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#include "amdgpu_xgmi.h"
32
33/* GPU Processor ID base for dGPUs for which VCRAT needs to be created.
34 * GPU processor ID are expressed with Bit[31]=1.
35 * The base is set to 0x8000_0000 + 0x1000 to avoid collision with GPU IDs
36 * used in the CRAT.
37 */
38static uint32_t gpu_processor_id_low = 0x80001000;
39
40/* Return the next available gpu_processor_id and increment it for next GPU
41 * @total_cu_count - Total CUs present in the GPU including ones
42 * masked off
43 */
44static inline unsigned int get_and_inc_gpu_processor_id(
45 unsigned int total_cu_count)
46{
47 int current_id = gpu_processor_id_low;
48
49 gpu_processor_id_low += total_cu_count;
50 return current_id;
51}
52
53
54static struct kfd_gpu_cache_info kaveri_cache_info[] = {
55 {
56 /* TCP L1 Cache per CU */
57 .cache_size = 16,
58 .cache_level = 1,
59 .cache_line_size = 64,
60 .flags = (CRAT_CACHE_FLAGS_ENABLED |
61 CRAT_CACHE_FLAGS_DATA_CACHE |
62 CRAT_CACHE_FLAGS_SIMD_CACHE),
63 .num_cu_shared = 1,
64 },
65 {
66 /* Scalar L1 Instruction Cache (in SQC module) per bank */
67 .cache_size = 16,
68 .cache_level = 1,
69 .cache_line_size = 64,
70 .flags = (CRAT_CACHE_FLAGS_ENABLED |
71 CRAT_CACHE_FLAGS_INST_CACHE |
72 CRAT_CACHE_FLAGS_SIMD_CACHE),
73 .num_cu_shared = 2,
74 },
75 {
76 /* Scalar L1 Data Cache (in SQC module) per bank */
77 .cache_size = 8,
78 .cache_level = 1,
79 .cache_line_size = 64,
80 .flags = (CRAT_CACHE_FLAGS_ENABLED |
81 CRAT_CACHE_FLAGS_DATA_CACHE |
82 CRAT_CACHE_FLAGS_SIMD_CACHE),
83 .num_cu_shared = 2,
84 },
85
86 /* TODO: Add L2 Cache information */
87};
88
89
90static struct kfd_gpu_cache_info carrizo_cache_info[] = {
91 {
92 /* TCP L1 Cache per CU */
93 .cache_size = 16,
94 .cache_level = 1,
95 .cache_line_size = 64,
96 .flags = (CRAT_CACHE_FLAGS_ENABLED |
97 CRAT_CACHE_FLAGS_DATA_CACHE |
98 CRAT_CACHE_FLAGS_SIMD_CACHE),
99 .num_cu_shared = 1,
100 },
101 {
102 /* Scalar L1 Instruction Cache (in SQC module) per bank */
103 .cache_size = 32,
104 .cache_level = 1,
105 .cache_line_size = 64,
106 .flags = (CRAT_CACHE_FLAGS_ENABLED |
107 CRAT_CACHE_FLAGS_INST_CACHE |
108 CRAT_CACHE_FLAGS_SIMD_CACHE),
109 .num_cu_shared = 4,
110 },
111 {
112 /* Scalar L1 Data Cache (in SQC module) per bank. */
113 .cache_size = 16,
114 .cache_level = 1,
115 .cache_line_size = 64,
116 .flags = (CRAT_CACHE_FLAGS_ENABLED |
117 CRAT_CACHE_FLAGS_DATA_CACHE |
118 CRAT_CACHE_FLAGS_SIMD_CACHE),
119 .num_cu_shared = 4,
120 },
121
122 /* TODO: Add L2 Cache information */
123};
124
125#define hawaii_cache_info kaveri_cache_info
126#define tonga_cache_info carrizo_cache_info
127#define fiji_cache_info carrizo_cache_info
128#define polaris10_cache_info carrizo_cache_info
129#define polaris11_cache_info carrizo_cache_info
130#define polaris12_cache_info carrizo_cache_info
131#define vegam_cache_info carrizo_cache_info
132
133/* NOTE: L1 cache information has been updated and L2/L3
134 * cache information has been added for Vega10 and
135 * newer ASICs. The unit for cache_size is KiB.
136 * In future, check & update cache details
137 * for every new ASIC is required.
138 */
139
140static struct kfd_gpu_cache_info vega10_cache_info[] = {
141 {
142 /* TCP L1 Cache per CU */
143 .cache_size = 16,
144 .cache_level = 1,
145 .cache_line_size = 64,
146 .flags = (CRAT_CACHE_FLAGS_ENABLED |
147 CRAT_CACHE_FLAGS_DATA_CACHE |
148 CRAT_CACHE_FLAGS_SIMD_CACHE),
149 .num_cu_shared = 1,
150 },
151 {
152 /* Scalar L1 Instruction Cache per SQC */
153 .cache_size = 32,
154 .cache_level = 1,
155 .cache_line_size = 64,
156 .flags = (CRAT_CACHE_FLAGS_ENABLED |
157 CRAT_CACHE_FLAGS_INST_CACHE |
158 CRAT_CACHE_FLAGS_SIMD_CACHE),
159 .num_cu_shared = 3,
160 },
161 {
162 /* Scalar L1 Data Cache per SQC */
163 .cache_size = 16,
164 .cache_level = 1,
165 .cache_line_size = 64,
166 .flags = (CRAT_CACHE_FLAGS_ENABLED |
167 CRAT_CACHE_FLAGS_DATA_CACHE |
168 CRAT_CACHE_FLAGS_SIMD_CACHE),
169 .num_cu_shared = 3,
170 },
171 {
172 /* L2 Data Cache per GPU (Total Tex Cache) */
173 .cache_size = 4096,
174 .cache_level = 2,
175 .cache_line_size = 64,
176 .flags = (CRAT_CACHE_FLAGS_ENABLED |
177 CRAT_CACHE_FLAGS_DATA_CACHE |
178 CRAT_CACHE_FLAGS_SIMD_CACHE),
179 .num_cu_shared = 16,
180 },
181};
182
183static struct kfd_gpu_cache_info raven_cache_info[] = {
184 {
185 /* TCP L1 Cache per CU */
186 .cache_size = 16,
187 .cache_level = 1,
188 .cache_line_size = 64,
189 .flags = (CRAT_CACHE_FLAGS_ENABLED |
190 CRAT_CACHE_FLAGS_DATA_CACHE |
191 CRAT_CACHE_FLAGS_SIMD_CACHE),
192 .num_cu_shared = 1,
193 },
194 {
195 /* Scalar L1 Instruction Cache per SQC */
196 .cache_size = 32,
197 .cache_level = 1,
198 .cache_line_size = 64,
199 .flags = (CRAT_CACHE_FLAGS_ENABLED |
200 CRAT_CACHE_FLAGS_INST_CACHE |
201 CRAT_CACHE_FLAGS_SIMD_CACHE),
202 .num_cu_shared = 3,
203 },
204 {
205 /* Scalar L1 Data Cache per SQC */
206 .cache_size = 16,
207 .cache_level = 1,
208 .cache_line_size = 64,
209 .flags = (CRAT_CACHE_FLAGS_ENABLED |
210 CRAT_CACHE_FLAGS_DATA_CACHE |
211 CRAT_CACHE_FLAGS_SIMD_CACHE),
212 .num_cu_shared = 3,
213 },
214 {
215 /* L2 Data Cache per GPU (Total Tex Cache) */
216 .cache_size = 1024,
217 .cache_level = 2,
218 .cache_line_size = 64,
219 .flags = (CRAT_CACHE_FLAGS_ENABLED |
220 CRAT_CACHE_FLAGS_DATA_CACHE |
221 CRAT_CACHE_FLAGS_SIMD_CACHE),
222 .num_cu_shared = 11,
223 },
224};
225
226static struct kfd_gpu_cache_info renoir_cache_info[] = {
227 {
228 /* TCP L1 Cache per CU */
229 .cache_size = 16,
230 .cache_level = 1,
231 .cache_line_size = 64,
232 .flags = (CRAT_CACHE_FLAGS_ENABLED |
233 CRAT_CACHE_FLAGS_DATA_CACHE |
234 CRAT_CACHE_FLAGS_SIMD_CACHE),
235 .num_cu_shared = 1,
236 },
237 {
238 /* Scalar L1 Instruction Cache per SQC */
239 .cache_size = 32,
240 .cache_level = 1,
241 .cache_line_size = 64,
242 .flags = (CRAT_CACHE_FLAGS_ENABLED |
243 CRAT_CACHE_FLAGS_INST_CACHE |
244 CRAT_CACHE_FLAGS_SIMD_CACHE),
245 .num_cu_shared = 3,
246 },
247 {
248 /* Scalar L1 Data Cache per SQC */
249 .cache_size = 16,
250 .cache_level = 1,
251 .cache_line_size = 64,
252 .flags = (CRAT_CACHE_FLAGS_ENABLED |
253 CRAT_CACHE_FLAGS_DATA_CACHE |
254 CRAT_CACHE_FLAGS_SIMD_CACHE),
255 .num_cu_shared = 3,
256 },
257 {
258 /* L2 Data Cache per GPU (Total Tex Cache) */
259 .cache_size = 1024,
260 .cache_level = 2,
261 .cache_line_size = 64,
262 .flags = (CRAT_CACHE_FLAGS_ENABLED |
263 CRAT_CACHE_FLAGS_DATA_CACHE |
264 CRAT_CACHE_FLAGS_SIMD_CACHE),
265 .num_cu_shared = 8,
266 },
267};
268
269static struct kfd_gpu_cache_info vega12_cache_info[] = {
270 {
271 /* TCP L1 Cache per CU */
272 .cache_size = 16,
273 .cache_level = 1,
274 .cache_line_size = 64,
275 .flags = (CRAT_CACHE_FLAGS_ENABLED |
276 CRAT_CACHE_FLAGS_DATA_CACHE |
277 CRAT_CACHE_FLAGS_SIMD_CACHE),
278 .num_cu_shared = 1,
279 },
280 {
281 /* Scalar L1 Instruction Cache per SQC */
282 .cache_size = 32,
283 .cache_level = 1,
284 .cache_line_size = 64,
285 .flags = (CRAT_CACHE_FLAGS_ENABLED |
286 CRAT_CACHE_FLAGS_INST_CACHE |
287 CRAT_CACHE_FLAGS_SIMD_CACHE),
288 .num_cu_shared = 3,
289 },
290 {
291 /* Scalar L1 Data Cache per SQC */
292 .cache_size = 16,
293 .cache_level = 1,
294 .cache_line_size = 64,
295 .flags = (CRAT_CACHE_FLAGS_ENABLED |
296 CRAT_CACHE_FLAGS_DATA_CACHE |
297 CRAT_CACHE_FLAGS_SIMD_CACHE),
298 .num_cu_shared = 3,
299 },
300 {
301 /* L2 Data Cache per GPU (Total Tex Cache) */
302 .cache_size = 2048,
303 .cache_level = 2,
304 .cache_line_size = 64,
305 .flags = (CRAT_CACHE_FLAGS_ENABLED |
306 CRAT_CACHE_FLAGS_DATA_CACHE |
307 CRAT_CACHE_FLAGS_SIMD_CACHE),
308 .num_cu_shared = 5,
309 },
310};
311
312static struct kfd_gpu_cache_info vega20_cache_info[] = {
313 {
314 /* TCP L1 Cache per CU */
315 .cache_size = 16,
316 .cache_level = 1,
317 .cache_line_size = 64,
318 .flags = (CRAT_CACHE_FLAGS_ENABLED |
319 CRAT_CACHE_FLAGS_DATA_CACHE |
320 CRAT_CACHE_FLAGS_SIMD_CACHE),
321 .num_cu_shared = 1,
322 },
323 {
324 /* Scalar L1 Instruction Cache per SQC */
325 .cache_size = 32,
326 .cache_level = 1,
327 .cache_line_size = 64,
328 .flags = (CRAT_CACHE_FLAGS_ENABLED |
329 CRAT_CACHE_FLAGS_INST_CACHE |
330 CRAT_CACHE_FLAGS_SIMD_CACHE),
331 .num_cu_shared = 3,
332 },
333 {
334 /* Scalar L1 Data Cache per SQC */
335 .cache_size = 16,
336 .cache_level = 1,
337 .cache_line_size = 64,
338 .flags = (CRAT_CACHE_FLAGS_ENABLED |
339 CRAT_CACHE_FLAGS_DATA_CACHE |
340 CRAT_CACHE_FLAGS_SIMD_CACHE),
341 .num_cu_shared = 3,
342 },
343 {
344 /* L2 Data Cache per GPU (Total Tex Cache) */
345 .cache_size = 8192,
346 .cache_level = 2,
347 .cache_line_size = 64,
348 .flags = (CRAT_CACHE_FLAGS_ENABLED |
349 CRAT_CACHE_FLAGS_DATA_CACHE |
350 CRAT_CACHE_FLAGS_SIMD_CACHE),
351 .num_cu_shared = 16,
352 },
353};
354
355static struct kfd_gpu_cache_info aldebaran_cache_info[] = {
356 {
357 /* TCP L1 Cache per CU */
358 .cache_size = 16,
359 .cache_level = 1,
360 .cache_line_size = 64,
361 .flags = (CRAT_CACHE_FLAGS_ENABLED |
362 CRAT_CACHE_FLAGS_DATA_CACHE |
363 CRAT_CACHE_FLAGS_SIMD_CACHE),
364 .num_cu_shared = 1,
365 },
366 {
367 /* Scalar L1 Instruction Cache per SQC */
368 .cache_size = 32,
369 .cache_level = 1,
370 .cache_line_size = 64,
371 .flags = (CRAT_CACHE_FLAGS_ENABLED |
372 CRAT_CACHE_FLAGS_INST_CACHE |
373 CRAT_CACHE_FLAGS_SIMD_CACHE),
374 .num_cu_shared = 2,
375 },
376 {
377 /* Scalar L1 Data Cache per SQC */
378 .cache_size = 16,
379 .cache_level = 1,
380 .cache_line_size = 64,
381 .flags = (CRAT_CACHE_FLAGS_ENABLED |
382 CRAT_CACHE_FLAGS_DATA_CACHE |
383 CRAT_CACHE_FLAGS_SIMD_CACHE),
384 .num_cu_shared = 2,
385 },
386 {
387 /* L2 Data Cache per GPU (Total Tex Cache) */
388 .cache_size = 8192,
389 .cache_level = 2,
390 .cache_line_size = 128,
391 .flags = (CRAT_CACHE_FLAGS_ENABLED |
392 CRAT_CACHE_FLAGS_DATA_CACHE |
393 CRAT_CACHE_FLAGS_SIMD_CACHE),
394 .num_cu_shared = 14,
395 },
396};
397
398static struct kfd_gpu_cache_info navi10_cache_info[] = {
399 {
400 /* TCP L1 Cache per CU */
401 .cache_size = 16,
402 .cache_level = 1,
403 .cache_line_size = 128,
404 .flags = (CRAT_CACHE_FLAGS_ENABLED |
405 CRAT_CACHE_FLAGS_DATA_CACHE |
406 CRAT_CACHE_FLAGS_SIMD_CACHE),
407 .num_cu_shared = 1,
408 },
409 {
410 /* Scalar L1 Instruction Cache per SQC */
411 .cache_size = 32,
412 .cache_level = 1,
413 .cache_line_size = 64,
414 .flags = (CRAT_CACHE_FLAGS_ENABLED |
415 CRAT_CACHE_FLAGS_INST_CACHE |
416 CRAT_CACHE_FLAGS_SIMD_CACHE),
417 .num_cu_shared = 2,
418 },
419 {
420 /* Scalar L1 Data Cache per SQC */
421 .cache_size = 16,
422 .cache_level = 1,
423 .cache_line_size = 64,
424 .flags = (CRAT_CACHE_FLAGS_ENABLED |
425 CRAT_CACHE_FLAGS_DATA_CACHE |
426 CRAT_CACHE_FLAGS_SIMD_CACHE),
427 .num_cu_shared = 2,
428 },
429 {
430 /* GL1 Data Cache per SA */
431 .cache_size = 128,
432 .cache_level = 1,
433 .cache_line_size = 128,
434 .flags = (CRAT_CACHE_FLAGS_ENABLED |
435 CRAT_CACHE_FLAGS_DATA_CACHE |
436 CRAT_CACHE_FLAGS_SIMD_CACHE),
437 .num_cu_shared = 10,
438 },
439 {
440 /* L2 Data Cache per GPU (Total Tex Cache) */
441 .cache_size = 4096,
442 .cache_level = 2,
443 .cache_line_size = 128,
444 .flags = (CRAT_CACHE_FLAGS_ENABLED |
445 CRAT_CACHE_FLAGS_DATA_CACHE |
446 CRAT_CACHE_FLAGS_SIMD_CACHE),
447 .num_cu_shared = 10,
448 },
449};
450
451static struct kfd_gpu_cache_info vangogh_cache_info[] = {
452 {
453 /* TCP L1 Cache per CU */
454 .cache_size = 16,
455 .cache_level = 1,
456 .cache_line_size = 128,
457 .flags = (CRAT_CACHE_FLAGS_ENABLED |
458 CRAT_CACHE_FLAGS_DATA_CACHE |
459 CRAT_CACHE_FLAGS_SIMD_CACHE),
460 .num_cu_shared = 1,
461 },
462 {
463 /* Scalar L1 Instruction Cache per SQC */
464 .cache_size = 32,
465 .cache_level = 1,
466 .cache_line_size = 64,
467 .flags = (CRAT_CACHE_FLAGS_ENABLED |
468 CRAT_CACHE_FLAGS_INST_CACHE |
469 CRAT_CACHE_FLAGS_SIMD_CACHE),
470 .num_cu_shared = 2,
471 },
472 {
473 /* Scalar L1 Data Cache per SQC */
474 .cache_size = 16,
475 .cache_level = 1,
476 .cache_line_size = 64,
477 .flags = (CRAT_CACHE_FLAGS_ENABLED |
478 CRAT_CACHE_FLAGS_DATA_CACHE |
479 CRAT_CACHE_FLAGS_SIMD_CACHE),
480 .num_cu_shared = 2,
481 },
482 {
483 /* GL1 Data Cache per SA */
484 .cache_size = 128,
485 .cache_level = 1,
486 .cache_line_size = 128,
487 .flags = (CRAT_CACHE_FLAGS_ENABLED |
488 CRAT_CACHE_FLAGS_DATA_CACHE |
489 CRAT_CACHE_FLAGS_SIMD_CACHE),
490 .num_cu_shared = 8,
491 },
492 {
493 /* L2 Data Cache per GPU (Total Tex Cache) */
494 .cache_size = 1024,
495 .cache_level = 2,
496 .cache_line_size = 128,
497 .flags = (CRAT_CACHE_FLAGS_ENABLED |
498 CRAT_CACHE_FLAGS_DATA_CACHE |
499 CRAT_CACHE_FLAGS_SIMD_CACHE),
500 .num_cu_shared = 8,
501 },
502};
503
504static struct kfd_gpu_cache_info navi14_cache_info[] = {
505 {
506 /* TCP L1 Cache per CU */
507 .cache_size = 16,
508 .cache_level = 1,
509 .cache_line_size = 128,
510 .flags = (CRAT_CACHE_FLAGS_ENABLED |
511 CRAT_CACHE_FLAGS_DATA_CACHE |
512 CRAT_CACHE_FLAGS_SIMD_CACHE),
513 .num_cu_shared = 1,
514 },
515 {
516 /* Scalar L1 Instruction Cache per SQC */
517 .cache_size = 32,
518 .cache_level = 1,
519 .cache_line_size = 64,
520 .flags = (CRAT_CACHE_FLAGS_ENABLED |
521 CRAT_CACHE_FLAGS_INST_CACHE |
522 CRAT_CACHE_FLAGS_SIMD_CACHE),
523 .num_cu_shared = 2,
524 },
525 {
526 /* Scalar L1 Data Cache per SQC */
527 .cache_size = 16,
528 .cache_level = 1,
529 .cache_line_size = 64,
530 .flags = (CRAT_CACHE_FLAGS_ENABLED |
531 CRAT_CACHE_FLAGS_DATA_CACHE |
532 CRAT_CACHE_FLAGS_SIMD_CACHE),
533 .num_cu_shared = 2,
534 },
535 {
536 /* GL1 Data Cache per SA */
537 .cache_size = 128,
538 .cache_level = 1,
539 .cache_line_size = 128,
540 .flags = (CRAT_CACHE_FLAGS_ENABLED |
541 CRAT_CACHE_FLAGS_DATA_CACHE |
542 CRAT_CACHE_FLAGS_SIMD_CACHE),
543 .num_cu_shared = 12,
544 },
545 {
546 /* L2 Data Cache per GPU (Total Tex Cache) */
547 .cache_size = 2048,
548 .cache_level = 2,
549 .cache_line_size = 128,
550 .flags = (CRAT_CACHE_FLAGS_ENABLED |
551 CRAT_CACHE_FLAGS_DATA_CACHE |
552 CRAT_CACHE_FLAGS_SIMD_CACHE),
553 .num_cu_shared = 12,
554 },
555};
556
557static struct kfd_gpu_cache_info sienna_cichlid_cache_info[] = {
558 {
559 /* TCP L1 Cache per CU */
560 .cache_size = 16,
561 .cache_level = 1,
562 .cache_line_size = 128,
563 .flags = (CRAT_CACHE_FLAGS_ENABLED |
564 CRAT_CACHE_FLAGS_DATA_CACHE |
565 CRAT_CACHE_FLAGS_SIMD_CACHE),
566 .num_cu_shared = 1,
567 },
568 {
569 /* Scalar L1 Instruction Cache per SQC */
570 .cache_size = 32,
571 .cache_level = 1,
572 .cache_line_size = 64,
573 .flags = (CRAT_CACHE_FLAGS_ENABLED |
574 CRAT_CACHE_FLAGS_INST_CACHE |
575 CRAT_CACHE_FLAGS_SIMD_CACHE),
576 .num_cu_shared = 2,
577 },
578 {
579 /* Scalar L1 Data Cache per SQC */
580 .cache_size = 16,
581 .cache_level = 1,
582 .cache_line_size = 64,
583 .flags = (CRAT_CACHE_FLAGS_ENABLED |
584 CRAT_CACHE_FLAGS_DATA_CACHE |
585 CRAT_CACHE_FLAGS_SIMD_CACHE),
586 .num_cu_shared = 2,
587 },
588 {
589 /* GL1 Data Cache per SA */
590 .cache_size = 128,
591 .cache_level = 1,
592 .cache_line_size = 128,
593 .flags = (CRAT_CACHE_FLAGS_ENABLED |
594 CRAT_CACHE_FLAGS_DATA_CACHE |
595 CRAT_CACHE_FLAGS_SIMD_CACHE),
596 .num_cu_shared = 10,
597 },
598 {
599 /* L2 Data Cache per GPU (Total Tex Cache) */
600 .cache_size = 4096,
601 .cache_level = 2,
602 .cache_line_size = 128,
603 .flags = (CRAT_CACHE_FLAGS_ENABLED |
604 CRAT_CACHE_FLAGS_DATA_CACHE |
605 CRAT_CACHE_FLAGS_SIMD_CACHE),
606 .num_cu_shared = 10,
607 },
608 {
609 /* L3 Data Cache per GPU */
610 .cache_size = 128*1024,
611 .cache_level = 3,
612 .cache_line_size = 64,
613 .flags = (CRAT_CACHE_FLAGS_ENABLED |
614 CRAT_CACHE_FLAGS_DATA_CACHE |
615 CRAT_CACHE_FLAGS_SIMD_CACHE),
616 .num_cu_shared = 10,
617 },
618};
619
620static struct kfd_gpu_cache_info navy_flounder_cache_info[] = {
621 {
622 /* TCP L1 Cache per CU */
623 .cache_size = 16,
624 .cache_level = 1,
625 .cache_line_size = 128,
626 .flags = (CRAT_CACHE_FLAGS_ENABLED |
627 CRAT_CACHE_FLAGS_DATA_CACHE |
628 CRAT_CACHE_FLAGS_SIMD_CACHE),
629 .num_cu_shared = 1,
630 },
631 {
632 /* Scalar L1 Instruction Cache per SQC */
633 .cache_size = 32,
634 .cache_level = 1,
635 .cache_line_size = 64,
636 .flags = (CRAT_CACHE_FLAGS_ENABLED |
637 CRAT_CACHE_FLAGS_INST_CACHE |
638 CRAT_CACHE_FLAGS_SIMD_CACHE),
639 .num_cu_shared = 2,
640 },
641 {
642 /* Scalar L1 Data Cache per SQC */
643 .cache_size = 16,
644 .cache_level = 1,
645 .cache_line_size = 64,
646 .flags = (CRAT_CACHE_FLAGS_ENABLED |
647 CRAT_CACHE_FLAGS_DATA_CACHE |
648 CRAT_CACHE_FLAGS_SIMD_CACHE),
649 .num_cu_shared = 2,
650 },
651 {
652 /* GL1 Data Cache per SA */
653 .cache_size = 128,
654 .cache_level = 1,
655 .cache_line_size = 128,
656 .flags = (CRAT_CACHE_FLAGS_ENABLED |
657 CRAT_CACHE_FLAGS_DATA_CACHE |
658 CRAT_CACHE_FLAGS_SIMD_CACHE),
659 .num_cu_shared = 10,
660 },
661 {
662 /* L2 Data Cache per GPU (Total Tex Cache) */
663 .cache_size = 3072,
664 .cache_level = 2,
665 .cache_line_size = 128,
666 .flags = (CRAT_CACHE_FLAGS_ENABLED |
667 CRAT_CACHE_FLAGS_DATA_CACHE |
668 CRAT_CACHE_FLAGS_SIMD_CACHE),
669 .num_cu_shared = 10,
670 },
671 {
672 /* L3 Data Cache per GPU */
673 .cache_size = 96*1024,
674 .cache_level = 3,
675 .cache_line_size = 64,
676 .flags = (CRAT_CACHE_FLAGS_ENABLED |
677 CRAT_CACHE_FLAGS_DATA_CACHE |
678 CRAT_CACHE_FLAGS_SIMD_CACHE),
679 .num_cu_shared = 10,
680 },
681};
682
683static struct kfd_gpu_cache_info dimgrey_cavefish_cache_info[] = {
684 {
685 /* TCP L1 Cache per CU */
686 .cache_size = 16,
687 .cache_level = 1,
688 .cache_line_size = 128,
689 .flags = (CRAT_CACHE_FLAGS_ENABLED |
690 CRAT_CACHE_FLAGS_DATA_CACHE |
691 CRAT_CACHE_FLAGS_SIMD_CACHE),
692 .num_cu_shared = 1,
693 },
694 {
695 /* Scalar L1 Instruction Cache per SQC */
696 .cache_size = 32,
697 .cache_level = 1,
698 .cache_line_size = 64,
699 .flags = (CRAT_CACHE_FLAGS_ENABLED |
700 CRAT_CACHE_FLAGS_INST_CACHE |
701 CRAT_CACHE_FLAGS_SIMD_CACHE),
702 .num_cu_shared = 2,
703 },
704 {
705 /* Scalar L1 Data Cache per SQC */
706 .cache_size = 16,
707 .cache_level = 1,
708 .cache_line_size = 64,
709 .flags = (CRAT_CACHE_FLAGS_ENABLED |
710 CRAT_CACHE_FLAGS_DATA_CACHE |
711 CRAT_CACHE_FLAGS_SIMD_CACHE),
712 .num_cu_shared = 2,
713 },
714 {
715 /* GL1 Data Cache per SA */
716 .cache_size = 128,
717 .cache_level = 1,
718 .cache_line_size = 128,
719 .flags = (CRAT_CACHE_FLAGS_ENABLED |
720 CRAT_CACHE_FLAGS_DATA_CACHE |
721 CRAT_CACHE_FLAGS_SIMD_CACHE),
722 .num_cu_shared = 8,
723 },
724 {
725 /* L2 Data Cache per GPU (Total Tex Cache) */
726 .cache_size = 2048,
727 .cache_level = 2,
728 .cache_line_size = 128,
729 .flags = (CRAT_CACHE_FLAGS_ENABLED |
730 CRAT_CACHE_FLAGS_DATA_CACHE |
731 CRAT_CACHE_FLAGS_SIMD_CACHE),
732 .num_cu_shared = 8,
733 },
734 {
735 /* L3 Data Cache per GPU */
736 .cache_size = 32*1024,
737 .cache_level = 3,
738 .cache_line_size = 64,
739 .flags = (CRAT_CACHE_FLAGS_ENABLED |
740 CRAT_CACHE_FLAGS_DATA_CACHE |
741 CRAT_CACHE_FLAGS_SIMD_CACHE),
742 .num_cu_shared = 8,
743 },
744};
745
746static struct kfd_gpu_cache_info beige_goby_cache_info[] = {
747 {
748 /* TCP L1 Cache per CU */
749 .cache_size = 16,
750 .cache_level = 1,
751 .cache_line_size = 128,
752 .flags = (CRAT_CACHE_FLAGS_ENABLED |
753 CRAT_CACHE_FLAGS_DATA_CACHE |
754 CRAT_CACHE_FLAGS_SIMD_CACHE),
755 .num_cu_shared = 1,
756 },
757 {
758 /* Scalar L1 Instruction Cache per SQC */
759 .cache_size = 32,
760 .cache_level = 1,
761 .cache_line_size = 64,
762 .flags = (CRAT_CACHE_FLAGS_ENABLED |
763 CRAT_CACHE_FLAGS_INST_CACHE |
764 CRAT_CACHE_FLAGS_SIMD_CACHE),
765 .num_cu_shared = 2,
766 },
767 {
768 /* Scalar L1 Data Cache per SQC */
769 .cache_size = 16,
770 .cache_level = 1,
771 .cache_line_size = 64,
772 .flags = (CRAT_CACHE_FLAGS_ENABLED |
773 CRAT_CACHE_FLAGS_DATA_CACHE |
774 CRAT_CACHE_FLAGS_SIMD_CACHE),
775 .num_cu_shared = 2,
776 },
777 {
778 /* GL1 Data Cache per SA */
779 .cache_size = 128,
780 .cache_level = 1,
781 .cache_line_size = 128,
782 .flags = (CRAT_CACHE_FLAGS_ENABLED |
783 CRAT_CACHE_FLAGS_DATA_CACHE |
784 CRAT_CACHE_FLAGS_SIMD_CACHE),
785 .num_cu_shared = 8,
786 },
787 {
788 /* L2 Data Cache per GPU (Total Tex Cache) */
789 .cache_size = 1024,
790 .cache_level = 2,
791 .cache_line_size = 128,
792 .flags = (CRAT_CACHE_FLAGS_ENABLED |
793 CRAT_CACHE_FLAGS_DATA_CACHE |
794 CRAT_CACHE_FLAGS_SIMD_CACHE),
795 .num_cu_shared = 8,
796 },
797 {
798 /* L3 Data Cache per GPU */
799 .cache_size = 16*1024,
800 .cache_level = 3,
801 .cache_line_size = 64,
802 .flags = (CRAT_CACHE_FLAGS_ENABLED |
803 CRAT_CACHE_FLAGS_DATA_CACHE |
804 CRAT_CACHE_FLAGS_SIMD_CACHE),
805 .num_cu_shared = 8,
806 },
807};
808
809static struct kfd_gpu_cache_info yellow_carp_cache_info[] = {
810 {
811 /* TCP L1 Cache per CU */
812 .cache_size = 16,
813 .cache_level = 1,
814 .cache_line_size = 128,
815 .flags = (CRAT_CACHE_FLAGS_ENABLED |
816 CRAT_CACHE_FLAGS_DATA_CACHE |
817 CRAT_CACHE_FLAGS_SIMD_CACHE),
818 .num_cu_shared = 1,
819 },
820 {
821 /* Scalar L1 Instruction Cache per SQC */
822 .cache_size = 32,
823 .cache_level = 1,
824 .cache_line_size = 64,
825 .flags = (CRAT_CACHE_FLAGS_ENABLED |
826 CRAT_CACHE_FLAGS_INST_CACHE |
827 CRAT_CACHE_FLAGS_SIMD_CACHE),
828 .num_cu_shared = 2,
829 },
830 {
831 /* Scalar L1 Data Cache per SQC */
832 .cache_size = 16,
833 .cache_level = 1,
834 .cache_line_size = 64,
835 .flags = (CRAT_CACHE_FLAGS_ENABLED |
836 CRAT_CACHE_FLAGS_DATA_CACHE |
837 CRAT_CACHE_FLAGS_SIMD_CACHE),
838 .num_cu_shared = 2,
839 },
840 {
841 /* GL1 Data Cache per SA */
842 .cache_size = 128,
843 .cache_level = 1,
844 .cache_line_size = 128,
845 .flags = (CRAT_CACHE_FLAGS_ENABLED |
846 CRAT_CACHE_FLAGS_DATA_CACHE |
847 CRAT_CACHE_FLAGS_SIMD_CACHE),
848 .num_cu_shared = 6,
849 },
850 {
851 /* L2 Data Cache per GPU (Total Tex Cache) */
852 .cache_size = 2048,
853 .cache_level = 2,
854 .cache_line_size = 128,
855 .flags = (CRAT_CACHE_FLAGS_ENABLED |
856 CRAT_CACHE_FLAGS_DATA_CACHE |
857 CRAT_CACHE_FLAGS_SIMD_CACHE),
858 .num_cu_shared = 6,
859 },
860};
861
862static struct kfd_gpu_cache_info gfx1037_cache_info[] = {
863 {
864 /* TCP L1 Cache per CU */
865 .cache_size = 16,
866 .cache_level = 1,
867 .cache_line_size = 128,
868 .flags = (CRAT_CACHE_FLAGS_ENABLED |
869 CRAT_CACHE_FLAGS_DATA_CACHE |
870 CRAT_CACHE_FLAGS_SIMD_CACHE),
871 .num_cu_shared = 1,
872 },
873 {
874 /* Scalar L1 Instruction Cache per SQC */
875 .cache_size = 32,
876 .cache_level = 1,
877 .cache_line_size = 64,
878 .flags = (CRAT_CACHE_FLAGS_ENABLED |
879 CRAT_CACHE_FLAGS_INST_CACHE |
880 CRAT_CACHE_FLAGS_SIMD_CACHE),
881 .num_cu_shared = 2,
882 },
883 {
884 /* Scalar L1 Data Cache per SQC */
885 .cache_size = 16,
886 .cache_level = 1,
887 .cache_line_size = 64,
888 .flags = (CRAT_CACHE_FLAGS_ENABLED |
889 CRAT_CACHE_FLAGS_DATA_CACHE |
890 CRAT_CACHE_FLAGS_SIMD_CACHE),
891 .num_cu_shared = 2,
892 },
893 {
894 /* GL1 Data Cache per SA */
895 .cache_size = 128,
896 .cache_level = 1,
897 .cache_line_size = 128,
898 .flags = (CRAT_CACHE_FLAGS_ENABLED |
899 CRAT_CACHE_FLAGS_DATA_CACHE |
900 CRAT_CACHE_FLAGS_SIMD_CACHE),
901 .num_cu_shared = 2,
902 },
903 {
904 /* L2 Data Cache per GPU (Total Tex Cache) */
905 .cache_size = 256,
906 .cache_level = 2,
907 .cache_line_size = 128,
908 .flags = (CRAT_CACHE_FLAGS_ENABLED |
909 CRAT_CACHE_FLAGS_DATA_CACHE |
910 CRAT_CACHE_FLAGS_SIMD_CACHE),
911 .num_cu_shared = 2,
912 },
913};
914
915static struct kfd_gpu_cache_info gc_10_3_6_cache_info[] = {
916 {
917 /* TCP L1 Cache per CU */
918 .cache_size = 16,
919 .cache_level = 1,
920 .cache_line_size = 128,
921 .flags = (CRAT_CACHE_FLAGS_ENABLED |
922 CRAT_CACHE_FLAGS_DATA_CACHE |
923 CRAT_CACHE_FLAGS_SIMD_CACHE),
924 .num_cu_shared = 1,
925 },
926 {
927 /* Scalar L1 Instruction Cache per SQC */
928 .cache_size = 32,
929 .cache_level = 1,
930 .cache_line_size = 64,
931 .flags = (CRAT_CACHE_FLAGS_ENABLED |
932 CRAT_CACHE_FLAGS_INST_CACHE |
933 CRAT_CACHE_FLAGS_SIMD_CACHE),
934 .num_cu_shared = 2,
935 },
936 {
937 /* Scalar L1 Data Cache per SQC */
938 .cache_size = 16,
939 .cache_level = 1,
940 .cache_line_size = 64,
941 .flags = (CRAT_CACHE_FLAGS_ENABLED |
942 CRAT_CACHE_FLAGS_DATA_CACHE |
943 CRAT_CACHE_FLAGS_SIMD_CACHE),
944 .num_cu_shared = 2,
945 },
946 {
947 /* GL1 Data Cache per SA */
948 .cache_size = 128,
949 .cache_level = 1,
950 .cache_line_size = 128,
951 .flags = (CRAT_CACHE_FLAGS_ENABLED |
952 CRAT_CACHE_FLAGS_DATA_CACHE |
953 CRAT_CACHE_FLAGS_SIMD_CACHE),
954 .num_cu_shared = 2,
955 },
956 {
957 /* L2 Data Cache per GPU (Total Tex Cache) */
958 .cache_size = 256,
959 .cache_level = 2,
960 .cache_line_size = 128,
961 .flags = (CRAT_CACHE_FLAGS_ENABLED |
962 CRAT_CACHE_FLAGS_DATA_CACHE |
963 CRAT_CACHE_FLAGS_SIMD_CACHE),
964 .num_cu_shared = 2,
965 },
966};
967
968static struct kfd_gpu_cache_info dummy_cache_info[] = {
969 {
970 /* TCP L1 Cache per CU */
971 .cache_size = 16,
972 .cache_level = 1,
973 .cache_line_size = 64,
974 .flags = (CRAT_CACHE_FLAGS_ENABLED |
975 CRAT_CACHE_FLAGS_DATA_CACHE |
976 CRAT_CACHE_FLAGS_SIMD_CACHE),
977 .num_cu_shared = 1,
978 },
979 {
980 /* Scalar L1 Instruction Cache per SQC */
981 .cache_size = 32,
982 .cache_level = 1,
983 .cache_line_size = 64,
984 .flags = (CRAT_CACHE_FLAGS_ENABLED |
985 CRAT_CACHE_FLAGS_INST_CACHE |
986 CRAT_CACHE_FLAGS_SIMD_CACHE),
987 .num_cu_shared = 2,
988 },
989 {
990 /* Scalar L1 Data Cache per SQC */
991 .cache_size = 16,
992 .cache_level = 1,
993 .cache_line_size = 64,
994 .flags = (CRAT_CACHE_FLAGS_ENABLED |
995 CRAT_CACHE_FLAGS_DATA_CACHE |
996 CRAT_CACHE_FLAGS_SIMD_CACHE),
997 .num_cu_shared = 2,
998 },
999 {
1000 /* GL1 Data Cache per SA */
1001 .cache_size = 128,
1002 .cache_level = 1,
1003 .cache_line_size = 64,
1004 .flags = (CRAT_CACHE_FLAGS_ENABLED |
1005 CRAT_CACHE_FLAGS_DATA_CACHE |
1006 CRAT_CACHE_FLAGS_SIMD_CACHE),
1007 .num_cu_shared = 6,
1008 },
1009 {
1010 /* L2 Data Cache per GPU (Total Tex Cache) */
1011 .cache_size = 2048,
1012 .cache_level = 2,
1013 .cache_line_size = 64,
1014 .flags = (CRAT_CACHE_FLAGS_ENABLED |
1015 CRAT_CACHE_FLAGS_DATA_CACHE |
1016 CRAT_CACHE_FLAGS_SIMD_CACHE),
1017 .num_cu_shared = 6,
1018 },
1019};
1020
1021static void kfd_populated_cu_info_cpu(struct kfd_topology_device *dev,
1022 struct crat_subtype_computeunit *cu)
1023{
1024 dev->node_props.cpu_cores_count = cu->num_cpu_cores;
1025 dev->node_props.cpu_core_id_base = cu->processor_id_low;
1026 if (cu->hsa_capability & CRAT_CU_FLAGS_IOMMU_PRESENT)
1027 dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
1028
1029 pr_debug("CU CPU: cores=%d id_base=%d\n", cu->num_cpu_cores,
1030 cu->processor_id_low);
1031}
1032
1033static void kfd_populated_cu_info_gpu(struct kfd_topology_device *dev,
1034 struct crat_subtype_computeunit *cu)
1035{
1036 dev->node_props.simd_id_base = cu->processor_id_low;
1037 dev->node_props.simd_count = cu->num_simd_cores;
1038 dev->node_props.lds_size_in_kb = cu->lds_size_in_kb;
1039 dev->node_props.max_waves_per_simd = cu->max_waves_simd;
1040 dev->node_props.wave_front_size = cu->wave_front_size;
1041 dev->node_props.array_count = cu->array_count;
1042 dev->node_props.cu_per_simd_array = cu->num_cu_per_array;
1043 dev->node_props.simd_per_cu = cu->num_simd_per_cu;
1044 dev->node_props.max_slots_scratch_cu = cu->max_slots_scatch_cu;
1045 if (cu->hsa_capability & CRAT_CU_FLAGS_HOT_PLUGGABLE)
1046 dev->node_props.capability |= HSA_CAP_HOT_PLUGGABLE;
1047 pr_debug("CU GPU: id_base=%d\n", cu->processor_id_low);
1048}
1049
1050/* kfd_parse_subtype_cu - parse compute unit subtypes and attach it to correct
1051 * topology device present in the device_list
1052 */
1053static int kfd_parse_subtype_cu(struct crat_subtype_computeunit *cu,
1054 struct list_head *device_list)
1055{
1056 struct kfd_topology_device *dev;
1057
1058 pr_debug("Found CU entry in CRAT table with proximity_domain=%d caps=%x\n",
1059 cu->proximity_domain, cu->hsa_capability);
1060 list_for_each_entry(dev, device_list, list) {
1061 if (cu->proximity_domain == dev->proximity_domain) {
1062 if (cu->flags & CRAT_CU_FLAGS_CPU_PRESENT)
1063 kfd_populated_cu_info_cpu(dev, cu);
1064
1065 if (cu->flags & CRAT_CU_FLAGS_GPU_PRESENT)
1066 kfd_populated_cu_info_gpu(dev, cu);
1067 break;
1068 }
1069 }
1070
1071 return 0;
1072}
1073
1074static struct kfd_mem_properties *
1075find_subtype_mem(uint32_t heap_type, uint32_t flags, uint32_t width,
1076 struct kfd_topology_device *dev)
1077{
1078 struct kfd_mem_properties *props;
1079
1080 list_for_each_entry(props, &dev->mem_props, list) {
1081 if (props->heap_type == heap_type
1082 && props->flags == flags
1083 && props->width == width)
1084 return props;
1085 }
1086
1087 return NULL;
1088}
1089/* kfd_parse_subtype_mem - parse memory subtypes and attach it to correct
1090 * topology device present in the device_list
1091 */
1092static int kfd_parse_subtype_mem(struct crat_subtype_memory *mem,
1093 struct list_head *device_list)
1094{
1095 struct kfd_mem_properties *props;
1096 struct kfd_topology_device *dev;
1097 uint32_t heap_type;
1098 uint64_t size_in_bytes;
1099 uint32_t flags = 0;
1100 uint32_t width;
1101
1102 pr_debug("Found memory entry in CRAT table with proximity_domain=%d\n",
1103 mem->proximity_domain);
1104 list_for_each_entry(dev, device_list, list) {
1105 if (mem->proximity_domain == dev->proximity_domain) {
1106 /* We're on GPU node */
1107 if (dev->node_props.cpu_cores_count == 0) {
1108 /* APU */
1109 if (mem->visibility_type == 0)
1110 heap_type =
1111 HSA_MEM_HEAP_TYPE_FB_PRIVATE;
1112 /* dGPU */
1113 else
1114 heap_type = mem->visibility_type;
1115 } else
1116 heap_type = HSA_MEM_HEAP_TYPE_SYSTEM;
1117
1118 if (mem->flags & CRAT_MEM_FLAGS_HOT_PLUGGABLE)
1119 flags |= HSA_MEM_FLAGS_HOT_PLUGGABLE;
1120 if (mem->flags & CRAT_MEM_FLAGS_NON_VOLATILE)
1121 flags |= HSA_MEM_FLAGS_NON_VOLATILE;
1122
1123 size_in_bytes =
1124 ((uint64_t)mem->length_high << 32) +
1125 mem->length_low;
1126 width = mem->width;
1127
1128 /* Multiple banks of the same type are aggregated into
1129 * one. User mode doesn't care about multiple physical
1130 * memory segments. It's managed as a single virtual
1131 * heap for user mode.
1132 */
1133 props = find_subtype_mem(heap_type, flags, width, dev);
1134 if (props) {
1135 props->size_in_bytes += size_in_bytes;
1136 break;
1137 }
1138
1139 props = kfd_alloc_struct(props);
1140 if (!props)
1141 return -ENOMEM;
1142
1143 props->heap_type = heap_type;
1144 props->flags = flags;
1145 props->size_in_bytes = size_in_bytes;
1146 props->width = width;
1147
1148 dev->node_props.mem_banks_count++;
1149 list_add_tail(&props->list, &dev->mem_props);
1150
1151 break;
1152 }
1153 }
1154
1155 return 0;
1156}
1157
1158/* kfd_parse_subtype_cache - parse cache subtypes and attach it to correct
1159 * topology device present in the device_list
1160 */
1161static int kfd_parse_subtype_cache(struct crat_subtype_cache *cache,
1162 struct list_head *device_list)
1163{
1164 struct kfd_cache_properties *props;
1165 struct kfd_topology_device *dev;
1166 uint32_t id;
1167 uint32_t total_num_of_cu;
1168
1169 id = cache->processor_id_low;
1170
1171 pr_debug("Found cache entry in CRAT table with processor_id=%d\n", id);
1172 list_for_each_entry(dev, device_list, list) {
1173 total_num_of_cu = (dev->node_props.array_count *
1174 dev->node_props.cu_per_simd_array);
1175
1176 /* Cache infomration in CRAT doesn't have proximity_domain
1177 * information as it is associated with a CPU core or GPU
1178 * Compute Unit. So map the cache using CPU core Id or SIMD
1179 * (GPU) ID.
1180 * TODO: This works because currently we can safely assume that
1181 * Compute Units are parsed before caches are parsed. In
1182 * future, remove this dependency
1183 */
1184 if ((id >= dev->node_props.cpu_core_id_base &&
1185 id <= dev->node_props.cpu_core_id_base +
1186 dev->node_props.cpu_cores_count) ||
1187 (id >= dev->node_props.simd_id_base &&
1188 id < dev->node_props.simd_id_base +
1189 total_num_of_cu)) {
1190 props = kfd_alloc_struct(props);
1191 if (!props)
1192 return -ENOMEM;
1193
1194 props->processor_id_low = id;
1195 props->cache_level = cache->cache_level;
1196 props->cache_size = cache->cache_size;
1197 props->cacheline_size = cache->cache_line_size;
1198 props->cachelines_per_tag = cache->lines_per_tag;
1199 props->cache_assoc = cache->associativity;
1200 props->cache_latency = cache->cache_latency;
1201
1202 memcpy(props->sibling_map, cache->sibling_map,
1203 CRAT_SIBLINGMAP_SIZE);
1204
1205 /* set the sibling_map_size as 32 for CRAT from ACPI */
1206 props->sibling_map_size = CRAT_SIBLINGMAP_SIZE;
1207
1208 if (cache->flags & CRAT_CACHE_FLAGS_DATA_CACHE)
1209 props->cache_type |= HSA_CACHE_TYPE_DATA;
1210 if (cache->flags & CRAT_CACHE_FLAGS_INST_CACHE)
1211 props->cache_type |= HSA_CACHE_TYPE_INSTRUCTION;
1212 if (cache->flags & CRAT_CACHE_FLAGS_CPU_CACHE)
1213 props->cache_type |= HSA_CACHE_TYPE_CPU;
1214 if (cache->flags & CRAT_CACHE_FLAGS_SIMD_CACHE)
1215 props->cache_type |= HSA_CACHE_TYPE_HSACU;
1216
1217 dev->node_props.caches_count++;
1218 list_add_tail(&props->list, &dev->cache_props);
1219
1220 break;
1221 }
1222 }
1223
1224 return 0;
1225}
1226
1227/* kfd_parse_subtype_iolink - parse iolink subtypes and attach it to correct
1228 * topology device present in the device_list
1229 */
1230static int kfd_parse_subtype_iolink(struct crat_subtype_iolink *iolink,
1231 struct list_head *device_list)
1232{
1233 struct kfd_iolink_properties *props = NULL, *props2;
1234 struct kfd_topology_device *dev, *to_dev;
1235 uint32_t id_from;
1236 uint32_t id_to;
1237
1238 id_from = iolink->proximity_domain_from;
1239 id_to = iolink->proximity_domain_to;
1240
1241 pr_debug("Found IO link entry in CRAT table with id_from=%d, id_to %d\n",
1242 id_from, id_to);
1243 list_for_each_entry(dev, device_list, list) {
1244 if (id_from == dev->proximity_domain) {
1245 props = kfd_alloc_struct(props);
1246 if (!props)
1247 return -ENOMEM;
1248
1249 props->node_from = id_from;
1250 props->node_to = id_to;
1251 props->ver_maj = iolink->version_major;
1252 props->ver_min = iolink->version_minor;
1253 props->iolink_type = iolink->io_interface_type;
1254
1255 if (props->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS)
1256 props->weight = 20;
1257 else if (props->iolink_type == CRAT_IOLINK_TYPE_XGMI)
1258 props->weight = iolink->weight_xgmi;
1259 else
1260 props->weight = node_distance(id_from, id_to);
1261
1262 props->min_latency = iolink->minimum_latency;
1263 props->max_latency = iolink->maximum_latency;
1264 props->min_bandwidth = iolink->minimum_bandwidth_mbs;
1265 props->max_bandwidth = iolink->maximum_bandwidth_mbs;
1266 props->rec_transfer_size =
1267 iolink->recommended_transfer_size;
1268
1269 dev->node_props.io_links_count++;
1270 list_add_tail(&props->list, &dev->io_link_props);
1271 break;
1272 }
1273 }
1274
1275 /* CPU topology is created before GPUs are detected, so CPU->GPU
1276 * links are not built at that time. If a PCIe type is discovered, it
1277 * means a GPU is detected and we are adding GPU->CPU to the topology.
1278 * At this time, also add the corresponded CPU->GPU link if GPU
1279 * is large bar.
1280 * For xGMI, we only added the link with one direction in the crat
1281 * table, add corresponded reversed direction link now.
1282 */
1283 if (props && (iolink->flags & CRAT_IOLINK_FLAGS_BI_DIRECTIONAL)) {
1284 to_dev = kfd_topology_device_by_proximity_domain_no_lock(id_to);
1285 if (!to_dev)
1286 return -ENODEV;
1287 /* same everything but the other direction */
1288 props2 = kmemdup(props, sizeof(*props2), GFP_KERNEL);
1289 if (!props2)
1290 return -ENOMEM;
1291
1292 props2->node_from = id_to;
1293 props2->node_to = id_from;
1294 props2->kobj = NULL;
1295 to_dev->node_props.io_links_count++;
1296 list_add_tail(&props2->list, &to_dev->io_link_props);
1297 }
1298
1299 return 0;
1300}
1301
1302/* kfd_parse_subtype - parse subtypes and attach it to correct topology device
1303 * present in the device_list
1304 * @sub_type_hdr - subtype section of crat_image
1305 * @device_list - list of topology devices present in this crat_image
1306 */
1307static int kfd_parse_subtype(struct crat_subtype_generic *sub_type_hdr,
1308 struct list_head *device_list)
1309{
1310 struct crat_subtype_computeunit *cu;
1311 struct crat_subtype_memory *mem;
1312 struct crat_subtype_cache *cache;
1313 struct crat_subtype_iolink *iolink;
1314 int ret = 0;
1315
1316 switch (sub_type_hdr->type) {
1317 case CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY:
1318 cu = (struct crat_subtype_computeunit *)sub_type_hdr;
1319 ret = kfd_parse_subtype_cu(cu, device_list);
1320 break;
1321 case CRAT_SUBTYPE_MEMORY_AFFINITY:
1322 mem = (struct crat_subtype_memory *)sub_type_hdr;
1323 ret = kfd_parse_subtype_mem(mem, device_list);
1324 break;
1325 case CRAT_SUBTYPE_CACHE_AFFINITY:
1326 cache = (struct crat_subtype_cache *)sub_type_hdr;
1327 ret = kfd_parse_subtype_cache(cache, device_list);
1328 break;
1329 case CRAT_SUBTYPE_TLB_AFFINITY:
1330 /*
1331 * For now, nothing to do here
1332 */
1333 pr_debug("Found TLB entry in CRAT table (not processing)\n");
1334 break;
1335 case CRAT_SUBTYPE_CCOMPUTE_AFFINITY:
1336 /*
1337 * For now, nothing to do here
1338 */
1339 pr_debug("Found CCOMPUTE entry in CRAT table (not processing)\n");
1340 break;
1341 case CRAT_SUBTYPE_IOLINK_AFFINITY:
1342 iolink = (struct crat_subtype_iolink *)sub_type_hdr;
1343 ret = kfd_parse_subtype_iolink(iolink, device_list);
1344 break;
1345 default:
1346 pr_warn("Unknown subtype %d in CRAT\n",
1347 sub_type_hdr->type);
1348 }
1349
1350 return ret;
1351}
1352
1353/* kfd_parse_crat_table - parse CRAT table. For each node present in CRAT
1354 * create a kfd_topology_device and add in to device_list. Also parse
1355 * CRAT subtypes and attach it to appropriate kfd_topology_device
1356 * @crat_image - input image containing CRAT
1357 * @device_list - [OUT] list of kfd_topology_device generated after
1358 * parsing crat_image
1359 * @proximity_domain - Proximity domain of the first device in the table
1360 *
1361 * Return - 0 if successful else -ve value
1362 */
1363int kfd_parse_crat_table(void *crat_image, struct list_head *device_list,
1364 uint32_t proximity_domain)
1365{
1366 struct kfd_topology_device *top_dev = NULL;
1367 struct crat_subtype_generic *sub_type_hdr;
1368 uint16_t node_id;
1369 int ret = 0;
1370 struct crat_header *crat_table = (struct crat_header *)crat_image;
1371 uint16_t num_nodes;
1372 uint32_t image_len;
1373
1374 if (!crat_image)
1375 return -EINVAL;
1376
1377 if (!list_empty(device_list)) {
1378 pr_warn("Error device list should be empty\n");
1379 return -EINVAL;
1380 }
1381
1382 num_nodes = crat_table->num_domains;
1383 image_len = crat_table->length;
1384
1385 pr_debug("Parsing CRAT table with %d nodes\n", num_nodes);
1386
1387 for (node_id = 0; node_id < num_nodes; node_id++) {
1388 top_dev = kfd_create_topology_device(device_list);
1389 if (!top_dev)
1390 break;
1391 top_dev->proximity_domain = proximity_domain++;
1392 }
1393
1394 if (!top_dev) {
1395 ret = -ENOMEM;
1396 goto err;
1397 }
1398
1399 memcpy(top_dev->oem_id, crat_table->oem_id, CRAT_OEMID_LENGTH);
1400 memcpy(top_dev->oem_table_id, crat_table->oem_table_id,
1401 CRAT_OEMTABLEID_LENGTH);
1402 top_dev->oem_revision = crat_table->oem_revision;
1403
1404 sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1);
1405 while ((char *)sub_type_hdr + sizeof(struct crat_subtype_generic) <
1406 ((char *)crat_image) + image_len) {
1407 if (sub_type_hdr->flags & CRAT_SUBTYPE_FLAGS_ENABLED) {
1408 ret = kfd_parse_subtype(sub_type_hdr, device_list);
1409 if (ret)
1410 break;
1411 }
1412
1413 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1414 sub_type_hdr->length);
1415 }
1416
1417err:
1418 if (ret)
1419 kfd_release_topology_device_list(device_list);
1420
1421 return ret;
1422}
1423
1424
1425static int kfd_fill_gpu_cache_info_from_gfx_config(struct kfd_dev *kdev,
1426 bool cache_line_size_missing,
1427 struct kfd_gpu_cache_info *pcache_info)
1428{
1429 struct amdgpu_device *adev = kdev->adev;
1430 int i = 0;
1431
1432 /* TCP L1 Cache per CU */
1433 if (adev->gfx.config.gc_tcp_l1_size) {
1434 pcache_info[i].cache_size = adev->gfx.config.gc_tcp_l1_size;
1435 pcache_info[i].cache_level = 1;
1436 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1437 CRAT_CACHE_FLAGS_DATA_CACHE |
1438 CRAT_CACHE_FLAGS_SIMD_CACHE);
1439 pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_tcp_per_wpg / 2;
1440 pcache_info[i].cache_line_size = adev->gfx.config.gc_tcp_cache_line_size;
1441 if (cache_line_size_missing && !pcache_info[i].cache_line_size)
1442 pcache_info[i].cache_line_size = 128;
1443 i++;
1444 }
1445 /* Scalar L1 Instruction Cache per SQC */
1446 if (adev->gfx.config.gc_l1_instruction_cache_size_per_sqc) {
1447 pcache_info[i].cache_size =
1448 adev->gfx.config.gc_l1_instruction_cache_size_per_sqc;
1449 pcache_info[i].cache_level = 1;
1450 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1451 CRAT_CACHE_FLAGS_INST_CACHE |
1452 CRAT_CACHE_FLAGS_SIMD_CACHE);
1453 pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_sqc_per_wgp * 2;
1454 pcache_info[i].cache_line_size = adev->gfx.config.gc_instruction_cache_line_size;
1455 if (cache_line_size_missing && !pcache_info[i].cache_line_size)
1456 pcache_info[i].cache_line_size = 128;
1457 i++;
1458 }
1459 /* Scalar L1 Data Cache per SQC */
1460 if (adev->gfx.config.gc_l1_data_cache_size_per_sqc) {
1461 pcache_info[i].cache_size = adev->gfx.config.gc_l1_data_cache_size_per_sqc;
1462 pcache_info[i].cache_level = 1;
1463 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1464 CRAT_CACHE_FLAGS_DATA_CACHE |
1465 CRAT_CACHE_FLAGS_SIMD_CACHE);
1466 pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_sqc_per_wgp * 2;
1467 pcache_info[i].cache_line_size = adev->gfx.config.gc_scalar_data_cache_line_size;
1468 if (cache_line_size_missing && !pcache_info[i].cache_line_size)
1469 pcache_info[i].cache_line_size = 64;
1470 i++;
1471 }
1472 /* GL1 Data Cache per SA */
1473 if (adev->gfx.config.gc_gl1c_per_sa &&
1474 adev->gfx.config.gc_gl1c_size_per_instance) {
1475 pcache_info[i].cache_size = adev->gfx.config.gc_gl1c_per_sa *
1476 adev->gfx.config.gc_gl1c_size_per_instance;
1477 pcache_info[i].cache_level = 1;
1478 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1479 CRAT_CACHE_FLAGS_DATA_CACHE |
1480 CRAT_CACHE_FLAGS_SIMD_CACHE);
1481 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
1482 if (cache_line_size_missing)
1483 pcache_info[i].cache_line_size = 128;
1484 i++;
1485 }
1486 /* L2 Data Cache per GPU (Total Tex Cache) */
1487 if (adev->gfx.config.gc_gl2c_per_gpu) {
1488 pcache_info[i].cache_size = adev->gfx.config.gc_gl2c_per_gpu;
1489 pcache_info[i].cache_level = 2;
1490 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1491 CRAT_CACHE_FLAGS_DATA_CACHE |
1492 CRAT_CACHE_FLAGS_SIMD_CACHE);
1493 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
1494 pcache_info[i].cache_line_size = adev->gfx.config.gc_tcc_cache_line_size;
1495 if (cache_line_size_missing && !pcache_info[i].cache_line_size)
1496 pcache_info[i].cache_line_size = 128;
1497 i++;
1498 }
1499 /* L3 Data Cache per GPU */
1500 if (adev->gmc.mall_size) {
1501 pcache_info[i].cache_size = adev->gmc.mall_size / 1024;
1502 pcache_info[i].cache_level = 3;
1503 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1504 CRAT_CACHE_FLAGS_DATA_CACHE |
1505 CRAT_CACHE_FLAGS_SIMD_CACHE);
1506 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
1507 pcache_info[i].cache_line_size = 64;
1508 i++;
1509 }
1510 return i;
1511}
1512
1513static int kfd_fill_gpu_cache_info_from_gfx_config_v2(struct kfd_dev *kdev,
1514 struct kfd_gpu_cache_info *pcache_info)
1515{
1516 struct amdgpu_device *adev = kdev->adev;
1517 int i = 0;
1518
1519 /* TCP L1 Cache per CU */
1520 if (adev->gfx.config.gc_tcp_size_per_cu) {
1521 pcache_info[i].cache_size = adev->gfx.config.gc_tcp_size_per_cu;
1522 pcache_info[i].cache_level = 1;
1523 /* Cacheline size not available in IP discovery for gc943,gc944 */
1524 pcache_info[i].cache_line_size = 128;
1525 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1526 CRAT_CACHE_FLAGS_DATA_CACHE |
1527 CRAT_CACHE_FLAGS_SIMD_CACHE);
1528 pcache_info[i].num_cu_shared = 1;
1529 i++;
1530 }
1531 /* Scalar L1 Instruction Cache per SQC */
1532 if (adev->gfx.config.gc_l1_instruction_cache_size_per_sqc) {
1533 pcache_info[i].cache_size =
1534 adev->gfx.config.gc_l1_instruction_cache_size_per_sqc;
1535 pcache_info[i].cache_level = 1;
1536 pcache_info[i].cache_line_size = 64;
1537 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1538 CRAT_CACHE_FLAGS_INST_CACHE |
1539 CRAT_CACHE_FLAGS_SIMD_CACHE);
1540 pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_cu_per_sqc;
1541 i++;
1542 }
1543 /* Scalar L1 Data Cache per SQC */
1544 if (adev->gfx.config.gc_l1_data_cache_size_per_sqc) {
1545 pcache_info[i].cache_size = adev->gfx.config.gc_l1_data_cache_size_per_sqc;
1546 pcache_info[i].cache_level = 1;
1547 pcache_info[i].cache_line_size = 64;
1548 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1549 CRAT_CACHE_FLAGS_DATA_CACHE |
1550 CRAT_CACHE_FLAGS_SIMD_CACHE);
1551 pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_cu_per_sqc;
1552 i++;
1553 }
1554 /* L2 Data Cache per GPU (Total Tex Cache) */
1555 if (adev->gfx.config.gc_tcc_size) {
1556 pcache_info[i].cache_size = adev->gfx.config.gc_tcc_size;
1557 pcache_info[i].cache_level = 2;
1558 pcache_info[i].cache_line_size = 128;
1559 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1560 CRAT_CACHE_FLAGS_DATA_CACHE |
1561 CRAT_CACHE_FLAGS_SIMD_CACHE);
1562 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
1563 i++;
1564 }
1565 /* L3 Data Cache per GPU */
1566 if (adev->gmc.mall_size) {
1567 pcache_info[i].cache_size = adev->gmc.mall_size / 1024;
1568 pcache_info[i].cache_level = 3;
1569 pcache_info[i].cache_line_size = 64;
1570 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1571 CRAT_CACHE_FLAGS_DATA_CACHE |
1572 CRAT_CACHE_FLAGS_SIMD_CACHE);
1573 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
1574 i++;
1575 }
1576 return i;
1577}
1578
1579int kfd_get_gpu_cache_info(struct kfd_node *kdev, struct kfd_gpu_cache_info **pcache_info)
1580{
1581 int num_of_cache_types = 0;
1582 bool cache_line_size_missing = false;
1583
1584 switch (kdev->adev->asic_type) {
1585 case CHIP_KAVERI:
1586 *pcache_info = kaveri_cache_info;
1587 num_of_cache_types = ARRAY_SIZE(kaveri_cache_info);
1588 break;
1589 case CHIP_HAWAII:
1590 *pcache_info = hawaii_cache_info;
1591 num_of_cache_types = ARRAY_SIZE(hawaii_cache_info);
1592 break;
1593 case CHIP_CARRIZO:
1594 *pcache_info = carrizo_cache_info;
1595 num_of_cache_types = ARRAY_SIZE(carrizo_cache_info);
1596 break;
1597 case CHIP_TONGA:
1598 *pcache_info = tonga_cache_info;
1599 num_of_cache_types = ARRAY_SIZE(tonga_cache_info);
1600 break;
1601 case CHIP_FIJI:
1602 *pcache_info = fiji_cache_info;
1603 num_of_cache_types = ARRAY_SIZE(fiji_cache_info);
1604 break;
1605 case CHIP_POLARIS10:
1606 *pcache_info = polaris10_cache_info;
1607 num_of_cache_types = ARRAY_SIZE(polaris10_cache_info);
1608 break;
1609 case CHIP_POLARIS11:
1610 *pcache_info = polaris11_cache_info;
1611 num_of_cache_types = ARRAY_SIZE(polaris11_cache_info);
1612 break;
1613 case CHIP_POLARIS12:
1614 *pcache_info = polaris12_cache_info;
1615 num_of_cache_types = ARRAY_SIZE(polaris12_cache_info);
1616 break;
1617 case CHIP_VEGAM:
1618 *pcache_info = vegam_cache_info;
1619 num_of_cache_types = ARRAY_SIZE(vegam_cache_info);
1620 break;
1621 default:
1622 switch (KFD_GC_VERSION(kdev)) {
1623 case IP_VERSION(9, 0, 1):
1624 *pcache_info = vega10_cache_info;
1625 num_of_cache_types = ARRAY_SIZE(vega10_cache_info);
1626 break;
1627 case IP_VERSION(9, 2, 1):
1628 *pcache_info = vega12_cache_info;
1629 num_of_cache_types = ARRAY_SIZE(vega12_cache_info);
1630 break;
1631 case IP_VERSION(9, 4, 0):
1632 case IP_VERSION(9, 4, 1):
1633 *pcache_info = vega20_cache_info;
1634 num_of_cache_types = ARRAY_SIZE(vega20_cache_info);
1635 break;
1636 case IP_VERSION(9, 4, 2):
1637 *pcache_info = aldebaran_cache_info;
1638 num_of_cache_types = ARRAY_SIZE(aldebaran_cache_info);
1639 break;
1640 case IP_VERSION(9, 4, 3):
1641 case IP_VERSION(9, 4, 4):
1642 num_of_cache_types =
1643 kfd_fill_gpu_cache_info_from_gfx_config_v2(kdev->kfd,
1644 *pcache_info);
1645 break;
1646 case IP_VERSION(9, 1, 0):
1647 case IP_VERSION(9, 2, 2):
1648 *pcache_info = raven_cache_info;
1649 num_of_cache_types = ARRAY_SIZE(raven_cache_info);
1650 break;
1651 case IP_VERSION(9, 3, 0):
1652 *pcache_info = renoir_cache_info;
1653 num_of_cache_types = ARRAY_SIZE(renoir_cache_info);
1654 break;
1655 case IP_VERSION(10, 1, 10):
1656 case IP_VERSION(10, 1, 2):
1657 case IP_VERSION(10, 1, 3):
1658 case IP_VERSION(10, 1, 4):
1659 *pcache_info = navi10_cache_info;
1660 num_of_cache_types = ARRAY_SIZE(navi10_cache_info);
1661 break;
1662 case IP_VERSION(10, 1, 1):
1663 *pcache_info = navi14_cache_info;
1664 num_of_cache_types = ARRAY_SIZE(navi14_cache_info);
1665 break;
1666 case IP_VERSION(10, 3, 0):
1667 *pcache_info = sienna_cichlid_cache_info;
1668 num_of_cache_types = ARRAY_SIZE(sienna_cichlid_cache_info);
1669 break;
1670 case IP_VERSION(10, 3, 2):
1671 *pcache_info = navy_flounder_cache_info;
1672 num_of_cache_types = ARRAY_SIZE(navy_flounder_cache_info);
1673 break;
1674 case IP_VERSION(10, 3, 4):
1675 *pcache_info = dimgrey_cavefish_cache_info;
1676 num_of_cache_types = ARRAY_SIZE(dimgrey_cavefish_cache_info);
1677 break;
1678 case IP_VERSION(10, 3, 1):
1679 *pcache_info = vangogh_cache_info;
1680 num_of_cache_types = ARRAY_SIZE(vangogh_cache_info);
1681 break;
1682 case IP_VERSION(10, 3, 5):
1683 *pcache_info = beige_goby_cache_info;
1684 num_of_cache_types = ARRAY_SIZE(beige_goby_cache_info);
1685 break;
1686 case IP_VERSION(10, 3, 3):
1687 *pcache_info = yellow_carp_cache_info;
1688 num_of_cache_types = ARRAY_SIZE(yellow_carp_cache_info);
1689 break;
1690 case IP_VERSION(10, 3, 6):
1691 *pcache_info = gc_10_3_6_cache_info;
1692 num_of_cache_types = ARRAY_SIZE(gc_10_3_6_cache_info);
1693 break;
1694 case IP_VERSION(10, 3, 7):
1695 *pcache_info = gfx1037_cache_info;
1696 num_of_cache_types = ARRAY_SIZE(gfx1037_cache_info);
1697 break;
1698 case IP_VERSION(11, 0, 0):
1699 case IP_VERSION(11, 0, 1):
1700 case IP_VERSION(11, 0, 2):
1701 case IP_VERSION(11, 0, 3):
1702 case IP_VERSION(11, 0, 4):
1703 case IP_VERSION(11, 5, 0):
1704 case IP_VERSION(11, 5, 1):
1705 case IP_VERSION(11, 5, 2):
1706 /* Cacheline size not available in IP discovery for gc11.
1707 * kfd_fill_gpu_cache_info_from_gfx_config to hard code it
1708 */
1709 cache_line_size_missing = true;
1710 fallthrough;
1711 case IP_VERSION(12, 0, 0):
1712 case IP_VERSION(12, 0, 1):
1713 num_of_cache_types =
1714 kfd_fill_gpu_cache_info_from_gfx_config(kdev->kfd,
1715 cache_line_size_missing,
1716 *pcache_info);
1717 break;
1718 default:
1719 *pcache_info = dummy_cache_info;
1720 num_of_cache_types = ARRAY_SIZE(dummy_cache_info);
1721 pr_warn("dummy cache info is used temporarily and real cache info need update later.\n");
1722 break;
1723 }
1724 }
1725 return num_of_cache_types;
1726}
1727
1728/* Memory required to create Virtual CRAT.
1729 * Since there is no easy way to predict the amount of memory required, the
1730 * following amount is allocated for GPU Virtual CRAT. This is
1731 * expected to cover all known conditions. But to be safe additional check
1732 * is put in the code to ensure we don't overwrite.
1733 */
1734#define VCRAT_SIZE_FOR_GPU (4 * PAGE_SIZE)
1735
1736/* kfd_fill_cu_for_cpu - Fill in Compute info for the given CPU NUMA node
1737 *
1738 * @numa_node_id: CPU NUMA node id
1739 * @avail_size: Available size in the memory
1740 * @sub_type_hdr: Memory into which compute info will be filled in
1741 *
1742 * Return 0 if successful else return -ve value
1743 */
1744static int kfd_fill_cu_for_cpu(int numa_node_id, int *avail_size,
1745 int proximity_domain,
1746 struct crat_subtype_computeunit *sub_type_hdr)
1747{
1748 const struct cpumask *cpumask;
1749
1750 *avail_size -= sizeof(struct crat_subtype_computeunit);
1751 if (*avail_size < 0)
1752 return -ENOMEM;
1753
1754 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit));
1755
1756 /* Fill in subtype header data */
1757 sub_type_hdr->type = CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY;
1758 sub_type_hdr->length = sizeof(struct crat_subtype_computeunit);
1759 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
1760
1761 cpumask = cpumask_of_node(numa_node_id);
1762
1763 /* Fill in CU data */
1764 sub_type_hdr->flags |= CRAT_CU_FLAGS_CPU_PRESENT;
1765 sub_type_hdr->proximity_domain = proximity_domain;
1766 sub_type_hdr->processor_id_low = kfd_numa_node_to_apic_id(numa_node_id);
1767 if (sub_type_hdr->processor_id_low == -1)
1768 return -EINVAL;
1769
1770 sub_type_hdr->num_cpu_cores = cpumask_weight(cpumask);
1771
1772 return 0;
1773}
1774
1775/* kfd_fill_mem_info_for_cpu - Fill in Memory info for the given CPU NUMA node
1776 *
1777 * @numa_node_id: CPU NUMA node id
1778 * @avail_size: Available size in the memory
1779 * @sub_type_hdr: Memory into which compute info will be filled in
1780 *
1781 * Return 0 if successful else return -ve value
1782 */
1783static int kfd_fill_mem_info_for_cpu(int numa_node_id, int *avail_size,
1784 int proximity_domain,
1785 struct crat_subtype_memory *sub_type_hdr)
1786{
1787 uint64_t mem_in_bytes = 0;
1788 pg_data_t *pgdat;
1789 int zone_type;
1790
1791 *avail_size -= sizeof(struct crat_subtype_memory);
1792 if (*avail_size < 0)
1793 return -ENOMEM;
1794
1795 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_memory));
1796
1797 /* Fill in subtype header data */
1798 sub_type_hdr->type = CRAT_SUBTYPE_MEMORY_AFFINITY;
1799 sub_type_hdr->length = sizeof(struct crat_subtype_memory);
1800 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
1801
1802 /* Fill in Memory Subunit data */
1803
1804 /* Unlike si_meminfo, si_meminfo_node is not exported. So
1805 * the following lines are duplicated from si_meminfo_node
1806 * function
1807 */
1808 pgdat = NODE_DATA(numa_node_id);
1809 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
1810 mem_in_bytes += zone_managed_pages(&pgdat->node_zones[zone_type]);
1811 mem_in_bytes <<= PAGE_SHIFT;
1812
1813 sub_type_hdr->length_low = lower_32_bits(mem_in_bytes);
1814 sub_type_hdr->length_high = upper_32_bits(mem_in_bytes);
1815 sub_type_hdr->proximity_domain = proximity_domain;
1816
1817 return 0;
1818}
1819
1820#ifdef CONFIG_X86_64
1821static int kfd_fill_iolink_info_for_cpu(int numa_node_id, int *avail_size,
1822 uint32_t *num_entries,
1823 struct crat_subtype_iolink *sub_type_hdr)
1824{
1825 int nid;
1826 struct cpuinfo_x86 *c = &cpu_data(0);
1827 uint8_t link_type;
1828
1829 if (c->x86_vendor == X86_VENDOR_AMD)
1830 link_type = CRAT_IOLINK_TYPE_HYPERTRANSPORT;
1831 else
1832 link_type = CRAT_IOLINK_TYPE_QPI_1_1;
1833
1834 *num_entries = 0;
1835
1836 /* Create IO links from this node to other CPU nodes */
1837 for_each_online_node(nid) {
1838 if (nid == numa_node_id) /* node itself */
1839 continue;
1840
1841 *avail_size -= sizeof(struct crat_subtype_iolink);
1842 if (*avail_size < 0)
1843 return -ENOMEM;
1844
1845 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
1846
1847 /* Fill in subtype header data */
1848 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
1849 sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
1850 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
1851
1852 /* Fill in IO link data */
1853 sub_type_hdr->proximity_domain_from = numa_node_id;
1854 sub_type_hdr->proximity_domain_to = nid;
1855 sub_type_hdr->io_interface_type = link_type;
1856
1857 (*num_entries)++;
1858 sub_type_hdr++;
1859 }
1860
1861 return 0;
1862}
1863#endif
1864
1865/* kfd_create_vcrat_image_cpu - Create Virtual CRAT for CPU
1866 *
1867 * @pcrat_image: Fill in VCRAT for CPU
1868 * @size: [IN] allocated size of crat_image.
1869 * [OUT] actual size of data filled in crat_image
1870 */
1871static int kfd_create_vcrat_image_cpu(void *pcrat_image, size_t *size)
1872{
1873 struct crat_header *crat_table = (struct crat_header *)pcrat_image;
1874 struct acpi_table_header *acpi_table;
1875 acpi_status status;
1876 struct crat_subtype_generic *sub_type_hdr;
1877 int avail_size = *size;
1878 int numa_node_id;
1879#ifdef CONFIG_X86_64
1880 uint32_t entries = 0;
1881#endif
1882 int ret = 0;
1883
1884 if (!pcrat_image)
1885 return -EINVAL;
1886
1887 /* Fill in CRAT Header.
1888 * Modify length and total_entries as subunits are added.
1889 */
1890 avail_size -= sizeof(struct crat_header);
1891 if (avail_size < 0)
1892 return -ENOMEM;
1893
1894 memset(crat_table, 0, sizeof(struct crat_header));
1895 memcpy(&crat_table->signature, CRAT_SIGNATURE,
1896 sizeof(crat_table->signature));
1897 crat_table->length = sizeof(struct crat_header);
1898
1899 status = acpi_get_table("DSDT", 0, &acpi_table);
1900 if (status != AE_OK)
1901 pr_warn("DSDT table not found for OEM information\n");
1902 else {
1903 crat_table->oem_revision = acpi_table->revision;
1904 memcpy(crat_table->oem_id, acpi_table->oem_id,
1905 CRAT_OEMID_LENGTH);
1906 memcpy(crat_table->oem_table_id, acpi_table->oem_table_id,
1907 CRAT_OEMTABLEID_LENGTH);
1908 acpi_put_table(acpi_table);
1909 }
1910 crat_table->total_entries = 0;
1911 crat_table->num_domains = 0;
1912
1913 sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1);
1914
1915 for_each_online_node(numa_node_id) {
1916 if (kfd_numa_node_to_apic_id(numa_node_id) == -1)
1917 continue;
1918
1919 /* Fill in Subtype: Compute Unit */
1920 ret = kfd_fill_cu_for_cpu(numa_node_id, &avail_size,
1921 crat_table->num_domains,
1922 (struct crat_subtype_computeunit *)sub_type_hdr);
1923 if (ret < 0)
1924 return ret;
1925 crat_table->length += sub_type_hdr->length;
1926 crat_table->total_entries++;
1927
1928 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1929 sub_type_hdr->length);
1930
1931 /* Fill in Subtype: Memory */
1932 ret = kfd_fill_mem_info_for_cpu(numa_node_id, &avail_size,
1933 crat_table->num_domains,
1934 (struct crat_subtype_memory *)sub_type_hdr);
1935 if (ret < 0)
1936 return ret;
1937 crat_table->length += sub_type_hdr->length;
1938 crat_table->total_entries++;
1939
1940 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1941 sub_type_hdr->length);
1942
1943 /* Fill in Subtype: IO Link */
1944#ifdef CONFIG_X86_64
1945 ret = kfd_fill_iolink_info_for_cpu(numa_node_id, &avail_size,
1946 &entries,
1947 (struct crat_subtype_iolink *)sub_type_hdr);
1948 if (ret < 0)
1949 return ret;
1950
1951 if (entries) {
1952 crat_table->length += (sub_type_hdr->length * entries);
1953 crat_table->total_entries += entries;
1954
1955 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1956 sub_type_hdr->length * entries);
1957 }
1958#else
1959 pr_info("IO link not available for non x86 platforms\n");
1960#endif
1961
1962 crat_table->num_domains++;
1963 }
1964
1965 /* TODO: Add cache Subtype for CPU.
1966 * Currently, CPU cache information is available in function
1967 * detect_cache_attributes(cpu) defined in the file
1968 * ./arch/x86/kernel/cpu/intel_cacheinfo.c. This function is not
1969 * exported and to get the same information the code needs to be
1970 * duplicated.
1971 */
1972
1973 *size = crat_table->length;
1974 pr_info("Virtual CRAT table created for CPU\n");
1975
1976 return 0;
1977}
1978
1979static int kfd_fill_gpu_memory_affinity(int *avail_size,
1980 struct kfd_node *kdev, uint8_t type, uint64_t size,
1981 struct crat_subtype_memory *sub_type_hdr,
1982 uint32_t proximity_domain,
1983 const struct kfd_local_mem_info *local_mem_info)
1984{
1985 *avail_size -= sizeof(struct crat_subtype_memory);
1986 if (*avail_size < 0)
1987 return -ENOMEM;
1988
1989 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_memory));
1990 sub_type_hdr->type = CRAT_SUBTYPE_MEMORY_AFFINITY;
1991 sub_type_hdr->length = sizeof(struct crat_subtype_memory);
1992 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED;
1993
1994 sub_type_hdr->proximity_domain = proximity_domain;
1995
1996 pr_debug("Fill gpu memory affinity - type 0x%x size 0x%llx\n",
1997 type, size);
1998
1999 sub_type_hdr->length_low = lower_32_bits(size);
2000 sub_type_hdr->length_high = upper_32_bits(size);
2001
2002 sub_type_hdr->width = local_mem_info->vram_width;
2003 sub_type_hdr->visibility_type = type;
2004
2005 return 0;
2006}
2007
2008#ifdef CONFIG_ACPI_NUMA
2009static void kfd_find_numa_node_in_srat(struct kfd_node *kdev)
2010{
2011 struct acpi_table_header *table_header = NULL;
2012 struct acpi_subtable_header *sub_header = NULL;
2013 unsigned long table_end, subtable_len;
2014 u32 pci_id = pci_domain_nr(kdev->adev->pdev->bus) << 16 |
2015 pci_dev_id(kdev->adev->pdev);
2016 u32 bdf;
2017 acpi_status status;
2018 struct acpi_srat_cpu_affinity *cpu;
2019 struct acpi_srat_generic_affinity *gpu;
2020 int pxm = 0, max_pxm = 0;
2021 int numa_node = NUMA_NO_NODE;
2022 bool found = false;
2023
2024 /* Fetch the SRAT table from ACPI */
2025 status = acpi_get_table(ACPI_SIG_SRAT, 0, &table_header);
2026 if (status == AE_NOT_FOUND) {
2027 pr_warn("SRAT table not found\n");
2028 return;
2029 } else if (ACPI_FAILURE(status)) {
2030 const char *err = acpi_format_exception(status);
2031 pr_err("SRAT table error: %s\n", err);
2032 return;
2033 }
2034
2035 table_end = (unsigned long)table_header + table_header->length;
2036
2037 /* Parse all entries looking for a match. */
2038 sub_header = (struct acpi_subtable_header *)
2039 ((unsigned long)table_header +
2040 sizeof(struct acpi_table_srat));
2041 subtable_len = sub_header->length;
2042
2043 while (((unsigned long)sub_header) + subtable_len < table_end) {
2044 /*
2045 * If length is 0, break from this loop to avoid
2046 * infinite loop.
2047 */
2048 if (subtable_len == 0) {
2049 pr_err("SRAT invalid zero length\n");
2050 break;
2051 }
2052
2053 switch (sub_header->type) {
2054 case ACPI_SRAT_TYPE_CPU_AFFINITY:
2055 cpu = (struct acpi_srat_cpu_affinity *)sub_header;
2056 pxm = *((u32 *)cpu->proximity_domain_hi) << 8 |
2057 cpu->proximity_domain_lo;
2058 if (pxm > max_pxm)
2059 max_pxm = pxm;
2060 break;
2061 case ACPI_SRAT_TYPE_GENERIC_AFFINITY:
2062 gpu = (struct acpi_srat_generic_affinity *)sub_header;
2063 bdf = *((u16 *)(&gpu->device_handle[0])) << 16 |
2064 *((u16 *)(&gpu->device_handle[2]));
2065 if (bdf == pci_id) {
2066 found = true;
2067 numa_node = pxm_to_node(gpu->proximity_domain);
2068 }
2069 break;
2070 default:
2071 break;
2072 }
2073
2074 if (found)
2075 break;
2076
2077 sub_header = (struct acpi_subtable_header *)
2078 ((unsigned long)sub_header + subtable_len);
2079 subtable_len = sub_header->length;
2080 }
2081
2082 acpi_put_table(table_header);
2083
2084 /* Workaround bad cpu-gpu binding case */
2085 if (found && (numa_node < 0 ||
2086 numa_node > pxm_to_node(max_pxm)))
2087 numa_node = 0;
2088
2089 if (numa_node != NUMA_NO_NODE)
2090 set_dev_node(&kdev->adev->pdev->dev, numa_node);
2091}
2092#endif
2093
2094#define KFD_CRAT_INTRA_SOCKET_WEIGHT 13
2095#define KFD_CRAT_XGMI_WEIGHT 15
2096
2097/* kfd_fill_gpu_direct_io_link - Fill in direct io link from GPU
2098 * to its NUMA node
2099 * @avail_size: Available size in the memory
2100 * @kdev - [IN] GPU device
2101 * @sub_type_hdr: Memory into which io link info will be filled in
2102 * @proximity_domain - proximity domain of the GPU node
2103 *
2104 * Return 0 if successful else return -ve value
2105 */
2106static int kfd_fill_gpu_direct_io_link_to_cpu(int *avail_size,
2107 struct kfd_node *kdev,
2108 struct crat_subtype_iolink *sub_type_hdr,
2109 uint32_t proximity_domain)
2110{
2111 *avail_size -= sizeof(struct crat_subtype_iolink);
2112 if (*avail_size < 0)
2113 return -ENOMEM;
2114
2115 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
2116
2117 /* Fill in subtype header data */
2118 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
2119 sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
2120 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED;
2121 if (kfd_dev_is_large_bar(kdev))
2122 sub_type_hdr->flags |= CRAT_IOLINK_FLAGS_BI_DIRECTIONAL;
2123
2124 /* Fill in IOLINK subtype.
2125 * TODO: Fill-in other fields of iolink subtype
2126 */
2127 if (kdev->adev->gmc.xgmi.connected_to_cpu ||
2128 (KFD_GC_VERSION(kdev) == IP_VERSION(9, 4, 3) &&
2129 kdev->adev->smuio.funcs->get_pkg_type(kdev->adev) ==
2130 AMDGPU_PKG_TYPE_APU)) {
2131 bool ext_cpu = KFD_GC_VERSION(kdev) != IP_VERSION(9, 4, 3);
2132 int mem_bw = 819200, weight = ext_cpu ? KFD_CRAT_XGMI_WEIGHT :
2133 KFD_CRAT_INTRA_SOCKET_WEIGHT;
2134 uint32_t bandwidth = ext_cpu ? amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(
2135 kdev->adev, NULL, true) : mem_bw;
2136
2137 /*
2138 * with host gpu xgmi link, host can access gpu memory whether
2139 * or not pcie bar type is large, so always create bidirectional
2140 * io link.
2141 */
2142 sub_type_hdr->flags |= CRAT_IOLINK_FLAGS_BI_DIRECTIONAL;
2143 sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_XGMI;
2144 sub_type_hdr->weight_xgmi = weight;
2145 sub_type_hdr->minimum_bandwidth_mbs = bandwidth;
2146 sub_type_hdr->maximum_bandwidth_mbs = bandwidth;
2147 } else {
2148 sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_PCIEXPRESS;
2149 sub_type_hdr->minimum_bandwidth_mbs =
2150 amdgpu_amdkfd_get_pcie_bandwidth_mbytes(kdev->adev, true);
2151 sub_type_hdr->maximum_bandwidth_mbs =
2152 amdgpu_amdkfd_get_pcie_bandwidth_mbytes(kdev->adev, false);
2153 }
2154
2155 sub_type_hdr->proximity_domain_from = proximity_domain;
2156
2157#ifdef CONFIG_ACPI_NUMA
2158 if (kdev->adev->pdev->dev.numa_node == NUMA_NO_NODE &&
2159 num_possible_nodes() > 1)
2160 kfd_find_numa_node_in_srat(kdev);
2161#endif
2162#ifdef CONFIG_NUMA
2163 if (kdev->adev->pdev->dev.numa_node == NUMA_NO_NODE)
2164 sub_type_hdr->proximity_domain_to = 0;
2165 else
2166 sub_type_hdr->proximity_domain_to = kdev->adev->pdev->dev.numa_node;
2167#else
2168 sub_type_hdr->proximity_domain_to = 0;
2169#endif
2170 return 0;
2171}
2172
2173static int kfd_fill_gpu_xgmi_link_to_gpu(int *avail_size,
2174 struct kfd_node *kdev,
2175 struct kfd_node *peer_kdev,
2176 struct crat_subtype_iolink *sub_type_hdr,
2177 uint32_t proximity_domain_from,
2178 uint32_t proximity_domain_to)
2179{
2180 bool use_ta_info = kdev->kfd->num_nodes == 1;
2181
2182 *avail_size -= sizeof(struct crat_subtype_iolink);
2183 if (*avail_size < 0)
2184 return -ENOMEM;
2185
2186 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
2187
2188 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
2189 sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
2190 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED |
2191 CRAT_IOLINK_FLAGS_BI_DIRECTIONAL;
2192
2193 sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_XGMI;
2194 sub_type_hdr->proximity_domain_from = proximity_domain_from;
2195 sub_type_hdr->proximity_domain_to = proximity_domain_to;
2196
2197 if (use_ta_info) {
2198 sub_type_hdr->weight_xgmi = KFD_CRAT_XGMI_WEIGHT *
2199 amdgpu_amdkfd_get_xgmi_hops_count(kdev->adev, peer_kdev->adev);
2200 sub_type_hdr->maximum_bandwidth_mbs =
2201 amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(kdev->adev,
2202 peer_kdev->adev, false);
2203 sub_type_hdr->minimum_bandwidth_mbs = sub_type_hdr->maximum_bandwidth_mbs ?
2204 amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(kdev->adev, NULL, true) : 0;
2205 } else {
2206 bool is_single_hop = kdev->kfd == peer_kdev->kfd;
2207 int weight = is_single_hop ? KFD_CRAT_INTRA_SOCKET_WEIGHT :
2208 (2 * KFD_CRAT_INTRA_SOCKET_WEIGHT) + KFD_CRAT_XGMI_WEIGHT;
2209 int mem_bw = 819200;
2210
2211 sub_type_hdr->weight_xgmi = weight;
2212 sub_type_hdr->maximum_bandwidth_mbs = is_single_hop ? mem_bw : 0;
2213 sub_type_hdr->minimum_bandwidth_mbs = is_single_hop ? mem_bw : 0;
2214 }
2215
2216 return 0;
2217}
2218
2219/* kfd_create_vcrat_image_gpu - Create Virtual CRAT for CPU
2220 *
2221 * @pcrat_image: Fill in VCRAT for GPU
2222 * @size: [IN] allocated size of crat_image.
2223 * [OUT] actual size of data filled in crat_image
2224 */
2225static int kfd_create_vcrat_image_gpu(void *pcrat_image,
2226 size_t *size, struct kfd_node *kdev,
2227 uint32_t proximity_domain)
2228{
2229 struct crat_header *crat_table = (struct crat_header *)pcrat_image;
2230 struct amdgpu_gfx_config *gfx_info = &kdev->adev->gfx.config;
2231 struct amdgpu_cu_info *cu_info = &kdev->adev->gfx.cu_info;
2232 struct crat_subtype_generic *sub_type_hdr;
2233 struct kfd_local_mem_info local_mem_info;
2234 struct kfd_topology_device *peer_dev;
2235 struct crat_subtype_computeunit *cu;
2236 int avail_size = *size;
2237 uint32_t total_num_of_cu;
2238 uint32_t nid = 0;
2239 int ret = 0;
2240
2241 if (!pcrat_image || avail_size < VCRAT_SIZE_FOR_GPU)
2242 return -EINVAL;
2243
2244 /* Fill the CRAT Header.
2245 * Modify length and total_entries as subunits are added.
2246 */
2247 avail_size -= sizeof(struct crat_header);
2248 memset(crat_table, 0, sizeof(struct crat_header));
2249
2250 memcpy(&crat_table->signature, CRAT_SIGNATURE,
2251 sizeof(crat_table->signature));
2252 /* Change length as we add more subtypes*/
2253 crat_table->length = sizeof(struct crat_header);
2254 crat_table->num_domains = 1;
2255 crat_table->total_entries = 0;
2256
2257 /* Fill in Subtype: Compute Unit
2258 * First fill in the sub type header and then sub type data
2259 */
2260 avail_size -= sizeof(struct crat_subtype_computeunit);
2261 sub_type_hdr = (struct crat_subtype_generic *)(crat_table + 1);
2262 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit));
2263
2264 sub_type_hdr->type = CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY;
2265 sub_type_hdr->length = sizeof(struct crat_subtype_computeunit);
2266 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
2267
2268 /* Fill CU subtype data */
2269 cu = (struct crat_subtype_computeunit *)sub_type_hdr;
2270 cu->flags |= CRAT_CU_FLAGS_GPU_PRESENT;
2271 cu->proximity_domain = proximity_domain;
2272
2273 cu->num_simd_per_cu = cu_info->simd_per_cu;
2274 cu->num_simd_cores = cu_info->simd_per_cu *
2275 (cu_info->number / kdev->kfd->num_nodes);
2276 cu->max_waves_simd = cu_info->max_waves_per_simd;
2277
2278 cu->wave_front_size = cu_info->wave_front_size;
2279 cu->array_count = gfx_info->max_sh_per_se *
2280 gfx_info->max_shader_engines;
2281 total_num_of_cu = (cu->array_count * gfx_info->max_cu_per_sh);
2282 cu->processor_id_low = get_and_inc_gpu_processor_id(total_num_of_cu);
2283 cu->num_cu_per_array = gfx_info->max_cu_per_sh;
2284 cu->max_slots_scatch_cu = cu_info->max_scratch_slots_per_cu;
2285 cu->num_banks = gfx_info->max_shader_engines;
2286 cu->lds_size_in_kb = cu_info->lds_size;
2287
2288 cu->hsa_capability = 0;
2289
2290 crat_table->length += sub_type_hdr->length;
2291 crat_table->total_entries++;
2292
2293 /* Fill in Subtype: Memory. Only on systems with large BAR (no
2294 * private FB), report memory as public. On other systems
2295 * report the total FB size (public+private) as a single
2296 * private heap.
2297 */
2298 local_mem_info = kdev->local_mem_info;
2299 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
2300 sub_type_hdr->length);
2301
2302 if (kdev->adev->debug_largebar)
2303 local_mem_info.local_mem_size_private = 0;
2304
2305 if (local_mem_info.local_mem_size_private == 0)
2306 ret = kfd_fill_gpu_memory_affinity(&avail_size,
2307 kdev, HSA_MEM_HEAP_TYPE_FB_PUBLIC,
2308 local_mem_info.local_mem_size_public,
2309 (struct crat_subtype_memory *)sub_type_hdr,
2310 proximity_domain,
2311 &local_mem_info);
2312 else
2313 ret = kfd_fill_gpu_memory_affinity(&avail_size,
2314 kdev, HSA_MEM_HEAP_TYPE_FB_PRIVATE,
2315 local_mem_info.local_mem_size_public +
2316 local_mem_info.local_mem_size_private,
2317 (struct crat_subtype_memory *)sub_type_hdr,
2318 proximity_domain,
2319 &local_mem_info);
2320 if (ret < 0)
2321 return ret;
2322
2323 crat_table->length += sizeof(struct crat_subtype_memory);
2324 crat_table->total_entries++;
2325
2326 /* Fill in Subtype: IO_LINKS
2327 * Only direct links are added here which is Link from GPU to
2328 * its NUMA node. Indirect links are added by userspace.
2329 */
2330 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
2331 sub_type_hdr->length);
2332 ret = kfd_fill_gpu_direct_io_link_to_cpu(&avail_size, kdev,
2333 (struct crat_subtype_iolink *)sub_type_hdr, proximity_domain);
2334
2335 if (ret < 0)
2336 return ret;
2337
2338 crat_table->length += sub_type_hdr->length;
2339 crat_table->total_entries++;
2340
2341
2342 /* Fill in Subtype: IO_LINKS
2343 * Direct links from GPU to other GPUs through xGMI.
2344 * We will loop GPUs that already be processed (with lower value
2345 * of proximity_domain), add the link for the GPUs with same
2346 * hive id (from this GPU to other GPU) . The reversed iolink
2347 * (from other GPU to this GPU) will be added
2348 * in kfd_parse_subtype_iolink.
2349 */
2350 if (kdev->kfd->hive_id) {
2351 for (nid = 0; nid < proximity_domain; ++nid) {
2352 peer_dev = kfd_topology_device_by_proximity_domain_no_lock(nid);
2353 if (!peer_dev->gpu)
2354 continue;
2355 if (peer_dev->gpu->kfd->hive_id != kdev->kfd->hive_id)
2356 continue;
2357 if (!amdgpu_xgmi_get_is_sharing_enabled(kdev->adev, peer_dev->gpu->adev))
2358 continue;
2359 sub_type_hdr = (typeof(sub_type_hdr))(
2360 (char *)sub_type_hdr +
2361 sizeof(struct crat_subtype_iolink));
2362 ret = kfd_fill_gpu_xgmi_link_to_gpu(
2363 &avail_size, kdev, peer_dev->gpu,
2364 (struct crat_subtype_iolink *)sub_type_hdr,
2365 proximity_domain, nid);
2366 if (ret < 0)
2367 return ret;
2368 crat_table->length += sub_type_hdr->length;
2369 crat_table->total_entries++;
2370 }
2371 }
2372 *size = crat_table->length;
2373 pr_info("Virtual CRAT table created for GPU\n");
2374
2375 return ret;
2376}
2377
2378/* kfd_create_crat_image_virtual - Allocates memory for CRAT image and
2379 * creates a Virtual CRAT (VCRAT) image
2380 *
2381 * NOTE: Call kfd_destroy_crat_image to free CRAT image memory
2382 *
2383 * @crat_image: VCRAT image created because ACPI does not have a
2384 * CRAT for this device
2385 * @size: [OUT] size of virtual crat_image
2386 * @flags: COMPUTE_UNIT_CPU - Create VCRAT for CPU device
2387 * COMPUTE_UNIT_GPU - Create VCRAT for GPU
2388 * (COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU) - Create VCRAT for APU
2389 * -- this option is not currently implemented.
2390 * The assumption is that all AMD APUs will have CRAT
2391 * @kdev: Valid kfd_node required if flags contain COMPUTE_UNIT_GPU
2392 *
2393 * Return 0 if successful else return -ve value
2394 */
2395int kfd_create_crat_image_virtual(void **crat_image, size_t *size,
2396 int flags, struct kfd_node *kdev,
2397 uint32_t proximity_domain)
2398{
2399 void *pcrat_image = NULL;
2400 int ret = 0, num_nodes;
2401 size_t dyn_size;
2402
2403 if (!crat_image)
2404 return -EINVAL;
2405
2406 *crat_image = NULL;
2407
2408 /* Allocate the CPU Virtual CRAT size based on the number of online
2409 * nodes. Allocate VCRAT_SIZE_FOR_GPU for GPU virtual CRAT image.
2410 * This should cover all the current conditions. A check is put not
2411 * to overwrite beyond allocated size for GPUs
2412 */
2413 switch (flags) {
2414 case COMPUTE_UNIT_CPU:
2415 num_nodes = num_online_nodes();
2416 dyn_size = sizeof(struct crat_header) +
2417 num_nodes * (sizeof(struct crat_subtype_computeunit) +
2418 sizeof(struct crat_subtype_memory) +
2419 (num_nodes - 1) * sizeof(struct crat_subtype_iolink));
2420 pcrat_image = kvmalloc(dyn_size, GFP_KERNEL);
2421 if (!pcrat_image)
2422 return -ENOMEM;
2423 *size = dyn_size;
2424 pr_debug("CRAT size is %ld", dyn_size);
2425 ret = kfd_create_vcrat_image_cpu(pcrat_image, size);
2426 break;
2427 case COMPUTE_UNIT_GPU:
2428 if (!kdev)
2429 return -EINVAL;
2430 pcrat_image = kvmalloc(VCRAT_SIZE_FOR_GPU, GFP_KERNEL);
2431 if (!pcrat_image)
2432 return -ENOMEM;
2433 *size = VCRAT_SIZE_FOR_GPU;
2434 ret = kfd_create_vcrat_image_gpu(pcrat_image, size, kdev,
2435 proximity_domain);
2436 break;
2437 case (COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU):
2438 /* TODO: */
2439 ret = -EINVAL;
2440 pr_err("VCRAT not implemented for APU\n");
2441 break;
2442 default:
2443 ret = -EINVAL;
2444 }
2445
2446 if (!ret)
2447 *crat_image = pcrat_image;
2448 else
2449 kvfree(pcrat_image);
2450
2451 return ret;
2452}
2453
2454
2455/* kfd_destroy_crat_image
2456 *
2457 * @crat_image: [IN] - crat_image from kfd_create_crat_image_xxx(..)
2458 *
2459 */
2460void kfd_destroy_crat_image(void *crat_image)
2461{
2462 kvfree(crat_image);
2463}