xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_topology.c (revision bab2c80e5a6c855657482eac9e97f5f3eedb509a)
1 /*
2  * Copyright 2014 Advanced Micro Devices, Inc.
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20  * OTHER DEALINGS IN THE SOFTWARE.
21  */
22 
23 #include <linux/types.h>
24 #include <linux/kernel.h>
25 #include <linux/pci.h>
26 #include <linux/errno.h>
27 #include <linux/acpi.h>
28 #include <linux/hash.h>
29 #include <linux/cpufreq.h>
30 #include <linux/log2.h>
31 #include <linux/dmi.h>
32 #include <linux/atomic.h>
33 
34 #include "kfd_priv.h"
35 #include "kfd_crat.h"
36 #include "kfd_topology.h"
37 #include "kfd_device_queue_manager.h"
38 #include "kfd_iommu.h"
39 
40 /* topology_device_list - Master list of all topology devices */
41 static struct list_head topology_device_list;
42 static struct kfd_system_properties sys_props;
43 
44 static DECLARE_RWSEM(topology_lock);
45 static atomic_t topology_crat_proximity_domain;
46 
47 struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
48 						uint32_t proximity_domain)
49 {
50 	struct kfd_topology_device *top_dev;
51 	struct kfd_topology_device *device = NULL;
52 
53 	down_read(&topology_lock);
54 
55 	list_for_each_entry(top_dev, &topology_device_list, list)
56 		if (top_dev->proximity_domain == proximity_domain) {
57 			device = top_dev;
58 			break;
59 		}
60 
61 	up_read(&topology_lock);
62 
63 	return device;
64 }
65 
66 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id)
67 {
68 	struct kfd_topology_device *top_dev;
69 	struct kfd_dev *device = NULL;
70 
71 	down_read(&topology_lock);
72 
73 	list_for_each_entry(top_dev, &topology_device_list, list)
74 		if (top_dev->gpu_id == gpu_id) {
75 			device = top_dev->gpu;
76 			break;
77 		}
78 
79 	up_read(&topology_lock);
80 
81 	return device;
82 }
83 
84 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev)
85 {
86 	struct kfd_topology_device *top_dev;
87 	struct kfd_dev *device = NULL;
88 
89 	down_read(&topology_lock);
90 
91 	list_for_each_entry(top_dev, &topology_device_list, list)
92 		if (top_dev->gpu->pdev == pdev) {
93 			device = top_dev->gpu;
94 			break;
95 		}
96 
97 	up_read(&topology_lock);
98 
99 	return device;
100 }
101 
102 /* Called with write topology_lock acquired */
103 static void kfd_release_topology_device(struct kfd_topology_device *dev)
104 {
105 	struct kfd_mem_properties *mem;
106 	struct kfd_cache_properties *cache;
107 	struct kfd_iolink_properties *iolink;
108 	struct kfd_perf_properties *perf;
109 
110 	list_del(&dev->list);
111 
112 	while (dev->mem_props.next != &dev->mem_props) {
113 		mem = container_of(dev->mem_props.next,
114 				struct kfd_mem_properties, list);
115 		list_del(&mem->list);
116 		kfree(mem);
117 	}
118 
119 	while (dev->cache_props.next != &dev->cache_props) {
120 		cache = container_of(dev->cache_props.next,
121 				struct kfd_cache_properties, list);
122 		list_del(&cache->list);
123 		kfree(cache);
124 	}
125 
126 	while (dev->io_link_props.next != &dev->io_link_props) {
127 		iolink = container_of(dev->io_link_props.next,
128 				struct kfd_iolink_properties, list);
129 		list_del(&iolink->list);
130 		kfree(iolink);
131 	}
132 
133 	while (dev->perf_props.next != &dev->perf_props) {
134 		perf = container_of(dev->perf_props.next,
135 				struct kfd_perf_properties, list);
136 		list_del(&perf->list);
137 		kfree(perf);
138 	}
139 
140 	kfree(dev);
141 }
142 
143 void kfd_release_topology_device_list(struct list_head *device_list)
144 {
145 	struct kfd_topology_device *dev;
146 
147 	while (!list_empty(device_list)) {
148 		dev = list_first_entry(device_list,
149 				       struct kfd_topology_device, list);
150 		kfd_release_topology_device(dev);
151 	}
152 }
153 
154 static void kfd_release_live_view(void)
155 {
156 	kfd_release_topology_device_list(&topology_device_list);
157 	memset(&sys_props, 0, sizeof(sys_props));
158 }
159 
160 struct kfd_topology_device *kfd_create_topology_device(
161 				struct list_head *device_list)
162 {
163 	struct kfd_topology_device *dev;
164 
165 	dev = kfd_alloc_struct(dev);
166 	if (!dev) {
167 		pr_err("No memory to allocate a topology device");
168 		return NULL;
169 	}
170 
171 	INIT_LIST_HEAD(&dev->mem_props);
172 	INIT_LIST_HEAD(&dev->cache_props);
173 	INIT_LIST_HEAD(&dev->io_link_props);
174 	INIT_LIST_HEAD(&dev->perf_props);
175 
176 	list_add_tail(&dev->list, device_list);
177 
178 	return dev;
179 }
180 
181 
182 #define sysfs_show_gen_prop(buffer, fmt, ...) \
183 		snprintf(buffer, PAGE_SIZE, "%s"fmt, buffer, __VA_ARGS__)
184 #define sysfs_show_32bit_prop(buffer, name, value) \
185 		sysfs_show_gen_prop(buffer, "%s %u\n", name, value)
186 #define sysfs_show_64bit_prop(buffer, name, value) \
187 		sysfs_show_gen_prop(buffer, "%s %llu\n", name, value)
188 #define sysfs_show_32bit_val(buffer, value) \
189 		sysfs_show_gen_prop(buffer, "%u\n", value)
190 #define sysfs_show_str_val(buffer, value) \
191 		sysfs_show_gen_prop(buffer, "%s\n", value)
192 
193 static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr,
194 		char *buffer)
195 {
196 	ssize_t ret;
197 
198 	/* Making sure that the buffer is an empty string */
199 	buffer[0] = 0;
200 
201 	if (attr == &sys_props.attr_genid) {
202 		ret = sysfs_show_32bit_val(buffer, sys_props.generation_count);
203 	} else if (attr == &sys_props.attr_props) {
204 		sysfs_show_64bit_prop(buffer, "platform_oem",
205 				sys_props.platform_oem);
206 		sysfs_show_64bit_prop(buffer, "platform_id",
207 				sys_props.platform_id);
208 		ret = sysfs_show_64bit_prop(buffer, "platform_rev",
209 				sys_props.platform_rev);
210 	} else {
211 		ret = -EINVAL;
212 	}
213 
214 	return ret;
215 }
216 
217 static void kfd_topology_kobj_release(struct kobject *kobj)
218 {
219 	kfree(kobj);
220 }
221 
222 static const struct sysfs_ops sysprops_ops = {
223 	.show = sysprops_show,
224 };
225 
226 static struct kobj_type sysprops_type = {
227 	.release = kfd_topology_kobj_release,
228 	.sysfs_ops = &sysprops_ops,
229 };
230 
231 static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr,
232 		char *buffer)
233 {
234 	ssize_t ret;
235 	struct kfd_iolink_properties *iolink;
236 
237 	/* Making sure that the buffer is an empty string */
238 	buffer[0] = 0;
239 
240 	iolink = container_of(attr, struct kfd_iolink_properties, attr);
241 	sysfs_show_32bit_prop(buffer, "type", iolink->iolink_type);
242 	sysfs_show_32bit_prop(buffer, "version_major", iolink->ver_maj);
243 	sysfs_show_32bit_prop(buffer, "version_minor", iolink->ver_min);
244 	sysfs_show_32bit_prop(buffer, "node_from", iolink->node_from);
245 	sysfs_show_32bit_prop(buffer, "node_to", iolink->node_to);
246 	sysfs_show_32bit_prop(buffer, "weight", iolink->weight);
247 	sysfs_show_32bit_prop(buffer, "min_latency", iolink->min_latency);
248 	sysfs_show_32bit_prop(buffer, "max_latency", iolink->max_latency);
249 	sysfs_show_32bit_prop(buffer, "min_bandwidth", iolink->min_bandwidth);
250 	sysfs_show_32bit_prop(buffer, "max_bandwidth", iolink->max_bandwidth);
251 	sysfs_show_32bit_prop(buffer, "recommended_transfer_size",
252 			iolink->rec_transfer_size);
253 	ret = sysfs_show_32bit_prop(buffer, "flags", iolink->flags);
254 
255 	return ret;
256 }
257 
258 static const struct sysfs_ops iolink_ops = {
259 	.show = iolink_show,
260 };
261 
262 static struct kobj_type iolink_type = {
263 	.release = kfd_topology_kobj_release,
264 	.sysfs_ops = &iolink_ops,
265 };
266 
267 static ssize_t mem_show(struct kobject *kobj, struct attribute *attr,
268 		char *buffer)
269 {
270 	ssize_t ret;
271 	struct kfd_mem_properties *mem;
272 
273 	/* Making sure that the buffer is an empty string */
274 	buffer[0] = 0;
275 
276 	mem = container_of(attr, struct kfd_mem_properties, attr);
277 	sysfs_show_32bit_prop(buffer, "heap_type", mem->heap_type);
278 	sysfs_show_64bit_prop(buffer, "size_in_bytes", mem->size_in_bytes);
279 	sysfs_show_32bit_prop(buffer, "flags", mem->flags);
280 	sysfs_show_32bit_prop(buffer, "width", mem->width);
281 	ret = sysfs_show_32bit_prop(buffer, "mem_clk_max", mem->mem_clk_max);
282 
283 	return ret;
284 }
285 
286 static const struct sysfs_ops mem_ops = {
287 	.show = mem_show,
288 };
289 
290 static struct kobj_type mem_type = {
291 	.release = kfd_topology_kobj_release,
292 	.sysfs_ops = &mem_ops,
293 };
294 
295 static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr,
296 		char *buffer)
297 {
298 	ssize_t ret;
299 	uint32_t i, j;
300 	struct kfd_cache_properties *cache;
301 
302 	/* Making sure that the buffer is an empty string */
303 	buffer[0] = 0;
304 
305 	cache = container_of(attr, struct kfd_cache_properties, attr);
306 	sysfs_show_32bit_prop(buffer, "processor_id_low",
307 			cache->processor_id_low);
308 	sysfs_show_32bit_prop(buffer, "level", cache->cache_level);
309 	sysfs_show_32bit_prop(buffer, "size", cache->cache_size);
310 	sysfs_show_32bit_prop(buffer, "cache_line_size", cache->cacheline_size);
311 	sysfs_show_32bit_prop(buffer, "cache_lines_per_tag",
312 			cache->cachelines_per_tag);
313 	sysfs_show_32bit_prop(buffer, "association", cache->cache_assoc);
314 	sysfs_show_32bit_prop(buffer, "latency", cache->cache_latency);
315 	sysfs_show_32bit_prop(buffer, "type", cache->cache_type);
316 	snprintf(buffer, PAGE_SIZE, "%ssibling_map ", buffer);
317 	for (i = 0; i < CRAT_SIBLINGMAP_SIZE; i++)
318 		for (j = 0; j < sizeof(cache->sibling_map[0])*8; j++) {
319 			/* Check each bit */
320 			if (cache->sibling_map[i] & (1 << j))
321 				ret = snprintf(buffer, PAGE_SIZE,
322 					 "%s%d%s", buffer, 1, ",");
323 			else
324 				ret = snprintf(buffer, PAGE_SIZE,
325 					 "%s%d%s", buffer, 0, ",");
326 		}
327 	/* Replace the last "," with end of line */
328 	*(buffer + strlen(buffer) - 1) = 0xA;
329 	return ret;
330 }
331 
332 static const struct sysfs_ops cache_ops = {
333 	.show = kfd_cache_show,
334 };
335 
336 static struct kobj_type cache_type = {
337 	.release = kfd_topology_kobj_release,
338 	.sysfs_ops = &cache_ops,
339 };
340 
341 /****** Sysfs of Performance Counters ******/
342 
343 struct kfd_perf_attr {
344 	struct kobj_attribute attr;
345 	uint32_t data;
346 };
347 
348 static ssize_t perf_show(struct kobject *kobj, struct kobj_attribute *attrs,
349 			char *buf)
350 {
351 	struct kfd_perf_attr *attr;
352 
353 	buf[0] = 0;
354 	attr = container_of(attrs, struct kfd_perf_attr, attr);
355 	if (!attr->data) /* invalid data for PMC */
356 		return 0;
357 	else
358 		return sysfs_show_32bit_val(buf, attr->data);
359 }
360 
361 #define KFD_PERF_DESC(_name, _data)			\
362 {							\
363 	.attr  = __ATTR(_name, 0444, perf_show, NULL),	\
364 	.data = _data,					\
365 }
366 
367 static struct kfd_perf_attr perf_attr_iommu[] = {
368 	KFD_PERF_DESC(max_concurrent, 0),
369 	KFD_PERF_DESC(num_counters, 0),
370 	KFD_PERF_DESC(counter_ids, 0),
371 };
372 /****************************************/
373 
374 static ssize_t node_show(struct kobject *kobj, struct attribute *attr,
375 		char *buffer)
376 {
377 	struct kfd_topology_device *dev;
378 	char public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE];
379 	uint32_t i;
380 	uint32_t log_max_watch_addr;
381 
382 	/* Making sure that the buffer is an empty string */
383 	buffer[0] = 0;
384 
385 	if (strcmp(attr->name, "gpu_id") == 0) {
386 		dev = container_of(attr, struct kfd_topology_device,
387 				attr_gpuid);
388 		return sysfs_show_32bit_val(buffer, dev->gpu_id);
389 	}
390 
391 	if (strcmp(attr->name, "name") == 0) {
392 		dev = container_of(attr, struct kfd_topology_device,
393 				attr_name);
394 		for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE; i++) {
395 			public_name[i] =
396 					(char)dev->node_props.marketing_name[i];
397 			if (dev->node_props.marketing_name[i] == 0)
398 				break;
399 		}
400 		public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1] = 0x0;
401 		return sysfs_show_str_val(buffer, public_name);
402 	}
403 
404 	dev = container_of(attr, struct kfd_topology_device,
405 			attr_props);
406 	sysfs_show_32bit_prop(buffer, "cpu_cores_count",
407 			dev->node_props.cpu_cores_count);
408 	sysfs_show_32bit_prop(buffer, "simd_count",
409 			dev->node_props.simd_count);
410 	sysfs_show_32bit_prop(buffer, "mem_banks_count",
411 			dev->node_props.mem_banks_count);
412 	sysfs_show_32bit_prop(buffer, "caches_count",
413 			dev->node_props.caches_count);
414 	sysfs_show_32bit_prop(buffer, "io_links_count",
415 			dev->node_props.io_links_count);
416 	sysfs_show_32bit_prop(buffer, "cpu_core_id_base",
417 			dev->node_props.cpu_core_id_base);
418 	sysfs_show_32bit_prop(buffer, "simd_id_base",
419 			dev->node_props.simd_id_base);
420 	sysfs_show_32bit_prop(buffer, "max_waves_per_simd",
421 			dev->node_props.max_waves_per_simd);
422 	sysfs_show_32bit_prop(buffer, "lds_size_in_kb",
423 			dev->node_props.lds_size_in_kb);
424 	sysfs_show_32bit_prop(buffer, "gds_size_in_kb",
425 			dev->node_props.gds_size_in_kb);
426 	sysfs_show_32bit_prop(buffer, "wave_front_size",
427 			dev->node_props.wave_front_size);
428 	sysfs_show_32bit_prop(buffer, "array_count",
429 			dev->node_props.array_count);
430 	sysfs_show_32bit_prop(buffer, "simd_arrays_per_engine",
431 			dev->node_props.simd_arrays_per_engine);
432 	sysfs_show_32bit_prop(buffer, "cu_per_simd_array",
433 			dev->node_props.cu_per_simd_array);
434 	sysfs_show_32bit_prop(buffer, "simd_per_cu",
435 			dev->node_props.simd_per_cu);
436 	sysfs_show_32bit_prop(buffer, "max_slots_scratch_cu",
437 			dev->node_props.max_slots_scratch_cu);
438 	sysfs_show_32bit_prop(buffer, "vendor_id",
439 			dev->node_props.vendor_id);
440 	sysfs_show_32bit_prop(buffer, "device_id",
441 			dev->node_props.device_id);
442 	sysfs_show_32bit_prop(buffer, "location_id",
443 			dev->node_props.location_id);
444 	sysfs_show_32bit_prop(buffer, "drm_render_minor",
445 			dev->node_props.drm_render_minor);
446 
447 	if (dev->gpu) {
448 		log_max_watch_addr =
449 			__ilog2_u32(dev->gpu->device_info->num_of_watch_points);
450 
451 		if (log_max_watch_addr) {
452 			dev->node_props.capability |=
453 					HSA_CAP_WATCH_POINTS_SUPPORTED;
454 
455 			dev->node_props.capability |=
456 				((log_max_watch_addr <<
457 					HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) &
458 				HSA_CAP_WATCH_POINTS_TOTALBITS_MASK);
459 		}
460 
461 		if (dev->gpu->device_info->asic_family == CHIP_TONGA)
462 			dev->node_props.capability |=
463 					HSA_CAP_AQL_QUEUE_DOUBLE_MAP;
464 
465 		sysfs_show_32bit_prop(buffer, "max_engine_clk_fcompute",
466 			dev->node_props.max_engine_clk_fcompute);
467 
468 		sysfs_show_64bit_prop(buffer, "local_mem_size",
469 				(unsigned long long int) 0);
470 
471 		sysfs_show_32bit_prop(buffer, "fw_version",
472 			dev->gpu->kfd2kgd->get_fw_version(
473 						dev->gpu->kgd,
474 						KGD_ENGINE_MEC1));
475 		sysfs_show_32bit_prop(buffer, "capability",
476 				dev->node_props.capability);
477 	}
478 
479 	return sysfs_show_32bit_prop(buffer, "max_engine_clk_ccompute",
480 					cpufreq_quick_get_max(0)/1000);
481 }
482 
483 static const struct sysfs_ops node_ops = {
484 	.show = node_show,
485 };
486 
487 static struct kobj_type node_type = {
488 	.release = kfd_topology_kobj_release,
489 	.sysfs_ops = &node_ops,
490 };
491 
492 static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr)
493 {
494 	sysfs_remove_file(kobj, attr);
495 	kobject_del(kobj);
496 	kobject_put(kobj);
497 }
498 
499 static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev)
500 {
501 	struct kfd_iolink_properties *iolink;
502 	struct kfd_cache_properties *cache;
503 	struct kfd_mem_properties *mem;
504 	struct kfd_perf_properties *perf;
505 
506 	if (dev->kobj_iolink) {
507 		list_for_each_entry(iolink, &dev->io_link_props, list)
508 			if (iolink->kobj) {
509 				kfd_remove_sysfs_file(iolink->kobj,
510 							&iolink->attr);
511 				iolink->kobj = NULL;
512 			}
513 		kobject_del(dev->kobj_iolink);
514 		kobject_put(dev->kobj_iolink);
515 		dev->kobj_iolink = NULL;
516 	}
517 
518 	if (dev->kobj_cache) {
519 		list_for_each_entry(cache, &dev->cache_props, list)
520 			if (cache->kobj) {
521 				kfd_remove_sysfs_file(cache->kobj,
522 							&cache->attr);
523 				cache->kobj = NULL;
524 			}
525 		kobject_del(dev->kobj_cache);
526 		kobject_put(dev->kobj_cache);
527 		dev->kobj_cache = NULL;
528 	}
529 
530 	if (dev->kobj_mem) {
531 		list_for_each_entry(mem, &dev->mem_props, list)
532 			if (mem->kobj) {
533 				kfd_remove_sysfs_file(mem->kobj, &mem->attr);
534 				mem->kobj = NULL;
535 			}
536 		kobject_del(dev->kobj_mem);
537 		kobject_put(dev->kobj_mem);
538 		dev->kobj_mem = NULL;
539 	}
540 
541 	if (dev->kobj_perf) {
542 		list_for_each_entry(perf, &dev->perf_props, list) {
543 			kfree(perf->attr_group);
544 			perf->attr_group = NULL;
545 		}
546 		kobject_del(dev->kobj_perf);
547 		kobject_put(dev->kobj_perf);
548 		dev->kobj_perf = NULL;
549 	}
550 
551 	if (dev->kobj_node) {
552 		sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid);
553 		sysfs_remove_file(dev->kobj_node, &dev->attr_name);
554 		sysfs_remove_file(dev->kobj_node, &dev->attr_props);
555 		kobject_del(dev->kobj_node);
556 		kobject_put(dev->kobj_node);
557 		dev->kobj_node = NULL;
558 	}
559 }
560 
561 static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev,
562 		uint32_t id)
563 {
564 	struct kfd_iolink_properties *iolink;
565 	struct kfd_cache_properties *cache;
566 	struct kfd_mem_properties *mem;
567 	struct kfd_perf_properties *perf;
568 	int ret;
569 	uint32_t i, num_attrs;
570 	struct attribute **attrs;
571 
572 	if (WARN_ON(dev->kobj_node))
573 		return -EEXIST;
574 
575 	/*
576 	 * Creating the sysfs folders
577 	 */
578 	dev->kobj_node = kfd_alloc_struct(dev->kobj_node);
579 	if (!dev->kobj_node)
580 		return -ENOMEM;
581 
582 	ret = kobject_init_and_add(dev->kobj_node, &node_type,
583 			sys_props.kobj_nodes, "%d", id);
584 	if (ret < 0)
585 		return ret;
586 
587 	dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node);
588 	if (!dev->kobj_mem)
589 		return -ENOMEM;
590 
591 	dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node);
592 	if (!dev->kobj_cache)
593 		return -ENOMEM;
594 
595 	dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node);
596 	if (!dev->kobj_iolink)
597 		return -ENOMEM;
598 
599 	dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node);
600 	if (!dev->kobj_perf)
601 		return -ENOMEM;
602 
603 	/*
604 	 * Creating sysfs files for node properties
605 	 */
606 	dev->attr_gpuid.name = "gpu_id";
607 	dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE;
608 	sysfs_attr_init(&dev->attr_gpuid);
609 	dev->attr_name.name = "name";
610 	dev->attr_name.mode = KFD_SYSFS_FILE_MODE;
611 	sysfs_attr_init(&dev->attr_name);
612 	dev->attr_props.name = "properties";
613 	dev->attr_props.mode = KFD_SYSFS_FILE_MODE;
614 	sysfs_attr_init(&dev->attr_props);
615 	ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid);
616 	if (ret < 0)
617 		return ret;
618 	ret = sysfs_create_file(dev->kobj_node, &dev->attr_name);
619 	if (ret < 0)
620 		return ret;
621 	ret = sysfs_create_file(dev->kobj_node, &dev->attr_props);
622 	if (ret < 0)
623 		return ret;
624 
625 	i = 0;
626 	list_for_each_entry(mem, &dev->mem_props, list) {
627 		mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
628 		if (!mem->kobj)
629 			return -ENOMEM;
630 		ret = kobject_init_and_add(mem->kobj, &mem_type,
631 				dev->kobj_mem, "%d", i);
632 		if (ret < 0)
633 			return ret;
634 
635 		mem->attr.name = "properties";
636 		mem->attr.mode = KFD_SYSFS_FILE_MODE;
637 		sysfs_attr_init(&mem->attr);
638 		ret = sysfs_create_file(mem->kobj, &mem->attr);
639 		if (ret < 0)
640 			return ret;
641 		i++;
642 	}
643 
644 	i = 0;
645 	list_for_each_entry(cache, &dev->cache_props, list) {
646 		cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
647 		if (!cache->kobj)
648 			return -ENOMEM;
649 		ret = kobject_init_and_add(cache->kobj, &cache_type,
650 				dev->kobj_cache, "%d", i);
651 		if (ret < 0)
652 			return ret;
653 
654 		cache->attr.name = "properties";
655 		cache->attr.mode = KFD_SYSFS_FILE_MODE;
656 		sysfs_attr_init(&cache->attr);
657 		ret = sysfs_create_file(cache->kobj, &cache->attr);
658 		if (ret < 0)
659 			return ret;
660 		i++;
661 	}
662 
663 	i = 0;
664 	list_for_each_entry(iolink, &dev->io_link_props, list) {
665 		iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
666 		if (!iolink->kobj)
667 			return -ENOMEM;
668 		ret = kobject_init_and_add(iolink->kobj, &iolink_type,
669 				dev->kobj_iolink, "%d", i);
670 		if (ret < 0)
671 			return ret;
672 
673 		iolink->attr.name = "properties";
674 		iolink->attr.mode = KFD_SYSFS_FILE_MODE;
675 		sysfs_attr_init(&iolink->attr);
676 		ret = sysfs_create_file(iolink->kobj, &iolink->attr);
677 		if (ret < 0)
678 			return ret;
679 		i++;
680 	}
681 
682 	/* All hardware blocks have the same number of attributes. */
683 	num_attrs = ARRAY_SIZE(perf_attr_iommu);
684 	list_for_each_entry(perf, &dev->perf_props, list) {
685 		perf->attr_group = kzalloc(sizeof(struct kfd_perf_attr)
686 			* num_attrs + sizeof(struct attribute_group),
687 			GFP_KERNEL);
688 		if (!perf->attr_group)
689 			return -ENOMEM;
690 
691 		attrs = (struct attribute **)(perf->attr_group + 1);
692 		if (!strcmp(perf->block_name, "iommu")) {
693 		/* Information of IOMMU's num_counters and counter_ids is shown
694 		 * under /sys/bus/event_source/devices/amd_iommu. We don't
695 		 * duplicate here.
696 		 */
697 			perf_attr_iommu[0].data = perf->max_concurrent;
698 			for (i = 0; i < num_attrs; i++)
699 				attrs[i] = &perf_attr_iommu[i].attr.attr;
700 		}
701 		perf->attr_group->name = perf->block_name;
702 		perf->attr_group->attrs = attrs;
703 		ret = sysfs_create_group(dev->kobj_perf, perf->attr_group);
704 		if (ret < 0)
705 			return ret;
706 	}
707 
708 	return 0;
709 }
710 
711 /* Called with write topology lock acquired */
712 static int kfd_build_sysfs_node_tree(void)
713 {
714 	struct kfd_topology_device *dev;
715 	int ret;
716 	uint32_t i = 0;
717 
718 	list_for_each_entry(dev, &topology_device_list, list) {
719 		ret = kfd_build_sysfs_node_entry(dev, i);
720 		if (ret < 0)
721 			return ret;
722 		i++;
723 	}
724 
725 	return 0;
726 }
727 
728 /* Called with write topology lock acquired */
729 static void kfd_remove_sysfs_node_tree(void)
730 {
731 	struct kfd_topology_device *dev;
732 
733 	list_for_each_entry(dev, &topology_device_list, list)
734 		kfd_remove_sysfs_node_entry(dev);
735 }
736 
737 static int kfd_topology_update_sysfs(void)
738 {
739 	int ret;
740 
741 	pr_info("Creating topology SYSFS entries\n");
742 	if (!sys_props.kobj_topology) {
743 		sys_props.kobj_topology =
744 				kfd_alloc_struct(sys_props.kobj_topology);
745 		if (!sys_props.kobj_topology)
746 			return -ENOMEM;
747 
748 		ret = kobject_init_and_add(sys_props.kobj_topology,
749 				&sysprops_type,  &kfd_device->kobj,
750 				"topology");
751 		if (ret < 0)
752 			return ret;
753 
754 		sys_props.kobj_nodes = kobject_create_and_add("nodes",
755 				sys_props.kobj_topology);
756 		if (!sys_props.kobj_nodes)
757 			return -ENOMEM;
758 
759 		sys_props.attr_genid.name = "generation_id";
760 		sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE;
761 		sysfs_attr_init(&sys_props.attr_genid);
762 		ret = sysfs_create_file(sys_props.kobj_topology,
763 				&sys_props.attr_genid);
764 		if (ret < 0)
765 			return ret;
766 
767 		sys_props.attr_props.name = "system_properties";
768 		sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE;
769 		sysfs_attr_init(&sys_props.attr_props);
770 		ret = sysfs_create_file(sys_props.kobj_topology,
771 				&sys_props.attr_props);
772 		if (ret < 0)
773 			return ret;
774 	}
775 
776 	kfd_remove_sysfs_node_tree();
777 
778 	return kfd_build_sysfs_node_tree();
779 }
780 
781 static void kfd_topology_release_sysfs(void)
782 {
783 	kfd_remove_sysfs_node_tree();
784 	if (sys_props.kobj_topology) {
785 		sysfs_remove_file(sys_props.kobj_topology,
786 				&sys_props.attr_genid);
787 		sysfs_remove_file(sys_props.kobj_topology,
788 				&sys_props.attr_props);
789 		if (sys_props.kobj_nodes) {
790 			kobject_del(sys_props.kobj_nodes);
791 			kobject_put(sys_props.kobj_nodes);
792 			sys_props.kobj_nodes = NULL;
793 		}
794 		kobject_del(sys_props.kobj_topology);
795 		kobject_put(sys_props.kobj_topology);
796 		sys_props.kobj_topology = NULL;
797 	}
798 }
799 
800 /* Called with write topology_lock acquired */
801 static void kfd_topology_update_device_list(struct list_head *temp_list,
802 					struct list_head *master_list)
803 {
804 	while (!list_empty(temp_list)) {
805 		list_move_tail(temp_list->next, master_list);
806 		sys_props.num_devices++;
807 	}
808 }
809 
810 static void kfd_debug_print_topology(void)
811 {
812 	struct kfd_topology_device *dev;
813 
814 	down_read(&topology_lock);
815 
816 	dev = list_last_entry(&topology_device_list,
817 			struct kfd_topology_device, list);
818 	if (dev) {
819 		if (dev->node_props.cpu_cores_count &&
820 				dev->node_props.simd_count) {
821 			pr_info("Topology: Add APU node [0x%0x:0x%0x]\n",
822 				dev->node_props.device_id,
823 				dev->node_props.vendor_id);
824 		} else if (dev->node_props.cpu_cores_count)
825 			pr_info("Topology: Add CPU node\n");
826 		else if (dev->node_props.simd_count)
827 			pr_info("Topology: Add dGPU node [0x%0x:0x%0x]\n",
828 				dev->node_props.device_id,
829 				dev->node_props.vendor_id);
830 	}
831 	up_read(&topology_lock);
832 }
833 
834 /* Helper function for intializing platform_xx members of
835  * kfd_system_properties. Uses OEM info from the last CPU/APU node.
836  */
837 static void kfd_update_system_properties(void)
838 {
839 	struct kfd_topology_device *dev;
840 
841 	down_read(&topology_lock);
842 	dev = list_last_entry(&topology_device_list,
843 			struct kfd_topology_device, list);
844 	if (dev) {
845 		sys_props.platform_id =
846 			(*((uint64_t *)dev->oem_id)) & CRAT_OEMID_64BIT_MASK;
847 		sys_props.platform_oem = *((uint64_t *)dev->oem_table_id);
848 		sys_props.platform_rev = dev->oem_revision;
849 	}
850 	up_read(&topology_lock);
851 }
852 
853 static void find_system_memory(const struct dmi_header *dm,
854 	void *private)
855 {
856 	struct kfd_mem_properties *mem;
857 	u16 mem_width, mem_clock;
858 	struct kfd_topology_device *kdev =
859 		(struct kfd_topology_device *)private;
860 	const u8 *dmi_data = (const u8 *)(dm + 1);
861 
862 	if (dm->type == DMI_ENTRY_MEM_DEVICE && dm->length >= 0x15) {
863 		mem_width = (u16)(*(const u16 *)(dmi_data + 0x6));
864 		mem_clock = (u16)(*(const u16 *)(dmi_data + 0x11));
865 		list_for_each_entry(mem, &kdev->mem_props, list) {
866 			if (mem_width != 0xFFFF && mem_width != 0)
867 				mem->width = mem_width;
868 			if (mem_clock != 0)
869 				mem->mem_clk_max = mem_clock;
870 		}
871 	}
872 }
873 
874 /*
875  * Performance counters information is not part of CRAT but we would like to
876  * put them in the sysfs under topology directory for Thunk to get the data.
877  * This function is called before updating the sysfs.
878  */
879 static int kfd_add_perf_to_topology(struct kfd_topology_device *kdev)
880 {
881 	/* These are the only counters supported so far */
882 	return kfd_iommu_add_perf_counters(kdev);
883 }
884 
885 /* kfd_add_non_crat_information - Add information that is not currently
886  *	defined in CRAT but is necessary for KFD topology
887  * @dev - topology device to which addition info is added
888  */
889 static void kfd_add_non_crat_information(struct kfd_topology_device *kdev)
890 {
891 	/* Check if CPU only node. */
892 	if (!kdev->gpu) {
893 		/* Add system memory information */
894 		dmi_walk(find_system_memory, kdev);
895 	}
896 	/* TODO: For GPU node, rearrange code from kfd_topology_add_device */
897 }
898 
899 /* kfd_is_acpi_crat_invalid - CRAT from ACPI is valid only for AMD APU devices.
900  *	Ignore CRAT for all other devices. AMD APU is identified if both CPU
901  *	and GPU cores are present.
902  * @device_list - topology device list created by parsing ACPI CRAT table.
903  * @return - TRUE if invalid, FALSE is valid.
904  */
905 static bool kfd_is_acpi_crat_invalid(struct list_head *device_list)
906 {
907 	struct kfd_topology_device *dev;
908 
909 	list_for_each_entry(dev, device_list, list) {
910 		if (dev->node_props.cpu_cores_count &&
911 			dev->node_props.simd_count)
912 			return false;
913 	}
914 	pr_info("Ignoring ACPI CRAT on non-APU system\n");
915 	return true;
916 }
917 
918 int kfd_topology_init(void)
919 {
920 	void *crat_image = NULL;
921 	size_t image_size = 0;
922 	int ret;
923 	struct list_head temp_topology_device_list;
924 	int cpu_only_node = 0;
925 	struct kfd_topology_device *kdev;
926 	int proximity_domain;
927 
928 	/* topology_device_list - Master list of all topology devices
929 	 * temp_topology_device_list - temporary list created while parsing CRAT
930 	 * or VCRAT. Once parsing is complete the contents of list is moved to
931 	 * topology_device_list
932 	 */
933 
934 	/* Initialize the head for the both the lists */
935 	INIT_LIST_HEAD(&topology_device_list);
936 	INIT_LIST_HEAD(&temp_topology_device_list);
937 	init_rwsem(&topology_lock);
938 
939 	memset(&sys_props, 0, sizeof(sys_props));
940 
941 	/* Proximity domains in ACPI CRAT tables start counting at
942 	 * 0. The same should be true for virtual CRAT tables created
943 	 * at this stage. GPUs added later in kfd_topology_add_device
944 	 * use a counter.
945 	 */
946 	proximity_domain = 0;
947 
948 	/*
949 	 * Get the CRAT image from the ACPI. If ACPI doesn't have one
950 	 * or if ACPI CRAT is invalid create a virtual CRAT.
951 	 * NOTE: The current implementation expects all AMD APUs to have
952 	 *	CRAT. If no CRAT is available, it is assumed to be a CPU
953 	 */
954 	ret = kfd_create_crat_image_acpi(&crat_image, &image_size);
955 	if (!ret) {
956 		ret = kfd_parse_crat_table(crat_image,
957 					   &temp_topology_device_list,
958 					   proximity_domain);
959 		if (ret ||
960 		    kfd_is_acpi_crat_invalid(&temp_topology_device_list)) {
961 			kfd_release_topology_device_list(
962 				&temp_topology_device_list);
963 			kfd_destroy_crat_image(crat_image);
964 			crat_image = NULL;
965 		}
966 	}
967 
968 	if (!crat_image) {
969 		ret = kfd_create_crat_image_virtual(&crat_image, &image_size,
970 						    COMPUTE_UNIT_CPU, NULL,
971 						    proximity_domain);
972 		cpu_only_node = 1;
973 		if (ret) {
974 			pr_err("Error creating VCRAT table for CPU\n");
975 			return ret;
976 		}
977 
978 		ret = kfd_parse_crat_table(crat_image,
979 					   &temp_topology_device_list,
980 					   proximity_domain);
981 		if (ret) {
982 			pr_err("Error parsing VCRAT table for CPU\n");
983 			goto err;
984 		}
985 	}
986 
987 	kdev = list_first_entry(&temp_topology_device_list,
988 				struct kfd_topology_device, list);
989 	kfd_add_perf_to_topology(kdev);
990 
991 	down_write(&topology_lock);
992 	kfd_topology_update_device_list(&temp_topology_device_list,
993 					&topology_device_list);
994 	atomic_set(&topology_crat_proximity_domain, sys_props.num_devices-1);
995 	ret = kfd_topology_update_sysfs();
996 	up_write(&topology_lock);
997 
998 	if (!ret) {
999 		sys_props.generation_count++;
1000 		kfd_update_system_properties();
1001 		kfd_debug_print_topology();
1002 		pr_info("Finished initializing topology\n");
1003 	} else
1004 		pr_err("Failed to update topology in sysfs ret=%d\n", ret);
1005 
1006 	/* For nodes with GPU, this information gets added
1007 	 * when GPU is detected (kfd_topology_add_device).
1008 	 */
1009 	if (cpu_only_node) {
1010 		/* Add additional information to CPU only node created above */
1011 		down_write(&topology_lock);
1012 		kdev = list_first_entry(&topology_device_list,
1013 				struct kfd_topology_device, list);
1014 		up_write(&topology_lock);
1015 		kfd_add_non_crat_information(kdev);
1016 	}
1017 
1018 err:
1019 	kfd_destroy_crat_image(crat_image);
1020 	return ret;
1021 }
1022 
1023 void kfd_topology_shutdown(void)
1024 {
1025 	down_write(&topology_lock);
1026 	kfd_topology_release_sysfs();
1027 	kfd_release_live_view();
1028 	up_write(&topology_lock);
1029 }
1030 
1031 static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu)
1032 {
1033 	uint32_t hashout;
1034 	uint32_t buf[7];
1035 	uint64_t local_mem_size;
1036 	int i;
1037 	struct kfd_local_mem_info local_mem_info;
1038 
1039 	if (!gpu)
1040 		return 0;
1041 
1042 	gpu->kfd2kgd->get_local_mem_info(gpu->kgd, &local_mem_info);
1043 
1044 	local_mem_size = local_mem_info.local_mem_size_private +
1045 			local_mem_info.local_mem_size_public;
1046 
1047 	buf[0] = gpu->pdev->devfn;
1048 	buf[1] = gpu->pdev->subsystem_vendor;
1049 	buf[2] = gpu->pdev->subsystem_device;
1050 	buf[3] = gpu->pdev->device;
1051 	buf[4] = gpu->pdev->bus->number;
1052 	buf[5] = lower_32_bits(local_mem_size);
1053 	buf[6] = upper_32_bits(local_mem_size);
1054 
1055 	for (i = 0, hashout = 0; i < 7; i++)
1056 		hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH);
1057 
1058 	return hashout;
1059 }
1060 /* kfd_assign_gpu - Attach @gpu to the correct kfd topology device. If
1061  *		the GPU device is not already present in the topology device
1062  *		list then return NULL. This means a new topology device has to
1063  *		be created for this GPU.
1064  * TODO: Rather than assiging @gpu to first topology device withtout
1065  *		gpu attached, it will better to have more stringent check.
1066  */
1067 static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu)
1068 {
1069 	struct kfd_topology_device *dev;
1070 	struct kfd_topology_device *out_dev = NULL;
1071 
1072 	down_write(&topology_lock);
1073 	list_for_each_entry(dev, &topology_device_list, list)
1074 		if (!dev->gpu && (dev->node_props.simd_count > 0)) {
1075 			dev->gpu = gpu;
1076 			out_dev = dev;
1077 			break;
1078 		}
1079 	up_write(&topology_lock);
1080 	return out_dev;
1081 }
1082 
1083 static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival)
1084 {
1085 	/*
1086 	 * TODO: Generate an event for thunk about the arrival/removal
1087 	 * of the GPU
1088 	 */
1089 }
1090 
1091 /* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info,
1092  *		patch this after CRAT parsing.
1093  */
1094 static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev)
1095 {
1096 	struct kfd_mem_properties *mem;
1097 	struct kfd_local_mem_info local_mem_info;
1098 
1099 	if (!dev)
1100 		return;
1101 
1102 	/* Currently, amdgpu driver (amdgpu_mc) deals only with GPUs with
1103 	 * single bank of VRAM local memory.
1104 	 * for dGPUs - VCRAT reports only one bank of Local Memory
1105 	 * for APUs - If CRAT from ACPI reports more than one bank, then
1106 	 *	all the banks will report the same mem_clk_max information
1107 	 */
1108 	dev->gpu->kfd2kgd->get_local_mem_info(dev->gpu->kgd,
1109 		&local_mem_info);
1110 
1111 	list_for_each_entry(mem, &dev->mem_props, list)
1112 		mem->mem_clk_max = local_mem_info.mem_clk_max;
1113 }
1114 
1115 static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev)
1116 {
1117 	struct kfd_iolink_properties *link;
1118 
1119 	if (!dev || !dev->gpu)
1120 		return;
1121 
1122 	/* GPU only creates direck links so apply flags setting to all */
1123 	if (dev->gpu->device_info->asic_family == CHIP_HAWAII)
1124 		list_for_each_entry(link, &dev->io_link_props, list)
1125 			link->flags = CRAT_IOLINK_FLAGS_ENABLED |
1126 				CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1127 				CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1128 }
1129 
1130 int kfd_topology_add_device(struct kfd_dev *gpu)
1131 {
1132 	uint32_t gpu_id;
1133 	struct kfd_topology_device *dev;
1134 	struct kfd_cu_info cu_info;
1135 	int res = 0;
1136 	struct list_head temp_topology_device_list;
1137 	void *crat_image = NULL;
1138 	size_t image_size = 0;
1139 	int proximity_domain;
1140 
1141 	INIT_LIST_HEAD(&temp_topology_device_list);
1142 
1143 	gpu_id = kfd_generate_gpu_id(gpu);
1144 
1145 	pr_debug("Adding new GPU (ID: 0x%x) to topology\n", gpu_id);
1146 
1147 	proximity_domain = atomic_inc_return(&topology_crat_proximity_domain);
1148 
1149 	/* Check to see if this gpu device exists in the topology_device_list.
1150 	 * If so, assign the gpu to that device,
1151 	 * else create a Virtual CRAT for this gpu device and then parse that
1152 	 * CRAT to create a new topology device. Once created assign the gpu to
1153 	 * that topology device
1154 	 */
1155 	dev = kfd_assign_gpu(gpu);
1156 	if (!dev) {
1157 		res = kfd_create_crat_image_virtual(&crat_image, &image_size,
1158 						    COMPUTE_UNIT_GPU, gpu,
1159 						    proximity_domain);
1160 		if (res) {
1161 			pr_err("Error creating VCRAT for GPU (ID: 0x%x)\n",
1162 			       gpu_id);
1163 			return res;
1164 		}
1165 		res = kfd_parse_crat_table(crat_image,
1166 					   &temp_topology_device_list,
1167 					   proximity_domain);
1168 		if (res) {
1169 			pr_err("Error parsing VCRAT for GPU (ID: 0x%x)\n",
1170 			       gpu_id);
1171 			goto err;
1172 		}
1173 
1174 		down_write(&topology_lock);
1175 		kfd_topology_update_device_list(&temp_topology_device_list,
1176 			&topology_device_list);
1177 
1178 		/* Update the SYSFS tree, since we added another topology
1179 		 * device
1180 		 */
1181 		res = kfd_topology_update_sysfs();
1182 		up_write(&topology_lock);
1183 
1184 		if (!res)
1185 			sys_props.generation_count++;
1186 		else
1187 			pr_err("Failed to update GPU (ID: 0x%x) to sysfs topology. res=%d\n",
1188 						gpu_id, res);
1189 		dev = kfd_assign_gpu(gpu);
1190 		if (WARN_ON(!dev)) {
1191 			res = -ENODEV;
1192 			goto err;
1193 		}
1194 	}
1195 
1196 	dev->gpu_id = gpu_id;
1197 	gpu->id = gpu_id;
1198 
1199 	/* TODO: Move the following lines to function
1200 	 *	kfd_add_non_crat_information
1201 	 */
1202 
1203 	/* Fill-in additional information that is not available in CRAT but
1204 	 * needed for the topology
1205 	 */
1206 
1207 	dev->gpu->kfd2kgd->get_cu_info(dev->gpu->kgd, &cu_info);
1208 	dev->node_props.simd_arrays_per_engine =
1209 		cu_info.num_shader_arrays_per_engine;
1210 
1211 	dev->node_props.vendor_id = gpu->pdev->vendor;
1212 	dev->node_props.device_id = gpu->pdev->device;
1213 	dev->node_props.location_id = PCI_DEVID(gpu->pdev->bus->number,
1214 		gpu->pdev->devfn);
1215 	dev->node_props.max_engine_clk_fcompute =
1216 		dev->gpu->kfd2kgd->get_max_engine_clock_in_mhz(dev->gpu->kgd);
1217 	dev->node_props.max_engine_clk_ccompute =
1218 		cpufreq_quick_get_max(0) / 1000;
1219 	dev->node_props.drm_render_minor =
1220 		gpu->shared_resources.drm_render_minor;
1221 
1222 	kfd_fill_mem_clk_max_info(dev);
1223 	kfd_fill_iolink_non_crat_info(dev);
1224 
1225 	switch (dev->gpu->device_info->asic_family) {
1226 	case CHIP_KAVERI:
1227 	case CHIP_HAWAII:
1228 	case CHIP_TONGA:
1229 		dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_PRE_1_0 <<
1230 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1231 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1232 		break;
1233 	case CHIP_CARRIZO:
1234 	case CHIP_FIJI:
1235 	case CHIP_POLARIS10:
1236 	case CHIP_POLARIS11:
1237 		pr_debug("Adding doorbell packet type capability\n");
1238 		dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 <<
1239 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1240 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1241 		break;
1242 	case CHIP_VEGA10:
1243 	case CHIP_RAVEN:
1244 		dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 <<
1245 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1246 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1247 		break;
1248 	default:
1249 		WARN(1, "Unexpected ASIC family %u",
1250 		     dev->gpu->device_info->asic_family);
1251 	}
1252 
1253 	/* Fix errors in CZ CRAT.
1254 	 * simd_count: Carrizo CRAT reports wrong simd_count, probably
1255 	 *		because it doesn't consider masked out CUs
1256 	 * max_waves_per_simd: Carrizo reports wrong max_waves_per_simd
1257 	 * capability flag: Carrizo CRAT doesn't report IOMMU flags
1258 	 */
1259 	if (dev->gpu->device_info->asic_family == CHIP_CARRIZO) {
1260 		dev->node_props.simd_count =
1261 			cu_info.simd_per_cu * cu_info.cu_active_number;
1262 		dev->node_props.max_waves_per_simd = 10;
1263 		dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
1264 	}
1265 
1266 	kfd_debug_print_topology();
1267 
1268 	if (!res)
1269 		kfd_notify_gpu_change(gpu_id, 1);
1270 err:
1271 	kfd_destroy_crat_image(crat_image);
1272 	return res;
1273 }
1274 
1275 int kfd_topology_remove_device(struct kfd_dev *gpu)
1276 {
1277 	struct kfd_topology_device *dev, *tmp;
1278 	uint32_t gpu_id;
1279 	int res = -ENODEV;
1280 
1281 	down_write(&topology_lock);
1282 
1283 	list_for_each_entry_safe(dev, tmp, &topology_device_list, list)
1284 		if (dev->gpu == gpu) {
1285 			gpu_id = dev->gpu_id;
1286 			kfd_remove_sysfs_node_entry(dev);
1287 			kfd_release_topology_device(dev);
1288 			sys_props.num_devices--;
1289 			res = 0;
1290 			if (kfd_topology_update_sysfs() < 0)
1291 				kfd_topology_release_sysfs();
1292 			break;
1293 		}
1294 
1295 	up_write(&topology_lock);
1296 
1297 	if (!res)
1298 		kfd_notify_gpu_change(gpu_id, 0);
1299 
1300 	return res;
1301 }
1302 
1303 /* kfd_topology_enum_kfd_devices - Enumerate through all devices in KFD
1304  *	topology. If GPU device is found @idx, then valid kfd_dev pointer is
1305  *	returned through @kdev
1306  * Return -	0: On success (@kdev will be NULL for non GPU nodes)
1307  *		-1: If end of list
1308  */
1309 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev)
1310 {
1311 
1312 	struct kfd_topology_device *top_dev;
1313 	uint8_t device_idx = 0;
1314 
1315 	*kdev = NULL;
1316 	down_read(&topology_lock);
1317 
1318 	list_for_each_entry(top_dev, &topology_device_list, list) {
1319 		if (device_idx == idx) {
1320 			*kdev = top_dev->gpu;
1321 			up_read(&topology_lock);
1322 			return 0;
1323 		}
1324 
1325 		device_idx++;
1326 	}
1327 
1328 	up_read(&topology_lock);
1329 
1330 	return -1;
1331 
1332 }
1333 
1334 static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask)
1335 {
1336 	const struct cpuinfo_x86 *cpuinfo;
1337 	int first_cpu_of_numa_node;
1338 
1339 	if (!cpumask || cpumask == cpu_none_mask)
1340 		return -1;
1341 	first_cpu_of_numa_node = cpumask_first(cpumask);
1342 	if (first_cpu_of_numa_node >= nr_cpu_ids)
1343 		return -1;
1344 	cpuinfo = &cpu_data(first_cpu_of_numa_node);
1345 
1346 	return cpuinfo->apicid;
1347 }
1348 
1349 /* kfd_numa_node_to_apic_id - Returns the APIC ID of the first logical processor
1350  *	of the given NUMA node (numa_node_id)
1351  * Return -1 on failure
1352  */
1353 int kfd_numa_node_to_apic_id(int numa_node_id)
1354 {
1355 	if (numa_node_id == -1) {
1356 		pr_warn("Invalid NUMA Node. Use online CPU mask\n");
1357 		return kfd_cpumask_to_apic_id(cpu_online_mask);
1358 	}
1359 	return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id));
1360 }
1361 
1362 #if defined(CONFIG_DEBUG_FS)
1363 
1364 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data)
1365 {
1366 	struct kfd_topology_device *dev;
1367 	unsigned int i = 0;
1368 	int r = 0;
1369 
1370 	down_read(&topology_lock);
1371 
1372 	list_for_each_entry(dev, &topology_device_list, list) {
1373 		if (!dev->gpu) {
1374 			i++;
1375 			continue;
1376 		}
1377 
1378 		seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1379 		r = dqm_debugfs_hqds(m, dev->gpu->dqm);
1380 		if (r)
1381 			break;
1382 	}
1383 
1384 	up_read(&topology_lock);
1385 
1386 	return r;
1387 }
1388 
1389 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data)
1390 {
1391 	struct kfd_topology_device *dev;
1392 	unsigned int i = 0;
1393 	int r = 0;
1394 
1395 	down_read(&topology_lock);
1396 
1397 	list_for_each_entry(dev, &topology_device_list, list) {
1398 		if (!dev->gpu) {
1399 			i++;
1400 			continue;
1401 		}
1402 
1403 		seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1404 		r = pm_debugfs_runlist(m, &dev->gpu->dqm->packets);
1405 		if (r)
1406 			break;
1407 	}
1408 
1409 	up_read(&topology_lock);
1410 
1411 	return r;
1412 }
1413 
1414 #endif
1415