Loading...
1/*
2 * Copyright 2014 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/types.h>
24#include <linux/kernel.h>
25#include <linux/pci.h>
26#include <linux/errno.h>
27#include <linux/acpi.h>
28#include <linux/hash.h>
29#include <linux/cpufreq.h>
30#include <linux/log2.h>
31
32#include "kfd_priv.h"
33#include "kfd_crat.h"
34#include "kfd_topology.h"
35
36static struct list_head topology_device_list;
37static int topology_crat_parsed;
38static struct kfd_system_properties sys_props;
39
40static DECLARE_RWSEM(topology_lock);
41
42struct kfd_dev *kfd_device_by_id(uint32_t gpu_id)
43{
44 struct kfd_topology_device *top_dev;
45 struct kfd_dev *device = NULL;
46
47 down_read(&topology_lock);
48
49 list_for_each_entry(top_dev, &topology_device_list, list)
50 if (top_dev->gpu_id == gpu_id) {
51 device = top_dev->gpu;
52 break;
53 }
54
55 up_read(&topology_lock);
56
57 return device;
58}
59
60struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev)
61{
62 struct kfd_topology_device *top_dev;
63 struct kfd_dev *device = NULL;
64
65 down_read(&topology_lock);
66
67 list_for_each_entry(top_dev, &topology_device_list, list)
68 if (top_dev->gpu->pdev == pdev) {
69 device = top_dev->gpu;
70 break;
71 }
72
73 up_read(&topology_lock);
74
75 return device;
76}
77
78static int kfd_topology_get_crat_acpi(void *crat_image, size_t *size)
79{
80 struct acpi_table_header *crat_table;
81 acpi_status status;
82
83 if (!size)
84 return -EINVAL;
85
86 /*
87 * Fetch the CRAT table from ACPI
88 */
89 status = acpi_get_table(CRAT_SIGNATURE, 0, &crat_table);
90 if (status == AE_NOT_FOUND) {
91 pr_warn("CRAT table not found\n");
92 return -ENODATA;
93 } else if (ACPI_FAILURE(status)) {
94 const char *err = acpi_format_exception(status);
95
96 pr_err("CRAT table error: %s\n", err);
97 return -EINVAL;
98 }
99
100 if (*size >= crat_table->length && crat_image != NULL)
101 memcpy(crat_image, crat_table, crat_table->length);
102
103 *size = crat_table->length;
104
105 return 0;
106}
107
108static void kfd_populated_cu_info_cpu(struct kfd_topology_device *dev,
109 struct crat_subtype_computeunit *cu)
110{
111 BUG_ON(!dev);
112 BUG_ON(!cu);
113
114 dev->node_props.cpu_cores_count = cu->num_cpu_cores;
115 dev->node_props.cpu_core_id_base = cu->processor_id_low;
116 if (cu->hsa_capability & CRAT_CU_FLAGS_IOMMU_PRESENT)
117 dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
118
119 pr_info("CU CPU: cores=%d id_base=%d\n", cu->num_cpu_cores,
120 cu->processor_id_low);
121}
122
123static void kfd_populated_cu_info_gpu(struct kfd_topology_device *dev,
124 struct crat_subtype_computeunit *cu)
125{
126 BUG_ON(!dev);
127 BUG_ON(!cu);
128
129 dev->node_props.simd_id_base = cu->processor_id_low;
130 dev->node_props.simd_count = cu->num_simd_cores;
131 dev->node_props.lds_size_in_kb = cu->lds_size_in_kb;
132 dev->node_props.max_waves_per_simd = cu->max_waves_simd;
133 dev->node_props.wave_front_size = cu->wave_front_size;
134 dev->node_props.mem_banks_count = cu->num_banks;
135 dev->node_props.array_count = cu->num_arrays;
136 dev->node_props.cu_per_simd_array = cu->num_cu_per_array;
137 dev->node_props.simd_per_cu = cu->num_simd_per_cu;
138 dev->node_props.max_slots_scratch_cu = cu->max_slots_scatch_cu;
139 if (cu->hsa_capability & CRAT_CU_FLAGS_HOT_PLUGGABLE)
140 dev->node_props.capability |= HSA_CAP_HOT_PLUGGABLE;
141 pr_info("CU GPU: simds=%d id_base=%d\n", cu->num_simd_cores,
142 cu->processor_id_low);
143}
144
145/* kfd_parse_subtype_cu is called when the topology mutex is already acquired */
146static int kfd_parse_subtype_cu(struct crat_subtype_computeunit *cu)
147{
148 struct kfd_topology_device *dev;
149 int i = 0;
150
151 BUG_ON(!cu);
152
153 pr_info("Found CU entry in CRAT table with proximity_domain=%d caps=%x\n",
154 cu->proximity_domain, cu->hsa_capability);
155 list_for_each_entry(dev, &topology_device_list, list) {
156 if (cu->proximity_domain == i) {
157 if (cu->flags & CRAT_CU_FLAGS_CPU_PRESENT)
158 kfd_populated_cu_info_cpu(dev, cu);
159
160 if (cu->flags & CRAT_CU_FLAGS_GPU_PRESENT)
161 kfd_populated_cu_info_gpu(dev, cu);
162 break;
163 }
164 i++;
165 }
166
167 return 0;
168}
169
170/*
171 * kfd_parse_subtype_mem is called when the topology mutex is
172 * already acquired
173 */
174static int kfd_parse_subtype_mem(struct crat_subtype_memory *mem)
175{
176 struct kfd_mem_properties *props;
177 struct kfd_topology_device *dev;
178 int i = 0;
179
180 BUG_ON(!mem);
181
182 pr_info("Found memory entry in CRAT table with proximity_domain=%d\n",
183 mem->promixity_domain);
184 list_for_each_entry(dev, &topology_device_list, list) {
185 if (mem->promixity_domain == i) {
186 props = kfd_alloc_struct(props);
187 if (props == NULL)
188 return -ENOMEM;
189
190 if (dev->node_props.cpu_cores_count == 0)
191 props->heap_type = HSA_MEM_HEAP_TYPE_FB_PRIVATE;
192 else
193 props->heap_type = HSA_MEM_HEAP_TYPE_SYSTEM;
194
195 if (mem->flags & CRAT_MEM_FLAGS_HOT_PLUGGABLE)
196 props->flags |= HSA_MEM_FLAGS_HOT_PLUGGABLE;
197 if (mem->flags & CRAT_MEM_FLAGS_NON_VOLATILE)
198 props->flags |= HSA_MEM_FLAGS_NON_VOLATILE;
199
200 props->size_in_bytes =
201 ((uint64_t)mem->length_high << 32) +
202 mem->length_low;
203 props->width = mem->width;
204
205 dev->mem_bank_count++;
206 list_add_tail(&props->list, &dev->mem_props);
207
208 break;
209 }
210 i++;
211 }
212
213 return 0;
214}
215
216/*
217 * kfd_parse_subtype_cache is called when the topology mutex
218 * is already acquired
219 */
220static int kfd_parse_subtype_cache(struct crat_subtype_cache *cache)
221{
222 struct kfd_cache_properties *props;
223 struct kfd_topology_device *dev;
224 uint32_t id;
225
226 BUG_ON(!cache);
227
228 id = cache->processor_id_low;
229
230 pr_info("Found cache entry in CRAT table with processor_id=%d\n", id);
231 list_for_each_entry(dev, &topology_device_list, list)
232 if (id == dev->node_props.cpu_core_id_base ||
233 id == dev->node_props.simd_id_base) {
234 props = kfd_alloc_struct(props);
235 if (props == NULL)
236 return -ENOMEM;
237
238 props->processor_id_low = id;
239 props->cache_level = cache->cache_level;
240 props->cache_size = cache->cache_size;
241 props->cacheline_size = cache->cache_line_size;
242 props->cachelines_per_tag = cache->lines_per_tag;
243 props->cache_assoc = cache->associativity;
244 props->cache_latency = cache->cache_latency;
245
246 if (cache->flags & CRAT_CACHE_FLAGS_DATA_CACHE)
247 props->cache_type |= HSA_CACHE_TYPE_DATA;
248 if (cache->flags & CRAT_CACHE_FLAGS_INST_CACHE)
249 props->cache_type |= HSA_CACHE_TYPE_INSTRUCTION;
250 if (cache->flags & CRAT_CACHE_FLAGS_CPU_CACHE)
251 props->cache_type |= HSA_CACHE_TYPE_CPU;
252 if (cache->flags & CRAT_CACHE_FLAGS_SIMD_CACHE)
253 props->cache_type |= HSA_CACHE_TYPE_HSACU;
254
255 dev->cache_count++;
256 dev->node_props.caches_count++;
257 list_add_tail(&props->list, &dev->cache_props);
258
259 break;
260 }
261
262 return 0;
263}
264
265/*
266 * kfd_parse_subtype_iolink is called when the topology mutex
267 * is already acquired
268 */
269static int kfd_parse_subtype_iolink(struct crat_subtype_iolink *iolink)
270{
271 struct kfd_iolink_properties *props;
272 struct kfd_topology_device *dev;
273 uint32_t i = 0;
274 uint32_t id_from;
275 uint32_t id_to;
276
277 BUG_ON(!iolink);
278
279 id_from = iolink->proximity_domain_from;
280 id_to = iolink->proximity_domain_to;
281
282 pr_info("Found IO link entry in CRAT table with id_from=%d\n", id_from);
283 list_for_each_entry(dev, &topology_device_list, list) {
284 if (id_from == i) {
285 props = kfd_alloc_struct(props);
286 if (props == NULL)
287 return -ENOMEM;
288
289 props->node_from = id_from;
290 props->node_to = id_to;
291 props->ver_maj = iolink->version_major;
292 props->ver_min = iolink->version_minor;
293
294 /*
295 * weight factor (derived from CDIR), currently always 1
296 */
297 props->weight = 1;
298
299 props->min_latency = iolink->minimum_latency;
300 props->max_latency = iolink->maximum_latency;
301 props->min_bandwidth = iolink->minimum_bandwidth_mbs;
302 props->max_bandwidth = iolink->maximum_bandwidth_mbs;
303 props->rec_transfer_size =
304 iolink->recommended_transfer_size;
305
306 dev->io_link_count++;
307 dev->node_props.io_links_count++;
308 list_add_tail(&props->list, &dev->io_link_props);
309
310 break;
311 }
312 i++;
313 }
314
315 return 0;
316}
317
318static int kfd_parse_subtype(struct crat_subtype_generic *sub_type_hdr)
319{
320 struct crat_subtype_computeunit *cu;
321 struct crat_subtype_memory *mem;
322 struct crat_subtype_cache *cache;
323 struct crat_subtype_iolink *iolink;
324 int ret = 0;
325
326 BUG_ON(!sub_type_hdr);
327
328 switch (sub_type_hdr->type) {
329 case CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY:
330 cu = (struct crat_subtype_computeunit *)sub_type_hdr;
331 ret = kfd_parse_subtype_cu(cu);
332 break;
333 case CRAT_SUBTYPE_MEMORY_AFFINITY:
334 mem = (struct crat_subtype_memory *)sub_type_hdr;
335 ret = kfd_parse_subtype_mem(mem);
336 break;
337 case CRAT_SUBTYPE_CACHE_AFFINITY:
338 cache = (struct crat_subtype_cache *)sub_type_hdr;
339 ret = kfd_parse_subtype_cache(cache);
340 break;
341 case CRAT_SUBTYPE_TLB_AFFINITY:
342 /*
343 * For now, nothing to do here
344 */
345 pr_info("Found TLB entry in CRAT table (not processing)\n");
346 break;
347 case CRAT_SUBTYPE_CCOMPUTE_AFFINITY:
348 /*
349 * For now, nothing to do here
350 */
351 pr_info("Found CCOMPUTE entry in CRAT table (not processing)\n");
352 break;
353 case CRAT_SUBTYPE_IOLINK_AFFINITY:
354 iolink = (struct crat_subtype_iolink *)sub_type_hdr;
355 ret = kfd_parse_subtype_iolink(iolink);
356 break;
357 default:
358 pr_warn("Unknown subtype (%d) in CRAT\n",
359 sub_type_hdr->type);
360 }
361
362 return ret;
363}
364
365static void kfd_release_topology_device(struct kfd_topology_device *dev)
366{
367 struct kfd_mem_properties *mem;
368 struct kfd_cache_properties *cache;
369 struct kfd_iolink_properties *iolink;
370
371 BUG_ON(!dev);
372
373 list_del(&dev->list);
374
375 while (dev->mem_props.next != &dev->mem_props) {
376 mem = container_of(dev->mem_props.next,
377 struct kfd_mem_properties, list);
378 list_del(&mem->list);
379 kfree(mem);
380 }
381
382 while (dev->cache_props.next != &dev->cache_props) {
383 cache = container_of(dev->cache_props.next,
384 struct kfd_cache_properties, list);
385 list_del(&cache->list);
386 kfree(cache);
387 }
388
389 while (dev->io_link_props.next != &dev->io_link_props) {
390 iolink = container_of(dev->io_link_props.next,
391 struct kfd_iolink_properties, list);
392 list_del(&iolink->list);
393 kfree(iolink);
394 }
395
396 kfree(dev);
397
398 sys_props.num_devices--;
399}
400
401static void kfd_release_live_view(void)
402{
403 struct kfd_topology_device *dev;
404
405 while (topology_device_list.next != &topology_device_list) {
406 dev = container_of(topology_device_list.next,
407 struct kfd_topology_device, list);
408 kfd_release_topology_device(dev);
409}
410
411 memset(&sys_props, 0, sizeof(sys_props));
412}
413
414static struct kfd_topology_device *kfd_create_topology_device(void)
415{
416 struct kfd_topology_device *dev;
417
418 dev = kfd_alloc_struct(dev);
419 if (dev == NULL) {
420 pr_err("No memory to allocate a topology device");
421 return NULL;
422 }
423
424 INIT_LIST_HEAD(&dev->mem_props);
425 INIT_LIST_HEAD(&dev->cache_props);
426 INIT_LIST_HEAD(&dev->io_link_props);
427
428 list_add_tail(&dev->list, &topology_device_list);
429 sys_props.num_devices++;
430
431 return dev;
432}
433
434static int kfd_parse_crat_table(void *crat_image)
435{
436 struct kfd_topology_device *top_dev;
437 struct crat_subtype_generic *sub_type_hdr;
438 uint16_t node_id;
439 int ret;
440 struct crat_header *crat_table = (struct crat_header *)crat_image;
441 uint16_t num_nodes;
442 uint32_t image_len;
443
444 if (!crat_image)
445 return -EINVAL;
446
447 num_nodes = crat_table->num_domains;
448 image_len = crat_table->length;
449
450 pr_info("Parsing CRAT table with %d nodes\n", num_nodes);
451
452 for (node_id = 0; node_id < num_nodes; node_id++) {
453 top_dev = kfd_create_topology_device();
454 if (!top_dev) {
455 kfd_release_live_view();
456 return -ENOMEM;
457 }
458 }
459
460 sys_props.platform_id =
461 (*((uint64_t *)crat_table->oem_id)) & CRAT_OEMID_64BIT_MASK;
462 sys_props.platform_oem = *((uint64_t *)crat_table->oem_table_id);
463 sys_props.platform_rev = crat_table->revision;
464
465 sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1);
466 while ((char *)sub_type_hdr + sizeof(struct crat_subtype_generic) <
467 ((char *)crat_image) + image_len) {
468 if (sub_type_hdr->flags & CRAT_SUBTYPE_FLAGS_ENABLED) {
469 ret = kfd_parse_subtype(sub_type_hdr);
470 if (ret != 0) {
471 kfd_release_live_view();
472 return ret;
473 }
474 }
475
476 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
477 sub_type_hdr->length);
478 }
479
480 sys_props.generation_count++;
481 topology_crat_parsed = 1;
482
483 return 0;
484}
485
486
487#define sysfs_show_gen_prop(buffer, fmt, ...) \
488 snprintf(buffer, PAGE_SIZE, "%s"fmt, buffer, __VA_ARGS__)
489#define sysfs_show_32bit_prop(buffer, name, value) \
490 sysfs_show_gen_prop(buffer, "%s %u\n", name, value)
491#define sysfs_show_64bit_prop(buffer, name, value) \
492 sysfs_show_gen_prop(buffer, "%s %llu\n", name, value)
493#define sysfs_show_32bit_val(buffer, value) \
494 sysfs_show_gen_prop(buffer, "%u\n", value)
495#define sysfs_show_str_val(buffer, value) \
496 sysfs_show_gen_prop(buffer, "%s\n", value)
497
498static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr,
499 char *buffer)
500{
501 ssize_t ret;
502
503 /* Making sure that the buffer is an empty string */
504 buffer[0] = 0;
505
506 if (attr == &sys_props.attr_genid) {
507 ret = sysfs_show_32bit_val(buffer, sys_props.generation_count);
508 } else if (attr == &sys_props.attr_props) {
509 sysfs_show_64bit_prop(buffer, "platform_oem",
510 sys_props.platform_oem);
511 sysfs_show_64bit_prop(buffer, "platform_id",
512 sys_props.platform_id);
513 ret = sysfs_show_64bit_prop(buffer, "platform_rev",
514 sys_props.platform_rev);
515 } else {
516 ret = -EINVAL;
517 }
518
519 return ret;
520}
521
522static const struct sysfs_ops sysprops_ops = {
523 .show = sysprops_show,
524};
525
526static struct kobj_type sysprops_type = {
527 .sysfs_ops = &sysprops_ops,
528};
529
530static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr,
531 char *buffer)
532{
533 ssize_t ret;
534 struct kfd_iolink_properties *iolink;
535
536 /* Making sure that the buffer is an empty string */
537 buffer[0] = 0;
538
539 iolink = container_of(attr, struct kfd_iolink_properties, attr);
540 sysfs_show_32bit_prop(buffer, "type", iolink->iolink_type);
541 sysfs_show_32bit_prop(buffer, "version_major", iolink->ver_maj);
542 sysfs_show_32bit_prop(buffer, "version_minor", iolink->ver_min);
543 sysfs_show_32bit_prop(buffer, "node_from", iolink->node_from);
544 sysfs_show_32bit_prop(buffer, "node_to", iolink->node_to);
545 sysfs_show_32bit_prop(buffer, "weight", iolink->weight);
546 sysfs_show_32bit_prop(buffer, "min_latency", iolink->min_latency);
547 sysfs_show_32bit_prop(buffer, "max_latency", iolink->max_latency);
548 sysfs_show_32bit_prop(buffer, "min_bandwidth", iolink->min_bandwidth);
549 sysfs_show_32bit_prop(buffer, "max_bandwidth", iolink->max_bandwidth);
550 sysfs_show_32bit_prop(buffer, "recommended_transfer_size",
551 iolink->rec_transfer_size);
552 ret = sysfs_show_32bit_prop(buffer, "flags", iolink->flags);
553
554 return ret;
555}
556
557static const struct sysfs_ops iolink_ops = {
558 .show = iolink_show,
559};
560
561static struct kobj_type iolink_type = {
562 .sysfs_ops = &iolink_ops,
563};
564
565static ssize_t mem_show(struct kobject *kobj, struct attribute *attr,
566 char *buffer)
567{
568 ssize_t ret;
569 struct kfd_mem_properties *mem;
570
571 /* Making sure that the buffer is an empty string */
572 buffer[0] = 0;
573
574 mem = container_of(attr, struct kfd_mem_properties, attr);
575 sysfs_show_32bit_prop(buffer, "heap_type", mem->heap_type);
576 sysfs_show_64bit_prop(buffer, "size_in_bytes", mem->size_in_bytes);
577 sysfs_show_32bit_prop(buffer, "flags", mem->flags);
578 sysfs_show_32bit_prop(buffer, "width", mem->width);
579 ret = sysfs_show_32bit_prop(buffer, "mem_clk_max", mem->mem_clk_max);
580
581 return ret;
582}
583
584static const struct sysfs_ops mem_ops = {
585 .show = mem_show,
586};
587
588static struct kobj_type mem_type = {
589 .sysfs_ops = &mem_ops,
590};
591
592static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr,
593 char *buffer)
594{
595 ssize_t ret;
596 uint32_t i;
597 struct kfd_cache_properties *cache;
598
599 /* Making sure that the buffer is an empty string */
600 buffer[0] = 0;
601
602 cache = container_of(attr, struct kfd_cache_properties, attr);
603 sysfs_show_32bit_prop(buffer, "processor_id_low",
604 cache->processor_id_low);
605 sysfs_show_32bit_prop(buffer, "level", cache->cache_level);
606 sysfs_show_32bit_prop(buffer, "size", cache->cache_size);
607 sysfs_show_32bit_prop(buffer, "cache_line_size", cache->cacheline_size);
608 sysfs_show_32bit_prop(buffer, "cache_lines_per_tag",
609 cache->cachelines_per_tag);
610 sysfs_show_32bit_prop(buffer, "association", cache->cache_assoc);
611 sysfs_show_32bit_prop(buffer, "latency", cache->cache_latency);
612 sysfs_show_32bit_prop(buffer, "type", cache->cache_type);
613 snprintf(buffer, PAGE_SIZE, "%ssibling_map ", buffer);
614 for (i = 0; i < KFD_TOPOLOGY_CPU_SIBLINGS; i++)
615 ret = snprintf(buffer, PAGE_SIZE, "%s%d%s",
616 buffer, cache->sibling_map[i],
617 (i == KFD_TOPOLOGY_CPU_SIBLINGS-1) ?
618 "\n" : ",");
619
620 return ret;
621}
622
623static const struct sysfs_ops cache_ops = {
624 .show = kfd_cache_show,
625};
626
627static struct kobj_type cache_type = {
628 .sysfs_ops = &cache_ops,
629};
630
631static ssize_t node_show(struct kobject *kobj, struct attribute *attr,
632 char *buffer)
633{
634 struct kfd_topology_device *dev;
635 char public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE];
636 uint32_t i;
637 uint32_t log_max_watch_addr;
638
639 /* Making sure that the buffer is an empty string */
640 buffer[0] = 0;
641
642 if (strcmp(attr->name, "gpu_id") == 0) {
643 dev = container_of(attr, struct kfd_topology_device,
644 attr_gpuid);
645 return sysfs_show_32bit_val(buffer, dev->gpu_id);
646 }
647
648 if (strcmp(attr->name, "name") == 0) {
649 dev = container_of(attr, struct kfd_topology_device,
650 attr_name);
651 for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE; i++) {
652 public_name[i] =
653 (char)dev->node_props.marketing_name[i];
654 if (dev->node_props.marketing_name[i] == 0)
655 break;
656 }
657 public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1] = 0x0;
658 return sysfs_show_str_val(buffer, public_name);
659 }
660
661 dev = container_of(attr, struct kfd_topology_device,
662 attr_props);
663 sysfs_show_32bit_prop(buffer, "cpu_cores_count",
664 dev->node_props.cpu_cores_count);
665 sysfs_show_32bit_prop(buffer, "simd_count",
666 dev->node_props.simd_count);
667
668 if (dev->mem_bank_count < dev->node_props.mem_banks_count) {
669 pr_warn("kfd: mem_banks_count truncated from %d to %d\n",
670 dev->node_props.mem_banks_count,
671 dev->mem_bank_count);
672 sysfs_show_32bit_prop(buffer, "mem_banks_count",
673 dev->mem_bank_count);
674 } else {
675 sysfs_show_32bit_prop(buffer, "mem_banks_count",
676 dev->node_props.mem_banks_count);
677 }
678
679 sysfs_show_32bit_prop(buffer, "caches_count",
680 dev->node_props.caches_count);
681 sysfs_show_32bit_prop(buffer, "io_links_count",
682 dev->node_props.io_links_count);
683 sysfs_show_32bit_prop(buffer, "cpu_core_id_base",
684 dev->node_props.cpu_core_id_base);
685 sysfs_show_32bit_prop(buffer, "simd_id_base",
686 dev->node_props.simd_id_base);
687 sysfs_show_32bit_prop(buffer, "max_waves_per_simd",
688 dev->node_props.max_waves_per_simd);
689 sysfs_show_32bit_prop(buffer, "lds_size_in_kb",
690 dev->node_props.lds_size_in_kb);
691 sysfs_show_32bit_prop(buffer, "gds_size_in_kb",
692 dev->node_props.gds_size_in_kb);
693 sysfs_show_32bit_prop(buffer, "wave_front_size",
694 dev->node_props.wave_front_size);
695 sysfs_show_32bit_prop(buffer, "array_count",
696 dev->node_props.array_count);
697 sysfs_show_32bit_prop(buffer, "simd_arrays_per_engine",
698 dev->node_props.simd_arrays_per_engine);
699 sysfs_show_32bit_prop(buffer, "cu_per_simd_array",
700 dev->node_props.cu_per_simd_array);
701 sysfs_show_32bit_prop(buffer, "simd_per_cu",
702 dev->node_props.simd_per_cu);
703 sysfs_show_32bit_prop(buffer, "max_slots_scratch_cu",
704 dev->node_props.max_slots_scratch_cu);
705 sysfs_show_32bit_prop(buffer, "vendor_id",
706 dev->node_props.vendor_id);
707 sysfs_show_32bit_prop(buffer, "device_id",
708 dev->node_props.device_id);
709 sysfs_show_32bit_prop(buffer, "location_id",
710 dev->node_props.location_id);
711
712 if (dev->gpu) {
713 log_max_watch_addr =
714 __ilog2_u32(dev->gpu->device_info->num_of_watch_points);
715
716 if (log_max_watch_addr) {
717 dev->node_props.capability |=
718 HSA_CAP_WATCH_POINTS_SUPPORTED;
719
720 dev->node_props.capability |=
721 ((log_max_watch_addr <<
722 HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) &
723 HSA_CAP_WATCH_POINTS_TOTALBITS_MASK);
724 }
725
726 sysfs_show_32bit_prop(buffer, "max_engine_clk_fcompute",
727 dev->gpu->kfd2kgd->get_max_engine_clock_in_mhz(
728 dev->gpu->kgd));
729
730 sysfs_show_64bit_prop(buffer, "local_mem_size",
731 (unsigned long long int) 0);
732
733 sysfs_show_32bit_prop(buffer, "fw_version",
734 dev->gpu->kfd2kgd->get_fw_version(
735 dev->gpu->kgd,
736 KGD_ENGINE_MEC1));
737 sysfs_show_32bit_prop(buffer, "capability",
738 dev->node_props.capability);
739 }
740
741 return sysfs_show_32bit_prop(buffer, "max_engine_clk_ccompute",
742 cpufreq_quick_get_max(0)/1000);
743}
744
745static const struct sysfs_ops node_ops = {
746 .show = node_show,
747};
748
749static struct kobj_type node_type = {
750 .sysfs_ops = &node_ops,
751};
752
753static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr)
754{
755 sysfs_remove_file(kobj, attr);
756 kobject_del(kobj);
757 kobject_put(kobj);
758}
759
760static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev)
761{
762 struct kfd_iolink_properties *iolink;
763 struct kfd_cache_properties *cache;
764 struct kfd_mem_properties *mem;
765
766 BUG_ON(!dev);
767
768 if (dev->kobj_iolink) {
769 list_for_each_entry(iolink, &dev->io_link_props, list)
770 if (iolink->kobj) {
771 kfd_remove_sysfs_file(iolink->kobj,
772 &iolink->attr);
773 iolink->kobj = NULL;
774 }
775 kobject_del(dev->kobj_iolink);
776 kobject_put(dev->kobj_iolink);
777 dev->kobj_iolink = NULL;
778 }
779
780 if (dev->kobj_cache) {
781 list_for_each_entry(cache, &dev->cache_props, list)
782 if (cache->kobj) {
783 kfd_remove_sysfs_file(cache->kobj,
784 &cache->attr);
785 cache->kobj = NULL;
786 }
787 kobject_del(dev->kobj_cache);
788 kobject_put(dev->kobj_cache);
789 dev->kobj_cache = NULL;
790 }
791
792 if (dev->kobj_mem) {
793 list_for_each_entry(mem, &dev->mem_props, list)
794 if (mem->kobj) {
795 kfd_remove_sysfs_file(mem->kobj, &mem->attr);
796 mem->kobj = NULL;
797 }
798 kobject_del(dev->kobj_mem);
799 kobject_put(dev->kobj_mem);
800 dev->kobj_mem = NULL;
801 }
802
803 if (dev->kobj_node) {
804 sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid);
805 sysfs_remove_file(dev->kobj_node, &dev->attr_name);
806 sysfs_remove_file(dev->kobj_node, &dev->attr_props);
807 kobject_del(dev->kobj_node);
808 kobject_put(dev->kobj_node);
809 dev->kobj_node = NULL;
810 }
811}
812
813static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev,
814 uint32_t id)
815{
816 struct kfd_iolink_properties *iolink;
817 struct kfd_cache_properties *cache;
818 struct kfd_mem_properties *mem;
819 int ret;
820 uint32_t i;
821
822 BUG_ON(!dev);
823
824 /*
825 * Creating the sysfs folders
826 */
827 BUG_ON(dev->kobj_node);
828 dev->kobj_node = kfd_alloc_struct(dev->kobj_node);
829 if (!dev->kobj_node)
830 return -ENOMEM;
831
832 ret = kobject_init_and_add(dev->kobj_node, &node_type,
833 sys_props.kobj_nodes, "%d", id);
834 if (ret < 0)
835 return ret;
836
837 dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node);
838 if (!dev->kobj_mem)
839 return -ENOMEM;
840
841 dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node);
842 if (!dev->kobj_cache)
843 return -ENOMEM;
844
845 dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node);
846 if (!dev->kobj_iolink)
847 return -ENOMEM;
848
849 /*
850 * Creating sysfs files for node properties
851 */
852 dev->attr_gpuid.name = "gpu_id";
853 dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE;
854 sysfs_attr_init(&dev->attr_gpuid);
855 dev->attr_name.name = "name";
856 dev->attr_name.mode = KFD_SYSFS_FILE_MODE;
857 sysfs_attr_init(&dev->attr_name);
858 dev->attr_props.name = "properties";
859 dev->attr_props.mode = KFD_SYSFS_FILE_MODE;
860 sysfs_attr_init(&dev->attr_props);
861 ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid);
862 if (ret < 0)
863 return ret;
864 ret = sysfs_create_file(dev->kobj_node, &dev->attr_name);
865 if (ret < 0)
866 return ret;
867 ret = sysfs_create_file(dev->kobj_node, &dev->attr_props);
868 if (ret < 0)
869 return ret;
870
871 i = 0;
872 list_for_each_entry(mem, &dev->mem_props, list) {
873 mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
874 if (!mem->kobj)
875 return -ENOMEM;
876 ret = kobject_init_and_add(mem->kobj, &mem_type,
877 dev->kobj_mem, "%d", i);
878 if (ret < 0)
879 return ret;
880
881 mem->attr.name = "properties";
882 mem->attr.mode = KFD_SYSFS_FILE_MODE;
883 sysfs_attr_init(&mem->attr);
884 ret = sysfs_create_file(mem->kobj, &mem->attr);
885 if (ret < 0)
886 return ret;
887 i++;
888 }
889
890 i = 0;
891 list_for_each_entry(cache, &dev->cache_props, list) {
892 cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
893 if (!cache->kobj)
894 return -ENOMEM;
895 ret = kobject_init_and_add(cache->kobj, &cache_type,
896 dev->kobj_cache, "%d", i);
897 if (ret < 0)
898 return ret;
899
900 cache->attr.name = "properties";
901 cache->attr.mode = KFD_SYSFS_FILE_MODE;
902 sysfs_attr_init(&cache->attr);
903 ret = sysfs_create_file(cache->kobj, &cache->attr);
904 if (ret < 0)
905 return ret;
906 i++;
907 }
908
909 i = 0;
910 list_for_each_entry(iolink, &dev->io_link_props, list) {
911 iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
912 if (!iolink->kobj)
913 return -ENOMEM;
914 ret = kobject_init_and_add(iolink->kobj, &iolink_type,
915 dev->kobj_iolink, "%d", i);
916 if (ret < 0)
917 return ret;
918
919 iolink->attr.name = "properties";
920 iolink->attr.mode = KFD_SYSFS_FILE_MODE;
921 sysfs_attr_init(&iolink->attr);
922 ret = sysfs_create_file(iolink->kobj, &iolink->attr);
923 if (ret < 0)
924 return ret;
925 i++;
926}
927
928 return 0;
929}
930
931static int kfd_build_sysfs_node_tree(void)
932{
933 struct kfd_topology_device *dev;
934 int ret;
935 uint32_t i = 0;
936
937 list_for_each_entry(dev, &topology_device_list, list) {
938 ret = kfd_build_sysfs_node_entry(dev, i);
939 if (ret < 0)
940 return ret;
941 i++;
942 }
943
944 return 0;
945}
946
947static void kfd_remove_sysfs_node_tree(void)
948{
949 struct kfd_topology_device *dev;
950
951 list_for_each_entry(dev, &topology_device_list, list)
952 kfd_remove_sysfs_node_entry(dev);
953}
954
955static int kfd_topology_update_sysfs(void)
956{
957 int ret;
958
959 pr_info("Creating topology SYSFS entries\n");
960 if (sys_props.kobj_topology == NULL) {
961 sys_props.kobj_topology =
962 kfd_alloc_struct(sys_props.kobj_topology);
963 if (!sys_props.kobj_topology)
964 return -ENOMEM;
965
966 ret = kobject_init_and_add(sys_props.kobj_topology,
967 &sysprops_type, &kfd_device->kobj,
968 "topology");
969 if (ret < 0)
970 return ret;
971
972 sys_props.kobj_nodes = kobject_create_and_add("nodes",
973 sys_props.kobj_topology);
974 if (!sys_props.kobj_nodes)
975 return -ENOMEM;
976
977 sys_props.attr_genid.name = "generation_id";
978 sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE;
979 sysfs_attr_init(&sys_props.attr_genid);
980 ret = sysfs_create_file(sys_props.kobj_topology,
981 &sys_props.attr_genid);
982 if (ret < 0)
983 return ret;
984
985 sys_props.attr_props.name = "system_properties";
986 sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE;
987 sysfs_attr_init(&sys_props.attr_props);
988 ret = sysfs_create_file(sys_props.kobj_topology,
989 &sys_props.attr_props);
990 if (ret < 0)
991 return ret;
992 }
993
994 kfd_remove_sysfs_node_tree();
995
996 return kfd_build_sysfs_node_tree();
997}
998
999static void kfd_topology_release_sysfs(void)
1000{
1001 kfd_remove_sysfs_node_tree();
1002 if (sys_props.kobj_topology) {
1003 sysfs_remove_file(sys_props.kobj_topology,
1004 &sys_props.attr_genid);
1005 sysfs_remove_file(sys_props.kobj_topology,
1006 &sys_props.attr_props);
1007 if (sys_props.kobj_nodes) {
1008 kobject_del(sys_props.kobj_nodes);
1009 kobject_put(sys_props.kobj_nodes);
1010 sys_props.kobj_nodes = NULL;
1011 }
1012 kobject_del(sys_props.kobj_topology);
1013 kobject_put(sys_props.kobj_topology);
1014 sys_props.kobj_topology = NULL;
1015 }
1016}
1017
1018int kfd_topology_init(void)
1019{
1020 void *crat_image = NULL;
1021 size_t image_size = 0;
1022 int ret;
1023
1024 /*
1025 * Initialize the head for the topology device list
1026 */
1027 INIT_LIST_HEAD(&topology_device_list);
1028 init_rwsem(&topology_lock);
1029 topology_crat_parsed = 0;
1030
1031 memset(&sys_props, 0, sizeof(sys_props));
1032
1033 /*
1034 * Get the CRAT image from the ACPI
1035 */
1036 ret = kfd_topology_get_crat_acpi(crat_image, &image_size);
1037 if (ret == 0 && image_size > 0) {
1038 pr_info("Found CRAT image with size=%zd\n", image_size);
1039 crat_image = kmalloc(image_size, GFP_KERNEL);
1040 if (!crat_image) {
1041 ret = -ENOMEM;
1042 pr_err("No memory for allocating CRAT image\n");
1043 goto err;
1044 }
1045 ret = kfd_topology_get_crat_acpi(crat_image, &image_size);
1046
1047 if (ret == 0) {
1048 down_write(&topology_lock);
1049 ret = kfd_parse_crat_table(crat_image);
1050 if (ret == 0)
1051 ret = kfd_topology_update_sysfs();
1052 up_write(&topology_lock);
1053 } else {
1054 pr_err("Couldn't get CRAT table size from ACPI\n");
1055 }
1056 kfree(crat_image);
1057 } else if (ret == -ENODATA) {
1058 ret = 0;
1059 } else {
1060 pr_err("Couldn't get CRAT table size from ACPI\n");
1061 }
1062
1063err:
1064 pr_info("Finished initializing topology ret=%d\n", ret);
1065 return ret;
1066}
1067
1068void kfd_topology_shutdown(void)
1069{
1070 kfd_topology_release_sysfs();
1071 kfd_release_live_view();
1072}
1073
1074static void kfd_debug_print_topology(void)
1075{
1076 struct kfd_topology_device *dev;
1077 uint32_t i = 0;
1078
1079 pr_info("DEBUG PRINT OF TOPOLOGY:");
1080 list_for_each_entry(dev, &topology_device_list, list) {
1081 pr_info("Node: %d\n", i);
1082 pr_info("\tGPU assigned: %s\n", (dev->gpu ? "yes" : "no"));
1083 pr_info("\tCPU count: %d\n", dev->node_props.cpu_cores_count);
1084 pr_info("\tSIMD count: %d", dev->node_props.simd_count);
1085 i++;
1086 }
1087}
1088
1089static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu)
1090{
1091 uint32_t hashout;
1092 uint32_t buf[7];
1093 int i;
1094
1095 if (!gpu)
1096 return 0;
1097
1098 buf[0] = gpu->pdev->devfn;
1099 buf[1] = gpu->pdev->subsystem_vendor;
1100 buf[2] = gpu->pdev->subsystem_device;
1101 buf[3] = gpu->pdev->device;
1102 buf[4] = gpu->pdev->bus->number;
1103 buf[5] = (uint32_t)(gpu->kfd2kgd->get_vmem_size(gpu->kgd)
1104 & 0xffffffff);
1105 buf[6] = (uint32_t)(gpu->kfd2kgd->get_vmem_size(gpu->kgd) >> 32);
1106
1107 for (i = 0, hashout = 0; i < 7; i++)
1108 hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH);
1109
1110 return hashout;
1111}
1112
1113static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu)
1114{
1115 struct kfd_topology_device *dev;
1116 struct kfd_topology_device *out_dev = NULL;
1117
1118 BUG_ON(!gpu);
1119
1120 list_for_each_entry(dev, &topology_device_list, list)
1121 if (dev->gpu == NULL && dev->node_props.simd_count > 0) {
1122 dev->gpu = gpu;
1123 out_dev = dev;
1124 break;
1125 }
1126
1127 return out_dev;
1128}
1129
1130static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival)
1131{
1132 /*
1133 * TODO: Generate an event for thunk about the arrival/removal
1134 * of the GPU
1135 */
1136}
1137
1138int kfd_topology_add_device(struct kfd_dev *gpu)
1139{
1140 uint32_t gpu_id;
1141 struct kfd_topology_device *dev;
1142 int res;
1143
1144 BUG_ON(!gpu);
1145
1146 gpu_id = kfd_generate_gpu_id(gpu);
1147
1148 pr_debug("kfd: Adding new GPU (ID: 0x%x) to topology\n", gpu_id);
1149
1150 down_write(&topology_lock);
1151 /*
1152 * Try to assign the GPU to existing topology device (generated from
1153 * CRAT table
1154 */
1155 dev = kfd_assign_gpu(gpu);
1156 if (!dev) {
1157 pr_info("GPU was not found in the current topology. Extending.\n");
1158 kfd_debug_print_topology();
1159 dev = kfd_create_topology_device();
1160 if (!dev) {
1161 res = -ENOMEM;
1162 goto err;
1163 }
1164 dev->gpu = gpu;
1165
1166 /*
1167 * TODO: Make a call to retrieve topology information from the
1168 * GPU vBIOS
1169 */
1170
1171 /*
1172 * Update the SYSFS tree, since we added another topology device
1173 */
1174 if (kfd_topology_update_sysfs() < 0)
1175 kfd_topology_release_sysfs();
1176
1177 }
1178
1179 dev->gpu_id = gpu_id;
1180 gpu->id = gpu_id;
1181 dev->node_props.vendor_id = gpu->pdev->vendor;
1182 dev->node_props.device_id = gpu->pdev->device;
1183 dev->node_props.location_id = (gpu->pdev->bus->number << 24) +
1184 (gpu->pdev->devfn & 0xffffff);
1185 /*
1186 * TODO: Retrieve max engine clock values from KGD
1187 */
1188
1189 if (dev->gpu->device_info->asic_family == CHIP_CARRIZO) {
1190 dev->node_props.capability |= HSA_CAP_DOORBELL_PACKET_TYPE;
1191 pr_info("amdkfd: adding doorbell packet type capability\n");
1192 }
1193
1194 res = 0;
1195
1196err:
1197 up_write(&topology_lock);
1198
1199 if (res == 0)
1200 kfd_notify_gpu_change(gpu_id, 1);
1201
1202 return res;
1203}
1204
1205int kfd_topology_remove_device(struct kfd_dev *gpu)
1206{
1207 struct kfd_topology_device *dev;
1208 uint32_t gpu_id;
1209 int res = -ENODEV;
1210
1211 BUG_ON(!gpu);
1212
1213 down_write(&topology_lock);
1214
1215 list_for_each_entry(dev, &topology_device_list, list)
1216 if (dev->gpu == gpu) {
1217 gpu_id = dev->gpu_id;
1218 kfd_remove_sysfs_node_entry(dev);
1219 kfd_release_topology_device(dev);
1220 res = 0;
1221 if (kfd_topology_update_sysfs() < 0)
1222 kfd_topology_release_sysfs();
1223 break;
1224 }
1225
1226 up_write(&topology_lock);
1227
1228 if (res == 0)
1229 kfd_notify_gpu_change(gpu_id, 0);
1230
1231 return res;
1232}
1233
1234/*
1235 * When idx is out of bounds, the function will return NULL
1236 */
1237struct kfd_dev *kfd_topology_enum_kfd_devices(uint8_t idx)
1238{
1239
1240 struct kfd_topology_device *top_dev;
1241 struct kfd_dev *device = NULL;
1242 uint8_t device_idx = 0;
1243
1244 down_read(&topology_lock);
1245
1246 list_for_each_entry(top_dev, &topology_device_list, list) {
1247 if (device_idx == idx) {
1248 device = top_dev->gpu;
1249 break;
1250 }
1251
1252 device_idx++;
1253 }
1254
1255 up_read(&topology_lock);
1256
1257 return device;
1258
1259}
1// SPDX-License-Identifier: GPL-2.0 OR MIT
2/*
3 * Copyright 2014-2022 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/types.h>
25#include <linux/kernel.h>
26#include <linux/pci.h>
27#include <linux/errno.h>
28#include <linux/acpi.h>
29#include <linux/hash.h>
30#include <linux/cpufreq.h>
31#include <linux/log2.h>
32#include <linux/dmi.h>
33#include <linux/atomic.h>
34#include <linux/crc16.h>
35
36#include "kfd_priv.h"
37#include "kfd_crat.h"
38#include "kfd_topology.h"
39#include "kfd_device_queue_manager.h"
40#include "kfd_svm.h"
41#include "kfd_debug.h"
42#include "amdgpu_amdkfd.h"
43#include "amdgpu_ras.h"
44#include "amdgpu.h"
45
46/* topology_device_list - Master list of all topology devices */
47static struct list_head topology_device_list;
48static struct kfd_system_properties sys_props;
49
50static DECLARE_RWSEM(topology_lock);
51static uint32_t topology_crat_proximity_domain;
52
53struct kfd_topology_device *kfd_topology_device_by_proximity_domain_no_lock(
54 uint32_t proximity_domain)
55{
56 struct kfd_topology_device *top_dev;
57 struct kfd_topology_device *device = NULL;
58
59 list_for_each_entry(top_dev, &topology_device_list, list)
60 if (top_dev->proximity_domain == proximity_domain) {
61 device = top_dev;
62 break;
63 }
64
65 return device;
66}
67
68struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
69 uint32_t proximity_domain)
70{
71 struct kfd_topology_device *device = NULL;
72
73 down_read(&topology_lock);
74
75 device = kfd_topology_device_by_proximity_domain_no_lock(
76 proximity_domain);
77 up_read(&topology_lock);
78
79 return device;
80}
81
82struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id)
83{
84 struct kfd_topology_device *top_dev = NULL;
85 struct kfd_topology_device *ret = NULL;
86
87 down_read(&topology_lock);
88
89 list_for_each_entry(top_dev, &topology_device_list, list)
90 if (top_dev->gpu_id == gpu_id) {
91 ret = top_dev;
92 break;
93 }
94
95 up_read(&topology_lock);
96
97 return ret;
98}
99
100struct kfd_node *kfd_device_by_id(uint32_t gpu_id)
101{
102 struct kfd_topology_device *top_dev;
103
104 top_dev = kfd_topology_device_by_id(gpu_id);
105 if (!top_dev)
106 return NULL;
107
108 return top_dev->gpu;
109}
110
111struct kfd_node *kfd_device_by_pci_dev(const struct pci_dev *pdev)
112{
113 struct kfd_topology_device *top_dev;
114 struct kfd_node *device = NULL;
115
116 down_read(&topology_lock);
117
118 list_for_each_entry(top_dev, &topology_device_list, list)
119 if (top_dev->gpu && top_dev->gpu->adev->pdev == pdev) {
120 device = top_dev->gpu;
121 break;
122 }
123
124 up_read(&topology_lock);
125
126 return device;
127}
128
129/* Called with write topology_lock acquired */
130static void kfd_release_topology_device(struct kfd_topology_device *dev)
131{
132 struct kfd_mem_properties *mem;
133 struct kfd_cache_properties *cache;
134 struct kfd_iolink_properties *iolink;
135 struct kfd_iolink_properties *p2plink;
136 struct kfd_perf_properties *perf;
137
138 list_del(&dev->list);
139
140 while (dev->mem_props.next != &dev->mem_props) {
141 mem = container_of(dev->mem_props.next,
142 struct kfd_mem_properties, list);
143 list_del(&mem->list);
144 kfree(mem);
145 }
146
147 while (dev->cache_props.next != &dev->cache_props) {
148 cache = container_of(dev->cache_props.next,
149 struct kfd_cache_properties, list);
150 list_del(&cache->list);
151 kfree(cache);
152 }
153
154 while (dev->io_link_props.next != &dev->io_link_props) {
155 iolink = container_of(dev->io_link_props.next,
156 struct kfd_iolink_properties, list);
157 list_del(&iolink->list);
158 kfree(iolink);
159 }
160
161 while (dev->p2p_link_props.next != &dev->p2p_link_props) {
162 p2plink = container_of(dev->p2p_link_props.next,
163 struct kfd_iolink_properties, list);
164 list_del(&p2plink->list);
165 kfree(p2plink);
166 }
167
168 while (dev->perf_props.next != &dev->perf_props) {
169 perf = container_of(dev->perf_props.next,
170 struct kfd_perf_properties, list);
171 list_del(&perf->list);
172 kfree(perf);
173 }
174
175 kfree(dev);
176}
177
178void kfd_release_topology_device_list(struct list_head *device_list)
179{
180 struct kfd_topology_device *dev;
181
182 while (!list_empty(device_list)) {
183 dev = list_first_entry(device_list,
184 struct kfd_topology_device, list);
185 kfd_release_topology_device(dev);
186 }
187}
188
189static void kfd_release_live_view(void)
190{
191 kfd_release_topology_device_list(&topology_device_list);
192 memset(&sys_props, 0, sizeof(sys_props));
193}
194
195struct kfd_topology_device *kfd_create_topology_device(
196 struct list_head *device_list)
197{
198 struct kfd_topology_device *dev;
199
200 dev = kfd_alloc_struct(dev);
201 if (!dev) {
202 pr_err("No memory to allocate a topology device");
203 return NULL;
204 }
205
206 INIT_LIST_HEAD(&dev->mem_props);
207 INIT_LIST_HEAD(&dev->cache_props);
208 INIT_LIST_HEAD(&dev->io_link_props);
209 INIT_LIST_HEAD(&dev->p2p_link_props);
210 INIT_LIST_HEAD(&dev->perf_props);
211
212 list_add_tail(&dev->list, device_list);
213
214 return dev;
215}
216
217
218#define sysfs_show_gen_prop(buffer, offs, fmt, ...) \
219 (offs += snprintf(buffer+offs, PAGE_SIZE-offs, \
220 fmt, __VA_ARGS__))
221#define sysfs_show_32bit_prop(buffer, offs, name, value) \
222 sysfs_show_gen_prop(buffer, offs, "%s %u\n", name, value)
223#define sysfs_show_64bit_prop(buffer, offs, name, value) \
224 sysfs_show_gen_prop(buffer, offs, "%s %llu\n", name, value)
225#define sysfs_show_32bit_val(buffer, offs, value) \
226 sysfs_show_gen_prop(buffer, offs, "%u\n", value)
227#define sysfs_show_str_val(buffer, offs, value) \
228 sysfs_show_gen_prop(buffer, offs, "%s\n", value)
229
230static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr,
231 char *buffer)
232{
233 int offs = 0;
234
235 /* Making sure that the buffer is an empty string */
236 buffer[0] = 0;
237
238 if (attr == &sys_props.attr_genid) {
239 sysfs_show_32bit_val(buffer, offs,
240 sys_props.generation_count);
241 } else if (attr == &sys_props.attr_props) {
242 sysfs_show_64bit_prop(buffer, offs, "platform_oem",
243 sys_props.platform_oem);
244 sysfs_show_64bit_prop(buffer, offs, "platform_id",
245 sys_props.platform_id);
246 sysfs_show_64bit_prop(buffer, offs, "platform_rev",
247 sys_props.platform_rev);
248 } else {
249 offs = -EINVAL;
250 }
251
252 return offs;
253}
254
255static void kfd_topology_kobj_release(struct kobject *kobj)
256{
257 kfree(kobj);
258}
259
260static const struct sysfs_ops sysprops_ops = {
261 .show = sysprops_show,
262};
263
264static const struct kobj_type sysprops_type = {
265 .release = kfd_topology_kobj_release,
266 .sysfs_ops = &sysprops_ops,
267};
268
269static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr,
270 char *buffer)
271{
272 int offs = 0;
273 struct kfd_iolink_properties *iolink;
274
275 /* Making sure that the buffer is an empty string */
276 buffer[0] = 0;
277
278 iolink = container_of(attr, struct kfd_iolink_properties, attr);
279 if (iolink->gpu && kfd_devcgroup_check_permission(iolink->gpu))
280 return -EPERM;
281 sysfs_show_32bit_prop(buffer, offs, "type", iolink->iolink_type);
282 sysfs_show_32bit_prop(buffer, offs, "version_major", iolink->ver_maj);
283 sysfs_show_32bit_prop(buffer, offs, "version_minor", iolink->ver_min);
284 sysfs_show_32bit_prop(buffer, offs, "node_from", iolink->node_from);
285 sysfs_show_32bit_prop(buffer, offs, "node_to", iolink->node_to);
286 sysfs_show_32bit_prop(buffer, offs, "weight", iolink->weight);
287 sysfs_show_32bit_prop(buffer, offs, "min_latency", iolink->min_latency);
288 sysfs_show_32bit_prop(buffer, offs, "max_latency", iolink->max_latency);
289 sysfs_show_32bit_prop(buffer, offs, "min_bandwidth",
290 iolink->min_bandwidth);
291 sysfs_show_32bit_prop(buffer, offs, "max_bandwidth",
292 iolink->max_bandwidth);
293 sysfs_show_32bit_prop(buffer, offs, "recommended_transfer_size",
294 iolink->rec_transfer_size);
295 sysfs_show_32bit_prop(buffer, offs, "recommended_sdma_engine_id_mask",
296 iolink->rec_sdma_eng_id_mask);
297 sysfs_show_32bit_prop(buffer, offs, "flags", iolink->flags);
298
299 return offs;
300}
301
302static const struct sysfs_ops iolink_ops = {
303 .show = iolink_show,
304};
305
306static const struct kobj_type iolink_type = {
307 .release = kfd_topology_kobj_release,
308 .sysfs_ops = &iolink_ops,
309};
310
311static ssize_t mem_show(struct kobject *kobj, struct attribute *attr,
312 char *buffer)
313{
314 int offs = 0;
315 struct kfd_mem_properties *mem;
316
317 /* Making sure that the buffer is an empty string */
318 buffer[0] = 0;
319
320 mem = container_of(attr, struct kfd_mem_properties, attr);
321 if (mem->gpu && kfd_devcgroup_check_permission(mem->gpu))
322 return -EPERM;
323 sysfs_show_32bit_prop(buffer, offs, "heap_type", mem->heap_type);
324 sysfs_show_64bit_prop(buffer, offs, "size_in_bytes",
325 mem->size_in_bytes);
326 sysfs_show_32bit_prop(buffer, offs, "flags", mem->flags);
327 sysfs_show_32bit_prop(buffer, offs, "width", mem->width);
328 sysfs_show_32bit_prop(buffer, offs, "mem_clk_max",
329 mem->mem_clk_max);
330
331 return offs;
332}
333
334static const struct sysfs_ops mem_ops = {
335 .show = mem_show,
336};
337
338static const struct kobj_type mem_type = {
339 .release = kfd_topology_kobj_release,
340 .sysfs_ops = &mem_ops,
341};
342
343static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr,
344 char *buffer)
345{
346 int offs = 0;
347 uint32_t i, j;
348 struct kfd_cache_properties *cache;
349
350 /* Making sure that the buffer is an empty string */
351 buffer[0] = 0;
352 cache = container_of(attr, struct kfd_cache_properties, attr);
353 if (cache->gpu && kfd_devcgroup_check_permission(cache->gpu))
354 return -EPERM;
355 sysfs_show_32bit_prop(buffer, offs, "processor_id_low",
356 cache->processor_id_low);
357 sysfs_show_32bit_prop(buffer, offs, "level", cache->cache_level);
358 sysfs_show_32bit_prop(buffer, offs, "size", cache->cache_size);
359 sysfs_show_32bit_prop(buffer, offs, "cache_line_size",
360 cache->cacheline_size);
361 sysfs_show_32bit_prop(buffer, offs, "cache_lines_per_tag",
362 cache->cachelines_per_tag);
363 sysfs_show_32bit_prop(buffer, offs, "association", cache->cache_assoc);
364 sysfs_show_32bit_prop(buffer, offs, "latency", cache->cache_latency);
365 sysfs_show_32bit_prop(buffer, offs, "type", cache->cache_type);
366
367 offs += snprintf(buffer+offs, PAGE_SIZE-offs, "sibling_map ");
368 for (i = 0; i < cache->sibling_map_size; i++)
369 for (j = 0; j < sizeof(cache->sibling_map[0])*8; j++)
370 /* Check each bit */
371 offs += snprintf(buffer+offs, PAGE_SIZE-offs, "%d,",
372 (cache->sibling_map[i] >> j) & 1);
373
374 /* Replace the last "," with end of line */
375 buffer[offs-1] = '\n';
376 return offs;
377}
378
379static const struct sysfs_ops cache_ops = {
380 .show = kfd_cache_show,
381};
382
383static const struct kobj_type cache_type = {
384 .release = kfd_topology_kobj_release,
385 .sysfs_ops = &cache_ops,
386};
387
388/****** Sysfs of Performance Counters ******/
389
390struct kfd_perf_attr {
391 struct kobj_attribute attr;
392 uint32_t data;
393};
394
395static ssize_t perf_show(struct kobject *kobj, struct kobj_attribute *attrs,
396 char *buf)
397{
398 int offs = 0;
399 struct kfd_perf_attr *attr;
400
401 buf[0] = 0;
402 attr = container_of(attrs, struct kfd_perf_attr, attr);
403 if (!attr->data) /* invalid data for PMC */
404 return 0;
405 else
406 return sysfs_show_32bit_val(buf, offs, attr->data);
407}
408
409#define KFD_PERF_DESC(_name, _data) \
410{ \
411 .attr = __ATTR(_name, 0444, perf_show, NULL), \
412 .data = _data, \
413}
414
415static struct kfd_perf_attr perf_attr_iommu[] = {
416 KFD_PERF_DESC(max_concurrent, 0),
417 KFD_PERF_DESC(num_counters, 0),
418 KFD_PERF_DESC(counter_ids, 0),
419};
420/****************************************/
421
422static ssize_t node_show(struct kobject *kobj, struct attribute *attr,
423 char *buffer)
424{
425 int offs = 0;
426 struct kfd_topology_device *dev;
427 uint32_t log_max_watch_addr;
428
429 /* Making sure that the buffer is an empty string */
430 buffer[0] = 0;
431
432 if (strcmp(attr->name, "gpu_id") == 0) {
433 dev = container_of(attr, struct kfd_topology_device,
434 attr_gpuid);
435 if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu))
436 return -EPERM;
437 return sysfs_show_32bit_val(buffer, offs, dev->gpu_id);
438 }
439
440 if (strcmp(attr->name, "name") == 0) {
441 dev = container_of(attr, struct kfd_topology_device,
442 attr_name);
443
444 if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu))
445 return -EPERM;
446 return sysfs_show_str_val(buffer, offs, dev->node_props.name);
447 }
448
449 dev = container_of(attr, struct kfd_topology_device,
450 attr_props);
451 if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu))
452 return -EPERM;
453 sysfs_show_32bit_prop(buffer, offs, "cpu_cores_count",
454 dev->node_props.cpu_cores_count);
455 sysfs_show_32bit_prop(buffer, offs, "simd_count",
456 dev->gpu ? dev->node_props.simd_count : 0);
457 sysfs_show_32bit_prop(buffer, offs, "mem_banks_count",
458 dev->node_props.mem_banks_count);
459 sysfs_show_32bit_prop(buffer, offs, "caches_count",
460 dev->node_props.caches_count);
461 sysfs_show_32bit_prop(buffer, offs, "io_links_count",
462 dev->node_props.io_links_count);
463 sysfs_show_32bit_prop(buffer, offs, "p2p_links_count",
464 dev->node_props.p2p_links_count);
465 sysfs_show_32bit_prop(buffer, offs, "cpu_core_id_base",
466 dev->node_props.cpu_core_id_base);
467 sysfs_show_32bit_prop(buffer, offs, "simd_id_base",
468 dev->node_props.simd_id_base);
469 sysfs_show_32bit_prop(buffer, offs, "max_waves_per_simd",
470 dev->node_props.max_waves_per_simd);
471 sysfs_show_32bit_prop(buffer, offs, "lds_size_in_kb",
472 dev->node_props.lds_size_in_kb);
473 sysfs_show_32bit_prop(buffer, offs, "gds_size_in_kb",
474 dev->node_props.gds_size_in_kb);
475 sysfs_show_32bit_prop(buffer, offs, "num_gws",
476 dev->node_props.num_gws);
477 sysfs_show_32bit_prop(buffer, offs, "wave_front_size",
478 dev->node_props.wave_front_size);
479 sysfs_show_32bit_prop(buffer, offs, "array_count",
480 dev->gpu ? (dev->node_props.array_count *
481 NUM_XCC(dev->gpu->xcc_mask)) : 0);
482 sysfs_show_32bit_prop(buffer, offs, "simd_arrays_per_engine",
483 dev->node_props.simd_arrays_per_engine);
484 sysfs_show_32bit_prop(buffer, offs, "cu_per_simd_array",
485 dev->node_props.cu_per_simd_array);
486 sysfs_show_32bit_prop(buffer, offs, "simd_per_cu",
487 dev->node_props.simd_per_cu);
488 sysfs_show_32bit_prop(buffer, offs, "max_slots_scratch_cu",
489 dev->node_props.max_slots_scratch_cu);
490 sysfs_show_32bit_prop(buffer, offs, "gfx_target_version",
491 dev->node_props.gfx_target_version);
492 sysfs_show_32bit_prop(buffer, offs, "vendor_id",
493 dev->node_props.vendor_id);
494 sysfs_show_32bit_prop(buffer, offs, "device_id",
495 dev->node_props.device_id);
496 sysfs_show_32bit_prop(buffer, offs, "location_id",
497 dev->node_props.location_id);
498 sysfs_show_32bit_prop(buffer, offs, "domain",
499 dev->node_props.domain);
500 sysfs_show_32bit_prop(buffer, offs, "drm_render_minor",
501 dev->node_props.drm_render_minor);
502 sysfs_show_64bit_prop(buffer, offs, "hive_id",
503 dev->node_props.hive_id);
504 sysfs_show_32bit_prop(buffer, offs, "num_sdma_engines",
505 dev->node_props.num_sdma_engines);
506 sysfs_show_32bit_prop(buffer, offs, "num_sdma_xgmi_engines",
507 dev->node_props.num_sdma_xgmi_engines);
508 sysfs_show_32bit_prop(buffer, offs, "num_sdma_queues_per_engine",
509 dev->node_props.num_sdma_queues_per_engine);
510 sysfs_show_32bit_prop(buffer, offs, "num_cp_queues",
511 dev->node_props.num_cp_queues);
512
513 if (dev->gpu) {
514 log_max_watch_addr =
515 __ilog2_u32(dev->gpu->kfd->device_info.num_of_watch_points);
516
517 if (log_max_watch_addr) {
518 dev->node_props.capability |=
519 HSA_CAP_WATCH_POINTS_SUPPORTED;
520
521 dev->node_props.capability |=
522 ((log_max_watch_addr <<
523 HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) &
524 HSA_CAP_WATCH_POINTS_TOTALBITS_MASK);
525 }
526
527 if (dev->gpu->adev->asic_type == CHIP_TONGA)
528 dev->node_props.capability |=
529 HSA_CAP_AQL_QUEUE_DOUBLE_MAP;
530
531 sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_fcompute",
532 dev->node_props.max_engine_clk_fcompute);
533
534 sysfs_show_64bit_prop(buffer, offs, "local_mem_size", 0ULL);
535
536 sysfs_show_32bit_prop(buffer, offs, "fw_version",
537 dev->gpu->kfd->mec_fw_version);
538 sysfs_show_32bit_prop(buffer, offs, "capability",
539 dev->node_props.capability);
540 sysfs_show_64bit_prop(buffer, offs, "debug_prop",
541 dev->node_props.debug_prop);
542 sysfs_show_32bit_prop(buffer, offs, "sdma_fw_version",
543 dev->gpu->kfd->sdma_fw_version);
544 sysfs_show_64bit_prop(buffer, offs, "unique_id",
545 dev->gpu->adev->unique_id);
546 sysfs_show_32bit_prop(buffer, offs, "num_xcc",
547 NUM_XCC(dev->gpu->xcc_mask));
548 }
549
550 return sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_ccompute",
551 cpufreq_quick_get_max(0)/1000);
552}
553
554static const struct sysfs_ops node_ops = {
555 .show = node_show,
556};
557
558static const struct kobj_type node_type = {
559 .release = kfd_topology_kobj_release,
560 .sysfs_ops = &node_ops,
561};
562
563static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr)
564{
565 sysfs_remove_file(kobj, attr);
566 kobject_del(kobj);
567 kobject_put(kobj);
568}
569
570static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev)
571{
572 struct kfd_iolink_properties *p2plink;
573 struct kfd_iolink_properties *iolink;
574 struct kfd_cache_properties *cache;
575 struct kfd_mem_properties *mem;
576 struct kfd_perf_properties *perf;
577
578 if (dev->kobj_iolink) {
579 list_for_each_entry(iolink, &dev->io_link_props, list)
580 if (iolink->kobj) {
581 kfd_remove_sysfs_file(iolink->kobj,
582 &iolink->attr);
583 iolink->kobj = NULL;
584 }
585 kobject_del(dev->kobj_iolink);
586 kobject_put(dev->kobj_iolink);
587 dev->kobj_iolink = NULL;
588 }
589
590 if (dev->kobj_p2plink) {
591 list_for_each_entry(p2plink, &dev->p2p_link_props, list)
592 if (p2plink->kobj) {
593 kfd_remove_sysfs_file(p2plink->kobj,
594 &p2plink->attr);
595 p2plink->kobj = NULL;
596 }
597 kobject_del(dev->kobj_p2plink);
598 kobject_put(dev->kobj_p2plink);
599 dev->kobj_p2plink = NULL;
600 }
601
602 if (dev->kobj_cache) {
603 list_for_each_entry(cache, &dev->cache_props, list)
604 if (cache->kobj) {
605 kfd_remove_sysfs_file(cache->kobj,
606 &cache->attr);
607 cache->kobj = NULL;
608 }
609 kobject_del(dev->kobj_cache);
610 kobject_put(dev->kobj_cache);
611 dev->kobj_cache = NULL;
612 }
613
614 if (dev->kobj_mem) {
615 list_for_each_entry(mem, &dev->mem_props, list)
616 if (mem->kobj) {
617 kfd_remove_sysfs_file(mem->kobj, &mem->attr);
618 mem->kobj = NULL;
619 }
620 kobject_del(dev->kobj_mem);
621 kobject_put(dev->kobj_mem);
622 dev->kobj_mem = NULL;
623 }
624
625 if (dev->kobj_perf) {
626 list_for_each_entry(perf, &dev->perf_props, list) {
627 kfree(perf->attr_group);
628 perf->attr_group = NULL;
629 }
630 kobject_del(dev->kobj_perf);
631 kobject_put(dev->kobj_perf);
632 dev->kobj_perf = NULL;
633 }
634
635 if (dev->kobj_node) {
636 sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid);
637 sysfs_remove_file(dev->kobj_node, &dev->attr_name);
638 sysfs_remove_file(dev->kobj_node, &dev->attr_props);
639 kobject_del(dev->kobj_node);
640 kobject_put(dev->kobj_node);
641 dev->kobj_node = NULL;
642 }
643}
644
645static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev,
646 uint32_t id)
647{
648 struct kfd_iolink_properties *p2plink;
649 struct kfd_iolink_properties *iolink;
650 struct kfd_cache_properties *cache;
651 struct kfd_mem_properties *mem;
652 struct kfd_perf_properties *perf;
653 int ret;
654 uint32_t i, num_attrs;
655 struct attribute **attrs;
656
657 if (WARN_ON(dev->kobj_node))
658 return -EEXIST;
659
660 /*
661 * Creating the sysfs folders
662 */
663 dev->kobj_node = kfd_alloc_struct(dev->kobj_node);
664 if (!dev->kobj_node)
665 return -ENOMEM;
666
667 ret = kobject_init_and_add(dev->kobj_node, &node_type,
668 sys_props.kobj_nodes, "%d", id);
669 if (ret < 0) {
670 kobject_put(dev->kobj_node);
671 return ret;
672 }
673
674 dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node);
675 if (!dev->kobj_mem)
676 return -ENOMEM;
677
678 dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node);
679 if (!dev->kobj_cache)
680 return -ENOMEM;
681
682 dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node);
683 if (!dev->kobj_iolink)
684 return -ENOMEM;
685
686 dev->kobj_p2plink = kobject_create_and_add("p2p_links", dev->kobj_node);
687 if (!dev->kobj_p2plink)
688 return -ENOMEM;
689
690 dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node);
691 if (!dev->kobj_perf)
692 return -ENOMEM;
693
694 /*
695 * Creating sysfs files for node properties
696 */
697 dev->attr_gpuid.name = "gpu_id";
698 dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE;
699 sysfs_attr_init(&dev->attr_gpuid);
700 dev->attr_name.name = "name";
701 dev->attr_name.mode = KFD_SYSFS_FILE_MODE;
702 sysfs_attr_init(&dev->attr_name);
703 dev->attr_props.name = "properties";
704 dev->attr_props.mode = KFD_SYSFS_FILE_MODE;
705 sysfs_attr_init(&dev->attr_props);
706 ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid);
707 if (ret < 0)
708 return ret;
709 ret = sysfs_create_file(dev->kobj_node, &dev->attr_name);
710 if (ret < 0)
711 return ret;
712 ret = sysfs_create_file(dev->kobj_node, &dev->attr_props);
713 if (ret < 0)
714 return ret;
715
716 i = 0;
717 list_for_each_entry(mem, &dev->mem_props, list) {
718 mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
719 if (!mem->kobj)
720 return -ENOMEM;
721 ret = kobject_init_and_add(mem->kobj, &mem_type,
722 dev->kobj_mem, "%d", i);
723 if (ret < 0) {
724 kobject_put(mem->kobj);
725 return ret;
726 }
727
728 mem->attr.name = "properties";
729 mem->attr.mode = KFD_SYSFS_FILE_MODE;
730 sysfs_attr_init(&mem->attr);
731 ret = sysfs_create_file(mem->kobj, &mem->attr);
732 if (ret < 0)
733 return ret;
734 i++;
735 }
736
737 i = 0;
738 list_for_each_entry(cache, &dev->cache_props, list) {
739 cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
740 if (!cache->kobj)
741 return -ENOMEM;
742 ret = kobject_init_and_add(cache->kobj, &cache_type,
743 dev->kobj_cache, "%d", i);
744 if (ret < 0) {
745 kobject_put(cache->kobj);
746 return ret;
747 }
748
749 cache->attr.name = "properties";
750 cache->attr.mode = KFD_SYSFS_FILE_MODE;
751 sysfs_attr_init(&cache->attr);
752 ret = sysfs_create_file(cache->kobj, &cache->attr);
753 if (ret < 0)
754 return ret;
755 i++;
756 }
757
758 i = 0;
759 list_for_each_entry(iolink, &dev->io_link_props, list) {
760 iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
761 if (!iolink->kobj)
762 return -ENOMEM;
763 ret = kobject_init_and_add(iolink->kobj, &iolink_type,
764 dev->kobj_iolink, "%d", i);
765 if (ret < 0) {
766 kobject_put(iolink->kobj);
767 return ret;
768 }
769
770 iolink->attr.name = "properties";
771 iolink->attr.mode = KFD_SYSFS_FILE_MODE;
772 sysfs_attr_init(&iolink->attr);
773 ret = sysfs_create_file(iolink->kobj, &iolink->attr);
774 if (ret < 0)
775 return ret;
776 i++;
777 }
778
779 i = 0;
780 list_for_each_entry(p2plink, &dev->p2p_link_props, list) {
781 p2plink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
782 if (!p2plink->kobj)
783 return -ENOMEM;
784 ret = kobject_init_and_add(p2plink->kobj, &iolink_type,
785 dev->kobj_p2plink, "%d", i);
786 if (ret < 0) {
787 kobject_put(p2plink->kobj);
788 return ret;
789 }
790
791 p2plink->attr.name = "properties";
792 p2plink->attr.mode = KFD_SYSFS_FILE_MODE;
793 sysfs_attr_init(&p2plink->attr);
794 ret = sysfs_create_file(p2plink->kobj, &p2plink->attr);
795 if (ret < 0)
796 return ret;
797 i++;
798 }
799
800 /* All hardware blocks have the same number of attributes. */
801 num_attrs = ARRAY_SIZE(perf_attr_iommu);
802 list_for_each_entry(perf, &dev->perf_props, list) {
803 perf->attr_group = kzalloc(sizeof(struct kfd_perf_attr)
804 * num_attrs + sizeof(struct attribute_group),
805 GFP_KERNEL);
806 if (!perf->attr_group)
807 return -ENOMEM;
808
809 attrs = (struct attribute **)(perf->attr_group + 1);
810 if (!strcmp(perf->block_name, "iommu")) {
811 /* Information of IOMMU's num_counters and counter_ids is shown
812 * under /sys/bus/event_source/devices/amd_iommu. We don't
813 * duplicate here.
814 */
815 perf_attr_iommu[0].data = perf->max_concurrent;
816 for (i = 0; i < num_attrs; i++)
817 attrs[i] = &perf_attr_iommu[i].attr.attr;
818 }
819 perf->attr_group->name = perf->block_name;
820 perf->attr_group->attrs = attrs;
821 ret = sysfs_create_group(dev->kobj_perf, perf->attr_group);
822 if (ret < 0)
823 return ret;
824 }
825
826 return 0;
827}
828
829/* Called with write topology lock acquired */
830static int kfd_build_sysfs_node_tree(void)
831{
832 struct kfd_topology_device *dev;
833 int ret;
834 uint32_t i = 0;
835
836 list_for_each_entry(dev, &topology_device_list, list) {
837 ret = kfd_build_sysfs_node_entry(dev, i);
838 if (ret < 0)
839 return ret;
840 i++;
841 }
842
843 return 0;
844}
845
846/* Called with write topology lock acquired */
847static void kfd_remove_sysfs_node_tree(void)
848{
849 struct kfd_topology_device *dev;
850
851 list_for_each_entry(dev, &topology_device_list, list)
852 kfd_remove_sysfs_node_entry(dev);
853}
854
855static int kfd_topology_update_sysfs(void)
856{
857 int ret;
858
859 if (!sys_props.kobj_topology) {
860 sys_props.kobj_topology =
861 kfd_alloc_struct(sys_props.kobj_topology);
862 if (!sys_props.kobj_topology)
863 return -ENOMEM;
864
865 ret = kobject_init_and_add(sys_props.kobj_topology,
866 &sysprops_type, &kfd_device->kobj,
867 "topology");
868 if (ret < 0) {
869 kobject_put(sys_props.kobj_topology);
870 return ret;
871 }
872
873 sys_props.kobj_nodes = kobject_create_and_add("nodes",
874 sys_props.kobj_topology);
875 if (!sys_props.kobj_nodes)
876 return -ENOMEM;
877
878 sys_props.attr_genid.name = "generation_id";
879 sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE;
880 sysfs_attr_init(&sys_props.attr_genid);
881 ret = sysfs_create_file(sys_props.kobj_topology,
882 &sys_props.attr_genid);
883 if (ret < 0)
884 return ret;
885
886 sys_props.attr_props.name = "system_properties";
887 sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE;
888 sysfs_attr_init(&sys_props.attr_props);
889 ret = sysfs_create_file(sys_props.kobj_topology,
890 &sys_props.attr_props);
891 if (ret < 0)
892 return ret;
893 }
894
895 kfd_remove_sysfs_node_tree();
896
897 return kfd_build_sysfs_node_tree();
898}
899
900static void kfd_topology_release_sysfs(void)
901{
902 kfd_remove_sysfs_node_tree();
903 if (sys_props.kobj_topology) {
904 sysfs_remove_file(sys_props.kobj_topology,
905 &sys_props.attr_genid);
906 sysfs_remove_file(sys_props.kobj_topology,
907 &sys_props.attr_props);
908 if (sys_props.kobj_nodes) {
909 kobject_del(sys_props.kobj_nodes);
910 kobject_put(sys_props.kobj_nodes);
911 sys_props.kobj_nodes = NULL;
912 }
913 kobject_del(sys_props.kobj_topology);
914 kobject_put(sys_props.kobj_topology);
915 sys_props.kobj_topology = NULL;
916 }
917}
918
919/* Called with write topology_lock acquired */
920static void kfd_topology_update_device_list(struct list_head *temp_list,
921 struct list_head *master_list)
922{
923 while (!list_empty(temp_list)) {
924 list_move_tail(temp_list->next, master_list);
925 sys_props.num_devices++;
926 }
927}
928
929static void kfd_debug_print_topology(void)
930{
931 struct kfd_topology_device *dev;
932
933 down_read(&topology_lock);
934
935 dev = list_last_entry(&topology_device_list,
936 struct kfd_topology_device, list);
937 if (dev) {
938 if (dev->node_props.cpu_cores_count &&
939 dev->node_props.simd_count) {
940 pr_info("Topology: Add APU node [0x%0x:0x%0x]\n",
941 dev->node_props.device_id,
942 dev->node_props.vendor_id);
943 } else if (dev->node_props.cpu_cores_count)
944 pr_info("Topology: Add CPU node\n");
945 else if (dev->node_props.simd_count)
946 pr_info("Topology: Add dGPU node [0x%0x:0x%0x]\n",
947 dev->node_props.device_id,
948 dev->node_props.vendor_id);
949 }
950 up_read(&topology_lock);
951}
952
953/* Helper function for intializing platform_xx members of
954 * kfd_system_properties. Uses OEM info from the last CPU/APU node.
955 */
956static void kfd_update_system_properties(void)
957{
958 struct kfd_topology_device *dev;
959
960 down_read(&topology_lock);
961 dev = list_last_entry(&topology_device_list,
962 struct kfd_topology_device, list);
963 if (dev) {
964 sys_props.platform_id = dev->oem_id64;
965 sys_props.platform_oem = *((uint64_t *)dev->oem_table_id);
966 sys_props.platform_rev = dev->oem_revision;
967 }
968 up_read(&topology_lock);
969}
970
971static void find_system_memory(const struct dmi_header *dm,
972 void *private)
973{
974 struct kfd_mem_properties *mem;
975 u16 mem_width, mem_clock;
976 struct kfd_topology_device *kdev =
977 (struct kfd_topology_device *)private;
978 const u8 *dmi_data = (const u8 *)(dm + 1);
979
980 if (dm->type == DMI_ENTRY_MEM_DEVICE && dm->length >= 0x15) {
981 mem_width = (u16)(*(const u16 *)(dmi_data + 0x6));
982 mem_clock = (u16)(*(const u16 *)(dmi_data + 0x11));
983 list_for_each_entry(mem, &kdev->mem_props, list) {
984 if (mem_width != 0xFFFF && mem_width != 0)
985 mem->width = mem_width;
986 if (mem_clock != 0)
987 mem->mem_clk_max = mem_clock;
988 }
989 }
990}
991
992/* kfd_add_non_crat_information - Add information that is not currently
993 * defined in CRAT but is necessary for KFD topology
994 * @dev - topology device to which addition info is added
995 */
996static void kfd_add_non_crat_information(struct kfd_topology_device *kdev)
997{
998 /* Check if CPU only node. */
999 if (!kdev->gpu) {
1000 /* Add system memory information */
1001 dmi_walk(find_system_memory, kdev);
1002 }
1003 /* TODO: For GPU node, rearrange code from kfd_topology_add_device */
1004}
1005
1006int kfd_topology_init(void)
1007{
1008 void *crat_image = NULL;
1009 size_t image_size = 0;
1010 int ret;
1011 struct list_head temp_topology_device_list;
1012 int cpu_only_node = 0;
1013 struct kfd_topology_device *kdev;
1014 int proximity_domain;
1015
1016 /* topology_device_list - Master list of all topology devices
1017 * temp_topology_device_list - temporary list created while parsing CRAT
1018 * or VCRAT. Once parsing is complete the contents of list is moved to
1019 * topology_device_list
1020 */
1021
1022 /* Initialize the head for the both the lists */
1023 INIT_LIST_HEAD(&topology_device_list);
1024 INIT_LIST_HEAD(&temp_topology_device_list);
1025 init_rwsem(&topology_lock);
1026
1027 memset(&sys_props, 0, sizeof(sys_props));
1028
1029 /* Proximity domains in ACPI CRAT tables start counting at
1030 * 0. The same should be true for virtual CRAT tables created
1031 * at this stage. GPUs added later in kfd_topology_add_device
1032 * use a counter.
1033 */
1034 proximity_domain = 0;
1035
1036 ret = kfd_create_crat_image_virtual(&crat_image, &image_size,
1037 COMPUTE_UNIT_CPU, NULL,
1038 proximity_domain);
1039 cpu_only_node = 1;
1040 if (ret) {
1041 pr_err("Error creating VCRAT table for CPU\n");
1042 return ret;
1043 }
1044
1045 ret = kfd_parse_crat_table(crat_image,
1046 &temp_topology_device_list,
1047 proximity_domain);
1048 if (ret) {
1049 pr_err("Error parsing VCRAT table for CPU\n");
1050 goto err;
1051 }
1052
1053 kdev = list_first_entry(&temp_topology_device_list,
1054 struct kfd_topology_device, list);
1055
1056 down_write(&topology_lock);
1057 kfd_topology_update_device_list(&temp_topology_device_list,
1058 &topology_device_list);
1059 topology_crat_proximity_domain = sys_props.num_devices-1;
1060 ret = kfd_topology_update_sysfs();
1061 up_write(&topology_lock);
1062
1063 if (!ret) {
1064 sys_props.generation_count++;
1065 kfd_update_system_properties();
1066 kfd_debug_print_topology();
1067 } else
1068 pr_err("Failed to update topology in sysfs ret=%d\n", ret);
1069
1070 /* For nodes with GPU, this information gets added
1071 * when GPU is detected (kfd_topology_add_device).
1072 */
1073 if (cpu_only_node) {
1074 /* Add additional information to CPU only node created above */
1075 down_write(&topology_lock);
1076 kdev = list_first_entry(&topology_device_list,
1077 struct kfd_topology_device, list);
1078 up_write(&topology_lock);
1079 kfd_add_non_crat_information(kdev);
1080 }
1081
1082err:
1083 kfd_destroy_crat_image(crat_image);
1084 return ret;
1085}
1086
1087void kfd_topology_shutdown(void)
1088{
1089 down_write(&topology_lock);
1090 kfd_topology_release_sysfs();
1091 kfd_release_live_view();
1092 up_write(&topology_lock);
1093}
1094
1095static uint32_t kfd_generate_gpu_id(struct kfd_node *gpu)
1096{
1097 uint32_t gpu_id;
1098 uint32_t buf[8];
1099 uint64_t local_mem_size;
1100 struct kfd_topology_device *dev;
1101 bool is_unique;
1102 uint8_t *crc_buf;
1103
1104 if (!gpu)
1105 return 0;
1106
1107 crc_buf = (uint8_t *)&buf;
1108 local_mem_size = gpu->local_mem_info.local_mem_size_private +
1109 gpu->local_mem_info.local_mem_size_public;
1110 buf[0] = gpu->adev->pdev->devfn;
1111 buf[1] = gpu->adev->pdev->subsystem_vendor |
1112 (gpu->adev->pdev->subsystem_device << 16);
1113 buf[2] = pci_domain_nr(gpu->adev->pdev->bus);
1114 buf[3] = gpu->adev->pdev->device;
1115 buf[4] = gpu->adev->pdev->bus->number;
1116 buf[5] = lower_32_bits(local_mem_size);
1117 buf[6] = upper_32_bits(local_mem_size);
1118 buf[7] = (ffs(gpu->xcc_mask) - 1) | (NUM_XCC(gpu->xcc_mask) << 16);
1119
1120 gpu_id = crc16(0, crc_buf, sizeof(buf)) &
1121 ((1 << KFD_GPU_ID_HASH_WIDTH) - 1);
1122
1123 /* There is a very small possibility when generating a
1124 * 16 (KFD_GPU_ID_HASH_WIDTH) bit value from 8 word buffer
1125 * that the value could be 0 or non-unique. So, check if
1126 * it is unique and non-zero. If not unique increment till
1127 * unique one is found. In case of overflow, restart from 1
1128 */
1129
1130 down_read(&topology_lock);
1131 do {
1132 is_unique = true;
1133 if (!gpu_id)
1134 gpu_id = 1;
1135 list_for_each_entry(dev, &topology_device_list, list) {
1136 if (dev->gpu && dev->gpu_id == gpu_id) {
1137 is_unique = false;
1138 break;
1139 }
1140 }
1141 if (unlikely(!is_unique))
1142 gpu_id = (gpu_id + 1) &
1143 ((1 << KFD_GPU_ID_HASH_WIDTH) - 1);
1144 } while (!is_unique);
1145 up_read(&topology_lock);
1146
1147 return gpu_id;
1148}
1149/* kfd_assign_gpu - Attach @gpu to the correct kfd topology device. If
1150 * the GPU device is not already present in the topology device
1151 * list then return NULL. This means a new topology device has to
1152 * be created for this GPU.
1153 */
1154static struct kfd_topology_device *kfd_assign_gpu(struct kfd_node *gpu)
1155{
1156 struct kfd_topology_device *dev;
1157 struct kfd_topology_device *out_dev = NULL;
1158 struct kfd_mem_properties *mem;
1159 struct kfd_cache_properties *cache;
1160 struct kfd_iolink_properties *iolink;
1161 struct kfd_iolink_properties *p2plink;
1162
1163 list_for_each_entry(dev, &topology_device_list, list) {
1164 /* Discrete GPUs need their own topology device list
1165 * entries. Don't assign them to CPU/APU nodes.
1166 */
1167 if (dev->node_props.cpu_cores_count)
1168 continue;
1169
1170 if (!dev->gpu && (dev->node_props.simd_count > 0)) {
1171 dev->gpu = gpu;
1172 out_dev = dev;
1173
1174 list_for_each_entry(mem, &dev->mem_props, list)
1175 mem->gpu = dev->gpu;
1176 list_for_each_entry(cache, &dev->cache_props, list)
1177 cache->gpu = dev->gpu;
1178 list_for_each_entry(iolink, &dev->io_link_props, list)
1179 iolink->gpu = dev->gpu;
1180 list_for_each_entry(p2plink, &dev->p2p_link_props, list)
1181 p2plink->gpu = dev->gpu;
1182 break;
1183 }
1184 }
1185 return out_dev;
1186}
1187
1188static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival)
1189{
1190 /*
1191 * TODO: Generate an event for thunk about the arrival/removal
1192 * of the GPU
1193 */
1194}
1195
1196/* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info,
1197 * patch this after CRAT parsing.
1198 */
1199static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev)
1200{
1201 struct kfd_mem_properties *mem;
1202 struct kfd_local_mem_info local_mem_info;
1203
1204 if (!dev)
1205 return;
1206
1207 /* Currently, amdgpu driver (amdgpu_mc) deals only with GPUs with
1208 * single bank of VRAM local memory.
1209 * for dGPUs - VCRAT reports only one bank of Local Memory
1210 * for APUs - If CRAT from ACPI reports more than one bank, then
1211 * all the banks will report the same mem_clk_max information
1212 */
1213 amdgpu_amdkfd_get_local_mem_info(dev->gpu->adev, &local_mem_info,
1214 dev->gpu->xcp);
1215
1216 list_for_each_entry(mem, &dev->mem_props, list)
1217 mem->mem_clk_max = local_mem_info.mem_clk_max;
1218}
1219
1220static void kfd_set_iolink_no_atomics(struct kfd_topology_device *dev,
1221 struct kfd_topology_device *target_gpu_dev,
1222 struct kfd_iolink_properties *link)
1223{
1224 /* xgmi always supports atomics between links. */
1225 if (link->iolink_type == CRAT_IOLINK_TYPE_XGMI)
1226 return;
1227
1228 /* check pcie support to set cpu(dev) flags for target_gpu_dev link. */
1229 if (target_gpu_dev) {
1230 uint32_t cap;
1231
1232 pcie_capability_read_dword(target_gpu_dev->gpu->adev->pdev,
1233 PCI_EXP_DEVCAP2, &cap);
1234
1235 if (!(cap & (PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
1236 PCI_EXP_DEVCAP2_ATOMIC_COMP64)))
1237 link->flags |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1238 CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1239 /* set gpu (dev) flags. */
1240 } else {
1241 if (!dev->gpu->kfd->pci_atomic_requested ||
1242 dev->gpu->adev->asic_type == CHIP_HAWAII)
1243 link->flags |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1244 CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1245 }
1246}
1247
1248static void kfd_set_iolink_non_coherent(struct kfd_topology_device *to_dev,
1249 struct kfd_iolink_properties *outbound_link,
1250 struct kfd_iolink_properties *inbound_link)
1251{
1252 /* CPU -> GPU with PCIe */
1253 if (!to_dev->gpu &&
1254 inbound_link->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS)
1255 inbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT;
1256
1257 if (to_dev->gpu) {
1258 /* GPU <-> GPU with PCIe and
1259 * Vega20 with XGMI
1260 */
1261 if (inbound_link->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS ||
1262 (inbound_link->iolink_type == CRAT_IOLINK_TYPE_XGMI &&
1263 KFD_GC_VERSION(to_dev->gpu) == IP_VERSION(9, 4, 0))) {
1264 outbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT;
1265 inbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT;
1266 }
1267 }
1268}
1269
1270#define REC_SDMA_NUM_GPU 8
1271static const int rec_sdma_eng_map[REC_SDMA_NUM_GPU][REC_SDMA_NUM_GPU] = {
1272 { -1, 14, 12, 2, 4, 8, 10, 6 },
1273 { 14, -1, 2, 10, 8, 4, 6, 12 },
1274 { 10, 2, -1, 12, 14, 6, 4, 8 },
1275 { 2, 12, 10, -1, 6, 14, 8, 4 },
1276 { 4, 8, 14, 6, -1, 10, 12, 2 },
1277 { 8, 4, 6, 14, 12, -1, 2, 10 },
1278 { 10, 6, 4, 8, 12, 2, -1, 14 },
1279 { 6, 12, 8, 4, 2, 10, 14, -1 }};
1280
1281static void kfd_set_recommended_sdma_engines(struct kfd_topology_device *to_dev,
1282 struct kfd_iolink_properties *outbound_link,
1283 struct kfd_iolink_properties *inbound_link)
1284{
1285 struct kfd_node *gpu = outbound_link->gpu;
1286 struct amdgpu_device *adev = gpu->adev;
1287 int num_xgmi_nodes = adev->gmc.xgmi.num_physical_nodes;
1288 bool support_rec_eng = !amdgpu_sriov_vf(adev) && to_dev->gpu &&
1289 adev->aid_mask && num_xgmi_nodes && gpu->kfd->num_nodes == 1 &&
1290 kfd_get_num_xgmi_sdma_engines(gpu) >= 14 &&
1291 (!(adev->flags & AMD_IS_APU) && num_xgmi_nodes == 8);
1292
1293 if (support_rec_eng) {
1294 int src_socket_id = adev->gmc.xgmi.physical_node_id;
1295 int dst_socket_id = to_dev->gpu->adev->gmc.xgmi.physical_node_id;
1296
1297 outbound_link->rec_sdma_eng_id_mask =
1298 1 << rec_sdma_eng_map[src_socket_id][dst_socket_id];
1299 inbound_link->rec_sdma_eng_id_mask =
1300 1 << rec_sdma_eng_map[dst_socket_id][src_socket_id];
1301 } else {
1302 int num_sdma_eng = kfd_get_num_sdma_engines(gpu);
1303 int i, eng_offset = 0;
1304
1305 if (outbound_link->iolink_type == CRAT_IOLINK_TYPE_XGMI &&
1306 kfd_get_num_xgmi_sdma_engines(gpu) && to_dev->gpu) {
1307 eng_offset = num_sdma_eng;
1308 num_sdma_eng = kfd_get_num_xgmi_sdma_engines(gpu);
1309 }
1310
1311 for (i = 0; i < num_sdma_eng; i++) {
1312 outbound_link->rec_sdma_eng_id_mask |= (1 << (i + eng_offset));
1313 inbound_link->rec_sdma_eng_id_mask |= (1 << (i + eng_offset));
1314 }
1315 }
1316}
1317
1318static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev)
1319{
1320 struct kfd_iolink_properties *link, *inbound_link;
1321 struct kfd_topology_device *peer_dev;
1322
1323 if (!dev || !dev->gpu)
1324 return;
1325
1326 /* GPU only creates direct links so apply flags setting to all */
1327 list_for_each_entry(link, &dev->io_link_props, list) {
1328 link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1329 kfd_set_iolink_no_atomics(dev, NULL, link);
1330 peer_dev = kfd_topology_device_by_proximity_domain(
1331 link->node_to);
1332
1333 if (!peer_dev)
1334 continue;
1335
1336 /* Include the CPU peer in GPU hive if connected over xGMI. */
1337 if (!peer_dev->gpu &&
1338 link->iolink_type == CRAT_IOLINK_TYPE_XGMI) {
1339 /*
1340 * If the GPU is not part of a GPU hive, use its pci
1341 * device location as the hive ID to bind with the CPU.
1342 */
1343 if (!dev->node_props.hive_id)
1344 dev->node_props.hive_id = pci_dev_id(dev->gpu->adev->pdev);
1345 peer_dev->node_props.hive_id = dev->node_props.hive_id;
1346 }
1347
1348 list_for_each_entry(inbound_link, &peer_dev->io_link_props,
1349 list) {
1350 if (inbound_link->node_to != link->node_from)
1351 continue;
1352
1353 inbound_link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1354 kfd_set_iolink_no_atomics(peer_dev, dev, inbound_link);
1355 kfd_set_iolink_non_coherent(peer_dev, link, inbound_link);
1356 kfd_set_recommended_sdma_engines(peer_dev, link, inbound_link);
1357 }
1358 }
1359
1360 /* Create indirect links so apply flags setting to all */
1361 list_for_each_entry(link, &dev->p2p_link_props, list) {
1362 link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1363 kfd_set_iolink_no_atomics(dev, NULL, link);
1364 peer_dev = kfd_topology_device_by_proximity_domain(
1365 link->node_to);
1366
1367 if (!peer_dev)
1368 continue;
1369
1370 list_for_each_entry(inbound_link, &peer_dev->p2p_link_props,
1371 list) {
1372 if (inbound_link->node_to != link->node_from)
1373 continue;
1374
1375 inbound_link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1376 kfd_set_iolink_no_atomics(peer_dev, dev, inbound_link);
1377 kfd_set_iolink_non_coherent(peer_dev, link, inbound_link);
1378 }
1379 }
1380}
1381
1382static int kfd_build_p2p_node_entry(struct kfd_topology_device *dev,
1383 struct kfd_iolink_properties *p2plink)
1384{
1385 int ret;
1386
1387 p2plink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
1388 if (!p2plink->kobj)
1389 return -ENOMEM;
1390
1391 ret = kobject_init_and_add(p2plink->kobj, &iolink_type,
1392 dev->kobj_p2plink, "%d", dev->node_props.p2p_links_count - 1);
1393 if (ret < 0) {
1394 kobject_put(p2plink->kobj);
1395 return ret;
1396 }
1397
1398 p2plink->attr.name = "properties";
1399 p2plink->attr.mode = KFD_SYSFS_FILE_MODE;
1400 sysfs_attr_init(&p2plink->attr);
1401 ret = sysfs_create_file(p2plink->kobj, &p2plink->attr);
1402 if (ret < 0)
1403 return ret;
1404
1405 return 0;
1406}
1407
1408static int kfd_create_indirect_link_prop(struct kfd_topology_device *kdev, int gpu_node)
1409{
1410 struct kfd_iolink_properties *gpu_link, *tmp_link, *cpu_link;
1411 struct kfd_iolink_properties *props = NULL, *props2 = NULL;
1412 struct kfd_topology_device *cpu_dev;
1413 int ret = 0;
1414 int i, num_cpu;
1415
1416 num_cpu = 0;
1417 list_for_each_entry(cpu_dev, &topology_device_list, list) {
1418 if (cpu_dev->gpu)
1419 break;
1420 num_cpu++;
1421 }
1422
1423 if (list_empty(&kdev->io_link_props))
1424 return -ENODATA;
1425
1426 gpu_link = list_first_entry(&kdev->io_link_props,
1427 struct kfd_iolink_properties, list);
1428
1429 for (i = 0; i < num_cpu; i++) {
1430 /* CPU <--> GPU */
1431 if (gpu_link->node_to == i)
1432 continue;
1433
1434 /* find CPU <--> CPU links */
1435 cpu_link = NULL;
1436 cpu_dev = kfd_topology_device_by_proximity_domain(i);
1437 if (cpu_dev) {
1438 list_for_each_entry(tmp_link,
1439 &cpu_dev->io_link_props, list) {
1440 if (tmp_link->node_to == gpu_link->node_to) {
1441 cpu_link = tmp_link;
1442 break;
1443 }
1444 }
1445 }
1446
1447 if (!cpu_link)
1448 return -ENOMEM;
1449
1450 /* CPU <--> CPU <--> GPU, GPU node*/
1451 props = kfd_alloc_struct(props);
1452 if (!props)
1453 return -ENOMEM;
1454
1455 memcpy(props, gpu_link, sizeof(struct kfd_iolink_properties));
1456 props->weight = gpu_link->weight + cpu_link->weight;
1457 props->min_latency = gpu_link->min_latency + cpu_link->min_latency;
1458 props->max_latency = gpu_link->max_latency + cpu_link->max_latency;
1459 props->min_bandwidth = min(gpu_link->min_bandwidth, cpu_link->min_bandwidth);
1460 props->max_bandwidth = min(gpu_link->max_bandwidth, cpu_link->max_bandwidth);
1461
1462 props->node_from = gpu_node;
1463 props->node_to = i;
1464 kdev->node_props.p2p_links_count++;
1465 list_add_tail(&props->list, &kdev->p2p_link_props);
1466 ret = kfd_build_p2p_node_entry(kdev, props);
1467 if (ret < 0)
1468 return ret;
1469
1470 /* for small Bar, no CPU --> GPU in-direct links */
1471 if (kfd_dev_is_large_bar(kdev->gpu)) {
1472 /* CPU <--> CPU <--> GPU, CPU node*/
1473 props2 = kfd_alloc_struct(props2);
1474 if (!props2)
1475 return -ENOMEM;
1476
1477 memcpy(props2, props, sizeof(struct kfd_iolink_properties));
1478 props2->node_from = i;
1479 props2->node_to = gpu_node;
1480 props2->kobj = NULL;
1481 cpu_dev->node_props.p2p_links_count++;
1482 list_add_tail(&props2->list, &cpu_dev->p2p_link_props);
1483 ret = kfd_build_p2p_node_entry(cpu_dev, props2);
1484 if (ret < 0)
1485 return ret;
1486 }
1487 }
1488 return ret;
1489}
1490
1491#if defined(CONFIG_HSA_AMD_P2P)
1492static int kfd_add_peer_prop(struct kfd_topology_device *kdev,
1493 struct kfd_topology_device *peer, int from, int to)
1494{
1495 struct kfd_iolink_properties *props = NULL;
1496 struct kfd_iolink_properties *iolink1, *iolink2, *iolink3;
1497 struct kfd_topology_device *cpu_dev;
1498 int ret = 0;
1499
1500 if (!amdgpu_device_is_peer_accessible(
1501 kdev->gpu->adev,
1502 peer->gpu->adev))
1503 return ret;
1504
1505 if (list_empty(&kdev->io_link_props))
1506 return -ENODATA;
1507
1508 iolink1 = list_first_entry(&kdev->io_link_props,
1509 struct kfd_iolink_properties, list);
1510
1511 if (list_empty(&peer->io_link_props))
1512 return -ENODATA;
1513
1514 iolink2 = list_first_entry(&peer->io_link_props,
1515 struct kfd_iolink_properties, list);
1516
1517 props = kfd_alloc_struct(props);
1518 if (!props)
1519 return -ENOMEM;
1520
1521 memcpy(props, iolink1, sizeof(struct kfd_iolink_properties));
1522
1523 props->weight = iolink1->weight + iolink2->weight;
1524 props->min_latency = iolink1->min_latency + iolink2->min_latency;
1525 props->max_latency = iolink1->max_latency + iolink2->max_latency;
1526 props->min_bandwidth = min(iolink1->min_bandwidth, iolink2->min_bandwidth);
1527 props->max_bandwidth = min(iolink2->max_bandwidth, iolink2->max_bandwidth);
1528
1529 if (iolink1->node_to != iolink2->node_to) {
1530 /* CPU->CPU link*/
1531 cpu_dev = kfd_topology_device_by_proximity_domain(iolink1->node_to);
1532 if (cpu_dev) {
1533 list_for_each_entry(iolink3, &cpu_dev->io_link_props, list) {
1534 if (iolink3->node_to != iolink2->node_to)
1535 continue;
1536
1537 props->weight += iolink3->weight;
1538 props->min_latency += iolink3->min_latency;
1539 props->max_latency += iolink3->max_latency;
1540 props->min_bandwidth = min(props->min_bandwidth,
1541 iolink3->min_bandwidth);
1542 props->max_bandwidth = min(props->max_bandwidth,
1543 iolink3->max_bandwidth);
1544 break;
1545 }
1546 } else {
1547 WARN(1, "CPU node not found");
1548 }
1549 }
1550
1551 props->node_from = from;
1552 props->node_to = to;
1553 peer->node_props.p2p_links_count++;
1554 list_add_tail(&props->list, &peer->p2p_link_props);
1555 ret = kfd_build_p2p_node_entry(peer, props);
1556
1557 return ret;
1558}
1559#endif
1560
1561static int kfd_dev_create_p2p_links(void)
1562{
1563 struct kfd_topology_device *dev;
1564 struct kfd_topology_device *new_dev;
1565#if defined(CONFIG_HSA_AMD_P2P)
1566 uint32_t i;
1567#endif
1568 uint32_t k;
1569 int ret = 0;
1570
1571 k = 0;
1572 list_for_each_entry(dev, &topology_device_list, list)
1573 k++;
1574 if (k < 2)
1575 return 0;
1576
1577 new_dev = list_last_entry(&topology_device_list, struct kfd_topology_device, list);
1578 if (WARN_ON(!new_dev->gpu))
1579 return 0;
1580
1581 k--;
1582
1583 /* create in-direct links */
1584 ret = kfd_create_indirect_link_prop(new_dev, k);
1585 if (ret < 0)
1586 goto out;
1587
1588 /* create p2p links */
1589#if defined(CONFIG_HSA_AMD_P2P)
1590 i = 0;
1591 list_for_each_entry(dev, &topology_device_list, list) {
1592 if (dev == new_dev)
1593 break;
1594 if (!dev->gpu || !dev->gpu->adev ||
1595 (dev->gpu->kfd->hive_id &&
1596 dev->gpu->kfd->hive_id == new_dev->gpu->kfd->hive_id))
1597 goto next;
1598
1599 /* check if node(s) is/are peer accessible in one direction or bi-direction */
1600 ret = kfd_add_peer_prop(new_dev, dev, i, k);
1601 if (ret < 0)
1602 goto out;
1603
1604 ret = kfd_add_peer_prop(dev, new_dev, k, i);
1605 if (ret < 0)
1606 goto out;
1607next:
1608 i++;
1609 }
1610#endif
1611
1612out:
1613 return ret;
1614}
1615
1616/* Helper function. See kfd_fill_gpu_cache_info for parameter description */
1617static int fill_in_l1_pcache(struct kfd_cache_properties **props_ext,
1618 struct kfd_gpu_cache_info *pcache_info,
1619 int cu_bitmask,
1620 int cache_type, unsigned int cu_processor_id,
1621 int cu_block)
1622{
1623 unsigned int cu_sibling_map_mask;
1624 int first_active_cu;
1625 struct kfd_cache_properties *pcache = NULL;
1626
1627 cu_sibling_map_mask = cu_bitmask;
1628 cu_sibling_map_mask >>= cu_block;
1629 cu_sibling_map_mask &= ((1 << pcache_info[cache_type].num_cu_shared) - 1);
1630 first_active_cu = ffs(cu_sibling_map_mask);
1631
1632 /* CU could be inactive. In case of shared cache find the first active
1633 * CU. and incase of non-shared cache check if the CU is inactive. If
1634 * inactive active skip it
1635 */
1636 if (first_active_cu) {
1637 pcache = kfd_alloc_struct(pcache);
1638 if (!pcache)
1639 return -ENOMEM;
1640
1641 memset(pcache, 0, sizeof(struct kfd_cache_properties));
1642 pcache->processor_id_low = cu_processor_id + (first_active_cu - 1);
1643 pcache->cache_level = pcache_info[cache_type].cache_level;
1644 pcache->cache_size = pcache_info[cache_type].cache_size;
1645 pcache->cacheline_size = pcache_info[cache_type].cache_line_size;
1646
1647 if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_DATA_CACHE)
1648 pcache->cache_type |= HSA_CACHE_TYPE_DATA;
1649 if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_INST_CACHE)
1650 pcache->cache_type |= HSA_CACHE_TYPE_INSTRUCTION;
1651 if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_CPU_CACHE)
1652 pcache->cache_type |= HSA_CACHE_TYPE_CPU;
1653 if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_SIMD_CACHE)
1654 pcache->cache_type |= HSA_CACHE_TYPE_HSACU;
1655
1656 /* Sibling map is w.r.t processor_id_low, so shift out
1657 * inactive CU
1658 */
1659 cu_sibling_map_mask =
1660 cu_sibling_map_mask >> (first_active_cu - 1);
1661
1662 pcache->sibling_map[0] = (uint8_t)(cu_sibling_map_mask & 0xFF);
1663 pcache->sibling_map[1] =
1664 (uint8_t)((cu_sibling_map_mask >> 8) & 0xFF);
1665 pcache->sibling_map[2] =
1666 (uint8_t)((cu_sibling_map_mask >> 16) & 0xFF);
1667 pcache->sibling_map[3] =
1668 (uint8_t)((cu_sibling_map_mask >> 24) & 0xFF);
1669
1670 pcache->sibling_map_size = 4;
1671 *props_ext = pcache;
1672
1673 return 0;
1674 }
1675 return 1;
1676}
1677
1678/* Helper function. See kfd_fill_gpu_cache_info for parameter description */
1679static int fill_in_l2_l3_pcache(struct kfd_cache_properties **props_ext,
1680 struct kfd_gpu_cache_info *pcache_info,
1681 struct amdgpu_cu_info *cu_info,
1682 struct amdgpu_gfx_config *gfx_info,
1683 int cache_type, unsigned int cu_processor_id,
1684 struct kfd_node *knode)
1685{
1686 unsigned int cu_sibling_map_mask;
1687 int first_active_cu;
1688 int i, j, k, xcc, start, end;
1689 int num_xcc = NUM_XCC(knode->xcc_mask);
1690 struct kfd_cache_properties *pcache = NULL;
1691 enum amdgpu_memory_partition mode;
1692 struct amdgpu_device *adev = knode->adev;
1693
1694 start = ffs(knode->xcc_mask) - 1;
1695 end = start + num_xcc;
1696 cu_sibling_map_mask = cu_info->bitmap[start][0][0];
1697 cu_sibling_map_mask &=
1698 ((1 << pcache_info[cache_type].num_cu_shared) - 1);
1699 first_active_cu = ffs(cu_sibling_map_mask);
1700
1701 /* CU could be inactive. In case of shared cache find the first active
1702 * CU. and incase of non-shared cache check if the CU is inactive. If
1703 * inactive active skip it
1704 */
1705 if (first_active_cu) {
1706 pcache = kfd_alloc_struct(pcache);
1707 if (!pcache)
1708 return -ENOMEM;
1709
1710 memset(pcache, 0, sizeof(struct kfd_cache_properties));
1711 pcache->processor_id_low = cu_processor_id
1712 + (first_active_cu - 1);
1713 pcache->cache_level = pcache_info[cache_type].cache_level;
1714 pcache->cacheline_size = pcache_info[cache_type].cache_line_size;
1715
1716 if (KFD_GC_VERSION(knode) == IP_VERSION(9, 4, 3) ||
1717 KFD_GC_VERSION(knode) == IP_VERSION(9, 4, 4))
1718 mode = adev->gmc.gmc_funcs->query_mem_partition_mode(adev);
1719 else
1720 mode = UNKNOWN_MEMORY_PARTITION_MODE;
1721
1722 pcache->cache_size = pcache_info[cache_type].cache_size;
1723 /* Partition mode only affects L3 cache size */
1724 if (mode && pcache->cache_level == 3)
1725 pcache->cache_size /= mode;
1726
1727 if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_DATA_CACHE)
1728 pcache->cache_type |= HSA_CACHE_TYPE_DATA;
1729 if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_INST_CACHE)
1730 pcache->cache_type |= HSA_CACHE_TYPE_INSTRUCTION;
1731 if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_CPU_CACHE)
1732 pcache->cache_type |= HSA_CACHE_TYPE_CPU;
1733 if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_SIMD_CACHE)
1734 pcache->cache_type |= HSA_CACHE_TYPE_HSACU;
1735
1736 /* Sibling map is w.r.t processor_id_low, so shift out
1737 * inactive CU
1738 */
1739 cu_sibling_map_mask = cu_sibling_map_mask >> (first_active_cu - 1);
1740 k = 0;
1741
1742 for (xcc = start; xcc < end; xcc++) {
1743 for (i = 0; i < gfx_info->max_shader_engines; i++) {
1744 for (j = 0; j < gfx_info->max_sh_per_se; j++) {
1745 pcache->sibling_map[k] = (uint8_t)(cu_sibling_map_mask & 0xFF);
1746 pcache->sibling_map[k+1] = (uint8_t)((cu_sibling_map_mask >> 8) & 0xFF);
1747 pcache->sibling_map[k+2] = (uint8_t)((cu_sibling_map_mask >> 16) & 0xFF);
1748 pcache->sibling_map[k+3] = (uint8_t)((cu_sibling_map_mask >> 24) & 0xFF);
1749 k += 4;
1750
1751 cu_sibling_map_mask = cu_info->bitmap[xcc][i % 4][j + i / 4];
1752 cu_sibling_map_mask &= ((1 << pcache_info[cache_type].num_cu_shared) - 1);
1753 }
1754 }
1755 }
1756 pcache->sibling_map_size = k;
1757 *props_ext = pcache;
1758 return 0;
1759 }
1760 return 1;
1761}
1762
1763#define KFD_MAX_CACHE_TYPES 6
1764
1765/* kfd_fill_cache_non_crat_info - Fill GPU cache info using kfd_gpu_cache_info
1766 * tables
1767 */
1768static void kfd_fill_cache_non_crat_info(struct kfd_topology_device *dev, struct kfd_node *kdev)
1769{
1770 struct kfd_gpu_cache_info *pcache_info = NULL;
1771 int i, j, k, xcc, start, end;
1772 int ct = 0;
1773 unsigned int cu_processor_id;
1774 int ret;
1775 unsigned int num_cu_shared;
1776 struct amdgpu_cu_info *cu_info = &kdev->adev->gfx.cu_info;
1777 struct amdgpu_gfx_config *gfx_info = &kdev->adev->gfx.config;
1778 int gpu_processor_id;
1779 struct kfd_cache_properties *props_ext;
1780 int num_of_entries = 0;
1781 int num_of_cache_types = 0;
1782 struct kfd_gpu_cache_info cache_info[KFD_MAX_CACHE_TYPES];
1783
1784
1785 gpu_processor_id = dev->node_props.simd_id_base;
1786
1787 memset(cache_info, 0, sizeof(cache_info));
1788 pcache_info = cache_info;
1789 num_of_cache_types = kfd_get_gpu_cache_info(kdev, &pcache_info);
1790 if (!num_of_cache_types) {
1791 pr_warn("no cache info found\n");
1792 return;
1793 }
1794
1795 /* For each type of cache listed in the kfd_gpu_cache_info table,
1796 * go through all available Compute Units.
1797 * The [i,j,k] loop will
1798 * if kfd_gpu_cache_info.num_cu_shared = 1
1799 * will parse through all available CU
1800 * If (kfd_gpu_cache_info.num_cu_shared != 1)
1801 * then it will consider only one CU from
1802 * the shared unit
1803 */
1804 start = ffs(kdev->xcc_mask) - 1;
1805 end = start + NUM_XCC(kdev->xcc_mask);
1806
1807 for (ct = 0; ct < num_of_cache_types; ct++) {
1808 cu_processor_id = gpu_processor_id;
1809 if (pcache_info[ct].cache_level == 1) {
1810 for (xcc = start; xcc < end; xcc++) {
1811 for (i = 0; i < gfx_info->max_shader_engines; i++) {
1812 for (j = 0; j < gfx_info->max_sh_per_se; j++) {
1813 for (k = 0; k < gfx_info->max_cu_per_sh; k += pcache_info[ct].num_cu_shared) {
1814
1815 ret = fill_in_l1_pcache(&props_ext, pcache_info,
1816 cu_info->bitmap[xcc][i % 4][j + i / 4], ct,
1817 cu_processor_id, k);
1818
1819 if (ret < 0)
1820 break;
1821
1822 if (!ret) {
1823 num_of_entries++;
1824 list_add_tail(&props_ext->list, &dev->cache_props);
1825 }
1826
1827 /* Move to next CU block */
1828 num_cu_shared = ((k + pcache_info[ct].num_cu_shared) <=
1829 gfx_info->max_cu_per_sh) ?
1830 pcache_info[ct].num_cu_shared :
1831 (gfx_info->max_cu_per_sh - k);
1832 cu_processor_id += num_cu_shared;
1833 }
1834 }
1835 }
1836 }
1837 } else {
1838 ret = fill_in_l2_l3_pcache(&props_ext, pcache_info,
1839 cu_info, gfx_info, ct, cu_processor_id, kdev);
1840
1841 if (ret < 0)
1842 break;
1843
1844 if (!ret) {
1845 num_of_entries++;
1846 list_add_tail(&props_ext->list, &dev->cache_props);
1847 }
1848 }
1849 }
1850 dev->node_props.caches_count += num_of_entries;
1851 pr_debug("Added [%d] GPU cache entries\n", num_of_entries);
1852}
1853
1854static int kfd_topology_add_device_locked(struct kfd_node *gpu,
1855 struct kfd_topology_device **dev)
1856{
1857 int proximity_domain = ++topology_crat_proximity_domain;
1858 struct list_head temp_topology_device_list;
1859 void *crat_image = NULL;
1860 size_t image_size = 0;
1861 int res;
1862
1863 res = kfd_create_crat_image_virtual(&crat_image, &image_size,
1864 COMPUTE_UNIT_GPU, gpu,
1865 proximity_domain);
1866 if (res) {
1867 dev_err(gpu->adev->dev, "Error creating VCRAT\n");
1868 topology_crat_proximity_domain--;
1869 goto err;
1870 }
1871
1872 INIT_LIST_HEAD(&temp_topology_device_list);
1873
1874 res = kfd_parse_crat_table(crat_image,
1875 &temp_topology_device_list,
1876 proximity_domain);
1877 if (res) {
1878 dev_err(gpu->adev->dev, "Error parsing VCRAT\n");
1879 topology_crat_proximity_domain--;
1880 goto err;
1881 }
1882
1883 kfd_topology_update_device_list(&temp_topology_device_list,
1884 &topology_device_list);
1885
1886 *dev = kfd_assign_gpu(gpu);
1887 if (WARN_ON(!*dev)) {
1888 res = -ENODEV;
1889 goto err;
1890 }
1891
1892 /* Fill the cache affinity information here for the GPUs
1893 * using VCRAT
1894 */
1895 kfd_fill_cache_non_crat_info(*dev, gpu);
1896
1897 /* Update the SYSFS tree, since we added another topology
1898 * device
1899 */
1900 res = kfd_topology_update_sysfs();
1901 if (!res)
1902 sys_props.generation_count++;
1903 else
1904 dev_err(gpu->adev->dev, "Failed to update GPU to sysfs topology. res=%d\n",
1905 res);
1906
1907err:
1908 kfd_destroy_crat_image(crat_image);
1909 return res;
1910}
1911
1912static void kfd_topology_set_dbg_firmware_support(struct kfd_topology_device *dev)
1913{
1914 bool firmware_supported = true;
1915
1916 if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(11, 0, 0) &&
1917 KFD_GC_VERSION(dev->gpu) < IP_VERSION(12, 0, 0)) {
1918 uint32_t mes_api_rev = (dev->gpu->adev->mes.sched_version &
1919 AMDGPU_MES_API_VERSION_MASK) >>
1920 AMDGPU_MES_API_VERSION_SHIFT;
1921 uint32_t mes_rev = dev->gpu->adev->mes.sched_version &
1922 AMDGPU_MES_VERSION_MASK;
1923
1924 firmware_supported = (mes_api_rev >= 14) && (mes_rev >= 64);
1925 goto out;
1926 }
1927
1928 /*
1929 * Note: Any unlisted devices here are assumed to support exception handling.
1930 * Add additional checks here as needed.
1931 */
1932 switch (KFD_GC_VERSION(dev->gpu)) {
1933 case IP_VERSION(9, 0, 1):
1934 firmware_supported = dev->gpu->kfd->mec_fw_version >= 459 + 32768;
1935 break;
1936 case IP_VERSION(9, 1, 0):
1937 case IP_VERSION(9, 2, 1):
1938 case IP_VERSION(9, 2, 2):
1939 case IP_VERSION(9, 3, 0):
1940 case IP_VERSION(9, 4, 0):
1941 firmware_supported = dev->gpu->kfd->mec_fw_version >= 459;
1942 break;
1943 case IP_VERSION(9, 4, 1):
1944 firmware_supported = dev->gpu->kfd->mec_fw_version >= 60;
1945 break;
1946 case IP_VERSION(9, 4, 2):
1947 firmware_supported = dev->gpu->kfd->mec_fw_version >= 51;
1948 break;
1949 case IP_VERSION(10, 1, 10):
1950 case IP_VERSION(10, 1, 2):
1951 case IP_VERSION(10, 1, 1):
1952 firmware_supported = dev->gpu->kfd->mec_fw_version >= 144;
1953 break;
1954 case IP_VERSION(10, 3, 0):
1955 case IP_VERSION(10, 3, 2):
1956 case IP_VERSION(10, 3, 1):
1957 case IP_VERSION(10, 3, 4):
1958 case IP_VERSION(10, 3, 5):
1959 firmware_supported = dev->gpu->kfd->mec_fw_version >= 89;
1960 break;
1961 case IP_VERSION(10, 1, 3):
1962 case IP_VERSION(10, 3, 3):
1963 firmware_supported = false;
1964 break;
1965 default:
1966 break;
1967 }
1968
1969out:
1970 if (firmware_supported)
1971 dev->node_props.capability |= HSA_CAP_TRAP_DEBUG_FIRMWARE_SUPPORTED;
1972}
1973
1974static void kfd_topology_set_capabilities(struct kfd_topology_device *dev)
1975{
1976 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 <<
1977 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1978 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1979
1980 dev->node_props.capability |= HSA_CAP_TRAP_DEBUG_SUPPORT |
1981 HSA_CAP_TRAP_DEBUG_WAVE_LAUNCH_TRAP_OVERRIDE_SUPPORTED |
1982 HSA_CAP_TRAP_DEBUG_WAVE_LAUNCH_MODE_SUPPORTED;
1983
1984 if (kfd_dbg_has_ttmps_always_setup(dev->gpu))
1985 dev->node_props.debug_prop |= HSA_DBG_DISPATCH_INFO_ALWAYS_VALID;
1986
1987 if (KFD_GC_VERSION(dev->gpu) < IP_VERSION(10, 0, 0)) {
1988 if (KFD_GC_VERSION(dev->gpu) == IP_VERSION(9, 4, 3) ||
1989 KFD_GC_VERSION(dev->gpu) == IP_VERSION(9, 4, 4))
1990 dev->node_props.debug_prop |=
1991 HSA_DBG_WATCH_ADDR_MASK_LO_BIT_GFX9_4_3 |
1992 HSA_DBG_WATCH_ADDR_MASK_HI_BIT_GFX9_4_3;
1993 else
1994 dev->node_props.debug_prop |=
1995 HSA_DBG_WATCH_ADDR_MASK_LO_BIT_GFX9 |
1996 HSA_DBG_WATCH_ADDR_MASK_HI_BIT;
1997
1998 if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(9, 4, 2))
1999 dev->node_props.capability |=
2000 HSA_CAP_TRAP_DEBUG_PRECISE_MEMORY_OPERATIONS_SUPPORTED;
2001
2002 dev->node_props.capability |= HSA_CAP_PER_QUEUE_RESET_SUPPORTED;
2003 } else {
2004 dev->node_props.debug_prop |= HSA_DBG_WATCH_ADDR_MASK_LO_BIT_GFX10 |
2005 HSA_DBG_WATCH_ADDR_MASK_HI_BIT;
2006
2007 if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(11, 0, 0))
2008 dev->node_props.capability |=
2009 HSA_CAP_TRAP_DEBUG_PRECISE_MEMORY_OPERATIONS_SUPPORTED;
2010
2011 if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(12, 0, 0))
2012 dev->node_props.capability |=
2013 HSA_CAP_TRAP_DEBUG_PRECISE_ALU_OPERATIONS_SUPPORTED;
2014 }
2015
2016 kfd_topology_set_dbg_firmware_support(dev);
2017}
2018
2019int kfd_topology_add_device(struct kfd_node *gpu)
2020{
2021 uint32_t gpu_id;
2022 struct kfd_topology_device *dev;
2023 int res = 0;
2024 int i;
2025 const char *asic_name = amdgpu_asic_name[gpu->adev->asic_type];
2026 struct amdgpu_gfx_config *gfx_info = &gpu->adev->gfx.config;
2027 struct amdgpu_cu_info *cu_info = &gpu->adev->gfx.cu_info;
2028
2029 if (gpu->xcp && !gpu->xcp->ddev) {
2030 dev_warn(gpu->adev->dev,
2031 "Won't add GPU to topology since it has no drm node assigned.");
2032 return 0;
2033 } else {
2034 dev_dbg(gpu->adev->dev, "Adding new GPU to topology\n");
2035 }
2036
2037 /* Check to see if this gpu device exists in the topology_device_list.
2038 * If so, assign the gpu to that device,
2039 * else create a Virtual CRAT for this gpu device and then parse that
2040 * CRAT to create a new topology device. Once created assign the gpu to
2041 * that topology device
2042 */
2043 down_write(&topology_lock);
2044 dev = kfd_assign_gpu(gpu);
2045 if (!dev)
2046 res = kfd_topology_add_device_locked(gpu, &dev);
2047 up_write(&topology_lock);
2048 if (res)
2049 return res;
2050
2051 gpu_id = kfd_generate_gpu_id(gpu);
2052 dev->gpu_id = gpu_id;
2053 gpu->id = gpu_id;
2054
2055 kfd_dev_create_p2p_links();
2056
2057 /* TODO: Move the following lines to function
2058 * kfd_add_non_crat_information
2059 */
2060
2061 /* Fill-in additional information that is not available in CRAT but
2062 * needed for the topology
2063 */
2064 for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1; i++) {
2065 dev->node_props.name[i] = __tolower(asic_name[i]);
2066 if (asic_name[i] == '\0')
2067 break;
2068 }
2069 dev->node_props.name[i] = '\0';
2070
2071 dev->node_props.simd_arrays_per_engine =
2072 gfx_info->max_sh_per_se;
2073
2074 dev->node_props.gfx_target_version =
2075 gpu->kfd->device_info.gfx_target_version;
2076 dev->node_props.vendor_id = gpu->adev->pdev->vendor;
2077 dev->node_props.device_id = gpu->adev->pdev->device;
2078 dev->node_props.capability |=
2079 ((dev->gpu->adev->rev_id << HSA_CAP_ASIC_REVISION_SHIFT) &
2080 HSA_CAP_ASIC_REVISION_MASK);
2081
2082 dev->node_props.location_id = pci_dev_id(gpu->adev->pdev);
2083 if (gpu->kfd->num_nodes > 1)
2084 dev->node_props.location_id |= dev->gpu->node_id;
2085
2086 dev->node_props.domain = pci_domain_nr(gpu->adev->pdev->bus);
2087 dev->node_props.max_engine_clk_fcompute =
2088 amdgpu_amdkfd_get_max_engine_clock_in_mhz(dev->gpu->adev);
2089 dev->node_props.max_engine_clk_ccompute =
2090 cpufreq_quick_get_max(0) / 1000;
2091
2092 if (gpu->xcp)
2093 dev->node_props.drm_render_minor = gpu->xcp->ddev->render->index;
2094 else
2095 dev->node_props.drm_render_minor =
2096 gpu->kfd->shared_resources.drm_render_minor;
2097
2098 dev->node_props.hive_id = gpu->kfd->hive_id;
2099 dev->node_props.num_sdma_engines = kfd_get_num_sdma_engines(gpu);
2100 dev->node_props.num_sdma_xgmi_engines =
2101 kfd_get_num_xgmi_sdma_engines(gpu);
2102 dev->node_props.num_sdma_queues_per_engine =
2103 gpu->kfd->device_info.num_sdma_queues_per_engine -
2104 gpu->kfd->device_info.num_reserved_sdma_queues_per_engine;
2105 dev->node_props.num_gws = (dev->gpu->gws &&
2106 dev->gpu->dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) ?
2107 dev->gpu->adev->gds.gws_size : 0;
2108 dev->node_props.num_cp_queues = get_cp_queues_num(dev->gpu->dqm);
2109
2110 kfd_fill_mem_clk_max_info(dev);
2111 kfd_fill_iolink_non_crat_info(dev);
2112
2113 switch (dev->gpu->adev->asic_type) {
2114 case CHIP_KAVERI:
2115 case CHIP_HAWAII:
2116 case CHIP_TONGA:
2117 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_PRE_1_0 <<
2118 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
2119 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
2120 break;
2121 case CHIP_CARRIZO:
2122 case CHIP_FIJI:
2123 case CHIP_POLARIS10:
2124 case CHIP_POLARIS11:
2125 case CHIP_POLARIS12:
2126 case CHIP_VEGAM:
2127 pr_debug("Adding doorbell packet type capability\n");
2128 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 <<
2129 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
2130 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
2131 break;
2132 default:
2133 if (KFD_GC_VERSION(dev->gpu) < IP_VERSION(9, 0, 1))
2134 WARN(1, "Unexpected ASIC family %u",
2135 dev->gpu->adev->asic_type);
2136 else
2137 kfd_topology_set_capabilities(dev);
2138 }
2139
2140 /*
2141 * Overwrite ATS capability according to needs_iommu_device to fix
2142 * potential missing corresponding bit in CRAT of BIOS.
2143 */
2144 dev->node_props.capability &= ~HSA_CAP_ATS_PRESENT;
2145
2146 /* Fix errors in CZ CRAT.
2147 * simd_count: Carrizo CRAT reports wrong simd_count, probably
2148 * because it doesn't consider masked out CUs
2149 * max_waves_per_simd: Carrizo reports wrong max_waves_per_simd
2150 */
2151 if (dev->gpu->adev->asic_type == CHIP_CARRIZO) {
2152 dev->node_props.simd_count =
2153 cu_info->simd_per_cu * cu_info->number;
2154 dev->node_props.max_waves_per_simd = 10;
2155 }
2156
2157 /* kfd only concerns sram ecc on GFX and HBM ecc on UMC */
2158 dev->node_props.capability |=
2159 ((dev->gpu->adev->ras_enabled & BIT(AMDGPU_RAS_BLOCK__GFX)) != 0) ?
2160 HSA_CAP_SRAM_EDCSUPPORTED : 0;
2161 dev->node_props.capability |=
2162 ((dev->gpu->adev->ras_enabled & BIT(AMDGPU_RAS_BLOCK__UMC)) != 0) ?
2163 HSA_CAP_MEM_EDCSUPPORTED : 0;
2164
2165 if (KFD_GC_VERSION(dev->gpu) != IP_VERSION(9, 0, 1))
2166 dev->node_props.capability |= (dev->gpu->adev->ras_enabled != 0) ?
2167 HSA_CAP_RASEVENTNOTIFY : 0;
2168
2169 if (KFD_IS_SVM_API_SUPPORTED(dev->gpu->adev))
2170 dev->node_props.capability |= HSA_CAP_SVMAPI_SUPPORTED;
2171
2172 if (dev->gpu->adev->gmc.is_app_apu ||
2173 dev->gpu->adev->gmc.xgmi.connected_to_cpu)
2174 dev->node_props.capability |= HSA_CAP_FLAGS_COHERENTHOSTACCESS;
2175
2176 kfd_queue_ctx_save_restore_size(dev);
2177
2178 kfd_debug_print_topology();
2179
2180 kfd_notify_gpu_change(gpu_id, 1);
2181
2182 return 0;
2183}
2184
2185/**
2186 * kfd_topology_update_io_links() - Update IO links after device removal.
2187 * @proximity_domain: Proximity domain value of the dev being removed.
2188 *
2189 * The topology list currently is arranged in increasing order of
2190 * proximity domain.
2191 *
2192 * Two things need to be done when a device is removed:
2193 * 1. All the IO links to this device need to be removed.
2194 * 2. All nodes after the current device node need to move
2195 * up once this device node is removed from the topology
2196 * list. As a result, the proximity domain values for
2197 * all nodes after the node being deleted reduce by 1.
2198 * This would also cause the proximity domain values for
2199 * io links to be updated based on new proximity domain
2200 * values.
2201 *
2202 * Context: The caller must hold write topology_lock.
2203 */
2204static void kfd_topology_update_io_links(int proximity_domain)
2205{
2206 struct kfd_topology_device *dev;
2207 struct kfd_iolink_properties *iolink, *p2plink, *tmp;
2208
2209 list_for_each_entry(dev, &topology_device_list, list) {
2210 if (dev->proximity_domain > proximity_domain)
2211 dev->proximity_domain--;
2212
2213 list_for_each_entry_safe(iolink, tmp, &dev->io_link_props, list) {
2214 /*
2215 * If there is an io link to the dev being deleted
2216 * then remove that IO link also.
2217 */
2218 if (iolink->node_to == proximity_domain) {
2219 list_del(&iolink->list);
2220 dev->node_props.io_links_count--;
2221 } else {
2222 if (iolink->node_from > proximity_domain)
2223 iolink->node_from--;
2224 if (iolink->node_to > proximity_domain)
2225 iolink->node_to--;
2226 }
2227 }
2228
2229 list_for_each_entry_safe(p2plink, tmp, &dev->p2p_link_props, list) {
2230 /*
2231 * If there is a p2p link to the dev being deleted
2232 * then remove that p2p link also.
2233 */
2234 if (p2plink->node_to == proximity_domain) {
2235 list_del(&p2plink->list);
2236 dev->node_props.p2p_links_count--;
2237 } else {
2238 if (p2plink->node_from > proximity_domain)
2239 p2plink->node_from--;
2240 if (p2plink->node_to > proximity_domain)
2241 p2plink->node_to--;
2242 }
2243 }
2244 }
2245}
2246
2247int kfd_topology_remove_device(struct kfd_node *gpu)
2248{
2249 struct kfd_topology_device *dev, *tmp;
2250 uint32_t gpu_id;
2251 int res = -ENODEV;
2252 int i = 0;
2253
2254 down_write(&topology_lock);
2255
2256 list_for_each_entry_safe(dev, tmp, &topology_device_list, list) {
2257 if (dev->gpu == gpu) {
2258 gpu_id = dev->gpu_id;
2259 kfd_remove_sysfs_node_entry(dev);
2260 kfd_release_topology_device(dev);
2261 sys_props.num_devices--;
2262 kfd_topology_update_io_links(i);
2263 topology_crat_proximity_domain = sys_props.num_devices-1;
2264 sys_props.generation_count++;
2265 res = 0;
2266 if (kfd_topology_update_sysfs() < 0)
2267 kfd_topology_release_sysfs();
2268 break;
2269 }
2270 i++;
2271 }
2272
2273 up_write(&topology_lock);
2274
2275 if (!res)
2276 kfd_notify_gpu_change(gpu_id, 0);
2277
2278 return res;
2279}
2280
2281/* kfd_topology_enum_kfd_devices - Enumerate through all devices in KFD
2282 * topology. If GPU device is found @idx, then valid kfd_dev pointer is
2283 * returned through @kdev
2284 * Return - 0: On success (@kdev will be NULL for non GPU nodes)
2285 * -1: If end of list
2286 */
2287int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_node **kdev)
2288{
2289
2290 struct kfd_topology_device *top_dev;
2291 uint8_t device_idx = 0;
2292
2293 *kdev = NULL;
2294 down_read(&topology_lock);
2295
2296 list_for_each_entry(top_dev, &topology_device_list, list) {
2297 if (device_idx == idx) {
2298 *kdev = top_dev->gpu;
2299 up_read(&topology_lock);
2300 return 0;
2301 }
2302
2303 device_idx++;
2304 }
2305
2306 up_read(&topology_lock);
2307
2308 return -1;
2309
2310}
2311
2312static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask)
2313{
2314 int first_cpu_of_numa_node;
2315
2316 if (!cpumask || cpumask == cpu_none_mask)
2317 return -1;
2318 first_cpu_of_numa_node = cpumask_first(cpumask);
2319 if (first_cpu_of_numa_node >= nr_cpu_ids)
2320 return -1;
2321#ifdef CONFIG_X86_64
2322 return cpu_data(first_cpu_of_numa_node).topo.apicid;
2323#else
2324 return first_cpu_of_numa_node;
2325#endif
2326}
2327
2328/* kfd_numa_node_to_apic_id - Returns the APIC ID of the first logical processor
2329 * of the given NUMA node (numa_node_id)
2330 * Return -1 on failure
2331 */
2332int kfd_numa_node_to_apic_id(int numa_node_id)
2333{
2334 if (numa_node_id == -1) {
2335 pr_warn("Invalid NUMA Node. Use online CPU mask\n");
2336 return kfd_cpumask_to_apic_id(cpu_online_mask);
2337 }
2338 return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id));
2339}
2340
2341#if defined(CONFIG_DEBUG_FS)
2342
2343int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data)
2344{
2345 struct kfd_topology_device *dev;
2346 unsigned int i = 0;
2347 int r = 0;
2348
2349 down_read(&topology_lock);
2350
2351 list_for_each_entry(dev, &topology_device_list, list) {
2352 if (!dev->gpu) {
2353 i++;
2354 continue;
2355 }
2356
2357 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
2358 r = dqm_debugfs_hqds(m, dev->gpu->dqm);
2359 if (r)
2360 break;
2361 }
2362
2363 up_read(&topology_lock);
2364
2365 return r;
2366}
2367
2368int kfd_debugfs_rls_by_device(struct seq_file *m, void *data)
2369{
2370 struct kfd_topology_device *dev;
2371 unsigned int i = 0;
2372 int r = 0;
2373
2374 down_read(&topology_lock);
2375
2376 list_for_each_entry(dev, &topology_device_list, list) {
2377 if (!dev->gpu) {
2378 i++;
2379 continue;
2380 }
2381
2382 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
2383 r = pm_debugfs_runlist(m, &dev->gpu->dqm->packet_mgr);
2384 if (r)
2385 break;
2386 }
2387
2388 up_read(&topology_lock);
2389
2390 return r;
2391}
2392
2393#endif