat_subtype_iolink *sub_type_hdr,
2206 uint32_t proximity_domain_from,
2207 uint32_t proximity_domain_to)
2208 {
2209 bool use_ta_info = kdev->kfd->num_nodes == 1;
2210
2211 *avail_size -= sizeof(struct crat_subtype_iolink);
2212 if (*avail_size < 0)
2213 return -ENOMEM;
2214
2215 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
2216
2217 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
2218 sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
2219 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED |
2220 CRAT_IOLINK_FLAGS_BI_DIRECTIONAL;
2221
2222 sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_XGMI;
2223 sub_type_hdr->proximity_domain_from = proximity_domain_from;
2224 sub_type_hdr->proximity_domain_to = proximity_domain_to;
2225
2226 if (use_ta_info) {
2227 sub_type_hdr->weight_xgmi = KFD_CRAT_XGMI_WEIGHT *
2228 amdgpu_xgmi_get_hops_count(kdev->adev, peer_kdev->adev);
2229 amdgpu_xgmi_get_bandwidth(kdev->adev, peer_kdev->adev,
2230 AMDGPU_XGMI_BW_MODE_PER_PEER,
2231 AMDGPU_XGMI_BW_UNIT_MBYTES,
2232 &sub_type_hdr->minimum_bandwidth_mbs,
2233 &sub_type_hdr->maximum_bandwidth_mbs);
2234 } else {
2235 bool is_single_hop = kdev->kfd == peer_kdev->kfd;
2236 int weight = is_single_hop ? KFD_CRAT_INTRA_SOCKET_WEIGHT :
2237 (2 * KFD_CRAT_INTRA_SOCKET_WEIGHT) + KFD_CRAT_XGMI_WEIGHT;
2238 int mem_bw = 819200;
2239
2240 sub_type_hdr->weight_xgmi = weight;
2241 sub_type_hdr->maximum_bandwidth_mbs = is_single_hop ? mem_bw : 0;
2242 sub_type_hdr->minimum_bandwidth_mbs = is_single_hop ? mem_bw : 0;
2243 }
2244
2245 return 0;
2246 }
2247
2248 /* kfd_create_vcrat_image_gpu - Create Virtual CRAT for CPU
2249 *
2250 * @pcrat_image: Fill in VCRAT for GPU
2251 * @size: [IN] allocated size of crat_image.
2252 * [OUT] actual size of data filled in crat_image
2253 */
kfd_create_vcrat_image_gpu(void * pcrat_image,size_t * size,struct kfd_node * kdev,uint32_t proximity_domain)2254 static int kfd_create_vcrat_image_gpu(void *pcrat_image,
2255 size_t *size, struct kfd_node *kdev,
2256 uint32_t proximity_domain)
2257 {
2258 struct crat_header *crat_table = (struct crat_header *)pcrat_image;
2259 struct amdgpu_gfx_config *gfx_info = &kdev->adev->gfx.config;
2260 struct amdgpu_cu_info *cu_info = &kdev->adev->gfx.cu_info;
2261 struct crat_subtype_generic *sub_type_hdr;
2262 struct kfd_local_mem_info local_mem_info;
2263 struct kfd_topology_device *peer_dev;
2264 struct crat_subtype_computeunit *cu;
2265 int avail_size = *size;
2266 uint32_t total_num_of_cu;
2267 uint32_t nid = 0;
2268 int ret = 0;
2269
2270 if (!pcrat_image || avail_size < VCRAT_SIZE_FOR_GPU)
2271 return -EINVAL;
2272
2273 /* Fill the CRAT Header.
2274 * Modify length and total_entries as subunits are added.
2275 */
2276 avail_size -= sizeof(struct crat_header);
2277 memset(crat_table, 0, sizeof(struct crat_header));
2278
2279 memcpy(&crat_table->signature, CRAT_SIGNATURE,
2280 sizeof(crat_table->signature));
2281 /* Change length as we add more subtypes*/
2282 crat_table->length = sizeof(struct crat_header);
2283 crat_table->num_domains = 1;
2284 crat_table->total_entries = 0;
2285
2286 /* Fill in Subtype: Compute Unit
2287 * First fill in the sub type header and then sub type data
2288 */
2289 avail_size -= sizeof(struct crat_subtype_computeunit);
2290 sub_type_hdr = (struct crat_subtype_generic *)(crat_table + 1);
2291 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit));
2292
2293 sub_type_hdr->type = CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY;
2294 sub_type_hdr->length = sizeof(struct crat_subtype_computeunit);
2295 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
2296
2297 /* Fill CU subtype data */
2298 cu = (struct crat_subtype_computeunit *)sub_type_hdr;
2299 cu->flags |= CRAT_CU_FLAGS_GPU_PRESENT;
2300 cu->proximity_domain = proximity_domain;
2301
2302 cu->num_simd_per_cu = cu_info->simd_per_cu;
2303 cu->num_simd_cores = cu_info->simd_per_cu *
2304 (cu_info->number / kdev->kfd->num_nodes);
2305 cu->max_waves_simd = cu_info->max_waves_per_simd;
2306
2307 cu->wave_front_size = cu_info->wave_front_size;
2308 cu->array_count = gfx_info->max_sh_per_se *
2309 gfx_info->max_shader_engines;
2310 total_num_of_cu = (cu->array_count * gfx_info->max_cu_per_sh);
2311 cu->processor_id_low = get_and_inc_gpu_processor_id(total_num_of_cu);
2312 cu->num_cu_per_array = gfx_info->max_cu_per_sh;
2313 cu->max_slots_scatch_cu = cu_info->max_scratch_slots_per_cu;
2314 cu->num_banks = gfx_info->max_shader_engines;
2315 cu->lds_size_in_kb = cu_info->lds_size;
2316
2317 cu->hsa_capability = 0;
2318
2319 crat_table->length += sub_type_hdr->length;
2320 crat_table->total_entries++;
2321
2322 /* Fill in Subtype: Memory. Only on systems with large BAR (no
2323 * private FB), report memory as public. On other systems
2324 * report the total FB size (public+private) as a single
2325 * private heap.
2326 */
2327 local_mem_info = kdev->local_mem_info;
2328 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
2329 sub_type_hdr->length);
2330
2331 if (kdev->adev->debug_largebar)
2332 local_mem_info.local_mem_size_private = 0;
2333
2334 if (local_mem_info.local_mem_size_private == 0)
2335 ret = kfd_fill_gpu_memory_affinity(&avail_size,
2336 kdev, HSA_MEM_HEAP_TYPE_FB_PUBLIC,
2337 local_mem_info.local_mem_size_public,
2338 (struct crat_subtype_memory *)sub_type_hdr,
2339 proximity_domain,
2340 &local_mem_info);
2341 else
2342 ret = kfd_fill_gpu_memory_affinity(&avail_size,
2343 kdev, HSA_MEM_HEAP_TYPE_FB_PRIVATE,
2344 local_mem_info.local_mem_size_public +
2345 local_mem_info.local_mem_size_private,
2346 (struct crat_subtype_memory *)sub_type_hdr,
2347 proximity_domain,
2348 &local_mem_info);
2349 if (ret < 0)
2350 return ret;
2351
2352 crat_table->length += sizeof(struct crat_subtype_memory);
2353 crat_table->total_entries++;
2354
2355 /* Fill in Subtype: IO_LINKS
2356 * Only direct links are added here which is Link from GPU to
2357 * its NUMA node. Indirect links are added by userspace.
2358 */
2359 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
2360 sub_type_hdr->length);
2361 ret = kfd_fill_gpu_direct_io_link_to_cpu(&avail_size, kdev,
2362 (struct crat_subtype_iolink *)sub_type_hdr, proximity_domain);
2363
2364 if (ret < 0)
2365 return ret;
2366
2367 crat_table->length += sub_type_hdr->length;
2368 crat_table->total_entries++;
2369
2370
2371 /* Fill in Subtype: IO_LINKS
2372 * Direct links from GPU to other GPUs through xGMI.
2373 * We will loop GPUs that already be processed (with lower value
2374 * of proximity_domain), add the link for the GPUs with same
2375 * hive id (from this GPU to other GPU) . The reversed iolink
2376 * (from other GPU to this GPU) will be added
2377 * in kfd_parse_subtype_iolink.
2378 */
2379 if (kdev->kfd->hive_id) {
2380 for (nid = 0; nid < proximity_domain; ++nid) {
2381 peer_dev = kfd_topology_device_by_proximity_domain_no_lock(nid);
2382 if (!peer_dev || !peer_dev->gpu)
2383 continue;
2384 if (peer_dev->gpu->kfd->hive_id != kdev->kfd->hive_id)
2385 continue;
2386 if (!amdgpu_xgmi_get_is_sharing_enabled(kdev->adev, peer_dev->gpu->adev))
2387 continue;
2388 sub_type_hdr = (typeof(sub_type_hdr))(
2389 (char *)sub_type_hdr +
2390 sizeof(struct crat_subtype_iolink));
2391 ret = kfd_fill_gpu_xgmi_link_to_gpu(
2392 &avail_size, kdev, peer_dev->gpu,
2393 (struct crat_subtype_iolink *)sub_type_hdr,
2394 proximity_domain, nid);
2395 if (ret < 0)
2396 return ret;
2397 crat_table->length += sub_type_hdr->length;
2398 crat_table->total_entries++;
2399 }
2400 }
2401 *size = crat_table->length;
2402 pr_info("Virtual CRAT table created for GPU\n");
2403
2404 return ret;
2405 }
2406
2407 /* kfd_create_crat_image_virtual - Allocates memory for CRAT image and
2408 * creates a Virtual CRAT (VCRAT) image
2409 *
2410 * NOTE: Call kfd_destroy_crat_image to free CRAT image memory
2411 *
2412 * @crat_image: VCRAT image created because ACPI does not have a
2413 * CRAT for this device
2414 * @size: [OUT] size of virtual crat_image
2415 * @flags: COMPUTE_UNIT_CPU - Create VCRAT for CPU device
2416 * COMPUTE_UNIT_GPU - Create VCRAT for GPU
2417 * (COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU) - Create VCRAT for APU
2418 * -- this option is not currently implemented.
2419 * The assumption is that all AMD APUs will have CRAT
2420 * @kdev: Valid kfd_node required if flags contain COMPUTE_UNIT_GPU
2421 *
2422 * Return 0 if successful else return -ve value
2423 */
kfd_create_crat_image_virtual(void ** crat_image,size_t * size,int flags,struct kfd_node * kdev,uint32_t proximity_domain)2424 int kfd_create_crat_image_virtual(void **crat_image, size_t *size,
2425 int flags, struct kfd_node *kdev,
2426 uint32_t proximity_domain)
2427 {
2428 void *pcrat_image = NULL;
2429 int ret = 0, num_nodes;
2430 size_t dyn_size;
2431
2432 if (!crat_image)
2433 return -EINVAL;
2434
2435 *crat_image = NULL;
2436
2437 /* Allocate the CPU Virtual CRAT size based on the number of online
2438 * nodes. Allocate VCRAT_SIZE_FOR_GPU for GPU virtual CRAT image.
2439 * This should cover all the current conditions. A check is put not
2440 * to overwrite beyond allocated size for GPUs
2441 */
2442 switch (flags) {
2443 case COMPUTE_UNIT_CPU:
2444 num_nodes = num_online_nodes();
2445 dyn_size = sizeof(struct crat_header) +
2446 num_nodes * (sizeof(struct crat_subtype_computeunit) +
2447 sizeof(struct crat_subtype_memory) +
2448 (num_nodes - 1) * sizeof(struct crat_subtype_iolink));
2449 pcrat_image = kvmalloc(dyn_size, GFP_KERNEL);
2450 if (!pcrat_image)
2451 return -ENOMEM;
2452 *size = dyn_size;
2453 pr_debug("CRAT size is %ld", dyn_size);
2454 ret = kfd_create_vcrat_image_cpu(pcrat_image, size);
2455 break;
2456 case COMPUTE_UNIT_GPU:
2457 if (!kdev)
2458 return -EINVAL;
2459 pcrat_image = kvmalloc(VCRAT_SIZE_FOR_GPU, GFP_KERNEL);
2460 if (!pcrat_image)
2461 return -ENOMEM;
2462 *size = VCRAT_SIZE_FOR_GPU;
2463 ret = kfd_create_vcrat_image_gpu(pcrat_image, size, kdev,
2464 proximity_domain);
2465 break;
2466 case (COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU):
2467 /* TODO: */
2468 ret = -EINVAL;
2469 pr_err("VCRAT not implemented for APU\n");
2470 break;
2471 default:
2472 ret = -EINVAL;
2473 }
2474
2475 if (!ret)
2476 *crat_image = pcrat_image;
2477 else
2478 kvfree(pcrat_image);
2479
2480 return ret;
2481 }
2482
2483
2484 /* kfd_destroy_crat_image
2485 *
2486 * @crat_image: [IN] - crat_image from kfd_create_crat_image_xxx(..)
2487 *
2488 */
kfd_destroy_crat_image(void * crat_image)2489 void kfd_destroy_crat_image(void *crat_image)
2490 {
2491 kvfree(crat_image);
2492 }
2493