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_svm.h"
40 #include "kfd_debug.h"
41 #include "amdgpu_amdkfd.h"
42 #include "amdgpu_ras.h"
43 #include "amdgpu.h"
44 
45 /* topology_device_list - Master list of all topology devices */
46 static struct list_head topology_device_list;
47 static struct kfd_system_properties sys_props;
48 
49 static DECLARE_RWSEM(topology_lock);
50 static uint32_t topology_crat_proximity_domain;
51 
52 struct kfd_topology_device *kfd_topology_device_by_proximity_domain_no_lock(
53 						uint32_t proximity_domain)
54 {
55 	struct kfd_topology_device *top_dev;
56 	struct kfd_topology_device *device = NULL;
57 
58 	list_for_each_entry(top_dev, &topology_device_list, list)
59 		if (top_dev->proximity_domain == proximity_domain) {
60 			device = top_dev;
61 			break;
62 		}
63 
64 	return device;
65 }
66 
67 struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
68 						uint32_t proximity_domain)
69 {
70 	struct kfd_topology_device *device = NULL;
71 
72 	down_read(&topology_lock);
73 
74 	device = kfd_topology_device_by_proximity_domain_no_lock(
75 							proximity_domain);
76 	up_read(&topology_lock);
77 
78 	return device;
79 }
80 
81 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id)
82 {
83 	struct kfd_topology_device *top_dev = NULL;
84 	struct kfd_topology_device *ret = NULL;
85 
86 	down_read(&topology_lock);
87 
88 	list_for_each_entry(top_dev, &topology_device_list, list)
89 		if (top_dev->gpu_id == gpu_id) {
90 			ret = top_dev;
91 			break;
92 		}
93 
94 	up_read(&topology_lock);
95 
96 	return ret;
97 }
98 
99 struct kfd_node *kfd_device_by_id(uint32_t gpu_id)
100 {
101 	struct kfd_topology_device *top_dev;
102 
103 	top_dev = kfd_topology_device_by_id(gpu_id);
104 	if (!top_dev)
105 		return NULL;
106 
107 	return top_dev->gpu;
108 }
109 
110 struct kfd_node *kfd_device_by_pci_dev(const struct pci_dev *pdev)
111 {
112 	struct kfd_topology_device *top_dev;
113 	struct kfd_node *device = NULL;
114 
115 	down_read(&topology_lock);
116 
117 	list_for_each_entry(top_dev, &topology_device_list, list)
118 		if (top_dev->gpu && top_dev->gpu->adev->pdev == pdev) {
119 			device = top_dev->gpu;
120 			break;
121 		}
122 
123 	up_read(&topology_lock);
124 
125 	return device;
126 }
127 
128 /* Called with write topology_lock acquired */
129 static void kfd_release_topology_device(struct kfd_topology_device *dev)
130 {
131 	struct kfd_mem_properties *mem;
132 	struct kfd_cache_properties *cache;
133 	struct kfd_iolink_properties *iolink;
134 	struct kfd_iolink_properties *p2plink;
135 	struct kfd_perf_properties *perf;
136 
137 	list_del(&dev->list);
138 
139 	while (dev->mem_props.next != &dev->mem_props) {
140 		mem = container_of(dev->mem_props.next,
141 				struct kfd_mem_properties, list);
142 		list_del(&mem->list);
143 		kfree(mem);
144 	}
145 
146 	while (dev->cache_props.next != &dev->cache_props) {
147 		cache = container_of(dev->cache_props.next,
148 				struct kfd_cache_properties, list);
149 		list_del(&cache->list);
150 		kfree(cache);
151 	}
152 
153 	while (dev->io_link_props.next != &dev->io_link_props) {
154 		iolink = container_of(dev->io_link_props.next,
155 				struct kfd_iolink_properties, list);
156 		list_del(&iolink->list);
157 		kfree(iolink);
158 	}
159 
160 	while (dev->p2p_link_props.next != &dev->p2p_link_props) {
161 		p2plink = container_of(dev->p2p_link_props.next,
162 				struct kfd_iolink_properties, list);
163 		list_del(&p2plink->list);
164 		kfree(p2plink);
165 	}
166 
167 	while (dev->perf_props.next != &dev->perf_props) {
168 		perf = container_of(dev->perf_props.next,
169 				struct kfd_perf_properties, list);
170 		list_del(&perf->list);
171 		kfree(perf);
172 	}
173 
174 	kfree(dev);
175 }
176 
177 void kfd_release_topology_device_list(struct list_head *device_list)
178 {
179 	struct kfd_topology_device *dev;
180 
181 	while (!list_empty(device_list)) {
182 		dev = list_first_entry(device_list,
183 				       struct kfd_topology_device, list);
184 		kfd_release_topology_device(dev);
185 	}
186 }
187 
188 static void kfd_release_live_view(void)
189 {
190 	kfd_release_topology_device_list(&topology_device_list);
191 	memset(&sys_props, 0, sizeof(sys_props));
192 }
193 
194 struct kfd_topology_device *kfd_create_topology_device(
195 				struct list_head *device_list)
196 {
197 	struct kfd_topology_device *dev;
198 
199 	dev = kfd_alloc_struct(dev);
200 	if (!dev) {
201 		pr_err("No memory to allocate a topology device");
202 		return NULL;
203 	}
204 
205 	INIT_LIST_HEAD(&dev->mem_props);
206 	INIT_LIST_HEAD(&dev->cache_props);
207 	INIT_LIST_HEAD(&dev->io_link_props);
208 	INIT_LIST_HEAD(&dev->p2p_link_props);
209 	INIT_LIST_HEAD(&dev->perf_props);
210 
211 	list_add_tail(&dev->list, device_list);
212 
213 	return dev;
214 }
215 
216 
217 #define sysfs_show_gen_prop(buffer, offs, fmt, ...)		\
218 		(offs += snprintf(buffer+offs, PAGE_SIZE-offs,	\
219 				  fmt, __VA_ARGS__))
220 #define sysfs_show_32bit_prop(buffer, offs, name, value) \
221 		sysfs_show_gen_prop(buffer, offs, "%s %u\n", name, value)
222 #define sysfs_show_64bit_prop(buffer, offs, name, value) \
223 		sysfs_show_gen_prop(buffer, offs, "%s %llu\n", name, value)
224 #define sysfs_show_32bit_val(buffer, offs, value) \
225 		sysfs_show_gen_prop(buffer, offs, "%u\n", value)
226 #define sysfs_show_str_val(buffer, offs, value) \
227 		sysfs_show_gen_prop(buffer, offs, "%s\n", value)
228 
229 static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr,
230 		char *buffer)
231 {
232 	int offs = 0;
233 
234 	/* Making sure that the buffer is an empty string */
235 	buffer[0] = 0;
236 
237 	if (attr == &sys_props.attr_genid) {
238 		sysfs_show_32bit_val(buffer, offs,
239 				     sys_props.generation_count);
240 	} else if (attr == &sys_props.attr_props) {
241 		sysfs_show_64bit_prop(buffer, offs, "platform_oem",
242 				      sys_props.platform_oem);
243 		sysfs_show_64bit_prop(buffer, offs, "platform_id",
244 				      sys_props.platform_id);
245 		sysfs_show_64bit_prop(buffer, offs, "platform_rev",
246 				      sys_props.platform_rev);
247 	} else {
248 		offs = -EINVAL;
249 	}
250 
251 	return offs;
252 }
253 
254 static void kfd_topology_kobj_release(struct kobject *kobj)
255 {
256 	kfree(kobj);
257 }
258 
259 static const struct sysfs_ops sysprops_ops = {
260 	.show = sysprops_show,
261 };
262 
263 static const struct kobj_type sysprops_type = {
264 	.release = kfd_topology_kobj_release,
265 	.sysfs_ops = &sysprops_ops,
266 };
267 
268 static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr,
269 		char *buffer)
270 {
271 	int offs = 0;
272 	struct kfd_iolink_properties *iolink;
273 
274 	/* Making sure that the buffer is an empty string */
275 	buffer[0] = 0;
276 
277 	iolink = container_of(attr, struct kfd_iolink_properties, attr);
278 	if (iolink->gpu && kfd_devcgroup_check_permission(iolink->gpu))
279 		return -EPERM;
280 	sysfs_show_32bit_prop(buffer, offs, "type", iolink->iolink_type);
281 	sysfs_show_32bit_prop(buffer, offs, "version_major", iolink->ver_maj);
282 	sysfs_show_32bit_prop(buffer, offs, "version_minor", iolink->ver_min);
283 	sysfs_show_32bit_prop(buffer, offs, "node_from", iolink->node_from);
284 	sysfs_show_32bit_prop(buffer, offs, "node_to", iolink->node_to);
285 	sysfs_show_32bit_prop(buffer, offs, "weight", iolink->weight);
286 	sysfs_show_32bit_prop(buffer, offs, "min_latency", iolink->min_latency);
287 	sysfs_show_32bit_prop(buffer, offs, "max_latency", iolink->max_latency);
288 	sysfs_show_32bit_prop(buffer, offs, "min_bandwidth",
289 			      iolink->min_bandwidth);
290 	sysfs_show_32bit_prop(buffer, offs, "max_bandwidth",
291 			      iolink->max_bandwidth);
292 	sysfs_show_32bit_prop(buffer, offs, "recommended_transfer_size",
293 			      iolink->rec_transfer_size);
294 	sysfs_show_32bit_prop(buffer, offs, "flags", iolink->flags);
295 
296 	return offs;
297 }
298 
299 static const struct sysfs_ops iolink_ops = {
300 	.show = iolink_show,
301 };
302 
303 static const struct kobj_type iolink_type = {
304 	.release = kfd_topology_kobj_release,
305 	.sysfs_ops = &iolink_ops,
306 };
307 
308 static ssize_t mem_show(struct kobject *kobj, struct attribute *attr,
309 		char *buffer)
310 {
311 	int offs = 0;
312 	struct kfd_mem_properties *mem;
313 
314 	/* Making sure that the buffer is an empty string */
315 	buffer[0] = 0;
316 
317 	mem = container_of(attr, struct kfd_mem_properties, attr);
318 	if (mem->gpu && kfd_devcgroup_check_permission(mem->gpu))
319 		return -EPERM;
320 	sysfs_show_32bit_prop(buffer, offs, "heap_type", mem->heap_type);
321 	sysfs_show_64bit_prop(buffer, offs, "size_in_bytes",
322 			      mem->size_in_bytes);
323 	sysfs_show_32bit_prop(buffer, offs, "flags", mem->flags);
324 	sysfs_show_32bit_prop(buffer, offs, "width", mem->width);
325 	sysfs_show_32bit_prop(buffer, offs, "mem_clk_max",
326 			      mem->mem_clk_max);
327 
328 	return offs;
329 }
330 
331 static const struct sysfs_ops mem_ops = {
332 	.show = mem_show,
333 };
334 
335 static const struct kobj_type mem_type = {
336 	.release = kfd_topology_kobj_release,
337 	.sysfs_ops = &mem_ops,
338 };
339 
340 static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr,
341 		char *buffer)
342 {
343 	int offs = 0;
344 	uint32_t i, j;
345 	struct kfd_cache_properties *cache;
346 
347 	/* Making sure that the buffer is an empty string */
348 	buffer[0] = 0;
349 	cache = container_of(attr, struct kfd_cache_properties, attr);
350 	if (cache->gpu && kfd_devcgroup_check_permission(cache->gpu))
351 		return -EPERM;
352 	sysfs_show_32bit_prop(buffer, offs, "processor_id_low",
353 			cache->processor_id_low);
354 	sysfs_show_32bit_prop(buffer, offs, "level", cache->cache_level);
355 	sysfs_show_32bit_prop(buffer, offs, "size", cache->cache_size);
356 	sysfs_show_32bit_prop(buffer, offs, "cache_line_size",
357 			      cache->cacheline_size);
358 	sysfs_show_32bit_prop(buffer, offs, "cache_lines_per_tag",
359 			      cache->cachelines_per_tag);
360 	sysfs_show_32bit_prop(buffer, offs, "association", cache->cache_assoc);
361 	sysfs_show_32bit_prop(buffer, offs, "latency", cache->cache_latency);
362 	sysfs_show_32bit_prop(buffer, offs, "type", cache->cache_type);
363 
364 	offs += snprintf(buffer+offs, PAGE_SIZE-offs, "sibling_map ");
365 	for (i = 0; i < cache->sibling_map_size; i++)
366 		for (j = 0; j < sizeof(cache->sibling_map[0])*8; j++)
367 			/* Check each bit */
368 			offs += snprintf(buffer+offs, PAGE_SIZE-offs, "%d,",
369 						(cache->sibling_map[i] >> j) & 1);
370 
371 	/* Replace the last "," with end of line */
372 	buffer[offs-1] = '\n';
373 	return offs;
374 }
375 
376 static const struct sysfs_ops cache_ops = {
377 	.show = kfd_cache_show,
378 };
379 
380 static const struct kobj_type cache_type = {
381 	.release = kfd_topology_kobj_release,
382 	.sysfs_ops = &cache_ops,
383 };
384 
385 /****** Sysfs of Performance Counters ******/
386 
387 struct kfd_perf_attr {
388 	struct kobj_attribute attr;
389 	uint32_t data;
390 };
391 
392 static ssize_t perf_show(struct kobject *kobj, struct kobj_attribute *attrs,
393 			char *buf)
394 {
395 	int offs = 0;
396 	struct kfd_perf_attr *attr;
397 
398 	buf[0] = 0;
399 	attr = container_of(attrs, struct kfd_perf_attr, attr);
400 	if (!attr->data) /* invalid data for PMC */
401 		return 0;
402 	else
403 		return sysfs_show_32bit_val(buf, offs, attr->data);
404 }
405 
406 #define KFD_PERF_DESC(_name, _data)			\
407 {							\
408 	.attr  = __ATTR(_name, 0444, perf_show, NULL),	\
409 	.data = _data,					\
410 }
411 
412 static struct kfd_perf_attr perf_attr_iommu[] = {
413 	KFD_PERF_DESC(max_concurrent, 0),
414 	KFD_PERF_DESC(num_counters, 0),
415 	KFD_PERF_DESC(counter_ids, 0),
416 };
417 /****************************************/
418 
419 static ssize_t node_show(struct kobject *kobj, struct attribute *attr,
420 		char *buffer)
421 {
422 	int offs = 0;
423 	struct kfd_topology_device *dev;
424 	uint32_t log_max_watch_addr;
425 
426 	/* Making sure that the buffer is an empty string */
427 	buffer[0] = 0;
428 
429 	if (strcmp(attr->name, "gpu_id") == 0) {
430 		dev = container_of(attr, struct kfd_topology_device,
431 				attr_gpuid);
432 		if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu))
433 			return -EPERM;
434 		return sysfs_show_32bit_val(buffer, offs, dev->gpu_id);
435 	}
436 
437 	if (strcmp(attr->name, "name") == 0) {
438 		dev = container_of(attr, struct kfd_topology_device,
439 				attr_name);
440 
441 		if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu))
442 			return -EPERM;
443 		return sysfs_show_str_val(buffer, offs, dev->node_props.name);
444 	}
445 
446 	dev = container_of(attr, struct kfd_topology_device,
447 			attr_props);
448 	if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu))
449 		return -EPERM;
450 	sysfs_show_32bit_prop(buffer, offs, "cpu_cores_count",
451 			      dev->node_props.cpu_cores_count);
452 	sysfs_show_32bit_prop(buffer, offs, "simd_count",
453 			      dev->gpu ? dev->node_props.simd_count : 0);
454 	sysfs_show_32bit_prop(buffer, offs, "mem_banks_count",
455 			      dev->node_props.mem_banks_count);
456 	sysfs_show_32bit_prop(buffer, offs, "caches_count",
457 			      dev->node_props.caches_count);
458 	sysfs_show_32bit_prop(buffer, offs, "io_links_count",
459 			      dev->node_props.io_links_count);
460 	sysfs_show_32bit_prop(buffer, offs, "p2p_links_count",
461 			      dev->node_props.p2p_links_count);
462 	sysfs_show_32bit_prop(buffer, offs, "cpu_core_id_base",
463 			      dev->node_props.cpu_core_id_base);
464 	sysfs_show_32bit_prop(buffer, offs, "simd_id_base",
465 			      dev->node_props.simd_id_base);
466 	sysfs_show_32bit_prop(buffer, offs, "max_waves_per_simd",
467 			      dev->node_props.max_waves_per_simd);
468 	sysfs_show_32bit_prop(buffer, offs, "lds_size_in_kb",
469 			      dev->node_props.lds_size_in_kb);
470 	sysfs_show_32bit_prop(buffer, offs, "gds_size_in_kb",
471 			      dev->node_props.gds_size_in_kb);
472 	sysfs_show_32bit_prop(buffer, offs, "num_gws",
473 			      dev->node_props.num_gws);
474 	sysfs_show_32bit_prop(buffer, offs, "wave_front_size",
475 			      dev->node_props.wave_front_size);
476 	sysfs_show_32bit_prop(buffer, offs, "array_count",
477 			      dev->gpu ? (dev->node_props.array_count *
478 					  NUM_XCC(dev->gpu->xcc_mask)) : 0);
479 	sysfs_show_32bit_prop(buffer, offs, "simd_arrays_per_engine",
480 			      dev->node_props.simd_arrays_per_engine);
481 	sysfs_show_32bit_prop(buffer, offs, "cu_per_simd_array",
482 			      dev->node_props.cu_per_simd_array);
483 	sysfs_show_32bit_prop(buffer, offs, "simd_per_cu",
484 			      dev->node_props.simd_per_cu);
485 	sysfs_show_32bit_prop(buffer, offs, "max_slots_scratch_cu",
486 			      dev->node_props.max_slots_scratch_cu);
487 	sysfs_show_32bit_prop(buffer, offs, "gfx_target_version",
488 			      dev->node_props.gfx_target_version);
489 	sysfs_show_32bit_prop(buffer, offs, "vendor_id",
490 			      dev->node_props.vendor_id);
491 	sysfs_show_32bit_prop(buffer, offs, "device_id",
492 			      dev->node_props.device_id);
493 	sysfs_show_32bit_prop(buffer, offs, "location_id",
494 			      dev->node_props.location_id);
495 	sysfs_show_32bit_prop(buffer, offs, "domain",
496 			      dev->node_props.domain);
497 	sysfs_show_32bit_prop(buffer, offs, "drm_render_minor",
498 			      dev->node_props.drm_render_minor);
499 	sysfs_show_64bit_prop(buffer, offs, "hive_id",
500 			      dev->node_props.hive_id);
501 	sysfs_show_32bit_prop(buffer, offs, "num_sdma_engines",
502 			      dev->node_props.num_sdma_engines);
503 	sysfs_show_32bit_prop(buffer, offs, "num_sdma_xgmi_engines",
504 			      dev->node_props.num_sdma_xgmi_engines);
505 	sysfs_show_32bit_prop(buffer, offs, "num_sdma_queues_per_engine",
506 			      dev->node_props.num_sdma_queues_per_engine);
507 	sysfs_show_32bit_prop(buffer, offs, "num_cp_queues",
508 			      dev->node_props.num_cp_queues);
509 
510 	if (dev->gpu) {
511 		log_max_watch_addr =
512 			__ilog2_u32(dev->gpu->kfd->device_info.num_of_watch_points);
513 
514 		if (log_max_watch_addr) {
515 			dev->node_props.capability |=
516 					HSA_CAP_WATCH_POINTS_SUPPORTED;
517 
518 			dev->node_props.capability |=
519 				((log_max_watch_addr <<
520 					HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) &
521 				HSA_CAP_WATCH_POINTS_TOTALBITS_MASK);
522 		}
523 
524 		if (dev->gpu->adev->asic_type == CHIP_TONGA)
525 			dev->node_props.capability |=
526 					HSA_CAP_AQL_QUEUE_DOUBLE_MAP;
527 
528 		sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_fcompute",
529 			dev->node_props.max_engine_clk_fcompute);
530 
531 		sysfs_show_64bit_prop(buffer, offs, "local_mem_size", 0ULL);
532 
533 		sysfs_show_32bit_prop(buffer, offs, "fw_version",
534 				      dev->gpu->kfd->mec_fw_version);
535 		sysfs_show_32bit_prop(buffer, offs, "capability",
536 				      dev->node_props.capability);
537 		sysfs_show_64bit_prop(buffer, offs, "debug_prop",
538 				      dev->node_props.debug_prop);
539 		sysfs_show_32bit_prop(buffer, offs, "sdma_fw_version",
540 				      dev->gpu->kfd->sdma_fw_version);
541 		sysfs_show_64bit_prop(buffer, offs, "unique_id",
542 				      dev->gpu->adev->unique_id);
543 		sysfs_show_32bit_prop(buffer, offs, "num_xcc",
544 				      NUM_XCC(dev->gpu->xcc_mask));
545 	}
546 
547 	return sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_ccompute",
548 				     cpufreq_quick_get_max(0)/1000);
549 }
550 
551 static const struct sysfs_ops node_ops = {
552 	.show = node_show,
553 };
554 
555 static const struct kobj_type node_type = {
556 	.release = kfd_topology_kobj_release,
557 	.sysfs_ops = &node_ops,
558 };
559 
560 static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr)
561 {
562 	sysfs_remove_file(kobj, attr);
563 	kobject_del(kobj);
564 	kobject_put(kobj);
565 }
566 
567 static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev)
568 {
569 	struct kfd_iolink_properties *p2plink;
570 	struct kfd_iolink_properties *iolink;
571 	struct kfd_cache_properties *cache;
572 	struct kfd_mem_properties *mem;
573 	struct kfd_perf_properties *perf;
574 
575 	if (dev->kobj_iolink) {
576 		list_for_each_entry(iolink, &dev->io_link_props, list)
577 			if (iolink->kobj) {
578 				kfd_remove_sysfs_file(iolink->kobj,
579 							&iolink->attr);
580 				iolink->kobj = NULL;
581 			}
582 		kobject_del(dev->kobj_iolink);
583 		kobject_put(dev->kobj_iolink);
584 		dev->kobj_iolink = NULL;
585 	}
586 
587 	if (dev->kobj_p2plink) {
588 		list_for_each_entry(p2plink, &dev->p2p_link_props, list)
589 			if (p2plink->kobj) {
590 				kfd_remove_sysfs_file(p2plink->kobj,
591 							&p2plink->attr);
592 				p2plink->kobj = NULL;
593 			}
594 		kobject_del(dev->kobj_p2plink);
595 		kobject_put(dev->kobj_p2plink);
596 		dev->kobj_p2plink = NULL;
597 	}
598 
599 	if (dev->kobj_cache) {
600 		list_for_each_entry(cache, &dev->cache_props, list)
601 			if (cache->kobj) {
602 				kfd_remove_sysfs_file(cache->kobj,
603 							&cache->attr);
604 				cache->kobj = NULL;
605 			}
606 		kobject_del(dev->kobj_cache);
607 		kobject_put(dev->kobj_cache);
608 		dev->kobj_cache = NULL;
609 	}
610 
611 	if (dev->kobj_mem) {
612 		list_for_each_entry(mem, &dev->mem_props, list)
613 			if (mem->kobj) {
614 				kfd_remove_sysfs_file(mem->kobj, &mem->attr);
615 				mem->kobj = NULL;
616 			}
617 		kobject_del(dev->kobj_mem);
618 		kobject_put(dev->kobj_mem);
619 		dev->kobj_mem = NULL;
620 	}
621 
622 	if (dev->kobj_perf) {
623 		list_for_each_entry(perf, &dev->perf_props, list) {
624 			kfree(perf->attr_group);
625 			perf->attr_group = NULL;
626 		}
627 		kobject_del(dev->kobj_perf);
628 		kobject_put(dev->kobj_perf);
629 		dev->kobj_perf = NULL;
630 	}
631 
632 	if (dev->kobj_node) {
633 		sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid);
634 		sysfs_remove_file(dev->kobj_node, &dev->attr_name);
635 		sysfs_remove_file(dev->kobj_node, &dev->attr_props);
636 		kobject_del(dev->kobj_node);
637 		kobject_put(dev->kobj_node);
638 		dev->kobj_node = NULL;
639 	}
640 }
641 
642 static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev,
643 		uint32_t id)
644 {
645 	struct kfd_iolink_properties *p2plink;
646 	struct kfd_iolink_properties *iolink;
647 	struct kfd_cache_properties *cache;
648 	struct kfd_mem_properties *mem;
649 	struct kfd_perf_properties *perf;
650 	int ret;
651 	uint32_t i, num_attrs;
652 	struct attribute **attrs;
653 
654 	if (WARN_ON(dev->kobj_node))
655 		return -EEXIST;
656 
657 	/*
658 	 * Creating the sysfs folders
659 	 */
660 	dev->kobj_node = kfd_alloc_struct(dev->kobj_node);
661 	if (!dev->kobj_node)
662 		return -ENOMEM;
663 
664 	ret = kobject_init_and_add(dev->kobj_node, &node_type,
665 			sys_props.kobj_nodes, "%d", id);
666 	if (ret < 0) {
667 		kobject_put(dev->kobj_node);
668 		return ret;
669 	}
670 
671 	dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node);
672 	if (!dev->kobj_mem)
673 		return -ENOMEM;
674 
675 	dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node);
676 	if (!dev->kobj_cache)
677 		return -ENOMEM;
678 
679 	dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node);
680 	if (!dev->kobj_iolink)
681 		return -ENOMEM;
682 
683 	dev->kobj_p2plink = kobject_create_and_add("p2p_links", dev->kobj_node);
684 	if (!dev->kobj_p2plink)
685 		return -ENOMEM;
686 
687 	dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node);
688 	if (!dev->kobj_perf)
689 		return -ENOMEM;
690 
691 	/*
692 	 * Creating sysfs files for node properties
693 	 */
694 	dev->attr_gpuid.name = "gpu_id";
695 	dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE;
696 	sysfs_attr_init(&dev->attr_gpuid);
697 	dev->attr_name.name = "name";
698 	dev->attr_name.mode = KFD_SYSFS_FILE_MODE;
699 	sysfs_attr_init(&dev->attr_name);
700 	dev->attr_props.name = "properties";
701 	dev->attr_props.mode = KFD_SYSFS_FILE_MODE;
702 	sysfs_attr_init(&dev->attr_props);
703 	ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid);
704 	if (ret < 0)
705 		return ret;
706 	ret = sysfs_create_file(dev->kobj_node, &dev->attr_name);
707 	if (ret < 0)
708 		return ret;
709 	ret = sysfs_create_file(dev->kobj_node, &dev->attr_props);
710 	if (ret < 0)
711 		return ret;
712 
713 	i = 0;
714 	list_for_each_entry(mem, &dev->mem_props, list) {
715 		mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
716 		if (!mem->kobj)
717 			return -ENOMEM;
718 		ret = kobject_init_and_add(mem->kobj, &mem_type,
719 				dev->kobj_mem, "%d", i);
720 		if (ret < 0) {
721 			kobject_put(mem->kobj);
722 			return ret;
723 		}
724 
725 		mem->attr.name = "properties";
726 		mem->attr.mode = KFD_SYSFS_FILE_MODE;
727 		sysfs_attr_init(&mem->attr);
728 		ret = sysfs_create_file(mem->kobj, &mem->attr);
729 		if (ret < 0)
730 			return ret;
731 		i++;
732 	}
733 
734 	i = 0;
735 	list_for_each_entry(cache, &dev->cache_props, list) {
736 		cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
737 		if (!cache->kobj)
738 			return -ENOMEM;
739 		ret = kobject_init_and_add(cache->kobj, &cache_type,
740 				dev->kobj_cache, "%d", i);
741 		if (ret < 0) {
742 			kobject_put(cache->kobj);
743 			return ret;
744 		}
745 
746 		cache->attr.name = "properties";
747 		cache->attr.mode = KFD_SYSFS_FILE_MODE;
748 		sysfs_attr_init(&cache->attr);
749 		ret = sysfs_create_file(cache->kobj, &cache->attr);
750 		if (ret < 0)
751 			return ret;
752 		i++;
753 	}
754 
755 	i = 0;
756 	list_for_each_entry(iolink, &dev->io_link_props, list) {
757 		iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
758 		if (!iolink->kobj)
759 			return -ENOMEM;
760 		ret = kobject_init_and_add(iolink->kobj, &iolink_type,
761 				dev->kobj_iolink, "%d", i);
762 		if (ret < 0) {
763 			kobject_put(iolink->kobj);
764 			return ret;
765 		}
766 
767 		iolink->attr.name = "properties";
768 		iolink->attr.mode = KFD_SYSFS_FILE_MODE;
769 		sysfs_attr_init(&iolink->attr);
770 		ret = sysfs_create_file(iolink->kobj, &iolink->attr);
771 		if (ret < 0)
772 			return ret;
773 		i++;
774 	}
775 
776 	i = 0;
777 	list_for_each_entry(p2plink, &dev->p2p_link_props, list) {
778 		p2plink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
779 		if (!p2plink->kobj)
780 			return -ENOMEM;
781 		ret = kobject_init_and_add(p2plink->kobj, &iolink_type,
782 				dev->kobj_p2plink, "%d", i);
783 		if (ret < 0) {
784 			kobject_put(p2plink->kobj);
785 			return ret;
786 		}
787 
788 		p2plink->attr.name = "properties";
789 		p2plink->attr.mode = KFD_SYSFS_FILE_MODE;
790 		sysfs_attr_init(&p2plink->attr);
791 		ret = sysfs_create_file(p2plink->kobj, &p2plink->attr);
792 		if (ret < 0)
793 			return ret;
794 		i++;
795 	}
796 
797 	/* All hardware blocks have the same number of attributes. */
798 	num_attrs = ARRAY_SIZE(perf_attr_iommu);
799 	list_for_each_entry(perf, &dev->perf_props, list) {
800 		perf->attr_group = kzalloc(sizeof(struct kfd_perf_attr)
801 			* num_attrs + sizeof(struct attribute_group),
802 			GFP_KERNEL);
803 		if (!perf->attr_group)
804 			return -ENOMEM;
805 
806 		attrs = (struct attribute **)(perf->attr_group + 1);
807 		if (!strcmp(perf->block_name, "iommu")) {
808 		/* Information of IOMMU's num_counters and counter_ids is shown
809 		 * under /sys/bus/event_source/devices/amd_iommu. We don't
810 		 * duplicate here.
811 		 */
812 			perf_attr_iommu[0].data = perf->max_concurrent;
813 			for (i = 0; i < num_attrs; i++)
814 				attrs[i] = &perf_attr_iommu[i].attr.attr;
815 		}
816 		perf->attr_group->name = perf->block_name;
817 		perf->attr_group->attrs = attrs;
818 		ret = sysfs_create_group(dev->kobj_perf, perf->attr_group);
819 		if (ret < 0)
820 			return ret;
821 	}
822 
823 	return 0;
824 }
825 
826 /* Called with write topology lock acquired */
827 static int kfd_build_sysfs_node_tree(void)
828 {
829 	struct kfd_topology_device *dev;
830 	int ret;
831 	uint32_t i = 0;
832 
833 	list_for_each_entry(dev, &topology_device_list, list) {
834 		ret = kfd_build_sysfs_node_entry(dev, i);
835 		if (ret < 0)
836 			return ret;
837 		i++;
838 	}
839 
840 	return 0;
841 }
842 
843 /* Called with write topology lock acquired */
844 static void kfd_remove_sysfs_node_tree(void)
845 {
846 	struct kfd_topology_device *dev;
847 
848 	list_for_each_entry(dev, &topology_device_list, list)
849 		kfd_remove_sysfs_node_entry(dev);
850 }
851 
852 static int kfd_topology_update_sysfs(void)
853 {
854 	int ret;
855 
856 	if (!sys_props.kobj_topology) {
857 		sys_props.kobj_topology =
858 				kfd_alloc_struct(sys_props.kobj_topology);
859 		if (!sys_props.kobj_topology)
860 			return -ENOMEM;
861 
862 		ret = kobject_init_and_add(sys_props.kobj_topology,
863 				&sysprops_type,  &kfd_device->kobj,
864 				"topology");
865 		if (ret < 0) {
866 			kobject_put(sys_props.kobj_topology);
867 			return ret;
868 		}
869 
870 		sys_props.kobj_nodes = kobject_create_and_add("nodes",
871 				sys_props.kobj_topology);
872 		if (!sys_props.kobj_nodes)
873 			return -ENOMEM;
874 
875 		sys_props.attr_genid.name = "generation_id";
876 		sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE;
877 		sysfs_attr_init(&sys_props.attr_genid);
878 		ret = sysfs_create_file(sys_props.kobj_topology,
879 				&sys_props.attr_genid);
880 		if (ret < 0)
881 			return ret;
882 
883 		sys_props.attr_props.name = "system_properties";
884 		sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE;
885 		sysfs_attr_init(&sys_props.attr_props);
886 		ret = sysfs_create_file(sys_props.kobj_topology,
887 				&sys_props.attr_props);
888 		if (ret < 0)
889 			return ret;
890 	}
891 
892 	kfd_remove_sysfs_node_tree();
893 
894 	return kfd_build_sysfs_node_tree();
895 }
896 
897 static void kfd_topology_release_sysfs(void)
898 {
899 	kfd_remove_sysfs_node_tree();
900 	if (sys_props.kobj_topology) {
901 		sysfs_remove_file(sys_props.kobj_topology,
902 				&sys_props.attr_genid);
903 		sysfs_remove_file(sys_props.kobj_topology,
904 				&sys_props.attr_props);
905 		if (sys_props.kobj_nodes) {
906 			kobject_del(sys_props.kobj_nodes);
907 			kobject_put(sys_props.kobj_nodes);
908 			sys_props.kobj_nodes = NULL;
909 		}
910 		kobject_del(sys_props.kobj_topology);
911 		kobject_put(sys_props.kobj_topology);
912 		sys_props.kobj_topology = NULL;
913 	}
914 }
915 
916 /* Called with write topology_lock acquired */
917 static void kfd_topology_update_device_list(struct list_head *temp_list,
918 					struct list_head *master_list)
919 {
920 	while (!list_empty(temp_list)) {
921 		list_move_tail(temp_list->next, master_list);
922 		sys_props.num_devices++;
923 	}
924 }
925 
926 static void kfd_debug_print_topology(void)
927 {
928 	struct kfd_topology_device *dev;
929 
930 	down_read(&topology_lock);
931 
932 	dev = list_last_entry(&topology_device_list,
933 			struct kfd_topology_device, list);
934 	if (dev) {
935 		if (dev->node_props.cpu_cores_count &&
936 				dev->node_props.simd_count) {
937 			pr_info("Topology: Add APU node [0x%0x:0x%0x]\n",
938 				dev->node_props.device_id,
939 				dev->node_props.vendor_id);
940 		} else if (dev->node_props.cpu_cores_count)
941 			pr_info("Topology: Add CPU node\n");
942 		else if (dev->node_props.simd_count)
943 			pr_info("Topology: Add dGPU node [0x%0x:0x%0x]\n",
944 				dev->node_props.device_id,
945 				dev->node_props.vendor_id);
946 	}
947 	up_read(&topology_lock);
948 }
949 
950 /* Helper function for intializing platform_xx members of
951  * kfd_system_properties. Uses OEM info from the last CPU/APU node.
952  */
953 static void kfd_update_system_properties(void)
954 {
955 	struct kfd_topology_device *dev;
956 
957 	down_read(&topology_lock);
958 	dev = list_last_entry(&topology_device_list,
959 			struct kfd_topology_device, list);
960 	if (dev) {
961 		sys_props.platform_id = dev->oem_id64;
962 		sys_props.platform_oem = *((uint64_t *)dev->oem_table_id);
963 		sys_props.platform_rev = dev->oem_revision;
964 	}
965 	up_read(&topology_lock);
966 }
967 
968 static void find_system_memory(const struct dmi_header *dm,
969 	void *private)
970 {
971 	struct kfd_mem_properties *mem;
972 	u16 mem_width, mem_clock;
973 	struct kfd_topology_device *kdev =
974 		(struct kfd_topology_device *)private;
975 	const u8 *dmi_data = (const u8 *)(dm + 1);
976 
977 	if (dm->type == DMI_ENTRY_MEM_DEVICE && dm->length >= 0x15) {
978 		mem_width = (u16)(*(const u16 *)(dmi_data + 0x6));
979 		mem_clock = (u16)(*(const u16 *)(dmi_data + 0x11));
980 		list_for_each_entry(mem, &kdev->mem_props, list) {
981 			if (mem_width != 0xFFFF && mem_width != 0)
982 				mem->width = mem_width;
983 			if (mem_clock != 0)
984 				mem->mem_clk_max = mem_clock;
985 		}
986 	}
987 }
988 
989 /* kfd_add_non_crat_information - Add information that is not currently
990  *	defined in CRAT but is necessary for KFD topology
991  * @dev - topology device to which addition info is added
992  */
993 static void kfd_add_non_crat_information(struct kfd_topology_device *kdev)
994 {
995 	/* Check if CPU only node. */
996 	if (!kdev->gpu) {
997 		/* Add system memory information */
998 		dmi_walk(find_system_memory, kdev);
999 	}
1000 	/* TODO: For GPU node, rearrange code from kfd_topology_add_device */
1001 }
1002 
1003 int kfd_topology_init(void)
1004 {
1005 	void *crat_image = NULL;
1006 	size_t image_size = 0;
1007 	int ret;
1008 	struct list_head temp_topology_device_list;
1009 	int cpu_only_node = 0;
1010 	struct kfd_topology_device *kdev;
1011 	int proximity_domain;
1012 
1013 	/* topology_device_list - Master list of all topology devices
1014 	 * temp_topology_device_list - temporary list created while parsing CRAT
1015 	 * or VCRAT. Once parsing is complete the contents of list is moved to
1016 	 * topology_device_list
1017 	 */
1018 
1019 	/* Initialize the head for the both the lists */
1020 	INIT_LIST_HEAD(&topology_device_list);
1021 	INIT_LIST_HEAD(&temp_topology_device_list);
1022 	init_rwsem(&topology_lock);
1023 
1024 	memset(&sys_props, 0, sizeof(sys_props));
1025 
1026 	/* Proximity domains in ACPI CRAT tables start counting at
1027 	 * 0. The same should be true for virtual CRAT tables created
1028 	 * at this stage. GPUs added later in kfd_topology_add_device
1029 	 * use a counter.
1030 	 */
1031 	proximity_domain = 0;
1032 
1033 	ret = kfd_create_crat_image_virtual(&crat_image, &image_size,
1034 					    COMPUTE_UNIT_CPU, NULL,
1035 					    proximity_domain);
1036 	cpu_only_node = 1;
1037 	if (ret) {
1038 		pr_err("Error creating VCRAT table for CPU\n");
1039 		return ret;
1040 	}
1041 
1042 	ret = kfd_parse_crat_table(crat_image,
1043 				   &temp_topology_device_list,
1044 				   proximity_domain);
1045 	if (ret) {
1046 		pr_err("Error parsing VCRAT table for CPU\n");
1047 		goto err;
1048 	}
1049 
1050 	kdev = list_first_entry(&temp_topology_device_list,
1051 				struct kfd_topology_device, list);
1052 
1053 	down_write(&topology_lock);
1054 	kfd_topology_update_device_list(&temp_topology_device_list,
1055 					&topology_device_list);
1056 	topology_crat_proximity_domain = sys_props.num_devices-1;
1057 	ret = kfd_topology_update_sysfs();
1058 	up_write(&topology_lock);
1059 
1060 	if (!ret) {
1061 		sys_props.generation_count++;
1062 		kfd_update_system_properties();
1063 		kfd_debug_print_topology();
1064 	} else
1065 		pr_err("Failed to update topology in sysfs ret=%d\n", ret);
1066 
1067 	/* For nodes with GPU, this information gets added
1068 	 * when GPU is detected (kfd_topology_add_device).
1069 	 */
1070 	if (cpu_only_node) {
1071 		/* Add additional information to CPU only node created above */
1072 		down_write(&topology_lock);
1073 		kdev = list_first_entry(&topology_device_list,
1074 				struct kfd_topology_device, list);
1075 		up_write(&topology_lock);
1076 		kfd_add_non_crat_information(kdev);
1077 	}
1078 
1079 err:
1080 	kfd_destroy_crat_image(crat_image);
1081 	return ret;
1082 }
1083 
1084 void kfd_topology_shutdown(void)
1085 {
1086 	down_write(&topology_lock);
1087 	kfd_topology_release_sysfs();
1088 	kfd_release_live_view();
1089 	up_write(&topology_lock);
1090 }
1091 
1092 static uint32_t kfd_generate_gpu_id(struct kfd_node *gpu)
1093 {
1094 	uint32_t hashout;
1095 	uint32_t buf[8];
1096 	uint64_t local_mem_size;
1097 	int i;
1098 
1099 	if (!gpu)
1100 		return 0;
1101 
1102 	local_mem_size = gpu->local_mem_info.local_mem_size_private +
1103 			gpu->local_mem_info.local_mem_size_public;
1104 	buf[0] = gpu->adev->pdev->devfn;
1105 	buf[1] = gpu->adev->pdev->subsystem_vendor |
1106 		(gpu->adev->pdev->subsystem_device << 16);
1107 	buf[2] = pci_domain_nr(gpu->adev->pdev->bus);
1108 	buf[3] = gpu->adev->pdev->device;
1109 	buf[4] = gpu->adev->pdev->bus->number;
1110 	buf[5] = lower_32_bits(local_mem_size);
1111 	buf[6] = upper_32_bits(local_mem_size);
1112 	buf[7] = (ffs(gpu->xcc_mask) - 1) | (NUM_XCC(gpu->xcc_mask) << 16);
1113 
1114 	for (i = 0, hashout = 0; i < 8; i++)
1115 		hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH);
1116 
1117 	return hashout;
1118 }
1119 /* kfd_assign_gpu - Attach @gpu to the correct kfd topology device. If
1120  *		the GPU device is not already present in the topology device
1121  *		list then return NULL. This means a new topology device has to
1122  *		be created for this GPU.
1123  */
1124 static struct kfd_topology_device *kfd_assign_gpu(struct kfd_node *gpu)
1125 {
1126 	struct kfd_topology_device *dev;
1127 	struct kfd_topology_device *out_dev = NULL;
1128 	struct kfd_mem_properties *mem;
1129 	struct kfd_cache_properties *cache;
1130 	struct kfd_iolink_properties *iolink;
1131 	struct kfd_iolink_properties *p2plink;
1132 
1133 	list_for_each_entry(dev, &topology_device_list, list) {
1134 		/* Discrete GPUs need their own topology device list
1135 		 * entries. Don't assign them to CPU/APU nodes.
1136 		 */
1137 		if (dev->node_props.cpu_cores_count)
1138 			continue;
1139 
1140 		if (!dev->gpu && (dev->node_props.simd_count > 0)) {
1141 			dev->gpu = gpu;
1142 			out_dev = dev;
1143 
1144 			list_for_each_entry(mem, &dev->mem_props, list)
1145 				mem->gpu = dev->gpu;
1146 			list_for_each_entry(cache, &dev->cache_props, list)
1147 				cache->gpu = dev->gpu;
1148 			list_for_each_entry(iolink, &dev->io_link_props, list)
1149 				iolink->gpu = dev->gpu;
1150 			list_for_each_entry(p2plink, &dev->p2p_link_props, list)
1151 				p2plink->gpu = dev->gpu;
1152 			break;
1153 		}
1154 	}
1155 	return out_dev;
1156 }
1157 
1158 static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival)
1159 {
1160 	/*
1161 	 * TODO: Generate an event for thunk about the arrival/removal
1162 	 * of the GPU
1163 	 */
1164 }
1165 
1166 /* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info,
1167  *		patch this after CRAT parsing.
1168  */
1169 static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev)
1170 {
1171 	struct kfd_mem_properties *mem;
1172 	struct kfd_local_mem_info local_mem_info;
1173 
1174 	if (!dev)
1175 		return;
1176 
1177 	/* Currently, amdgpu driver (amdgpu_mc) deals only with GPUs with
1178 	 * single bank of VRAM local memory.
1179 	 * for dGPUs - VCRAT reports only one bank of Local Memory
1180 	 * for APUs - If CRAT from ACPI reports more than one bank, then
1181 	 *	all the banks will report the same mem_clk_max information
1182 	 */
1183 	amdgpu_amdkfd_get_local_mem_info(dev->gpu->adev, &local_mem_info,
1184 					 dev->gpu->xcp);
1185 
1186 	list_for_each_entry(mem, &dev->mem_props, list)
1187 		mem->mem_clk_max = local_mem_info.mem_clk_max;
1188 }
1189 
1190 static void kfd_set_iolink_no_atomics(struct kfd_topology_device *dev,
1191 					struct kfd_topology_device *target_gpu_dev,
1192 					struct kfd_iolink_properties *link)
1193 {
1194 	/* xgmi always supports atomics between links. */
1195 	if (link->iolink_type == CRAT_IOLINK_TYPE_XGMI)
1196 		return;
1197 
1198 	/* check pcie support to set cpu(dev) flags for target_gpu_dev link. */
1199 	if (target_gpu_dev) {
1200 		uint32_t cap;
1201 
1202 		pcie_capability_read_dword(target_gpu_dev->gpu->adev->pdev,
1203 				PCI_EXP_DEVCAP2, &cap);
1204 
1205 		if (!(cap & (PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
1206 			     PCI_EXP_DEVCAP2_ATOMIC_COMP64)))
1207 			link->flags |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1208 				CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1209 	/* set gpu (dev) flags. */
1210 	} else {
1211 		if (!dev->gpu->kfd->pci_atomic_requested ||
1212 				dev->gpu->adev->asic_type == CHIP_HAWAII)
1213 			link->flags |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1214 				CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1215 	}
1216 }
1217 
1218 static void kfd_set_iolink_non_coherent(struct kfd_topology_device *to_dev,
1219 		struct kfd_iolink_properties *outbound_link,
1220 		struct kfd_iolink_properties *inbound_link)
1221 {
1222 	/* CPU -> GPU with PCIe */
1223 	if (!to_dev->gpu &&
1224 	    inbound_link->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS)
1225 		inbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT;
1226 
1227 	if (to_dev->gpu) {
1228 		/* GPU <-> GPU with PCIe and
1229 		 * Vega20 with XGMI
1230 		 */
1231 		if (inbound_link->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS ||
1232 		    (inbound_link->iolink_type == CRAT_IOLINK_TYPE_XGMI &&
1233 		    KFD_GC_VERSION(to_dev->gpu) == IP_VERSION(9, 4, 0))) {
1234 			outbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT;
1235 			inbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT;
1236 		}
1237 	}
1238 }
1239 
1240 static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev)
1241 {
1242 	struct kfd_iolink_properties *link, *inbound_link;
1243 	struct kfd_topology_device *peer_dev;
1244 
1245 	if (!dev || !dev->gpu)
1246 		return;
1247 
1248 	/* GPU only creates direct links so apply flags setting to all */
1249 	list_for_each_entry(link, &dev->io_link_props, list) {
1250 		link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1251 		kfd_set_iolink_no_atomics(dev, NULL, link);
1252 		peer_dev = kfd_topology_device_by_proximity_domain(
1253 				link->node_to);
1254 
1255 		if (!peer_dev)
1256 			continue;
1257 
1258 		/* Include the CPU peer in GPU hive if connected over xGMI. */
1259 		if (!peer_dev->gpu &&
1260 		    link->iolink_type == CRAT_IOLINK_TYPE_XGMI) {
1261 			/*
1262 			 * If the GPU is not part of a GPU hive, use its pci
1263 			 * device location as the hive ID to bind with the CPU.
1264 			 */
1265 			if (!dev->node_props.hive_id)
1266 				dev->node_props.hive_id = pci_dev_id(dev->gpu->adev->pdev);
1267 			peer_dev->node_props.hive_id = dev->node_props.hive_id;
1268 		}
1269 
1270 		list_for_each_entry(inbound_link, &peer_dev->io_link_props,
1271 									list) {
1272 			if (inbound_link->node_to != link->node_from)
1273 				continue;
1274 
1275 			inbound_link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1276 			kfd_set_iolink_no_atomics(peer_dev, dev, inbound_link);
1277 			kfd_set_iolink_non_coherent(peer_dev, link, inbound_link);
1278 		}
1279 	}
1280 
1281 	/* Create indirect links so apply flags setting to all */
1282 	list_for_each_entry(link, &dev->p2p_link_props, list) {
1283 		link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1284 		kfd_set_iolink_no_atomics(dev, NULL, link);
1285 		peer_dev = kfd_topology_device_by_proximity_domain(
1286 				link->node_to);
1287 
1288 		if (!peer_dev)
1289 			continue;
1290 
1291 		list_for_each_entry(inbound_link, &peer_dev->p2p_link_props,
1292 									list) {
1293 			if (inbound_link->node_to != link->node_from)
1294 				continue;
1295 
1296 			inbound_link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1297 			kfd_set_iolink_no_atomics(peer_dev, dev, inbound_link);
1298 			kfd_set_iolink_non_coherent(peer_dev, link, inbound_link);
1299 		}
1300 	}
1301 }
1302 
1303 static int kfd_build_p2p_node_entry(struct kfd_topology_device *dev,
1304 				struct kfd_iolink_properties *p2plink)
1305 {
1306 	int ret;
1307 
1308 	p2plink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
1309 	if (!p2plink->kobj)
1310 		return -ENOMEM;
1311 
1312 	ret = kobject_init_and_add(p2plink->kobj, &iolink_type,
1313 			dev->kobj_p2plink, "%d", dev->node_props.p2p_links_count - 1);
1314 	if (ret < 0) {
1315 		kobject_put(p2plink->kobj);
1316 		return ret;
1317 	}
1318 
1319 	p2plink->attr.name = "properties";
1320 	p2plink->attr.mode = KFD_SYSFS_FILE_MODE;
1321 	sysfs_attr_init(&p2plink->attr);
1322 	ret = sysfs_create_file(p2plink->kobj, &p2plink->attr);
1323 	if (ret < 0)
1324 		return ret;
1325 
1326 	return 0;
1327 }
1328 
1329 static int kfd_create_indirect_link_prop(struct kfd_topology_device *kdev, int gpu_node)
1330 {
1331 	struct kfd_iolink_properties *gpu_link, *tmp_link, *cpu_link;
1332 	struct kfd_iolink_properties *props = NULL, *props2 = NULL;
1333 	struct kfd_topology_device *cpu_dev;
1334 	int ret = 0;
1335 	int i, num_cpu;
1336 
1337 	num_cpu = 0;
1338 	list_for_each_entry(cpu_dev, &topology_device_list, list) {
1339 		if (cpu_dev->gpu)
1340 			break;
1341 		num_cpu++;
1342 	}
1343 
1344 	if (list_empty(&kdev->io_link_props))
1345 		return -ENODATA;
1346 
1347 	gpu_link = list_first_entry(&kdev->io_link_props,
1348 				    struct kfd_iolink_properties, list);
1349 
1350 	for (i = 0; i < num_cpu; i++) {
1351 		/* CPU <--> GPU */
1352 		if (gpu_link->node_to == i)
1353 			continue;
1354 
1355 		/* find CPU <-->  CPU links */
1356 		cpu_link = NULL;
1357 		cpu_dev = kfd_topology_device_by_proximity_domain(i);
1358 		if (cpu_dev) {
1359 			list_for_each_entry(tmp_link,
1360 					&cpu_dev->io_link_props, list) {
1361 				if (tmp_link->node_to == gpu_link->node_to) {
1362 					cpu_link = tmp_link;
1363 					break;
1364 				}
1365 			}
1366 		}
1367 
1368 		if (!cpu_link)
1369 			return -ENOMEM;
1370 
1371 		/* CPU <--> CPU <--> GPU, GPU node*/
1372 		props = kfd_alloc_struct(props);
1373 		if (!props)
1374 			return -ENOMEM;
1375 
1376 		memcpy(props, gpu_link, sizeof(struct kfd_iolink_properties));
1377 		props->weight = gpu_link->weight + cpu_link->weight;
1378 		props->min_latency = gpu_link->min_latency + cpu_link->min_latency;
1379 		props->max_latency = gpu_link->max_latency + cpu_link->max_latency;
1380 		props->min_bandwidth = min(gpu_link->min_bandwidth, cpu_link->min_bandwidth);
1381 		props->max_bandwidth = min(gpu_link->max_bandwidth, cpu_link->max_bandwidth);
1382 
1383 		props->node_from = gpu_node;
1384 		props->node_to = i;
1385 		kdev->node_props.p2p_links_count++;
1386 		list_add_tail(&props->list, &kdev->p2p_link_props);
1387 		ret = kfd_build_p2p_node_entry(kdev, props);
1388 		if (ret < 0)
1389 			return ret;
1390 
1391 		/* for small Bar, no CPU --> GPU in-direct links */
1392 		if (kfd_dev_is_large_bar(kdev->gpu)) {
1393 			/* CPU <--> CPU <--> GPU, CPU node*/
1394 			props2 = kfd_alloc_struct(props2);
1395 			if (!props2)
1396 				return -ENOMEM;
1397 
1398 			memcpy(props2, props, sizeof(struct kfd_iolink_properties));
1399 			props2->node_from = i;
1400 			props2->node_to = gpu_node;
1401 			props2->kobj = NULL;
1402 			cpu_dev->node_props.p2p_links_count++;
1403 			list_add_tail(&props2->list, &cpu_dev->p2p_link_props);
1404 			ret = kfd_build_p2p_node_entry(cpu_dev, props2);
1405 			if (ret < 0)
1406 				return ret;
1407 		}
1408 	}
1409 	return ret;
1410 }
1411 
1412 #if defined(CONFIG_HSA_AMD_P2P)
1413 static int kfd_add_peer_prop(struct kfd_topology_device *kdev,
1414 		struct kfd_topology_device *peer, int from, int to)
1415 {
1416 	struct kfd_iolink_properties *props = NULL;
1417 	struct kfd_iolink_properties *iolink1, *iolink2, *iolink3;
1418 	struct kfd_topology_device *cpu_dev;
1419 	int ret = 0;
1420 
1421 	if (!amdgpu_device_is_peer_accessible(
1422 				kdev->gpu->adev,
1423 				peer->gpu->adev))
1424 		return ret;
1425 
1426 	if (list_empty(&kdev->io_link_props))
1427 		return -ENODATA;
1428 
1429 	iolink1 = list_first_entry(&kdev->io_link_props,
1430 				   struct kfd_iolink_properties, list);
1431 
1432 	if (list_empty(&peer->io_link_props))
1433 		return -ENODATA;
1434 
1435 	iolink2 = list_first_entry(&peer->io_link_props,
1436 				   struct kfd_iolink_properties, list);
1437 
1438 	props = kfd_alloc_struct(props);
1439 	if (!props)
1440 		return -ENOMEM;
1441 
1442 	memcpy(props, iolink1, sizeof(struct kfd_iolink_properties));
1443 
1444 	props->weight = iolink1->weight + iolink2->weight;
1445 	props->min_latency = iolink1->min_latency + iolink2->min_latency;
1446 	props->max_latency = iolink1->max_latency + iolink2->max_latency;
1447 	props->min_bandwidth = min(iolink1->min_bandwidth, iolink2->min_bandwidth);
1448 	props->max_bandwidth = min(iolink2->max_bandwidth, iolink2->max_bandwidth);
1449 
1450 	if (iolink1->node_to != iolink2->node_to) {
1451 		/* CPU->CPU  link*/
1452 		cpu_dev = kfd_topology_device_by_proximity_domain(iolink1->node_to);
1453 		if (cpu_dev) {
1454 			list_for_each_entry(iolink3, &cpu_dev->io_link_props, list) {
1455 				if (iolink3->node_to != iolink2->node_to)
1456 					continue;
1457 
1458 				props->weight += iolink3->weight;
1459 				props->min_latency += iolink3->min_latency;
1460 				props->max_latency += iolink3->max_latency;
1461 				props->min_bandwidth = min(props->min_bandwidth,
1462 							   iolink3->min_bandwidth);
1463 				props->max_bandwidth = min(props->max_bandwidth,
1464 							   iolink3->max_bandwidth);
1465 				break;
1466 			}
1467 		} else {
1468 			WARN(1, "CPU node not found");
1469 		}
1470 	}
1471 
1472 	props->node_from = from;
1473 	props->node_to = to;
1474 	peer->node_props.p2p_links_count++;
1475 	list_add_tail(&props->list, &peer->p2p_link_props);
1476 	ret = kfd_build_p2p_node_entry(peer, props);
1477 
1478 	return ret;
1479 }
1480 #endif
1481 
1482 static int kfd_dev_create_p2p_links(void)
1483 {
1484 	struct kfd_topology_device *dev;
1485 	struct kfd_topology_device *new_dev;
1486 #if defined(CONFIG_HSA_AMD_P2P)
1487 	uint32_t i;
1488 #endif
1489 	uint32_t k;
1490 	int ret = 0;
1491 
1492 	k = 0;
1493 	list_for_each_entry(dev, &topology_device_list, list)
1494 		k++;
1495 	if (k < 2)
1496 		return 0;
1497 
1498 	new_dev = list_last_entry(&topology_device_list, struct kfd_topology_device, list);
1499 	if (WARN_ON(!new_dev->gpu))
1500 		return 0;
1501 
1502 	k--;
1503 
1504 	/* create in-direct links */
1505 	ret = kfd_create_indirect_link_prop(new_dev, k);
1506 	if (ret < 0)
1507 		goto out;
1508 
1509 	/* create p2p links */
1510 #if defined(CONFIG_HSA_AMD_P2P)
1511 	i = 0;
1512 	list_for_each_entry(dev, &topology_device_list, list) {
1513 		if (dev == new_dev)
1514 			break;
1515 		if (!dev->gpu || !dev->gpu->adev ||
1516 		    (dev->gpu->kfd->hive_id &&
1517 		     dev->gpu->kfd->hive_id == new_dev->gpu->kfd->hive_id))
1518 			goto next;
1519 
1520 		/* check if node(s) is/are peer accessible in one direction or bi-direction */
1521 		ret = kfd_add_peer_prop(new_dev, dev, i, k);
1522 		if (ret < 0)
1523 			goto out;
1524 
1525 		ret = kfd_add_peer_prop(dev, new_dev, k, i);
1526 		if (ret < 0)
1527 			goto out;
1528 next:
1529 		i++;
1530 	}
1531 #endif
1532 
1533 out:
1534 	return ret;
1535 }
1536 
1537 /* Helper function. See kfd_fill_gpu_cache_info for parameter description */
1538 static int fill_in_l1_pcache(struct kfd_cache_properties **props_ext,
1539 				struct kfd_gpu_cache_info *pcache_info,
1540 				struct kfd_cu_info *cu_info,
1541 				int cu_bitmask,
1542 				int cache_type, unsigned int cu_processor_id,
1543 				int cu_block)
1544 {
1545 	unsigned int cu_sibling_map_mask;
1546 	int first_active_cu;
1547 	struct kfd_cache_properties *pcache = NULL;
1548 
1549 	cu_sibling_map_mask = cu_bitmask;
1550 	cu_sibling_map_mask >>= cu_block;
1551 	cu_sibling_map_mask &= ((1 << pcache_info[cache_type].num_cu_shared) - 1);
1552 	first_active_cu = ffs(cu_sibling_map_mask);
1553 
1554 	/* CU could be inactive. In case of shared cache find the first active
1555 	 * CU. and incase of non-shared cache check if the CU is inactive. If
1556 	 * inactive active skip it
1557 	 */
1558 	if (first_active_cu) {
1559 		pcache = kfd_alloc_struct(pcache);
1560 		if (!pcache)
1561 			return -ENOMEM;
1562 
1563 		memset(pcache, 0, sizeof(struct kfd_cache_properties));
1564 		pcache->processor_id_low = cu_processor_id + (first_active_cu - 1);
1565 		pcache->cache_level = pcache_info[cache_type].cache_level;
1566 		pcache->cache_size = pcache_info[cache_type].cache_size;
1567 
1568 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_DATA_CACHE)
1569 			pcache->cache_type |= HSA_CACHE_TYPE_DATA;
1570 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_INST_CACHE)
1571 			pcache->cache_type |= HSA_CACHE_TYPE_INSTRUCTION;
1572 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_CPU_CACHE)
1573 			pcache->cache_type |= HSA_CACHE_TYPE_CPU;
1574 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_SIMD_CACHE)
1575 			pcache->cache_type |= HSA_CACHE_TYPE_HSACU;
1576 
1577 		/* Sibling map is w.r.t processor_id_low, so shift out
1578 		 * inactive CU
1579 		 */
1580 		cu_sibling_map_mask =
1581 			cu_sibling_map_mask >> (first_active_cu - 1);
1582 
1583 		pcache->sibling_map[0] = (uint8_t)(cu_sibling_map_mask & 0xFF);
1584 		pcache->sibling_map[1] =
1585 				(uint8_t)((cu_sibling_map_mask >> 8) & 0xFF);
1586 		pcache->sibling_map[2] =
1587 				(uint8_t)((cu_sibling_map_mask >> 16) & 0xFF);
1588 		pcache->sibling_map[3] =
1589 				(uint8_t)((cu_sibling_map_mask >> 24) & 0xFF);
1590 
1591 		pcache->sibling_map_size = 4;
1592 		*props_ext = pcache;
1593 
1594 		return 0;
1595 	}
1596 	return 1;
1597 }
1598 
1599 /* Helper function. See kfd_fill_gpu_cache_info for parameter description */
1600 static int fill_in_l2_l3_pcache(struct kfd_cache_properties **props_ext,
1601 				struct kfd_gpu_cache_info *pcache_info,
1602 				struct kfd_cu_info *cu_info,
1603 				int cache_type, unsigned int cu_processor_id,
1604 				struct kfd_node *knode)
1605 {
1606 	unsigned int cu_sibling_map_mask;
1607 	int first_active_cu;
1608 	int i, j, k, xcc, start, end;
1609 	struct kfd_cache_properties *pcache = NULL;
1610 
1611 	start = ffs(knode->xcc_mask) - 1;
1612 	end = start + NUM_XCC(knode->xcc_mask);
1613 	cu_sibling_map_mask = cu_info->cu_bitmap[start][0][0];
1614 	cu_sibling_map_mask &=
1615 		((1 << pcache_info[cache_type].num_cu_shared) - 1);
1616 	first_active_cu = ffs(cu_sibling_map_mask);
1617 
1618 	/* CU could be inactive. In case of shared cache find the first active
1619 	 * CU. and incase of non-shared cache check if the CU is inactive. If
1620 	 * inactive active skip it
1621 	 */
1622 	if (first_active_cu) {
1623 		pcache = kfd_alloc_struct(pcache);
1624 		if (!pcache)
1625 			return -ENOMEM;
1626 
1627 		memset(pcache, 0, sizeof(struct kfd_cache_properties));
1628 		pcache->processor_id_low = cu_processor_id
1629 					+ (first_active_cu - 1);
1630 		pcache->cache_level = pcache_info[cache_type].cache_level;
1631 		pcache->cache_size = pcache_info[cache_type].cache_size;
1632 
1633 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_DATA_CACHE)
1634 			pcache->cache_type |= HSA_CACHE_TYPE_DATA;
1635 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_INST_CACHE)
1636 			pcache->cache_type |= HSA_CACHE_TYPE_INSTRUCTION;
1637 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_CPU_CACHE)
1638 			pcache->cache_type |= HSA_CACHE_TYPE_CPU;
1639 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_SIMD_CACHE)
1640 			pcache->cache_type |= HSA_CACHE_TYPE_HSACU;
1641 
1642 		/* Sibling map is w.r.t processor_id_low, so shift out
1643 		 * inactive CU
1644 		 */
1645 		cu_sibling_map_mask = cu_sibling_map_mask >> (first_active_cu - 1);
1646 		k = 0;
1647 
1648 		for (xcc = start; xcc < end; xcc++) {
1649 			for (i = 0; i < cu_info->num_shader_engines; i++) {
1650 				for (j = 0; j < cu_info->num_shader_arrays_per_engine; j++) {
1651 					pcache->sibling_map[k] = (uint8_t)(cu_sibling_map_mask & 0xFF);
1652 					pcache->sibling_map[k+1] = (uint8_t)((cu_sibling_map_mask >> 8) & 0xFF);
1653 					pcache->sibling_map[k+2] = (uint8_t)((cu_sibling_map_mask >> 16) & 0xFF);
1654 					pcache->sibling_map[k+3] = (uint8_t)((cu_sibling_map_mask >> 24) & 0xFF);
1655 					k += 4;
1656 
1657 					cu_sibling_map_mask = cu_info->cu_bitmap[xcc][i % 4][j + i / 4];
1658 					cu_sibling_map_mask &= ((1 << pcache_info[cache_type].num_cu_shared) - 1);
1659 				}
1660 			}
1661 		}
1662 		pcache->sibling_map_size = k;
1663 		*props_ext = pcache;
1664 		return 0;
1665 	}
1666 	return 1;
1667 }
1668 
1669 #define KFD_MAX_CACHE_TYPES 6
1670 
1671 /* kfd_fill_cache_non_crat_info - Fill GPU cache info using kfd_gpu_cache_info
1672  * tables
1673  */
1674 static void kfd_fill_cache_non_crat_info(struct kfd_topology_device *dev, struct kfd_node *kdev)
1675 {
1676 	struct kfd_gpu_cache_info *pcache_info = NULL;
1677 	int i, j, k, xcc, start, end;
1678 	int ct = 0;
1679 	unsigned int cu_processor_id;
1680 	int ret;
1681 	unsigned int num_cu_shared;
1682 	struct kfd_cu_info cu_info;
1683 	struct kfd_cu_info *pcu_info;
1684 	int gpu_processor_id;
1685 	struct kfd_cache_properties *props_ext;
1686 	int num_of_entries = 0;
1687 	int num_of_cache_types = 0;
1688 	struct kfd_gpu_cache_info cache_info[KFD_MAX_CACHE_TYPES];
1689 
1690 	amdgpu_amdkfd_get_cu_info(kdev->adev, &cu_info);
1691 	pcu_info = &cu_info;
1692 
1693 	gpu_processor_id = dev->node_props.simd_id_base;
1694 
1695 	pcache_info = cache_info;
1696 	num_of_cache_types = kfd_get_gpu_cache_info(kdev, &pcache_info);
1697 	if (!num_of_cache_types) {
1698 		pr_warn("no cache info found\n");
1699 		return;
1700 	}
1701 
1702 	/* For each type of cache listed in the kfd_gpu_cache_info table,
1703 	 * go through all available Compute Units.
1704 	 * The [i,j,k] loop will
1705 	 *		if kfd_gpu_cache_info.num_cu_shared = 1
1706 	 *			will parse through all available CU
1707 	 *		If (kfd_gpu_cache_info.num_cu_shared != 1)
1708 	 *			then it will consider only one CU from
1709 	 *			the shared unit
1710 	 */
1711 	start = ffs(kdev->xcc_mask) - 1;
1712 	end = start + NUM_XCC(kdev->xcc_mask);
1713 
1714 	for (ct = 0; ct < num_of_cache_types; ct++) {
1715 		cu_processor_id = gpu_processor_id;
1716 		if (pcache_info[ct].cache_level == 1) {
1717 			for (xcc = start; xcc < end; xcc++) {
1718 				for (i = 0; i < pcu_info->num_shader_engines; i++) {
1719 					for (j = 0; j < pcu_info->num_shader_arrays_per_engine; j++) {
1720 						for (k = 0; k < pcu_info->num_cu_per_sh; k += pcache_info[ct].num_cu_shared) {
1721 
1722 							ret = fill_in_l1_pcache(&props_ext, pcache_info, pcu_info,
1723 										pcu_info->cu_bitmap[xcc][i % 4][j + i / 4], ct,
1724 										cu_processor_id, k);
1725 
1726 							if (ret < 0)
1727 								break;
1728 
1729 							if (!ret) {
1730 								num_of_entries++;
1731 								list_add_tail(&props_ext->list, &dev->cache_props);
1732 							}
1733 
1734 							/* Move to next CU block */
1735 							num_cu_shared = ((k + pcache_info[ct].num_cu_shared) <=
1736 								pcu_info->num_cu_per_sh) ?
1737 								pcache_info[ct].num_cu_shared :
1738 								(pcu_info->num_cu_per_sh - k);
1739 							cu_processor_id += num_cu_shared;
1740 						}
1741 					}
1742 				}
1743 			}
1744 		} else {
1745 			ret = fill_in_l2_l3_pcache(&props_ext, pcache_info,
1746 					pcu_info, ct, cu_processor_id, kdev);
1747 
1748 			if (ret < 0)
1749 				break;
1750 
1751 			if (!ret) {
1752 				num_of_entries++;
1753 				list_add_tail(&props_ext->list, &dev->cache_props);
1754 			}
1755 		}
1756 	}
1757 	dev->node_props.caches_count += num_of_entries;
1758 	pr_debug("Added [%d] GPU cache entries\n", num_of_entries);
1759 }
1760 
1761 static int kfd_topology_add_device_locked(struct kfd_node *gpu, uint32_t gpu_id,
1762 					  struct kfd_topology_device **dev)
1763 {
1764 	int proximity_domain = ++topology_crat_proximity_domain;
1765 	struct list_head temp_topology_device_list;
1766 	void *crat_image = NULL;
1767 	size_t image_size = 0;
1768 	int res;
1769 
1770 	res = kfd_create_crat_image_virtual(&crat_image, &image_size,
1771 					    COMPUTE_UNIT_GPU, gpu,
1772 					    proximity_domain);
1773 	if (res) {
1774 		pr_err("Error creating VCRAT for GPU (ID: 0x%x)\n",
1775 		       gpu_id);
1776 		topology_crat_proximity_domain--;
1777 		goto err;
1778 	}
1779 
1780 	INIT_LIST_HEAD(&temp_topology_device_list);
1781 
1782 	res = kfd_parse_crat_table(crat_image,
1783 				   &temp_topology_device_list,
1784 				   proximity_domain);
1785 	if (res) {
1786 		pr_err("Error parsing VCRAT for GPU (ID: 0x%x)\n",
1787 		       gpu_id);
1788 		topology_crat_proximity_domain--;
1789 		goto err;
1790 	}
1791 
1792 	kfd_topology_update_device_list(&temp_topology_device_list,
1793 					&topology_device_list);
1794 
1795 	*dev = kfd_assign_gpu(gpu);
1796 	if (WARN_ON(!*dev)) {
1797 		res = -ENODEV;
1798 		goto err;
1799 	}
1800 
1801 	/* Fill the cache affinity information here for the GPUs
1802 	 * using VCRAT
1803 	 */
1804 	kfd_fill_cache_non_crat_info(*dev, gpu);
1805 
1806 	/* Update the SYSFS tree, since we added another topology
1807 	 * device
1808 	 */
1809 	res = kfd_topology_update_sysfs();
1810 	if (!res)
1811 		sys_props.generation_count++;
1812 	else
1813 		pr_err("Failed to update GPU (ID: 0x%x) to sysfs topology. res=%d\n",
1814 		       gpu_id, res);
1815 
1816 err:
1817 	kfd_destroy_crat_image(crat_image);
1818 	return res;
1819 }
1820 
1821 static void kfd_topology_set_dbg_firmware_support(struct kfd_topology_device *dev)
1822 {
1823 	bool firmware_supported = true;
1824 
1825 	if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(11, 0, 0) &&
1826 			KFD_GC_VERSION(dev->gpu) < IP_VERSION(12, 0, 0)) {
1827 		uint32_t mes_api_rev = (dev->gpu->adev->mes.sched_version &
1828 						AMDGPU_MES_API_VERSION_MASK) >>
1829 						AMDGPU_MES_API_VERSION_SHIFT;
1830 		uint32_t mes_rev = dev->gpu->adev->mes.sched_version &
1831 						AMDGPU_MES_VERSION_MASK;
1832 
1833 		firmware_supported = (mes_api_rev >= 14) && (mes_rev >= 64);
1834 		goto out;
1835 	}
1836 
1837 	/*
1838 	 * Note: Any unlisted devices here are assumed to support exception handling.
1839 	 * Add additional checks here as needed.
1840 	 */
1841 	switch (KFD_GC_VERSION(dev->gpu)) {
1842 	case IP_VERSION(9, 0, 1):
1843 		firmware_supported = dev->gpu->kfd->mec_fw_version >= 459 + 32768;
1844 		break;
1845 	case IP_VERSION(9, 1, 0):
1846 	case IP_VERSION(9, 2, 1):
1847 	case IP_VERSION(9, 2, 2):
1848 	case IP_VERSION(9, 3, 0):
1849 	case IP_VERSION(9, 4, 0):
1850 		firmware_supported = dev->gpu->kfd->mec_fw_version >= 459;
1851 		break;
1852 	case IP_VERSION(9, 4, 1):
1853 		firmware_supported = dev->gpu->kfd->mec_fw_version >= 60;
1854 		break;
1855 	case IP_VERSION(9, 4, 2):
1856 		firmware_supported = dev->gpu->kfd->mec_fw_version >= 51;
1857 		break;
1858 	case IP_VERSION(10, 1, 10):
1859 	case IP_VERSION(10, 1, 2):
1860 	case IP_VERSION(10, 1, 1):
1861 		firmware_supported = dev->gpu->kfd->mec_fw_version >= 144;
1862 		break;
1863 	case IP_VERSION(10, 3, 0):
1864 	case IP_VERSION(10, 3, 2):
1865 	case IP_VERSION(10, 3, 1):
1866 	case IP_VERSION(10, 3, 4):
1867 	case IP_VERSION(10, 3, 5):
1868 		firmware_supported = dev->gpu->kfd->mec_fw_version >= 89;
1869 		break;
1870 	case IP_VERSION(10, 1, 3):
1871 	case IP_VERSION(10, 3, 3):
1872 		firmware_supported = false;
1873 		break;
1874 	default:
1875 		break;
1876 	}
1877 
1878 out:
1879 	if (firmware_supported)
1880 		dev->node_props.capability |= HSA_CAP_TRAP_DEBUG_FIRMWARE_SUPPORTED;
1881 }
1882 
1883 static void kfd_topology_set_capabilities(struct kfd_topology_device *dev)
1884 {
1885 	dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 <<
1886 				HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1887 				HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1888 
1889 	dev->node_props.capability |= HSA_CAP_TRAP_DEBUG_SUPPORT |
1890 			HSA_CAP_TRAP_DEBUG_WAVE_LAUNCH_TRAP_OVERRIDE_SUPPORTED |
1891 			HSA_CAP_TRAP_DEBUG_WAVE_LAUNCH_MODE_SUPPORTED;
1892 
1893 	if (kfd_dbg_has_ttmps_always_setup(dev->gpu))
1894 		dev->node_props.debug_prop |= HSA_DBG_DISPATCH_INFO_ALWAYS_VALID;
1895 
1896 	if (KFD_GC_VERSION(dev->gpu) < IP_VERSION(10, 0, 0)) {
1897 		if (KFD_GC_VERSION(dev->gpu) == IP_VERSION(9, 4, 3))
1898 			dev->node_props.debug_prop |=
1899 				HSA_DBG_WATCH_ADDR_MASK_LO_BIT_GFX9_4_3 |
1900 				HSA_DBG_WATCH_ADDR_MASK_HI_BIT_GFX9_4_3;
1901 		else
1902 			dev->node_props.debug_prop |=
1903 				HSA_DBG_WATCH_ADDR_MASK_LO_BIT_GFX9 |
1904 				HSA_DBG_WATCH_ADDR_MASK_HI_BIT;
1905 
1906 		if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(9, 4, 2))
1907 			dev->node_props.capability |=
1908 				HSA_CAP_TRAP_DEBUG_PRECISE_MEMORY_OPERATIONS_SUPPORTED;
1909 	} else {
1910 		dev->node_props.debug_prop |= HSA_DBG_WATCH_ADDR_MASK_LO_BIT_GFX10 |
1911 					HSA_DBG_WATCH_ADDR_MASK_HI_BIT;
1912 
1913 		if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(11, 0, 0))
1914 			dev->node_props.capability |=
1915 				HSA_CAP_TRAP_DEBUG_PRECISE_MEMORY_OPERATIONS_SUPPORTED;
1916 	}
1917 
1918 	kfd_topology_set_dbg_firmware_support(dev);
1919 }
1920 
1921 int kfd_topology_add_device(struct kfd_node *gpu)
1922 {
1923 	uint32_t gpu_id;
1924 	struct kfd_topology_device *dev;
1925 	struct kfd_cu_info cu_info;
1926 	int res = 0;
1927 	int i;
1928 	const char *asic_name = amdgpu_asic_name[gpu->adev->asic_type];
1929 
1930 	gpu_id = kfd_generate_gpu_id(gpu);
1931 	if (gpu->xcp && !gpu->xcp->ddev) {
1932 		dev_warn(gpu->adev->dev,
1933 		"Won't add GPU (ID: 0x%x) to topology since it has no drm node assigned.",
1934 		gpu_id);
1935 		return 0;
1936 	} else {
1937 		pr_debug("Adding new GPU (ID: 0x%x) to topology\n", gpu_id);
1938 	}
1939 
1940 	/* Check to see if this gpu device exists in the topology_device_list.
1941 	 * If so, assign the gpu to that device,
1942 	 * else create a Virtual CRAT for this gpu device and then parse that
1943 	 * CRAT to create a new topology device. Once created assign the gpu to
1944 	 * that topology device
1945 	 */
1946 	down_write(&topology_lock);
1947 	dev = kfd_assign_gpu(gpu);
1948 	if (!dev)
1949 		res = kfd_topology_add_device_locked(gpu, gpu_id, &dev);
1950 	up_write(&topology_lock);
1951 	if (res)
1952 		return res;
1953 
1954 	dev->gpu_id = gpu_id;
1955 	gpu->id = gpu_id;
1956 
1957 	kfd_dev_create_p2p_links();
1958 
1959 	/* TODO: Move the following lines to function
1960 	 *	kfd_add_non_crat_information
1961 	 */
1962 
1963 	/* Fill-in additional information that is not available in CRAT but
1964 	 * needed for the topology
1965 	 */
1966 
1967 	amdgpu_amdkfd_get_cu_info(dev->gpu->adev, &cu_info);
1968 
1969 	for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1; i++) {
1970 		dev->node_props.name[i] = __tolower(asic_name[i]);
1971 		if (asic_name[i] == '\0')
1972 			break;
1973 	}
1974 	dev->node_props.name[i] = '\0';
1975 
1976 	dev->node_props.simd_arrays_per_engine =
1977 		cu_info.num_shader_arrays_per_engine;
1978 
1979 	dev->node_props.gfx_target_version =
1980 				gpu->kfd->device_info.gfx_target_version;
1981 	dev->node_props.vendor_id = gpu->adev->pdev->vendor;
1982 	dev->node_props.device_id = gpu->adev->pdev->device;
1983 	dev->node_props.capability |=
1984 		((dev->gpu->adev->rev_id << HSA_CAP_ASIC_REVISION_SHIFT) &
1985 			HSA_CAP_ASIC_REVISION_MASK);
1986 
1987 	dev->node_props.location_id = pci_dev_id(gpu->adev->pdev);
1988 	if (KFD_GC_VERSION(dev->gpu->kfd) == IP_VERSION(9, 4, 3))
1989 		dev->node_props.location_id |= dev->gpu->node_id;
1990 
1991 	dev->node_props.domain = pci_domain_nr(gpu->adev->pdev->bus);
1992 	dev->node_props.max_engine_clk_fcompute =
1993 		amdgpu_amdkfd_get_max_engine_clock_in_mhz(dev->gpu->adev);
1994 	dev->node_props.max_engine_clk_ccompute =
1995 		cpufreq_quick_get_max(0) / 1000;
1996 
1997 	if (gpu->xcp)
1998 		dev->node_props.drm_render_minor = gpu->xcp->ddev->render->index;
1999 	else
2000 		dev->node_props.drm_render_minor =
2001 				gpu->kfd->shared_resources.drm_render_minor;
2002 
2003 	dev->node_props.hive_id = gpu->kfd->hive_id;
2004 	dev->node_props.num_sdma_engines = kfd_get_num_sdma_engines(gpu);
2005 	dev->node_props.num_sdma_xgmi_engines =
2006 					kfd_get_num_xgmi_sdma_engines(gpu);
2007 	dev->node_props.num_sdma_queues_per_engine =
2008 				gpu->kfd->device_info.num_sdma_queues_per_engine -
2009 				gpu->kfd->device_info.num_reserved_sdma_queues_per_engine;
2010 	dev->node_props.num_gws = (dev->gpu->gws &&
2011 		dev->gpu->dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) ?
2012 		dev->gpu->adev->gds.gws_size : 0;
2013 	dev->node_props.num_cp_queues = get_cp_queues_num(dev->gpu->dqm);
2014 
2015 	kfd_fill_mem_clk_max_info(dev);
2016 	kfd_fill_iolink_non_crat_info(dev);
2017 
2018 	switch (dev->gpu->adev->asic_type) {
2019 	case CHIP_KAVERI:
2020 	case CHIP_HAWAII:
2021 	case CHIP_TONGA:
2022 		dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_PRE_1_0 <<
2023 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
2024 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
2025 		break;
2026 	case CHIP_CARRIZO:
2027 	case CHIP_FIJI:
2028 	case CHIP_POLARIS10:
2029 	case CHIP_POLARIS11:
2030 	case CHIP_POLARIS12:
2031 	case CHIP_VEGAM:
2032 		pr_debug("Adding doorbell packet type capability\n");
2033 		dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 <<
2034 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
2035 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
2036 		break;
2037 	default:
2038 		if (KFD_GC_VERSION(dev->gpu) < IP_VERSION(9, 0, 1))
2039 			WARN(1, "Unexpected ASIC family %u",
2040 			     dev->gpu->adev->asic_type);
2041 		else
2042 			kfd_topology_set_capabilities(dev);
2043 	}
2044 
2045 	/*
2046 	 * Overwrite ATS capability according to needs_iommu_device to fix
2047 	 * potential missing corresponding bit in CRAT of BIOS.
2048 	 */
2049 	dev->node_props.capability &= ~HSA_CAP_ATS_PRESENT;
2050 
2051 	/* Fix errors in CZ CRAT.
2052 	 * simd_count: Carrizo CRAT reports wrong simd_count, probably
2053 	 *		because it doesn't consider masked out CUs
2054 	 * max_waves_per_simd: Carrizo reports wrong max_waves_per_simd
2055 	 */
2056 	if (dev->gpu->adev->asic_type == CHIP_CARRIZO) {
2057 		dev->node_props.simd_count =
2058 			cu_info.simd_per_cu * cu_info.cu_active_number;
2059 		dev->node_props.max_waves_per_simd = 10;
2060 	}
2061 
2062 	/* kfd only concerns sram ecc on GFX and HBM ecc on UMC */
2063 	dev->node_props.capability |=
2064 		((dev->gpu->adev->ras_enabled & BIT(AMDGPU_RAS_BLOCK__GFX)) != 0) ?
2065 		HSA_CAP_SRAM_EDCSUPPORTED : 0;
2066 	dev->node_props.capability |=
2067 		((dev->gpu->adev->ras_enabled & BIT(AMDGPU_RAS_BLOCK__UMC)) != 0) ?
2068 		HSA_CAP_MEM_EDCSUPPORTED : 0;
2069 
2070 	if (KFD_GC_VERSION(dev->gpu) != IP_VERSION(9, 0, 1))
2071 		dev->node_props.capability |= (dev->gpu->adev->ras_enabled != 0) ?
2072 			HSA_CAP_RASEVENTNOTIFY : 0;
2073 
2074 	if (KFD_IS_SVM_API_SUPPORTED(dev->gpu->adev))
2075 		dev->node_props.capability |= HSA_CAP_SVMAPI_SUPPORTED;
2076 
2077 	if (dev->gpu->adev->gmc.is_app_apu ||
2078 		dev->gpu->adev->gmc.xgmi.connected_to_cpu)
2079 		dev->node_props.capability |= HSA_CAP_FLAGS_COHERENTHOSTACCESS;
2080 
2081 	kfd_debug_print_topology();
2082 
2083 	kfd_notify_gpu_change(gpu_id, 1);
2084 
2085 	return 0;
2086 }
2087 
2088 /**
2089  * kfd_topology_update_io_links() - Update IO links after device removal.
2090  * @proximity_domain: Proximity domain value of the dev being removed.
2091  *
2092  * The topology list currently is arranged in increasing order of
2093  * proximity domain.
2094  *
2095  * Two things need to be done when a device is removed:
2096  * 1. All the IO links to this device need to be removed.
2097  * 2. All nodes after the current device node need to move
2098  *    up once this device node is removed from the topology
2099  *    list. As a result, the proximity domain values for
2100  *    all nodes after the node being deleted reduce by 1.
2101  *    This would also cause the proximity domain values for
2102  *    io links to be updated based on new proximity domain
2103  *    values.
2104  *
2105  * Context: The caller must hold write topology_lock.
2106  */
2107 static void kfd_topology_update_io_links(int proximity_domain)
2108 {
2109 	struct kfd_topology_device *dev;
2110 	struct kfd_iolink_properties *iolink, *p2plink, *tmp;
2111 
2112 	list_for_each_entry(dev, &topology_device_list, list) {
2113 		if (dev->proximity_domain > proximity_domain)
2114 			dev->proximity_domain--;
2115 
2116 		list_for_each_entry_safe(iolink, tmp, &dev->io_link_props, list) {
2117 			/*
2118 			 * If there is an io link to the dev being deleted
2119 			 * then remove that IO link also.
2120 			 */
2121 			if (iolink->node_to == proximity_domain) {
2122 				list_del(&iolink->list);
2123 				dev->node_props.io_links_count--;
2124 			} else {
2125 				if (iolink->node_from > proximity_domain)
2126 					iolink->node_from--;
2127 				if (iolink->node_to > proximity_domain)
2128 					iolink->node_to--;
2129 			}
2130 		}
2131 
2132 		list_for_each_entry_safe(p2plink, tmp, &dev->p2p_link_props, list) {
2133 			/*
2134 			 * If there is a p2p link to the dev being deleted
2135 			 * then remove that p2p link also.
2136 			 */
2137 			if (p2plink->node_to == proximity_domain) {
2138 				list_del(&p2plink->list);
2139 				dev->node_props.p2p_links_count--;
2140 			} else {
2141 				if (p2plink->node_from > proximity_domain)
2142 					p2plink->node_from--;
2143 				if (p2plink->node_to > proximity_domain)
2144 					p2plink->node_to--;
2145 			}
2146 		}
2147 	}
2148 }
2149 
2150 int kfd_topology_remove_device(struct kfd_node *gpu)
2151 {
2152 	struct kfd_topology_device *dev, *tmp;
2153 	uint32_t gpu_id;
2154 	int res = -ENODEV;
2155 	int i = 0;
2156 
2157 	down_write(&topology_lock);
2158 
2159 	list_for_each_entry_safe(dev, tmp, &topology_device_list, list) {
2160 		if (dev->gpu == gpu) {
2161 			gpu_id = dev->gpu_id;
2162 			kfd_remove_sysfs_node_entry(dev);
2163 			kfd_release_topology_device(dev);
2164 			sys_props.num_devices--;
2165 			kfd_topology_update_io_links(i);
2166 			topology_crat_proximity_domain = sys_props.num_devices-1;
2167 			sys_props.generation_count++;
2168 			res = 0;
2169 			if (kfd_topology_update_sysfs() < 0)
2170 				kfd_topology_release_sysfs();
2171 			break;
2172 		}
2173 		i++;
2174 	}
2175 
2176 	up_write(&topology_lock);
2177 
2178 	if (!res)
2179 		kfd_notify_gpu_change(gpu_id, 0);
2180 
2181 	return res;
2182 }
2183 
2184 /* kfd_topology_enum_kfd_devices - Enumerate through all devices in KFD
2185  *	topology. If GPU device is found @idx, then valid kfd_dev pointer is
2186  *	returned through @kdev
2187  * Return -	0: On success (@kdev will be NULL for non GPU nodes)
2188  *		-1: If end of list
2189  */
2190 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_node **kdev)
2191 {
2192 
2193 	struct kfd_topology_device *top_dev;
2194 	uint8_t device_idx = 0;
2195 
2196 	*kdev = NULL;
2197 	down_read(&topology_lock);
2198 
2199 	list_for_each_entry(top_dev, &topology_device_list, list) {
2200 		if (device_idx == idx) {
2201 			*kdev = top_dev->gpu;
2202 			up_read(&topology_lock);
2203 			return 0;
2204 		}
2205 
2206 		device_idx++;
2207 	}
2208 
2209 	up_read(&topology_lock);
2210 
2211 	return -1;
2212 
2213 }
2214 
2215 static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask)
2216 {
2217 	int first_cpu_of_numa_node;
2218 
2219 	if (!cpumask || cpumask == cpu_none_mask)
2220 		return -1;
2221 	first_cpu_of_numa_node = cpumask_first(cpumask);
2222 	if (first_cpu_of_numa_node >= nr_cpu_ids)
2223 		return -1;
2224 #ifdef CONFIG_X86_64
2225 	return cpu_data(first_cpu_of_numa_node).apicid;
2226 #else
2227 	return first_cpu_of_numa_node;
2228 #endif
2229 }
2230 
2231 /* kfd_numa_node_to_apic_id - Returns the APIC ID of the first logical processor
2232  *	of the given NUMA node (numa_node_id)
2233  * Return -1 on failure
2234  */
2235 int kfd_numa_node_to_apic_id(int numa_node_id)
2236 {
2237 	if (numa_node_id == -1) {
2238 		pr_warn("Invalid NUMA Node. Use online CPU mask\n");
2239 		return kfd_cpumask_to_apic_id(cpu_online_mask);
2240 	}
2241 	return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id));
2242 }
2243 
2244 #if defined(CONFIG_DEBUG_FS)
2245 
2246 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data)
2247 {
2248 	struct kfd_topology_device *dev;
2249 	unsigned int i = 0;
2250 	int r = 0;
2251 
2252 	down_read(&topology_lock);
2253 
2254 	list_for_each_entry(dev, &topology_device_list, list) {
2255 		if (!dev->gpu) {
2256 			i++;
2257 			continue;
2258 		}
2259 
2260 		seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
2261 		r = dqm_debugfs_hqds(m, dev->gpu->dqm);
2262 		if (r)
2263 			break;
2264 	}
2265 
2266 	up_read(&topology_lock);
2267 
2268 	return r;
2269 }
2270 
2271 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data)
2272 {
2273 	struct kfd_topology_device *dev;
2274 	unsigned int i = 0;
2275 	int r = 0;
2276 
2277 	down_read(&topology_lock);
2278 
2279 	list_for_each_entry(dev, &topology_device_list, list) {
2280 		if (!dev->gpu) {
2281 			i++;
2282 			continue;
2283 		}
2284 
2285 		seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
2286 		r = pm_debugfs_runlist(m, &dev->gpu->dqm->packet_mgr);
2287 		if (r)
2288 			break;
2289 	}
2290 
2291 	up_read(&topology_lock);
2292 
2293 	return r;
2294 }
2295 
2296 #endif
2297