Loading...
1/*
2 * Copyright 2015-2017 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
21 */
22
23#include <linux/pci.h>
24#include <linux/acpi.h>
25#include "kfd_crat.h"
26#include "kfd_priv.h"
27#include "kfd_topology.h"
28#include "kfd_iommu.h"
29#include "amdgpu_amdkfd.h"
30
31/* GPU Processor ID base for dGPUs for which VCRAT needs to be created.
32 * GPU processor ID are expressed with Bit[31]=1.
33 * The base is set to 0x8000_0000 + 0x1000 to avoid collision with GPU IDs
34 * used in the CRAT.
35 */
36static uint32_t gpu_processor_id_low = 0x80001000;
37
38/* Return the next available gpu_processor_id and increment it for next GPU
39 * @total_cu_count - Total CUs present in the GPU including ones
40 * masked off
41 */
42static inline unsigned int get_and_inc_gpu_processor_id(
43 unsigned int total_cu_count)
44{
45 int current_id = gpu_processor_id_low;
46
47 gpu_processor_id_low += total_cu_count;
48 return current_id;
49}
50
51/* Static table to describe GPU Cache information */
52struct kfd_gpu_cache_info {
53 uint32_t cache_size;
54 uint32_t cache_level;
55 uint32_t flags;
56 /* Indicates how many Compute Units share this cache
57 * Value = 1 indicates the cache is not shared
58 */
59 uint32_t num_cu_shared;
60};
61
62static struct kfd_gpu_cache_info kaveri_cache_info[] = {
63 {
64 /* TCP L1 Cache per CU */
65 .cache_size = 16,
66 .cache_level = 1,
67 .flags = (CRAT_CACHE_FLAGS_ENABLED |
68 CRAT_CACHE_FLAGS_DATA_CACHE |
69 CRAT_CACHE_FLAGS_SIMD_CACHE),
70 .num_cu_shared = 1,
71
72 },
73 {
74 /* Scalar L1 Instruction Cache (in SQC module) per bank */
75 .cache_size = 16,
76 .cache_level = 1,
77 .flags = (CRAT_CACHE_FLAGS_ENABLED |
78 CRAT_CACHE_FLAGS_INST_CACHE |
79 CRAT_CACHE_FLAGS_SIMD_CACHE),
80 .num_cu_shared = 2,
81 },
82 {
83 /* Scalar L1 Data Cache (in SQC module) per bank */
84 .cache_size = 8,
85 .cache_level = 1,
86 .flags = (CRAT_CACHE_FLAGS_ENABLED |
87 CRAT_CACHE_FLAGS_DATA_CACHE |
88 CRAT_CACHE_FLAGS_SIMD_CACHE),
89 .num_cu_shared = 2,
90 },
91
92 /* TODO: Add L2 Cache information */
93};
94
95
96static struct kfd_gpu_cache_info carrizo_cache_info[] = {
97 {
98 /* TCP L1 Cache per CU */
99 .cache_size = 16,
100 .cache_level = 1,
101 .flags = (CRAT_CACHE_FLAGS_ENABLED |
102 CRAT_CACHE_FLAGS_DATA_CACHE |
103 CRAT_CACHE_FLAGS_SIMD_CACHE),
104 .num_cu_shared = 1,
105 },
106 {
107 /* Scalar L1 Instruction Cache (in SQC module) per bank */
108 .cache_size = 8,
109 .cache_level = 1,
110 .flags = (CRAT_CACHE_FLAGS_ENABLED |
111 CRAT_CACHE_FLAGS_INST_CACHE |
112 CRAT_CACHE_FLAGS_SIMD_CACHE),
113 .num_cu_shared = 4,
114 },
115 {
116 /* Scalar L1 Data Cache (in SQC module) per bank. */
117 .cache_size = 4,
118 .cache_level = 1,
119 .flags = (CRAT_CACHE_FLAGS_ENABLED |
120 CRAT_CACHE_FLAGS_DATA_CACHE |
121 CRAT_CACHE_FLAGS_SIMD_CACHE),
122 .num_cu_shared = 4,
123 },
124
125 /* TODO: Add L2 Cache information */
126};
127
128/* NOTE: In future if more information is added to struct kfd_gpu_cache_info
129 * the following ASICs may need a separate table.
130 */
131#define hawaii_cache_info kaveri_cache_info
132#define tonga_cache_info carrizo_cache_info
133#define fiji_cache_info carrizo_cache_info
134#define polaris10_cache_info carrizo_cache_info
135#define polaris11_cache_info carrizo_cache_info
136#define polaris12_cache_info carrizo_cache_info
137#define vegam_cache_info carrizo_cache_info
138/* TODO - check & update Vega10 cache details */
139#define vega10_cache_info carrizo_cache_info
140#define raven_cache_info carrizo_cache_info
141#define renoir_cache_info carrizo_cache_info
142/* TODO - check & update Navi10 cache details */
143#define navi10_cache_info carrizo_cache_info
144
145static void kfd_populated_cu_info_cpu(struct kfd_topology_device *dev,
146 struct crat_subtype_computeunit *cu)
147{
148 dev->node_props.cpu_cores_count = cu->num_cpu_cores;
149 dev->node_props.cpu_core_id_base = cu->processor_id_low;
150 if (cu->hsa_capability & CRAT_CU_FLAGS_IOMMU_PRESENT)
151 dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
152
153 pr_debug("CU CPU: cores=%d id_base=%d\n", cu->num_cpu_cores,
154 cu->processor_id_low);
155}
156
157static void kfd_populated_cu_info_gpu(struct kfd_topology_device *dev,
158 struct crat_subtype_computeunit *cu)
159{
160 dev->node_props.simd_id_base = cu->processor_id_low;
161 dev->node_props.simd_count = cu->num_simd_cores;
162 dev->node_props.lds_size_in_kb = cu->lds_size_in_kb;
163 dev->node_props.max_waves_per_simd = cu->max_waves_simd;
164 dev->node_props.wave_front_size = cu->wave_front_size;
165 dev->node_props.array_count = cu->array_count;
166 dev->node_props.cu_per_simd_array = cu->num_cu_per_array;
167 dev->node_props.simd_per_cu = cu->num_simd_per_cu;
168 dev->node_props.max_slots_scratch_cu = cu->max_slots_scatch_cu;
169 if (cu->hsa_capability & CRAT_CU_FLAGS_HOT_PLUGGABLE)
170 dev->node_props.capability |= HSA_CAP_HOT_PLUGGABLE;
171 pr_debug("CU GPU: id_base=%d\n", cu->processor_id_low);
172}
173
174/* kfd_parse_subtype_cu - parse compute unit subtypes and attach it to correct
175 * topology device present in the device_list
176 */
177static int kfd_parse_subtype_cu(struct crat_subtype_computeunit *cu,
178 struct list_head *device_list)
179{
180 struct kfd_topology_device *dev;
181
182 pr_debug("Found CU entry in CRAT table with proximity_domain=%d caps=%x\n",
183 cu->proximity_domain, cu->hsa_capability);
184 list_for_each_entry(dev, device_list, list) {
185 if (cu->proximity_domain == dev->proximity_domain) {
186 if (cu->flags & CRAT_CU_FLAGS_CPU_PRESENT)
187 kfd_populated_cu_info_cpu(dev, cu);
188
189 if (cu->flags & CRAT_CU_FLAGS_GPU_PRESENT)
190 kfd_populated_cu_info_gpu(dev, cu);
191 break;
192 }
193 }
194
195 return 0;
196}
197
198static struct kfd_mem_properties *
199find_subtype_mem(uint32_t heap_type, uint32_t flags, uint32_t width,
200 struct kfd_topology_device *dev)
201{
202 struct kfd_mem_properties *props;
203
204 list_for_each_entry(props, &dev->mem_props, list) {
205 if (props->heap_type == heap_type
206 && props->flags == flags
207 && props->width == width)
208 return props;
209 }
210
211 return NULL;
212}
213/* kfd_parse_subtype_mem - parse memory subtypes and attach it to correct
214 * topology device present in the device_list
215 */
216static int kfd_parse_subtype_mem(struct crat_subtype_memory *mem,
217 struct list_head *device_list)
218{
219 struct kfd_mem_properties *props;
220 struct kfd_topology_device *dev;
221 uint32_t heap_type;
222 uint64_t size_in_bytes;
223 uint32_t flags = 0;
224 uint32_t width;
225
226 pr_debug("Found memory entry in CRAT table with proximity_domain=%d\n",
227 mem->proximity_domain);
228 list_for_each_entry(dev, device_list, list) {
229 if (mem->proximity_domain == dev->proximity_domain) {
230 /* We're on GPU node */
231 if (dev->node_props.cpu_cores_count == 0) {
232 /* APU */
233 if (mem->visibility_type == 0)
234 heap_type =
235 HSA_MEM_HEAP_TYPE_FB_PRIVATE;
236 /* dGPU */
237 else
238 heap_type = mem->visibility_type;
239 } else
240 heap_type = HSA_MEM_HEAP_TYPE_SYSTEM;
241
242 if (mem->flags & CRAT_MEM_FLAGS_HOT_PLUGGABLE)
243 flags |= HSA_MEM_FLAGS_HOT_PLUGGABLE;
244 if (mem->flags & CRAT_MEM_FLAGS_NON_VOLATILE)
245 flags |= HSA_MEM_FLAGS_NON_VOLATILE;
246
247 size_in_bytes =
248 ((uint64_t)mem->length_high << 32) +
249 mem->length_low;
250 width = mem->width;
251
252 /* Multiple banks of the same type are aggregated into
253 * one. User mode doesn't care about multiple physical
254 * memory segments. It's managed as a single virtual
255 * heap for user mode.
256 */
257 props = find_subtype_mem(heap_type, flags, width, dev);
258 if (props) {
259 props->size_in_bytes += size_in_bytes;
260 break;
261 }
262
263 props = kfd_alloc_struct(props);
264 if (!props)
265 return -ENOMEM;
266
267 props->heap_type = heap_type;
268 props->flags = flags;
269 props->size_in_bytes = size_in_bytes;
270 props->width = width;
271
272 dev->node_props.mem_banks_count++;
273 list_add_tail(&props->list, &dev->mem_props);
274
275 break;
276 }
277 }
278
279 return 0;
280}
281
282/* kfd_parse_subtype_cache - parse cache subtypes and attach it to correct
283 * topology device present in the device_list
284 */
285static int kfd_parse_subtype_cache(struct crat_subtype_cache *cache,
286 struct list_head *device_list)
287{
288 struct kfd_cache_properties *props;
289 struct kfd_topology_device *dev;
290 uint32_t id;
291 uint32_t total_num_of_cu;
292
293 id = cache->processor_id_low;
294
295 pr_debug("Found cache entry in CRAT table with processor_id=%d\n", id);
296 list_for_each_entry(dev, device_list, list) {
297 total_num_of_cu = (dev->node_props.array_count *
298 dev->node_props.cu_per_simd_array);
299
300 /* Cache infomration in CRAT doesn't have proximity_domain
301 * information as it is associated with a CPU core or GPU
302 * Compute Unit. So map the cache using CPU core Id or SIMD
303 * (GPU) ID.
304 * TODO: This works because currently we can safely assume that
305 * Compute Units are parsed before caches are parsed. In
306 * future, remove this dependency
307 */
308 if ((id >= dev->node_props.cpu_core_id_base &&
309 id <= dev->node_props.cpu_core_id_base +
310 dev->node_props.cpu_cores_count) ||
311 (id >= dev->node_props.simd_id_base &&
312 id < dev->node_props.simd_id_base +
313 total_num_of_cu)) {
314 props = kfd_alloc_struct(props);
315 if (!props)
316 return -ENOMEM;
317
318 props->processor_id_low = id;
319 props->cache_level = cache->cache_level;
320 props->cache_size = cache->cache_size;
321 props->cacheline_size = cache->cache_line_size;
322 props->cachelines_per_tag = cache->lines_per_tag;
323 props->cache_assoc = cache->associativity;
324 props->cache_latency = cache->cache_latency;
325 memcpy(props->sibling_map, cache->sibling_map,
326 sizeof(props->sibling_map));
327
328 if (cache->flags & CRAT_CACHE_FLAGS_DATA_CACHE)
329 props->cache_type |= HSA_CACHE_TYPE_DATA;
330 if (cache->flags & CRAT_CACHE_FLAGS_INST_CACHE)
331 props->cache_type |= HSA_CACHE_TYPE_INSTRUCTION;
332 if (cache->flags & CRAT_CACHE_FLAGS_CPU_CACHE)
333 props->cache_type |= HSA_CACHE_TYPE_CPU;
334 if (cache->flags & CRAT_CACHE_FLAGS_SIMD_CACHE)
335 props->cache_type |= HSA_CACHE_TYPE_HSACU;
336
337 dev->cache_count++;
338 dev->node_props.caches_count++;
339 list_add_tail(&props->list, &dev->cache_props);
340
341 break;
342 }
343 }
344
345 return 0;
346}
347
348/* kfd_parse_subtype_iolink - parse iolink subtypes and attach it to correct
349 * topology device present in the device_list
350 */
351static int kfd_parse_subtype_iolink(struct crat_subtype_iolink *iolink,
352 struct list_head *device_list)
353{
354 struct kfd_iolink_properties *props = NULL, *props2;
355 struct kfd_topology_device *dev, *to_dev;
356 uint32_t id_from;
357 uint32_t id_to;
358
359 id_from = iolink->proximity_domain_from;
360 id_to = iolink->proximity_domain_to;
361
362 pr_debug("Found IO link entry in CRAT table with id_from=%d, id_to %d\n",
363 id_from, id_to);
364 list_for_each_entry(dev, device_list, list) {
365 if (id_from == dev->proximity_domain) {
366 props = kfd_alloc_struct(props);
367 if (!props)
368 return -ENOMEM;
369
370 props->node_from = id_from;
371 props->node_to = id_to;
372 props->ver_maj = iolink->version_major;
373 props->ver_min = iolink->version_minor;
374 props->iolink_type = iolink->io_interface_type;
375
376 if (props->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS)
377 props->weight = 20;
378 else if (props->iolink_type == CRAT_IOLINK_TYPE_XGMI)
379 props->weight = 15 * iolink->num_hops_xgmi;
380 else
381 props->weight = node_distance(id_from, id_to);
382
383 props->min_latency = iolink->minimum_latency;
384 props->max_latency = iolink->maximum_latency;
385 props->min_bandwidth = iolink->minimum_bandwidth_mbs;
386 props->max_bandwidth = iolink->maximum_bandwidth_mbs;
387 props->rec_transfer_size =
388 iolink->recommended_transfer_size;
389
390 dev->io_link_count++;
391 dev->node_props.io_links_count++;
392 list_add_tail(&props->list, &dev->io_link_props);
393 break;
394 }
395 }
396
397 /* CPU topology is created before GPUs are detected, so CPU->GPU
398 * links are not built at that time. If a PCIe type is discovered, it
399 * means a GPU is detected and we are adding GPU->CPU to the topology.
400 * At this time, also add the corresponded CPU->GPU link if GPU
401 * is large bar.
402 * For xGMI, we only added the link with one direction in the crat
403 * table, add corresponded reversed direction link now.
404 */
405 if (props && (iolink->flags & CRAT_IOLINK_FLAGS_BI_DIRECTIONAL)) {
406 to_dev = kfd_topology_device_by_proximity_domain(id_to);
407 if (!to_dev)
408 return -ENODEV;
409 /* same everything but the other direction */
410 props2 = kmemdup(props, sizeof(*props2), GFP_KERNEL);
411 props2->node_from = id_to;
412 props2->node_to = id_from;
413 props2->kobj = NULL;
414 to_dev->io_link_count++;
415 to_dev->node_props.io_links_count++;
416 list_add_tail(&props2->list, &to_dev->io_link_props);
417 }
418
419 return 0;
420}
421
422/* kfd_parse_subtype - parse subtypes and attach it to correct topology device
423 * present in the device_list
424 * @sub_type_hdr - subtype section of crat_image
425 * @device_list - list of topology devices present in this crat_image
426 */
427static int kfd_parse_subtype(struct crat_subtype_generic *sub_type_hdr,
428 struct list_head *device_list)
429{
430 struct crat_subtype_computeunit *cu;
431 struct crat_subtype_memory *mem;
432 struct crat_subtype_cache *cache;
433 struct crat_subtype_iolink *iolink;
434 int ret = 0;
435
436 switch (sub_type_hdr->type) {
437 case CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY:
438 cu = (struct crat_subtype_computeunit *)sub_type_hdr;
439 ret = kfd_parse_subtype_cu(cu, device_list);
440 break;
441 case CRAT_SUBTYPE_MEMORY_AFFINITY:
442 mem = (struct crat_subtype_memory *)sub_type_hdr;
443 ret = kfd_parse_subtype_mem(mem, device_list);
444 break;
445 case CRAT_SUBTYPE_CACHE_AFFINITY:
446 cache = (struct crat_subtype_cache *)sub_type_hdr;
447 ret = kfd_parse_subtype_cache(cache, device_list);
448 break;
449 case CRAT_SUBTYPE_TLB_AFFINITY:
450 /*
451 * For now, nothing to do here
452 */
453 pr_debug("Found TLB entry in CRAT table (not processing)\n");
454 break;
455 case CRAT_SUBTYPE_CCOMPUTE_AFFINITY:
456 /*
457 * For now, nothing to do here
458 */
459 pr_debug("Found CCOMPUTE entry in CRAT table (not processing)\n");
460 break;
461 case CRAT_SUBTYPE_IOLINK_AFFINITY:
462 iolink = (struct crat_subtype_iolink *)sub_type_hdr;
463 ret = kfd_parse_subtype_iolink(iolink, device_list);
464 break;
465 default:
466 pr_warn("Unknown subtype %d in CRAT\n",
467 sub_type_hdr->type);
468 }
469
470 return ret;
471}
472
473/* kfd_parse_crat_table - parse CRAT table. For each node present in CRAT
474 * create a kfd_topology_device and add in to device_list. Also parse
475 * CRAT subtypes and attach it to appropriate kfd_topology_device
476 * @crat_image - input image containing CRAT
477 * @device_list - [OUT] list of kfd_topology_device generated after
478 * parsing crat_image
479 * @proximity_domain - Proximity domain of the first device in the table
480 *
481 * Return - 0 if successful else -ve value
482 */
483int kfd_parse_crat_table(void *crat_image, struct list_head *device_list,
484 uint32_t proximity_domain)
485{
486 struct kfd_topology_device *top_dev = NULL;
487 struct crat_subtype_generic *sub_type_hdr;
488 uint16_t node_id;
489 int ret = 0;
490 struct crat_header *crat_table = (struct crat_header *)crat_image;
491 uint16_t num_nodes;
492 uint32_t image_len;
493
494 if (!crat_image)
495 return -EINVAL;
496
497 if (!list_empty(device_list)) {
498 pr_warn("Error device list should be empty\n");
499 return -EINVAL;
500 }
501
502 num_nodes = crat_table->num_domains;
503 image_len = crat_table->length;
504
505 pr_debug("Parsing CRAT table with %d nodes\n", num_nodes);
506
507 for (node_id = 0; node_id < num_nodes; node_id++) {
508 top_dev = kfd_create_topology_device(device_list);
509 if (!top_dev)
510 break;
511 top_dev->proximity_domain = proximity_domain++;
512 }
513
514 if (!top_dev) {
515 ret = -ENOMEM;
516 goto err;
517 }
518
519 memcpy(top_dev->oem_id, crat_table->oem_id, CRAT_OEMID_LENGTH);
520 memcpy(top_dev->oem_table_id, crat_table->oem_table_id,
521 CRAT_OEMTABLEID_LENGTH);
522 top_dev->oem_revision = crat_table->oem_revision;
523
524 sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1);
525 while ((char *)sub_type_hdr + sizeof(struct crat_subtype_generic) <
526 ((char *)crat_image) + image_len) {
527 if (sub_type_hdr->flags & CRAT_SUBTYPE_FLAGS_ENABLED) {
528 ret = kfd_parse_subtype(sub_type_hdr, device_list);
529 if (ret)
530 break;
531 }
532
533 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
534 sub_type_hdr->length);
535 }
536
537err:
538 if (ret)
539 kfd_release_topology_device_list(device_list);
540
541 return ret;
542}
543
544/* Helper function. See kfd_fill_gpu_cache_info for parameter description */
545static int fill_in_pcache(struct crat_subtype_cache *pcache,
546 struct kfd_gpu_cache_info *pcache_info,
547 struct kfd_cu_info *cu_info,
548 int mem_available,
549 int cu_bitmask,
550 int cache_type, unsigned int cu_processor_id,
551 int cu_block)
552{
553 unsigned int cu_sibling_map_mask;
554 int first_active_cu;
555
556 /* First check if enough memory is available */
557 if (sizeof(struct crat_subtype_cache) > mem_available)
558 return -ENOMEM;
559
560 cu_sibling_map_mask = cu_bitmask;
561 cu_sibling_map_mask >>= cu_block;
562 cu_sibling_map_mask &=
563 ((1 << pcache_info[cache_type].num_cu_shared) - 1);
564 first_active_cu = ffs(cu_sibling_map_mask);
565
566 /* CU could be inactive. In case of shared cache find the first active
567 * CU. and incase of non-shared cache check if the CU is inactive. If
568 * inactive active skip it
569 */
570 if (first_active_cu) {
571 memset(pcache, 0, sizeof(struct crat_subtype_cache));
572 pcache->type = CRAT_SUBTYPE_CACHE_AFFINITY;
573 pcache->length = sizeof(struct crat_subtype_cache);
574 pcache->flags = pcache_info[cache_type].flags;
575 pcache->processor_id_low = cu_processor_id
576 + (first_active_cu - 1);
577 pcache->cache_level = pcache_info[cache_type].cache_level;
578 pcache->cache_size = pcache_info[cache_type].cache_size;
579
580 /* Sibling map is w.r.t processor_id_low, so shift out
581 * inactive CU
582 */
583 cu_sibling_map_mask =
584 cu_sibling_map_mask >> (first_active_cu - 1);
585
586 pcache->sibling_map[0] = (uint8_t)(cu_sibling_map_mask & 0xFF);
587 pcache->sibling_map[1] =
588 (uint8_t)((cu_sibling_map_mask >> 8) & 0xFF);
589 pcache->sibling_map[2] =
590 (uint8_t)((cu_sibling_map_mask >> 16) & 0xFF);
591 pcache->sibling_map[3] =
592 (uint8_t)((cu_sibling_map_mask >> 24) & 0xFF);
593 return 0;
594 }
595 return 1;
596}
597
598/* kfd_fill_gpu_cache_info - Fill GPU cache info using kfd_gpu_cache_info
599 * tables
600 *
601 * @kdev - [IN] GPU device
602 * @gpu_processor_id - [IN] GPU processor ID to which these caches
603 * associate
604 * @available_size - [IN] Amount of memory available in pcache
605 * @cu_info - [IN] Compute Unit info obtained from KGD
606 * @pcache - [OUT] memory into which cache data is to be filled in.
607 * @size_filled - [OUT] amount of data used up in pcache.
608 * @num_of_entries - [OUT] number of caches added
609 */
610static int kfd_fill_gpu_cache_info(struct kfd_dev *kdev,
611 int gpu_processor_id,
612 int available_size,
613 struct kfd_cu_info *cu_info,
614 struct crat_subtype_cache *pcache,
615 int *size_filled,
616 int *num_of_entries)
617{
618 struct kfd_gpu_cache_info *pcache_info;
619 int num_of_cache_types = 0;
620 int i, j, k;
621 int ct = 0;
622 int mem_available = available_size;
623 unsigned int cu_processor_id;
624 int ret;
625
626 switch (kdev->device_info->asic_family) {
627 case CHIP_KAVERI:
628 pcache_info = kaveri_cache_info;
629 num_of_cache_types = ARRAY_SIZE(kaveri_cache_info);
630 break;
631 case CHIP_HAWAII:
632 pcache_info = hawaii_cache_info;
633 num_of_cache_types = ARRAY_SIZE(hawaii_cache_info);
634 break;
635 case CHIP_CARRIZO:
636 pcache_info = carrizo_cache_info;
637 num_of_cache_types = ARRAY_SIZE(carrizo_cache_info);
638 break;
639 case CHIP_TONGA:
640 pcache_info = tonga_cache_info;
641 num_of_cache_types = ARRAY_SIZE(tonga_cache_info);
642 break;
643 case CHIP_FIJI:
644 pcache_info = fiji_cache_info;
645 num_of_cache_types = ARRAY_SIZE(fiji_cache_info);
646 break;
647 case CHIP_POLARIS10:
648 pcache_info = polaris10_cache_info;
649 num_of_cache_types = ARRAY_SIZE(polaris10_cache_info);
650 break;
651 case CHIP_POLARIS11:
652 pcache_info = polaris11_cache_info;
653 num_of_cache_types = ARRAY_SIZE(polaris11_cache_info);
654 break;
655 case CHIP_POLARIS12:
656 pcache_info = polaris12_cache_info;
657 num_of_cache_types = ARRAY_SIZE(polaris12_cache_info);
658 break;
659 case CHIP_VEGAM:
660 pcache_info = vegam_cache_info;
661 num_of_cache_types = ARRAY_SIZE(vegam_cache_info);
662 break;
663 case CHIP_VEGA10:
664 case CHIP_VEGA12:
665 case CHIP_VEGA20:
666 case CHIP_ARCTURUS:
667 pcache_info = vega10_cache_info;
668 num_of_cache_types = ARRAY_SIZE(vega10_cache_info);
669 break;
670 case CHIP_RAVEN:
671 pcache_info = raven_cache_info;
672 num_of_cache_types = ARRAY_SIZE(raven_cache_info);
673 break;
674 case CHIP_RENOIR:
675 pcache_info = renoir_cache_info;
676 num_of_cache_types = ARRAY_SIZE(renoir_cache_info);
677 break;
678 case CHIP_NAVI10:
679 case CHIP_NAVI12:
680 case CHIP_NAVI14:
681 case CHIP_SIENNA_CICHLID:
682 case CHIP_NAVY_FLOUNDER:
683 pcache_info = navi10_cache_info;
684 num_of_cache_types = ARRAY_SIZE(navi10_cache_info);
685 break;
686 default:
687 return -EINVAL;
688 }
689
690 *size_filled = 0;
691 *num_of_entries = 0;
692
693 /* For each type of cache listed in the kfd_gpu_cache_info table,
694 * go through all available Compute Units.
695 * The [i,j,k] loop will
696 * if kfd_gpu_cache_info.num_cu_shared = 1
697 * will parse through all available CU
698 * If (kfd_gpu_cache_info.num_cu_shared != 1)
699 * then it will consider only one CU from
700 * the shared unit
701 */
702
703 for (ct = 0; ct < num_of_cache_types; ct++) {
704 cu_processor_id = gpu_processor_id;
705 for (i = 0; i < cu_info->num_shader_engines; i++) {
706 for (j = 0; j < cu_info->num_shader_arrays_per_engine;
707 j++) {
708 for (k = 0; k < cu_info->num_cu_per_sh;
709 k += pcache_info[ct].num_cu_shared) {
710
711 ret = fill_in_pcache(pcache,
712 pcache_info,
713 cu_info,
714 mem_available,
715 cu_info->cu_bitmap[i % 4][j + i / 4],
716 ct,
717 cu_processor_id,
718 k);
719
720 if (ret < 0)
721 break;
722
723 if (!ret) {
724 pcache++;
725 (*num_of_entries)++;
726 mem_available -=
727 sizeof(*pcache);
728 (*size_filled) +=
729 sizeof(*pcache);
730 }
731
732 /* Move to next CU block */
733 cu_processor_id +=
734 pcache_info[ct].num_cu_shared;
735 }
736 }
737 }
738 }
739
740 pr_debug("Added [%d] GPU cache entries\n", *num_of_entries);
741
742 return 0;
743}
744
745/*
746 * kfd_create_crat_image_acpi - Allocates memory for CRAT image and
747 * copies CRAT from ACPI (if available).
748 * NOTE: Call kfd_destroy_crat_image to free CRAT image memory
749 *
750 * @crat_image: CRAT read from ACPI. If no CRAT in ACPI then
751 * crat_image will be NULL
752 * @size: [OUT] size of crat_image
753 *
754 * Return 0 if successful else return error code
755 */
756int kfd_create_crat_image_acpi(void **crat_image, size_t *size)
757{
758 struct acpi_table_header *crat_table;
759 acpi_status status;
760 void *pcrat_image;
761
762 if (!crat_image)
763 return -EINVAL;
764
765 *crat_image = NULL;
766
767 /* Fetch the CRAT table from ACPI */
768 status = acpi_get_table(CRAT_SIGNATURE, 0, &crat_table);
769 if (status == AE_NOT_FOUND) {
770 pr_warn("CRAT table not found\n");
771 return -ENODATA;
772 } else if (ACPI_FAILURE(status)) {
773 const char *err = acpi_format_exception(status);
774
775 pr_err("CRAT table error: %s\n", err);
776 return -EINVAL;
777 }
778
779 if (ignore_crat) {
780 pr_info("CRAT table disabled by module option\n");
781 return -ENODATA;
782 }
783
784 pcrat_image = kmemdup(crat_table, crat_table->length, GFP_KERNEL);
785 if (!pcrat_image)
786 return -ENOMEM;
787
788 *crat_image = pcrat_image;
789 *size = crat_table->length;
790
791 return 0;
792}
793
794/* Memory required to create Virtual CRAT.
795 * Since there is no easy way to predict the amount of memory required, the
796 * following amount are allocated for CPU and GPU Virtual CRAT. This is
797 * expected to cover all known conditions. But to be safe additional check
798 * is put in the code to ensure we don't overwrite.
799 */
800#define VCRAT_SIZE_FOR_CPU (2 * PAGE_SIZE)
801#define VCRAT_SIZE_FOR_GPU (4 * PAGE_SIZE)
802
803/* kfd_fill_cu_for_cpu - Fill in Compute info for the given CPU NUMA node
804 *
805 * @numa_node_id: CPU NUMA node id
806 * @avail_size: Available size in the memory
807 * @sub_type_hdr: Memory into which compute info will be filled in
808 *
809 * Return 0 if successful else return -ve value
810 */
811static int kfd_fill_cu_for_cpu(int numa_node_id, int *avail_size,
812 int proximity_domain,
813 struct crat_subtype_computeunit *sub_type_hdr)
814{
815 const struct cpumask *cpumask;
816
817 *avail_size -= sizeof(struct crat_subtype_computeunit);
818 if (*avail_size < 0)
819 return -ENOMEM;
820
821 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit));
822
823 /* Fill in subtype header data */
824 sub_type_hdr->type = CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY;
825 sub_type_hdr->length = sizeof(struct crat_subtype_computeunit);
826 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
827
828 cpumask = cpumask_of_node(numa_node_id);
829
830 /* Fill in CU data */
831 sub_type_hdr->flags |= CRAT_CU_FLAGS_CPU_PRESENT;
832 sub_type_hdr->proximity_domain = proximity_domain;
833 sub_type_hdr->processor_id_low = kfd_numa_node_to_apic_id(numa_node_id);
834 if (sub_type_hdr->processor_id_low == -1)
835 return -EINVAL;
836
837 sub_type_hdr->num_cpu_cores = cpumask_weight(cpumask);
838
839 return 0;
840}
841
842/* kfd_fill_mem_info_for_cpu - Fill in Memory info for the given CPU NUMA node
843 *
844 * @numa_node_id: CPU NUMA node id
845 * @avail_size: Available size in the memory
846 * @sub_type_hdr: Memory into which compute info will be filled in
847 *
848 * Return 0 if successful else return -ve value
849 */
850static int kfd_fill_mem_info_for_cpu(int numa_node_id, int *avail_size,
851 int proximity_domain,
852 struct crat_subtype_memory *sub_type_hdr)
853{
854 uint64_t mem_in_bytes = 0;
855 pg_data_t *pgdat;
856 int zone_type;
857
858 *avail_size -= sizeof(struct crat_subtype_memory);
859 if (*avail_size < 0)
860 return -ENOMEM;
861
862 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_memory));
863
864 /* Fill in subtype header data */
865 sub_type_hdr->type = CRAT_SUBTYPE_MEMORY_AFFINITY;
866 sub_type_hdr->length = sizeof(struct crat_subtype_memory);
867 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
868
869 /* Fill in Memory Subunit data */
870
871 /* Unlike si_meminfo, si_meminfo_node is not exported. So
872 * the following lines are duplicated from si_meminfo_node
873 * function
874 */
875 pgdat = NODE_DATA(numa_node_id);
876 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
877 mem_in_bytes += zone_managed_pages(&pgdat->node_zones[zone_type]);
878 mem_in_bytes <<= PAGE_SHIFT;
879
880 sub_type_hdr->length_low = lower_32_bits(mem_in_bytes);
881 sub_type_hdr->length_high = upper_32_bits(mem_in_bytes);
882 sub_type_hdr->proximity_domain = proximity_domain;
883
884 return 0;
885}
886
887#ifdef CONFIG_X86_64
888static int kfd_fill_iolink_info_for_cpu(int numa_node_id, int *avail_size,
889 uint32_t *num_entries,
890 struct crat_subtype_iolink *sub_type_hdr)
891{
892 int nid;
893 struct cpuinfo_x86 *c = &cpu_data(0);
894 uint8_t link_type;
895
896 if (c->x86_vendor == X86_VENDOR_AMD)
897 link_type = CRAT_IOLINK_TYPE_HYPERTRANSPORT;
898 else
899 link_type = CRAT_IOLINK_TYPE_QPI_1_1;
900
901 *num_entries = 0;
902
903 /* Create IO links from this node to other CPU nodes */
904 for_each_online_node(nid) {
905 if (nid == numa_node_id) /* node itself */
906 continue;
907
908 *avail_size -= sizeof(struct crat_subtype_iolink);
909 if (*avail_size < 0)
910 return -ENOMEM;
911
912 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
913
914 /* Fill in subtype header data */
915 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
916 sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
917 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
918
919 /* Fill in IO link data */
920 sub_type_hdr->proximity_domain_from = numa_node_id;
921 sub_type_hdr->proximity_domain_to = nid;
922 sub_type_hdr->io_interface_type = link_type;
923
924 (*num_entries)++;
925 sub_type_hdr++;
926 }
927
928 return 0;
929}
930#endif
931
932/* kfd_create_vcrat_image_cpu - Create Virtual CRAT for CPU
933 *
934 * @pcrat_image: Fill in VCRAT for CPU
935 * @size: [IN] allocated size of crat_image.
936 * [OUT] actual size of data filled in crat_image
937 */
938static int kfd_create_vcrat_image_cpu(void *pcrat_image, size_t *size)
939{
940 struct crat_header *crat_table = (struct crat_header *)pcrat_image;
941 struct acpi_table_header *acpi_table;
942 acpi_status status;
943 struct crat_subtype_generic *sub_type_hdr;
944 int avail_size = *size;
945 int numa_node_id;
946#ifdef CONFIG_X86_64
947 uint32_t entries = 0;
948#endif
949 int ret = 0;
950
951 if (!pcrat_image || avail_size < VCRAT_SIZE_FOR_CPU)
952 return -EINVAL;
953
954 /* Fill in CRAT Header.
955 * Modify length and total_entries as subunits are added.
956 */
957 avail_size -= sizeof(struct crat_header);
958 if (avail_size < 0)
959 return -ENOMEM;
960
961 memset(crat_table, 0, sizeof(struct crat_header));
962 memcpy(&crat_table->signature, CRAT_SIGNATURE,
963 sizeof(crat_table->signature));
964 crat_table->length = sizeof(struct crat_header);
965
966 status = acpi_get_table("DSDT", 0, &acpi_table);
967 if (status != AE_OK)
968 pr_warn("DSDT table not found for OEM information\n");
969 else {
970 crat_table->oem_revision = acpi_table->revision;
971 memcpy(crat_table->oem_id, acpi_table->oem_id,
972 CRAT_OEMID_LENGTH);
973 memcpy(crat_table->oem_table_id, acpi_table->oem_table_id,
974 CRAT_OEMTABLEID_LENGTH);
975 }
976 crat_table->total_entries = 0;
977 crat_table->num_domains = 0;
978
979 sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1);
980
981 for_each_online_node(numa_node_id) {
982 if (kfd_numa_node_to_apic_id(numa_node_id) == -1)
983 continue;
984
985 /* Fill in Subtype: Compute Unit */
986 ret = kfd_fill_cu_for_cpu(numa_node_id, &avail_size,
987 crat_table->num_domains,
988 (struct crat_subtype_computeunit *)sub_type_hdr);
989 if (ret < 0)
990 return ret;
991 crat_table->length += sub_type_hdr->length;
992 crat_table->total_entries++;
993
994 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
995 sub_type_hdr->length);
996
997 /* Fill in Subtype: Memory */
998 ret = kfd_fill_mem_info_for_cpu(numa_node_id, &avail_size,
999 crat_table->num_domains,
1000 (struct crat_subtype_memory *)sub_type_hdr);
1001 if (ret < 0)
1002 return ret;
1003 crat_table->length += sub_type_hdr->length;
1004 crat_table->total_entries++;
1005
1006 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1007 sub_type_hdr->length);
1008
1009 /* Fill in Subtype: IO Link */
1010#ifdef CONFIG_X86_64
1011 ret = kfd_fill_iolink_info_for_cpu(numa_node_id, &avail_size,
1012 &entries,
1013 (struct crat_subtype_iolink *)sub_type_hdr);
1014 if (ret < 0)
1015 return ret;
1016 crat_table->length += (sub_type_hdr->length * entries);
1017 crat_table->total_entries += entries;
1018
1019 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1020 sub_type_hdr->length * entries);
1021#else
1022 pr_info("IO link not available for non x86 platforms\n");
1023#endif
1024
1025 crat_table->num_domains++;
1026 }
1027
1028 /* TODO: Add cache Subtype for CPU.
1029 * Currently, CPU cache information is available in function
1030 * detect_cache_attributes(cpu) defined in the file
1031 * ./arch/x86/kernel/cpu/intel_cacheinfo.c. This function is not
1032 * exported and to get the same information the code needs to be
1033 * duplicated.
1034 */
1035
1036 *size = crat_table->length;
1037 pr_info("Virtual CRAT table created for CPU\n");
1038
1039 return 0;
1040}
1041
1042static int kfd_fill_gpu_memory_affinity(int *avail_size,
1043 struct kfd_dev *kdev, uint8_t type, uint64_t size,
1044 struct crat_subtype_memory *sub_type_hdr,
1045 uint32_t proximity_domain,
1046 const struct kfd_local_mem_info *local_mem_info)
1047{
1048 *avail_size -= sizeof(struct crat_subtype_memory);
1049 if (*avail_size < 0)
1050 return -ENOMEM;
1051
1052 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_memory));
1053 sub_type_hdr->type = CRAT_SUBTYPE_MEMORY_AFFINITY;
1054 sub_type_hdr->length = sizeof(struct crat_subtype_memory);
1055 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED;
1056
1057 sub_type_hdr->proximity_domain = proximity_domain;
1058
1059 pr_debug("Fill gpu memory affinity - type 0x%x size 0x%llx\n",
1060 type, size);
1061
1062 sub_type_hdr->length_low = lower_32_bits(size);
1063 sub_type_hdr->length_high = upper_32_bits(size);
1064
1065 sub_type_hdr->width = local_mem_info->vram_width;
1066 sub_type_hdr->visibility_type = type;
1067
1068 return 0;
1069}
1070
1071/* kfd_fill_gpu_direct_io_link - Fill in direct io link from GPU
1072 * to its NUMA node
1073 * @avail_size: Available size in the memory
1074 * @kdev - [IN] GPU device
1075 * @sub_type_hdr: Memory into which io link info will be filled in
1076 * @proximity_domain - proximity domain of the GPU node
1077 *
1078 * Return 0 if successful else return -ve value
1079 */
1080static int kfd_fill_gpu_direct_io_link_to_cpu(int *avail_size,
1081 struct kfd_dev *kdev,
1082 struct crat_subtype_iolink *sub_type_hdr,
1083 uint32_t proximity_domain)
1084{
1085 *avail_size -= sizeof(struct crat_subtype_iolink);
1086 if (*avail_size < 0)
1087 return -ENOMEM;
1088
1089 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
1090
1091 /* Fill in subtype header data */
1092 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
1093 sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
1094 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED;
1095 if (kfd_dev_is_large_bar(kdev))
1096 sub_type_hdr->flags |= CRAT_IOLINK_FLAGS_BI_DIRECTIONAL;
1097
1098 /* Fill in IOLINK subtype.
1099 * TODO: Fill-in other fields of iolink subtype
1100 */
1101 sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_PCIEXPRESS;
1102 sub_type_hdr->proximity_domain_from = proximity_domain;
1103#ifdef CONFIG_NUMA
1104 if (kdev->pdev->dev.numa_node == NUMA_NO_NODE)
1105 sub_type_hdr->proximity_domain_to = 0;
1106 else
1107 sub_type_hdr->proximity_domain_to = kdev->pdev->dev.numa_node;
1108#else
1109 sub_type_hdr->proximity_domain_to = 0;
1110#endif
1111 return 0;
1112}
1113
1114static int kfd_fill_gpu_xgmi_link_to_gpu(int *avail_size,
1115 struct kfd_dev *kdev,
1116 struct kfd_dev *peer_kdev,
1117 struct crat_subtype_iolink *sub_type_hdr,
1118 uint32_t proximity_domain_from,
1119 uint32_t proximity_domain_to)
1120{
1121 *avail_size -= sizeof(struct crat_subtype_iolink);
1122 if (*avail_size < 0)
1123 return -ENOMEM;
1124
1125 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
1126
1127 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
1128 sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
1129 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED |
1130 CRAT_IOLINK_FLAGS_BI_DIRECTIONAL;
1131
1132 sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_XGMI;
1133 sub_type_hdr->proximity_domain_from = proximity_domain_from;
1134 sub_type_hdr->proximity_domain_to = proximity_domain_to;
1135 sub_type_hdr->num_hops_xgmi =
1136 amdgpu_amdkfd_get_xgmi_hops_count(kdev->kgd, peer_kdev->kgd);
1137 return 0;
1138}
1139
1140/* kfd_create_vcrat_image_gpu - Create Virtual CRAT for CPU
1141 *
1142 * @pcrat_image: Fill in VCRAT for GPU
1143 * @size: [IN] allocated size of crat_image.
1144 * [OUT] actual size of data filled in crat_image
1145 */
1146static int kfd_create_vcrat_image_gpu(void *pcrat_image,
1147 size_t *size, struct kfd_dev *kdev,
1148 uint32_t proximity_domain)
1149{
1150 struct crat_header *crat_table = (struct crat_header *)pcrat_image;
1151 struct crat_subtype_generic *sub_type_hdr;
1152 struct kfd_local_mem_info local_mem_info;
1153 struct kfd_topology_device *peer_dev;
1154 struct crat_subtype_computeunit *cu;
1155 struct kfd_cu_info cu_info;
1156 int avail_size = *size;
1157 uint32_t total_num_of_cu;
1158 int num_of_cache_entries = 0;
1159 int cache_mem_filled = 0;
1160 uint32_t nid = 0;
1161 int ret = 0;
1162
1163 if (!pcrat_image || avail_size < VCRAT_SIZE_FOR_GPU)
1164 return -EINVAL;
1165
1166 /* Fill the CRAT Header.
1167 * Modify length and total_entries as subunits are added.
1168 */
1169 avail_size -= sizeof(struct crat_header);
1170 if (avail_size < 0)
1171 return -ENOMEM;
1172
1173 memset(crat_table, 0, sizeof(struct crat_header));
1174
1175 memcpy(&crat_table->signature, CRAT_SIGNATURE,
1176 sizeof(crat_table->signature));
1177 /* Change length as we add more subtypes*/
1178 crat_table->length = sizeof(struct crat_header);
1179 crat_table->num_domains = 1;
1180 crat_table->total_entries = 0;
1181
1182 /* Fill in Subtype: Compute Unit
1183 * First fill in the sub type header and then sub type data
1184 */
1185 avail_size -= sizeof(struct crat_subtype_computeunit);
1186 if (avail_size < 0)
1187 return -ENOMEM;
1188
1189 sub_type_hdr = (struct crat_subtype_generic *)(crat_table + 1);
1190 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit));
1191
1192 sub_type_hdr->type = CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY;
1193 sub_type_hdr->length = sizeof(struct crat_subtype_computeunit);
1194 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
1195
1196 /* Fill CU subtype data */
1197 cu = (struct crat_subtype_computeunit *)sub_type_hdr;
1198 cu->flags |= CRAT_CU_FLAGS_GPU_PRESENT;
1199 cu->proximity_domain = proximity_domain;
1200
1201 amdgpu_amdkfd_get_cu_info(kdev->kgd, &cu_info);
1202 cu->num_simd_per_cu = cu_info.simd_per_cu;
1203 cu->num_simd_cores = cu_info.simd_per_cu * cu_info.cu_active_number;
1204 cu->max_waves_simd = cu_info.max_waves_per_simd;
1205
1206 cu->wave_front_size = cu_info.wave_front_size;
1207 cu->array_count = cu_info.num_shader_arrays_per_engine *
1208 cu_info.num_shader_engines;
1209 total_num_of_cu = (cu->array_count * cu_info.num_cu_per_sh);
1210 cu->processor_id_low = get_and_inc_gpu_processor_id(total_num_of_cu);
1211 cu->num_cu_per_array = cu_info.num_cu_per_sh;
1212 cu->max_slots_scatch_cu = cu_info.max_scratch_slots_per_cu;
1213 cu->num_banks = cu_info.num_shader_engines;
1214 cu->lds_size_in_kb = cu_info.lds_size;
1215
1216 cu->hsa_capability = 0;
1217
1218 /* Check if this node supports IOMMU. During parsing this flag will
1219 * translate to HSA_CAP_ATS_PRESENT
1220 */
1221 if (!kfd_iommu_check_device(kdev))
1222 cu->hsa_capability |= CRAT_CU_FLAGS_IOMMU_PRESENT;
1223
1224 crat_table->length += sub_type_hdr->length;
1225 crat_table->total_entries++;
1226
1227 /* Fill in Subtype: Memory. Only on systems with large BAR (no
1228 * private FB), report memory as public. On other systems
1229 * report the total FB size (public+private) as a single
1230 * private heap.
1231 */
1232 amdgpu_amdkfd_get_local_mem_info(kdev->kgd, &local_mem_info);
1233 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1234 sub_type_hdr->length);
1235
1236 if (debug_largebar)
1237 local_mem_info.local_mem_size_private = 0;
1238
1239 if (local_mem_info.local_mem_size_private == 0)
1240 ret = kfd_fill_gpu_memory_affinity(&avail_size,
1241 kdev, HSA_MEM_HEAP_TYPE_FB_PUBLIC,
1242 local_mem_info.local_mem_size_public,
1243 (struct crat_subtype_memory *)sub_type_hdr,
1244 proximity_domain,
1245 &local_mem_info);
1246 else
1247 ret = kfd_fill_gpu_memory_affinity(&avail_size,
1248 kdev, HSA_MEM_HEAP_TYPE_FB_PRIVATE,
1249 local_mem_info.local_mem_size_public +
1250 local_mem_info.local_mem_size_private,
1251 (struct crat_subtype_memory *)sub_type_hdr,
1252 proximity_domain,
1253 &local_mem_info);
1254 if (ret < 0)
1255 return ret;
1256
1257 crat_table->length += sizeof(struct crat_subtype_memory);
1258 crat_table->total_entries++;
1259
1260 /* TODO: Fill in cache information. This information is NOT readily
1261 * available in KGD
1262 */
1263 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1264 sub_type_hdr->length);
1265 ret = kfd_fill_gpu_cache_info(kdev, cu->processor_id_low,
1266 avail_size,
1267 &cu_info,
1268 (struct crat_subtype_cache *)sub_type_hdr,
1269 &cache_mem_filled,
1270 &num_of_cache_entries);
1271
1272 if (ret < 0)
1273 return ret;
1274
1275 crat_table->length += cache_mem_filled;
1276 crat_table->total_entries += num_of_cache_entries;
1277 avail_size -= cache_mem_filled;
1278
1279 /* Fill in Subtype: IO_LINKS
1280 * Only direct links are added here which is Link from GPU to
1281 * to its NUMA node. Indirect links are added by userspace.
1282 */
1283 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1284 cache_mem_filled);
1285 ret = kfd_fill_gpu_direct_io_link_to_cpu(&avail_size, kdev,
1286 (struct crat_subtype_iolink *)sub_type_hdr, proximity_domain);
1287
1288 if (ret < 0)
1289 return ret;
1290
1291 crat_table->length += sub_type_hdr->length;
1292 crat_table->total_entries++;
1293
1294
1295 /* Fill in Subtype: IO_LINKS
1296 * Direct links from GPU to other GPUs through xGMI.
1297 * We will loop GPUs that already be processed (with lower value
1298 * of proximity_domain), add the link for the GPUs with same
1299 * hive id (from this GPU to other GPU) . The reversed iolink
1300 * (from other GPU to this GPU) will be added
1301 * in kfd_parse_subtype_iolink.
1302 */
1303 if (kdev->hive_id) {
1304 for (nid = 0; nid < proximity_domain; ++nid) {
1305 peer_dev = kfd_topology_device_by_proximity_domain(nid);
1306 if (!peer_dev->gpu)
1307 continue;
1308 if (peer_dev->gpu->hive_id != kdev->hive_id)
1309 continue;
1310 sub_type_hdr = (typeof(sub_type_hdr))(
1311 (char *)sub_type_hdr +
1312 sizeof(struct crat_subtype_iolink));
1313 ret = kfd_fill_gpu_xgmi_link_to_gpu(
1314 &avail_size, kdev, peer_dev->gpu,
1315 (struct crat_subtype_iolink *)sub_type_hdr,
1316 proximity_domain, nid);
1317 if (ret < 0)
1318 return ret;
1319 crat_table->length += sub_type_hdr->length;
1320 crat_table->total_entries++;
1321 }
1322 }
1323 *size = crat_table->length;
1324 pr_info("Virtual CRAT table created for GPU\n");
1325
1326 return ret;
1327}
1328
1329/* kfd_create_crat_image_virtual - Allocates memory for CRAT image and
1330 * creates a Virtual CRAT (VCRAT) image
1331 *
1332 * NOTE: Call kfd_destroy_crat_image to free CRAT image memory
1333 *
1334 * @crat_image: VCRAT image created because ACPI does not have a
1335 * CRAT for this device
1336 * @size: [OUT] size of virtual crat_image
1337 * @flags: COMPUTE_UNIT_CPU - Create VCRAT for CPU device
1338 * COMPUTE_UNIT_GPU - Create VCRAT for GPU
1339 * (COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU) - Create VCRAT for APU
1340 * -- this option is not currently implemented.
1341 * The assumption is that all AMD APUs will have CRAT
1342 * @kdev: Valid kfd_device required if flags contain COMPUTE_UNIT_GPU
1343 *
1344 * Return 0 if successful else return -ve value
1345 */
1346int kfd_create_crat_image_virtual(void **crat_image, size_t *size,
1347 int flags, struct kfd_dev *kdev,
1348 uint32_t proximity_domain)
1349{
1350 void *pcrat_image = NULL;
1351 int ret = 0;
1352
1353 if (!crat_image)
1354 return -EINVAL;
1355
1356 *crat_image = NULL;
1357
1358 /* Allocate one VCRAT_SIZE_FOR_CPU for CPU virtual CRAT image and
1359 * VCRAT_SIZE_FOR_GPU for GPU virtual CRAT image. This should cover
1360 * all the current conditions. A check is put not to overwrite beyond
1361 * allocated size
1362 */
1363 switch (flags) {
1364 case COMPUTE_UNIT_CPU:
1365 pcrat_image = kmalloc(VCRAT_SIZE_FOR_CPU, GFP_KERNEL);
1366 if (!pcrat_image)
1367 return -ENOMEM;
1368 *size = VCRAT_SIZE_FOR_CPU;
1369 ret = kfd_create_vcrat_image_cpu(pcrat_image, size);
1370 break;
1371 case COMPUTE_UNIT_GPU:
1372 if (!kdev)
1373 return -EINVAL;
1374 pcrat_image = kmalloc(VCRAT_SIZE_FOR_GPU, GFP_KERNEL);
1375 if (!pcrat_image)
1376 return -ENOMEM;
1377 *size = VCRAT_SIZE_FOR_GPU;
1378 ret = kfd_create_vcrat_image_gpu(pcrat_image, size, kdev,
1379 proximity_domain);
1380 break;
1381 case (COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU):
1382 /* TODO: */
1383 ret = -EINVAL;
1384 pr_err("VCRAT not implemented for APU\n");
1385 break;
1386 default:
1387 ret = -EINVAL;
1388 }
1389
1390 if (!ret)
1391 *crat_image = pcrat_image;
1392 else
1393 kfree(pcrat_image);
1394
1395 return ret;
1396}
1397
1398
1399/* kfd_destroy_crat_image
1400 *
1401 * @crat_image: [IN] - crat_image from kfd_create_crat_image_xxx(..)
1402 *
1403 */
1404void kfd_destroy_crat_image(void *crat_image)
1405{
1406 kfree(crat_image);
1407}
1// SPDX-License-Identifier: GPL-2.0 OR MIT
2/*
3 * Copyright 2015-2022 Advanced Micro Devices, Inc.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice shall be included in
13 * all copies or substantial portions of the Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21 * OTHER DEALINGS IN THE SOFTWARE.
22 */
23
24#include <linux/pci.h>
25#include <linux/acpi.h>
26#include "kfd_crat.h"
27#include "kfd_priv.h"
28#include "kfd_topology.h"
29#include "amdgpu.h"
30#include "amdgpu_amdkfd.h"
31
32/* GPU Processor ID base for dGPUs for which VCRAT needs to be created.
33 * GPU processor ID are expressed with Bit[31]=1.
34 * The base is set to 0x8000_0000 + 0x1000 to avoid collision with GPU IDs
35 * used in the CRAT.
36 */
37static uint32_t gpu_processor_id_low = 0x80001000;
38
39/* Return the next available gpu_processor_id and increment it for next GPU
40 * @total_cu_count - Total CUs present in the GPU including ones
41 * masked off
42 */
43static inline unsigned int get_and_inc_gpu_processor_id(
44 unsigned int total_cu_count)
45{
46 int current_id = gpu_processor_id_low;
47
48 gpu_processor_id_low += total_cu_count;
49 return current_id;
50}
51
52
53static struct kfd_gpu_cache_info kaveri_cache_info[] = {
54 {
55 /* TCP L1 Cache per CU */
56 .cache_size = 16,
57 .cache_level = 1,
58 .flags = (CRAT_CACHE_FLAGS_ENABLED |
59 CRAT_CACHE_FLAGS_DATA_CACHE |
60 CRAT_CACHE_FLAGS_SIMD_CACHE),
61 .num_cu_shared = 1,
62 },
63 {
64 /* Scalar L1 Instruction Cache (in SQC module) per bank */
65 .cache_size = 16,
66 .cache_level = 1,
67 .flags = (CRAT_CACHE_FLAGS_ENABLED |
68 CRAT_CACHE_FLAGS_INST_CACHE |
69 CRAT_CACHE_FLAGS_SIMD_CACHE),
70 .num_cu_shared = 2,
71 },
72 {
73 /* Scalar L1 Data Cache (in SQC module) per bank */
74 .cache_size = 8,
75 .cache_level = 1,
76 .flags = (CRAT_CACHE_FLAGS_ENABLED |
77 CRAT_CACHE_FLAGS_DATA_CACHE |
78 CRAT_CACHE_FLAGS_SIMD_CACHE),
79 .num_cu_shared = 2,
80 },
81
82 /* TODO: Add L2 Cache information */
83};
84
85
86static struct kfd_gpu_cache_info carrizo_cache_info[] = {
87 {
88 /* TCP L1 Cache per CU */
89 .cache_size = 16,
90 .cache_level = 1,
91 .flags = (CRAT_CACHE_FLAGS_ENABLED |
92 CRAT_CACHE_FLAGS_DATA_CACHE |
93 CRAT_CACHE_FLAGS_SIMD_CACHE),
94 .num_cu_shared = 1,
95 },
96 {
97 /* Scalar L1 Instruction Cache (in SQC module) per bank */
98 .cache_size = 8,
99 .cache_level = 1,
100 .flags = (CRAT_CACHE_FLAGS_ENABLED |
101 CRAT_CACHE_FLAGS_INST_CACHE |
102 CRAT_CACHE_FLAGS_SIMD_CACHE),
103 .num_cu_shared = 4,
104 },
105 {
106 /* Scalar L1 Data Cache (in SQC module) per bank. */
107 .cache_size = 4,
108 .cache_level = 1,
109 .flags = (CRAT_CACHE_FLAGS_ENABLED |
110 CRAT_CACHE_FLAGS_DATA_CACHE |
111 CRAT_CACHE_FLAGS_SIMD_CACHE),
112 .num_cu_shared = 4,
113 },
114
115 /* TODO: Add L2 Cache information */
116};
117
118#define hawaii_cache_info kaveri_cache_info
119#define tonga_cache_info carrizo_cache_info
120#define fiji_cache_info carrizo_cache_info
121#define polaris10_cache_info carrizo_cache_info
122#define polaris11_cache_info carrizo_cache_info
123#define polaris12_cache_info carrizo_cache_info
124#define vegam_cache_info carrizo_cache_info
125
126/* NOTE: L1 cache information has been updated and L2/L3
127 * cache information has been added for Vega10 and
128 * newer ASICs. The unit for cache_size is KiB.
129 * In future, check & update cache details
130 * for every new ASIC is required.
131 */
132
133static struct kfd_gpu_cache_info vega10_cache_info[] = {
134 {
135 /* TCP L1 Cache per CU */
136 .cache_size = 16,
137 .cache_level = 1,
138 .flags = (CRAT_CACHE_FLAGS_ENABLED |
139 CRAT_CACHE_FLAGS_DATA_CACHE |
140 CRAT_CACHE_FLAGS_SIMD_CACHE),
141 .num_cu_shared = 1,
142 },
143 {
144 /* Scalar L1 Instruction Cache per SQC */
145 .cache_size = 32,
146 .cache_level = 1,
147 .flags = (CRAT_CACHE_FLAGS_ENABLED |
148 CRAT_CACHE_FLAGS_INST_CACHE |
149 CRAT_CACHE_FLAGS_SIMD_CACHE),
150 .num_cu_shared = 3,
151 },
152 {
153 /* Scalar L1 Data Cache per SQC */
154 .cache_size = 16,
155 .cache_level = 1,
156 .flags = (CRAT_CACHE_FLAGS_ENABLED |
157 CRAT_CACHE_FLAGS_DATA_CACHE |
158 CRAT_CACHE_FLAGS_SIMD_CACHE),
159 .num_cu_shared = 3,
160 },
161 {
162 /* L2 Data Cache per GPU (Total Tex Cache) */
163 .cache_size = 4096,
164 .cache_level = 2,
165 .flags = (CRAT_CACHE_FLAGS_ENABLED |
166 CRAT_CACHE_FLAGS_DATA_CACHE |
167 CRAT_CACHE_FLAGS_SIMD_CACHE),
168 .num_cu_shared = 16,
169 },
170};
171
172static struct kfd_gpu_cache_info raven_cache_info[] = {
173 {
174 /* TCP L1 Cache per CU */
175 .cache_size = 16,
176 .cache_level = 1,
177 .flags = (CRAT_CACHE_FLAGS_ENABLED |
178 CRAT_CACHE_FLAGS_DATA_CACHE |
179 CRAT_CACHE_FLAGS_SIMD_CACHE),
180 .num_cu_shared = 1,
181 },
182 {
183 /* Scalar L1 Instruction Cache per SQC */
184 .cache_size = 32,
185 .cache_level = 1,
186 .flags = (CRAT_CACHE_FLAGS_ENABLED |
187 CRAT_CACHE_FLAGS_INST_CACHE |
188 CRAT_CACHE_FLAGS_SIMD_CACHE),
189 .num_cu_shared = 3,
190 },
191 {
192 /* Scalar L1 Data Cache per SQC */
193 .cache_size = 16,
194 .cache_level = 1,
195 .flags = (CRAT_CACHE_FLAGS_ENABLED |
196 CRAT_CACHE_FLAGS_DATA_CACHE |
197 CRAT_CACHE_FLAGS_SIMD_CACHE),
198 .num_cu_shared = 3,
199 },
200 {
201 /* L2 Data Cache per GPU (Total Tex Cache) */
202 .cache_size = 1024,
203 .cache_level = 2,
204 .flags = (CRAT_CACHE_FLAGS_ENABLED |
205 CRAT_CACHE_FLAGS_DATA_CACHE |
206 CRAT_CACHE_FLAGS_SIMD_CACHE),
207 .num_cu_shared = 11,
208 },
209};
210
211static struct kfd_gpu_cache_info renoir_cache_info[] = {
212 {
213 /* TCP L1 Cache per CU */
214 .cache_size = 16,
215 .cache_level = 1,
216 .flags = (CRAT_CACHE_FLAGS_ENABLED |
217 CRAT_CACHE_FLAGS_DATA_CACHE |
218 CRAT_CACHE_FLAGS_SIMD_CACHE),
219 .num_cu_shared = 1,
220 },
221 {
222 /* Scalar L1 Instruction Cache per SQC */
223 .cache_size = 32,
224 .cache_level = 1,
225 .flags = (CRAT_CACHE_FLAGS_ENABLED |
226 CRAT_CACHE_FLAGS_INST_CACHE |
227 CRAT_CACHE_FLAGS_SIMD_CACHE),
228 .num_cu_shared = 3,
229 },
230 {
231 /* Scalar L1 Data Cache per SQC */
232 .cache_size = 16,
233 .cache_level = 1,
234 .flags = (CRAT_CACHE_FLAGS_ENABLED |
235 CRAT_CACHE_FLAGS_DATA_CACHE |
236 CRAT_CACHE_FLAGS_SIMD_CACHE),
237 .num_cu_shared = 3,
238 },
239 {
240 /* L2 Data Cache per GPU (Total Tex Cache) */
241 .cache_size = 1024,
242 .cache_level = 2,
243 .flags = (CRAT_CACHE_FLAGS_ENABLED |
244 CRAT_CACHE_FLAGS_DATA_CACHE |
245 CRAT_CACHE_FLAGS_SIMD_CACHE),
246 .num_cu_shared = 8,
247 },
248};
249
250static struct kfd_gpu_cache_info vega12_cache_info[] = {
251 {
252 /* TCP L1 Cache per CU */
253 .cache_size = 16,
254 .cache_level = 1,
255 .flags = (CRAT_CACHE_FLAGS_ENABLED |
256 CRAT_CACHE_FLAGS_DATA_CACHE |
257 CRAT_CACHE_FLAGS_SIMD_CACHE),
258 .num_cu_shared = 1,
259 },
260 {
261 /* Scalar L1 Instruction Cache per SQC */
262 .cache_size = 32,
263 .cache_level = 1,
264 .flags = (CRAT_CACHE_FLAGS_ENABLED |
265 CRAT_CACHE_FLAGS_INST_CACHE |
266 CRAT_CACHE_FLAGS_SIMD_CACHE),
267 .num_cu_shared = 3,
268 },
269 {
270 /* Scalar L1 Data Cache per SQC */
271 .cache_size = 16,
272 .cache_level = 1,
273 .flags = (CRAT_CACHE_FLAGS_ENABLED |
274 CRAT_CACHE_FLAGS_DATA_CACHE |
275 CRAT_CACHE_FLAGS_SIMD_CACHE),
276 .num_cu_shared = 3,
277 },
278 {
279 /* L2 Data Cache per GPU (Total Tex Cache) */
280 .cache_size = 2048,
281 .cache_level = 2,
282 .flags = (CRAT_CACHE_FLAGS_ENABLED |
283 CRAT_CACHE_FLAGS_DATA_CACHE |
284 CRAT_CACHE_FLAGS_SIMD_CACHE),
285 .num_cu_shared = 5,
286 },
287};
288
289static struct kfd_gpu_cache_info vega20_cache_info[] = {
290 {
291 /* TCP L1 Cache per CU */
292 .cache_size = 16,
293 .cache_level = 1,
294 .flags = (CRAT_CACHE_FLAGS_ENABLED |
295 CRAT_CACHE_FLAGS_DATA_CACHE |
296 CRAT_CACHE_FLAGS_SIMD_CACHE),
297 .num_cu_shared = 1,
298 },
299 {
300 /* Scalar L1 Instruction Cache per SQC */
301 .cache_size = 32,
302 .cache_level = 1,
303 .flags = (CRAT_CACHE_FLAGS_ENABLED |
304 CRAT_CACHE_FLAGS_INST_CACHE |
305 CRAT_CACHE_FLAGS_SIMD_CACHE),
306 .num_cu_shared = 3,
307 },
308 {
309 /* Scalar L1 Data Cache per SQC */
310 .cache_size = 16,
311 .cache_level = 1,
312 .flags = (CRAT_CACHE_FLAGS_ENABLED |
313 CRAT_CACHE_FLAGS_DATA_CACHE |
314 CRAT_CACHE_FLAGS_SIMD_CACHE),
315 .num_cu_shared = 3,
316 },
317 {
318 /* L2 Data Cache per GPU (Total Tex Cache) */
319 .cache_size = 8192,
320 .cache_level = 2,
321 .flags = (CRAT_CACHE_FLAGS_ENABLED |
322 CRAT_CACHE_FLAGS_DATA_CACHE |
323 CRAT_CACHE_FLAGS_SIMD_CACHE),
324 .num_cu_shared = 16,
325 },
326};
327
328static struct kfd_gpu_cache_info aldebaran_cache_info[] = {
329 {
330 /* TCP L1 Cache per CU */
331 .cache_size = 16,
332 .cache_level = 1,
333 .flags = (CRAT_CACHE_FLAGS_ENABLED |
334 CRAT_CACHE_FLAGS_DATA_CACHE |
335 CRAT_CACHE_FLAGS_SIMD_CACHE),
336 .num_cu_shared = 1,
337 },
338 {
339 /* Scalar L1 Instruction Cache per SQC */
340 .cache_size = 32,
341 .cache_level = 1,
342 .flags = (CRAT_CACHE_FLAGS_ENABLED |
343 CRAT_CACHE_FLAGS_INST_CACHE |
344 CRAT_CACHE_FLAGS_SIMD_CACHE),
345 .num_cu_shared = 2,
346 },
347 {
348 /* Scalar L1 Data Cache per SQC */
349 .cache_size = 16,
350 .cache_level = 1,
351 .flags = (CRAT_CACHE_FLAGS_ENABLED |
352 CRAT_CACHE_FLAGS_DATA_CACHE |
353 CRAT_CACHE_FLAGS_SIMD_CACHE),
354 .num_cu_shared = 2,
355 },
356 {
357 /* L2 Data Cache per GPU (Total Tex Cache) */
358 .cache_size = 8192,
359 .cache_level = 2,
360 .flags = (CRAT_CACHE_FLAGS_ENABLED |
361 CRAT_CACHE_FLAGS_DATA_CACHE |
362 CRAT_CACHE_FLAGS_SIMD_CACHE),
363 .num_cu_shared = 14,
364 },
365};
366
367static struct kfd_gpu_cache_info navi10_cache_info[] = {
368 {
369 /* TCP L1 Cache per CU */
370 .cache_size = 16,
371 .cache_level = 1,
372 .flags = (CRAT_CACHE_FLAGS_ENABLED |
373 CRAT_CACHE_FLAGS_DATA_CACHE |
374 CRAT_CACHE_FLAGS_SIMD_CACHE),
375 .num_cu_shared = 1,
376 },
377 {
378 /* Scalar L1 Instruction Cache per SQC */
379 .cache_size = 32,
380 .cache_level = 1,
381 .flags = (CRAT_CACHE_FLAGS_ENABLED |
382 CRAT_CACHE_FLAGS_INST_CACHE |
383 CRAT_CACHE_FLAGS_SIMD_CACHE),
384 .num_cu_shared = 2,
385 },
386 {
387 /* Scalar L1 Data Cache per SQC */
388 .cache_size = 16,
389 .cache_level = 1,
390 .flags = (CRAT_CACHE_FLAGS_ENABLED |
391 CRAT_CACHE_FLAGS_DATA_CACHE |
392 CRAT_CACHE_FLAGS_SIMD_CACHE),
393 .num_cu_shared = 2,
394 },
395 {
396 /* GL1 Data Cache per SA */
397 .cache_size = 128,
398 .cache_level = 1,
399 .flags = (CRAT_CACHE_FLAGS_ENABLED |
400 CRAT_CACHE_FLAGS_DATA_CACHE |
401 CRAT_CACHE_FLAGS_SIMD_CACHE),
402 .num_cu_shared = 10,
403 },
404 {
405 /* L2 Data Cache per GPU (Total Tex Cache) */
406 .cache_size = 4096,
407 .cache_level = 2,
408 .flags = (CRAT_CACHE_FLAGS_ENABLED |
409 CRAT_CACHE_FLAGS_DATA_CACHE |
410 CRAT_CACHE_FLAGS_SIMD_CACHE),
411 .num_cu_shared = 10,
412 },
413};
414
415static struct kfd_gpu_cache_info vangogh_cache_info[] = {
416 {
417 /* TCP L1 Cache per CU */
418 .cache_size = 16,
419 .cache_level = 1,
420 .flags = (CRAT_CACHE_FLAGS_ENABLED |
421 CRAT_CACHE_FLAGS_DATA_CACHE |
422 CRAT_CACHE_FLAGS_SIMD_CACHE),
423 .num_cu_shared = 1,
424 },
425 {
426 /* Scalar L1 Instruction Cache per SQC */
427 .cache_size = 32,
428 .cache_level = 1,
429 .flags = (CRAT_CACHE_FLAGS_ENABLED |
430 CRAT_CACHE_FLAGS_INST_CACHE |
431 CRAT_CACHE_FLAGS_SIMD_CACHE),
432 .num_cu_shared = 2,
433 },
434 {
435 /* Scalar L1 Data Cache per SQC */
436 .cache_size = 16,
437 .cache_level = 1,
438 .flags = (CRAT_CACHE_FLAGS_ENABLED |
439 CRAT_CACHE_FLAGS_DATA_CACHE |
440 CRAT_CACHE_FLAGS_SIMD_CACHE),
441 .num_cu_shared = 2,
442 },
443 {
444 /* GL1 Data Cache per SA */
445 .cache_size = 128,
446 .cache_level = 1,
447 .flags = (CRAT_CACHE_FLAGS_ENABLED |
448 CRAT_CACHE_FLAGS_DATA_CACHE |
449 CRAT_CACHE_FLAGS_SIMD_CACHE),
450 .num_cu_shared = 8,
451 },
452 {
453 /* L2 Data Cache per GPU (Total Tex Cache) */
454 .cache_size = 1024,
455 .cache_level = 2,
456 .flags = (CRAT_CACHE_FLAGS_ENABLED |
457 CRAT_CACHE_FLAGS_DATA_CACHE |
458 CRAT_CACHE_FLAGS_SIMD_CACHE),
459 .num_cu_shared = 8,
460 },
461};
462
463static struct kfd_gpu_cache_info navi14_cache_info[] = {
464 {
465 /* TCP L1 Cache per CU */
466 .cache_size = 16,
467 .cache_level = 1,
468 .flags = (CRAT_CACHE_FLAGS_ENABLED |
469 CRAT_CACHE_FLAGS_DATA_CACHE |
470 CRAT_CACHE_FLAGS_SIMD_CACHE),
471 .num_cu_shared = 1,
472 },
473 {
474 /* Scalar L1 Instruction Cache per SQC */
475 .cache_size = 32,
476 .cache_level = 1,
477 .flags = (CRAT_CACHE_FLAGS_ENABLED |
478 CRAT_CACHE_FLAGS_INST_CACHE |
479 CRAT_CACHE_FLAGS_SIMD_CACHE),
480 .num_cu_shared = 2,
481 },
482 {
483 /* Scalar L1 Data Cache per SQC */
484 .cache_size = 16,
485 .cache_level = 1,
486 .flags = (CRAT_CACHE_FLAGS_ENABLED |
487 CRAT_CACHE_FLAGS_DATA_CACHE |
488 CRAT_CACHE_FLAGS_SIMD_CACHE),
489 .num_cu_shared = 2,
490 },
491 {
492 /* GL1 Data Cache per SA */
493 .cache_size = 128,
494 .cache_level = 1,
495 .flags = (CRAT_CACHE_FLAGS_ENABLED |
496 CRAT_CACHE_FLAGS_DATA_CACHE |
497 CRAT_CACHE_FLAGS_SIMD_CACHE),
498 .num_cu_shared = 12,
499 },
500 {
501 /* L2 Data Cache per GPU (Total Tex Cache) */
502 .cache_size = 2048,
503 .cache_level = 2,
504 .flags = (CRAT_CACHE_FLAGS_ENABLED |
505 CRAT_CACHE_FLAGS_DATA_CACHE |
506 CRAT_CACHE_FLAGS_SIMD_CACHE),
507 .num_cu_shared = 12,
508 },
509};
510
511static struct kfd_gpu_cache_info sienna_cichlid_cache_info[] = {
512 {
513 /* TCP L1 Cache per CU */
514 .cache_size = 16,
515 .cache_level = 1,
516 .flags = (CRAT_CACHE_FLAGS_ENABLED |
517 CRAT_CACHE_FLAGS_DATA_CACHE |
518 CRAT_CACHE_FLAGS_SIMD_CACHE),
519 .num_cu_shared = 1,
520 },
521 {
522 /* Scalar L1 Instruction Cache per SQC */
523 .cache_size = 32,
524 .cache_level = 1,
525 .flags = (CRAT_CACHE_FLAGS_ENABLED |
526 CRAT_CACHE_FLAGS_INST_CACHE |
527 CRAT_CACHE_FLAGS_SIMD_CACHE),
528 .num_cu_shared = 2,
529 },
530 {
531 /* Scalar L1 Data Cache per SQC */
532 .cache_size = 16,
533 .cache_level = 1,
534 .flags = (CRAT_CACHE_FLAGS_ENABLED |
535 CRAT_CACHE_FLAGS_DATA_CACHE |
536 CRAT_CACHE_FLAGS_SIMD_CACHE),
537 .num_cu_shared = 2,
538 },
539 {
540 /* GL1 Data Cache per SA */
541 .cache_size = 128,
542 .cache_level = 1,
543 .flags = (CRAT_CACHE_FLAGS_ENABLED |
544 CRAT_CACHE_FLAGS_DATA_CACHE |
545 CRAT_CACHE_FLAGS_SIMD_CACHE),
546 .num_cu_shared = 10,
547 },
548 {
549 /* L2 Data Cache per GPU (Total Tex Cache) */
550 .cache_size = 4096,
551 .cache_level = 2,
552 .flags = (CRAT_CACHE_FLAGS_ENABLED |
553 CRAT_CACHE_FLAGS_DATA_CACHE |
554 CRAT_CACHE_FLAGS_SIMD_CACHE),
555 .num_cu_shared = 10,
556 },
557 {
558 /* L3 Data Cache per GPU */
559 .cache_size = 128*1024,
560 .cache_level = 3,
561 .flags = (CRAT_CACHE_FLAGS_ENABLED |
562 CRAT_CACHE_FLAGS_DATA_CACHE |
563 CRAT_CACHE_FLAGS_SIMD_CACHE),
564 .num_cu_shared = 10,
565 },
566};
567
568static struct kfd_gpu_cache_info navy_flounder_cache_info[] = {
569 {
570 /* TCP L1 Cache per CU */
571 .cache_size = 16,
572 .cache_level = 1,
573 .flags = (CRAT_CACHE_FLAGS_ENABLED |
574 CRAT_CACHE_FLAGS_DATA_CACHE |
575 CRAT_CACHE_FLAGS_SIMD_CACHE),
576 .num_cu_shared = 1,
577 },
578 {
579 /* Scalar L1 Instruction Cache per SQC */
580 .cache_size = 32,
581 .cache_level = 1,
582 .flags = (CRAT_CACHE_FLAGS_ENABLED |
583 CRAT_CACHE_FLAGS_INST_CACHE |
584 CRAT_CACHE_FLAGS_SIMD_CACHE),
585 .num_cu_shared = 2,
586 },
587 {
588 /* Scalar L1 Data Cache per SQC */
589 .cache_size = 16,
590 .cache_level = 1,
591 .flags = (CRAT_CACHE_FLAGS_ENABLED |
592 CRAT_CACHE_FLAGS_DATA_CACHE |
593 CRAT_CACHE_FLAGS_SIMD_CACHE),
594 .num_cu_shared = 2,
595 },
596 {
597 /* GL1 Data Cache per SA */
598 .cache_size = 128,
599 .cache_level = 1,
600 .flags = (CRAT_CACHE_FLAGS_ENABLED |
601 CRAT_CACHE_FLAGS_DATA_CACHE |
602 CRAT_CACHE_FLAGS_SIMD_CACHE),
603 .num_cu_shared = 10,
604 },
605 {
606 /* L2 Data Cache per GPU (Total Tex Cache) */
607 .cache_size = 3072,
608 .cache_level = 2,
609 .flags = (CRAT_CACHE_FLAGS_ENABLED |
610 CRAT_CACHE_FLAGS_DATA_CACHE |
611 CRAT_CACHE_FLAGS_SIMD_CACHE),
612 .num_cu_shared = 10,
613 },
614 {
615 /* L3 Data Cache per GPU */
616 .cache_size = 96*1024,
617 .cache_level = 3,
618 .flags = (CRAT_CACHE_FLAGS_ENABLED |
619 CRAT_CACHE_FLAGS_DATA_CACHE |
620 CRAT_CACHE_FLAGS_SIMD_CACHE),
621 .num_cu_shared = 10,
622 },
623};
624
625static struct kfd_gpu_cache_info dimgrey_cavefish_cache_info[] = {
626 {
627 /* TCP L1 Cache per CU */
628 .cache_size = 16,
629 .cache_level = 1,
630 .flags = (CRAT_CACHE_FLAGS_ENABLED |
631 CRAT_CACHE_FLAGS_DATA_CACHE |
632 CRAT_CACHE_FLAGS_SIMD_CACHE),
633 .num_cu_shared = 1,
634 },
635 {
636 /* Scalar L1 Instruction Cache per SQC */
637 .cache_size = 32,
638 .cache_level = 1,
639 .flags = (CRAT_CACHE_FLAGS_ENABLED |
640 CRAT_CACHE_FLAGS_INST_CACHE |
641 CRAT_CACHE_FLAGS_SIMD_CACHE),
642 .num_cu_shared = 2,
643 },
644 {
645 /* Scalar L1 Data Cache per SQC */
646 .cache_size = 16,
647 .cache_level = 1,
648 .flags = (CRAT_CACHE_FLAGS_ENABLED |
649 CRAT_CACHE_FLAGS_DATA_CACHE |
650 CRAT_CACHE_FLAGS_SIMD_CACHE),
651 .num_cu_shared = 2,
652 },
653 {
654 /* GL1 Data Cache per SA */
655 .cache_size = 128,
656 .cache_level = 1,
657 .flags = (CRAT_CACHE_FLAGS_ENABLED |
658 CRAT_CACHE_FLAGS_DATA_CACHE |
659 CRAT_CACHE_FLAGS_SIMD_CACHE),
660 .num_cu_shared = 8,
661 },
662 {
663 /* L2 Data Cache per GPU (Total Tex Cache) */
664 .cache_size = 2048,
665 .cache_level = 2,
666 .flags = (CRAT_CACHE_FLAGS_ENABLED |
667 CRAT_CACHE_FLAGS_DATA_CACHE |
668 CRAT_CACHE_FLAGS_SIMD_CACHE),
669 .num_cu_shared = 8,
670 },
671 {
672 /* L3 Data Cache per GPU */
673 .cache_size = 32*1024,
674 .cache_level = 3,
675 .flags = (CRAT_CACHE_FLAGS_ENABLED |
676 CRAT_CACHE_FLAGS_DATA_CACHE |
677 CRAT_CACHE_FLAGS_SIMD_CACHE),
678 .num_cu_shared = 8,
679 },
680};
681
682static struct kfd_gpu_cache_info beige_goby_cache_info[] = {
683 {
684 /* TCP L1 Cache per CU */
685 .cache_size = 16,
686 .cache_level = 1,
687 .flags = (CRAT_CACHE_FLAGS_ENABLED |
688 CRAT_CACHE_FLAGS_DATA_CACHE |
689 CRAT_CACHE_FLAGS_SIMD_CACHE),
690 .num_cu_shared = 1,
691 },
692 {
693 /* Scalar L1 Instruction Cache per SQC */
694 .cache_size = 32,
695 .cache_level = 1,
696 .flags = (CRAT_CACHE_FLAGS_ENABLED |
697 CRAT_CACHE_FLAGS_INST_CACHE |
698 CRAT_CACHE_FLAGS_SIMD_CACHE),
699 .num_cu_shared = 2,
700 },
701 {
702 /* Scalar L1 Data Cache per SQC */
703 .cache_size = 16,
704 .cache_level = 1,
705 .flags = (CRAT_CACHE_FLAGS_ENABLED |
706 CRAT_CACHE_FLAGS_DATA_CACHE |
707 CRAT_CACHE_FLAGS_SIMD_CACHE),
708 .num_cu_shared = 2,
709 },
710 {
711 /* GL1 Data Cache per SA */
712 .cache_size = 128,
713 .cache_level = 1,
714 .flags = (CRAT_CACHE_FLAGS_ENABLED |
715 CRAT_CACHE_FLAGS_DATA_CACHE |
716 CRAT_CACHE_FLAGS_SIMD_CACHE),
717 .num_cu_shared = 8,
718 },
719 {
720 /* L2 Data Cache per GPU (Total Tex Cache) */
721 .cache_size = 1024,
722 .cache_level = 2,
723 .flags = (CRAT_CACHE_FLAGS_ENABLED |
724 CRAT_CACHE_FLAGS_DATA_CACHE |
725 CRAT_CACHE_FLAGS_SIMD_CACHE),
726 .num_cu_shared = 8,
727 },
728 {
729 /* L3 Data Cache per GPU */
730 .cache_size = 16*1024,
731 .cache_level = 3,
732 .flags = (CRAT_CACHE_FLAGS_ENABLED |
733 CRAT_CACHE_FLAGS_DATA_CACHE |
734 CRAT_CACHE_FLAGS_SIMD_CACHE),
735 .num_cu_shared = 8,
736 },
737};
738
739static struct kfd_gpu_cache_info yellow_carp_cache_info[] = {
740 {
741 /* TCP L1 Cache per CU */
742 .cache_size = 16,
743 .cache_level = 1,
744 .flags = (CRAT_CACHE_FLAGS_ENABLED |
745 CRAT_CACHE_FLAGS_DATA_CACHE |
746 CRAT_CACHE_FLAGS_SIMD_CACHE),
747 .num_cu_shared = 1,
748 },
749 {
750 /* Scalar L1 Instruction Cache per SQC */
751 .cache_size = 32,
752 .cache_level = 1,
753 .flags = (CRAT_CACHE_FLAGS_ENABLED |
754 CRAT_CACHE_FLAGS_INST_CACHE |
755 CRAT_CACHE_FLAGS_SIMD_CACHE),
756 .num_cu_shared = 2,
757 },
758 {
759 /* Scalar L1 Data Cache per SQC */
760 .cache_size = 16,
761 .cache_level = 1,
762 .flags = (CRAT_CACHE_FLAGS_ENABLED |
763 CRAT_CACHE_FLAGS_DATA_CACHE |
764 CRAT_CACHE_FLAGS_SIMD_CACHE),
765 .num_cu_shared = 2,
766 },
767 {
768 /* GL1 Data Cache per SA */
769 .cache_size = 128,
770 .cache_level = 1,
771 .flags = (CRAT_CACHE_FLAGS_ENABLED |
772 CRAT_CACHE_FLAGS_DATA_CACHE |
773 CRAT_CACHE_FLAGS_SIMD_CACHE),
774 .num_cu_shared = 6,
775 },
776 {
777 /* L2 Data Cache per GPU (Total Tex Cache) */
778 .cache_size = 2048,
779 .cache_level = 2,
780 .flags = (CRAT_CACHE_FLAGS_ENABLED |
781 CRAT_CACHE_FLAGS_DATA_CACHE |
782 CRAT_CACHE_FLAGS_SIMD_CACHE),
783 .num_cu_shared = 6,
784 },
785};
786
787static struct kfd_gpu_cache_info gfx1037_cache_info[] = {
788 {
789 /* TCP L1 Cache per CU */
790 .cache_size = 16,
791 .cache_level = 1,
792 .flags = (CRAT_CACHE_FLAGS_ENABLED |
793 CRAT_CACHE_FLAGS_DATA_CACHE |
794 CRAT_CACHE_FLAGS_SIMD_CACHE),
795 .num_cu_shared = 1,
796 },
797 {
798 /* Scalar L1 Instruction Cache per SQC */
799 .cache_size = 32,
800 .cache_level = 1,
801 .flags = (CRAT_CACHE_FLAGS_ENABLED |
802 CRAT_CACHE_FLAGS_INST_CACHE |
803 CRAT_CACHE_FLAGS_SIMD_CACHE),
804 .num_cu_shared = 2,
805 },
806 {
807 /* Scalar L1 Data Cache per SQC */
808 .cache_size = 16,
809 .cache_level = 1,
810 .flags = (CRAT_CACHE_FLAGS_ENABLED |
811 CRAT_CACHE_FLAGS_DATA_CACHE |
812 CRAT_CACHE_FLAGS_SIMD_CACHE),
813 .num_cu_shared = 2,
814 },
815 {
816 /* GL1 Data Cache per SA */
817 .cache_size = 128,
818 .cache_level = 1,
819 .flags = (CRAT_CACHE_FLAGS_ENABLED |
820 CRAT_CACHE_FLAGS_DATA_CACHE |
821 CRAT_CACHE_FLAGS_SIMD_CACHE),
822 .num_cu_shared = 2,
823 },
824 {
825 /* L2 Data Cache per GPU (Total Tex Cache) */
826 .cache_size = 256,
827 .cache_level = 2,
828 .flags = (CRAT_CACHE_FLAGS_ENABLED |
829 CRAT_CACHE_FLAGS_DATA_CACHE |
830 CRAT_CACHE_FLAGS_SIMD_CACHE),
831 .num_cu_shared = 2,
832 },
833};
834
835static struct kfd_gpu_cache_info gc_10_3_6_cache_info[] = {
836 {
837 /* TCP L1 Cache per CU */
838 .cache_size = 16,
839 .cache_level = 1,
840 .flags = (CRAT_CACHE_FLAGS_ENABLED |
841 CRAT_CACHE_FLAGS_DATA_CACHE |
842 CRAT_CACHE_FLAGS_SIMD_CACHE),
843 .num_cu_shared = 1,
844 },
845 {
846 /* Scalar L1 Instruction Cache per SQC */
847 .cache_size = 32,
848 .cache_level = 1,
849 .flags = (CRAT_CACHE_FLAGS_ENABLED |
850 CRAT_CACHE_FLAGS_INST_CACHE |
851 CRAT_CACHE_FLAGS_SIMD_CACHE),
852 .num_cu_shared = 2,
853 },
854 {
855 /* Scalar L1 Data Cache per SQC */
856 .cache_size = 16,
857 .cache_level = 1,
858 .flags = (CRAT_CACHE_FLAGS_ENABLED |
859 CRAT_CACHE_FLAGS_DATA_CACHE |
860 CRAT_CACHE_FLAGS_SIMD_CACHE),
861 .num_cu_shared = 2,
862 },
863 {
864 /* GL1 Data Cache per SA */
865 .cache_size = 128,
866 .cache_level = 1,
867 .flags = (CRAT_CACHE_FLAGS_ENABLED |
868 CRAT_CACHE_FLAGS_DATA_CACHE |
869 CRAT_CACHE_FLAGS_SIMD_CACHE),
870 .num_cu_shared = 2,
871 },
872 {
873 /* L2 Data Cache per GPU (Total Tex Cache) */
874 .cache_size = 256,
875 .cache_level = 2,
876 .flags = (CRAT_CACHE_FLAGS_ENABLED |
877 CRAT_CACHE_FLAGS_DATA_CACHE |
878 CRAT_CACHE_FLAGS_SIMD_CACHE),
879 .num_cu_shared = 2,
880 },
881};
882
883static struct kfd_gpu_cache_info dummy_cache_info[] = {
884 {
885 /* TCP L1 Cache per CU */
886 .cache_size = 16,
887 .cache_level = 1,
888 .flags = (CRAT_CACHE_FLAGS_ENABLED |
889 CRAT_CACHE_FLAGS_DATA_CACHE |
890 CRAT_CACHE_FLAGS_SIMD_CACHE),
891 .num_cu_shared = 1,
892 },
893 {
894 /* Scalar L1 Instruction Cache per SQC */
895 .cache_size = 32,
896 .cache_level = 1,
897 .flags = (CRAT_CACHE_FLAGS_ENABLED |
898 CRAT_CACHE_FLAGS_INST_CACHE |
899 CRAT_CACHE_FLAGS_SIMD_CACHE),
900 .num_cu_shared = 2,
901 },
902 {
903 /* Scalar L1 Data Cache per SQC */
904 .cache_size = 16,
905 .cache_level = 1,
906 .flags = (CRAT_CACHE_FLAGS_ENABLED |
907 CRAT_CACHE_FLAGS_DATA_CACHE |
908 CRAT_CACHE_FLAGS_SIMD_CACHE),
909 .num_cu_shared = 2,
910 },
911 {
912 /* GL1 Data Cache per SA */
913 .cache_size = 128,
914 .cache_level = 1,
915 .flags = (CRAT_CACHE_FLAGS_ENABLED |
916 CRAT_CACHE_FLAGS_DATA_CACHE |
917 CRAT_CACHE_FLAGS_SIMD_CACHE),
918 .num_cu_shared = 6,
919 },
920 {
921 /* L2 Data Cache per GPU (Total Tex Cache) */
922 .cache_size = 2048,
923 .cache_level = 2,
924 .flags = (CRAT_CACHE_FLAGS_ENABLED |
925 CRAT_CACHE_FLAGS_DATA_CACHE |
926 CRAT_CACHE_FLAGS_SIMD_CACHE),
927 .num_cu_shared = 6,
928 },
929};
930
931static void kfd_populated_cu_info_cpu(struct kfd_topology_device *dev,
932 struct crat_subtype_computeunit *cu)
933{
934 dev->node_props.cpu_cores_count = cu->num_cpu_cores;
935 dev->node_props.cpu_core_id_base = cu->processor_id_low;
936 if (cu->hsa_capability & CRAT_CU_FLAGS_IOMMU_PRESENT)
937 dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
938
939 pr_debug("CU CPU: cores=%d id_base=%d\n", cu->num_cpu_cores,
940 cu->processor_id_low);
941}
942
943static void kfd_populated_cu_info_gpu(struct kfd_topology_device *dev,
944 struct crat_subtype_computeunit *cu)
945{
946 dev->node_props.simd_id_base = cu->processor_id_low;
947 dev->node_props.simd_count = cu->num_simd_cores;
948 dev->node_props.lds_size_in_kb = cu->lds_size_in_kb;
949 dev->node_props.max_waves_per_simd = cu->max_waves_simd;
950 dev->node_props.wave_front_size = cu->wave_front_size;
951 dev->node_props.array_count = cu->array_count;
952 dev->node_props.cu_per_simd_array = cu->num_cu_per_array;
953 dev->node_props.simd_per_cu = cu->num_simd_per_cu;
954 dev->node_props.max_slots_scratch_cu = cu->max_slots_scatch_cu;
955 if (cu->hsa_capability & CRAT_CU_FLAGS_HOT_PLUGGABLE)
956 dev->node_props.capability |= HSA_CAP_HOT_PLUGGABLE;
957 pr_debug("CU GPU: id_base=%d\n", cu->processor_id_low);
958}
959
960/* kfd_parse_subtype_cu - parse compute unit subtypes and attach it to correct
961 * topology device present in the device_list
962 */
963static int kfd_parse_subtype_cu(struct crat_subtype_computeunit *cu,
964 struct list_head *device_list)
965{
966 struct kfd_topology_device *dev;
967
968 pr_debug("Found CU entry in CRAT table with proximity_domain=%d caps=%x\n",
969 cu->proximity_domain, cu->hsa_capability);
970 list_for_each_entry(dev, device_list, list) {
971 if (cu->proximity_domain == dev->proximity_domain) {
972 if (cu->flags & CRAT_CU_FLAGS_CPU_PRESENT)
973 kfd_populated_cu_info_cpu(dev, cu);
974
975 if (cu->flags & CRAT_CU_FLAGS_GPU_PRESENT)
976 kfd_populated_cu_info_gpu(dev, cu);
977 break;
978 }
979 }
980
981 return 0;
982}
983
984static struct kfd_mem_properties *
985find_subtype_mem(uint32_t heap_type, uint32_t flags, uint32_t width,
986 struct kfd_topology_device *dev)
987{
988 struct kfd_mem_properties *props;
989
990 list_for_each_entry(props, &dev->mem_props, list) {
991 if (props->heap_type == heap_type
992 && props->flags == flags
993 && props->width == width)
994 return props;
995 }
996
997 return NULL;
998}
999/* kfd_parse_subtype_mem - parse memory subtypes and attach it to correct
1000 * topology device present in the device_list
1001 */
1002static int kfd_parse_subtype_mem(struct crat_subtype_memory *mem,
1003 struct list_head *device_list)
1004{
1005 struct kfd_mem_properties *props;
1006 struct kfd_topology_device *dev;
1007 uint32_t heap_type;
1008 uint64_t size_in_bytes;
1009 uint32_t flags = 0;
1010 uint32_t width;
1011
1012 pr_debug("Found memory entry in CRAT table with proximity_domain=%d\n",
1013 mem->proximity_domain);
1014 list_for_each_entry(dev, device_list, list) {
1015 if (mem->proximity_domain == dev->proximity_domain) {
1016 /* We're on GPU node */
1017 if (dev->node_props.cpu_cores_count == 0) {
1018 /* APU */
1019 if (mem->visibility_type == 0)
1020 heap_type =
1021 HSA_MEM_HEAP_TYPE_FB_PRIVATE;
1022 /* dGPU */
1023 else
1024 heap_type = mem->visibility_type;
1025 } else
1026 heap_type = HSA_MEM_HEAP_TYPE_SYSTEM;
1027
1028 if (mem->flags & CRAT_MEM_FLAGS_HOT_PLUGGABLE)
1029 flags |= HSA_MEM_FLAGS_HOT_PLUGGABLE;
1030 if (mem->flags & CRAT_MEM_FLAGS_NON_VOLATILE)
1031 flags |= HSA_MEM_FLAGS_NON_VOLATILE;
1032
1033 size_in_bytes =
1034 ((uint64_t)mem->length_high << 32) +
1035 mem->length_low;
1036 width = mem->width;
1037
1038 /* Multiple banks of the same type are aggregated into
1039 * one. User mode doesn't care about multiple physical
1040 * memory segments. It's managed as a single virtual
1041 * heap for user mode.
1042 */
1043 props = find_subtype_mem(heap_type, flags, width, dev);
1044 if (props) {
1045 props->size_in_bytes += size_in_bytes;
1046 break;
1047 }
1048
1049 props = kfd_alloc_struct(props);
1050 if (!props)
1051 return -ENOMEM;
1052
1053 props->heap_type = heap_type;
1054 props->flags = flags;
1055 props->size_in_bytes = size_in_bytes;
1056 props->width = width;
1057
1058 dev->node_props.mem_banks_count++;
1059 list_add_tail(&props->list, &dev->mem_props);
1060
1061 break;
1062 }
1063 }
1064
1065 return 0;
1066}
1067
1068/* kfd_parse_subtype_cache - parse cache subtypes and attach it to correct
1069 * topology device present in the device_list
1070 */
1071static int kfd_parse_subtype_cache(struct crat_subtype_cache *cache,
1072 struct list_head *device_list)
1073{
1074 struct kfd_cache_properties *props;
1075 struct kfd_topology_device *dev;
1076 uint32_t id;
1077 uint32_t total_num_of_cu;
1078
1079 id = cache->processor_id_low;
1080
1081 pr_debug("Found cache entry in CRAT table with processor_id=%d\n", id);
1082 list_for_each_entry(dev, device_list, list) {
1083 total_num_of_cu = (dev->node_props.array_count *
1084 dev->node_props.cu_per_simd_array);
1085
1086 /* Cache infomration in CRAT doesn't have proximity_domain
1087 * information as it is associated with a CPU core or GPU
1088 * Compute Unit. So map the cache using CPU core Id or SIMD
1089 * (GPU) ID.
1090 * TODO: This works because currently we can safely assume that
1091 * Compute Units are parsed before caches are parsed. In
1092 * future, remove this dependency
1093 */
1094 if ((id >= dev->node_props.cpu_core_id_base &&
1095 id <= dev->node_props.cpu_core_id_base +
1096 dev->node_props.cpu_cores_count) ||
1097 (id >= dev->node_props.simd_id_base &&
1098 id < dev->node_props.simd_id_base +
1099 total_num_of_cu)) {
1100 props = kfd_alloc_struct(props);
1101 if (!props)
1102 return -ENOMEM;
1103
1104 props->processor_id_low = id;
1105 props->cache_level = cache->cache_level;
1106 props->cache_size = cache->cache_size;
1107 props->cacheline_size = cache->cache_line_size;
1108 props->cachelines_per_tag = cache->lines_per_tag;
1109 props->cache_assoc = cache->associativity;
1110 props->cache_latency = cache->cache_latency;
1111
1112 memcpy(props->sibling_map, cache->sibling_map,
1113 CRAT_SIBLINGMAP_SIZE);
1114
1115 /* set the sibling_map_size as 32 for CRAT from ACPI */
1116 props->sibling_map_size = CRAT_SIBLINGMAP_SIZE;
1117
1118 if (cache->flags & CRAT_CACHE_FLAGS_DATA_CACHE)
1119 props->cache_type |= HSA_CACHE_TYPE_DATA;
1120 if (cache->flags & CRAT_CACHE_FLAGS_INST_CACHE)
1121 props->cache_type |= HSA_CACHE_TYPE_INSTRUCTION;
1122 if (cache->flags & CRAT_CACHE_FLAGS_CPU_CACHE)
1123 props->cache_type |= HSA_CACHE_TYPE_CPU;
1124 if (cache->flags & CRAT_CACHE_FLAGS_SIMD_CACHE)
1125 props->cache_type |= HSA_CACHE_TYPE_HSACU;
1126
1127 dev->node_props.caches_count++;
1128 list_add_tail(&props->list, &dev->cache_props);
1129
1130 break;
1131 }
1132 }
1133
1134 return 0;
1135}
1136
1137/* kfd_parse_subtype_iolink - parse iolink subtypes and attach it to correct
1138 * topology device present in the device_list
1139 */
1140static int kfd_parse_subtype_iolink(struct crat_subtype_iolink *iolink,
1141 struct list_head *device_list)
1142{
1143 struct kfd_iolink_properties *props = NULL, *props2;
1144 struct kfd_topology_device *dev, *to_dev;
1145 uint32_t id_from;
1146 uint32_t id_to;
1147
1148 id_from = iolink->proximity_domain_from;
1149 id_to = iolink->proximity_domain_to;
1150
1151 pr_debug("Found IO link entry in CRAT table with id_from=%d, id_to %d\n",
1152 id_from, id_to);
1153 list_for_each_entry(dev, device_list, list) {
1154 if (id_from == dev->proximity_domain) {
1155 props = kfd_alloc_struct(props);
1156 if (!props)
1157 return -ENOMEM;
1158
1159 props->node_from = id_from;
1160 props->node_to = id_to;
1161 props->ver_maj = iolink->version_major;
1162 props->ver_min = iolink->version_minor;
1163 props->iolink_type = iolink->io_interface_type;
1164
1165 if (props->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS)
1166 props->weight = 20;
1167 else if (props->iolink_type == CRAT_IOLINK_TYPE_XGMI)
1168 props->weight = iolink->weight_xgmi;
1169 else
1170 props->weight = node_distance(id_from, id_to);
1171
1172 props->min_latency = iolink->minimum_latency;
1173 props->max_latency = iolink->maximum_latency;
1174 props->min_bandwidth = iolink->minimum_bandwidth_mbs;
1175 props->max_bandwidth = iolink->maximum_bandwidth_mbs;
1176 props->rec_transfer_size =
1177 iolink->recommended_transfer_size;
1178
1179 dev->node_props.io_links_count++;
1180 list_add_tail(&props->list, &dev->io_link_props);
1181 break;
1182 }
1183 }
1184
1185 /* CPU topology is created before GPUs are detected, so CPU->GPU
1186 * links are not built at that time. If a PCIe type is discovered, it
1187 * means a GPU is detected and we are adding GPU->CPU to the topology.
1188 * At this time, also add the corresponded CPU->GPU link if GPU
1189 * is large bar.
1190 * For xGMI, we only added the link with one direction in the crat
1191 * table, add corresponded reversed direction link now.
1192 */
1193 if (props && (iolink->flags & CRAT_IOLINK_FLAGS_BI_DIRECTIONAL)) {
1194 to_dev = kfd_topology_device_by_proximity_domain_no_lock(id_to);
1195 if (!to_dev)
1196 return -ENODEV;
1197 /* same everything but the other direction */
1198 props2 = kmemdup(props, sizeof(*props2), GFP_KERNEL);
1199 if (!props2)
1200 return -ENOMEM;
1201
1202 props2->node_from = id_to;
1203 props2->node_to = id_from;
1204 props2->kobj = NULL;
1205 to_dev->node_props.io_links_count++;
1206 list_add_tail(&props2->list, &to_dev->io_link_props);
1207 }
1208
1209 return 0;
1210}
1211
1212/* kfd_parse_subtype - parse subtypes and attach it to correct topology device
1213 * present in the device_list
1214 * @sub_type_hdr - subtype section of crat_image
1215 * @device_list - list of topology devices present in this crat_image
1216 */
1217static int kfd_parse_subtype(struct crat_subtype_generic *sub_type_hdr,
1218 struct list_head *device_list)
1219{
1220 struct crat_subtype_computeunit *cu;
1221 struct crat_subtype_memory *mem;
1222 struct crat_subtype_cache *cache;
1223 struct crat_subtype_iolink *iolink;
1224 int ret = 0;
1225
1226 switch (sub_type_hdr->type) {
1227 case CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY:
1228 cu = (struct crat_subtype_computeunit *)sub_type_hdr;
1229 ret = kfd_parse_subtype_cu(cu, device_list);
1230 break;
1231 case CRAT_SUBTYPE_MEMORY_AFFINITY:
1232 mem = (struct crat_subtype_memory *)sub_type_hdr;
1233 ret = kfd_parse_subtype_mem(mem, device_list);
1234 break;
1235 case CRAT_SUBTYPE_CACHE_AFFINITY:
1236 cache = (struct crat_subtype_cache *)sub_type_hdr;
1237 ret = kfd_parse_subtype_cache(cache, device_list);
1238 break;
1239 case CRAT_SUBTYPE_TLB_AFFINITY:
1240 /*
1241 * For now, nothing to do here
1242 */
1243 pr_debug("Found TLB entry in CRAT table (not processing)\n");
1244 break;
1245 case CRAT_SUBTYPE_CCOMPUTE_AFFINITY:
1246 /*
1247 * For now, nothing to do here
1248 */
1249 pr_debug("Found CCOMPUTE entry in CRAT table (not processing)\n");
1250 break;
1251 case CRAT_SUBTYPE_IOLINK_AFFINITY:
1252 iolink = (struct crat_subtype_iolink *)sub_type_hdr;
1253 ret = kfd_parse_subtype_iolink(iolink, device_list);
1254 break;
1255 default:
1256 pr_warn("Unknown subtype %d in CRAT\n",
1257 sub_type_hdr->type);
1258 }
1259
1260 return ret;
1261}
1262
1263/* kfd_parse_crat_table - parse CRAT table. For each node present in CRAT
1264 * create a kfd_topology_device and add in to device_list. Also parse
1265 * CRAT subtypes and attach it to appropriate kfd_topology_device
1266 * @crat_image - input image containing CRAT
1267 * @device_list - [OUT] list of kfd_topology_device generated after
1268 * parsing crat_image
1269 * @proximity_domain - Proximity domain of the first device in the table
1270 *
1271 * Return - 0 if successful else -ve value
1272 */
1273int kfd_parse_crat_table(void *crat_image, struct list_head *device_list,
1274 uint32_t proximity_domain)
1275{
1276 struct kfd_topology_device *top_dev = NULL;
1277 struct crat_subtype_generic *sub_type_hdr;
1278 uint16_t node_id;
1279 int ret = 0;
1280 struct crat_header *crat_table = (struct crat_header *)crat_image;
1281 uint16_t num_nodes;
1282 uint32_t image_len;
1283
1284 if (!crat_image)
1285 return -EINVAL;
1286
1287 if (!list_empty(device_list)) {
1288 pr_warn("Error device list should be empty\n");
1289 return -EINVAL;
1290 }
1291
1292 num_nodes = crat_table->num_domains;
1293 image_len = crat_table->length;
1294
1295 pr_debug("Parsing CRAT table with %d nodes\n", num_nodes);
1296
1297 for (node_id = 0; node_id < num_nodes; node_id++) {
1298 top_dev = kfd_create_topology_device(device_list);
1299 if (!top_dev)
1300 break;
1301 top_dev->proximity_domain = proximity_domain++;
1302 }
1303
1304 if (!top_dev) {
1305 ret = -ENOMEM;
1306 goto err;
1307 }
1308
1309 memcpy(top_dev->oem_id, crat_table->oem_id, CRAT_OEMID_LENGTH);
1310 memcpy(top_dev->oem_table_id, crat_table->oem_table_id,
1311 CRAT_OEMTABLEID_LENGTH);
1312 top_dev->oem_revision = crat_table->oem_revision;
1313
1314 sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1);
1315 while ((char *)sub_type_hdr + sizeof(struct crat_subtype_generic) <
1316 ((char *)crat_image) + image_len) {
1317 if (sub_type_hdr->flags & CRAT_SUBTYPE_FLAGS_ENABLED) {
1318 ret = kfd_parse_subtype(sub_type_hdr, device_list);
1319 if (ret)
1320 break;
1321 }
1322
1323 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1324 sub_type_hdr->length);
1325 }
1326
1327err:
1328 if (ret)
1329 kfd_release_topology_device_list(device_list);
1330
1331 return ret;
1332}
1333
1334
1335static int kfd_fill_gpu_cache_info_from_gfx_config(struct kfd_dev *kdev,
1336 struct kfd_gpu_cache_info *pcache_info)
1337{
1338 struct amdgpu_device *adev = kdev->adev;
1339 int i = 0;
1340
1341 /* TCP L1 Cache per CU */
1342 if (adev->gfx.config.gc_tcp_l1_size) {
1343 pcache_info[i].cache_size = adev->gfx.config.gc_tcp_l1_size;
1344 pcache_info[i].cache_level = 1;
1345 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1346 CRAT_CACHE_FLAGS_DATA_CACHE |
1347 CRAT_CACHE_FLAGS_SIMD_CACHE);
1348 pcache_info[0].num_cu_shared = adev->gfx.config.gc_num_tcp_per_wpg / 2;
1349 i++;
1350 }
1351 /* Scalar L1 Instruction Cache per SQC */
1352 if (adev->gfx.config.gc_l1_instruction_cache_size_per_sqc) {
1353 pcache_info[i].cache_size =
1354 adev->gfx.config.gc_l1_instruction_cache_size_per_sqc;
1355 pcache_info[i].cache_level = 1;
1356 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1357 CRAT_CACHE_FLAGS_INST_CACHE |
1358 CRAT_CACHE_FLAGS_SIMD_CACHE);
1359 pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_sqc_per_wgp * 2;
1360 i++;
1361 }
1362 /* Scalar L1 Data Cache per SQC */
1363 if (adev->gfx.config.gc_l1_data_cache_size_per_sqc) {
1364 pcache_info[i].cache_size = adev->gfx.config.gc_l1_data_cache_size_per_sqc;
1365 pcache_info[i].cache_level = 1;
1366 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1367 CRAT_CACHE_FLAGS_DATA_CACHE |
1368 CRAT_CACHE_FLAGS_SIMD_CACHE);
1369 pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_sqc_per_wgp * 2;
1370 i++;
1371 }
1372 /* GL1 Data Cache per SA */
1373 if (adev->gfx.config.gc_gl1c_per_sa &&
1374 adev->gfx.config.gc_gl1c_size_per_instance) {
1375 pcache_info[i].cache_size = adev->gfx.config.gc_gl1c_per_sa *
1376 adev->gfx.config.gc_gl1c_size_per_instance;
1377 pcache_info[i].cache_level = 1;
1378 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1379 CRAT_CACHE_FLAGS_DATA_CACHE |
1380 CRAT_CACHE_FLAGS_SIMD_CACHE);
1381 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
1382 i++;
1383 }
1384 /* L2 Data Cache per GPU (Total Tex Cache) */
1385 if (adev->gfx.config.gc_gl2c_per_gpu) {
1386 pcache_info[i].cache_size = adev->gfx.config.gc_gl2c_per_gpu;
1387 pcache_info[i].cache_level = 2;
1388 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1389 CRAT_CACHE_FLAGS_DATA_CACHE |
1390 CRAT_CACHE_FLAGS_SIMD_CACHE);
1391 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
1392 i++;
1393 }
1394 /* L3 Data Cache per GPU */
1395 if (adev->gmc.mall_size) {
1396 pcache_info[i].cache_size = adev->gmc.mall_size / 1024;
1397 pcache_info[i].cache_level = 3;
1398 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1399 CRAT_CACHE_FLAGS_DATA_CACHE |
1400 CRAT_CACHE_FLAGS_SIMD_CACHE);
1401 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
1402 i++;
1403 }
1404 return i;
1405}
1406
1407static int kfd_fill_gpu_cache_info_from_gfx_config_v2(struct kfd_dev *kdev,
1408 struct kfd_gpu_cache_info *pcache_info)
1409{
1410 struct amdgpu_device *adev = kdev->adev;
1411 int i = 0;
1412
1413 /* TCP L1 Cache per CU */
1414 if (adev->gfx.config.gc_tcp_size_per_cu) {
1415 pcache_info[i].cache_size = adev->gfx.config.gc_tcp_size_per_cu;
1416 pcache_info[i].cache_level = 1;
1417 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1418 CRAT_CACHE_FLAGS_DATA_CACHE |
1419 CRAT_CACHE_FLAGS_SIMD_CACHE);
1420 pcache_info[i].num_cu_shared = 1;
1421 i++;
1422 }
1423 /* Scalar L1 Instruction Cache per SQC */
1424 if (adev->gfx.config.gc_l1_instruction_cache_size_per_sqc) {
1425 pcache_info[i].cache_size =
1426 adev->gfx.config.gc_l1_instruction_cache_size_per_sqc;
1427 pcache_info[i].cache_level = 1;
1428 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1429 CRAT_CACHE_FLAGS_INST_CACHE |
1430 CRAT_CACHE_FLAGS_SIMD_CACHE);
1431 pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_cu_per_sqc;
1432 i++;
1433 }
1434 /* Scalar L1 Data Cache per SQC */
1435 if (adev->gfx.config.gc_l1_data_cache_size_per_sqc) {
1436 pcache_info[i].cache_size = adev->gfx.config.gc_l1_data_cache_size_per_sqc;
1437 pcache_info[i].cache_level = 1;
1438 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1439 CRAT_CACHE_FLAGS_DATA_CACHE |
1440 CRAT_CACHE_FLAGS_SIMD_CACHE);
1441 pcache_info[i].num_cu_shared = adev->gfx.config.gc_num_cu_per_sqc;
1442 i++;
1443 }
1444 /* L2 Data Cache per GPU (Total Tex Cache) */
1445 if (adev->gfx.config.gc_tcc_size) {
1446 pcache_info[i].cache_size = adev->gfx.config.gc_tcc_size;
1447 pcache_info[i].cache_level = 2;
1448 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1449 CRAT_CACHE_FLAGS_DATA_CACHE |
1450 CRAT_CACHE_FLAGS_SIMD_CACHE);
1451 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
1452 i++;
1453 }
1454 /* L3 Data Cache per GPU */
1455 if (adev->gmc.mall_size) {
1456 pcache_info[i].cache_size = adev->gmc.mall_size / 1024;
1457 pcache_info[i].cache_level = 3;
1458 pcache_info[i].flags = (CRAT_CACHE_FLAGS_ENABLED |
1459 CRAT_CACHE_FLAGS_DATA_CACHE |
1460 CRAT_CACHE_FLAGS_SIMD_CACHE);
1461 pcache_info[i].num_cu_shared = adev->gfx.config.max_cu_per_sh;
1462 i++;
1463 }
1464 return i;
1465}
1466
1467int kfd_get_gpu_cache_info(struct kfd_node *kdev, struct kfd_gpu_cache_info **pcache_info)
1468{
1469 int num_of_cache_types = 0;
1470
1471 switch (kdev->adev->asic_type) {
1472 case CHIP_KAVERI:
1473 *pcache_info = kaveri_cache_info;
1474 num_of_cache_types = ARRAY_SIZE(kaveri_cache_info);
1475 break;
1476 case CHIP_HAWAII:
1477 *pcache_info = hawaii_cache_info;
1478 num_of_cache_types = ARRAY_SIZE(hawaii_cache_info);
1479 break;
1480 case CHIP_CARRIZO:
1481 *pcache_info = carrizo_cache_info;
1482 num_of_cache_types = ARRAY_SIZE(carrizo_cache_info);
1483 break;
1484 case CHIP_TONGA:
1485 *pcache_info = tonga_cache_info;
1486 num_of_cache_types = ARRAY_SIZE(tonga_cache_info);
1487 break;
1488 case CHIP_FIJI:
1489 *pcache_info = fiji_cache_info;
1490 num_of_cache_types = ARRAY_SIZE(fiji_cache_info);
1491 break;
1492 case CHIP_POLARIS10:
1493 *pcache_info = polaris10_cache_info;
1494 num_of_cache_types = ARRAY_SIZE(polaris10_cache_info);
1495 break;
1496 case CHIP_POLARIS11:
1497 *pcache_info = polaris11_cache_info;
1498 num_of_cache_types = ARRAY_SIZE(polaris11_cache_info);
1499 break;
1500 case CHIP_POLARIS12:
1501 *pcache_info = polaris12_cache_info;
1502 num_of_cache_types = ARRAY_SIZE(polaris12_cache_info);
1503 break;
1504 case CHIP_VEGAM:
1505 *pcache_info = vegam_cache_info;
1506 num_of_cache_types = ARRAY_SIZE(vegam_cache_info);
1507 break;
1508 default:
1509 switch (KFD_GC_VERSION(kdev)) {
1510 case IP_VERSION(9, 0, 1):
1511 *pcache_info = vega10_cache_info;
1512 num_of_cache_types = ARRAY_SIZE(vega10_cache_info);
1513 break;
1514 case IP_VERSION(9, 2, 1):
1515 *pcache_info = vega12_cache_info;
1516 num_of_cache_types = ARRAY_SIZE(vega12_cache_info);
1517 break;
1518 case IP_VERSION(9, 4, 0):
1519 case IP_VERSION(9, 4, 1):
1520 *pcache_info = vega20_cache_info;
1521 num_of_cache_types = ARRAY_SIZE(vega20_cache_info);
1522 break;
1523 case IP_VERSION(9, 4, 2):
1524 *pcache_info = aldebaran_cache_info;
1525 num_of_cache_types = ARRAY_SIZE(aldebaran_cache_info);
1526 break;
1527 case IP_VERSION(9, 4, 3):
1528 num_of_cache_types =
1529 kfd_fill_gpu_cache_info_from_gfx_config_v2(kdev->kfd,
1530 *pcache_info);
1531 break;
1532 case IP_VERSION(9, 1, 0):
1533 case IP_VERSION(9, 2, 2):
1534 *pcache_info = raven_cache_info;
1535 num_of_cache_types = ARRAY_SIZE(raven_cache_info);
1536 break;
1537 case IP_VERSION(9, 3, 0):
1538 *pcache_info = renoir_cache_info;
1539 num_of_cache_types = ARRAY_SIZE(renoir_cache_info);
1540 break;
1541 case IP_VERSION(10, 1, 10):
1542 case IP_VERSION(10, 1, 2):
1543 case IP_VERSION(10, 1, 3):
1544 case IP_VERSION(10, 1, 4):
1545 *pcache_info = navi10_cache_info;
1546 num_of_cache_types = ARRAY_SIZE(navi10_cache_info);
1547 break;
1548 case IP_VERSION(10, 1, 1):
1549 *pcache_info = navi14_cache_info;
1550 num_of_cache_types = ARRAY_SIZE(navi14_cache_info);
1551 break;
1552 case IP_VERSION(10, 3, 0):
1553 *pcache_info = sienna_cichlid_cache_info;
1554 num_of_cache_types = ARRAY_SIZE(sienna_cichlid_cache_info);
1555 break;
1556 case IP_VERSION(10, 3, 2):
1557 *pcache_info = navy_flounder_cache_info;
1558 num_of_cache_types = ARRAY_SIZE(navy_flounder_cache_info);
1559 break;
1560 case IP_VERSION(10, 3, 4):
1561 *pcache_info = dimgrey_cavefish_cache_info;
1562 num_of_cache_types = ARRAY_SIZE(dimgrey_cavefish_cache_info);
1563 break;
1564 case IP_VERSION(10, 3, 1):
1565 *pcache_info = vangogh_cache_info;
1566 num_of_cache_types = ARRAY_SIZE(vangogh_cache_info);
1567 break;
1568 case IP_VERSION(10, 3, 5):
1569 *pcache_info = beige_goby_cache_info;
1570 num_of_cache_types = ARRAY_SIZE(beige_goby_cache_info);
1571 break;
1572 case IP_VERSION(10, 3, 3):
1573 *pcache_info = yellow_carp_cache_info;
1574 num_of_cache_types = ARRAY_SIZE(yellow_carp_cache_info);
1575 break;
1576 case IP_VERSION(10, 3, 6):
1577 *pcache_info = gc_10_3_6_cache_info;
1578 num_of_cache_types = ARRAY_SIZE(gc_10_3_6_cache_info);
1579 break;
1580 case IP_VERSION(10, 3, 7):
1581 *pcache_info = gfx1037_cache_info;
1582 num_of_cache_types = ARRAY_SIZE(gfx1037_cache_info);
1583 break;
1584 case IP_VERSION(11, 0, 0):
1585 case IP_VERSION(11, 0, 1):
1586 case IP_VERSION(11, 0, 2):
1587 case IP_VERSION(11, 0, 3):
1588 case IP_VERSION(11, 0, 4):
1589 case IP_VERSION(11, 5, 0):
1590 num_of_cache_types =
1591 kfd_fill_gpu_cache_info_from_gfx_config(kdev->kfd, *pcache_info);
1592 break;
1593 default:
1594 *pcache_info = dummy_cache_info;
1595 num_of_cache_types = ARRAY_SIZE(dummy_cache_info);
1596 pr_warn("dummy cache info is used temporarily and real cache info need update later.\n");
1597 break;
1598 }
1599 }
1600 return num_of_cache_types;
1601}
1602
1603/* Memory required to create Virtual CRAT.
1604 * Since there is no easy way to predict the amount of memory required, the
1605 * following amount is allocated for GPU Virtual CRAT. This is
1606 * expected to cover all known conditions. But to be safe additional check
1607 * is put in the code to ensure we don't overwrite.
1608 */
1609#define VCRAT_SIZE_FOR_GPU (4 * PAGE_SIZE)
1610
1611/* kfd_fill_cu_for_cpu - Fill in Compute info for the given CPU NUMA node
1612 *
1613 * @numa_node_id: CPU NUMA node id
1614 * @avail_size: Available size in the memory
1615 * @sub_type_hdr: Memory into which compute info will be filled in
1616 *
1617 * Return 0 if successful else return -ve value
1618 */
1619static int kfd_fill_cu_for_cpu(int numa_node_id, int *avail_size,
1620 int proximity_domain,
1621 struct crat_subtype_computeunit *sub_type_hdr)
1622{
1623 const struct cpumask *cpumask;
1624
1625 *avail_size -= sizeof(struct crat_subtype_computeunit);
1626 if (*avail_size < 0)
1627 return -ENOMEM;
1628
1629 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit));
1630
1631 /* Fill in subtype header data */
1632 sub_type_hdr->type = CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY;
1633 sub_type_hdr->length = sizeof(struct crat_subtype_computeunit);
1634 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
1635
1636 cpumask = cpumask_of_node(numa_node_id);
1637
1638 /* Fill in CU data */
1639 sub_type_hdr->flags |= CRAT_CU_FLAGS_CPU_PRESENT;
1640 sub_type_hdr->proximity_domain = proximity_domain;
1641 sub_type_hdr->processor_id_low = kfd_numa_node_to_apic_id(numa_node_id);
1642 if (sub_type_hdr->processor_id_low == -1)
1643 return -EINVAL;
1644
1645 sub_type_hdr->num_cpu_cores = cpumask_weight(cpumask);
1646
1647 return 0;
1648}
1649
1650/* kfd_fill_mem_info_for_cpu - Fill in Memory info for the given CPU NUMA node
1651 *
1652 * @numa_node_id: CPU NUMA node id
1653 * @avail_size: Available size in the memory
1654 * @sub_type_hdr: Memory into which compute info will be filled in
1655 *
1656 * Return 0 if successful else return -ve value
1657 */
1658static int kfd_fill_mem_info_for_cpu(int numa_node_id, int *avail_size,
1659 int proximity_domain,
1660 struct crat_subtype_memory *sub_type_hdr)
1661{
1662 uint64_t mem_in_bytes = 0;
1663 pg_data_t *pgdat;
1664 int zone_type;
1665
1666 *avail_size -= sizeof(struct crat_subtype_memory);
1667 if (*avail_size < 0)
1668 return -ENOMEM;
1669
1670 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_memory));
1671
1672 /* Fill in subtype header data */
1673 sub_type_hdr->type = CRAT_SUBTYPE_MEMORY_AFFINITY;
1674 sub_type_hdr->length = sizeof(struct crat_subtype_memory);
1675 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
1676
1677 /* Fill in Memory Subunit data */
1678
1679 /* Unlike si_meminfo, si_meminfo_node is not exported. So
1680 * the following lines are duplicated from si_meminfo_node
1681 * function
1682 */
1683 pgdat = NODE_DATA(numa_node_id);
1684 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
1685 mem_in_bytes += zone_managed_pages(&pgdat->node_zones[zone_type]);
1686 mem_in_bytes <<= PAGE_SHIFT;
1687
1688 sub_type_hdr->length_low = lower_32_bits(mem_in_bytes);
1689 sub_type_hdr->length_high = upper_32_bits(mem_in_bytes);
1690 sub_type_hdr->proximity_domain = proximity_domain;
1691
1692 return 0;
1693}
1694
1695#ifdef CONFIG_X86_64
1696static int kfd_fill_iolink_info_for_cpu(int numa_node_id, int *avail_size,
1697 uint32_t *num_entries,
1698 struct crat_subtype_iolink *sub_type_hdr)
1699{
1700 int nid;
1701 struct cpuinfo_x86 *c = &cpu_data(0);
1702 uint8_t link_type;
1703
1704 if (c->x86_vendor == X86_VENDOR_AMD)
1705 link_type = CRAT_IOLINK_TYPE_HYPERTRANSPORT;
1706 else
1707 link_type = CRAT_IOLINK_TYPE_QPI_1_1;
1708
1709 *num_entries = 0;
1710
1711 /* Create IO links from this node to other CPU nodes */
1712 for_each_online_node(nid) {
1713 if (nid == numa_node_id) /* node itself */
1714 continue;
1715
1716 *avail_size -= sizeof(struct crat_subtype_iolink);
1717 if (*avail_size < 0)
1718 return -ENOMEM;
1719
1720 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
1721
1722 /* Fill in subtype header data */
1723 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
1724 sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
1725 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
1726
1727 /* Fill in IO link data */
1728 sub_type_hdr->proximity_domain_from = numa_node_id;
1729 sub_type_hdr->proximity_domain_to = nid;
1730 sub_type_hdr->io_interface_type = link_type;
1731
1732 (*num_entries)++;
1733 sub_type_hdr++;
1734 }
1735
1736 return 0;
1737}
1738#endif
1739
1740/* kfd_create_vcrat_image_cpu - Create Virtual CRAT for CPU
1741 *
1742 * @pcrat_image: Fill in VCRAT for CPU
1743 * @size: [IN] allocated size of crat_image.
1744 * [OUT] actual size of data filled in crat_image
1745 */
1746static int kfd_create_vcrat_image_cpu(void *pcrat_image, size_t *size)
1747{
1748 struct crat_header *crat_table = (struct crat_header *)pcrat_image;
1749 struct acpi_table_header *acpi_table;
1750 acpi_status status;
1751 struct crat_subtype_generic *sub_type_hdr;
1752 int avail_size = *size;
1753 int numa_node_id;
1754#ifdef CONFIG_X86_64
1755 uint32_t entries = 0;
1756#endif
1757 int ret = 0;
1758
1759 if (!pcrat_image)
1760 return -EINVAL;
1761
1762 /* Fill in CRAT Header.
1763 * Modify length and total_entries as subunits are added.
1764 */
1765 avail_size -= sizeof(struct crat_header);
1766 if (avail_size < 0)
1767 return -ENOMEM;
1768
1769 memset(crat_table, 0, sizeof(struct crat_header));
1770 memcpy(&crat_table->signature, CRAT_SIGNATURE,
1771 sizeof(crat_table->signature));
1772 crat_table->length = sizeof(struct crat_header);
1773
1774 status = acpi_get_table("DSDT", 0, &acpi_table);
1775 if (status != AE_OK)
1776 pr_warn("DSDT table not found for OEM information\n");
1777 else {
1778 crat_table->oem_revision = acpi_table->revision;
1779 memcpy(crat_table->oem_id, acpi_table->oem_id,
1780 CRAT_OEMID_LENGTH);
1781 memcpy(crat_table->oem_table_id, acpi_table->oem_table_id,
1782 CRAT_OEMTABLEID_LENGTH);
1783 acpi_put_table(acpi_table);
1784 }
1785 crat_table->total_entries = 0;
1786 crat_table->num_domains = 0;
1787
1788 sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1);
1789
1790 for_each_online_node(numa_node_id) {
1791 if (kfd_numa_node_to_apic_id(numa_node_id) == -1)
1792 continue;
1793
1794 /* Fill in Subtype: Compute Unit */
1795 ret = kfd_fill_cu_for_cpu(numa_node_id, &avail_size,
1796 crat_table->num_domains,
1797 (struct crat_subtype_computeunit *)sub_type_hdr);
1798 if (ret < 0)
1799 return ret;
1800 crat_table->length += sub_type_hdr->length;
1801 crat_table->total_entries++;
1802
1803 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1804 sub_type_hdr->length);
1805
1806 /* Fill in Subtype: Memory */
1807 ret = kfd_fill_mem_info_for_cpu(numa_node_id, &avail_size,
1808 crat_table->num_domains,
1809 (struct crat_subtype_memory *)sub_type_hdr);
1810 if (ret < 0)
1811 return ret;
1812 crat_table->length += sub_type_hdr->length;
1813 crat_table->total_entries++;
1814
1815 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1816 sub_type_hdr->length);
1817
1818 /* Fill in Subtype: IO Link */
1819#ifdef CONFIG_X86_64
1820 ret = kfd_fill_iolink_info_for_cpu(numa_node_id, &avail_size,
1821 &entries,
1822 (struct crat_subtype_iolink *)sub_type_hdr);
1823 if (ret < 0)
1824 return ret;
1825
1826 if (entries) {
1827 crat_table->length += (sub_type_hdr->length * entries);
1828 crat_table->total_entries += entries;
1829
1830 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
1831 sub_type_hdr->length * entries);
1832 }
1833#else
1834 pr_info("IO link not available for non x86 platforms\n");
1835#endif
1836
1837 crat_table->num_domains++;
1838 }
1839
1840 /* TODO: Add cache Subtype for CPU.
1841 * Currently, CPU cache information is available in function
1842 * detect_cache_attributes(cpu) defined in the file
1843 * ./arch/x86/kernel/cpu/intel_cacheinfo.c. This function is not
1844 * exported and to get the same information the code needs to be
1845 * duplicated.
1846 */
1847
1848 *size = crat_table->length;
1849 pr_info("Virtual CRAT table created for CPU\n");
1850
1851 return 0;
1852}
1853
1854static int kfd_fill_gpu_memory_affinity(int *avail_size,
1855 struct kfd_node *kdev, uint8_t type, uint64_t size,
1856 struct crat_subtype_memory *sub_type_hdr,
1857 uint32_t proximity_domain,
1858 const struct kfd_local_mem_info *local_mem_info)
1859{
1860 *avail_size -= sizeof(struct crat_subtype_memory);
1861 if (*avail_size < 0)
1862 return -ENOMEM;
1863
1864 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_memory));
1865 sub_type_hdr->type = CRAT_SUBTYPE_MEMORY_AFFINITY;
1866 sub_type_hdr->length = sizeof(struct crat_subtype_memory);
1867 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED;
1868
1869 sub_type_hdr->proximity_domain = proximity_domain;
1870
1871 pr_debug("Fill gpu memory affinity - type 0x%x size 0x%llx\n",
1872 type, size);
1873
1874 sub_type_hdr->length_low = lower_32_bits(size);
1875 sub_type_hdr->length_high = upper_32_bits(size);
1876
1877 sub_type_hdr->width = local_mem_info->vram_width;
1878 sub_type_hdr->visibility_type = type;
1879
1880 return 0;
1881}
1882
1883#ifdef CONFIG_ACPI_NUMA
1884static void kfd_find_numa_node_in_srat(struct kfd_node *kdev)
1885{
1886 struct acpi_table_header *table_header = NULL;
1887 struct acpi_subtable_header *sub_header = NULL;
1888 unsigned long table_end, subtable_len;
1889 u32 pci_id = pci_domain_nr(kdev->adev->pdev->bus) << 16 |
1890 pci_dev_id(kdev->adev->pdev);
1891 u32 bdf;
1892 acpi_status status;
1893 struct acpi_srat_cpu_affinity *cpu;
1894 struct acpi_srat_generic_affinity *gpu;
1895 int pxm = 0, max_pxm = 0;
1896 int numa_node = NUMA_NO_NODE;
1897 bool found = false;
1898
1899 /* Fetch the SRAT table from ACPI */
1900 status = acpi_get_table(ACPI_SIG_SRAT, 0, &table_header);
1901 if (status == AE_NOT_FOUND) {
1902 pr_warn("SRAT table not found\n");
1903 return;
1904 } else if (ACPI_FAILURE(status)) {
1905 const char *err = acpi_format_exception(status);
1906 pr_err("SRAT table error: %s\n", err);
1907 return;
1908 }
1909
1910 table_end = (unsigned long)table_header + table_header->length;
1911
1912 /* Parse all entries looking for a match. */
1913 sub_header = (struct acpi_subtable_header *)
1914 ((unsigned long)table_header +
1915 sizeof(struct acpi_table_srat));
1916 subtable_len = sub_header->length;
1917
1918 while (((unsigned long)sub_header) + subtable_len < table_end) {
1919 /*
1920 * If length is 0, break from this loop to avoid
1921 * infinite loop.
1922 */
1923 if (subtable_len == 0) {
1924 pr_err("SRAT invalid zero length\n");
1925 break;
1926 }
1927
1928 switch (sub_header->type) {
1929 case ACPI_SRAT_TYPE_CPU_AFFINITY:
1930 cpu = (struct acpi_srat_cpu_affinity *)sub_header;
1931 pxm = *((u32 *)cpu->proximity_domain_hi) << 8 |
1932 cpu->proximity_domain_lo;
1933 if (pxm > max_pxm)
1934 max_pxm = pxm;
1935 break;
1936 case ACPI_SRAT_TYPE_GENERIC_AFFINITY:
1937 gpu = (struct acpi_srat_generic_affinity *)sub_header;
1938 bdf = *((u16 *)(&gpu->device_handle[0])) << 16 |
1939 *((u16 *)(&gpu->device_handle[2]));
1940 if (bdf == pci_id) {
1941 found = true;
1942 numa_node = pxm_to_node(gpu->proximity_domain);
1943 }
1944 break;
1945 default:
1946 break;
1947 }
1948
1949 if (found)
1950 break;
1951
1952 sub_header = (struct acpi_subtable_header *)
1953 ((unsigned long)sub_header + subtable_len);
1954 subtable_len = sub_header->length;
1955 }
1956
1957 acpi_put_table(table_header);
1958
1959 /* Workaround bad cpu-gpu binding case */
1960 if (found && (numa_node < 0 ||
1961 numa_node > pxm_to_node(max_pxm)))
1962 numa_node = 0;
1963
1964 if (numa_node != NUMA_NO_NODE)
1965 set_dev_node(&kdev->adev->pdev->dev, numa_node);
1966}
1967#endif
1968
1969#define KFD_CRAT_INTRA_SOCKET_WEIGHT 13
1970#define KFD_CRAT_XGMI_WEIGHT 15
1971
1972/* kfd_fill_gpu_direct_io_link - Fill in direct io link from GPU
1973 * to its NUMA node
1974 * @avail_size: Available size in the memory
1975 * @kdev - [IN] GPU device
1976 * @sub_type_hdr: Memory into which io link info will be filled in
1977 * @proximity_domain - proximity domain of the GPU node
1978 *
1979 * Return 0 if successful else return -ve value
1980 */
1981static int kfd_fill_gpu_direct_io_link_to_cpu(int *avail_size,
1982 struct kfd_node *kdev,
1983 struct crat_subtype_iolink *sub_type_hdr,
1984 uint32_t proximity_domain)
1985{
1986 *avail_size -= sizeof(struct crat_subtype_iolink);
1987 if (*avail_size < 0)
1988 return -ENOMEM;
1989
1990 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
1991
1992 /* Fill in subtype header data */
1993 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
1994 sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
1995 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED;
1996 if (kfd_dev_is_large_bar(kdev))
1997 sub_type_hdr->flags |= CRAT_IOLINK_FLAGS_BI_DIRECTIONAL;
1998
1999 /* Fill in IOLINK subtype.
2000 * TODO: Fill-in other fields of iolink subtype
2001 */
2002 if (kdev->adev->gmc.xgmi.connected_to_cpu ||
2003 (KFD_GC_VERSION(kdev) == IP_VERSION(9, 4, 3) &&
2004 kdev->adev->smuio.funcs->get_pkg_type(kdev->adev) ==
2005 AMDGPU_PKG_TYPE_APU)) {
2006 bool ext_cpu = KFD_GC_VERSION(kdev) != IP_VERSION(9, 4, 3);
2007 int mem_bw = 819200, weight = ext_cpu ? KFD_CRAT_XGMI_WEIGHT :
2008 KFD_CRAT_INTRA_SOCKET_WEIGHT;
2009 uint32_t bandwidth = ext_cpu ? amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(
2010 kdev->adev, NULL, true) : mem_bw;
2011
2012 /*
2013 * with host gpu xgmi link, host can access gpu memory whether
2014 * or not pcie bar type is large, so always create bidirectional
2015 * io link.
2016 */
2017 sub_type_hdr->flags |= CRAT_IOLINK_FLAGS_BI_DIRECTIONAL;
2018 sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_XGMI;
2019 sub_type_hdr->weight_xgmi = weight;
2020 sub_type_hdr->minimum_bandwidth_mbs = bandwidth;
2021 sub_type_hdr->maximum_bandwidth_mbs = bandwidth;
2022 } else {
2023 sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_PCIEXPRESS;
2024 sub_type_hdr->minimum_bandwidth_mbs =
2025 amdgpu_amdkfd_get_pcie_bandwidth_mbytes(kdev->adev, true);
2026 sub_type_hdr->maximum_bandwidth_mbs =
2027 amdgpu_amdkfd_get_pcie_bandwidth_mbytes(kdev->adev, false);
2028 }
2029
2030 sub_type_hdr->proximity_domain_from = proximity_domain;
2031
2032#ifdef CONFIG_ACPI_NUMA
2033 if (kdev->adev->pdev->dev.numa_node == NUMA_NO_NODE &&
2034 num_possible_nodes() > 1)
2035 kfd_find_numa_node_in_srat(kdev);
2036#endif
2037#ifdef CONFIG_NUMA
2038 if (kdev->adev->pdev->dev.numa_node == NUMA_NO_NODE)
2039 sub_type_hdr->proximity_domain_to = 0;
2040 else
2041 sub_type_hdr->proximity_domain_to = kdev->adev->pdev->dev.numa_node;
2042#else
2043 sub_type_hdr->proximity_domain_to = 0;
2044#endif
2045 return 0;
2046}
2047
2048static int kfd_fill_gpu_xgmi_link_to_gpu(int *avail_size,
2049 struct kfd_node *kdev,
2050 struct kfd_node *peer_kdev,
2051 struct crat_subtype_iolink *sub_type_hdr,
2052 uint32_t proximity_domain_from,
2053 uint32_t proximity_domain_to)
2054{
2055 bool use_ta_info = kdev->kfd->num_nodes == 1;
2056
2057 *avail_size -= sizeof(struct crat_subtype_iolink);
2058 if (*avail_size < 0)
2059 return -ENOMEM;
2060
2061 memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
2062
2063 sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
2064 sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
2065 sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED |
2066 CRAT_IOLINK_FLAGS_BI_DIRECTIONAL;
2067
2068 sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_XGMI;
2069 sub_type_hdr->proximity_domain_from = proximity_domain_from;
2070 sub_type_hdr->proximity_domain_to = proximity_domain_to;
2071
2072 if (use_ta_info) {
2073 sub_type_hdr->weight_xgmi = KFD_CRAT_XGMI_WEIGHT *
2074 amdgpu_amdkfd_get_xgmi_hops_count(kdev->adev, peer_kdev->adev);
2075 sub_type_hdr->maximum_bandwidth_mbs =
2076 amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(kdev->adev,
2077 peer_kdev->adev, false);
2078 sub_type_hdr->minimum_bandwidth_mbs = sub_type_hdr->maximum_bandwidth_mbs ?
2079 amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(kdev->adev, NULL, true) : 0;
2080 } else {
2081 bool is_single_hop = kdev->kfd == peer_kdev->kfd;
2082 int weight = is_single_hop ? KFD_CRAT_INTRA_SOCKET_WEIGHT :
2083 (2 * KFD_CRAT_INTRA_SOCKET_WEIGHT) + KFD_CRAT_XGMI_WEIGHT;
2084 int mem_bw = 819200;
2085
2086 sub_type_hdr->weight_xgmi = weight;
2087 sub_type_hdr->maximum_bandwidth_mbs = is_single_hop ? mem_bw : 0;
2088 sub_type_hdr->minimum_bandwidth_mbs = is_single_hop ? mem_bw : 0;
2089 }
2090
2091 return 0;
2092}
2093
2094/* kfd_create_vcrat_image_gpu - Create Virtual CRAT for CPU
2095 *
2096 * @pcrat_image: Fill in VCRAT for GPU
2097 * @size: [IN] allocated size of crat_image.
2098 * [OUT] actual size of data filled in crat_image
2099 */
2100static int kfd_create_vcrat_image_gpu(void *pcrat_image,
2101 size_t *size, struct kfd_node *kdev,
2102 uint32_t proximity_domain)
2103{
2104 struct crat_header *crat_table = (struct crat_header *)pcrat_image;
2105 struct amdgpu_gfx_config *gfx_info = &kdev->adev->gfx.config;
2106 struct amdgpu_cu_info *cu_info = &kdev->adev->gfx.cu_info;
2107 struct crat_subtype_generic *sub_type_hdr;
2108 struct kfd_local_mem_info local_mem_info;
2109 struct kfd_topology_device *peer_dev;
2110 struct crat_subtype_computeunit *cu;
2111 int avail_size = *size;
2112 uint32_t total_num_of_cu;
2113 uint32_t nid = 0;
2114 int ret = 0;
2115
2116 if (!pcrat_image || avail_size < VCRAT_SIZE_FOR_GPU)
2117 return -EINVAL;
2118
2119 /* Fill the CRAT Header.
2120 * Modify length and total_entries as subunits are added.
2121 */
2122 avail_size -= sizeof(struct crat_header);
2123 if (avail_size < 0)
2124 return -ENOMEM;
2125
2126 memset(crat_table, 0, sizeof(struct crat_header));
2127
2128 memcpy(&crat_table->signature, CRAT_SIGNATURE,
2129 sizeof(crat_table->signature));
2130 /* Change length as we add more subtypes*/
2131 crat_table->length = sizeof(struct crat_header);
2132 crat_table->num_domains = 1;
2133 crat_table->total_entries = 0;
2134
2135 /* Fill in Subtype: Compute Unit
2136 * First fill in the sub type header and then sub type data
2137 */
2138 avail_size -= sizeof(struct crat_subtype_computeunit);
2139 if (avail_size < 0)
2140 return -ENOMEM;
2141
2142 sub_type_hdr = (struct crat_subtype_generic *)(crat_table + 1);
2143 memset(sub_type_hdr, 0, sizeof(struct crat_subtype_computeunit));
2144
2145 sub_type_hdr->type = CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY;
2146 sub_type_hdr->length = sizeof(struct crat_subtype_computeunit);
2147 sub_type_hdr->flags = CRAT_SUBTYPE_FLAGS_ENABLED;
2148
2149 /* Fill CU subtype data */
2150 cu = (struct crat_subtype_computeunit *)sub_type_hdr;
2151 cu->flags |= CRAT_CU_FLAGS_GPU_PRESENT;
2152 cu->proximity_domain = proximity_domain;
2153
2154 cu->num_simd_per_cu = cu_info->simd_per_cu;
2155 cu->num_simd_cores = cu_info->simd_per_cu *
2156 (cu_info->number / kdev->kfd->num_nodes);
2157 cu->max_waves_simd = cu_info->max_waves_per_simd;
2158
2159 cu->wave_front_size = cu_info->wave_front_size;
2160 cu->array_count = gfx_info->max_sh_per_se *
2161 gfx_info->max_shader_engines;
2162 total_num_of_cu = (cu->array_count * gfx_info->max_cu_per_sh);
2163 cu->processor_id_low = get_and_inc_gpu_processor_id(total_num_of_cu);
2164 cu->num_cu_per_array = gfx_info->max_cu_per_sh;
2165 cu->max_slots_scatch_cu = cu_info->max_scratch_slots_per_cu;
2166 cu->num_banks = gfx_info->max_shader_engines;
2167 cu->lds_size_in_kb = cu_info->lds_size;
2168
2169 cu->hsa_capability = 0;
2170
2171 crat_table->length += sub_type_hdr->length;
2172 crat_table->total_entries++;
2173
2174 /* Fill in Subtype: Memory. Only on systems with large BAR (no
2175 * private FB), report memory as public. On other systems
2176 * report the total FB size (public+private) as a single
2177 * private heap.
2178 */
2179 local_mem_info = kdev->local_mem_info;
2180 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
2181 sub_type_hdr->length);
2182
2183 if (kdev->adev->debug_largebar)
2184 local_mem_info.local_mem_size_private = 0;
2185
2186 if (local_mem_info.local_mem_size_private == 0)
2187 ret = kfd_fill_gpu_memory_affinity(&avail_size,
2188 kdev, HSA_MEM_HEAP_TYPE_FB_PUBLIC,
2189 local_mem_info.local_mem_size_public,
2190 (struct crat_subtype_memory *)sub_type_hdr,
2191 proximity_domain,
2192 &local_mem_info);
2193 else
2194 ret = kfd_fill_gpu_memory_affinity(&avail_size,
2195 kdev, HSA_MEM_HEAP_TYPE_FB_PRIVATE,
2196 local_mem_info.local_mem_size_public +
2197 local_mem_info.local_mem_size_private,
2198 (struct crat_subtype_memory *)sub_type_hdr,
2199 proximity_domain,
2200 &local_mem_info);
2201 if (ret < 0)
2202 return ret;
2203
2204 crat_table->length += sizeof(struct crat_subtype_memory);
2205 crat_table->total_entries++;
2206
2207 /* Fill in Subtype: IO_LINKS
2208 * Only direct links are added here which is Link from GPU to
2209 * its NUMA node. Indirect links are added by userspace.
2210 */
2211 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
2212 sub_type_hdr->length);
2213 ret = kfd_fill_gpu_direct_io_link_to_cpu(&avail_size, kdev,
2214 (struct crat_subtype_iolink *)sub_type_hdr, proximity_domain);
2215
2216 if (ret < 0)
2217 return ret;
2218
2219 crat_table->length += sub_type_hdr->length;
2220 crat_table->total_entries++;
2221
2222
2223 /* Fill in Subtype: IO_LINKS
2224 * Direct links from GPU to other GPUs through xGMI.
2225 * We will loop GPUs that already be processed (with lower value
2226 * of proximity_domain), add the link for the GPUs with same
2227 * hive id (from this GPU to other GPU) . The reversed iolink
2228 * (from other GPU to this GPU) will be added
2229 * in kfd_parse_subtype_iolink.
2230 */
2231 if (kdev->kfd->hive_id) {
2232 for (nid = 0; nid < proximity_domain; ++nid) {
2233 peer_dev = kfd_topology_device_by_proximity_domain_no_lock(nid);
2234 if (!peer_dev->gpu)
2235 continue;
2236 if (peer_dev->gpu->kfd->hive_id != kdev->kfd->hive_id)
2237 continue;
2238 sub_type_hdr = (typeof(sub_type_hdr))(
2239 (char *)sub_type_hdr +
2240 sizeof(struct crat_subtype_iolink));
2241 ret = kfd_fill_gpu_xgmi_link_to_gpu(
2242 &avail_size, kdev, peer_dev->gpu,
2243 (struct crat_subtype_iolink *)sub_type_hdr,
2244 proximity_domain, nid);
2245 if (ret < 0)
2246 return ret;
2247 crat_table->length += sub_type_hdr->length;
2248 crat_table->total_entries++;
2249 }
2250 }
2251 *size = crat_table->length;
2252 pr_info("Virtual CRAT table created for GPU\n");
2253
2254 return ret;
2255}
2256
2257/* kfd_create_crat_image_virtual - Allocates memory for CRAT image and
2258 * creates a Virtual CRAT (VCRAT) image
2259 *
2260 * NOTE: Call kfd_destroy_crat_image to free CRAT image memory
2261 *
2262 * @crat_image: VCRAT image created because ACPI does not have a
2263 * CRAT for this device
2264 * @size: [OUT] size of virtual crat_image
2265 * @flags: COMPUTE_UNIT_CPU - Create VCRAT for CPU device
2266 * COMPUTE_UNIT_GPU - Create VCRAT for GPU
2267 * (COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU) - Create VCRAT for APU
2268 * -- this option is not currently implemented.
2269 * The assumption is that all AMD APUs will have CRAT
2270 * @kdev: Valid kfd_node required if flags contain COMPUTE_UNIT_GPU
2271 *
2272 * Return 0 if successful else return -ve value
2273 */
2274int kfd_create_crat_image_virtual(void **crat_image, size_t *size,
2275 int flags, struct kfd_node *kdev,
2276 uint32_t proximity_domain)
2277{
2278 void *pcrat_image = NULL;
2279 int ret = 0, num_nodes;
2280 size_t dyn_size;
2281
2282 if (!crat_image)
2283 return -EINVAL;
2284
2285 *crat_image = NULL;
2286
2287 /* Allocate the CPU Virtual CRAT size based on the number of online
2288 * nodes. Allocate VCRAT_SIZE_FOR_GPU for GPU virtual CRAT image.
2289 * This should cover all the current conditions. A check is put not
2290 * to overwrite beyond allocated size for GPUs
2291 */
2292 switch (flags) {
2293 case COMPUTE_UNIT_CPU:
2294 num_nodes = num_online_nodes();
2295 dyn_size = sizeof(struct crat_header) +
2296 num_nodes * (sizeof(struct crat_subtype_computeunit) +
2297 sizeof(struct crat_subtype_memory) +
2298 (num_nodes - 1) * sizeof(struct crat_subtype_iolink));
2299 pcrat_image = kvmalloc(dyn_size, GFP_KERNEL);
2300 if (!pcrat_image)
2301 return -ENOMEM;
2302 *size = dyn_size;
2303 pr_debug("CRAT size is %ld", dyn_size);
2304 ret = kfd_create_vcrat_image_cpu(pcrat_image, size);
2305 break;
2306 case COMPUTE_UNIT_GPU:
2307 if (!kdev)
2308 return -EINVAL;
2309 pcrat_image = kvmalloc(VCRAT_SIZE_FOR_GPU, GFP_KERNEL);
2310 if (!pcrat_image)
2311 return -ENOMEM;
2312 *size = VCRAT_SIZE_FOR_GPU;
2313 ret = kfd_create_vcrat_image_gpu(pcrat_image, size, kdev,
2314 proximity_domain);
2315 break;
2316 case (COMPUTE_UNIT_CPU | COMPUTE_UNIT_GPU):
2317 /* TODO: */
2318 ret = -EINVAL;
2319 pr_err("VCRAT not implemented for APU\n");
2320 break;
2321 default:
2322 ret = -EINVAL;
2323 }
2324
2325 if (!ret)
2326 *crat_image = pcrat_image;
2327 else
2328 kvfree(pcrat_image);
2329
2330 return ret;
2331}
2332
2333
2334/* kfd_destroy_crat_image
2335 *
2336 * @crat_image: [IN] - crat_image from kfd_create_crat_image_xxx(..)
2337 *
2338 */
2339void kfd_destroy_crat_image(void *crat_image)
2340{
2341 kvfree(crat_image);
2342}