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"
40 /* topology_device_list - Master list of all topology devices */
41 static struct list_head topology_device_list;
42 static struct kfd_system_properties sys_props;
44 static DECLARE_RWSEM(topology_lock);
45 static atomic_t topology_crat_proximity_domain;
47 struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
48 uint32_t proximity_domain)
50 struct kfd_topology_device *top_dev;
51 struct kfd_topology_device *device = NULL;
53 down_read(&topology_lock);
55 list_for_each_entry(top_dev, &topology_device_list, list)
56 if (top_dev->proximity_domain == proximity_domain) {
61 up_read(&topology_lock);
66 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id)
68 struct kfd_topology_device *top_dev = NULL;
69 struct kfd_topology_device *ret = NULL;
71 down_read(&topology_lock);
73 list_for_each_entry(top_dev, &topology_device_list, list)
74 if (top_dev->gpu_id == gpu_id) {
79 up_read(&topology_lock);
84 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id)
86 struct kfd_topology_device *top_dev;
88 top_dev = kfd_topology_device_by_id(gpu_id);
95 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev)
97 struct kfd_topology_device *top_dev;
98 struct kfd_dev *device = NULL;
100 down_read(&topology_lock);
102 list_for_each_entry(top_dev, &topology_device_list, list)
103 if (top_dev->gpu->pdev == pdev) {
104 device = top_dev->gpu;
108 up_read(&topology_lock);
113 /* Called with write topology_lock acquired */
114 static void kfd_release_topology_device(struct kfd_topology_device *dev)
116 struct kfd_mem_properties *mem;
117 struct kfd_cache_properties *cache;
118 struct kfd_iolink_properties *iolink;
119 struct kfd_perf_properties *perf;
121 list_del(&dev->list);
123 while (dev->mem_props.next != &dev->mem_props) {
124 mem = container_of(dev->mem_props.next,
125 struct kfd_mem_properties, list);
126 list_del(&mem->list);
130 while (dev->cache_props.next != &dev->cache_props) {
131 cache = container_of(dev->cache_props.next,
132 struct kfd_cache_properties, list);
133 list_del(&cache->list);
137 while (dev->io_link_props.next != &dev->io_link_props) {
138 iolink = container_of(dev->io_link_props.next,
139 struct kfd_iolink_properties, list);
140 list_del(&iolink->list);
144 while (dev->perf_props.next != &dev->perf_props) {
145 perf = container_of(dev->perf_props.next,
146 struct kfd_perf_properties, list);
147 list_del(&perf->list);
154 void kfd_release_topology_device_list(struct list_head *device_list)
156 struct kfd_topology_device *dev;
158 while (!list_empty(device_list)) {
159 dev = list_first_entry(device_list,
160 struct kfd_topology_device, list);
161 kfd_release_topology_device(dev);
165 static void kfd_release_live_view(void)
167 kfd_release_topology_device_list(&topology_device_list);
168 memset(&sys_props, 0, sizeof(sys_props));
171 struct kfd_topology_device *kfd_create_topology_device(
172 struct list_head *device_list)
174 struct kfd_topology_device *dev;
176 dev = kfd_alloc_struct(dev);
178 pr_err("No memory to allocate a topology device");
182 INIT_LIST_HEAD(&dev->mem_props);
183 INIT_LIST_HEAD(&dev->cache_props);
184 INIT_LIST_HEAD(&dev->io_link_props);
185 INIT_LIST_HEAD(&dev->perf_props);
187 list_add_tail(&dev->list, device_list);
193 #define sysfs_show_gen_prop(buffer, fmt, ...) \
194 snprintf(buffer, PAGE_SIZE, "%s"fmt, buffer, __VA_ARGS__)
195 #define sysfs_show_32bit_prop(buffer, name, value) \
196 sysfs_show_gen_prop(buffer, "%s %u\n", name, value)
197 #define sysfs_show_64bit_prop(buffer, name, value) \
198 sysfs_show_gen_prop(buffer, "%s %llu\n", name, value)
199 #define sysfs_show_32bit_val(buffer, value) \
200 sysfs_show_gen_prop(buffer, "%u\n", value)
201 #define sysfs_show_str_val(buffer, value) \
202 sysfs_show_gen_prop(buffer, "%s\n", value)
204 static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr,
209 /* Making sure that the buffer is an empty string */
212 if (attr == &sys_props.attr_genid) {
213 ret = sysfs_show_32bit_val(buffer, sys_props.generation_count);
214 } else if (attr == &sys_props.attr_props) {
215 sysfs_show_64bit_prop(buffer, "platform_oem",
216 sys_props.platform_oem);
217 sysfs_show_64bit_prop(buffer, "platform_id",
218 sys_props.platform_id);
219 ret = sysfs_show_64bit_prop(buffer, "platform_rev",
220 sys_props.platform_rev);
228 static void kfd_topology_kobj_release(struct kobject *kobj)
233 static const struct sysfs_ops sysprops_ops = {
234 .show = sysprops_show,
237 static struct kobj_type sysprops_type = {
238 .release = kfd_topology_kobj_release,
239 .sysfs_ops = &sysprops_ops,
242 static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr,
246 struct kfd_iolink_properties *iolink;
248 /* Making sure that the buffer is an empty string */
251 iolink = container_of(attr, struct kfd_iolink_properties, attr);
252 sysfs_show_32bit_prop(buffer, "type", iolink->iolink_type);
253 sysfs_show_32bit_prop(buffer, "version_major", iolink->ver_maj);
254 sysfs_show_32bit_prop(buffer, "version_minor", iolink->ver_min);
255 sysfs_show_32bit_prop(buffer, "node_from", iolink->node_from);
256 sysfs_show_32bit_prop(buffer, "node_to", iolink->node_to);
257 sysfs_show_32bit_prop(buffer, "weight", iolink->weight);
258 sysfs_show_32bit_prop(buffer, "min_latency", iolink->min_latency);
259 sysfs_show_32bit_prop(buffer, "max_latency", iolink->max_latency);
260 sysfs_show_32bit_prop(buffer, "min_bandwidth", iolink->min_bandwidth);
261 sysfs_show_32bit_prop(buffer, "max_bandwidth", iolink->max_bandwidth);
262 sysfs_show_32bit_prop(buffer, "recommended_transfer_size",
263 iolink->rec_transfer_size);
264 ret = sysfs_show_32bit_prop(buffer, "flags", iolink->flags);
269 static const struct sysfs_ops iolink_ops = {
273 static struct kobj_type iolink_type = {
274 .release = kfd_topology_kobj_release,
275 .sysfs_ops = &iolink_ops,
278 static ssize_t mem_show(struct kobject *kobj, struct attribute *attr,
282 struct kfd_mem_properties *mem;
284 /* Making sure that the buffer is an empty string */
287 mem = container_of(attr, struct kfd_mem_properties, attr);
288 sysfs_show_32bit_prop(buffer, "heap_type", mem->heap_type);
289 sysfs_show_64bit_prop(buffer, "size_in_bytes", mem->size_in_bytes);
290 sysfs_show_32bit_prop(buffer, "flags", mem->flags);
291 sysfs_show_32bit_prop(buffer, "width", mem->width);
292 ret = sysfs_show_32bit_prop(buffer, "mem_clk_max", mem->mem_clk_max);
297 static const struct sysfs_ops mem_ops = {
301 static struct kobj_type mem_type = {
302 .release = kfd_topology_kobj_release,
303 .sysfs_ops = &mem_ops,
306 static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr,
311 struct kfd_cache_properties *cache;
313 /* Making sure that the buffer is an empty string */
316 cache = container_of(attr, struct kfd_cache_properties, attr);
317 sysfs_show_32bit_prop(buffer, "processor_id_low",
318 cache->processor_id_low);
319 sysfs_show_32bit_prop(buffer, "level", cache->cache_level);
320 sysfs_show_32bit_prop(buffer, "size", cache->cache_size);
321 sysfs_show_32bit_prop(buffer, "cache_line_size", cache->cacheline_size);
322 sysfs_show_32bit_prop(buffer, "cache_lines_per_tag",
323 cache->cachelines_per_tag);
324 sysfs_show_32bit_prop(buffer, "association", cache->cache_assoc);
325 sysfs_show_32bit_prop(buffer, "latency", cache->cache_latency);
326 sysfs_show_32bit_prop(buffer, "type", cache->cache_type);
327 snprintf(buffer, PAGE_SIZE, "%ssibling_map ", buffer);
328 for (i = 0; i < CRAT_SIBLINGMAP_SIZE; i++)
329 for (j = 0; j < sizeof(cache->sibling_map[0])*8; j++) {
331 if (cache->sibling_map[i] & (1 << j))
332 ret = snprintf(buffer, PAGE_SIZE,
333 "%s%d%s", buffer, 1, ",");
335 ret = snprintf(buffer, PAGE_SIZE,
336 "%s%d%s", buffer, 0, ",");
338 /* Replace the last "," with end of line */
339 *(buffer + strlen(buffer) - 1) = 0xA;
343 static const struct sysfs_ops cache_ops = {
344 .show = kfd_cache_show,
347 static struct kobj_type cache_type = {
348 .release = kfd_topology_kobj_release,
349 .sysfs_ops = &cache_ops,
352 /****** Sysfs of Performance Counters ******/
354 struct kfd_perf_attr {
355 struct kobj_attribute attr;
359 static ssize_t perf_show(struct kobject *kobj, struct kobj_attribute *attrs,
362 struct kfd_perf_attr *attr;
365 attr = container_of(attrs, struct kfd_perf_attr, attr);
366 if (!attr->data) /* invalid data for PMC */
369 return sysfs_show_32bit_val(buf, attr->data);
372 #define KFD_PERF_DESC(_name, _data) \
374 .attr = __ATTR(_name, 0444, perf_show, NULL), \
378 static struct kfd_perf_attr perf_attr_iommu[] = {
379 KFD_PERF_DESC(max_concurrent, 0),
380 KFD_PERF_DESC(num_counters, 0),
381 KFD_PERF_DESC(counter_ids, 0),
383 /****************************************/
385 static ssize_t node_show(struct kobject *kobj, struct attribute *attr,
388 struct kfd_topology_device *dev;
389 char public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE];
391 uint32_t log_max_watch_addr;
393 /* Making sure that the buffer is an empty string */
396 if (strcmp(attr->name, "gpu_id") == 0) {
397 dev = container_of(attr, struct kfd_topology_device,
399 return sysfs_show_32bit_val(buffer, dev->gpu_id);
402 if (strcmp(attr->name, "name") == 0) {
403 dev = container_of(attr, struct kfd_topology_device,
405 for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE; i++) {
407 (char)dev->node_props.marketing_name[i];
408 if (dev->node_props.marketing_name[i] == 0)
411 public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1] = 0x0;
412 return sysfs_show_str_val(buffer, public_name);
415 dev = container_of(attr, struct kfd_topology_device,
417 sysfs_show_32bit_prop(buffer, "cpu_cores_count",
418 dev->node_props.cpu_cores_count);
419 sysfs_show_32bit_prop(buffer, "simd_count",
420 dev->node_props.simd_count);
421 sysfs_show_32bit_prop(buffer, "mem_banks_count",
422 dev->node_props.mem_banks_count);
423 sysfs_show_32bit_prop(buffer, "caches_count",
424 dev->node_props.caches_count);
425 sysfs_show_32bit_prop(buffer, "io_links_count",
426 dev->node_props.io_links_count);
427 sysfs_show_32bit_prop(buffer, "cpu_core_id_base",
428 dev->node_props.cpu_core_id_base);
429 sysfs_show_32bit_prop(buffer, "simd_id_base",
430 dev->node_props.simd_id_base);
431 sysfs_show_32bit_prop(buffer, "max_waves_per_simd",
432 dev->node_props.max_waves_per_simd);
433 sysfs_show_32bit_prop(buffer, "lds_size_in_kb",
434 dev->node_props.lds_size_in_kb);
435 sysfs_show_32bit_prop(buffer, "gds_size_in_kb",
436 dev->node_props.gds_size_in_kb);
437 sysfs_show_32bit_prop(buffer, "wave_front_size",
438 dev->node_props.wave_front_size);
439 sysfs_show_32bit_prop(buffer, "array_count",
440 dev->node_props.array_count);
441 sysfs_show_32bit_prop(buffer, "simd_arrays_per_engine",
442 dev->node_props.simd_arrays_per_engine);
443 sysfs_show_32bit_prop(buffer, "cu_per_simd_array",
444 dev->node_props.cu_per_simd_array);
445 sysfs_show_32bit_prop(buffer, "simd_per_cu",
446 dev->node_props.simd_per_cu);
447 sysfs_show_32bit_prop(buffer, "max_slots_scratch_cu",
448 dev->node_props.max_slots_scratch_cu);
449 sysfs_show_32bit_prop(buffer, "vendor_id",
450 dev->node_props.vendor_id);
451 sysfs_show_32bit_prop(buffer, "device_id",
452 dev->node_props.device_id);
453 sysfs_show_32bit_prop(buffer, "location_id",
454 dev->node_props.location_id);
455 sysfs_show_32bit_prop(buffer, "drm_render_minor",
456 dev->node_props.drm_render_minor);
460 __ilog2_u32(dev->gpu->device_info->num_of_watch_points);
462 if (log_max_watch_addr) {
463 dev->node_props.capability |=
464 HSA_CAP_WATCH_POINTS_SUPPORTED;
466 dev->node_props.capability |=
467 ((log_max_watch_addr <<
468 HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) &
469 HSA_CAP_WATCH_POINTS_TOTALBITS_MASK);
472 if (dev->gpu->device_info->asic_family == CHIP_TONGA)
473 dev->node_props.capability |=
474 HSA_CAP_AQL_QUEUE_DOUBLE_MAP;
476 sysfs_show_32bit_prop(buffer, "max_engine_clk_fcompute",
477 dev->node_props.max_engine_clk_fcompute);
479 sysfs_show_64bit_prop(buffer, "local_mem_size",
480 (unsigned long long int) 0);
482 sysfs_show_32bit_prop(buffer, "fw_version",
483 dev->gpu->kfd2kgd->get_fw_version(
486 sysfs_show_32bit_prop(buffer, "capability",
487 dev->node_props.capability);
490 return sysfs_show_32bit_prop(buffer, "max_engine_clk_ccompute",
491 cpufreq_quick_get_max(0)/1000);
494 static const struct sysfs_ops node_ops = {
498 static struct kobj_type node_type = {
499 .release = kfd_topology_kobj_release,
500 .sysfs_ops = &node_ops,
503 static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr)
505 sysfs_remove_file(kobj, attr);
510 static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev)
512 struct kfd_iolink_properties *iolink;
513 struct kfd_cache_properties *cache;
514 struct kfd_mem_properties *mem;
515 struct kfd_perf_properties *perf;
517 if (dev->kobj_iolink) {
518 list_for_each_entry(iolink, &dev->io_link_props, list)
520 kfd_remove_sysfs_file(iolink->kobj,
524 kobject_del(dev->kobj_iolink);
525 kobject_put(dev->kobj_iolink);
526 dev->kobj_iolink = NULL;
529 if (dev->kobj_cache) {
530 list_for_each_entry(cache, &dev->cache_props, list)
532 kfd_remove_sysfs_file(cache->kobj,
536 kobject_del(dev->kobj_cache);
537 kobject_put(dev->kobj_cache);
538 dev->kobj_cache = NULL;
542 list_for_each_entry(mem, &dev->mem_props, list)
544 kfd_remove_sysfs_file(mem->kobj, &mem->attr);
547 kobject_del(dev->kobj_mem);
548 kobject_put(dev->kobj_mem);
549 dev->kobj_mem = NULL;
552 if (dev->kobj_perf) {
553 list_for_each_entry(perf, &dev->perf_props, list) {
554 kfree(perf->attr_group);
555 perf->attr_group = NULL;
557 kobject_del(dev->kobj_perf);
558 kobject_put(dev->kobj_perf);
559 dev->kobj_perf = NULL;
562 if (dev->kobj_node) {
563 sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid);
564 sysfs_remove_file(dev->kobj_node, &dev->attr_name);
565 sysfs_remove_file(dev->kobj_node, &dev->attr_props);
566 kobject_del(dev->kobj_node);
567 kobject_put(dev->kobj_node);
568 dev->kobj_node = NULL;
572 static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev,
575 struct kfd_iolink_properties *iolink;
576 struct kfd_cache_properties *cache;
577 struct kfd_mem_properties *mem;
578 struct kfd_perf_properties *perf;
580 uint32_t i, num_attrs;
581 struct attribute **attrs;
583 if (WARN_ON(dev->kobj_node))
587 * Creating the sysfs folders
589 dev->kobj_node = kfd_alloc_struct(dev->kobj_node);
593 ret = kobject_init_and_add(dev->kobj_node, &node_type,
594 sys_props.kobj_nodes, "%d", id);
596 kobject_put(dev->kobj_node);
600 dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node);
604 dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node);
605 if (!dev->kobj_cache)
608 dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node);
609 if (!dev->kobj_iolink)
612 dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node);
617 * Creating sysfs files for node properties
619 dev->attr_gpuid.name = "gpu_id";
620 dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE;
621 sysfs_attr_init(&dev->attr_gpuid);
622 dev->attr_name.name = "name";
623 dev->attr_name.mode = KFD_SYSFS_FILE_MODE;
624 sysfs_attr_init(&dev->attr_name);
625 dev->attr_props.name = "properties";
626 dev->attr_props.mode = KFD_SYSFS_FILE_MODE;
627 sysfs_attr_init(&dev->attr_props);
628 ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid);
631 ret = sysfs_create_file(dev->kobj_node, &dev->attr_name);
634 ret = sysfs_create_file(dev->kobj_node, &dev->attr_props);
639 list_for_each_entry(mem, &dev->mem_props, list) {
640 mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
643 ret = kobject_init_and_add(mem->kobj, &mem_type,
644 dev->kobj_mem, "%d", i);
646 kobject_put(mem->kobj);
650 mem->attr.name = "properties";
651 mem->attr.mode = KFD_SYSFS_FILE_MODE;
652 sysfs_attr_init(&mem->attr);
653 ret = sysfs_create_file(mem->kobj, &mem->attr);
660 list_for_each_entry(cache, &dev->cache_props, list) {
661 cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
664 ret = kobject_init_and_add(cache->kobj, &cache_type,
665 dev->kobj_cache, "%d", i);
667 kobject_put(cache->kobj);
671 cache->attr.name = "properties";
672 cache->attr.mode = KFD_SYSFS_FILE_MODE;
673 sysfs_attr_init(&cache->attr);
674 ret = sysfs_create_file(cache->kobj, &cache->attr);
681 list_for_each_entry(iolink, &dev->io_link_props, list) {
682 iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
685 ret = kobject_init_and_add(iolink->kobj, &iolink_type,
686 dev->kobj_iolink, "%d", i);
688 kobject_put(iolink->kobj);
692 iolink->attr.name = "properties";
693 iolink->attr.mode = KFD_SYSFS_FILE_MODE;
694 sysfs_attr_init(&iolink->attr);
695 ret = sysfs_create_file(iolink->kobj, &iolink->attr);
701 /* All hardware blocks have the same number of attributes. */
702 num_attrs = ARRAY_SIZE(perf_attr_iommu);
703 list_for_each_entry(perf, &dev->perf_props, list) {
704 perf->attr_group = kzalloc(sizeof(struct kfd_perf_attr)
705 * num_attrs + sizeof(struct attribute_group),
707 if (!perf->attr_group)
710 attrs = (struct attribute **)(perf->attr_group + 1);
711 if (!strcmp(perf->block_name, "iommu")) {
712 /* Information of IOMMU's num_counters and counter_ids is shown
713 * under /sys/bus/event_source/devices/amd_iommu. We don't
716 perf_attr_iommu[0].data = perf->max_concurrent;
717 for (i = 0; i < num_attrs; i++)
718 attrs[i] = &perf_attr_iommu[i].attr.attr;
720 perf->attr_group->name = perf->block_name;
721 perf->attr_group->attrs = attrs;
722 ret = sysfs_create_group(dev->kobj_perf, perf->attr_group);
730 /* Called with write topology lock acquired */
731 static int kfd_build_sysfs_node_tree(void)
733 struct kfd_topology_device *dev;
737 list_for_each_entry(dev, &topology_device_list, list) {
738 ret = kfd_build_sysfs_node_entry(dev, i);
747 /* Called with write topology lock acquired */
748 static void kfd_remove_sysfs_node_tree(void)
750 struct kfd_topology_device *dev;
752 list_for_each_entry(dev, &topology_device_list, list)
753 kfd_remove_sysfs_node_entry(dev);
756 static int kfd_topology_update_sysfs(void)
760 pr_info("Creating topology SYSFS entries\n");
761 if (!sys_props.kobj_topology) {
762 sys_props.kobj_topology =
763 kfd_alloc_struct(sys_props.kobj_topology);
764 if (!sys_props.kobj_topology)
767 ret = kobject_init_and_add(sys_props.kobj_topology,
768 &sysprops_type, &kfd_device->kobj,
771 kobject_put(sys_props.kobj_topology);
775 sys_props.kobj_nodes = kobject_create_and_add("nodes",
776 sys_props.kobj_topology);
777 if (!sys_props.kobj_nodes)
780 sys_props.attr_genid.name = "generation_id";
781 sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE;
782 sysfs_attr_init(&sys_props.attr_genid);
783 ret = sysfs_create_file(sys_props.kobj_topology,
784 &sys_props.attr_genid);
788 sys_props.attr_props.name = "system_properties";
789 sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE;
790 sysfs_attr_init(&sys_props.attr_props);
791 ret = sysfs_create_file(sys_props.kobj_topology,
792 &sys_props.attr_props);
797 kfd_remove_sysfs_node_tree();
799 return kfd_build_sysfs_node_tree();
802 static void kfd_topology_release_sysfs(void)
804 kfd_remove_sysfs_node_tree();
805 if (sys_props.kobj_topology) {
806 sysfs_remove_file(sys_props.kobj_topology,
807 &sys_props.attr_genid);
808 sysfs_remove_file(sys_props.kobj_topology,
809 &sys_props.attr_props);
810 if (sys_props.kobj_nodes) {
811 kobject_del(sys_props.kobj_nodes);
812 kobject_put(sys_props.kobj_nodes);
813 sys_props.kobj_nodes = NULL;
815 kobject_del(sys_props.kobj_topology);
816 kobject_put(sys_props.kobj_topology);
817 sys_props.kobj_topology = NULL;
821 /* Called with write topology_lock acquired */
822 static void kfd_topology_update_device_list(struct list_head *temp_list,
823 struct list_head *master_list)
825 while (!list_empty(temp_list)) {
826 list_move_tail(temp_list->next, master_list);
827 sys_props.num_devices++;
831 static void kfd_debug_print_topology(void)
833 struct kfd_topology_device *dev;
835 down_read(&topology_lock);
837 dev = list_last_entry(&topology_device_list,
838 struct kfd_topology_device, list);
840 if (dev->node_props.cpu_cores_count &&
841 dev->node_props.simd_count) {
842 pr_info("Topology: Add APU node [0x%0x:0x%0x]\n",
843 dev->node_props.device_id,
844 dev->node_props.vendor_id);
845 } else if (dev->node_props.cpu_cores_count)
846 pr_info("Topology: Add CPU node\n");
847 else if (dev->node_props.simd_count)
848 pr_info("Topology: Add dGPU node [0x%0x:0x%0x]\n",
849 dev->node_props.device_id,
850 dev->node_props.vendor_id);
852 up_read(&topology_lock);
855 /* Helper function for intializing platform_xx members of
856 * kfd_system_properties. Uses OEM info from the last CPU/APU node.
858 static void kfd_update_system_properties(void)
860 struct kfd_topology_device *dev;
862 down_read(&topology_lock);
863 dev = list_last_entry(&topology_device_list,
864 struct kfd_topology_device, list);
866 sys_props.platform_id =
867 (*((uint64_t *)dev->oem_id)) & CRAT_OEMID_64BIT_MASK;
868 sys_props.platform_oem = *((uint64_t *)dev->oem_table_id);
869 sys_props.platform_rev = dev->oem_revision;
871 up_read(&topology_lock);
874 static void find_system_memory(const struct dmi_header *dm,
877 struct kfd_mem_properties *mem;
878 u16 mem_width, mem_clock;
879 struct kfd_topology_device *kdev =
880 (struct kfd_topology_device *)private;
881 const u8 *dmi_data = (const u8 *)(dm + 1);
883 if (dm->type == DMI_ENTRY_MEM_DEVICE && dm->length >= 0x15) {
884 mem_width = (u16)(*(const u16 *)(dmi_data + 0x6));
885 mem_clock = (u16)(*(const u16 *)(dmi_data + 0x11));
886 list_for_each_entry(mem, &kdev->mem_props, list) {
887 if (mem_width != 0xFFFF && mem_width != 0)
888 mem->width = mem_width;
890 mem->mem_clk_max = mem_clock;
896 * Performance counters information is not part of CRAT but we would like to
897 * put them in the sysfs under topology directory for Thunk to get the data.
898 * This function is called before updating the sysfs.
900 static int kfd_add_perf_to_topology(struct kfd_topology_device *kdev)
902 /* These are the only counters supported so far */
903 return kfd_iommu_add_perf_counters(kdev);
906 /* kfd_add_non_crat_information - Add information that is not currently
907 * defined in CRAT but is necessary for KFD topology
908 * @dev - topology device to which addition info is added
910 static void kfd_add_non_crat_information(struct kfd_topology_device *kdev)
912 /* Check if CPU only node. */
914 /* Add system memory information */
915 dmi_walk(find_system_memory, kdev);
917 /* TODO: For GPU node, rearrange code from kfd_topology_add_device */
920 /* kfd_is_acpi_crat_invalid - CRAT from ACPI is valid only for AMD APU devices.
921 * Ignore CRAT for all other devices. AMD APU is identified if both CPU
922 * and GPU cores are present.
923 * @device_list - topology device list created by parsing ACPI CRAT table.
924 * @return - TRUE if invalid, FALSE is valid.
926 static bool kfd_is_acpi_crat_invalid(struct list_head *device_list)
928 struct kfd_topology_device *dev;
930 list_for_each_entry(dev, device_list, list) {
931 if (dev->node_props.cpu_cores_count &&
932 dev->node_props.simd_count)
935 pr_info("Ignoring ACPI CRAT on non-APU system\n");
939 int kfd_topology_init(void)
941 void *crat_image = NULL;
942 size_t image_size = 0;
944 struct list_head temp_topology_device_list;
945 int cpu_only_node = 0;
946 struct kfd_topology_device *kdev;
947 int proximity_domain;
949 /* topology_device_list - Master list of all topology devices
950 * temp_topology_device_list - temporary list created while parsing CRAT
951 * or VCRAT. Once parsing is complete the contents of list is moved to
952 * topology_device_list
955 /* Initialize the head for the both the lists */
956 INIT_LIST_HEAD(&topology_device_list);
957 INIT_LIST_HEAD(&temp_topology_device_list);
958 init_rwsem(&topology_lock);
960 memset(&sys_props, 0, sizeof(sys_props));
962 /* Proximity domains in ACPI CRAT tables start counting at
963 * 0. The same should be true for virtual CRAT tables created
964 * at this stage. GPUs added later in kfd_topology_add_device
967 proximity_domain = 0;
970 * Get the CRAT image from the ACPI. If ACPI doesn't have one
971 * or if ACPI CRAT is invalid create a virtual CRAT.
972 * NOTE: The current implementation expects all AMD APUs to have
973 * CRAT. If no CRAT is available, it is assumed to be a CPU
975 ret = kfd_create_crat_image_acpi(&crat_image, &image_size);
977 ret = kfd_parse_crat_table(crat_image,
978 &temp_topology_device_list,
981 kfd_is_acpi_crat_invalid(&temp_topology_device_list)) {
982 kfd_release_topology_device_list(
983 &temp_topology_device_list);
984 kfd_destroy_crat_image(crat_image);
990 ret = kfd_create_crat_image_virtual(&crat_image, &image_size,
991 COMPUTE_UNIT_CPU, NULL,
995 pr_err("Error creating VCRAT table for CPU\n");
999 ret = kfd_parse_crat_table(crat_image,
1000 &temp_topology_device_list,
1003 pr_err("Error parsing VCRAT table for CPU\n");
1008 kdev = list_first_entry(&temp_topology_device_list,
1009 struct kfd_topology_device, list);
1010 kfd_add_perf_to_topology(kdev);
1012 down_write(&topology_lock);
1013 kfd_topology_update_device_list(&temp_topology_device_list,
1014 &topology_device_list);
1015 atomic_set(&topology_crat_proximity_domain, sys_props.num_devices-1);
1016 ret = kfd_topology_update_sysfs();
1017 up_write(&topology_lock);
1020 sys_props.generation_count++;
1021 kfd_update_system_properties();
1022 kfd_debug_print_topology();
1023 pr_info("Finished initializing topology\n");
1025 pr_err("Failed to update topology in sysfs ret=%d\n", ret);
1027 /* For nodes with GPU, this information gets added
1028 * when GPU is detected (kfd_topology_add_device).
1030 if (cpu_only_node) {
1031 /* Add additional information to CPU only node created above */
1032 down_write(&topology_lock);
1033 kdev = list_first_entry(&topology_device_list,
1034 struct kfd_topology_device, list);
1035 up_write(&topology_lock);
1036 kfd_add_non_crat_information(kdev);
1040 kfd_destroy_crat_image(crat_image);
1044 void kfd_topology_shutdown(void)
1046 down_write(&topology_lock);
1047 kfd_topology_release_sysfs();
1048 kfd_release_live_view();
1049 up_write(&topology_lock);
1052 static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu)
1056 uint64_t local_mem_size;
1058 struct kfd_local_mem_info local_mem_info;
1063 gpu->kfd2kgd->get_local_mem_info(gpu->kgd, &local_mem_info);
1065 local_mem_size = local_mem_info.local_mem_size_private +
1066 local_mem_info.local_mem_size_public;
1068 buf[0] = gpu->pdev->devfn;
1069 buf[1] = gpu->pdev->subsystem_vendor;
1070 buf[2] = gpu->pdev->subsystem_device;
1071 buf[3] = gpu->pdev->device;
1072 buf[4] = gpu->pdev->bus->number;
1073 buf[5] = lower_32_bits(local_mem_size);
1074 buf[6] = upper_32_bits(local_mem_size);
1076 for (i = 0, hashout = 0; i < 7; i++)
1077 hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH);
1081 /* kfd_assign_gpu - Attach @gpu to the correct kfd topology device. If
1082 * the GPU device is not already present in the topology device
1083 * list then return NULL. This means a new topology device has to
1084 * be created for this GPU.
1086 static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu)
1088 struct kfd_topology_device *dev;
1089 struct kfd_topology_device *out_dev = NULL;
1091 down_write(&topology_lock);
1092 list_for_each_entry(dev, &topology_device_list, list) {
1093 /* Discrete GPUs need their own topology device list
1094 * entries. Don't assign them to CPU/APU nodes.
1096 if (!gpu->device_info->needs_iommu_device &&
1097 dev->node_props.cpu_cores_count)
1100 if (!dev->gpu && (dev->node_props.simd_count > 0)) {
1106 up_write(&topology_lock);
1110 static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival)
1113 * TODO: Generate an event for thunk about the arrival/removal
1118 /* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info,
1119 * patch this after CRAT parsing.
1121 static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev)
1123 struct kfd_mem_properties *mem;
1124 struct kfd_local_mem_info local_mem_info;
1129 /* Currently, amdgpu driver (amdgpu_mc) deals only with GPUs with
1130 * single bank of VRAM local memory.
1131 * for dGPUs - VCRAT reports only one bank of Local Memory
1132 * for APUs - If CRAT from ACPI reports more than one bank, then
1133 * all the banks will report the same mem_clk_max information
1135 dev->gpu->kfd2kgd->get_local_mem_info(dev->gpu->kgd,
1138 list_for_each_entry(mem, &dev->mem_props, list)
1139 mem->mem_clk_max = local_mem_info.mem_clk_max;
1142 static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev)
1144 struct kfd_iolink_properties *link;
1146 if (!dev || !dev->gpu)
1149 /* GPU only creates direck links so apply flags setting to all */
1150 if (dev->gpu->device_info->asic_family == CHIP_HAWAII)
1151 list_for_each_entry(link, &dev->io_link_props, list)
1152 link->flags = CRAT_IOLINK_FLAGS_ENABLED |
1153 CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1154 CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1157 int kfd_topology_add_device(struct kfd_dev *gpu)
1160 struct kfd_topology_device *dev;
1161 struct kfd_cu_info cu_info;
1163 struct list_head temp_topology_device_list;
1164 void *crat_image = NULL;
1165 size_t image_size = 0;
1166 int proximity_domain;
1168 INIT_LIST_HEAD(&temp_topology_device_list);
1170 gpu_id = kfd_generate_gpu_id(gpu);
1172 pr_debug("Adding new GPU (ID: 0x%x) to topology\n", gpu_id);
1174 proximity_domain = atomic_inc_return(&topology_crat_proximity_domain);
1176 /* Check to see if this gpu device exists in the topology_device_list.
1177 * If so, assign the gpu to that device,
1178 * else create a Virtual CRAT for this gpu device and then parse that
1179 * CRAT to create a new topology device. Once created assign the gpu to
1180 * that topology device
1182 dev = kfd_assign_gpu(gpu);
1184 res = kfd_create_crat_image_virtual(&crat_image, &image_size,
1185 COMPUTE_UNIT_GPU, gpu,
1188 pr_err("Error creating VCRAT for GPU (ID: 0x%x)\n",
1192 res = kfd_parse_crat_table(crat_image,
1193 &temp_topology_device_list,
1196 pr_err("Error parsing VCRAT for GPU (ID: 0x%x)\n",
1201 down_write(&topology_lock);
1202 kfd_topology_update_device_list(&temp_topology_device_list,
1203 &topology_device_list);
1205 /* Update the SYSFS tree, since we added another topology
1208 res = kfd_topology_update_sysfs();
1209 up_write(&topology_lock);
1212 sys_props.generation_count++;
1214 pr_err("Failed to update GPU (ID: 0x%x) to sysfs topology. res=%d\n",
1216 dev = kfd_assign_gpu(gpu);
1217 if (WARN_ON(!dev)) {
1223 dev->gpu_id = gpu_id;
1226 /* TODO: Move the following lines to function
1227 * kfd_add_non_crat_information
1230 /* Fill-in additional information that is not available in CRAT but
1231 * needed for the topology
1234 dev->gpu->kfd2kgd->get_cu_info(dev->gpu->kgd, &cu_info);
1235 dev->node_props.simd_arrays_per_engine =
1236 cu_info.num_shader_arrays_per_engine;
1238 dev->node_props.vendor_id = gpu->pdev->vendor;
1239 dev->node_props.device_id = gpu->pdev->device;
1240 dev->node_props.location_id = PCI_DEVID(gpu->pdev->bus->number,
1242 dev->node_props.max_engine_clk_fcompute =
1243 dev->gpu->kfd2kgd->get_max_engine_clock_in_mhz(dev->gpu->kgd);
1244 dev->node_props.max_engine_clk_ccompute =
1245 cpufreq_quick_get_max(0) / 1000;
1246 dev->node_props.drm_render_minor =
1247 gpu->shared_resources.drm_render_minor;
1249 kfd_fill_mem_clk_max_info(dev);
1250 kfd_fill_iolink_non_crat_info(dev);
1252 switch (dev->gpu->device_info->asic_family) {
1256 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_PRE_1_0 <<
1257 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1258 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1262 case CHIP_POLARIS10:
1263 case CHIP_POLARIS11:
1264 pr_debug("Adding doorbell packet type capability\n");
1265 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 <<
1266 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1267 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1271 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 <<
1272 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1273 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1276 WARN(1, "Unexpected ASIC family %u",
1277 dev->gpu->device_info->asic_family);
1280 /* Fix errors in CZ CRAT.
1281 * simd_count: Carrizo CRAT reports wrong simd_count, probably
1282 * because it doesn't consider masked out CUs
1283 * max_waves_per_simd: Carrizo reports wrong max_waves_per_simd
1284 * capability flag: Carrizo CRAT doesn't report IOMMU flags
1286 if (dev->gpu->device_info->asic_family == CHIP_CARRIZO) {
1287 dev->node_props.simd_count =
1288 cu_info.simd_per_cu * cu_info.cu_active_number;
1289 dev->node_props.max_waves_per_simd = 10;
1290 dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
1293 kfd_debug_print_topology();
1296 kfd_notify_gpu_change(gpu_id, 1);
1298 kfd_destroy_crat_image(crat_image);
1302 int kfd_topology_remove_device(struct kfd_dev *gpu)
1304 struct kfd_topology_device *dev, *tmp;
1308 down_write(&topology_lock);
1310 list_for_each_entry_safe(dev, tmp, &topology_device_list, list)
1311 if (dev->gpu == gpu) {
1312 gpu_id = dev->gpu_id;
1313 kfd_remove_sysfs_node_entry(dev);
1314 kfd_release_topology_device(dev);
1315 sys_props.num_devices--;
1317 if (kfd_topology_update_sysfs() < 0)
1318 kfd_topology_release_sysfs();
1322 up_write(&topology_lock);
1325 kfd_notify_gpu_change(gpu_id, 0);
1330 /* kfd_topology_enum_kfd_devices - Enumerate through all devices in KFD
1331 * topology. If GPU device is found @idx, then valid kfd_dev pointer is
1332 * returned through @kdev
1333 * Return - 0: On success (@kdev will be NULL for non GPU nodes)
1334 * -1: If end of list
1336 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev)
1339 struct kfd_topology_device *top_dev;
1340 uint8_t device_idx = 0;
1343 down_read(&topology_lock);
1345 list_for_each_entry(top_dev, &topology_device_list, list) {
1346 if (device_idx == idx) {
1347 *kdev = top_dev->gpu;
1348 up_read(&topology_lock);
1355 up_read(&topology_lock);
1361 static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask)
1363 const struct cpuinfo_x86 *cpuinfo;
1364 int first_cpu_of_numa_node;
1366 if (!cpumask || cpumask == cpu_none_mask)
1368 first_cpu_of_numa_node = cpumask_first(cpumask);
1369 if (first_cpu_of_numa_node >= nr_cpu_ids)
1371 cpuinfo = &cpu_data(first_cpu_of_numa_node);
1373 return cpuinfo->apicid;
1376 /* kfd_numa_node_to_apic_id - Returns the APIC ID of the first logical processor
1377 * of the given NUMA node (numa_node_id)
1378 * Return -1 on failure
1380 int kfd_numa_node_to_apic_id(int numa_node_id)
1382 if (numa_node_id == -1) {
1383 pr_warn("Invalid NUMA Node. Use online CPU mask\n");
1384 return kfd_cpumask_to_apic_id(cpu_online_mask);
1386 return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id));
1389 #if defined(CONFIG_DEBUG_FS)
1391 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data)
1393 struct kfd_topology_device *dev;
1397 down_read(&topology_lock);
1399 list_for_each_entry(dev, &topology_device_list, list) {
1405 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1406 r = dqm_debugfs_hqds(m, dev->gpu->dqm);
1411 up_read(&topology_lock);
1416 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data)
1418 struct kfd_topology_device *dev;
1422 down_read(&topology_lock);
1424 list_for_each_entry(dev, &topology_device_list, list) {
1430 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1431 r = pm_debugfs_runlist(m, &dev->gpu->dqm->packets);
1436 up_read(&topology_lock);