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>
31 #include <linux/dmi.h>
32 #include <linux/atomic.h>
36 #include "kfd_topology.h"
37 #include "kfd_device_queue_manager.h"
38 #include "kfd_iommu.h"
39 #include "amdgpu_amdkfd.h"
41 /* topology_device_list - Master list of all topology devices */
42 static struct list_head topology_device_list;
43 static struct kfd_system_properties sys_props;
45 static DECLARE_RWSEM(topology_lock);
46 static atomic_t topology_crat_proximity_domain;
48 struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
49 uint32_t proximity_domain)
51 struct kfd_topology_device *top_dev;
52 struct kfd_topology_device *device = NULL;
54 down_read(&topology_lock);
56 list_for_each_entry(top_dev, &topology_device_list, list)
57 if (top_dev->proximity_domain == proximity_domain) {
62 up_read(&topology_lock);
67 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id)
69 struct kfd_topology_device *top_dev = NULL;
70 struct kfd_topology_device *ret = NULL;
72 down_read(&topology_lock);
74 list_for_each_entry(top_dev, &topology_device_list, list)
75 if (top_dev->gpu_id == gpu_id) {
80 up_read(&topology_lock);
85 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id)
87 struct kfd_topology_device *top_dev;
89 top_dev = kfd_topology_device_by_id(gpu_id);
96 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev)
98 struct kfd_topology_device *top_dev;
99 struct kfd_dev *device = NULL;
101 down_read(&topology_lock);
103 list_for_each_entry(top_dev, &topology_device_list, list)
104 if (top_dev->gpu && top_dev->gpu->pdev == pdev) {
105 device = top_dev->gpu;
109 up_read(&topology_lock);
114 struct kfd_dev *kfd_device_by_kgd(const struct kgd_dev *kgd)
116 struct kfd_topology_device *top_dev;
117 struct kfd_dev *device = NULL;
119 down_read(&topology_lock);
121 list_for_each_entry(top_dev, &topology_device_list, list)
122 if (top_dev->gpu && top_dev->gpu->kgd == kgd) {
123 device = top_dev->gpu;
127 up_read(&topology_lock);
132 /* Called with write topology_lock acquired */
133 static void kfd_release_topology_device(struct kfd_topology_device *dev)
135 struct kfd_mem_properties *mem;
136 struct kfd_cache_properties *cache;
137 struct kfd_iolink_properties *iolink;
138 struct kfd_perf_properties *perf;
140 list_del(&dev->list);
142 while (dev->mem_props.next != &dev->mem_props) {
143 mem = container_of(dev->mem_props.next,
144 struct kfd_mem_properties, list);
145 list_del(&mem->list);
149 while (dev->cache_props.next != &dev->cache_props) {
150 cache = container_of(dev->cache_props.next,
151 struct kfd_cache_properties, list);
152 list_del(&cache->list);
156 while (dev->io_link_props.next != &dev->io_link_props) {
157 iolink = container_of(dev->io_link_props.next,
158 struct kfd_iolink_properties, list);
159 list_del(&iolink->list);
163 while (dev->perf_props.next != &dev->perf_props) {
164 perf = container_of(dev->perf_props.next,
165 struct kfd_perf_properties, list);
166 list_del(&perf->list);
173 void kfd_release_topology_device_list(struct list_head *device_list)
175 struct kfd_topology_device *dev;
177 while (!list_empty(device_list)) {
178 dev = list_first_entry(device_list,
179 struct kfd_topology_device, list);
180 kfd_release_topology_device(dev);
184 static void kfd_release_live_view(void)
186 kfd_release_topology_device_list(&topology_device_list);
187 memset(&sys_props, 0, sizeof(sys_props));
190 struct kfd_topology_device *kfd_create_topology_device(
191 struct list_head *device_list)
193 struct kfd_topology_device *dev;
195 dev = kfd_alloc_struct(dev);
197 pr_err("No memory to allocate a topology device");
201 INIT_LIST_HEAD(&dev->mem_props);
202 INIT_LIST_HEAD(&dev->cache_props);
203 INIT_LIST_HEAD(&dev->io_link_props);
204 INIT_LIST_HEAD(&dev->perf_props);
206 list_add_tail(&dev->list, device_list);
212 #define sysfs_show_gen_prop(buffer, fmt, ...) \
213 snprintf(buffer, PAGE_SIZE, "%s"fmt, buffer, __VA_ARGS__)
214 #define sysfs_show_32bit_prop(buffer, name, value) \
215 sysfs_show_gen_prop(buffer, "%s %u\n", name, value)
216 #define sysfs_show_64bit_prop(buffer, name, value) \
217 sysfs_show_gen_prop(buffer, "%s %llu\n", name, value)
218 #define sysfs_show_32bit_val(buffer, value) \
219 sysfs_show_gen_prop(buffer, "%u\n", value)
220 #define sysfs_show_str_val(buffer, value) \
221 sysfs_show_gen_prop(buffer, "%s\n", value)
223 static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr,
228 /* Making sure that the buffer is an empty string */
231 if (attr == &sys_props.attr_genid) {
232 ret = sysfs_show_32bit_val(buffer, sys_props.generation_count);
233 } else if (attr == &sys_props.attr_props) {
234 sysfs_show_64bit_prop(buffer, "platform_oem",
235 sys_props.platform_oem);
236 sysfs_show_64bit_prop(buffer, "platform_id",
237 sys_props.platform_id);
238 ret = sysfs_show_64bit_prop(buffer, "platform_rev",
239 sys_props.platform_rev);
247 static void kfd_topology_kobj_release(struct kobject *kobj)
252 static const struct sysfs_ops sysprops_ops = {
253 .show = sysprops_show,
256 static struct kobj_type sysprops_type = {
257 .release = kfd_topology_kobj_release,
258 .sysfs_ops = &sysprops_ops,
261 static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr,
265 struct kfd_iolink_properties *iolink;
267 /* Making sure that the buffer is an empty string */
270 iolink = container_of(attr, struct kfd_iolink_properties, attr);
271 sysfs_show_32bit_prop(buffer, "type", iolink->iolink_type);
272 sysfs_show_32bit_prop(buffer, "version_major", iolink->ver_maj);
273 sysfs_show_32bit_prop(buffer, "version_minor", iolink->ver_min);
274 sysfs_show_32bit_prop(buffer, "node_from", iolink->node_from);
275 sysfs_show_32bit_prop(buffer, "node_to", iolink->node_to);
276 sysfs_show_32bit_prop(buffer, "weight", iolink->weight);
277 sysfs_show_32bit_prop(buffer, "min_latency", iolink->min_latency);
278 sysfs_show_32bit_prop(buffer, "max_latency", iolink->max_latency);
279 sysfs_show_32bit_prop(buffer, "min_bandwidth", iolink->min_bandwidth);
280 sysfs_show_32bit_prop(buffer, "max_bandwidth", iolink->max_bandwidth);
281 sysfs_show_32bit_prop(buffer, "recommended_transfer_size",
282 iolink->rec_transfer_size);
283 ret = sysfs_show_32bit_prop(buffer, "flags", iolink->flags);
288 static const struct sysfs_ops iolink_ops = {
292 static struct kobj_type iolink_type = {
293 .release = kfd_topology_kobj_release,
294 .sysfs_ops = &iolink_ops,
297 static ssize_t mem_show(struct kobject *kobj, struct attribute *attr,
301 struct kfd_mem_properties *mem;
303 /* Making sure that the buffer is an empty string */
306 mem = container_of(attr, struct kfd_mem_properties, attr);
307 sysfs_show_32bit_prop(buffer, "heap_type", mem->heap_type);
308 sysfs_show_64bit_prop(buffer, "size_in_bytes", mem->size_in_bytes);
309 sysfs_show_32bit_prop(buffer, "flags", mem->flags);
310 sysfs_show_32bit_prop(buffer, "width", mem->width);
311 ret = sysfs_show_32bit_prop(buffer, "mem_clk_max", mem->mem_clk_max);
316 static const struct sysfs_ops mem_ops = {
320 static struct kobj_type mem_type = {
321 .release = kfd_topology_kobj_release,
322 .sysfs_ops = &mem_ops,
325 static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr,
330 struct kfd_cache_properties *cache;
332 /* Making sure that the buffer is an empty string */
335 cache = container_of(attr, struct kfd_cache_properties, attr);
336 sysfs_show_32bit_prop(buffer, "processor_id_low",
337 cache->processor_id_low);
338 sysfs_show_32bit_prop(buffer, "level", cache->cache_level);
339 sysfs_show_32bit_prop(buffer, "size", cache->cache_size);
340 sysfs_show_32bit_prop(buffer, "cache_line_size", cache->cacheline_size);
341 sysfs_show_32bit_prop(buffer, "cache_lines_per_tag",
342 cache->cachelines_per_tag);
343 sysfs_show_32bit_prop(buffer, "association", cache->cache_assoc);
344 sysfs_show_32bit_prop(buffer, "latency", cache->cache_latency);
345 sysfs_show_32bit_prop(buffer, "type", cache->cache_type);
346 snprintf(buffer, PAGE_SIZE, "%ssibling_map ", buffer);
347 for (i = 0; i < CRAT_SIBLINGMAP_SIZE; i++)
348 for (j = 0; j < sizeof(cache->sibling_map[0])*8; j++) {
350 if (cache->sibling_map[i] & (1 << j))
351 ret = snprintf(buffer, PAGE_SIZE,
352 "%s%d%s", buffer, 1, ",");
354 ret = snprintf(buffer, PAGE_SIZE,
355 "%s%d%s", buffer, 0, ",");
357 /* Replace the last "," with end of line */
358 *(buffer + strlen(buffer) - 1) = 0xA;
362 static const struct sysfs_ops cache_ops = {
363 .show = kfd_cache_show,
366 static struct kobj_type cache_type = {
367 .release = kfd_topology_kobj_release,
368 .sysfs_ops = &cache_ops,
371 /****** Sysfs of Performance Counters ******/
373 struct kfd_perf_attr {
374 struct kobj_attribute attr;
378 static ssize_t perf_show(struct kobject *kobj, struct kobj_attribute *attrs,
381 struct kfd_perf_attr *attr;
384 attr = container_of(attrs, struct kfd_perf_attr, attr);
385 if (!attr->data) /* invalid data for PMC */
388 return sysfs_show_32bit_val(buf, attr->data);
391 #define KFD_PERF_DESC(_name, _data) \
393 .attr = __ATTR(_name, 0444, perf_show, NULL), \
397 static struct kfd_perf_attr perf_attr_iommu[] = {
398 KFD_PERF_DESC(max_concurrent, 0),
399 KFD_PERF_DESC(num_counters, 0),
400 KFD_PERF_DESC(counter_ids, 0),
402 /****************************************/
404 static ssize_t node_show(struct kobject *kobj, struct attribute *attr,
407 struct kfd_topology_device *dev;
408 char public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE];
410 uint32_t log_max_watch_addr;
412 /* Making sure that the buffer is an empty string */
415 if (strcmp(attr->name, "gpu_id") == 0) {
416 dev = container_of(attr, struct kfd_topology_device,
418 return sysfs_show_32bit_val(buffer, dev->gpu_id);
421 if (strcmp(attr->name, "name") == 0) {
422 dev = container_of(attr, struct kfd_topology_device,
424 for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE; i++) {
426 (char)dev->node_props.marketing_name[i];
427 if (dev->node_props.marketing_name[i] == 0)
430 public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1] = 0x0;
431 return sysfs_show_str_val(buffer, public_name);
434 dev = container_of(attr, struct kfd_topology_device,
436 sysfs_show_32bit_prop(buffer, "cpu_cores_count",
437 dev->node_props.cpu_cores_count);
438 sysfs_show_32bit_prop(buffer, "simd_count",
439 dev->node_props.simd_count);
440 sysfs_show_32bit_prop(buffer, "mem_banks_count",
441 dev->node_props.mem_banks_count);
442 sysfs_show_32bit_prop(buffer, "caches_count",
443 dev->node_props.caches_count);
444 sysfs_show_32bit_prop(buffer, "io_links_count",
445 dev->node_props.io_links_count);
446 sysfs_show_32bit_prop(buffer, "cpu_core_id_base",
447 dev->node_props.cpu_core_id_base);
448 sysfs_show_32bit_prop(buffer, "simd_id_base",
449 dev->node_props.simd_id_base);
450 sysfs_show_32bit_prop(buffer, "max_waves_per_simd",
451 dev->node_props.max_waves_per_simd);
452 sysfs_show_32bit_prop(buffer, "lds_size_in_kb",
453 dev->node_props.lds_size_in_kb);
454 sysfs_show_32bit_prop(buffer, "gds_size_in_kb",
455 dev->node_props.gds_size_in_kb);
456 sysfs_show_32bit_prop(buffer, "wave_front_size",
457 dev->node_props.wave_front_size);
458 sysfs_show_32bit_prop(buffer, "array_count",
459 dev->node_props.array_count);
460 sysfs_show_32bit_prop(buffer, "simd_arrays_per_engine",
461 dev->node_props.simd_arrays_per_engine);
462 sysfs_show_32bit_prop(buffer, "cu_per_simd_array",
463 dev->node_props.cu_per_simd_array);
464 sysfs_show_32bit_prop(buffer, "simd_per_cu",
465 dev->node_props.simd_per_cu);
466 sysfs_show_32bit_prop(buffer, "max_slots_scratch_cu",
467 dev->node_props.max_slots_scratch_cu);
468 sysfs_show_32bit_prop(buffer, "vendor_id",
469 dev->node_props.vendor_id);
470 sysfs_show_32bit_prop(buffer, "device_id",
471 dev->node_props.device_id);
472 sysfs_show_32bit_prop(buffer, "location_id",
473 dev->node_props.location_id);
474 sysfs_show_32bit_prop(buffer, "drm_render_minor",
475 dev->node_props.drm_render_minor);
476 sysfs_show_64bit_prop(buffer, "hive_id",
477 dev->node_props.hive_id);
481 __ilog2_u32(dev->gpu->device_info->num_of_watch_points);
483 if (log_max_watch_addr) {
484 dev->node_props.capability |=
485 HSA_CAP_WATCH_POINTS_SUPPORTED;
487 dev->node_props.capability |=
488 ((log_max_watch_addr <<
489 HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) &
490 HSA_CAP_WATCH_POINTS_TOTALBITS_MASK);
493 if (dev->gpu->device_info->asic_family == CHIP_TONGA)
494 dev->node_props.capability |=
495 HSA_CAP_AQL_QUEUE_DOUBLE_MAP;
497 sysfs_show_32bit_prop(buffer, "max_engine_clk_fcompute",
498 dev->node_props.max_engine_clk_fcompute);
500 sysfs_show_64bit_prop(buffer, "local_mem_size",
501 (unsigned long long int) 0);
503 sysfs_show_32bit_prop(buffer, "fw_version",
504 dev->gpu->mec_fw_version);
505 sysfs_show_32bit_prop(buffer, "capability",
506 dev->node_props.capability);
507 sysfs_show_32bit_prop(buffer, "sdma_fw_version",
508 dev->gpu->sdma_fw_version);
511 return sysfs_show_32bit_prop(buffer, "max_engine_clk_ccompute",
512 cpufreq_quick_get_max(0)/1000);
515 static const struct sysfs_ops node_ops = {
519 static struct kobj_type node_type = {
520 .release = kfd_topology_kobj_release,
521 .sysfs_ops = &node_ops,
524 static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr)
526 sysfs_remove_file(kobj, attr);
531 static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev)
533 struct kfd_iolink_properties *iolink;
534 struct kfd_cache_properties *cache;
535 struct kfd_mem_properties *mem;
536 struct kfd_perf_properties *perf;
538 if (dev->kobj_iolink) {
539 list_for_each_entry(iolink, &dev->io_link_props, list)
541 kfd_remove_sysfs_file(iolink->kobj,
545 kobject_del(dev->kobj_iolink);
546 kobject_put(dev->kobj_iolink);
547 dev->kobj_iolink = NULL;
550 if (dev->kobj_cache) {
551 list_for_each_entry(cache, &dev->cache_props, list)
553 kfd_remove_sysfs_file(cache->kobj,
557 kobject_del(dev->kobj_cache);
558 kobject_put(dev->kobj_cache);
559 dev->kobj_cache = NULL;
563 list_for_each_entry(mem, &dev->mem_props, list)
565 kfd_remove_sysfs_file(mem->kobj, &mem->attr);
568 kobject_del(dev->kobj_mem);
569 kobject_put(dev->kobj_mem);
570 dev->kobj_mem = NULL;
573 if (dev->kobj_perf) {
574 list_for_each_entry(perf, &dev->perf_props, list) {
575 kfree(perf->attr_group);
576 perf->attr_group = NULL;
578 kobject_del(dev->kobj_perf);
579 kobject_put(dev->kobj_perf);
580 dev->kobj_perf = NULL;
583 if (dev->kobj_node) {
584 sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid);
585 sysfs_remove_file(dev->kobj_node, &dev->attr_name);
586 sysfs_remove_file(dev->kobj_node, &dev->attr_props);
587 kobject_del(dev->kobj_node);
588 kobject_put(dev->kobj_node);
589 dev->kobj_node = NULL;
593 static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev,
596 struct kfd_iolink_properties *iolink;
597 struct kfd_cache_properties *cache;
598 struct kfd_mem_properties *mem;
599 struct kfd_perf_properties *perf;
601 uint32_t i, num_attrs;
602 struct attribute **attrs;
604 if (WARN_ON(dev->kobj_node))
608 * Creating the sysfs folders
610 dev->kobj_node = kfd_alloc_struct(dev->kobj_node);
614 ret = kobject_init_and_add(dev->kobj_node, &node_type,
615 sys_props.kobj_nodes, "%d", id);
619 dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node);
623 dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node);
624 if (!dev->kobj_cache)
627 dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node);
628 if (!dev->kobj_iolink)
631 dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node);
636 * Creating sysfs files for node properties
638 dev->attr_gpuid.name = "gpu_id";
639 dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE;
640 sysfs_attr_init(&dev->attr_gpuid);
641 dev->attr_name.name = "name";
642 dev->attr_name.mode = KFD_SYSFS_FILE_MODE;
643 sysfs_attr_init(&dev->attr_name);
644 dev->attr_props.name = "properties";
645 dev->attr_props.mode = KFD_SYSFS_FILE_MODE;
646 sysfs_attr_init(&dev->attr_props);
647 ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid);
650 ret = sysfs_create_file(dev->kobj_node, &dev->attr_name);
653 ret = sysfs_create_file(dev->kobj_node, &dev->attr_props);
658 list_for_each_entry(mem, &dev->mem_props, list) {
659 mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
662 ret = kobject_init_and_add(mem->kobj, &mem_type,
663 dev->kobj_mem, "%d", i);
667 mem->attr.name = "properties";
668 mem->attr.mode = KFD_SYSFS_FILE_MODE;
669 sysfs_attr_init(&mem->attr);
670 ret = sysfs_create_file(mem->kobj, &mem->attr);
677 list_for_each_entry(cache, &dev->cache_props, list) {
678 cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
681 ret = kobject_init_and_add(cache->kobj, &cache_type,
682 dev->kobj_cache, "%d", i);
686 cache->attr.name = "properties";
687 cache->attr.mode = KFD_SYSFS_FILE_MODE;
688 sysfs_attr_init(&cache->attr);
689 ret = sysfs_create_file(cache->kobj, &cache->attr);
696 list_for_each_entry(iolink, &dev->io_link_props, list) {
697 iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
700 ret = kobject_init_and_add(iolink->kobj, &iolink_type,
701 dev->kobj_iolink, "%d", i);
705 iolink->attr.name = "properties";
706 iolink->attr.mode = KFD_SYSFS_FILE_MODE;
707 sysfs_attr_init(&iolink->attr);
708 ret = sysfs_create_file(iolink->kobj, &iolink->attr);
714 /* All hardware blocks have the same number of attributes. */
715 num_attrs = ARRAY_SIZE(perf_attr_iommu);
716 list_for_each_entry(perf, &dev->perf_props, list) {
717 perf->attr_group = kzalloc(sizeof(struct kfd_perf_attr)
718 * num_attrs + sizeof(struct attribute_group),
720 if (!perf->attr_group)
723 attrs = (struct attribute **)(perf->attr_group + 1);
724 if (!strcmp(perf->block_name, "iommu")) {
725 /* Information of IOMMU's num_counters and counter_ids is shown
726 * under /sys/bus/event_source/devices/amd_iommu. We don't
729 perf_attr_iommu[0].data = perf->max_concurrent;
730 for (i = 0; i < num_attrs; i++)
731 attrs[i] = &perf_attr_iommu[i].attr.attr;
733 perf->attr_group->name = perf->block_name;
734 perf->attr_group->attrs = attrs;
735 ret = sysfs_create_group(dev->kobj_perf, perf->attr_group);
743 /* Called with write topology lock acquired */
744 static int kfd_build_sysfs_node_tree(void)
746 struct kfd_topology_device *dev;
750 list_for_each_entry(dev, &topology_device_list, list) {
751 ret = kfd_build_sysfs_node_entry(dev, i);
760 /* Called with write topology lock acquired */
761 static void kfd_remove_sysfs_node_tree(void)
763 struct kfd_topology_device *dev;
765 list_for_each_entry(dev, &topology_device_list, list)
766 kfd_remove_sysfs_node_entry(dev);
769 static int kfd_topology_update_sysfs(void)
773 pr_info("Creating topology SYSFS entries\n");
774 if (!sys_props.kobj_topology) {
775 sys_props.kobj_topology =
776 kfd_alloc_struct(sys_props.kobj_topology);
777 if (!sys_props.kobj_topology)
780 ret = kobject_init_and_add(sys_props.kobj_topology,
781 &sysprops_type, &kfd_device->kobj,
786 sys_props.kobj_nodes = kobject_create_and_add("nodes",
787 sys_props.kobj_topology);
788 if (!sys_props.kobj_nodes)
791 sys_props.attr_genid.name = "generation_id";
792 sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE;
793 sysfs_attr_init(&sys_props.attr_genid);
794 ret = sysfs_create_file(sys_props.kobj_topology,
795 &sys_props.attr_genid);
799 sys_props.attr_props.name = "system_properties";
800 sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE;
801 sysfs_attr_init(&sys_props.attr_props);
802 ret = sysfs_create_file(sys_props.kobj_topology,
803 &sys_props.attr_props);
808 kfd_remove_sysfs_node_tree();
810 return kfd_build_sysfs_node_tree();
813 static void kfd_topology_release_sysfs(void)
815 kfd_remove_sysfs_node_tree();
816 if (sys_props.kobj_topology) {
817 sysfs_remove_file(sys_props.kobj_topology,
818 &sys_props.attr_genid);
819 sysfs_remove_file(sys_props.kobj_topology,
820 &sys_props.attr_props);
821 if (sys_props.kobj_nodes) {
822 kobject_del(sys_props.kobj_nodes);
823 kobject_put(sys_props.kobj_nodes);
824 sys_props.kobj_nodes = NULL;
826 kobject_del(sys_props.kobj_topology);
827 kobject_put(sys_props.kobj_topology);
828 sys_props.kobj_topology = NULL;
832 /* Called with write topology_lock acquired */
833 static void kfd_topology_update_device_list(struct list_head *temp_list,
834 struct list_head *master_list)
836 while (!list_empty(temp_list)) {
837 list_move_tail(temp_list->next, master_list);
838 sys_props.num_devices++;
842 static void kfd_debug_print_topology(void)
844 struct kfd_topology_device *dev;
846 down_read(&topology_lock);
848 dev = list_last_entry(&topology_device_list,
849 struct kfd_topology_device, list);
851 if (dev->node_props.cpu_cores_count &&
852 dev->node_props.simd_count) {
853 pr_info("Topology: Add APU node [0x%0x:0x%0x]\n",
854 dev->node_props.device_id,
855 dev->node_props.vendor_id);
856 } else if (dev->node_props.cpu_cores_count)
857 pr_info("Topology: Add CPU node\n");
858 else if (dev->node_props.simd_count)
859 pr_info("Topology: Add dGPU node [0x%0x:0x%0x]\n",
860 dev->node_props.device_id,
861 dev->node_props.vendor_id);
863 up_read(&topology_lock);
866 /* Helper function for intializing platform_xx members of
867 * kfd_system_properties. Uses OEM info from the last CPU/APU node.
869 static void kfd_update_system_properties(void)
871 struct kfd_topology_device *dev;
873 down_read(&topology_lock);
874 dev = list_last_entry(&topology_device_list,
875 struct kfd_topology_device, list);
877 sys_props.platform_id =
878 (*((uint64_t *)dev->oem_id)) & CRAT_OEMID_64BIT_MASK;
879 sys_props.platform_oem = *((uint64_t *)dev->oem_table_id);
880 sys_props.platform_rev = dev->oem_revision;
882 up_read(&topology_lock);
885 static void find_system_memory(const struct dmi_header *dm,
888 struct kfd_mem_properties *mem;
889 u16 mem_width, mem_clock;
890 struct kfd_topology_device *kdev =
891 (struct kfd_topology_device *)private;
892 const u8 *dmi_data = (const u8 *)(dm + 1);
894 if (dm->type == DMI_ENTRY_MEM_DEVICE && dm->length >= 0x15) {
895 mem_width = (u16)(*(const u16 *)(dmi_data + 0x6));
896 mem_clock = (u16)(*(const u16 *)(dmi_data + 0x11));
897 list_for_each_entry(mem, &kdev->mem_props, list) {
898 if (mem_width != 0xFFFF && mem_width != 0)
899 mem->width = mem_width;
901 mem->mem_clk_max = mem_clock;
907 * Performance counters information is not part of CRAT but we would like to
908 * put them in the sysfs under topology directory for Thunk to get the data.
909 * This function is called before updating the sysfs.
911 static int kfd_add_perf_to_topology(struct kfd_topology_device *kdev)
913 /* These are the only counters supported so far */
914 return kfd_iommu_add_perf_counters(kdev);
917 /* kfd_add_non_crat_information - Add information that is not currently
918 * defined in CRAT but is necessary for KFD topology
919 * @dev - topology device to which addition info is added
921 static void kfd_add_non_crat_information(struct kfd_topology_device *kdev)
923 /* Check if CPU only node. */
925 /* Add system memory information */
926 dmi_walk(find_system_memory, kdev);
928 /* TODO: For GPU node, rearrange code from kfd_topology_add_device */
931 /* kfd_is_acpi_crat_invalid - CRAT from ACPI is valid only for AMD APU devices.
932 * Ignore CRAT for all other devices. AMD APU is identified if both CPU
933 * and GPU cores are present.
934 * @device_list - topology device list created by parsing ACPI CRAT table.
935 * @return - TRUE if invalid, FALSE is valid.
937 static bool kfd_is_acpi_crat_invalid(struct list_head *device_list)
939 struct kfd_topology_device *dev;
941 list_for_each_entry(dev, device_list, list) {
942 if (dev->node_props.cpu_cores_count &&
943 dev->node_props.simd_count)
946 pr_info("Ignoring ACPI CRAT on non-APU system\n");
950 int kfd_topology_init(void)
952 void *crat_image = NULL;
953 size_t image_size = 0;
955 struct list_head temp_topology_device_list;
956 int cpu_only_node = 0;
957 struct kfd_topology_device *kdev;
958 int proximity_domain;
960 /* topology_device_list - Master list of all topology devices
961 * temp_topology_device_list - temporary list created while parsing CRAT
962 * or VCRAT. Once parsing is complete the contents of list is moved to
963 * topology_device_list
966 /* Initialize the head for the both the lists */
967 INIT_LIST_HEAD(&topology_device_list);
968 INIT_LIST_HEAD(&temp_topology_device_list);
969 init_rwsem(&topology_lock);
971 memset(&sys_props, 0, sizeof(sys_props));
973 /* Proximity domains in ACPI CRAT tables start counting at
974 * 0. The same should be true for virtual CRAT tables created
975 * at this stage. GPUs added later in kfd_topology_add_device
978 proximity_domain = 0;
981 * Get the CRAT image from the ACPI. If ACPI doesn't have one
982 * or if ACPI CRAT is invalid create a virtual CRAT.
983 * NOTE: The current implementation expects all AMD APUs to have
984 * CRAT. If no CRAT is available, it is assumed to be a CPU
986 ret = kfd_create_crat_image_acpi(&crat_image, &image_size);
988 ret = kfd_parse_crat_table(crat_image,
989 &temp_topology_device_list,
992 kfd_is_acpi_crat_invalid(&temp_topology_device_list)) {
993 kfd_release_topology_device_list(
994 &temp_topology_device_list);
995 kfd_destroy_crat_image(crat_image);
1001 ret = kfd_create_crat_image_virtual(&crat_image, &image_size,
1002 COMPUTE_UNIT_CPU, NULL,
1006 pr_err("Error creating VCRAT table for CPU\n");
1010 ret = kfd_parse_crat_table(crat_image,
1011 &temp_topology_device_list,
1014 pr_err("Error parsing VCRAT table for CPU\n");
1019 kdev = list_first_entry(&temp_topology_device_list,
1020 struct kfd_topology_device, list);
1021 kfd_add_perf_to_topology(kdev);
1023 down_write(&topology_lock);
1024 kfd_topology_update_device_list(&temp_topology_device_list,
1025 &topology_device_list);
1026 atomic_set(&topology_crat_proximity_domain, sys_props.num_devices-1);
1027 ret = kfd_topology_update_sysfs();
1028 up_write(&topology_lock);
1031 sys_props.generation_count++;
1032 kfd_update_system_properties();
1033 kfd_debug_print_topology();
1034 pr_info("Finished initializing topology\n");
1036 pr_err("Failed to update topology in sysfs ret=%d\n", ret);
1038 /* For nodes with GPU, this information gets added
1039 * when GPU is detected (kfd_topology_add_device).
1041 if (cpu_only_node) {
1042 /* Add additional information to CPU only node created above */
1043 down_write(&topology_lock);
1044 kdev = list_first_entry(&topology_device_list,
1045 struct kfd_topology_device, list);
1046 up_write(&topology_lock);
1047 kfd_add_non_crat_information(kdev);
1051 kfd_destroy_crat_image(crat_image);
1055 void kfd_topology_shutdown(void)
1057 down_write(&topology_lock);
1058 kfd_topology_release_sysfs();
1059 kfd_release_live_view();
1060 up_write(&topology_lock);
1063 static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu)
1067 uint64_t local_mem_size;
1069 struct kfd_local_mem_info local_mem_info;
1074 amdgpu_amdkfd_get_local_mem_info(gpu->kgd, &local_mem_info);
1076 local_mem_size = local_mem_info.local_mem_size_private +
1077 local_mem_info.local_mem_size_public;
1079 buf[0] = gpu->pdev->devfn;
1080 buf[1] = gpu->pdev->subsystem_vendor;
1081 buf[2] = gpu->pdev->subsystem_device;
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);
1087 for (i = 0, hashout = 0; i < 7; i++)
1088 hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH);
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.
1097 static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu)
1099 struct kfd_topology_device *dev;
1100 struct kfd_topology_device *out_dev = NULL;
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.
1107 if (!gpu->device_info->needs_iommu_device &&
1108 dev->node_props.cpu_cores_count)
1111 if (!dev->gpu && (dev->node_props.simd_count > 0)) {
1117 up_write(&topology_lock);
1121 static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival)
1124 * TODO: Generate an event for thunk about the arrival/removal
1129 /* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info,
1130 * patch this after CRAT parsing.
1132 static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev)
1134 struct kfd_mem_properties *mem;
1135 struct kfd_local_mem_info local_mem_info;
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
1146 amdgpu_amdkfd_get_local_mem_info(dev->gpu->kgd, &local_mem_info);
1148 list_for_each_entry(mem, &dev->mem_props, list)
1149 mem->mem_clk_max = local_mem_info.mem_clk_max;
1152 static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev)
1154 struct kfd_iolink_properties *link, *cpu_link;
1155 struct kfd_topology_device *cpu_dev;
1157 uint32_t cpu_flag = CRAT_IOLINK_FLAGS_ENABLED;
1158 uint32_t flag = CRAT_IOLINK_FLAGS_ENABLED;
1160 if (!dev || !dev->gpu)
1163 pcie_capability_read_dword(dev->gpu->pdev,
1164 PCI_EXP_DEVCAP2, &cap);
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;
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;
1176 /* GPU only creates direct links so apply flags setting to all */
1177 list_for_each_entry(link, &dev->io_link_props, list) {
1179 cpu_dev = kfd_topology_device_by_proximity_domain(
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;
1190 int kfd_topology_add_device(struct kfd_dev *gpu)
1193 struct kfd_topology_device *dev;
1194 struct kfd_cu_info cu_info;
1196 struct list_head temp_topology_device_list;
1197 void *crat_image = NULL;
1198 size_t image_size = 0;
1199 int proximity_domain;
1201 INIT_LIST_HEAD(&temp_topology_device_list);
1203 gpu_id = kfd_generate_gpu_id(gpu);
1205 pr_debug("Adding new GPU (ID: 0x%x) to topology\n", gpu_id);
1207 proximity_domain = atomic_inc_return(&topology_crat_proximity_domain);
1209 /* Check to see if this gpu device exists in the topology_device_list.
1210 * If so, assign the gpu to that device,
1211 * else create a Virtual CRAT for this gpu device and then parse that
1212 * CRAT to create a new topology device. Once created assign the gpu to
1213 * that topology device
1215 dev = kfd_assign_gpu(gpu);
1217 res = kfd_create_crat_image_virtual(&crat_image, &image_size,
1218 COMPUTE_UNIT_GPU, gpu,
1221 pr_err("Error creating VCRAT for GPU (ID: 0x%x)\n",
1225 res = kfd_parse_crat_table(crat_image,
1226 &temp_topology_device_list,
1229 pr_err("Error parsing VCRAT for GPU (ID: 0x%x)\n",
1234 down_write(&topology_lock);
1235 kfd_topology_update_device_list(&temp_topology_device_list,
1236 &topology_device_list);
1238 /* Update the SYSFS tree, since we added another topology
1241 res = kfd_topology_update_sysfs();
1242 up_write(&topology_lock);
1245 sys_props.generation_count++;
1247 pr_err("Failed to update GPU (ID: 0x%x) to sysfs topology. res=%d\n",
1249 dev = kfd_assign_gpu(gpu);
1250 if (WARN_ON(!dev)) {
1256 dev->gpu_id = gpu_id;
1259 /* TODO: Move the following lines to function
1260 * kfd_add_non_crat_information
1263 /* Fill-in additional information that is not available in CRAT but
1264 * needed for the topology
1267 amdgpu_amdkfd_get_cu_info(dev->gpu->kgd, &cu_info);
1268 dev->node_props.simd_arrays_per_engine =
1269 cu_info.num_shader_arrays_per_engine;
1271 dev->node_props.vendor_id = gpu->pdev->vendor;
1272 dev->node_props.device_id = gpu->pdev->device;
1273 dev->node_props.location_id = PCI_DEVID(gpu->pdev->bus->number,
1275 dev->node_props.max_engine_clk_fcompute =
1276 amdgpu_amdkfd_get_max_engine_clock_in_mhz(dev->gpu->kgd);
1277 dev->node_props.max_engine_clk_ccompute =
1278 cpufreq_quick_get_max(0) / 1000;
1279 dev->node_props.drm_render_minor =
1280 gpu->shared_resources.drm_render_minor;
1282 dev->node_props.hive_id = gpu->hive_id;
1284 kfd_fill_mem_clk_max_info(dev);
1285 kfd_fill_iolink_non_crat_info(dev);
1287 switch (dev->gpu->device_info->asic_family) {
1291 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_PRE_1_0 <<
1292 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1293 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1297 case CHIP_POLARIS10:
1298 case CHIP_POLARIS11:
1299 case CHIP_POLARIS12:
1300 pr_debug("Adding doorbell packet type capability\n");
1301 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 <<
1302 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1303 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1309 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 <<
1310 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1311 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1314 WARN(1, "Unexpected ASIC family %u",
1315 dev->gpu->device_info->asic_family);
1318 /* Fix errors in CZ CRAT.
1319 * simd_count: Carrizo CRAT reports wrong simd_count, probably
1320 * because it doesn't consider masked out CUs
1321 * max_waves_per_simd: Carrizo reports wrong max_waves_per_simd
1322 * capability flag: Carrizo CRAT doesn't report IOMMU flags
1324 if (dev->gpu->device_info->asic_family == CHIP_CARRIZO) {
1325 dev->node_props.simd_count =
1326 cu_info.simd_per_cu * cu_info.cu_active_number;
1327 dev->node_props.max_waves_per_simd = 10;
1328 dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
1331 kfd_debug_print_topology();
1334 kfd_notify_gpu_change(gpu_id, 1);
1336 kfd_destroy_crat_image(crat_image);
1340 int kfd_topology_remove_device(struct kfd_dev *gpu)
1342 struct kfd_topology_device *dev, *tmp;
1346 down_write(&topology_lock);
1348 list_for_each_entry_safe(dev, tmp, &topology_device_list, list)
1349 if (dev->gpu == gpu) {
1350 gpu_id = dev->gpu_id;
1351 kfd_remove_sysfs_node_entry(dev);
1352 kfd_release_topology_device(dev);
1353 sys_props.num_devices--;
1355 if (kfd_topology_update_sysfs() < 0)
1356 kfd_topology_release_sysfs();
1360 up_write(&topology_lock);
1363 kfd_notify_gpu_change(gpu_id, 0);
1368 /* kfd_topology_enum_kfd_devices - Enumerate through all devices in KFD
1369 * topology. If GPU device is found @idx, then valid kfd_dev pointer is
1370 * returned through @kdev
1371 * Return - 0: On success (@kdev will be NULL for non GPU nodes)
1372 * -1: If end of list
1374 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev)
1377 struct kfd_topology_device *top_dev;
1378 uint8_t device_idx = 0;
1381 down_read(&topology_lock);
1383 list_for_each_entry(top_dev, &topology_device_list, list) {
1384 if (device_idx == idx) {
1385 *kdev = top_dev->gpu;
1386 up_read(&topology_lock);
1393 up_read(&topology_lock);
1399 static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask)
1401 int first_cpu_of_numa_node;
1403 if (!cpumask || cpumask == cpu_none_mask)
1405 first_cpu_of_numa_node = cpumask_first(cpumask);
1406 if (first_cpu_of_numa_node >= nr_cpu_ids)
1408 #ifdef CONFIG_X86_64
1409 return cpu_data(first_cpu_of_numa_node).apicid;
1411 return first_cpu_of_numa_node;
1415 /* kfd_numa_node_to_apic_id - Returns the APIC ID of the first logical processor
1416 * of the given NUMA node (numa_node_id)
1417 * Return -1 on failure
1419 int kfd_numa_node_to_apic_id(int numa_node_id)
1421 if (numa_node_id == -1) {
1422 pr_warn("Invalid NUMA Node. Use online CPU mask\n");
1423 return kfd_cpumask_to_apic_id(cpu_online_mask);
1425 return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id));
1428 #if defined(CONFIG_DEBUG_FS)
1430 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data)
1432 struct kfd_topology_device *dev;
1436 down_read(&topology_lock);
1438 list_for_each_entry(dev, &topology_device_list, list) {
1444 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1445 r = dqm_debugfs_hqds(m, dev->gpu->dqm);
1450 up_read(&topology_lock);
1455 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data)
1457 struct kfd_topology_device *dev;
1461 down_read(&topology_lock);
1463 list_for_each_entry(dev, &topology_device_list, list) {
1469 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1470 r = pm_debugfs_runlist(m, &dev->gpu->dqm->packets);
1475 up_read(&topology_lock);