• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 #include "amdgpu.h"
44 
45 /* topology_device_list - Master list of all topology devices */
46 static struct list_head topology_device_list;
47 static struct kfd_system_properties sys_props;
48 
49 static DECLARE_RWSEM(topology_lock);
50 static uint32_t topology_crat_proximity_domain;
51 
kfd_topology_device_by_proximity_domain_no_lock(uint32_t proximity_domain)52 struct kfd_topology_device *kfd_topology_device_by_proximity_domain_no_lock(
53 						uint32_t proximity_domain)
54 {
55 	struct kfd_topology_device *top_dev;
56 	struct kfd_topology_device *device = NULL;
57 
58 	list_for_each_entry(top_dev, &topology_device_list, list)
59 		if (top_dev->proximity_domain == proximity_domain) {
60 			device = top_dev;
61 			break;
62 		}
63 
64 	return device;
65 }
66 
kfd_topology_device_by_proximity_domain(uint32_t proximity_domain)67 struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
68 						uint32_t proximity_domain)
69 {
70 	struct kfd_topology_device *device = NULL;
71 
72 	down_read(&topology_lock);
73 
74 	device = kfd_topology_device_by_proximity_domain_no_lock(
75 							proximity_domain);
76 	up_read(&topology_lock);
77 
78 	return device;
79 }
80 
kfd_topology_device_by_id(uint32_t gpu_id)81 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id)
82 {
83 	struct kfd_topology_device *top_dev = NULL;
84 	struct kfd_topology_device *ret = NULL;
85 
86 	down_read(&topology_lock);
87 
88 	list_for_each_entry(top_dev, &topology_device_list, list)
89 		if (top_dev->gpu_id == gpu_id) {
90 			ret = top_dev;
91 			break;
92 		}
93 
94 	up_read(&topology_lock);
95 
96 	return ret;
97 }
98 
kfd_device_by_id(uint32_t gpu_id)99 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id)
100 {
101 	struct kfd_topology_device *top_dev;
102 
103 	top_dev = kfd_topology_device_by_id(gpu_id);
104 	if (!top_dev)
105 		return NULL;
106 
107 	return top_dev->gpu;
108 }
109 
kfd_device_by_pci_dev(const struct pci_dev * pdev)110 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev)
111 {
112 	struct kfd_topology_device *top_dev;
113 	struct kfd_dev *device = NULL;
114 
115 	down_read(&topology_lock);
116 
117 	list_for_each_entry(top_dev, &topology_device_list, list)
118 		if (top_dev->gpu && top_dev->gpu->pdev == pdev) {
119 			device = top_dev->gpu;
120 			break;
121 		}
122 
123 	up_read(&topology_lock);
124 
125 	return device;
126 }
127 
kfd_device_by_adev(const struct amdgpu_device * adev)128 struct kfd_dev *kfd_device_by_adev(const struct amdgpu_device *adev)
129 {
130 	struct kfd_topology_device *top_dev;
131 	struct kfd_dev *device = NULL;
132 
133 	down_read(&topology_lock);
134 
135 	list_for_each_entry(top_dev, &topology_device_list, list)
136 		if (top_dev->gpu && top_dev->gpu->adev == adev) {
137 			device = top_dev->gpu;
138 			break;
139 		}
140 
141 	up_read(&topology_lock);
142 
143 	return device;
144 }
145 
146 /* Called with write topology_lock acquired */
kfd_release_topology_device(struct kfd_topology_device * dev)147 static void kfd_release_topology_device(struct kfd_topology_device *dev)
148 {
149 	struct kfd_mem_properties *mem;
150 	struct kfd_cache_properties *cache;
151 	struct kfd_iolink_properties *iolink;
152 	struct kfd_iolink_properties *p2plink;
153 	struct kfd_perf_properties *perf;
154 
155 	list_del(&dev->list);
156 
157 	while (dev->mem_props.next != &dev->mem_props) {
158 		mem = container_of(dev->mem_props.next,
159 				struct kfd_mem_properties, list);
160 		list_del(&mem->list);
161 		kfree(mem);
162 	}
163 
164 	while (dev->cache_props.next != &dev->cache_props) {
165 		cache = container_of(dev->cache_props.next,
166 				struct kfd_cache_properties, list);
167 		list_del(&cache->list);
168 		kfree(cache);
169 	}
170 
171 	while (dev->io_link_props.next != &dev->io_link_props) {
172 		iolink = container_of(dev->io_link_props.next,
173 				struct kfd_iolink_properties, list);
174 		list_del(&iolink->list);
175 		kfree(iolink);
176 	}
177 
178 	while (dev->p2p_link_props.next != &dev->p2p_link_props) {
179 		p2plink = container_of(dev->p2p_link_props.next,
180 				struct kfd_iolink_properties, list);
181 		list_del(&p2plink->list);
182 		kfree(p2plink);
183 	}
184 
185 	while (dev->perf_props.next != &dev->perf_props) {
186 		perf = container_of(dev->perf_props.next,
187 				struct kfd_perf_properties, list);
188 		list_del(&perf->list);
189 		kfree(perf);
190 	}
191 
192 	kfree(dev);
193 }
194 
kfd_release_topology_device_list(struct list_head * device_list)195 void kfd_release_topology_device_list(struct list_head *device_list)
196 {
197 	struct kfd_topology_device *dev;
198 
199 	while (!list_empty(device_list)) {
200 		dev = list_first_entry(device_list,
201 				       struct kfd_topology_device, list);
202 		kfd_release_topology_device(dev);
203 	}
204 }
205 
kfd_release_live_view(void)206 static void kfd_release_live_view(void)
207 {
208 	kfd_release_topology_device_list(&topology_device_list);
209 	memset(&sys_props, 0, sizeof(sys_props));
210 }
211 
kfd_create_topology_device(struct list_head * device_list)212 struct kfd_topology_device *kfd_create_topology_device(
213 				struct list_head *device_list)
214 {
215 	struct kfd_topology_device *dev;
216 
217 	dev = kfd_alloc_struct(dev);
218 	if (!dev) {
219 		pr_err("No memory to allocate a topology device");
220 		return NULL;
221 	}
222 
223 	INIT_LIST_HEAD(&dev->mem_props);
224 	INIT_LIST_HEAD(&dev->cache_props);
225 	INIT_LIST_HEAD(&dev->io_link_props);
226 	INIT_LIST_HEAD(&dev->p2p_link_props);
227 	INIT_LIST_HEAD(&dev->perf_props);
228 
229 	list_add_tail(&dev->list, device_list);
230 
231 	return dev;
232 }
233 
234 
235 #define sysfs_show_gen_prop(buffer, offs, fmt, ...)		\
236 		(offs += snprintf(buffer+offs, PAGE_SIZE-offs,	\
237 				  fmt, __VA_ARGS__))
238 #define sysfs_show_32bit_prop(buffer, offs, name, value) \
239 		sysfs_show_gen_prop(buffer, offs, "%s %u\n", name, value)
240 #define sysfs_show_64bit_prop(buffer, offs, name, value) \
241 		sysfs_show_gen_prop(buffer, offs, "%s %llu\n", name, value)
242 #define sysfs_show_32bit_val(buffer, offs, value) \
243 		sysfs_show_gen_prop(buffer, offs, "%u\n", value)
244 #define sysfs_show_str_val(buffer, offs, value) \
245 		sysfs_show_gen_prop(buffer, offs, "%s\n", value)
246 
sysprops_show(struct kobject * kobj,struct attribute * attr,char * buffer)247 static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr,
248 		char *buffer)
249 {
250 	int offs = 0;
251 
252 	/* Making sure that the buffer is an empty string */
253 	buffer[0] = 0;
254 
255 	if (attr == &sys_props.attr_genid) {
256 		sysfs_show_32bit_val(buffer, offs,
257 				     sys_props.generation_count);
258 	} else if (attr == &sys_props.attr_props) {
259 		sysfs_show_64bit_prop(buffer, offs, "platform_oem",
260 				      sys_props.platform_oem);
261 		sysfs_show_64bit_prop(buffer, offs, "platform_id",
262 				      sys_props.platform_id);
263 		sysfs_show_64bit_prop(buffer, offs, "platform_rev",
264 				      sys_props.platform_rev);
265 	} else {
266 		offs = -EINVAL;
267 	}
268 
269 	return offs;
270 }
271 
kfd_topology_kobj_release(struct kobject * kobj)272 static void kfd_topology_kobj_release(struct kobject *kobj)
273 {
274 	kfree(kobj);
275 }
276 
277 static const struct sysfs_ops sysprops_ops = {
278 	.show = sysprops_show,
279 };
280 
281 static struct kobj_type sysprops_type = {
282 	.release = kfd_topology_kobj_release,
283 	.sysfs_ops = &sysprops_ops,
284 };
285 
iolink_show(struct kobject * kobj,struct attribute * attr,char * buffer)286 static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr,
287 		char *buffer)
288 {
289 	int offs = 0;
290 	struct kfd_iolink_properties *iolink;
291 
292 	/* Making sure that the buffer is an empty string */
293 	buffer[0] = 0;
294 
295 	iolink = container_of(attr, struct kfd_iolink_properties, attr);
296 	if (iolink->gpu && kfd_devcgroup_check_permission(iolink->gpu))
297 		return -EPERM;
298 	sysfs_show_32bit_prop(buffer, offs, "type", iolink->iolink_type);
299 	sysfs_show_32bit_prop(buffer, offs, "version_major", iolink->ver_maj);
300 	sysfs_show_32bit_prop(buffer, offs, "version_minor", iolink->ver_min);
301 	sysfs_show_32bit_prop(buffer, offs, "node_from", iolink->node_from);
302 	sysfs_show_32bit_prop(buffer, offs, "node_to", iolink->node_to);
303 	sysfs_show_32bit_prop(buffer, offs, "weight", iolink->weight);
304 	sysfs_show_32bit_prop(buffer, offs, "min_latency", iolink->min_latency);
305 	sysfs_show_32bit_prop(buffer, offs, "max_latency", iolink->max_latency);
306 	sysfs_show_32bit_prop(buffer, offs, "min_bandwidth",
307 			      iolink->min_bandwidth);
308 	sysfs_show_32bit_prop(buffer, offs, "max_bandwidth",
309 			      iolink->max_bandwidth);
310 	sysfs_show_32bit_prop(buffer, offs, "recommended_transfer_size",
311 			      iolink->rec_transfer_size);
312 	sysfs_show_32bit_prop(buffer, offs, "flags", iolink->flags);
313 
314 	return offs;
315 }
316 
317 static const struct sysfs_ops iolink_ops = {
318 	.show = iolink_show,
319 };
320 
321 static struct kobj_type iolink_type = {
322 	.release = kfd_topology_kobj_release,
323 	.sysfs_ops = &iolink_ops,
324 };
325 
mem_show(struct kobject * kobj,struct attribute * attr,char * buffer)326 static ssize_t mem_show(struct kobject *kobj, struct attribute *attr,
327 		char *buffer)
328 {
329 	int offs = 0;
330 	struct kfd_mem_properties *mem;
331 
332 	/* Making sure that the buffer is an empty string */
333 	buffer[0] = 0;
334 
335 	mem = container_of(attr, struct kfd_mem_properties, attr);
336 	if (mem->gpu && kfd_devcgroup_check_permission(mem->gpu))
337 		return -EPERM;
338 	sysfs_show_32bit_prop(buffer, offs, "heap_type", mem->heap_type);
339 	sysfs_show_64bit_prop(buffer, offs, "size_in_bytes",
340 			      mem->size_in_bytes);
341 	sysfs_show_32bit_prop(buffer, offs, "flags", mem->flags);
342 	sysfs_show_32bit_prop(buffer, offs, "width", mem->width);
343 	sysfs_show_32bit_prop(buffer, offs, "mem_clk_max",
344 			      mem->mem_clk_max);
345 
346 	return offs;
347 }
348 
349 static const struct sysfs_ops mem_ops = {
350 	.show = mem_show,
351 };
352 
353 static struct kobj_type mem_type = {
354 	.release = kfd_topology_kobj_release,
355 	.sysfs_ops = &mem_ops,
356 };
357 
kfd_cache_show(struct kobject * kobj,struct attribute * attr,char * buffer)358 static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr,
359 		char *buffer)
360 {
361 	int offs = 0;
362 	uint32_t i, j;
363 	struct kfd_cache_properties *cache;
364 
365 	/* Making sure that the buffer is an empty string */
366 	buffer[0] = 0;
367 	cache = container_of(attr, struct kfd_cache_properties, attr);
368 	if (cache->gpu && kfd_devcgroup_check_permission(cache->gpu))
369 		return -EPERM;
370 	sysfs_show_32bit_prop(buffer, offs, "processor_id_low",
371 			cache->processor_id_low);
372 	sysfs_show_32bit_prop(buffer, offs, "level", cache->cache_level);
373 	sysfs_show_32bit_prop(buffer, offs, "size", cache->cache_size);
374 	sysfs_show_32bit_prop(buffer, offs, "cache_line_size",
375 			      cache->cacheline_size);
376 	sysfs_show_32bit_prop(buffer, offs, "cache_lines_per_tag",
377 			      cache->cachelines_per_tag);
378 	sysfs_show_32bit_prop(buffer, offs, "association", cache->cache_assoc);
379 	sysfs_show_32bit_prop(buffer, offs, "latency", cache->cache_latency);
380 	sysfs_show_32bit_prop(buffer, offs, "type", cache->cache_type);
381 
382 	offs += snprintf(buffer+offs, PAGE_SIZE-offs, "sibling_map ");
383 	for (i = 0; i < cache->sibling_map_size; i++)
384 		for (j = 0; j < sizeof(cache->sibling_map[0])*8; j++)
385 			/* Check each bit */
386 			offs += snprintf(buffer+offs, PAGE_SIZE-offs, "%d,",
387 						(cache->sibling_map[i] >> j) & 1);
388 
389 	/* Replace the last "," with end of line */
390 	buffer[offs-1] = '\n';
391 	return offs;
392 }
393 
394 static const struct sysfs_ops cache_ops = {
395 	.show = kfd_cache_show,
396 };
397 
398 static struct kobj_type cache_type = {
399 	.release = kfd_topology_kobj_release,
400 	.sysfs_ops = &cache_ops,
401 };
402 
403 /****** Sysfs of Performance Counters ******/
404 
405 struct kfd_perf_attr {
406 	struct kobj_attribute attr;
407 	uint32_t data;
408 };
409 
perf_show(struct kobject * kobj,struct kobj_attribute * attrs,char * buf)410 static ssize_t perf_show(struct kobject *kobj, struct kobj_attribute *attrs,
411 			char *buf)
412 {
413 	int offs = 0;
414 	struct kfd_perf_attr *attr;
415 
416 	buf[0] = 0;
417 	attr = container_of(attrs, struct kfd_perf_attr, attr);
418 	if (!attr->data) /* invalid data for PMC */
419 		return 0;
420 	else
421 		return sysfs_show_32bit_val(buf, offs, attr->data);
422 }
423 
424 #define KFD_PERF_DESC(_name, _data)			\
425 {							\
426 	.attr  = __ATTR(_name, 0444, perf_show, NULL),	\
427 	.data = _data,					\
428 }
429 
430 static struct kfd_perf_attr perf_attr_iommu[] = {
431 	KFD_PERF_DESC(max_concurrent, 0),
432 	KFD_PERF_DESC(num_counters, 0),
433 	KFD_PERF_DESC(counter_ids, 0),
434 };
435 /****************************************/
436 
node_show(struct kobject * kobj,struct attribute * attr,char * buffer)437 static ssize_t node_show(struct kobject *kobj, struct attribute *attr,
438 		char *buffer)
439 {
440 	int offs = 0;
441 	struct kfd_topology_device *dev;
442 	uint32_t log_max_watch_addr;
443 
444 	/* Making sure that the buffer is an empty string */
445 	buffer[0] = 0;
446 
447 	if (strcmp(attr->name, "gpu_id") == 0) {
448 		dev = container_of(attr, struct kfd_topology_device,
449 				attr_gpuid);
450 		if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu))
451 			return -EPERM;
452 		return sysfs_show_32bit_val(buffer, offs, dev->gpu_id);
453 	}
454 
455 	if (strcmp(attr->name, "name") == 0) {
456 		dev = container_of(attr, struct kfd_topology_device,
457 				attr_name);
458 
459 		if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu))
460 			return -EPERM;
461 		return sysfs_show_str_val(buffer, offs, dev->node_props.name);
462 	}
463 
464 	dev = container_of(attr, struct kfd_topology_device,
465 			attr_props);
466 	if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu))
467 		return -EPERM;
468 	sysfs_show_32bit_prop(buffer, offs, "cpu_cores_count",
469 			      dev->node_props.cpu_cores_count);
470 	sysfs_show_32bit_prop(buffer, offs, "simd_count",
471 			      dev->gpu ? dev->node_props.simd_count : 0);
472 	sysfs_show_32bit_prop(buffer, offs, "mem_banks_count",
473 			      dev->node_props.mem_banks_count);
474 	sysfs_show_32bit_prop(buffer, offs, "caches_count",
475 			      dev->node_props.caches_count);
476 	sysfs_show_32bit_prop(buffer, offs, "io_links_count",
477 			      dev->node_props.io_links_count);
478 	sysfs_show_32bit_prop(buffer, offs, "p2p_links_count",
479 			      dev->node_props.p2p_links_count);
480 	sysfs_show_32bit_prop(buffer, offs, "cpu_core_id_base",
481 			      dev->node_props.cpu_core_id_base);
482 	sysfs_show_32bit_prop(buffer, offs, "simd_id_base",
483 			      dev->node_props.simd_id_base);
484 	sysfs_show_32bit_prop(buffer, offs, "max_waves_per_simd",
485 			      dev->node_props.max_waves_per_simd);
486 	sysfs_show_32bit_prop(buffer, offs, "lds_size_in_kb",
487 			      dev->node_props.lds_size_in_kb);
488 	sysfs_show_32bit_prop(buffer, offs, "gds_size_in_kb",
489 			      dev->node_props.gds_size_in_kb);
490 	sysfs_show_32bit_prop(buffer, offs, "num_gws",
491 			      dev->node_props.num_gws);
492 	sysfs_show_32bit_prop(buffer, offs, "wave_front_size",
493 			      dev->node_props.wave_front_size);
494 	sysfs_show_32bit_prop(buffer, offs, "array_count",
495 			      dev->node_props.array_count);
496 	sysfs_show_32bit_prop(buffer, offs, "simd_arrays_per_engine",
497 			      dev->node_props.simd_arrays_per_engine);
498 	sysfs_show_32bit_prop(buffer, offs, "cu_per_simd_array",
499 			      dev->node_props.cu_per_simd_array);
500 	sysfs_show_32bit_prop(buffer, offs, "simd_per_cu",
501 			      dev->node_props.simd_per_cu);
502 	sysfs_show_32bit_prop(buffer, offs, "max_slots_scratch_cu",
503 			      dev->node_props.max_slots_scratch_cu);
504 	sysfs_show_32bit_prop(buffer, offs, "gfx_target_version",
505 			      dev->node_props.gfx_target_version);
506 	sysfs_show_32bit_prop(buffer, offs, "vendor_id",
507 			      dev->node_props.vendor_id);
508 	sysfs_show_32bit_prop(buffer, offs, "device_id",
509 			      dev->node_props.device_id);
510 	sysfs_show_32bit_prop(buffer, offs, "location_id",
511 			      dev->node_props.location_id);
512 	sysfs_show_32bit_prop(buffer, offs, "domain",
513 			      dev->node_props.domain);
514 	sysfs_show_32bit_prop(buffer, offs, "drm_render_minor",
515 			      dev->node_props.drm_render_minor);
516 	sysfs_show_64bit_prop(buffer, offs, "hive_id",
517 			      dev->node_props.hive_id);
518 	sysfs_show_32bit_prop(buffer, offs, "num_sdma_engines",
519 			      dev->node_props.num_sdma_engines);
520 	sysfs_show_32bit_prop(buffer, offs, "num_sdma_xgmi_engines",
521 			      dev->node_props.num_sdma_xgmi_engines);
522 	sysfs_show_32bit_prop(buffer, offs, "num_sdma_queues_per_engine",
523 			      dev->node_props.num_sdma_queues_per_engine);
524 	sysfs_show_32bit_prop(buffer, offs, "num_cp_queues",
525 			      dev->node_props.num_cp_queues);
526 
527 	if (dev->gpu) {
528 		log_max_watch_addr =
529 			__ilog2_u32(dev->gpu->device_info.num_of_watch_points);
530 
531 		if (log_max_watch_addr) {
532 			dev->node_props.capability |=
533 					HSA_CAP_WATCH_POINTS_SUPPORTED;
534 
535 			dev->node_props.capability |=
536 				((log_max_watch_addr <<
537 					HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) &
538 				HSA_CAP_WATCH_POINTS_TOTALBITS_MASK);
539 		}
540 
541 		if (dev->gpu->adev->asic_type == CHIP_TONGA)
542 			dev->node_props.capability |=
543 					HSA_CAP_AQL_QUEUE_DOUBLE_MAP;
544 
545 		sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_fcompute",
546 			dev->node_props.max_engine_clk_fcompute);
547 
548 		sysfs_show_64bit_prop(buffer, offs, "local_mem_size", 0ULL);
549 
550 		sysfs_show_32bit_prop(buffer, offs, "fw_version",
551 				      dev->gpu->mec_fw_version);
552 		sysfs_show_32bit_prop(buffer, offs, "capability",
553 				      dev->node_props.capability);
554 		sysfs_show_32bit_prop(buffer, offs, "sdma_fw_version",
555 				      dev->gpu->sdma_fw_version);
556 		sysfs_show_64bit_prop(buffer, offs, "unique_id",
557 				      dev->gpu->adev->unique_id);
558 
559 	}
560 
561 	return sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_ccompute",
562 				     cpufreq_quick_get_max(0)/1000);
563 }
564 
565 static const struct sysfs_ops node_ops = {
566 	.show = node_show,
567 };
568 
569 static struct kobj_type node_type = {
570 	.release = kfd_topology_kobj_release,
571 	.sysfs_ops = &node_ops,
572 };
573 
kfd_remove_sysfs_file(struct kobject * kobj,struct attribute * attr)574 static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr)
575 {
576 	sysfs_remove_file(kobj, attr);
577 	kobject_del(kobj);
578 	kobject_put(kobj);
579 }
580 
kfd_remove_sysfs_node_entry(struct kfd_topology_device * dev)581 static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev)
582 {
583 	struct kfd_iolink_properties *p2plink;
584 	struct kfd_iolink_properties *iolink;
585 	struct kfd_cache_properties *cache;
586 	struct kfd_mem_properties *mem;
587 	struct kfd_perf_properties *perf;
588 
589 	if (dev->kobj_iolink) {
590 		list_for_each_entry(iolink, &dev->io_link_props, list)
591 			if (iolink->kobj) {
592 				kfd_remove_sysfs_file(iolink->kobj,
593 							&iolink->attr);
594 				iolink->kobj = NULL;
595 			}
596 		kobject_del(dev->kobj_iolink);
597 		kobject_put(dev->kobj_iolink);
598 		dev->kobj_iolink = NULL;
599 	}
600 
601 	if (dev->kobj_p2plink) {
602 		list_for_each_entry(p2plink, &dev->p2p_link_props, list)
603 			if (p2plink->kobj) {
604 				kfd_remove_sysfs_file(p2plink->kobj,
605 							&p2plink->attr);
606 				p2plink->kobj = NULL;
607 			}
608 		kobject_del(dev->kobj_p2plink);
609 		kobject_put(dev->kobj_p2plink);
610 		dev->kobj_p2plink = NULL;
611 	}
612 
613 	if (dev->kobj_cache) {
614 		list_for_each_entry(cache, &dev->cache_props, list)
615 			if (cache->kobj) {
616 				kfd_remove_sysfs_file(cache->kobj,
617 							&cache->attr);
618 				cache->kobj = NULL;
619 			}
620 		kobject_del(dev->kobj_cache);
621 		kobject_put(dev->kobj_cache);
622 		dev->kobj_cache = NULL;
623 	}
624 
625 	if (dev->kobj_mem) {
626 		list_for_each_entry(mem, &dev->mem_props, list)
627 			if (mem->kobj) {
628 				kfd_remove_sysfs_file(mem->kobj, &mem->attr);
629 				mem->kobj = NULL;
630 			}
631 		kobject_del(dev->kobj_mem);
632 		kobject_put(dev->kobj_mem);
633 		dev->kobj_mem = NULL;
634 	}
635 
636 	if (dev->kobj_perf) {
637 		list_for_each_entry(perf, &dev->perf_props, list) {
638 			kfree(perf->attr_group);
639 			perf->attr_group = NULL;
640 		}
641 		kobject_del(dev->kobj_perf);
642 		kobject_put(dev->kobj_perf);
643 		dev->kobj_perf = NULL;
644 	}
645 
646 	if (dev->kobj_node) {
647 		sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid);
648 		sysfs_remove_file(dev->kobj_node, &dev->attr_name);
649 		sysfs_remove_file(dev->kobj_node, &dev->attr_props);
650 		kobject_del(dev->kobj_node);
651 		kobject_put(dev->kobj_node);
652 		dev->kobj_node = NULL;
653 	}
654 }
655 
kfd_build_sysfs_node_entry(struct kfd_topology_device * dev,uint32_t id)656 static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev,
657 		uint32_t id)
658 {
659 	struct kfd_iolink_properties *p2plink;
660 	struct kfd_iolink_properties *iolink;
661 	struct kfd_cache_properties *cache;
662 	struct kfd_mem_properties *mem;
663 	struct kfd_perf_properties *perf;
664 	int ret;
665 	uint32_t i, num_attrs;
666 	struct attribute **attrs;
667 
668 	if (WARN_ON(dev->kobj_node))
669 		return -EEXIST;
670 
671 	/*
672 	 * Creating the sysfs folders
673 	 */
674 	dev->kobj_node = kfd_alloc_struct(dev->kobj_node);
675 	if (!dev->kobj_node)
676 		return -ENOMEM;
677 
678 	ret = kobject_init_and_add(dev->kobj_node, &node_type,
679 			sys_props.kobj_nodes, "%d", id);
680 	if (ret < 0) {
681 		kobject_put(dev->kobj_node);
682 		return ret;
683 	}
684 
685 	dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node);
686 	if (!dev->kobj_mem)
687 		return -ENOMEM;
688 
689 	dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node);
690 	if (!dev->kobj_cache)
691 		return -ENOMEM;
692 
693 	dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node);
694 	if (!dev->kobj_iolink)
695 		return -ENOMEM;
696 
697 	dev->kobj_p2plink = kobject_create_and_add("p2p_links", dev->kobj_node);
698 	if (!dev->kobj_p2plink)
699 		return -ENOMEM;
700 
701 	dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node);
702 	if (!dev->kobj_perf)
703 		return -ENOMEM;
704 
705 	/*
706 	 * Creating sysfs files for node properties
707 	 */
708 	dev->attr_gpuid.name = "gpu_id";
709 	dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE;
710 	sysfs_attr_init(&dev->attr_gpuid);
711 	dev->attr_name.name = "name";
712 	dev->attr_name.mode = KFD_SYSFS_FILE_MODE;
713 	sysfs_attr_init(&dev->attr_name);
714 	dev->attr_props.name = "properties";
715 	dev->attr_props.mode = KFD_SYSFS_FILE_MODE;
716 	sysfs_attr_init(&dev->attr_props);
717 	ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid);
718 	if (ret < 0)
719 		return ret;
720 	ret = sysfs_create_file(dev->kobj_node, &dev->attr_name);
721 	if (ret < 0)
722 		return ret;
723 	ret = sysfs_create_file(dev->kobj_node, &dev->attr_props);
724 	if (ret < 0)
725 		return ret;
726 
727 	i = 0;
728 	list_for_each_entry(mem, &dev->mem_props, list) {
729 		mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
730 		if (!mem->kobj)
731 			return -ENOMEM;
732 		ret = kobject_init_and_add(mem->kobj, &mem_type,
733 				dev->kobj_mem, "%d", i);
734 		if (ret < 0) {
735 			kobject_put(mem->kobj);
736 			return ret;
737 		}
738 
739 		mem->attr.name = "properties";
740 		mem->attr.mode = KFD_SYSFS_FILE_MODE;
741 		sysfs_attr_init(&mem->attr);
742 		ret = sysfs_create_file(mem->kobj, &mem->attr);
743 		if (ret < 0)
744 			return ret;
745 		i++;
746 	}
747 
748 	i = 0;
749 	list_for_each_entry(cache, &dev->cache_props, list) {
750 		cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
751 		if (!cache->kobj)
752 			return -ENOMEM;
753 		ret = kobject_init_and_add(cache->kobj, &cache_type,
754 				dev->kobj_cache, "%d", i);
755 		if (ret < 0) {
756 			kobject_put(cache->kobj);
757 			return ret;
758 		}
759 
760 		cache->attr.name = "properties";
761 		cache->attr.mode = KFD_SYSFS_FILE_MODE;
762 		sysfs_attr_init(&cache->attr);
763 		ret = sysfs_create_file(cache->kobj, &cache->attr);
764 		if (ret < 0)
765 			return ret;
766 		i++;
767 	}
768 
769 	i = 0;
770 	list_for_each_entry(iolink, &dev->io_link_props, list) {
771 		iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
772 		if (!iolink->kobj)
773 			return -ENOMEM;
774 		ret = kobject_init_and_add(iolink->kobj, &iolink_type,
775 				dev->kobj_iolink, "%d", i);
776 		if (ret < 0) {
777 			kobject_put(iolink->kobj);
778 			return ret;
779 		}
780 
781 		iolink->attr.name = "properties";
782 		iolink->attr.mode = KFD_SYSFS_FILE_MODE;
783 		sysfs_attr_init(&iolink->attr);
784 		ret = sysfs_create_file(iolink->kobj, &iolink->attr);
785 		if (ret < 0)
786 			return ret;
787 		i++;
788 	}
789 
790 	i = 0;
791 	list_for_each_entry(p2plink, &dev->p2p_link_props, list) {
792 		p2plink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
793 		if (!p2plink->kobj)
794 			return -ENOMEM;
795 		ret = kobject_init_and_add(p2plink->kobj, &iolink_type,
796 				dev->kobj_p2plink, "%d", i);
797 		if (ret < 0) {
798 			kobject_put(p2plink->kobj);
799 			return ret;
800 		}
801 
802 		p2plink->attr.name = "properties";
803 		p2plink->attr.mode = KFD_SYSFS_FILE_MODE;
804 		sysfs_attr_init(&p2plink->attr);
805 		ret = sysfs_create_file(p2plink->kobj, &p2plink->attr);
806 		if (ret < 0)
807 			return ret;
808 		i++;
809 	}
810 
811 	/* All hardware blocks have the same number of attributes. */
812 	num_attrs = ARRAY_SIZE(perf_attr_iommu);
813 	list_for_each_entry(perf, &dev->perf_props, list) {
814 		perf->attr_group = kzalloc(sizeof(struct kfd_perf_attr)
815 			* num_attrs + sizeof(struct attribute_group),
816 			GFP_KERNEL);
817 		if (!perf->attr_group)
818 			return -ENOMEM;
819 
820 		attrs = (struct attribute **)(perf->attr_group + 1);
821 		if (!strcmp(perf->block_name, "iommu")) {
822 		/* Information of IOMMU's num_counters and counter_ids is shown
823 		 * under /sys/bus/event_source/devices/amd_iommu. We don't
824 		 * duplicate here.
825 		 */
826 			perf_attr_iommu[0].data = perf->max_concurrent;
827 			for (i = 0; i < num_attrs; i++)
828 				attrs[i] = &perf_attr_iommu[i].attr.attr;
829 		}
830 		perf->attr_group->name = perf->block_name;
831 		perf->attr_group->attrs = attrs;
832 		ret = sysfs_create_group(dev->kobj_perf, perf->attr_group);
833 		if (ret < 0)
834 			return ret;
835 	}
836 
837 	return 0;
838 }
839 
840 /* Called with write topology lock acquired */
kfd_build_sysfs_node_tree(void)841 static int kfd_build_sysfs_node_tree(void)
842 {
843 	struct kfd_topology_device *dev;
844 	int ret;
845 	uint32_t i = 0;
846 
847 	list_for_each_entry(dev, &topology_device_list, list) {
848 		ret = kfd_build_sysfs_node_entry(dev, i);
849 		if (ret < 0)
850 			return ret;
851 		i++;
852 	}
853 
854 	return 0;
855 }
856 
857 /* Called with write topology lock acquired */
kfd_remove_sysfs_node_tree(void)858 static void kfd_remove_sysfs_node_tree(void)
859 {
860 	struct kfd_topology_device *dev;
861 
862 	list_for_each_entry(dev, &topology_device_list, list)
863 		kfd_remove_sysfs_node_entry(dev);
864 }
865 
kfd_topology_update_sysfs(void)866 static int kfd_topology_update_sysfs(void)
867 {
868 	int ret;
869 
870 	if (!sys_props.kobj_topology) {
871 		sys_props.kobj_topology =
872 				kfd_alloc_struct(sys_props.kobj_topology);
873 		if (!sys_props.kobj_topology)
874 			return -ENOMEM;
875 
876 		ret = kobject_init_and_add(sys_props.kobj_topology,
877 				&sysprops_type,  &kfd_device->kobj,
878 				"topology");
879 		if (ret < 0) {
880 			kobject_put(sys_props.kobj_topology);
881 			return ret;
882 		}
883 
884 		sys_props.kobj_nodes = kobject_create_and_add("nodes",
885 				sys_props.kobj_topology);
886 		if (!sys_props.kobj_nodes)
887 			return -ENOMEM;
888 
889 		sys_props.attr_genid.name = "generation_id";
890 		sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE;
891 		sysfs_attr_init(&sys_props.attr_genid);
892 		ret = sysfs_create_file(sys_props.kobj_topology,
893 				&sys_props.attr_genid);
894 		if (ret < 0)
895 			return ret;
896 
897 		sys_props.attr_props.name = "system_properties";
898 		sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE;
899 		sysfs_attr_init(&sys_props.attr_props);
900 		ret = sysfs_create_file(sys_props.kobj_topology,
901 				&sys_props.attr_props);
902 		if (ret < 0)
903 			return ret;
904 	}
905 
906 	kfd_remove_sysfs_node_tree();
907 
908 	return kfd_build_sysfs_node_tree();
909 }
910 
kfd_topology_release_sysfs(void)911 static void kfd_topology_release_sysfs(void)
912 {
913 	kfd_remove_sysfs_node_tree();
914 	if (sys_props.kobj_topology) {
915 		sysfs_remove_file(sys_props.kobj_topology,
916 				&sys_props.attr_genid);
917 		sysfs_remove_file(sys_props.kobj_topology,
918 				&sys_props.attr_props);
919 		if (sys_props.kobj_nodes) {
920 			kobject_del(sys_props.kobj_nodes);
921 			kobject_put(sys_props.kobj_nodes);
922 			sys_props.kobj_nodes = NULL;
923 		}
924 		kobject_del(sys_props.kobj_topology);
925 		kobject_put(sys_props.kobj_topology);
926 		sys_props.kobj_topology = NULL;
927 	}
928 }
929 
930 /* Called with write topology_lock acquired */
kfd_topology_update_device_list(struct list_head * temp_list,struct list_head * master_list)931 static void kfd_topology_update_device_list(struct list_head *temp_list,
932 					struct list_head *master_list)
933 {
934 	while (!list_empty(temp_list)) {
935 		list_move_tail(temp_list->next, master_list);
936 		sys_props.num_devices++;
937 	}
938 }
939 
kfd_debug_print_topology(void)940 static void kfd_debug_print_topology(void)
941 {
942 	struct kfd_topology_device *dev;
943 
944 	down_read(&topology_lock);
945 
946 	dev = list_last_entry(&topology_device_list,
947 			struct kfd_topology_device, list);
948 	if (dev) {
949 		if (dev->node_props.cpu_cores_count &&
950 				dev->node_props.simd_count) {
951 			pr_info("Topology: Add APU node [0x%0x:0x%0x]\n",
952 				dev->node_props.device_id,
953 				dev->node_props.vendor_id);
954 		} else if (dev->node_props.cpu_cores_count)
955 			pr_info("Topology: Add CPU node\n");
956 		else if (dev->node_props.simd_count)
957 			pr_info("Topology: Add dGPU node [0x%0x:0x%0x]\n",
958 				dev->node_props.device_id,
959 				dev->node_props.vendor_id);
960 	}
961 	up_read(&topology_lock);
962 }
963 
964 /* Helper function for intializing platform_xx members of
965  * kfd_system_properties. Uses OEM info from the last CPU/APU node.
966  */
kfd_update_system_properties(void)967 static void kfd_update_system_properties(void)
968 {
969 	struct kfd_topology_device *dev;
970 
971 	down_read(&topology_lock);
972 	dev = list_last_entry(&topology_device_list,
973 			struct kfd_topology_device, list);
974 	if (dev) {
975 		sys_props.platform_id =
976 			(*((uint64_t *)dev->oem_id)) & CRAT_OEMID_64BIT_MASK;
977 		sys_props.platform_oem = *((uint64_t *)dev->oem_table_id);
978 		sys_props.platform_rev = dev->oem_revision;
979 	}
980 	up_read(&topology_lock);
981 }
982 
find_system_memory(const struct dmi_header * dm,void * private)983 static void find_system_memory(const struct dmi_header *dm,
984 	void *private)
985 {
986 	struct kfd_mem_properties *mem;
987 	u16 mem_width, mem_clock;
988 	struct kfd_topology_device *kdev =
989 		(struct kfd_topology_device *)private;
990 	const u8 *dmi_data = (const u8 *)(dm + 1);
991 
992 	if (dm->type == DMI_ENTRY_MEM_DEVICE && dm->length >= 0x15) {
993 		mem_width = (u16)(*(const u16 *)(dmi_data + 0x6));
994 		mem_clock = (u16)(*(const u16 *)(dmi_data + 0x11));
995 		list_for_each_entry(mem, &kdev->mem_props, list) {
996 			if (mem_width != 0xFFFF && mem_width != 0)
997 				mem->width = mem_width;
998 			if (mem_clock != 0)
999 				mem->mem_clk_max = mem_clock;
1000 		}
1001 	}
1002 }
1003 
1004 /*
1005  * Performance counters information is not part of CRAT but we would like to
1006  * put them in the sysfs under topology directory for Thunk to get the data.
1007  * This function is called before updating the sysfs.
1008  */
kfd_add_perf_to_topology(struct kfd_topology_device * kdev)1009 static int kfd_add_perf_to_topology(struct kfd_topology_device *kdev)
1010 {
1011 	/* These are the only counters supported so far */
1012 	return kfd_iommu_add_perf_counters(kdev);
1013 }
1014 
1015 /* kfd_add_non_crat_information - Add information that is not currently
1016  *	defined in CRAT but is necessary for KFD topology
1017  * @dev - topology device to which addition info is added
1018  */
kfd_add_non_crat_information(struct kfd_topology_device * kdev)1019 static void kfd_add_non_crat_information(struct kfd_topology_device *kdev)
1020 {
1021 	/* Check if CPU only node. */
1022 	if (!kdev->gpu) {
1023 		/* Add system memory information */
1024 		dmi_walk(find_system_memory, kdev);
1025 	}
1026 	/* TODO: For GPU node, rearrange code from kfd_topology_add_device */
1027 }
1028 
1029 /* kfd_is_acpi_crat_invalid - CRAT from ACPI is valid only for AMD APU devices.
1030  *	Ignore CRAT for all other devices. AMD APU is identified if both CPU
1031  *	and GPU cores are present.
1032  * @device_list - topology device list created by parsing ACPI CRAT table.
1033  * @return - TRUE if invalid, FALSE is valid.
1034  */
kfd_is_acpi_crat_invalid(struct list_head * device_list)1035 static bool kfd_is_acpi_crat_invalid(struct list_head *device_list)
1036 {
1037 	struct kfd_topology_device *dev;
1038 
1039 	list_for_each_entry(dev, device_list, list) {
1040 		if (dev->node_props.cpu_cores_count &&
1041 			dev->node_props.simd_count)
1042 			return false;
1043 	}
1044 	pr_info("Ignoring ACPI CRAT on non-APU system\n");
1045 	return true;
1046 }
1047 
kfd_topology_init(void)1048 int kfd_topology_init(void)
1049 {
1050 	void *crat_image = NULL;
1051 	size_t image_size = 0;
1052 	int ret;
1053 	struct list_head temp_topology_device_list;
1054 	int cpu_only_node = 0;
1055 	struct kfd_topology_device *kdev;
1056 	int proximity_domain;
1057 
1058 	/* topology_device_list - Master list of all topology devices
1059 	 * temp_topology_device_list - temporary list created while parsing CRAT
1060 	 * or VCRAT. Once parsing is complete the contents of list is moved to
1061 	 * topology_device_list
1062 	 */
1063 
1064 	/* Initialize the head for the both the lists */
1065 	INIT_LIST_HEAD(&topology_device_list);
1066 	INIT_LIST_HEAD(&temp_topology_device_list);
1067 	init_rwsem(&topology_lock);
1068 
1069 	memset(&sys_props, 0, sizeof(sys_props));
1070 
1071 	/* Proximity domains in ACPI CRAT tables start counting at
1072 	 * 0. The same should be true for virtual CRAT tables created
1073 	 * at this stage. GPUs added later in kfd_topology_add_device
1074 	 * use a counter.
1075 	 */
1076 	proximity_domain = 0;
1077 
1078 	/*
1079 	 * Get the CRAT image from the ACPI. If ACPI doesn't have one
1080 	 * or if ACPI CRAT is invalid create a virtual CRAT.
1081 	 * NOTE: The current implementation expects all AMD APUs to have
1082 	 *	CRAT. If no CRAT is available, it is assumed to be a CPU
1083 	 */
1084 	ret = kfd_create_crat_image_acpi(&crat_image, &image_size);
1085 	if (!ret) {
1086 		ret = kfd_parse_crat_table(crat_image,
1087 					   &temp_topology_device_list,
1088 					   proximity_domain);
1089 		if (ret ||
1090 		    kfd_is_acpi_crat_invalid(&temp_topology_device_list)) {
1091 			kfd_release_topology_device_list(
1092 				&temp_topology_device_list);
1093 			kfd_destroy_crat_image(crat_image);
1094 			crat_image = NULL;
1095 		}
1096 	}
1097 
1098 	if (!crat_image) {
1099 		ret = kfd_create_crat_image_virtual(&crat_image, &image_size,
1100 						    COMPUTE_UNIT_CPU, NULL,
1101 						    proximity_domain);
1102 		cpu_only_node = 1;
1103 		if (ret) {
1104 			pr_err("Error creating VCRAT table for CPU\n");
1105 			return ret;
1106 		}
1107 
1108 		ret = kfd_parse_crat_table(crat_image,
1109 					   &temp_topology_device_list,
1110 					   proximity_domain);
1111 		if (ret) {
1112 			pr_err("Error parsing VCRAT table for CPU\n");
1113 			goto err;
1114 		}
1115 	}
1116 
1117 	kdev = list_first_entry(&temp_topology_device_list,
1118 				struct kfd_topology_device, list);
1119 	kfd_add_perf_to_topology(kdev);
1120 
1121 	down_write(&topology_lock);
1122 	kfd_topology_update_device_list(&temp_topology_device_list,
1123 					&topology_device_list);
1124 	topology_crat_proximity_domain = sys_props.num_devices-1;
1125 	ret = kfd_topology_update_sysfs();
1126 	up_write(&topology_lock);
1127 
1128 	if (!ret) {
1129 		sys_props.generation_count++;
1130 		kfd_update_system_properties();
1131 		kfd_debug_print_topology();
1132 	} else
1133 		pr_err("Failed to update topology in sysfs ret=%d\n", ret);
1134 
1135 	/* For nodes with GPU, this information gets added
1136 	 * when GPU is detected (kfd_topology_add_device).
1137 	 */
1138 	if (cpu_only_node) {
1139 		/* Add additional information to CPU only node created above */
1140 		down_write(&topology_lock);
1141 		kdev = list_first_entry(&topology_device_list,
1142 				struct kfd_topology_device, list);
1143 		up_write(&topology_lock);
1144 		kfd_add_non_crat_information(kdev);
1145 	}
1146 
1147 err:
1148 	kfd_destroy_crat_image(crat_image);
1149 	return ret;
1150 }
1151 
kfd_topology_shutdown(void)1152 void kfd_topology_shutdown(void)
1153 {
1154 	down_write(&topology_lock);
1155 	kfd_topology_release_sysfs();
1156 	kfd_release_live_view();
1157 	up_write(&topology_lock);
1158 }
1159 
kfd_generate_gpu_id(struct kfd_dev * gpu)1160 static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu)
1161 {
1162 	uint32_t hashout;
1163 	uint32_t buf[7];
1164 	uint64_t local_mem_size;
1165 	int i;
1166 
1167 	if (!gpu)
1168 		return 0;
1169 
1170 	local_mem_size = gpu->local_mem_info.local_mem_size_private +
1171 			gpu->local_mem_info.local_mem_size_public;
1172 
1173 	buf[0] = gpu->pdev->devfn;
1174 	buf[1] = gpu->pdev->subsystem_vendor |
1175 		(gpu->pdev->subsystem_device << 16);
1176 	buf[2] = pci_domain_nr(gpu->pdev->bus);
1177 	buf[3] = gpu->pdev->device;
1178 	buf[4] = gpu->pdev->bus->number;
1179 	buf[5] = lower_32_bits(local_mem_size);
1180 	buf[6] = upper_32_bits(local_mem_size);
1181 
1182 	for (i = 0, hashout = 0; i < 7; i++)
1183 		hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH);
1184 
1185 	return hashout;
1186 }
1187 /* kfd_assign_gpu - Attach @gpu to the correct kfd topology device. If
1188  *		the GPU device is not already present in the topology device
1189  *		list then return NULL. This means a new topology device has to
1190  *		be created for this GPU.
1191  */
kfd_assign_gpu(struct kfd_dev * gpu)1192 static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu)
1193 {
1194 	struct kfd_topology_device *dev;
1195 	struct kfd_topology_device *out_dev = NULL;
1196 	struct kfd_mem_properties *mem;
1197 	struct kfd_cache_properties *cache;
1198 	struct kfd_iolink_properties *iolink;
1199 	struct kfd_iolink_properties *p2plink;
1200 
1201 	list_for_each_entry(dev, &topology_device_list, list) {
1202 		/* Discrete GPUs need their own topology device list
1203 		 * entries. Don't assign them to CPU/APU nodes.
1204 		 */
1205 		if (!gpu->use_iommu_v2 &&
1206 		    dev->node_props.cpu_cores_count)
1207 			continue;
1208 
1209 		if (!dev->gpu && (dev->node_props.simd_count > 0)) {
1210 			dev->gpu = gpu;
1211 			out_dev = dev;
1212 
1213 			list_for_each_entry(mem, &dev->mem_props, list)
1214 				mem->gpu = dev->gpu;
1215 			list_for_each_entry(cache, &dev->cache_props, list)
1216 				cache->gpu = dev->gpu;
1217 			list_for_each_entry(iolink, &dev->io_link_props, list)
1218 				iolink->gpu = dev->gpu;
1219 			list_for_each_entry(p2plink, &dev->p2p_link_props, list)
1220 				p2plink->gpu = dev->gpu;
1221 			break;
1222 		}
1223 	}
1224 	return out_dev;
1225 }
1226 
kfd_notify_gpu_change(uint32_t gpu_id,int arrival)1227 static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival)
1228 {
1229 	/*
1230 	 * TODO: Generate an event for thunk about the arrival/removal
1231 	 * of the GPU
1232 	 */
1233 }
1234 
1235 /* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info,
1236  *		patch this after CRAT parsing.
1237  */
kfd_fill_mem_clk_max_info(struct kfd_topology_device * dev)1238 static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev)
1239 {
1240 	struct kfd_mem_properties *mem;
1241 	struct kfd_local_mem_info local_mem_info;
1242 
1243 	if (!dev)
1244 		return;
1245 
1246 	/* Currently, amdgpu driver (amdgpu_mc) deals only with GPUs with
1247 	 * single bank of VRAM local memory.
1248 	 * for dGPUs - VCRAT reports only one bank of Local Memory
1249 	 * for APUs - If CRAT from ACPI reports more than one bank, then
1250 	 *	all the banks will report the same mem_clk_max information
1251 	 */
1252 	amdgpu_amdkfd_get_local_mem_info(dev->gpu->adev, &local_mem_info);
1253 
1254 	list_for_each_entry(mem, &dev->mem_props, list)
1255 		mem->mem_clk_max = local_mem_info.mem_clk_max;
1256 }
1257 
kfd_set_iolink_no_atomics(struct kfd_topology_device * dev,struct kfd_topology_device * target_gpu_dev,struct kfd_iolink_properties * link)1258 static void kfd_set_iolink_no_atomics(struct kfd_topology_device *dev,
1259 					struct kfd_topology_device *target_gpu_dev,
1260 					struct kfd_iolink_properties *link)
1261 {
1262 	/* xgmi always supports atomics between links. */
1263 	if (link->iolink_type == CRAT_IOLINK_TYPE_XGMI)
1264 		return;
1265 
1266 	/* check pcie support to set cpu(dev) flags for target_gpu_dev link. */
1267 	if (target_gpu_dev) {
1268 		uint32_t cap;
1269 
1270 		pcie_capability_read_dword(target_gpu_dev->gpu->pdev,
1271 				PCI_EXP_DEVCAP2, &cap);
1272 
1273 		if (!(cap & (PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
1274 			     PCI_EXP_DEVCAP2_ATOMIC_COMP64)))
1275 			link->flags |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1276 				CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1277 	/* set gpu (dev) flags. */
1278 	} else {
1279 		if (!dev->gpu->pci_atomic_requested ||
1280 				dev->gpu->adev->asic_type == CHIP_HAWAII)
1281 			link->flags |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1282 				CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1283 	}
1284 }
1285 
kfd_set_iolink_non_coherent(struct kfd_topology_device * to_dev,struct kfd_iolink_properties * outbound_link,struct kfd_iolink_properties * inbound_link)1286 static void kfd_set_iolink_non_coherent(struct kfd_topology_device *to_dev,
1287 		struct kfd_iolink_properties *outbound_link,
1288 		struct kfd_iolink_properties *inbound_link)
1289 {
1290 	/* CPU -> GPU with PCIe */
1291 	if (!to_dev->gpu &&
1292 	    inbound_link->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS)
1293 		inbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT;
1294 
1295 	if (to_dev->gpu) {
1296 		/* GPU <-> GPU with PCIe and
1297 		 * Vega20 with XGMI
1298 		 */
1299 		if (inbound_link->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS ||
1300 		    (inbound_link->iolink_type == CRAT_IOLINK_TYPE_XGMI &&
1301 		    KFD_GC_VERSION(to_dev->gpu) == IP_VERSION(9, 4, 0))) {
1302 			outbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT;
1303 			inbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT;
1304 		}
1305 	}
1306 }
1307 
kfd_fill_iolink_non_crat_info(struct kfd_topology_device * dev)1308 static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev)
1309 {
1310 	struct kfd_iolink_properties *link, *inbound_link;
1311 	struct kfd_topology_device *peer_dev;
1312 
1313 	if (!dev || !dev->gpu)
1314 		return;
1315 
1316 	/* GPU only creates direct links so apply flags setting to all */
1317 	list_for_each_entry(link, &dev->io_link_props, list) {
1318 		link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1319 		kfd_set_iolink_no_atomics(dev, NULL, link);
1320 		peer_dev = kfd_topology_device_by_proximity_domain(
1321 				link->node_to);
1322 
1323 		if (!peer_dev)
1324 			continue;
1325 
1326 		/* Include the CPU peer in GPU hive if connected over xGMI. */
1327 		if (!peer_dev->gpu && !peer_dev->node_props.hive_id &&
1328 				dev->node_props.hive_id &&
1329 				dev->gpu->adev->gmc.xgmi.connected_to_cpu)
1330 			peer_dev->node_props.hive_id = dev->node_props.hive_id;
1331 
1332 		list_for_each_entry(inbound_link, &peer_dev->io_link_props,
1333 									list) {
1334 			if (inbound_link->node_to != link->node_from)
1335 				continue;
1336 
1337 			inbound_link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1338 			kfd_set_iolink_no_atomics(peer_dev, dev, inbound_link);
1339 			kfd_set_iolink_non_coherent(peer_dev, link, inbound_link);
1340 		}
1341 	}
1342 
1343 	/* Create indirect links so apply flags setting to all */
1344 	list_for_each_entry(link, &dev->p2p_link_props, list) {
1345 		link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1346 		kfd_set_iolink_no_atomics(dev, NULL, link);
1347 		peer_dev = kfd_topology_device_by_proximity_domain(
1348 				link->node_to);
1349 
1350 		if (!peer_dev)
1351 			continue;
1352 
1353 		list_for_each_entry(inbound_link, &peer_dev->p2p_link_props,
1354 									list) {
1355 			if (inbound_link->node_to != link->node_from)
1356 				continue;
1357 
1358 			inbound_link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1359 			kfd_set_iolink_no_atomics(peer_dev, dev, inbound_link);
1360 			kfd_set_iolink_non_coherent(peer_dev, link, inbound_link);
1361 		}
1362 	}
1363 }
1364 
kfd_build_p2p_node_entry(struct kfd_topology_device * dev,struct kfd_iolink_properties * p2plink)1365 static int kfd_build_p2p_node_entry(struct kfd_topology_device *dev,
1366 				struct kfd_iolink_properties *p2plink)
1367 {
1368 	int ret;
1369 
1370 	p2plink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
1371 	if (!p2plink->kobj)
1372 		return -ENOMEM;
1373 
1374 	ret = kobject_init_and_add(p2plink->kobj, &iolink_type,
1375 			dev->kobj_p2plink, "%d", dev->node_props.p2p_links_count - 1);
1376 	if (ret < 0) {
1377 		kobject_put(p2plink->kobj);
1378 		return ret;
1379 	}
1380 
1381 	p2plink->attr.name = "properties";
1382 	p2plink->attr.mode = KFD_SYSFS_FILE_MODE;
1383 	sysfs_attr_init(&p2plink->attr);
1384 	ret = sysfs_create_file(p2plink->kobj, &p2plink->attr);
1385 	if (ret < 0)
1386 		return ret;
1387 
1388 	return 0;
1389 }
1390 
kfd_create_indirect_link_prop(struct kfd_topology_device * kdev,int gpu_node)1391 static int kfd_create_indirect_link_prop(struct kfd_topology_device *kdev, int gpu_node)
1392 {
1393 	struct kfd_iolink_properties *gpu_link, *tmp_link, *cpu_link;
1394 	struct kfd_iolink_properties *props = NULL, *props2 = NULL;
1395 	struct kfd_topology_device *cpu_dev;
1396 	int ret = 0;
1397 	int i, num_cpu;
1398 
1399 	num_cpu = 0;
1400 	list_for_each_entry(cpu_dev, &topology_device_list, list) {
1401 		if (cpu_dev->gpu)
1402 			break;
1403 		num_cpu++;
1404 	}
1405 
1406 	if (list_empty(&kdev->io_link_props))
1407 		return -ENODATA;
1408 
1409 	gpu_link = list_first_entry(&kdev->io_link_props,
1410 				    struct kfd_iolink_properties, list);
1411 
1412 	for (i = 0; i < num_cpu; i++) {
1413 		/* CPU <--> GPU */
1414 		if (gpu_link->node_to == i)
1415 			continue;
1416 
1417 		/* find CPU <-->  CPU links */
1418 		cpu_link = NULL;
1419 		cpu_dev = kfd_topology_device_by_proximity_domain(i);
1420 		if (cpu_dev) {
1421 			list_for_each_entry(tmp_link,
1422 					&cpu_dev->io_link_props, list) {
1423 				if (tmp_link->node_to == gpu_link->node_to) {
1424 					cpu_link = tmp_link;
1425 					break;
1426 				}
1427 			}
1428 		}
1429 
1430 		if (!cpu_link)
1431 			return -ENOMEM;
1432 
1433 		/* CPU <--> CPU <--> GPU, GPU node*/
1434 		props = kfd_alloc_struct(props);
1435 		if (!props)
1436 			return -ENOMEM;
1437 
1438 		memcpy(props, gpu_link, sizeof(struct kfd_iolink_properties));
1439 		props->weight = gpu_link->weight + cpu_link->weight;
1440 		props->min_latency = gpu_link->min_latency + cpu_link->min_latency;
1441 		props->max_latency = gpu_link->max_latency + cpu_link->max_latency;
1442 		props->min_bandwidth = min(gpu_link->min_bandwidth, cpu_link->min_bandwidth);
1443 		props->max_bandwidth = min(gpu_link->max_bandwidth, cpu_link->max_bandwidth);
1444 
1445 		props->node_from = gpu_node;
1446 		props->node_to = i;
1447 		kdev->node_props.p2p_links_count++;
1448 		list_add_tail(&props->list, &kdev->p2p_link_props);
1449 		ret = kfd_build_p2p_node_entry(kdev, props);
1450 		if (ret < 0)
1451 			return ret;
1452 
1453 		/* for small Bar, no CPU --> GPU in-direct links */
1454 		if (kfd_dev_is_large_bar(kdev->gpu)) {
1455 			/* CPU <--> CPU <--> GPU, CPU node*/
1456 			props2 = kfd_alloc_struct(props2);
1457 			if (!props2)
1458 				return -ENOMEM;
1459 
1460 			memcpy(props2, props, sizeof(struct kfd_iolink_properties));
1461 			props2->node_from = i;
1462 			props2->node_to = gpu_node;
1463 			props2->kobj = NULL;
1464 			cpu_dev->node_props.p2p_links_count++;
1465 			list_add_tail(&props2->list, &cpu_dev->p2p_link_props);
1466 			ret = kfd_build_p2p_node_entry(cpu_dev, props2);
1467 			if (ret < 0)
1468 				return ret;
1469 		}
1470 	}
1471 	return ret;
1472 }
1473 
1474 #if defined(CONFIG_HSA_AMD_P2P)
kfd_add_peer_prop(struct kfd_topology_device * kdev,struct kfd_topology_device * peer,int from,int to)1475 static int kfd_add_peer_prop(struct kfd_topology_device *kdev,
1476 		struct kfd_topology_device *peer, int from, int to)
1477 {
1478 	struct kfd_iolink_properties *props = NULL;
1479 	struct kfd_iolink_properties *iolink1, *iolink2, *iolink3;
1480 	struct kfd_topology_device *cpu_dev;
1481 	int ret = 0;
1482 
1483 	if (!amdgpu_device_is_peer_accessible(
1484 				kdev->gpu->adev,
1485 				peer->gpu->adev))
1486 		return ret;
1487 
1488 	if (list_empty(&kdev->io_link_props))
1489 		return -ENODATA;
1490 
1491 	iolink1 = list_first_entry(&kdev->io_link_props,
1492 				   struct kfd_iolink_properties, list);
1493 
1494 	if (list_empty(&peer->io_link_props))
1495 		return -ENODATA;
1496 
1497 	iolink2 = list_first_entry(&peer->io_link_props,
1498 				   struct kfd_iolink_properties, list);
1499 
1500 	props = kfd_alloc_struct(props);
1501 	if (!props)
1502 		return -ENOMEM;
1503 
1504 	memcpy(props, iolink1, sizeof(struct kfd_iolink_properties));
1505 
1506 	props->weight = iolink1->weight + iolink2->weight;
1507 	props->min_latency = iolink1->min_latency + iolink2->min_latency;
1508 	props->max_latency = iolink1->max_latency + iolink2->max_latency;
1509 	props->min_bandwidth = min(iolink1->min_bandwidth, iolink2->min_bandwidth);
1510 	props->max_bandwidth = min(iolink2->max_bandwidth, iolink2->max_bandwidth);
1511 
1512 	if (iolink1->node_to != iolink2->node_to) {
1513 		/* CPU->CPU  link*/
1514 		cpu_dev = kfd_topology_device_by_proximity_domain(iolink1->node_to);
1515 		if (cpu_dev) {
1516 			list_for_each_entry(iolink3, &cpu_dev->io_link_props, list) {
1517 				if (iolink3->node_to != iolink2->node_to)
1518 					continue;
1519 
1520 				props->weight += iolink3->weight;
1521 				props->min_latency += iolink3->min_latency;
1522 				props->max_latency += iolink3->max_latency;
1523 				props->min_bandwidth = min(props->min_bandwidth,
1524 							   iolink3->min_bandwidth);
1525 				props->max_bandwidth = min(props->max_bandwidth,
1526 							   iolink3->max_bandwidth);
1527 				break;
1528 			}
1529 		} else {
1530 			WARN(1, "CPU node not found");
1531 		}
1532 	}
1533 
1534 	props->node_from = from;
1535 	props->node_to = to;
1536 	peer->node_props.p2p_links_count++;
1537 	list_add_tail(&props->list, &peer->p2p_link_props);
1538 	ret = kfd_build_p2p_node_entry(peer, props);
1539 
1540 	return ret;
1541 }
1542 #endif
1543 
kfd_dev_create_p2p_links(void)1544 static int kfd_dev_create_p2p_links(void)
1545 {
1546 	struct kfd_topology_device *dev;
1547 	struct kfd_topology_device *new_dev;
1548 #if defined(CONFIG_HSA_AMD_P2P)
1549 	uint32_t i;
1550 #endif
1551 	uint32_t k;
1552 	int ret = 0;
1553 
1554 	k = 0;
1555 	list_for_each_entry(dev, &topology_device_list, list)
1556 		k++;
1557 	if (k < 2)
1558 		return 0;
1559 
1560 	new_dev = list_last_entry(&topology_device_list, struct kfd_topology_device, list);
1561 	if (WARN_ON(!new_dev->gpu))
1562 		return 0;
1563 
1564 	k--;
1565 
1566 	/* create in-direct links */
1567 	ret = kfd_create_indirect_link_prop(new_dev, k);
1568 	if (ret < 0)
1569 		goto out;
1570 
1571 	/* create p2p links */
1572 #if defined(CONFIG_HSA_AMD_P2P)
1573 	i = 0;
1574 	list_for_each_entry(dev, &topology_device_list, list) {
1575 		if (dev == new_dev)
1576 			break;
1577 		if (!dev->gpu || !dev->gpu->adev ||
1578 		    (dev->gpu->hive_id &&
1579 		     dev->gpu->hive_id == new_dev->gpu->hive_id))
1580 			goto next;
1581 
1582 		/* check if node(s) is/are peer accessible in one direction or bi-direction */
1583 		ret = kfd_add_peer_prop(new_dev, dev, i, k);
1584 		if (ret < 0)
1585 			goto out;
1586 
1587 		ret = kfd_add_peer_prop(dev, new_dev, k, i);
1588 		if (ret < 0)
1589 			goto out;
1590 next:
1591 		i++;
1592 	}
1593 #endif
1594 
1595 out:
1596 	return ret;
1597 }
1598 
1599 
1600 /* Helper function. See kfd_fill_gpu_cache_info for parameter description */
fill_in_l1_pcache(struct kfd_cache_properties ** props_ext,struct kfd_gpu_cache_info * pcache_info,struct kfd_cu_info * cu_info,int cu_bitmask,int cache_type,unsigned int cu_processor_id,int cu_block)1601 static int fill_in_l1_pcache(struct kfd_cache_properties **props_ext,
1602 				struct kfd_gpu_cache_info *pcache_info,
1603 				struct kfd_cu_info *cu_info,
1604 				int cu_bitmask,
1605 				int cache_type, unsigned int cu_processor_id,
1606 				int cu_block)
1607 {
1608 	unsigned int cu_sibling_map_mask;
1609 	int first_active_cu;
1610 	struct kfd_cache_properties *pcache = NULL;
1611 
1612 	cu_sibling_map_mask = cu_bitmask;
1613 	cu_sibling_map_mask >>= cu_block;
1614 	cu_sibling_map_mask &= ((1 << pcache_info[cache_type].num_cu_shared) - 1);
1615 	first_active_cu = ffs(cu_sibling_map_mask);
1616 
1617 	/* CU could be inactive. In case of shared cache find the first active
1618 	 * CU. and incase of non-shared cache check if the CU is inactive. If
1619 	 * inactive active skip it
1620 	 */
1621 	if (first_active_cu) {
1622 		pcache = kfd_alloc_struct(pcache);
1623 		if (!pcache)
1624 			return -ENOMEM;
1625 
1626 		memset(pcache, 0, sizeof(struct kfd_cache_properties));
1627 		pcache->processor_id_low = cu_processor_id + (first_active_cu - 1);
1628 		pcache->cache_level = pcache_info[cache_type].cache_level;
1629 		pcache->cache_size = pcache_info[cache_type].cache_size;
1630 
1631 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_DATA_CACHE)
1632 			pcache->cache_type |= HSA_CACHE_TYPE_DATA;
1633 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_INST_CACHE)
1634 			pcache->cache_type |= HSA_CACHE_TYPE_INSTRUCTION;
1635 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_CPU_CACHE)
1636 			pcache->cache_type |= HSA_CACHE_TYPE_CPU;
1637 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_SIMD_CACHE)
1638 			pcache->cache_type |= HSA_CACHE_TYPE_HSACU;
1639 
1640 		/* Sibling map is w.r.t processor_id_low, so shift out
1641 		 * inactive CU
1642 		 */
1643 		cu_sibling_map_mask =
1644 			cu_sibling_map_mask >> (first_active_cu - 1);
1645 
1646 		pcache->sibling_map[0] = (uint8_t)(cu_sibling_map_mask & 0xFF);
1647 		pcache->sibling_map[1] =
1648 				(uint8_t)((cu_sibling_map_mask >> 8) & 0xFF);
1649 		pcache->sibling_map[2] =
1650 				(uint8_t)((cu_sibling_map_mask >> 16) & 0xFF);
1651 		pcache->sibling_map[3] =
1652 				(uint8_t)((cu_sibling_map_mask >> 24) & 0xFF);
1653 
1654 		pcache->sibling_map_size = 4;
1655 		*props_ext = pcache;
1656 
1657 		return 0;
1658 	}
1659 	return 1;
1660 }
1661 
1662 /* Helper function. See kfd_fill_gpu_cache_info for parameter description */
fill_in_l2_l3_pcache(struct kfd_cache_properties ** props_ext,struct kfd_gpu_cache_info * pcache_info,struct kfd_cu_info * cu_info,int cache_type,unsigned int cu_processor_id)1663 static int fill_in_l2_l3_pcache(struct kfd_cache_properties **props_ext,
1664 				struct kfd_gpu_cache_info *pcache_info,
1665 				struct kfd_cu_info *cu_info,
1666 				int cache_type, unsigned int cu_processor_id)
1667 {
1668 	unsigned int cu_sibling_map_mask;
1669 	int first_active_cu;
1670 	int i, j, k;
1671 	struct kfd_cache_properties *pcache = NULL;
1672 
1673 	cu_sibling_map_mask = cu_info->cu_bitmap[0][0];
1674 	cu_sibling_map_mask &=
1675 		((1 << pcache_info[cache_type].num_cu_shared) - 1);
1676 	first_active_cu = ffs(cu_sibling_map_mask);
1677 
1678 	/* CU could be inactive. In case of shared cache find the first active
1679 	 * CU. and incase of non-shared cache check if the CU is inactive. If
1680 	 * inactive active skip it
1681 	 */
1682 	if (first_active_cu) {
1683 		pcache = kfd_alloc_struct(pcache);
1684 		if (!pcache)
1685 			return -ENOMEM;
1686 
1687 		memset(pcache, 0, sizeof(struct kfd_cache_properties));
1688 		pcache->processor_id_low = cu_processor_id
1689 					+ (first_active_cu - 1);
1690 		pcache->cache_level = pcache_info[cache_type].cache_level;
1691 		pcache->cache_size = pcache_info[cache_type].cache_size;
1692 
1693 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_DATA_CACHE)
1694 			pcache->cache_type |= HSA_CACHE_TYPE_DATA;
1695 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_INST_CACHE)
1696 			pcache->cache_type |= HSA_CACHE_TYPE_INSTRUCTION;
1697 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_CPU_CACHE)
1698 			pcache->cache_type |= HSA_CACHE_TYPE_CPU;
1699 		if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_SIMD_CACHE)
1700 			pcache->cache_type |= HSA_CACHE_TYPE_HSACU;
1701 
1702 		/* Sibling map is w.r.t processor_id_low, so shift out
1703 		 * inactive CU
1704 		 */
1705 		cu_sibling_map_mask = cu_sibling_map_mask >> (first_active_cu - 1);
1706 		k = 0;
1707 
1708 		for (i = 0; i < cu_info->num_shader_engines; i++) {
1709 			for (j = 0; j < cu_info->num_shader_arrays_per_engine; j++) {
1710 				pcache->sibling_map[k] = (uint8_t)(cu_sibling_map_mask & 0xFF);
1711 				pcache->sibling_map[k+1] = (uint8_t)((cu_sibling_map_mask >> 8) & 0xFF);
1712 				pcache->sibling_map[k+2] = (uint8_t)((cu_sibling_map_mask >> 16) & 0xFF);
1713 				pcache->sibling_map[k+3] = (uint8_t)((cu_sibling_map_mask >> 24) & 0xFF);
1714 				k += 4;
1715 
1716 				cu_sibling_map_mask = cu_info->cu_bitmap[i % 4][j + i / 4];
1717 				cu_sibling_map_mask &= ((1 << pcache_info[cache_type].num_cu_shared) - 1);
1718 			}
1719 		}
1720 		pcache->sibling_map_size = k;
1721 		*props_ext = pcache;
1722 		return 0;
1723 	}
1724 	return 1;
1725 }
1726 
1727 #define KFD_MAX_CACHE_TYPES 6
1728 
1729 /* kfd_fill_cache_non_crat_info - Fill GPU cache info using kfd_gpu_cache_info
1730  * tables
1731  */
kfd_fill_cache_non_crat_info(struct kfd_topology_device * dev,struct kfd_dev * kdev)1732 void kfd_fill_cache_non_crat_info(struct kfd_topology_device *dev, struct kfd_dev *kdev)
1733 {
1734 	struct kfd_gpu_cache_info *pcache_info = NULL;
1735 	int i, j, k;
1736 	int ct = 0;
1737 	unsigned int cu_processor_id;
1738 	int ret;
1739 	unsigned int num_cu_shared;
1740 	struct kfd_cu_info cu_info;
1741 	struct kfd_cu_info *pcu_info;
1742 	int gpu_processor_id;
1743 	struct kfd_cache_properties *props_ext;
1744 	int num_of_entries = 0;
1745 	int num_of_cache_types = 0;
1746 	struct kfd_gpu_cache_info cache_info[KFD_MAX_CACHE_TYPES];
1747 
1748 	amdgpu_amdkfd_get_cu_info(kdev->adev, &cu_info);
1749 	pcu_info = &cu_info;
1750 
1751 	gpu_processor_id = dev->node_props.simd_id_base;
1752 
1753 	pcache_info = cache_info;
1754 	num_of_cache_types = kfd_get_gpu_cache_info(kdev, &pcache_info);
1755 	if (!num_of_cache_types) {
1756 		pr_warn("no cache info found\n");
1757 		return;
1758 	}
1759 
1760 	/* For each type of cache listed in the kfd_gpu_cache_info table,
1761 	 * go through all available Compute Units.
1762 	 * The [i,j,k] loop will
1763 	 *		if kfd_gpu_cache_info.num_cu_shared = 1
1764 	 *			will parse through all available CU
1765 	 *		If (kfd_gpu_cache_info.num_cu_shared != 1)
1766 	 *			then it will consider only one CU from
1767 	 *			the shared unit
1768 	 */
1769 	for (ct = 0; ct < num_of_cache_types; ct++) {
1770 		cu_processor_id = gpu_processor_id;
1771 		if (pcache_info[ct].cache_level == 1) {
1772 			for (i = 0; i < pcu_info->num_shader_engines; i++) {
1773 				for (j = 0; j < pcu_info->num_shader_arrays_per_engine; j++) {
1774 					for (k = 0; k < pcu_info->num_cu_per_sh; k += pcache_info[ct].num_cu_shared) {
1775 
1776 						ret = fill_in_l1_pcache(&props_ext, pcache_info, pcu_info,
1777 										pcu_info->cu_bitmap[i % 4][j + i / 4], ct,
1778 										cu_processor_id, k);
1779 
1780 						if (ret < 0)
1781 							break;
1782 
1783 						if (!ret) {
1784 							num_of_entries++;
1785 							list_add_tail(&props_ext->list, &dev->cache_props);
1786 						}
1787 
1788 						/* Move to next CU block */
1789 						num_cu_shared = ((k + pcache_info[ct].num_cu_shared) <=
1790 							pcu_info->num_cu_per_sh) ?
1791 							pcache_info[ct].num_cu_shared :
1792 							(pcu_info->num_cu_per_sh - k);
1793 						cu_processor_id += num_cu_shared;
1794 					}
1795 				}
1796 			}
1797 		} else {
1798 			ret = fill_in_l2_l3_pcache(&props_ext, pcache_info,
1799 								pcu_info, ct, cu_processor_id);
1800 
1801 			if (ret < 0)
1802 				break;
1803 
1804 			if (!ret) {
1805 				num_of_entries++;
1806 				list_add_tail(&props_ext->list, &dev->cache_props);
1807 			}
1808 		}
1809 	}
1810 	dev->node_props.caches_count += num_of_entries;
1811 	pr_debug("Added [%d] GPU cache entries\n", num_of_entries);
1812 }
1813 
kfd_topology_add_device(struct kfd_dev * gpu)1814 int kfd_topology_add_device(struct kfd_dev *gpu)
1815 {
1816 	uint32_t gpu_id;
1817 	struct kfd_topology_device *dev;
1818 	struct kfd_cu_info cu_info;
1819 	int res = 0;
1820 	struct list_head temp_topology_device_list;
1821 	void *crat_image = NULL;
1822 	size_t image_size = 0;
1823 	int proximity_domain;
1824 	int i;
1825 	const char *asic_name = amdgpu_asic_name[gpu->adev->asic_type];
1826 
1827 	INIT_LIST_HEAD(&temp_topology_device_list);
1828 
1829 	gpu_id = kfd_generate_gpu_id(gpu);
1830 	pr_debug("Adding new GPU (ID: 0x%x) to topology\n", gpu_id);
1831 
1832 	/* Check to see if this gpu device exists in the topology_device_list.
1833 	 * If so, assign the gpu to that device,
1834 	 * else create a Virtual CRAT for this gpu device and then parse that
1835 	 * CRAT to create a new topology device. Once created assign the gpu to
1836 	 * that topology device
1837 	 */
1838 	down_write(&topology_lock);
1839 	dev = kfd_assign_gpu(gpu);
1840 	if (!dev) {
1841 		proximity_domain = ++topology_crat_proximity_domain;
1842 
1843 		res = kfd_create_crat_image_virtual(&crat_image, &image_size,
1844 						    COMPUTE_UNIT_GPU, gpu,
1845 						    proximity_domain);
1846 		if (res) {
1847 			pr_err("Error creating VCRAT for GPU (ID: 0x%x)\n",
1848 			       gpu_id);
1849 			topology_crat_proximity_domain--;
1850 			return res;
1851 		}
1852 
1853 		res = kfd_parse_crat_table(crat_image,
1854 					   &temp_topology_device_list,
1855 					   proximity_domain);
1856 		if (res) {
1857 			pr_err("Error parsing VCRAT for GPU (ID: 0x%x)\n",
1858 			       gpu_id);
1859 			topology_crat_proximity_domain--;
1860 			goto err;
1861 		}
1862 
1863 		kfd_topology_update_device_list(&temp_topology_device_list,
1864 			&topology_device_list);
1865 
1866 		dev = kfd_assign_gpu(gpu);
1867 		if (WARN_ON(!dev)) {
1868 			res = -ENODEV;
1869 			goto err;
1870 		}
1871 
1872 		/* Fill the cache affinity information here for the GPUs
1873 		 * using VCRAT
1874 		 */
1875 		kfd_fill_cache_non_crat_info(dev, gpu);
1876 
1877 		/* Update the SYSFS tree, since we added another topology
1878 		 * device
1879 		 */
1880 		res = kfd_topology_update_sysfs();
1881 		if (!res)
1882 			sys_props.generation_count++;
1883 		else
1884 			pr_err("Failed to update GPU (ID: 0x%x) to sysfs topology. res=%d\n",
1885 						gpu_id, res);
1886 	}
1887 	up_write(&topology_lock);
1888 
1889 	dev->gpu_id = gpu_id;
1890 	gpu->id = gpu_id;
1891 
1892 	kfd_dev_create_p2p_links();
1893 
1894 	/* TODO: Move the following lines to function
1895 	 *	kfd_add_non_crat_information
1896 	 */
1897 
1898 	/* Fill-in additional information that is not available in CRAT but
1899 	 * needed for the topology
1900 	 */
1901 
1902 	amdgpu_amdkfd_get_cu_info(dev->gpu->adev, &cu_info);
1903 
1904 	for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1; i++) {
1905 		dev->node_props.name[i] = __tolower(asic_name[i]);
1906 		if (asic_name[i] == '\0')
1907 			break;
1908 	}
1909 	dev->node_props.name[i] = '\0';
1910 
1911 	dev->node_props.simd_arrays_per_engine =
1912 		cu_info.num_shader_arrays_per_engine;
1913 
1914 	dev->node_props.gfx_target_version = gpu->device_info.gfx_target_version;
1915 	dev->node_props.vendor_id = gpu->pdev->vendor;
1916 	dev->node_props.device_id = gpu->pdev->device;
1917 	dev->node_props.capability |=
1918 		((dev->gpu->adev->rev_id << HSA_CAP_ASIC_REVISION_SHIFT) &
1919 			HSA_CAP_ASIC_REVISION_MASK);
1920 	dev->node_props.location_id = pci_dev_id(gpu->pdev);
1921 	dev->node_props.domain = pci_domain_nr(gpu->pdev->bus);
1922 	dev->node_props.max_engine_clk_fcompute =
1923 		amdgpu_amdkfd_get_max_engine_clock_in_mhz(dev->gpu->adev);
1924 	dev->node_props.max_engine_clk_ccompute =
1925 		cpufreq_quick_get_max(0) / 1000;
1926 	dev->node_props.drm_render_minor =
1927 		gpu->shared_resources.drm_render_minor;
1928 
1929 	dev->node_props.hive_id = gpu->hive_id;
1930 	dev->node_props.num_sdma_engines = kfd_get_num_sdma_engines(gpu);
1931 	dev->node_props.num_sdma_xgmi_engines =
1932 					kfd_get_num_xgmi_sdma_engines(gpu);
1933 	dev->node_props.num_sdma_queues_per_engine =
1934 				gpu->device_info.num_sdma_queues_per_engine -
1935 				gpu->device_info.num_reserved_sdma_queues_per_engine;
1936 	dev->node_props.num_gws = (dev->gpu->gws &&
1937 		dev->gpu->dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) ?
1938 		dev->gpu->adev->gds.gws_size : 0;
1939 	dev->node_props.num_cp_queues = get_cp_queues_num(dev->gpu->dqm);
1940 
1941 	kfd_fill_mem_clk_max_info(dev);
1942 	kfd_fill_iolink_non_crat_info(dev);
1943 
1944 	switch (dev->gpu->adev->asic_type) {
1945 	case CHIP_KAVERI:
1946 	case CHIP_HAWAII:
1947 	case CHIP_TONGA:
1948 		dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_PRE_1_0 <<
1949 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1950 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1951 		break;
1952 	case CHIP_CARRIZO:
1953 	case CHIP_FIJI:
1954 	case CHIP_POLARIS10:
1955 	case CHIP_POLARIS11:
1956 	case CHIP_POLARIS12:
1957 	case CHIP_VEGAM:
1958 		pr_debug("Adding doorbell packet type capability\n");
1959 		dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 <<
1960 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1961 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1962 		break;
1963 	default:
1964 		if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(9, 0, 1))
1965 			dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 <<
1966 				HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1967 				HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1968 		else
1969 			WARN(1, "Unexpected ASIC family %u",
1970 			     dev->gpu->adev->asic_type);
1971 	}
1972 
1973 	/*
1974 	 * Overwrite ATS capability according to needs_iommu_device to fix
1975 	 * potential missing corresponding bit in CRAT of BIOS.
1976 	 */
1977 	if (dev->gpu->use_iommu_v2)
1978 		dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
1979 	else
1980 		dev->node_props.capability &= ~HSA_CAP_ATS_PRESENT;
1981 
1982 	/* Fix errors in CZ CRAT.
1983 	 * simd_count: Carrizo CRAT reports wrong simd_count, probably
1984 	 *		because it doesn't consider masked out CUs
1985 	 * max_waves_per_simd: Carrizo reports wrong max_waves_per_simd
1986 	 */
1987 	if (dev->gpu->adev->asic_type == CHIP_CARRIZO) {
1988 		dev->node_props.simd_count =
1989 			cu_info.simd_per_cu * cu_info.cu_active_number;
1990 		dev->node_props.max_waves_per_simd = 10;
1991 	}
1992 
1993 	/* kfd only concerns sram ecc on GFX and HBM ecc on UMC */
1994 	dev->node_props.capability |=
1995 		((dev->gpu->adev->ras_enabled & BIT(AMDGPU_RAS_BLOCK__GFX)) != 0) ?
1996 		HSA_CAP_SRAM_EDCSUPPORTED : 0;
1997 	dev->node_props.capability |=
1998 		((dev->gpu->adev->ras_enabled & BIT(AMDGPU_RAS_BLOCK__UMC)) != 0) ?
1999 		HSA_CAP_MEM_EDCSUPPORTED : 0;
2000 
2001 	if (KFD_GC_VERSION(dev->gpu) != IP_VERSION(9, 0, 1))
2002 		dev->node_props.capability |= (dev->gpu->adev->ras_enabled != 0) ?
2003 			HSA_CAP_RASEVENTNOTIFY : 0;
2004 
2005 	if (KFD_IS_SVM_API_SUPPORTED(dev->gpu->adev->kfd.dev))
2006 		dev->node_props.capability |= HSA_CAP_SVMAPI_SUPPORTED;
2007 
2008 	kfd_debug_print_topology();
2009 
2010 	if (!res)
2011 		kfd_notify_gpu_change(gpu_id, 1);
2012 err:
2013 	kfd_destroy_crat_image(crat_image);
2014 	return res;
2015 }
2016 
2017 /**
2018  * kfd_topology_update_io_links() - Update IO links after device removal.
2019  * @proximity_domain: Proximity domain value of the dev being removed.
2020  *
2021  * The topology list currently is arranged in increasing order of
2022  * proximity domain.
2023  *
2024  * Two things need to be done when a device is removed:
2025  * 1. All the IO links to this device need to be removed.
2026  * 2. All nodes after the current device node need to move
2027  *    up once this device node is removed from the topology
2028  *    list. As a result, the proximity domain values for
2029  *    all nodes after the node being deleted reduce by 1.
2030  *    This would also cause the proximity domain values for
2031  *    io links to be updated based on new proximity domain
2032  *    values.
2033  *
2034  * Context: The caller must hold write topology_lock.
2035  */
kfd_topology_update_io_links(int proximity_domain)2036 static void kfd_topology_update_io_links(int proximity_domain)
2037 {
2038 	struct kfd_topology_device *dev;
2039 	struct kfd_iolink_properties *iolink, *p2plink, *tmp;
2040 
2041 	list_for_each_entry(dev, &topology_device_list, list) {
2042 		if (dev->proximity_domain > proximity_domain)
2043 			dev->proximity_domain--;
2044 
2045 		list_for_each_entry_safe(iolink, tmp, &dev->io_link_props, list) {
2046 			/*
2047 			 * If there is an io link to the dev being deleted
2048 			 * then remove that IO link also.
2049 			 */
2050 			if (iolink->node_to == proximity_domain) {
2051 				list_del(&iolink->list);
2052 				dev->node_props.io_links_count--;
2053 			} else {
2054 				if (iolink->node_from > proximity_domain)
2055 					iolink->node_from--;
2056 				if (iolink->node_to > proximity_domain)
2057 					iolink->node_to--;
2058 			}
2059 		}
2060 
2061 		list_for_each_entry_safe(p2plink, tmp, &dev->p2p_link_props, list) {
2062 			/*
2063 			 * If there is a p2p link to the dev being deleted
2064 			 * then remove that p2p link also.
2065 			 */
2066 			if (p2plink->node_to == proximity_domain) {
2067 				list_del(&p2plink->list);
2068 				dev->node_props.p2p_links_count--;
2069 			} else {
2070 				if (p2plink->node_from > proximity_domain)
2071 					p2plink->node_from--;
2072 				if (p2plink->node_to > proximity_domain)
2073 					p2plink->node_to--;
2074 			}
2075 		}
2076 	}
2077 }
2078 
kfd_topology_remove_device(struct kfd_dev * gpu)2079 int kfd_topology_remove_device(struct kfd_dev *gpu)
2080 {
2081 	struct kfd_topology_device *dev, *tmp;
2082 	uint32_t gpu_id;
2083 	int res = -ENODEV;
2084 	int i = 0;
2085 
2086 	down_write(&topology_lock);
2087 
2088 	list_for_each_entry_safe(dev, tmp, &topology_device_list, list) {
2089 		if (dev->gpu == gpu) {
2090 			gpu_id = dev->gpu_id;
2091 			kfd_remove_sysfs_node_entry(dev);
2092 			kfd_release_topology_device(dev);
2093 			sys_props.num_devices--;
2094 			kfd_topology_update_io_links(i);
2095 			topology_crat_proximity_domain = sys_props.num_devices-1;
2096 			sys_props.generation_count++;
2097 			res = 0;
2098 			if (kfd_topology_update_sysfs() < 0)
2099 				kfd_topology_release_sysfs();
2100 			break;
2101 		}
2102 		i++;
2103 	}
2104 
2105 	up_write(&topology_lock);
2106 
2107 	if (!res)
2108 		kfd_notify_gpu_change(gpu_id, 0);
2109 
2110 	return res;
2111 }
2112 
2113 /* kfd_topology_enum_kfd_devices - Enumerate through all devices in KFD
2114  *	topology. If GPU device is found @idx, then valid kfd_dev pointer is
2115  *	returned through @kdev
2116  * Return -	0: On success (@kdev will be NULL for non GPU nodes)
2117  *		-1: If end of list
2118  */
kfd_topology_enum_kfd_devices(uint8_t idx,struct kfd_dev ** kdev)2119 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev)
2120 {
2121 
2122 	struct kfd_topology_device *top_dev;
2123 	uint8_t device_idx = 0;
2124 
2125 	*kdev = NULL;
2126 	down_read(&topology_lock);
2127 
2128 	list_for_each_entry(top_dev, &topology_device_list, list) {
2129 		if (device_idx == idx) {
2130 			*kdev = top_dev->gpu;
2131 			up_read(&topology_lock);
2132 			return 0;
2133 		}
2134 
2135 		device_idx++;
2136 	}
2137 
2138 	up_read(&topology_lock);
2139 
2140 	return -1;
2141 
2142 }
2143 
kfd_cpumask_to_apic_id(const struct cpumask * cpumask)2144 static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask)
2145 {
2146 	int first_cpu_of_numa_node;
2147 
2148 	if (!cpumask || cpumask == cpu_none_mask)
2149 		return -1;
2150 	first_cpu_of_numa_node = cpumask_first(cpumask);
2151 	if (first_cpu_of_numa_node >= nr_cpu_ids)
2152 		return -1;
2153 #ifdef CONFIG_X86_64
2154 	return cpu_data(first_cpu_of_numa_node).apicid;
2155 #else
2156 	return first_cpu_of_numa_node;
2157 #endif
2158 }
2159 
2160 /* kfd_numa_node_to_apic_id - Returns the APIC ID of the first logical processor
2161  *	of the given NUMA node (numa_node_id)
2162  * Return -1 on failure
2163  */
kfd_numa_node_to_apic_id(int numa_node_id)2164 int kfd_numa_node_to_apic_id(int numa_node_id)
2165 {
2166 	if (numa_node_id == -1) {
2167 		pr_warn("Invalid NUMA Node. Use online CPU mask\n");
2168 		return kfd_cpumask_to_apic_id(cpu_online_mask);
2169 	}
2170 	return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id));
2171 }
2172 
kfd_double_confirm_iommu_support(struct kfd_dev * gpu)2173 void kfd_double_confirm_iommu_support(struct kfd_dev *gpu)
2174 {
2175 	struct kfd_topology_device *dev;
2176 
2177 	gpu->use_iommu_v2 = false;
2178 
2179 	if (!gpu->device_info.needs_iommu_device)
2180 		return;
2181 
2182 	down_read(&topology_lock);
2183 
2184 	/* Only use IOMMUv2 if there is an APU topology node with no GPU
2185 	 * assigned yet. This GPU will be assigned to it.
2186 	 */
2187 	list_for_each_entry(dev, &topology_device_list, list)
2188 		if (dev->node_props.cpu_cores_count &&
2189 		    dev->node_props.simd_count &&
2190 		    !dev->gpu)
2191 			gpu->use_iommu_v2 = true;
2192 
2193 	up_read(&topology_lock);
2194 }
2195 
2196 #if defined(CONFIG_DEBUG_FS)
2197 
kfd_debugfs_hqds_by_device(struct seq_file * m,void * data)2198 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data)
2199 {
2200 	struct kfd_topology_device *dev;
2201 	unsigned int i = 0;
2202 	int r = 0;
2203 
2204 	down_read(&topology_lock);
2205 
2206 	list_for_each_entry(dev, &topology_device_list, list) {
2207 		if (!dev->gpu) {
2208 			i++;
2209 			continue;
2210 		}
2211 
2212 		seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
2213 		r = dqm_debugfs_hqds(m, dev->gpu->dqm);
2214 		if (r)
2215 			break;
2216 	}
2217 
2218 	up_read(&topology_lock);
2219 
2220 	return r;
2221 }
2222 
kfd_debugfs_rls_by_device(struct seq_file * m,void * data)2223 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data)
2224 {
2225 	struct kfd_topology_device *dev;
2226 	unsigned int i = 0;
2227 	int r = 0;
2228 
2229 	down_read(&topology_lock);
2230 
2231 	list_for_each_entry(dev, &topology_device_list, list) {
2232 		if (!dev->gpu) {
2233 			i++;
2234 			continue;
2235 		}
2236 
2237 		seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
2238 		r = pm_debugfs_runlist(m, &dev->gpu->dqm->packet_mgr);
2239 		if (r)
2240 			break;
2241 	}
2242 
2243 	up_read(&topology_lock);
2244 
2245 	return r;
2246 }
2247 
2248 #endif
2249