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