xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_topology.c (revision 46e6acfe3501fa938af9c5bd730f0020235b08a2)
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 */
47 static struct list_head topology_device_list;
48 static struct kfd_system_properties sys_props;
49 
50 static DECLARE_RWSEM(topology_lock);
51 static uint32_t topology_crat_proximity_domain;
52 
53 struct 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 
68 struct 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 
82 struct 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 
100 struct 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 
111 struct 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 */
130 static 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 
178 void 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 
189 static 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 
195 struct 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 
230 static 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 
255 static void kfd_topology_kobj_release(struct kobject *kobj)
256 {
257 	kfree(kobj);
258 }
259 
260 static const struct sysfs_ops sysprops_ops = {
261 	.show = sysprops_show,
262 };
263 
264 static const struct kobj_type sysprops_type = {
265 	.release = kfd_topology_kobj_release,
266 	.sysfs_ops = &sysprops_ops,
267 };
268 
269 static 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, "flags", iolink->flags);
296 
297 	return offs;
298 }
299 
300 static const struct sysfs_ops iolink_ops = {
301 	.show = iolink_show,
302 };
303 
304 static const struct kobj_type iolink_type = {
305 	.release = kfd_topology_kobj_release,
306 	.sysfs_ops = &iolink_ops,
307 };
308 
309 static ssize_t mem_show(struct kobject *kobj, struct attribute *attr,
310 		char *buffer)
311 {
312 	int offs = 0;
313 	struct kfd_mem_properties *mem;
314 
315 	/* Making sure that the buffer is an empty string */
316 	buffer[0] = 0;
317 
318 	mem = container_of(attr, struct kfd_mem_properties, attr);
319 	if (mem->gpu && kfd_devcgroup_check_permission(mem->gpu))
320 		return -EPERM;
321 	sysfs_show_32bit_prop(buffer, offs, "heap_type", mem->heap_type);
322 	sysfs_show_64bit_prop(buffer, offs, "size_in_bytes",
323 			      mem->size_in_bytes);
324 	sysfs_show_32bit_prop(buffer, offs, "flags", mem->flags);
325 	sysfs_show_32bit_prop(buffer, offs, "width", mem->width);
326 	sysfs_show_32bit_prop(buffer, offs, "mem_clk_max",
327 			      mem->mem_clk_max);
328 
329 	return offs;
330 }
331 
332 static const struct sysfs_ops mem_ops = {
333 	.show = mem_show,
334 };
335 
336 static const struct kobj_type mem_type = {
337 	.release = kfd_topology_kobj_release,
338 	.sysfs_ops = &mem_ops,
339 };
340 
341 static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr,
342 		char *buffer)
343 {
344 	int offs = 0;
345 	uint32_t i, j;
346 	struct kfd_cache_properties *cache;
347 
348 	/* Making sure that the buffer is an empty string */
349 	buffer[0] = 0;
350 	cache = container_of(attr, struct kfd_cache_properties, attr);
351 	if (cache->gpu && kfd_devcgroup_check_permission(cache->gpu))
352 		return -EPERM;
353 	sysfs_show_32bit_prop(buffer, offs, "processor_id_low",
354 			cache->processor_id_low);
355 	sysfs_show_32bit_prop(buffer, offs, "level", cache->cache_level);
356 	sysfs_show_32bit_prop(buffer, offs, "size", cache->cache_size);
357 	sysfs_show_32bit_prop(buffer, offs, "cache_line_size",
358 			      cache->cacheline_size);
359 	sysfs_show_32bit_prop(buffer, offs, "cache_lines_per_tag",
360 			      cache->cachelines_per_tag);
361 	sysfs_show_32bit_prop(buffer, offs, "association", cache->cache_assoc);
362 	sysfs_show_32bit_prop(buffer, offs, "latency", cache->cache_latency);
363 	sysfs_show_32bit_prop(buffer, offs, "type", cache->cache_type);
364 
365 	offs += snprintf(buffer+offs, PAGE_SIZE-offs, "sibling_map ");
366 	for (i = 0; i < cache->sibling_map_size; i++)
367 		for (j = 0; j < sizeof(cache->sibling_map[0])*8; j++)
368 			/* Check each bit */
369 			offs += snprintf(buffer+offs, PAGE_SIZE-offs, "%d,",
370 						(cache->sibling_map[i] >> j) & 1);
371 
372 	/* Replace the last "," with end of line */
373 	buffer[offs-1] = '\n';
374 	return offs;
375 }
376 
377 static const struct sysfs_ops cache_ops = {
378 	.show = kfd_cache_show,
379 };
380 
381 static const struct kobj_type cache_type = {
382 	.release = kfd_topology_kobj_release,
383 	.sysfs_ops = &cache_ops,
384 };
385 
386 /****** Sysfs of Performance Counters ******/
387 
388 struct kfd_perf_attr {
389 	struct kobj_attribute attr;
390 	uint32_t data;
391 };
392 
393 static ssize_t perf_show(struct kobject *kobj, struct kobj_attribute *attrs,
394 			char *buf)
395 {
396 	int offs = 0;
397 	struct kfd_perf_attr *attr;
398 
399 	buf[0] = 0;
400 	attr = container_of(attrs, struct kfd_perf_attr, attr);
401 	if (!attr->data) /* invalid data for PMC */
402 		return 0;
403 	else
404 		return sysfs_show_32bit_val(buf, offs, attr->data);
405 }
406 
407 #define KFD_PERF_DESC(_name, _data)			\
408 {							\
409 	.attr  = __ATTR(_name, 0444, perf_show, NULL),	\
410 	.data = _data,					\
411 }
412 
413 static struct kfd_perf_attr perf_attr_iommu[] = {
414 	KFD_PERF_DESC(max_concurrent, 0),
415 	KFD_PERF_DESC(num_counters, 0),
416 	KFD_PERF_DESC(counter_ids, 0),
417 };
418 /****************************************/
419 
420 static ssize_t node_show(struct kobject *kobj, struct attribute *attr,
421 		char *buffer)
422 {
423 	int offs = 0;
424 	struct kfd_topology_device *dev;
425 	uint32_t log_max_watch_addr;
426 
427 	/* Making sure that the buffer is an empty string */
428 	buffer[0] = 0;
429 
430 	if (strcmp(attr->name, "gpu_id") == 0) {
431 		dev = container_of(attr, struct kfd_topology_device,
432 				attr_gpuid);
433 		if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu))
434 			return -EPERM;
435 		return sysfs_show_32bit_val(buffer, offs, dev->gpu_id);
436 	}
437 
438 	if (strcmp(attr->name, "name") == 0) {
439 		dev = container_of(attr, struct kfd_topology_device,
440 				attr_name);
441 
442 		if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu))
443 			return -EPERM;
444 		return sysfs_show_str_val(buffer, offs, dev->node_props.name);
445 	}
446 
447 	dev = container_of(attr, struct kfd_topology_device,
448 			attr_props);
449 	if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu))
450 		return -EPERM;
451 	sysfs_show_32bit_prop(buffer, offs, "cpu_cores_count",
452 			      dev->node_props.cpu_cores_count);
453 	sysfs_show_32bit_prop(buffer, offs, "simd_count",
454 			      dev->gpu ? dev->node_props.simd_count : 0);
455 	sysfs_show_32bit_prop(buffer, offs, "mem_banks_count",
456 			      dev->node_props.mem_banks_count);
457 	sysfs_show_32bit_prop(buffer, offs, "caches_count",
458 			      dev->node_props.caches_count);
459 	sysfs_show_32bit_prop(buffer, offs, "io_links_count",
460 			      dev->node_props.io_links_count);
461 	sysfs_show_32bit_prop(buffer, offs, "p2p_links_count",
462 			      dev->node_props.p2p_links_count);
463 	sysfs_show_32bit_prop(buffer, offs, "cpu_core_id_base",
464 			      dev->node_props.cpu_core_id_base);
465 	sysfs_show_32bit_prop(buffer, offs, "simd_id_base",
466 			      dev->node_props.simd_id_base);
467 	sysfs_show_32bit_prop(buffer, offs, "max_waves_per_simd",
468 			      dev->node_props.max_waves_per_simd);
469 	sysfs_show_32bit_prop(buffer, offs, "lds_size_in_kb",
470 			      dev->node_props.lds_size_in_kb);
471 	sysfs_show_32bit_prop(buffer, offs, "gds_size_in_kb",
472 			      dev->node_props.gds_size_in_kb);
473 	sysfs_show_32bit_prop(buffer, offs, "num_gws",
474 			      dev->node_props.num_gws);
475 	sysfs_show_32bit_prop(buffer, offs, "wave_front_size",
476 			      dev->node_props.wave_front_size);
477 	sysfs_show_32bit_prop(buffer, offs, "array_count",
478 			      dev->gpu ? (dev->node_props.array_count *
479 					  NUM_XCC(dev->gpu->xcc_mask)) : 0);
480 	sysfs_show_32bit_prop(buffer, offs, "simd_arrays_per_engine",
481 			      dev->node_props.simd_arrays_per_engine);
482 	sysfs_show_32bit_prop(buffer, offs, "cu_per_simd_array",
483 			      dev->node_props.cu_per_simd_array);
484 	sysfs_show_32bit_prop(buffer, offs, "simd_per_cu",
485 			      dev->node_props.simd_per_cu);
486 	sysfs_show_32bit_prop(buffer, offs, "max_slots_scratch_cu",
487 			      dev->node_props.max_slots_scratch_cu);
488 	sysfs_show_32bit_prop(buffer, offs, "gfx_target_version",
489 			      dev->node_props.gfx_target_version);
490 	sysfs_show_32bit_prop(buffer, offs, "vendor_id",
491 			      dev->node_props.vendor_id);
492 	sysfs_show_32bit_prop(buffer, offs, "device_id",
493 			      dev->node_props.device_id);
494 	sysfs_show_32bit_prop(buffer, offs, "location_id",
495 			      dev->node_props.location_id);
496 	sysfs_show_32bit_prop(buffer, offs, "domain",
497 			      dev->node_props.domain);
498 	sysfs_show_32bit_prop(buffer, offs, "drm_render_minor",
499 			      dev->node_props.drm_render_minor);
500 	sysfs_show_64bit_prop(buffer, offs, "hive_id",
501 			      dev->node_props.hive_id);
502 	sysfs_show_32bit_prop(buffer, offs, "num_sdma_engines",
503 			      dev->node_props.num_sdma_engines);
504 	sysfs_show_32bit_prop(buffer, offs, "num_sdma_xgmi_engines",
505 			      dev->node_props.num_sdma_xgmi_engines);
506 	sysfs_show_32bit_prop(buffer, offs, "num_sdma_queues_per_engine",
507 			      dev->node_props.num_sdma_queues_per_engine);
508 	sysfs_show_32bit_prop(buffer, offs, "num_cp_queues",
509 			      dev->node_props.num_cp_queues);
510 
511 	if (dev->gpu) {
512 		log_max_watch_addr =
513 			__ilog2_u32(dev->gpu->kfd->device_info.num_of_watch_points);
514 
515 		if (log_max_watch_addr) {
516 			dev->node_props.capability |=
517 					HSA_CAP_WATCH_POINTS_SUPPORTED;
518 
519 			dev->node_props.capability |=
520 				((log_max_watch_addr <<
521 					HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) &
522 				HSA_CAP_WATCH_POINTS_TOTALBITS_MASK);
523 		}
524 
525 		if (dev->gpu->adev->asic_type == CHIP_TONGA)
526 			dev->node_props.capability |=
527 					HSA_CAP_AQL_QUEUE_DOUBLE_MAP;
528 
529 		sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_fcompute",
530 			dev->node_props.max_engine_clk_fcompute);
531 
532 		sysfs_show_64bit_prop(buffer, offs, "local_mem_size", 0ULL);
533 
534 		sysfs_show_32bit_prop(buffer, offs, "fw_version",
535 				      dev->gpu->kfd->mec_fw_version);
536 		sysfs_show_32bit_prop(buffer, offs, "capability",
537 				      dev->node_props.capability);
538 		sysfs_show_64bit_prop(buffer, offs, "debug_prop",
539 				      dev->node_props.debug_prop);
540 		sysfs_show_32bit_prop(buffer, offs, "sdma_fw_version",
541 				      dev->gpu->kfd->sdma_fw_version);
542 		sysfs_show_64bit_prop(buffer, offs, "unique_id",
543 				      dev->gpu->adev->unique_id);
544 		sysfs_show_32bit_prop(buffer, offs, "num_xcc",
545 				      NUM_XCC(dev->gpu->xcc_mask));
546 	}
547 
548 	return sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_ccompute",
549 				     cpufreq_quick_get_max(0)/1000);
550 }
551 
552 static const struct sysfs_ops node_ops = {
553 	.show = node_show,
554 };
555 
556 static const struct kobj_type node_type = {
557 	.release = kfd_topology_kobj_release,
558 	.sysfs_ops = &node_ops,
559 };
560 
561 static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr)
562 {
563 	sysfs_remove_file(kobj, attr);
564 	kobject_del(kobj);
565 	kobject_put(kobj);
566 }
567 
568 static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev)
569 {
570 	struct kfd_iolink_properties *p2plink;
571 	struct kfd_iolink_properties *iolink;
572 	struct kfd_cache_properties *cache;
573 	struct kfd_mem_properties *mem;
574 	struct kfd_perf_properties *perf;
575 
576 	if (dev->kobj_iolink) {
577 		list_for_each_entry(iolink, &dev->io_link_props, list)
578 			if (iolink->kobj) {
579 				kfd_remove_sysfs_file(iolink->kobj,
580 							&iolink->attr);
581 				iolink->kobj = NULL;
582 			}
583 		kobject_del(dev->kobj_iolink);
584 		kobject_put(dev->kobj_iolink);
585 		dev->kobj_iolink = NULL;
586 	}
587 
588 	if (dev->kobj_p2plink) {
589 		list_for_each_entry(p2plink, &dev->p2p_link_props, list)
590 			if (p2plink->kobj) {
591 				kfd_remove_sysfs_file(p2plink->kobj,
592 							&p2plink->attr);
593 				p2plink->kobj = NULL;
594 			}
595 		kobject_del(dev->kobj_p2plink);
596 		kobject_put(dev->kobj_p2plink);
597 		dev->kobj_p2plink = NULL;
598 	}
599 
600 	if (dev->kobj_cache) {
601 		list_for_each_entry(cache, &dev->cache_props, list)
602 			if (cache->kobj) {
603 				kfd_remove_sysfs_file(cache->kobj,
604 							&cache->attr);
605 				cache->kobj = NULL;
606 			}
607 		kobject_del(dev->kobj_cache);
608 		kobject_put(dev->kobj_cache);
609 		dev->kobj_cache = NULL;
610 	}
611 
612 	if (dev->kobj_mem) {
613 		list_for_each_entry(mem, &dev->mem_props, list)
614 			if (mem->kobj) {
615 				kfd_remove_sysfs_file(mem->kobj, &mem->attr);
616 				mem->kobj = NULL;
617 			}
618 		kobject_del(dev->kobj_mem);
619 		kobject_put(dev->kobj_mem);
620 		dev->kobj_mem = NULL;
621 	}
622 
623 	if (dev->kobj_perf) {
624 		list_for_each_entry(perf, &dev->perf_props, list) {
625 			kfree(perf->attr_group);
626 			perf->attr_group = NULL;
627 		}
628 		kobject_del(dev->kobj_perf);
629 		kobject_put(dev->kobj_perf);
630 		dev->kobj_perf = NULL;
631 	}
632 
633 	if (dev->kobj_node) {
634 		sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid);
635 		sysfs_remove_file(dev->kobj_node, &dev->attr_name);
636 		sysfs_remove_file(dev->kobj_node, &dev->attr_props);
637 		kobject_del(dev->kobj_node);
638 		kobject_put(dev->kobj_node);
639 		dev->kobj_node = NULL;
640 	}
641 }
642 
643 static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev,
644 		uint32_t id)
645 {
646 	struct kfd_iolink_properties *p2plink;
647 	struct kfd_iolink_properties *iolink;
648 	struct kfd_cache_properties *cache;
649 	struct kfd_mem_properties *mem;
650 	struct kfd_perf_properties *perf;
651 	int ret;
652 	uint32_t i, num_attrs;
653 	struct attribute **attrs;
654 
655 	if (WARN_ON(dev->kobj_node))
656 		return -EEXIST;
657 
658 	/*
659 	 * Creating the sysfs folders
660 	 */
661 	dev->kobj_node = kfd_alloc_struct(dev->kobj_node);
662 	if (!dev->kobj_node)
663 		return -ENOMEM;
664 
665 	ret = kobject_init_and_add(dev->kobj_node, &node_type,
666 			sys_props.kobj_nodes, "%d", id);
667 	if (ret < 0) {
668 		kobject_put(dev->kobj_node);
669 		return ret;
670 	}
671 
672 	dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node);
673 	if (!dev->kobj_mem)
674 		return -ENOMEM;
675 
676 	dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node);
677 	if (!dev->kobj_cache)
678 		return -ENOMEM;
679 
680 	dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node);
681 	if (!dev->kobj_iolink)
682 		return -ENOMEM;
683 
684 	dev->kobj_p2plink = kobject_create_and_add("p2p_links", dev->kobj_node);
685 	if (!dev->kobj_p2plink)
686 		return -ENOMEM;
687 
688 	dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node);
689 	if (!dev->kobj_perf)
690 		return -ENOMEM;
691 
692 	/*
693 	 * Creating sysfs files for node properties
694 	 */
695 	dev->attr_gpuid.name = "gpu_id";
696 	dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE;
697 	sysfs_attr_init(&dev->attr_gpuid);
698 	dev->attr_name.name = "name";
699 	dev->attr_name.mode = KFD_SYSFS_FILE_MODE;
700 	sysfs_attr_init(&dev->attr_name);
701 	dev->attr_props.name = "properties";
702 	dev->attr_props.mode = KFD_SYSFS_FILE_MODE;
703 	sysfs_attr_init(&dev->attr_props);
704 	ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid);
705 	if (ret < 0)
706 		return ret;
707 	ret = sysfs_create_file(dev->kobj_node, &dev->attr_name);
708 	if (ret < 0)
709 		return ret;
710 	ret = sysfs_create_file(dev->kobj_node, &dev->attr_props);
711 	if (ret < 0)
712 		return ret;
713 
714 	i = 0;
715 	list_for_each_entry(mem, &dev->mem_props, list) {
716 		mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
717 		if (!mem->kobj)
718 			return -ENOMEM;
719 		ret = kobject_init_and_add(mem->kobj, &mem_type,
720 				dev->kobj_mem, "%d", i);
721 		if (ret < 0) {
722 			kobject_put(mem->kobj);
723 			return ret;
724 		}
725 
726 		mem->attr.name = "properties";
727 		mem->attr.mode = KFD_SYSFS_FILE_MODE;
728 		sysfs_attr_init(&mem->attr);
729 		ret = sysfs_create_file(mem->kobj, &mem->attr);
730 		if (ret < 0)
731 			return ret;
732 		i++;
733 	}
734 
735 	i = 0;
736 	list_for_each_entry(cache, &dev->cache_props, list) {
737 		cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
738 		if (!cache->kobj)
739 			return -ENOMEM;
740 		ret = kobject_init_and_add(cache->kobj, &cache_type,
741 				dev->kobj_cache, "%d", i);
742 		if (ret < 0) {
743 			kobject_put(cache->kobj);
744 			return ret;
745 		}
746 
747 		cache->attr.name = "properties";
748 		cache->attr.mode = KFD_SYSFS_FILE_MODE;
749 		sysfs_attr_init(&cache->attr);
750 		ret = sysfs_create_file(cache->kobj, &cache->attr);
751 		if (ret < 0)
752 			return ret;
753 		i++;
754 	}
755 
756 	i = 0;
757 	list_for_each_entry(iolink, &dev->io_link_props, list) {
758 		iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
759 		if (!iolink->kobj)
760 			return -ENOMEM;
761 		ret = kobject_init_and_add(iolink->kobj, &iolink_type,
762 				dev->kobj_iolink, "%d", i);
763 		if (ret < 0) {
764 			kobject_put(iolink->kobj);
765 			return ret;
766 		}
767 
768 		iolink->attr.name = "properties";
769 		iolink->attr.mode = KFD_SYSFS_FILE_MODE;
770 		sysfs_attr_init(&iolink->attr);
771 		ret = sysfs_create_file(iolink->kobj, &iolink->attr);
772 		if (ret < 0)
773 			return ret;
774 		i++;
775 	}
776 
777 	i = 0;
778 	list_for_each_entry(p2plink, &dev->p2p_link_props, list) {
779 		p2plink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
780 		if (!p2plink->kobj)
781 			return -ENOMEM;
782 		ret = kobject_init_and_add(p2plink->kobj, &iolink_type,
783 				dev->kobj_p2plink, "%d", i);
784 		if (ret < 0) {
785 			kobject_put(p2plink->kobj);
786 			return ret;
787 		}
788 
789 		p2plink->attr.name = "properties";
790 		p2plink->attr.mode = KFD_SYSFS_FILE_MODE;
791 		sysfs_attr_init(&p2plink->attr);
792 		ret = sysfs_create_file(p2plink->kobj, &p2plink->attr);
793 		if (ret < 0)
794 			return ret;
795 		i++;
796 	}
797 
798 	/* All hardware blocks have the same number of attributes. */
799 	num_attrs = ARRAY_SIZE(perf_attr_iommu);
800 	list_for_each_entry(perf, &dev->perf_props, list) {
801 		perf->attr_group = kzalloc(sizeof(struct kfd_perf_attr)
802 			* num_attrs + sizeof(struct attribute_group),
803 			GFP_KERNEL);
804 		if (!perf->attr_group)
805 			return -ENOMEM;
806 
807 		attrs = (struct attribute **)(perf->attr_group + 1);
808 		if (!strcmp(perf->block_name, "iommu")) {
809 		/* Information of IOMMU's num_counters and counter_ids is shown
810 		 * under /sys/bus/event_source/devices/amd_iommu. We don't
811 		 * duplicate here.
812 		 */
813 			perf_attr_iommu[0].data = perf->max_concurrent;
814 			for (i = 0; i < num_attrs; i++)
815 				attrs[i] = &perf_attr_iommu[i].attr.attr;
816 		}
817 		perf->attr_group->name = perf->block_name;
818 		perf->attr_group->attrs = attrs;
819 		ret = sysfs_create_group(dev->kobj_perf, perf->attr_group);
820 		if (ret < 0)
821 			return ret;
822 	}
823 
824 	return 0;
825 }
826 
827 /* Called with write topology lock acquired */
828 static int kfd_build_sysfs_node_tree(void)
829 {
830 	struct kfd_topology_device *dev;
831 	int ret;
832 	uint32_t i = 0;
833 
834 	list_for_each_entry(dev, &topology_device_list, list) {
835 		ret = kfd_build_sysfs_node_entry(dev, i);
836 		if (ret < 0)
837 			return ret;
838 		i++;
839 	}
840 
841 	return 0;
842 }
843 
844 /* Called with write topology lock acquired */
845 static void kfd_remove_sysfs_node_tree(void)
846 {
847 	struct kfd_topology_device *dev;
848 
849 	list_for_each_entry(dev, &topology_device_list, list)
850 		kfd_remove_sysfs_node_entry(dev);
851 }
852 
853 static int kfd_topology_update_sysfs(void)
854 {
855 	int ret;
856 
857 	if (!sys_props.kobj_topology) {
858 		sys_props.kobj_topology =
859 				kfd_alloc_struct(sys_props.kobj_topology);
860 		if (!sys_props.kobj_topology)
861 			return -ENOMEM;
862 
863 		ret = kobject_init_and_add(sys_props.kobj_topology,
864 				&sysprops_type,  &kfd_device->kobj,
865 				"topology");
866 		if (ret < 0) {
867 			kobject_put(sys_props.kobj_topology);
868 			return ret;
869 		}
870 
871 		sys_props.kobj_nodes = kobject_create_and_add("nodes",
872 				sys_props.kobj_topology);
873 		if (!sys_props.kobj_nodes)
874 			return -ENOMEM;
875 
876 		sys_props.attr_genid.name = "generation_id";
877 		sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE;
878 		sysfs_attr_init(&sys_props.attr_genid);
879 		ret = sysfs_create_file(sys_props.kobj_topology,
880 				&sys_props.attr_genid);
881 		if (ret < 0)
882 			return ret;
883 
884 		sys_props.attr_props.name = "system_properties";
885 		sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE;
886 		sysfs_attr_init(&sys_props.attr_props);
887 		ret = sysfs_create_file(sys_props.kobj_topology,
888 				&sys_props.attr_props);
889 		if (ret < 0)
890 			return ret;
891 	}
892 
893 	kfd_remove_sysfs_node_tree();
894 
895 	return kfd_build_sysfs_node_tree();
896 }
897 
898 static void kfd_topology_release_sysfs(void)
899 {
900 	kfd_remove_sysfs_node_tree();
901 	if (sys_props.kobj_topology) {
902 		sysfs_remove_file(sys_props.kobj_topology,
903 				&sys_props.attr_genid);
904 		sysfs_remove_file(sys_props.kobj_topology,
905 				&sys_props.attr_props);
906 		if (sys_props.kobj_nodes) {
907 			kobject_del(sys_props.kobj_nodes);
908 			kobject_put(sys_props.kobj_nodes);
909 			sys_props.kobj_nodes = NULL;
910 		}
911 		kobject_del(sys_props.kobj_topology);
912 		kobject_put(sys_props.kobj_topology);
913 		sys_props.kobj_topology = NULL;
914 	}
915 }
916 
917 /* Called with write topology_lock acquired */
918 static void kfd_topology_update_device_list(struct list_head *temp_list,
919 					struct list_head *master_list)
920 {
921 	while (!list_empty(temp_list)) {
922 		list_move_tail(temp_list->next, master_list);
923 		sys_props.num_devices++;
924 	}
925 }
926 
927 static void kfd_debug_print_topology(void)
928 {
929 	struct kfd_topology_device *dev;
930 
931 	down_read(&topology_lock);
932 
933 	dev = list_last_entry(&topology_device_list,
934 			struct kfd_topology_device, list);
935 	if (dev) {
936 		if (dev->node_props.cpu_cores_count &&
937 				dev->node_props.simd_count) {
938 			pr_info("Topology: Add APU node [0x%0x:0x%0x]\n",
939 				dev->node_props.device_id,
940 				dev->node_props.vendor_id);
941 		} else if (dev->node_props.cpu_cores_count)
942 			pr_info("Topology: Add CPU node\n");
943 		else if (dev->node_props.simd_count)
944 			pr_info("Topology: Add dGPU node [0x%0x:0x%0x]\n",
945 				dev->node_props.device_id,
946 				dev->node_props.vendor_id);
947 	}
948 	up_read(&topology_lock);
949 }
950 
951 /* Helper function for intializing platform_xx members of
952  * kfd_system_properties. Uses OEM info from the last CPU/APU node.
953  */
954 static void kfd_update_system_properties(void)
955 {
956 	struct kfd_topology_device *dev;
957 
958 	down_read(&topology_lock);
959 	dev = list_last_entry(&topology_device_list,
960 			struct kfd_topology_device, list);
961 	if (dev) {
962 		sys_props.platform_id = dev->oem_id64;
963 		sys_props.platform_oem = *((uint64_t *)dev->oem_table_id);
964 		sys_props.platform_rev = dev->oem_revision;
965 	}
966 	up_read(&topology_lock);
967 }
968 
969 static void find_system_memory(const struct dmi_header *dm,
970 	void *private)
971 {
972 	struct kfd_mem_properties *mem;
973 	u16 mem_width, mem_clock;
974 	struct kfd_topology_device *kdev =
975 		(struct kfd_topology_device *)private;
976 	const u8 *dmi_data = (const u8 *)(dm + 1);
977 
978 	if (dm->type == DMI_ENTRY_MEM_DEVICE && dm->length >= 0x15) {
979 		mem_width = (u16)(*(const u16 *)(dmi_data + 0x6));
980 		mem_clock = (u16)(*(const u16 *)(dmi_data + 0x11));
981 		list_for_each_entry(mem, &kdev->mem_props, list) {
982 			if (mem_width != 0xFFFF && mem_width != 0)
983 				mem->width = mem_width;
984 			if (mem_clock != 0)
985 				mem->mem_clk_max = mem_clock;
986 		}
987 	}
988 }
989 
990 /* kfd_add_non_crat_information - Add information that is not currently
991  *	defined in CRAT but is necessary for KFD topology
992  * @dev - topology device to which addition info is added
993  */
994 static void kfd_add_non_crat_information(struct kfd_topology_device *kdev)
995 {
996 	/* Check if CPU only node. */
997 	if (!kdev->gpu) {
998 		/* Add system memory information */
999 		dmi_walk(find_system_memory, kdev);
1000 	}
1001 	/* TODO: For GPU node, rearrange code from kfd_topology_add_device */
1002 }
1003 
1004 int kfd_topology_init(void)
1005 {
1006 	void *crat_image = NULL;
1007 	size_t image_size = 0;
1008 	int ret;
1009 	struct list_head temp_topology_device_list;
1010 	int cpu_only_node = 0;
1011 	struct kfd_topology_device *kdev;
1012 	int proximity_domain;
1013 
1014 	/* topology_device_list - Master list of all topology devices
1015 	 * temp_topology_device_list - temporary list created while parsing CRAT
1016 	 * or VCRAT. Once parsing is complete the contents of list is moved to
1017 	 * topology_device_list
1018 	 */
1019 
1020 	/* Initialize the head for the both the lists */
1021 	INIT_LIST_HEAD(&topology_device_list);
1022 	INIT_LIST_HEAD(&temp_topology_device_list);
1023 	init_rwsem(&topology_lock);
1024 
1025 	memset(&sys_props, 0, sizeof(sys_props));
1026 
1027 	/* Proximity domains in ACPI CRAT tables start counting at
1028 	 * 0. The same should be true for virtual CRAT tables created
1029 	 * at this stage. GPUs added later in kfd_topology_add_device
1030 	 * use a counter.
1031 	 */
1032 	proximity_domain = 0;
1033 
1034 	ret = kfd_create_crat_image_virtual(&crat_image, &image_size,
1035 					    COMPUTE_UNIT_CPU, NULL,
1036 					    proximity_domain);
1037 	cpu_only_node = 1;
1038 	if (ret) {
1039 		pr_err("Error creating VCRAT table for CPU\n");
1040 		return ret;
1041 	}
1042 
1043 	ret = kfd_parse_crat_table(crat_image,
1044 				   &temp_topology_device_list,
1045 				   proximity_domain);
1046 	if (ret) {
1047 		pr_err("Error parsing VCRAT table for CPU\n");
1048 		goto err;
1049 	}
1050 
1051 	kdev = list_first_entry(&temp_topology_device_list,
1052 				struct kfd_topology_device, list);
1053 
1054 	down_write(&topology_lock);
1055 	kfd_topology_update_device_list(&temp_topology_device_list,
1056 					&topology_device_list);
1057 	topology_crat_proximity_domain = sys_props.num_devices-1;
1058 	ret = kfd_topology_update_sysfs();
1059 	up_write(&topology_lock);
1060 
1061 	if (!ret) {
1062 		sys_props.generation_count++;
1063 		kfd_update_system_properties();
1064 		kfd_debug_print_topology();
1065 	} else
1066 		pr_err("Failed to update topology in sysfs ret=%d\n", ret);
1067 
1068 	/* For nodes with GPU, this information gets added
1069 	 * when GPU is detected (kfd_topology_add_device).
1070 	 */
1071 	if (cpu_only_node) {
1072 		/* Add additional information to CPU only node created above */
1073 		down_write(&topology_lock);
1074 		kdev = list_first_entry(&topology_device_list,
1075 				struct kfd_topology_device, list);
1076 		up_write(&topology_lock);
1077 		kfd_add_non_crat_information(kdev);
1078 	}
1079 
1080 err:
1081 	kfd_destroy_crat_image(crat_image);
1082 	return ret;
1083 }
1084 
1085 void kfd_topology_shutdown(void)
1086 {
1087 	down_write(&topology_lock);
1088 	kfd_topology_release_sysfs();
1089 	kfd_release_live_view();
1090 	up_write(&topology_lock);
1091 }
1092 
1093 static uint32_t kfd_generate_gpu_id(struct kfd_node *gpu)
1094 {
1095 	uint32_t gpu_id;
1096 	uint32_t buf[8];
1097 	uint64_t local_mem_size;
1098 	struct kfd_topology_device *dev;
1099 	bool is_unique;
1100 	uint8_t *crc_buf;
1101 
1102 	if (!gpu)
1103 		return 0;
1104 
1105 	crc_buf = (uint8_t *)&buf;
1106 	local_mem_size = gpu->local_mem_info.local_mem_size_private +
1107 			gpu->local_mem_info.local_mem_size_public;
1108 	buf[0] = gpu->adev->pdev->devfn;
1109 	buf[1] = gpu->adev->pdev->subsystem_vendor |
1110 		(gpu->adev->pdev->subsystem_device << 16);
1111 	buf[2] = pci_domain_nr(gpu->adev->pdev->bus);
1112 	buf[3] = gpu->adev->pdev->device;
1113 	buf[4] = gpu->adev->pdev->bus->number;
1114 	buf[5] = lower_32_bits(local_mem_size);
1115 	buf[6] = upper_32_bits(local_mem_size);
1116 	buf[7] = (ffs(gpu->xcc_mask) - 1) | (NUM_XCC(gpu->xcc_mask) << 16);
1117 
1118 	gpu_id = crc16(0, crc_buf, sizeof(buf)) &
1119 		 ((1 << KFD_GPU_ID_HASH_WIDTH) - 1);
1120 
1121 	/* There is a very small possibility when generating a
1122 	 * 16 (KFD_GPU_ID_HASH_WIDTH) bit value from 8 word buffer
1123 	 * that the value could be 0 or non-unique. So, check if
1124 	 * it is unique and non-zero. If not unique increment till
1125 	 * unique one is found. In case of overflow, restart from 1
1126 	 */
1127 
1128 	down_read(&topology_lock);
1129 	do {
1130 		is_unique = true;
1131 		if (!gpu_id)
1132 			gpu_id = 1;
1133 		list_for_each_entry(dev, &topology_device_list, list) {
1134 			if (dev->gpu && dev->gpu_id == gpu_id) {
1135 				is_unique = false;
1136 				break;
1137 			}
1138 		}
1139 		if (unlikely(!is_unique))
1140 			gpu_id = (gpu_id + 1) &
1141 				  ((1 << KFD_GPU_ID_HASH_WIDTH) - 1);
1142 	} while (!is_unique);
1143 	up_read(&topology_lock);
1144 
1145 	return gpu_id;
1146 }
1147 /* kfd_assign_gpu - Attach @gpu to the correct kfd topology device. If
1148  *		the GPU device is not already present in the topology device
1149  *		list then return NULL. This means a new topology device has to
1150  *		be created for this GPU.
1151  */
1152 static struct kfd_topology_device *kfd_assign_gpu(struct kfd_node *gpu)
1153 {
1154 	struct kfd_topology_device *dev;
1155 	struct kfd_topology_device *out_dev = NULL;
1156 	struct kfd_mem_properties *mem;
1157 	struct kfd_cache_properties *cache;
1158 	struct kfd_iolink_properties *iolink;
1159 	struct kfd_iolink_properties *p2plink;
1160 
1161 	list_for_each_entry(dev, &topology_device_list, list) {
1162 		/* Discrete GPUs need their own topology device list
1163 		 * entries. Don't assign them to CPU/APU nodes.
1164 		 */
1165 		if (dev->node_props.cpu_cores_count)
1166 			continue;
1167 
1168 		if (!dev->gpu && (dev->node_props.simd_count > 0)) {
1169 			dev->gpu = gpu;
1170 			out_dev = dev;
1171 
1172 			list_for_each_entry(mem, &dev->mem_props, list)
1173 				mem->gpu = dev->gpu;
1174 			list_for_each_entry(cache, &dev->cache_props, list)
1175 				cache->gpu = dev->gpu;
1176 			list_for_each_entry(iolink, &dev->io_link_props, list)
1177 				iolink->gpu = dev->gpu;
1178 			list_for_each_entry(p2plink, &dev->p2p_link_props, list)
1179 				p2plink->gpu = dev->gpu;
1180 			break;
1181 		}
1182 	}
1183 	return out_dev;
1184 }
1185 
1186 static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival)
1187 {
1188 	/*
1189 	 * TODO: Generate an event for thunk about the arrival/removal
1190 	 * of the GPU
1191 	 */
1192 }
1193 
1194 /* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info,
1195  *		patch this after CRAT parsing.
1196  */
1197 static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev)
1198 {
1199 	struct kfd_mem_properties *mem;
1200 	struct kfd_local_mem_info local_mem_info;
1201 
1202 	if (!dev)
1203 		return;
1204 
1205 	/* Currently, amdgpu driver (amdgpu_mc) deals only with GPUs with
1206 	 * single bank of VRAM local memory.
1207 	 * for dGPUs - VCRAT reports only one bank of Local Memory
1208 	 * for APUs - If CRAT from ACPI reports more than one bank, then
1209 	 *	all the banks will report the same mem_clk_max information
1210 	 */
1211 	amdgpu_amdkfd_get_local_mem_info(dev->gpu->adev, &local_mem_info,
1212 					 dev->gpu->xcp);
1213 
1214 	list_for_each_entry(mem, &dev->mem_props, list)
1215 		mem->mem_clk_max = local_mem_info.mem_clk_max;
1216 }
1217 
1218 static void kfd_set_iolink_no_atomics(struct kfd_topology_device *dev,
1219 					struct kfd_topology_device *target_gpu_dev,
1220 					struct kfd_iolink_properties *link)
1221 {
1222 	/* xgmi always supports atomics between links. */
1223 	if (link->iolink_type == CRAT_IOLINK_TYPE_XGMI)
1224 		return;
1225 
1226 	/* check pcie support to set cpu(dev) flags for target_gpu_dev link. */
1227 	if (target_gpu_dev) {
1228 		uint32_t cap;
1229 
1230 		pcie_capability_read_dword(target_gpu_dev->gpu->adev->pdev,
1231 				PCI_EXP_DEVCAP2, &cap);
1232 
1233 		if (!(cap & (PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
1234 			     PCI_EXP_DEVCAP2_ATOMIC_COMP64)))
1235 			link->flags |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1236 				CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1237 	/* set gpu (dev) flags. */
1238 	} else {
1239 		if (!dev->gpu->kfd->pci_atomic_requested ||
1240 				dev->gpu->adev->asic_type == CHIP_HAWAII)
1241 			link->flags |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1242 				CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1243 	}
1244 }
1245 
1246 static void kfd_set_iolink_non_coherent(struct kfd_topology_device *to_dev,
1247 		struct kfd_iolink_properties *outbound_link,
1248 		struct kfd_iolink_properties *inbound_link)
1249 {
1250 	/* CPU -> GPU with PCIe */
1251 	if (!to_dev->gpu &&
1252 	    inbound_link->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS)
1253 		inbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT;
1254 
1255 	if (to_dev->gpu) {
1256 		/* GPU <-> GPU with PCIe and
1257 		 * Vega20 with XGMI
1258 		 */
1259 		if (inbound_link->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS ||
1260 		    (inbound_link->iolink_type == CRAT_IOLINK_TYPE_XGMI &&
1261 		    KFD_GC_VERSION(to_dev->gpu) == IP_VERSION(9, 4, 0))) {
1262 			outbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT;
1263 			inbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT;
1264 		}
1265 	}
1266 }
1267 
1268 static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev)
1269 {
1270 	struct kfd_iolink_properties *link, *inbound_link;
1271 	struct kfd_topology_device *peer_dev;
1272 
1273 	if (!dev || !dev->gpu)
1274 		return;
1275 
1276 	/* GPU only creates direct links so apply flags setting to all */
1277 	list_for_each_entry(link, &dev->io_link_props, list) {
1278 		link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1279 		kfd_set_iolink_no_atomics(dev, NULL, link);
1280 		peer_dev = kfd_topology_device_by_proximity_domain(
1281 				link->node_to);
1282 
1283 		if (!peer_dev)
1284 			continue;
1285 
1286 		/* Include the CPU peer in GPU hive if connected over xGMI. */
1287 		if (!peer_dev->gpu &&
1288 		    link->iolink_type == CRAT_IOLINK_TYPE_XGMI) {
1289 			/*
1290 			 * If the GPU is not part of a GPU hive, use its pci
1291 			 * device location as the hive ID to bind with the CPU.
1292 			 */
1293 			if (!dev->node_props.hive_id)
1294 				dev->node_props.hive_id = pci_dev_id(dev->gpu->adev->pdev);
1295 			peer_dev->node_props.hive_id = dev->node_props.hive_id;
1296 		}
1297 
1298 		list_for_each_entry(inbound_link, &peer_dev->io_link_props,
1299 									list) {
1300 			if (inbound_link->node_to != link->node_from)
1301 				continue;
1302 
1303 			inbound_link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1304 			kfd_set_iolink_no_atomics(peer_dev, dev, inbound_link);
1305 			kfd_set_iolink_non_coherent(peer_dev, link, inbound_link);
1306 		}
1307 	}
1308 
1309 	/* Create indirect links so apply flags setting to all */
1310 	list_for_each_entry(link, &dev->p2p_link_props, list) {
1311 		link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1312 		kfd_set_iolink_no_atomics(dev, NULL, link);
1313 		peer_dev = kfd_topology_device_by_proximity_domain(
1314 				link->node_to);
1315 
1316 		if (!peer_dev)
1317 			continue;
1318 
1319 		list_for_each_entry(inbound_link, &peer_dev->p2p_link_props,
1320 									list) {
1321 			if (inbound_link->node_to != link->node_from)
1322 				continue;
1323 
1324 			inbound_link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1325 			kfd_set_iolink_no_atomics(peer_dev, dev, inbound_link);
1326 			kfd_set_iolink_non_coherent(peer_dev, link, inbound_link);
1327 		}
1328 	}
1329 }
1330 
1331 static int kfd_build_p2p_node_entry(struct kfd_topology_device *dev,
1332 				struct kfd_iolink_properties *p2plink)
1333 {
1334 	int ret;
1335 
1336 	p2plink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
1337 	if (!p2plink->kobj)
1338 		return -ENOMEM;
1339 
1340 	ret = kobject_init_and_add(p2plink->kobj, &iolink_type,
1341 			dev->kobj_p2plink, "%d", dev->node_props.p2p_links_count - 1);
1342 	if (ret < 0) {
1343 		kobject_put(p2plink->kobj);
1344 		return ret;
1345 	}
1346 
1347 	p2plink->attr.name = "properties";
1348 	p2plink->attr.mode = KFD_SYSFS_FILE_MODE;
1349 	sysfs_attr_init(&p2plink->attr);
1350 	ret = sysfs_create_file(p2plink->kobj, &p2plink->attr);
1351 	if (ret < 0)
1352 		return ret;
1353 
1354 	return 0;
1355 }
1356 
1357 static int kfd_create_indirect_link_prop(struct kfd_topology_device *kdev, int gpu_node)
1358 {
1359 	struct kfd_iolink_properties *gpu_link, *tmp_link, *cpu_link;
1360 	struct kfd_iolink_properties *props = NULL, *props2 = NULL;
1361 	struct kfd_topology_device *cpu_dev;
1362 	int ret = 0;
1363 	int i, num_cpu;
1364 
1365 	num_cpu = 0;
1366 	list_for_each_entry(cpu_dev, &topology_device_list, list) {
1367 		if (cpu_dev->gpu)
1368 			break;
1369 		num_cpu++;
1370 	}
1371 
1372 	if (list_empty(&kdev->io_link_props))
1373 		return -ENODATA;
1374 
1375 	gpu_link = list_first_entry(&kdev->io_link_props,
1376 				    struct kfd_iolink_properties, list);
1377 
1378 	for (i = 0; i < num_cpu; i++) {
1379 		/* CPU <--> GPU */
1380 		if (gpu_link->node_to == i)
1381 			continue;
1382 
1383 		/* find CPU <-->  CPU links */
1384 		cpu_link = NULL;
1385 		cpu_dev = kfd_topology_device_by_proximity_domain(i);
1386 		if (cpu_dev) {
1387 			list_for_each_entry(tmp_link,
1388 					&cpu_dev->io_link_props, list) {
1389 				if (tmp_link->node_to == gpu_link->node_to) {
1390 					cpu_link = tmp_link;
1391 					break;
1392 				}
1393 			}
1394 		}
1395 
1396 		if (!cpu_link)
1397 			return -ENOMEM;
1398 
1399 		/* CPU <--> CPU <--> GPU, GPU node*/
1400 		props = kfd_alloc_struct(props);
1401 		if (!props)
1402 			return -ENOMEM;
1403 
1404 		memcpy(props, gpu_link, sizeof(struct kfd_iolink_properties));
1405 		props->weight = gpu_link->weight + cpu_link->weight;
1406 		props->min_latency = gpu_link->min_latency + cpu_link->min_latency;
1407 		props->max_latency = gpu_link->max_latency + cpu_link->max_latency;
1408 		props->min_bandwidth = min(gpu_link->min_bandwidth, cpu_link->min_bandwidth);
1409 		props->max_bandwidth = min(gpu_link->max_bandwidth, cpu_link->max_bandwidth);
1410 
1411 		props->node_from = gpu_node;
1412 		props->node_to = i;
1413 		kdev->node_props.p2p_links_count++;
1414 		list_add_tail(&props->list, &kdev->p2p_link_props);
1415 		ret = kfd_build_p2p_node_entry(kdev, props);
1416 		if (ret < 0)
1417 			return ret;
1418 
1419 		/* for small Bar, no CPU --> GPU in-direct links */
1420 		if (kfd_dev_is_large_bar(kdev->gpu)) {
1421 			/* CPU <--> CPU <--> GPU, CPU node*/
1422 			props2 = kfd_alloc_struct(props2);
1423 			if (!props2)
1424 				return -ENOMEM;
1425 
1426 			memcpy(props2, props, sizeof(struct kfd_iolink_properties));
1427 			props2->node_from = i;
1428 			props2->node_to = gpu_node;
1429 			props2->kobj = NULL;
1430 			cpu_dev->node_props.p2p_links_count++;
1431 			list_add_tail(&props2->list, &cpu_dev->p2p_link_props);
1432 			ret = kfd_build_p2p_node_entry(cpu_dev, props2);
1433 			if (ret < 0)
1434 				return ret;
1435 		}
1436 	}
1437 	return ret;
1438 }
1439 
1440 #if defined(CONFIG_HSA_AMD_P2P)
1441 static int kfd_add_peer_prop(struct kfd_topology_device *kdev,
1442 		struct kfd_topology_device *peer, int from, int to)
1443 {
1444 	struct kfd_iolink_properties *props = NULL;
1445 	struct kfd_iolink_properties *iolink1, *iolink2, *iolink3;
1446 	struct kfd_topology_device *cpu_dev;
1447 	int ret = 0;
1448 
1449 	if (!amdgpu_device_is_peer_accessible(
1450 				kdev->gpu->adev,
1451 				peer->gpu->adev))
1452 		return ret;
1453 
1454 	if (list_empty(&kdev->io_link_props))
1455 		return -ENODATA;
1456 
1457 	iolink1 = list_first_entry(&kdev->io_link_props,
1458 				   struct kfd_iolink_properties, list);
1459 
1460 	if (list_empty(&peer->io_link_props))
1461 		return -ENODATA;
1462 
1463 	iolink2 = list_first_entry(&peer->io_link_props,
1464 				   struct kfd_iolink_properties, list);
1465 
1466 	props = kfd_alloc_struct(props);
1467 	if (!props)
1468 		return -ENOMEM;
1469 
1470 	memcpy(props, iolink1, sizeof(struct kfd_iolink_properties));
1471 
1472 	props->weight = iolink1->weight + iolink2->weight;
1473 	props->min_latency = iolink1->min_latency + iolink2->min_latency;
1474 	props->max_latency = iolink1->max_latency + iolink2->max_latency;
1475 	props->min_bandwidth = min(iolink1->min_bandwidth, iolink2->min_bandwidth);
1476 	props->max_bandwidth = min(iolink2->max_bandwidth, iolink2->max_bandwidth);
1477 
1478 	if (iolink1->node_to != iolink2->node_to) {
1479 		/* CPU->CPU  link*/
1480 		cpu_dev = kfd_topology_device_by_proximity_domain(iolink1->node_to);
1481 		if (cpu_dev) {
1482 			list_for_each_entry(iolink3, &cpu_dev->io_link_props, list) {
1483 				if (iolink3->node_to != iolink2->node_to)
1484 					continue;
1485 
1486 				props->weight += iolink3->weight;
1487 				props->min_latency += iolink3->min_latency;
1488 				props->max_latency += iolink3->max_latency;
1489 				props->min_bandwidth = min(props->min_bandwidth,
1490 							   iolink3->min_bandwidth);
1491 				props->max_bandwidth = min(props->max_bandwidth,
1492 							   iolink3->max_bandwidth);
1493 				break;
1494 			}
1495 		} else {
1496 			WARN(1, "CPU node not found");
1497 		}
1498 	}
1499 
1500 	props->node_from = from;
1501 	props->node_to = to;
1502 	peer->node_props.p2p_links_count++;
1503 	list_add_tail(&props->list, &peer->p2p_link_props);
1504 	ret = kfd_build_p2p_node_entry(peer, props);
1505 
1506 	return ret;
1507 }
1508 #endif
1509 
1510 static int kfd_dev_create_p2p_links(void)
1511 {
1512 	struct kfd_topology_device *dev;
1513 	struct kfd_topology_device *new_dev;
1514 #if defined(CONFIG_HSA_AMD_P2P)
1515 	uint32_t i;
1516 #endif
1517 	uint32_t k;
1518 	int ret = 0;
1519 
1520 	k = 0;
1521 	list_for_each_entry(dev, &topology_device_list, list)
1522 		k++;
1523 	if (k < 2)
1524 		return 0;
1525 
1526 	new_dev = list_last_entry(&topology_device_list, struct kfd_topology_device, list);
1527 	if (WARN_ON(!new_dev->gpu))
1528 		return 0;
1529 
1530 	k--;
1531 
1532 	/* create in-direct links */
1533 	ret = kfd_create_indirect_link_prop(new_dev, k);
1534 	if (ret < 0)
1535 		goto out;
1536 
1537 	/* create p2p links */
1538 #if defined(CONFIG_HSA_AMD_P2P)
1539 	i = 0;
1540 	list_for_each_entry(dev, &topology_device_list, list) {
1541 		if (dev == new_dev)
1542 			break;
1543 		if (!dev->gpu || !dev->gpu->adev ||
1544 		    (dev->gpu->kfd->hive_id &&
1545 		     dev->gpu->kfd->hive_id == new_dev->gpu->kfd->hive_id))
1546 			goto next;
1547 
1548 		/* check if node(s) is/are peer accessible in one direction or bi-direction */
1549 		ret = kfd_add_peer_prop(new_dev, dev, i, k);
1550 		if (ret < 0)
1551 			goto out;
1552 
1553 		ret = kfd_add_peer_prop(dev, new_dev, k, i);
1554 		if (ret < 0)
1555 			goto out;
1556 next:
1557 		i++;
1558 	}
1559 #endif
1560 
1561 out:
1562 	return ret;
1563 }
1564 
1565 /* Helper function. See kfd_fill_gpu_cache_info for parameter description */
1566 static int fill_in_l1_pcache(struct kfd_cache_properties **props_ext,
1567 				struct kfd_gpu_cache_info *pcache_info,
1568 				int cu_bitmask,
1569 				int cache_type, unsigned int cu_processor_id,
1570 				int cu_block)
1571 {
1572 	unsigned int cu_sibling_map_mask;
1573 	int first_active_cu;
1574 	struct kfd_cache_properties *pcache = NULL;
1575 
1576 	cu_sibling_map_mask = cu_bitmask;
1577 	cu_sibling_map_mask >>= cu_block;
1578 	cu_sibling_map_mask &= ((1 << pcache_info[cache_type].num_cu_shared) - 1);
1579 	first_active_cu = ffs(cu_sibling_map_mask);
1580 
1581 	/* CU could be inactive. In case of shared cache find the first active
1582 	 * CU. and incase of non-shared cache check if the CU is inactive. If
1583 	 * inactive active skip it
1584 	 */
1585 	if (first_active_cu) {
1586 		pcache = kfd_alloc_struct(pcache);
1587 		if (!pcache)
1588 			return -ENOMEM;
1589 
1590 		memset(pcache, 0, sizeof(struct kfd_cache_properties));
1591 		pcache->processor_id_low = cu_processor_id + (first_active_cu - 1);
1592 		pcache->cache_level = pcache_info[cache_type].cache_level;
1593 		pcache->cache_size = pcache_info[cache_type].cache_size;
1594 		pcache->cacheline_size = pcache_info[cache_type].cache_line_size;
1595 
1596 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_DATA_CACHE)
1597 			pcache->cache_type |= HSA_CACHE_TYPE_DATA;
1598 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_INST_CACHE)
1599 			pcache->cache_type |= HSA_CACHE_TYPE_INSTRUCTION;
1600 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_CPU_CACHE)
1601 			pcache->cache_type |= HSA_CACHE_TYPE_CPU;
1602 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_SIMD_CACHE)
1603 			pcache->cache_type |= HSA_CACHE_TYPE_HSACU;
1604 
1605 		/* Sibling map is w.r.t processor_id_low, so shift out
1606 		 * inactive CU
1607 		 */
1608 		cu_sibling_map_mask =
1609 			cu_sibling_map_mask >> (first_active_cu - 1);
1610 
1611 		pcache->sibling_map[0] = (uint8_t)(cu_sibling_map_mask & 0xFF);
1612 		pcache->sibling_map[1] =
1613 				(uint8_t)((cu_sibling_map_mask >> 8) & 0xFF);
1614 		pcache->sibling_map[2] =
1615 				(uint8_t)((cu_sibling_map_mask >> 16) & 0xFF);
1616 		pcache->sibling_map[3] =
1617 				(uint8_t)((cu_sibling_map_mask >> 24) & 0xFF);
1618 
1619 		pcache->sibling_map_size = 4;
1620 		*props_ext = pcache;
1621 
1622 		return 0;
1623 	}
1624 	return 1;
1625 }
1626 
1627 /* Helper function. See kfd_fill_gpu_cache_info for parameter description */
1628 static int fill_in_l2_l3_pcache(struct kfd_cache_properties **props_ext,
1629 				struct kfd_gpu_cache_info *pcache_info,
1630 				struct amdgpu_cu_info *cu_info,
1631 				struct amdgpu_gfx_config *gfx_info,
1632 				int cache_type, unsigned int cu_processor_id,
1633 				struct kfd_node *knode)
1634 {
1635 	unsigned int cu_sibling_map_mask;
1636 	int first_active_cu;
1637 	int i, j, k, xcc, start, end;
1638 	int num_xcc = NUM_XCC(knode->xcc_mask);
1639 	struct kfd_cache_properties *pcache = NULL;
1640 	enum amdgpu_memory_partition mode;
1641 	struct amdgpu_device *adev = knode->adev;
1642 
1643 	start = ffs(knode->xcc_mask) - 1;
1644 	end = start + num_xcc;
1645 	cu_sibling_map_mask = cu_info->bitmap[start][0][0];
1646 	cu_sibling_map_mask &=
1647 		((1 << pcache_info[cache_type].num_cu_shared) - 1);
1648 	first_active_cu = ffs(cu_sibling_map_mask);
1649 
1650 	/* CU could be inactive. In case of shared cache find the first active
1651 	 * CU. and incase of non-shared cache check if the CU is inactive. If
1652 	 * inactive active skip it
1653 	 */
1654 	if (first_active_cu) {
1655 		pcache = kfd_alloc_struct(pcache);
1656 		if (!pcache)
1657 			return -ENOMEM;
1658 
1659 		memset(pcache, 0, sizeof(struct kfd_cache_properties));
1660 		pcache->processor_id_low = cu_processor_id
1661 					+ (first_active_cu - 1);
1662 		pcache->cache_level = pcache_info[cache_type].cache_level;
1663 		pcache->cacheline_size = pcache_info[cache_type].cache_line_size;
1664 
1665 		if (KFD_GC_VERSION(knode) == IP_VERSION(9, 4, 3) ||
1666 		    KFD_GC_VERSION(knode) == IP_VERSION(9, 4, 4))
1667 			mode = adev->gmc.gmc_funcs->query_mem_partition_mode(adev);
1668 		else
1669 			mode = UNKNOWN_MEMORY_PARTITION_MODE;
1670 
1671 		pcache->cache_size = pcache_info[cache_type].cache_size;
1672 		/* Partition mode only affects L3 cache size */
1673 		if (mode && pcache->cache_level == 3)
1674 			pcache->cache_size /= mode;
1675 
1676 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_DATA_CACHE)
1677 			pcache->cache_type |= HSA_CACHE_TYPE_DATA;
1678 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_INST_CACHE)
1679 			pcache->cache_type |= HSA_CACHE_TYPE_INSTRUCTION;
1680 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_CPU_CACHE)
1681 			pcache->cache_type |= HSA_CACHE_TYPE_CPU;
1682 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_SIMD_CACHE)
1683 			pcache->cache_type |= HSA_CACHE_TYPE_HSACU;
1684 
1685 		/* Sibling map is w.r.t processor_id_low, so shift out
1686 		 * inactive CU
1687 		 */
1688 		cu_sibling_map_mask = cu_sibling_map_mask >> (first_active_cu - 1);
1689 		k = 0;
1690 
1691 		for (xcc = start; xcc < end; xcc++) {
1692 			for (i = 0; i < gfx_info->max_shader_engines; i++) {
1693 				for (j = 0; j < gfx_info->max_sh_per_se; j++) {
1694 					pcache->sibling_map[k] = (uint8_t)(cu_sibling_map_mask & 0xFF);
1695 					pcache->sibling_map[k+1] = (uint8_t)((cu_sibling_map_mask >> 8) & 0xFF);
1696 					pcache->sibling_map[k+2] = (uint8_t)((cu_sibling_map_mask >> 16) & 0xFF);
1697 					pcache->sibling_map[k+3] = (uint8_t)((cu_sibling_map_mask >> 24) & 0xFF);
1698 					k += 4;
1699 
1700 					cu_sibling_map_mask = cu_info->bitmap[xcc][i % 4][j + i / 4];
1701 					cu_sibling_map_mask &= ((1 << pcache_info[cache_type].num_cu_shared) - 1);
1702 				}
1703 			}
1704 		}
1705 		pcache->sibling_map_size = k;
1706 		*props_ext = pcache;
1707 		return 0;
1708 	}
1709 	return 1;
1710 }
1711 
1712 #define KFD_MAX_CACHE_TYPES 6
1713 
1714 /* kfd_fill_cache_non_crat_info - Fill GPU cache info using kfd_gpu_cache_info
1715  * tables
1716  */
1717 static void kfd_fill_cache_non_crat_info(struct kfd_topology_device *dev, struct kfd_node *kdev)
1718 {
1719 	struct kfd_gpu_cache_info *pcache_info = NULL;
1720 	int i, j, k, xcc, start, end;
1721 	int ct = 0;
1722 	unsigned int cu_processor_id;
1723 	int ret;
1724 	unsigned int num_cu_shared;
1725 	struct amdgpu_cu_info *cu_info = &kdev->adev->gfx.cu_info;
1726 	struct amdgpu_gfx_config *gfx_info = &kdev->adev->gfx.config;
1727 	int gpu_processor_id;
1728 	struct kfd_cache_properties *props_ext;
1729 	int num_of_entries = 0;
1730 	int num_of_cache_types = 0;
1731 	struct kfd_gpu_cache_info cache_info[KFD_MAX_CACHE_TYPES];
1732 
1733 
1734 	gpu_processor_id = dev->node_props.simd_id_base;
1735 
1736 	memset(cache_info, 0, sizeof(cache_info));
1737 	pcache_info = cache_info;
1738 	num_of_cache_types = kfd_get_gpu_cache_info(kdev, &pcache_info);
1739 	if (!num_of_cache_types) {
1740 		pr_warn("no cache info found\n");
1741 		return;
1742 	}
1743 
1744 	/* For each type of cache listed in the kfd_gpu_cache_info table,
1745 	 * go through all available Compute Units.
1746 	 * The [i,j,k] loop will
1747 	 *		if kfd_gpu_cache_info.num_cu_shared = 1
1748 	 *			will parse through all available CU
1749 	 *		If (kfd_gpu_cache_info.num_cu_shared != 1)
1750 	 *			then it will consider only one CU from
1751 	 *			the shared unit
1752 	 */
1753 	start = ffs(kdev->xcc_mask) - 1;
1754 	end = start + NUM_XCC(kdev->xcc_mask);
1755 
1756 	for (ct = 0; ct < num_of_cache_types; ct++) {
1757 		cu_processor_id = gpu_processor_id;
1758 		if (pcache_info[ct].cache_level == 1) {
1759 			for (xcc = start; xcc < end; xcc++) {
1760 				for (i = 0; i < gfx_info->max_shader_engines; i++) {
1761 					for (j = 0; j < gfx_info->max_sh_per_se; j++) {
1762 						for (k = 0; k < gfx_info->max_cu_per_sh; k += pcache_info[ct].num_cu_shared) {
1763 
1764 							ret = fill_in_l1_pcache(&props_ext, pcache_info,
1765 										cu_info->bitmap[xcc][i % 4][j + i / 4], ct,
1766 										cu_processor_id, k);
1767 
1768 							if (ret < 0)
1769 								break;
1770 
1771 							if (!ret) {
1772 								num_of_entries++;
1773 								list_add_tail(&props_ext->list, &dev->cache_props);
1774 							}
1775 
1776 							/* Move to next CU block */
1777 							num_cu_shared = ((k + pcache_info[ct].num_cu_shared) <=
1778 								gfx_info->max_cu_per_sh) ?
1779 								pcache_info[ct].num_cu_shared :
1780 								(gfx_info->max_cu_per_sh - k);
1781 							cu_processor_id += num_cu_shared;
1782 						}
1783 					}
1784 				}
1785 			}
1786 		} else {
1787 			ret = fill_in_l2_l3_pcache(&props_ext, pcache_info,
1788 						   cu_info, gfx_info, ct, cu_processor_id, kdev);
1789 
1790 			if (ret < 0)
1791 				break;
1792 
1793 			if (!ret) {
1794 				num_of_entries++;
1795 				list_add_tail(&props_ext->list, &dev->cache_props);
1796 			}
1797 		}
1798 	}
1799 	dev->node_props.caches_count += num_of_entries;
1800 	pr_debug("Added [%d] GPU cache entries\n", num_of_entries);
1801 }
1802 
1803 static int kfd_topology_add_device_locked(struct kfd_node *gpu,
1804 					  struct kfd_topology_device **dev)
1805 {
1806 	int proximity_domain = ++topology_crat_proximity_domain;
1807 	struct list_head temp_topology_device_list;
1808 	void *crat_image = NULL;
1809 	size_t image_size = 0;
1810 	int res;
1811 
1812 	res = kfd_create_crat_image_virtual(&crat_image, &image_size,
1813 					    COMPUTE_UNIT_GPU, gpu,
1814 					    proximity_domain);
1815 	if (res) {
1816 		dev_err(gpu->adev->dev, "Error creating VCRAT\n");
1817 		topology_crat_proximity_domain--;
1818 		goto err;
1819 	}
1820 
1821 	INIT_LIST_HEAD(&temp_topology_device_list);
1822 
1823 	res = kfd_parse_crat_table(crat_image,
1824 				   &temp_topology_device_list,
1825 				   proximity_domain);
1826 	if (res) {
1827 		dev_err(gpu->adev->dev, "Error parsing VCRAT\n");
1828 		topology_crat_proximity_domain--;
1829 		goto err;
1830 	}
1831 
1832 	kfd_topology_update_device_list(&temp_topology_device_list,
1833 					&topology_device_list);
1834 
1835 	*dev = kfd_assign_gpu(gpu);
1836 	if (WARN_ON(!*dev)) {
1837 		res = -ENODEV;
1838 		goto err;
1839 	}
1840 
1841 	/* Fill the cache affinity information here for the GPUs
1842 	 * using VCRAT
1843 	 */
1844 	kfd_fill_cache_non_crat_info(*dev, gpu);
1845 
1846 	/* Update the SYSFS tree, since we added another topology
1847 	 * device
1848 	 */
1849 	res = kfd_topology_update_sysfs();
1850 	if (!res)
1851 		sys_props.generation_count++;
1852 	else
1853 		dev_err(gpu->adev->dev, "Failed to update GPU to sysfs topology. res=%d\n",
1854 			res);
1855 
1856 err:
1857 	kfd_destroy_crat_image(crat_image);
1858 	return res;
1859 }
1860 
1861 static void kfd_topology_set_dbg_firmware_support(struct kfd_topology_device *dev)
1862 {
1863 	bool firmware_supported = true;
1864 
1865 	if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(11, 0, 0) &&
1866 			KFD_GC_VERSION(dev->gpu) < IP_VERSION(12, 0, 0)) {
1867 		uint32_t mes_api_rev = (dev->gpu->adev->mes.sched_version &
1868 						AMDGPU_MES_API_VERSION_MASK) >>
1869 						AMDGPU_MES_API_VERSION_SHIFT;
1870 		uint32_t mes_rev = dev->gpu->adev->mes.sched_version &
1871 						AMDGPU_MES_VERSION_MASK;
1872 
1873 		firmware_supported = (mes_api_rev >= 14) && (mes_rev >= 64);
1874 		goto out;
1875 	}
1876 
1877 	/*
1878 	 * Note: Any unlisted devices here are assumed to support exception handling.
1879 	 * Add additional checks here as needed.
1880 	 */
1881 	switch (KFD_GC_VERSION(dev->gpu)) {
1882 	case IP_VERSION(9, 0, 1):
1883 		firmware_supported = dev->gpu->kfd->mec_fw_version >= 459 + 32768;
1884 		break;
1885 	case IP_VERSION(9, 1, 0):
1886 	case IP_VERSION(9, 2, 1):
1887 	case IP_VERSION(9, 2, 2):
1888 	case IP_VERSION(9, 3, 0):
1889 	case IP_VERSION(9, 4, 0):
1890 		firmware_supported = dev->gpu->kfd->mec_fw_version >= 459;
1891 		break;
1892 	case IP_VERSION(9, 4, 1):
1893 		firmware_supported = dev->gpu->kfd->mec_fw_version >= 60;
1894 		break;
1895 	case IP_VERSION(9, 4, 2):
1896 		firmware_supported = dev->gpu->kfd->mec_fw_version >= 51;
1897 		break;
1898 	case IP_VERSION(10, 1, 10):
1899 	case IP_VERSION(10, 1, 2):
1900 	case IP_VERSION(10, 1, 1):
1901 		firmware_supported = dev->gpu->kfd->mec_fw_version >= 144;
1902 		break;
1903 	case IP_VERSION(10, 3, 0):
1904 	case IP_VERSION(10, 3, 2):
1905 	case IP_VERSION(10, 3, 1):
1906 	case IP_VERSION(10, 3, 4):
1907 	case IP_VERSION(10, 3, 5):
1908 		firmware_supported = dev->gpu->kfd->mec_fw_version >= 89;
1909 		break;
1910 	case IP_VERSION(10, 1, 3):
1911 	case IP_VERSION(10, 3, 3):
1912 		firmware_supported = false;
1913 		break;
1914 	default:
1915 		break;
1916 	}
1917 
1918 out:
1919 	if (firmware_supported)
1920 		dev->node_props.capability |= HSA_CAP_TRAP_DEBUG_FIRMWARE_SUPPORTED;
1921 }
1922 
1923 static void kfd_topology_set_capabilities(struct kfd_topology_device *dev)
1924 {
1925 	dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 <<
1926 				HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1927 				HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1928 
1929 	dev->node_props.capability |= HSA_CAP_TRAP_DEBUG_SUPPORT |
1930 			HSA_CAP_TRAP_DEBUG_WAVE_LAUNCH_TRAP_OVERRIDE_SUPPORTED |
1931 			HSA_CAP_TRAP_DEBUG_WAVE_LAUNCH_MODE_SUPPORTED;
1932 
1933 	if (kfd_dbg_has_ttmps_always_setup(dev->gpu))
1934 		dev->node_props.debug_prop |= HSA_DBG_DISPATCH_INFO_ALWAYS_VALID;
1935 
1936 	if (KFD_GC_VERSION(dev->gpu) < IP_VERSION(10, 0, 0)) {
1937 		if (KFD_GC_VERSION(dev->gpu) == IP_VERSION(9, 4, 3) ||
1938 		    KFD_GC_VERSION(dev->gpu) == IP_VERSION(9, 4, 4))
1939 			dev->node_props.debug_prop |=
1940 				HSA_DBG_WATCH_ADDR_MASK_LO_BIT_GFX9_4_3 |
1941 				HSA_DBG_WATCH_ADDR_MASK_HI_BIT_GFX9_4_3;
1942 		else
1943 			dev->node_props.debug_prop |=
1944 				HSA_DBG_WATCH_ADDR_MASK_LO_BIT_GFX9 |
1945 				HSA_DBG_WATCH_ADDR_MASK_HI_BIT;
1946 
1947 		if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(9, 4, 2))
1948 			dev->node_props.capability |=
1949 				HSA_CAP_TRAP_DEBUG_PRECISE_MEMORY_OPERATIONS_SUPPORTED;
1950 	} else {
1951 		dev->node_props.debug_prop |= HSA_DBG_WATCH_ADDR_MASK_LO_BIT_GFX10 |
1952 					HSA_DBG_WATCH_ADDR_MASK_HI_BIT;
1953 
1954 		if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(11, 0, 0))
1955 			dev->node_props.capability |=
1956 				HSA_CAP_TRAP_DEBUG_PRECISE_MEMORY_OPERATIONS_SUPPORTED;
1957 
1958 		if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(12, 0, 0))
1959 			dev->node_props.capability |=
1960 				HSA_CAP_TRAP_DEBUG_PRECISE_ALU_OPERATIONS_SUPPORTED;
1961 	}
1962 
1963 	kfd_topology_set_dbg_firmware_support(dev);
1964 }
1965 
1966 int kfd_topology_add_device(struct kfd_node *gpu)
1967 {
1968 	uint32_t gpu_id;
1969 	struct kfd_topology_device *dev;
1970 	int res = 0;
1971 	int i;
1972 	const char *asic_name = amdgpu_asic_name[gpu->adev->asic_type];
1973 	struct amdgpu_gfx_config *gfx_info = &gpu->adev->gfx.config;
1974 	struct amdgpu_cu_info *cu_info = &gpu->adev->gfx.cu_info;
1975 
1976 	if (gpu->xcp && !gpu->xcp->ddev) {
1977 		dev_warn(gpu->adev->dev,
1978 			 "Won't add GPU to topology since it has no drm node assigned.");
1979 		return 0;
1980 	} else {
1981 		dev_dbg(gpu->adev->dev, "Adding new GPU to topology\n");
1982 	}
1983 
1984 	/* Check to see if this gpu device exists in the topology_device_list.
1985 	 * If so, assign the gpu to that device,
1986 	 * else create a Virtual CRAT for this gpu device and then parse that
1987 	 * CRAT to create a new topology device. Once created assign the gpu to
1988 	 * that topology device
1989 	 */
1990 	down_write(&topology_lock);
1991 	dev = kfd_assign_gpu(gpu);
1992 	if (!dev)
1993 		res = kfd_topology_add_device_locked(gpu, &dev);
1994 	up_write(&topology_lock);
1995 	if (res)
1996 		return res;
1997 
1998 	gpu_id = kfd_generate_gpu_id(gpu);
1999 	dev->gpu_id = gpu_id;
2000 	gpu->id = gpu_id;
2001 
2002 	kfd_dev_create_p2p_links();
2003 
2004 	/* TODO: Move the following lines to function
2005 	 *	kfd_add_non_crat_information
2006 	 */
2007 
2008 	/* Fill-in additional information that is not available in CRAT but
2009 	 * needed for the topology
2010 	 */
2011 	for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1; i++) {
2012 		dev->node_props.name[i] = __tolower(asic_name[i]);
2013 		if (asic_name[i] == '\0')
2014 			break;
2015 	}
2016 	dev->node_props.name[i] = '\0';
2017 
2018 	dev->node_props.simd_arrays_per_engine =
2019 		gfx_info->max_sh_per_se;
2020 
2021 	dev->node_props.gfx_target_version =
2022 				gpu->kfd->device_info.gfx_target_version;
2023 	dev->node_props.vendor_id = gpu->adev->pdev->vendor;
2024 	dev->node_props.device_id = gpu->adev->pdev->device;
2025 	dev->node_props.capability |=
2026 		((dev->gpu->adev->rev_id << HSA_CAP_ASIC_REVISION_SHIFT) &
2027 			HSA_CAP_ASIC_REVISION_MASK);
2028 
2029 	dev->node_props.location_id = pci_dev_id(gpu->adev->pdev);
2030 	if (KFD_GC_VERSION(dev->gpu->kfd) == IP_VERSION(9, 4, 3))
2031 		dev->node_props.location_id |= dev->gpu->node_id;
2032 
2033 	dev->node_props.domain = pci_domain_nr(gpu->adev->pdev->bus);
2034 	dev->node_props.max_engine_clk_fcompute =
2035 		amdgpu_amdkfd_get_max_engine_clock_in_mhz(dev->gpu->adev);
2036 	dev->node_props.max_engine_clk_ccompute =
2037 		cpufreq_quick_get_max(0) / 1000;
2038 
2039 	if (gpu->xcp)
2040 		dev->node_props.drm_render_minor = gpu->xcp->ddev->render->index;
2041 	else
2042 		dev->node_props.drm_render_minor =
2043 				gpu->kfd->shared_resources.drm_render_minor;
2044 
2045 	dev->node_props.hive_id = gpu->kfd->hive_id;
2046 	dev->node_props.num_sdma_engines = kfd_get_num_sdma_engines(gpu);
2047 	dev->node_props.num_sdma_xgmi_engines =
2048 					kfd_get_num_xgmi_sdma_engines(gpu);
2049 	dev->node_props.num_sdma_queues_per_engine =
2050 				gpu->kfd->device_info.num_sdma_queues_per_engine -
2051 				gpu->kfd->device_info.num_reserved_sdma_queues_per_engine;
2052 	dev->node_props.num_gws = (dev->gpu->gws &&
2053 		dev->gpu->dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) ?
2054 		dev->gpu->adev->gds.gws_size : 0;
2055 	dev->node_props.num_cp_queues = get_cp_queues_num(dev->gpu->dqm);
2056 
2057 	kfd_fill_mem_clk_max_info(dev);
2058 	kfd_fill_iolink_non_crat_info(dev);
2059 
2060 	switch (dev->gpu->adev->asic_type) {
2061 	case CHIP_KAVERI:
2062 	case CHIP_HAWAII:
2063 	case CHIP_TONGA:
2064 		dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_PRE_1_0 <<
2065 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
2066 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
2067 		break;
2068 	case CHIP_CARRIZO:
2069 	case CHIP_FIJI:
2070 	case CHIP_POLARIS10:
2071 	case CHIP_POLARIS11:
2072 	case CHIP_POLARIS12:
2073 	case CHIP_VEGAM:
2074 		pr_debug("Adding doorbell packet type capability\n");
2075 		dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 <<
2076 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
2077 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
2078 		break;
2079 	default:
2080 		if (KFD_GC_VERSION(dev->gpu) < IP_VERSION(9, 0, 1))
2081 			WARN(1, "Unexpected ASIC family %u",
2082 			     dev->gpu->adev->asic_type);
2083 		else
2084 			kfd_topology_set_capabilities(dev);
2085 	}
2086 
2087 	/*
2088 	 * Overwrite ATS capability according to needs_iommu_device to fix
2089 	 * potential missing corresponding bit in CRAT of BIOS.
2090 	 */
2091 	dev->node_props.capability &= ~HSA_CAP_ATS_PRESENT;
2092 
2093 	/* Fix errors in CZ CRAT.
2094 	 * simd_count: Carrizo CRAT reports wrong simd_count, probably
2095 	 *		because it doesn't consider masked out CUs
2096 	 * max_waves_per_simd: Carrizo reports wrong max_waves_per_simd
2097 	 */
2098 	if (dev->gpu->adev->asic_type == CHIP_CARRIZO) {
2099 		dev->node_props.simd_count =
2100 			cu_info->simd_per_cu * cu_info->number;
2101 		dev->node_props.max_waves_per_simd = 10;
2102 	}
2103 
2104 	/* kfd only concerns sram ecc on GFX and HBM ecc on UMC */
2105 	dev->node_props.capability |=
2106 		((dev->gpu->adev->ras_enabled & BIT(AMDGPU_RAS_BLOCK__GFX)) != 0) ?
2107 		HSA_CAP_SRAM_EDCSUPPORTED : 0;
2108 	dev->node_props.capability |=
2109 		((dev->gpu->adev->ras_enabled & BIT(AMDGPU_RAS_BLOCK__UMC)) != 0) ?
2110 		HSA_CAP_MEM_EDCSUPPORTED : 0;
2111 
2112 	if (KFD_GC_VERSION(dev->gpu) != IP_VERSION(9, 0, 1))
2113 		dev->node_props.capability |= (dev->gpu->adev->ras_enabled != 0) ?
2114 			HSA_CAP_RASEVENTNOTIFY : 0;
2115 
2116 	if (KFD_IS_SVM_API_SUPPORTED(dev->gpu->adev))
2117 		dev->node_props.capability |= HSA_CAP_SVMAPI_SUPPORTED;
2118 
2119 	if (dev->gpu->adev->gmc.is_app_apu ||
2120 		dev->gpu->adev->gmc.xgmi.connected_to_cpu)
2121 		dev->node_props.capability |= HSA_CAP_FLAGS_COHERENTHOSTACCESS;
2122 
2123 	kfd_debug_print_topology();
2124 
2125 	kfd_notify_gpu_change(gpu_id, 1);
2126 
2127 	return 0;
2128 }
2129 
2130 /**
2131  * kfd_topology_update_io_links() - Update IO links after device removal.
2132  * @proximity_domain: Proximity domain value of the dev being removed.
2133  *
2134  * The topology list currently is arranged in increasing order of
2135  * proximity domain.
2136  *
2137  * Two things need to be done when a device is removed:
2138  * 1. All the IO links to this device need to be removed.
2139  * 2. All nodes after the current device node need to move
2140  *    up once this device node is removed from the topology
2141  *    list. As a result, the proximity domain values for
2142  *    all nodes after the node being deleted reduce by 1.
2143  *    This would also cause the proximity domain values for
2144  *    io links to be updated based on new proximity domain
2145  *    values.
2146  *
2147  * Context: The caller must hold write topology_lock.
2148  */
2149 static void kfd_topology_update_io_links(int proximity_domain)
2150 {
2151 	struct kfd_topology_device *dev;
2152 	struct kfd_iolink_properties *iolink, *p2plink, *tmp;
2153 
2154 	list_for_each_entry(dev, &topology_device_list, list) {
2155 		if (dev->proximity_domain > proximity_domain)
2156 			dev->proximity_domain--;
2157 
2158 		list_for_each_entry_safe(iolink, tmp, &dev->io_link_props, list) {
2159 			/*
2160 			 * If there is an io link to the dev being deleted
2161 			 * then remove that IO link also.
2162 			 */
2163 			if (iolink->node_to == proximity_domain) {
2164 				list_del(&iolink->list);
2165 				dev->node_props.io_links_count--;
2166 			} else {
2167 				if (iolink->node_from > proximity_domain)
2168 					iolink->node_from--;
2169 				if (iolink->node_to > proximity_domain)
2170 					iolink->node_to--;
2171 			}
2172 		}
2173 
2174 		list_for_each_entry_safe(p2plink, tmp, &dev->p2p_link_props, list) {
2175 			/*
2176 			 * If there is a p2p link to the dev being deleted
2177 			 * then remove that p2p link also.
2178 			 */
2179 			if (p2plink->node_to == proximity_domain) {
2180 				list_del(&p2plink->list);
2181 				dev->node_props.p2p_links_count--;
2182 			} else {
2183 				if (p2plink->node_from > proximity_domain)
2184 					p2plink->node_from--;
2185 				if (p2plink->node_to > proximity_domain)
2186 					p2plink->node_to--;
2187 			}
2188 		}
2189 	}
2190 }
2191 
2192 int kfd_topology_remove_device(struct kfd_node *gpu)
2193 {
2194 	struct kfd_topology_device *dev, *tmp;
2195 	uint32_t gpu_id;
2196 	int res = -ENODEV;
2197 	int i = 0;
2198 
2199 	down_write(&topology_lock);
2200 
2201 	list_for_each_entry_safe(dev, tmp, &topology_device_list, list) {
2202 		if (dev->gpu == gpu) {
2203 			gpu_id = dev->gpu_id;
2204 			kfd_remove_sysfs_node_entry(dev);
2205 			kfd_release_topology_device(dev);
2206 			sys_props.num_devices--;
2207 			kfd_topology_update_io_links(i);
2208 			topology_crat_proximity_domain = sys_props.num_devices-1;
2209 			sys_props.generation_count++;
2210 			res = 0;
2211 			if (kfd_topology_update_sysfs() < 0)
2212 				kfd_topology_release_sysfs();
2213 			break;
2214 		}
2215 		i++;
2216 	}
2217 
2218 	up_write(&topology_lock);
2219 
2220 	if (!res)
2221 		kfd_notify_gpu_change(gpu_id, 0);
2222 
2223 	return res;
2224 }
2225 
2226 /* kfd_topology_enum_kfd_devices - Enumerate through all devices in KFD
2227  *	topology. If GPU device is found @idx, then valid kfd_dev pointer is
2228  *	returned through @kdev
2229  * Return -	0: On success (@kdev will be NULL for non GPU nodes)
2230  *		-1: If end of list
2231  */
2232 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_node **kdev)
2233 {
2234 
2235 	struct kfd_topology_device *top_dev;
2236 	uint8_t device_idx = 0;
2237 
2238 	*kdev = NULL;
2239 	down_read(&topology_lock);
2240 
2241 	list_for_each_entry(top_dev, &topology_device_list, list) {
2242 		if (device_idx == idx) {
2243 			*kdev = top_dev->gpu;
2244 			up_read(&topology_lock);
2245 			return 0;
2246 		}
2247 
2248 		device_idx++;
2249 	}
2250 
2251 	up_read(&topology_lock);
2252 
2253 	return -1;
2254 
2255 }
2256 
2257 static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask)
2258 {
2259 	int first_cpu_of_numa_node;
2260 
2261 	if (!cpumask || cpumask == cpu_none_mask)
2262 		return -1;
2263 	first_cpu_of_numa_node = cpumask_first(cpumask);
2264 	if (first_cpu_of_numa_node >= nr_cpu_ids)
2265 		return -1;
2266 #ifdef CONFIG_X86_64
2267 	return cpu_data(first_cpu_of_numa_node).topo.apicid;
2268 #else
2269 	return first_cpu_of_numa_node;
2270 #endif
2271 }
2272 
2273 /* kfd_numa_node_to_apic_id - Returns the APIC ID of the first logical processor
2274  *	of the given NUMA node (numa_node_id)
2275  * Return -1 on failure
2276  */
2277 int kfd_numa_node_to_apic_id(int numa_node_id)
2278 {
2279 	if (numa_node_id == -1) {
2280 		pr_warn("Invalid NUMA Node. Use online CPU mask\n");
2281 		return kfd_cpumask_to_apic_id(cpu_online_mask);
2282 	}
2283 	return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id));
2284 }
2285 
2286 #if defined(CONFIG_DEBUG_FS)
2287 
2288 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data)
2289 {
2290 	struct kfd_topology_device *dev;
2291 	unsigned int i = 0;
2292 	int r = 0;
2293 
2294 	down_read(&topology_lock);
2295 
2296 	list_for_each_entry(dev, &topology_device_list, list) {
2297 		if (!dev->gpu) {
2298 			i++;
2299 			continue;
2300 		}
2301 
2302 		seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
2303 		r = dqm_debugfs_hqds(m, dev->gpu->dqm);
2304 		if (r)
2305 			break;
2306 	}
2307 
2308 	up_read(&topology_lock);
2309 
2310 	return r;
2311 }
2312 
2313 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data)
2314 {
2315 	struct kfd_topology_device *dev;
2316 	unsigned int i = 0;
2317 	int r = 0;
2318 
2319 	down_read(&topology_lock);
2320 
2321 	list_for_each_entry(dev, &topology_device_list, list) {
2322 		if (!dev->gpu) {
2323 			i++;
2324 			continue;
2325 		}
2326 
2327 		seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
2328 		r = pm_debugfs_runlist(m, &dev->gpu->dqm->packet_mgr);
2329 		if (r)
2330 			break;
2331 	}
2332 
2333 	up_read(&topology_lock);
2334 
2335 	return r;
2336 }
2337 
2338 #endif
2339