xref: /netbsd-src/sys/external/bsd/drm2/dist/drm/amd/amdkfd/kfd_topology.c (revision 41ec02673d281bbb3d38e6c78504ce6e30c228c1)
1 /*	$NetBSD: kfd_topology.c,v 1.3 2021/12/18 23:44:59 riastradh Exp $	*/
2 
3 /*
4  * Copyright 2014 Advanced Micro Devices, Inc.
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a
7  * copy of this software and associated documentation files (the "Software"),
8  * to deal in the Software without restriction, including without limitation
9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10  * and/or sell copies of the Software, and to permit persons to whom the
11  * Software is furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in
14  * all copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22  * OTHER DEALINGS IN THE SOFTWARE.
23  */
24 
25 #include <sys/cdefs.h>
26 __KERNEL_RCSID(0, "$NetBSD: kfd_topology.c,v 1.3 2021/12/18 23:44:59 riastradh Exp $");
27 
28 #include <linux/types.h>
29 #include <linux/kernel.h>
30 #include <linux/pci.h>
31 #include <linux/errno.h>
32 #include <linux/acpi.h>
33 #include <linux/hash.h>
34 #include <linux/cpufreq.h>
35 #include <linux/log2.h>
36 #include <linux/dmi.h>
37 #include <linux/atomic.h>
38 
39 #include "kfd_priv.h"
40 #include "kfd_crat.h"
41 #include "kfd_topology.h"
42 #include "kfd_device_queue_manager.h"
43 #include "kfd_iommu.h"
44 #include "amdgpu_amdkfd.h"
45 #include "amdgpu_ras.h"
46 
47 /* topology_device_list - Master list of all topology devices */
48 static struct list_head topology_device_list;
49 static struct kfd_system_properties sys_props;
50 
51 static DECLARE_RWSEM(topology_lock);
52 static atomic_t topology_crat_proximity_domain;
53 
kfd_topology_device_by_proximity_domain(uint32_t proximity_domain)54 struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
55 						uint32_t proximity_domain)
56 {
57 	struct kfd_topology_device *top_dev;
58 	struct kfd_topology_device *device = NULL;
59 
60 	down_read(&topology_lock);
61 
62 	list_for_each_entry(top_dev, &topology_device_list, list)
63 		if (top_dev->proximity_domain == proximity_domain) {
64 			device = top_dev;
65 			break;
66 		}
67 
68 	up_read(&topology_lock);
69 
70 	return device;
71 }
72 
kfd_topology_device_by_id(uint32_t gpu_id)73 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id)
74 {
75 	struct kfd_topology_device *top_dev = NULL;
76 	struct kfd_topology_device *ret = NULL;
77 
78 	down_read(&topology_lock);
79 
80 	list_for_each_entry(top_dev, &topology_device_list, list)
81 		if (top_dev->gpu_id == gpu_id) {
82 			ret = top_dev;
83 			break;
84 		}
85 
86 	up_read(&topology_lock);
87 
88 	return ret;
89 }
90 
kfd_device_by_id(uint32_t gpu_id)91 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id)
92 {
93 	struct kfd_topology_device *top_dev;
94 
95 	top_dev = kfd_topology_device_by_id(gpu_id);
96 	if (!top_dev)
97 		return NULL;
98 
99 	return top_dev->gpu;
100 }
101 
kfd_device_by_pci_dev(const struct pci_dev * pdev)102 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev)
103 {
104 	struct kfd_topology_device *top_dev;
105 	struct kfd_dev *device = NULL;
106 
107 	down_read(&topology_lock);
108 
109 	list_for_each_entry(top_dev, &topology_device_list, list)
110 		if (top_dev->gpu && top_dev->gpu->pdev == pdev) {
111 			device = top_dev->gpu;
112 			break;
113 		}
114 
115 	up_read(&topology_lock);
116 
117 	return device;
118 }
119 
kfd_device_by_kgd(const struct kgd_dev * kgd)120 struct kfd_dev *kfd_device_by_kgd(const struct kgd_dev *kgd)
121 {
122 	struct kfd_topology_device *top_dev;
123 	struct kfd_dev *device = NULL;
124 
125 	down_read(&topology_lock);
126 
127 	list_for_each_entry(top_dev, &topology_device_list, list)
128 		if (top_dev->gpu && top_dev->gpu->kgd == kgd) {
129 			device = top_dev->gpu;
130 			break;
131 		}
132 
133 	up_read(&topology_lock);
134 
135 	return device;
136 }
137 
138 /* Called with write topology_lock acquired */
kfd_release_topology_device(struct kfd_topology_device * dev)139 static void kfd_release_topology_device(struct kfd_topology_device *dev)
140 {
141 	struct kfd_mem_properties *mem;
142 	struct kfd_cache_properties *cache;
143 	struct kfd_iolink_properties *iolink;
144 	struct kfd_perf_properties *perf;
145 
146 	list_del(&dev->list);
147 
148 	while (dev->mem_props.next != &dev->mem_props) {
149 		mem = container_of(dev->mem_props.next,
150 				struct kfd_mem_properties, list);
151 		list_del(&mem->list);
152 		kfree(mem);
153 	}
154 
155 	while (dev->cache_props.next != &dev->cache_props) {
156 		cache = container_of(dev->cache_props.next,
157 				struct kfd_cache_properties, list);
158 		list_del(&cache->list);
159 		kfree(cache);
160 	}
161 
162 	while (dev->io_link_props.next != &dev->io_link_props) {
163 		iolink = container_of(dev->io_link_props.next,
164 				struct kfd_iolink_properties, list);
165 		list_del(&iolink->list);
166 		kfree(iolink);
167 	}
168 
169 	while (dev->perf_props.next != &dev->perf_props) {
170 		perf = container_of(dev->perf_props.next,
171 				struct kfd_perf_properties, list);
172 		list_del(&perf->list);
173 		kfree(perf);
174 	}
175 
176 	kfree(dev);
177 }
178 
kfd_release_topology_device_list(struct list_head * device_list)179 void kfd_release_topology_device_list(struct list_head *device_list)
180 {
181 	struct kfd_topology_device *dev;
182 
183 	while (!list_empty(device_list)) {
184 		dev = list_first_entry(device_list,
185 				       struct kfd_topology_device, list);
186 		kfd_release_topology_device(dev);
187 	}
188 }
189 
kfd_release_live_view(void)190 static void kfd_release_live_view(void)
191 {
192 	kfd_release_topology_device_list(&topology_device_list);
193 	memset(&sys_props, 0, sizeof(sys_props));
194 }
195 
kfd_create_topology_device(struct list_head * device_list)196 struct kfd_topology_device *kfd_create_topology_device(
197 				struct list_head *device_list)
198 {
199 	struct kfd_topology_device *dev;
200 
201 	dev = kfd_alloc_struct(dev);
202 	if (!dev) {
203 		pr_err("No memory to allocate a topology device");
204 		return NULL;
205 	}
206 
207 	INIT_LIST_HEAD(&dev->mem_props);
208 	INIT_LIST_HEAD(&dev->cache_props);
209 	INIT_LIST_HEAD(&dev->io_link_props);
210 	INIT_LIST_HEAD(&dev->perf_props);
211 
212 	list_add_tail(&dev->list, device_list);
213 
214 	return dev;
215 }
216 
217 
218 #define sysfs_show_gen_prop(buffer, fmt, ...) \
219 		snprintf(buffer, PAGE_SIZE, "%s"fmt, buffer, __VA_ARGS__)
220 #define sysfs_show_32bit_prop(buffer, name, value) \
221 		sysfs_show_gen_prop(buffer, "%s %u\n", name, value)
222 #define sysfs_show_64bit_prop(buffer, name, value) \
223 		sysfs_show_gen_prop(buffer, "%s %llu\n", name, value)
224 #define sysfs_show_32bit_val(buffer, value) \
225 		sysfs_show_gen_prop(buffer, "%u\n", value)
226 #define sysfs_show_str_val(buffer, value) \
227 		sysfs_show_gen_prop(buffer, "%s\n", value)
228 
sysprops_show(struct kobject * kobj,struct attribute * attr,char * buffer)229 static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr,
230 		char *buffer)
231 {
232 	ssize_t ret;
233 
234 	/* Making sure that the buffer is an empty string */
235 	buffer[0] = 0;
236 
237 	if (attr == &sys_props.attr_genid) {
238 		ret = sysfs_show_32bit_val(buffer, sys_props.generation_count);
239 	} else if (attr == &sys_props.attr_props) {
240 		sysfs_show_64bit_prop(buffer, "platform_oem",
241 				sys_props.platform_oem);
242 		sysfs_show_64bit_prop(buffer, "platform_id",
243 				sys_props.platform_id);
244 		ret = sysfs_show_64bit_prop(buffer, "platform_rev",
245 				sys_props.platform_rev);
246 	} else {
247 		ret = -EINVAL;
248 	}
249 
250 	return ret;
251 }
252 
kfd_topology_kobj_release(struct kobject * kobj)253 static void kfd_topology_kobj_release(struct kobject *kobj)
254 {
255 	kfree(kobj);
256 }
257 
258 static const struct sysfs_ops sysprops_ops = {
259 	.show = sysprops_show,
260 };
261 
262 static struct kobj_type sysprops_type = {
263 	.release = kfd_topology_kobj_release,
264 	.sysfs_ops = &sysprops_ops,
265 };
266 
iolink_show(struct kobject * kobj,struct attribute * attr,char * buffer)267 static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr,
268 		char *buffer)
269 {
270 	ssize_t ret;
271 	struct kfd_iolink_properties *iolink;
272 
273 	/* Making sure that the buffer is an empty string */
274 	buffer[0] = 0;
275 
276 	iolink = container_of(attr, struct kfd_iolink_properties, attr);
277 	if (iolink->gpu && kfd_devcgroup_check_permission(iolink->gpu))
278 		return -EPERM;
279 	sysfs_show_32bit_prop(buffer, "type", iolink->iolink_type);
280 	sysfs_show_32bit_prop(buffer, "version_major", iolink->ver_maj);
281 	sysfs_show_32bit_prop(buffer, "version_minor", iolink->ver_min);
282 	sysfs_show_32bit_prop(buffer, "node_from", iolink->node_from);
283 	sysfs_show_32bit_prop(buffer, "node_to", iolink->node_to);
284 	sysfs_show_32bit_prop(buffer, "weight", iolink->weight);
285 	sysfs_show_32bit_prop(buffer, "min_latency", iolink->min_latency);
286 	sysfs_show_32bit_prop(buffer, "max_latency", iolink->max_latency);
287 	sysfs_show_32bit_prop(buffer, "min_bandwidth", iolink->min_bandwidth);
288 	sysfs_show_32bit_prop(buffer, "max_bandwidth", iolink->max_bandwidth);
289 	sysfs_show_32bit_prop(buffer, "recommended_transfer_size",
290 			iolink->rec_transfer_size);
291 	ret = sysfs_show_32bit_prop(buffer, "flags", iolink->flags);
292 
293 	return ret;
294 }
295 
296 static const struct sysfs_ops iolink_ops = {
297 	.show = iolink_show,
298 };
299 
300 static struct kobj_type iolink_type = {
301 	.release = kfd_topology_kobj_release,
302 	.sysfs_ops = &iolink_ops,
303 };
304 
mem_show(struct kobject * kobj,struct attribute * attr,char * buffer)305 static ssize_t mem_show(struct kobject *kobj, struct attribute *attr,
306 		char *buffer)
307 {
308 	ssize_t ret;
309 	struct kfd_mem_properties *mem;
310 
311 	/* Making sure that the buffer is an empty string */
312 	buffer[0] = 0;
313 
314 	mem = container_of(attr, struct kfd_mem_properties, attr);
315 	if (mem->gpu && kfd_devcgroup_check_permission(mem->gpu))
316 		return -EPERM;
317 	sysfs_show_32bit_prop(buffer, "heap_type", mem->heap_type);
318 	sysfs_show_64bit_prop(buffer, "size_in_bytes", mem->size_in_bytes);
319 	sysfs_show_32bit_prop(buffer, "flags", mem->flags);
320 	sysfs_show_32bit_prop(buffer, "width", mem->width);
321 	ret = sysfs_show_32bit_prop(buffer, "mem_clk_max", mem->mem_clk_max);
322 
323 	return ret;
324 }
325 
326 static const struct sysfs_ops mem_ops = {
327 	.show = mem_show,
328 };
329 
330 static struct kobj_type mem_type = {
331 	.release = kfd_topology_kobj_release,
332 	.sysfs_ops = &mem_ops,
333 };
334 
kfd_cache_show(struct kobject * kobj,struct attribute * attr,char * buffer)335 static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr,
336 		char *buffer)
337 {
338 	ssize_t ret;
339 	uint32_t i, j;
340 	struct kfd_cache_properties *cache;
341 
342 	/* Making sure that the buffer is an empty string */
343 	buffer[0] = 0;
344 
345 	cache = container_of(attr, struct kfd_cache_properties, attr);
346 	if (cache->gpu && kfd_devcgroup_check_permission(cache->gpu))
347 		return -EPERM;
348 	sysfs_show_32bit_prop(buffer, "processor_id_low",
349 			cache->processor_id_low);
350 	sysfs_show_32bit_prop(buffer, "level", cache->cache_level);
351 	sysfs_show_32bit_prop(buffer, "size", cache->cache_size);
352 	sysfs_show_32bit_prop(buffer, "cache_line_size", cache->cacheline_size);
353 	sysfs_show_32bit_prop(buffer, "cache_lines_per_tag",
354 			cache->cachelines_per_tag);
355 	sysfs_show_32bit_prop(buffer, "association", cache->cache_assoc);
356 	sysfs_show_32bit_prop(buffer, "latency", cache->cache_latency);
357 	sysfs_show_32bit_prop(buffer, "type", cache->cache_type);
358 	snprintf(buffer, PAGE_SIZE, "%ssibling_map ", buffer);
359 	for (i = 0; i < CRAT_SIBLINGMAP_SIZE; i++)
360 		for (j = 0; j < sizeof(cache->sibling_map[0])*8; j++) {
361 			/* Check each bit */
362 			if (cache->sibling_map[i] & (1 << j))
363 				ret = snprintf(buffer, PAGE_SIZE,
364 					 "%s%d%s", buffer, 1, ",");
365 			else
366 				ret = snprintf(buffer, PAGE_SIZE,
367 					 "%s%d%s", buffer, 0, ",");
368 		}
369 	/* Replace the last "," with end of line */
370 	*(buffer + strlen(buffer) - 1) = 0xA;
371 	return ret;
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 
perf_show(struct kobject * kobj,struct kobj_attribute * attrs,char * buf)390 static ssize_t perf_show(struct kobject *kobj, struct kobj_attribute *attrs,
391 			char *buf)
392 {
393 	struct kfd_perf_attr *attr;
394 
395 	buf[0] = 0;
396 	attr = container_of(attrs, struct kfd_perf_attr, attr);
397 	if (!attr->data) /* invalid data for PMC */
398 		return 0;
399 	else
400 		return sysfs_show_32bit_val(buf, attr->data);
401 }
402 
403 #define KFD_PERF_DESC(_name, _data)			\
404 {							\
405 	.attr  = __ATTR(_name, 0444, perf_show, NULL),	\
406 	.data = _data,					\
407 }
408 
409 static struct kfd_perf_attr perf_attr_iommu[] = {
410 	KFD_PERF_DESC(max_concurrent, 0),
411 	KFD_PERF_DESC(num_counters, 0),
412 	KFD_PERF_DESC(counter_ids, 0),
413 };
414 /****************************************/
415 
node_show(struct kobject * kobj,struct attribute * attr,char * buffer)416 static ssize_t node_show(struct kobject *kobj, struct attribute *attr,
417 		char *buffer)
418 {
419 	struct kfd_topology_device *dev;
420 	uint32_t log_max_watch_addr;
421 
422 	/* Making sure that the buffer is an empty string */
423 	buffer[0] = 0;
424 
425 	if (strcmp(attr->name, "gpu_id") == 0) {
426 		dev = container_of(attr, struct kfd_topology_device,
427 				attr_gpuid);
428 		if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu))
429 			return -EPERM;
430 		return sysfs_show_32bit_val(buffer, dev->gpu_id);
431 	}
432 
433 	if (strcmp(attr->name, "name") == 0) {
434 		dev = container_of(attr, struct kfd_topology_device,
435 				attr_name);
436 
437 		if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu))
438 			return -EPERM;
439 		return sysfs_show_str_val(buffer, dev->node_props.name);
440 	}
441 
442 	dev = container_of(attr, struct kfd_topology_device,
443 			attr_props);
444 	if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu))
445 		return -EPERM;
446 	sysfs_show_32bit_prop(buffer, "cpu_cores_count",
447 			dev->node_props.cpu_cores_count);
448 	sysfs_show_32bit_prop(buffer, "simd_count",
449 			dev->node_props.simd_count);
450 	sysfs_show_32bit_prop(buffer, "mem_banks_count",
451 			dev->node_props.mem_banks_count);
452 	sysfs_show_32bit_prop(buffer, "caches_count",
453 			dev->node_props.caches_count);
454 	sysfs_show_32bit_prop(buffer, "io_links_count",
455 			dev->node_props.io_links_count);
456 	sysfs_show_32bit_prop(buffer, "cpu_core_id_base",
457 			dev->node_props.cpu_core_id_base);
458 	sysfs_show_32bit_prop(buffer, "simd_id_base",
459 			dev->node_props.simd_id_base);
460 	sysfs_show_32bit_prop(buffer, "max_waves_per_simd",
461 			dev->node_props.max_waves_per_simd);
462 	sysfs_show_32bit_prop(buffer, "lds_size_in_kb",
463 			dev->node_props.lds_size_in_kb);
464 	sysfs_show_32bit_prop(buffer, "gds_size_in_kb",
465 			dev->node_props.gds_size_in_kb);
466 	sysfs_show_32bit_prop(buffer, "num_gws",
467 			dev->node_props.num_gws);
468 	sysfs_show_32bit_prop(buffer, "wave_front_size",
469 			dev->node_props.wave_front_size);
470 	sysfs_show_32bit_prop(buffer, "array_count",
471 			dev->node_props.array_count);
472 	sysfs_show_32bit_prop(buffer, "simd_arrays_per_engine",
473 			dev->node_props.simd_arrays_per_engine);
474 	sysfs_show_32bit_prop(buffer, "cu_per_simd_array",
475 			dev->node_props.cu_per_simd_array);
476 	sysfs_show_32bit_prop(buffer, "simd_per_cu",
477 			dev->node_props.simd_per_cu);
478 	sysfs_show_32bit_prop(buffer, "max_slots_scratch_cu",
479 			dev->node_props.max_slots_scratch_cu);
480 	sysfs_show_32bit_prop(buffer, "vendor_id",
481 			dev->node_props.vendor_id);
482 	sysfs_show_32bit_prop(buffer, "device_id",
483 			dev->node_props.device_id);
484 	sysfs_show_32bit_prop(buffer, "location_id",
485 			dev->node_props.location_id);
486 	sysfs_show_32bit_prop(buffer, "drm_render_minor",
487 			dev->node_props.drm_render_minor);
488 	sysfs_show_64bit_prop(buffer, "hive_id",
489 			dev->node_props.hive_id);
490 	sysfs_show_32bit_prop(buffer, "num_sdma_engines",
491 			dev->node_props.num_sdma_engines);
492 	sysfs_show_32bit_prop(buffer, "num_sdma_xgmi_engines",
493 			dev->node_props.num_sdma_xgmi_engines);
494 	sysfs_show_32bit_prop(buffer, "num_sdma_queues_per_engine",
495 			dev->node_props.num_sdma_queues_per_engine);
496 	sysfs_show_32bit_prop(buffer, "num_cp_queues",
497 			dev->node_props.num_cp_queues);
498 
499 	if (dev->gpu) {
500 		log_max_watch_addr =
501 			__ilog2_u32(dev->gpu->device_info->num_of_watch_points);
502 
503 		if (log_max_watch_addr) {
504 			dev->node_props.capability |=
505 					HSA_CAP_WATCH_POINTS_SUPPORTED;
506 
507 			dev->node_props.capability |=
508 				((log_max_watch_addr <<
509 					HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) &
510 				HSA_CAP_WATCH_POINTS_TOTALBITS_MASK);
511 		}
512 
513 		if (dev->gpu->device_info->asic_family == CHIP_TONGA)
514 			dev->node_props.capability |=
515 					HSA_CAP_AQL_QUEUE_DOUBLE_MAP;
516 
517 		sysfs_show_32bit_prop(buffer, "max_engine_clk_fcompute",
518 			dev->node_props.max_engine_clk_fcompute);
519 
520 		sysfs_show_64bit_prop(buffer, "local_mem_size",
521 				(unsigned long long int) 0);
522 
523 		sysfs_show_32bit_prop(buffer, "fw_version",
524 				dev->gpu->mec_fw_version);
525 		sysfs_show_32bit_prop(buffer, "capability",
526 				dev->node_props.capability);
527 		sysfs_show_32bit_prop(buffer, "sdma_fw_version",
528 				dev->gpu->sdma_fw_version);
529 	}
530 
531 	return sysfs_show_32bit_prop(buffer, "max_engine_clk_ccompute",
532 					cpufreq_quick_get_max(0)/1000);
533 }
534 
535 static const struct sysfs_ops node_ops = {
536 	.show = node_show,
537 };
538 
539 static struct kobj_type node_type = {
540 	.release = kfd_topology_kobj_release,
541 	.sysfs_ops = &node_ops,
542 };
543 
kfd_remove_sysfs_file(struct kobject * kobj,struct attribute * attr)544 static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr)
545 {
546 	sysfs_remove_file(kobj, attr);
547 	kobject_del(kobj);
548 	kobject_put(kobj);
549 }
550 
kfd_remove_sysfs_node_entry(struct kfd_topology_device * dev)551 static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev)
552 {
553 	struct kfd_iolink_properties *iolink;
554 	struct kfd_cache_properties *cache;
555 	struct kfd_mem_properties *mem;
556 	struct kfd_perf_properties *perf;
557 
558 	if (dev->kobj_iolink) {
559 		list_for_each_entry(iolink, &dev->io_link_props, list)
560 			if (iolink->kobj) {
561 				kfd_remove_sysfs_file(iolink->kobj,
562 							&iolink->attr);
563 				iolink->kobj = NULL;
564 			}
565 		kobject_del(dev->kobj_iolink);
566 		kobject_put(dev->kobj_iolink);
567 		dev->kobj_iolink = NULL;
568 	}
569 
570 	if (dev->kobj_cache) {
571 		list_for_each_entry(cache, &dev->cache_props, list)
572 			if (cache->kobj) {
573 				kfd_remove_sysfs_file(cache->kobj,
574 							&cache->attr);
575 				cache->kobj = NULL;
576 			}
577 		kobject_del(dev->kobj_cache);
578 		kobject_put(dev->kobj_cache);
579 		dev->kobj_cache = NULL;
580 	}
581 
582 	if (dev->kobj_mem) {
583 		list_for_each_entry(mem, &dev->mem_props, list)
584 			if (mem->kobj) {
585 				kfd_remove_sysfs_file(mem->kobj, &mem->attr);
586 				mem->kobj = NULL;
587 			}
588 		kobject_del(dev->kobj_mem);
589 		kobject_put(dev->kobj_mem);
590 		dev->kobj_mem = NULL;
591 	}
592 
593 	if (dev->kobj_perf) {
594 		list_for_each_entry(perf, &dev->perf_props, list) {
595 			kfree(perf->attr_group);
596 			perf->attr_group = NULL;
597 		}
598 		kobject_del(dev->kobj_perf);
599 		kobject_put(dev->kobj_perf);
600 		dev->kobj_perf = NULL;
601 	}
602 
603 	if (dev->kobj_node) {
604 		sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid);
605 		sysfs_remove_file(dev->kobj_node, &dev->attr_name);
606 		sysfs_remove_file(dev->kobj_node, &dev->attr_props);
607 		kobject_del(dev->kobj_node);
608 		kobject_put(dev->kobj_node);
609 		dev->kobj_node = NULL;
610 	}
611 }
612 
kfd_build_sysfs_node_entry(struct kfd_topology_device * dev,uint32_t id)613 static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev,
614 		uint32_t id)
615 {
616 	struct kfd_iolink_properties *iolink;
617 	struct kfd_cache_properties *cache;
618 	struct kfd_mem_properties *mem;
619 	struct kfd_perf_properties *perf;
620 	int ret;
621 	uint32_t i, num_attrs;
622 	struct attribute **attrs;
623 
624 	if (WARN_ON(dev->kobj_node))
625 		return -EEXIST;
626 
627 	/*
628 	 * Creating the sysfs folders
629 	 */
630 	dev->kobj_node = kfd_alloc_struct(dev->kobj_node);
631 	if (!dev->kobj_node)
632 		return -ENOMEM;
633 
634 	ret = kobject_init_and_add(dev->kobj_node, &node_type,
635 			sys_props.kobj_nodes, "%d", id);
636 	if (ret < 0)
637 		return ret;
638 
639 	dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node);
640 	if (!dev->kobj_mem)
641 		return -ENOMEM;
642 
643 	dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node);
644 	if (!dev->kobj_cache)
645 		return -ENOMEM;
646 
647 	dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node);
648 	if (!dev->kobj_iolink)
649 		return -ENOMEM;
650 
651 	dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node);
652 	if (!dev->kobj_perf)
653 		return -ENOMEM;
654 
655 	/*
656 	 * Creating sysfs files for node properties
657 	 */
658 	dev->attr_gpuid.name = "gpu_id";
659 	dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE;
660 	sysfs_attr_init(&dev->attr_gpuid);
661 	dev->attr_name.name = "name";
662 	dev->attr_name.mode = KFD_SYSFS_FILE_MODE;
663 	sysfs_attr_init(&dev->attr_name);
664 	dev->attr_props.name = "properties";
665 	dev->attr_props.mode = KFD_SYSFS_FILE_MODE;
666 	sysfs_attr_init(&dev->attr_props);
667 	ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid);
668 	if (ret < 0)
669 		return ret;
670 	ret = sysfs_create_file(dev->kobj_node, &dev->attr_name);
671 	if (ret < 0)
672 		return ret;
673 	ret = sysfs_create_file(dev->kobj_node, &dev->attr_props);
674 	if (ret < 0)
675 		return ret;
676 
677 	i = 0;
678 	list_for_each_entry(mem, &dev->mem_props, list) {
679 		mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
680 		if (!mem->kobj)
681 			return -ENOMEM;
682 		ret = kobject_init_and_add(mem->kobj, &mem_type,
683 				dev->kobj_mem, "%d", i);
684 		if (ret < 0)
685 			return ret;
686 
687 		mem->attr.name = "properties";
688 		mem->attr.mode = KFD_SYSFS_FILE_MODE;
689 		sysfs_attr_init(&mem->attr);
690 		ret = sysfs_create_file(mem->kobj, &mem->attr);
691 		if (ret < 0)
692 			return ret;
693 		i++;
694 	}
695 
696 	i = 0;
697 	list_for_each_entry(cache, &dev->cache_props, list) {
698 		cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
699 		if (!cache->kobj)
700 			return -ENOMEM;
701 		ret = kobject_init_and_add(cache->kobj, &cache_type,
702 				dev->kobj_cache, "%d", i);
703 		if (ret < 0)
704 			return ret;
705 
706 		cache->attr.name = "properties";
707 		cache->attr.mode = KFD_SYSFS_FILE_MODE;
708 		sysfs_attr_init(&cache->attr);
709 		ret = sysfs_create_file(cache->kobj, &cache->attr);
710 		if (ret < 0)
711 			return ret;
712 		i++;
713 	}
714 
715 	i = 0;
716 	list_for_each_entry(iolink, &dev->io_link_props, list) {
717 		iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
718 		if (!iolink->kobj)
719 			return -ENOMEM;
720 		ret = kobject_init_and_add(iolink->kobj, &iolink_type,
721 				dev->kobj_iolink, "%d", i);
722 		if (ret < 0)
723 			return ret;
724 
725 		iolink->attr.name = "properties";
726 		iolink->attr.mode = KFD_SYSFS_FILE_MODE;
727 		sysfs_attr_init(&iolink->attr);
728 		ret = sysfs_create_file(iolink->kobj, &iolink->attr);
729 		if (ret < 0)
730 			return ret;
731 		i++;
732 	}
733 
734 	/* All hardware blocks have the same number of attributes. */
735 	num_attrs = ARRAY_SIZE(perf_attr_iommu);
736 	list_for_each_entry(perf, &dev->perf_props, list) {
737 		perf->attr_group = kzalloc(sizeof(struct kfd_perf_attr)
738 			* num_attrs + sizeof(struct attribute_group),
739 			GFP_KERNEL);
740 		if (!perf->attr_group)
741 			return -ENOMEM;
742 
743 		attrs = (struct attribute **)(perf->attr_group + 1);
744 		if (!strcmp(perf->block_name, "iommu")) {
745 		/* Information of IOMMU's num_counters and counter_ids is shown
746 		 * under /sys/bus/event_source/devices/amd_iommu. We don't
747 		 * duplicate here.
748 		 */
749 			perf_attr_iommu[0].data = perf->max_concurrent;
750 			for (i = 0; i < num_attrs; i++)
751 				attrs[i] = &perf_attr_iommu[i].attr.attr;
752 		}
753 		perf->attr_group->name = perf->block_name;
754 		perf->attr_group->attrs = attrs;
755 		ret = sysfs_create_group(dev->kobj_perf, perf->attr_group);
756 		if (ret < 0)
757 			return ret;
758 	}
759 
760 	return 0;
761 }
762 
763 /* Called with write topology lock acquired */
kfd_build_sysfs_node_tree(void)764 static int kfd_build_sysfs_node_tree(void)
765 {
766 	struct kfd_topology_device *dev;
767 	int ret;
768 	uint32_t i = 0;
769 
770 	list_for_each_entry(dev, &topology_device_list, list) {
771 		ret = kfd_build_sysfs_node_entry(dev, i);
772 		if (ret < 0)
773 			return ret;
774 		i++;
775 	}
776 
777 	return 0;
778 }
779 
780 /* Called with write topology lock acquired */
kfd_remove_sysfs_node_tree(void)781 static void kfd_remove_sysfs_node_tree(void)
782 {
783 	struct kfd_topology_device *dev;
784 
785 	list_for_each_entry(dev, &topology_device_list, list)
786 		kfd_remove_sysfs_node_entry(dev);
787 }
788 
kfd_topology_update_sysfs(void)789 static int kfd_topology_update_sysfs(void)
790 {
791 	int ret;
792 
793 	pr_info("Creating topology SYSFS entries\n");
794 	if (!sys_props.kobj_topology) {
795 		sys_props.kobj_topology =
796 				kfd_alloc_struct(sys_props.kobj_topology);
797 		if (!sys_props.kobj_topology)
798 			return -ENOMEM;
799 
800 		ret = kobject_init_and_add(sys_props.kobj_topology,
801 				&sysprops_type,  &kfd_device->kobj,
802 				"topology");
803 		if (ret < 0)
804 			return ret;
805 
806 		sys_props.kobj_nodes = kobject_create_and_add("nodes",
807 				sys_props.kobj_topology);
808 		if (!sys_props.kobj_nodes)
809 			return -ENOMEM;
810 
811 		sys_props.attr_genid.name = "generation_id";
812 		sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE;
813 		sysfs_attr_init(&sys_props.attr_genid);
814 		ret = sysfs_create_file(sys_props.kobj_topology,
815 				&sys_props.attr_genid);
816 		if (ret < 0)
817 			return ret;
818 
819 		sys_props.attr_props.name = "system_properties";
820 		sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE;
821 		sysfs_attr_init(&sys_props.attr_props);
822 		ret = sysfs_create_file(sys_props.kobj_topology,
823 				&sys_props.attr_props);
824 		if (ret < 0)
825 			return ret;
826 	}
827 
828 	kfd_remove_sysfs_node_tree();
829 
830 	return kfd_build_sysfs_node_tree();
831 }
832 
kfd_topology_release_sysfs(void)833 static void kfd_topology_release_sysfs(void)
834 {
835 	kfd_remove_sysfs_node_tree();
836 	if (sys_props.kobj_topology) {
837 		sysfs_remove_file(sys_props.kobj_topology,
838 				&sys_props.attr_genid);
839 		sysfs_remove_file(sys_props.kobj_topology,
840 				&sys_props.attr_props);
841 		if (sys_props.kobj_nodes) {
842 			kobject_del(sys_props.kobj_nodes);
843 			kobject_put(sys_props.kobj_nodes);
844 			sys_props.kobj_nodes = NULL;
845 		}
846 		kobject_del(sys_props.kobj_topology);
847 		kobject_put(sys_props.kobj_topology);
848 		sys_props.kobj_topology = NULL;
849 	}
850 }
851 
852 /* Called with write topology_lock acquired */
kfd_topology_update_device_list(struct list_head * temp_list,struct list_head * master_list)853 static void kfd_topology_update_device_list(struct list_head *temp_list,
854 					struct list_head *master_list)
855 {
856 	while (!list_empty(temp_list)) {
857 		list_move_tail(temp_list->next, master_list);
858 		sys_props.num_devices++;
859 	}
860 }
861 
kfd_debug_print_topology(void)862 static void kfd_debug_print_topology(void)
863 {
864 	struct kfd_topology_device *dev;
865 
866 	down_read(&topology_lock);
867 
868 	dev = list_last_entry(&topology_device_list,
869 			struct kfd_topology_device, list);
870 	if (dev) {
871 		if (dev->node_props.cpu_cores_count &&
872 				dev->node_props.simd_count) {
873 			pr_info("Topology: Add APU node [0x%0x:0x%0x]\n",
874 				dev->node_props.device_id,
875 				dev->node_props.vendor_id);
876 		} else if (dev->node_props.cpu_cores_count)
877 			pr_info("Topology: Add CPU node\n");
878 		else if (dev->node_props.simd_count)
879 			pr_info("Topology: Add dGPU node [0x%0x:0x%0x]\n",
880 				dev->node_props.device_id,
881 				dev->node_props.vendor_id);
882 	}
883 	up_read(&topology_lock);
884 }
885 
886 /* Helper function for intializing platform_xx members of
887  * kfd_system_properties. Uses OEM info from the last CPU/APU node.
888  */
kfd_update_system_properties(void)889 static void kfd_update_system_properties(void)
890 {
891 	struct kfd_topology_device *dev;
892 
893 	down_read(&topology_lock);
894 	dev = list_last_entry(&topology_device_list,
895 			struct kfd_topology_device, list);
896 	if (dev) {
897 		sys_props.platform_id =
898 			(*((uint64_t *)dev->oem_id)) & CRAT_OEMID_64BIT_MASK;
899 		sys_props.platform_oem = *((uint64_t *)dev->oem_table_id);
900 		sys_props.platform_rev = dev->oem_revision;
901 	}
902 	up_read(&topology_lock);
903 }
904 
find_system_memory(const struct dmi_header * dm,void * private)905 static void find_system_memory(const struct dmi_header *dm,
906 	void *private)
907 {
908 	struct kfd_mem_properties *mem;
909 	u16 mem_width, mem_clock;
910 	struct kfd_topology_device *kdev =
911 		(struct kfd_topology_device *)private;
912 	const u8 *dmi_data = (const u8 *)(dm + 1);
913 
914 	if (dm->type == DMI_ENTRY_MEM_DEVICE && dm->length >= 0x15) {
915 		mem_width = (u16)(*(const u16 *)(dmi_data + 0x6));
916 		mem_clock = (u16)(*(const u16 *)(dmi_data + 0x11));
917 		list_for_each_entry(mem, &kdev->mem_props, list) {
918 			if (mem_width != 0xFFFF && mem_width != 0)
919 				mem->width = mem_width;
920 			if (mem_clock != 0)
921 				mem->mem_clk_max = mem_clock;
922 		}
923 	}
924 }
925 
926 /*
927  * Performance counters information is not part of CRAT but we would like to
928  * put them in the sysfs under topology directory for Thunk to get the data.
929  * This function is called before updating the sysfs.
930  */
kfd_add_perf_to_topology(struct kfd_topology_device * kdev)931 static int kfd_add_perf_to_topology(struct kfd_topology_device *kdev)
932 {
933 	/* These are the only counters supported so far */
934 	return kfd_iommu_add_perf_counters(kdev);
935 }
936 
937 /* kfd_add_non_crat_information - Add information that is not currently
938  *	defined in CRAT but is necessary for KFD topology
939  * @dev - topology device to which addition info is added
940  */
kfd_add_non_crat_information(struct kfd_topology_device * kdev)941 static void kfd_add_non_crat_information(struct kfd_topology_device *kdev)
942 {
943 	/* Check if CPU only node. */
944 	if (!kdev->gpu) {
945 		/* Add system memory information */
946 		dmi_walk(find_system_memory, kdev);
947 	}
948 	/* TODO: For GPU node, rearrange code from kfd_topology_add_device */
949 }
950 
951 /* kfd_is_acpi_crat_invalid - CRAT from ACPI is valid only for AMD APU devices.
952  *	Ignore CRAT for all other devices. AMD APU is identified if both CPU
953  *	and GPU cores are present.
954  * @device_list - topology device list created by parsing ACPI CRAT table.
955  * @return - TRUE if invalid, FALSE is valid.
956  */
kfd_is_acpi_crat_invalid(struct list_head * device_list)957 static bool kfd_is_acpi_crat_invalid(struct list_head *device_list)
958 {
959 	struct kfd_topology_device *dev;
960 
961 	list_for_each_entry(dev, device_list, list) {
962 		if (dev->node_props.cpu_cores_count &&
963 			dev->node_props.simd_count)
964 			return false;
965 	}
966 	pr_info("Ignoring ACPI CRAT on non-APU system\n");
967 	return true;
968 }
969 
kfd_topology_init(void)970 int kfd_topology_init(void)
971 {
972 	void *crat_image = NULL;
973 	size_t image_size = 0;
974 	int ret;
975 	struct list_head temp_topology_device_list;
976 	int cpu_only_node = 0;
977 	struct kfd_topology_device *kdev;
978 	int proximity_domain;
979 
980 	/* topology_device_list - Master list of all topology devices
981 	 * temp_topology_device_list - temporary list created while parsing CRAT
982 	 * or VCRAT. Once parsing is complete the contents of list is moved to
983 	 * topology_device_list
984 	 */
985 
986 	/* Initialize the head for the both the lists */
987 	INIT_LIST_HEAD(&topology_device_list);
988 	INIT_LIST_HEAD(&temp_topology_device_list);
989 	init_rwsem(&topology_lock);
990 
991 	memset(&sys_props, 0, sizeof(sys_props));
992 
993 	/* Proximity domains in ACPI CRAT tables start counting at
994 	 * 0. The same should be true for virtual CRAT tables created
995 	 * at this stage. GPUs added later in kfd_topology_add_device
996 	 * use a counter.
997 	 */
998 	proximity_domain = 0;
999 
1000 	/*
1001 	 * Get the CRAT image from the ACPI. If ACPI doesn't have one
1002 	 * or if ACPI CRAT is invalid create a virtual CRAT.
1003 	 * NOTE: The current implementation expects all AMD APUs to have
1004 	 *	CRAT. If no CRAT is available, it is assumed to be a CPU
1005 	 */
1006 	ret = kfd_create_crat_image_acpi(&crat_image, &image_size);
1007 	if (!ret) {
1008 		ret = kfd_parse_crat_table(crat_image,
1009 					   &temp_topology_device_list,
1010 					   proximity_domain);
1011 		if (ret ||
1012 		    kfd_is_acpi_crat_invalid(&temp_topology_device_list)) {
1013 			kfd_release_topology_device_list(
1014 				&temp_topology_device_list);
1015 			kfd_destroy_crat_image(crat_image);
1016 			crat_image = NULL;
1017 		}
1018 	}
1019 
1020 	if (!crat_image) {
1021 		ret = kfd_create_crat_image_virtual(&crat_image, &image_size,
1022 						    COMPUTE_UNIT_CPU, NULL,
1023 						    proximity_domain);
1024 		cpu_only_node = 1;
1025 		if (ret) {
1026 			pr_err("Error creating VCRAT table for CPU\n");
1027 			return ret;
1028 		}
1029 
1030 		ret = kfd_parse_crat_table(crat_image,
1031 					   &temp_topology_device_list,
1032 					   proximity_domain);
1033 		if (ret) {
1034 			pr_err("Error parsing VCRAT table for CPU\n");
1035 			goto err;
1036 		}
1037 	}
1038 
1039 	kdev = list_first_entry(&temp_topology_device_list,
1040 				struct kfd_topology_device, list);
1041 	kfd_add_perf_to_topology(kdev);
1042 
1043 	down_write(&topology_lock);
1044 	kfd_topology_update_device_list(&temp_topology_device_list,
1045 					&topology_device_list);
1046 	atomic_set(&topology_crat_proximity_domain, sys_props.num_devices-1);
1047 	ret = kfd_topology_update_sysfs();
1048 	up_write(&topology_lock);
1049 
1050 	if (!ret) {
1051 		sys_props.generation_count++;
1052 		kfd_update_system_properties();
1053 		kfd_debug_print_topology();
1054 		pr_info("Finished initializing topology\n");
1055 	} else
1056 		pr_err("Failed to update topology in sysfs ret=%d\n", ret);
1057 
1058 	/* For nodes with GPU, this information gets added
1059 	 * when GPU is detected (kfd_topology_add_device).
1060 	 */
1061 	if (cpu_only_node) {
1062 		/* Add additional information to CPU only node created above */
1063 		down_write(&topology_lock);
1064 		kdev = list_first_entry(&topology_device_list,
1065 				struct kfd_topology_device, list);
1066 		up_write(&topology_lock);
1067 		kfd_add_non_crat_information(kdev);
1068 	}
1069 
1070 err:
1071 	kfd_destroy_crat_image(crat_image);
1072 	return ret;
1073 }
1074 
kfd_topology_shutdown(void)1075 void kfd_topology_shutdown(void)
1076 {
1077 	down_write(&topology_lock);
1078 	kfd_topology_release_sysfs();
1079 	kfd_release_live_view();
1080 	up_write(&topology_lock);
1081 }
1082 
kfd_generate_gpu_id(struct kfd_dev * gpu)1083 static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu)
1084 {
1085 	uint32_t hashout;
1086 	uint32_t buf[7];
1087 	uint64_t local_mem_size;
1088 	int i;
1089 	struct kfd_local_mem_info local_mem_info;
1090 
1091 	if (!gpu)
1092 		return 0;
1093 
1094 	amdgpu_amdkfd_get_local_mem_info(gpu->kgd, &local_mem_info);
1095 
1096 	local_mem_size = local_mem_info.local_mem_size_private +
1097 			local_mem_info.local_mem_size_public;
1098 
1099 	buf[0] = gpu->pdev->devfn;
1100 	buf[1] = gpu->pdev->subsystem_vendor |
1101 		(gpu->pdev->subsystem_device << 16);
1102 	buf[2] = pci_domain_nr(gpu->pdev->bus);
1103 	buf[3] = gpu->pdev->device;
1104 	buf[4] = gpu->pdev->bus->number;
1105 	buf[5] = lower_32_bits(local_mem_size);
1106 	buf[6] = upper_32_bits(local_mem_size);
1107 
1108 	for (i = 0, hashout = 0; i < 7; i++)
1109 		hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH);
1110 
1111 	return hashout;
1112 }
1113 /* kfd_assign_gpu - Attach @gpu to the correct kfd topology device. If
1114  *		the GPU device is not already present in the topology device
1115  *		list then return NULL. This means a new topology device has to
1116  *		be created for this GPU.
1117  */
kfd_assign_gpu(struct kfd_dev * gpu)1118 static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu)
1119 {
1120 	struct kfd_topology_device *dev;
1121 	struct kfd_topology_device *out_dev = NULL;
1122 	struct kfd_mem_properties *mem;
1123 	struct kfd_cache_properties *cache;
1124 	struct kfd_iolink_properties *iolink;
1125 
1126 	down_write(&topology_lock);
1127 	list_for_each_entry(dev, &topology_device_list, list) {
1128 		/* Discrete GPUs need their own topology device list
1129 		 * entries. Don't assign them to CPU/APU nodes.
1130 		 */
1131 		if (!gpu->device_info->needs_iommu_device &&
1132 		    dev->node_props.cpu_cores_count)
1133 			continue;
1134 
1135 		if (!dev->gpu && (dev->node_props.simd_count > 0)) {
1136 			dev->gpu = gpu;
1137 			out_dev = dev;
1138 
1139 			list_for_each_entry(mem, &dev->mem_props, list)
1140 				mem->gpu = dev->gpu;
1141 			list_for_each_entry(cache, &dev->cache_props, list)
1142 				cache->gpu = dev->gpu;
1143 			list_for_each_entry(iolink, &dev->io_link_props, list)
1144 				iolink->gpu = dev->gpu;
1145 			break;
1146 		}
1147 	}
1148 	up_write(&topology_lock);
1149 	return out_dev;
1150 }
1151 
kfd_notify_gpu_change(uint32_t gpu_id,int arrival)1152 static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival)
1153 {
1154 	/*
1155 	 * TODO: Generate an event for thunk about the arrival/removal
1156 	 * of the GPU
1157 	 */
1158 }
1159 
1160 /* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info,
1161  *		patch this after CRAT parsing.
1162  */
kfd_fill_mem_clk_max_info(struct kfd_topology_device * dev)1163 static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev)
1164 {
1165 	struct kfd_mem_properties *mem;
1166 	struct kfd_local_mem_info local_mem_info;
1167 
1168 	if (!dev)
1169 		return;
1170 
1171 	/* Currently, amdgpu driver (amdgpu_mc) deals only with GPUs with
1172 	 * single bank of VRAM local memory.
1173 	 * for dGPUs - VCRAT reports only one bank of Local Memory
1174 	 * for APUs - If CRAT from ACPI reports more than one bank, then
1175 	 *	all the banks will report the same mem_clk_max information
1176 	 */
1177 	amdgpu_amdkfd_get_local_mem_info(dev->gpu->kgd, &local_mem_info);
1178 
1179 	list_for_each_entry(mem, &dev->mem_props, list)
1180 		mem->mem_clk_max = local_mem_info.mem_clk_max;
1181 }
1182 
kfd_fill_iolink_non_crat_info(struct kfd_topology_device * dev)1183 static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev)
1184 {
1185 	struct kfd_iolink_properties *link, *cpu_link;
1186 	struct kfd_topology_device *cpu_dev;
1187 	uint32_t cap;
1188 	uint32_t cpu_flag = CRAT_IOLINK_FLAGS_ENABLED;
1189 	uint32_t flag = CRAT_IOLINK_FLAGS_ENABLED;
1190 
1191 	if (!dev || !dev->gpu)
1192 		return;
1193 
1194 	pcie_capability_read_dword(dev->gpu->pdev,
1195 			PCI_EXP_DEVCAP2, &cap);
1196 
1197 	if (!(cap & (PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
1198 		     PCI_EXP_DEVCAP2_ATOMIC_COMP64)))
1199 		cpu_flag |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1200 			CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1201 
1202 	if (!dev->gpu->pci_atomic_requested ||
1203 	    dev->gpu->device_info->asic_family == CHIP_HAWAII)
1204 		flag |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1205 			CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1206 
1207 	/* GPU only creates direct links so apply flags setting to all */
1208 	list_for_each_entry(link, &dev->io_link_props, list) {
1209 		link->flags = flag;
1210 		cpu_dev = kfd_topology_device_by_proximity_domain(
1211 				link->node_to);
1212 		if (cpu_dev) {
1213 			list_for_each_entry(cpu_link,
1214 					    &cpu_dev->io_link_props, list)
1215 				if (cpu_link->node_to == link->node_from)
1216 					cpu_link->flags = cpu_flag;
1217 		}
1218 	}
1219 }
1220 
kfd_topology_add_device(struct kfd_dev * gpu)1221 int kfd_topology_add_device(struct kfd_dev *gpu)
1222 {
1223 	uint32_t gpu_id;
1224 	struct kfd_topology_device *dev;
1225 	struct kfd_cu_info cu_info;
1226 	int res = 0;
1227 	struct list_head temp_topology_device_list;
1228 	void *crat_image = NULL;
1229 	size_t image_size = 0;
1230 	int proximity_domain;
1231 	struct amdgpu_ras *ctx;
1232 
1233 	INIT_LIST_HEAD(&temp_topology_device_list);
1234 
1235 	gpu_id = kfd_generate_gpu_id(gpu);
1236 
1237 	pr_debug("Adding new GPU (ID: 0x%x) to topology\n", gpu_id);
1238 
1239 	proximity_domain = atomic_inc_return(&topology_crat_proximity_domain);
1240 
1241 	/* Check to see if this gpu device exists in the topology_device_list.
1242 	 * If so, assign the gpu to that device,
1243 	 * else create a Virtual CRAT for this gpu device and then parse that
1244 	 * CRAT to create a new topology device. Once created assign the gpu to
1245 	 * that topology device
1246 	 */
1247 	dev = kfd_assign_gpu(gpu);
1248 	if (!dev) {
1249 		res = kfd_create_crat_image_virtual(&crat_image, &image_size,
1250 						    COMPUTE_UNIT_GPU, gpu,
1251 						    proximity_domain);
1252 		if (res) {
1253 			pr_err("Error creating VCRAT for GPU (ID: 0x%x)\n",
1254 			       gpu_id);
1255 			return res;
1256 		}
1257 		res = kfd_parse_crat_table(crat_image,
1258 					   &temp_topology_device_list,
1259 					   proximity_domain);
1260 		if (res) {
1261 			pr_err("Error parsing VCRAT for GPU (ID: 0x%x)\n",
1262 			       gpu_id);
1263 			goto err;
1264 		}
1265 
1266 		down_write(&topology_lock);
1267 		kfd_topology_update_device_list(&temp_topology_device_list,
1268 			&topology_device_list);
1269 
1270 		/* Update the SYSFS tree, since we added another topology
1271 		 * device
1272 		 */
1273 		res = kfd_topology_update_sysfs();
1274 		up_write(&topology_lock);
1275 
1276 		if (!res)
1277 			sys_props.generation_count++;
1278 		else
1279 			pr_err("Failed to update GPU (ID: 0x%x) to sysfs topology. res=%d\n",
1280 						gpu_id, res);
1281 		dev = kfd_assign_gpu(gpu);
1282 		if (WARN_ON(!dev)) {
1283 			res = -ENODEV;
1284 			goto err;
1285 		}
1286 	}
1287 
1288 	dev->gpu_id = gpu_id;
1289 	gpu->id = gpu_id;
1290 
1291 	/* TODO: Move the following lines to function
1292 	 *	kfd_add_non_crat_information
1293 	 */
1294 
1295 	/* Fill-in additional information that is not available in CRAT but
1296 	 * needed for the topology
1297 	 */
1298 
1299 	amdgpu_amdkfd_get_cu_info(dev->gpu->kgd, &cu_info);
1300 
1301 	strncpy(dev->node_props.name, gpu->device_info->asic_name,
1302 			KFD_TOPOLOGY_PUBLIC_NAME_SIZE);
1303 
1304 	dev->node_props.simd_arrays_per_engine =
1305 		cu_info.num_shader_arrays_per_engine;
1306 
1307 	dev->node_props.vendor_id = gpu->pdev->vendor;
1308 	dev->node_props.device_id = gpu->pdev->device;
1309 	dev->node_props.location_id = pci_dev_id(gpu->pdev);
1310 	dev->node_props.max_engine_clk_fcompute =
1311 		amdgpu_amdkfd_get_max_engine_clock_in_mhz(dev->gpu->kgd);
1312 	dev->node_props.max_engine_clk_ccompute =
1313 		cpufreq_quick_get_max(0) / 1000;
1314 	dev->node_props.drm_render_minor =
1315 		gpu->shared_resources.drm_render_minor;
1316 
1317 	dev->node_props.hive_id = gpu->hive_id;
1318 	dev->node_props.num_sdma_engines = gpu->device_info->num_sdma_engines;
1319 	dev->node_props.num_sdma_xgmi_engines =
1320 				gpu->device_info->num_xgmi_sdma_engines;
1321 	dev->node_props.num_sdma_queues_per_engine =
1322 				gpu->device_info->num_sdma_queues_per_engine;
1323 	dev->node_props.num_gws = (hws_gws_support &&
1324 		dev->gpu->dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) ?
1325 		amdgpu_amdkfd_get_num_gws(dev->gpu->kgd) : 0;
1326 	dev->node_props.num_cp_queues = get_queues_num(dev->gpu->dqm);
1327 
1328 	kfd_fill_mem_clk_max_info(dev);
1329 	kfd_fill_iolink_non_crat_info(dev);
1330 
1331 	switch (dev->gpu->device_info->asic_family) {
1332 	case CHIP_KAVERI:
1333 	case CHIP_HAWAII:
1334 	case CHIP_TONGA:
1335 		dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_PRE_1_0 <<
1336 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1337 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1338 		break;
1339 	case CHIP_CARRIZO:
1340 	case CHIP_FIJI:
1341 	case CHIP_POLARIS10:
1342 	case CHIP_POLARIS11:
1343 	case CHIP_POLARIS12:
1344 	case CHIP_VEGAM:
1345 		pr_debug("Adding doorbell packet type capability\n");
1346 		dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 <<
1347 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1348 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1349 		break;
1350 	case CHIP_VEGA10:
1351 	case CHIP_VEGA12:
1352 	case CHIP_VEGA20:
1353 	case CHIP_RAVEN:
1354 	case CHIP_RENOIR:
1355 	case CHIP_ARCTURUS:
1356 	case CHIP_NAVI10:
1357 	case CHIP_NAVI12:
1358 	case CHIP_NAVI14:
1359 		dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 <<
1360 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1361 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1362 		break;
1363 	default:
1364 		WARN(1, "Unexpected ASIC family %u",
1365 		     dev->gpu->device_info->asic_family);
1366 	}
1367 
1368 	/*
1369 	* Overwrite ATS capability according to needs_iommu_device to fix
1370 	* potential missing corresponding bit in CRAT of BIOS.
1371 	*/
1372 	if (dev->gpu->device_info->needs_iommu_device)
1373 		dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
1374 	else
1375 		dev->node_props.capability &= ~HSA_CAP_ATS_PRESENT;
1376 
1377 	/* Fix errors in CZ CRAT.
1378 	 * simd_count: Carrizo CRAT reports wrong simd_count, probably
1379 	 *		because it doesn't consider masked out CUs
1380 	 * max_waves_per_simd: Carrizo reports wrong max_waves_per_simd
1381 	 */
1382 	if (dev->gpu->device_info->asic_family == CHIP_CARRIZO) {
1383 		dev->node_props.simd_count =
1384 			cu_info.simd_per_cu * cu_info.cu_active_number;
1385 		dev->node_props.max_waves_per_simd = 10;
1386 	}
1387 
1388 	ctx = amdgpu_ras_get_context((struct amdgpu_device *)(dev->gpu->kgd));
1389 	if (ctx) {
1390 		/* kfd only concerns sram ecc on GFX/SDMA and HBM ecc on UMC */
1391 		dev->node_props.capability |=
1392 			(((ctx->features & BIT(AMDGPU_RAS_BLOCK__SDMA)) != 0) ||
1393 			 ((ctx->features & BIT(AMDGPU_RAS_BLOCK__GFX)) != 0)) ?
1394 			HSA_CAP_SRAM_EDCSUPPORTED : 0;
1395 		dev->node_props.capability |= ((ctx->features & BIT(AMDGPU_RAS_BLOCK__UMC)) != 0) ?
1396 			HSA_CAP_MEM_EDCSUPPORTED : 0;
1397 
1398 		dev->node_props.capability |= (ctx->features != 0) ?
1399 			HSA_CAP_RASEVENTNOTIFY : 0;
1400 	}
1401 
1402 	kfd_debug_print_topology();
1403 
1404 	if (!res)
1405 		kfd_notify_gpu_change(gpu_id, 1);
1406 err:
1407 	kfd_destroy_crat_image(crat_image);
1408 	return res;
1409 }
1410 
kfd_topology_remove_device(struct kfd_dev * gpu)1411 int kfd_topology_remove_device(struct kfd_dev *gpu)
1412 {
1413 	struct kfd_topology_device *dev, *tmp;
1414 	uint32_t gpu_id;
1415 	int res = -ENODEV;
1416 
1417 	down_write(&topology_lock);
1418 
1419 	list_for_each_entry_safe(dev, tmp, &topology_device_list, list)
1420 		if (dev->gpu == gpu) {
1421 			gpu_id = dev->gpu_id;
1422 			kfd_remove_sysfs_node_entry(dev);
1423 			kfd_release_topology_device(dev);
1424 			sys_props.num_devices--;
1425 			res = 0;
1426 			if (kfd_topology_update_sysfs() < 0)
1427 				kfd_topology_release_sysfs();
1428 			break;
1429 		}
1430 
1431 	up_write(&topology_lock);
1432 
1433 	if (!res)
1434 		kfd_notify_gpu_change(gpu_id, 0);
1435 
1436 	return res;
1437 }
1438 
1439 /* kfd_topology_enum_kfd_devices - Enumerate through all devices in KFD
1440  *	topology. If GPU device is found @idx, then valid kfd_dev pointer is
1441  *	returned through @kdev
1442  * Return -	0: On success (@kdev will be NULL for non GPU nodes)
1443  *		-1: If end of list
1444  */
kfd_topology_enum_kfd_devices(uint8_t idx,struct kfd_dev ** kdev)1445 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev)
1446 {
1447 
1448 	struct kfd_topology_device *top_dev;
1449 	uint8_t device_idx = 0;
1450 
1451 	*kdev = NULL;
1452 	down_read(&topology_lock);
1453 
1454 	list_for_each_entry(top_dev, &topology_device_list, list) {
1455 		if (device_idx == idx) {
1456 			*kdev = top_dev->gpu;
1457 			up_read(&topology_lock);
1458 			return 0;
1459 		}
1460 
1461 		device_idx++;
1462 	}
1463 
1464 	up_read(&topology_lock);
1465 
1466 	return -1;
1467 
1468 }
1469 
kfd_cpumask_to_apic_id(const struct cpumask * cpumask)1470 static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask)
1471 {
1472 	int first_cpu_of_numa_node;
1473 
1474 	if (!cpumask || cpumask == cpu_none_mask)
1475 		return -1;
1476 	first_cpu_of_numa_node = cpumask_first(cpumask);
1477 	if (first_cpu_of_numa_node >= nr_cpu_ids)
1478 		return -1;
1479 #ifdef CONFIG_X86_64
1480 	return cpu_data(first_cpu_of_numa_node).apicid;
1481 #else
1482 	return first_cpu_of_numa_node;
1483 #endif
1484 }
1485 
1486 /* kfd_numa_node_to_apic_id - Returns the APIC ID of the first logical processor
1487  *	of the given NUMA node (numa_node_id)
1488  * Return -1 on failure
1489  */
kfd_numa_node_to_apic_id(int numa_node_id)1490 int kfd_numa_node_to_apic_id(int numa_node_id)
1491 {
1492 	if (numa_node_id == -1) {
1493 		pr_warn("Invalid NUMA Node. Use online CPU mask\n");
1494 		return kfd_cpumask_to_apic_id(cpu_online_mask);
1495 	}
1496 	return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id));
1497 }
1498 
1499 #if defined(CONFIG_DEBUG_FS)
1500 
kfd_debugfs_hqds_by_device(struct seq_file * m,void * data)1501 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data)
1502 {
1503 	struct kfd_topology_device *dev;
1504 	unsigned int i = 0;
1505 	int r = 0;
1506 
1507 	down_read(&topology_lock);
1508 
1509 	list_for_each_entry(dev, &topology_device_list, list) {
1510 		if (!dev->gpu) {
1511 			i++;
1512 			continue;
1513 		}
1514 
1515 		seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1516 		r = dqm_debugfs_hqds(m, dev->gpu->dqm);
1517 		if (r)
1518 			break;
1519 	}
1520 
1521 	up_read(&topology_lock);
1522 
1523 	return r;
1524 }
1525 
kfd_debugfs_rls_by_device(struct seq_file * m,void * data)1526 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data)
1527 {
1528 	struct kfd_topology_device *dev;
1529 	unsigned int i = 0;
1530 	int r = 0;
1531 
1532 	down_read(&topology_lock);
1533 
1534 	list_for_each_entry(dev, &topology_device_list, list) {
1535 		if (!dev->gpu) {
1536 			i++;
1537 			continue;
1538 		}
1539 
1540 		seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1541 		r = pm_debugfs_runlist(m, &dev->gpu->dqm->packets);
1542 		if (r)
1543 			break;
1544 	}
1545 
1546 	up_read(&topology_lock);
1547 
1548 	return r;
1549 }
1550 
1551 #endif
1552