forked from ~ljy/RK356X_SDK_RELEASE

blame | history | patch | commit | commitdiff
hc
2023-12-11 d2ccde1c8e90d38cee87a1b0309ad2827f3fd30d 
 kernel/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
....@@ -36,6 +36,8 @@
3636 #include "kfd_topology.h"
3737 #include "kfd_device_queue_manager.h"
3838 #include "kfd_iommu.h"
39
+#include "amdgpu_amdkfd.h"
40
+#include "amdgpu_ras.h"
3941 
4042 /* topology_device_list - Master list of all topology devices */
4143 static struct list_head topology_device_list;
....@@ -100,7 +102,25 @@
100102 	down_read(&topology_lock);
101103 
102104 	list_for_each_entry(top_dev, &topology_device_list, list)
103
-		if (top_dev->gpu->pdev == pdev) {
105
+		if (top_dev->gpu && top_dev->gpu->pdev == pdev) {
106
+			device = top_dev->gpu;
107
+			break;
108
+		}
109
+
110
+	up_read(&topology_lock);
111
+
112
+	return device;
113
+}
114
+
115
+struct kfd_dev *kfd_device_by_kgd(const struct kgd_dev *kgd)
116
+{
117
+	struct kfd_topology_device *top_dev;
118
+	struct kfd_dev *device = NULL;
119
+
120
+	down_read(&topology_lock);
121
+
122
+	list_for_each_entry(top_dev, &topology_device_list, list)
123
+		if (top_dev->gpu && top_dev->gpu->kgd == kgd) {
104124 			device = top_dev->gpu;
105125 			break;
106126 		}
....@@ -190,39 +210,41 @@
190210 }
191211 
192212 
193
-#define sysfs_show_gen_prop(buffer, fmt, ...) \
194
-		snprintf(buffer, PAGE_SIZE, "%s"fmt, buffer, __VA_ARGS__)
195
-#define sysfs_show_32bit_prop(buffer, name, value) \
196
-		sysfs_show_gen_prop(buffer, "%s %u\n", name, value)
197
-#define sysfs_show_64bit_prop(buffer, name, value) \
198
-		sysfs_show_gen_prop(buffer, "%s %llu\n", name, value)
199
-#define sysfs_show_32bit_val(buffer, value) \
200
-		sysfs_show_gen_prop(buffer, "%u\n", value)
201
-#define sysfs_show_str_val(buffer, value) \
202
-		sysfs_show_gen_prop(buffer, "%s\n", value)
213
+#define sysfs_show_gen_prop(buffer, offs, fmt, ...)		\
214
+		(offs += snprintf(buffer+offs, PAGE_SIZE-offs,	\
215
+				  fmt, __VA_ARGS__))
216
+#define sysfs_show_32bit_prop(buffer, offs, name, value) \
217
+		sysfs_show_gen_prop(buffer, offs, "%s %u\n", name, value)
218
+#define sysfs_show_64bit_prop(buffer, offs, name, value) \
219
+		sysfs_show_gen_prop(buffer, offs, "%s %llu\n", name, value)
220
+#define sysfs_show_32bit_val(buffer, offs, value) \
221
+		sysfs_show_gen_prop(buffer, offs, "%u\n", value)
222
+#define sysfs_show_str_val(buffer, offs, value) \
223
+		sysfs_show_gen_prop(buffer, offs, "%s\n", value)
203224 
204225 static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr,
205226 		char *buffer)
206227 {
207
-	ssize_t ret;
228
+	int offs = 0;
208229 
209230 	/* Making sure that the buffer is an empty string */
210231 	buffer[0] = 0;
211232 
212233 	if (attr == &sys_props.attr_genid) {
213
-		ret = sysfs_show_32bit_val(buffer, sys_props.generation_count);
234
+		sysfs_show_32bit_val(buffer, offs,
235
+				     sys_props.generation_count);
214236 	} else if (attr == &sys_props.attr_props) {
215
-		sysfs_show_64bit_prop(buffer, "platform_oem",
216
-				sys_props.platform_oem);
217
-		sysfs_show_64bit_prop(buffer, "platform_id",
218
-				sys_props.platform_id);
219
-		ret = sysfs_show_64bit_prop(buffer, "platform_rev",
220
-				sys_props.platform_rev);
237
+		sysfs_show_64bit_prop(buffer, offs, "platform_oem",
238
+				      sys_props.platform_oem);
239
+		sysfs_show_64bit_prop(buffer, offs, "platform_id",
240
+				      sys_props.platform_id);
241
+		sysfs_show_64bit_prop(buffer, offs, "platform_rev",
242
+				      sys_props.platform_rev);
221243 	} else {
222
-		ret = -EINVAL;
244
+		offs = -EINVAL;
223245 	}
224246 
225
-	return ret;
247
+	return offs;
226248 }
227249 
228250 static void kfd_topology_kobj_release(struct kobject *kobj)
....@@ -242,28 +264,32 @@
242264 static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr,
243265 		char *buffer)
244266 {
245
-	ssize_t ret;
267
+	int offs = 0;
246268 	struct kfd_iolink_properties *iolink;
247269 
248270 	/* Making sure that the buffer is an empty string */
249271 	buffer[0] = 0;
250272 
251273 	iolink = container_of(attr, struct kfd_iolink_properties, attr);
252
-	sysfs_show_32bit_prop(buffer, "type", iolink->iolink_type);
253
-	sysfs_show_32bit_prop(buffer, "version_major", iolink->ver_maj);
254
-	sysfs_show_32bit_prop(buffer, "version_minor", iolink->ver_min);
255
-	sysfs_show_32bit_prop(buffer, "node_from", iolink->node_from);
256
-	sysfs_show_32bit_prop(buffer, "node_to", iolink->node_to);
257
-	sysfs_show_32bit_prop(buffer, "weight", iolink->weight);
258
-	sysfs_show_32bit_prop(buffer, "min_latency", iolink->min_latency);
259
-	sysfs_show_32bit_prop(buffer, "max_latency", iolink->max_latency);
260
-	sysfs_show_32bit_prop(buffer, "min_bandwidth", iolink->min_bandwidth);
261
-	sysfs_show_32bit_prop(buffer, "max_bandwidth", iolink->max_bandwidth);
262
-	sysfs_show_32bit_prop(buffer, "recommended_transfer_size",
263
-			iolink->rec_transfer_size);
264
-	ret = sysfs_show_32bit_prop(buffer, "flags", iolink->flags);
274
+	if (iolink->gpu && kfd_devcgroup_check_permission(iolink->gpu))
275
+		return -EPERM;
276
+	sysfs_show_32bit_prop(buffer, offs, "type", iolink->iolink_type);
277
+	sysfs_show_32bit_prop(buffer, offs, "version_major", iolink->ver_maj);
278
+	sysfs_show_32bit_prop(buffer, offs, "version_minor", iolink->ver_min);
279
+	sysfs_show_32bit_prop(buffer, offs, "node_from", iolink->node_from);
280
+	sysfs_show_32bit_prop(buffer, offs, "node_to", iolink->node_to);
281
+	sysfs_show_32bit_prop(buffer, offs, "weight", iolink->weight);
282
+	sysfs_show_32bit_prop(buffer, offs, "min_latency", iolink->min_latency);
283
+	sysfs_show_32bit_prop(buffer, offs, "max_latency", iolink->max_latency);
284
+	sysfs_show_32bit_prop(buffer, offs, "min_bandwidth",
285
+			      iolink->min_bandwidth);
286
+	sysfs_show_32bit_prop(buffer, offs, "max_bandwidth",
287
+			      iolink->max_bandwidth);
288
+	sysfs_show_32bit_prop(buffer, offs, "recommended_transfer_size",
289
+			      iolink->rec_transfer_size);
290
+	sysfs_show_32bit_prop(buffer, offs, "flags", iolink->flags);
265291 
266
-	return ret;
292
+	return offs;
267293 }
268294 
269295 static const struct sysfs_ops iolink_ops = {
....@@ -278,20 +304,24 @@
278304 static ssize_t mem_show(struct kobject *kobj, struct attribute *attr,
279305 		char *buffer)
280306 {
281
-	ssize_t ret;
307
+	int offs = 0;
282308 	struct kfd_mem_properties *mem;
283309 
284310 	/* Making sure that the buffer is an empty string */
285311 	buffer[0] = 0;
286312 
287313 	mem = container_of(attr, struct kfd_mem_properties, attr);
288
-	sysfs_show_32bit_prop(buffer, "heap_type", mem->heap_type);
289
-	sysfs_show_64bit_prop(buffer, "size_in_bytes", mem->size_in_bytes);
290
-	sysfs_show_32bit_prop(buffer, "flags", mem->flags);
291
-	sysfs_show_32bit_prop(buffer, "width", mem->width);
292
-	ret = sysfs_show_32bit_prop(buffer, "mem_clk_max", mem->mem_clk_max);
314
+	if (mem->gpu && kfd_devcgroup_check_permission(mem->gpu))
315
+		return -EPERM;
316
+	sysfs_show_32bit_prop(buffer, offs, "heap_type", mem->heap_type);
317
+	sysfs_show_64bit_prop(buffer, offs, "size_in_bytes",
318
+			      mem->size_in_bytes);
319
+	sysfs_show_32bit_prop(buffer, offs, "flags", mem->flags);
320
+	sysfs_show_32bit_prop(buffer, offs, "width", mem->width);
321
+	sysfs_show_32bit_prop(buffer, offs, "mem_clk_max",
322
+			      mem->mem_clk_max);
293323 
294
-	return ret;
324
+	return offs;
295325 }
296326 
297327 static const struct sysfs_ops mem_ops = {
....@@ -306,7 +336,7 @@
306336 static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr,
307337 		char *buffer)
308338 {
309
-	ssize_t ret;
339
+	int offs = 0;
310340 	uint32_t i, j;
311341 	struct kfd_cache_properties *cache;
312342 
....@@ -314,30 +344,29 @@
314344 	buffer[0] = 0;
315345 
316346 	cache = container_of(attr, struct kfd_cache_properties, attr);
317
-	sysfs_show_32bit_prop(buffer, "processor_id_low",
347
+	if (cache->gpu && kfd_devcgroup_check_permission(cache->gpu))
348
+		return -EPERM;
349
+	sysfs_show_32bit_prop(buffer, offs, "processor_id_low",
318350 			cache->processor_id_low);
319
-	sysfs_show_32bit_prop(buffer, "level", cache->cache_level);
320
-	sysfs_show_32bit_prop(buffer, "size", cache->cache_size);
321
-	sysfs_show_32bit_prop(buffer, "cache_line_size", cache->cacheline_size);
322
-	sysfs_show_32bit_prop(buffer, "cache_lines_per_tag",
323
-			cache->cachelines_per_tag);
324
-	sysfs_show_32bit_prop(buffer, "association", cache->cache_assoc);
325
-	sysfs_show_32bit_prop(buffer, "latency", cache->cache_latency);
326
-	sysfs_show_32bit_prop(buffer, "type", cache->cache_type);
327
-	snprintf(buffer, PAGE_SIZE, "%ssibling_map ", buffer);
351
+	sysfs_show_32bit_prop(buffer, offs, "level", cache->cache_level);
352
+	sysfs_show_32bit_prop(buffer, offs, "size", cache->cache_size);
353
+	sysfs_show_32bit_prop(buffer, offs, "cache_line_size",
354
+			      cache->cacheline_size);
355
+	sysfs_show_32bit_prop(buffer, offs, "cache_lines_per_tag",
356
+			      cache->cachelines_per_tag);
357
+	sysfs_show_32bit_prop(buffer, offs, "association", cache->cache_assoc);
358
+	sysfs_show_32bit_prop(buffer, offs, "latency", cache->cache_latency);
359
+	sysfs_show_32bit_prop(buffer, offs, "type", cache->cache_type);
360
+	offs += snprintf(buffer+offs, PAGE_SIZE-offs, "sibling_map ");
328361 	for (i = 0; i < CRAT_SIBLINGMAP_SIZE; i++)
329
-		for (j = 0; j < sizeof(cache->sibling_map[0])*8; j++) {
362
+		for (j = 0; j < sizeof(cache->sibling_map[0])*8; j++)
330363 			/* Check each bit */
331
-			if (cache->sibling_map[i] & (1 << j))
332
-				ret = snprintf(buffer, PAGE_SIZE,
333
-					 "%s%d%s", buffer, 1, ",");
334
-			else
335
-				ret = snprintf(buffer, PAGE_SIZE,
336
-					 "%s%d%s", buffer, 0, ",");
337
-		}
364
+			offs += snprintf(buffer+offs, PAGE_SIZE-offs, "%d,",
365
+					 (cache->sibling_map[i] >> j) & 1);
366
+
338367 	/* Replace the last "," with end of line */
339
-	*(buffer + strlen(buffer) - 1) = 0xA;
340
-	return ret;
368
+	buffer[offs-1] = '\n';
369
+	return offs;
341370 }
342371 
343372 static const struct sysfs_ops cache_ops = {
....@@ -359,6 +388,7 @@
359388 static ssize_t perf_show(struct kobject *kobj, struct kobj_attribute *attrs,
360389 			char *buf)
361390 {
391
+	int offs = 0;
362392 	struct kfd_perf_attr *attr;
363393 
364394 	buf[0] = 0;
....@@ -366,7 +396,7 @@
366396 	if (!attr->data) /* invalid data for PMC */
367397 		return 0;
368398 	else
369
-		return sysfs_show_32bit_val(buf, attr->data);
399
+		return sysfs_show_32bit_val(buf, offs, attr->data);
370400 }
371401 
372402 #define KFD_PERF_DESC(_name, _data)			\
....@@ -385,9 +415,8 @@
385415 static ssize_t node_show(struct kobject *kobj, struct attribute *attr,
386416 		char *buffer)
387417 {
418
+	int offs = 0;
388419 	struct kfd_topology_device *dev;
389
-	char public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE];
390
-	uint32_t i;
391420 	uint32_t log_max_watch_addr;
392421 
393422 	/* Making sure that the buffer is an empty string */
....@@ -396,64 +425,80 @@
396425 	if (strcmp(attr->name, "gpu_id") == 0) {
397426 		dev = container_of(attr, struct kfd_topology_device,
398427 				attr_gpuid);
399
-		return sysfs_show_32bit_val(buffer, dev->gpu_id);
428
+		if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu))
429
+			return -EPERM;
430
+		return sysfs_show_32bit_val(buffer, offs, dev->gpu_id);
400431 	}
401432 
402433 	if (strcmp(attr->name, "name") == 0) {
403434 		dev = container_of(attr, struct kfd_topology_device,
404435 				attr_name);
405
-		for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE; i++) {
406
-			public_name[i] =
407
-					(char)dev->node_props.marketing_name[i];
408
-			if (dev->node_props.marketing_name[i] == 0)
409
-				break;
410
-		}
411
-		public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1] = 0x0;
412
-		return sysfs_show_str_val(buffer, public_name);
436
+
437
+		if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu))
438
+			return -EPERM;
439
+		return sysfs_show_str_val(buffer, offs, dev->node_props.name);
413440 	}
414441 
415442 	dev = container_of(attr, struct kfd_topology_device,
416443 			attr_props);
417
-	sysfs_show_32bit_prop(buffer, "cpu_cores_count",
418
-			dev->node_props.cpu_cores_count);
419
-	sysfs_show_32bit_prop(buffer, "simd_count",
420
-			dev->node_props.simd_count);
421
-	sysfs_show_32bit_prop(buffer, "mem_banks_count",
422
-			dev->node_props.mem_banks_count);
423
-	sysfs_show_32bit_prop(buffer, "caches_count",
424
-			dev->node_props.caches_count);
425
-	sysfs_show_32bit_prop(buffer, "io_links_count",
426
-			dev->node_props.io_links_count);
427
-	sysfs_show_32bit_prop(buffer, "cpu_core_id_base",
428
-			dev->node_props.cpu_core_id_base);
429
-	sysfs_show_32bit_prop(buffer, "simd_id_base",
430
-			dev->node_props.simd_id_base);
431
-	sysfs_show_32bit_prop(buffer, "max_waves_per_simd",
432
-			dev->node_props.max_waves_per_simd);
433
-	sysfs_show_32bit_prop(buffer, "lds_size_in_kb",
434
-			dev->node_props.lds_size_in_kb);
435
-	sysfs_show_32bit_prop(buffer, "gds_size_in_kb",
436
-			dev->node_props.gds_size_in_kb);
437
-	sysfs_show_32bit_prop(buffer, "wave_front_size",
438
-			dev->node_props.wave_front_size);
439
-	sysfs_show_32bit_prop(buffer, "array_count",
440
-			dev->node_props.array_count);
441
-	sysfs_show_32bit_prop(buffer, "simd_arrays_per_engine",
442
-			dev->node_props.simd_arrays_per_engine);
443
-	sysfs_show_32bit_prop(buffer, "cu_per_simd_array",
444
-			dev->node_props.cu_per_simd_array);
445
-	sysfs_show_32bit_prop(buffer, "simd_per_cu",
446
-			dev->node_props.simd_per_cu);
447
-	sysfs_show_32bit_prop(buffer, "max_slots_scratch_cu",
448
-			dev->node_props.max_slots_scratch_cu);
449
-	sysfs_show_32bit_prop(buffer, "vendor_id",
450
-			dev->node_props.vendor_id);
451
-	sysfs_show_32bit_prop(buffer, "device_id",
452
-			dev->node_props.device_id);
453
-	sysfs_show_32bit_prop(buffer, "location_id",
454
-			dev->node_props.location_id);
455
-	sysfs_show_32bit_prop(buffer, "drm_render_minor",
456
-			dev->node_props.drm_render_minor);
444
+	if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu))
445
+		return -EPERM;
446
+	sysfs_show_32bit_prop(buffer, offs, "cpu_cores_count",
447
+			      dev->node_props.cpu_cores_count);
448
+	sysfs_show_32bit_prop(buffer, offs, "simd_count",
449
+			      dev->gpu ? dev->node_props.simd_count : 0);
450
+	sysfs_show_32bit_prop(buffer, offs, "mem_banks_count",
451
+			      dev->node_props.mem_banks_count);
452
+	sysfs_show_32bit_prop(buffer, offs, "caches_count",
453
+			      dev->node_props.caches_count);
454
+	sysfs_show_32bit_prop(buffer, offs, "io_links_count",
455
+			      dev->node_props.io_links_count);
456
+	sysfs_show_32bit_prop(buffer, offs, "cpu_core_id_base",
457
+			      dev->node_props.cpu_core_id_base);
458
+	sysfs_show_32bit_prop(buffer, offs, "simd_id_base",
459
+			      dev->node_props.simd_id_base);
460
+	sysfs_show_32bit_prop(buffer, offs, "max_waves_per_simd",
461
+			      dev->node_props.max_waves_per_simd);
462
+	sysfs_show_32bit_prop(buffer, offs, "lds_size_in_kb",
463
+			      dev->node_props.lds_size_in_kb);
464
+	sysfs_show_32bit_prop(buffer, offs, "gds_size_in_kb",
465
+			      dev->node_props.gds_size_in_kb);
466
+	sysfs_show_32bit_prop(buffer, offs, "num_gws",
467
+			      dev->node_props.num_gws);
468
+	sysfs_show_32bit_prop(buffer, offs, "wave_front_size",
469
+			      dev->node_props.wave_front_size);
470
+	sysfs_show_32bit_prop(buffer, offs, "array_count",
471
+			      dev->node_props.array_count);
472
+	sysfs_show_32bit_prop(buffer, offs, "simd_arrays_per_engine",
473
+			      dev->node_props.simd_arrays_per_engine);
474
+	sysfs_show_32bit_prop(buffer, offs, "cu_per_simd_array",
475
+			      dev->node_props.cu_per_simd_array);
476
+	sysfs_show_32bit_prop(buffer, offs, "simd_per_cu",
477
+			      dev->node_props.simd_per_cu);
478
+	sysfs_show_32bit_prop(buffer, offs, "max_slots_scratch_cu",
479
+			      dev->node_props.max_slots_scratch_cu);
480
+	sysfs_show_32bit_prop(buffer, offs, "vendor_id",
481
+			      dev->node_props.vendor_id);
482
+	sysfs_show_32bit_prop(buffer, offs, "device_id",
483
+			      dev->node_props.device_id);
484
+	sysfs_show_32bit_prop(buffer, offs, "location_id",
485
+			      dev->node_props.location_id);
486
+	sysfs_show_32bit_prop(buffer, offs, "domain",
487
+			      dev->node_props.domain);
488
+	sysfs_show_32bit_prop(buffer, offs, "drm_render_minor",
489
+			      dev->node_props.drm_render_minor);
490
+	sysfs_show_64bit_prop(buffer, offs, "hive_id",
491
+			      dev->node_props.hive_id);
492
+	sysfs_show_32bit_prop(buffer, offs, "num_sdma_engines",
493
+			      dev->node_props.num_sdma_engines);
494
+	sysfs_show_32bit_prop(buffer, offs, "num_sdma_xgmi_engines",
495
+			      dev->node_props.num_sdma_xgmi_engines);
496
+	sysfs_show_32bit_prop(buffer, offs, "num_sdma_queues_per_engine",
497
+			      dev->node_props.num_sdma_queues_per_engine);
498
+	sysfs_show_32bit_prop(buffer, offs, "num_cp_queues",
499
+			      dev->node_props.num_cp_queues);
500
+	sysfs_show_64bit_prop(buffer, offs, "unique_id",
501
+			      dev->node_props.unique_id);
457502 
458503 	if (dev->gpu) {
459504 		log_max_watch_addr =
....@@ -473,22 +518,21 @@
473518 			dev->node_props.capability |=
474519 					HSA_CAP_AQL_QUEUE_DOUBLE_MAP;
475520 
476
-		sysfs_show_32bit_prop(buffer, "max_engine_clk_fcompute",
521
+		sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_fcompute",
477522 			dev->node_props.max_engine_clk_fcompute);
478523 
479
-		sysfs_show_64bit_prop(buffer, "local_mem_size",
480
-				(unsigned long long int) 0);
524
+		sysfs_show_64bit_prop(buffer, offs, "local_mem_size", 0ULL);
481525 
482
-		sysfs_show_32bit_prop(buffer, "fw_version",
483
-			dev->gpu->kfd2kgd->get_fw_version(
484
-						dev->gpu->kgd,
485
-						KGD_ENGINE_MEC1));
486
-		sysfs_show_32bit_prop(buffer, "capability",
487
-				dev->node_props.capability);
526
+		sysfs_show_32bit_prop(buffer, offs, "fw_version",
527
+				      dev->gpu->mec_fw_version);
528
+		sysfs_show_32bit_prop(buffer, offs, "capability",
529
+				      dev->node_props.capability);
530
+		sysfs_show_32bit_prop(buffer, offs, "sdma_fw_version",
531
+				      dev->gpu->sdma_fw_version);
488532 	}
489533 
490
-	return sysfs_show_32bit_prop(buffer, "max_engine_clk_ccompute",
491
-					cpufreq_quick_get_max(0)/1000);
534
+	return sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_ccompute",
535
+				     cpufreq_quick_get_max(0)/1000);
492536 }
493537 
494538 static const struct sysfs_ops node_ops = {
....@@ -757,7 +801,6 @@
757801 {
758802 	int ret;
759803 
760
-	pr_info("Creating topology SYSFS entries\n");
761804 	if (!sys_props.kobj_topology) {
762805 		sys_props.kobj_topology =
763806 				kfd_alloc_struct(sys_props.kobj_topology);
....@@ -1020,7 +1063,6 @@
10201063 		sys_props.generation_count++;
10211064 		kfd_update_system_properties();
10221065 		kfd_debug_print_topology();
1023
-		pr_info("Finished initializing topology\n");
10241066 	} else
10251067 		pr_err("Failed to update topology in sysfs ret=%d\n", ret);
10261068 
....@@ -1060,14 +1102,15 @@
10601102 	if (!gpu)
10611103 		return 0;
10621104 
1063
-	gpu->kfd2kgd->get_local_mem_info(gpu->kgd, &local_mem_info);
1105
+	amdgpu_amdkfd_get_local_mem_info(gpu->kgd, &local_mem_info);
10641106 
10651107 	local_mem_size = local_mem_info.local_mem_size_private +
10661108 			local_mem_info.local_mem_size_public;
10671109 
10681110 	buf[0] = gpu->pdev->devfn;
1069
-	buf[1] = gpu->pdev->subsystem_vendor;
1070
-	buf[2] = gpu->pdev->subsystem_device;
1111
+	buf[1] = gpu->pdev->subsystem_vendor |
1112
+		(gpu->pdev->subsystem_device << 16);
1113
+	buf[2] = pci_domain_nr(gpu->pdev->bus);
10711114 	buf[3] = gpu->pdev->device;
10721115 	buf[4] = gpu->pdev->bus->number;
10731116 	buf[5] = lower_32_bits(local_mem_size);
....@@ -1087,19 +1130,29 @@
10871130 {
10881131 	struct kfd_topology_device *dev;
10891132 	struct kfd_topology_device *out_dev = NULL;
1133
+	struct kfd_mem_properties *mem;
1134
+	struct kfd_cache_properties *cache;
1135
+	struct kfd_iolink_properties *iolink;
10901136 
10911137 	down_write(&topology_lock);
10921138 	list_for_each_entry(dev, &topology_device_list, list) {
10931139 		/* Discrete GPUs need their own topology device list
10941140 		 * entries. Don't assign them to CPU/APU nodes.
10951141 		 */
1096
-		if (!gpu->device_info->needs_iommu_device &&
1142
+		if (!gpu->use_iommu_v2 &&
10971143 		    dev->node_props.cpu_cores_count)
10981144 			continue;
10991145 
11001146 		if (!dev->gpu && (dev->node_props.simd_count > 0)) {
11011147 			dev->gpu = gpu;
11021148 			out_dev = dev;
1149
+
1150
+			list_for_each_entry(mem, &dev->mem_props, list)
1151
+				mem->gpu = dev->gpu;
1152
+			list_for_each_entry(cache, &dev->cache_props, list)
1153
+				cache->gpu = dev->gpu;
1154
+			list_for_each_entry(iolink, &dev->io_link_props, list)
1155
+				iolink->gpu = dev->gpu;
11031156 			break;
11041157 		}
11051158 	}
....@@ -1132,8 +1185,7 @@
11321185 	 * for APUs - If CRAT from ACPI reports more than one bank, then
11331186 	 *	all the banks will report the same mem_clk_max information
11341187 	 */
1135
-	dev->gpu->kfd2kgd->get_local_mem_info(dev->gpu->kgd,
1136
-		&local_mem_info);
1188
+	amdgpu_amdkfd_get_local_mem_info(dev->gpu->kgd, &local_mem_info);
11371189 
11381190 	list_for_each_entry(mem, &dev->mem_props, list)
11391191 		mem->mem_clk_max = local_mem_info.mem_clk_max;
....@@ -1141,17 +1193,40 @@
11411193 
11421194 static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev)
11431195 {
1144
-	struct kfd_iolink_properties *link;
1196
+	struct kfd_iolink_properties *link, *cpu_link;
1197
+	struct kfd_topology_device *cpu_dev;
1198
+	uint32_t cap;
1199
+	uint32_t cpu_flag = CRAT_IOLINK_FLAGS_ENABLED;
1200
+	uint32_t flag = CRAT_IOLINK_FLAGS_ENABLED;
11451201 
11461202 	if (!dev || !dev->gpu)
11471203 		return;
11481204 
1149
-	/* GPU only creates direck links so apply flags setting to all */
1150
-	if (dev->gpu->device_info->asic_family == CHIP_HAWAII)
1151
-		list_for_each_entry(link, &dev->io_link_props, list)
1152
-			link->flags = CRAT_IOLINK_FLAGS_ENABLED |
1153
-				CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1154
-				CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1205
+	pcie_capability_read_dword(dev->gpu->pdev,
1206
+			PCI_EXP_DEVCAP2, &cap);
1207
+
1208
+	if (!(cap & (PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
1209
+		     PCI_EXP_DEVCAP2_ATOMIC_COMP64)))
1210
+		cpu_flag |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1211
+			CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1212
+
1213
+	if (!dev->gpu->pci_atomic_requested ||
1214
+	    dev->gpu->device_info->asic_family == CHIP_HAWAII)
1215
+		flag |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1216
+			CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1217
+
1218
+	/* GPU only creates direct links so apply flags setting to all */
1219
+	list_for_each_entry(link, &dev->io_link_props, list) {
1220
+		link->flags = flag;
1221
+		cpu_dev = kfd_topology_device_by_proximity_domain(
1222
+				link->node_to);
1223
+		if (cpu_dev) {
1224
+			list_for_each_entry(cpu_link,
1225
+					    &cpu_dev->io_link_props, list)
1226
+				if (cpu_link->node_to == link->node_from)
1227
+					cpu_link->flags = cpu_flag;
1228
+		}
1229
+	}
11551230 }
11561231 
11571232 int kfd_topology_add_device(struct kfd_dev *gpu)
....@@ -1164,6 +1239,7 @@
11641239 	void *crat_image = NULL;
11651240 	size_t image_size = 0;
11661241 	int proximity_domain;
1242
+	struct amdgpu_device *adev;
11671243 
11681244 	INIT_LIST_HEAD(&temp_topology_device_list);
11691245 
....@@ -1231,20 +1307,40 @@
12311307 	 * needed for the topology
12321308 	 */
12331309 
1234
-	dev->gpu->kfd2kgd->get_cu_info(dev->gpu->kgd, &cu_info);
1310
+	amdgpu_amdkfd_get_cu_info(dev->gpu->kgd, &cu_info);
1311
+
1312
+	strncpy(dev->node_props.name, gpu->device_info->asic_name,
1313
+			KFD_TOPOLOGY_PUBLIC_NAME_SIZE);
1314
+
12351315 	dev->node_props.simd_arrays_per_engine =
12361316 		cu_info.num_shader_arrays_per_engine;
12371317 
12381318 	dev->node_props.vendor_id = gpu->pdev->vendor;
12391319 	dev->node_props.device_id = gpu->pdev->device;
1240
-	dev->node_props.location_id = PCI_DEVID(gpu->pdev->bus->number,
1241
-		gpu->pdev->devfn);
1320
+	dev->node_props.capability |=
1321
+		((amdgpu_amdkfd_get_asic_rev_id(dev->gpu->kgd) <<
1322
+			HSA_CAP_ASIC_REVISION_SHIFT) &
1323
+			HSA_CAP_ASIC_REVISION_MASK);
1324
+	dev->node_props.location_id = pci_dev_id(gpu->pdev);
1325
+	dev->node_props.domain = pci_domain_nr(gpu->pdev->bus);
12421326 	dev->node_props.max_engine_clk_fcompute =
1243
-		dev->gpu->kfd2kgd->get_max_engine_clock_in_mhz(dev->gpu->kgd);
1327
+		amdgpu_amdkfd_get_max_engine_clock_in_mhz(dev->gpu->kgd);
12441328 	dev->node_props.max_engine_clk_ccompute =
12451329 		cpufreq_quick_get_max(0) / 1000;
12461330 	dev->node_props.drm_render_minor =
12471331 		gpu->shared_resources.drm_render_minor;
1332
+
1333
+	dev->node_props.hive_id = gpu->hive_id;
1334
+	dev->node_props.num_sdma_engines = gpu->device_info->num_sdma_engines;
1335
+	dev->node_props.num_sdma_xgmi_engines =
1336
+				gpu->device_info->num_xgmi_sdma_engines;
1337
+	dev->node_props.num_sdma_queues_per_engine =
1338
+				gpu->device_info->num_sdma_queues_per_engine;
1339
+	dev->node_props.num_gws = (dev->gpu->gws &&
1340
+		dev->gpu->dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) ?
1341
+		amdgpu_amdkfd_get_num_gws(dev->gpu->kgd) : 0;
1342
+	dev->node_props.num_cp_queues = get_cp_queues_num(dev->gpu->dqm);
1343
+	dev->node_props.unique_id = gpu->unique_id;
12481344 
12491345 	kfd_fill_mem_clk_max_info(dev);
12501346 	kfd_fill_iolink_non_crat_info(dev);
....@@ -1261,13 +1357,24 @@
12611357 	case CHIP_FIJI:
12621358 	case CHIP_POLARIS10:
12631359 	case CHIP_POLARIS11:
1360
+	case CHIP_POLARIS12:
1361
+	case CHIP_VEGAM:
12641362 		pr_debug("Adding doorbell packet type capability\n");
12651363 		dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 <<
12661364 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
12671365 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
12681366 		break;
12691367 	case CHIP_VEGA10:
1368
+	case CHIP_VEGA12:
1369
+	case CHIP_VEGA20:
12701370 	case CHIP_RAVEN:
1371
+	case CHIP_RENOIR:
1372
+	case CHIP_ARCTURUS:
1373
+	case CHIP_NAVI10:
1374
+	case CHIP_NAVI12:
1375
+	case CHIP_NAVI14:
1376
+	case CHIP_SIENNA_CICHLID:
1377
+	case CHIP_NAVY_FLOUNDER:
12711378 		dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 <<
12721379 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
12731380 			HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
....@@ -1277,18 +1384,37 @@
12771384 		     dev->gpu->device_info->asic_family);
12781385 	}
12791386 
1387
+	/*
1388
+	* Overwrite ATS capability according to needs_iommu_device to fix
1389
+	* potential missing corresponding bit in CRAT of BIOS.
1390
+	*/
1391
+	if (dev->gpu->use_iommu_v2)
1392
+		dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
1393
+	else
1394
+		dev->node_props.capability &= ~HSA_CAP_ATS_PRESENT;
1395
+
12801396 	/* Fix errors in CZ CRAT.
12811397 	 * simd_count: Carrizo CRAT reports wrong simd_count, probably
12821398 	 *		because it doesn't consider masked out CUs
12831399 	 * max_waves_per_simd: Carrizo reports wrong max_waves_per_simd
1284
-	 * capability flag: Carrizo CRAT doesn't report IOMMU flags
12851400 	 */
12861401 	if (dev->gpu->device_info->asic_family == CHIP_CARRIZO) {
12871402 		dev->node_props.simd_count =
12881403 			cu_info.simd_per_cu * cu_info.cu_active_number;
12891404 		dev->node_props.max_waves_per_simd = 10;
1290
-		dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
12911405 	}
1406
+
1407
+	adev = (struct amdgpu_device *)(dev->gpu->kgd);
1408
+	/* kfd only concerns sram ecc on GFX and HBM ecc on UMC */
1409
+	dev->node_props.capability |=
1410
+		((adev->ras_features & BIT(AMDGPU_RAS_BLOCK__GFX)) != 0) ?
1411
+		HSA_CAP_SRAM_EDCSUPPORTED : 0;
1412
+	dev->node_props.capability |= ((adev->ras_features & BIT(AMDGPU_RAS_BLOCK__UMC)) != 0) ?
1413
+		HSA_CAP_MEM_EDCSUPPORTED : 0;
1414
+
1415
+	if (adev->asic_type != CHIP_VEGA10)
1416
+		dev->node_props.capability |= (adev->ras_features != 0) ?
1417
+			HSA_CAP_RASEVENTNOTIFY : 0;
12921418 
12931419 	kfd_debug_print_topology();
12941420 
....@@ -1360,7 +1486,6 @@
13601486 
13611487 static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask)
13621488 {
1363
-	const struct cpuinfo_x86 *cpuinfo;
13641489 	int first_cpu_of_numa_node;
13651490 
13661491 	if (!cpumask || cpumask == cpu_none_mask)
....@@ -1368,9 +1493,11 @@
13681493 	first_cpu_of_numa_node = cpumask_first(cpumask);
13691494 	if (first_cpu_of_numa_node >= nr_cpu_ids)
13701495 		return -1;
1371
-	cpuinfo = &cpu_data(first_cpu_of_numa_node);
1372
-
1373
-	return cpuinfo->apicid;
1496
+#ifdef CONFIG_X86_64
1497
+	return cpu_data(first_cpu_of_numa_node).apicid;
1498
+#else
1499
+	return first_cpu_of_numa_node;
1500
+#endif
13741501 }
13751502 
13761503 /* kfd_numa_node_to_apic_id - Returns the APIC ID of the first logical processor
....@@ -1386,6 +1513,29 @@
13861513 	return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id));
13871514 }
13881515 
1516
+void kfd_double_confirm_iommu_support(struct kfd_dev *gpu)
1517
+{
1518
+	struct kfd_topology_device *dev;
1519
+
1520
+	gpu->use_iommu_v2 = false;
1521
+
1522
+	if (!gpu->device_info->needs_iommu_device)
1523
+		return;
1524
+
1525
+	down_read(&topology_lock);
1526
+
1527
+	/* Only use IOMMUv2 if there is an APU topology node with no GPU
1528
+	 * assigned yet. This GPU will be assigned to it.
1529
+	 */
1530
+	list_for_each_entry(dev, &topology_device_list, list)
1531
+		if (dev->node_props.cpu_cores_count &&
1532
+		    dev->node_props.simd_count &&
1533
+		    !dev->gpu)
1534
+			gpu->use_iommu_v2 = true;
1535
+
1536
+	up_read(&topology_lock);
1537
+}
1538
+
13891539 #if defined(CONFIG_DEBUG_FS)
13901540 
13911541 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data)