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 
35 #include "kfd_priv.h"
36 #include "kfd_crat.h"
37 #include "kfd_topology.h"
38 #include "kfd_device_queue_manager.h"
39 #include "kfd_iommu.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 *
455 					  NUM_XCC(dev->gpu->xcc_mask)) : 0);
456 	sysfs_show_32bit_prop(buffer, offs, "mem_banks_count",
457 			      dev->node_props.mem_banks_count);
458 	sysfs_show_32bit_prop(buffer, offs, "caches_count",
459 			      dev->node_props.caches_count);
460 	sysfs_show_32bit_prop(buffer, offs, "io_links_count",
461 			      dev->node_props.io_links_count);
462 	sysfs_show_32bit_prop(buffer, offs, "p2p_links_count",
463 			      dev->node_props.p2p_links_count);
464 	sysfs_show_32bit_prop(buffer, offs, "cpu_core_id_base",
465 			      dev->node_props.cpu_core_id_base);
466 	sysfs_show_32bit_prop(buffer, offs, "simd_id_base",
467 			      dev->node_props.simd_id_base);
468 	sysfs_show_32bit_prop(buffer, offs, "max_waves_per_simd",
469 			      dev->node_props.max_waves_per_simd);
470 	sysfs_show_32bit_prop(buffer, offs, "lds_size_in_kb",
471 			      dev->node_props.lds_size_in_kb);
472 	sysfs_show_32bit_prop(buffer, offs, "gds_size_in_kb",
473 			      dev->node_props.gds_size_in_kb);
474 	sysfs_show_32bit_prop(buffer, offs, "num_gws",
475 			      dev->node_props.num_gws);
476 	sysfs_show_32bit_prop(buffer, offs, "wave_front_size",
477 			      dev->node_props.wave_front_size);
478 	sysfs_show_32bit_prop(buffer, offs, "array_count",
479 			      dev->gpu ? (dev->node_props.array_count *
480 					  NUM_XCC(dev->gpu->xcc_mask)) : 0);
481 	sysfs_show_32bit_prop(buffer, offs, "simd_arrays_per_engine",
482 			      dev->node_props.simd_arrays_per_engine);
483 	sysfs_show_32bit_prop(buffer, offs, "cu_per_simd_array",
484 			      dev->node_props.cu_per_simd_array);
485 	sysfs_show_32bit_prop(buffer, offs, "simd_per_cu",
486 			      dev->node_props.simd_per_cu);
487 	sysfs_show_32bit_prop(buffer, offs, "max_slots_scratch_cu",
488 			      dev->node_props.max_slots_scratch_cu);
489 	sysfs_show_32bit_prop(buffer, offs, "gfx_target_version",
490 			      dev->node_props.gfx_target_version);
491 	sysfs_show_32bit_prop(buffer, offs, "vendor_id",
492 			      dev->node_props.vendor_id);
493 	sysfs_show_32bit_prop(buffer, offs, "device_id",
494 			      dev->node_props.device_id);
495 	sysfs_show_32bit_prop(buffer, offs, "location_id",
496 			      dev->node_props.location_id);
497 	sysfs_show_32bit_prop(buffer, offs, "domain",
498 			      dev->node_props.domain);
499 	sysfs_show_32bit_prop(buffer, offs, "drm_render_minor",
500 			      dev->node_props.drm_render_minor);
501 	sysfs_show_64bit_prop(buffer, offs, "hive_id",
502 			      dev->node_props.hive_id);
503 	sysfs_show_32bit_prop(buffer, offs, "num_sdma_engines",
504 			      dev->node_props.num_sdma_engines);
505 	sysfs_show_32bit_prop(buffer, offs, "num_sdma_xgmi_engines",
506 			      dev->node_props.num_sdma_xgmi_engines);
507 	sysfs_show_32bit_prop(buffer, offs, "num_sdma_queues_per_engine",
508 			      dev->node_props.num_sdma_queues_per_engine);
509 	sysfs_show_32bit_prop(buffer, offs, "num_cp_queues",
510 			      dev->node_props.num_cp_queues);
511 
512 	if (dev->gpu) {
513 		log_max_watch_addr =
514 			__ilog2_u32(dev->gpu->kfd->device_info.num_of_watch_points);
515 
516 		if (log_max_watch_addr) {
517 			dev->node_props.capability |=
518 					HSA_CAP_WATCH_POINTS_SUPPORTED;
519 
520 			dev->node_props.capability |=
521 				((log_max_watch_addr <<
522 					HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) &
523 				HSA_CAP_WATCH_POINTS_TOTALBITS_MASK);
524 		}
525 
526 		if (dev->gpu->adev->asic_type == CHIP_TONGA)
527 			dev->node_props.capability |=
528 					HSA_CAP_AQL_QUEUE_DOUBLE_MAP;
529 
530 		sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_fcompute",
531 			dev->node_props.max_engine_clk_fcompute);
532 
533 		sysfs_show_64bit_prop(buffer, offs, "local_mem_size", 0ULL);
534 
535 		sysfs_show_32bit_prop(buffer, offs, "fw_version",
536 				      dev->gpu->kfd->mec_fw_version);
537 		sysfs_show_32bit_prop(buffer, offs, "capability",
538 				      dev->node_props.capability);
539 		sysfs_show_64bit_prop(buffer, offs, "debug_prop",
540 				      dev->node_props.debug_prop);
541 		sysfs_show_32bit_prop(buffer, offs, "sdma_fw_version",
542 				      dev->gpu->kfd->sdma_fw_version);
543 		sysfs_show_64bit_prop(buffer, offs, "unique_id",
544 				      dev->gpu->adev->unique_id);
545 		sysfs_show_32bit_prop(buffer, offs, "num_xcc",
546 				      NUM_XCC(dev->gpu->xcc_mask));
547 	}
548 
549 	return sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_ccompute",
550 				     cpufreq_quick_get_max(0)/1000);
551 }
552 
553 static const struct sysfs_ops node_ops = {
554 	.show = node_show,
555 };
556 
557 static const struct kobj_type node_type = {
558 	.release = kfd_topology_kobj_release,
559 	.sysfs_ops = &node_ops,
560 };
561 
562 static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr)
563 {
564 	sysfs_remove_file(kobj, attr);
565 	kobject_del(kobj);
566 	kobject_put(kobj);
567 }
568 
569 static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev)
570 {
571 	struct kfd_iolink_properties *p2plink;
572 	struct kfd_iolink_properties *iolink;
573 	struct kfd_cache_properties *cache;
574 	struct kfd_mem_properties *mem;
575 	struct kfd_perf_properties *perf;
576 
577 	if (dev->kobj_iolink) {
578 		list_for_each_entry(iolink, &dev->io_link_props, list)
579 			if (iolink->kobj) {
580 				kfd_remove_sysfs_file(iolink->kobj,
581 							&iolink->attr);
582 				iolink->kobj = NULL;
583 			}
584 		kobject_del(dev->kobj_iolink);
585 		kobject_put(dev->kobj_iolink);
586 		dev->kobj_iolink = NULL;
587 	}
588 
589 	if (dev->kobj_p2plink) {
590 		list_for_each_entry(p2plink, &dev->p2p_link_props, list)
591 			if (p2plink->kobj) {
592 				kfd_remove_sysfs_file(p2plink->kobj,
593 							&p2plink->attr);
594 				p2plink->kobj = NULL;
595 			}
596 		kobject_del(dev->kobj_p2plink);
597 		kobject_put(dev->kobj_p2plink);
598 		dev->kobj_p2plink = NULL;
599 	}
600 
601 	if (dev->kobj_cache) {
602 		list_for_each_entry(cache, &dev->cache_props, list)
603 			if (cache->kobj) {
604 				kfd_remove_sysfs_file(cache->kobj,
605 							&cache->attr);
606 				cache->kobj = NULL;
607 			}
608 		kobject_del(dev->kobj_cache);
609 		kobject_put(dev->kobj_cache);
610 		dev->kobj_cache = NULL;
611 	}
612 
613 	if (dev->kobj_mem) {
614 		list_for_each_entry(mem, &dev->mem_props, list)
615 			if (mem->kobj) {
616 				kfd_remove_sysfs_file(mem->kobj, &mem->attr);
617 				mem->kobj = NULL;
618 			}
619 		kobject_del(dev->kobj_mem);
620 		kobject_put(dev->kobj_mem);
621 		dev->kobj_mem = NULL;
622 	}
623 
624 	if (dev->kobj_perf) {
625 		list_for_each_entry(perf, &dev->perf_props, list) {
626 			kfree(perf->attr_group);
627 			perf->attr_group = NULL;
628 		}
629 		kobject_del(dev->kobj_perf);
630 		kobject_put(dev->kobj_perf);
631 		dev->kobj_perf = NULL;
632 	}
633 
634 	if (dev->kobj_node) {
635 		sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid);
636 		sysfs_remove_file(dev->kobj_node, &dev->attr_name);
637 		sysfs_remove_file(dev->kobj_node, &dev->attr_props);
638 		kobject_del(dev->kobj_node);
639 		kobject_put(dev->kobj_node);
640 		dev->kobj_node = NULL;
641 	}
642 }
643 
644 static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev,
645 		uint32_t id)
646 {
647 	struct kfd_iolink_properties *p2plink;
648 	struct kfd_iolink_properties *iolink;
649 	struct kfd_cache_properties *cache;
650 	struct kfd_mem_properties *mem;
651 	struct kfd_perf_properties *perf;
652 	int ret;
653 	uint32_t i, num_attrs;
654 	struct attribute **attrs;
655 
656 	if (WARN_ON(dev->kobj_node))
657 		return -EEXIST;
658 
659 	/*
660 	 * Creating the sysfs folders
661 	 */
662 	dev->kobj_node = kfd_alloc_struct(dev->kobj_node);
663 	if (!dev->kobj_node)
664 		return -ENOMEM;
665 
666 	ret = kobject_init_and_add(dev->kobj_node, &node_type,
667 			sys_props.kobj_nodes, "%d", id);
668 	if (ret < 0) {
669 		kobject_put(dev->kobj_node);
670 		return ret;
671 	}
672 
673 	dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node);
674 	if (!dev->kobj_mem)
675 		return -ENOMEM;
676 
677 	dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node);
678 	if (!dev->kobj_cache)
679 		return -ENOMEM;
680 
681 	dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node);
682 	if (!dev->kobj_iolink)
683 		return -ENOMEM;
684 
685 	dev->kobj_p2plink = kobject_create_and_add("p2p_links", dev->kobj_node);
686 	if (!dev->kobj_p2plink)
687 		return -ENOMEM;
688 
689 	dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node);
690 	if (!dev->kobj_perf)
691 		return -ENOMEM;
692 
693 	/*
694 	 * Creating sysfs files for node properties
695 	 */
696 	dev->attr_gpuid.name = "gpu_id";
697 	dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE;
698 	sysfs_attr_init(&dev->attr_gpuid);
699 	dev->attr_name.name = "name";
700 	dev->attr_name.mode = KFD_SYSFS_FILE_MODE;
701 	sysfs_attr_init(&dev->attr_name);
702 	dev->attr_props.name = "properties";
703 	dev->attr_props.mode = KFD_SYSFS_FILE_MODE;
704 	sysfs_attr_init(&dev->attr_props);
705 	ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid);
706 	if (ret < 0)
707 		return ret;
708 	ret = sysfs_create_file(dev->kobj_node, &dev->attr_name);
709 	if (ret < 0)
710 		return ret;
711 	ret = sysfs_create_file(dev->kobj_node, &dev->attr_props);
712 	if (ret < 0)
713 		return ret;
714 
715 	i = 0;
716 	list_for_each_entry(mem, &dev->mem_props, list) {
717 		mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
718 		if (!mem->kobj)
719 			return -ENOMEM;
720 		ret = kobject_init_and_add(mem->kobj, &mem_type,
721 				dev->kobj_mem, "%d", i);
722 		if (ret < 0) {
723 			kobject_put(mem->kobj);
724 			return ret;
725 		}
726 
727 		mem->attr.name = "properties";
728 		mem->attr.mode = KFD_SYSFS_FILE_MODE;
729 		sysfs_attr_init(&mem->attr);
730 		ret = sysfs_create_file(mem->kobj, &mem->attr);
731 		if (ret < 0)
732 			return ret;
733 		i++;
734 	}
735 
736 	i = 0;
737 	list_for_each_entry(cache, &dev->cache_props, list) {
738 		cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
739 		if (!cache->kobj)
740 			return -ENOMEM;
741 		ret = kobject_init_and_add(cache->kobj, &cache_type,
742 				dev->kobj_cache, "%d", i);
743 		if (ret < 0) {
744 			kobject_put(cache->kobj);
745 			return ret;
746 		}
747 
748 		cache->attr.name = "properties";
749 		cache->attr.mode = KFD_SYSFS_FILE_MODE;
750 		sysfs_attr_init(&cache->attr);
751 		ret = sysfs_create_file(cache->kobj, &cache->attr);
752 		if (ret < 0)
753 			return ret;
754 		i++;
755 	}
756 
757 	i = 0;
758 	list_for_each_entry(iolink, &dev->io_link_props, list) {
759 		iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
760 		if (!iolink->kobj)
761 			return -ENOMEM;
762 		ret = kobject_init_and_add(iolink->kobj, &iolink_type,
763 				dev->kobj_iolink, "%d", i);
764 		if (ret < 0) {
765 			kobject_put(iolink->kobj);
766 			return ret;
767 		}
768 
769 		iolink->attr.name = "properties";
770 		iolink->attr.mode = KFD_SYSFS_FILE_MODE;
771 		sysfs_attr_init(&iolink->attr);
772 		ret = sysfs_create_file(iolink->kobj, &iolink->attr);
773 		if (ret < 0)
774 			return ret;
775 		i++;
776 	}
777 
778 	i = 0;
779 	list_for_each_entry(p2plink, &dev->p2p_link_props, list) {
780 		p2plink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
781 		if (!p2plink->kobj)
782 			return -ENOMEM;
783 		ret = kobject_init_and_add(p2plink->kobj, &iolink_type,
784 				dev->kobj_p2plink, "%d", i);
785 		if (ret < 0) {
786 			kobject_put(p2plink->kobj);
787 			return ret;
788 		}
789 
790 		p2plink->attr.name = "properties";
791 		p2plink->attr.mode = KFD_SYSFS_FILE_MODE;
792 		sysfs_attr_init(&p2plink->attr);
793 		ret = sysfs_create_file(p2plink->kobj, &p2plink->attr);
794 		if (ret < 0)
795 			return ret;
796 		i++;
797 	}
798 
799 	/* All hardware blocks have the same number of attributes. */
800 	num_attrs = ARRAY_SIZE(perf_attr_iommu);
801 	list_for_each_entry(perf, &dev->perf_props, list) {
802 		perf->attr_group = kzalloc(sizeof(struct kfd_perf_attr)
803 			* num_attrs + sizeof(struct attribute_group),
804 			GFP_KERNEL);
805 		if (!perf->attr_group)
806 			return -ENOMEM;
807 
808 		attrs = (struct attribute **)(perf->attr_group + 1);
809 		if (!strcmp(perf->block_name, "iommu")) {
810 		/* Information of IOMMU's num_counters and counter_ids is shown
811 		 * under /sys/bus/event_source/devices/amd_iommu. We don't
812 		 * duplicate here.
813 		 */
814 			perf_attr_iommu[0].data = perf->max_concurrent;
815 			for (i = 0; i < num_attrs; i++)
816 				attrs[i] = &perf_attr_iommu[i].attr.attr;
817 		}
818 		perf->attr_group->name = perf->block_name;
819 		perf->attr_group->attrs = attrs;
820 		ret = sysfs_create_group(dev->kobj_perf, perf->attr_group);
821 		if (ret < 0)
822 			return ret;
823 	}
824 
825 	return 0;
826 }
827 
828 /* Called with write topology lock acquired */
829 static int kfd_build_sysfs_node_tree(void)
830 {
831 	struct kfd_topology_device *dev;
832 	int ret;
833 	uint32_t i = 0;
834 
835 	list_for_each_entry(dev, &topology_device_list, list) {
836 		ret = kfd_build_sysfs_node_entry(dev, i);
837 		if (ret < 0)
838 			return ret;
839 		i++;
840 	}
841 
842 	return 0;
843 }
844 
845 /* Called with write topology lock acquired */
846 static void kfd_remove_sysfs_node_tree(void)
847 {
848 	struct kfd_topology_device *dev;
849 
850 	list_for_each_entry(dev, &topology_device_list, list)
851 		kfd_remove_sysfs_node_entry(dev);
852 }
853 
854 static int kfd_topology_update_sysfs(void)
855 {
856 	int ret;
857 
858 	if (!sys_props.kobj_topology) {
859 		sys_props.kobj_topology =
860 				kfd_alloc_struct(sys_props.kobj_topology);
861 		if (!sys_props.kobj_topology)
862 			return -ENOMEM;
863 
864 		ret = kobject_init_and_add(sys_props.kobj_topology,
865 				&sysprops_type,  &kfd_device->kobj,
866 				"topology");
867 		if (ret < 0) {
868 			kobject_put(sys_props.kobj_topology);
869 			return ret;
870 		}
871 
872 		sys_props.kobj_nodes = kobject_create_and_add("nodes",
873 				sys_props.kobj_topology);
874 		if (!sys_props.kobj_nodes)
875 			return -ENOMEM;
876 
877 		sys_props.attr_genid.name = "generation_id";
878 		sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE;
879 		sysfs_attr_init(&sys_props.attr_genid);
880 		ret = sysfs_create_file(sys_props.kobj_topology,
881 				&sys_props.attr_genid);
882 		if (ret < 0)
883 			return ret;
884 
885 		sys_props.attr_props.name = "system_properties";
886 		sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE;
887 		sysfs_attr_init(&sys_props.attr_props);
888 		ret = sysfs_create_file(sys_props.kobj_topology,
889 				&sys_props.attr_props);
890 		if (ret < 0)
891 			return ret;
892 	}
893 
894 	kfd_remove_sysfs_node_tree();
895 
896 	return kfd_build_sysfs_node_tree();
897 }
898 
899 static void kfd_topology_release_sysfs(void)
900 {
901 	kfd_remove_sysfs_node_tree();
902 	if (sys_props.kobj_topology) {
903 		sysfs_remove_file(sys_props.kobj_topology,
904 				&sys_props.attr_genid);
905 		sysfs_remove_file(sys_props.kobj_topology,
906 				&sys_props.attr_props);
907 		if (sys_props.kobj_nodes) {
908 			kobject_del(sys_props.kobj_nodes);
909 			kobject_put(sys_props.kobj_nodes);
910 			sys_props.kobj_nodes = NULL;
911 		}
912 		kobject_del(sys_props.kobj_topology);
913 		kobject_put(sys_props.kobj_topology);
914 		sys_props.kobj_topology = NULL;
915 	}
916 }
917 
918 /* Called with write topology_lock acquired */
919 static void kfd_topology_update_device_list(struct list_head *temp_list,
920 					struct list_head *master_list)
921 {
922 	while (!list_empty(temp_list)) {
923 		list_move_tail(temp_list->next, master_list);
924 		sys_props.num_devices++;
925 	}
926 }
927 
928 static void kfd_debug_print_topology(void)
929 {
930 	struct kfd_topology_device *dev;
931 
932 	down_read(&topology_lock);
933 
934 	dev = list_last_entry(&topology_device_list,
935 			struct kfd_topology_device, list);
936 	if (dev) {
937 		if (dev->node_props.cpu_cores_count &&
938 				dev->node_props.simd_count) {
939 			pr_info("Topology: Add APU node [0x%0x:0x%0x]\n",
940 				dev->node_props.device_id,
941 				dev->node_props.vendor_id);
942 		} else if (dev->node_props.cpu_cores_count)
943 			pr_info("Topology: Add CPU node\n");
944 		else if (dev->node_props.simd_count)
945 			pr_info("Topology: Add dGPU node [0x%0x:0x%0x]\n",
946 				dev->node_props.device_id,
947 				dev->node_props.vendor_id);
948 	}
949 	up_read(&topology_lock);
950 }
951 
952 /* Helper function for intializing platform_xx members of
953  * kfd_system_properties. Uses OEM info from the last CPU/APU node.
954  */
955 static void kfd_update_system_properties(void)
956 {
957 	struct kfd_topology_device *dev;
958 
959 	down_read(&topology_lock);
960 	dev = list_last_entry(&topology_device_list,
961 			struct kfd_topology_device, list);
962 	if (dev) {
963 		sys_props.platform_id =
964 			(*((uint64_t *)dev->oem_id)) & CRAT_OEMID_64BIT_MASK;
965 		sys_props.platform_oem = *((uint64_t *)dev->oem_table_id);
966 		sys_props.platform_rev = dev->oem_revision;
967 	}
968 	up_read(&topology_lock);
969 }
970 
971 static void find_system_memory(const struct dmi_header *dm,
972 	void *private)
973 {
974 	struct kfd_mem_properties *mem;
975 	u16 mem_width, mem_clock;
976 	struct kfd_topology_device *kdev =
977 		(struct kfd_topology_device *)private;
978 	const u8 *dmi_data = (const u8 *)(dm + 1);
979 
980 	if (dm->type == DMI_ENTRY_MEM_DEVICE && dm->length >= 0x15) {
981 		mem_width = (u16)(*(const u16 *)(dmi_data + 0x6));
982 		mem_clock = (u16)(*(const u16 *)(dmi_data + 0x11));
983 		list_for_each_entry(mem, &kdev->mem_props, list) {
984 			if (mem_width != 0xFFFF && mem_width != 0)
985 				mem->width = mem_width;
986 			if (mem_clock != 0)
987 				mem->mem_clk_max = mem_clock;
988 		}
989 	}
990 }
991 
992 /*
993  * Performance counters information is not part of CRAT but we would like to
994  * put them in the sysfs under topology directory for Thunk to get the data.
995  * This function is called before updating the sysfs.
996  */
997 static int kfd_add_perf_to_topology(struct kfd_topology_device *kdev)
998 {
999 	/* These are the only counters supported so far */
1000 	return kfd_iommu_add_perf_counters(kdev);
1001 }
1002 
1003 /* kfd_add_non_crat_information - Add information that is not currently
1004  *	defined in CRAT but is necessary for KFD topology
1005  * @dev - topology device to which addition info is added
1006  */
1007 static void kfd_add_non_crat_information(struct kfd_topology_device *kdev)
1008 {
1009 	/* Check if CPU only node. */
1010 	if (!kdev->gpu) {
1011 		/* Add system memory information */
1012 		dmi_walk(find_system_memory, kdev);
1013 	}
1014 	/* TODO: For GPU node, rearrange code from kfd_topology_add_device */
1015 }
1016 
1017 /* kfd_is_acpi_crat_invalid - CRAT from ACPI is valid only for AMD APU devices.
1018  *	Ignore CRAT for all other devices. AMD APU is identified if both CPU
1019  *	and GPU cores are present.
1020  * @device_list - topology device list created by parsing ACPI CRAT table.
1021  * @return - TRUE if invalid, FALSE is valid.
1022  */
1023 static bool kfd_is_acpi_crat_invalid(struct list_head *device_list)
1024 {
1025 	struct kfd_topology_device *dev;
1026 
1027 	list_for_each_entry(dev, device_list, list) {
1028 		if (dev->node_props.cpu_cores_count &&
1029 			dev->node_props.simd_count)
1030 			return false;
1031 	}
1032 	pr_info("Ignoring ACPI CRAT on non-APU system\n");
1033 	return true;
1034 }
1035 
1036 int kfd_topology_init(void)
1037 {
1038 	void *crat_image = NULL;
1039 	size_t image_size = 0;
1040 	int ret;
1041 	struct list_head temp_topology_device_list;
1042 	int cpu_only_node = 0;
1043 	struct kfd_topology_device *kdev;
1044 	int proximity_domain;
1045 
1046 	/* topology_device_list - Master list of all topology devices
1047 	 * temp_topology_device_list - temporary list created while parsing CRAT
1048 	 * or VCRAT. Once parsing is complete the contents of list is moved to
1049 	 * topology_device_list
1050 	 */
1051 
1052 	/* Initialize the head for the both the lists */
1053 	INIT_LIST_HEAD(&topology_device_list);
1054 	INIT_LIST_HEAD(&temp_topology_device_list);
1055 	init_rwsem(&topology_lock);
1056 
1057 	memset(&sys_props, 0, sizeof(sys_props));
1058 
1059 	/* Proximity domains in ACPI CRAT tables start counting at
1060 	 * 0. The same should be true for virtual CRAT tables created
1061 	 * at this stage. GPUs added later in kfd_topology_add_device
1062 	 * use a counter.
1063 	 */
1064 	proximity_domain = 0;
1065 
1066 	/*
1067 	 * Get the CRAT image from the ACPI. If ACPI doesn't have one
1068 	 * or if ACPI CRAT is invalid create a virtual CRAT.
1069 	 * NOTE: The current implementation expects all AMD APUs to have
1070 	 *	CRAT. If no CRAT is available, it is assumed to be a CPU
1071 	 */
1072 	ret = kfd_create_crat_image_acpi(&crat_image, &image_size);
1073 	if (!ret) {
1074 		ret = kfd_parse_crat_table(crat_image,
1075 					   &temp_topology_device_list,
1076 					   proximity_domain);
1077 		if (ret ||
1078 		    kfd_is_acpi_crat_invalid(&temp_topology_device_list)) {
1079 			kfd_release_topology_device_list(
1080 				&temp_topology_device_list);
1081 			kfd_destroy_crat_image(crat_image);
1082 			crat_image = NULL;
1083 		}
1084 	}
1085 
1086 	if (!crat_image) {
1087 		ret = kfd_create_crat_image_virtual(&crat_image, &image_size,
1088 						    COMPUTE_UNIT_CPU, NULL,
1089 						    proximity_domain);
1090 		cpu_only_node = 1;
1091 		if (ret) {
1092 			pr_err("Error creating VCRAT table for CPU\n");
1093 			return ret;
1094 		}
1095 
1096 		ret = kfd_parse_crat_table(crat_image,
1097 					   &temp_topology_device_list,
1098 					   proximity_domain);
1099 		if (ret) {
1100 			pr_err("Error parsing VCRAT table for CPU\n");
1101 			goto err;
1102 		}
1103 	}
1104 
1105 	kdev = list_first_entry(&temp_topology_device_list,
1106 				struct kfd_topology_device, list);
1107 	kfd_add_perf_to_topology(kdev);
1108 
1109 	down_write(&topology_lock);
1110 	kfd_topology_update_device_list(&temp_topology_device_list,
1111 					&topology_device_list);
1112 	topology_crat_proximity_domain = sys_props.num_devices-1;
1113 	ret = kfd_topology_update_sysfs();
1114 	up_write(&topology_lock);
1115 
1116 	if (!ret) {
1117 		sys_props.generation_count++;
1118 		kfd_update_system_properties();
1119 		kfd_debug_print_topology();
1120 	} else
1121 		pr_err("Failed to update topology in sysfs ret=%d\n", ret);
1122 
1123 	/* For nodes with GPU, this information gets added
1124 	 * when GPU is detected (kfd_topology_add_device).
1125 	 */
1126 	if (cpu_only_node) {
1127 		/* Add additional information to CPU only node created above */
1128 		down_write(&topology_lock);
1129 		kdev = list_first_entry(&topology_device_list,
1130 				struct kfd_topology_device, list);
1131 		up_write(&topology_lock);
1132 		kfd_add_non_crat_information(kdev);
1133 	}
1134 
1135 err:
1136 	kfd_destroy_crat_image(crat_image);
1137 	return ret;
1138 }
1139 
1140 void kfd_topology_shutdown(void)
1141 {
1142 	down_write(&topology_lock);
1143 	kfd_topology_release_sysfs();
1144 	kfd_release_live_view();
1145 	up_write(&topology_lock);
1146 }
1147 
1148 static uint32_t kfd_generate_gpu_id(struct kfd_node *gpu)
1149 {
1150 	uint32_t hashout;
1151 	uint32_t buf[8];
1152 	uint64_t local_mem_size;
1153 	int i;
1154 
1155 	if (!gpu)
1156 		return 0;
1157 
1158 	local_mem_size = gpu->local_mem_info.local_mem_size_private +
1159 			gpu->local_mem_info.local_mem_size_public;
1160 	buf[0] = gpu->adev->pdev->devfn;
1161 	buf[1] = gpu->adev->pdev->subsystem_vendor |
1162 		(gpu->adev->pdev->subsystem_device << 16);
1163 	buf[2] = pci_domain_nr(gpu->adev->pdev->bus);
1164 	buf[3] = gpu->adev->pdev->device;
1165 	buf[4] = gpu->adev->pdev->bus->number;
1166 	buf[5] = lower_32_bits(local_mem_size);
1167 	buf[6] = upper_32_bits(local_mem_size);
1168 	buf[7] = (ffs(gpu->xcc_mask) - 1) | (NUM_XCC(gpu->xcc_mask) << 16);
1169 
1170 	for (i = 0, hashout = 0; i < 8; i++)
1171 		hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH);
1172 
1173 	return hashout;
1174 }
1175 /* kfd_assign_gpu - Attach @gpu to the correct kfd topology device. If
1176  *		the GPU device is not already present in the topology device
1177  *		list then return NULL. This means a new topology device has to
1178  *		be created for this GPU.
1179  */
1180 static struct kfd_topology_device *kfd_assign_gpu(struct kfd_node *gpu)
1181 {
1182 	struct kfd_topology_device *dev;
1183 	struct kfd_topology_device *out_dev = NULL;
1184 	struct kfd_mem_properties *mem;
1185 	struct kfd_cache_properties *cache;
1186 	struct kfd_iolink_properties *iolink;
1187 	struct kfd_iolink_properties *p2plink;
1188 
1189 	list_for_each_entry(dev, &topology_device_list, list) {
1190 		/* Discrete GPUs need their own topology device list
1191 		 * entries. Don't assign them to CPU/APU nodes.
1192 		 */
1193 		if (!gpu->kfd->use_iommu_v2 &&
1194 		    dev->node_props.cpu_cores_count)
1195 			continue;
1196 
1197 		if (!dev->gpu && (dev->node_props.simd_count > 0)) {
1198 			dev->gpu = gpu;
1199 			out_dev = dev;
1200 
1201 			list_for_each_entry(mem, &dev->mem_props, list)
1202 				mem->gpu = dev->gpu;
1203 			list_for_each_entry(cache, &dev->cache_props, list)
1204 				cache->gpu = dev->gpu;
1205 			list_for_each_entry(iolink, &dev->io_link_props, list)
1206 				iolink->gpu = dev->gpu;
1207 			list_for_each_entry(p2plink, &dev->p2p_link_props, list)
1208 				p2plink->gpu = dev->gpu;
1209 			break;
1210 		}
1211 	}
1212 	return out_dev;
1213 }
1214 
1215 static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival)
1216 {
1217 	/*
1218 	 * TODO: Generate an event for thunk about the arrival/removal
1219 	 * of the GPU
1220 	 */
1221 }
1222 
1223 /* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info,
1224  *		patch this after CRAT parsing.
1225  */
1226 static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev)
1227 {
1228 	struct kfd_mem_properties *mem;
1229 	struct kfd_local_mem_info local_mem_info;
1230 
1231 	if (!dev)
1232 		return;
1233 
1234 	/* Currently, amdgpu driver (amdgpu_mc) deals only with GPUs with
1235 	 * single bank of VRAM local memory.
1236 	 * for dGPUs - VCRAT reports only one bank of Local Memory
1237 	 * for APUs - If CRAT from ACPI reports more than one bank, then
1238 	 *	all the banks will report the same mem_clk_max information
1239 	 */
1240 	amdgpu_amdkfd_get_local_mem_info(dev->gpu->adev, &local_mem_info,
1241 					 dev->gpu->xcp);
1242 
1243 	list_for_each_entry(mem, &dev->mem_props, list)
1244 		mem->mem_clk_max = local_mem_info.mem_clk_max;
1245 }
1246 
1247 static void kfd_set_iolink_no_atomics(struct kfd_topology_device *dev,
1248 					struct kfd_topology_device *target_gpu_dev,
1249 					struct kfd_iolink_properties *link)
1250 {
1251 	/* xgmi always supports atomics between links. */
1252 	if (link->iolink_type == CRAT_IOLINK_TYPE_XGMI)
1253 		return;
1254 
1255 	/* check pcie support to set cpu(dev) flags for target_gpu_dev link. */
1256 	if (target_gpu_dev) {
1257 		uint32_t cap;
1258 
1259 		pcie_capability_read_dword(target_gpu_dev->gpu->adev->pdev,
1260 				PCI_EXP_DEVCAP2, &cap);
1261 
1262 		if (!(cap & (PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
1263 			     PCI_EXP_DEVCAP2_ATOMIC_COMP64)))
1264 			link->flags |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1265 				CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1266 	/* set gpu (dev) flags. */
1267 	} else {
1268 		if (!dev->gpu->kfd->pci_atomic_requested ||
1269 				dev->gpu->adev->asic_type == CHIP_HAWAII)
1270 			link->flags |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1271 				CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1272 	}
1273 }
1274 
1275 static void kfd_set_iolink_non_coherent(struct kfd_topology_device *to_dev,
1276 		struct kfd_iolink_properties *outbound_link,
1277 		struct kfd_iolink_properties *inbound_link)
1278 {
1279 	/* CPU -> GPU with PCIe */
1280 	if (!to_dev->gpu &&
1281 	    inbound_link->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS)
1282 		inbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT;
1283 
1284 	if (to_dev->gpu) {
1285 		/* GPU <-> GPU with PCIe and
1286 		 * Vega20 with XGMI
1287 		 */
1288 		if (inbound_link->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS ||
1289 		    (inbound_link->iolink_type == CRAT_IOLINK_TYPE_XGMI &&
1290 		    KFD_GC_VERSION(to_dev->gpu) == IP_VERSION(9, 4, 0))) {
1291 			outbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT;
1292 			inbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT;
1293 		}
1294 	}
1295 }
1296 
1297 static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev)
1298 {
1299 	struct kfd_iolink_properties *link, *inbound_link;
1300 	struct kfd_topology_device *peer_dev;
1301 
1302 	if (!dev || !dev->gpu)
1303 		return;
1304 
1305 	/* GPU only creates direct links so apply flags setting to all */
1306 	list_for_each_entry(link, &dev->io_link_props, list) {
1307 		link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1308 		kfd_set_iolink_no_atomics(dev, NULL, link);
1309 		peer_dev = kfd_topology_device_by_proximity_domain(
1310 				link->node_to);
1311 
1312 		if (!peer_dev)
1313 			continue;
1314 
1315 		/* Include the CPU peer in GPU hive if connected over xGMI. */
1316 		if (!peer_dev->gpu &&
1317 		    link->iolink_type == CRAT_IOLINK_TYPE_XGMI) {
1318 			/*
1319 			 * If the GPU is not part of a GPU hive, use its pci
1320 			 * device location as the hive ID to bind with the CPU.
1321 			 */
1322 			if (!dev->node_props.hive_id)
1323 				dev->node_props.hive_id = pci_dev_id(dev->gpu->adev->pdev);
1324 			peer_dev->node_props.hive_id = dev->node_props.hive_id;
1325 		}
1326 
1327 		list_for_each_entry(inbound_link, &peer_dev->io_link_props,
1328 									list) {
1329 			if (inbound_link->node_to != link->node_from)
1330 				continue;
1331 
1332 			inbound_link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1333 			kfd_set_iolink_no_atomics(peer_dev, dev, inbound_link);
1334 			kfd_set_iolink_non_coherent(peer_dev, link, inbound_link);
1335 		}
1336 	}
1337 
1338 	/* Create indirect links so apply flags setting to all */
1339 	list_for_each_entry(link, &dev->p2p_link_props, list) {
1340 		link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1341 		kfd_set_iolink_no_atomics(dev, NULL, link);
1342 		peer_dev = kfd_topology_device_by_proximity_domain(
1343 				link->node_to);
1344 
1345 		if (!peer_dev)
1346 			continue;
1347 
1348 		list_for_each_entry(inbound_link, &peer_dev->p2p_link_props,
1349 									list) {
1350 			if (inbound_link->node_to != link->node_from)
1351 				continue;
1352 
1353 			inbound_link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1354 			kfd_set_iolink_no_atomics(peer_dev, dev, inbound_link);
1355 			kfd_set_iolink_non_coherent(peer_dev, link, inbound_link);
1356 		}
1357 	}
1358 }
1359 
1360 static int kfd_build_p2p_node_entry(struct kfd_topology_device *dev,
1361 				struct kfd_iolink_properties *p2plink)
1362 {
1363 	int ret;
1364 
1365 	p2plink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
1366 	if (!p2plink->kobj)
1367 		return -ENOMEM;
1368 
1369 	ret = kobject_init_and_add(p2plink->kobj, &iolink_type,
1370 			dev->kobj_p2plink, "%d", dev->node_props.p2p_links_count - 1);
1371 	if (ret < 0) {
1372 		kobject_put(p2plink->kobj);
1373 		return ret;
1374 	}
1375 
1376 	p2plink->attr.name = "properties";
1377 	p2plink->attr.mode = KFD_SYSFS_FILE_MODE;
1378 	sysfs_attr_init(&p2plink->attr);
1379 	ret = sysfs_create_file(p2plink->kobj, &p2plink->attr);
1380 	if (ret < 0)
1381 		return ret;
1382 
1383 	return 0;
1384 }
1385 
1386 static int kfd_create_indirect_link_prop(struct kfd_topology_device *kdev, int gpu_node)
1387 {
1388 	struct kfd_iolink_properties *gpu_link, *tmp_link, *cpu_link;
1389 	struct kfd_iolink_properties *props = NULL, *props2 = NULL;
1390 	struct kfd_topology_device *cpu_dev;
1391 	int ret = 0;
1392 	int i, num_cpu;
1393 
1394 	num_cpu = 0;
1395 	list_for_each_entry(cpu_dev, &topology_device_list, list) {
1396 		if (cpu_dev->gpu)
1397 			break;
1398 		num_cpu++;
1399 	}
1400 
1401 	gpu_link = list_first_entry(&kdev->io_link_props,
1402 					struct kfd_iolink_properties, list);
1403 	if (!gpu_link)
1404 		return -ENOMEM;
1405 
1406 	for (i = 0; i < num_cpu; i++) {
1407 		/* CPU <--> GPU */
1408 		if (gpu_link->node_to == i)
1409 			continue;
1410 
1411 		/* find CPU <-->  CPU links */
1412 		cpu_link = NULL;
1413 		cpu_dev = kfd_topology_device_by_proximity_domain(i);
1414 		if (cpu_dev) {
1415 			list_for_each_entry(tmp_link,
1416 					&cpu_dev->io_link_props, list) {
1417 				if (tmp_link->node_to == gpu_link->node_to) {
1418 					cpu_link = tmp_link;
1419 					break;
1420 				}
1421 			}
1422 		}
1423 
1424 		if (!cpu_link)
1425 			return -ENOMEM;
1426 
1427 		/* CPU <--> CPU <--> GPU, GPU node*/
1428 		props = kfd_alloc_struct(props);
1429 		if (!props)
1430 			return -ENOMEM;
1431 
1432 		memcpy(props, gpu_link, sizeof(struct kfd_iolink_properties));
1433 		props->weight = gpu_link->weight + cpu_link->weight;
1434 		props->min_latency = gpu_link->min_latency + cpu_link->min_latency;
1435 		props->max_latency = gpu_link->max_latency + cpu_link->max_latency;
1436 		props->min_bandwidth = min(gpu_link->min_bandwidth, cpu_link->min_bandwidth);
1437 		props->max_bandwidth = min(gpu_link->max_bandwidth, cpu_link->max_bandwidth);
1438 
1439 		props->node_from = gpu_node;
1440 		props->node_to = i;
1441 		kdev->node_props.p2p_links_count++;
1442 		list_add_tail(&props->list, &kdev->p2p_link_props);
1443 		ret = kfd_build_p2p_node_entry(kdev, props);
1444 		if (ret < 0)
1445 			return ret;
1446 
1447 		/* for small Bar, no CPU --> GPU in-direct links */
1448 		if (kfd_dev_is_large_bar(kdev->gpu)) {
1449 			/* CPU <--> CPU <--> GPU, CPU node*/
1450 			props2 = kfd_alloc_struct(props2);
1451 			if (!props2)
1452 				return -ENOMEM;
1453 
1454 			memcpy(props2, props, sizeof(struct kfd_iolink_properties));
1455 			props2->node_from = i;
1456 			props2->node_to = gpu_node;
1457 			props2->kobj = NULL;
1458 			cpu_dev->node_props.p2p_links_count++;
1459 			list_add_tail(&props2->list, &cpu_dev->p2p_link_props);
1460 			ret = kfd_build_p2p_node_entry(cpu_dev, props2);
1461 			if (ret < 0)
1462 				return ret;
1463 		}
1464 	}
1465 	return ret;
1466 }
1467 
1468 #if defined(CONFIG_HSA_AMD_P2P)
1469 static int kfd_add_peer_prop(struct kfd_topology_device *kdev,
1470 		struct kfd_topology_device *peer, int from, int to)
1471 {
1472 	struct kfd_iolink_properties *props = NULL;
1473 	struct kfd_iolink_properties *iolink1, *iolink2, *iolink3;
1474 	struct kfd_topology_device *cpu_dev;
1475 	int ret = 0;
1476 
1477 	if (!amdgpu_device_is_peer_accessible(
1478 				kdev->gpu->adev,
1479 				peer->gpu->adev))
1480 		return ret;
1481 
1482 	iolink1 = list_first_entry(&kdev->io_link_props,
1483 							struct kfd_iolink_properties, list);
1484 	if (!iolink1)
1485 		return -ENOMEM;
1486 
1487 	iolink2 = list_first_entry(&peer->io_link_props,
1488 							struct kfd_iolink_properties, list);
1489 	if (!iolink2)
1490 		return -ENOMEM;
1491 
1492 	props = kfd_alloc_struct(props);
1493 	if (!props)
1494 		return -ENOMEM;
1495 
1496 	memcpy(props, iolink1, sizeof(struct kfd_iolink_properties));
1497 
1498 	props->weight = iolink1->weight + iolink2->weight;
1499 	props->min_latency = iolink1->min_latency + iolink2->min_latency;
1500 	props->max_latency = iolink1->max_latency + iolink2->max_latency;
1501 	props->min_bandwidth = min(iolink1->min_bandwidth, iolink2->min_bandwidth);
1502 	props->max_bandwidth = min(iolink2->max_bandwidth, iolink2->max_bandwidth);
1503 
1504 	if (iolink1->node_to != iolink2->node_to) {
1505 		/* CPU->CPU  link*/
1506 		cpu_dev = kfd_topology_device_by_proximity_domain(iolink1->node_to);
1507 		if (cpu_dev) {
1508 			list_for_each_entry(iolink3, &cpu_dev->io_link_props, list)
1509 				if (iolink3->node_to == iolink2->node_to)
1510 					break;
1511 
1512 			props->weight += iolink3->weight;
1513 			props->min_latency += iolink3->min_latency;
1514 			props->max_latency += iolink3->max_latency;
1515 			props->min_bandwidth = min(props->min_bandwidth,
1516 							iolink3->min_bandwidth);
1517 			props->max_bandwidth = min(props->max_bandwidth,
1518 							iolink3->max_bandwidth);
1519 		} else {
1520 			WARN(1, "CPU node not found");
1521 		}
1522 	}
1523 
1524 	props->node_from = from;
1525 	props->node_to = to;
1526 	peer->node_props.p2p_links_count++;
1527 	list_add_tail(&props->list, &peer->p2p_link_props);
1528 	ret = kfd_build_p2p_node_entry(peer, props);
1529 
1530 	return ret;
1531 }
1532 #endif
1533 
1534 static int kfd_dev_create_p2p_links(void)
1535 {
1536 	struct kfd_topology_device *dev;
1537 	struct kfd_topology_device *new_dev;
1538 #if defined(CONFIG_HSA_AMD_P2P)
1539 	uint32_t i;
1540 #endif
1541 	uint32_t k;
1542 	int ret = 0;
1543 
1544 	k = 0;
1545 	list_for_each_entry(dev, &topology_device_list, list)
1546 		k++;
1547 	if (k < 2)
1548 		return 0;
1549 
1550 	new_dev = list_last_entry(&topology_device_list, struct kfd_topology_device, list);
1551 	if (WARN_ON(!new_dev->gpu))
1552 		return 0;
1553 
1554 	k--;
1555 
1556 	/* create in-direct links */
1557 	ret = kfd_create_indirect_link_prop(new_dev, k);
1558 	if (ret < 0)
1559 		goto out;
1560 
1561 	/* create p2p links */
1562 #if defined(CONFIG_HSA_AMD_P2P)
1563 	i = 0;
1564 	list_for_each_entry(dev, &topology_device_list, list) {
1565 		if (dev == new_dev)
1566 			break;
1567 		if (!dev->gpu || !dev->gpu->adev ||
1568 		    (dev->gpu->kfd->hive_id &&
1569 		     dev->gpu->kfd->hive_id == new_dev->gpu->kfd->hive_id))
1570 			goto next;
1571 
1572 		/* check if node(s) is/are peer accessible in one direction or bi-direction */
1573 		ret = kfd_add_peer_prop(new_dev, dev, i, k);
1574 		if (ret < 0)
1575 			goto out;
1576 
1577 		ret = kfd_add_peer_prop(dev, new_dev, k, i);
1578 		if (ret < 0)
1579 			goto out;
1580 next:
1581 		i++;
1582 	}
1583 #endif
1584 
1585 out:
1586 	return ret;
1587 }
1588 
1589 /* Helper function. See kfd_fill_gpu_cache_info for parameter description */
1590 static int fill_in_l1_pcache(struct kfd_cache_properties **props_ext,
1591 				struct kfd_gpu_cache_info *pcache_info,
1592 				struct kfd_cu_info *cu_info,
1593 				int cu_bitmask,
1594 				int cache_type, unsigned int cu_processor_id,
1595 				int cu_block)
1596 {
1597 	unsigned int cu_sibling_map_mask;
1598 	int first_active_cu;
1599 	struct kfd_cache_properties *pcache = NULL;
1600 
1601 	cu_sibling_map_mask = cu_bitmask;
1602 	cu_sibling_map_mask >>= cu_block;
1603 	cu_sibling_map_mask &= ((1 << pcache_info[cache_type].num_cu_shared) - 1);
1604 	first_active_cu = ffs(cu_sibling_map_mask);
1605 
1606 	/* CU could be inactive. In case of shared cache find the first active
1607 	 * CU. and incase of non-shared cache check if the CU is inactive. If
1608 	 * inactive active skip it
1609 	 */
1610 	if (first_active_cu) {
1611 		pcache = kfd_alloc_struct(pcache);
1612 		if (!pcache)
1613 			return -ENOMEM;
1614 
1615 		memset(pcache, 0, sizeof(struct kfd_cache_properties));
1616 		pcache->processor_id_low = cu_processor_id + (first_active_cu - 1);
1617 		pcache->cache_level = pcache_info[cache_type].cache_level;
1618 		pcache->cache_size = pcache_info[cache_type].cache_size;
1619 
1620 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_DATA_CACHE)
1621 			pcache->cache_type |= HSA_CACHE_TYPE_DATA;
1622 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_INST_CACHE)
1623 			pcache->cache_type |= HSA_CACHE_TYPE_INSTRUCTION;
1624 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_CPU_CACHE)
1625 			pcache->cache_type |= HSA_CACHE_TYPE_CPU;
1626 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_SIMD_CACHE)
1627 			pcache->cache_type |= HSA_CACHE_TYPE_HSACU;
1628 
1629 		/* Sibling map is w.r.t processor_id_low, so shift out
1630 		 * inactive CU
1631 		 */
1632 		cu_sibling_map_mask =
1633 			cu_sibling_map_mask >> (first_active_cu - 1);
1634 
1635 		pcache->sibling_map[0] = (uint8_t)(cu_sibling_map_mask & 0xFF);
1636 		pcache->sibling_map[1] =
1637 				(uint8_t)((cu_sibling_map_mask >> 8) & 0xFF);
1638 		pcache->sibling_map[2] =
1639 				(uint8_t)((cu_sibling_map_mask >> 16) & 0xFF);
1640 		pcache->sibling_map[3] =
1641 				(uint8_t)((cu_sibling_map_mask >> 24) & 0xFF);
1642 
1643 		pcache->sibling_map_size = 4;
1644 		*props_ext = pcache;
1645 
1646 		return 0;
1647 	}
1648 	return 1;
1649 }
1650 
1651 /* Helper function. See kfd_fill_gpu_cache_info for parameter description */
1652 static int fill_in_l2_l3_pcache(struct kfd_cache_properties **props_ext,
1653 				struct kfd_gpu_cache_info *pcache_info,
1654 				struct kfd_cu_info *cu_info,
1655 				int cache_type, unsigned int cu_processor_id)
1656 {
1657 	unsigned int cu_sibling_map_mask;
1658 	int first_active_cu;
1659 	int i, j, k;
1660 	struct kfd_cache_properties *pcache = NULL;
1661 
1662 	cu_sibling_map_mask = cu_info->cu_bitmap[0][0];
1663 	cu_sibling_map_mask &=
1664 		((1 << pcache_info[cache_type].num_cu_shared) - 1);
1665 	first_active_cu = ffs(cu_sibling_map_mask);
1666 
1667 	/* CU could be inactive. In case of shared cache find the first active
1668 	 * CU. and incase of non-shared cache check if the CU is inactive. If
1669 	 * inactive active skip it
1670 	 */
1671 	if (first_active_cu) {
1672 		pcache = kfd_alloc_struct(pcache);
1673 		if (!pcache)
1674 			return -ENOMEM;
1675 
1676 		memset(pcache, 0, sizeof(struct kfd_cache_properties));
1677 		pcache->processor_id_low = cu_processor_id
1678 					+ (first_active_cu - 1);
1679 		pcache->cache_level = pcache_info[cache_type].cache_level;
1680 		pcache->cache_size = pcache_info[cache_type].cache_size;
1681 
1682 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_DATA_CACHE)
1683 			pcache->cache_type |= HSA_CACHE_TYPE_DATA;
1684 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_INST_CACHE)
1685 			pcache->cache_type |= HSA_CACHE_TYPE_INSTRUCTION;
1686 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_CPU_CACHE)
1687 			pcache->cache_type |= HSA_CACHE_TYPE_CPU;
1688 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_SIMD_CACHE)
1689 			pcache->cache_type |= HSA_CACHE_TYPE_HSACU;
1690 
1691 		/* Sibling map is w.r.t processor_id_low, so shift out
1692 		 * inactive CU
1693 		 */
1694 		cu_sibling_map_mask = cu_sibling_map_mask >> (first_active_cu - 1);
1695 		k = 0;
1696 
1697 		for (i = 0; i < cu_info->num_shader_engines; i++) {
1698 			for (j = 0; j < cu_info->num_shader_arrays_per_engine; j++) {
1699 				pcache->sibling_map[k] = (uint8_t)(cu_sibling_map_mask & 0xFF);
1700 				pcache->sibling_map[k+1] = (uint8_t)((cu_sibling_map_mask >> 8) & 0xFF);
1701 				pcache->sibling_map[k+2] = (uint8_t)((cu_sibling_map_mask >> 16) & 0xFF);
1702 				pcache->sibling_map[k+3] = (uint8_t)((cu_sibling_map_mask >> 24) & 0xFF);
1703 				k += 4;
1704 
1705 				cu_sibling_map_mask = cu_info->cu_bitmap[i % 4][j + i / 4];
1706 				cu_sibling_map_mask &= ((1 << pcache_info[cache_type].num_cu_shared) - 1);
1707 			}
1708 		}
1709 		pcache->sibling_map_size = k;
1710 		*props_ext = pcache;
1711 		return 0;
1712 	}
1713 	return 1;
1714 }
1715 
1716 #define KFD_MAX_CACHE_TYPES 6
1717 
1718 /* kfd_fill_cache_non_crat_info - Fill GPU cache info using kfd_gpu_cache_info
1719  * tables
1720  */
1721 static void kfd_fill_cache_non_crat_info(struct kfd_topology_device *dev, struct kfd_node *kdev)
1722 {
1723 	struct kfd_gpu_cache_info *pcache_info = NULL;
1724 	int i, j, k;
1725 	int ct = 0;
1726 	unsigned int cu_processor_id;
1727 	int ret;
1728 	unsigned int num_cu_shared;
1729 	struct kfd_cu_info cu_info;
1730 	struct kfd_cu_info *pcu_info;
1731 	int gpu_processor_id;
1732 	struct kfd_cache_properties *props_ext;
1733 	int num_of_entries = 0;
1734 	int num_of_cache_types = 0;
1735 	struct kfd_gpu_cache_info cache_info[KFD_MAX_CACHE_TYPES];
1736 
1737 	amdgpu_amdkfd_get_cu_info(kdev->adev, &cu_info);
1738 	pcu_info = &cu_info;
1739 
1740 	gpu_processor_id = dev->node_props.simd_id_base;
1741 
1742 	pcache_info = cache_info;
1743 	num_of_cache_types = kfd_get_gpu_cache_info(kdev, &pcache_info);
1744 	if (!num_of_cache_types) {
1745 		pr_warn("no cache info found\n");
1746 		return;
1747 	}
1748 
1749 	/* For each type of cache listed in the kfd_gpu_cache_info table,
1750 	 * go through all available Compute Units.
1751 	 * The [i,j,k] loop will
1752 	 *		if kfd_gpu_cache_info.num_cu_shared = 1
1753 	 *			will parse through all available CU
1754 	 *		If (kfd_gpu_cache_info.num_cu_shared != 1)
1755 	 *			then it will consider only one CU from
1756 	 *			the shared unit
1757 	 */
1758 	for (ct = 0; ct < num_of_cache_types; ct++) {
1759 		cu_processor_id = gpu_processor_id;
1760 		if (pcache_info[ct].cache_level == 1) {
1761 			for (i = 0; i < pcu_info->num_shader_engines; i++) {
1762 				for (j = 0; j < pcu_info->num_shader_arrays_per_engine; j++) {
1763 					for (k = 0; k < pcu_info->num_cu_per_sh; k += pcache_info[ct].num_cu_shared) {
1764 
1765 						ret = fill_in_l1_pcache(&props_ext, pcache_info, pcu_info,
1766 										pcu_info->cu_bitmap[i % 4][j + i / 4], ct,
1767 										cu_processor_id, k);
1768 
1769 						if (ret < 0)
1770 							break;
1771 
1772 						if (!ret) {
1773 							num_of_entries++;
1774 							list_add_tail(&props_ext->list, &dev->cache_props);
1775 						}
1776 
1777 						/* Move to next CU block */
1778 						num_cu_shared = ((k + pcache_info[ct].num_cu_shared) <=
1779 							pcu_info->num_cu_per_sh) ?
1780 							pcache_info[ct].num_cu_shared :
1781 							(pcu_info->num_cu_per_sh - k);
1782 						cu_processor_id += num_cu_shared;
1783 					}
1784 				}
1785 			}
1786 		} else {
1787 			ret = fill_in_l2_l3_pcache(&props_ext, pcache_info,
1788 								pcu_info, ct, cu_processor_id);
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, uint32_t gpu_id,
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 		pr_err("Error creating VCRAT for GPU (ID: 0x%x)\n",
1817 		       gpu_id);
1818 		topology_crat_proximity_domain--;
1819 		goto err;
1820 	}
1821 
1822 	INIT_LIST_HEAD(&temp_topology_device_list);
1823 
1824 	res = kfd_parse_crat_table(crat_image,
1825 				   &temp_topology_device_list,
1826 				   proximity_domain);
1827 	if (res) {
1828 		pr_err("Error parsing VCRAT for GPU (ID: 0x%x)\n",
1829 		       gpu_id);
1830 		topology_crat_proximity_domain--;
1831 		goto err;
1832 	}
1833 
1834 	kfd_topology_update_device_list(&temp_topology_device_list,
1835 					&topology_device_list);
1836 
1837 	*dev = kfd_assign_gpu(gpu);
1838 	if (WARN_ON(!*dev)) {
1839 		res = -ENODEV;
1840 		goto err;
1841 	}
1842 
1843 	/* Fill the cache affinity information here for the GPUs
1844 	 * using VCRAT
1845 	 */
1846 	kfd_fill_cache_non_crat_info(*dev, gpu);
1847 
1848 	/* Update the SYSFS tree, since we added another topology
1849 	 * device
1850 	 */
1851 	res = kfd_topology_update_sysfs();
1852 	if (!res)
1853 		sys_props.generation_count++;
1854 	else
1855 		pr_err("Failed to update GPU (ID: 0x%x) to sysfs topology. res=%d\n",
1856 		       gpu_id, res);
1857 
1858 err:
1859 	kfd_destroy_crat_image(crat_image);
1860 	return res;
1861 }
1862 
1863 static void kfd_topology_set_dbg_firmware_support(struct kfd_topology_device *dev)
1864 {
1865 	bool firmware_supported = true;
1866 
1867 	if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(11, 0, 0) &&
1868 			KFD_GC_VERSION(dev->gpu) < IP_VERSION(12, 0, 0)) {
1869 		uint32_t mes_api_rev = (dev->gpu->adev->mes.sched_version &
1870 						AMDGPU_MES_API_VERSION_MASK) >>
1871 						AMDGPU_MES_API_VERSION_SHIFT;
1872 		uint32_t mes_rev = dev->gpu->adev->mes.sched_version &
1873 						AMDGPU_MES_VERSION_MASK;
1874 
1875 		firmware_supported = (mes_api_rev >= 14) && (mes_rev >= 64);
1876 		goto out;
1877 	}
1878 
1879 	/*
1880 	 * Note: Any unlisted devices here are assumed to support exception handling.
1881 	 * Add additional checks here as needed.
1882 	 */
1883 	switch (KFD_GC_VERSION(dev->gpu)) {
1884 	case IP_VERSION(9, 0, 1):
1885 		firmware_supported = dev->gpu->kfd->mec_fw_version >= 459 + 32768;
1886 		break;
1887 	case IP_VERSION(9, 1, 0):
1888 	case IP_VERSION(9, 2, 1):
1889 	case IP_VERSION(9, 2, 2):
1890 	case IP_VERSION(9, 3, 0):
1891 	case IP_VERSION(9, 4, 0):
1892 		firmware_supported = dev->gpu->kfd->mec_fw_version >= 459;
1893 		break;
1894 	case IP_VERSION(9, 4, 1):
1895 		firmware_supported = dev->gpu->kfd->mec_fw_version >= 60;
1896 		break;
1897 	case IP_VERSION(9, 4, 2):
1898 		firmware_supported = dev->gpu->kfd->mec_fw_version >= 51;
1899 		break;
1900 	case IP_VERSION(10, 1, 10):
1901 	case IP_VERSION(10, 1, 2):
1902 	case IP_VERSION(10, 1, 1):
1903 		firmware_supported = dev->gpu->kfd->mec_fw_version >= 144;
1904 		break;
1905 	case IP_VERSION(10, 3, 0):
1906 	case IP_VERSION(10, 3, 2):
1907 	case IP_VERSION(10, 3, 1):
1908 	case IP_VERSION(10, 3, 4):
1909 	case IP_VERSION(10, 3, 5):
1910 		firmware_supported = dev->gpu->kfd->mec_fw_version >= 89;
1911 		break;
1912 	case IP_VERSION(10, 1, 3):
1913 	case IP_VERSION(10, 3, 3):
1914 		firmware_supported = false;
1915 		break;
1916 	default:
1917 		break;
1918 	}
1919 
1920 out:
1921 	if (firmware_supported)
1922 		dev->node_props.capability |= HSA_CAP_TRAP_DEBUG_FIRMWARE_SUPPORTED;
1923 }
1924 
1925 static void kfd_topology_set_capabilities(struct kfd_topology_device *dev)
1926 {
1927 	dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 <<
1928 				HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1929 				HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1930 
1931 	dev->node_props.capability |= HSA_CAP_TRAP_DEBUG_SUPPORT |
1932 			HSA_CAP_TRAP_DEBUG_WAVE_LAUNCH_TRAP_OVERRIDE_SUPPORTED |
1933 			HSA_CAP_TRAP_DEBUG_WAVE_LAUNCH_MODE_SUPPORTED;
1934 
1935 	if (kfd_dbg_has_ttmps_always_setup(dev->gpu))
1936 		dev->node_props.debug_prop |= HSA_DBG_DISPATCH_INFO_ALWAYS_VALID;
1937 
1938 	if (KFD_GC_VERSION(dev->gpu) < IP_VERSION(10, 0, 0)) {
1939 		if (KFD_GC_VERSION(dev->gpu) == IP_VERSION(9, 4, 3))
1940 			dev->node_props.debug_prop |=
1941 				HSA_DBG_WATCH_ADDR_MASK_LO_BIT_GFX9_4_3 |
1942 				HSA_DBG_WATCH_ADDR_MASK_HI_BIT_GFX9_4_3;
1943 		else
1944 			dev->node_props.debug_prop |=
1945 				HSA_DBG_WATCH_ADDR_MASK_LO_BIT_GFX9 |
1946 				HSA_DBG_WATCH_ADDR_MASK_HI_BIT;
1947 
1948 		if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(9, 4, 2))
1949 			dev->node_props.capability |=
1950 				HSA_CAP_TRAP_DEBUG_PRECISE_MEMORY_OPERATIONS_SUPPORTED;
1951 	} else {
1952 		dev->node_props.debug_prop |= HSA_DBG_WATCH_ADDR_MASK_LO_BIT_GFX10 |
1953 					HSA_DBG_WATCH_ADDR_MASK_HI_BIT;
1954 
1955 		if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(11, 0, 0))
1956 			dev->node_props.capability |=
1957 				HSA_CAP_TRAP_DEBUG_PRECISE_MEMORY_OPERATIONS_SUPPORTED;
1958 	}
1959 
1960 	kfd_topology_set_dbg_firmware_support(dev);
1961 }
1962 
1963 int kfd_topology_add_device(struct kfd_node *gpu)
1964 {
1965 	uint32_t gpu_id;
1966 	struct kfd_topology_device *dev;
1967 	struct kfd_cu_info cu_info;
1968 	int res = 0;
1969 	int i;
1970 	const char *asic_name = amdgpu_asic_name[gpu->adev->asic_type];
1971 
1972 	gpu_id = kfd_generate_gpu_id(gpu);
1973 	pr_debug("Adding new GPU (ID: 0x%x) to topology\n", gpu_id);
1974 
1975 	/* Check to see if this gpu device exists in the topology_device_list.
1976 	 * If so, assign the gpu to that device,
1977 	 * else create a Virtual CRAT for this gpu device and then parse that
1978 	 * CRAT to create a new topology device. Once created assign the gpu to
1979 	 * that topology device
1980 	 */
1981 	down_write(&topology_lock);
1982 	dev = kfd_assign_gpu(gpu);
1983 	if (!dev)
1984 		res = kfd_topology_add_device_locked(gpu, gpu_id, &dev);
1985 	up_write(&topology_lock);
1986 	if (res)
1987 		return res;
1988 
1989 	dev->gpu_id = gpu_id;
1990 	gpu->id = gpu_id;
1991 
1992 	kfd_dev_create_p2p_links();
1993 
1994 	/* TODO: Move the following lines to function
1995 	 *	kfd_add_non_crat_information
1996 	 */
1997 
1998 	/* Fill-in additional information that is not available in CRAT but
1999 	 * needed for the topology
2000 	 */
2001 
2002 	amdgpu_amdkfd_get_cu_info(dev->gpu->adev, &cu_info);
2003 
2004 	for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1; i++) {
2005 		dev->node_props.name[i] = __tolower(asic_name[i]);
2006 		if (asic_name[i] == '\0')
2007 			break;
2008 	}
2009 	dev->node_props.name[i] = '\0';
2010 
2011 	dev->node_props.simd_arrays_per_engine =
2012 		cu_info.num_shader_arrays_per_engine;
2013 
2014 	dev->node_props.gfx_target_version =
2015 				gpu->kfd->device_info.gfx_target_version;
2016 	dev->node_props.vendor_id = gpu->adev->pdev->vendor;
2017 	dev->node_props.device_id = gpu->adev->pdev->device;
2018 	dev->node_props.capability |=
2019 		((dev->gpu->adev->rev_id << HSA_CAP_ASIC_REVISION_SHIFT) &
2020 			HSA_CAP_ASIC_REVISION_MASK);
2021 
2022 	dev->node_props.location_id = pci_dev_id(gpu->adev->pdev);
2023 	if (KFD_GC_VERSION(dev->gpu->kfd) == IP_VERSION(9, 4, 3))
2024 		dev->node_props.location_id |= dev->gpu->node_id;
2025 
2026 	dev->node_props.domain = pci_domain_nr(gpu->adev->pdev->bus);
2027 	dev->node_props.max_engine_clk_fcompute =
2028 		amdgpu_amdkfd_get_max_engine_clock_in_mhz(dev->gpu->adev);
2029 	dev->node_props.max_engine_clk_ccompute =
2030 		cpufreq_quick_get_max(0) / 1000;
2031 
2032 	if (gpu->xcp)
2033 		dev->node_props.drm_render_minor = gpu->xcp->ddev->render->index;
2034 	else
2035 		dev->node_props.drm_render_minor =
2036 				gpu->kfd->shared_resources.drm_render_minor;
2037 
2038 	dev->node_props.hive_id = gpu->kfd->hive_id;
2039 	dev->node_props.num_sdma_engines = kfd_get_num_sdma_engines(gpu);
2040 	dev->node_props.num_sdma_xgmi_engines =
2041 					kfd_get_num_xgmi_sdma_engines(gpu);
2042 	dev->node_props.num_sdma_queues_per_engine =
2043 				gpu->kfd->device_info.num_sdma_queues_per_engine -
2044 				gpu->kfd->device_info.num_reserved_sdma_queues_per_engine;
2045 	dev->node_props.num_gws = (dev->gpu->gws &&
2046 		dev->gpu->dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) ?
2047 		dev->gpu->adev->gds.gws_size : 0;
2048 	dev->node_props.num_cp_queues = get_cp_queues_num(dev->gpu->dqm);
2049 
2050 	kfd_fill_mem_clk_max_info(dev);
2051 	kfd_fill_iolink_non_crat_info(dev);
2052 
2053 	switch (dev->gpu->adev->asic_type) {
2054 	case CHIP_KAVERI:
2055 	case CHIP_HAWAII:
2056 	case CHIP_TONGA:
2057 		dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_PRE_1_0 <<
2058 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
2059 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
2060 		break;
2061 	case CHIP_CARRIZO:
2062 	case CHIP_FIJI:
2063 	case CHIP_POLARIS10:
2064 	case CHIP_POLARIS11:
2065 	case CHIP_POLARIS12:
2066 	case CHIP_VEGAM:
2067 		pr_debug("Adding doorbell packet type capability\n");
2068 		dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 <<
2069 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
2070 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
2071 		break;
2072 	default:
2073 		if (KFD_GC_VERSION(dev->gpu) < IP_VERSION(9, 0, 1))
2074 			WARN(1, "Unexpected ASIC family %u",
2075 			     dev->gpu->adev->asic_type);
2076 		else
2077 			kfd_topology_set_capabilities(dev);
2078 	}
2079 
2080 	/*
2081 	 * Overwrite ATS capability according to needs_iommu_device to fix
2082 	 * potential missing corresponding bit in CRAT of BIOS.
2083 	 */
2084 	if (dev->gpu->kfd->use_iommu_v2)
2085 		dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
2086 	else
2087 		dev->node_props.capability &= ~HSA_CAP_ATS_PRESENT;
2088 
2089 	/* Fix errors in CZ CRAT.
2090 	 * simd_count: Carrizo CRAT reports wrong simd_count, probably
2091 	 *		because it doesn't consider masked out CUs
2092 	 * max_waves_per_simd: Carrizo reports wrong max_waves_per_simd
2093 	 */
2094 	if (dev->gpu->adev->asic_type == CHIP_CARRIZO) {
2095 		dev->node_props.simd_count =
2096 			cu_info.simd_per_cu * cu_info.cu_active_number;
2097 		dev->node_props.max_waves_per_simd = 10;
2098 	}
2099 
2100 	/* kfd only concerns sram ecc on GFX and HBM ecc on UMC */
2101 	dev->node_props.capability |=
2102 		((dev->gpu->adev->ras_enabled & BIT(AMDGPU_RAS_BLOCK__GFX)) != 0) ?
2103 		HSA_CAP_SRAM_EDCSUPPORTED : 0;
2104 	dev->node_props.capability |=
2105 		((dev->gpu->adev->ras_enabled & BIT(AMDGPU_RAS_BLOCK__UMC)) != 0) ?
2106 		HSA_CAP_MEM_EDCSUPPORTED : 0;
2107 
2108 	if (KFD_GC_VERSION(dev->gpu) != IP_VERSION(9, 0, 1))
2109 		dev->node_props.capability |= (dev->gpu->adev->ras_enabled != 0) ?
2110 			HSA_CAP_RASEVENTNOTIFY : 0;
2111 
2112 	if (KFD_IS_SVM_API_SUPPORTED(dev->gpu->adev))
2113 		dev->node_props.capability |= HSA_CAP_SVMAPI_SUPPORTED;
2114 
2115 	if (dev->gpu->adev->gmc.is_app_apu ||
2116 		dev->gpu->adev->gmc.xgmi.connected_to_cpu)
2117 		dev->node_props.capability |= HSA_CAP_FLAGS_COHERENTHOSTACCESS;
2118 
2119 	kfd_debug_print_topology();
2120 
2121 	kfd_notify_gpu_change(gpu_id, 1);
2122 
2123 	return 0;
2124 }
2125 
2126 /**
2127  * kfd_topology_update_io_links() - Update IO links after device removal.
2128  * @proximity_domain: Proximity domain value of the dev being removed.
2129  *
2130  * The topology list currently is arranged in increasing order of
2131  * proximity domain.
2132  *
2133  * Two things need to be done when a device is removed:
2134  * 1. All the IO links to this device need to be removed.
2135  * 2. All nodes after the current device node need to move
2136  *    up once this device node is removed from the topology
2137  *    list. As a result, the proximity domain values for
2138  *    all nodes after the node being deleted reduce by 1.
2139  *    This would also cause the proximity domain values for
2140  *    io links to be updated based on new proximity domain
2141  *    values.
2142  *
2143  * Context: The caller must hold write topology_lock.
2144  */
2145 static void kfd_topology_update_io_links(int proximity_domain)
2146 {
2147 	struct kfd_topology_device *dev;
2148 	struct kfd_iolink_properties *iolink, *p2plink, *tmp;
2149 
2150 	list_for_each_entry(dev, &topology_device_list, list) {
2151 		if (dev->proximity_domain > proximity_domain)
2152 			dev->proximity_domain--;
2153 
2154 		list_for_each_entry_safe(iolink, tmp, &dev->io_link_props, list) {
2155 			/*
2156 			 * If there is an io link to the dev being deleted
2157 			 * then remove that IO link also.
2158 			 */
2159 			if (iolink->node_to == proximity_domain) {
2160 				list_del(&iolink->list);
2161 				dev->node_props.io_links_count--;
2162 			} else {
2163 				if (iolink->node_from > proximity_domain)
2164 					iolink->node_from--;
2165 				if (iolink->node_to > proximity_domain)
2166 					iolink->node_to--;
2167 			}
2168 		}
2169 
2170 		list_for_each_entry_safe(p2plink, tmp, &dev->p2p_link_props, list) {
2171 			/*
2172 			 * If there is a p2p link to the dev being deleted
2173 			 * then remove that p2p link also.
2174 			 */
2175 			if (p2plink->node_to == proximity_domain) {
2176 				list_del(&p2plink->list);
2177 				dev->node_props.p2p_links_count--;
2178 			} else {
2179 				if (p2plink->node_from > proximity_domain)
2180 					p2plink->node_from--;
2181 				if (p2plink->node_to > proximity_domain)
2182 					p2plink->node_to--;
2183 			}
2184 		}
2185 	}
2186 }
2187 
2188 int kfd_topology_remove_device(struct kfd_node *gpu)
2189 {
2190 	struct kfd_topology_device *dev, *tmp;
2191 	uint32_t gpu_id;
2192 	int res = -ENODEV;
2193 	int i = 0;
2194 
2195 	down_write(&topology_lock);
2196 
2197 	list_for_each_entry_safe(dev, tmp, &topology_device_list, list) {
2198 		if (dev->gpu == gpu) {
2199 			gpu_id = dev->gpu_id;
2200 			kfd_remove_sysfs_node_entry(dev);
2201 			kfd_release_topology_device(dev);
2202 			sys_props.num_devices--;
2203 			kfd_topology_update_io_links(i);
2204 			topology_crat_proximity_domain = sys_props.num_devices-1;
2205 			sys_props.generation_count++;
2206 			res = 0;
2207 			if (kfd_topology_update_sysfs() < 0)
2208 				kfd_topology_release_sysfs();
2209 			break;
2210 		}
2211 		i++;
2212 	}
2213 
2214 	up_write(&topology_lock);
2215 
2216 	if (!res)
2217 		kfd_notify_gpu_change(gpu_id, 0);
2218 
2219 	return res;
2220 }
2221 
2222 /* kfd_topology_enum_kfd_devices - Enumerate through all devices in KFD
2223  *	topology. If GPU device is found @idx, then valid kfd_dev pointer is
2224  *	returned through @kdev
2225  * Return -	0: On success (@kdev will be NULL for non GPU nodes)
2226  *		-1: If end of list
2227  */
2228 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_node **kdev)
2229 {
2230 
2231 	struct kfd_topology_device *top_dev;
2232 	uint8_t device_idx = 0;
2233 
2234 	*kdev = NULL;
2235 	down_read(&topology_lock);
2236 
2237 	list_for_each_entry(top_dev, &topology_device_list, list) {
2238 		if (device_idx == idx) {
2239 			*kdev = top_dev->gpu;
2240 			up_read(&topology_lock);
2241 			return 0;
2242 		}
2243 
2244 		device_idx++;
2245 	}
2246 
2247 	up_read(&topology_lock);
2248 
2249 	return -1;
2250 
2251 }
2252 
2253 static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask)
2254 {
2255 	int first_cpu_of_numa_node;
2256 
2257 	if (!cpumask || cpumask == cpu_none_mask)
2258 		return -1;
2259 	first_cpu_of_numa_node = cpumask_first(cpumask);
2260 	if (first_cpu_of_numa_node >= nr_cpu_ids)
2261 		return -1;
2262 #ifdef CONFIG_X86_64
2263 	return cpu_data(first_cpu_of_numa_node).apicid;
2264 #else
2265 	return first_cpu_of_numa_node;
2266 #endif
2267 }
2268 
2269 /* kfd_numa_node_to_apic_id - Returns the APIC ID of the first logical processor
2270  *	of the given NUMA node (numa_node_id)
2271  * Return -1 on failure
2272  */
2273 int kfd_numa_node_to_apic_id(int numa_node_id)
2274 {
2275 	if (numa_node_id == -1) {
2276 		pr_warn("Invalid NUMA Node. Use online CPU mask\n");
2277 		return kfd_cpumask_to_apic_id(cpu_online_mask);
2278 	}
2279 	return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id));
2280 }
2281 
2282 void kfd_double_confirm_iommu_support(struct kfd_dev *gpu)
2283 {
2284 	struct kfd_topology_device *dev;
2285 
2286 	gpu->use_iommu_v2 = false;
2287 
2288 	if (!gpu->device_info.needs_iommu_device)
2289 		return;
2290 
2291 	down_read(&topology_lock);
2292 
2293 	/* Only use IOMMUv2 if there is an APU topology node with no GPU
2294 	 * assigned yet. This GPU will be assigned to it.
2295 	 */
2296 	list_for_each_entry(dev, &topology_device_list, list)
2297 		if (dev->node_props.cpu_cores_count &&
2298 		    dev->node_props.simd_count &&
2299 		    !dev->gpu)
2300 			gpu->use_iommu_v2 = true;
2301 
2302 	up_read(&topology_lock);
2303 }
2304 
2305 #if defined(CONFIG_DEBUG_FS)
2306 
2307 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data)
2308 {
2309 	struct kfd_topology_device *dev;
2310 	unsigned int i = 0;
2311 	int r = 0;
2312 
2313 	down_read(&topology_lock);
2314 
2315 	list_for_each_entry(dev, &topology_device_list, list) {
2316 		if (!dev->gpu) {
2317 			i++;
2318 			continue;
2319 		}
2320 
2321 		seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
2322 		r = dqm_debugfs_hqds(m, dev->gpu->dqm);
2323 		if (r)
2324 			break;
2325 	}
2326 
2327 	up_read(&topology_lock);
2328 
2329 	return r;
2330 }
2331 
2332 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data)
2333 {
2334 	struct kfd_topology_device *dev;
2335 	unsigned int i = 0;
2336 	int r = 0;
2337 
2338 	down_read(&topology_lock);
2339 
2340 	list_for_each_entry(dev, &topology_device_list, list) {
2341 		if (!dev->gpu) {
2342 			i++;
2343 			continue;
2344 		}
2345 
2346 		seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
2347 		r = pm_debugfs_runlist(m, &dev->gpu->dqm->packet_mgr);
2348 		if (r)
2349 			break;
2350 	}
2351 
2352 	up_read(&topology_lock);
2353 
2354 	return r;
2355 }
2356 
2357 #endif
2358