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#include <linux/dmi.h>
32#include <linux/atomic.h>
33
34#include "kfd_priv.h"
35#include "kfd_crat.h"
36#include "kfd_topology.h"
37#include "kfd_device_queue_manager.h"
38#include "kfd_iommu.h"
39
40/* topology_device_list - Master list of all topology devices */
41static struct list_head topology_device_list;
42static struct kfd_system_properties sys_props;
43
44static DECLARE_RWSEM(topology_lock);
45static atomic_t topology_crat_proximity_domain;
46
47struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
48 uint32_t proximity_domain)
49{
50 struct kfd_topology_device *top_dev;
51 struct kfd_topology_device *device = NULL;
52
53 down_read(&topology_lock);
54
55 list_for_each_entry(top_dev, &topology_device_list, list)
56 if (top_dev->proximity_domain == proximity_domain) {
57 device = top_dev;
58 break;
59 }
60
61 up_read(&topology_lock);
62
63 return device;
64}
65
66struct kfd_dev *kfd_device_by_id(uint32_t gpu_id)
67{
68 struct kfd_topology_device *top_dev;
69 struct kfd_dev *device = NULL;
70
71 down_read(&topology_lock);
72
73 list_for_each_entry(top_dev, &topology_device_list, list)
74 if (top_dev->gpu_id == gpu_id) {
75 device = top_dev->gpu;
76 break;
77 }
78
79 up_read(&topology_lock);
80
81 return device;
82}
83
84struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev)
85{
86 struct kfd_topology_device *top_dev;
87 struct kfd_dev *device = NULL;
88
89 down_read(&topology_lock);
90
91 list_for_each_entry(top_dev, &topology_device_list, list)
92 if (top_dev->gpu->pdev == pdev) {
93 device = top_dev->gpu;
94 break;
95 }
96
97 up_read(&topology_lock);
98
99 return device;
100}
101
102/* Called with write topology_lock acquired */
103static void kfd_release_topology_device(struct kfd_topology_device *dev)
104{
105 struct kfd_mem_properties *mem;
106 struct kfd_cache_properties *cache;
107 struct kfd_iolink_properties *iolink;
108 struct kfd_perf_properties *perf;
109
110 list_del(&dev->list);
111
112 while (dev->mem_props.next != &dev->mem_props) {
113 mem = container_of(dev->mem_props.next,
114 struct kfd_mem_properties, list);
115 list_del(&mem->list);
116 kfree(mem);
117 }
118
119 while (dev->cache_props.next != &dev->cache_props) {
120 cache = container_of(dev->cache_props.next,
121 struct kfd_cache_properties, list);
122 list_del(&cache->list);
123 kfree(cache);
124 }
125
126 while (dev->io_link_props.next != &dev->io_link_props) {
127 iolink = container_of(dev->io_link_props.next,
128 struct kfd_iolink_properties, list);
129 list_del(&iolink->list);
130 kfree(iolink);
131 }
132
133 while (dev->perf_props.next != &dev->perf_props) {
134 perf = container_of(dev->perf_props.next,
135 struct kfd_perf_properties, list);
136 list_del(&perf->list);
137 kfree(perf);
138 }
139
140 kfree(dev);
141}
142
143void kfd_release_topology_device_list(struct list_head *device_list)
144{
145 struct kfd_topology_device *dev;
146
147 while (!list_empty(device_list)) {
148 dev = list_first_entry(device_list,
149 struct kfd_topology_device, list);
150 kfd_release_topology_device(dev);
151 }
152}
153
154static void kfd_release_live_view(void)
155{
156 kfd_release_topology_device_list(&topology_device_list);
157 memset(&sys_props, 0, sizeof(sys_props));
158}
159
160struct kfd_topology_device *kfd_create_topology_device(
161 struct list_head *device_list)
162{
163 struct kfd_topology_device *dev;
164
165 dev = kfd_alloc_struct(dev);
166 if (!dev) {
167 pr_err("No memory to allocate a topology device");
168 return NULL;
169 }
170
171 INIT_LIST_HEAD(&dev->mem_props);
172 INIT_LIST_HEAD(&dev->cache_props);
173 INIT_LIST_HEAD(&dev->io_link_props);
174 INIT_LIST_HEAD(&dev->perf_props);
175
176 list_add_tail(&dev->list, device_list);
177
178 return dev;
179}
180
181
182#define sysfs_show_gen_prop(buffer, fmt, ...) \
183 snprintf(buffer, PAGE_SIZE, "%s"fmt, buffer, __VA_ARGS__)
184#define sysfs_show_32bit_prop(buffer, name, value) \
185 sysfs_show_gen_prop(buffer, "%s %u\n", name, value)
186#define sysfs_show_64bit_prop(buffer, name, value) \
187 sysfs_show_gen_prop(buffer, "%s %llu\n", name, value)
188#define sysfs_show_32bit_val(buffer, value) \
189 sysfs_show_gen_prop(buffer, "%u\n", value)
190#define sysfs_show_str_val(buffer, value) \
191 sysfs_show_gen_prop(buffer, "%s\n", value)
192
193static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr,
194 char *buffer)
195{
196 ssize_t ret;
197
198 /* Making sure that the buffer is an empty string */
199 buffer[0] = 0;
200
201 if (attr == &sys_props.attr_genid) {
202 ret = sysfs_show_32bit_val(buffer, sys_props.generation_count);
203 } else if (attr == &sys_props.attr_props) {
204 sysfs_show_64bit_prop(buffer, "platform_oem",
205 sys_props.platform_oem);
206 sysfs_show_64bit_prop(buffer, "platform_id",
207 sys_props.platform_id);
208 ret = sysfs_show_64bit_prop(buffer, "platform_rev",
209 sys_props.platform_rev);
210 } else {
211 ret = -EINVAL;
212 }
213
214 return ret;
215}
216
217static void kfd_topology_kobj_release(struct kobject *kobj)
218{
219 kfree(kobj);
220}
221
222static const struct sysfs_ops sysprops_ops = {
223 .show = sysprops_show,
224};
225
226static struct kobj_type sysprops_type = {
227 .release = kfd_topology_kobj_release,
228 .sysfs_ops = &sysprops_ops,
229};
230
231static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr,
232 char *buffer)
233{
234 ssize_t ret;
235 struct kfd_iolink_properties *iolink;
236
237 /* Making sure that the buffer is an empty string */
238 buffer[0] = 0;
239
240 iolink = container_of(attr, struct kfd_iolink_properties, attr);
241 sysfs_show_32bit_prop(buffer, "type", iolink->iolink_type);
242 sysfs_show_32bit_prop(buffer, "version_major", iolink->ver_maj);
243 sysfs_show_32bit_prop(buffer, "version_minor", iolink->ver_min);
244 sysfs_show_32bit_prop(buffer, "node_from", iolink->node_from);
245 sysfs_show_32bit_prop(buffer, "node_to", iolink->node_to);
246 sysfs_show_32bit_prop(buffer, "weight", iolink->weight);
247 sysfs_show_32bit_prop(buffer, "min_latency", iolink->min_latency);
248 sysfs_show_32bit_prop(buffer, "max_latency", iolink->max_latency);
249 sysfs_show_32bit_prop(buffer, "min_bandwidth", iolink->min_bandwidth);
250 sysfs_show_32bit_prop(buffer, "max_bandwidth", iolink->max_bandwidth);
251 sysfs_show_32bit_prop(buffer, "recommended_transfer_size",
252 iolink->rec_transfer_size);
253 ret = sysfs_show_32bit_prop(buffer, "flags", iolink->flags);
254
255 return ret;
256}
257
258static const struct sysfs_ops iolink_ops = {
259 .show = iolink_show,
260};
261
262static struct kobj_type iolink_type = {
263 .release = kfd_topology_kobj_release,
264 .sysfs_ops = &iolink_ops,
265};
266
267static ssize_t mem_show(struct kobject *kobj, struct attribute *attr,
268 char *buffer)
269{
270 ssize_t ret;
271 struct kfd_mem_properties *mem;
272
273 /* Making sure that the buffer is an empty string */
274 buffer[0] = 0;
275
276 mem = container_of(attr, struct kfd_mem_properties, attr);
277 sysfs_show_32bit_prop(buffer, "heap_type", mem->heap_type);
278 sysfs_show_64bit_prop(buffer, "size_in_bytes", mem->size_in_bytes);
279 sysfs_show_32bit_prop(buffer, "flags", mem->flags);
280 sysfs_show_32bit_prop(buffer, "width", mem->width);
281 ret = sysfs_show_32bit_prop(buffer, "mem_clk_max", mem->mem_clk_max);
282
283 return ret;
284}
285
286static const struct sysfs_ops mem_ops = {
287 .show = mem_show,
288};
289
290static struct kobj_type mem_type = {
291 .release = kfd_topology_kobj_release,
292 .sysfs_ops = &mem_ops,
293};
294
295static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr,
296 char *buffer)
297{
298 ssize_t ret;
299 uint32_t i, j;
300 struct kfd_cache_properties *cache;
301
302 /* Making sure that the buffer is an empty string */
303 buffer[0] = 0;
304
305 cache = container_of(attr, struct kfd_cache_properties, attr);
306 sysfs_show_32bit_prop(buffer, "processor_id_low",
307 cache->processor_id_low);
308 sysfs_show_32bit_prop(buffer, "level", cache->cache_level);
309 sysfs_show_32bit_prop(buffer, "size", cache->cache_size);
310 sysfs_show_32bit_prop(buffer, "cache_line_size", cache->cacheline_size);
311 sysfs_show_32bit_prop(buffer, "cache_lines_per_tag",
312 cache->cachelines_per_tag);
313 sysfs_show_32bit_prop(buffer, "association", cache->cache_assoc);
314 sysfs_show_32bit_prop(buffer, "latency", cache->cache_latency);
315 sysfs_show_32bit_prop(buffer, "type", cache->cache_type);
316 snprintf(buffer, PAGE_SIZE, "%ssibling_map ", buffer);
317 for (i = 0; i < CRAT_SIBLINGMAP_SIZE; i++)
318 for (j = 0; j < sizeof(cache->sibling_map[0])*8; j++) {
319 /* Check each bit */
320 if (cache->sibling_map[i] & (1 << j))
321 ret = snprintf(buffer, PAGE_SIZE,
322 "%s%d%s", buffer, 1, ",");
323 else
324 ret = snprintf(buffer, PAGE_SIZE,
325 "%s%d%s", buffer, 0, ",");
326 }
327 /* Replace the last "," with end of line */
328 *(buffer + strlen(buffer) - 1) = 0xA;
329 return ret;
330}
331
332static const struct sysfs_ops cache_ops = {
333 .show = kfd_cache_show,
334};
335
336static struct kobj_type cache_type = {
337 .release = kfd_topology_kobj_release,
338 .sysfs_ops = &cache_ops,
339};
340
341/****** Sysfs of Performance Counters ******/
342
343struct kfd_perf_attr {
344 struct kobj_attribute attr;
345 uint32_t data;
346};
347
348static ssize_t perf_show(struct kobject *kobj, struct kobj_attribute *attrs,
349 char *buf)
350{
351 struct kfd_perf_attr *attr;
352
353 buf[0] = 0;
354 attr = container_of(attrs, struct kfd_perf_attr, attr);
355 if (!attr->data) /* invalid data for PMC */
356 return 0;
357 else
358 return sysfs_show_32bit_val(buf, attr->data);
359}
360
361#define KFD_PERF_DESC(_name, _data) \
362{ \
363 .attr = __ATTR(_name, 0444, perf_show, NULL), \
364 .data = _data, \
365}
366
367static struct kfd_perf_attr perf_attr_iommu[] = {
368 KFD_PERF_DESC(max_concurrent, 0),
369 KFD_PERF_DESC(num_counters, 0),
370 KFD_PERF_DESC(counter_ids, 0),
371};
372/****************************************/
373
374static ssize_t node_show(struct kobject *kobj, struct attribute *attr,
375 char *buffer)
376{
377 struct kfd_topology_device *dev;
378 char public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE];
379 uint32_t i;
380 uint32_t log_max_watch_addr;
381
382 /* Making sure that the buffer is an empty string */
383 buffer[0] = 0;
384
385 if (strcmp(attr->name, "gpu_id") == 0) {
386 dev = container_of(attr, struct kfd_topology_device,
387 attr_gpuid);
388 return sysfs_show_32bit_val(buffer, dev->gpu_id);
389 }
390
391 if (strcmp(attr->name, "name") == 0) {
392 dev = container_of(attr, struct kfd_topology_device,
393 attr_name);
394 for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE; i++) {
395 public_name[i] =
396 (char)dev->node_props.marketing_name[i];
397 if (dev->node_props.marketing_name[i] == 0)
398 break;
399 }
400 public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1] = 0x0;
401 return sysfs_show_str_val(buffer, public_name);
402 }
403
404 dev = container_of(attr, struct kfd_topology_device,
405 attr_props);
406 sysfs_show_32bit_prop(buffer, "cpu_cores_count",
407 dev->node_props.cpu_cores_count);
408 sysfs_show_32bit_prop(buffer, "simd_count",
409 dev->node_props.simd_count);
410 sysfs_show_32bit_prop(buffer, "mem_banks_count",
411 dev->node_props.mem_banks_count);
412 sysfs_show_32bit_prop(buffer, "caches_count",
413 dev->node_props.caches_count);
414 sysfs_show_32bit_prop(buffer, "io_links_count",
415 dev->node_props.io_links_count);
416 sysfs_show_32bit_prop(buffer, "cpu_core_id_base",
417 dev->node_props.cpu_core_id_base);
418 sysfs_show_32bit_prop(buffer, "simd_id_base",
419 dev->node_props.simd_id_base);
420 sysfs_show_32bit_prop(buffer, "max_waves_per_simd",
421 dev->node_props.max_waves_per_simd);
422 sysfs_show_32bit_prop(buffer, "lds_size_in_kb",
423 dev->node_props.lds_size_in_kb);
424 sysfs_show_32bit_prop(buffer, "gds_size_in_kb",
425 dev->node_props.gds_size_in_kb);
426 sysfs_show_32bit_prop(buffer, "wave_front_size",
427 dev->node_props.wave_front_size);
428 sysfs_show_32bit_prop(buffer, "array_count",
429 dev->node_props.array_count);
430 sysfs_show_32bit_prop(buffer, "simd_arrays_per_engine",
431 dev->node_props.simd_arrays_per_engine);
432 sysfs_show_32bit_prop(buffer, "cu_per_simd_array",
433 dev->node_props.cu_per_simd_array);
434 sysfs_show_32bit_prop(buffer, "simd_per_cu",
435 dev->node_props.simd_per_cu);
436 sysfs_show_32bit_prop(buffer, "max_slots_scratch_cu",
437 dev->node_props.max_slots_scratch_cu);
438 sysfs_show_32bit_prop(buffer, "vendor_id",
439 dev->node_props.vendor_id);
440 sysfs_show_32bit_prop(buffer, "device_id",
441 dev->node_props.device_id);
442 sysfs_show_32bit_prop(buffer, "location_id",
443 dev->node_props.location_id);
444 sysfs_show_32bit_prop(buffer, "drm_render_minor",
445 dev->node_props.drm_render_minor);
446
447 if (dev->gpu) {
448 log_max_watch_addr =
449 __ilog2_u32(dev->gpu->device_info->num_of_watch_points);
450
451 if (log_max_watch_addr) {
452 dev->node_props.capability |=
453 HSA_CAP_WATCH_POINTS_SUPPORTED;
454
455 dev->node_props.capability |=
456 ((log_max_watch_addr <<
457 HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) &
458 HSA_CAP_WATCH_POINTS_TOTALBITS_MASK);
459 }
460
461 if (dev->gpu->device_info->asic_family == CHIP_TONGA)
462 dev->node_props.capability |=
463 HSA_CAP_AQL_QUEUE_DOUBLE_MAP;
464
465 sysfs_show_32bit_prop(buffer, "max_engine_clk_fcompute",
466 dev->node_props.max_engine_clk_fcompute);
467
468 sysfs_show_64bit_prop(buffer, "local_mem_size",
469 (unsigned long long int) 0);
470
471 sysfs_show_32bit_prop(buffer, "fw_version",
472 dev->gpu->kfd2kgd->get_fw_version(
473 dev->gpu->kgd,
474 KGD_ENGINE_MEC1));
475 sysfs_show_32bit_prop(buffer, "capability",
476 dev->node_props.capability);
477 }
478
479 return sysfs_show_32bit_prop(buffer, "max_engine_clk_ccompute",
480 cpufreq_quick_get_max(0)/1000);
481}
482
483static const struct sysfs_ops node_ops = {
484 .show = node_show,
485};
486
487static struct kobj_type node_type = {
488 .release = kfd_topology_kobj_release,
489 .sysfs_ops = &node_ops,
490};
491
492static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr)
493{
494 sysfs_remove_file(kobj, attr);
495 kobject_del(kobj);
496 kobject_put(kobj);
497}
498
499static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev)
500{
501 struct kfd_iolink_properties *iolink;
502 struct kfd_cache_properties *cache;
503 struct kfd_mem_properties *mem;
504 struct kfd_perf_properties *perf;
505
506 if (dev->kobj_iolink) {
507 list_for_each_entry(iolink, &dev->io_link_props, list)
508 if (iolink->kobj) {
509 kfd_remove_sysfs_file(iolink->kobj,
510 &iolink->attr);
511 iolink->kobj = NULL;
512 }
513 kobject_del(dev->kobj_iolink);
514 kobject_put(dev->kobj_iolink);
515 dev->kobj_iolink = NULL;
516 }
517
518 if (dev->kobj_cache) {
519 list_for_each_entry(cache, &dev->cache_props, list)
520 if (cache->kobj) {
521 kfd_remove_sysfs_file(cache->kobj,
522 &cache->attr);
523 cache->kobj = NULL;
524 }
525 kobject_del(dev->kobj_cache);
526 kobject_put(dev->kobj_cache);
527 dev->kobj_cache = NULL;
528 }
529
530 if (dev->kobj_mem) {
531 list_for_each_entry(mem, &dev->mem_props, list)
532 if (mem->kobj) {
533 kfd_remove_sysfs_file(mem->kobj, &mem->attr);
534 mem->kobj = NULL;
535 }
536 kobject_del(dev->kobj_mem);
537 kobject_put(dev->kobj_mem);
538 dev->kobj_mem = NULL;
539 }
540
541 if (dev->kobj_perf) {
542 list_for_each_entry(perf, &dev->perf_props, list) {
543 kfree(perf->attr_group);
544 perf->attr_group = NULL;
545 }
546 kobject_del(dev->kobj_perf);
547 kobject_put(dev->kobj_perf);
548 dev->kobj_perf = NULL;
549 }
550
551 if (dev->kobj_node) {
552 sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid);
553 sysfs_remove_file(dev->kobj_node, &dev->attr_name);
554 sysfs_remove_file(dev->kobj_node, &dev->attr_props);
555 kobject_del(dev->kobj_node);
556 kobject_put(dev->kobj_node);
557 dev->kobj_node = NULL;
558 }
559}
560
561static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev,
562 uint32_t id)
563{
564 struct kfd_iolink_properties *iolink;
565 struct kfd_cache_properties *cache;
566 struct kfd_mem_properties *mem;
567 struct kfd_perf_properties *perf;
568 int ret;
569 uint32_t i, num_attrs;
570 struct attribute **attrs;
571
572 if (WARN_ON(dev->kobj_node))
573 return -EEXIST;
574
575 /*
576 * Creating the sysfs folders
577 */
578 dev->kobj_node = kfd_alloc_struct(dev->kobj_node);
579 if (!dev->kobj_node)
580 return -ENOMEM;
581
582 ret = kobject_init_and_add(dev->kobj_node, &node_type,
583 sys_props.kobj_nodes, "%d", id);
584 if (ret < 0)
585 return ret;
586
587 dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node);
588 if (!dev->kobj_mem)
589 return -ENOMEM;
590
591 dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node);
592 if (!dev->kobj_cache)
593 return -ENOMEM;
594
595 dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node);
596 if (!dev->kobj_iolink)
597 return -ENOMEM;
598
599 dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node);
600 if (!dev->kobj_perf)
601 return -ENOMEM;
602
603 /*
604 * Creating sysfs files for node properties
605 */
606 dev->attr_gpuid.name = "gpu_id";
607 dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE;
608 sysfs_attr_init(&dev->attr_gpuid);
609 dev->attr_name.name = "name";
610 dev->attr_name.mode = KFD_SYSFS_FILE_MODE;
611 sysfs_attr_init(&dev->attr_name);
612 dev->attr_props.name = "properties";
613 dev->attr_props.mode = KFD_SYSFS_FILE_MODE;
614 sysfs_attr_init(&dev->attr_props);
615 ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid);
616 if (ret < 0)
617 return ret;
618 ret = sysfs_create_file(dev->kobj_node, &dev->attr_name);
619 if (ret < 0)
620 return ret;
621 ret = sysfs_create_file(dev->kobj_node, &dev->attr_props);
622 if (ret < 0)
623 return ret;
624
625 i = 0;
626 list_for_each_entry(mem, &dev->mem_props, list) {
627 mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
628 if (!mem->kobj)
629 return -ENOMEM;
630 ret = kobject_init_and_add(mem->kobj, &mem_type,
631 dev->kobj_mem, "%d", i);
632 if (ret < 0)
633 return ret;
634
635 mem->attr.name = "properties";
636 mem->attr.mode = KFD_SYSFS_FILE_MODE;
637 sysfs_attr_init(&mem->attr);
638 ret = sysfs_create_file(mem->kobj, &mem->attr);
639 if (ret < 0)
640 return ret;
641 i++;
642 }
643
644 i = 0;
645 list_for_each_entry(cache, &dev->cache_props, list) {
646 cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
647 if (!cache->kobj)
648 return -ENOMEM;
649 ret = kobject_init_and_add(cache->kobj, &cache_type,
650 dev->kobj_cache, "%d", i);
651 if (ret < 0)
652 return ret;
653
654 cache->attr.name = "properties";
655 cache->attr.mode = KFD_SYSFS_FILE_MODE;
656 sysfs_attr_init(&cache->attr);
657 ret = sysfs_create_file(cache->kobj, &cache->attr);
658 if (ret < 0)
659 return ret;
660 i++;
661 }
662
663 i = 0;
664 list_for_each_entry(iolink, &dev->io_link_props, list) {
665 iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
666 if (!iolink->kobj)
667 return -ENOMEM;
668 ret = kobject_init_and_add(iolink->kobj, &iolink_type,
669 dev->kobj_iolink, "%d", i);
670 if (ret < 0)
671 return ret;
672
673 iolink->attr.name = "properties";
674 iolink->attr.mode = KFD_SYSFS_FILE_MODE;
675 sysfs_attr_init(&iolink->attr);
676 ret = sysfs_create_file(iolink->kobj, &iolink->attr);
677 if (ret < 0)
678 return ret;
679 i++;
680 }
681
682 /* All hardware blocks have the same number of attributes. */
683 num_attrs = ARRAY_SIZE(perf_attr_iommu);
684 list_for_each_entry(perf, &dev->perf_props, list) {
685 perf->attr_group = kzalloc(sizeof(struct kfd_perf_attr)
686 * num_attrs + sizeof(struct attribute_group),
687 GFP_KERNEL);
688 if (!perf->attr_group)
689 return -ENOMEM;
690
691 attrs = (struct attribute **)(perf->attr_group + 1);
692 if (!strcmp(perf->block_name, "iommu")) {
693 /* Information of IOMMU's num_counters and counter_ids is shown
694 * under /sys/bus/event_source/devices/amd_iommu. We don't
695 * duplicate here.
696 */
697 perf_attr_iommu[0].data = perf->max_concurrent;
698 for (i = 0; i < num_attrs; i++)
699 attrs[i] = &perf_attr_iommu[i].attr.attr;
700 }
701 perf->attr_group->name = perf->block_name;
702 perf->attr_group->attrs = attrs;
703 ret = sysfs_create_group(dev->kobj_perf, perf->attr_group);
704 if (ret < 0)
705 return ret;
706 }
707
708 return 0;
709}
710
711/* Called with write topology lock acquired */
712static int kfd_build_sysfs_node_tree(void)
713{
714 struct kfd_topology_device *dev;
715 int ret;
716 uint32_t i = 0;
717
718 list_for_each_entry(dev, &topology_device_list, list) {
719 ret = kfd_build_sysfs_node_entry(dev, i);
720 if (ret < 0)
721 return ret;
722 i++;
723 }
724
725 return 0;
726}
727
728/* Called with write topology lock acquired */
729static void kfd_remove_sysfs_node_tree(void)
730{
731 struct kfd_topology_device *dev;
732
733 list_for_each_entry(dev, &topology_device_list, list)
734 kfd_remove_sysfs_node_entry(dev);
735}
736
737static int kfd_topology_update_sysfs(void)
738{
739 int ret;
740
741 pr_info("Creating topology SYSFS entries\n");
742 if (!sys_props.kobj_topology) {
743 sys_props.kobj_topology =
744 kfd_alloc_struct(sys_props.kobj_topology);
745 if (!sys_props.kobj_topology)
746 return -ENOMEM;
747
748 ret = kobject_init_and_add(sys_props.kobj_topology,
749 &sysprops_type, &kfd_device->kobj,
750 "topology");
751 if (ret < 0)
752 return ret;
753
754 sys_props.kobj_nodes = kobject_create_and_add("nodes",
755 sys_props.kobj_topology);
756 if (!sys_props.kobj_nodes)
757 return -ENOMEM;
758
759 sys_props.attr_genid.name = "generation_id";
760 sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE;
761 sysfs_attr_init(&sys_props.attr_genid);
762 ret = sysfs_create_file(sys_props.kobj_topology,
763 &sys_props.attr_genid);
764 if (ret < 0)
765 return ret;
766
767 sys_props.attr_props.name = "system_properties";
768 sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE;
769 sysfs_attr_init(&sys_props.attr_props);
770 ret = sysfs_create_file(sys_props.kobj_topology,
771 &sys_props.attr_props);
772 if (ret < 0)
773 return ret;
774 }
775
776 kfd_remove_sysfs_node_tree();
777
778 return kfd_build_sysfs_node_tree();
779}
780
781static void kfd_topology_release_sysfs(void)
782{
783 kfd_remove_sysfs_node_tree();
784 if (sys_props.kobj_topology) {
785 sysfs_remove_file(sys_props.kobj_topology,
786 &sys_props.attr_genid);
787 sysfs_remove_file(sys_props.kobj_topology,
788 &sys_props.attr_props);
789 if (sys_props.kobj_nodes) {
790 kobject_del(sys_props.kobj_nodes);
791 kobject_put(sys_props.kobj_nodes);
792 sys_props.kobj_nodes = NULL;
793 }
794 kobject_del(sys_props.kobj_topology);
795 kobject_put(sys_props.kobj_topology);
796 sys_props.kobj_topology = NULL;
797 }
798}
799
800/* Called with write topology_lock acquired */
801static void kfd_topology_update_device_list(struct list_head *temp_list,
802 struct list_head *master_list)
803{
804 while (!list_empty(temp_list)) {
805 list_move_tail(temp_list->next, master_list);
806 sys_props.num_devices++;
807 }
808}
809
810static void kfd_debug_print_topology(void)
811{
812 struct kfd_topology_device *dev;
813
814 down_read(&topology_lock);
815
816 dev = list_last_entry(&topology_device_list,
817 struct kfd_topology_device, list);
818 if (dev) {
819 if (dev->node_props.cpu_cores_count &&
820 dev->node_props.simd_count) {
821 pr_info("Topology: Add APU node [0x%0x:0x%0x]\n",
822 dev->node_props.device_id,
823 dev->node_props.vendor_id);
824 } else if (dev->node_props.cpu_cores_count)
825 pr_info("Topology: Add CPU node\n");
826 else if (dev->node_props.simd_count)
827 pr_info("Topology: Add dGPU node [0x%0x:0x%0x]\n",
828 dev->node_props.device_id,
829 dev->node_props.vendor_id);
830 }
831 up_read(&topology_lock);
832}
833
834/* Helper function for intializing platform_xx members of
835 * kfd_system_properties. Uses OEM info from the last CPU/APU node.
836 */
837static void kfd_update_system_properties(void)
838{
839 struct kfd_topology_device *dev;
840
841 down_read(&topology_lock);
842 dev = list_last_entry(&topology_device_list,
843 struct kfd_topology_device, list);
844 if (dev) {
845 sys_props.platform_id =
846 (*((uint64_t *)dev->oem_id)) & CRAT_OEMID_64BIT_MASK;
847 sys_props.platform_oem = *((uint64_t *)dev->oem_table_id);
848 sys_props.platform_rev = dev->oem_revision;
849 }
850 up_read(&topology_lock);
851}
852
853static void find_system_memory(const struct dmi_header *dm,
854 void *private)
855{
856 struct kfd_mem_properties *mem;
857 u16 mem_width, mem_clock;
858 struct kfd_topology_device *kdev =
859 (struct kfd_topology_device *)private;
860 const u8 *dmi_data = (const u8 *)(dm + 1);
861
862 if (dm->type == DMI_ENTRY_MEM_DEVICE && dm->length >= 0x15) {
863 mem_width = (u16)(*(const u16 *)(dmi_data + 0x6));
864 mem_clock = (u16)(*(const u16 *)(dmi_data + 0x11));
865 list_for_each_entry(mem, &kdev->mem_props, list) {
866 if (mem_width != 0xFFFF && mem_width != 0)
867 mem->width = mem_width;
868 if (mem_clock != 0)
869 mem->mem_clk_max = mem_clock;
870 }
871 }
872}
873
874/*
875 * Performance counters information is not part of CRAT but we would like to
876 * put them in the sysfs under topology directory for Thunk to get the data.
877 * This function is called before updating the sysfs.
878 */
879static int kfd_add_perf_to_topology(struct kfd_topology_device *kdev)
880{
881 /* These are the only counters supported so far */
882 return kfd_iommu_add_perf_counters(kdev);
883}
884
885/* kfd_add_non_crat_information - Add information that is not currently
886 * defined in CRAT but is necessary for KFD topology
887 * @dev - topology device to which addition info is added
888 */
889static void kfd_add_non_crat_information(struct kfd_topology_device *kdev)
890{
891 /* Check if CPU only node. */
892 if (!kdev->gpu) {
893 /* Add system memory information */
894 dmi_walk(find_system_memory, kdev);
895 }
896 /* TODO: For GPU node, rearrange code from kfd_topology_add_device */
897}
898
899/* kfd_is_acpi_crat_invalid - CRAT from ACPI is valid only for AMD APU devices.
900 * Ignore CRAT for all other devices. AMD APU is identified if both CPU
901 * and GPU cores are present.
902 * @device_list - topology device list created by parsing ACPI CRAT table.
903 * @return - TRUE if invalid, FALSE is valid.
904 */
905static bool kfd_is_acpi_crat_invalid(struct list_head *device_list)
906{
907 struct kfd_topology_device *dev;
908
909 list_for_each_entry(dev, device_list, list) {
910 if (dev->node_props.cpu_cores_count &&
911 dev->node_props.simd_count)
912 return false;
913 }
914 pr_info("Ignoring ACPI CRAT on non-APU system\n");
915 return true;
916}
917
918int kfd_topology_init(void)
919{
920 void *crat_image = NULL;
921 size_t image_size = 0;
922 int ret;
923 struct list_head temp_topology_device_list;
924 int cpu_only_node = 0;
925 struct kfd_topology_device *kdev;
926 int proximity_domain;
927
928 /* topology_device_list - Master list of all topology devices
929 * temp_topology_device_list - temporary list created while parsing CRAT
930 * or VCRAT. Once parsing is complete the contents of list is moved to
931 * topology_device_list
932 */
933
934 /* Initialize the head for the both the lists */
935 INIT_LIST_HEAD(&topology_device_list);
936 INIT_LIST_HEAD(&temp_topology_device_list);
937 init_rwsem(&topology_lock);
938
939 memset(&sys_props, 0, sizeof(sys_props));
940
941 /* Proximity domains in ACPI CRAT tables start counting at
942 * 0. The same should be true for virtual CRAT tables created
943 * at this stage. GPUs added later in kfd_topology_add_device
944 * use a counter.
945 */
946 proximity_domain = 0;
947
948 /*
949 * Get the CRAT image from the ACPI. If ACPI doesn't have one
950 * or if ACPI CRAT is invalid create a virtual CRAT.
951 * NOTE: The current implementation expects all AMD APUs to have
952 * CRAT. If no CRAT is available, it is assumed to be a CPU
953 */
954 ret = kfd_create_crat_image_acpi(&crat_image, &image_size);
955 if (!ret) {
956 ret = kfd_parse_crat_table(crat_image,
957 &temp_topology_device_list,
958 proximity_domain);
959 if (ret ||
960 kfd_is_acpi_crat_invalid(&temp_topology_device_list)) {
961 kfd_release_topology_device_list(
962 &temp_topology_device_list);
963 kfd_destroy_crat_image(crat_image);
964 crat_image = NULL;
965 }
966 }
967
968 if (!crat_image) {
969 ret = kfd_create_crat_image_virtual(&crat_image, &image_size,
970 COMPUTE_UNIT_CPU, NULL,
971 proximity_domain);
972 cpu_only_node = 1;
973 if (ret) {
974 pr_err("Error creating VCRAT table for CPU\n");
975 return ret;
976 }
977
978 ret = kfd_parse_crat_table(crat_image,
979 &temp_topology_device_list,
980 proximity_domain);
981 if (ret) {
982 pr_err("Error parsing VCRAT table for CPU\n");
983 goto err;
984 }
985 }
986
987 kdev = list_first_entry(&temp_topology_device_list,
988 struct kfd_topology_device, list);
989 kfd_add_perf_to_topology(kdev);
990
991 down_write(&topology_lock);
992 kfd_topology_update_device_list(&temp_topology_device_list,
993 &topology_device_list);
994 atomic_set(&topology_crat_proximity_domain, sys_props.num_devices-1);
995 ret = kfd_topology_update_sysfs();
996 up_write(&topology_lock);
997
998 if (!ret) {
999 sys_props.generation_count++;
1000 kfd_update_system_properties();
1001 kfd_debug_print_topology();
1002 pr_info("Finished initializing topology\n");
1003 } else
1004 pr_err("Failed to update topology in sysfs ret=%d\n", ret);
1005
1006 /* For nodes with GPU, this information gets added
1007 * when GPU is detected (kfd_topology_add_device).
1008 */
1009 if (cpu_only_node) {
1010 /* Add additional information to CPU only node created above */
1011 down_write(&topology_lock);
1012 kdev = list_first_entry(&topology_device_list,
1013 struct kfd_topology_device, list);
1014 up_write(&topology_lock);
1015 kfd_add_non_crat_information(kdev);
1016 }
1017
1018err:
1019 kfd_destroy_crat_image(crat_image);
1020 return ret;
1021}
1022
1023void kfd_topology_shutdown(void)
1024{
1025 down_write(&topology_lock);
1026 kfd_topology_release_sysfs();
1027 kfd_release_live_view();
1028 up_write(&topology_lock);
1029}
1030
1031static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu)
1032{
1033 uint32_t hashout;
1034 uint32_t buf[7];
1035 uint64_t local_mem_size;
1036 int i;
1037 struct kfd_local_mem_info local_mem_info;
1038
1039 if (!gpu)
1040 return 0;
1041
1042 gpu->kfd2kgd->get_local_mem_info(gpu->kgd, &local_mem_info);
1043
1044 local_mem_size = local_mem_info.local_mem_size_private +
1045 local_mem_info.local_mem_size_public;
1046
1047 buf[0] = gpu->pdev->devfn;
1048 buf[1] = gpu->pdev->subsystem_vendor;
1049 buf[2] = gpu->pdev->subsystem_device;
1050 buf[3] = gpu->pdev->device;
1051 buf[4] = gpu->pdev->bus->number;
1052 buf[5] = lower_32_bits(local_mem_size);
1053 buf[6] = upper_32_bits(local_mem_size);
1054
1055 for (i = 0, hashout = 0; i < 7; i++)
1056 hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH);
1057
1058 return hashout;
1059}
1060/* kfd_assign_gpu - Attach @gpu to the correct kfd topology device. If
1061 * the GPU device is not already present in the topology device
1062 * list then return NULL. This means a new topology device has to
1063 * be created for this GPU.
1064 * TODO: Rather than assiging @gpu to first topology device withtout
1065 * gpu attached, it will better to have more stringent check.
1066 */
1067static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu)
1068{
1069 struct kfd_topology_device *dev;
1070 struct kfd_topology_device *out_dev = NULL;
1071
1072 down_write(&topology_lock);
1073 list_for_each_entry(dev, &topology_device_list, list)
1074 if (!dev->gpu && (dev->node_props.simd_count > 0)) {
1075 dev->gpu = gpu;
1076 out_dev = dev;
1077 break;
1078 }
1079 up_write(&topology_lock);
1080 return out_dev;
1081}
1082
1083static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival)
1084{
1085 /*
1086 * TODO: Generate an event for thunk about the arrival/removal
1087 * of the GPU
1088 */
1089}
1090
1091/* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info,
1092 * patch this after CRAT parsing.
1093 */
1094static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev)
1095{
1096 struct kfd_mem_properties *mem;
1097 struct kfd_local_mem_info local_mem_info;
1098
1099 if (!dev)
1100 return;
1101
1102 /* Currently, amdgpu driver (amdgpu_mc) deals only with GPUs with
1103 * single bank of VRAM local memory.
1104 * for dGPUs - VCRAT reports only one bank of Local Memory
1105 * for APUs - If CRAT from ACPI reports more than one bank, then
1106 * all the banks will report the same mem_clk_max information
1107 */
1108 dev->gpu->kfd2kgd->get_local_mem_info(dev->gpu->kgd,
1109 &local_mem_info);
1110
1111 list_for_each_entry(mem, &dev->mem_props, list)
1112 mem->mem_clk_max = local_mem_info.mem_clk_max;
1113}
1114
1115static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev)
1116{
1117 struct kfd_iolink_properties *link;
1118
1119 if (!dev || !dev->gpu)
1120 return;
1121
1122 /* GPU only creates direck links so apply flags setting to all */
1123 if (dev->gpu->device_info->asic_family == CHIP_HAWAII)
1124 list_for_each_entry(link, &dev->io_link_props, list)
1125 link->flags = CRAT_IOLINK_FLAGS_ENABLED |
1126 CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1127 CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1128}
1129
1130int kfd_topology_add_device(struct kfd_dev *gpu)
1131{
1132 uint32_t gpu_id;
1133 struct kfd_topology_device *dev;
1134 struct kfd_cu_info cu_info;
1135 int res = 0;
1136 struct list_head temp_topology_device_list;
1137 void *crat_image = NULL;
1138 size_t image_size = 0;
1139 int proximity_domain;
1140
1141 INIT_LIST_HEAD(&temp_topology_device_list);
1142
1143 gpu_id = kfd_generate_gpu_id(gpu);
1144
1145 pr_debug("Adding new GPU (ID: 0x%x) to topology\n", gpu_id);
1146
1147 proximity_domain = atomic_inc_return(&topology_crat_proximity_domain);
1148
1149 /* Check to see if this gpu device exists in the topology_device_list.
1150 * If so, assign the gpu to that device,
1151 * else create a Virtual CRAT for this gpu device and then parse that
1152 * CRAT to create a new topology device. Once created assign the gpu to
1153 * that topology device
1154 */
1155 dev = kfd_assign_gpu(gpu);
1156 if (!dev) {
1157 res = kfd_create_crat_image_virtual(&crat_image, &image_size,
1158 COMPUTE_UNIT_GPU, gpu,
1159 proximity_domain);
1160 if (res) {
1161 pr_err("Error creating VCRAT for GPU (ID: 0x%x)\n",
1162 gpu_id);
1163 return res;
1164 }
1165 res = kfd_parse_crat_table(crat_image,
1166 &temp_topology_device_list,
1167 proximity_domain);
1168 if (res) {
1169 pr_err("Error parsing VCRAT for GPU (ID: 0x%x)\n",
1170 gpu_id);
1171 goto err;
1172 }
1173
1174 down_write(&topology_lock);
1175 kfd_topology_update_device_list(&temp_topology_device_list,
1176 &topology_device_list);
1177
1178 /* Update the SYSFS tree, since we added another topology
1179 * device
1180 */
1181 res = kfd_topology_update_sysfs();
1182 up_write(&topology_lock);
1183
1184 if (!res)
1185 sys_props.generation_count++;
1186 else
1187 pr_err("Failed to update GPU (ID: 0x%x) to sysfs topology. res=%d\n",
1188 gpu_id, res);
1189 dev = kfd_assign_gpu(gpu);
1190 if (WARN_ON(!dev)) {
1191 res = -ENODEV;
1192 goto err;
1193 }
1194 }
1195
1196 dev->gpu_id = gpu_id;
1197 gpu->id = gpu_id;
1198
1199 /* TODO: Move the following lines to function
1200 * kfd_add_non_crat_information
1201 */
1202
1203 /* Fill-in additional information that is not available in CRAT but
1204 * needed for the topology
1205 */
1206
1207 dev->gpu->kfd2kgd->get_cu_info(dev->gpu->kgd, &cu_info);
1208 dev->node_props.simd_arrays_per_engine =
1209 cu_info.num_shader_arrays_per_engine;
1210
1211 dev->node_props.vendor_id = gpu->pdev->vendor;
1212 dev->node_props.device_id = gpu->pdev->device;
1213 dev->node_props.location_id = PCI_DEVID(gpu->pdev->bus->number,
1214 gpu->pdev->devfn);
1215 dev->node_props.max_engine_clk_fcompute =
1216 dev->gpu->kfd2kgd->get_max_engine_clock_in_mhz(dev->gpu->kgd);
1217 dev->node_props.max_engine_clk_ccompute =
1218 cpufreq_quick_get_max(0) / 1000;
1219 dev->node_props.drm_render_minor =
1220 gpu->shared_resources.drm_render_minor;
1221
1222 kfd_fill_mem_clk_max_info(dev);
1223 kfd_fill_iolink_non_crat_info(dev);
1224
1225 switch (dev->gpu->device_info->asic_family) {
1226 case CHIP_KAVERI:
1227 case CHIP_HAWAII:
1228 case CHIP_TONGA:
1229 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_PRE_1_0 <<
1230 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1231 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1232 break;
1233 case CHIP_CARRIZO:
1234 case CHIP_FIJI:
1235 case CHIP_POLARIS10:
1236 case CHIP_POLARIS11:
1237 pr_debug("Adding doorbell packet type capability\n");
1238 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 <<
1239 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1240 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1241 break;
1242 default:
1243 WARN(1, "Unexpected ASIC family %u",
1244 dev->gpu->device_info->asic_family);
1245 }
1246
1247 /* Fix errors in CZ CRAT.
1248 * simd_count: Carrizo CRAT reports wrong simd_count, probably
1249 * because it doesn't consider masked out CUs
1250 * max_waves_per_simd: Carrizo reports wrong max_waves_per_simd
1251 * capability flag: Carrizo CRAT doesn't report IOMMU flags
1252 */
1253 if (dev->gpu->device_info->asic_family == CHIP_CARRIZO) {
1254 dev->node_props.simd_count =
1255 cu_info.simd_per_cu * cu_info.cu_active_number;
1256 dev->node_props.max_waves_per_simd = 10;
1257 dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
1258 }
1259
1260 kfd_debug_print_topology();
1261
1262 if (!res)
1263 kfd_notify_gpu_change(gpu_id, 1);
1264err:
1265 kfd_destroy_crat_image(crat_image);
1266 return res;
1267}
1268
1269int kfd_topology_remove_device(struct kfd_dev *gpu)
1270{
1271 struct kfd_topology_device *dev, *tmp;
1272 uint32_t gpu_id;
1273 int res = -ENODEV;
1274
1275 down_write(&topology_lock);
1276
1277 list_for_each_entry_safe(dev, tmp, &topology_device_list, list)
1278 if (dev->gpu == gpu) {
1279 gpu_id = dev->gpu_id;
1280 kfd_remove_sysfs_node_entry(dev);
1281 kfd_release_topology_device(dev);
1282 sys_props.num_devices--;
1283 res = 0;
1284 if (kfd_topology_update_sysfs() < 0)
1285 kfd_topology_release_sysfs();
1286 break;
1287 }
1288
1289 up_write(&topology_lock);
1290
1291 if (!res)
1292 kfd_notify_gpu_change(gpu_id, 0);
1293
1294 return res;
1295}
1296
1297/* kfd_topology_enum_kfd_devices - Enumerate through all devices in KFD
1298 * topology. If GPU device is found @idx, then valid kfd_dev pointer is
1299 * returned through @kdev
1300 * Return - 0: On success (@kdev will be NULL for non GPU nodes)
1301 * -1: If end of list
1302 */
1303int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev)
1304{
1305
1306 struct kfd_topology_device *top_dev;
1307 uint8_t device_idx = 0;
1308
1309 *kdev = NULL;
1310 down_read(&topology_lock);
1311
1312 list_for_each_entry(top_dev, &topology_device_list, list) {
1313 if (device_idx == idx) {
1314 *kdev = top_dev->gpu;
1315 up_read(&topology_lock);
1316 return 0;
1317 }
1318
1319 device_idx++;
1320 }
1321
1322 up_read(&topology_lock);
1323
1324 return -1;
1325
1326}
1327
1328static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask)
1329{
1330 const struct cpuinfo_x86 *cpuinfo;
1331 int first_cpu_of_numa_node;
1332
1333 if (!cpumask || cpumask == cpu_none_mask)
1334 return -1;
1335 first_cpu_of_numa_node = cpumask_first(cpumask);
1336 if (first_cpu_of_numa_node >= nr_cpu_ids)
1337 return -1;
1338 cpuinfo = &cpu_data(first_cpu_of_numa_node);
1339
1340 return cpuinfo->apicid;
1341}
1342
1343/* kfd_numa_node_to_apic_id - Returns the APIC ID of the first logical processor
1344 * of the given NUMA node (numa_node_id)
1345 * Return -1 on failure
1346 */
1347int kfd_numa_node_to_apic_id(int numa_node_id)
1348{
1349 if (numa_node_id == -1) {
1350 pr_warn("Invalid NUMA Node. Use online CPU mask\n");
1351 return kfd_cpumask_to_apic_id(cpu_online_mask);
1352 }
1353 return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id));
1354}
1355
1356#if defined(CONFIG_DEBUG_FS)
1357
1358int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data)
1359{
1360 struct kfd_topology_device *dev;
1361 unsigned int i = 0;
1362 int r = 0;
1363
1364 down_read(&topology_lock);
1365
1366 list_for_each_entry(dev, &topology_device_list, list) {
1367 if (!dev->gpu) {
1368 i++;
1369 continue;
1370 }
1371
1372 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1373 r = dqm_debugfs_hqds(m, dev->gpu->dqm);
1374 if (r)
1375 break;
1376 }
1377
1378 up_read(&topology_lock);
1379
1380 return r;
1381}
1382
1383int kfd_debugfs_rls_by_device(struct seq_file *m, void *data)
1384{
1385 struct kfd_topology_device *dev;
1386 unsigned int i = 0;
1387 int r = 0;
1388
1389 down_read(&topology_lock);
1390
1391 list_for_each_entry(dev, &topology_device_list, list) {
1392 if (!dev->gpu) {
1393 i++;
1394 continue;
1395 }
1396
1397 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1398 r = pm_debugfs_runlist(m, &dev->gpu->dqm->packets);
1399 if (r)
1400 break;
1401 }
1402
1403 up_read(&topology_lock);
1404
1405 return r;
1406}
1407
1408#endif
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