Linux Audio

Check our new training course

Loading...
v4.6
   1/*
   2 * Copyright 2014 Advanced Micro Devices, Inc.
   3 *
   4 * Permission is hereby granted, free of charge, to any person obtaining a
   5 * copy of this software and associated documentation files (the "Software"),
   6 * to deal in the Software without restriction, including without limitation
   7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8 * and/or sell copies of the Software, and to permit persons to whom the
   9 * Software is furnished to do so, subject to the following conditions:
  10 *
  11 * The above copyright notice and this permission notice shall be included in
  12 * all copies or substantial portions of the Software.
  13 *
  14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20 * OTHER DEALINGS IN THE SOFTWARE.
  21 */
  22
  23#include <linux/types.h>
  24#include <linux/kernel.h>
  25#include <linux/pci.h>
  26#include <linux/errno.h>
  27#include <linux/acpi.h>
  28#include <linux/hash.h>
  29#include <linux/cpufreq.h>
  30#include <linux/log2.h>
 
 
  31
  32#include "kfd_priv.h"
  33#include "kfd_crat.h"
  34#include "kfd_topology.h"
 
 
 
 
  35
 
  36static struct list_head topology_device_list;
  37static int topology_crat_parsed;
  38static struct kfd_system_properties sys_props;
  39
  40static DECLARE_RWSEM(topology_lock);
 
  41
  42struct kfd_dev *kfd_device_by_id(uint32_t gpu_id)
 
  43{
  44	struct kfd_topology_device *top_dev;
  45	struct kfd_dev *device = NULL;
  46
  47	down_read(&topology_lock);
  48
  49	list_for_each_entry(top_dev, &topology_device_list, list)
  50		if (top_dev->gpu_id == gpu_id) {
  51			device = top_dev->gpu;
  52			break;
  53		}
  54
  55	up_read(&topology_lock);
  56
  57	return device;
  58}
  59
  60struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev)
  61{
  62	struct kfd_topology_device *top_dev;
  63	struct kfd_dev *device = NULL;
  64
  65	down_read(&topology_lock);
  66
  67	list_for_each_entry(top_dev, &topology_device_list, list)
  68		if (top_dev->gpu->pdev == pdev) {
  69			device = top_dev->gpu;
  70			break;
  71		}
  72
  73	up_read(&topology_lock);
  74
  75	return device;
  76}
  77
  78static int kfd_topology_get_crat_acpi(void *crat_image, size_t *size)
  79{
  80	struct acpi_table_header *crat_table;
  81	acpi_status status;
  82
  83	if (!size)
  84		return -EINVAL;
  85
  86	/*
  87	 * Fetch the CRAT table from ACPI
  88	 */
  89	status = acpi_get_table(CRAT_SIGNATURE, 0, &crat_table);
  90	if (status == AE_NOT_FOUND) {
  91		pr_warn("CRAT table not found\n");
  92		return -ENODATA;
  93	} else if (ACPI_FAILURE(status)) {
  94		const char *err = acpi_format_exception(status);
  95
  96		pr_err("CRAT table error: %s\n", err);
  97		return -EINVAL;
  98	}
  99
 100	if (*size >= crat_table->length && crat_image != NULL)
 101		memcpy(crat_image, crat_table, crat_table->length);
 102
 103	*size = crat_table->length;
 104
 105	return 0;
 106}
 107
 108static void kfd_populated_cu_info_cpu(struct kfd_topology_device *dev,
 109		struct crat_subtype_computeunit *cu)
 110{
 111	BUG_ON(!dev);
 112	BUG_ON(!cu);
 113
 114	dev->node_props.cpu_cores_count = cu->num_cpu_cores;
 115	dev->node_props.cpu_core_id_base = cu->processor_id_low;
 116	if (cu->hsa_capability & CRAT_CU_FLAGS_IOMMU_PRESENT)
 117		dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
 118
 119	pr_info("CU CPU: cores=%d id_base=%d\n", cu->num_cpu_cores,
 120			cu->processor_id_low);
 121}
 122
 123static void kfd_populated_cu_info_gpu(struct kfd_topology_device *dev,
 124		struct crat_subtype_computeunit *cu)
 125{
 126	BUG_ON(!dev);
 127	BUG_ON(!cu);
 128
 129	dev->node_props.simd_id_base = cu->processor_id_low;
 130	dev->node_props.simd_count = cu->num_simd_cores;
 131	dev->node_props.lds_size_in_kb = cu->lds_size_in_kb;
 132	dev->node_props.max_waves_per_simd = cu->max_waves_simd;
 133	dev->node_props.wave_front_size = cu->wave_front_size;
 134	dev->node_props.mem_banks_count = cu->num_banks;
 135	dev->node_props.array_count = cu->num_arrays;
 136	dev->node_props.cu_per_simd_array = cu->num_cu_per_array;
 137	dev->node_props.simd_per_cu = cu->num_simd_per_cu;
 138	dev->node_props.max_slots_scratch_cu = cu->max_slots_scatch_cu;
 139	if (cu->hsa_capability & CRAT_CU_FLAGS_HOT_PLUGGABLE)
 140		dev->node_props.capability |= HSA_CAP_HOT_PLUGGABLE;
 141	pr_info("CU GPU: simds=%d id_base=%d\n", cu->num_simd_cores,
 142				cu->processor_id_low);
 143}
 144
 145/* kfd_parse_subtype_cu is called when the topology mutex is already acquired */
 146static int kfd_parse_subtype_cu(struct crat_subtype_computeunit *cu)
 147{
 148	struct kfd_topology_device *dev;
 149	int i = 0;
 150
 151	BUG_ON(!cu);
 152
 153	pr_info("Found CU entry in CRAT table with proximity_domain=%d caps=%x\n",
 154			cu->proximity_domain, cu->hsa_capability);
 155	list_for_each_entry(dev, &topology_device_list, list) {
 156		if (cu->proximity_domain == i) {
 157			if (cu->flags & CRAT_CU_FLAGS_CPU_PRESENT)
 158				kfd_populated_cu_info_cpu(dev, cu);
 159
 160			if (cu->flags & CRAT_CU_FLAGS_GPU_PRESENT)
 161				kfd_populated_cu_info_gpu(dev, cu);
 
 162			break;
 163		}
 164		i++;
 165	}
 166
 167	return 0;
 168}
 169
 170/*
 171 * kfd_parse_subtype_mem is called when the topology mutex is
 172 * already acquired
 173 */
 174static int kfd_parse_subtype_mem(struct crat_subtype_memory *mem)
 175{
 176	struct kfd_mem_properties *props;
 177	struct kfd_topology_device *dev;
 178	int i = 0;
 179
 180	BUG_ON(!mem);
 181
 182	pr_info("Found memory entry in CRAT table with proximity_domain=%d\n",
 183			mem->promixity_domain);
 184	list_for_each_entry(dev, &topology_device_list, list) {
 185		if (mem->promixity_domain == i) {
 186			props = kfd_alloc_struct(props);
 187			if (props == NULL)
 188				return -ENOMEM;
 189
 190			if (dev->node_props.cpu_cores_count == 0)
 191				props->heap_type = HSA_MEM_HEAP_TYPE_FB_PRIVATE;
 192			else
 193				props->heap_type = HSA_MEM_HEAP_TYPE_SYSTEM;
 194
 195			if (mem->flags & CRAT_MEM_FLAGS_HOT_PLUGGABLE)
 196				props->flags |= HSA_MEM_FLAGS_HOT_PLUGGABLE;
 197			if (mem->flags & CRAT_MEM_FLAGS_NON_VOLATILE)
 198				props->flags |= HSA_MEM_FLAGS_NON_VOLATILE;
 199
 200			props->size_in_bytes =
 201				((uint64_t)mem->length_high << 32) +
 202							mem->length_low;
 203			props->width = mem->width;
 204
 205			dev->mem_bank_count++;
 206			list_add_tail(&props->list, &dev->mem_props);
 207
 208			break;
 209		}
 210		i++;
 211	}
 212
 213	return 0;
 214}
 215
 216/*
 217 * kfd_parse_subtype_cache is called when the topology mutex
 218 * is already acquired
 219 */
 220static int kfd_parse_subtype_cache(struct crat_subtype_cache *cache)
 221{
 222	struct kfd_cache_properties *props;
 223	struct kfd_topology_device *dev;
 224	uint32_t id;
 225
 226	BUG_ON(!cache);
 227
 228	id = cache->processor_id_low;
 229
 230	pr_info("Found cache entry in CRAT table with processor_id=%d\n", id);
 231	list_for_each_entry(dev, &topology_device_list, list)
 232		if (id == dev->node_props.cpu_core_id_base ||
 233		    id == dev->node_props.simd_id_base) {
 234			props = kfd_alloc_struct(props);
 235			if (props == NULL)
 236				return -ENOMEM;
 237
 238			props->processor_id_low = id;
 239			props->cache_level = cache->cache_level;
 240			props->cache_size = cache->cache_size;
 241			props->cacheline_size = cache->cache_line_size;
 242			props->cachelines_per_tag = cache->lines_per_tag;
 243			props->cache_assoc = cache->associativity;
 244			props->cache_latency = cache->cache_latency;
 245
 246			if (cache->flags & CRAT_CACHE_FLAGS_DATA_CACHE)
 247				props->cache_type |= HSA_CACHE_TYPE_DATA;
 248			if (cache->flags & CRAT_CACHE_FLAGS_INST_CACHE)
 249				props->cache_type |= HSA_CACHE_TYPE_INSTRUCTION;
 250			if (cache->flags & CRAT_CACHE_FLAGS_CPU_CACHE)
 251				props->cache_type |= HSA_CACHE_TYPE_CPU;
 252			if (cache->flags & CRAT_CACHE_FLAGS_SIMD_CACHE)
 253				props->cache_type |= HSA_CACHE_TYPE_HSACU;
 254
 255			dev->cache_count++;
 256			dev->node_props.caches_count++;
 257			list_add_tail(&props->list, &dev->cache_props);
 258
 
 
 
 259			break;
 260		}
 261
 262	return 0;
 263}
 264
 265/*
 266 * kfd_parse_subtype_iolink is called when the topology mutex
 267 * is already acquired
 268 */
 269static int kfd_parse_subtype_iolink(struct crat_subtype_iolink *iolink)
 270{
 271	struct kfd_iolink_properties *props;
 272	struct kfd_topology_device *dev;
 273	uint32_t i = 0;
 274	uint32_t id_from;
 275	uint32_t id_to;
 276
 277	BUG_ON(!iolink);
 278
 279	id_from = iolink->proximity_domain_from;
 280	id_to = iolink->proximity_domain_to;
 281
 282	pr_info("Found IO link entry in CRAT table with id_from=%d\n", id_from);
 283	list_for_each_entry(dev, &topology_device_list, list) {
 284		if (id_from == i) {
 285			props = kfd_alloc_struct(props);
 286			if (props == NULL)
 287				return -ENOMEM;
 288
 289			props->node_from = id_from;
 290			props->node_to = id_to;
 291			props->ver_maj = iolink->version_major;
 292			props->ver_min = iolink->version_minor;
 293
 294			/*
 295			 * weight factor (derived from CDIR), currently always 1
 296			 */
 297			props->weight = 1;
 298
 299			props->min_latency = iolink->minimum_latency;
 300			props->max_latency = iolink->maximum_latency;
 301			props->min_bandwidth = iolink->minimum_bandwidth_mbs;
 302			props->max_bandwidth = iolink->maximum_bandwidth_mbs;
 303			props->rec_transfer_size =
 304					iolink->recommended_transfer_size;
 305
 306			dev->io_link_count++;
 307			dev->node_props.io_links_count++;
 308			list_add_tail(&props->list, &dev->io_link_props);
 309
 310			break;
 311		}
 312		i++;
 313	}
 314
 315	return 0;
 316}
 317
 318static int kfd_parse_subtype(struct crat_subtype_generic *sub_type_hdr)
 319{
 320	struct crat_subtype_computeunit *cu;
 321	struct crat_subtype_memory *mem;
 322	struct crat_subtype_cache *cache;
 323	struct crat_subtype_iolink *iolink;
 324	int ret = 0;
 325
 326	BUG_ON(!sub_type_hdr);
 327
 328	switch (sub_type_hdr->type) {
 329	case CRAT_SUBTYPE_COMPUTEUNIT_AFFINITY:
 330		cu = (struct crat_subtype_computeunit *)sub_type_hdr;
 331		ret = kfd_parse_subtype_cu(cu);
 332		break;
 333	case CRAT_SUBTYPE_MEMORY_AFFINITY:
 334		mem = (struct crat_subtype_memory *)sub_type_hdr;
 335		ret = kfd_parse_subtype_mem(mem);
 336		break;
 337	case CRAT_SUBTYPE_CACHE_AFFINITY:
 338		cache = (struct crat_subtype_cache *)sub_type_hdr;
 339		ret = kfd_parse_subtype_cache(cache);
 340		break;
 341	case CRAT_SUBTYPE_TLB_AFFINITY:
 342		/*
 343		 * For now, nothing to do here
 344		 */
 345		pr_info("Found TLB entry in CRAT table (not processing)\n");
 346		break;
 347	case CRAT_SUBTYPE_CCOMPUTE_AFFINITY:
 348		/*
 349		 * For now, nothing to do here
 350		 */
 351		pr_info("Found CCOMPUTE entry in CRAT table (not processing)\n");
 352		break;
 353	case CRAT_SUBTYPE_IOLINK_AFFINITY:
 354		iolink = (struct crat_subtype_iolink *)sub_type_hdr;
 355		ret = kfd_parse_subtype_iolink(iolink);
 356		break;
 357	default:
 358		pr_warn("Unknown subtype (%d) in CRAT\n",
 359				sub_type_hdr->type);
 360	}
 361
 362	return ret;
 363}
 364
 
 365static void kfd_release_topology_device(struct kfd_topology_device *dev)
 366{
 367	struct kfd_mem_properties *mem;
 368	struct kfd_cache_properties *cache;
 369	struct kfd_iolink_properties *iolink;
 370
 371	BUG_ON(!dev);
 372
 373	list_del(&dev->list);
 374
 375	while (dev->mem_props.next != &dev->mem_props) {
 376		mem = container_of(dev->mem_props.next,
 377				struct kfd_mem_properties, list);
 378		list_del(&mem->list);
 379		kfree(mem);
 380	}
 381
 382	while (dev->cache_props.next != &dev->cache_props) {
 383		cache = container_of(dev->cache_props.next,
 384				struct kfd_cache_properties, list);
 385		list_del(&cache->list);
 386		kfree(cache);
 387	}
 388
 389	while (dev->io_link_props.next != &dev->io_link_props) {
 390		iolink = container_of(dev->io_link_props.next,
 391				struct kfd_iolink_properties, list);
 392		list_del(&iolink->list);
 393		kfree(iolink);
 394	}
 395
 396	kfree(dev);
 
 
 
 
 
 397
 398	sys_props.num_devices--;
 399}
 400
 401static void kfd_release_live_view(void)
 402{
 403	struct kfd_topology_device *dev;
 404
 405	while (topology_device_list.next != &topology_device_list) {
 406		dev = container_of(topology_device_list.next,
 407				 struct kfd_topology_device, list);
 408		kfd_release_topology_device(dev);
 
 409}
 410
 
 
 
 411	memset(&sys_props, 0, sizeof(sys_props));
 412}
 413
 414static struct kfd_topology_device *kfd_create_topology_device(void)
 
 415{
 416	struct kfd_topology_device *dev;
 417
 418	dev = kfd_alloc_struct(dev);
 419	if (dev == NULL) {
 420		pr_err("No memory to allocate a topology device");
 421		return NULL;
 422	}
 423
 424	INIT_LIST_HEAD(&dev->mem_props);
 425	INIT_LIST_HEAD(&dev->cache_props);
 426	INIT_LIST_HEAD(&dev->io_link_props);
 
 427
 428	list_add_tail(&dev->list, &topology_device_list);
 429	sys_props.num_devices++;
 430
 431	return dev;
 432}
 433
 434static int kfd_parse_crat_table(void *crat_image)
 435{
 436	struct kfd_topology_device *top_dev;
 437	struct crat_subtype_generic *sub_type_hdr;
 438	uint16_t node_id;
 439	int ret;
 440	struct crat_header *crat_table = (struct crat_header *)crat_image;
 441	uint16_t num_nodes;
 442	uint32_t image_len;
 443
 444	if (!crat_image)
 445		return -EINVAL;
 446
 447	num_nodes = crat_table->num_domains;
 448	image_len = crat_table->length;
 449
 450	pr_info("Parsing CRAT table with %d nodes\n", num_nodes);
 451
 452	for (node_id = 0; node_id < num_nodes; node_id++) {
 453		top_dev = kfd_create_topology_device();
 454		if (!top_dev) {
 455			kfd_release_live_view();
 456			return -ENOMEM;
 457		}
 458	}
 459
 460	sys_props.platform_id =
 461		(*((uint64_t *)crat_table->oem_id)) & CRAT_OEMID_64BIT_MASK;
 462	sys_props.platform_oem = *((uint64_t *)crat_table->oem_table_id);
 463	sys_props.platform_rev = crat_table->revision;
 464
 465	sub_type_hdr = (struct crat_subtype_generic *)(crat_table+1);
 466	while ((char *)sub_type_hdr + sizeof(struct crat_subtype_generic) <
 467			((char *)crat_image) + image_len) {
 468		if (sub_type_hdr->flags & CRAT_SUBTYPE_FLAGS_ENABLED) {
 469			ret = kfd_parse_subtype(sub_type_hdr);
 470			if (ret != 0) {
 471				kfd_release_live_view();
 472				return ret;
 473			}
 474		}
 475
 476		sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
 477				sub_type_hdr->length);
 478	}
 479
 480	sys_props.generation_count++;
 481	topology_crat_parsed = 1;
 482
 483	return 0;
 484}
 485
 486
 487#define sysfs_show_gen_prop(buffer, fmt, ...) \
 488		snprintf(buffer, PAGE_SIZE, "%s"fmt, buffer, __VA_ARGS__)
 489#define sysfs_show_32bit_prop(buffer, name, value) \
 490		sysfs_show_gen_prop(buffer, "%s %u\n", name, value)
 491#define sysfs_show_64bit_prop(buffer, name, value) \
 492		sysfs_show_gen_prop(buffer, "%s %llu\n", name, value)
 493#define sysfs_show_32bit_val(buffer, value) \
 494		sysfs_show_gen_prop(buffer, "%u\n", value)
 495#define sysfs_show_str_val(buffer, value) \
 496		sysfs_show_gen_prop(buffer, "%s\n", value)
 497
 498static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr,
 499		char *buffer)
 500{
 501	ssize_t ret;
 502
 503	/* Making sure that the buffer is an empty string */
 504	buffer[0] = 0;
 505
 506	if (attr == &sys_props.attr_genid) {
 507		ret = sysfs_show_32bit_val(buffer, sys_props.generation_count);
 508	} else if (attr == &sys_props.attr_props) {
 509		sysfs_show_64bit_prop(buffer, "platform_oem",
 510				sys_props.platform_oem);
 511		sysfs_show_64bit_prop(buffer, "platform_id",
 512				sys_props.platform_id);
 513		ret = sysfs_show_64bit_prop(buffer, "platform_rev",
 514				sys_props.platform_rev);
 515	} else {
 516		ret = -EINVAL;
 517	}
 518
 519	return ret;
 520}
 521
 
 
 
 
 
 522static const struct sysfs_ops sysprops_ops = {
 523	.show = sysprops_show,
 524};
 525
 526static struct kobj_type sysprops_type = {
 
 527	.sysfs_ops = &sysprops_ops,
 528};
 529
 530static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr,
 531		char *buffer)
 532{
 533	ssize_t ret;
 534	struct kfd_iolink_properties *iolink;
 535
 536	/* Making sure that the buffer is an empty string */
 537	buffer[0] = 0;
 538
 539	iolink = container_of(attr, struct kfd_iolink_properties, attr);
 540	sysfs_show_32bit_prop(buffer, "type", iolink->iolink_type);
 541	sysfs_show_32bit_prop(buffer, "version_major", iolink->ver_maj);
 542	sysfs_show_32bit_prop(buffer, "version_minor", iolink->ver_min);
 543	sysfs_show_32bit_prop(buffer, "node_from", iolink->node_from);
 544	sysfs_show_32bit_prop(buffer, "node_to", iolink->node_to);
 545	sysfs_show_32bit_prop(buffer, "weight", iolink->weight);
 546	sysfs_show_32bit_prop(buffer, "min_latency", iolink->min_latency);
 547	sysfs_show_32bit_prop(buffer, "max_latency", iolink->max_latency);
 548	sysfs_show_32bit_prop(buffer, "min_bandwidth", iolink->min_bandwidth);
 549	sysfs_show_32bit_prop(buffer, "max_bandwidth", iolink->max_bandwidth);
 550	sysfs_show_32bit_prop(buffer, "recommended_transfer_size",
 551			iolink->rec_transfer_size);
 552	ret = sysfs_show_32bit_prop(buffer, "flags", iolink->flags);
 553
 554	return ret;
 555}
 556
 557static const struct sysfs_ops iolink_ops = {
 558	.show = iolink_show,
 559};
 560
 561static struct kobj_type iolink_type = {
 
 562	.sysfs_ops = &iolink_ops,
 563};
 564
 565static ssize_t mem_show(struct kobject *kobj, struct attribute *attr,
 566		char *buffer)
 567{
 568	ssize_t ret;
 569	struct kfd_mem_properties *mem;
 570
 571	/* Making sure that the buffer is an empty string */
 572	buffer[0] = 0;
 573
 574	mem = container_of(attr, struct kfd_mem_properties, attr);
 575	sysfs_show_32bit_prop(buffer, "heap_type", mem->heap_type);
 576	sysfs_show_64bit_prop(buffer, "size_in_bytes", mem->size_in_bytes);
 577	sysfs_show_32bit_prop(buffer, "flags", mem->flags);
 578	sysfs_show_32bit_prop(buffer, "width", mem->width);
 579	ret = sysfs_show_32bit_prop(buffer, "mem_clk_max", mem->mem_clk_max);
 580
 581	return ret;
 582}
 583
 584static const struct sysfs_ops mem_ops = {
 585	.show = mem_show,
 586};
 587
 588static struct kobj_type mem_type = {
 
 589	.sysfs_ops = &mem_ops,
 590};
 591
 592static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr,
 593		char *buffer)
 594{
 595	ssize_t ret;
 596	uint32_t i;
 597	struct kfd_cache_properties *cache;
 598
 599	/* Making sure that the buffer is an empty string */
 600	buffer[0] = 0;
 601
 602	cache = container_of(attr, struct kfd_cache_properties, attr);
 603	sysfs_show_32bit_prop(buffer, "processor_id_low",
 604			cache->processor_id_low);
 605	sysfs_show_32bit_prop(buffer, "level", cache->cache_level);
 606	sysfs_show_32bit_prop(buffer, "size", cache->cache_size);
 607	sysfs_show_32bit_prop(buffer, "cache_line_size", cache->cacheline_size);
 608	sysfs_show_32bit_prop(buffer, "cache_lines_per_tag",
 609			cache->cachelines_per_tag);
 610	sysfs_show_32bit_prop(buffer, "association", cache->cache_assoc);
 611	sysfs_show_32bit_prop(buffer, "latency", cache->cache_latency);
 612	sysfs_show_32bit_prop(buffer, "type", cache->cache_type);
 613	snprintf(buffer, PAGE_SIZE, "%ssibling_map ", buffer);
 614	for (i = 0; i < KFD_TOPOLOGY_CPU_SIBLINGS; i++)
 615		ret = snprintf(buffer, PAGE_SIZE, "%s%d%s",
 616				buffer, cache->sibling_map[i],
 617				(i == KFD_TOPOLOGY_CPU_SIBLINGS-1) ?
 618						"\n" : ",");
 619
 
 
 
 
 
 
 620	return ret;
 621}
 622
 623static const struct sysfs_ops cache_ops = {
 624	.show = kfd_cache_show,
 625};
 626
 627static struct kobj_type cache_type = {
 
 628	.sysfs_ops = &cache_ops,
 629};
 630
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 631static ssize_t node_show(struct kobject *kobj, struct attribute *attr,
 632		char *buffer)
 633{
 634	struct kfd_topology_device *dev;
 635	char public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE];
 636	uint32_t i;
 637	uint32_t log_max_watch_addr;
 638
 639	/* Making sure that the buffer is an empty string */
 640	buffer[0] = 0;
 641
 642	if (strcmp(attr->name, "gpu_id") == 0) {
 643		dev = container_of(attr, struct kfd_topology_device,
 644				attr_gpuid);
 645		return sysfs_show_32bit_val(buffer, dev->gpu_id);
 646	}
 647
 648	if (strcmp(attr->name, "name") == 0) {
 649		dev = container_of(attr, struct kfd_topology_device,
 650				attr_name);
 651		for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE; i++) {
 652			public_name[i] =
 653					(char)dev->node_props.marketing_name[i];
 654			if (dev->node_props.marketing_name[i] == 0)
 655				break;
 656		}
 657		public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1] = 0x0;
 658		return sysfs_show_str_val(buffer, public_name);
 659	}
 660
 661	dev = container_of(attr, struct kfd_topology_device,
 662			attr_props);
 663	sysfs_show_32bit_prop(buffer, "cpu_cores_count",
 664			dev->node_props.cpu_cores_count);
 665	sysfs_show_32bit_prop(buffer, "simd_count",
 666			dev->node_props.simd_count);
 667
 668	if (dev->mem_bank_count < dev->node_props.mem_banks_count) {
 669		pr_warn("kfd: mem_banks_count truncated from %d to %d\n",
 670				dev->node_props.mem_banks_count,
 671				dev->mem_bank_count);
 672		sysfs_show_32bit_prop(buffer, "mem_banks_count",
 673				dev->mem_bank_count);
 674	} else {
 675		sysfs_show_32bit_prop(buffer, "mem_banks_count",
 676				dev->node_props.mem_banks_count);
 677	}
 678
 679	sysfs_show_32bit_prop(buffer, "caches_count",
 680			dev->node_props.caches_count);
 681	sysfs_show_32bit_prop(buffer, "io_links_count",
 682			dev->node_props.io_links_count);
 683	sysfs_show_32bit_prop(buffer, "cpu_core_id_base",
 684			dev->node_props.cpu_core_id_base);
 685	sysfs_show_32bit_prop(buffer, "simd_id_base",
 686			dev->node_props.simd_id_base);
 687	sysfs_show_32bit_prop(buffer, "max_waves_per_simd",
 688			dev->node_props.max_waves_per_simd);
 689	sysfs_show_32bit_prop(buffer, "lds_size_in_kb",
 690			dev->node_props.lds_size_in_kb);
 691	sysfs_show_32bit_prop(buffer, "gds_size_in_kb",
 692			dev->node_props.gds_size_in_kb);
 
 
 693	sysfs_show_32bit_prop(buffer, "wave_front_size",
 694			dev->node_props.wave_front_size);
 695	sysfs_show_32bit_prop(buffer, "array_count",
 696			dev->node_props.array_count);
 697	sysfs_show_32bit_prop(buffer, "simd_arrays_per_engine",
 698			dev->node_props.simd_arrays_per_engine);
 699	sysfs_show_32bit_prop(buffer, "cu_per_simd_array",
 700			dev->node_props.cu_per_simd_array);
 701	sysfs_show_32bit_prop(buffer, "simd_per_cu",
 702			dev->node_props.simd_per_cu);
 703	sysfs_show_32bit_prop(buffer, "max_slots_scratch_cu",
 704			dev->node_props.max_slots_scratch_cu);
 705	sysfs_show_32bit_prop(buffer, "vendor_id",
 706			dev->node_props.vendor_id);
 707	sysfs_show_32bit_prop(buffer, "device_id",
 708			dev->node_props.device_id);
 709	sysfs_show_32bit_prop(buffer, "location_id",
 710			dev->node_props.location_id);
 
 
 
 
 
 
 
 
 711
 712	if (dev->gpu) {
 713		log_max_watch_addr =
 714			__ilog2_u32(dev->gpu->device_info->num_of_watch_points);
 715
 716		if (log_max_watch_addr) {
 717			dev->node_props.capability |=
 718					HSA_CAP_WATCH_POINTS_SUPPORTED;
 719
 720			dev->node_props.capability |=
 721				((log_max_watch_addr <<
 722					HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) &
 723				HSA_CAP_WATCH_POINTS_TOTALBITS_MASK);
 724		}
 725
 
 
 
 
 726		sysfs_show_32bit_prop(buffer, "max_engine_clk_fcompute",
 727			dev->gpu->kfd2kgd->get_max_engine_clock_in_mhz(
 728					dev->gpu->kgd));
 729
 730		sysfs_show_64bit_prop(buffer, "local_mem_size",
 731				(unsigned long long int) 0);
 732
 733		sysfs_show_32bit_prop(buffer, "fw_version",
 734			dev->gpu->kfd2kgd->get_fw_version(
 735						dev->gpu->kgd,
 736						KGD_ENGINE_MEC1));
 737		sysfs_show_32bit_prop(buffer, "capability",
 738				dev->node_props.capability);
 
 
 739	}
 740
 741	return sysfs_show_32bit_prop(buffer, "max_engine_clk_ccompute",
 742					cpufreq_quick_get_max(0)/1000);
 743}
 744
 745static const struct sysfs_ops node_ops = {
 746	.show = node_show,
 747};
 748
 749static struct kobj_type node_type = {
 
 750	.sysfs_ops = &node_ops,
 751};
 752
 753static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr)
 754{
 755	sysfs_remove_file(kobj, attr);
 756	kobject_del(kobj);
 757	kobject_put(kobj);
 758}
 759
 760static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev)
 761{
 762	struct kfd_iolink_properties *iolink;
 763	struct kfd_cache_properties *cache;
 764	struct kfd_mem_properties *mem;
 765
 766	BUG_ON(!dev);
 767
 768	if (dev->kobj_iolink) {
 769		list_for_each_entry(iolink, &dev->io_link_props, list)
 770			if (iolink->kobj) {
 771				kfd_remove_sysfs_file(iolink->kobj,
 772							&iolink->attr);
 773				iolink->kobj = NULL;
 774			}
 775		kobject_del(dev->kobj_iolink);
 776		kobject_put(dev->kobj_iolink);
 777		dev->kobj_iolink = NULL;
 778	}
 779
 780	if (dev->kobj_cache) {
 781		list_for_each_entry(cache, &dev->cache_props, list)
 782			if (cache->kobj) {
 783				kfd_remove_sysfs_file(cache->kobj,
 784							&cache->attr);
 785				cache->kobj = NULL;
 786			}
 787		kobject_del(dev->kobj_cache);
 788		kobject_put(dev->kobj_cache);
 789		dev->kobj_cache = NULL;
 790	}
 791
 792	if (dev->kobj_mem) {
 793		list_for_each_entry(mem, &dev->mem_props, list)
 794			if (mem->kobj) {
 795				kfd_remove_sysfs_file(mem->kobj, &mem->attr);
 796				mem->kobj = NULL;
 797			}
 798		kobject_del(dev->kobj_mem);
 799		kobject_put(dev->kobj_mem);
 800		dev->kobj_mem = NULL;
 801	}
 802
 
 
 
 
 
 
 
 
 
 
 803	if (dev->kobj_node) {
 804		sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid);
 805		sysfs_remove_file(dev->kobj_node, &dev->attr_name);
 806		sysfs_remove_file(dev->kobj_node, &dev->attr_props);
 807		kobject_del(dev->kobj_node);
 808		kobject_put(dev->kobj_node);
 809		dev->kobj_node = NULL;
 810	}
 811}
 812
 813static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev,
 814		uint32_t id)
 815{
 816	struct kfd_iolink_properties *iolink;
 817	struct kfd_cache_properties *cache;
 818	struct kfd_mem_properties *mem;
 
 819	int ret;
 820	uint32_t i;
 
 821
 822	BUG_ON(!dev);
 
 823
 824	/*
 825	 * Creating the sysfs folders
 826	 */
 827	BUG_ON(dev->kobj_node);
 828	dev->kobj_node = kfd_alloc_struct(dev->kobj_node);
 829	if (!dev->kobj_node)
 830		return -ENOMEM;
 831
 832	ret = kobject_init_and_add(dev->kobj_node, &node_type,
 833			sys_props.kobj_nodes, "%d", id);
 834	if (ret < 0)
 835		return ret;
 836
 837	dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node);
 838	if (!dev->kobj_mem)
 839		return -ENOMEM;
 840
 841	dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node);
 842	if (!dev->kobj_cache)
 843		return -ENOMEM;
 844
 845	dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node);
 846	if (!dev->kobj_iolink)
 847		return -ENOMEM;
 848
 
 
 
 
 849	/*
 850	 * Creating sysfs files for node properties
 851	 */
 852	dev->attr_gpuid.name = "gpu_id";
 853	dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE;
 854	sysfs_attr_init(&dev->attr_gpuid);
 855	dev->attr_name.name = "name";
 856	dev->attr_name.mode = KFD_SYSFS_FILE_MODE;
 857	sysfs_attr_init(&dev->attr_name);
 858	dev->attr_props.name = "properties";
 859	dev->attr_props.mode = KFD_SYSFS_FILE_MODE;
 860	sysfs_attr_init(&dev->attr_props);
 861	ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid);
 862	if (ret < 0)
 863		return ret;
 864	ret = sysfs_create_file(dev->kobj_node, &dev->attr_name);
 865	if (ret < 0)
 866		return ret;
 867	ret = sysfs_create_file(dev->kobj_node, &dev->attr_props);
 868	if (ret < 0)
 869		return ret;
 870
 871	i = 0;
 872	list_for_each_entry(mem, &dev->mem_props, list) {
 873		mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
 874		if (!mem->kobj)
 875			return -ENOMEM;
 876		ret = kobject_init_and_add(mem->kobj, &mem_type,
 877				dev->kobj_mem, "%d", i);
 878		if (ret < 0)
 879			return ret;
 880
 881		mem->attr.name = "properties";
 882		mem->attr.mode = KFD_SYSFS_FILE_MODE;
 883		sysfs_attr_init(&mem->attr);
 884		ret = sysfs_create_file(mem->kobj, &mem->attr);
 885		if (ret < 0)
 886			return ret;
 887		i++;
 888	}
 889
 890	i = 0;
 891	list_for_each_entry(cache, &dev->cache_props, list) {
 892		cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
 893		if (!cache->kobj)
 894			return -ENOMEM;
 895		ret = kobject_init_and_add(cache->kobj, &cache_type,
 896				dev->kobj_cache, "%d", i);
 897		if (ret < 0)
 898			return ret;
 899
 900		cache->attr.name = "properties";
 901		cache->attr.mode = KFD_SYSFS_FILE_MODE;
 902		sysfs_attr_init(&cache->attr);
 903		ret = sysfs_create_file(cache->kobj, &cache->attr);
 904		if (ret < 0)
 905			return ret;
 906		i++;
 907	}
 908
 909	i = 0;
 910	list_for_each_entry(iolink, &dev->io_link_props, list) {
 911		iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
 912		if (!iolink->kobj)
 913			return -ENOMEM;
 914		ret = kobject_init_and_add(iolink->kobj, &iolink_type,
 915				dev->kobj_iolink, "%d", i);
 916		if (ret < 0)
 917			return ret;
 918
 919		iolink->attr.name = "properties";
 920		iolink->attr.mode = KFD_SYSFS_FILE_MODE;
 921		sysfs_attr_init(&iolink->attr);
 922		ret = sysfs_create_file(iolink->kobj, &iolink->attr);
 923		if (ret < 0)
 924			return ret;
 925		i++;
 926}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 927
 928	return 0;
 929}
 930
 
 931static int kfd_build_sysfs_node_tree(void)
 932{
 933	struct kfd_topology_device *dev;
 934	int ret;
 935	uint32_t i = 0;
 936
 937	list_for_each_entry(dev, &topology_device_list, list) {
 938		ret = kfd_build_sysfs_node_entry(dev, i);
 939		if (ret < 0)
 940			return ret;
 941		i++;
 942	}
 943
 944	return 0;
 945}
 946
 
 947static void kfd_remove_sysfs_node_tree(void)
 948{
 949	struct kfd_topology_device *dev;
 950
 951	list_for_each_entry(dev, &topology_device_list, list)
 952		kfd_remove_sysfs_node_entry(dev);
 953}
 954
 955static int kfd_topology_update_sysfs(void)
 956{
 957	int ret;
 958
 959	pr_info("Creating topology SYSFS entries\n");
 960	if (sys_props.kobj_topology == NULL) {
 961		sys_props.kobj_topology =
 962				kfd_alloc_struct(sys_props.kobj_topology);
 963		if (!sys_props.kobj_topology)
 964			return -ENOMEM;
 965
 966		ret = kobject_init_and_add(sys_props.kobj_topology,
 967				&sysprops_type,  &kfd_device->kobj,
 968				"topology");
 969		if (ret < 0)
 970			return ret;
 971
 972		sys_props.kobj_nodes = kobject_create_and_add("nodes",
 973				sys_props.kobj_topology);
 974		if (!sys_props.kobj_nodes)
 975			return -ENOMEM;
 976
 977		sys_props.attr_genid.name = "generation_id";
 978		sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE;
 979		sysfs_attr_init(&sys_props.attr_genid);
 980		ret = sysfs_create_file(sys_props.kobj_topology,
 981				&sys_props.attr_genid);
 982		if (ret < 0)
 983			return ret;
 984
 985		sys_props.attr_props.name = "system_properties";
 986		sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE;
 987		sysfs_attr_init(&sys_props.attr_props);
 988		ret = sysfs_create_file(sys_props.kobj_topology,
 989				&sys_props.attr_props);
 990		if (ret < 0)
 991			return ret;
 992	}
 993
 994	kfd_remove_sysfs_node_tree();
 995
 996	return kfd_build_sysfs_node_tree();
 997}
 998
 999static void kfd_topology_release_sysfs(void)
1000{
1001	kfd_remove_sysfs_node_tree();
1002	if (sys_props.kobj_topology) {
1003		sysfs_remove_file(sys_props.kobj_topology,
1004				&sys_props.attr_genid);
1005		sysfs_remove_file(sys_props.kobj_topology,
1006				&sys_props.attr_props);
1007		if (sys_props.kobj_nodes) {
1008			kobject_del(sys_props.kobj_nodes);
1009			kobject_put(sys_props.kobj_nodes);
1010			sys_props.kobj_nodes = NULL;
1011		}
1012		kobject_del(sys_props.kobj_topology);
1013		kobject_put(sys_props.kobj_topology);
1014		sys_props.kobj_topology = NULL;
1015	}
1016}
1017
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1018int kfd_topology_init(void)
1019{
1020	void *crat_image = NULL;
1021	size_t image_size = 0;
1022	int ret;
1023
1024	/*
1025	 * Initialize the head for the topology device list
 
 
 
 
 
 
1026	 */
 
 
1027	INIT_LIST_HEAD(&topology_device_list);
 
1028	init_rwsem(&topology_lock);
1029	topology_crat_parsed = 0;
1030
1031	memset(&sys_props, 0, sizeof(sys_props));
1032
 
 
 
 
 
 
 
1033	/*
1034	 * Get the CRAT image from the ACPI
 
 
 
1035	 */
1036	ret = kfd_topology_get_crat_acpi(crat_image, &image_size);
1037	if (ret == 0 && image_size > 0) {
1038		pr_info("Found CRAT image with size=%zd\n", image_size);
1039		crat_image = kmalloc(image_size, GFP_KERNEL);
1040		if (!crat_image) {
1041			ret = -ENOMEM;
1042			pr_err("No memory for allocating CRAT image\n");
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1043			goto err;
1044		}
1045		ret = kfd_topology_get_crat_acpi(crat_image, &image_size);
1046
1047		if (ret == 0) {
1048			down_write(&topology_lock);
1049			ret = kfd_parse_crat_table(crat_image);
1050			if (ret == 0)
1051				ret = kfd_topology_update_sysfs();
1052			up_write(&topology_lock);
1053		} else {
1054			pr_err("Couldn't get CRAT table size from ACPI\n");
1055		}
1056		kfree(crat_image);
1057	} else if (ret == -ENODATA) {
1058		ret = 0;
1059	} else {
1060		pr_err("Couldn't get CRAT table size from ACPI\n");
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1061	}
1062
1063err:
1064	pr_info("Finished initializing topology ret=%d\n", ret);
1065	return ret;
1066}
1067
1068void kfd_topology_shutdown(void)
1069{
 
1070	kfd_topology_release_sysfs();
1071	kfd_release_live_view();
1072}
1073
1074static void kfd_debug_print_topology(void)
1075{
1076	struct kfd_topology_device *dev;
1077	uint32_t i = 0;
1078
1079	pr_info("DEBUG PRINT OF TOPOLOGY:");
1080	list_for_each_entry(dev, &topology_device_list, list) {
1081		pr_info("Node: %d\n", i);
1082		pr_info("\tGPU assigned: %s\n", (dev->gpu ? "yes" : "no"));
1083		pr_info("\tCPU count: %d\n", dev->node_props.cpu_cores_count);
1084		pr_info("\tSIMD count: %d", dev->node_props.simd_count);
1085		i++;
1086	}
1087}
1088
1089static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu)
1090{
1091	uint32_t hashout;
1092	uint32_t buf[7];
 
1093	int i;
 
1094
1095	if (!gpu)
1096		return 0;
1097
 
 
 
 
 
1098	buf[0] = gpu->pdev->devfn;
1099	buf[1] = gpu->pdev->subsystem_vendor;
1100	buf[2] = gpu->pdev->subsystem_device;
 
1101	buf[3] = gpu->pdev->device;
1102	buf[4] = gpu->pdev->bus->number;
1103	buf[5] = (uint32_t)(gpu->kfd2kgd->get_vmem_size(gpu->kgd)
1104			& 0xffffffff);
1105	buf[6] = (uint32_t)(gpu->kfd2kgd->get_vmem_size(gpu->kgd) >> 32);
1106
1107	for (i = 0, hashout = 0; i < 7; i++)
1108		hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH);
1109
1110	return hashout;
1111}
1112
 
 
 
 
1113static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu)
1114{
1115	struct kfd_topology_device *dev;
1116	struct kfd_topology_device *out_dev = NULL;
1117
1118	BUG_ON(!gpu);
 
 
 
 
 
 
 
1119
1120	list_for_each_entry(dev, &topology_device_list, list)
1121		if (dev->gpu == NULL && dev->node_props.simd_count > 0) {
1122			dev->gpu = gpu;
1123			out_dev = dev;
1124			break;
1125		}
1126
 
1127	return out_dev;
1128}
1129
1130static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival)
1131{
1132	/*
1133	 * TODO: Generate an event for thunk about the arrival/removal
1134	 * of the GPU
1135	 */
1136}
1137
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1138int kfd_topology_add_device(struct kfd_dev *gpu)
1139{
1140	uint32_t gpu_id;
1141	struct kfd_topology_device *dev;
1142	int res;
 
 
 
 
 
 
1143
1144	BUG_ON(!gpu);
1145
1146	gpu_id = kfd_generate_gpu_id(gpu);
1147
1148	pr_debug("kfd: Adding new GPU (ID: 0x%x) to topology\n", gpu_id);
1149
1150	down_write(&topology_lock);
1151	/*
1152	 * Try to assign the GPU to existing topology device (generated from
1153	 * CRAT table
 
 
 
1154	 */
1155	dev = kfd_assign_gpu(gpu);
1156	if (!dev) {
1157		pr_info("GPU was not found in the current topology. Extending.\n");
1158		kfd_debug_print_topology();
1159		dev = kfd_create_topology_device();
1160		if (!dev) {
1161			res = -ENOMEM;
 
 
 
 
 
 
 
 
 
1162			goto err;
1163		}
1164		dev->gpu = gpu;
1165
1166		/*
1167		 * TODO: Make a call to retrieve topology information from the
1168		 * GPU vBIOS
1169		 */
1170
1171		/*
1172		 * Update the SYSFS tree, since we added another topology device
1173		 */
1174		if (kfd_topology_update_sysfs() < 0)
1175			kfd_topology_release_sysfs();
1176
 
 
 
 
 
 
 
 
 
 
1177	}
1178
1179	dev->gpu_id = gpu_id;
1180	gpu->id = gpu_id;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1181	dev->node_props.vendor_id = gpu->pdev->vendor;
1182	dev->node_props.device_id = gpu->pdev->device;
1183	dev->node_props.location_id = (gpu->pdev->bus->number << 24) +
1184			(gpu->pdev->devfn & 0xffffff);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1185	/*
1186	 * TODO: Retrieve max engine clock values from KGD
1187	 */
 
 
 
 
 
1188
 
 
 
 
 
1189	if (dev->gpu->device_info->asic_family == CHIP_CARRIZO) {
1190		dev->node_props.capability |= HSA_CAP_DOORBELL_PACKET_TYPE;
1191		pr_info("amdkfd: adding doorbell packet type capability\n");
 
1192	}
1193
1194	res = 0;
 
 
 
 
 
 
 
 
1195
1196err:
1197	up_write(&topology_lock);
 
1198
1199	if (res == 0)
1200		kfd_notify_gpu_change(gpu_id, 1);
1201
 
 
 
 
1202	return res;
1203}
1204
1205int kfd_topology_remove_device(struct kfd_dev *gpu)
1206{
1207	struct kfd_topology_device *dev;
1208	uint32_t gpu_id;
1209	int res = -ENODEV;
1210
1211	BUG_ON(!gpu);
1212
1213	down_write(&topology_lock);
1214
1215	list_for_each_entry(dev, &topology_device_list, list)
1216		if (dev->gpu == gpu) {
1217			gpu_id = dev->gpu_id;
1218			kfd_remove_sysfs_node_entry(dev);
1219			kfd_release_topology_device(dev);
 
1220			res = 0;
1221			if (kfd_topology_update_sysfs() < 0)
1222				kfd_topology_release_sysfs();
1223			break;
1224		}
1225
1226	up_write(&topology_lock);
1227
1228	if (res == 0)
1229		kfd_notify_gpu_change(gpu_id, 0);
1230
1231	return res;
1232}
1233
1234/*
1235 * When idx is out of bounds, the function will return NULL
 
 
 
1236 */
1237struct kfd_dev *kfd_topology_enum_kfd_devices(uint8_t idx)
1238{
1239
1240	struct kfd_topology_device *top_dev;
1241	struct kfd_dev *device = NULL;
1242	uint8_t device_idx = 0;
1243
 
1244	down_read(&topology_lock);
1245
1246	list_for_each_entry(top_dev, &topology_device_list, list) {
1247		if (device_idx == idx) {
1248			device = top_dev->gpu;
1249			break;
 
1250		}
1251
1252		device_idx++;
1253	}
1254
1255	up_read(&topology_lock);
1256
1257	return device;
 
 
1258
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1259}
v5.4
   1/*
   2 * Copyright 2014 Advanced Micro Devices, Inc.
   3 *
   4 * Permission is hereby granted, free of charge, to any person obtaining a
   5 * copy of this software and associated documentation files (the "Software"),
   6 * to deal in the Software without restriction, including without limitation
   7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8 * and/or sell copies of the Software, and to permit persons to whom the
   9 * Software is furnished to do so, subject to the following conditions:
  10 *
  11 * The above copyright notice and this permission notice shall be included in
  12 * all copies or substantial portions of the Software.
  13 *
  14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20 * OTHER DEALINGS IN THE SOFTWARE.
  21 */
  22
  23#include <linux/types.h>
  24#include <linux/kernel.h>
  25#include <linux/pci.h>
  26#include <linux/errno.h>
  27#include <linux/acpi.h>
  28#include <linux/hash.h>
  29#include <linux/cpufreq.h>
  30#include <linux/log2.h>
  31#include <linux/dmi.h>
  32#include <linux/atomic.h>
  33
  34#include "kfd_priv.h"
  35#include "kfd_crat.h"
  36#include "kfd_topology.h"
  37#include "kfd_device_queue_manager.h"
  38#include "kfd_iommu.h"
  39#include "amdgpu_amdkfd.h"
  40#include "amdgpu_ras.h"
  41
  42/* topology_device_list - Master list of all topology devices */
  43static struct list_head topology_device_list;
 
  44static struct kfd_system_properties sys_props;
  45
  46static DECLARE_RWSEM(topology_lock);
  47static atomic_t topology_crat_proximity_domain;
  48
  49struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
  50						uint32_t proximity_domain)
  51{
  52	struct kfd_topology_device *top_dev;
  53	struct kfd_topology_device *device = NULL;
  54
  55	down_read(&topology_lock);
  56
  57	list_for_each_entry(top_dev, &topology_device_list, list)
  58		if (top_dev->proximity_domain == proximity_domain) {
  59			device = top_dev;
  60			break;
  61		}
  62
  63	up_read(&topology_lock);
  64
  65	return device;
  66}
  67
  68struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id)
  69{
  70	struct kfd_topology_device *top_dev = NULL;
  71	struct kfd_topology_device *ret = NULL;
  72
  73	down_read(&topology_lock);
  74
  75	list_for_each_entry(top_dev, &topology_device_list, list)
  76		if (top_dev->gpu_id == gpu_id) {
  77			ret = top_dev;
  78			break;
  79		}
  80
  81	up_read(&topology_lock);
  82
  83	return ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  84}
  85
  86struct kfd_dev *kfd_device_by_id(uint32_t gpu_id)
 
  87{
  88	struct kfd_topology_device *top_dev;
 
 
 
 
 
 
  89
  90	top_dev = kfd_topology_device_by_id(gpu_id);
  91	if (!top_dev)
  92		return NULL;
  93
  94	return top_dev->gpu;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  95}
  96
  97struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev)
 
  98{
  99	struct kfd_topology_device *top_dev;
 100	struct kfd_dev *device = NULL;
 
 
 101
 102	down_read(&topology_lock);
 
 
 
 
 
 103
 104	list_for_each_entry(top_dev, &topology_device_list, list)
 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
 115struct 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) {
 124			device = top_dev->gpu;
 125			break;
 126		}
 127
 128	up_read(&topology_lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 129
 130	return device;
 131}
 132
 133/* Called with write topology_lock acquired */
 134static void kfd_release_topology_device(struct kfd_topology_device *dev)
 135{
 136	struct kfd_mem_properties *mem;
 137	struct kfd_cache_properties *cache;
 138	struct kfd_iolink_properties *iolink;
 139	struct kfd_perf_properties *perf;
 
 140
 141	list_del(&dev->list);
 142
 143	while (dev->mem_props.next != &dev->mem_props) {
 144		mem = container_of(dev->mem_props.next,
 145				struct kfd_mem_properties, list);
 146		list_del(&mem->list);
 147		kfree(mem);
 148	}
 149
 150	while (dev->cache_props.next != &dev->cache_props) {
 151		cache = container_of(dev->cache_props.next,
 152				struct kfd_cache_properties, list);
 153		list_del(&cache->list);
 154		kfree(cache);
 155	}
 156
 157	while (dev->io_link_props.next != &dev->io_link_props) {
 158		iolink = container_of(dev->io_link_props.next,
 159				struct kfd_iolink_properties, list);
 160		list_del(&iolink->list);
 161		kfree(iolink);
 162	}
 163
 164	while (dev->perf_props.next != &dev->perf_props) {
 165		perf = container_of(dev->perf_props.next,
 166				struct kfd_perf_properties, list);
 167		list_del(&perf->list);
 168		kfree(perf);
 169	}
 170
 171	kfree(dev);
 172}
 173
 174void kfd_release_topology_device_list(struct list_head *device_list)
 175{
 176	struct kfd_topology_device *dev;
 177
 178	while (!list_empty(device_list)) {
 179		dev = list_first_entry(device_list,
 180				       struct kfd_topology_device, list);
 181		kfd_release_topology_device(dev);
 182	}
 183}
 184
 185static void kfd_release_live_view(void)
 186{
 187	kfd_release_topology_device_list(&topology_device_list);
 188	memset(&sys_props, 0, sizeof(sys_props));
 189}
 190
 191struct kfd_topology_device *kfd_create_topology_device(
 192				struct list_head *device_list)
 193{
 194	struct kfd_topology_device *dev;
 195
 196	dev = kfd_alloc_struct(dev);
 197	if (!dev) {
 198		pr_err("No memory to allocate a topology device");
 199		return NULL;
 200	}
 201
 202	INIT_LIST_HEAD(&dev->mem_props);
 203	INIT_LIST_HEAD(&dev->cache_props);
 204	INIT_LIST_HEAD(&dev->io_link_props);
 205	INIT_LIST_HEAD(&dev->perf_props);
 206
 207	list_add_tail(&dev->list, device_list);
 
 208
 209	return dev;
 210}
 211
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 212
 213#define sysfs_show_gen_prop(buffer, fmt, ...) \
 214		snprintf(buffer, PAGE_SIZE, "%s"fmt, buffer, __VA_ARGS__)
 215#define sysfs_show_32bit_prop(buffer, name, value) \
 216		sysfs_show_gen_prop(buffer, "%s %u\n", name, value)
 217#define sysfs_show_64bit_prop(buffer, name, value) \
 218		sysfs_show_gen_prop(buffer, "%s %llu\n", name, value)
 219#define sysfs_show_32bit_val(buffer, value) \
 220		sysfs_show_gen_prop(buffer, "%u\n", value)
 221#define sysfs_show_str_val(buffer, value) \
 222		sysfs_show_gen_prop(buffer, "%s\n", value)
 223
 224static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr,
 225		char *buffer)
 226{
 227	ssize_t ret;
 228
 229	/* Making sure that the buffer is an empty string */
 230	buffer[0] = 0;
 231
 232	if (attr == &sys_props.attr_genid) {
 233		ret = sysfs_show_32bit_val(buffer, sys_props.generation_count);
 234	} else if (attr == &sys_props.attr_props) {
 235		sysfs_show_64bit_prop(buffer, "platform_oem",
 236				sys_props.platform_oem);
 237		sysfs_show_64bit_prop(buffer, "platform_id",
 238				sys_props.platform_id);
 239		ret = sysfs_show_64bit_prop(buffer, "platform_rev",
 240				sys_props.platform_rev);
 241	} else {
 242		ret = -EINVAL;
 243	}
 244
 245	return ret;
 246}
 247
 248static void kfd_topology_kobj_release(struct kobject *kobj)
 249{
 250	kfree(kobj);
 251}
 252
 253static const struct sysfs_ops sysprops_ops = {
 254	.show = sysprops_show,
 255};
 256
 257static struct kobj_type sysprops_type = {
 258	.release = kfd_topology_kobj_release,
 259	.sysfs_ops = &sysprops_ops,
 260};
 261
 262static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr,
 263		char *buffer)
 264{
 265	ssize_t ret;
 266	struct kfd_iolink_properties *iolink;
 267
 268	/* Making sure that the buffer is an empty string */
 269	buffer[0] = 0;
 270
 271	iolink = container_of(attr, struct kfd_iolink_properties, attr);
 272	sysfs_show_32bit_prop(buffer, "type", iolink->iolink_type);
 273	sysfs_show_32bit_prop(buffer, "version_major", iolink->ver_maj);
 274	sysfs_show_32bit_prop(buffer, "version_minor", iolink->ver_min);
 275	sysfs_show_32bit_prop(buffer, "node_from", iolink->node_from);
 276	sysfs_show_32bit_prop(buffer, "node_to", iolink->node_to);
 277	sysfs_show_32bit_prop(buffer, "weight", iolink->weight);
 278	sysfs_show_32bit_prop(buffer, "min_latency", iolink->min_latency);
 279	sysfs_show_32bit_prop(buffer, "max_latency", iolink->max_latency);
 280	sysfs_show_32bit_prop(buffer, "min_bandwidth", iolink->min_bandwidth);
 281	sysfs_show_32bit_prop(buffer, "max_bandwidth", iolink->max_bandwidth);
 282	sysfs_show_32bit_prop(buffer, "recommended_transfer_size",
 283			iolink->rec_transfer_size);
 284	ret = sysfs_show_32bit_prop(buffer, "flags", iolink->flags);
 285
 286	return ret;
 287}
 288
 289static const struct sysfs_ops iolink_ops = {
 290	.show = iolink_show,
 291};
 292
 293static struct kobj_type iolink_type = {
 294	.release = kfd_topology_kobj_release,
 295	.sysfs_ops = &iolink_ops,
 296};
 297
 298static ssize_t mem_show(struct kobject *kobj, struct attribute *attr,
 299		char *buffer)
 300{
 301	ssize_t ret;
 302	struct kfd_mem_properties *mem;
 303
 304	/* Making sure that the buffer is an empty string */
 305	buffer[0] = 0;
 306
 307	mem = container_of(attr, struct kfd_mem_properties, attr);
 308	sysfs_show_32bit_prop(buffer, "heap_type", mem->heap_type);
 309	sysfs_show_64bit_prop(buffer, "size_in_bytes", mem->size_in_bytes);
 310	sysfs_show_32bit_prop(buffer, "flags", mem->flags);
 311	sysfs_show_32bit_prop(buffer, "width", mem->width);
 312	ret = sysfs_show_32bit_prop(buffer, "mem_clk_max", mem->mem_clk_max);
 313
 314	return ret;
 315}
 316
 317static const struct sysfs_ops mem_ops = {
 318	.show = mem_show,
 319};
 320
 321static struct kobj_type mem_type = {
 322	.release = kfd_topology_kobj_release,
 323	.sysfs_ops = &mem_ops,
 324};
 325
 326static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr,
 327		char *buffer)
 328{
 329	ssize_t ret;
 330	uint32_t i, j;
 331	struct kfd_cache_properties *cache;
 332
 333	/* Making sure that the buffer is an empty string */
 334	buffer[0] = 0;
 335
 336	cache = container_of(attr, struct kfd_cache_properties, attr);
 337	sysfs_show_32bit_prop(buffer, "processor_id_low",
 338			cache->processor_id_low);
 339	sysfs_show_32bit_prop(buffer, "level", cache->cache_level);
 340	sysfs_show_32bit_prop(buffer, "size", cache->cache_size);
 341	sysfs_show_32bit_prop(buffer, "cache_line_size", cache->cacheline_size);
 342	sysfs_show_32bit_prop(buffer, "cache_lines_per_tag",
 343			cache->cachelines_per_tag);
 344	sysfs_show_32bit_prop(buffer, "association", cache->cache_assoc);
 345	sysfs_show_32bit_prop(buffer, "latency", cache->cache_latency);
 346	sysfs_show_32bit_prop(buffer, "type", cache->cache_type);
 347	snprintf(buffer, PAGE_SIZE, "%ssibling_map ", buffer);
 348	for (i = 0; i < CRAT_SIBLINGMAP_SIZE; i++)
 349		for (j = 0; j < sizeof(cache->sibling_map[0])*8; j++) {
 350			/* Check each bit */
 351			if (cache->sibling_map[i] & (1 << j))
 352				ret = snprintf(buffer, PAGE_SIZE,
 353					 "%s%d%s", buffer, 1, ",");
 354			else
 355				ret = snprintf(buffer, PAGE_SIZE,
 356					 "%s%d%s", buffer, 0, ",");
 357		}
 358	/* Replace the last "," with end of line */
 359	*(buffer + strlen(buffer) - 1) = 0xA;
 360	return ret;
 361}
 362
 363static const struct sysfs_ops cache_ops = {
 364	.show = kfd_cache_show,
 365};
 366
 367static struct kobj_type cache_type = {
 368	.release = kfd_topology_kobj_release,
 369	.sysfs_ops = &cache_ops,
 370};
 371
 372/****** Sysfs of Performance Counters ******/
 373
 374struct kfd_perf_attr {
 375	struct kobj_attribute attr;
 376	uint32_t data;
 377};
 378
 379static ssize_t perf_show(struct kobject *kobj, struct kobj_attribute *attrs,
 380			char *buf)
 381{
 382	struct kfd_perf_attr *attr;
 383
 384	buf[0] = 0;
 385	attr = container_of(attrs, struct kfd_perf_attr, attr);
 386	if (!attr->data) /* invalid data for PMC */
 387		return 0;
 388	else
 389		return sysfs_show_32bit_val(buf, attr->data);
 390}
 391
 392#define KFD_PERF_DESC(_name, _data)			\
 393{							\
 394	.attr  = __ATTR(_name, 0444, perf_show, NULL),	\
 395	.data = _data,					\
 396}
 397
 398static struct kfd_perf_attr perf_attr_iommu[] = {
 399	KFD_PERF_DESC(max_concurrent, 0),
 400	KFD_PERF_DESC(num_counters, 0),
 401	KFD_PERF_DESC(counter_ids, 0),
 402};
 403/****************************************/
 404
 405static ssize_t node_show(struct kobject *kobj, struct attribute *attr,
 406		char *buffer)
 407{
 408	struct kfd_topology_device *dev;
 
 
 409	uint32_t log_max_watch_addr;
 410
 411	/* Making sure that the buffer is an empty string */
 412	buffer[0] = 0;
 413
 414	if (strcmp(attr->name, "gpu_id") == 0) {
 415		dev = container_of(attr, struct kfd_topology_device,
 416				attr_gpuid);
 417		return sysfs_show_32bit_val(buffer, dev->gpu_id);
 418	}
 419
 420	if (strcmp(attr->name, "name") == 0) {
 421		dev = container_of(attr, struct kfd_topology_device,
 422				attr_name);
 423
 424		return sysfs_show_str_val(buffer, dev->node_props.name);
 
 
 
 
 
 
 425	}
 426
 427	dev = container_of(attr, struct kfd_topology_device,
 428			attr_props);
 429	sysfs_show_32bit_prop(buffer, "cpu_cores_count",
 430			dev->node_props.cpu_cores_count);
 431	sysfs_show_32bit_prop(buffer, "simd_count",
 432			dev->node_props.simd_count);
 433	sysfs_show_32bit_prop(buffer, "mem_banks_count",
 434			dev->node_props.mem_banks_count);
 
 
 
 
 
 
 
 
 
 
 435	sysfs_show_32bit_prop(buffer, "caches_count",
 436			dev->node_props.caches_count);
 437	sysfs_show_32bit_prop(buffer, "io_links_count",
 438			dev->node_props.io_links_count);
 439	sysfs_show_32bit_prop(buffer, "cpu_core_id_base",
 440			dev->node_props.cpu_core_id_base);
 441	sysfs_show_32bit_prop(buffer, "simd_id_base",
 442			dev->node_props.simd_id_base);
 443	sysfs_show_32bit_prop(buffer, "max_waves_per_simd",
 444			dev->node_props.max_waves_per_simd);
 445	sysfs_show_32bit_prop(buffer, "lds_size_in_kb",
 446			dev->node_props.lds_size_in_kb);
 447	sysfs_show_32bit_prop(buffer, "gds_size_in_kb",
 448			dev->node_props.gds_size_in_kb);
 449	sysfs_show_32bit_prop(buffer, "num_gws",
 450			dev->node_props.num_gws);
 451	sysfs_show_32bit_prop(buffer, "wave_front_size",
 452			dev->node_props.wave_front_size);
 453	sysfs_show_32bit_prop(buffer, "array_count",
 454			dev->node_props.array_count);
 455	sysfs_show_32bit_prop(buffer, "simd_arrays_per_engine",
 456			dev->node_props.simd_arrays_per_engine);
 457	sysfs_show_32bit_prop(buffer, "cu_per_simd_array",
 458			dev->node_props.cu_per_simd_array);
 459	sysfs_show_32bit_prop(buffer, "simd_per_cu",
 460			dev->node_props.simd_per_cu);
 461	sysfs_show_32bit_prop(buffer, "max_slots_scratch_cu",
 462			dev->node_props.max_slots_scratch_cu);
 463	sysfs_show_32bit_prop(buffer, "vendor_id",
 464			dev->node_props.vendor_id);
 465	sysfs_show_32bit_prop(buffer, "device_id",
 466			dev->node_props.device_id);
 467	sysfs_show_32bit_prop(buffer, "location_id",
 468			dev->node_props.location_id);
 469	sysfs_show_32bit_prop(buffer, "drm_render_minor",
 470			dev->node_props.drm_render_minor);
 471	sysfs_show_64bit_prop(buffer, "hive_id",
 472			dev->node_props.hive_id);
 473	sysfs_show_32bit_prop(buffer, "num_sdma_engines",
 474			dev->node_props.num_sdma_engines);
 475	sysfs_show_32bit_prop(buffer, "num_sdma_xgmi_engines",
 476			dev->node_props.num_sdma_xgmi_engines);
 477
 478	if (dev->gpu) {
 479		log_max_watch_addr =
 480			__ilog2_u32(dev->gpu->device_info->num_of_watch_points);
 481
 482		if (log_max_watch_addr) {
 483			dev->node_props.capability |=
 484					HSA_CAP_WATCH_POINTS_SUPPORTED;
 485
 486			dev->node_props.capability |=
 487				((log_max_watch_addr <<
 488					HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) &
 489				HSA_CAP_WATCH_POINTS_TOTALBITS_MASK);
 490		}
 491
 492		if (dev->gpu->device_info->asic_family == CHIP_TONGA)
 493			dev->node_props.capability |=
 494					HSA_CAP_AQL_QUEUE_DOUBLE_MAP;
 495
 496		sysfs_show_32bit_prop(buffer, "max_engine_clk_fcompute",
 497			dev->node_props.max_engine_clk_fcompute);
 
 498
 499		sysfs_show_64bit_prop(buffer, "local_mem_size",
 500				(unsigned long long int) 0);
 501
 502		sysfs_show_32bit_prop(buffer, "fw_version",
 503				dev->gpu->mec_fw_version);
 
 
 504		sysfs_show_32bit_prop(buffer, "capability",
 505				dev->node_props.capability);
 506		sysfs_show_32bit_prop(buffer, "sdma_fw_version",
 507				dev->gpu->sdma_fw_version);
 508	}
 509
 510	return sysfs_show_32bit_prop(buffer, "max_engine_clk_ccompute",
 511					cpufreq_quick_get_max(0)/1000);
 512}
 513
 514static const struct sysfs_ops node_ops = {
 515	.show = node_show,
 516};
 517
 518static struct kobj_type node_type = {
 519	.release = kfd_topology_kobj_release,
 520	.sysfs_ops = &node_ops,
 521};
 522
 523static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr)
 524{
 525	sysfs_remove_file(kobj, attr);
 526	kobject_del(kobj);
 527	kobject_put(kobj);
 528}
 529
 530static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev)
 531{
 532	struct kfd_iolink_properties *iolink;
 533	struct kfd_cache_properties *cache;
 534	struct kfd_mem_properties *mem;
 535	struct kfd_perf_properties *perf;
 
 536
 537	if (dev->kobj_iolink) {
 538		list_for_each_entry(iolink, &dev->io_link_props, list)
 539			if (iolink->kobj) {
 540				kfd_remove_sysfs_file(iolink->kobj,
 541							&iolink->attr);
 542				iolink->kobj = NULL;
 543			}
 544		kobject_del(dev->kobj_iolink);
 545		kobject_put(dev->kobj_iolink);
 546		dev->kobj_iolink = NULL;
 547	}
 548
 549	if (dev->kobj_cache) {
 550		list_for_each_entry(cache, &dev->cache_props, list)
 551			if (cache->kobj) {
 552				kfd_remove_sysfs_file(cache->kobj,
 553							&cache->attr);
 554				cache->kobj = NULL;
 555			}
 556		kobject_del(dev->kobj_cache);
 557		kobject_put(dev->kobj_cache);
 558		dev->kobj_cache = NULL;
 559	}
 560
 561	if (dev->kobj_mem) {
 562		list_for_each_entry(mem, &dev->mem_props, list)
 563			if (mem->kobj) {
 564				kfd_remove_sysfs_file(mem->kobj, &mem->attr);
 565				mem->kobj = NULL;
 566			}
 567		kobject_del(dev->kobj_mem);
 568		kobject_put(dev->kobj_mem);
 569		dev->kobj_mem = NULL;
 570	}
 571
 572	if (dev->kobj_perf) {
 573		list_for_each_entry(perf, &dev->perf_props, list) {
 574			kfree(perf->attr_group);
 575			perf->attr_group = NULL;
 576		}
 577		kobject_del(dev->kobj_perf);
 578		kobject_put(dev->kobj_perf);
 579		dev->kobj_perf = NULL;
 580	}
 581
 582	if (dev->kobj_node) {
 583		sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid);
 584		sysfs_remove_file(dev->kobj_node, &dev->attr_name);
 585		sysfs_remove_file(dev->kobj_node, &dev->attr_props);
 586		kobject_del(dev->kobj_node);
 587		kobject_put(dev->kobj_node);
 588		dev->kobj_node = NULL;
 589	}
 590}
 591
 592static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev,
 593		uint32_t id)
 594{
 595	struct kfd_iolink_properties *iolink;
 596	struct kfd_cache_properties *cache;
 597	struct kfd_mem_properties *mem;
 598	struct kfd_perf_properties *perf;
 599	int ret;
 600	uint32_t i, num_attrs;
 601	struct attribute **attrs;
 602
 603	if (WARN_ON(dev->kobj_node))
 604		return -EEXIST;
 605
 606	/*
 607	 * Creating the sysfs folders
 608	 */
 
 609	dev->kobj_node = kfd_alloc_struct(dev->kobj_node);
 610	if (!dev->kobj_node)
 611		return -ENOMEM;
 612
 613	ret = kobject_init_and_add(dev->kobj_node, &node_type,
 614			sys_props.kobj_nodes, "%d", id);
 615	if (ret < 0)
 616		return ret;
 617
 618	dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node);
 619	if (!dev->kobj_mem)
 620		return -ENOMEM;
 621
 622	dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node);
 623	if (!dev->kobj_cache)
 624		return -ENOMEM;
 625
 626	dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node);
 627	if (!dev->kobj_iolink)
 628		return -ENOMEM;
 629
 630	dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node);
 631	if (!dev->kobj_perf)
 632		return -ENOMEM;
 633
 634	/*
 635	 * Creating sysfs files for node properties
 636	 */
 637	dev->attr_gpuid.name = "gpu_id";
 638	dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE;
 639	sysfs_attr_init(&dev->attr_gpuid);
 640	dev->attr_name.name = "name";
 641	dev->attr_name.mode = KFD_SYSFS_FILE_MODE;
 642	sysfs_attr_init(&dev->attr_name);
 643	dev->attr_props.name = "properties";
 644	dev->attr_props.mode = KFD_SYSFS_FILE_MODE;
 645	sysfs_attr_init(&dev->attr_props);
 646	ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid);
 647	if (ret < 0)
 648		return ret;
 649	ret = sysfs_create_file(dev->kobj_node, &dev->attr_name);
 650	if (ret < 0)
 651		return ret;
 652	ret = sysfs_create_file(dev->kobj_node, &dev->attr_props);
 653	if (ret < 0)
 654		return ret;
 655
 656	i = 0;
 657	list_for_each_entry(mem, &dev->mem_props, list) {
 658		mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
 659		if (!mem->kobj)
 660			return -ENOMEM;
 661		ret = kobject_init_and_add(mem->kobj, &mem_type,
 662				dev->kobj_mem, "%d", i);
 663		if (ret < 0)
 664			return ret;
 665
 666		mem->attr.name = "properties";
 667		mem->attr.mode = KFD_SYSFS_FILE_MODE;
 668		sysfs_attr_init(&mem->attr);
 669		ret = sysfs_create_file(mem->kobj, &mem->attr);
 670		if (ret < 0)
 671			return ret;
 672		i++;
 673	}
 674
 675	i = 0;
 676	list_for_each_entry(cache, &dev->cache_props, list) {
 677		cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
 678		if (!cache->kobj)
 679			return -ENOMEM;
 680		ret = kobject_init_and_add(cache->kobj, &cache_type,
 681				dev->kobj_cache, "%d", i);
 682		if (ret < 0)
 683			return ret;
 684
 685		cache->attr.name = "properties";
 686		cache->attr.mode = KFD_SYSFS_FILE_MODE;
 687		sysfs_attr_init(&cache->attr);
 688		ret = sysfs_create_file(cache->kobj, &cache->attr);
 689		if (ret < 0)
 690			return ret;
 691		i++;
 692	}
 693
 694	i = 0;
 695	list_for_each_entry(iolink, &dev->io_link_props, list) {
 696		iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
 697		if (!iolink->kobj)
 698			return -ENOMEM;
 699		ret = kobject_init_and_add(iolink->kobj, &iolink_type,
 700				dev->kobj_iolink, "%d", i);
 701		if (ret < 0)
 702			return ret;
 703
 704		iolink->attr.name = "properties";
 705		iolink->attr.mode = KFD_SYSFS_FILE_MODE;
 706		sysfs_attr_init(&iolink->attr);
 707		ret = sysfs_create_file(iolink->kobj, &iolink->attr);
 708		if (ret < 0)
 709			return ret;
 710		i++;
 711	}
 712
 713	/* All hardware blocks have the same number of attributes. */
 714	num_attrs = ARRAY_SIZE(perf_attr_iommu);
 715	list_for_each_entry(perf, &dev->perf_props, list) {
 716		perf->attr_group = kzalloc(sizeof(struct kfd_perf_attr)
 717			* num_attrs + sizeof(struct attribute_group),
 718			GFP_KERNEL);
 719		if (!perf->attr_group)
 720			return -ENOMEM;
 721
 722		attrs = (struct attribute **)(perf->attr_group + 1);
 723		if (!strcmp(perf->block_name, "iommu")) {
 724		/* Information of IOMMU's num_counters and counter_ids is shown
 725		 * under /sys/bus/event_source/devices/amd_iommu. We don't
 726		 * duplicate here.
 727		 */
 728			perf_attr_iommu[0].data = perf->max_concurrent;
 729			for (i = 0; i < num_attrs; i++)
 730				attrs[i] = &perf_attr_iommu[i].attr.attr;
 731		}
 732		perf->attr_group->name = perf->block_name;
 733		perf->attr_group->attrs = attrs;
 734		ret = sysfs_create_group(dev->kobj_perf, perf->attr_group);
 735		if (ret < 0)
 736			return ret;
 737	}
 738
 739	return 0;
 740}
 741
 742/* Called with write topology lock acquired */
 743static int kfd_build_sysfs_node_tree(void)
 744{
 745	struct kfd_topology_device *dev;
 746	int ret;
 747	uint32_t i = 0;
 748
 749	list_for_each_entry(dev, &topology_device_list, list) {
 750		ret = kfd_build_sysfs_node_entry(dev, i);
 751		if (ret < 0)
 752			return ret;
 753		i++;
 754	}
 755
 756	return 0;
 757}
 758
 759/* Called with write topology lock acquired */
 760static void kfd_remove_sysfs_node_tree(void)
 761{
 762	struct kfd_topology_device *dev;
 763
 764	list_for_each_entry(dev, &topology_device_list, list)
 765		kfd_remove_sysfs_node_entry(dev);
 766}
 767
 768static int kfd_topology_update_sysfs(void)
 769{
 770	int ret;
 771
 772	pr_info("Creating topology SYSFS entries\n");
 773	if (!sys_props.kobj_topology) {
 774		sys_props.kobj_topology =
 775				kfd_alloc_struct(sys_props.kobj_topology);
 776		if (!sys_props.kobj_topology)
 777			return -ENOMEM;
 778
 779		ret = kobject_init_and_add(sys_props.kobj_topology,
 780				&sysprops_type,  &kfd_device->kobj,
 781				"topology");
 782		if (ret < 0)
 783			return ret;
 784
 785		sys_props.kobj_nodes = kobject_create_and_add("nodes",
 786				sys_props.kobj_topology);
 787		if (!sys_props.kobj_nodes)
 788			return -ENOMEM;
 789
 790		sys_props.attr_genid.name = "generation_id";
 791		sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE;
 792		sysfs_attr_init(&sys_props.attr_genid);
 793		ret = sysfs_create_file(sys_props.kobj_topology,
 794				&sys_props.attr_genid);
 795		if (ret < 0)
 796			return ret;
 797
 798		sys_props.attr_props.name = "system_properties";
 799		sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE;
 800		sysfs_attr_init(&sys_props.attr_props);
 801		ret = sysfs_create_file(sys_props.kobj_topology,
 802				&sys_props.attr_props);
 803		if (ret < 0)
 804			return ret;
 805	}
 806
 807	kfd_remove_sysfs_node_tree();
 808
 809	return kfd_build_sysfs_node_tree();
 810}
 811
 812static void kfd_topology_release_sysfs(void)
 813{
 814	kfd_remove_sysfs_node_tree();
 815	if (sys_props.kobj_topology) {
 816		sysfs_remove_file(sys_props.kobj_topology,
 817				&sys_props.attr_genid);
 818		sysfs_remove_file(sys_props.kobj_topology,
 819				&sys_props.attr_props);
 820		if (sys_props.kobj_nodes) {
 821			kobject_del(sys_props.kobj_nodes);
 822			kobject_put(sys_props.kobj_nodes);
 823			sys_props.kobj_nodes = NULL;
 824		}
 825		kobject_del(sys_props.kobj_topology);
 826		kobject_put(sys_props.kobj_topology);
 827		sys_props.kobj_topology = NULL;
 828	}
 829}
 830
 831/* Called with write topology_lock acquired */
 832static void kfd_topology_update_device_list(struct list_head *temp_list,
 833					struct list_head *master_list)
 834{
 835	while (!list_empty(temp_list)) {
 836		list_move_tail(temp_list->next, master_list);
 837		sys_props.num_devices++;
 838	}
 839}
 840
 841static void kfd_debug_print_topology(void)
 842{
 843	struct kfd_topology_device *dev;
 844
 845	down_read(&topology_lock);
 846
 847	dev = list_last_entry(&topology_device_list,
 848			struct kfd_topology_device, list);
 849	if (dev) {
 850		if (dev->node_props.cpu_cores_count &&
 851				dev->node_props.simd_count) {
 852			pr_info("Topology: Add APU node [0x%0x:0x%0x]\n",
 853				dev->node_props.device_id,
 854				dev->node_props.vendor_id);
 855		} else if (dev->node_props.cpu_cores_count)
 856			pr_info("Topology: Add CPU node\n");
 857		else if (dev->node_props.simd_count)
 858			pr_info("Topology: Add dGPU node [0x%0x:0x%0x]\n",
 859				dev->node_props.device_id,
 860				dev->node_props.vendor_id);
 861	}
 862	up_read(&topology_lock);
 863}
 864
 865/* Helper function for intializing platform_xx members of
 866 * kfd_system_properties. Uses OEM info from the last CPU/APU node.
 867 */
 868static void kfd_update_system_properties(void)
 869{
 870	struct kfd_topology_device *dev;
 871
 872	down_read(&topology_lock);
 873	dev = list_last_entry(&topology_device_list,
 874			struct kfd_topology_device, list);
 875	if (dev) {
 876		sys_props.platform_id =
 877			(*((uint64_t *)dev->oem_id)) & CRAT_OEMID_64BIT_MASK;
 878		sys_props.platform_oem = *((uint64_t *)dev->oem_table_id);
 879		sys_props.platform_rev = dev->oem_revision;
 880	}
 881	up_read(&topology_lock);
 882}
 883
 884static void find_system_memory(const struct dmi_header *dm,
 885	void *private)
 886{
 887	struct kfd_mem_properties *mem;
 888	u16 mem_width, mem_clock;
 889	struct kfd_topology_device *kdev =
 890		(struct kfd_topology_device *)private;
 891	const u8 *dmi_data = (const u8 *)(dm + 1);
 892
 893	if (dm->type == DMI_ENTRY_MEM_DEVICE && dm->length >= 0x15) {
 894		mem_width = (u16)(*(const u16 *)(dmi_data + 0x6));
 895		mem_clock = (u16)(*(const u16 *)(dmi_data + 0x11));
 896		list_for_each_entry(mem, &kdev->mem_props, list) {
 897			if (mem_width != 0xFFFF && mem_width != 0)
 898				mem->width = mem_width;
 899			if (mem_clock != 0)
 900				mem->mem_clk_max = mem_clock;
 901		}
 902	}
 903}
 904
 905/*
 906 * Performance counters information is not part of CRAT but we would like to
 907 * put them in the sysfs under topology directory for Thunk to get the data.
 908 * This function is called before updating the sysfs.
 909 */
 910static int kfd_add_perf_to_topology(struct kfd_topology_device *kdev)
 911{
 912	/* These are the only counters supported so far */
 913	return kfd_iommu_add_perf_counters(kdev);
 914}
 915
 916/* kfd_add_non_crat_information - Add information that is not currently
 917 *	defined in CRAT but is necessary for KFD topology
 918 * @dev - topology device to which addition info is added
 919 */
 920static void kfd_add_non_crat_information(struct kfd_topology_device *kdev)
 921{
 922	/* Check if CPU only node. */
 923	if (!kdev->gpu) {
 924		/* Add system memory information */
 925		dmi_walk(find_system_memory, kdev);
 926	}
 927	/* TODO: For GPU node, rearrange code from kfd_topology_add_device */
 928}
 929
 930/* kfd_is_acpi_crat_invalid - CRAT from ACPI is valid only for AMD APU devices.
 931 *	Ignore CRAT for all other devices. AMD APU is identified if both CPU
 932 *	and GPU cores are present.
 933 * @device_list - topology device list created by parsing ACPI CRAT table.
 934 * @return - TRUE if invalid, FALSE is valid.
 935 */
 936static bool kfd_is_acpi_crat_invalid(struct list_head *device_list)
 937{
 938	struct kfd_topology_device *dev;
 939
 940	list_for_each_entry(dev, device_list, list) {
 941		if (dev->node_props.cpu_cores_count &&
 942			dev->node_props.simd_count)
 943			return false;
 944	}
 945	pr_info("Ignoring ACPI CRAT on non-APU system\n");
 946	return true;
 947}
 948
 949int kfd_topology_init(void)
 950{
 951	void *crat_image = NULL;
 952	size_t image_size = 0;
 953	int ret;
 954	struct list_head temp_topology_device_list;
 955	int cpu_only_node = 0;
 956	struct kfd_topology_device *kdev;
 957	int proximity_domain;
 958
 959	/* topology_device_list - Master list of all topology devices
 960	 * temp_topology_device_list - temporary list created while parsing CRAT
 961	 * or VCRAT. Once parsing is complete the contents of list is moved to
 962	 * topology_device_list
 963	 */
 964
 965	/* Initialize the head for the both the lists */
 966	INIT_LIST_HEAD(&topology_device_list);
 967	INIT_LIST_HEAD(&temp_topology_device_list);
 968	init_rwsem(&topology_lock);
 
 969
 970	memset(&sys_props, 0, sizeof(sys_props));
 971
 972	/* Proximity domains in ACPI CRAT tables start counting at
 973	 * 0. The same should be true for virtual CRAT tables created
 974	 * at this stage. GPUs added later in kfd_topology_add_device
 975	 * use a counter.
 976	 */
 977	proximity_domain = 0;
 978
 979	/*
 980	 * Get the CRAT image from the ACPI. If ACPI doesn't have one
 981	 * or if ACPI CRAT is invalid create a virtual CRAT.
 982	 * NOTE: The current implementation expects all AMD APUs to have
 983	 *	CRAT. If no CRAT is available, it is assumed to be a CPU
 984	 */
 985	ret = kfd_create_crat_image_acpi(&crat_image, &image_size);
 986	if (!ret) {
 987		ret = kfd_parse_crat_table(crat_image,
 988					   &temp_topology_device_list,
 989					   proximity_domain);
 990		if (ret ||
 991		    kfd_is_acpi_crat_invalid(&temp_topology_device_list)) {
 992			kfd_release_topology_device_list(
 993				&temp_topology_device_list);
 994			kfd_destroy_crat_image(crat_image);
 995			crat_image = NULL;
 996		}
 997	}
 998
 999	if (!crat_image) {
1000		ret = kfd_create_crat_image_virtual(&crat_image, &image_size,
1001						    COMPUTE_UNIT_CPU, NULL,
1002						    proximity_domain);
1003		cpu_only_node = 1;
1004		if (ret) {
1005			pr_err("Error creating VCRAT table for CPU\n");
1006			return ret;
1007		}
1008
1009		ret = kfd_parse_crat_table(crat_image,
1010					   &temp_topology_device_list,
1011					   proximity_domain);
1012		if (ret) {
1013			pr_err("Error parsing VCRAT table for CPU\n");
1014			goto err;
1015		}
1016	}
1017
1018	kdev = list_first_entry(&temp_topology_device_list,
1019				struct kfd_topology_device, list);
1020	kfd_add_perf_to_topology(kdev);
1021
1022	down_write(&topology_lock);
1023	kfd_topology_update_device_list(&temp_topology_device_list,
1024					&topology_device_list);
1025	atomic_set(&topology_crat_proximity_domain, sys_props.num_devices-1);
1026	ret = kfd_topology_update_sysfs();
1027	up_write(&topology_lock);
1028
1029	if (!ret) {
1030		sys_props.generation_count++;
1031		kfd_update_system_properties();
1032		kfd_debug_print_topology();
1033		pr_info("Finished initializing topology\n");
1034	} else
1035		pr_err("Failed to update topology in sysfs ret=%d\n", ret);
1036
1037	/* For nodes with GPU, this information gets added
1038	 * when GPU is detected (kfd_topology_add_device).
1039	 */
1040	if (cpu_only_node) {
1041		/* Add additional information to CPU only node created above */
1042		down_write(&topology_lock);
1043		kdev = list_first_entry(&topology_device_list,
1044				struct kfd_topology_device, list);
1045		up_write(&topology_lock);
1046		kfd_add_non_crat_information(kdev);
1047	}
1048
1049err:
1050	kfd_destroy_crat_image(crat_image);
1051	return ret;
1052}
1053
1054void kfd_topology_shutdown(void)
1055{
1056	down_write(&topology_lock);
1057	kfd_topology_release_sysfs();
1058	kfd_release_live_view();
1059	up_write(&topology_lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1060}
1061
1062static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu)
1063{
1064	uint32_t hashout;
1065	uint32_t buf[7];
1066	uint64_t local_mem_size;
1067	int i;
1068	struct kfd_local_mem_info local_mem_info;
1069
1070	if (!gpu)
1071		return 0;
1072
1073	amdgpu_amdkfd_get_local_mem_info(gpu->kgd, &local_mem_info);
1074
1075	local_mem_size = local_mem_info.local_mem_size_private +
1076			local_mem_info.local_mem_size_public;
1077
1078	buf[0] = gpu->pdev->devfn;
1079	buf[1] = gpu->pdev->subsystem_vendor |
1080		(gpu->pdev->subsystem_device << 16);
1081	buf[2] = pci_domain_nr(gpu->pdev->bus);
1082	buf[3] = gpu->pdev->device;
1083	buf[4] = gpu->pdev->bus->number;
1084	buf[5] = lower_32_bits(local_mem_size);
1085	buf[6] = upper_32_bits(local_mem_size);
 
1086
1087	for (i = 0, hashout = 0; i < 7; i++)
1088		hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH);
1089
1090	return hashout;
1091}
1092/* kfd_assign_gpu - Attach @gpu to the correct kfd topology device. If
1093 *		the GPU device is not already present in the topology device
1094 *		list then return NULL. This means a new topology device has to
1095 *		be created for this GPU.
1096 */
1097static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu)
1098{
1099	struct kfd_topology_device *dev;
1100	struct kfd_topology_device *out_dev = NULL;
1101
1102	down_write(&topology_lock);
1103	list_for_each_entry(dev, &topology_device_list, list) {
1104		/* Discrete GPUs need their own topology device list
1105		 * entries. Don't assign them to CPU/APU nodes.
1106		 */
1107		if (!gpu->device_info->needs_iommu_device &&
1108		    dev->node_props.cpu_cores_count)
1109			continue;
1110
1111		if (!dev->gpu && (dev->node_props.simd_count > 0)) {
 
1112			dev->gpu = gpu;
1113			out_dev = dev;
1114			break;
1115		}
1116	}
1117	up_write(&topology_lock);
1118	return out_dev;
1119}
1120
1121static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival)
1122{
1123	/*
1124	 * TODO: Generate an event for thunk about the arrival/removal
1125	 * of the GPU
1126	 */
1127}
1128
1129/* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info,
1130 *		patch this after CRAT parsing.
1131 */
1132static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev)
1133{
1134	struct kfd_mem_properties *mem;
1135	struct kfd_local_mem_info local_mem_info;
1136
1137	if (!dev)
1138		return;
1139
1140	/* Currently, amdgpu driver (amdgpu_mc) deals only with GPUs with
1141	 * single bank of VRAM local memory.
1142	 * for dGPUs - VCRAT reports only one bank of Local Memory
1143	 * for APUs - If CRAT from ACPI reports more than one bank, then
1144	 *	all the banks will report the same mem_clk_max information
1145	 */
1146	amdgpu_amdkfd_get_local_mem_info(dev->gpu->kgd, &local_mem_info);
1147
1148	list_for_each_entry(mem, &dev->mem_props, list)
1149		mem->mem_clk_max = local_mem_info.mem_clk_max;
1150}
1151
1152static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev)
1153{
1154	struct kfd_iolink_properties *link, *cpu_link;
1155	struct kfd_topology_device *cpu_dev;
1156	uint32_t cap;
1157	uint32_t cpu_flag = CRAT_IOLINK_FLAGS_ENABLED;
1158	uint32_t flag = CRAT_IOLINK_FLAGS_ENABLED;
1159
1160	if (!dev || !dev->gpu)
1161		return;
1162
1163	pcie_capability_read_dword(dev->gpu->pdev,
1164			PCI_EXP_DEVCAP2, &cap);
1165
1166	if (!(cap & (PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
1167		     PCI_EXP_DEVCAP2_ATOMIC_COMP64)))
1168		cpu_flag |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1169			CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1170
1171	if (!dev->gpu->pci_atomic_requested ||
1172	    dev->gpu->device_info->asic_family == CHIP_HAWAII)
1173		flag |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1174			CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1175
1176	/* GPU only creates direct links so apply flags setting to all */
1177	list_for_each_entry(link, &dev->io_link_props, list) {
1178		link->flags = flag;
1179		cpu_dev = kfd_topology_device_by_proximity_domain(
1180				link->node_to);
1181		if (cpu_dev) {
1182			list_for_each_entry(cpu_link,
1183					    &cpu_dev->io_link_props, list)
1184				if (cpu_link->node_to == link->node_from)
1185					cpu_link->flags = cpu_flag;
1186		}
1187	}
1188}
1189
1190int kfd_topology_add_device(struct kfd_dev *gpu)
1191{
1192	uint32_t gpu_id;
1193	struct kfd_topology_device *dev;
1194	struct kfd_cu_info cu_info;
1195	int res = 0;
1196	struct list_head temp_topology_device_list;
1197	void *crat_image = NULL;
1198	size_t image_size = 0;
1199	int proximity_domain;
1200	struct amdgpu_ras *ctx;
1201
1202	INIT_LIST_HEAD(&temp_topology_device_list);
1203
1204	gpu_id = kfd_generate_gpu_id(gpu);
1205
1206	pr_debug("Adding new GPU (ID: 0x%x) to topology\n", gpu_id);
1207
1208	proximity_domain = atomic_inc_return(&topology_crat_proximity_domain);
1209
1210	/* Check to see if this gpu device exists in the topology_device_list.
1211	 * If so, assign the gpu to that device,
1212	 * else create a Virtual CRAT for this gpu device and then parse that
1213	 * CRAT to create a new topology device. Once created assign the gpu to
1214	 * that topology device
1215	 */
1216	dev = kfd_assign_gpu(gpu);
1217	if (!dev) {
1218		res = kfd_create_crat_image_virtual(&crat_image, &image_size,
1219						    COMPUTE_UNIT_GPU, gpu,
1220						    proximity_domain);
1221		if (res) {
1222			pr_err("Error creating VCRAT for GPU (ID: 0x%x)\n",
1223			       gpu_id);
1224			return res;
1225		}
1226		res = kfd_parse_crat_table(crat_image,
1227					   &temp_topology_device_list,
1228					   proximity_domain);
1229		if (res) {
1230			pr_err("Error parsing VCRAT for GPU (ID: 0x%x)\n",
1231			       gpu_id);
1232			goto err;
1233		}
 
1234
1235		down_write(&topology_lock);
1236		kfd_topology_update_device_list(&temp_topology_device_list,
1237			&topology_device_list);
 
1238
1239		/* Update the SYSFS tree, since we added another topology
1240		 * device
1241		 */
1242		res = kfd_topology_update_sysfs();
1243		up_write(&topology_lock);
1244
1245		if (!res)
1246			sys_props.generation_count++;
1247		else
1248			pr_err("Failed to update GPU (ID: 0x%x) to sysfs topology. res=%d\n",
1249						gpu_id, res);
1250		dev = kfd_assign_gpu(gpu);
1251		if (WARN_ON(!dev)) {
1252			res = -ENODEV;
1253			goto err;
1254		}
1255	}
1256
1257	dev->gpu_id = gpu_id;
1258	gpu->id = gpu_id;
1259
1260	/* TODO: Move the following lines to function
1261	 *	kfd_add_non_crat_information
1262	 */
1263
1264	/* Fill-in additional information that is not available in CRAT but
1265	 * needed for the topology
1266	 */
1267
1268	amdgpu_amdkfd_get_cu_info(dev->gpu->kgd, &cu_info);
1269
1270	strncpy(dev->node_props.name, gpu->device_info->asic_name,
1271			KFD_TOPOLOGY_PUBLIC_NAME_SIZE);
1272
1273	dev->node_props.simd_arrays_per_engine =
1274		cu_info.num_shader_arrays_per_engine;
1275
1276	dev->node_props.vendor_id = gpu->pdev->vendor;
1277	dev->node_props.device_id = gpu->pdev->device;
1278	dev->node_props.location_id = pci_dev_id(gpu->pdev);
1279	dev->node_props.max_engine_clk_fcompute =
1280		amdgpu_amdkfd_get_max_engine_clock_in_mhz(dev->gpu->kgd);
1281	dev->node_props.max_engine_clk_ccompute =
1282		cpufreq_quick_get_max(0) / 1000;
1283	dev->node_props.drm_render_minor =
1284		gpu->shared_resources.drm_render_minor;
1285
1286	dev->node_props.hive_id = gpu->hive_id;
1287	dev->node_props.num_sdma_engines = gpu->device_info->num_sdma_engines;
1288	dev->node_props.num_sdma_xgmi_engines =
1289				gpu->device_info->num_xgmi_sdma_engines;
1290	dev->node_props.num_gws = (hws_gws_support &&
1291		dev->gpu->dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) ?
1292		amdgpu_amdkfd_get_num_gws(dev->gpu->kgd) : 0;
1293
1294	kfd_fill_mem_clk_max_info(dev);
1295	kfd_fill_iolink_non_crat_info(dev);
1296
1297	switch (dev->gpu->device_info->asic_family) {
1298	case CHIP_KAVERI:
1299	case CHIP_HAWAII:
1300	case CHIP_TONGA:
1301		dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_PRE_1_0 <<
1302			HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1303			HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1304		break;
1305	case CHIP_CARRIZO:
1306	case CHIP_FIJI:
1307	case CHIP_POLARIS10:
1308	case CHIP_POLARIS11:
1309	case CHIP_POLARIS12:
1310	case CHIP_VEGAM:
1311		pr_debug("Adding doorbell packet type capability\n");
1312		dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 <<
1313			HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1314			HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1315		break;
1316	case CHIP_VEGA10:
1317	case CHIP_VEGA12:
1318	case CHIP_VEGA20:
1319	case CHIP_RAVEN:
1320	case CHIP_ARCTURUS:
1321	case CHIP_NAVI10:
1322		dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 <<
1323			HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1324			HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1325		break;
1326	default:
1327		WARN(1, "Unexpected ASIC family %u",
1328		     dev->gpu->device_info->asic_family);
1329	}
1330
1331	/*
1332	* Overwrite ATS capability according to needs_iommu_device to fix
1333	* potential missing corresponding bit in CRAT of BIOS.
1334	*/
1335	if (dev->gpu->device_info->needs_iommu_device)
1336		dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
1337	else
1338		dev->node_props.capability &= ~HSA_CAP_ATS_PRESENT;
1339
1340	/* Fix errors in CZ CRAT.
1341	 * simd_count: Carrizo CRAT reports wrong simd_count, probably
1342	 *		because it doesn't consider masked out CUs
1343	 * max_waves_per_simd: Carrizo reports wrong max_waves_per_simd
1344	 */
1345	if (dev->gpu->device_info->asic_family == CHIP_CARRIZO) {
1346		dev->node_props.simd_count =
1347			cu_info.simd_per_cu * cu_info.cu_active_number;
1348		dev->node_props.max_waves_per_simd = 10;
1349	}
1350
1351	ctx = amdgpu_ras_get_context((struct amdgpu_device *)(dev->gpu->kgd));
1352	if (ctx) {
1353		/* kfd only concerns sram ecc on GFX/SDMA and HBM ecc on UMC */
1354		dev->node_props.capability |=
1355			(((ctx->features & BIT(AMDGPU_RAS_BLOCK__SDMA)) != 0) ||
1356			 ((ctx->features & BIT(AMDGPU_RAS_BLOCK__GFX)) != 0)) ?
1357			HSA_CAP_SRAM_EDCSUPPORTED : 0;
1358		dev->node_props.capability |= ((ctx->features & BIT(AMDGPU_RAS_BLOCK__UMC)) != 0) ?
1359			HSA_CAP_MEM_EDCSUPPORTED : 0;
1360
1361		dev->node_props.capability |= (ctx->features != 0) ?
1362			HSA_CAP_RASEVENTNOTIFY : 0;
1363	}
1364
1365	kfd_debug_print_topology();
 
1366
1367	if (!res)
1368		kfd_notify_gpu_change(gpu_id, 1);
1369err:
1370	kfd_destroy_crat_image(crat_image);
1371	return res;
1372}
1373
1374int kfd_topology_remove_device(struct kfd_dev *gpu)
1375{
1376	struct kfd_topology_device *dev, *tmp;
1377	uint32_t gpu_id;
1378	int res = -ENODEV;
1379
 
 
1380	down_write(&topology_lock);
1381
1382	list_for_each_entry_safe(dev, tmp, &topology_device_list, list)
1383		if (dev->gpu == gpu) {
1384			gpu_id = dev->gpu_id;
1385			kfd_remove_sysfs_node_entry(dev);
1386			kfd_release_topology_device(dev);
1387			sys_props.num_devices--;
1388			res = 0;
1389			if (kfd_topology_update_sysfs() < 0)
1390				kfd_topology_release_sysfs();
1391			break;
1392		}
1393
1394	up_write(&topology_lock);
1395
1396	if (!res)
1397		kfd_notify_gpu_change(gpu_id, 0);
1398
1399	return res;
1400}
1401
1402/* kfd_topology_enum_kfd_devices - Enumerate through all devices in KFD
1403 *	topology. If GPU device is found @idx, then valid kfd_dev pointer is
1404 *	returned through @kdev
1405 * Return -	0: On success (@kdev will be NULL for non GPU nodes)
1406 *		-1: If end of list
1407 */
1408int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev)
1409{
1410
1411	struct kfd_topology_device *top_dev;
 
1412	uint8_t device_idx = 0;
1413
1414	*kdev = NULL;
1415	down_read(&topology_lock);
1416
1417	list_for_each_entry(top_dev, &topology_device_list, list) {
1418		if (device_idx == idx) {
1419			*kdev = top_dev->gpu;
1420			up_read(&topology_lock);
1421			return 0;
1422		}
1423
1424		device_idx++;
1425	}
1426
1427	up_read(&topology_lock);
1428
1429	return -1;
1430
1431}
1432
1433static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask)
1434{
1435	int first_cpu_of_numa_node;
1436
1437	if (!cpumask || cpumask == cpu_none_mask)
1438		return -1;
1439	first_cpu_of_numa_node = cpumask_first(cpumask);
1440	if (first_cpu_of_numa_node >= nr_cpu_ids)
1441		return -1;
1442#ifdef CONFIG_X86_64
1443	return cpu_data(first_cpu_of_numa_node).apicid;
1444#else
1445	return first_cpu_of_numa_node;
1446#endif
1447}
1448
1449/* kfd_numa_node_to_apic_id - Returns the APIC ID of the first logical processor
1450 *	of the given NUMA node (numa_node_id)
1451 * Return -1 on failure
1452 */
1453int kfd_numa_node_to_apic_id(int numa_node_id)
1454{
1455	if (numa_node_id == -1) {
1456		pr_warn("Invalid NUMA Node. Use online CPU mask\n");
1457		return kfd_cpumask_to_apic_id(cpu_online_mask);
1458	}
1459	return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id));
1460}
1461
1462#if defined(CONFIG_DEBUG_FS)
1463
1464int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data)
1465{
1466	struct kfd_topology_device *dev;
1467	unsigned int i = 0;
1468	int r = 0;
1469
1470	down_read(&topology_lock);
1471
1472	list_for_each_entry(dev, &topology_device_list, list) {
1473		if (!dev->gpu) {
1474			i++;
1475			continue;
1476		}
1477
1478		seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1479		r = dqm_debugfs_hqds(m, dev->gpu->dqm);
1480		if (r)
1481			break;
1482	}
1483
1484	up_read(&topology_lock);
1485
1486	return r;
1487}
1488
1489int kfd_debugfs_rls_by_device(struct seq_file *m, void *data)
1490{
1491	struct kfd_topology_device *dev;
1492	unsigned int i = 0;
1493	int r = 0;
1494
1495	down_read(&topology_lock);
1496
1497	list_for_each_entry(dev, &topology_device_list, list) {
1498		if (!dev->gpu) {
1499			i++;
1500			continue;
1501		}
1502
1503		seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1504		r = pm_debugfs_runlist(m, &dev->gpu->dqm->packets);
1505		if (r)
1506			break;
1507	}
1508
1509	up_read(&topology_lock);
1510
1511	return r;
1512}
1513
1514#endif