Loading...
1/*
2 * Copyright 2015-2017 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
21 */
22
23#include <linux/pci.h>
24#include <linux/acpi.h>
25#include "kfd_crat.h"
26#include "kfd_priv.h"
27#include "kfd_topology.h"
28#include "kfd_iommu.h"
29#include "amdgpu_amdkfd.h"
30
31/* GPU Processor ID base for dGPUs for which VCRAT needs to be created.
32 * GPU processor ID are expressed with Bit[31]=1.
33 * The base is set to 0x8000_0000 + 0x1000 to avoid collision with GPU IDs
34 * used in the CRAT.
35 */
36static uint32_t gpu_processor_id_low = 0x80001000;
37
38/* Return the next available gpu_processor_id and increment it for next GPU
39 * @total_cu_count - Total CUs present in the GPU including ones
40 * masked off
41 */
42static inline unsigned int get_and_inc_gpu_processor_id(
43 unsigned int total_cu_count)
44{
45 int current_id = gpu_processor_id_low;
46
47 gpu_processor_id_low += total_cu_count;
48 return current_id;
49}
50
51/* Static table to describe GPU Cache information */
52struct kfd_gpu_cache_info {
53 uint32_t cache_size;
54 uint32_t cache_level;
55 uint32_t flags;
56 /* Indicates how many Compute Units share this cache
57 * Value = 1 indicates the cache is not shared
58 */
59 uint32_t num_cu_shared;
60};
61
62static struct kfd_gpu_cache_info kaveri_cache_info[] = {
63 {
64 /* TCP L1 Cache per CU */
65 .cache_size = 16,
66 .cache_level = 1,
67 .flags = (CRAT_CACHE_FLAGS_ENABLED |
68 CRAT_CACHE_FLAGS_DATA_CACHE |
69 CRAT_CACHE_FLAGS_SIMD_CACHE),
70 .num_cu_shared = 1,
71
72 },
73 {
74 /* Scalar L1 Instruction Cache (in SQC module) per bank */
75 .cache_size = 16,
76 .cache_level = 1,
77 .flags = (CRAT_CACHE_FLAGS_ENABLED |
78 CRAT_CACHE_FLAGS_INST_CACHE |
79 CRAT_CACHE_FLAGS_SIMD_CACHE),
80 .num_cu_shared = 2,
81 },
82 {
83 /* Scalar L1 Data Cache (in SQC module) per bank */
84 .cache_size = 8,
85 .cache_level = 1,
86 .flags = (CRAT_CACHE_FLAGS_ENABLED |
87 CRAT_CACHE_FLAGS_DATA_CACHE |
88 CRAT_CACHE_FLAGS_SIMD_CACHE),
89 .num_cu_shared = 2,
90 },
91
92 /* TODO: Add L2 Cache information */
93};
94
95
96static struct kfd_gpu_cache_info carrizo_cache_info[] = {
97 {
98 /* TCP L1 Cache per CU */
99 .cache_size = 16,
100 .cache_level = 1,
101 .flags = (CRAT_CACHE_FLAGS_ENABLED |
102 CRAT_CACHE_FLAGS_DATA_CACHE |
103 CRAT_CACHE_FLAGS_SIMD_CACHE),
104 .num_cu_shared = 1,
105 },
106 {
107 /* Scalar L1 Instruction Cache (in SQC module) per bank */
108 .cache_size = 8,
109 .cache_level = 1,
110 .flags = (CRAT_CACHE_FLAGS_ENABLED |
111 CRAT_CACHE_FLAGS_INST_CACHE |
112 CRAT_CACHE_FLAGS_SIMD_CACHE),
113 .num_cu_shared = 4,
114 },
115 {
116 /* Scalar L1 Data Cache (in SQC module) per bank. */
117 .cache_size = 4,
118 .cache_level = 1,
119 .flags = (CRAT_CACHE_FLAGS_ENABLED |
120 CRAT_CACHE_FLAGS_DATA_CACHE |
121 CRAT_CACHE_FLAGS_SIMD_CACHE),
122 .num_cu_shared = 4,
123 },
124
125 /* TODO: Add L2 Cache information */
126};
127
128/* NOTE: In future if more information is added to struct kfd_gpu_cache_info
129 * the following ASICs may need a separate table.
130 */
131#define hawaii_cache_info kaveri_cache_info
132#define tonga_cache_info carrizo_cache_info
133#define fiji_cache_info carrizo_cache_info
134#define polaris10_cache_info carrizo_cache_info
135#define polaris11_cache_info carrizo_cache_info
136#define polaris12_cache_info carrizo_cache_info
137#define vegam_cache_info carrizo_cache_info
138/* TODO - check & update Vega10 cache details */
139#define vega10_cache_info carrizo_cache_info
140#define raven_cache_info carrizo_cache_info
141#define renoir_cache_info carrizo_cache_info
142/* TODO - check & update Navi10 cache details */
143#define navi10_cache_info carrizo_cache_info
144
145static void kfd_populated_cu_info_cpu(struct kfd_topology_device *dev,
146 struct crat_subtype_computeunit *cu)
147{
148 dev->node_props.cpu_cores_count = cu->num_cpu_cores;
149 dev->node_props.cpu_core_id_base = cu->processor_id_low;
150 if (cu->hsa_capability & CRAT_CU_FLAGS_IOMMU_PRESENT)
151 dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
152
153 pr_debug("CU CPU: cores=%d id_base=%d\n", cu->num_cpu_cores,
154 cu->processor_id_low);
155}
156
157static void kfd_populated_cu_info_gpu(struct kfd_topology_device *dev,
158 struct crat_subtype_computeunit *cu)
159{
160 dev->node_props.simd_id_base = cu->processor_id_low;
161 dev->node_props.simd_count = cu->num_simd_cores;
162 dev->node_props.lds_size_in_kb = cu->lds_size_in_kb;
163 dev->node_props.max_waves_per_simd = cu->max_waves_simd;
164 dev->node_props.wave_front_size = cu->wave_front_size;
165 dev->node_props.array_count = cu->array_count;
166 dev->node_props.cu_per_simd_array = cu->num_cu_per_array;
167 dev->node_props.simd_per_cu = cu->num_simd_per_cu;
168 dev->node_props.max_slots_scratch_cu = cu->max_slots_scatch_cu;
169 if (cu->hsa_capability & CRAT_CU_FLAGS_HOT_PLUGGABLE)
170 dev->node_props.capability |= HSA_CAP_HOT_PLUGGABLE;
171 pr_debug("CU GPU: id_base=%d\n", cu->processor_id_low);
172}
173
174/* kfd_parse_subtype_cu - parse compute unit subtypes and attach it to correct
175 * topology device present in the device_list
176 */
177static int kfd_parse_subtype_cu(struct crat_subtype_computeunit *cu,
178 struct list_head *device_list)
179{
180 struct kfd_topology_device *dev;
181
182 pr_debug("Found CU entry in CRAT table with proximity_domain=%d caps=%x\n",
183 cu->proximity_domain, cu->hsa_capability);
184 list_for_each_entry(dev, device_list, list) {
185 if (cu->proximity_domain == dev->proximity_domain) {
186 if (cu->flags & CRAT_CU_FLAGS_CPU_PRESENT)
187 kfd_populated_cu_info_cpu(dev, cu);
188
189 if (cu->flags & CRAT_CU_FLAGS_GPU_PRESENT)
190 kfd_populated_cu_info_gpu(dev, cu);
191 break;
192 }
193 }
194
195 return 0;
196}
197
198static struct kfd_mem_properties *
199find_subtype_mem(uint32_t heap_type, uint32_t flags, uint32_t width,
200 struct kfd_topology_device *dev)
201{
202 struct kfd_mem_properties *props;
203
204 list_for_each_entry(props, &dev->mem_props, list) {
205 if (props->heap_type == heap_type
206 && props->flags == flags
207 && props->width == width)
208 return props;
209 }
210
211 return NULL;
212}
213/* kfd_parse_subtype_mem - parse memory subtypes and attach it to correct
214 * topology device present in the device_list
215 */
216static int kfd_parse_subtype_mem(struct crat_subtype_memory *mem,
217 struct list_head *device_list)
218{
219 struct kfd_mem_properties *props;
220 struct kfd_topology_device *dev;
221 uint32_t heap_type;
222 uint64_t size_in_bytes;
223 uint32_t flags = 0;
224 uint32_t width;
225
226 pr_debug("Found memory entry in CRAT table with proximity_domain=%d\n",
227 mem->proximity_domain);
228 list_for_each_entry(dev, device_list, list) {
229 if (mem->proximity_domain == dev->proximity_domain) {
230 /* We're on GPU node */
231 if (dev->node_props.cpu_cores_count == 0) {
232 /* APU */
233 if (mem->visibility_type == 0)
234 heap_type =
235 HSA_MEM_HEAP_TYPE_FB_PRIVATE;
236 /* dGPU */
237 else
238 heap_type = mem->visibility_type;
239 } else
240 heap_type = HSA_MEM_HEAP_TYPE_SYSTEM;
241
242 if (mem->flags & CRAT_MEM_FLAGS_HOT_PLUGGABLE)
243 flags |= HSA_MEM_FLAGS_HOT_PLUGGABLE;
244 if (mem->flags & CRAT_MEM_FLAGS_NON_VOLATILE)
245 flags |= HSA_MEM_FLAGS_NON_VOLATILE;
246
247 size_in_bytes =
248 ((uint64_t)mem->length_high << 32) +
249 mem->length_low;
250 width = mem->width;
251
252 /* Multiple banks of the same type are aggregated into
253 * one. User mode doesn't care about multiple physical
254 * memory segments. It's managed as a single virtual
255 * heap for user mode.
256 */
257 props = find_subtype_mem(heap_type, flags, width, dev);
258 if (props) {
259 props->size_in_bytes += size_in_bytes;
260 break;
261 }
262
263 props = kfd_alloc_struct(props);
264 if (!props)
265 return -ENOMEM;
266
267 props->heap_type = heap_type;
268 props->flags = flags;
269 props->size_in_bytes = size_in_bytes;
270 props->width = width;
271
272 dev->node_props.mem_banks_count++;
273 list_add_tail(&props->list, &dev->mem_props);
274
275 break;
276 }
277 }
278
279 return 0;
280}
281
282/* kfd_parse_subtype_cache - parse cache subtypes and attach it to correct
283 * topology device present in the device_list
284 */
285static int kfd_parse_subtype_cache(struct crat_subtype_cache *cache,
286 struct list_head *device_list)
287{
288 struct kfd_cache_properties *props;
289 struct kfd_topology_device *dev;
290 uint32_t id;
291 uint32_t total_num_of_cu;
292
293 id = cache->processor_id_low;
294
295 pr_debug("Found cache entry in CRAT table with processor_id=%d\n", id);
296 list_for_each_entry(dev, device_list, list) {
297 total_num_of_cu = (dev->node_props.array_count *
298 dev->node_props.cu_per_simd_array);
299
300 /* Cache infomration in CRAT doesn't have proximity_domain
301 * information as it is associated with a CPU core or GPU
302 * Compute Unit. So map the cache using CPU core Id or SIMD
303 * (GPU) ID.
304 * TODO: This works because currently we can safely assume that
305 * Compute Units are parsed before caches are parsed. In
306 * future, remove this dependency
307 */
308 if ((id >= dev->node_props.cpu_core_id_base &&
309 id <= dev->node_props.cpu_core_id_base +
310 dev->node_props.cpu_cores_count) ||
311 (id >= dev->node_props.simd_id_base &&
312 id < dev->node_props.simd_id_base +
313 total_num_of_cu)) {
314 props = kfd_alloc_struct(props);
315 if (!props)
316 return -ENOMEM;
317
318 props->processor_id_low = id;
319 props->cache_level = cache->cache_level;
320 props->cache_size = cache->cache_size;
321 props->cacheline_size = cache->cache_line_size;
322 props->cachelines_per_tag = cache->lines_per_tag;
323 props->cache_assoc = cache->associativity;
324 props->cache_latency = cache->cache_latency;
325 memcpy(props->sibling_map, cache->sibling_map,
326 sizeof(props->sibling_map));
327
328 if (cache->flags & CRAT_CACHE_FLAGS_DATA_CACHE)
329 props->cache_type |= HSA_CACHE_TYPE_DATA;
330 if (cache->flags & CRAT_CACHE_FLAGS_INST_CACHE)
331 props->cache_type |= HSA_CACHE_TYPE_INSTRUCTION;
332 if (cache->flags & CRAT_CACHE_FLAGS_CPU_CACHE)
333 props->cache_type |= HSA_CACHE_TYPE_CPU;
334 if (cache->flags & CRAT_CACHE_FLAGS_SIMD_CACHE)
335 props->cache_type |= HSA_CACHE_TYPE_HSACU;
336
337 dev->cache_count++;
338 dev->node_props.caches_count++;
339 list_add_tail(&props->list, &dev->cache_props);
340
341 break;
342 }
343 }
344
345 return 0;
346}
347
348/* kfd_parse_subtype_iolink - parse iolink subtypes and attach it to correct
349 * topology device present in the device_list
350 */
351static int kfd_parse_subtype_iolink(struct crat_subtype_iolink *iolink,
352 struct list_head *device_list)
353{
354 struct kfd_iolink_properties *props = NULL, *props2;
355 struct kfd_topology_device *dev, *to_dev;
356 uint32_t id_from;
357 uint32_t id_to;
358
359 id_from = iolink->proximity_domain_from;
360 id_to = iolink->proximity_domain_to;
361
362 pr_debug("Found IO link entry in CRAT table with id_from=%d, id_to %d\n",
363 id_from, id_to);
364 list_for_each_entry(dev, device_list, list) {
365 if (id_from == dev->proximity_domain) {
366 props = kfd_alloc_struct(props);
367 if (!props)
368 return -ENOMEM;
369
370 props->node_from = id_from;
371 props->node_to = id_to;
372 props->ver_maj = iolink->version_major;
373 props->ver_min = iolink->version_minor;
374 props->iolink_type = iolink->io_interface_type;
375
376 if (props->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS)
377 props->weight = 20;
378 else if (props->iolink_type == CRAT_IOLINK_TYPE_XGMI)
379 props->weight = 15 * iolink->num_hops_xgmi;
380 else
381 props->weight = node_distance(id_from, id_to);
382
383 props->min_latency = iolink->minimum_latency;
384 props->max_latency = iolink->maximum_latency;
385 props->min_bandwidth = iolink->minimum_bandwidth_mbs;
386 props->max_bandwidth = iolink->maximum_bandwidth_mbs;
387 props->rec_transfer_size =
388 iolink->recommended_transfer_size;
389
390 dev->io_link_count++;
391 dev->node_props.io_links_count++;
392 list_add_tail(&props->list, &dev->io_link_props);
393 break;
394 }
395 }
396
397 /* CPU topology is created before GPUs are detected, so CPU->GPU
398 * links are not built at that time. If a PCIe type is discovered, it
399 * means a GPU is detected and we are adding GPU->CPU to the topology.
400 * At this time, also add the corresponded CPU->GPU link if GPU
401 * is large bar.
402 * For xGMI, we only added the link with one direction in the crat
403 * table, add corresponded reversed direction link now.
404 */
405 if (props && (iolink->flags & CRAT_IOLINK_FLAGS_BI_DIRECTIONAL)) {
406 to_dev = kfd_topology_device_by_proximity_domain(id_to);
407 if (!to_dev)
408 return -ENODEV;
409 /* same everything but the other direction */
410 props2 = kmemdup(props, sizeof(*props2), GFP_KERNEL);
411 props2->node_from = id_to;
412 props2->node_to = id_from;
413 props2->kobj = NULL;
414 to_dev->io_link_count++;
415 to_dev->node_props.io_links_count++;
416 list_add_tail(&props2->list, &to_dev->io_link_props);
417 }
418
419 return 0;
420}
421
422/* kfd_parse_subtype - parse subtypes and attach it to correct topology device
423 * present in the device_list
424 * @sub_type_hdr - subtype section of crat_image
425 * @device_list - list of topology devices present in this crat_image
426 */
427static int kfd_parse_subtype(struct crat_subtype_generic *sub_type_hdr,
428 struct list_head *device_list)
429{
430 struct crat_subtype_computeunit *cu;
431 struct crat_subtype_memory *mem;
432 struct crat_subtype_cache *cache;
433 struct crat_subtype_iolink *iolink;
434 int ret = 0;
435
436 switch (sub_type_hdr->type) {
437 case CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY:
438 cu = (struct crat_subtype_computeunit *)sub_type_hdr;
439 ret = kfd_parse_subtype_cu(cu, device_list);
440 break;
441 case CRAT_SUBTYPE_MEMORY_AFFINITY:
442 mem = (struct crat_subtype_memory *)sub_type_hdr;
443 ret = kfd_parse_subtype_mem(mem, device_list);
444 break;
445 case CRAT_SUBTYPE_CACHE_AFFINITY:
446 cache = (struct crat_subtype_cache *)sub_type_hdr;
447 ret = kfd_parse_subtype_cache(cache, device_list);
448 break;
449 case CRAT_SUBTYPE_TLB_AFFINITY:
450 /*
451 * For now, nothing to do here
452 */
453 pr_debug("Found TLB entry in CRAT table (not processing)\n");
454 break;
455 case CRAT_SUBTYPE_CCOMPUTE_AFFINITY:
456 /*
457 * For now, nothing to do here
458 */
459 pr_debug("Found CCOMPUTE entry in CRAT table (not processing)\n");
460 break;
461 case CRAT_SUBTYPE_IOLINK_AFFINITY:
462 iolink = (struct crat_subtype_iolink *)sub_type_hdr;
463 ret = kfd_parse_subtype_iolink(iolink, device_list);
464 break;
465 default:
466 pr_warn("Unknown subtype %d in CRAT\n",
467 sub_type_hdr->type);
468 }
469
470 return ret;
471}
472
473/* kfd_parse_crat_table - parse CRAT table. For each node present in CRAT
474 * create a kfd_topology_device and add in to device_list. Also parse
475 * CRAT subtypes and attach it to appropriate kfd_topology_device
476 * @crat_image - input image containing CRAT
477 * @device_list - [OUT] list of kfd_topology_device generated after
478 * parsing crat_image
479 * @proximity_domain - Proximity domain of the first device in the table
480 *
481 * Return - 0 if successful else -ve value
482 */
483int kfd_parse_crat_table(void *crat_image, struct list_head *device_list,
484 uint32_t proximity_domain)
485{
486 struct kfd_topology_device *top_dev = NULL;
487 struct crat_subtype_generic *sub_type_hdr;
488 uint16_t node_id;
489 int ret = 0;
490 struct crat_header *crat_table = (struct crat_header *)crat_image;
491 uint16_t num_nodes;
492 uint32_t image_len;
493
494 if (!crat_image)
495 return -EINVAL;
496
497 if (!list_empty(device_list)) {
498 pr_warn("Error device list should be empty\n");
499 return -EINVAL;
500 }
501
502 num_nodes = crat_table->num_domains;
503 image_len = crat_table->length;
504
505 pr_debug("Parsing CRAT table with %d nodes\n", num_nodes);
506
507 for (node_id = 0; node_id < num_nodes; node_id++) {
508 top_dev = kfd_create_topology_device(device_list);
509 if (!top_dev)
510 break;
511 top_dev->proximity_domain = proximity_domain++;
512 }
513
514 if (!top_dev) {
515 ret = -ENOMEM;
516 goto err;
517 }
518
519 memcpy(top_dev->oem_id, crat_table->oem_id, CRAT_OEMID_LENGTH);
520 memcpy(top_dev->oem_table_id, crat_table->oem_table_id,
521 CRAT_OEMTABLEID_LENGTH);
522 top_dev->oem_revision = crat_table->oem_revision;
523
524 sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1);
525 while ((char *)sub_type_hdr + sizeof(struct crat_subtype_generic) <
526 ((char *)crat_image) + image_len) {
527 if (sub_type_hdr->flags & CRAT_SUBTYPE_FLAGS_ENABLED) {
528 ret = kfd_parse_subtype(sub_type_hdr, device_list);
529 if (ret)
530 break;
531 }
532
533 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
534 sub_type_hdr->length);
535 }
536
537err:
538 if (ret)
539 kfd_release_topology_device_list(device_list);
540
541 return ret;
542}
543
544/* Helper function. See kfd_fill_gpu_cache_info for parameter description */
545static int fill_in_pcache(struct crat_subtype_cache *pcache,
546 struct kfd_gpu_cache_info *pcache_info,
547 struct kfd_cu_info *cu_info,
548 int mem_available,
549 int cu_bitmask,
550 int cache_type, unsigned int cu_processor_id,
551 int cu_block)
552{
553 unsigned int cu_sibling_map_mask;
554 int first_active_cu;
555
556 /* First check if enough memory is available */
557 if (sizeof(struct crat_subtype_cache) > mem_available)
558 return -ENOMEM;
559
560 cu_sibling_map_mask = cu_bitmask;
561 cu_sibling_map_mask >>= cu_block;
562 cu_sibling_map_mask &=
563 ((1 << pcache_info[cache_type].num_cu_shared) - 1);
564 first_active_cu = ffs(cu_sibling_map_mask);
565
566 /* CU could be inactive. In case of shared cache find the first active
567 * CU. and incase of non-shared cache check if the CU is inactive. If
568 * inactive active skip it
569 */
570 if (first_active_cu) {
571 memset(pcache, 0, sizeof(struct crat_subtype_cache));
572 pcache->type = CRAT_SUBTYPE_CACHE_AFFINITY;
573 pcache->length = sizeof(struct crat_subtype_cache);
574 pcache->flags = pcache_info[cache_type].flags;
575 pcache->processor_id_low = cu_processor_id
576 + (first_active_cu - 1);
577 pcache->cache_level = pcache_info[cache_type].cache_level;
578 pcache->cache_size = pcache_info[cache_type].cache_size;
579
580 /* Sibling map is w.r.t processor_id_low, so shift out
581 * inactive CU
582 */
583 cu_sibling_map_mask =
584 cu_sibling_map_mask >> (first_active_cu - 1);
585
586 pcache->sibling_map[0] = (uint8_t)(cu_sibling_map_mask & 0xFF);
587 pcache->sibling_map[1] =
588 (uint8_t)((cu_sibling_map_mask >> 8) & 0xFF);
589 pcache->sibling_map[2] =
590 (uint8_t)((cu_sibling_map_mask >> 16) & 0xFF);
591 pcache->sibling_map[3] =
592 (uint8_t)((cu_sibling_map_mask >> 24) & 0xFF);
593 return 0;
594 }
595 return 1;
596}
597
598/* kfd_fill_gpu_cache_info - Fill GPU cache info using kfd_gpu_cache_info
599 * tables
600 *
601 * @kdev - [IN] GPU device
602 * @gpu_processor_id - [IN] GPU processor ID to which these caches
603 * associate
604 * @available_size - [IN] Amount of memory available in pcache
605 * @cu_info - [IN] Compute Unit info obtained from KGD
606 * @pcache - [OUT] memory into which cache data is to be filled in.
607 * @size_filled - [OUT] amount of data used up in pcache.
608 * @num_of_entries - [OUT] number of caches added
609 */
610static int kfd_fill_gpu_cache_info(struct kfd_dev *kdev,
611 int gpu_processor_id,
612 int available_size,
613 struct kfd_cu_info *cu_info,
614 struct crat_subtype_cache *pcache,
615 int *size_filled,
616 int *num_of_entries)
617{
618 struct kfd_gpu_cache_info *pcache_info;
619 int num_of_cache_types = 0;
620 int i, j, k;
621 int ct = 0;
622 int mem_available = available_size;
623 unsigned int cu_processor_id;
624 int ret;
625
626 switch (kdev->device_info->asic_family) {
627 case CHIP_KAVERI:
628 pcache_info = kaveri_cache_info;
629 num_of_cache_types = ARRAY_SIZE(kaveri_cache_info);
630 break;
631 case CHIP_HAWAII:
632 pcache_info = hawaii_cache_info;
633 num_of_cache_types = ARRAY_SIZE(hawaii_cache_info);
634 break;
635 case CHIP_CARRIZO:
636 pcache_info = carrizo_cache_info;
637 num_of_cache_types = ARRAY_SIZE(carrizo_cache_info);
638 break;
639 case CHIP_TONGA:
640 pcache_info = tonga_cache_info;
641 num_of_cache_types = ARRAY_SIZE(tonga_cache_info);
642 break;
643 case CHIP_FIJI:
644 pcache_info = fiji_cache_info;
645 num_of_cache_types = ARRAY_SIZE(fiji_cache_info);
646 break;
647 case CHIP_POLARIS10:
648 pcache_info = polaris10_cache_info;
649 num_of_cache_types = ARRAY_SIZE(polaris10_cache_info);
650 break;
651 case CHIP_POLARIS11:
652 pcache_info = polaris11_cache_info;
653 num_of_cache_types = ARRAY_SIZE(polaris11_cache_info);
654 break;
655 case CHIP_POLARIS12:
656 pcache_info = polaris12_cache_info;
657 num_of_cache_types = ARRAY_SIZE(polaris12_cache_info);
658 break;
659 case CHIP_VEGAM:
660 pcache_info = vegam_cache_info;
661 num_of_cache_types = ARRAY_SIZE(vegam_cache_info);
662 break;
663 case CHIP_VEGA10:
664 case CHIP_VEGA12:
665 case CHIP_VEGA20:
666 case CHIP_ARCTURUS:
667 pcache_info = vega10_cache_info;
668 num_of_cache_types = ARRAY_SIZE(vega10_cache_info);
669 break;
670 case CHIP_RAVEN:
671 pcache_info = raven_cache_info;
672 num_of_cache_types = ARRAY_SIZE(raven_cache_info);
673 break;
674 case CHIP_RENOIR:
675 pcache_info = renoir_cache_info;
676 num_of_cache_types = ARRAY_SIZE(renoir_cache_info);
677 break;
678 case CHIP_NAVI10:
679 case CHIP_NAVI12:
680 case CHIP_NAVI14:
681 case CHIP_SIENNA_CICHLID:
682 case CHIP_NAVY_FLOUNDER:
683 pcache_info = navi10_cache_info;
684 num_of_cache_types = ARRAY_SIZE(navi10_cache_info);
685 break;
686 default:
687 return -EINVAL;
688 }
689
690 *size_filled = 0;
691 *num_of_entries = 0;
692
693 /* For each type of cache listed in the kfd_gpu_cache_info table,
694 * go through all available Compute Units.
695 * The [i,j,k] loop will
696 * if kfd_gpu_cache_info.num_cu_shared = 1
697 * will parse through all available CU
698 * If (kfd_gpu_cache_info.num_cu_shared != 1)
699 * then it will consider only one CU from
700 * the shared unit
701 */
702
703 for (ct = 0; ct < num_of_cache_types; ct++) {
704 cu_processor_id = gpu_processor_id;
705 for (i = 0; i < cu_info->num_shader_engines; i++) {
706 for (j = 0; j < cu_info->num_shader_arrays_per_engine;
707 j++) {
708 for (k = 0; k < cu_info->num_cu_per_sh;
709 k += pcache_info[ct].num_cu_shared) {
710
711 ret = fill_in_pcache(pcache,
712 pcache_info,
713 cu_info,
714 mem_available,
715 cu_info->cu_bitmap[i % 4][j + i / 4],
716 ct,
717 cu_processor_id,
718 k);
719
720 if (ret < 0)
721 break;
722
723 if (!ret) {
724 pcache++;
725 (*num_of_entries)++;
726 mem_available -=
727 sizeof(*pcache);
728 (*size_filled) +=
729 sizeof(*pcache);
730 }
731
732 /* Move to next CU block */
733 cu_processor_id +=
734 pcache_info[ct].num_cu_shared;
735 }
736 }
737 }
738 }
739
740 pr_debug("Added [%d] GPU cache entries\n", *num_of_entries);
741
742 return 0;
743}
744
745/*
746 * kfd_create_crat_image_acpi - Allocates memory for CRAT image and
747 * copies CRAT from ACPI (if available).
748 * NOTE: Call kfd_destroy_crat_image to free CRAT image memory
749 *
750 * @crat_image: CRAT read from ACPI. If no CRAT in ACPI then
751 * crat_image will be NULL
752 * @size: [OUT] size of crat_image
753 *
754 * Return 0 if successful else return error code
755 */
756int kfd_create_crat_image_acpi(void **crat_image, size_t *size)
757{
758 struct acpi_table_header *crat_table;
759 acpi_status status;
760 void *pcrat_image;
761
762 if (!crat_image)
763 return -EINVAL;
764
765 *crat_image = NULL;
766
767 /* Fetch the CRAT table from ACPI */
768 status = acpi_get_table(CRAT_SIGNATURE, 0, &crat_table);
769 if (status == AE_NOT_FOUND) {
770 pr_warn("CRAT table not found\n");
771 return -ENODATA;
772 } else if (ACPI_FAILURE(status)) {
773 const char *err = acpi_format_exception(status);
774
775 pr_err("CRAT table error: %s\n", err);
776 return -EINVAL;
777 }
778
779 if (ignore_crat) {
780 pr_info("CRAT table disabled by module option\n");
781 return -ENODATA;
782 }
783
784 pcrat_image = kmemdup(crat_table, crat_table->length, GFP_KERNEL);
785 if (!pcrat_image)
786 return -ENOMEM;
787
788 *crat_image = pcrat_image;
789 *size = crat_table->length;
790
791 return 0;
792}
793
794/* Memory required to create Virtual CRAT.
795 * Since there is no easy way to predict the amount of memory required, the
796 * following amount are allocated for CPU and GPU Virtual CRAT. This is
797 * expected to cover all known conditions. But to be safe additional check
798 * is put in the code to ensure we don't overwrite.
799 */
800#define VCRAT_SIZE_FOR_CPU (2 * PAGE_SIZE)
801#define VCRAT_SIZE_FOR_GPU (4 * PAGE_SIZE)
802
803/* kfd_fill_cu_for_cpu - Fill in Compute info for the given CPU NUMA node
804 *
805 * @numa_node_id: CPU NUMA node id
806 * @avail_size: Available size in the memory
807 * @sub_type_hdr: Memory into which compute info will be filled in
808 *
809 * Return 0 if successful else return -ve value
810 */
811static int kfd_fill_cu_for_cpu(int numa_node_id, int *avail_size,
812 int proximity_domain,
813 struct crat_subtype_computeunit *sub_type_hdr)
814{
815 const struct cpumask *cpumask;
816
817 *avail_size -= sizeof(struct crat_subtype_computeunit);
818 if (*avail_size < 0)
819 return -ENOMEM;
820
821 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit));
822
823 /* Fill in subtype header data */
824 sub_type_hdr->type = CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY;
825 sub_type_hdr->length = sizeof(struct crat_subtype_computeunit);
826 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
827
828 cpumask = cpumask_of_node(numa_node_id);
829
830 /* Fill in CU data */
831 sub_type_hdr->flags |= CRAT_CU_FLAGS_CPU_PRESENT;
832 sub_type_hdr->proximity_domain = proximity_domain;
833 sub_type_hdr->processor_id_low = kfd_numa_node_to_apic_id(numa_node_id);
834 if (sub_type_hdr->processor_id_low == -1)
835 return -EINVAL;
836
837 sub_type_hdr->num_cpu_cores = cpumask_weight(cpumask);
838
839 return 0;
840}
841
842/* kfd_fill_mem_info_for_cpu - Fill in Memory info for the given CPU NUMA node
843 *
844 * @numa_node_id: CPU NUMA node id
845 * @avail_size: Available size in the memory
846 * @sub_type_hdr: Memory into which compute info will be filled in
847 *
848 * Return 0 if successful else return -ve value
849 */
850static int kfd_fill_mem_info_for_cpu(int numa_node_id, int *avail_size,
851 int proximity_domain,
852 struct crat_subtype_memory *sub_type_hdr)
853{
854 uint64_t mem_in_bytes = 0;
855 pg_data_t *pgdat;
856 int zone_type;
857
858 *avail_size -= sizeof(struct crat_subtype_memory);
859 if (*avail_size < 0)
860 return -ENOMEM;
861
862 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_memory));
863
864 /* Fill in subtype header data */
865 sub_type_hdr->type = CRAT_SUBTYPE_MEMORY_AFFINITY;
866 sub_type_hdr->length = sizeof(struct crat_subtype_memory);
867 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
868
869 /* Fill in Memory Subunit data */
870
871 /* Unlike si_meminfo, si_meminfo_node is not exported. So
872 * the following lines are duplicated from si_meminfo_node
873 * function
874 */
875 pgdat = NODE_DATA(numa_node_id);
876 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
877 mem_in_bytes += zone_managed_pages(&pgdat->node_zones[zone_type]);
878 mem_in_bytes <<= PAGE_SHIFT;
879
880 sub_type_hdr->length_low = lower_32_bits(mem_in_bytes);
881 sub_type_hdr->length_high = upper_32_bits(mem_in_bytes);
882 sub_type_hdr->proximity_domain = proximity_domain;
883
884 return 0;
885}
886
887#ifdef CONFIG_X86_64
888static int kfd_fill_iolink_info_for_cpu(int numa_node_id, int *avail_size,
889 uint32_t *num_entries,
890 struct crat_subtype_iolink *sub_type_hdr)
891{
892 int nid;
893 struct cpuinfo_x86 *c = &cpu_data(0);
894 uint8_t link_type;
895
896 if (c->x86_vendor == X86_VENDOR_AMD)
897 link_type = CRAT_IOLINK_TYPE_HYPERTRANSPORT;
898 else
899 link_type = CRAT_IOLINK_TYPE_QPI_1_1;
900
901 *num_entries = 0;
902
903 /* Create IO links from this node to other CPU nodes */
904 for_each_online_node(nid) {
905 if (nid == numa_node_id) /* node itself */
906 continue;
907
908 *avail_size -= sizeof(struct crat_subtype_iolink);
909 if (*avail_size < 0)
910 return -ENOMEM;
911
912 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
913
914 /* Fill in subtype header data */
915 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
916 sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
917 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
918
919 /* Fill in IO link data */
920 sub_type_hdr->proximity_domain_from = numa_node_id;
921 sub_type_hdr->proximity_domain_to = nid;
922 sub_type_hdr->io_interface_type = link_type;
923
924 (*num_entries)++;
925 sub_type_hdr++;
926 }
927
928 return 0;
929}
930#endif
931
932/* kfd_create_vcrat_image_cpu - Create Virtual CRAT for CPU
933 *
934 * @pcrat_image: Fill in VCRAT for CPU
935 * @size: [IN] allocated size of crat_image.
936 * [OUT] actual size of data filled in crat_image
937 */
938static int kfd_create_vcrat_image_cpu(void *pcrat_image, size_t *size)
939{
940 struct crat_header *crat_table = (struct crat_header *)pcrat_image;
941 struct acpi_table_header *acpi_table;
942 acpi_status status;
943 struct crat_subtype_generic *sub_type_hdr;
944 int avail_size = *size;
945 int numa_node_id;
946#ifdef CONFIG_X86_64
947 uint32_t entries = 0;
948#endif
949 int ret = 0;
950
951 if (!pcrat_image || avail_size < VCRAT_SIZE_FOR_CPU)
952 return -EINVAL;
953
954 /* Fill in CRAT Header.
955 * Modify length and total_entries as subunits are added.
956 */
957 avail_size -= sizeof(struct crat_header);
958 if (avail_size < 0)
959 return -ENOMEM;
960
961 memset(crat_table, 0, sizeof(struct crat_header));
962 memcpy(&crat_table->signature, CRAT_SIGNATURE,
963 sizeof(crat_table->signature));
964 crat_table->length = sizeof(struct crat_header);
965
966 status = acpi_get_table("DSDT", 0, &acpi_table);
967 if (status != AE_OK)
968 pr_warn("DSDT table not found for OEM information\n");
969 else {
970 crat_table->oem_revision = acpi_table->revision;
971 memcpy(crat_table->oem_id, acpi_table->oem_id,
972 CRAT_OEMID_LENGTH);
973 memcpy(crat_table->oem_table_id, acpi_table->oem_table_id,
974 CRAT_OEMTABLEID_LENGTH);
975 }
976 crat_table->total_entries = 0;
977 crat_table->num_domains = 0;
978
979 sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1);
980
981 for_each_online_node(numa_node_id) {
982 if (kfd_numa_node_to_apic_id(numa_node_id) == -1)
983 continue;
984
985 /* Fill in Subtype: Compute Unit */
986 ret = kfd_fill_cu_for_cpu(numa_node_id, &avail_size,
987 crat_table->num_domains,
988 (struct crat_subtype_computeunit *)sub_type_hdr);
989 if (ret < 0)
990 return ret;
991 crat_table->length += sub_type_hdr->length;
992 crat_table->total_entries++;
993
994 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
995 sub_type_hdr->length);
996
997 /* Fill in Subtype: Memory */
998 ret = kfd_fill_mem_info_for_cpu(numa_node_id, &avail_size,
999 crat_table->num_domains,
1000 (struct crat_subtype_memory *)sub_type_hdr);
1001 if (ret < 0)
1002 return ret;
1003 crat_table->length += sub_type_hdr->length;
1004 crat_table->total_entries++;
1005
1006 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1007 sub_type_hdr->length);
1008
1009 /* Fill in Subtype: IO Link */
1010#ifdef CONFIG_X86_64
1011 ret = kfd_fill_iolink_info_for_cpu(numa_node_id, &avail_size,
1012 &entries,
1013 (struct crat_subtype_iolink *)sub_type_hdr);
1014 if (ret < 0)
1015 return ret;
1016 crat_table->length += (sub_type_hdr->length * entries);
1017 crat_table->total_entries += entries;
1018
1019 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1020 sub_type_hdr->length * entries);
1021#else
1022 pr_info("IO link not available for non x86 platforms\n");
1023#endif
1024
1025 crat_table->num_domains++;
1026 }
1027
1028 /* TODO: Add cache Subtype for CPU.
1029 * Currently, CPU cache information is available in function
1030 * detect_cache_attributes(cpu) defined in the file
1031 * ./arch/x86/kernel/cpu/intel_cacheinfo.c. This function is not
1032 * exported and to get the same information the code needs to be
1033 * duplicated.
1034 */
1035
1036 *size = crat_table->length;
1037 pr_info("Virtual CRAT table created for CPU\n");
1038
1039 return 0;
1040}
1041
1042static int kfd_fill_gpu_memory_affinity(int *avail_size,
1043 struct kfd_dev *kdev, uint8_t type, uint64_t size,
1044 struct crat_subtype_memory *sub_type_hdr,
1045 uint32_t proximity_domain,
1046 const struct kfd_local_mem_info *local_mem_info)
1047{
1048 *avail_size -= sizeof(struct crat_subtype_memory);
1049 if (*avail_size < 0)
1050 return -ENOMEM;
1051
1052 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_memory));
1053 sub_type_hdr->type = CRAT_SUBTYPE_MEMORY_AFFINITY;
1054 sub_type_hdr->length = sizeof(struct crat_subtype_memory);
1055 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED;
1056
1057 sub_type_hdr->proximity_domain = proximity_domain;
1058
1059 pr_debug("Fill gpu memory affinity - type 0x%x size 0x%llx\n",
1060 type, size);
1061
1062 sub_type_hdr->length_low = lower_32_bits(size);
1063 sub_type_hdr->length_high = upper_32_bits(size);
1064
1065 sub_type_hdr->width = local_mem_info->vram_width;
1066 sub_type_hdr->visibility_type = type;
1067
1068 return 0;
1069}
1070
1071/* kfd_fill_gpu_direct_io_link - Fill in direct io link from GPU
1072 * to its NUMA node
1073 * @avail_size: Available size in the memory
1074 * @kdev - [IN] GPU device
1075 * @sub_type_hdr: Memory into which io link info will be filled in
1076 * @proximity_domain - proximity domain of the GPU node
1077 *
1078 * Return 0 if successful else return -ve value
1079 */
1080static int kfd_fill_gpu_direct_io_link_to_cpu(int *avail_size,
1081 struct kfd_dev *kdev,
1082 struct crat_subtype_iolink *sub_type_hdr,
1083 uint32_t proximity_domain)
1084{
1085 *avail_size -= sizeof(struct crat_subtype_iolink);
1086 if (*avail_size < 0)
1087 return -ENOMEM;
1088
1089 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
1090
1091 /* Fill in subtype header data */
1092 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
1093 sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
1094 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED;
1095 if (kfd_dev_is_large_bar(kdev))
1096 sub_type_hdr->flags |= CRAT_IOLINK_FLAGS_BI_DIRECTIONAL;
1097
1098 /* Fill in IOLINK subtype.
1099 * TODO: Fill-in other fields of iolink subtype
1100 */
1101 sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_PCIEXPRESS;
1102 sub_type_hdr->proximity_domain_from = proximity_domain;
1103#ifdef CONFIG_NUMA
1104 if (kdev->pdev->dev.numa_node == NUMA_NO_NODE)
1105 sub_type_hdr->proximity_domain_to = 0;
1106 else
1107 sub_type_hdr->proximity_domain_to = kdev->pdev->dev.numa_node;
1108#else
1109 sub_type_hdr->proximity_domain_to = 0;
1110#endif
1111 return 0;
1112}
1113
1114static int kfd_fill_gpu_xgmi_link_to_gpu(int *avail_size,
1115 struct kfd_dev *kdev,
1116 struct kfd_dev *peer_kdev,
1117 struct crat_subtype_iolink *sub_type_hdr,
1118 uint32_t proximity_domain_from,
1119 uint32_t proximity_domain_to)
1120{
1121 *avail_size -= sizeof(struct crat_subtype_iolink);
1122 if (*avail_size < 0)
1123 return -ENOMEM;
1124
1125 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
1126
1127 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
1128 sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
1129 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED |
1130 CRAT_IOLINK_FLAGS_BI_DIRECTIONAL;
1131
1132 sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_XGMI;
1133 sub_type_hdr->proximity_domain_from = proximity_domain_from;
1134 sub_type_hdr->proximity_domain_to = proximity_domain_to;
1135 sub_type_hdr->num_hops_xgmi =
1136 amdgpu_amdkfd_get_xgmi_hops_count(kdev->kgd, peer_kdev->kgd);
1137 return 0;
1138}
1139
1140/* kfd_create_vcrat_image_gpu - Create Virtual CRAT for CPU
1141 *
1142 * @pcrat_image: Fill in VCRAT for GPU
1143 * @size: [IN] allocated size of crat_image.
1144 * [OUT] actual size of data filled in crat_image
1145 */
1146static int kfd_create_vcrat_image_gpu(void *pcrat_image,
1147 size_t *size, struct kfd_dev *kdev,
1148 uint32_t proximity_domain)
1149{
1150 struct crat_header *crat_table = (struct crat_header *)pcrat_image;
1151 struct crat_subtype_generic *sub_type_hdr;
1152 struct kfd_local_mem_info local_mem_info;
1153 struct kfd_topology_device *peer_dev;
1154 struct crat_subtype_computeunit *cu;
1155 struct kfd_cu_info cu_info;
1156 int avail_size = *size;
1157 uint32_t total_num_of_cu;
1158 int num_of_cache_entries = 0;
1159 int cache_mem_filled = 0;
1160 uint32_t nid = 0;
1161 int ret = 0;
1162
1163 if (!pcrat_image || avail_size < VCRAT_SIZE_FOR_GPU)
1164 return -EINVAL;
1165
1166 /* Fill the CRAT Header.
1167 * Modify length and total_entries as subunits are added.
1168 */
1169 avail_size -= sizeof(struct crat_header);
1170 if (avail_size < 0)
1171 return -ENOMEM;
1172
1173 memset(crat_table, 0, sizeof(struct crat_header));
1174
1175 memcpy(&crat_table->signature, CRAT_SIGNATURE,
1176 sizeof(crat_table->signature));
1177 /* Change length as we add more subtypes*/
1178 crat_table->length = sizeof(struct crat_header);
1179 crat_table->num_domains = 1;
1180 crat_table->total_entries = 0;
1181
1182 /* Fill in Subtype: Compute Unit
1183 * First fill in the sub type header and then sub type data
1184 */
1185 avail_size -= sizeof(struct crat_subtype_computeunit);
1186 if (avail_size < 0)
1187 return -ENOMEM;
1188
1189 sub_type_hdr = (struct crat_subtype_generic *)(crat_table + 1);
1190 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit));
1191
1192 sub_type_hdr->type = CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY;
1193 sub_type_hdr->length = sizeof(struct crat_subtype_computeunit);
1194 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
1195
1196 /* Fill CU subtype data */
1197 cu = (struct crat_subtype_computeunit *)sub_type_hdr;
1198 cu->flags |= CRAT_CU_FLAGS_GPU_PRESENT;
1199 cu->proximity_domain = proximity_domain;
1200
1201 amdgpu_amdkfd_get_cu_info(kdev->kgd, &cu_info);
1202 cu->num_simd_per_cu = cu_info.simd_per_cu;
1203 cu->num_simd_cores = cu_info.simd_per_cu * cu_info.cu_active_number;
1204 cu->max_waves_simd = cu_info.max_waves_per_simd;
1205
1206 cu->wave_front_size = cu_info.wave_front_size;
1207 cu->array_count = cu_info.num_shader_arrays_per_engine *
1208 cu_info.num_shader_engines;
1209 total_num_of_cu = (cu->array_count * cu_info.num_cu_per_sh);
1210 cu->processor_id_low = get_and_inc_gpu_processor_id(total_num_of_cu);
1211 cu->num_cu_per_array = cu_info.num_cu_per_sh;
1212 cu->max_slots_scatch_cu = cu_info.max_scratch_slots_per_cu;
1213 cu->num_banks = cu_info.num_shader_engines;
1214 cu->lds_size_in_kb = cu_info.lds_size;
1215
1216 cu->hsa_capability = 0;
1217
1218 /* Check if this node supports IOMMU. During parsing this flag will
1219 * translate to HSA_CAP_ATS_PRESENT
1220 */
1221 if (!kfd_iommu_check_device(kdev))
1222 cu->hsa_capability |= CRAT_CU_FLAGS_IOMMU_PRESENT;
1223
1224 crat_table->length += sub_type_hdr->length;
1225 crat_table->total_entries++;
1226
1227 /* Fill in Subtype: Memory. Only on systems with large BAR (no
1228 * private FB), report memory as public. On other systems
1229 * report the total FB size (public+private) as a single
1230 * private heap.
1231 */
1232 amdgpu_amdkfd_get_local_mem_info(kdev->kgd, &local_mem_info);
1233 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1234 sub_type_hdr->length);
1235
1236 if (debug_largebar)
1237 local_mem_info.local_mem_size_private = 0;
1238
1239 if (local_mem_info.local_mem_size_private == 0)
1240 ret = kfd_fill_gpu_memory_affinity(&avail_size,
1241 kdev, HSA_MEM_HEAP_TYPE_FB_PUBLIC,
1242 local_mem_info.local_mem_size_public,
1243 (struct crat_subtype_memory *)sub_type_hdr,
1244 proximity_domain,
1245 &local_mem_info);
1246 else
1247 ret = kfd_fill_gpu_memory_affinity(&avail_size,
1248 kdev, HSA_MEM_HEAP_TYPE_FB_PRIVATE,
1249 local_mem_info.local_mem_size_public +
1250 local_mem_info.local_mem_size_private,
1251 (struct crat_subtype_memory *)sub_type_hdr,
1252 proximity_domain,
1253 &local_mem_info);
1254 if (ret < 0)
1255 return ret;
1256
1257 crat_table->length += sizeof(struct crat_subtype_memory);
1258 crat_table->total_entries++;
1259
1260 /* TODO: Fill in cache information. This information is NOT readily
1261 * available in KGD
1262 */
1263 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1264 sub_type_hdr->length);
1265 ret = kfd_fill_gpu_cache_info(kdev, cu->processor_id_low,
1266 avail_size,
1267 &cu_info,
1268 (struct crat_subtype_cache *)sub_type_hdr,
1269 &cache_mem_filled,
1270 &num_of_cache_entries);
1271
1272 if (ret < 0)
1273 return ret;
1274
1275 crat_table->length += cache_mem_filled;
1276 crat_table->total_entries += num_of_cache_entries;
1277 avail_size -= cache_mem_filled;
1278
1279 /* Fill in Subtype: IO_LINKS
1280 * Only direct links are added here which is Link from GPU to
1281 * to its NUMA node. Indirect links are added by userspace.
1282 */
1283 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1284 cache_mem_filled);
1285 ret = kfd_fill_gpu_direct_io_link_to_cpu(&avail_size, kdev,
1286 (struct crat_subtype_iolink *)sub_type_hdr, proximity_domain);
1287
1288 if (ret < 0)
1289 return ret;
1290
1291 crat_table->length += sub_type_hdr->length;
1292 crat_table->total_entries++;
1293
1294
1295 /* Fill in Subtype: IO_LINKS
1296 * Direct links from GPU to other GPUs through xGMI.
1297 * We will loop GPUs that already be processed (with lower value
1298 * of proximity_domain), add the link for the GPUs with same
1299 * hive id (from this GPU to other GPU) . The reversed iolink
1300 * (from other GPU to this GPU) will be added
1301 * in kfd_parse_subtype_iolink.
1302 */
1303 if (kdev->hive_id) {
1304 for (nid = 0; nid < proximity_domain; ++nid) {
1305 peer_dev = kfd_topology_device_by_proximity_domain(nid);
1306 if (!peer_dev->gpu)
1307 continue;
1308 if (peer_dev->gpu->hive_id != kdev->hive_id)
1309 continue;
1310 sub_type_hdr = (typeof(sub_type_hdr))(
1311 (char *)sub_type_hdr +
1312 sizeof(struct crat_subtype_iolink));
1313 ret = kfd_fill_gpu_xgmi_link_to_gpu(
1314 &avail_size, kdev, peer_dev->gpu,
1315 (struct crat_subtype_iolink *)sub_type_hdr,
1316 proximity_domain, nid);
1317 if (ret < 0)
1318 return ret;
1319 crat_table->length += sub_type_hdr->length;
1320 crat_table->total_entries++;
1321 }
1322 }
1323 *size = crat_table->length;
1324 pr_info("Virtual CRAT table created for GPU\n");
1325
1326 return ret;
1327}
1328
1329/* kfd_create_crat_image_virtual - Allocates memory for CRAT image and
1330 * creates a Virtual CRAT (VCRAT) image
1331 *
1332 * NOTE: Call kfd_destroy_crat_image to free CRAT image memory
1333 *
1334 * @crat_image: VCRAT image created because ACPI does not have a
1335 * CRAT for this device
1336 * @size: [OUT] size of virtual crat_image
1337 * @flags: COMPUTE_UNIT_CPU - Create VCRAT for CPU device
1338 * COMPUTE_UNIT_GPU - Create VCRAT for GPU
1339 * (COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU) - Create VCRAT for APU
1340 * -- this option is not currently implemented.
1341 * The assumption is that all AMD APUs will have CRAT
1342 * @kdev: Valid kfd_device required if flags contain COMPUTE_UNIT_GPU
1343 *
1344 * Return 0 if successful else return -ve value
1345 */
1346int kfd_create_crat_image_virtual(void **crat_image, size_t *size,
1347 int flags, struct kfd_dev *kdev,
1348 uint32_t proximity_domain)
1349{
1350 void *pcrat_image = NULL;
1351 int ret = 0;
1352
1353 if (!crat_image)
1354 return -EINVAL;
1355
1356 *crat_image = NULL;
1357
1358 /* Allocate one VCRAT_SIZE_FOR_CPU for CPU virtual CRAT image and
1359 * VCRAT_SIZE_FOR_GPU for GPU virtual CRAT image. This should cover
1360 * all the current conditions. A check is put not to overwrite beyond
1361 * allocated size
1362 */
1363 switch (flags) {
1364 case COMPUTE_UNIT_CPU:
1365 pcrat_image = kmalloc(VCRAT_SIZE_FOR_CPU, GFP_KERNEL);
1366 if (!pcrat_image)
1367 return -ENOMEM;
1368 *size = VCRAT_SIZE_FOR_CPU;
1369 ret = kfd_create_vcrat_image_cpu(pcrat_image, size);
1370 break;
1371 case COMPUTE_UNIT_GPU:
1372 if (!kdev)
1373 return -EINVAL;
1374 pcrat_image = kmalloc(VCRAT_SIZE_FOR_GPU, GFP_KERNEL);
1375 if (!pcrat_image)
1376 return -ENOMEM;
1377 *size = VCRAT_SIZE_FOR_GPU;
1378 ret = kfd_create_vcrat_image_gpu(pcrat_image, size, kdev,
1379 proximity_domain);
1380 break;
1381 case (COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU):
1382 /* TODO: */
1383 ret = -EINVAL;
1384 pr_err("VCRAT not implemented for APU\n");
1385 break;
1386 default:
1387 ret = -EINVAL;
1388 }
1389
1390 if (!ret)
1391 *crat_image = pcrat_image;
1392 else
1393 kfree(pcrat_image);
1394
1395 return ret;
1396}
1397
1398
1399/* kfd_destroy_crat_image
1400 *
1401 * @crat_image: [IN] - crat_image from kfd_create_crat_image_xxx(..)
1402 *
1403 */
1404void kfd_destroy_crat_image(void *crat_image)
1405{
1406 kfree(crat_image);
1407}
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
32/* GPU Processor ID base for dGPUs for which VCRAT needs to be created.
33 * GPU processor ID are expressed with Bit[31]=1.
34 * The base is set to 0x8000_0000 + 0x1000 to avoid collision with GPU IDs
35 * used in the CRAT.
36 */
37static uint32_t gpu_processor_id_low = 0x80001000;
38
39/* Return the next available gpu_processor_id and increment it for next GPU
40 * @total_cu_count - Total CUs present in the GPU including ones
41 * masked off
42 */
43static inline unsigned int get_and_inc_gpu_processor_id(
44 unsigned int total_cu_count)
45{
46 int current_id = gpu_processor_id_low;
47
48 gpu_processor_id_low += total_cu_count;
49 return current_id;
50}
51
52
53static struct kfd_gpu_cache_info kaveri_cache_info[] = {
54 {
55 /* TCP L1 Cache per CU */
56 .cache_size = 16,
57 .cache_level = 1,
58 .cache_line_size = 64,
59 .flags = (CRAT_CACHE_FLAGS_ENABLED |
60 CRAT_CACHE_FLAGS_DATA_CACHE |
61 CRAT_CACHE_FLAGS_SIMD_CACHE),
62 .num_cu_shared = 1,
63 },
64 {
65 /* Scalar L1 Instruction Cache (in SQC module) per bank */
66 .cache_size = 16,
67 .cache_level = 1,
68 .cache_line_size = 64,
69 .flags = (CRAT_CACHE_FLAGS_ENABLED |
70 CRAT_CACHE_FLAGS_INST_CACHE |
71 CRAT_CACHE_FLAGS_SIMD_CACHE),
72 .num_cu_shared = 2,
73 },
74 {
75 /* Scalar L1 Data Cache (in SQC module) per bank */
76 .cache_size = 8,
77 .cache_level = 1,
78 .cache_line_size = 64,
79 .flags = (CRAT_CACHE_FLAGS_ENABLED |
80 CRAT_CACHE_FLAGS_DATA_CACHE |
81 CRAT_CACHE_FLAGS_SIMD_CACHE),
82 .num_cu_shared = 2,
83 },
84
85 /* TODO: Add L2 Cache information */
86};
87
88
89static struct kfd_gpu_cache_info carrizo_cache_info[] = {
90 {
91 /* TCP L1 Cache per CU */
92 .cache_size = 16,
93 .cache_level = 1,
94 .cache_line_size = 64,
95 .flags = (CRAT_CACHE_FLAGS_ENABLED |
96 CRAT_CACHE_FLAGS_DATA_CACHE |
97 CRAT_CACHE_FLAGS_SIMD_CACHE),
98 .num_cu_shared = 1,
99 },
100 {
101 /* Scalar L1 Instruction Cache (in SQC module) per bank */
102 .cache_size = 32,
103 .cache_level = 1,
104 .cache_line_size = 64,
105 .flags = (CRAT_CACHE_FLAGS_ENABLED |
106 CRAT_CACHE_FLAGS_INST_CACHE |
107 CRAT_CACHE_FLAGS_SIMD_CACHE),
108 .num_cu_shared = 4,
109 },
110 {
111 /* Scalar L1 Data Cache (in SQC module) per bank. */
112 .cache_size = 16,
113 .cache_level = 1,
114 .cache_line_size = 64,
115 .flags = (CRAT_CACHE_FLAGS_ENABLED |
116 CRAT_CACHE_FLAGS_DATA_CACHE |
117 CRAT_CACHE_FLAGS_SIMD_CACHE),
118 .num_cu_shared = 4,
119 },
120
121 /* TODO: Add L2 Cache information */
122};
123
124#define hawaii_cache_info kaveri_cache_info
125#define tonga_cache_info carrizo_cache_info
126#define fiji_cache_info carrizo_cache_info
127#define polaris10_cache_info carrizo_cache_info
128#define polaris11_cache_info carrizo_cache_info
129#define polaris12_cache_info carrizo_cache_info
130#define vegam_cache_info carrizo_cache_info
131
132/* NOTE: L1 cache information has been updated and L2/L3
133 * cache information has been added for Vega10 and
134 * newer ASICs. The unit for cache_size is KiB.
135 * In future, check & update cache details
136 * for every new ASIC is required.
137 */
138
139static struct kfd_gpu_cache_info vega10_cache_info[] = {
140 {
141 /* TCP L1 Cache per CU */
142 .cache_size = 16,
143 .cache_level = 1,
144 .cache_line_size = 64,
145 .flags = (CRAT_CACHE_FLAGS_ENABLED |
146 CRAT_CACHE_FLAGS_DATA_CACHE |
147 CRAT_CACHE_FLAGS_SIMD_CACHE),
148 .num_cu_shared = 1,
149 },
150 {
151 /* Scalar L1 Instruction Cache per SQC */
152 .cache_size = 32,
153 .cache_level = 1,
154 .cache_line_size = 64,
155 .flags = (CRAT_CACHE_FLAGS_ENABLED |
156 CRAT_CACHE_FLAGS_INST_CACHE |
157 CRAT_CACHE_FLAGS_SIMD_CACHE),
158 .num_cu_shared = 3,
159 },
160 {
161 /* Scalar L1 Data Cache per SQC */
162 .cache_size = 16,
163 .cache_level = 1,
164 .cache_line_size = 64,
165 .flags = (CRAT_CACHE_FLAGS_ENABLED |
166 CRAT_CACHE_FLAGS_DATA_CACHE |
167 CRAT_CACHE_FLAGS_SIMD_CACHE),
168 .num_cu_shared = 3,
169 },
170 {
171 /* L2 Data Cache per GPU (Total Tex Cache) */
172 .cache_size = 4096,
173 .cache_level = 2,
174 .cache_line_size = 64,
175 .flags = (CRAT_CACHE_FLAGS_ENABLED |
176 CRAT_CACHE_FLAGS_DATA_CACHE |
177 CRAT_CACHE_FLAGS_SIMD_CACHE),
178 .num_cu_shared = 16,
179 },
180};
181
182static struct kfd_gpu_cache_info raven_cache_info[] = {
183 {
184 /* TCP L1 Cache per CU */
185 .cache_size = 16,
186 .cache_level = 1,
187 .cache_line_size = 64,
188 .flags = (CRAT_CACHE_FLAGS_ENABLED |
189 CRAT_CACHE_FLAGS_DATA_CACHE |
190 CRAT_CACHE_FLAGS_SIMD_CACHE),
191 .num_cu_shared = 1,
192 },
193 {
194 /* Scalar L1 Instruction Cache per SQC */
195 .cache_size = 32,
196 .cache_level = 1,
197 .cache_line_size = 64,
198 .flags = (CRAT_CACHE_FLAGS_ENABLED |
199 CRAT_CACHE_FLAGS_INST_CACHE |
200 CRAT_CACHE_FLAGS_SIMD_CACHE),
201 .num_cu_shared = 3,
202 },
203 {
204 /* Scalar L1 Data Cache per SQC */
205 .cache_size = 16,
206 .cache_level = 1,
207 .cache_line_size = 64,
208 .flags = (CRAT_CACHE_FLAGS_ENABLED |
209 CRAT_CACHE_FLAGS_DATA_CACHE |
210 CRAT_CACHE_FLAGS_SIMD_CACHE),
211 .num_cu_shared = 3,
212 },
213 {
214 /* L2 Data Cache per GPU (Total Tex Cache) */
215 .cache_size = 1024,
216 .cache_level = 2,
217 .cache_line_size = 64,
218 .flags = (CRAT_CACHE_FLAGS_ENABLED |
219 CRAT_CACHE_FLAGS_DATA_CACHE |
220 CRAT_CACHE_FLAGS_SIMD_CACHE),
221 .num_cu_shared = 11,
222 },
223};
224
225static struct kfd_gpu_cache_info renoir_cache_info[] = {
226 {
227 /* TCP L1 Cache per CU */
228 .cache_size = 16,
229 .cache_level = 1,
230 .cache_line_size = 64,
231 .flags = (CRAT_CACHE_FLAGS_ENABLED |
232 CRAT_CACHE_FLAGS_DATA_CACHE |
233 CRAT_CACHE_FLAGS_SIMD_CACHE),
234 .num_cu_shared = 1,
235 },
236 {
237 /* Scalar L1 Instruction Cache per SQC */
238 .cache_size = 32,
239 .cache_level = 1,
240 .cache_line_size = 64,
241 .flags = (CRAT_CACHE_FLAGS_ENABLED |
242 CRAT_CACHE_FLAGS_INST_CACHE |
243 CRAT_CACHE_FLAGS_SIMD_CACHE),
244 .num_cu_shared = 3,
245 },
246 {
247 /* Scalar L1 Data Cache per SQC */
248 .cache_size = 16,
249 .cache_level = 1,
250 .cache_line_size = 64,
251 .flags = (CRAT_CACHE_FLAGS_ENABLED |
252 CRAT_CACHE_FLAGS_DATA_CACHE |
253 CRAT_CACHE_FLAGS_SIMD_CACHE),
254 .num_cu_shared = 3,
255 },
256 {
257 /* L2 Data Cache per GPU (Total Tex Cache) */
258 .cache_size = 1024,
259 .cache_level = 2,
260 .cache_line_size = 64,
261 .flags = (CRAT_CACHE_FLAGS_ENABLED |
262 CRAT_CACHE_FLAGS_DATA_CACHE |
263 CRAT_CACHE_FLAGS_SIMD_CACHE),
264 .num_cu_shared = 8,
265 },
266};
267
268static struct kfd_gpu_cache_info vega12_cache_info[] = {
269 {
270 /* TCP L1 Cache per CU */
271 .cache_size = 16,
272 .cache_level = 1,
273 .cache_line_size = 64,
274 .flags = (CRAT_CACHE_FLAGS_ENABLED |
275 CRAT_CACHE_FLAGS_DATA_CACHE |
276 CRAT_CACHE_FLAGS_SIMD_CACHE),
277 .num_cu_shared = 1,
278 },
279 {
280 /* Scalar L1 Instruction Cache per SQC */
281 .cache_size = 32,
282 .cache_level = 1,
283 .cache_line_size = 64,
284 .flags = (CRAT_CACHE_FLAGS_ENABLED |
285 CRAT_CACHE_FLAGS_INST_CACHE |
286 CRAT_CACHE_FLAGS_SIMD_CACHE),
287 .num_cu_shared = 3,
288 },
289 {
290 /* Scalar L1 Data Cache per SQC */
291 .cache_size = 16,
292 .cache_level = 1,
293 .cache_line_size = 64,
294 .flags = (CRAT_CACHE_FLAGS_ENABLED |
295 CRAT_CACHE_FLAGS_DATA_CACHE |
296 CRAT_CACHE_FLAGS_SIMD_CACHE),
297 .num_cu_shared = 3,
298 },
299 {
300 /* L2 Data Cache per GPU (Total Tex Cache) */
301 .cache_size = 2048,
302 .cache_level = 2,
303 .cache_line_size = 64,
304 .flags = (CRAT_CACHE_FLAGS_ENABLED |
305 CRAT_CACHE_FLAGS_DATA_CACHE |
306 CRAT_CACHE_FLAGS_SIMD_CACHE),
307 .num_cu_shared = 5,
308 },
309};
310
311static struct kfd_gpu_cache_info vega20_cache_info[] = {
312 {
313 /* TCP L1 Cache per CU */
314 .cache_size = 16,
315 .cache_level = 1,
316 .cache_line_size = 64,
317 .flags = (CRAT_CACHE_FLAGS_ENABLED |
318 CRAT_CACHE_FLAGS_DATA_CACHE |
319 CRAT_CACHE_FLAGS_SIMD_CACHE),
320 .num_cu_shared = 1,
321 },
322 {
323 /* Scalar L1 Instruction Cache per SQC */
324 .cache_size = 32,
325 .cache_level = 1,
326 .cache_line_size = 64,
327 .flags = (CRAT_CACHE_FLAGS_ENABLED |
328 CRAT_CACHE_FLAGS_INST_CACHE |
329 CRAT_CACHE_FLAGS_SIMD_CACHE),
330 .num_cu_shared = 3,
331 },
332 {
333 /* Scalar L1 Data Cache per SQC */
334 .cache_size = 16,
335 .cache_level = 1,
336 .cache_line_size = 64,
337 .flags = (CRAT_CACHE_FLAGS_ENABLED |
338 CRAT_CACHE_FLAGS_DATA_CACHE |
339 CRAT_CACHE_FLAGS_SIMD_CACHE),
340 .num_cu_shared = 3,
341 },
342 {
343 /* L2 Data Cache per GPU (Total Tex Cache) */
344 .cache_size = 8192,
345 .cache_level = 2,
346 .cache_line_size = 64,
347 .flags = (CRAT_CACHE_FLAGS_ENABLED |
348 CRAT_CACHE_FLAGS_DATA_CACHE |
349 CRAT_CACHE_FLAGS_SIMD_CACHE),
350 .num_cu_shared = 16,
351 },
352};
353
354static struct kfd_gpu_cache_info aldebaran_cache_info[] = {
355 {
356 /* TCP L1 Cache per CU */
357 .cache_size = 16,
358 .cache_level = 1,
359 .cache_line_size = 64,
360 .flags = (CRAT_CACHE_FLAGS_ENABLED |
361 CRAT_CACHE_FLAGS_DATA_CACHE |
362 CRAT_CACHE_FLAGS_SIMD_CACHE),
363 .num_cu_shared = 1,
364 },
365 {
366 /* Scalar L1 Instruction Cache per SQC */
367 .cache_size = 32,
368 .cache_level = 1,
369 .cache_line_size = 64,
370 .flags = (CRAT_CACHE_FLAGS_ENABLED |
371 CRAT_CACHE_FLAGS_INST_CACHE |
372 CRAT_CACHE_FLAGS_SIMD_CACHE),
373 .num_cu_shared = 2,
374 },
375 {
376 /* Scalar L1 Data Cache per SQC */
377 .cache_size = 16,
378 .cache_level = 1,
379 .cache_line_size = 64,
380 .flags = (CRAT_CACHE_FLAGS_ENABLED |
381 CRAT_CACHE_FLAGS_DATA_CACHE |
382 CRAT_CACHE_FLAGS_SIMD_CACHE),
383 .num_cu_shared = 2,
384 },
385 {
386 /* L2 Data Cache per GPU (Total Tex Cache) */
387 .cache_size = 8192,
388 .cache_level = 2,
389 .cache_line_size = 128,
390 .flags = (CRAT_CACHE_FLAGS_ENABLED |
391 CRAT_CACHE_FLAGS_DATA_CACHE |
392 CRAT_CACHE_FLAGS_SIMD_CACHE),
393 .num_cu_shared = 14,
394 },
395};
396
397static struct kfd_gpu_cache_info navi10_cache_info[] = {
398 {
399 /* TCP L1 Cache per CU */
400 .cache_size = 16,
401 .cache_level = 1,
402 .cache_line_size = 128,
403 .flags = (CRAT_CACHE_FLAGS_ENABLED |
404 CRAT_CACHE_FLAGS_DATA_CACHE |
405 CRAT_CACHE_FLAGS_SIMD_CACHE),
406 .num_cu_shared = 1,
407 },
408 {
409 /* Scalar L1 Instruction Cache per SQC */
410 .cache_size = 32,
411 .cache_level = 1,
412 .cache_line_size = 64,
413 .flags = (CRAT_CACHE_FLAGS_ENABLED |
414 CRAT_CACHE_FLAGS_INST_CACHE |
415 CRAT_CACHE_FLAGS_SIMD_CACHE),
416 .num_cu_shared = 2,
417 },
418 {
419 /* Scalar L1 Data Cache per SQC */
420 .cache_size = 16,
421 .cache_level = 1,
422 .cache_line_size = 64,
423 .flags = (CRAT_CACHE_FLAGS_ENABLED |
424 CRAT_CACHE_FLAGS_DATA_CACHE |
425 CRAT_CACHE_FLAGS_SIMD_CACHE),
426 .num_cu_shared = 2,
427 },
428 {
429 /* GL1 Data Cache per SA */
430 .cache_size = 128,
431 .cache_level = 1,
432 .cache_line_size = 128,
433 .flags = (CRAT_CACHE_FLAGS_ENABLED |
434 CRAT_CACHE_FLAGS_DATA_CACHE |
435 CRAT_CACHE_FLAGS_SIMD_CACHE),
436 .num_cu_shared = 10,
437 },
438 {
439 /* L2 Data Cache per GPU (Total Tex Cache) */
440 .cache_size = 4096,
441 .cache_level = 2,
442 .cache_line_size = 128,
443 .flags = (CRAT_CACHE_FLAGS_ENABLED |
444 CRAT_CACHE_FLAGS_DATA_CACHE |
445 CRAT_CACHE_FLAGS_SIMD_CACHE),
446 .num_cu_shared = 10,
447 },
448};
449
450static struct kfd_gpu_cache_info vangogh_cache_info[] = {
451 {
452 /* TCP L1 Cache per CU */
453 .cache_size = 16,
454 .cache_level = 1,
455 .cache_line_size = 128,
456 .flags = (CRAT_CACHE_FLAGS_ENABLED |
457 CRAT_CACHE_FLAGS_DATA_CACHE |
458 CRAT_CACHE_FLAGS_SIMD_CACHE),
459 .num_cu_shared = 1,
460 },
461 {
462 /* Scalar L1 Instruction Cache per SQC */
463 .cache_size = 32,
464 .cache_level = 1,
465 .cache_line_size = 64,
466 .flags = (CRAT_CACHE_FLAGS_ENABLED |
467 CRAT_CACHE_FLAGS_INST_CACHE |
468 CRAT_CACHE_FLAGS_SIMD_CACHE),
469 .num_cu_shared = 2,
470 },
471 {
472 /* Scalar L1 Data Cache per SQC */
473 .cache_size = 16,
474 .cache_level = 1,
475 .cache_line_size = 64,
476 .flags = (CRAT_CACHE_FLAGS_ENABLED |
477 CRAT_CACHE_FLAGS_DATA_CACHE |
478 CRAT_CACHE_FLAGS_SIMD_CACHE),
479 .num_cu_shared = 2,
480 },
481 {
482 /* GL1 Data Cache per SA */
483 .cache_size = 128,
484 .cache_level = 1,
485 .cache_line_size = 128,
486 .flags = (CRAT_CACHE_FLAGS_ENABLED |
487 CRAT_CACHE_FLAGS_DATA_CACHE |
488 CRAT_CACHE_FLAGS_SIMD_CACHE),
489 .num_cu_shared = 8,
490 },
491 {
492 /* L2 Data Cache per GPU (Total Tex Cache) */
493 .cache_size = 1024,
494 .cache_level = 2,
495 .cache_line_size = 128,
496 .flags = (CRAT_CACHE_FLAGS_ENABLED |
497 CRAT_CACHE_FLAGS_DATA_CACHE |
498 CRAT_CACHE_FLAGS_SIMD_CACHE),
499 .num_cu_shared = 8,
500 },
501};
502
503static struct kfd_gpu_cache_info navi14_cache_info[] = {
504 {
505 /* TCP L1 Cache per CU */
506 .cache_size = 16,
507 .cache_level = 1,
508 .cache_line_size = 128,
509 .flags = (CRAT_CACHE_FLAGS_ENABLED |
510 CRAT_CACHE_FLAGS_DATA_CACHE |
511 CRAT_CACHE_FLAGS_SIMD_CACHE),
512 .num_cu_shared = 1,
513 },
514 {
515 /* Scalar L1 Instruction Cache per SQC */
516 .cache_size = 32,
517 .cache_level = 1,
518 .cache_line_size = 64,
519 .flags = (CRAT_CACHE_FLAGS_ENABLED |
520 CRAT_CACHE_FLAGS_INST_CACHE |
521 CRAT_CACHE_FLAGS_SIMD_CACHE),
522 .num_cu_shared = 2,
523 },
524 {
525 /* Scalar L1 Data Cache per SQC */
526 .cache_size = 16,
527 .cache_level = 1,
528 .cache_line_size = 64,
529 .flags = (CRAT_CACHE_FLAGS_ENABLED |
530 CRAT_CACHE_FLAGS_DATA_CACHE |
531 CRAT_CACHE_FLAGS_SIMD_CACHE),
532 .num_cu_shared = 2,
533 },
534 {
535 /* GL1 Data Cache per SA */
536 .cache_size = 128,
537 .cache_level = 1,
538 .cache_line_size = 128,
539 .flags = (CRAT_CACHE_FLAGS_ENABLED |
540 CRAT_CACHE_FLAGS_DATA_CACHE |
541 CRAT_CACHE_FLAGS_SIMD_CACHE),
542 .num_cu_shared = 12,
543 },
544 {
545 /* L2 Data Cache per GPU (Total Tex Cache) */
546 .cache_size = 2048,
547 .cache_level = 2,
548 .cache_line_size = 128,
549 .flags = (CRAT_CACHE_FLAGS_ENABLED |
550 CRAT_CACHE_FLAGS_DATA_CACHE |
551 CRAT_CACHE_FLAGS_SIMD_CACHE),
552 .num_cu_shared = 12,
553 },
554};
555
556static struct kfd_gpu_cache_info sienna_cichlid_cache_info[] = {
557 {
558 /* TCP L1 Cache per CU */
559 .cache_size = 16,
560 .cache_level = 1,
561 .cache_line_size = 128,
562 .flags = (CRAT_CACHE_FLAGS_ENABLED |
563 CRAT_CACHE_FLAGS_DATA_CACHE |
564 CRAT_CACHE_FLAGS_SIMD_CACHE),
565 .num_cu_shared = 1,
566 },
567 {
568 /* Scalar L1 Instruction Cache per SQC */
569 .cache_size = 32,
570 .cache_level = 1,
571 .cache_line_size = 64,
572 .flags = (CRAT_CACHE_FLAGS_ENABLED |
573 CRAT_CACHE_FLAGS_INST_CACHE |
574 CRAT_CACHE_FLAGS_SIMD_CACHE),
575 .num_cu_shared = 2,
576 },
577 {
578 /* Scalar L1 Data Cache per SQC */
579 .cache_size = 16,
580 .cache_level = 1,
581 .cache_line_size = 64,
582 .flags = (CRAT_CACHE_FLAGS_ENABLED |
583 CRAT_CACHE_FLAGS_DATA_CACHE |
584 CRAT_CACHE_FLAGS_SIMD_CACHE),
585 .num_cu_shared = 2,
586 },
587 {
588 /* GL1 Data Cache per SA */
589 .cache_size = 128,
590 .cache_level = 1,
591 .cache_line_size = 128,
592 .flags = (CRAT_CACHE_FLAGS_ENABLED |
593 CRAT_CACHE_FLAGS_DATA_CACHE |
594 CRAT_CACHE_FLAGS_SIMD_CACHE),
595 .num_cu_shared = 10,
596 },
597 {
598 /* L2 Data Cache per GPU (Total Tex Cache) */
599 .cache_size = 4096,
600 .cache_level = 2,
601 .cache_line_size = 128,
602 .flags = (CRAT_CACHE_FLAGS_ENABLED |
603 CRAT_CACHE_FLAGS_DATA_CACHE |
604 CRAT_CACHE_FLAGS_SIMD_CACHE),
605 .num_cu_shared = 10,
606 },
607 {
608 /* L3 Data Cache per GPU */
609 .cache_size = 128*1024,
610 .cache_level = 3,
611 .cache_line_size = 64,
612 .flags = (CRAT_CACHE_FLAGS_ENABLED |
613 CRAT_CACHE_FLAGS_DATA_CACHE |
614 CRAT_CACHE_FLAGS_SIMD_CACHE),
615 .num_cu_shared = 10,
616 },
617};
618
619static struct kfd_gpu_cache_info navy_flounder_cache_info[] = {
620 {
621 /* TCP L1 Cache per CU */
622 .cache_size = 16,
623 .cache_level = 1,
624 .cache_line_size = 128,
625 .flags = (CRAT_CACHE_FLAGS_ENABLED |
626 CRAT_CACHE_FLAGS_DATA_CACHE |
627 CRAT_CACHE_FLAGS_SIMD_CACHE),
628 .num_cu_shared = 1,
629 },
630 {
631 /* Scalar L1 Instruction Cache per SQC */
632 .cache_size = 32,
633 .cache_level = 1,
634 .cache_line_size = 64,
635 .flags = (CRAT_CACHE_FLAGS_ENABLED |
636 CRAT_CACHE_FLAGS_INST_CACHE |
637 CRAT_CACHE_FLAGS_SIMD_CACHE),
638 .num_cu_shared = 2,
639 },
640 {
641 /* Scalar L1 Data Cache per SQC */
642 .cache_size = 16,
643 .cache_level = 1,
644 .cache_line_size = 64,
645 .flags = (CRAT_CACHE_FLAGS_ENABLED |
646 CRAT_CACHE_FLAGS_DATA_CACHE |
647 CRAT_CACHE_FLAGS_SIMD_CACHE),
648 .num_cu_shared = 2,
649 },
650 {
651 /* GL1 Data Cache per SA */
652 .cache_size = 128,
653 .cache_level = 1,
654 .cache_line_size = 128,
655 .flags = (CRAT_CACHE_FLAGS_ENABLED |
656 CRAT_CACHE_FLAGS_DATA_CACHE |
657 CRAT_CACHE_FLAGS_SIMD_CACHE),
658 .num_cu_shared = 10,
659 },
660 {
661 /* L2 Data Cache per GPU (Total Tex Cache) */
662 .cache_size = 3072,
663 .cache_level = 2,
664 .cache_line_size = 128,
665 .flags = (CRAT_CACHE_FLAGS_ENABLED |
666 CRAT_CACHE_FLAGS_DATA_CACHE |
667 CRAT_CACHE_FLAGS_SIMD_CACHE),
668 .num_cu_shared = 10,
669 },
670 {
671 /* L3 Data Cache per GPU */
672 .cache_size = 96*1024,
673 .cache_level = 3,
674 .cache_line_size = 64,
675 .flags = (CRAT_CACHE_FLAGS_ENABLED |
676 CRAT_CACHE_FLAGS_DATA_CACHE |
677 CRAT_CACHE_FLAGS_SIMD_CACHE),
678 .num_cu_shared = 10,
679 },
680};
681
682static struct kfd_gpu_cache_info dimgrey_cavefish_cache_info[] = {
683 {
684 /* TCP L1 Cache per CU */
685 .cache_size = 16,
686 .cache_level = 1,
687 .cache_line_size = 128,
688 .flags = (CRAT_CACHE_FLAGS_ENABLED |
689 CRAT_CACHE_FLAGS_DATA_CACHE |
690 CRAT_CACHE_FLAGS_SIMD_CACHE),
691 .num_cu_shared = 1,
692 },
693 {
694 /* Scalar L1 Instruction Cache per SQC */
695 .cache_size = 32,
696 .cache_level = 1,
697 .cache_line_size = 64,
698 .flags = (CRAT_CACHE_FLAGS_ENABLED |
699 CRAT_CACHE_FLAGS_INST_CACHE |
700 CRAT_CACHE_FLAGS_SIMD_CACHE),
701 .num_cu_shared = 2,
702 },
703 {
704 /* Scalar L1 Data Cache per SQC */
705 .cache_size = 16,
706 .cache_level = 1,
707 .cache_line_size = 64,
708 .flags = (CRAT_CACHE_FLAGS_ENABLED |
709 CRAT_CACHE_FLAGS_DATA_CACHE |
710 CRAT_CACHE_FLAGS_SIMD_CACHE),
711 .num_cu_shared = 2,
712 },
713 {
714 /* GL1 Data Cache per SA */
715 .cache_size = 128,
716 .cache_level = 1,
717 .cache_line_size = 128,
718 .flags = (CRAT_CACHE_FLAGS_ENABLED |
719 CRAT_CACHE_FLAGS_DATA_CACHE |
720 CRAT_CACHE_FLAGS_SIMD_CACHE),
721 .num_cu_shared = 8,
722 },
723 {
724 /* L2 Data Cache per GPU (Total Tex Cache) */
725 .cache_size = 2048,
726 .cache_level = 2,
727 .cache_line_size = 128,
728 .flags = (CRAT_CACHE_FLAGS_ENABLED |
729 CRAT_CACHE_FLAGS_DATA_CACHE |
730 CRAT_CACHE_FLAGS_SIMD_CACHE),
731 .num_cu_shared = 8,
732 },
733 {
734 /* L3 Data Cache per GPU */
735 .cache_size = 32*1024,
736 .cache_level = 3,
737 .cache_line_size = 64,
738 .flags = (CRAT_CACHE_FLAGS_ENABLED |
739 CRAT_CACHE_FLAGS_DATA_CACHE |
740 CRAT_CACHE_FLAGS_SIMD_CACHE),
741 .num_cu_shared = 8,
742 },
743};
744
745static struct kfd_gpu_cache_info beige_goby_cache_info[] = {
746 {
747 /* TCP L1 Cache per CU */
748 .cache_size = 16,
749 .cache_level = 1,
750 .cache_line_size = 128,
751 .flags = (CRAT_CACHE_FLAGS_ENABLED |
752 CRAT_CACHE_FLAGS_DATA_CACHE |
753 CRAT_CACHE_FLAGS_SIMD_CACHE),
754 .num_cu_shared = 1,
755 },
756 {
757 /* Scalar L1 Instruction Cache per SQC */
758 .cache_size = 32,
759 .cache_level = 1,
760 .cache_line_size = 64,
761 .flags = (CRAT_CACHE_FLAGS_ENABLED |
762 CRAT_CACHE_FLAGS_INST_CACHE |
763 CRAT_CACHE_FLAGS_SIMD_CACHE),
764 .num_cu_shared = 2,
765 },
766 {
767 /* Scalar L1 Data Cache per SQC */
768 .cache_size = 16,
769 .cache_level = 1,
770 .cache_line_size = 64,
771 .flags = (CRAT_CACHE_FLAGS_ENABLED |
772 CRAT_CACHE_FLAGS_DATA_CACHE |
773 CRAT_CACHE_FLAGS_SIMD_CACHE),
774 .num_cu_shared = 2,
775 },
776 {
777 /* GL1 Data Cache per SA */
778 .cache_size = 128,
779 .cache_level = 1,
780 .cache_line_size = 128,
781 .flags = (CRAT_CACHE_FLAGS_ENABLED |
782 CRAT_CACHE_FLAGS_DATA_CACHE |
783 CRAT_CACHE_FLAGS_SIMD_CACHE),
784 .num_cu_shared = 8,
785 },
786 {
787 /* L2 Data Cache per GPU (Total Tex Cache) */
788 .cache_size = 1024,
789 .cache_level = 2,
790 .cache_line_size = 128,
791 .flags = (CRAT_CACHE_FLAGS_ENABLED |
792 CRAT_CACHE_FLAGS_DATA_CACHE |
793 CRAT_CACHE_FLAGS_SIMD_CACHE),
794 .num_cu_shared = 8,
795 },
796 {
797 /* L3 Data Cache per GPU */
798 .cache_size = 16*1024,
799 .cache_level = 3,
800 .cache_line_size = 64,
801 .flags = (CRAT_CACHE_FLAGS_ENABLED |
802 CRAT_CACHE_FLAGS_DATA_CACHE |
803 CRAT_CACHE_FLAGS_SIMD_CACHE),
804 .num_cu_shared = 8,
805 },
806};
807
808static struct kfd_gpu_cache_info yellow_carp_cache_info[] = {
809 {
810 /* TCP L1 Cache per CU */
811 .cache_size = 16,
812 .cache_level = 1,
813 .cache_line_size = 128,
814 .flags = (CRAT_CACHE_FLAGS_ENABLED |
815 CRAT_CACHE_FLAGS_DATA_CACHE |
816 CRAT_CACHE_FLAGS_SIMD_CACHE),
817 .num_cu_shared = 1,
818 },
819 {
820 /* Scalar L1 Instruction Cache per SQC */
821 .cache_size = 32,
822 .cache_level = 1,
823 .cache_line_size = 64,
824 .flags = (CRAT_CACHE_FLAGS_ENABLED |
825 CRAT_CACHE_FLAGS_INST_CACHE |
826 CRAT_CACHE_FLAGS_SIMD_CACHE),
827 .num_cu_shared = 2,
828 },
829 {
830 /* Scalar L1 Data Cache per SQC */
831 .cache_size = 16,
832 .cache_level = 1,
833 .cache_line_size = 64,
834 .flags = (CRAT_CACHE_FLAGS_ENABLED |
835 CRAT_CACHE_FLAGS_DATA_CACHE |
836 CRAT_CACHE_FLAGS_SIMD_CACHE),
837 .num_cu_shared = 2,
838 },
839 {
840 /* GL1 Data Cache per SA */
841 .cache_size = 128,
842 .cache_level = 1,
843 .cache_line_size = 128,
844 .flags = (CRAT_CACHE_FLAGS_ENABLED |
845 CRAT_CACHE_FLAGS_DATA_CACHE |
846 CRAT_CACHE_FLAGS_SIMD_CACHE),
847 .num_cu_shared = 6,
848 },
849 {
850 /* L2 Data Cache per GPU (Total Tex Cache) */
851 .cache_size = 2048,
852 .cache_level = 2,
853 .cache_line_size = 128,
854 .flags = (CRAT_CACHE_FLAGS_ENABLED |
855 CRAT_CACHE_FLAGS_DATA_CACHE |
856 CRAT_CACHE_FLAGS_SIMD_CACHE),
857 .num_cu_shared = 6,
858 },
859};
860
861static struct kfd_gpu_cache_info gfx1037_cache_info[] = {
862 {
863 /* TCP L1 Cache per CU */
864 .cache_size = 16,
865 .cache_level = 1,
866 .cache_line_size = 128,
867 .flags = (CRAT_CACHE_FLAGS_ENABLED |
868 CRAT_CACHE_FLAGS_DATA_CACHE |
869 CRAT_CACHE_FLAGS_SIMD_CACHE),
870 .num_cu_shared = 1,
871 },
872 {
873 /* Scalar L1 Instruction Cache per SQC */
874 .cache_size = 32,
875 .cache_level = 1,
876 .cache_line_size = 64,
877 .flags = (CRAT_CACHE_FLAGS_ENABLED |
878 CRAT_CACHE_FLAGS_INST_CACHE |
879 CRAT_CACHE_FLAGS_SIMD_CACHE),
880 .num_cu_shared = 2,
881 },
882 {
883 /* Scalar L1 Data Cache per SQC */
884 .cache_size = 16,
885 .cache_level = 1,
886 .cache_line_size = 64,
887 .flags = (CRAT_CACHE_FLAGS_ENABLED |
888 CRAT_CACHE_FLAGS_DATA_CACHE |
889 CRAT_CACHE_FLAGS_SIMD_CACHE),
890 .num_cu_shared = 2,
891 },
892 {
893 /* GL1 Data Cache per SA */
894 .cache_size = 128,
895 .cache_level = 1,
896 .cache_line_size = 128,
897 .flags = (CRAT_CACHE_FLAGS_ENABLED |
898 CRAT_CACHE_FLAGS_DATA_CACHE |
899 CRAT_CACHE_FLAGS_SIMD_CACHE),
900 .num_cu_shared = 2,
901 },
902 {
903 /* L2 Data Cache per GPU (Total Tex Cache) */
904 .cache_size = 256,
905 .cache_level = 2,
906 .cache_line_size = 128,
907 .flags = (CRAT_CACHE_FLAGS_ENABLED |
908 CRAT_CACHE_FLAGS_DATA_CACHE |
909 CRAT_CACHE_FLAGS_SIMD_CACHE),
910 .num_cu_shared = 2,
911 },
912};
913
914static struct kfd_gpu_cache_info gc_10_3_6_cache_info[] = {
915 {
916 /* TCP L1 Cache per CU */
917 .cache_size = 16,
918 .cache_level = 1,
919 .cache_line_size = 128,
920 .flags = (CRAT_CACHE_FLAGS_ENABLED |
921 CRAT_CACHE_FLAGS_DATA_CACHE |
922 CRAT_CACHE_FLAGS_SIMD_CACHE),
923 .num_cu_shared = 1,
924 },
925 {
926 /* Scalar L1 Instruction Cache per SQC */
927 .cache_size = 32,
928 .cache_level = 1,
929 .cache_line_size = 64,
930 .flags = (CRAT_CACHE_FLAGS_ENABLED |
931 CRAT_CACHE_FLAGS_INST_CACHE |
932 CRAT_CACHE_FLAGS_SIMD_CACHE),
933 .num_cu_shared = 2,
934 },
935 {
936 /* Scalar L1 Data Cache per SQC */
937 .cache_size = 16,
938 .cache_level = 1,
939 .cache_line_size = 64,
940 .flags = (CRAT_CACHE_FLAGS_ENABLED |
941 CRAT_CACHE_FLAGS_DATA_CACHE |
942 CRAT_CACHE_FLAGS_SIMD_CACHE),
943 .num_cu_shared = 2,
944 },
945 {
946 /* GL1 Data Cache per SA */
947 .cache_size = 128,
948 .cache_level = 1,
949 .cache_line_size = 128,
950 .flags = (CRAT_CACHE_FLAGS_ENABLED |
951 CRAT_CACHE_FLAGS_DATA_CACHE |
952 CRAT_CACHE_FLAGS_SIMD_CACHE),
953 .num_cu_shared = 2,
954 },
955 {
956 /* L2 Data Cache per GPU (Total Tex Cache) */
957 .cache_size = 256,
958 .cache_level = 2,
959 .cache_line_size = 128,
960 .flags = (CRAT_CACHE_FLAGS_ENABLED |
961 CRAT_CACHE_FLAGS_DATA_CACHE |
962 CRAT_CACHE_FLAGS_SIMD_CACHE),
963 .num_cu_shared = 2,
964 },
965};
966
967static struct kfd_gpu_cache_info dummy_cache_info[] = {
968 {
969 /* TCP L1 Cache per CU */
970 .cache_size = 16,
971 .cache_level = 1,
972 .cache_line_size = 64,
973 .flags = (CRAT_CACHE_FLAGS_ENABLED |
974 CRAT_CACHE_FLAGS_DATA_CACHE |
975 CRAT_CACHE_FLAGS_SIMD_CACHE),
976 .num_cu_shared = 1,
977 },
978 {
979 /* Scalar L1 Instruction Cache per SQC */
980 .cache_size = 32,
981 .cache_level = 1,
982 .cache_line_size = 64,
983 .flags = (CRAT_CACHE_FLAGS_ENABLED |
984 CRAT_CACHE_FLAGS_INST_CACHE |
985 CRAT_CACHE_FLAGS_SIMD_CACHE),
986 .num_cu_shared = 2,
987 },
988 {
989 /* Scalar L1 Data Cache per SQC */
990 .cache_size = 16,
991 .cache_level = 1,
992 .cache_line_size = 64,
993 .flags = (CRAT_CACHE_FLAGS_ENABLED |
994 CRAT_CACHE_FLAGS_DATA_CACHE |
995 CRAT_CACHE_FLAGS_SIMD_CACHE),
996 .num_cu_shared = 2,
997 },
998 {
999 /* GL1 Data Cache per SA */
1000 .cache_size = 128,
1001 .cache_level = 1,
1002 .cache_line_size = 64,
1003 .flags = (CRAT_CACHE_FLAGS_ENABLED |
1004 CRAT_CACHE_FLAGS_DATA_CACHE |
1005 CRAT_CACHE_FLAGS_SIMD_CACHE),
1006 .num_cu_shared = 6,
1007 },
1008 {
1009 /* L2 Data Cache per GPU (Total Tex Cache) */
1010 .cache_size = 2048,
1011 .cache_level = 2,
1012 .cache_line_size = 64,
1013 .flags = (CRAT_CACHE_FLAGS_ENABLED |
1014 CRAT_CACHE_FLAGS_DATA_CACHE |
1015 CRAT_CACHE_FLAGS_SIMD_CACHE),
1016 .num_cu_shared = 6,
1017 },
1018};
1019
1020static void kfd_populated_cu_info_cpu(struct kfd_topology_device *dev,
1021 struct crat_subtype_computeunit *cu)
1022{
1023 dev->node_props.cpu_cores_count = cu->num_cpu_cores;
1024 dev->node_props.cpu_core_id_base = cu->processor_id_low;
1025 if (cu->hsa_capability & CRAT_CU_FLAGS_IOMMU_PRESENT)
1026 dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
1027
1028 pr_debug("CU CPU: cores=%d id_base=%d\n", cu->num_cpu_cores,
1029 cu->processor_id_low);
1030}
1031
1032static void kfd_populated_cu_info_gpu(struct kfd_topology_device *dev,
1033 struct crat_subtype_computeunit *cu)
1034{
1035 dev->node_props.simd_id_base = cu->processor_id_low;
1036 dev->node_props.simd_count = cu->num_simd_cores;
1037 dev->node_props.lds_size_in_kb = cu->lds_size_in_kb;
1038 dev->node_props.max_waves_per_simd = cu->max_waves_simd;
1039 dev->node_props.wave_front_size = cu->wave_front_size;
1040 dev->node_props.array_count = cu->array_count;
1041 dev->node_props.cu_per_simd_array = cu->num_cu_per_array;
1042 dev->node_props.simd_per_cu = cu->num_simd_per_cu;
1043 dev->node_props.max_slots_scratch_cu = cu->max_slots_scatch_cu;
1044 if (cu->hsa_capability & CRAT_CU_FLAGS_HOT_PLUGGABLE)
1045 dev->node_props.capability |= HSA_CAP_HOT_PLUGGABLE;
1046 pr_debug("CU GPU: id_base=%d\n", cu->processor_id_low);
1047}
1048
1049/* kfd_parse_subtype_cu - parse compute unit subtypes and attach it to correct
1050 * topology device present in the device_list
1051 */
1052static int kfd_parse_subtype_cu(struct crat_subtype_computeunit *cu,
1053 struct list_head *device_list)
1054{
1055 struct kfd_topology_device *dev;
1056
1057 pr_debug("Found CU entry in CRAT table with proximity_domain=%d caps=%x\n",
1058 cu->proximity_domain, cu->hsa_capability);
1059 list_for_each_entry(dev, device_list, list) {
1060 if (cu->proximity_domain == dev->proximity_domain) {
1061 if (cu->flags & CRAT_CU_FLAGS_CPU_PRESENT)
1062 kfd_populated_cu_info_cpu(dev, cu);
1063
1064 if (cu->flags & CRAT_CU_FLAGS_GPU_PRESENT)
1065 kfd_populated_cu_info_gpu(dev, cu);
1066 break;
1067 }
1068 }
1069
1070 return 0;
1071}
1072
1073static struct kfd_mem_properties *
1074find_subtype_mem(uint32_t heap_type, uint32_t flags, uint32_t width,
1075 struct kfd_topology_device *dev)
1076{
1077 struct kfd_mem_properties *props;
1078
1079 list_for_each_entry(props, &dev->mem_props, list) {
1080 if (props->heap_type == heap_type
1081 && props->flags == flags
1082 && props->width == width)
1083 return props;
1084 }
1085
1086 return NULL;
1087}
1088/* kfd_parse_subtype_mem - parse memory subtypes and attach it to correct
1089 * topology device present in the device_list
1090 */
1091static int kfd_parse_subtype_mem(struct crat_subtype_memory *mem,
1092 struct list_head *device_list)
1093{
1094 struct kfd_mem_properties *props;
1095 struct kfd_topology_device *dev;
1096 uint32_t heap_type;
1097 uint64_t size_in_bytes;
1098 uint32_t flags = 0;
1099 uint32_t width;
1100
1101 pr_debug("Found memory entry in CRAT table with proximity_domain=%d\n",
1102 mem->proximity_domain);
1103 list_for_each_entry(dev, device_list, list) {
1104 if (mem->proximity_domain == dev->proximity_domain) {
1105 /* We're on GPU node */
1106 if (dev->node_props.cpu_cores_count == 0) {
1107 /* APU */
1108 if (mem->visibility_type == 0)
1109 heap_type =
1110 HSA_MEM_HEAP_TYPE_FB_PRIVATE;
1111 /* dGPU */
1112 else
1113 heap_type = mem->visibility_type;
1114 } else
1115 heap_type = HSA_MEM_HEAP_TYPE_SYSTEM;
1116
1117 if (mem->flags & CRAT_MEM_FLAGS_HOT_PLUGGABLE)
1118 flags |= HSA_MEM_FLAGS_HOT_PLUGGABLE;
1119 if (mem->flags & CRAT_MEM_FLAGS_NON_VOLATILE)
1120 flags |= HSA_MEM_FLAGS_NON_VOLATILE;
1121
1122 size_in_bytes =
1123 ((uint64_t)mem->length_high << 32) +
1124 mem->length_low;
1125 width = mem->width;
1126
1127 /* Multiple banks of the same type are aggregated into
1128 * one. User mode doesn't care about multiple physical
1129 * memory segments. It's managed as a single virtual
1130 * heap for user mode.
1131 */
1132 props = find_subtype_mem(heap_type, flags, width, dev);
1133 if (props) {
1134 props->size_in_bytes += size_in_bytes;
1135 break;
1136 }
1137
1138 props = kfd_alloc_struct(props);
1139 if (!props)
1140 return -ENOMEM;
1141
1142 props->heap_type = heap_type;
1143 props->flags = flags;
1144 props->size_in_bytes = size_in_bytes;
1145 props->width = width;
1146
1147 dev->node_props.mem_banks_count++;
1148 list_add_tail(&props->list, &dev->mem_props);
1149
1150 break;
1151 }
1152 }
1153
1154 return 0;
1155}
1156
1157/* kfd_parse_subtype_cache - parse cache subtypes and attach it to correct
1158 * topology device present in the device_list
1159 */
1160static int kfd_parse_subtype_cache(struct crat_subtype_cache *cache,
1161 struct list_head *device_list)
1162{
1163 struct kfd_cache_properties *props;
1164 struct kfd_topology_device *dev;
1165 uint32_t id;
1166 uint32_t total_num_of_cu;
1167
1168 id = cache->processor_id_low;
1169
1170 pr_debug("Found cache entry in CRAT table with processor_id=%d\n", id);
1171 list_for_each_entry(dev, device_list, list) {
1172 total_num_of_cu = (dev->node_props.array_count *
1173 dev->node_props.cu_per_simd_array);
1174
1175 /* Cache infomration in CRAT doesn't have proximity_domain
1176 * information as it is associated with a CPU core or GPU
1177 * Compute Unit. So map the cache using CPU core Id or SIMD
1178 * (GPU) ID.
1179 * TODO: This works because currently we can safely assume that
1180 * Compute Units are parsed before caches are parsed. In
1181 * future, remove this dependency
1182 */
1183 if ((id >= dev->node_props.cpu_core_id_base &&
1184 id <= dev->node_props.cpu_core_id_base +
1185 dev->node_props.cpu_cores_count) ||
1186 (id >= dev->node_props.simd_id_base &&
1187 id < dev->node_props.simd_id_base +
1188 total_num_of_cu)) {
1189 props = kfd_alloc_struct(props);
1190 if (!props)
1191 return -ENOMEM;
1192
1193 props->processor_id_low = id;
1194 props->cache_level = cache->cache_level;
1195 props->cache_size = cache->cache_size;
1196 props->cacheline_size = cache->cache_line_size;
1197 props->cachelines_per_tag = cache->lines_per_tag;
1198 props->cache_assoc = cache->associativity;
1199 props->cache_latency = cache->cache_latency;
1200
1201 memcpy(props->sibling_map, cache->sibling_map,
1202 CRAT_SIBLINGMAP_SIZE);
1203
1204 /* set the sibling_map_size as 32 for CRAT from ACPI */
1205 props->sibling_map_size = CRAT_SIBLINGMAP_SIZE;
1206
1207 if (cache->flags & CRAT_CACHE_FLAGS_DATA_CACHE)
1208 props->cache_type |= HSA_CACHE_TYPE_DATA;
1209 if (cache->flags & CRAT_CACHE_FLAGS_INST_CACHE)
1210 props->cache_type |= HSA_CACHE_TYPE_INSTRUCTION;
1211 if (cache->flags & CRAT_CACHE_FLAGS_CPU_CACHE)
1212 props->cache_type |= HSA_CACHE_TYPE_CPU;
1213 if (cache->flags & CRAT_CACHE_FLAGS_SIMD_CACHE)
1214 props->cache_type |= HSA_CACHE_TYPE_HSACU;
1215
1216 dev->node_props.caches_count++;
1217 list_add_tail(&props->list, &dev->cache_props);
1218
1219 break;
1220 }
1221 }
1222
1223 return 0;
1224}
1225
1226/* kfd_parse_subtype_iolink - parse iolink subtypes and attach it to correct
1227 * topology device present in the device_list
1228 */
1229static int kfd_parse_subtype_iolink(struct crat_subtype_iolink *iolink,
1230 struct list_head *device_list)
1231{
1232 struct kfd_iolink_properties *props = NULL, *props2;
1233 struct kfd_topology_device *dev, *to_dev;
1234 uint32_t id_from;
1235 uint32_t id_to;
1236
1237 id_from = iolink->proximity_domain_from;
1238 id_to = iolink->proximity_domain_to;
1239
1240 pr_debug("Found IO link entry in CRAT table with id_from=%d, id_to %d\n",
1241 id_from, id_to);
1242 list_for_each_entry(dev, device_list, list) {
1243 if (id_from == dev->proximity_domain) {
1244 props = kfd_alloc_struct(props);
1245 if (!props)
1246 return -ENOMEM;
1247
1248 props->node_from = id_from;
1249 props->node_to = id_to;
1250 props->ver_maj = iolink->version_major;
1251 props->ver_min = iolink->version_minor;
1252 props->iolink_type = iolink->io_interface_type;
1253
1254 if (props->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS)
1255 props->weight = 20;
1256 else if (props->iolink_type == CRAT_IOLINK_TYPE_XGMI)
1257 props->weight = iolink->weight_xgmi;
1258 else
1259 props->weight = node_distance(id_from, id_to);
1260
1261 props->min_latency = iolink->minimum_latency;
1262 props->max_latency = iolink->maximum_latency;
1263 props->min_bandwidth = iolink->minimum_bandwidth_mbs;
1264 props->max_bandwidth = iolink->maximum_bandwidth_mbs;
1265 props->rec_transfer_size =
1266 iolink->recommended_transfer_size;
1267
1268 dev->node_props.io_links_count++;
1269 list_add_tail(&props->list, &dev->io_link_props);
1270 break;
1271 }
1272 }
1273
1274 /* CPU topology is created before GPUs are detected, so CPU->GPU
1275 * links are not built at that time. If a PCIe type is discovered, it
1276 * means a GPU is detected and we are adding GPU->CPU to the topology.
1277 * At this time, also add the corresponded CPU->GPU link if GPU
1278 * is large bar.
1279 * For xGMI, we only added the link with one direction in the crat
1280 * table, add corresponded reversed direction link now.
1281 */
1282 if (props && (iolink->flags & CRAT_IOLINK_FLAGS_BI_DIRECTIONAL)) {
1283 to_dev = kfd_topology_device_by_proximity_domain_no_lock(id_to);
1284 if (!to_dev)
1285 return -ENODEV;
1286 /* same everything but the other direction */
1287 props2 = kmemdup(props, sizeof(*props2), GFP_KERNEL);
1288 if (!props2)
1289 return -ENOMEM;
1290
1291 props2->node_from = id_to;
1292 props2->node_to = id_from;
1293 props2->kobj = NULL;
1294 to_dev->node_props.io_links_count++;
1295 list_add_tail(&props2->list, &to_dev->io_link_props);
1296 }
1297
1298 return 0;
1299}
1300
1301/* kfd_parse_subtype - parse subtypes and attach it to correct topology device
1302 * present in the device_list
1303 * @sub_type_hdr - subtype section of crat_image
1304 * @device_list - list of topology devices present in this crat_image
1305 */
1306static int kfd_parse_subtype(struct crat_subtype_generic *sub_type_hdr,
1307 struct list_head *device_list)
1308{
1309 struct crat_subtype_computeunit *cu;
1310 struct crat_subtype_memory *mem;
1311 struct crat_subtype_cache *cache;
1312 struct crat_subtype_iolink *iolink;
1313 int ret = 0;
1314
1315 switch (sub_type_hdr->type) {
1316 case CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY:
1317 cu = (struct crat_subtype_computeunit *)sub_type_hdr;
1318 ret = kfd_parse_subtype_cu(cu, device_list);
1319 break;
1320 case CRAT_SUBTYPE_MEMORY_AFFINITY:
1321 mem = (struct crat_subtype_memory *)sub_type_hdr;
1322 ret = kfd_parse_subtype_mem(mem, device_list);
1323 break;
1324 case CRAT_SUBTYPE_CACHE_AFFINITY:
1325 cache = (struct crat_subtype_cache *)sub_type_hdr;
1326 ret = kfd_parse_subtype_cache(cache, device_list);
1327 break;
1328 case CRAT_SUBTYPE_TLB_AFFINITY:
1329 /*
1330 * For now, nothing to do here
1331 */
1332 pr_debug("Found TLB entry in CRAT table (not processing)\n");
1333 break;
1334 case CRAT_SUBTYPE_CCOMPUTE_AFFINITY:
1335 /*
1336 * For now, nothing to do here
1337 */
1338 pr_debug("Found CCOMPUTE entry in CRAT table (not processing)\n");
1339 break;
1340 case CRAT_SUBTYPE_IOLINK_AFFINITY:
1341 iolink = (struct crat_subtype_iolink *)sub_type_hdr;
1342 ret = kfd_parse_subtype_iolink(iolink, device_list);
1343 break;
1344 default:
1345 pr_warn("Unknown subtype %d in CRAT\n",
1346 sub_type_hdr->type);
1347 }
1348
1349 return ret;
1350}
1351
1352/* kfd_parse_crat_table - parse CRAT table. For each node present in CRAT
1353 * create a kfd_topology_device and add in to device_list. Also parse
1354 * CRAT subtypes and attach it to appropriate kfd_topology_device
1355 * @crat_image - input image containing CRAT
1356 * @device_list - [OUT] list of kfd_topology_device generated after
1357 * parsing crat_image
1358 * @proximity_domain - Proximity domain of the first device in the table
1359 *
1360 * Return - 0 if successful else -ve value
1361 */
1362int kfd_parse_crat_table(void *crat_image, struct list_head *device_list,
1363 uint32_t proximity_domain)
1364{
1365 struct kfd_topology_device *top_dev = NULL;
1366 struct crat_subtype_generic *sub_type_hdr;
1367 uint16_t node_id;
1368 int ret = 0;
1369 struct crat_header *crat_table = (struct crat_header *)crat_image;
1370 uint16_t num_nodes;
1371 uint32_t image_len;
1372
1373 if (!crat_image)
1374 return -EINVAL;
1375
1376 if (!list_empty(device_list)) {
1377 pr_warn("Error device list should be empty\n");
1378 return -EINVAL;
1379 }
1380
1381 num_nodes = crat_table->num_domains;
1382 image_len = crat_table->length;
1383
1384 pr_debug("Parsing CRAT table with %d nodes\n", num_nodes);
1385
1386 for (node_id = 0; node_id < num_nodes; node_id++) {
1387 top_dev = kfd_create_topology_device(device_list);
1388 if (!top_dev)
1389 break;
1390 top_dev->proximity_domain = proximity_domain++;
1391 }
1392
1393 if (!top_dev) {
1394 ret = -ENOMEM;
1395 goto err;
1396 }
1397
1398 memcpy(top_dev->oem_id, crat_table->oem_id, CRAT_OEMID_LENGTH);
1399 memcpy(top_dev->oem_table_id, crat_table->oem_table_id,
1400 CRAT_OEMTABLEID_LENGTH);
1401 top_dev->oem_revision = crat_table->oem_revision;
1402
1403 sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1);
1404 while ((char *)sub_type_hdr + sizeof(struct crat_subtype_generic) <
1405 ((char *)crat_image) + image_len) {
1406 if (sub_type_hdr->flags & CRAT_SUBTYPE_FLAGS_ENABLED) {
1407 ret = kfd_parse_subtype(sub_type_hdr, device_list);
1408 if (ret)
1409 break;
1410 }
1411
1412 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1413 sub_type_hdr->length);
1414 }
1415
1416err:
1417 if (ret)
1418 kfd_release_topology_device_list(device_list);
1419
1420 return ret;
1421}
1422
1423
1424static int kfd_fill_gpu_cache_info_from_gfx_config(struct kfd_dev *kdev,
1425 struct kfd_gpu_cache_info *pcache_info)
1426{
1427 struct amdgpu_device *adev = kdev->adev;
1428 int i = 0;
1429
1430 /* TCP L1 Cache per CU */
1431 if (adev->gfx.config.gc_tcp_l1_size) {
1432 pcache_info[i].cache_size = adev->gfx.config.gc_tcp_l1_size;
1433 pcache_info[i].cache_level = 1;
1434 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1435 CRAT_CACHE_FLAGS_DATA_CACHE |
1436 CRAT_CACHE_FLAGS_SIMD_CACHE);
1437 pcache_info[0].num_cu_shared = adev->gfx.config.gc_num_tcp_per_wpg / 2;
1438 i++;
1439 }
1440 /* Scalar L1 Instruction Cache per SQC */
1441 if (adev->gfx.config.gc_l1_instruction_cache_size_per_sqc) {
1442 pcache_info[i].cache_size =
1443 adev->gfx.config.gc_l1_instruction_cache_size_per_sqc;
1444 pcache_info[i].cache_level = 1;
1445 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1446 CRAT_CACHE_FLAGS_INST_CACHE |
1447 CRAT_CACHE_FLAGS_SIMD_CACHE);
1448 pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_sqc_per_wgp * 2;
1449 i++;
1450 }
1451 /* Scalar L1 Data Cache per SQC */
1452 if (adev->gfx.config.gc_l1_data_cache_size_per_sqc) {
1453 pcache_info[i].cache_size = adev->gfx.config.gc_l1_data_cache_size_per_sqc;
1454 pcache_info[i].cache_level = 1;
1455 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1456 CRAT_CACHE_FLAGS_DATA_CACHE |
1457 CRAT_CACHE_FLAGS_SIMD_CACHE);
1458 pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_sqc_per_wgp * 2;
1459 i++;
1460 }
1461 /* GL1 Data Cache per SA */
1462 if (adev->gfx.config.gc_gl1c_per_sa &&
1463 adev->gfx.config.gc_gl1c_size_per_instance) {
1464 pcache_info[i].cache_size = adev->gfx.config.gc_gl1c_per_sa *
1465 adev->gfx.config.gc_gl1c_size_per_instance;
1466 pcache_info[i].cache_level = 1;
1467 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1468 CRAT_CACHE_FLAGS_DATA_CACHE |
1469 CRAT_CACHE_FLAGS_SIMD_CACHE);
1470 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
1471 i++;
1472 }
1473 /* L2 Data Cache per GPU (Total Tex Cache) */
1474 if (adev->gfx.config.gc_gl2c_per_gpu) {
1475 pcache_info[i].cache_size = adev->gfx.config.gc_gl2c_per_gpu;
1476 pcache_info[i].cache_level = 2;
1477 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1478 CRAT_CACHE_FLAGS_DATA_CACHE |
1479 CRAT_CACHE_FLAGS_SIMD_CACHE);
1480 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
1481 i++;
1482 }
1483 /* L3 Data Cache per GPU */
1484 if (adev->gmc.mall_size) {
1485 pcache_info[i].cache_size = adev->gmc.mall_size / 1024;
1486 pcache_info[i].cache_level = 3;
1487 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1488 CRAT_CACHE_FLAGS_DATA_CACHE |
1489 CRAT_CACHE_FLAGS_SIMD_CACHE);
1490 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
1491 i++;
1492 }
1493 return i;
1494}
1495
1496static int kfd_fill_gpu_cache_info_from_gfx_config_v2(struct kfd_dev *kdev,
1497 struct kfd_gpu_cache_info *pcache_info)
1498{
1499 struct amdgpu_device *adev = kdev->adev;
1500 int i = 0;
1501
1502 /* TCP L1 Cache per CU */
1503 if (adev->gfx.config.gc_tcp_size_per_cu) {
1504 pcache_info[i].cache_size = adev->gfx.config.gc_tcp_size_per_cu;
1505 pcache_info[i].cache_level = 1;
1506 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1507 CRAT_CACHE_FLAGS_DATA_CACHE |
1508 CRAT_CACHE_FLAGS_SIMD_CACHE);
1509 pcache_info[i].num_cu_shared = 1;
1510 i++;
1511 }
1512 /* Scalar L1 Instruction Cache per SQC */
1513 if (adev->gfx.config.gc_l1_instruction_cache_size_per_sqc) {
1514 pcache_info[i].cache_size =
1515 adev->gfx.config.gc_l1_instruction_cache_size_per_sqc;
1516 pcache_info[i].cache_level = 1;
1517 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1518 CRAT_CACHE_FLAGS_INST_CACHE |
1519 CRAT_CACHE_FLAGS_SIMD_CACHE);
1520 pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_cu_per_sqc;
1521 i++;
1522 }
1523 /* Scalar L1 Data Cache per SQC */
1524 if (adev->gfx.config.gc_l1_data_cache_size_per_sqc) {
1525 pcache_info[i].cache_size = adev->gfx.config.gc_l1_data_cache_size_per_sqc;
1526 pcache_info[i].cache_level = 1;
1527 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1528 CRAT_CACHE_FLAGS_DATA_CACHE |
1529 CRAT_CACHE_FLAGS_SIMD_CACHE);
1530 pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_cu_per_sqc;
1531 i++;
1532 }
1533 /* L2 Data Cache per GPU (Total Tex Cache) */
1534 if (adev->gfx.config.gc_tcc_size) {
1535 pcache_info[i].cache_size = adev->gfx.config.gc_tcc_size;
1536 pcache_info[i].cache_level = 2;
1537 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1538 CRAT_CACHE_FLAGS_DATA_CACHE |
1539 CRAT_CACHE_FLAGS_SIMD_CACHE);
1540 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
1541 i++;
1542 }
1543 /* L3 Data Cache per GPU */
1544 if (adev->gmc.mall_size) {
1545 pcache_info[i].cache_size = adev->gmc.mall_size / 1024;
1546 pcache_info[i].cache_level = 3;
1547 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1548 CRAT_CACHE_FLAGS_DATA_CACHE |
1549 CRAT_CACHE_FLAGS_SIMD_CACHE);
1550 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
1551 i++;
1552 }
1553 return i;
1554}
1555
1556int kfd_get_gpu_cache_info(struct kfd_node *kdev, struct kfd_gpu_cache_info **pcache_info)
1557{
1558 int num_of_cache_types = 0;
1559
1560 switch (kdev->adev->asic_type) {
1561 case CHIP_KAVERI:
1562 *pcache_info = kaveri_cache_info;
1563 num_of_cache_types = ARRAY_SIZE(kaveri_cache_info);
1564 break;
1565 case CHIP_HAWAII:
1566 *pcache_info = hawaii_cache_info;
1567 num_of_cache_types = ARRAY_SIZE(hawaii_cache_info);
1568 break;
1569 case CHIP_CARRIZO:
1570 *pcache_info = carrizo_cache_info;
1571 num_of_cache_types = ARRAY_SIZE(carrizo_cache_info);
1572 break;
1573 case CHIP_TONGA:
1574 *pcache_info = tonga_cache_info;
1575 num_of_cache_types = ARRAY_SIZE(tonga_cache_info);
1576 break;
1577 case CHIP_FIJI:
1578 *pcache_info = fiji_cache_info;
1579 num_of_cache_types = ARRAY_SIZE(fiji_cache_info);
1580 break;
1581 case CHIP_POLARIS10:
1582 *pcache_info = polaris10_cache_info;
1583 num_of_cache_types = ARRAY_SIZE(polaris10_cache_info);
1584 break;
1585 case CHIP_POLARIS11:
1586 *pcache_info = polaris11_cache_info;
1587 num_of_cache_types = ARRAY_SIZE(polaris11_cache_info);
1588 break;
1589 case CHIP_POLARIS12:
1590 *pcache_info = polaris12_cache_info;
1591 num_of_cache_types = ARRAY_SIZE(polaris12_cache_info);
1592 break;
1593 case CHIP_VEGAM:
1594 *pcache_info = vegam_cache_info;
1595 num_of_cache_types = ARRAY_SIZE(vegam_cache_info);
1596 break;
1597 default:
1598 switch (KFD_GC_VERSION(kdev)) {
1599 case IP_VERSION(9, 0, 1):
1600 *pcache_info = vega10_cache_info;
1601 num_of_cache_types = ARRAY_SIZE(vega10_cache_info);
1602 break;
1603 case IP_VERSION(9, 2, 1):
1604 *pcache_info = vega12_cache_info;
1605 num_of_cache_types = ARRAY_SIZE(vega12_cache_info);
1606 break;
1607 case IP_VERSION(9, 4, 0):
1608 case IP_VERSION(9, 4, 1):
1609 *pcache_info = vega20_cache_info;
1610 num_of_cache_types = ARRAY_SIZE(vega20_cache_info);
1611 break;
1612 case IP_VERSION(9, 4, 2):
1613 *pcache_info = aldebaran_cache_info;
1614 num_of_cache_types = ARRAY_SIZE(aldebaran_cache_info);
1615 break;
1616 case IP_VERSION(9, 4, 3):
1617 num_of_cache_types =
1618 kfd_fill_gpu_cache_info_from_gfx_config_v2(kdev->kfd,
1619 *pcache_info);
1620 break;
1621 case IP_VERSION(9, 1, 0):
1622 case IP_VERSION(9, 2, 2):
1623 *pcache_info = raven_cache_info;
1624 num_of_cache_types = ARRAY_SIZE(raven_cache_info);
1625 break;
1626 case IP_VERSION(9, 3, 0):
1627 *pcache_info = renoir_cache_info;
1628 num_of_cache_types = ARRAY_SIZE(renoir_cache_info);
1629 break;
1630 case IP_VERSION(10, 1, 10):
1631 case IP_VERSION(10, 1, 2):
1632 case IP_VERSION(10, 1, 3):
1633 case IP_VERSION(10, 1, 4):
1634 *pcache_info = navi10_cache_info;
1635 num_of_cache_types = ARRAY_SIZE(navi10_cache_info);
1636 break;
1637 case IP_VERSION(10, 1, 1):
1638 *pcache_info = navi14_cache_info;
1639 num_of_cache_types = ARRAY_SIZE(navi14_cache_info);
1640 break;
1641 case IP_VERSION(10, 3, 0):
1642 *pcache_info = sienna_cichlid_cache_info;
1643 num_of_cache_types = ARRAY_SIZE(sienna_cichlid_cache_info);
1644 break;
1645 case IP_VERSION(10, 3, 2):
1646 *pcache_info = navy_flounder_cache_info;
1647 num_of_cache_types = ARRAY_SIZE(navy_flounder_cache_info);
1648 break;
1649 case IP_VERSION(10, 3, 4):
1650 *pcache_info = dimgrey_cavefish_cache_info;
1651 num_of_cache_types = ARRAY_SIZE(dimgrey_cavefish_cache_info);
1652 break;
1653 case IP_VERSION(10, 3, 1):
1654 *pcache_info = vangogh_cache_info;
1655 num_of_cache_types = ARRAY_SIZE(vangogh_cache_info);
1656 break;
1657 case IP_VERSION(10, 3, 5):
1658 *pcache_info = beige_goby_cache_info;
1659 num_of_cache_types = ARRAY_SIZE(beige_goby_cache_info);
1660 break;
1661 case IP_VERSION(10, 3, 3):
1662 *pcache_info = yellow_carp_cache_info;
1663 num_of_cache_types = ARRAY_SIZE(yellow_carp_cache_info);
1664 break;
1665 case IP_VERSION(10, 3, 6):
1666 *pcache_info = gc_10_3_6_cache_info;
1667 num_of_cache_types = ARRAY_SIZE(gc_10_3_6_cache_info);
1668 break;
1669 case IP_VERSION(10, 3, 7):
1670 *pcache_info = gfx1037_cache_info;
1671 num_of_cache_types = ARRAY_SIZE(gfx1037_cache_info);
1672 break;
1673 case IP_VERSION(11, 0, 0):
1674 case IP_VERSION(11, 0, 1):
1675 case IP_VERSION(11, 0, 2):
1676 case IP_VERSION(11, 0, 3):
1677 case IP_VERSION(11, 0, 4):
1678 case IP_VERSION(11, 5, 0):
1679 case IP_VERSION(11, 5, 1):
1680 num_of_cache_types =
1681 kfd_fill_gpu_cache_info_from_gfx_config(kdev->kfd, *pcache_info);
1682 break;
1683 default:
1684 *pcache_info = dummy_cache_info;
1685 num_of_cache_types = ARRAY_SIZE(dummy_cache_info);
1686 pr_warn("dummy cache info is used temporarily and real cache info need update later.\n");
1687 break;
1688 }
1689 }
1690 return num_of_cache_types;
1691}
1692
1693/* Memory required to create Virtual CRAT.
1694 * Since there is no easy way to predict the amount of memory required, the
1695 * following amount is allocated for GPU Virtual CRAT. This is
1696 * expected to cover all known conditions. But to be safe additional check
1697 * is put in the code to ensure we don't overwrite.
1698 */
1699#define VCRAT_SIZE_FOR_GPU (4 * PAGE_SIZE)
1700
1701/* kfd_fill_cu_for_cpu - Fill in Compute info for the given CPU NUMA node
1702 *
1703 * @numa_node_id: CPU NUMA node id
1704 * @avail_size: Available size in the memory
1705 * @sub_type_hdr: Memory into which compute info will be filled in
1706 *
1707 * Return 0 if successful else return -ve value
1708 */
1709static int kfd_fill_cu_for_cpu(int numa_node_id, int *avail_size,
1710 int proximity_domain,
1711 struct crat_subtype_computeunit *sub_type_hdr)
1712{
1713 const struct cpumask *cpumask;
1714
1715 *avail_size -= sizeof(struct crat_subtype_computeunit);
1716 if (*avail_size < 0)
1717 return -ENOMEM;
1718
1719 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit));
1720
1721 /* Fill in subtype header data */
1722 sub_type_hdr->type = CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY;
1723 sub_type_hdr->length = sizeof(struct crat_subtype_computeunit);
1724 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
1725
1726 cpumask = cpumask_of_node(numa_node_id);
1727
1728 /* Fill in CU data */
1729 sub_type_hdr->flags |= CRAT_CU_FLAGS_CPU_PRESENT;
1730 sub_type_hdr->proximity_domain = proximity_domain;
1731 sub_type_hdr->processor_id_low = kfd_numa_node_to_apic_id(numa_node_id);
1732 if (sub_type_hdr->processor_id_low == -1)
1733 return -EINVAL;
1734
1735 sub_type_hdr->num_cpu_cores = cpumask_weight(cpumask);
1736
1737 return 0;
1738}
1739
1740/* kfd_fill_mem_info_for_cpu - Fill in Memory info for the given CPU NUMA node
1741 *
1742 * @numa_node_id: CPU NUMA node id
1743 * @avail_size: Available size in the memory
1744 * @sub_type_hdr: Memory into which compute info will be filled in
1745 *
1746 * Return 0 if successful else return -ve value
1747 */
1748static int kfd_fill_mem_info_for_cpu(int numa_node_id, int *avail_size,
1749 int proximity_domain,
1750 struct crat_subtype_memory *sub_type_hdr)
1751{
1752 uint64_t mem_in_bytes = 0;
1753 pg_data_t *pgdat;
1754 int zone_type;
1755
1756 *avail_size -= sizeof(struct crat_subtype_memory);
1757 if (*avail_size < 0)
1758 return -ENOMEM;
1759
1760 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_memory));
1761
1762 /* Fill in subtype header data */
1763 sub_type_hdr->type = CRAT_SUBTYPE_MEMORY_AFFINITY;
1764 sub_type_hdr->length = sizeof(struct crat_subtype_memory);
1765 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
1766
1767 /* Fill in Memory Subunit data */
1768
1769 /* Unlike si_meminfo, si_meminfo_node is not exported. So
1770 * the following lines are duplicated from si_meminfo_node
1771 * function
1772 */
1773 pgdat = NODE_DATA(numa_node_id);
1774 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
1775 mem_in_bytes += zone_managed_pages(&pgdat->node_zones[zone_type]);
1776 mem_in_bytes <<= PAGE_SHIFT;
1777
1778 sub_type_hdr->length_low = lower_32_bits(mem_in_bytes);
1779 sub_type_hdr->length_high = upper_32_bits(mem_in_bytes);
1780 sub_type_hdr->proximity_domain = proximity_domain;
1781
1782 return 0;
1783}
1784
1785#ifdef CONFIG_X86_64
1786static int kfd_fill_iolink_info_for_cpu(int numa_node_id, int *avail_size,
1787 uint32_t *num_entries,
1788 struct crat_subtype_iolink *sub_type_hdr)
1789{
1790 int nid;
1791 struct cpuinfo_x86 *c = &cpu_data(0);
1792 uint8_t link_type;
1793
1794 if (c->x86_vendor == X86_VENDOR_AMD)
1795 link_type = CRAT_IOLINK_TYPE_HYPERTRANSPORT;
1796 else
1797 link_type = CRAT_IOLINK_TYPE_QPI_1_1;
1798
1799 *num_entries = 0;
1800
1801 /* Create IO links from this node to other CPU nodes */
1802 for_each_online_node(nid) {
1803 if (nid == numa_node_id) /* node itself */
1804 continue;
1805
1806 *avail_size -= sizeof(struct crat_subtype_iolink);
1807 if (*avail_size < 0)
1808 return -ENOMEM;
1809
1810 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
1811
1812 /* Fill in subtype header data */
1813 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
1814 sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
1815 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
1816
1817 /* Fill in IO link data */
1818 sub_type_hdr->proximity_domain_from = numa_node_id;
1819 sub_type_hdr->proximity_domain_to = nid;
1820 sub_type_hdr->io_interface_type = link_type;
1821
1822 (*num_entries)++;
1823 sub_type_hdr++;
1824 }
1825
1826 return 0;
1827}
1828#endif
1829
1830/* kfd_create_vcrat_image_cpu - Create Virtual CRAT for CPU
1831 *
1832 * @pcrat_image: Fill in VCRAT for CPU
1833 * @size: [IN] allocated size of crat_image.
1834 * [OUT] actual size of data filled in crat_image
1835 */
1836static int kfd_create_vcrat_image_cpu(void *pcrat_image, size_t *size)
1837{
1838 struct crat_header *crat_table = (struct crat_header *)pcrat_image;
1839 struct acpi_table_header *acpi_table;
1840 acpi_status status;
1841 struct crat_subtype_generic *sub_type_hdr;
1842 int avail_size = *size;
1843 int numa_node_id;
1844#ifdef CONFIG_X86_64
1845 uint32_t entries = 0;
1846#endif
1847 int ret = 0;
1848
1849 if (!pcrat_image)
1850 return -EINVAL;
1851
1852 /* Fill in CRAT Header.
1853 * Modify length and total_entries as subunits are added.
1854 */
1855 avail_size -= sizeof(struct crat_header);
1856 if (avail_size < 0)
1857 return -ENOMEM;
1858
1859 memset(crat_table, 0, sizeof(struct crat_header));
1860 memcpy(&crat_table->signature, CRAT_SIGNATURE,
1861 sizeof(crat_table->signature));
1862 crat_table->length = sizeof(struct crat_header);
1863
1864 status = acpi_get_table("DSDT", 0, &acpi_table);
1865 if (status != AE_OK)
1866 pr_warn("DSDT table not found for OEM information\n");
1867 else {
1868 crat_table->oem_revision = acpi_table->revision;
1869 memcpy(crat_table->oem_id, acpi_table->oem_id,
1870 CRAT_OEMID_LENGTH);
1871 memcpy(crat_table->oem_table_id, acpi_table->oem_table_id,
1872 CRAT_OEMTABLEID_LENGTH);
1873 acpi_put_table(acpi_table);
1874 }
1875 crat_table->total_entries = 0;
1876 crat_table->num_domains = 0;
1877
1878 sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1);
1879
1880 for_each_online_node(numa_node_id) {
1881 if (kfd_numa_node_to_apic_id(numa_node_id) == -1)
1882 continue;
1883
1884 /* Fill in Subtype: Compute Unit */
1885 ret = kfd_fill_cu_for_cpu(numa_node_id, &avail_size,
1886 crat_table->num_domains,
1887 (struct crat_subtype_computeunit *)sub_type_hdr);
1888 if (ret < 0)
1889 return ret;
1890 crat_table->length += sub_type_hdr->length;
1891 crat_table->total_entries++;
1892
1893 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1894 sub_type_hdr->length);
1895
1896 /* Fill in Subtype: Memory */
1897 ret = kfd_fill_mem_info_for_cpu(numa_node_id, &avail_size,
1898 crat_table->num_domains,
1899 (struct crat_subtype_memory *)sub_type_hdr);
1900 if (ret < 0)
1901 return ret;
1902 crat_table->length += sub_type_hdr->length;
1903 crat_table->total_entries++;
1904
1905 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1906 sub_type_hdr->length);
1907
1908 /* Fill in Subtype: IO Link */
1909#ifdef CONFIG_X86_64
1910 ret = kfd_fill_iolink_info_for_cpu(numa_node_id, &avail_size,
1911 &entries,
1912 (struct crat_subtype_iolink *)sub_type_hdr);
1913 if (ret < 0)
1914 return ret;
1915
1916 if (entries) {
1917 crat_table->length += (sub_type_hdr->length * entries);
1918 crat_table->total_entries += entries;
1919
1920 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1921 sub_type_hdr->length * entries);
1922 }
1923#else
1924 pr_info("IO link not available for non x86 platforms\n");
1925#endif
1926
1927 crat_table->num_domains++;
1928 }
1929
1930 /* TODO: Add cache Subtype for CPU.
1931 * Currently, CPU cache information is available in function
1932 * detect_cache_attributes(cpu) defined in the file
1933 * ./arch/x86/kernel/cpu/intel_cacheinfo.c. This function is not
1934 * exported and to get the same information the code needs to be
1935 * duplicated.
1936 */
1937
1938 *size = crat_table->length;
1939 pr_info("Virtual CRAT table created for CPU\n");
1940
1941 return 0;
1942}
1943
1944static int kfd_fill_gpu_memory_affinity(int *avail_size,
1945 struct kfd_node *kdev, uint8_t type, uint64_t size,
1946 struct crat_subtype_memory *sub_type_hdr,
1947 uint32_t proximity_domain,
1948 const struct kfd_local_mem_info *local_mem_info)
1949{
1950 *avail_size -= sizeof(struct crat_subtype_memory);
1951 if (*avail_size < 0)
1952 return -ENOMEM;
1953
1954 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_memory));
1955 sub_type_hdr->type = CRAT_SUBTYPE_MEMORY_AFFINITY;
1956 sub_type_hdr->length = sizeof(struct crat_subtype_memory);
1957 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED;
1958
1959 sub_type_hdr->proximity_domain = proximity_domain;
1960
1961 pr_debug("Fill gpu memory affinity - type 0x%x size 0x%llx\n",
1962 type, size);
1963
1964 sub_type_hdr->length_low = lower_32_bits(size);
1965 sub_type_hdr->length_high = upper_32_bits(size);
1966
1967 sub_type_hdr->width = local_mem_info->vram_width;
1968 sub_type_hdr->visibility_type = type;
1969
1970 return 0;
1971}
1972
1973#ifdef CONFIG_ACPI_NUMA
1974static void kfd_find_numa_node_in_srat(struct kfd_node *kdev)
1975{
1976 struct acpi_table_header *table_header = NULL;
1977 struct acpi_subtable_header *sub_header = NULL;
1978 unsigned long table_end, subtable_len;
1979 u32 pci_id = pci_domain_nr(kdev->adev->pdev->bus) << 16 |
1980 pci_dev_id(kdev->adev->pdev);
1981 u32 bdf;
1982 acpi_status status;
1983 struct acpi_srat_cpu_affinity *cpu;
1984 struct acpi_srat_generic_affinity *gpu;
1985 int pxm = 0, max_pxm = 0;
1986 int numa_node = NUMA_NO_NODE;
1987 bool found = false;
1988
1989 /* Fetch the SRAT table from ACPI */
1990 status = acpi_get_table(ACPI_SIG_SRAT, 0, &table_header);
1991 if (status == AE_NOT_FOUND) {
1992 pr_warn("SRAT table not found\n");
1993 return;
1994 } else if (ACPI_FAILURE(status)) {
1995 const char *err = acpi_format_exception(status);
1996 pr_err("SRAT table error: %s\n", err);
1997 return;
1998 }
1999
2000 table_end = (unsigned long)table_header + table_header->length;
2001
2002 /* Parse all entries looking for a match. */
2003 sub_header = (struct acpi_subtable_header *)
2004 ((unsigned long)table_header +
2005 sizeof(struct acpi_table_srat));
2006 subtable_len = sub_header->length;
2007
2008 while (((unsigned long)sub_header) + subtable_len < table_end) {
2009 /*
2010 * If length is 0, break from this loop to avoid
2011 * infinite loop.
2012 */
2013 if (subtable_len == 0) {
2014 pr_err("SRAT invalid zero length\n");
2015 break;
2016 }
2017
2018 switch (sub_header->type) {
2019 case ACPI_SRAT_TYPE_CPU_AFFINITY:
2020 cpu = (struct acpi_srat_cpu_affinity *)sub_header;
2021 pxm = *((u32 *)cpu->proximity_domain_hi) << 8 |
2022 cpu->proximity_domain_lo;
2023 if (pxm > max_pxm)
2024 max_pxm = pxm;
2025 break;
2026 case ACPI_SRAT_TYPE_GENERIC_AFFINITY:
2027 gpu = (struct acpi_srat_generic_affinity *)sub_header;
2028 bdf = *((u16 *)(&gpu->device_handle[0])) << 16 |
2029 *((u16 *)(&gpu->device_handle[2]));
2030 if (bdf == pci_id) {
2031 found = true;
2032 numa_node = pxm_to_node(gpu->proximity_domain);
2033 }
2034 break;
2035 default:
2036 break;
2037 }
2038
2039 if (found)
2040 break;
2041
2042 sub_header = (struct acpi_subtable_header *)
2043 ((unsigned long)sub_header + subtable_len);
2044 subtable_len = sub_header->length;
2045 }
2046
2047 acpi_put_table(table_header);
2048
2049 /* Workaround bad cpu-gpu binding case */
2050 if (found && (numa_node < 0 ||
2051 numa_node > pxm_to_node(max_pxm)))
2052 numa_node = 0;
2053
2054 if (numa_node != NUMA_NO_NODE)
2055 set_dev_node(&kdev->adev->pdev->dev, numa_node);
2056}
2057#endif
2058
2059#define KFD_CRAT_INTRA_SOCKET_WEIGHT 13
2060#define KFD_CRAT_XGMI_WEIGHT 15
2061
2062/* kfd_fill_gpu_direct_io_link - Fill in direct io link from GPU
2063 * to its NUMA node
2064 * @avail_size: Available size in the memory
2065 * @kdev - [IN] GPU device
2066 * @sub_type_hdr: Memory into which io link info will be filled in
2067 * @proximity_domain - proximity domain of the GPU node
2068 *
2069 * Return 0 if successful else return -ve value
2070 */
2071static int kfd_fill_gpu_direct_io_link_to_cpu(int *avail_size,
2072 struct kfd_node *kdev,
2073 struct crat_subtype_iolink *sub_type_hdr,
2074 uint32_t proximity_domain)
2075{
2076 *avail_size -= sizeof(struct crat_subtype_iolink);
2077 if (*avail_size < 0)
2078 return -ENOMEM;
2079
2080 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
2081
2082 /* Fill in subtype header data */
2083 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
2084 sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
2085 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED;
2086 if (kfd_dev_is_large_bar(kdev))
2087 sub_type_hdr->flags |= CRAT_IOLINK_FLAGS_BI_DIRECTIONAL;
2088
2089 /* Fill in IOLINK subtype.
2090 * TODO: Fill-in other fields of iolink subtype
2091 */
2092 if (kdev->adev->gmc.xgmi.connected_to_cpu ||
2093 (KFD_GC_VERSION(kdev) == IP_VERSION(9, 4, 3) &&
2094 kdev->adev->smuio.funcs->get_pkg_type(kdev->adev) ==
2095 AMDGPU_PKG_TYPE_APU)) {
2096 bool ext_cpu = KFD_GC_VERSION(kdev) != IP_VERSION(9, 4, 3);
2097 int mem_bw = 819200, weight = ext_cpu ? KFD_CRAT_XGMI_WEIGHT :
2098 KFD_CRAT_INTRA_SOCKET_WEIGHT;
2099 uint32_t bandwidth = ext_cpu ? amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(
2100 kdev->adev, NULL, true) : mem_bw;
2101
2102 /*
2103 * with host gpu xgmi link, host can access gpu memory whether
2104 * or not pcie bar type is large, so always create bidirectional
2105 * io link.
2106 */
2107 sub_type_hdr->flags |= CRAT_IOLINK_FLAGS_BI_DIRECTIONAL;
2108 sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_XGMI;
2109 sub_type_hdr->weight_xgmi = weight;
2110 sub_type_hdr->minimum_bandwidth_mbs = bandwidth;
2111 sub_type_hdr->maximum_bandwidth_mbs = bandwidth;
2112 } else {
2113 sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_PCIEXPRESS;
2114 sub_type_hdr->minimum_bandwidth_mbs =
2115 amdgpu_amdkfd_get_pcie_bandwidth_mbytes(kdev->adev, true);
2116 sub_type_hdr->maximum_bandwidth_mbs =
2117 amdgpu_amdkfd_get_pcie_bandwidth_mbytes(kdev->adev, false);
2118 }
2119
2120 sub_type_hdr->proximity_domain_from = proximity_domain;
2121
2122#ifdef CONFIG_ACPI_NUMA
2123 if (kdev->adev->pdev->dev.numa_node == NUMA_NO_NODE &&
2124 num_possible_nodes() > 1)
2125 kfd_find_numa_node_in_srat(kdev);
2126#endif
2127#ifdef CONFIG_NUMA
2128 if (kdev->adev->pdev->dev.numa_node == NUMA_NO_NODE)
2129 sub_type_hdr->proximity_domain_to = 0;
2130 else
2131 sub_type_hdr->proximity_domain_to = kdev->adev->pdev->dev.numa_node;
2132#else
2133 sub_type_hdr->proximity_domain_to = 0;
2134#endif
2135 return 0;
2136}
2137
2138static int kfd_fill_gpu_xgmi_link_to_gpu(int *avail_size,
2139 struct kfd_node *kdev,
2140 struct kfd_node *peer_kdev,
2141 struct crat_subtype_iolink *sub_type_hdr,
2142 uint32_t proximity_domain_from,
2143 uint32_t proximity_domain_to)
2144{
2145 bool use_ta_info = kdev->kfd->num_nodes == 1;
2146
2147 *avail_size -= sizeof(struct crat_subtype_iolink);
2148 if (*avail_size < 0)
2149 return -ENOMEM;
2150
2151 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
2152
2153 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
2154 sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
2155 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED |
2156 CRAT_IOLINK_FLAGS_BI_DIRECTIONAL;
2157
2158 sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_XGMI;
2159 sub_type_hdr->proximity_domain_from = proximity_domain_from;
2160 sub_type_hdr->proximity_domain_to = proximity_domain_to;
2161
2162 if (use_ta_info) {
2163 sub_type_hdr->weight_xgmi = KFD_CRAT_XGMI_WEIGHT *
2164 amdgpu_amdkfd_get_xgmi_hops_count(kdev->adev, peer_kdev->adev);
2165 sub_type_hdr->maximum_bandwidth_mbs =
2166 amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(kdev->adev,
2167 peer_kdev->adev, false);
2168 sub_type_hdr->minimum_bandwidth_mbs = sub_type_hdr->maximum_bandwidth_mbs ?
2169 amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(kdev->adev, NULL, true) : 0;
2170 } else {
2171 bool is_single_hop = kdev->kfd == peer_kdev->kfd;
2172 int weight = is_single_hop ? KFD_CRAT_INTRA_SOCKET_WEIGHT :
2173 (2 * KFD_CRAT_INTRA_SOCKET_WEIGHT) + KFD_CRAT_XGMI_WEIGHT;
2174 int mem_bw = 819200;
2175
2176 sub_type_hdr->weight_xgmi = weight;
2177 sub_type_hdr->maximum_bandwidth_mbs = is_single_hop ? mem_bw : 0;
2178 sub_type_hdr->minimum_bandwidth_mbs = is_single_hop ? mem_bw : 0;
2179 }
2180
2181 return 0;
2182}
2183
2184/* kfd_create_vcrat_image_gpu - Create Virtual CRAT for CPU
2185 *
2186 * @pcrat_image: Fill in VCRAT for GPU
2187 * @size: [IN] allocated size of crat_image.
2188 * [OUT] actual size of data filled in crat_image
2189 */
2190static int kfd_create_vcrat_image_gpu(void *pcrat_image,
2191 size_t *size, struct kfd_node *kdev,
2192 uint32_t proximity_domain)
2193{
2194 struct crat_header *crat_table = (struct crat_header *)pcrat_image;
2195 struct amdgpu_gfx_config *gfx_info = &kdev->adev->gfx.config;
2196 struct amdgpu_cu_info *cu_info = &kdev->adev->gfx.cu_info;
2197 struct crat_subtype_generic *sub_type_hdr;
2198 struct kfd_local_mem_info local_mem_info;
2199 struct kfd_topology_device *peer_dev;
2200 struct crat_subtype_computeunit *cu;
2201 int avail_size = *size;
2202 uint32_t total_num_of_cu;
2203 uint32_t nid = 0;
2204 int ret = 0;
2205
2206 if (!pcrat_image || avail_size < VCRAT_SIZE_FOR_GPU)
2207 return -EINVAL;
2208
2209 /* Fill the CRAT Header.
2210 * Modify length and total_entries as subunits are added.
2211 */
2212 avail_size -= sizeof(struct crat_header);
2213 if (avail_size < 0)
2214 return -ENOMEM;
2215
2216 memset(crat_table, 0, sizeof(struct crat_header));
2217
2218 memcpy(&crat_table->signature, CRAT_SIGNATURE,
2219 sizeof(crat_table->signature));
2220 /* Change length as we add more subtypes*/
2221 crat_table->length = sizeof(struct crat_header);
2222 crat_table->num_domains = 1;
2223 crat_table->total_entries = 0;
2224
2225 /* Fill in Subtype: Compute Unit
2226 * First fill in the sub type header and then sub type data
2227 */
2228 avail_size -= sizeof(struct crat_subtype_computeunit);
2229 if (avail_size < 0)
2230 return -ENOMEM;
2231
2232 sub_type_hdr = (struct crat_subtype_generic *)(crat_table + 1);
2233 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit));
2234
2235 sub_type_hdr->type = CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY;
2236 sub_type_hdr->length = sizeof(struct crat_subtype_computeunit);
2237 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
2238
2239 /* Fill CU subtype data */
2240 cu = (struct crat_subtype_computeunit *)sub_type_hdr;
2241 cu->flags |= CRAT_CU_FLAGS_GPU_PRESENT;
2242 cu->proximity_domain = proximity_domain;
2243
2244 cu->num_simd_per_cu = cu_info->simd_per_cu;
2245 cu->num_simd_cores = cu_info->simd_per_cu *
2246 (cu_info->number / kdev->kfd->num_nodes);
2247 cu->max_waves_simd = cu_info->max_waves_per_simd;
2248
2249 cu->wave_front_size = cu_info->wave_front_size;
2250 cu->array_count = gfx_info->max_sh_per_se *
2251 gfx_info->max_shader_engines;
2252 total_num_of_cu = (cu->array_count * gfx_info->max_cu_per_sh);
2253 cu->processor_id_low = get_and_inc_gpu_processor_id(total_num_of_cu);
2254 cu->num_cu_per_array = gfx_info->max_cu_per_sh;
2255 cu->max_slots_scatch_cu = cu_info->max_scratch_slots_per_cu;
2256 cu->num_banks = gfx_info->max_shader_engines;
2257 cu->lds_size_in_kb = cu_info->lds_size;
2258
2259 cu->hsa_capability = 0;
2260
2261 crat_table->length += sub_type_hdr->length;
2262 crat_table->total_entries++;
2263
2264 /* Fill in Subtype: Memory. Only on systems with large BAR (no
2265 * private FB), report memory as public. On other systems
2266 * report the total FB size (public+private) as a single
2267 * private heap.
2268 */
2269 local_mem_info = kdev->local_mem_info;
2270 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
2271 sub_type_hdr->length);
2272
2273 if (kdev->adev->debug_largebar)
2274 local_mem_info.local_mem_size_private = 0;
2275
2276 if (local_mem_info.local_mem_size_private == 0)
2277 ret = kfd_fill_gpu_memory_affinity(&avail_size,
2278 kdev, HSA_MEM_HEAP_TYPE_FB_PUBLIC,
2279 local_mem_info.local_mem_size_public,
2280 (struct crat_subtype_memory *)sub_type_hdr,
2281 proximity_domain,
2282 &local_mem_info);
2283 else
2284 ret = kfd_fill_gpu_memory_affinity(&avail_size,
2285 kdev, HSA_MEM_HEAP_TYPE_FB_PRIVATE,
2286 local_mem_info.local_mem_size_public +
2287 local_mem_info.local_mem_size_private,
2288 (struct crat_subtype_memory *)sub_type_hdr,
2289 proximity_domain,
2290 &local_mem_info);
2291 if (ret < 0)
2292 return ret;
2293
2294 crat_table->length += sizeof(struct crat_subtype_memory);
2295 crat_table->total_entries++;
2296
2297 /* Fill in Subtype: IO_LINKS
2298 * Only direct links are added here which is Link from GPU to
2299 * its NUMA node. Indirect links are added by userspace.
2300 */
2301 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
2302 sub_type_hdr->length);
2303 ret = kfd_fill_gpu_direct_io_link_to_cpu(&avail_size, kdev,
2304 (struct crat_subtype_iolink *)sub_type_hdr, proximity_domain);
2305
2306 if (ret < 0)
2307 return ret;
2308
2309 crat_table->length += sub_type_hdr->length;
2310 crat_table->total_entries++;
2311
2312
2313 /* Fill in Subtype: IO_LINKS
2314 * Direct links from GPU to other GPUs through xGMI.
2315 * We will loop GPUs that already be processed (with lower value
2316 * of proximity_domain), add the link for the GPUs with same
2317 * hive id (from this GPU to other GPU) . The reversed iolink
2318 * (from other GPU to this GPU) will be added
2319 * in kfd_parse_subtype_iolink.
2320 */
2321 if (kdev->kfd->hive_id) {
2322 for (nid = 0; nid < proximity_domain; ++nid) {
2323 peer_dev = kfd_topology_device_by_proximity_domain_no_lock(nid);
2324 if (!peer_dev->gpu)
2325 continue;
2326 if (peer_dev->gpu->kfd->hive_id != kdev->kfd->hive_id)
2327 continue;
2328 sub_type_hdr = (typeof(sub_type_hdr))(
2329 (char *)sub_type_hdr +
2330 sizeof(struct crat_subtype_iolink));
2331 ret = kfd_fill_gpu_xgmi_link_to_gpu(
2332 &avail_size, kdev, peer_dev->gpu,
2333 (struct crat_subtype_iolink *)sub_type_hdr,
2334 proximity_domain, nid);
2335 if (ret < 0)
2336 return ret;
2337 crat_table->length += sub_type_hdr->length;
2338 crat_table->total_entries++;
2339 }
2340 }
2341 *size = crat_table->length;
2342 pr_info("Virtual CRAT table created for GPU\n");
2343
2344 return ret;
2345}
2346
2347/* kfd_create_crat_image_virtual - Allocates memory for CRAT image and
2348 * creates a Virtual CRAT (VCRAT) image
2349 *
2350 * NOTE: Call kfd_destroy_crat_image to free CRAT image memory
2351 *
2352 * @crat_image: VCRAT image created because ACPI does not have a
2353 * CRAT for this device
2354 * @size: [OUT] size of virtual crat_image
2355 * @flags: COMPUTE_UNIT_CPU - Create VCRAT for CPU device
2356 * COMPUTE_UNIT_GPU - Create VCRAT for GPU
2357 * (COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU) - Create VCRAT for APU
2358 * -- this option is not currently implemented.
2359 * The assumption is that all AMD APUs will have CRAT
2360 * @kdev: Valid kfd_node required if flags contain COMPUTE_UNIT_GPU
2361 *
2362 * Return 0 if successful else return -ve value
2363 */
2364int kfd_create_crat_image_virtual(void **crat_image, size_t *size,
2365 int flags, struct kfd_node *kdev,
2366 uint32_t proximity_domain)
2367{
2368 void *pcrat_image = NULL;
2369 int ret = 0, num_nodes;
2370 size_t dyn_size;
2371
2372 if (!crat_image)
2373 return -EINVAL;
2374
2375 *crat_image = NULL;
2376
2377 /* Allocate the CPU Virtual CRAT size based on the number of online
2378 * nodes. Allocate VCRAT_SIZE_FOR_GPU for GPU virtual CRAT image.
2379 * This should cover all the current conditions. A check is put not
2380 * to overwrite beyond allocated size for GPUs
2381 */
2382 switch (flags) {
2383 case COMPUTE_UNIT_CPU:
2384 num_nodes = num_online_nodes();
2385 dyn_size = sizeof(struct crat_header) +
2386 num_nodes * (sizeof(struct crat_subtype_computeunit) +
2387 sizeof(struct crat_subtype_memory) +
2388 (num_nodes - 1) * sizeof(struct crat_subtype_iolink));
2389 pcrat_image = kvmalloc(dyn_size, GFP_KERNEL);
2390 if (!pcrat_image)
2391 return -ENOMEM;
2392 *size = dyn_size;
2393 pr_debug("CRAT size is %ld", dyn_size);
2394 ret = kfd_create_vcrat_image_cpu(pcrat_image, size);
2395 break;
2396 case COMPUTE_UNIT_GPU:
2397 if (!kdev)
2398 return -EINVAL;
2399 pcrat_image = kvmalloc(VCRAT_SIZE_FOR_GPU, GFP_KERNEL);
2400 if (!pcrat_image)
2401 return -ENOMEM;
2402 *size = VCRAT_SIZE_FOR_GPU;
2403 ret = kfd_create_vcrat_image_gpu(pcrat_image, size, kdev,
2404 proximity_domain);
2405 break;
2406 case (COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU):
2407 /* TODO: */
2408 ret = -EINVAL;
2409 pr_err("VCRAT not implemented for APU\n");
2410 break;
2411 default:
2412 ret = -EINVAL;
2413 }
2414
2415 if (!ret)
2416 *crat_image = pcrat_image;
2417 else
2418 kvfree(pcrat_image);
2419
2420 return ret;
2421}
2422
2423
2424/* kfd_destroy_crat_image
2425 *
2426 * @crat_image: [IN] - crat_image from kfd_create_crat_image_xxx(..)
2427 *
2428 */
2429void kfd_destroy_crat_image(void *crat_image)
2430{
2431 kvfree(crat_image);
2432}