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#include "amdgpu_amdkfd.h"
40
41/* topology_device_list - Master list of all topology devices */
42static struct list_head topology_device_list;
43static struct kfd_system_properties sys_props;
44
45static DECLARE_RWSEM(topology_lock);
46static atomic_t topology_crat_proximity_domain;
47
48struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
49 uint32_t proximity_domain)
50{
51 struct kfd_topology_device *top_dev;
52 struct kfd_topology_device *device = NULL;
53
54 down_read(&topology_lock);
55
56 list_for_each_entry(top_dev, &topology_device_list, list)
57 if (top_dev->proximity_domain == proximity_domain) {
58 device = top_dev;
59 break;
60 }
61
62 up_read(&topology_lock);
63
64 return device;
65}
66
67struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id)
68{
69 struct kfd_topology_device *top_dev = NULL;
70 struct kfd_topology_device *ret = NULL;
71
72 down_read(&topology_lock);
73
74 list_for_each_entry(top_dev, &topology_device_list, list)
75 if (top_dev->gpu_id == gpu_id) {
76 ret = top_dev;
77 break;
78 }
79
80 up_read(&topology_lock);
81
82 return ret;
83}
84
85struct kfd_dev *kfd_device_by_id(uint32_t gpu_id)
86{
87 struct kfd_topology_device *top_dev;
88
89 top_dev = kfd_topology_device_by_id(gpu_id);
90 if (!top_dev)
91 return NULL;
92
93 return top_dev->gpu;
94}
95
96struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev)
97{
98 struct kfd_topology_device *top_dev;
99 struct kfd_dev *device = NULL;
100
101 down_read(&topology_lock);
102
103 list_for_each_entry(top_dev, &topology_device_list, list)
104 if (top_dev->gpu && top_dev->gpu->pdev == pdev) {
105 device = top_dev->gpu;
106 break;
107 }
108
109 up_read(&topology_lock);
110
111 return device;
112}
113
114struct kfd_dev *kfd_device_by_kgd(const struct kgd_dev *kgd)
115{
116 struct kfd_topology_device *top_dev;
117 struct kfd_dev *device = NULL;
118
119 down_read(&topology_lock);
120
121 list_for_each_entry(top_dev, &topology_device_list, list)
122 if (top_dev->gpu && top_dev->gpu->kgd == kgd) {
123 device = top_dev->gpu;
124 break;
125 }
126
127 up_read(&topology_lock);
128
129 return device;
130}
131
132/* Called with write topology_lock acquired */
133static void kfd_release_topology_device(struct kfd_topology_device *dev)
134{
135 struct kfd_mem_properties *mem;
136 struct kfd_cache_properties *cache;
137 struct kfd_iolink_properties *iolink;
138 struct kfd_perf_properties *perf;
139
140 list_del(&dev->list);
141
142 while (dev->mem_props.next != &dev->mem_props) {
143 mem = container_of(dev->mem_props.next,
144 struct kfd_mem_properties, list);
145 list_del(&mem->list);
146 kfree(mem);
147 }
148
149 while (dev->cache_props.next != &dev->cache_props) {
150 cache = container_of(dev->cache_props.next,
151 struct kfd_cache_properties, list);
152 list_del(&cache->list);
153 kfree(cache);
154 }
155
156 while (dev->io_link_props.next != &dev->io_link_props) {
157 iolink = container_of(dev->io_link_props.next,
158 struct kfd_iolink_properties, list);
159 list_del(&iolink->list);
160 kfree(iolink);
161 }
162
163 while (dev->perf_props.next != &dev->perf_props) {
164 perf = container_of(dev->perf_props.next,
165 struct kfd_perf_properties, list);
166 list_del(&perf->list);
167 kfree(perf);
168 }
169
170 kfree(dev);
171}
172
173void kfd_release_topology_device_list(struct list_head *device_list)
174{
175 struct kfd_topology_device *dev;
176
177 while (!list_empty(device_list)) {
178 dev = list_first_entry(device_list,
179 struct kfd_topology_device, list);
180 kfd_release_topology_device(dev);
181 }
182}
183
184static void kfd_release_live_view(void)
185{
186 kfd_release_topology_device_list(&topology_device_list);
187 memset(&sys_props, 0, sizeof(sys_props));
188}
189
190struct kfd_topology_device *kfd_create_topology_device(
191 struct list_head *device_list)
192{
193 struct kfd_topology_device *dev;
194
195 dev = kfd_alloc_struct(dev);
196 if (!dev) {
197 pr_err("No memory to allocate a topology device");
198 return NULL;
199 }
200
201 INIT_LIST_HEAD(&dev->mem_props);
202 INIT_LIST_HEAD(&dev->cache_props);
203 INIT_LIST_HEAD(&dev->io_link_props);
204 INIT_LIST_HEAD(&dev->perf_props);
205
206 list_add_tail(&dev->list, device_list);
207
208 return dev;
209}
210
211
212#define sysfs_show_gen_prop(buffer, fmt, ...) \
213 snprintf(buffer, PAGE_SIZE, "%s"fmt, buffer, __VA_ARGS__)
214#define sysfs_show_32bit_prop(buffer, name, value) \
215 sysfs_show_gen_prop(buffer, "%s %u\n", name, value)
216#define sysfs_show_64bit_prop(buffer, name, value) \
217 sysfs_show_gen_prop(buffer, "%s %llu\n", name, value)
218#define sysfs_show_32bit_val(buffer, value) \
219 sysfs_show_gen_prop(buffer, "%u\n", value)
220#define sysfs_show_str_val(buffer, value) \
221 sysfs_show_gen_prop(buffer, "%s\n", value)
222
223static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr,
224 char *buffer)
225{
226 ssize_t ret;
227
228 /* Making sure that the buffer is an empty string */
229 buffer[0] = 0;
230
231 if (attr == &sys_props.attr_genid) {
232 ret = sysfs_show_32bit_val(buffer, sys_props.generation_count);
233 } else if (attr == &sys_props.attr_props) {
234 sysfs_show_64bit_prop(buffer, "platform_oem",
235 sys_props.platform_oem);
236 sysfs_show_64bit_prop(buffer, "platform_id",
237 sys_props.platform_id);
238 ret = sysfs_show_64bit_prop(buffer, "platform_rev",
239 sys_props.platform_rev);
240 } else {
241 ret = -EINVAL;
242 }
243
244 return ret;
245}
246
247static void kfd_topology_kobj_release(struct kobject *kobj)
248{
249 kfree(kobj);
250}
251
252static const struct sysfs_ops sysprops_ops = {
253 .show = sysprops_show,
254};
255
256static struct kobj_type sysprops_type = {
257 .release = kfd_topology_kobj_release,
258 .sysfs_ops = &sysprops_ops,
259};
260
261static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr,
262 char *buffer)
263{
264 ssize_t ret;
265 struct kfd_iolink_properties *iolink;
266
267 /* Making sure that the buffer is an empty string */
268 buffer[0] = 0;
269
270 iolink = container_of(attr, struct kfd_iolink_properties, attr);
271 sysfs_show_32bit_prop(buffer, "type", iolink->iolink_type);
272 sysfs_show_32bit_prop(buffer, "version_major", iolink->ver_maj);
273 sysfs_show_32bit_prop(buffer, "version_minor", iolink->ver_min);
274 sysfs_show_32bit_prop(buffer, "node_from", iolink->node_from);
275 sysfs_show_32bit_prop(buffer, "node_to", iolink->node_to);
276 sysfs_show_32bit_prop(buffer, "weight", iolink->weight);
277 sysfs_show_32bit_prop(buffer, "min_latency", iolink->min_latency);
278 sysfs_show_32bit_prop(buffer, "max_latency", iolink->max_latency);
279 sysfs_show_32bit_prop(buffer, "min_bandwidth", iolink->min_bandwidth);
280 sysfs_show_32bit_prop(buffer, "max_bandwidth", iolink->max_bandwidth);
281 sysfs_show_32bit_prop(buffer, "recommended_transfer_size",
282 iolink->rec_transfer_size);
283 ret = sysfs_show_32bit_prop(buffer, "flags", iolink->flags);
284
285 return ret;
286}
287
288static const struct sysfs_ops iolink_ops = {
289 .show = iolink_show,
290};
291
292static struct kobj_type iolink_type = {
293 .release = kfd_topology_kobj_release,
294 .sysfs_ops = &iolink_ops,
295};
296
297static ssize_t mem_show(struct kobject *kobj, struct attribute *attr,
298 char *buffer)
299{
300 ssize_t ret;
301 struct kfd_mem_properties *mem;
302
303 /* Making sure that the buffer is an empty string */
304 buffer[0] = 0;
305
306 mem = container_of(attr, struct kfd_mem_properties, attr);
307 sysfs_show_32bit_prop(buffer, "heap_type", mem->heap_type);
308 sysfs_show_64bit_prop(buffer, "size_in_bytes", mem->size_in_bytes);
309 sysfs_show_32bit_prop(buffer, "flags", mem->flags);
310 sysfs_show_32bit_prop(buffer, "width", mem->width);
311 ret = sysfs_show_32bit_prop(buffer, "mem_clk_max", mem->mem_clk_max);
312
313 return ret;
314}
315
316static const struct sysfs_ops mem_ops = {
317 .show = mem_show,
318};
319
320static struct kobj_type mem_type = {
321 .release = kfd_topology_kobj_release,
322 .sysfs_ops = &mem_ops,
323};
324
325static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr,
326 char *buffer)
327{
328 ssize_t ret;
329 uint32_t i, j;
330 struct kfd_cache_properties *cache;
331
332 /* Making sure that the buffer is an empty string */
333 buffer[0] = 0;
334
335 cache = container_of(attr, struct kfd_cache_properties, attr);
336 sysfs_show_32bit_prop(buffer, "processor_id_low",
337 cache->processor_id_low);
338 sysfs_show_32bit_prop(buffer, "level", cache->cache_level);
339 sysfs_show_32bit_prop(buffer, "size", cache->cache_size);
340 sysfs_show_32bit_prop(buffer, "cache_line_size", cache->cacheline_size);
341 sysfs_show_32bit_prop(buffer, "cache_lines_per_tag",
342 cache->cachelines_per_tag);
343 sysfs_show_32bit_prop(buffer, "association", cache->cache_assoc);
344 sysfs_show_32bit_prop(buffer, "latency", cache->cache_latency);
345 sysfs_show_32bit_prop(buffer, "type", cache->cache_type);
346 snprintf(buffer, PAGE_SIZE, "%ssibling_map ", buffer);
347 for (i = 0; i < CRAT_SIBLINGMAP_SIZE; i++)
348 for (j = 0; j < sizeof(cache->sibling_map[0])*8; j++) {
349 /* Check each bit */
350 if (cache->sibling_map[i] & (1 << j))
351 ret = snprintf(buffer, PAGE_SIZE,
352 "%s%d%s", buffer, 1, ",");
353 else
354 ret = snprintf(buffer, PAGE_SIZE,
355 "%s%d%s", buffer, 0, ",");
356 }
357 /* Replace the last "," with end of line */
358 *(buffer + strlen(buffer) - 1) = 0xA;
359 return ret;
360}
361
362static const struct sysfs_ops cache_ops = {
363 .show = kfd_cache_show,
364};
365
366static struct kobj_type cache_type = {
367 .release = kfd_topology_kobj_release,
368 .sysfs_ops = &cache_ops,
369};
370
371/****** Sysfs of Performance Counters ******/
372
373struct kfd_perf_attr {
374 struct kobj_attribute attr;
375 uint32_t data;
376};
377
378static ssize_t perf_show(struct kobject *kobj, struct kobj_attribute *attrs,
379 char *buf)
380{
381 struct kfd_perf_attr *attr;
382
383 buf[0] = 0;
384 attr = container_of(attrs, struct kfd_perf_attr, attr);
385 if (!attr->data) /* invalid data for PMC */
386 return 0;
387 else
388 return sysfs_show_32bit_val(buf, attr->data);
389}
390
391#define KFD_PERF_DESC(_name, _data) \
392{ \
393 .attr = __ATTR(_name, 0444, perf_show, NULL), \
394 .data = _data, \
395}
396
397static struct kfd_perf_attr perf_attr_iommu[] = {
398 KFD_PERF_DESC(max_concurrent, 0),
399 KFD_PERF_DESC(num_counters, 0),
400 KFD_PERF_DESC(counter_ids, 0),
401};
402/****************************************/
403
404static ssize_t node_show(struct kobject *kobj, struct attribute *attr,
405 char *buffer)
406{
407 struct kfd_topology_device *dev;
408 char public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE];
409 uint32_t i;
410 uint32_t log_max_watch_addr;
411
412 /* Making sure that the buffer is an empty string */
413 buffer[0] = 0;
414
415 if (strcmp(attr->name, "gpu_id") == 0) {
416 dev = container_of(attr, struct kfd_topology_device,
417 attr_gpuid);
418 return sysfs_show_32bit_val(buffer, dev->gpu_id);
419 }
420
421 if (strcmp(attr->name, "name") == 0) {
422 dev = container_of(attr, struct kfd_topology_device,
423 attr_name);
424 for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE; i++) {
425 public_name[i] =
426 (char)dev->node_props.marketing_name[i];
427 if (dev->node_props.marketing_name[i] == 0)
428 break;
429 }
430 public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1] = 0x0;
431 return sysfs_show_str_val(buffer, public_name);
432 }
433
434 dev = container_of(attr, struct kfd_topology_device,
435 attr_props);
436 sysfs_show_32bit_prop(buffer, "cpu_cores_count",
437 dev->node_props.cpu_cores_count);
438 sysfs_show_32bit_prop(buffer, "simd_count",
439 dev->node_props.simd_count);
440 sysfs_show_32bit_prop(buffer, "mem_banks_count",
441 dev->node_props.mem_banks_count);
442 sysfs_show_32bit_prop(buffer, "caches_count",
443 dev->node_props.caches_count);
444 sysfs_show_32bit_prop(buffer, "io_links_count",
445 dev->node_props.io_links_count);
446 sysfs_show_32bit_prop(buffer, "cpu_core_id_base",
447 dev->node_props.cpu_core_id_base);
448 sysfs_show_32bit_prop(buffer, "simd_id_base",
449 dev->node_props.simd_id_base);
450 sysfs_show_32bit_prop(buffer, "max_waves_per_simd",
451 dev->node_props.max_waves_per_simd);
452 sysfs_show_32bit_prop(buffer, "lds_size_in_kb",
453 dev->node_props.lds_size_in_kb);
454 sysfs_show_32bit_prop(buffer, "gds_size_in_kb",
455 dev->node_props.gds_size_in_kb);
456 sysfs_show_32bit_prop(buffer, "wave_front_size",
457 dev->node_props.wave_front_size);
458 sysfs_show_32bit_prop(buffer, "array_count",
459 dev->node_props.array_count);
460 sysfs_show_32bit_prop(buffer, "simd_arrays_per_engine",
461 dev->node_props.simd_arrays_per_engine);
462 sysfs_show_32bit_prop(buffer, "cu_per_simd_array",
463 dev->node_props.cu_per_simd_array);
464 sysfs_show_32bit_prop(buffer, "simd_per_cu",
465 dev->node_props.simd_per_cu);
466 sysfs_show_32bit_prop(buffer, "max_slots_scratch_cu",
467 dev->node_props.max_slots_scratch_cu);
468 sysfs_show_32bit_prop(buffer, "vendor_id",
469 dev->node_props.vendor_id);
470 sysfs_show_32bit_prop(buffer, "device_id",
471 dev->node_props.device_id);
472 sysfs_show_32bit_prop(buffer, "location_id",
473 dev->node_props.location_id);
474 sysfs_show_32bit_prop(buffer, "drm_render_minor",
475 dev->node_props.drm_render_minor);
476 sysfs_show_64bit_prop(buffer, "hive_id",
477 dev->node_props.hive_id);
478
479 if (dev->gpu) {
480 log_max_watch_addr =
481 __ilog2_u32(dev->gpu->device_info->num_of_watch_points);
482
483 if (log_max_watch_addr) {
484 dev->node_props.capability |=
485 HSA_CAP_WATCH_POINTS_SUPPORTED;
486
487 dev->node_props.capability |=
488 ((log_max_watch_addr <<
489 HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) &
490 HSA_CAP_WATCH_POINTS_TOTALBITS_MASK);
491 }
492
493 if (dev->gpu->device_info->asic_family == CHIP_TONGA)
494 dev->node_props.capability |=
495 HSA_CAP_AQL_QUEUE_DOUBLE_MAP;
496
497 sysfs_show_32bit_prop(buffer, "max_engine_clk_fcompute",
498 dev->node_props.max_engine_clk_fcompute);
499
500 sysfs_show_64bit_prop(buffer, "local_mem_size",
501 (unsigned long long int) 0);
502
503 sysfs_show_32bit_prop(buffer, "fw_version",
504 dev->gpu->mec_fw_version);
505 sysfs_show_32bit_prop(buffer, "capability",
506 dev->node_props.capability);
507 sysfs_show_32bit_prop(buffer, "sdma_fw_version",
508 dev->gpu->sdma_fw_version);
509 }
510
511 return sysfs_show_32bit_prop(buffer, "max_engine_clk_ccompute",
512 cpufreq_quick_get_max(0)/1000);
513}
514
515static const struct sysfs_ops node_ops = {
516 .show = node_show,
517};
518
519static struct kobj_type node_type = {
520 .release = kfd_topology_kobj_release,
521 .sysfs_ops = &node_ops,
522};
523
524static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr)
525{
526 sysfs_remove_file(kobj, attr);
527 kobject_del(kobj);
528 kobject_put(kobj);
529}
530
531static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev)
532{
533 struct kfd_iolink_properties *iolink;
534 struct kfd_cache_properties *cache;
535 struct kfd_mem_properties *mem;
536 struct kfd_perf_properties *perf;
537
538 if (dev->kobj_iolink) {
539 list_for_each_entry(iolink, &dev->io_link_props, list)
540 if (iolink->kobj) {
541 kfd_remove_sysfs_file(iolink->kobj,
542 &iolink->attr);
543 iolink->kobj = NULL;
544 }
545 kobject_del(dev->kobj_iolink);
546 kobject_put(dev->kobj_iolink);
547 dev->kobj_iolink = NULL;
548 }
549
550 if (dev->kobj_cache) {
551 list_for_each_entry(cache, &dev->cache_props, list)
552 if (cache->kobj) {
553 kfd_remove_sysfs_file(cache->kobj,
554 &cache->attr);
555 cache->kobj = NULL;
556 }
557 kobject_del(dev->kobj_cache);
558 kobject_put(dev->kobj_cache);
559 dev->kobj_cache = NULL;
560 }
561
562 if (dev->kobj_mem) {
563 list_for_each_entry(mem, &dev->mem_props, list)
564 if (mem->kobj) {
565 kfd_remove_sysfs_file(mem->kobj, &mem->attr);
566 mem->kobj = NULL;
567 }
568 kobject_del(dev->kobj_mem);
569 kobject_put(dev->kobj_mem);
570 dev->kobj_mem = NULL;
571 }
572
573 if (dev->kobj_perf) {
574 list_for_each_entry(perf, &dev->perf_props, list) {
575 kfree(perf->attr_group);
576 perf->attr_group = NULL;
577 }
578 kobject_del(dev->kobj_perf);
579 kobject_put(dev->kobj_perf);
580 dev->kobj_perf = NULL;
581 }
582
583 if (dev->kobj_node) {
584 sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid);
585 sysfs_remove_file(dev->kobj_node, &dev->attr_name);
586 sysfs_remove_file(dev->kobj_node, &dev->attr_props);
587 kobject_del(dev->kobj_node);
588 kobject_put(dev->kobj_node);
589 dev->kobj_node = NULL;
590 }
591}
592
593static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev,
594 uint32_t id)
595{
596 struct kfd_iolink_properties *iolink;
597 struct kfd_cache_properties *cache;
598 struct kfd_mem_properties *mem;
599 struct kfd_perf_properties *perf;
600 int ret;
601 uint32_t i, num_attrs;
602 struct attribute **attrs;
603
604 if (WARN_ON(dev->kobj_node))
605 return -EEXIST;
606
607 /*
608 * Creating the sysfs folders
609 */
610 dev->kobj_node = kfd_alloc_struct(dev->kobj_node);
611 if (!dev->kobj_node)
612 return -ENOMEM;
613
614 ret = kobject_init_and_add(dev->kobj_node, &node_type,
615 sys_props.kobj_nodes, "%d", id);
616 if (ret < 0)
617 return ret;
618
619 dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node);
620 if (!dev->kobj_mem)
621 return -ENOMEM;
622
623 dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node);
624 if (!dev->kobj_cache)
625 return -ENOMEM;
626
627 dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node);
628 if (!dev->kobj_iolink)
629 return -ENOMEM;
630
631 dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node);
632 if (!dev->kobj_perf)
633 return -ENOMEM;
634
635 /*
636 * Creating sysfs files for node properties
637 */
638 dev->attr_gpuid.name = "gpu_id";
639 dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE;
640 sysfs_attr_init(&dev->attr_gpuid);
641 dev->attr_name.name = "name";
642 dev->attr_name.mode = KFD_SYSFS_FILE_MODE;
643 sysfs_attr_init(&dev->attr_name);
644 dev->attr_props.name = "properties";
645 dev->attr_props.mode = KFD_SYSFS_FILE_MODE;
646 sysfs_attr_init(&dev->attr_props);
647 ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid);
648 if (ret < 0)
649 return ret;
650 ret = sysfs_create_file(dev->kobj_node, &dev->attr_name);
651 if (ret < 0)
652 return ret;
653 ret = sysfs_create_file(dev->kobj_node, &dev->attr_props);
654 if (ret < 0)
655 return ret;
656
657 i = 0;
658 list_for_each_entry(mem, &dev->mem_props, list) {
659 mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
660 if (!mem->kobj)
661 return -ENOMEM;
662 ret = kobject_init_and_add(mem->kobj, &mem_type,
663 dev->kobj_mem, "%d", i);
664 if (ret < 0)
665 return ret;
666
667 mem->attr.name = "properties";
668 mem->attr.mode = KFD_SYSFS_FILE_MODE;
669 sysfs_attr_init(&mem->attr);
670 ret = sysfs_create_file(mem->kobj, &mem->attr);
671 if (ret < 0)
672 return ret;
673 i++;
674 }
675
676 i = 0;
677 list_for_each_entry(cache, &dev->cache_props, list) {
678 cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
679 if (!cache->kobj)
680 return -ENOMEM;
681 ret = kobject_init_and_add(cache->kobj, &cache_type,
682 dev->kobj_cache, "%d", i);
683 if (ret < 0)
684 return ret;
685
686 cache->attr.name = "properties";
687 cache->attr.mode = KFD_SYSFS_FILE_MODE;
688 sysfs_attr_init(&cache->attr);
689 ret = sysfs_create_file(cache->kobj, &cache->attr);
690 if (ret < 0)
691 return ret;
692 i++;
693 }
694
695 i = 0;
696 list_for_each_entry(iolink, &dev->io_link_props, list) {
697 iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
698 if (!iolink->kobj)
699 return -ENOMEM;
700 ret = kobject_init_and_add(iolink->kobj, &iolink_type,
701 dev->kobj_iolink, "%d", i);
702 if (ret < 0)
703 return ret;
704
705 iolink->attr.name = "properties";
706 iolink->attr.mode = KFD_SYSFS_FILE_MODE;
707 sysfs_attr_init(&iolink->attr);
708 ret = sysfs_create_file(iolink->kobj, &iolink->attr);
709 if (ret < 0)
710 return ret;
711 i++;
712 }
713
714 /* All hardware blocks have the same number of attributes. */
715 num_attrs = ARRAY_SIZE(perf_attr_iommu);
716 list_for_each_entry(perf, &dev->perf_props, list) {
717 perf->attr_group = kzalloc(sizeof(struct kfd_perf_attr)
718 * num_attrs + sizeof(struct attribute_group),
719 GFP_KERNEL);
720 if (!perf->attr_group)
721 return -ENOMEM;
722
723 attrs = (struct attribute **)(perf->attr_group + 1);
724 if (!strcmp(perf->block_name, "iommu")) {
725 /* Information of IOMMU's num_counters and counter_ids is shown
726 * under /sys/bus/event_source/devices/amd_iommu. We don't
727 * duplicate here.
728 */
729 perf_attr_iommu[0].data = perf->max_concurrent;
730 for (i = 0; i < num_attrs; i++)
731 attrs[i] = &perf_attr_iommu[i].attr.attr;
732 }
733 perf->attr_group->name = perf->block_name;
734 perf->attr_group->attrs = attrs;
735 ret = sysfs_create_group(dev->kobj_perf, perf->attr_group);
736 if (ret < 0)
737 return ret;
738 }
739
740 return 0;
741}
742
743/* Called with write topology lock acquired */
744static int kfd_build_sysfs_node_tree(void)
745{
746 struct kfd_topology_device *dev;
747 int ret;
748 uint32_t i = 0;
749
750 list_for_each_entry(dev, &topology_device_list, list) {
751 ret = kfd_build_sysfs_node_entry(dev, i);
752 if (ret < 0)
753 return ret;
754 i++;
755 }
756
757 return 0;
758}
759
760/* Called with write topology lock acquired */
761static void kfd_remove_sysfs_node_tree(void)
762{
763 struct kfd_topology_device *dev;
764
765 list_for_each_entry(dev, &topology_device_list, list)
766 kfd_remove_sysfs_node_entry(dev);
767}
768
769static int kfd_topology_update_sysfs(void)
770{
771 int ret;
772
773 pr_info("Creating topology SYSFS entries\n");
774 if (!sys_props.kobj_topology) {
775 sys_props.kobj_topology =
776 kfd_alloc_struct(sys_props.kobj_topology);
777 if (!sys_props.kobj_topology)
778 return -ENOMEM;
779
780 ret = kobject_init_and_add(sys_props.kobj_topology,
781 &sysprops_type, &kfd_device->kobj,
782 "topology");
783 if (ret < 0)
784 return ret;
785
786 sys_props.kobj_nodes = kobject_create_and_add("nodes",
787 sys_props.kobj_topology);
788 if (!sys_props.kobj_nodes)
789 return -ENOMEM;
790
791 sys_props.attr_genid.name = "generation_id";
792 sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE;
793 sysfs_attr_init(&sys_props.attr_genid);
794 ret = sysfs_create_file(sys_props.kobj_topology,
795 &sys_props.attr_genid);
796 if (ret < 0)
797 return ret;
798
799 sys_props.attr_props.name = "system_properties";
800 sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE;
801 sysfs_attr_init(&sys_props.attr_props);
802 ret = sysfs_create_file(sys_props.kobj_topology,
803 &sys_props.attr_props);
804 if (ret < 0)
805 return ret;
806 }
807
808 kfd_remove_sysfs_node_tree();
809
810 return kfd_build_sysfs_node_tree();
811}
812
813static void kfd_topology_release_sysfs(void)
814{
815 kfd_remove_sysfs_node_tree();
816 if (sys_props.kobj_topology) {
817 sysfs_remove_file(sys_props.kobj_topology,
818 &sys_props.attr_genid);
819 sysfs_remove_file(sys_props.kobj_topology,
820 &sys_props.attr_props);
821 if (sys_props.kobj_nodes) {
822 kobject_del(sys_props.kobj_nodes);
823 kobject_put(sys_props.kobj_nodes);
824 sys_props.kobj_nodes = NULL;
825 }
826 kobject_del(sys_props.kobj_topology);
827 kobject_put(sys_props.kobj_topology);
828 sys_props.kobj_topology = NULL;
829 }
830}
831
832/* Called with write topology_lock acquired */
833static void kfd_topology_update_device_list(struct list_head *temp_list,
834 struct list_head *master_list)
835{
836 while (!list_empty(temp_list)) {
837 list_move_tail(temp_list->next, master_list);
838 sys_props.num_devices++;
839 }
840}
841
842static void kfd_debug_print_topology(void)
843{
844 struct kfd_topology_device *dev;
845
846 down_read(&topology_lock);
847
848 dev = list_last_entry(&topology_device_list,
849 struct kfd_topology_device, list);
850 if (dev) {
851 if (dev->node_props.cpu_cores_count &&
852 dev->node_props.simd_count) {
853 pr_info("Topology: Add APU node [0x%0x:0x%0x]\n",
854 dev->node_props.device_id,
855 dev->node_props.vendor_id);
856 } else if (dev->node_props.cpu_cores_count)
857 pr_info("Topology: Add CPU node\n");
858 else if (dev->node_props.simd_count)
859 pr_info("Topology: Add dGPU node [0x%0x:0x%0x]\n",
860 dev->node_props.device_id,
861 dev->node_props.vendor_id);
862 }
863 up_read(&topology_lock);
864}
865
866/* Helper function for intializing platform_xx members of
867 * kfd_system_properties. Uses OEM info from the last CPU/APU node.
868 */
869static void kfd_update_system_properties(void)
870{
871 struct kfd_topology_device *dev;
872
873 down_read(&topology_lock);
874 dev = list_last_entry(&topology_device_list,
875 struct kfd_topology_device, list);
876 if (dev) {
877 sys_props.platform_id =
878 (*((uint64_t *)dev->oem_id)) & CRAT_OEMID_64BIT_MASK;
879 sys_props.platform_oem = *((uint64_t *)dev->oem_table_id);
880 sys_props.platform_rev = dev->oem_revision;
881 }
882 up_read(&topology_lock);
883}
884
885static void find_system_memory(const struct dmi_header *dm,
886 void *private)
887{
888 struct kfd_mem_properties *mem;
889 u16 mem_width, mem_clock;
890 struct kfd_topology_device *kdev =
891 (struct kfd_topology_device *)private;
892 const u8 *dmi_data = (const u8 *)(dm + 1);
893
894 if (dm->type == DMI_ENTRY_MEM_DEVICE && dm->length >= 0x15) {
895 mem_width = (u16)(*(const u16 *)(dmi_data + 0x6));
896 mem_clock = (u16)(*(const u16 *)(dmi_data + 0x11));
897 list_for_each_entry(mem, &kdev->mem_props, list) {
898 if (mem_width != 0xFFFF && mem_width != 0)
899 mem->width = mem_width;
900 if (mem_clock != 0)
901 mem->mem_clk_max = mem_clock;
902 }
903 }
904}
905
906/*
907 * Performance counters information is not part of CRAT but we would like to
908 * put them in the sysfs under topology directory for Thunk to get the data.
909 * This function is called before updating the sysfs.
910 */
911static int kfd_add_perf_to_topology(struct kfd_topology_device *kdev)
912{
913 /* These are the only counters supported so far */
914 return kfd_iommu_add_perf_counters(kdev);
915}
916
917/* kfd_add_non_crat_information - Add information that is not currently
918 * defined in CRAT but is necessary for KFD topology
919 * @dev - topology device to which addition info is added
920 */
921static void kfd_add_non_crat_information(struct kfd_topology_device *kdev)
922{
923 /* Check if CPU only node. */
924 if (!kdev->gpu) {
925 /* Add system memory information */
926 dmi_walk(find_system_memory, kdev);
927 }
928 /* TODO: For GPU node, rearrange code from kfd_topology_add_device */
929}
930
931/* kfd_is_acpi_crat_invalid - CRAT from ACPI is valid only for AMD APU devices.
932 * Ignore CRAT for all other devices. AMD APU is identified if both CPU
933 * and GPU cores are present.
934 * @device_list - topology device list created by parsing ACPI CRAT table.
935 * @return - TRUE if invalid, FALSE is valid.
936 */
937static bool kfd_is_acpi_crat_invalid(struct list_head *device_list)
938{
939 struct kfd_topology_device *dev;
940
941 list_for_each_entry(dev, device_list, list) {
942 if (dev->node_props.cpu_cores_count &&
943 dev->node_props.simd_count)
944 return false;
945 }
946 pr_info("Ignoring ACPI CRAT on non-APU system\n");
947 return true;
948}
949
950int kfd_topology_init(void)
951{
952 void *crat_image = NULL;
953 size_t image_size = 0;
954 int ret;
955 struct list_head temp_topology_device_list;
956 int cpu_only_node = 0;
957 struct kfd_topology_device *kdev;
958 int proximity_domain;
959
960 /* topology_device_list - Master list of all topology devices
961 * temp_topology_device_list - temporary list created while parsing CRAT
962 * or VCRAT. Once parsing is complete the contents of list is moved to
963 * topology_device_list
964 */
965
966 /* Initialize the head for the both the lists */
967 INIT_LIST_HEAD(&topology_device_list);
968 INIT_LIST_HEAD(&temp_topology_device_list);
969 init_rwsem(&topology_lock);
970
971 memset(&sys_props, 0, sizeof(sys_props));
972
973 /* Proximity domains in ACPI CRAT tables start counting at
974 * 0. The same should be true for virtual CRAT tables created
975 * at this stage. GPUs added later in kfd_topology_add_device
976 * use a counter.
977 */
978 proximity_domain = 0;
979
980 /*
981 * Get the CRAT image from the ACPI. If ACPI doesn't have one
982 * or if ACPI CRAT is invalid create a virtual CRAT.
983 * NOTE: The current implementation expects all AMD APUs to have
984 * CRAT. If no CRAT is available, it is assumed to be a CPU
985 */
986 ret = kfd_create_crat_image_acpi(&crat_image, &image_size);
987 if (!ret) {
988 ret = kfd_parse_crat_table(crat_image,
989 &temp_topology_device_list,
990 proximity_domain);
991 if (ret ||
992 kfd_is_acpi_crat_invalid(&temp_topology_device_list)) {
993 kfd_release_topology_device_list(
994 &temp_topology_device_list);
995 kfd_destroy_crat_image(crat_image);
996 crat_image = NULL;
997 }
998 }
999
1000 if (!crat_image) {
1001 ret = kfd_create_crat_image_virtual(&crat_image, &image_size,
1002 COMPUTE_UNIT_CPU, NULL,
1003 proximity_domain);
1004 cpu_only_node = 1;
1005 if (ret) {
1006 pr_err("Error creating VCRAT table for CPU\n");
1007 return ret;
1008 }
1009
1010 ret = kfd_parse_crat_table(crat_image,
1011 &temp_topology_device_list,
1012 proximity_domain);
1013 if (ret) {
1014 pr_err("Error parsing VCRAT table for CPU\n");
1015 goto err;
1016 }
1017 }
1018
1019 kdev = list_first_entry(&temp_topology_device_list,
1020 struct kfd_topology_device, list);
1021 kfd_add_perf_to_topology(kdev);
1022
1023 down_write(&topology_lock);
1024 kfd_topology_update_device_list(&temp_topology_device_list,
1025 &topology_device_list);
1026 atomic_set(&topology_crat_proximity_domain, sys_props.num_devices-1);
1027 ret = kfd_topology_update_sysfs();
1028 up_write(&topology_lock);
1029
1030 if (!ret) {
1031 sys_props.generation_count++;
1032 kfd_update_system_properties();
1033 kfd_debug_print_topology();
1034 pr_info("Finished initializing topology\n");
1035 } else
1036 pr_err("Failed to update topology in sysfs ret=%d\n", ret);
1037
1038 /* For nodes with GPU, this information gets added
1039 * when GPU is detected (kfd_topology_add_device).
1040 */
1041 if (cpu_only_node) {
1042 /* Add additional information to CPU only node created above */
1043 down_write(&topology_lock);
1044 kdev = list_first_entry(&topology_device_list,
1045 struct kfd_topology_device, list);
1046 up_write(&topology_lock);
1047 kfd_add_non_crat_information(kdev);
1048 }
1049
1050err:
1051 kfd_destroy_crat_image(crat_image);
1052 return ret;
1053}
1054
1055void kfd_topology_shutdown(void)
1056{
1057 down_write(&topology_lock);
1058 kfd_topology_release_sysfs();
1059 kfd_release_live_view();
1060 up_write(&topology_lock);
1061}
1062
1063static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu)
1064{
1065 uint32_t hashout;
1066 uint32_t buf[7];
1067 uint64_t local_mem_size;
1068 int i;
1069 struct kfd_local_mem_info local_mem_info;
1070
1071 if (!gpu)
1072 return 0;
1073
1074 amdgpu_amdkfd_get_local_mem_info(gpu->kgd, &local_mem_info);
1075
1076 local_mem_size = local_mem_info.local_mem_size_private +
1077 local_mem_info.local_mem_size_public;
1078
1079 buf[0] = gpu->pdev->devfn;
1080 buf[1] = gpu->pdev->subsystem_vendor;
1081 buf[2] = gpu->pdev->subsystem_device;
1082 buf[3] = gpu->pdev->device;
1083 buf[4] = gpu->pdev->bus->number;
1084 buf[5] = lower_32_bits(local_mem_size);
1085 buf[6] = upper_32_bits(local_mem_size);
1086
1087 for (i = 0, hashout = 0; i < 7; i++)
1088 hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH);
1089
1090 return hashout;
1091}
1092/* kfd_assign_gpu - Attach @gpu to the correct kfd topology device. If
1093 * the GPU device is not already present in the topology device
1094 * list then return NULL. This means a new topology device has to
1095 * be created for this GPU.
1096 */
1097static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu)
1098{
1099 struct kfd_topology_device *dev;
1100 struct kfd_topology_device *out_dev = NULL;
1101
1102 down_write(&topology_lock);
1103 list_for_each_entry(dev, &topology_device_list, list) {
1104 /* Discrete GPUs need their own topology device list
1105 * entries. Don't assign them to CPU/APU nodes.
1106 */
1107 if (!gpu->device_info->needs_iommu_device &&
1108 dev->node_props.cpu_cores_count)
1109 continue;
1110
1111 if (!dev->gpu && (dev->node_props.simd_count > 0)) {
1112 dev->gpu = gpu;
1113 out_dev = dev;
1114 break;
1115 }
1116 }
1117 up_write(&topology_lock);
1118 return out_dev;
1119}
1120
1121static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival)
1122{
1123 /*
1124 * TODO: Generate an event for thunk about the arrival/removal
1125 * of the GPU
1126 */
1127}
1128
1129/* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info,
1130 * patch this after CRAT parsing.
1131 */
1132static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev)
1133{
1134 struct kfd_mem_properties *mem;
1135 struct kfd_local_mem_info local_mem_info;
1136
1137 if (!dev)
1138 return;
1139
1140 /* Currently, amdgpu driver (amdgpu_mc) deals only with GPUs with
1141 * single bank of VRAM local memory.
1142 * for dGPUs - VCRAT reports only one bank of Local Memory
1143 * for APUs - If CRAT from ACPI reports more than one bank, then
1144 * all the banks will report the same mem_clk_max information
1145 */
1146 amdgpu_amdkfd_get_local_mem_info(dev->gpu->kgd, &local_mem_info);
1147
1148 list_for_each_entry(mem, &dev->mem_props, list)
1149 mem->mem_clk_max = local_mem_info.mem_clk_max;
1150}
1151
1152static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev)
1153{
1154 struct kfd_iolink_properties *link, *cpu_link;
1155 struct kfd_topology_device *cpu_dev;
1156 uint32_t cap;
1157 uint32_t cpu_flag = CRAT_IOLINK_FLAGS_ENABLED;
1158 uint32_t flag = CRAT_IOLINK_FLAGS_ENABLED;
1159
1160 if (!dev || !dev->gpu)
1161 return;
1162
1163 pcie_capability_read_dword(dev->gpu->pdev,
1164 PCI_EXP_DEVCAP2, &cap);
1165
1166 if (!(cap & (PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
1167 PCI_EXP_DEVCAP2_ATOMIC_COMP64)))
1168 cpu_flag |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1169 CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1170
1171 if (!dev->gpu->pci_atomic_requested ||
1172 dev->gpu->device_info->asic_family == CHIP_HAWAII)
1173 flag |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1174 CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1175
1176 /* GPU only creates direct links so apply flags setting to all */
1177 list_for_each_entry(link, &dev->io_link_props, list) {
1178 link->flags = flag;
1179 cpu_dev = kfd_topology_device_by_proximity_domain(
1180 link->node_to);
1181 if (cpu_dev) {
1182 list_for_each_entry(cpu_link,
1183 &cpu_dev->io_link_props, list)
1184 if (cpu_link->node_to == link->node_from)
1185 cpu_link->flags = cpu_flag;
1186 }
1187 }
1188}
1189
1190int kfd_topology_add_device(struct kfd_dev *gpu)
1191{
1192 uint32_t gpu_id;
1193 struct kfd_topology_device *dev;
1194 struct kfd_cu_info cu_info;
1195 int res = 0;
1196 struct list_head temp_topology_device_list;
1197 void *crat_image = NULL;
1198 size_t image_size = 0;
1199 int proximity_domain;
1200
1201 INIT_LIST_HEAD(&temp_topology_device_list);
1202
1203 gpu_id = kfd_generate_gpu_id(gpu);
1204
1205 pr_debug("Adding new GPU (ID: 0x%x) to topology\n", gpu_id);
1206
1207 proximity_domain = atomic_inc_return(&topology_crat_proximity_domain);
1208
1209 /* Check to see if this gpu device exists in the topology_device_list.
1210 * If so, assign the gpu to that device,
1211 * else create a Virtual CRAT for this gpu device and then parse that
1212 * CRAT to create a new topology device. Once created assign the gpu to
1213 * that topology device
1214 */
1215 dev = kfd_assign_gpu(gpu);
1216 if (!dev) {
1217 res = kfd_create_crat_image_virtual(&crat_image, &image_size,
1218 COMPUTE_UNIT_GPU, gpu,
1219 proximity_domain);
1220 if (res) {
1221 pr_err("Error creating VCRAT for GPU (ID: 0x%x)\n",
1222 gpu_id);
1223 return res;
1224 }
1225 res = kfd_parse_crat_table(crat_image,
1226 &temp_topology_device_list,
1227 proximity_domain);
1228 if (res) {
1229 pr_err("Error parsing VCRAT for GPU (ID: 0x%x)\n",
1230 gpu_id);
1231 goto err;
1232 }
1233
1234 down_write(&topology_lock);
1235 kfd_topology_update_device_list(&temp_topology_device_list,
1236 &topology_device_list);
1237
1238 /* Update the SYSFS tree, since we added another topology
1239 * device
1240 */
1241 res = kfd_topology_update_sysfs();
1242 up_write(&topology_lock);
1243
1244 if (!res)
1245 sys_props.generation_count++;
1246 else
1247 pr_err("Failed to update GPU (ID: 0x%x) to sysfs topology. res=%d\n",
1248 gpu_id, res);
1249 dev = kfd_assign_gpu(gpu);
1250 if (WARN_ON(!dev)) {
1251 res = -ENODEV;
1252 goto err;
1253 }
1254 }
1255
1256 dev->gpu_id = gpu_id;
1257 gpu->id = gpu_id;
1258
1259 /* TODO: Move the following lines to function
1260 * kfd_add_non_crat_information
1261 */
1262
1263 /* Fill-in additional information that is not available in CRAT but
1264 * needed for the topology
1265 */
1266
1267 amdgpu_amdkfd_get_cu_info(dev->gpu->kgd, &cu_info);
1268 dev->node_props.simd_arrays_per_engine =
1269 cu_info.num_shader_arrays_per_engine;
1270
1271 dev->node_props.vendor_id = gpu->pdev->vendor;
1272 dev->node_props.device_id = gpu->pdev->device;
1273 dev->node_props.location_id = PCI_DEVID(gpu->pdev->bus->number,
1274 gpu->pdev->devfn);
1275 dev->node_props.max_engine_clk_fcompute =
1276 amdgpu_amdkfd_get_max_engine_clock_in_mhz(dev->gpu->kgd);
1277 dev->node_props.max_engine_clk_ccompute =
1278 cpufreq_quick_get_max(0) / 1000;
1279 dev->node_props.drm_render_minor =
1280 gpu->shared_resources.drm_render_minor;
1281
1282 dev->node_props.hive_id = gpu->hive_id;
1283
1284 kfd_fill_mem_clk_max_info(dev);
1285 kfd_fill_iolink_non_crat_info(dev);
1286
1287 switch (dev->gpu->device_info->asic_family) {
1288 case CHIP_KAVERI:
1289 case CHIP_HAWAII:
1290 case CHIP_TONGA:
1291 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_PRE_1_0 <<
1292 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1293 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1294 break;
1295 case CHIP_CARRIZO:
1296 case CHIP_FIJI:
1297 case CHIP_POLARIS10:
1298 case CHIP_POLARIS11:
1299 case CHIP_POLARIS12:
1300 pr_debug("Adding doorbell packet type capability\n");
1301 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 <<
1302 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1303 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1304 break;
1305 case CHIP_VEGA10:
1306 case CHIP_VEGA12:
1307 case CHIP_VEGA20:
1308 case CHIP_RAVEN:
1309 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 <<
1310 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1311 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1312 break;
1313 default:
1314 WARN(1, "Unexpected ASIC family %u",
1315 dev->gpu->device_info->asic_family);
1316 }
1317
1318 /* Fix errors in CZ CRAT.
1319 * simd_count: Carrizo CRAT reports wrong simd_count, probably
1320 * because it doesn't consider masked out CUs
1321 * max_waves_per_simd: Carrizo reports wrong max_waves_per_simd
1322 * capability flag: Carrizo CRAT doesn't report IOMMU flags
1323 */
1324 if (dev->gpu->device_info->asic_family == CHIP_CARRIZO) {
1325 dev->node_props.simd_count =
1326 cu_info.simd_per_cu * cu_info.cu_active_number;
1327 dev->node_props.max_waves_per_simd = 10;
1328 dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
1329 }
1330
1331 kfd_debug_print_topology();
1332
1333 if (!res)
1334 kfd_notify_gpu_change(gpu_id, 1);
1335err:
1336 kfd_destroy_crat_image(crat_image);
1337 return res;
1338}
1339
1340int kfd_topology_remove_device(struct kfd_dev *gpu)
1341{
1342 struct kfd_topology_device *dev, *tmp;
1343 uint32_t gpu_id;
1344 int res = -ENODEV;
1345
1346 down_write(&topology_lock);
1347
1348 list_for_each_entry_safe(dev, tmp, &topology_device_list, list)
1349 if (dev->gpu == gpu) {
1350 gpu_id = dev->gpu_id;
1351 kfd_remove_sysfs_node_entry(dev);
1352 kfd_release_topology_device(dev);
1353 sys_props.num_devices--;
1354 res = 0;
1355 if (kfd_topology_update_sysfs() < 0)
1356 kfd_topology_release_sysfs();
1357 break;
1358 }
1359
1360 up_write(&topology_lock);
1361
1362 if (!res)
1363 kfd_notify_gpu_change(gpu_id, 0);
1364
1365 return res;
1366}
1367
1368/* kfd_topology_enum_kfd_devices - Enumerate through all devices in KFD
1369 * topology. If GPU device is found @idx, then valid kfd_dev pointer is
1370 * returned through @kdev
1371 * Return - 0: On success (@kdev will be NULL for non GPU nodes)
1372 * -1: If end of list
1373 */
1374int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev)
1375{
1376
1377 struct kfd_topology_device *top_dev;
1378 uint8_t device_idx = 0;
1379
1380 *kdev = NULL;
1381 down_read(&topology_lock);
1382
1383 list_for_each_entry(top_dev, &topology_device_list, list) {
1384 if (device_idx == idx) {
1385 *kdev = top_dev->gpu;
1386 up_read(&topology_lock);
1387 return 0;
1388 }
1389
1390 device_idx++;
1391 }
1392
1393 up_read(&topology_lock);
1394
1395 return -1;
1396
1397}
1398
1399static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask)
1400{
1401 int first_cpu_of_numa_node;
1402
1403 if (!cpumask || cpumask == cpu_none_mask)
1404 return -1;
1405 first_cpu_of_numa_node = cpumask_first(cpumask);
1406 if (first_cpu_of_numa_node >= nr_cpu_ids)
1407 return -1;
1408#ifdef CONFIG_X86_64
1409 return cpu_data(first_cpu_of_numa_node).apicid;
1410#else
1411 return first_cpu_of_numa_node;
1412#endif
1413}
1414
1415/* kfd_numa_node_to_apic_id - Returns the APIC ID of the first logical processor
1416 * of the given NUMA node (numa_node_id)
1417 * Return -1 on failure
1418 */
1419int kfd_numa_node_to_apic_id(int numa_node_id)
1420{
1421 if (numa_node_id == -1) {
1422 pr_warn("Invalid NUMA Node. Use online CPU mask\n");
1423 return kfd_cpumask_to_apic_id(cpu_online_mask);
1424 }
1425 return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id));
1426}
1427
1428#if defined(CONFIG_DEBUG_FS)
1429
1430int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data)
1431{
1432 struct kfd_topology_device *dev;
1433 unsigned int i = 0;
1434 int r = 0;
1435
1436 down_read(&topology_lock);
1437
1438 list_for_each_entry(dev, &topology_device_list, list) {
1439 if (!dev->gpu) {
1440 i++;
1441 continue;
1442 }
1443
1444 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1445 r = dqm_debugfs_hqds(m, dev->gpu->dqm);
1446 if (r)
1447 break;
1448 }
1449
1450 up_read(&topology_lock);
1451
1452 return r;
1453}
1454
1455int kfd_debugfs_rls_by_device(struct seq_file *m, void *data)
1456{
1457 struct kfd_topology_device *dev;
1458 unsigned int i = 0;
1459 int r = 0;
1460
1461 down_read(&topology_lock);
1462
1463 list_for_each_entry(dev, &topology_device_list, list) {
1464 if (!dev->gpu) {
1465 i++;
1466 continue;
1467 }
1468
1469 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1470 r = pm_debugfs_runlist(m, &dev->gpu->dqm->packets);
1471 if (r)
1472 break;
1473 }
1474
1475 up_read(&topology_lock);
1476
1477 return r;
1478}
1479
1480#endif
1481