2 * Copyright 2014 Advanced Micro Devices, Inc.
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:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
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.
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>
34 #include "kfd_topology.h"
36 static struct list_head topology_device_list;
37 static int topology_crat_parsed;
38 static struct kfd_system_properties sys_props;
40 static DECLARE_RWSEM(topology_lock);
42 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id)
44 struct kfd_topology_device *top_dev;
45 struct kfd_dev *device = NULL;
47 down_read(&topology_lock);
49 list_for_each_entry(top_dev, &topology_device_list, list)
50 if (top_dev->gpu_id == gpu_id) {
51 device = top_dev->gpu;
55 up_read(&topology_lock);
60 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev)
62 struct kfd_topology_device *top_dev;
63 struct kfd_dev *device = NULL;
65 down_read(&topology_lock);
67 list_for_each_entry(top_dev, &topology_device_list, list)
68 if (top_dev->gpu->pdev == pdev) {
69 device = top_dev->gpu;
73 up_read(&topology_lock);
78 static int kfd_topology_get_crat_acpi(void *crat_image, size_t *size)
80 struct acpi_table_header *crat_table;
87 * Fetch the CRAT table from ACPI
89 status = acpi_get_table(CRAT_SIGNATURE, 0, &crat_table);
90 if (status == AE_NOT_FOUND) {
91 pr_warn("CRAT table not found\n");
93 } else if (ACPI_FAILURE(status)) {
94 const char *err = acpi_format_exception(status);
96 pr_err("CRAT table error: %s\n", err);
100 if (*size >= crat_table->length && crat_image != NULL)
101 memcpy(crat_image, crat_table, crat_table->length);
103 *size = crat_table->length;
108 static void kfd_populated_cu_info_cpu(struct kfd_topology_device *dev,
109 struct crat_subtype_computeunit *cu)
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;
119 pr_info("CU CPU: cores=%d id_base=%d\n", cu->num_cpu_cores,
120 cu->processor_id_low);
123 static void kfd_populated_cu_info_gpu(struct kfd_topology_device *dev,
124 struct crat_subtype_computeunit *cu)
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);
145 /* kfd_parse_subtype_cu is called when the topology mutex is already acquired */
146 static int kfd_parse_subtype_cu(struct crat_subtype_computeunit *cu)
148 struct kfd_topology_device *dev;
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);
160 if (cu->flags & CRAT_CU_FLAGS_GPU_PRESENT)
161 kfd_populated_cu_info_gpu(dev, cu);
171 * kfd_parse_subtype_mem is called when the topology mutex is
174 static int kfd_parse_subtype_mem(struct crat_subtype_memory *mem)
176 struct kfd_mem_properties *props;
177 struct kfd_topology_device *dev;
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);
190 if (dev->node_props.cpu_cores_count == 0)
191 props->heap_type = HSA_MEM_HEAP_TYPE_FB_PRIVATE;
193 props->heap_type = HSA_MEM_HEAP_TYPE_SYSTEM;
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;
200 props->size_in_bytes =
201 ((uint64_t)mem->length_high << 32) +
203 props->width = mem->width;
205 dev->mem_bank_count++;
206 list_add_tail(&props->list, &dev->mem_props);
217 * kfd_parse_subtype_cache is called when the topology mutex
218 * is already acquired
220 static int kfd_parse_subtype_cache(struct crat_subtype_cache *cache)
222 struct kfd_cache_properties *props;
223 struct kfd_topology_device *dev;
228 id = cache->processor_id_low;
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);
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;
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;
256 dev->node_props.caches_count++;
257 list_add_tail(&props->list, &dev->cache_props);
266 * kfd_parse_subtype_iolink is called when the topology mutex
267 * is already acquired
269 static int kfd_parse_subtype_iolink(struct crat_subtype_iolink *iolink)
271 struct kfd_iolink_properties *props;
272 struct kfd_topology_device *dev;
279 id_from = iolink->proximity_domain_from;
280 id_to = iolink->proximity_domain_to;
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) {
285 props = kfd_alloc_struct(props);
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;
295 * weight factor (derived from CDIR), currently always 1
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;
306 dev->io_link_count++;
307 dev->node_props.io_links_count++;
308 list_add_tail(&props->list, &dev->io_link_props);
318 static int kfd_parse_subtype(struct crat_subtype_generic *sub_type_hdr)
320 struct crat_subtype_computeunit *cu;
321 struct crat_subtype_memory *mem;
322 struct crat_subtype_cache *cache;
323 struct crat_subtype_iolink *iolink;
326 BUG_ON(!sub_type_hdr);
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);
333 case CRAT_SUBTYPE_MEMORY_AFFINITY:
334 mem = (struct crat_subtype_memory *)sub_type_hdr;
335 ret = kfd_parse_subtype_mem(mem);
337 case CRAT_SUBTYPE_CACHE_AFFINITY:
338 cache = (struct crat_subtype_cache *)sub_type_hdr;
339 ret = kfd_parse_subtype_cache(cache);
341 case CRAT_SUBTYPE_TLB_AFFINITY:
343 * For now, nothing to do here
345 pr_info("Found TLB entry in CRAT table (not processing)\n");
347 case CRAT_SUBTYPE_CCOMPUTE_AFFINITY:
349 * For now, nothing to do here
351 pr_info("Found CCOMPUTE entry in CRAT table (not processing)\n");
353 case CRAT_SUBTYPE_IOLINK_AFFINITY:
354 iolink = (struct crat_subtype_iolink *)sub_type_hdr;
355 ret = kfd_parse_subtype_iolink(iolink);
358 pr_warn("Unknown subtype (%d) in CRAT\n",
365 static void kfd_release_topology_device(struct kfd_topology_device *dev)
367 struct kfd_mem_properties *mem;
368 struct kfd_cache_properties *cache;
369 struct kfd_iolink_properties *iolink;
373 list_del(&dev->list);
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);
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);
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);
398 sys_props.num_devices--;
401 static void kfd_release_live_view(void)
403 struct kfd_topology_device *dev;
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);
411 memset(&sys_props, 0, sizeof(sys_props));
414 static struct kfd_topology_device *kfd_create_topology_device(void)
416 struct kfd_topology_device *dev;
418 dev = kfd_alloc_struct(dev);
420 pr_err("No memory to allocate a topology device");
424 INIT_LIST_HEAD(&dev->mem_props);
425 INIT_LIST_HEAD(&dev->cache_props);
426 INIT_LIST_HEAD(&dev->io_link_props);
428 list_add_tail(&dev->list, &topology_device_list);
429 sys_props.num_devices++;
434 static int kfd_parse_crat_table(void *crat_image)
436 struct kfd_topology_device *top_dev;
437 struct crat_subtype_generic *sub_type_hdr;
440 struct crat_header *crat_table = (struct crat_header *)crat_image;
447 num_nodes = crat_table->num_domains;
448 image_len = crat_table->length;
450 pr_info("Parsing CRAT table with %d nodes\n", num_nodes);
452 for (node_id = 0; node_id < num_nodes; node_id++) {
453 top_dev = kfd_create_topology_device();
455 kfd_release_live_view();
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;
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);
471 kfd_release_live_view();
476 sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
477 sub_type_hdr->length);
480 sys_props.generation_count++;
481 topology_crat_parsed = 1;
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)
498 static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr,
503 /* Making sure that the buffer is an empty string */
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);
522 static void kfd_topology_kobj_release(struct kobject *kobj)
527 static const struct sysfs_ops sysprops_ops = {
528 .show = sysprops_show,
531 static struct kobj_type sysprops_type = {
532 .release = kfd_topology_kobj_release,
533 .sysfs_ops = &sysprops_ops,
536 static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr,
540 struct kfd_iolink_properties *iolink;
542 /* Making sure that the buffer is an empty string */
545 iolink = container_of(attr, struct kfd_iolink_properties, attr);
546 sysfs_show_32bit_prop(buffer, "type", iolink->iolink_type);
547 sysfs_show_32bit_prop(buffer, "version_major", iolink->ver_maj);
548 sysfs_show_32bit_prop(buffer, "version_minor", iolink->ver_min);
549 sysfs_show_32bit_prop(buffer, "node_from", iolink->node_from);
550 sysfs_show_32bit_prop(buffer, "node_to", iolink->node_to);
551 sysfs_show_32bit_prop(buffer, "weight", iolink->weight);
552 sysfs_show_32bit_prop(buffer, "min_latency", iolink->min_latency);
553 sysfs_show_32bit_prop(buffer, "max_latency", iolink->max_latency);
554 sysfs_show_32bit_prop(buffer, "min_bandwidth", iolink->min_bandwidth);
555 sysfs_show_32bit_prop(buffer, "max_bandwidth", iolink->max_bandwidth);
556 sysfs_show_32bit_prop(buffer, "recommended_transfer_size",
557 iolink->rec_transfer_size);
558 ret = sysfs_show_32bit_prop(buffer, "flags", iolink->flags);
563 static const struct sysfs_ops iolink_ops = {
567 static struct kobj_type iolink_type = {
568 .release = kfd_topology_kobj_release,
569 .sysfs_ops = &iolink_ops,
572 static ssize_t mem_show(struct kobject *kobj, struct attribute *attr,
576 struct kfd_mem_properties *mem;
578 /* Making sure that the buffer is an empty string */
581 mem = container_of(attr, struct kfd_mem_properties, attr);
582 sysfs_show_32bit_prop(buffer, "heap_type", mem->heap_type);
583 sysfs_show_64bit_prop(buffer, "size_in_bytes", mem->size_in_bytes);
584 sysfs_show_32bit_prop(buffer, "flags", mem->flags);
585 sysfs_show_32bit_prop(buffer, "width", mem->width);
586 ret = sysfs_show_32bit_prop(buffer, "mem_clk_max", mem->mem_clk_max);
591 static const struct sysfs_ops mem_ops = {
595 static struct kobj_type mem_type = {
596 .release = kfd_topology_kobj_release,
597 .sysfs_ops = &mem_ops,
600 static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr,
605 struct kfd_cache_properties *cache;
607 /* Making sure that the buffer is an empty string */
610 cache = container_of(attr, struct kfd_cache_properties, attr);
611 sysfs_show_32bit_prop(buffer, "processor_id_low",
612 cache->processor_id_low);
613 sysfs_show_32bit_prop(buffer, "level", cache->cache_level);
614 sysfs_show_32bit_prop(buffer, "size", cache->cache_size);
615 sysfs_show_32bit_prop(buffer, "cache_line_size", cache->cacheline_size);
616 sysfs_show_32bit_prop(buffer, "cache_lines_per_tag",
617 cache->cachelines_per_tag);
618 sysfs_show_32bit_prop(buffer, "association", cache->cache_assoc);
619 sysfs_show_32bit_prop(buffer, "latency", cache->cache_latency);
620 sysfs_show_32bit_prop(buffer, "type", cache->cache_type);
621 snprintf(buffer, PAGE_SIZE, "%ssibling_map ", buffer);
622 for (i = 0; i < KFD_TOPOLOGY_CPU_SIBLINGS; i++)
623 ret = snprintf(buffer, PAGE_SIZE, "%s%d%s",
624 buffer, cache->sibling_map[i],
625 (i == KFD_TOPOLOGY_CPU_SIBLINGS-1) ?
631 static const struct sysfs_ops cache_ops = {
632 .show = kfd_cache_show,
635 static struct kobj_type cache_type = {
636 .release = kfd_topology_kobj_release,
637 .sysfs_ops = &cache_ops,
640 static ssize_t node_show(struct kobject *kobj, struct attribute *attr,
643 struct kfd_topology_device *dev;
644 char public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE];
646 uint32_t log_max_watch_addr;
648 /* Making sure that the buffer is an empty string */
651 if (strcmp(attr->name, "gpu_id") == 0) {
652 dev = container_of(attr, struct kfd_topology_device,
654 return sysfs_show_32bit_val(buffer, dev->gpu_id);
657 if (strcmp(attr->name, "name") == 0) {
658 dev = container_of(attr, struct kfd_topology_device,
660 for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE; i++) {
662 (char)dev->node_props.marketing_name[i];
663 if (dev->node_props.marketing_name[i] == 0)
666 public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1] = 0x0;
667 return sysfs_show_str_val(buffer, public_name);
670 dev = container_of(attr, struct kfd_topology_device,
672 sysfs_show_32bit_prop(buffer, "cpu_cores_count",
673 dev->node_props.cpu_cores_count);
674 sysfs_show_32bit_prop(buffer, "simd_count",
675 dev->node_props.simd_count);
677 if (dev->mem_bank_count < dev->node_props.mem_banks_count) {
678 pr_info_once("kfd: mem_banks_count truncated from %d to %d\n",
679 dev->node_props.mem_banks_count,
680 dev->mem_bank_count);
681 sysfs_show_32bit_prop(buffer, "mem_banks_count",
682 dev->mem_bank_count);
684 sysfs_show_32bit_prop(buffer, "mem_banks_count",
685 dev->node_props.mem_banks_count);
688 sysfs_show_32bit_prop(buffer, "caches_count",
689 dev->node_props.caches_count);
690 sysfs_show_32bit_prop(buffer, "io_links_count",
691 dev->node_props.io_links_count);
692 sysfs_show_32bit_prop(buffer, "cpu_core_id_base",
693 dev->node_props.cpu_core_id_base);
694 sysfs_show_32bit_prop(buffer, "simd_id_base",
695 dev->node_props.simd_id_base);
696 sysfs_show_32bit_prop(buffer, "max_waves_per_simd",
697 dev->node_props.max_waves_per_simd);
698 sysfs_show_32bit_prop(buffer, "lds_size_in_kb",
699 dev->node_props.lds_size_in_kb);
700 sysfs_show_32bit_prop(buffer, "gds_size_in_kb",
701 dev->node_props.gds_size_in_kb);
702 sysfs_show_32bit_prop(buffer, "wave_front_size",
703 dev->node_props.wave_front_size);
704 sysfs_show_32bit_prop(buffer, "array_count",
705 dev->node_props.array_count);
706 sysfs_show_32bit_prop(buffer, "simd_arrays_per_engine",
707 dev->node_props.simd_arrays_per_engine);
708 sysfs_show_32bit_prop(buffer, "cu_per_simd_array",
709 dev->node_props.cu_per_simd_array);
710 sysfs_show_32bit_prop(buffer, "simd_per_cu",
711 dev->node_props.simd_per_cu);
712 sysfs_show_32bit_prop(buffer, "max_slots_scratch_cu",
713 dev->node_props.max_slots_scratch_cu);
714 sysfs_show_32bit_prop(buffer, "vendor_id",
715 dev->node_props.vendor_id);
716 sysfs_show_32bit_prop(buffer, "device_id",
717 dev->node_props.device_id);
718 sysfs_show_32bit_prop(buffer, "location_id",
719 dev->node_props.location_id);
723 __ilog2_u32(dev->gpu->device_info->num_of_watch_points);
725 if (log_max_watch_addr) {
726 dev->node_props.capability |=
727 HSA_CAP_WATCH_POINTS_SUPPORTED;
729 dev->node_props.capability |=
730 ((log_max_watch_addr <<
731 HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) &
732 HSA_CAP_WATCH_POINTS_TOTALBITS_MASK);
735 sysfs_show_32bit_prop(buffer, "max_engine_clk_fcompute",
736 dev->gpu->kfd2kgd->get_max_engine_clock_in_mhz(
739 sysfs_show_64bit_prop(buffer, "local_mem_size",
740 (unsigned long long int) 0);
742 sysfs_show_32bit_prop(buffer, "fw_version",
743 dev->gpu->kfd2kgd->get_fw_version(
746 sysfs_show_32bit_prop(buffer, "capability",
747 dev->node_props.capability);
750 return sysfs_show_32bit_prop(buffer, "max_engine_clk_ccompute",
751 cpufreq_quick_get_max(0)/1000);
754 static const struct sysfs_ops node_ops = {
758 static struct kobj_type node_type = {
759 .release = kfd_topology_kobj_release,
760 .sysfs_ops = &node_ops,
763 static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr)
765 sysfs_remove_file(kobj, attr);
770 static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev)
772 struct kfd_iolink_properties *iolink;
773 struct kfd_cache_properties *cache;
774 struct kfd_mem_properties *mem;
778 if (dev->kobj_iolink) {
779 list_for_each_entry(iolink, &dev->io_link_props, list)
781 kfd_remove_sysfs_file(iolink->kobj,
785 kobject_del(dev->kobj_iolink);
786 kobject_put(dev->kobj_iolink);
787 dev->kobj_iolink = NULL;
790 if (dev->kobj_cache) {
791 list_for_each_entry(cache, &dev->cache_props, list)
793 kfd_remove_sysfs_file(cache->kobj,
797 kobject_del(dev->kobj_cache);
798 kobject_put(dev->kobj_cache);
799 dev->kobj_cache = NULL;
803 list_for_each_entry(mem, &dev->mem_props, list)
805 kfd_remove_sysfs_file(mem->kobj, &mem->attr);
808 kobject_del(dev->kobj_mem);
809 kobject_put(dev->kobj_mem);
810 dev->kobj_mem = NULL;
813 if (dev->kobj_node) {
814 sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid);
815 sysfs_remove_file(dev->kobj_node, &dev->attr_name);
816 sysfs_remove_file(dev->kobj_node, &dev->attr_props);
817 kobject_del(dev->kobj_node);
818 kobject_put(dev->kobj_node);
819 dev->kobj_node = NULL;
823 static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev,
826 struct kfd_iolink_properties *iolink;
827 struct kfd_cache_properties *cache;
828 struct kfd_mem_properties *mem;
835 * Creating the sysfs folders
837 BUG_ON(dev->kobj_node);
838 dev->kobj_node = kfd_alloc_struct(dev->kobj_node);
842 ret = kobject_init_and_add(dev->kobj_node, &node_type,
843 sys_props.kobj_nodes, "%d", id);
845 kobject_put(dev->kobj_node);
849 dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node);
853 dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node);
854 if (!dev->kobj_cache)
857 dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node);
858 if (!dev->kobj_iolink)
862 * Creating sysfs files for node properties
864 dev->attr_gpuid.name = "gpu_id";
865 dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE;
866 sysfs_attr_init(&dev->attr_gpuid);
867 dev->attr_name.name = "name";
868 dev->attr_name.mode = KFD_SYSFS_FILE_MODE;
869 sysfs_attr_init(&dev->attr_name);
870 dev->attr_props.name = "properties";
871 dev->attr_props.mode = KFD_SYSFS_FILE_MODE;
872 sysfs_attr_init(&dev->attr_props);
873 ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid);
876 ret = sysfs_create_file(dev->kobj_node, &dev->attr_name);
879 ret = sysfs_create_file(dev->kobj_node, &dev->attr_props);
884 list_for_each_entry(mem, &dev->mem_props, list) {
885 mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
888 ret = kobject_init_and_add(mem->kobj, &mem_type,
889 dev->kobj_mem, "%d", i);
891 kobject_put(mem->kobj);
895 mem->attr.name = "properties";
896 mem->attr.mode = KFD_SYSFS_FILE_MODE;
897 sysfs_attr_init(&mem->attr);
898 ret = sysfs_create_file(mem->kobj, &mem->attr);
905 list_for_each_entry(cache, &dev->cache_props, list) {
906 cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
909 ret = kobject_init_and_add(cache->kobj, &cache_type,
910 dev->kobj_cache, "%d", i);
912 kobject_put(cache->kobj);
916 cache->attr.name = "properties";
917 cache->attr.mode = KFD_SYSFS_FILE_MODE;
918 sysfs_attr_init(&cache->attr);
919 ret = sysfs_create_file(cache->kobj, &cache->attr);
926 list_for_each_entry(iolink, &dev->io_link_props, list) {
927 iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
930 ret = kobject_init_and_add(iolink->kobj, &iolink_type,
931 dev->kobj_iolink, "%d", i);
933 kobject_put(iolink->kobj);
937 iolink->attr.name = "properties";
938 iolink->attr.mode = KFD_SYSFS_FILE_MODE;
939 sysfs_attr_init(&iolink->attr);
940 ret = sysfs_create_file(iolink->kobj, &iolink->attr);
949 static int kfd_build_sysfs_node_tree(void)
951 struct kfd_topology_device *dev;
955 list_for_each_entry(dev, &topology_device_list, list) {
956 ret = kfd_build_sysfs_node_entry(dev, i);
965 static void kfd_remove_sysfs_node_tree(void)
967 struct kfd_topology_device *dev;
969 list_for_each_entry(dev, &topology_device_list, list)
970 kfd_remove_sysfs_node_entry(dev);
973 static int kfd_topology_update_sysfs(void)
977 pr_info("Creating topology SYSFS entries\n");
978 if (sys_props.kobj_topology == NULL) {
979 sys_props.kobj_topology =
980 kfd_alloc_struct(sys_props.kobj_topology);
981 if (!sys_props.kobj_topology)
984 ret = kobject_init_and_add(sys_props.kobj_topology,
985 &sysprops_type, &kfd_device->kobj,
988 kobject_put(sys_props.kobj_topology);
992 sys_props.kobj_nodes = kobject_create_and_add("nodes",
993 sys_props.kobj_topology);
994 if (!sys_props.kobj_nodes)
997 sys_props.attr_genid.name = "generation_id";
998 sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE;
999 sysfs_attr_init(&sys_props.attr_genid);
1000 ret = sysfs_create_file(sys_props.kobj_topology,
1001 &sys_props.attr_genid);
1005 sys_props.attr_props.name = "system_properties";
1006 sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE;
1007 sysfs_attr_init(&sys_props.attr_props);
1008 ret = sysfs_create_file(sys_props.kobj_topology,
1009 &sys_props.attr_props);
1014 kfd_remove_sysfs_node_tree();
1016 return kfd_build_sysfs_node_tree();
1019 static void kfd_topology_release_sysfs(void)
1021 kfd_remove_sysfs_node_tree();
1022 if (sys_props.kobj_topology) {
1023 sysfs_remove_file(sys_props.kobj_topology,
1024 &sys_props.attr_genid);
1025 sysfs_remove_file(sys_props.kobj_topology,
1026 &sys_props.attr_props);
1027 if (sys_props.kobj_nodes) {
1028 kobject_del(sys_props.kobj_nodes);
1029 kobject_put(sys_props.kobj_nodes);
1030 sys_props.kobj_nodes = NULL;
1032 kobject_del(sys_props.kobj_topology);
1033 kobject_put(sys_props.kobj_topology);
1034 sys_props.kobj_topology = NULL;
1038 int kfd_topology_init(void)
1040 void *crat_image = NULL;
1041 size_t image_size = 0;
1045 * Initialize the head for the topology device list
1047 INIT_LIST_HEAD(&topology_device_list);
1048 init_rwsem(&topology_lock);
1049 topology_crat_parsed = 0;
1051 memset(&sys_props, 0, sizeof(sys_props));
1054 * Get the CRAT image from the ACPI
1056 ret = kfd_topology_get_crat_acpi(crat_image, &image_size);
1057 if (ret == 0 && image_size > 0) {
1058 pr_info("Found CRAT image with size=%zd\n", image_size);
1059 crat_image = kmalloc(image_size, GFP_KERNEL);
1062 pr_err("No memory for allocating CRAT image\n");
1065 ret = kfd_topology_get_crat_acpi(crat_image, &image_size);
1068 down_write(&topology_lock);
1069 ret = kfd_parse_crat_table(crat_image);
1071 ret = kfd_topology_update_sysfs();
1072 up_write(&topology_lock);
1074 pr_err("Couldn't get CRAT table size from ACPI\n");
1077 } else if (ret == -ENODATA) {
1080 pr_err("Couldn't get CRAT table size from ACPI\n");
1084 pr_info("Finished initializing topology ret=%d\n", ret);
1088 void kfd_topology_shutdown(void)
1090 kfd_topology_release_sysfs();
1091 kfd_release_live_view();
1094 static void kfd_debug_print_topology(void)
1096 struct kfd_topology_device *dev;
1099 pr_info("DEBUG PRINT OF TOPOLOGY:");
1100 list_for_each_entry(dev, &topology_device_list, list) {
1101 pr_info("Node: %d\n", i);
1102 pr_info("\tGPU assigned: %s\n", (dev->gpu ? "yes" : "no"));
1103 pr_info("\tCPU count: %d\n", dev->node_props.cpu_cores_count);
1104 pr_info("\tSIMD count: %d", dev->node_props.simd_count);
1109 static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu)
1113 uint64_t local_mem_size;
1119 local_mem_size = gpu->kfd2kgd->get_vmem_size(gpu->kgd);
1121 buf[0] = gpu->pdev->devfn;
1122 buf[1] = gpu->pdev->subsystem_vendor;
1123 buf[2] = gpu->pdev->subsystem_device;
1124 buf[3] = gpu->pdev->device;
1125 buf[4] = gpu->pdev->bus->number;
1126 buf[5] = lower_32_bits(local_mem_size);
1127 buf[6] = upper_32_bits(local_mem_size);
1129 for (i = 0, hashout = 0; i < 7; i++)
1130 hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH);
1135 static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu)
1137 struct kfd_topology_device *dev;
1138 struct kfd_topology_device *out_dev = NULL;
1142 list_for_each_entry(dev, &topology_device_list, list)
1143 if (dev->gpu == NULL && dev->node_props.simd_count > 0) {
1152 static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival)
1155 * TODO: Generate an event for thunk about the arrival/removal
1160 int kfd_topology_add_device(struct kfd_dev *gpu)
1163 struct kfd_topology_device *dev;
1168 gpu_id = kfd_generate_gpu_id(gpu);
1170 pr_debug("kfd: Adding new GPU (ID: 0x%x) to topology\n", gpu_id);
1172 down_write(&topology_lock);
1174 * Try to assign the GPU to existing topology device (generated from
1177 dev = kfd_assign_gpu(gpu);
1179 pr_info("GPU was not found in the current topology. Extending.\n");
1180 kfd_debug_print_topology();
1181 dev = kfd_create_topology_device();
1189 * TODO: Make a call to retrieve topology information from the
1194 * Update the SYSFS tree, since we added another topology device
1196 if (kfd_topology_update_sysfs() < 0)
1197 kfd_topology_release_sysfs();
1201 dev->gpu_id = gpu_id;
1203 dev->node_props.vendor_id = gpu->pdev->vendor;
1204 dev->node_props.device_id = gpu->pdev->device;
1205 dev->node_props.location_id = (gpu->pdev->bus->number << 24) +
1206 (gpu->pdev->devfn & 0xffffff);
1208 * TODO: Retrieve max engine clock values from KGD
1211 if (dev->gpu->device_info->asic_family == CHIP_CARRIZO) {
1212 dev->node_props.capability |= HSA_CAP_DOORBELL_PACKET_TYPE;
1213 pr_info("amdkfd: adding doorbell packet type capability\n");
1219 up_write(&topology_lock);
1222 kfd_notify_gpu_change(gpu_id, 1);
1227 int kfd_topology_remove_device(struct kfd_dev *gpu)
1229 struct kfd_topology_device *dev;
1235 down_write(&topology_lock);
1237 list_for_each_entry(dev, &topology_device_list, list)
1238 if (dev->gpu == gpu) {
1239 gpu_id = dev->gpu_id;
1240 kfd_remove_sysfs_node_entry(dev);
1241 kfd_release_topology_device(dev);
1243 if (kfd_topology_update_sysfs() < 0)
1244 kfd_topology_release_sysfs();
1248 up_write(&topology_lock);
1251 kfd_notify_gpu_change(gpu_id, 0);
1257 * When idx is out of bounds, the function will return NULL
1259 struct kfd_dev *kfd_topology_enum_kfd_devices(uint8_t idx)
1262 struct kfd_topology_device *top_dev;
1263 struct kfd_dev *device = NULL;
1264 uint8_t device_idx = 0;
1266 down_read(&topology_lock);
1268 list_for_each_entry(top_dev, &topology_device_list, list) {
1269 if (device_idx == idx) {
1270 device = top_dev->gpu;
1277 up_read(&topology_lock);