1 // SPDX-License-Identifier: GPL-2.0 OR MIT
2 /*
3  * Copyright 2014-2022 Advanced Micro Devices, Inc.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be included in
13  * all copies or substantial portions of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21  * OTHER DEALINGS IN THE SOFTWARE.
22  */
23 
24 #include <linux/types.h>
25 #include <linux/kernel.h>
26 #include <linux/pci.h>
27 #include <linux/errno.h>
28 #include <linux/acpi.h>
29 #include <linux/hash.h>
30 #include <linux/cpufreq.h>
31 #include <linux/log2.h>
32 #include <linux/dmi.h>
33 #include <linux/atomic.h>
34 
35 #include "kfd_priv.h"
36 #include "kfd_crat.h"
37 #include "kfd_topology.h"
38 #include "kfd_device_queue_manager.h"
39 #include "kfd_iommu.h"
40 #include "kfd_svm.h"
41 #include "amdgpu_amdkfd.h"
42 #include "amdgpu_ras.h"
43 
44 /* topology_device_list - Master list of all topology devices */
45 static struct list_head topology_device_list;
46 static struct kfd_system_properties sys_props;
47 
48 static DECLARE_RWSEM(topology_lock);
49 static uint32_t topology_crat_proximity_domain;
50 
51 struct kfd_topology_device *kfd_topology_device_by_proximity_domain_no_lock(
52 						uint32_t proximity_domain)
53 {
54 	struct kfd_topology_device *top_dev;
55 	struct kfd_topology_device *device = NULL;
56 
57 	list_for_each_entry(top_dev, &topology_device_list, list)
58 		if (top_dev->proximity_domain == proximity_domain) {
59 			device = top_dev;
60 			break;
61 		}
62 
63 	return device;
64 }
65 
66 struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
67 						uint32_t proximity_domain)
68 {
69 	struct kfd_topology_device *device = NULL;
70 
71 	down_read(&topology_lock);
72 
73 	device = kfd_topology_device_by_proximity_domain_no_lock(
74 							proximity_domain);
75 	up_read(&topology_lock);
76 
77 	return device;
78 }
79 
80 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id)
81 {
82 	struct kfd_topology_device *top_dev = NULL;
83 	struct kfd_topology_device *ret = NULL;
84 
85 	down_read(&topology_lock);
86 
87 	list_for_each_entry(top_dev, &topology_device_list, list)
88 		if (top_dev->gpu_id == gpu_id) {
89 			ret = top_dev;
90 			break;
91 		}
92 
93 	up_read(&topology_lock);
94 
95 	return ret;
96 }
97 
98 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id)
99 {
100 	struct kfd_topology_device *top_dev;
101 
102 	top_dev = kfd_topology_device_by_id(gpu_id);
103 	if (!top_dev)
104 		return NULL;
105 
106 	return top_dev->gpu;
107 }
108 
109 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev)
110 {
111 	struct kfd_topology_device *top_dev;
112 	struct kfd_dev *device = NULL;
113 
114 	down_read(&topology_lock);
115 
116 	list_for_each_entry(top_dev, &topology_device_list, list)
117 		if (top_dev->gpu && top_dev->gpu->pdev == pdev) {
118 			device = top_dev->gpu;
119 			break;
120 		}
121 
122 	up_read(&topology_lock);
123 
124 	return device;
125 }
126 
127 struct kfd_dev *kfd_device_by_adev(const struct amdgpu_device *adev)
128 {
129 	struct kfd_topology_device *top_dev;
130 	struct kfd_dev *device = NULL;
131 
132 	down_read(&topology_lock);
133 
134 	list_for_each_entry(top_dev, &topology_device_list, list)
135 		if (top_dev->gpu && top_dev->gpu->adev == adev) {
136 			device = top_dev->gpu;
137 			break;
138 		}
139 
140 	up_read(&topology_lock);
141 
142 	return device;
143 }
144 
145 /* Called with write topology_lock acquired */
146 static void kfd_release_topology_device(struct kfd_topology_device *dev)
147 {
148 	struct kfd_mem_properties *mem;
149 	struct kfd_cache_properties *cache;
150 	struct kfd_iolink_properties *iolink;
151 	struct kfd_perf_properties *perf;
152 
153 	list_del(&dev->list);
154 
155 	while (dev->mem_props.next != &dev->mem_props) {
156 		mem = container_of(dev->mem_props.next,
157 				struct kfd_mem_properties, list);
158 		list_del(&mem->list);
159 		kfree(mem);
160 	}
161 
162 	while (dev->cache_props.next != &dev->cache_props) {
163 		cache = container_of(dev->cache_props.next,
164 				struct kfd_cache_properties, list);
165 		list_del(&cache->list);
166 		kfree(cache);
167 	}
168 
169 	while (dev->io_link_props.next != &dev->io_link_props) {
170 		iolink = container_of(dev->io_link_props.next,
171 				struct kfd_iolink_properties, list);
172 		list_del(&iolink->list);
173 		kfree(iolink);
174 	}
175 
176 	while (dev->perf_props.next != &dev->perf_props) {
177 		perf = container_of(dev->perf_props.next,
178 				struct kfd_perf_properties, list);
179 		list_del(&perf->list);
180 		kfree(perf);
181 	}
182 
183 	kfree(dev);
184 }
185 
186 void kfd_release_topology_device_list(struct list_head *device_list)
187 {
188 	struct kfd_topology_device *dev;
189 
190 	while (!list_empty(device_list)) {
191 		dev = list_first_entry(device_list,
192 				       struct kfd_topology_device, list);
193 		kfd_release_topology_device(dev);
194 	}
195 }
196 
197 static void kfd_release_live_view(void)
198 {
199 	kfd_release_topology_device_list(&topology_device_list);
200 	memset(&sys_props, 0, sizeof(sys_props));
201 }
202 
203 struct kfd_topology_device *kfd_create_topology_device(
204 				struct list_head *device_list)
205 {
206 	struct kfd_topology_device *dev;
207 
208 	dev = kfd_alloc_struct(dev);
209 	if (!dev) {
210 		pr_err("No memory to allocate a topology device");
211 		return NULL;
212 	}
213 
214 	INIT_LIST_HEAD(&dev->mem_props);
215 	INIT_LIST_HEAD(&dev->cache_props);
216 	INIT_LIST_HEAD(&dev->io_link_props);
217 	INIT_LIST_HEAD(&dev->perf_props);
218 
219 	list_add_tail(&dev->list, device_list);
220 
221 	return dev;
222 }
223 
224 
225 #define sysfs_show_gen_prop(buffer, offs, fmt, ...)		\
226 		(offs += snprintf(buffer+offs, PAGE_SIZE-offs,	\
227 				  fmt, __VA_ARGS__))
228 #define sysfs_show_32bit_prop(buffer, offs, name, value) \
229 		sysfs_show_gen_prop(buffer, offs, "%s %u\n", name, value)
230 #define sysfs_show_64bit_prop(buffer, offs, name, value) \
231 		sysfs_show_gen_prop(buffer, offs, "%s %llu\n", name, value)
232 #define sysfs_show_32bit_val(buffer, offs, value) \
233 		sysfs_show_gen_prop(buffer, offs, "%u\n", value)
234 #define sysfs_show_str_val(buffer, offs, value) \
235 		sysfs_show_gen_prop(buffer, offs, "%s\n", value)
236 
237 static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr,
238 		char *buffer)
239 {
240 	int offs = 0;
241 
242 	/* Making sure that the buffer is an empty string */
243 	buffer[0] = 0;
244 
245 	if (attr == &sys_props.attr_genid) {
246 		sysfs_show_32bit_val(buffer, offs,
247 				     sys_props.generation_count);
248 	} else if (attr == &sys_props.attr_props) {
249 		sysfs_show_64bit_prop(buffer, offs, "platform_oem",
250 				      sys_props.platform_oem);
251 		sysfs_show_64bit_prop(buffer, offs, "platform_id",
252 				      sys_props.platform_id);
253 		sysfs_show_64bit_prop(buffer, offs, "platform_rev",
254 				      sys_props.platform_rev);
255 	} else {
256 		offs = -EINVAL;
257 	}
258 
259 	return offs;
260 }
261 
262 static void kfd_topology_kobj_release(struct kobject *kobj)
263 {
264 	kfree(kobj);
265 }
266 
267 static const struct sysfs_ops sysprops_ops = {
268 	.show = sysprops_show,
269 };
270 
271 static struct kobj_type sysprops_type = {
272 	.release = kfd_topology_kobj_release,
273 	.sysfs_ops = &sysprops_ops,
274 };
275 
276 static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr,
277 		char *buffer)
278 {
279 	int offs = 0;
280 	struct kfd_iolink_properties *iolink;
281 
282 	/* Making sure that the buffer is an empty string */
283 	buffer[0] = 0;
284 
285 	iolink = container_of(attr, struct kfd_iolink_properties, attr);
286 	if (iolink->gpu && kfd_devcgroup_check_permission(iolink->gpu))
287 		return -EPERM;
288 	sysfs_show_32bit_prop(buffer, offs, "type", iolink->iolink_type);
289 	sysfs_show_32bit_prop(buffer, offs, "version_major", iolink->ver_maj);
290 	sysfs_show_32bit_prop(buffer, offs, "version_minor", iolink->ver_min);
291 	sysfs_show_32bit_prop(buffer, offs, "node_from", iolink->node_from);
292 	sysfs_show_32bit_prop(buffer, offs, "node_to", iolink->node_to);
293 	sysfs_show_32bit_prop(buffer, offs, "weight", iolink->weight);
294 	sysfs_show_32bit_prop(buffer, offs, "min_latency", iolink->min_latency);
295 	sysfs_show_32bit_prop(buffer, offs, "max_latency", iolink->max_latency);
296 	sysfs_show_32bit_prop(buffer, offs, "min_bandwidth",
297 			      iolink->min_bandwidth);
298 	sysfs_show_32bit_prop(buffer, offs, "max_bandwidth",
299 			      iolink->max_bandwidth);
300 	sysfs_show_32bit_prop(buffer, offs, "recommended_transfer_size",
301 			      iolink->rec_transfer_size);
302 	sysfs_show_32bit_prop(buffer, offs, "flags", iolink->flags);
303 
304 	return offs;
305 }
306 
307 static const struct sysfs_ops iolink_ops = {
308 	.show = iolink_show,
309 };
310 
311 static struct kobj_type iolink_type = {
312 	.release = kfd_topology_kobj_release,
313 	.sysfs_ops = &iolink_ops,
314 };
315 
316 static ssize_t mem_show(struct kobject *kobj, struct attribute *attr,
317 		char *buffer)
318 {
319 	int offs = 0;
320 	struct kfd_mem_properties *mem;
321 
322 	/* Making sure that the buffer is an empty string */
323 	buffer[0] = 0;
324 
325 	mem = container_of(attr, struct kfd_mem_properties, attr);
326 	if (mem->gpu && kfd_devcgroup_check_permission(mem->gpu))
327 		return -EPERM;
328 	sysfs_show_32bit_prop(buffer, offs, "heap_type", mem->heap_type);
329 	sysfs_show_64bit_prop(buffer, offs, "size_in_bytes",
330 			      mem->size_in_bytes);
331 	sysfs_show_32bit_prop(buffer, offs, "flags", mem->flags);
332 	sysfs_show_32bit_prop(buffer, offs, "width", mem->width);
333 	sysfs_show_32bit_prop(buffer, offs, "mem_clk_max",
334 			      mem->mem_clk_max);
335 
336 	return offs;
337 }
338 
339 static const struct sysfs_ops mem_ops = {
340 	.show = mem_show,
341 };
342 
343 static struct kobj_type mem_type = {
344 	.release = kfd_topology_kobj_release,
345 	.sysfs_ops = &mem_ops,
346 };
347 
348 static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr,
349 		char *buffer)
350 {
351 	int offs = 0;
352 	uint32_t i, j;
353 	struct kfd_cache_properties *cache;
354 
355 	/* Making sure that the buffer is an empty string */
356 	buffer[0] = 0;
357 
358 	cache = container_of(attr, struct kfd_cache_properties, attr);
359 	if (cache->gpu && kfd_devcgroup_check_permission(cache->gpu))
360 		return -EPERM;
361 	sysfs_show_32bit_prop(buffer, offs, "processor_id_low",
362 			cache->processor_id_low);
363 	sysfs_show_32bit_prop(buffer, offs, "level", cache->cache_level);
364 	sysfs_show_32bit_prop(buffer, offs, "size", cache->cache_size);
365 	sysfs_show_32bit_prop(buffer, offs, "cache_line_size",
366 			      cache->cacheline_size);
367 	sysfs_show_32bit_prop(buffer, offs, "cache_lines_per_tag",
368 			      cache->cachelines_per_tag);
369 	sysfs_show_32bit_prop(buffer, offs, "association", cache->cache_assoc);
370 	sysfs_show_32bit_prop(buffer, offs, "latency", cache->cache_latency);
371 	sysfs_show_32bit_prop(buffer, offs, "type", cache->cache_type);
372 	offs += snprintf(buffer+offs, PAGE_SIZE-offs, "sibling_map ");
373 	for (i = 0; i < CRAT_SIBLINGMAP_SIZE; i++)
374 		for (j = 0; j < sizeof(cache->sibling_map[0])*8; j++)
375 			/* Check each bit */
376 			offs += snprintf(buffer+offs, PAGE_SIZE-offs, "%d,",
377 					 (cache->sibling_map[i] >> j) & 1);
378 
379 	/* Replace the last "," with end of line */
380 	buffer[offs-1] = '\n';
381 	return offs;
382 }
383 
384 static const struct sysfs_ops cache_ops = {
385 	.show = kfd_cache_show,
386 };
387 
388 static struct kobj_type cache_type = {
389 	.release = kfd_topology_kobj_release,
390 	.sysfs_ops = &cache_ops,
391 };
392 
393 /****** Sysfs of Performance Counters ******/
394 
395 struct kfd_perf_attr {
396 	struct kobj_attribute attr;
397 	uint32_t data;
398 };
399 
400 static ssize_t perf_show(struct kobject *kobj, struct kobj_attribute *attrs,
401 			char *buf)
402 {
403 	int offs = 0;
404 	struct kfd_perf_attr *attr;
405 
406 	buf[0] = 0;
407 	attr = container_of(attrs, struct kfd_perf_attr, attr);
408 	if (!attr->data) /* invalid data for PMC */
409 		return 0;
410 	else
411 		return sysfs_show_32bit_val(buf, offs, attr->data);
412 }
413 
414 #define KFD_PERF_DESC(_name, _data)			\
415 {							\
416 	.attr  = __ATTR(_name, 0444, perf_show, NULL),	\
417 	.data = _data,					\
418 }
419 
420 static struct kfd_perf_attr perf_attr_iommu[] = {
421 	KFD_PERF_DESC(max_concurrent, 0),
422 	KFD_PERF_DESC(num_counters, 0),
423 	KFD_PERF_DESC(counter_ids, 0),
424 };
425 /****************************************/
426 
427 static ssize_t node_show(struct kobject *kobj, struct attribute *attr,
428 		char *buffer)
429 {
430 	int offs = 0;
431 	struct kfd_topology_device *dev;
432 	uint32_t log_max_watch_addr;
433 
434 	/* Making sure that the buffer is an empty string */
435 	buffer[0] = 0;
436 
437 	if (strcmp(attr->name, "gpu_id") == 0) {
438 		dev = container_of(attr, struct kfd_topology_device,
439 				attr_gpuid);
440 		if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu))
441 			return -EPERM;
442 		return sysfs_show_32bit_val(buffer, offs, dev->gpu_id);
443 	}
444 
445 	if (strcmp(attr->name, "name") == 0) {
446 		dev = container_of(attr, struct kfd_topology_device,
447 				attr_name);
448 
449 		if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu))
450 			return -EPERM;
451 		return sysfs_show_str_val(buffer, offs, dev->node_props.name);
452 	}
453 
454 	dev = container_of(attr, struct kfd_topology_device,
455 			attr_props);
456 	if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu))
457 		return -EPERM;
458 	sysfs_show_32bit_prop(buffer, offs, "cpu_cores_count",
459 			      dev->node_props.cpu_cores_count);
460 	sysfs_show_32bit_prop(buffer, offs, "simd_count",
461 			      dev->gpu ? dev->node_props.simd_count : 0);
462 	sysfs_show_32bit_prop(buffer, offs, "mem_banks_count",
463 			      dev->node_props.mem_banks_count);
464 	sysfs_show_32bit_prop(buffer, offs, "caches_count",
465 			      dev->node_props.caches_count);
466 	sysfs_show_32bit_prop(buffer, offs, "io_links_count",
467 			      dev->node_props.io_links_count);
468 	sysfs_show_32bit_prop(buffer, offs, "cpu_core_id_base",
469 			      dev->node_props.cpu_core_id_base);
470 	sysfs_show_32bit_prop(buffer, offs, "simd_id_base",
471 			      dev->node_props.simd_id_base);
472 	sysfs_show_32bit_prop(buffer, offs, "max_waves_per_simd",
473 			      dev->node_props.max_waves_per_simd);
474 	sysfs_show_32bit_prop(buffer, offs, "lds_size_in_kb",
475 			      dev->node_props.lds_size_in_kb);
476 	sysfs_show_32bit_prop(buffer, offs, "gds_size_in_kb",
477 			      dev->node_props.gds_size_in_kb);
478 	sysfs_show_32bit_prop(buffer, offs, "num_gws",
479 			      dev->node_props.num_gws);
480 	sysfs_show_32bit_prop(buffer, offs, "wave_front_size",
481 			      dev->node_props.wave_front_size);
482 	sysfs_show_32bit_prop(buffer, offs, "array_count",
483 			      dev->node_props.array_count);
484 	sysfs_show_32bit_prop(buffer, offs, "simd_arrays_per_engine",
485 			      dev->node_props.simd_arrays_per_engine);
486 	sysfs_show_32bit_prop(buffer, offs, "cu_per_simd_array",
487 			      dev->node_props.cu_per_simd_array);
488 	sysfs_show_32bit_prop(buffer, offs, "simd_per_cu",
489 			      dev->node_props.simd_per_cu);
490 	sysfs_show_32bit_prop(buffer, offs, "max_slots_scratch_cu",
491 			      dev->node_props.max_slots_scratch_cu);
492 	sysfs_show_32bit_prop(buffer, offs, "gfx_target_version",
493 			      dev->node_props.gfx_target_version);
494 	sysfs_show_32bit_prop(buffer, offs, "vendor_id",
495 			      dev->node_props.vendor_id);
496 	sysfs_show_32bit_prop(buffer, offs, "device_id",
497 			      dev->node_props.device_id);
498 	sysfs_show_32bit_prop(buffer, offs, "location_id",
499 			      dev->node_props.location_id);
500 	sysfs_show_32bit_prop(buffer, offs, "domain",
501 			      dev->node_props.domain);
502 	sysfs_show_32bit_prop(buffer, offs, "drm_render_minor",
503 			      dev->node_props.drm_render_minor);
504 	sysfs_show_64bit_prop(buffer, offs, "hive_id",
505 			      dev->node_props.hive_id);
506 	sysfs_show_32bit_prop(buffer, offs, "num_sdma_engines",
507 			      dev->node_props.num_sdma_engines);
508 	sysfs_show_32bit_prop(buffer, offs, "num_sdma_xgmi_engines",
509 			      dev->node_props.num_sdma_xgmi_engines);
510 	sysfs_show_32bit_prop(buffer, offs, "num_sdma_queues_per_engine",
511 			      dev->node_props.num_sdma_queues_per_engine);
512 	sysfs_show_32bit_prop(buffer, offs, "num_cp_queues",
513 			      dev->node_props.num_cp_queues);
514 
515 	if (dev->gpu) {
516 		log_max_watch_addr =
517 			__ilog2_u32(dev->gpu->device_info.num_of_watch_points);
518 
519 		if (log_max_watch_addr) {
520 			dev->node_props.capability |=
521 					HSA_CAP_WATCH_POINTS_SUPPORTED;
522 
523 			dev->node_props.capability |=
524 				((log_max_watch_addr <<
525 					HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) &
526 				HSA_CAP_WATCH_POINTS_TOTALBITS_MASK);
527 		}
528 
529 		if (dev->gpu->adev->asic_type == CHIP_TONGA)
530 			dev->node_props.capability |=
531 					HSA_CAP_AQL_QUEUE_DOUBLE_MAP;
532 
533 		sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_fcompute",
534 			dev->node_props.max_engine_clk_fcompute);
535 
536 		sysfs_show_64bit_prop(buffer, offs, "local_mem_size", 0ULL);
537 
538 		sysfs_show_32bit_prop(buffer, offs, "fw_version",
539 				      dev->gpu->mec_fw_version);
540 		sysfs_show_32bit_prop(buffer, offs, "capability",
541 				      dev->node_props.capability);
542 		sysfs_show_32bit_prop(buffer, offs, "sdma_fw_version",
543 				      dev->gpu->sdma_fw_version);
544 		sysfs_show_64bit_prop(buffer, offs, "unique_id",
545 				      dev->gpu->adev->unique_id);
546 
547 	}
548 
549 	return sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_ccompute",
550 				     cpufreq_quick_get_max(0)/1000);
551 }
552 
553 static const struct sysfs_ops node_ops = {
554 	.show = node_show,
555 };
556 
557 static struct kobj_type node_type = {
558 	.release = kfd_topology_kobj_release,
559 	.sysfs_ops = &node_ops,
560 };
561 
562 static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr)
563 {
564 	sysfs_remove_file(kobj, attr);
565 	kobject_del(kobj);
566 	kobject_put(kobj);
567 }
568 
569 static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev)
570 {
571 	struct kfd_iolink_properties *iolink;
572 	struct kfd_cache_properties *cache;
573 	struct kfd_mem_properties *mem;
574 	struct kfd_perf_properties *perf;
575 
576 	if (dev->kobj_iolink) {
577 		list_for_each_entry(iolink, &dev->io_link_props, list)
578 			if (iolink->kobj) {
579 				kfd_remove_sysfs_file(iolink->kobj,
580 							&iolink->attr);
581 				iolink->kobj = NULL;
582 			}
583 		kobject_del(dev->kobj_iolink);
584 		kobject_put(dev->kobj_iolink);
585 		dev->kobj_iolink = NULL;
586 	}
587 
588 	if (dev->kobj_cache) {
589 		list_for_each_entry(cache, &dev->cache_props, list)
590 			if (cache->kobj) {
591 				kfd_remove_sysfs_file(cache->kobj,
592 							&cache->attr);
593 				cache->kobj = NULL;
594 			}
595 		kobject_del(dev->kobj_cache);
596 		kobject_put(dev->kobj_cache);
597 		dev->kobj_cache = NULL;
598 	}
599 
600 	if (dev->kobj_mem) {
601 		list_for_each_entry(mem, &dev->mem_props, list)
602 			if (mem->kobj) {
603 				kfd_remove_sysfs_file(mem->kobj, &mem->attr);
604 				mem->kobj = NULL;
605 			}
606 		kobject_del(dev->kobj_mem);
607 		kobject_put(dev->kobj_mem);
608 		dev->kobj_mem = NULL;
609 	}
610 
611 	if (dev->kobj_perf) {
612 		list_for_each_entry(perf, &dev->perf_props, list) {
613 			kfree(perf->attr_group);
614 			perf->attr_group = NULL;
615 		}
616 		kobject_del(dev->kobj_perf);
617 		kobject_put(dev->kobj_perf);
618 		dev->kobj_perf = NULL;
619 	}
620 
621 	if (dev->kobj_node) {
622 		sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid);
623 		sysfs_remove_file(dev->kobj_node, &dev->attr_name);
624 		sysfs_remove_file(dev->kobj_node, &dev->attr_props);
625 		kobject_del(dev->kobj_node);
626 		kobject_put(dev->kobj_node);
627 		dev->kobj_node = NULL;
628 	}
629 }
630 
631 static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev,
632 		uint32_t id)
633 {
634 	struct kfd_iolink_properties *iolink;
635 	struct kfd_cache_properties *cache;
636 	struct kfd_mem_properties *mem;
637 	struct kfd_perf_properties *perf;
638 	int ret;
639 	uint32_t i, num_attrs;
640 	struct attribute **attrs;
641 
642 	if (WARN_ON(dev->kobj_node))
643 		return -EEXIST;
644 
645 	/*
646 	 * Creating the sysfs folders
647 	 */
648 	dev->kobj_node = kfd_alloc_struct(dev->kobj_node);
649 	if (!dev->kobj_node)
650 		return -ENOMEM;
651 
652 	ret = kobject_init_and_add(dev->kobj_node, &node_type,
653 			sys_props.kobj_nodes, "%d", id);
654 	if (ret < 0) {
655 		kobject_put(dev->kobj_node);
656 		return ret;
657 	}
658 
659 	dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node);
660 	if (!dev->kobj_mem)
661 		return -ENOMEM;
662 
663 	dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node);
664 	if (!dev->kobj_cache)
665 		return -ENOMEM;
666 
667 	dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node);
668 	if (!dev->kobj_iolink)
669 		return -ENOMEM;
670 
671 	dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node);
672 	if (!dev->kobj_perf)
673 		return -ENOMEM;
674 
675 	/*
676 	 * Creating sysfs files for node properties
677 	 */
678 	dev->attr_gpuid.name = "gpu_id";
679 	dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE;
680 	sysfs_attr_init(&dev->attr_gpuid);
681 	dev->attr_name.name = "name";
682 	dev->attr_name.mode = KFD_SYSFS_FILE_MODE;
683 	sysfs_attr_init(&dev->attr_name);
684 	dev->attr_props.name = "properties";
685 	dev->attr_props.mode = KFD_SYSFS_FILE_MODE;
686 	sysfs_attr_init(&dev->attr_props);
687 	ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid);
688 	if (ret < 0)
689 		return ret;
690 	ret = sysfs_create_file(dev->kobj_node, &dev->attr_name);
691 	if (ret < 0)
692 		return ret;
693 	ret = sysfs_create_file(dev->kobj_node, &dev->attr_props);
694 	if (ret < 0)
695 		return ret;
696 
697 	i = 0;
698 	list_for_each_entry(mem, &dev->mem_props, list) {
699 		mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
700 		if (!mem->kobj)
701 			return -ENOMEM;
702 		ret = kobject_init_and_add(mem->kobj, &mem_type,
703 				dev->kobj_mem, "%d", i);
704 		if (ret < 0) {
705 			kobject_put(mem->kobj);
706 			return ret;
707 		}
708 
709 		mem->attr.name = "properties";
710 		mem->attr.mode = KFD_SYSFS_FILE_MODE;
711 		sysfs_attr_init(&mem->attr);
712 		ret = sysfs_create_file(mem->kobj, &mem->attr);
713 		if (ret < 0)
714 			return ret;
715 		i++;
716 	}
717 
718 	i = 0;
719 	list_for_each_entry(cache, &dev->cache_props, list) {
720 		cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
721 		if (!cache->kobj)
722 			return -ENOMEM;
723 		ret = kobject_init_and_add(cache->kobj, &cache_type,
724 				dev->kobj_cache, "%d", i);
725 		if (ret < 0) {
726 			kobject_put(cache->kobj);
727 			return ret;
728 		}
729 
730 		cache->attr.name = "properties";
731 		cache->attr.mode = KFD_SYSFS_FILE_MODE;
732 		sysfs_attr_init(&cache->attr);
733 		ret = sysfs_create_file(cache->kobj, &cache->attr);
734 		if (ret < 0)
735 			return ret;
736 		i++;
737 	}
738 
739 	i = 0;
740 	list_for_each_entry(iolink, &dev->io_link_props, list) {
741 		iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
742 		if (!iolink->kobj)
743 			return -ENOMEM;
744 		ret = kobject_init_and_add(iolink->kobj, &iolink_type,
745 				dev->kobj_iolink, "%d", i);
746 		if (ret < 0) {
747 			kobject_put(iolink->kobj);
748 			return ret;
749 		}
750 
751 		iolink->attr.name = "properties";
752 		iolink->attr.mode = KFD_SYSFS_FILE_MODE;
753 		sysfs_attr_init(&iolink->attr);
754 		ret = sysfs_create_file(iolink->kobj, &iolink->attr);
755 		if (ret < 0)
756 			return ret;
757 		i++;
758 	}
759 
760 	/* All hardware blocks have the same number of attributes. */
761 	num_attrs = ARRAY_SIZE(perf_attr_iommu);
762 	list_for_each_entry(perf, &dev->perf_props, list) {
763 		perf->attr_group = kzalloc(sizeof(struct kfd_perf_attr)
764 			* num_attrs + sizeof(struct attribute_group),
765 			GFP_KERNEL);
766 		if (!perf->attr_group)
767 			return -ENOMEM;
768 
769 		attrs = (struct attribute **)(perf->attr_group + 1);
770 		if (!strcmp(perf->block_name, "iommu")) {
771 		/* Information of IOMMU's num_counters and counter_ids is shown
772 		 * under /sys/bus/event_source/devices/amd_iommu. We don't
773 		 * duplicate here.
774 		 */
775 			perf_attr_iommu[0].data = perf->max_concurrent;
776 			for (i = 0; i < num_attrs; i++)
777 				attrs[i] = &perf_attr_iommu[i].attr.attr;
778 		}
779 		perf->attr_group->name = perf->block_name;
780 		perf->attr_group->attrs = attrs;
781 		ret = sysfs_create_group(dev->kobj_perf, perf->attr_group);
782 		if (ret < 0)
783 			return ret;
784 	}
785 
786 	return 0;
787 }
788 
789 /* Called with write topology lock acquired */
790 static int kfd_build_sysfs_node_tree(void)
791 {
792 	struct kfd_topology_device *dev;
793 	int ret;
794 	uint32_t i = 0;
795 
796 	list_for_each_entry(dev, &topology_device_list, list) {
797 		ret = kfd_build_sysfs_node_entry(dev, i);
798 		if (ret < 0)
799 			return ret;
800 		i++;
801 	}
802 
803 	return 0;
804 }
805 
806 /* Called with write topology lock acquired */
807 static void kfd_remove_sysfs_node_tree(void)
808 {
809 	struct kfd_topology_device *dev;
810 
811 	list_for_each_entry(dev, &topology_device_list, list)
812 		kfd_remove_sysfs_node_entry(dev);
813 }
814 
815 static int kfd_topology_update_sysfs(void)
816 {
817 	int ret;
818 
819 	if (!sys_props.kobj_topology) {
820 		sys_props.kobj_topology =
821 				kfd_alloc_struct(sys_props.kobj_topology);
822 		if (!sys_props.kobj_topology)
823 			return -ENOMEM;
824 
825 		ret = kobject_init_and_add(sys_props.kobj_topology,
826 				&sysprops_type,  &kfd_device->kobj,
827 				"topology");
828 		if (ret < 0) {
829 			kobject_put(sys_props.kobj_topology);
830 			return ret;
831 		}
832 
833 		sys_props.kobj_nodes = kobject_create_and_add("nodes",
834 				sys_props.kobj_topology);
835 		if (!sys_props.kobj_nodes)
836 			return -ENOMEM;
837 
838 		sys_props.attr_genid.name = "generation_id";
839 		sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE;
840 		sysfs_attr_init(&sys_props.attr_genid);
841 		ret = sysfs_create_file(sys_props.kobj_topology,
842 				&sys_props.attr_genid);
843 		if (ret < 0)
844 			return ret;
845 
846 		sys_props.attr_props.name = "system_properties";
847 		sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE;
848 		sysfs_attr_init(&sys_props.attr_props);
849 		ret = sysfs_create_file(sys_props.kobj_topology,
850 				&sys_props.attr_props);
851 		if (ret < 0)
852 			return ret;
853 	}
854 
855 	kfd_remove_sysfs_node_tree();
856 
857 	return kfd_build_sysfs_node_tree();
858 }
859 
860 static void kfd_topology_release_sysfs(void)
861 {
862 	kfd_remove_sysfs_node_tree();
863 	if (sys_props.kobj_topology) {
864 		sysfs_remove_file(sys_props.kobj_topology,
865 				&sys_props.attr_genid);
866 		sysfs_remove_file(sys_props.kobj_topology,
867 				&sys_props.attr_props);
868 		if (sys_props.kobj_nodes) {
869 			kobject_del(sys_props.kobj_nodes);
870 			kobject_put(sys_props.kobj_nodes);
871 			sys_props.kobj_nodes = NULL;
872 		}
873 		kobject_del(sys_props.kobj_topology);
874 		kobject_put(sys_props.kobj_topology);
875 		sys_props.kobj_topology = NULL;
876 	}
877 }
878 
879 /* Called with write topology_lock acquired */
880 static void kfd_topology_update_device_list(struct list_head *temp_list,
881 					struct list_head *master_list)
882 {
883 	while (!list_empty(temp_list)) {
884 		list_move_tail(temp_list->next, master_list);
885 		sys_props.num_devices++;
886 	}
887 }
888 
889 static void kfd_debug_print_topology(void)
890 {
891 	struct kfd_topology_device *dev;
892 
893 	down_read(&topology_lock);
894 
895 	dev = list_last_entry(&topology_device_list,
896 			struct kfd_topology_device, list);
897 	if (dev) {
898 		if (dev->node_props.cpu_cores_count &&
899 				dev->node_props.simd_count) {
900 			pr_info("Topology: Add APU node [0x%0x:0x%0x]\n",
901 				dev->node_props.device_id,
902 				dev->node_props.vendor_id);
903 		} else if (dev->node_props.cpu_cores_count)
904 			pr_info("Topology: Add CPU node\n");
905 		else if (dev->node_props.simd_count)
906 			pr_info("Topology: Add dGPU node [0x%0x:0x%0x]\n",
907 				dev->node_props.device_id,
908 				dev->node_props.vendor_id);
909 	}
910 	up_read(&topology_lock);
911 }
912 
913 /* Helper function for intializing platform_xx members of
914  * kfd_system_properties. Uses OEM info from the last CPU/APU node.
915  */
916 static void kfd_update_system_properties(void)
917 {
918 	struct kfd_topology_device *dev;
919 
920 	down_read(&topology_lock);
921 	dev = list_last_entry(&topology_device_list,
922 			struct kfd_topology_device, list);
923 	if (dev) {
924 		sys_props.platform_id =
925 			(*((uint64_t *)dev->oem_id)) & CRAT_OEMID_64BIT_MASK;
926 		sys_props.platform_oem = *((uint64_t *)dev->oem_table_id);
927 		sys_props.platform_rev = dev->oem_revision;
928 	}
929 	up_read(&topology_lock);
930 }
931 
932 static void find_system_memory(const struct dmi_header *dm,
933 	void *private)
934 {
935 	struct kfd_mem_properties *mem;
936 	u16 mem_width, mem_clock;
937 	struct kfd_topology_device *kdev =
938 		(struct kfd_topology_device *)private;
939 	const u8 *dmi_data = (const u8 *)(dm + 1);
940 
941 	if (dm->type == DMI_ENTRY_MEM_DEVICE && dm->length >= 0x15) {
942 		mem_width = (u16)(*(const u16 *)(dmi_data + 0x6));
943 		mem_clock = (u16)(*(const u16 *)(dmi_data + 0x11));
944 		list_for_each_entry(mem, &kdev->mem_props, list) {
945 			if (mem_width != 0xFFFF && mem_width != 0)
946 				mem->width = mem_width;
947 			if (mem_clock != 0)
948 				mem->mem_clk_max = mem_clock;
949 		}
950 	}
951 }
952 
953 /*
954  * Performance counters information is not part of CRAT but we would like to
955  * put them in the sysfs under topology directory for Thunk to get the data.
956  * This function is called before updating the sysfs.
957  */
958 static int kfd_add_perf_to_topology(struct kfd_topology_device *kdev)
959 {
960 	/* These are the only counters supported so far */
961 	return kfd_iommu_add_perf_counters(kdev);
962 }
963 
964 /* kfd_add_non_crat_information - Add information that is not currently
965  *	defined in CRAT but is necessary for KFD topology
966  * @dev - topology device to which addition info is added
967  */
968 static void kfd_add_non_crat_information(struct kfd_topology_device *kdev)
969 {
970 	/* Check if CPU only node. */
971 	if (!kdev->gpu) {
972 		/* Add system memory information */
973 		dmi_walk(find_system_memory, kdev);
974 	}
975 	/* TODO: For GPU node, rearrange code from kfd_topology_add_device */
976 }
977 
978 /* kfd_is_acpi_crat_invalid - CRAT from ACPI is valid only for AMD APU devices.
979  *	Ignore CRAT for all other devices. AMD APU is identified if both CPU
980  *	and GPU cores are present.
981  * @device_list - topology device list created by parsing ACPI CRAT table.
982  * @return - TRUE if invalid, FALSE is valid.
983  */
984 static bool kfd_is_acpi_crat_invalid(struct list_head *device_list)
985 {
986 	struct kfd_topology_device *dev;
987 
988 	list_for_each_entry(dev, device_list, list) {
989 		if (dev->node_props.cpu_cores_count &&
990 			dev->node_props.simd_count)
991 			return false;
992 	}
993 	pr_info("Ignoring ACPI CRAT on non-APU system\n");
994 	return true;
995 }
996 
997 int kfd_topology_init(void)
998 {
999 	void *crat_image = NULL;
1000 	size_t image_size = 0;
1001 	int ret;
1002 	struct list_head temp_topology_device_list;
1003 	int cpu_only_node = 0;
1004 	struct kfd_topology_device *kdev;
1005 	int proximity_domain;
1006 
1007 	/* topology_device_list - Master list of all topology devices
1008 	 * temp_topology_device_list - temporary list created while parsing CRAT
1009 	 * or VCRAT. Once parsing is complete the contents of list is moved to
1010 	 * topology_device_list
1011 	 */
1012 
1013 	/* Initialize the head for the both the lists */
1014 	INIT_LIST_HEAD(&topology_device_list);
1015 	INIT_LIST_HEAD(&temp_topology_device_list);
1016 	init_rwsem(&topology_lock);
1017 
1018 	memset(&sys_props, 0, sizeof(sys_props));
1019 
1020 	/* Proximity domains in ACPI CRAT tables start counting at
1021 	 * 0. The same should be true for virtual CRAT tables created
1022 	 * at this stage. GPUs added later in kfd_topology_add_device
1023 	 * use a counter.
1024 	 */
1025 	proximity_domain = 0;
1026 
1027 	/*
1028 	 * Get the CRAT image from the ACPI. If ACPI doesn't have one
1029 	 * or if ACPI CRAT is invalid create a virtual CRAT.
1030 	 * NOTE: The current implementation expects all AMD APUs to have
1031 	 *	CRAT. If no CRAT is available, it is assumed to be a CPU
1032 	 */
1033 	ret = kfd_create_crat_image_acpi(&crat_image, &image_size);
1034 	if (!ret) {
1035 		ret = kfd_parse_crat_table(crat_image,
1036 					   &temp_topology_device_list,
1037 					   proximity_domain);
1038 		if (ret ||
1039 		    kfd_is_acpi_crat_invalid(&temp_topology_device_list)) {
1040 			kfd_release_topology_device_list(
1041 				&temp_topology_device_list);
1042 			kfd_destroy_crat_image(crat_image);
1043 			crat_image = NULL;
1044 		}
1045 	}
1046 
1047 	if (!crat_image) {
1048 		ret = kfd_create_crat_image_virtual(&crat_image, &image_size,
1049 						    COMPUTE_UNIT_CPU, NULL,
1050 						    proximity_domain);
1051 		cpu_only_node = 1;
1052 		if (ret) {
1053 			pr_err("Error creating VCRAT table for CPU\n");
1054 			return ret;
1055 		}
1056 
1057 		ret = kfd_parse_crat_table(crat_image,
1058 					   &temp_topology_device_list,
1059 					   proximity_domain);
1060 		if (ret) {
1061 			pr_err("Error parsing VCRAT table for CPU\n");
1062 			goto err;
1063 		}
1064 	}
1065 
1066 	kdev = list_first_entry(&temp_topology_device_list,
1067 				struct kfd_topology_device, list);
1068 	kfd_add_perf_to_topology(kdev);
1069 
1070 	down_write(&topology_lock);
1071 	kfd_topology_update_device_list(&temp_topology_device_list,
1072 					&topology_device_list);
1073 	topology_crat_proximity_domain = sys_props.num_devices-1;
1074 	ret = kfd_topology_update_sysfs();
1075 	up_write(&topology_lock);
1076 
1077 	if (!ret) {
1078 		sys_props.generation_count++;
1079 		kfd_update_system_properties();
1080 		kfd_debug_print_topology();
1081 	} else
1082 		pr_err("Failed to update topology in sysfs ret=%d\n", ret);
1083 
1084 	/* For nodes with GPU, this information gets added
1085 	 * when GPU is detected (kfd_topology_add_device).
1086 	 */
1087 	if (cpu_only_node) {
1088 		/* Add additional information to CPU only node created above */
1089 		down_write(&topology_lock);
1090 		kdev = list_first_entry(&topology_device_list,
1091 				struct kfd_topology_device, list);
1092 		up_write(&topology_lock);
1093 		kfd_add_non_crat_information(kdev);
1094 	}
1095 
1096 err:
1097 	kfd_destroy_crat_image(crat_image);
1098 	return ret;
1099 }
1100 
1101 void kfd_topology_shutdown(void)
1102 {
1103 	down_write(&topology_lock);
1104 	kfd_topology_release_sysfs();
1105 	kfd_release_live_view();
1106 	up_write(&topology_lock);
1107 }
1108 
1109 static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu)
1110 {
1111 	uint32_t hashout;
1112 	uint32_t buf[7];
1113 	uint64_t local_mem_size;
1114 	int i;
1115 
1116 	if (!gpu)
1117 		return 0;
1118 
1119 	local_mem_size = gpu->local_mem_info.local_mem_size_private +
1120 			gpu->local_mem_info.local_mem_size_public;
1121 
1122 	buf[0] = gpu->pdev->devfn;
1123 	buf[1] = gpu->pdev->subsystem_vendor |
1124 		(gpu->pdev->subsystem_device << 16);
1125 	buf[2] = pci_domain_nr(gpu->pdev->bus);
1126 	buf[3] = gpu->pdev->device;
1127 	buf[4] = gpu->pdev->bus->number;
1128 	buf[5] = lower_32_bits(local_mem_size);
1129 	buf[6] = upper_32_bits(local_mem_size);
1130 
1131 	for (i = 0, hashout = 0; i < 7; i++)
1132 		hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH);
1133 
1134 	return hashout;
1135 }
1136 /* kfd_assign_gpu - Attach @gpu to the correct kfd topology device. If
1137  *		the GPU device is not already present in the topology device
1138  *		list then return NULL. This means a new topology device has to
1139  *		be created for this GPU.
1140  */
1141 static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu)
1142 {
1143 	struct kfd_topology_device *dev;
1144 	struct kfd_topology_device *out_dev = NULL;
1145 	struct kfd_mem_properties *mem;
1146 	struct kfd_cache_properties *cache;
1147 	struct kfd_iolink_properties *iolink;
1148 
1149 	down_write(&topology_lock);
1150 	list_for_each_entry(dev, &topology_device_list, list) {
1151 		/* Discrete GPUs need their own topology device list
1152 		 * entries. Don't assign them to CPU/APU nodes.
1153 		 */
1154 		if (!gpu->use_iommu_v2 &&
1155 		    dev->node_props.cpu_cores_count)
1156 			continue;
1157 
1158 		if (!dev->gpu && (dev->node_props.simd_count > 0)) {
1159 			dev->gpu = gpu;
1160 			out_dev = dev;
1161 
1162 			list_for_each_entry(mem, &dev->mem_props, list)
1163 				mem->gpu = dev->gpu;
1164 			list_for_each_entry(cache, &dev->cache_props, list)
1165 				cache->gpu = dev->gpu;
1166 			list_for_each_entry(iolink, &dev->io_link_props, list)
1167 				iolink->gpu = dev->gpu;
1168 			break;
1169 		}
1170 	}
1171 	up_write(&topology_lock);
1172 	return out_dev;
1173 }
1174 
1175 static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival)
1176 {
1177 	/*
1178 	 * TODO: Generate an event for thunk about the arrival/removal
1179 	 * of the GPU
1180 	 */
1181 }
1182 
1183 /* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info,
1184  *		patch this after CRAT parsing.
1185  */
1186 static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev)
1187 {
1188 	struct kfd_mem_properties *mem;
1189 	struct kfd_local_mem_info local_mem_info;
1190 
1191 	if (!dev)
1192 		return;
1193 
1194 	/* Currently, amdgpu driver (amdgpu_mc) deals only with GPUs with
1195 	 * single bank of VRAM local memory.
1196 	 * for dGPUs - VCRAT reports only one bank of Local Memory
1197 	 * for APUs - If CRAT from ACPI reports more than one bank, then
1198 	 *	all the banks will report the same mem_clk_max information
1199 	 */
1200 	amdgpu_amdkfd_get_local_mem_info(dev->gpu->adev, &local_mem_info);
1201 
1202 	list_for_each_entry(mem, &dev->mem_props, list)
1203 		mem->mem_clk_max = local_mem_info.mem_clk_max;
1204 }
1205 
1206 static void kfd_set_iolink_no_atomics(struct kfd_topology_device *dev,
1207 					struct kfd_topology_device *target_gpu_dev,
1208 					struct kfd_iolink_properties *link)
1209 {
1210 	/* xgmi always supports atomics between links. */
1211 	if (link->iolink_type == CRAT_IOLINK_TYPE_XGMI)
1212 		return;
1213 
1214 	/* check pcie support to set cpu(dev) flags for target_gpu_dev link. */
1215 	if (target_gpu_dev) {
1216 		uint32_t cap;
1217 
1218 		pcie_capability_read_dword(target_gpu_dev->gpu->pdev,
1219 				PCI_EXP_DEVCAP2, &cap);
1220 
1221 		if (!(cap & (PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
1222 			     PCI_EXP_DEVCAP2_ATOMIC_COMP64)))
1223 			link->flags |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1224 				CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1225 	/* set gpu (dev) flags. */
1226 	} else {
1227 		if (!dev->gpu->pci_atomic_requested ||
1228 				dev->gpu->adev->asic_type == CHIP_HAWAII)
1229 			link->flags |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1230 				CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1231 	}
1232 }
1233 
1234 static void kfd_set_iolink_non_coherent(struct kfd_topology_device *to_dev,
1235 		struct kfd_iolink_properties *outbound_link,
1236 		struct kfd_iolink_properties *inbound_link)
1237 {
1238 	/* CPU -> GPU with PCIe */
1239 	if (!to_dev->gpu &&
1240 	    inbound_link->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS)
1241 		inbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT;
1242 
1243 	if (to_dev->gpu) {
1244 		/* GPU <-> GPU with PCIe and
1245 		 * Vega20 with XGMI
1246 		 */
1247 		if (inbound_link->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS ||
1248 		    (inbound_link->iolink_type == CRAT_IOLINK_TYPE_XGMI &&
1249 		    KFD_GC_VERSION(to_dev->gpu) == IP_VERSION(9, 4, 0))) {
1250 			outbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT;
1251 			inbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT;
1252 		}
1253 	}
1254 }
1255 
1256 static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev)
1257 {
1258 	struct kfd_iolink_properties *link, *inbound_link;
1259 	struct kfd_topology_device *peer_dev;
1260 
1261 	if (!dev || !dev->gpu)
1262 		return;
1263 
1264 	/* GPU only creates direct links so apply flags setting to all */
1265 	list_for_each_entry(link, &dev->io_link_props, list) {
1266 		link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1267 		kfd_set_iolink_no_atomics(dev, NULL, link);
1268 		peer_dev = kfd_topology_device_by_proximity_domain(
1269 				link->node_to);
1270 
1271 		if (!peer_dev)
1272 			continue;
1273 
1274 		/* Include the CPU peer in GPU hive if connected over xGMI. */
1275 		if (!peer_dev->gpu && !peer_dev->node_props.hive_id &&
1276 				dev->node_props.hive_id &&
1277 				dev->gpu->adev->gmc.xgmi.connected_to_cpu)
1278 			peer_dev->node_props.hive_id = dev->node_props.hive_id;
1279 
1280 		list_for_each_entry(inbound_link, &peer_dev->io_link_props,
1281 									list) {
1282 			if (inbound_link->node_to != link->node_from)
1283 				continue;
1284 
1285 			inbound_link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1286 			kfd_set_iolink_no_atomics(peer_dev, dev, inbound_link);
1287 			kfd_set_iolink_non_coherent(peer_dev, link, inbound_link);
1288 		}
1289 	}
1290 }
1291 
1292 int kfd_topology_add_device(struct kfd_dev *gpu)
1293 {
1294 	uint32_t gpu_id;
1295 	struct kfd_topology_device *dev;
1296 	struct kfd_cu_info cu_info;
1297 	int res = 0;
1298 	struct list_head temp_topology_device_list;
1299 	void *crat_image = NULL;
1300 	size_t image_size = 0;
1301 	int proximity_domain;
1302 	int i;
1303 	const char *asic_name = amdgpu_asic_name[gpu->adev->asic_type];
1304 
1305 	INIT_LIST_HEAD(&temp_topology_device_list);
1306 
1307 	gpu_id = kfd_generate_gpu_id(gpu);
1308 
1309 	pr_debug("Adding new GPU (ID: 0x%x) to topology\n", gpu_id);
1310 
1311 	/* Check to see if this gpu device exists in the topology_device_list.
1312 	 * If so, assign the gpu to that device,
1313 	 * else create a Virtual CRAT for this gpu device and then parse that
1314 	 * CRAT to create a new topology device. Once created assign the gpu to
1315 	 * that topology device
1316 	 */
1317 	dev = kfd_assign_gpu(gpu);
1318 	if (!dev) {
1319 		down_write(&topology_lock);
1320 		proximity_domain = ++topology_crat_proximity_domain;
1321 
1322 		res = kfd_create_crat_image_virtual(&crat_image, &image_size,
1323 						    COMPUTE_UNIT_GPU, gpu,
1324 						    proximity_domain);
1325 		if (res) {
1326 			pr_err("Error creating VCRAT for GPU (ID: 0x%x)\n",
1327 			       gpu_id);
1328 			topology_crat_proximity_domain--;
1329 			return res;
1330 		}
1331 		res = kfd_parse_crat_table(crat_image,
1332 					   &temp_topology_device_list,
1333 					   proximity_domain);
1334 		if (res) {
1335 			pr_err("Error parsing VCRAT for GPU (ID: 0x%x)\n",
1336 			       gpu_id);
1337 			topology_crat_proximity_domain--;
1338 			goto err;
1339 		}
1340 
1341 		kfd_topology_update_device_list(&temp_topology_device_list,
1342 			&topology_device_list);
1343 
1344 		/* Update the SYSFS tree, since we added another topology
1345 		 * device
1346 		 */
1347 		res = kfd_topology_update_sysfs();
1348 		up_write(&topology_lock);
1349 
1350 		if (!res)
1351 			sys_props.generation_count++;
1352 		else
1353 			pr_err("Failed to update GPU (ID: 0x%x) to sysfs topology. res=%d\n",
1354 						gpu_id, res);
1355 		dev = kfd_assign_gpu(gpu);
1356 		if (WARN_ON(!dev)) {
1357 			res = -ENODEV;
1358 			goto err;
1359 		}
1360 	}
1361 
1362 	dev->gpu_id = gpu_id;
1363 	gpu->id = gpu_id;
1364 
1365 	/* TODO: Move the following lines to function
1366 	 *	kfd_add_non_crat_information
1367 	 */
1368 
1369 	/* Fill-in additional information that is not available in CRAT but
1370 	 * needed for the topology
1371 	 */
1372 
1373 	amdgpu_amdkfd_get_cu_info(dev->gpu->adev, &cu_info);
1374 
1375 	for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1; i++) {
1376 		dev->node_props.name[i] = __tolower(asic_name[i]);
1377 		if (asic_name[i] == '\0')
1378 			break;
1379 	}
1380 	dev->node_props.name[i] = '\0';
1381 
1382 	dev->node_props.simd_arrays_per_engine =
1383 		cu_info.num_shader_arrays_per_engine;
1384 
1385 	dev->node_props.gfx_target_version = gpu->device_info.gfx_target_version;
1386 	dev->node_props.vendor_id = gpu->pdev->vendor;
1387 	dev->node_props.device_id = gpu->pdev->device;
1388 	dev->node_props.capability |=
1389 		((dev->gpu->adev->rev_id << HSA_CAP_ASIC_REVISION_SHIFT) &
1390 			HSA_CAP_ASIC_REVISION_MASK);
1391 	dev->node_props.location_id = pci_dev_id(gpu->pdev);
1392 	dev->node_props.domain = pci_domain_nr(gpu->pdev->bus);
1393 	dev->node_props.max_engine_clk_fcompute =
1394 		amdgpu_amdkfd_get_max_engine_clock_in_mhz(dev->gpu->adev);
1395 	dev->node_props.max_engine_clk_ccompute =
1396 		cpufreq_quick_get_max(0) / 1000;
1397 	dev->node_props.drm_render_minor =
1398 		gpu->shared_resources.drm_render_minor;
1399 
1400 	dev->node_props.hive_id = gpu->hive_id;
1401 	dev->node_props.num_sdma_engines = kfd_get_num_sdma_engines(gpu);
1402 	dev->node_props.num_sdma_xgmi_engines =
1403 					kfd_get_num_xgmi_sdma_engines(gpu);
1404 	dev->node_props.num_sdma_queues_per_engine =
1405 				gpu->device_info.num_sdma_queues_per_engine -
1406 				gpu->device_info.num_reserved_sdma_queues_per_engine;
1407 	dev->node_props.num_gws = (dev->gpu->gws &&
1408 		dev->gpu->dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) ?
1409 		dev->gpu->adev->gds.gws_size : 0;
1410 	dev->node_props.num_cp_queues = get_cp_queues_num(dev->gpu->dqm);
1411 
1412 	kfd_fill_mem_clk_max_info(dev);
1413 	kfd_fill_iolink_non_crat_info(dev);
1414 
1415 	switch (dev->gpu->adev->asic_type) {
1416 	case CHIP_KAVERI:
1417 	case CHIP_HAWAII:
1418 	case CHIP_TONGA:
1419 		dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_PRE_1_0 <<
1420 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1421 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1422 		break;
1423 	case CHIP_CARRIZO:
1424 	case CHIP_FIJI:
1425 	case CHIP_POLARIS10:
1426 	case CHIP_POLARIS11:
1427 	case CHIP_POLARIS12:
1428 	case CHIP_VEGAM:
1429 		pr_debug("Adding doorbell packet type capability\n");
1430 		dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 <<
1431 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1432 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1433 		break;
1434 	default:
1435 		if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(9, 0, 1))
1436 			dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 <<
1437 				HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1438 				HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1439 		else
1440 			WARN(1, "Unexpected ASIC family %u",
1441 			     dev->gpu->adev->asic_type);
1442 	}
1443 
1444 	/*
1445 	 * Overwrite ATS capability according to needs_iommu_device to fix
1446 	 * potential missing corresponding bit in CRAT of BIOS.
1447 	 */
1448 	if (dev->gpu->use_iommu_v2)
1449 		dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
1450 	else
1451 		dev->node_props.capability &= ~HSA_CAP_ATS_PRESENT;
1452 
1453 	/* Fix errors in CZ CRAT.
1454 	 * simd_count: Carrizo CRAT reports wrong simd_count, probably
1455 	 *		because it doesn't consider masked out CUs
1456 	 * max_waves_per_simd: Carrizo reports wrong max_waves_per_simd
1457 	 */
1458 	if (dev->gpu->adev->asic_type == CHIP_CARRIZO) {
1459 		dev->node_props.simd_count =
1460 			cu_info.simd_per_cu * cu_info.cu_active_number;
1461 		dev->node_props.max_waves_per_simd = 10;
1462 	}
1463 
1464 	/* kfd only concerns sram ecc on GFX and HBM ecc on UMC */
1465 	dev->node_props.capability |=
1466 		((dev->gpu->adev->ras_enabled & BIT(AMDGPU_RAS_BLOCK__GFX)) != 0) ?
1467 		HSA_CAP_SRAM_EDCSUPPORTED : 0;
1468 	dev->node_props.capability |=
1469 		((dev->gpu->adev->ras_enabled & BIT(AMDGPU_RAS_BLOCK__UMC)) != 0) ?
1470 		HSA_CAP_MEM_EDCSUPPORTED : 0;
1471 
1472 	if (KFD_GC_VERSION(dev->gpu) != IP_VERSION(9, 0, 1))
1473 		dev->node_props.capability |= (dev->gpu->adev->ras_enabled != 0) ?
1474 			HSA_CAP_RASEVENTNOTIFY : 0;
1475 
1476 	if (KFD_IS_SVM_API_SUPPORTED(dev->gpu->adev->kfd.dev))
1477 		dev->node_props.capability |= HSA_CAP_SVMAPI_SUPPORTED;
1478 
1479 	kfd_debug_print_topology();
1480 
1481 	if (!res)
1482 		kfd_notify_gpu_change(gpu_id, 1);
1483 err:
1484 	kfd_destroy_crat_image(crat_image);
1485 	return res;
1486 }
1487 
1488 /**
1489  * kfd_topology_update_io_links() - Update IO links after device removal.
1490  * @proximity_domain: Proximity domain value of the dev being removed.
1491  *
1492  * The topology list currently is arranged in increasing order of
1493  * proximity domain.
1494  *
1495  * Two things need to be done when a device is removed:
1496  * 1. All the IO links to this device need to be removed.
1497  * 2. All nodes after the current device node need to move
1498  *    up once this device node is removed from the topology
1499  *    list. As a result, the proximity domain values for
1500  *    all nodes after the node being deleted reduce by 1.
1501  *    This would also cause the proximity domain values for
1502  *    io links to be updated based on new proximity domain
1503  *    values.
1504  *
1505  * Context: The caller must hold write topology_lock.
1506  */
1507 static void kfd_topology_update_io_links(int proximity_domain)
1508 {
1509 	struct kfd_topology_device *dev;
1510 	struct kfd_iolink_properties *iolink, *tmp;
1511 
1512 	list_for_each_entry(dev, &topology_device_list, list) {
1513 		if (dev->proximity_domain > proximity_domain)
1514 			dev->proximity_domain--;
1515 
1516 		list_for_each_entry_safe(iolink, tmp, &dev->io_link_props, list) {
1517 			/*
1518 			 * If there is an io link to the dev being deleted
1519 			 * then remove that IO link also.
1520 			 */
1521 			if (iolink->node_to == proximity_domain) {
1522 				list_del(&iolink->list);
1523 				dev->io_link_count--;
1524 				dev->node_props.io_links_count--;
1525 			} else {
1526 				if (iolink->node_from > proximity_domain)
1527 					iolink->node_from--;
1528 				if (iolink->node_to > proximity_domain)
1529 					iolink->node_to--;
1530 			}
1531 		}
1532 	}
1533 }
1534 
1535 int kfd_topology_remove_device(struct kfd_dev *gpu)
1536 {
1537 	struct kfd_topology_device *dev, *tmp;
1538 	uint32_t gpu_id;
1539 	int res = -ENODEV;
1540 	int i = 0;
1541 
1542 	down_write(&topology_lock);
1543 
1544 	list_for_each_entry_safe(dev, tmp, &topology_device_list, list) {
1545 		if (dev->gpu == gpu) {
1546 			gpu_id = dev->gpu_id;
1547 			kfd_remove_sysfs_node_entry(dev);
1548 			kfd_release_topology_device(dev);
1549 			sys_props.num_devices--;
1550 			kfd_topology_update_io_links(i);
1551 			topology_crat_proximity_domain = sys_props.num_devices-1;
1552 			sys_props.generation_count++;
1553 			res = 0;
1554 			if (kfd_topology_update_sysfs() < 0)
1555 				kfd_topology_release_sysfs();
1556 			break;
1557 		}
1558 		i++;
1559 	}
1560 
1561 	up_write(&topology_lock);
1562 
1563 	if (!res)
1564 		kfd_notify_gpu_change(gpu_id, 0);
1565 
1566 	return res;
1567 }
1568 
1569 /* kfd_topology_enum_kfd_devices - Enumerate through all devices in KFD
1570  *	topology. If GPU device is found @idx, then valid kfd_dev pointer is
1571  *	returned through @kdev
1572  * Return -	0: On success (@kdev will be NULL for non GPU nodes)
1573  *		-1: If end of list
1574  */
1575 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev)
1576 {
1577 
1578 	struct kfd_topology_device *top_dev;
1579 	uint8_t device_idx = 0;
1580 
1581 	*kdev = NULL;
1582 	down_read(&topology_lock);
1583 
1584 	list_for_each_entry(top_dev, &topology_device_list, list) {
1585 		if (device_idx == idx) {
1586 			*kdev = top_dev->gpu;
1587 			up_read(&topology_lock);
1588 			return 0;
1589 		}
1590 
1591 		device_idx++;
1592 	}
1593 
1594 	up_read(&topology_lock);
1595 
1596 	return -1;
1597 
1598 }
1599 
1600 static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask)
1601 {
1602 	int first_cpu_of_numa_node;
1603 
1604 	if (!cpumask || cpumask == cpu_none_mask)
1605 		return -1;
1606 	first_cpu_of_numa_node = cpumask_first(cpumask);
1607 	if (first_cpu_of_numa_node >= nr_cpu_ids)
1608 		return -1;
1609 #ifdef CONFIG_X86_64
1610 	return cpu_data(first_cpu_of_numa_node).apicid;
1611 #else
1612 	return first_cpu_of_numa_node;
1613 #endif
1614 }
1615 
1616 /* kfd_numa_node_to_apic_id - Returns the APIC ID of the first logical processor
1617  *	of the given NUMA node (numa_node_id)
1618  * Return -1 on failure
1619  */
1620 int kfd_numa_node_to_apic_id(int numa_node_id)
1621 {
1622 	if (numa_node_id == -1) {
1623 		pr_warn("Invalid NUMA Node. Use online CPU mask\n");
1624 		return kfd_cpumask_to_apic_id(cpu_online_mask);
1625 	}
1626 	return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id));
1627 }
1628 
1629 void kfd_double_confirm_iommu_support(struct kfd_dev *gpu)
1630 {
1631 	struct kfd_topology_device *dev;
1632 
1633 	gpu->use_iommu_v2 = false;
1634 
1635 	if (!gpu->device_info.needs_iommu_device)
1636 		return;
1637 
1638 	down_read(&topology_lock);
1639 
1640 	/* Only use IOMMUv2 if there is an APU topology node with no GPU
1641 	 * assigned yet. This GPU will be assigned to it.
1642 	 */
1643 	list_for_each_entry(dev, &topology_device_list, list)
1644 		if (dev->node_props.cpu_cores_count &&
1645 		    dev->node_props.simd_count &&
1646 		    !dev->gpu)
1647 			gpu->use_iommu_v2 = true;
1648 
1649 	up_read(&topology_lock);
1650 }
1651 
1652 #if defined(CONFIG_DEBUG_FS)
1653 
1654 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data)
1655 {
1656 	struct kfd_topology_device *dev;
1657 	unsigned int i = 0;
1658 	int r = 0;
1659 
1660 	down_read(&topology_lock);
1661 
1662 	list_for_each_entry(dev, &topology_device_list, list) {
1663 		if (!dev->gpu) {
1664 			i++;
1665 			continue;
1666 		}
1667 
1668 		seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1669 		r = dqm_debugfs_hqds(m, dev->gpu->dqm);
1670 		if (r)
1671 			break;
1672 	}
1673 
1674 	up_read(&topology_lock);
1675 
1676 	return r;
1677 }
1678 
1679 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data)
1680 {
1681 	struct kfd_topology_device *dev;
1682 	unsigned int i = 0;
1683 	int r = 0;
1684 
1685 	down_read(&topology_lock);
1686 
1687 	list_for_each_entry(dev, &topology_device_list, list) {
1688 		if (!dev->gpu) {
1689 			i++;
1690 			continue;
1691 		}
1692 
1693 		seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1694 		r = pm_debugfs_runlist(m, &dev->gpu->dqm->packet_mgr);
1695 		if (r)
1696 			break;
1697 	}
1698 
1699 	up_read(&topology_lock);
1700 
1701 	return r;
1702 }
1703 
1704 #endif
1705