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 #include "amdgpu_amdkfd.h"
41 #include "amdgpu_ras.h"
43 /* topology_device_list - Master list of all topology devices */
44 static struct list_head topology_device_list;
45 static struct kfd_system_properties sys_props;
47 static DECLARE_RWSEM(topology_lock);
48 static atomic_t topology_crat_proximity_domain;
50 struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
51 uint32_t proximity_domain)
53 struct kfd_topology_device *top_dev;
54 struct kfd_topology_device *device = NULL;
56 down_read(&topology_lock);
58 list_for_each_entry(top_dev, &topology_device_list, list)
59 if (top_dev->proximity_domain == proximity_domain) {
64 up_read(&topology_lock);
69 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id)
71 struct kfd_topology_device *top_dev = NULL;
72 struct kfd_topology_device *ret = NULL;
74 down_read(&topology_lock);
76 list_for_each_entry(top_dev, &topology_device_list, list)
77 if (top_dev->gpu_id == gpu_id) {
82 up_read(&topology_lock);
87 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id)
89 struct kfd_topology_device *top_dev;
91 top_dev = kfd_topology_device_by_id(gpu_id);
98 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev)
100 struct kfd_topology_device *top_dev;
101 struct kfd_dev *device = NULL;
103 down_read(&topology_lock);
105 list_for_each_entry(top_dev, &topology_device_list, list)
106 if (top_dev->gpu && top_dev->gpu->pdev == pdev) {
107 device = top_dev->gpu;
111 up_read(&topology_lock);
116 struct kfd_dev *kfd_device_by_kgd(const struct kgd_dev *kgd)
118 struct kfd_topology_device *top_dev;
119 struct kfd_dev *device = NULL;
121 down_read(&topology_lock);
123 list_for_each_entry(top_dev, &topology_device_list, list)
124 if (top_dev->gpu && top_dev->gpu->kgd == kgd) {
125 device = top_dev->gpu;
129 up_read(&topology_lock);
134 /* Called with write topology_lock acquired */
135 static void kfd_release_topology_device(struct kfd_topology_device *dev)
137 struct kfd_mem_properties *mem;
138 struct kfd_cache_properties *cache;
139 struct kfd_iolink_properties *iolink;
140 struct kfd_perf_properties *perf;
142 list_del(&dev->list);
144 while (dev->mem_props.next != &dev->mem_props) {
145 mem = container_of(dev->mem_props.next,
146 struct kfd_mem_properties, list);
147 list_del(&mem->list);
151 while (dev->cache_props.next != &dev->cache_props) {
152 cache = container_of(dev->cache_props.next,
153 struct kfd_cache_properties, list);
154 list_del(&cache->list);
158 while (dev->io_link_props.next != &dev->io_link_props) {
159 iolink = container_of(dev->io_link_props.next,
160 struct kfd_iolink_properties, list);
161 list_del(&iolink->list);
165 while (dev->perf_props.next != &dev->perf_props) {
166 perf = container_of(dev->perf_props.next,
167 struct kfd_perf_properties, list);
168 list_del(&perf->list);
175 void kfd_release_topology_device_list(struct list_head *device_list)
177 struct kfd_topology_device *dev;
179 while (!list_empty(device_list)) {
180 dev = list_first_entry(device_list,
181 struct kfd_topology_device, list);
182 kfd_release_topology_device(dev);
186 static void kfd_release_live_view(void)
188 kfd_release_topology_device_list(&topology_device_list);
189 memset(&sys_props, 0, sizeof(sys_props));
192 struct kfd_topology_device *kfd_create_topology_device(
193 struct list_head *device_list)
195 struct kfd_topology_device *dev;
197 dev = kfd_alloc_struct(dev);
199 pr_err("No memory to allocate a topology device");
203 INIT_LIST_HEAD(&dev->mem_props);
204 INIT_LIST_HEAD(&dev->cache_props);
205 INIT_LIST_HEAD(&dev->io_link_props);
206 INIT_LIST_HEAD(&dev->perf_props);
208 list_add_tail(&dev->list, device_list);
214 #define sysfs_show_gen_prop(buffer, offs, fmt, ...) \
215 (offs += snprintf(buffer+offs, PAGE_SIZE-offs, \
217 #define sysfs_show_32bit_prop(buffer, offs, name, value) \
218 sysfs_show_gen_prop(buffer, offs, "%s %u\n", name, value)
219 #define sysfs_show_64bit_prop(buffer, offs, name, value) \
220 sysfs_show_gen_prop(buffer, offs, "%s %llu\n", name, value)
221 #define sysfs_show_32bit_val(buffer, offs, value) \
222 sysfs_show_gen_prop(buffer, offs, "%u\n", value)
223 #define sysfs_show_str_val(buffer, offs, value) \
224 sysfs_show_gen_prop(buffer, offs, "%s\n", value)
226 static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr,
231 /* Making sure that the buffer is an empty string */
234 if (attr == &sys_props.attr_genid) {
235 sysfs_show_32bit_val(buffer, offs,
236 sys_props.generation_count);
237 } else if (attr == &sys_props.attr_props) {
238 sysfs_show_64bit_prop(buffer, offs, "platform_oem",
239 sys_props.platform_oem);
240 sysfs_show_64bit_prop(buffer, offs, "platform_id",
241 sys_props.platform_id);
242 sysfs_show_64bit_prop(buffer, offs, "platform_rev",
243 sys_props.platform_rev);
251 static void kfd_topology_kobj_release(struct kobject *kobj)
256 static const struct sysfs_ops sysprops_ops = {
257 .show = sysprops_show,
260 static struct kobj_type sysprops_type = {
261 .release = kfd_topology_kobj_release,
262 .sysfs_ops = &sysprops_ops,
265 static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr,
269 struct kfd_iolink_properties *iolink;
271 /* Making sure that the buffer is an empty string */
274 iolink = container_of(attr, struct kfd_iolink_properties, attr);
275 if (iolink->gpu && kfd_devcgroup_check_permission(iolink->gpu))
277 sysfs_show_32bit_prop(buffer, offs, "type", iolink->iolink_type);
278 sysfs_show_32bit_prop(buffer, offs, "version_major", iolink->ver_maj);
279 sysfs_show_32bit_prop(buffer, offs, "version_minor", iolink->ver_min);
280 sysfs_show_32bit_prop(buffer, offs, "node_from", iolink->node_from);
281 sysfs_show_32bit_prop(buffer, offs, "node_to", iolink->node_to);
282 sysfs_show_32bit_prop(buffer, offs, "weight", iolink->weight);
283 sysfs_show_32bit_prop(buffer, offs, "min_latency", iolink->min_latency);
284 sysfs_show_32bit_prop(buffer, offs, "max_latency", iolink->max_latency);
285 sysfs_show_32bit_prop(buffer, offs, "min_bandwidth",
286 iolink->min_bandwidth);
287 sysfs_show_32bit_prop(buffer, offs, "max_bandwidth",
288 iolink->max_bandwidth);
289 sysfs_show_32bit_prop(buffer, offs, "recommended_transfer_size",
290 iolink->rec_transfer_size);
291 sysfs_show_32bit_prop(buffer, offs, "flags", iolink->flags);
296 static const struct sysfs_ops iolink_ops = {
300 static struct kobj_type iolink_type = {
301 .release = kfd_topology_kobj_release,
302 .sysfs_ops = &iolink_ops,
305 static ssize_t mem_show(struct kobject *kobj, struct attribute *attr,
309 struct kfd_mem_properties *mem;
311 /* Making sure that the buffer is an empty string */
314 mem = container_of(attr, struct kfd_mem_properties, attr);
315 if (mem->gpu && kfd_devcgroup_check_permission(mem->gpu))
317 sysfs_show_32bit_prop(buffer, offs, "heap_type", mem->heap_type);
318 sysfs_show_64bit_prop(buffer, offs, "size_in_bytes",
320 sysfs_show_32bit_prop(buffer, offs, "flags", mem->flags);
321 sysfs_show_32bit_prop(buffer, offs, "width", mem->width);
322 sysfs_show_32bit_prop(buffer, offs, "mem_clk_max",
328 static const struct sysfs_ops mem_ops = {
332 static struct kobj_type mem_type = {
333 .release = kfd_topology_kobj_release,
334 .sysfs_ops = &mem_ops,
337 static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr,
342 struct kfd_cache_properties *cache;
344 /* Making sure that the buffer is an empty string */
347 cache = container_of(attr, struct kfd_cache_properties, attr);
348 if (cache->gpu && kfd_devcgroup_check_permission(cache->gpu))
350 sysfs_show_32bit_prop(buffer, offs, "processor_id_low",
351 cache->processor_id_low);
352 sysfs_show_32bit_prop(buffer, offs, "level", cache->cache_level);
353 sysfs_show_32bit_prop(buffer, offs, "size", cache->cache_size);
354 sysfs_show_32bit_prop(buffer, offs, "cache_line_size",
355 cache->cacheline_size);
356 sysfs_show_32bit_prop(buffer, offs, "cache_lines_per_tag",
357 cache->cachelines_per_tag);
358 sysfs_show_32bit_prop(buffer, offs, "association", cache->cache_assoc);
359 sysfs_show_32bit_prop(buffer, offs, "latency", cache->cache_latency);
360 sysfs_show_32bit_prop(buffer, offs, "type", cache->cache_type);
361 offs += snprintf(buffer+offs, PAGE_SIZE-offs, "sibling_map ");
362 for (i = 0; i < CRAT_SIBLINGMAP_SIZE; i++)
363 for (j = 0; j < sizeof(cache->sibling_map[0])*8; j++)
365 offs += snprintf(buffer+offs, PAGE_SIZE-offs, "%d,",
366 (cache->sibling_map[i] >> j) & 1);
368 /* Replace the last "," with end of line */
369 buffer[offs-1] = '\n';
373 static const struct sysfs_ops cache_ops = {
374 .show = kfd_cache_show,
377 static struct kobj_type cache_type = {
378 .release = kfd_topology_kobj_release,
379 .sysfs_ops = &cache_ops,
382 /****** Sysfs of Performance Counters ******/
384 struct kfd_perf_attr {
385 struct kobj_attribute attr;
389 static ssize_t perf_show(struct kobject *kobj, struct kobj_attribute *attrs,
393 struct kfd_perf_attr *attr;
396 attr = container_of(attrs, struct kfd_perf_attr, attr);
397 if (!attr->data) /* invalid data for PMC */
400 return sysfs_show_32bit_val(buf, offs, attr->data);
403 #define KFD_PERF_DESC(_name, _data) \
405 .attr = __ATTR(_name, 0444, perf_show, NULL), \
409 static struct kfd_perf_attr perf_attr_iommu[] = {
410 KFD_PERF_DESC(max_concurrent, 0),
411 KFD_PERF_DESC(num_counters, 0),
412 KFD_PERF_DESC(counter_ids, 0),
414 /****************************************/
416 static ssize_t node_show(struct kobject *kobj, struct attribute *attr,
420 struct kfd_topology_device *dev;
421 uint32_t log_max_watch_addr;
423 /* Making sure that the buffer is an empty string */
426 if (strcmp(attr->name, "gpu_id") == 0) {
427 dev = container_of(attr, struct kfd_topology_device,
429 if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu))
431 return sysfs_show_32bit_val(buffer, offs, dev->gpu_id);
434 if (strcmp(attr->name, "name") == 0) {
435 dev = container_of(attr, struct kfd_topology_device,
438 if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu))
440 return sysfs_show_str_val(buffer, offs, dev->node_props.name);
443 dev = container_of(attr, struct kfd_topology_device,
445 if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu))
447 sysfs_show_32bit_prop(buffer, offs, "cpu_cores_count",
448 dev->node_props.cpu_cores_count);
449 sysfs_show_32bit_prop(buffer, offs, "simd_count",
450 dev->gpu ? dev->node_props.simd_count : 0);
451 sysfs_show_32bit_prop(buffer, offs, "mem_banks_count",
452 dev->node_props.mem_banks_count);
453 sysfs_show_32bit_prop(buffer, offs, "caches_count",
454 dev->node_props.caches_count);
455 sysfs_show_32bit_prop(buffer, offs, "io_links_count",
456 dev->node_props.io_links_count);
457 sysfs_show_32bit_prop(buffer, offs, "cpu_core_id_base",
458 dev->node_props.cpu_core_id_base);
459 sysfs_show_32bit_prop(buffer, offs, "simd_id_base",
460 dev->node_props.simd_id_base);
461 sysfs_show_32bit_prop(buffer, offs, "max_waves_per_simd",
462 dev->node_props.max_waves_per_simd);
463 sysfs_show_32bit_prop(buffer, offs, "lds_size_in_kb",
464 dev->node_props.lds_size_in_kb);
465 sysfs_show_32bit_prop(buffer, offs, "gds_size_in_kb",
466 dev->node_props.gds_size_in_kb);
467 sysfs_show_32bit_prop(buffer, offs, "num_gws",
468 dev->node_props.num_gws);
469 sysfs_show_32bit_prop(buffer, offs, "wave_front_size",
470 dev->node_props.wave_front_size);
471 sysfs_show_32bit_prop(buffer, offs, "array_count",
472 dev->node_props.array_count);
473 sysfs_show_32bit_prop(buffer, offs, "simd_arrays_per_engine",
474 dev->node_props.simd_arrays_per_engine);
475 sysfs_show_32bit_prop(buffer, offs, "cu_per_simd_array",
476 dev->node_props.cu_per_simd_array);
477 sysfs_show_32bit_prop(buffer, offs, "simd_per_cu",
478 dev->node_props.simd_per_cu);
479 sysfs_show_32bit_prop(buffer, offs, "max_slots_scratch_cu",
480 dev->node_props.max_slots_scratch_cu);
481 sysfs_show_32bit_prop(buffer, offs, "gfx_target_version",
482 dev->node_props.gfx_target_version);
483 sysfs_show_32bit_prop(buffer, offs, "vendor_id",
484 dev->node_props.vendor_id);
485 sysfs_show_32bit_prop(buffer, offs, "device_id",
486 dev->node_props.device_id);
487 sysfs_show_32bit_prop(buffer, offs, "location_id",
488 dev->node_props.location_id);
489 sysfs_show_32bit_prop(buffer, offs, "domain",
490 dev->node_props.domain);
491 sysfs_show_32bit_prop(buffer, offs, "drm_render_minor",
492 dev->node_props.drm_render_minor);
493 sysfs_show_64bit_prop(buffer, offs, "hive_id",
494 dev->node_props.hive_id);
495 sysfs_show_32bit_prop(buffer, offs, "num_sdma_engines",
496 dev->node_props.num_sdma_engines);
497 sysfs_show_32bit_prop(buffer, offs, "num_sdma_xgmi_engines",
498 dev->node_props.num_sdma_xgmi_engines);
499 sysfs_show_32bit_prop(buffer, offs, "num_sdma_queues_per_engine",
500 dev->node_props.num_sdma_queues_per_engine);
501 sysfs_show_32bit_prop(buffer, offs, "num_cp_queues",
502 dev->node_props.num_cp_queues);
506 __ilog2_u32(dev->gpu->device_info->num_of_watch_points);
508 if (log_max_watch_addr) {
509 dev->node_props.capability |=
510 HSA_CAP_WATCH_POINTS_SUPPORTED;
512 dev->node_props.capability |=
513 ((log_max_watch_addr <<
514 HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) &
515 HSA_CAP_WATCH_POINTS_TOTALBITS_MASK);
518 if (dev->gpu->device_info->asic_family == CHIP_TONGA)
519 dev->node_props.capability |=
520 HSA_CAP_AQL_QUEUE_DOUBLE_MAP;
522 sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_fcompute",
523 dev->node_props.max_engine_clk_fcompute);
525 sysfs_show_64bit_prop(buffer, offs, "local_mem_size", 0ULL);
527 sysfs_show_32bit_prop(buffer, offs, "fw_version",
528 dev->gpu->mec_fw_version);
529 sysfs_show_32bit_prop(buffer, offs, "capability",
530 dev->node_props.capability);
531 sysfs_show_32bit_prop(buffer, offs, "sdma_fw_version",
532 dev->gpu->sdma_fw_version);
533 sysfs_show_64bit_prop(buffer, offs, "unique_id",
534 amdgpu_amdkfd_get_unique_id(dev->gpu->kgd));
538 return sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_ccompute",
539 cpufreq_quick_get_max(0)/1000);
542 static const struct sysfs_ops node_ops = {
546 static struct kobj_type node_type = {
547 .release = kfd_topology_kobj_release,
548 .sysfs_ops = &node_ops,
551 static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr)
553 sysfs_remove_file(kobj, attr);
558 static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev)
560 struct kfd_iolink_properties *iolink;
561 struct kfd_cache_properties *cache;
562 struct kfd_mem_properties *mem;
563 struct kfd_perf_properties *perf;
565 if (dev->kobj_iolink) {
566 list_for_each_entry(iolink, &dev->io_link_props, list)
568 kfd_remove_sysfs_file(iolink->kobj,
572 kobject_del(dev->kobj_iolink);
573 kobject_put(dev->kobj_iolink);
574 dev->kobj_iolink = NULL;
577 if (dev->kobj_cache) {
578 list_for_each_entry(cache, &dev->cache_props, list)
580 kfd_remove_sysfs_file(cache->kobj,
584 kobject_del(dev->kobj_cache);
585 kobject_put(dev->kobj_cache);
586 dev->kobj_cache = NULL;
590 list_for_each_entry(mem, &dev->mem_props, list)
592 kfd_remove_sysfs_file(mem->kobj, &mem->attr);
595 kobject_del(dev->kobj_mem);
596 kobject_put(dev->kobj_mem);
597 dev->kobj_mem = NULL;
600 if (dev->kobj_perf) {
601 list_for_each_entry(perf, &dev->perf_props, list) {
602 kfree(perf->attr_group);
603 perf->attr_group = NULL;
605 kobject_del(dev->kobj_perf);
606 kobject_put(dev->kobj_perf);
607 dev->kobj_perf = NULL;
610 if (dev->kobj_node) {
611 sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid);
612 sysfs_remove_file(dev->kobj_node, &dev->attr_name);
613 sysfs_remove_file(dev->kobj_node, &dev->attr_props);
614 kobject_del(dev->kobj_node);
615 kobject_put(dev->kobj_node);
616 dev->kobj_node = NULL;
620 static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev,
623 struct kfd_iolink_properties *iolink;
624 struct kfd_cache_properties *cache;
625 struct kfd_mem_properties *mem;
626 struct kfd_perf_properties *perf;
628 uint32_t i, num_attrs;
629 struct attribute **attrs;
631 if (WARN_ON(dev->kobj_node))
635 * Creating the sysfs folders
637 dev->kobj_node = kfd_alloc_struct(dev->kobj_node);
641 ret = kobject_init_and_add(dev->kobj_node, &node_type,
642 sys_props.kobj_nodes, "%d", id);
644 kobject_put(dev->kobj_node);
648 dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node);
652 dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node);
653 if (!dev->kobj_cache)
656 dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node);
657 if (!dev->kobj_iolink)
660 dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node);
665 * Creating sysfs files for node properties
667 dev->attr_gpuid.name = "gpu_id";
668 dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE;
669 sysfs_attr_init(&dev->attr_gpuid);
670 dev->attr_name.name = "name";
671 dev->attr_name.mode = KFD_SYSFS_FILE_MODE;
672 sysfs_attr_init(&dev->attr_name);
673 dev->attr_props.name = "properties";
674 dev->attr_props.mode = KFD_SYSFS_FILE_MODE;
675 sysfs_attr_init(&dev->attr_props);
676 ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid);
679 ret = sysfs_create_file(dev->kobj_node, &dev->attr_name);
682 ret = sysfs_create_file(dev->kobj_node, &dev->attr_props);
687 list_for_each_entry(mem, &dev->mem_props, list) {
688 mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
691 ret = kobject_init_and_add(mem->kobj, &mem_type,
692 dev->kobj_mem, "%d", i);
694 kobject_put(mem->kobj);
698 mem->attr.name = "properties";
699 mem->attr.mode = KFD_SYSFS_FILE_MODE;
700 sysfs_attr_init(&mem->attr);
701 ret = sysfs_create_file(mem->kobj, &mem->attr);
708 list_for_each_entry(cache, &dev->cache_props, list) {
709 cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
712 ret = kobject_init_and_add(cache->kobj, &cache_type,
713 dev->kobj_cache, "%d", i);
715 kobject_put(cache->kobj);
719 cache->attr.name = "properties";
720 cache->attr.mode = KFD_SYSFS_FILE_MODE;
721 sysfs_attr_init(&cache->attr);
722 ret = sysfs_create_file(cache->kobj, &cache->attr);
729 list_for_each_entry(iolink, &dev->io_link_props, list) {
730 iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
733 ret = kobject_init_and_add(iolink->kobj, &iolink_type,
734 dev->kobj_iolink, "%d", i);
736 kobject_put(iolink->kobj);
740 iolink->attr.name = "properties";
741 iolink->attr.mode = KFD_SYSFS_FILE_MODE;
742 sysfs_attr_init(&iolink->attr);
743 ret = sysfs_create_file(iolink->kobj, &iolink->attr);
749 /* All hardware blocks have the same number of attributes. */
750 num_attrs = ARRAY_SIZE(perf_attr_iommu);
751 list_for_each_entry(perf, &dev->perf_props, list) {
752 perf->attr_group = kzalloc(sizeof(struct kfd_perf_attr)
753 * num_attrs + sizeof(struct attribute_group),
755 if (!perf->attr_group)
758 attrs = (struct attribute **)(perf->attr_group + 1);
759 if (!strcmp(perf->block_name, "iommu")) {
760 /* Information of IOMMU's num_counters and counter_ids is shown
761 * under /sys/bus/event_source/devices/amd_iommu. We don't
764 perf_attr_iommu[0].data = perf->max_concurrent;
765 for (i = 0; i < num_attrs; i++)
766 attrs[i] = &perf_attr_iommu[i].attr.attr;
768 perf->attr_group->name = perf->block_name;
769 perf->attr_group->attrs = attrs;
770 ret = sysfs_create_group(dev->kobj_perf, perf->attr_group);
778 /* Called with write topology lock acquired */
779 static int kfd_build_sysfs_node_tree(void)
781 struct kfd_topology_device *dev;
785 list_for_each_entry(dev, &topology_device_list, list) {
786 ret = kfd_build_sysfs_node_entry(dev, i);
795 /* Called with write topology lock acquired */
796 static void kfd_remove_sysfs_node_tree(void)
798 struct kfd_topology_device *dev;
800 list_for_each_entry(dev, &topology_device_list, list)
801 kfd_remove_sysfs_node_entry(dev);
804 static int kfd_topology_update_sysfs(void)
808 if (!sys_props.kobj_topology) {
809 sys_props.kobj_topology =
810 kfd_alloc_struct(sys_props.kobj_topology);
811 if (!sys_props.kobj_topology)
814 ret = kobject_init_and_add(sys_props.kobj_topology,
815 &sysprops_type, &kfd_device->kobj,
818 kobject_put(sys_props.kobj_topology);
822 sys_props.kobj_nodes = kobject_create_and_add("nodes",
823 sys_props.kobj_topology);
824 if (!sys_props.kobj_nodes)
827 sys_props.attr_genid.name = "generation_id";
828 sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE;
829 sysfs_attr_init(&sys_props.attr_genid);
830 ret = sysfs_create_file(sys_props.kobj_topology,
831 &sys_props.attr_genid);
835 sys_props.attr_props.name = "system_properties";
836 sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE;
837 sysfs_attr_init(&sys_props.attr_props);
838 ret = sysfs_create_file(sys_props.kobj_topology,
839 &sys_props.attr_props);
844 kfd_remove_sysfs_node_tree();
846 return kfd_build_sysfs_node_tree();
849 static void kfd_topology_release_sysfs(void)
851 kfd_remove_sysfs_node_tree();
852 if (sys_props.kobj_topology) {
853 sysfs_remove_file(sys_props.kobj_topology,
854 &sys_props.attr_genid);
855 sysfs_remove_file(sys_props.kobj_topology,
856 &sys_props.attr_props);
857 if (sys_props.kobj_nodes) {
858 kobject_del(sys_props.kobj_nodes);
859 kobject_put(sys_props.kobj_nodes);
860 sys_props.kobj_nodes = NULL;
862 kobject_del(sys_props.kobj_topology);
863 kobject_put(sys_props.kobj_topology);
864 sys_props.kobj_topology = NULL;
868 /* Called with write topology_lock acquired */
869 static void kfd_topology_update_device_list(struct list_head *temp_list,
870 struct list_head *master_list)
872 while (!list_empty(temp_list)) {
873 list_move_tail(temp_list->next, master_list);
874 sys_props.num_devices++;
878 static void kfd_debug_print_topology(void)
880 struct kfd_topology_device *dev;
882 down_read(&topology_lock);
884 dev = list_last_entry(&topology_device_list,
885 struct kfd_topology_device, list);
887 if (dev->node_props.cpu_cores_count &&
888 dev->node_props.simd_count) {
889 pr_info("Topology: Add APU node [0x%0x:0x%0x]\n",
890 dev->node_props.device_id,
891 dev->node_props.vendor_id);
892 } else if (dev->node_props.cpu_cores_count)
893 pr_info("Topology: Add CPU node\n");
894 else if (dev->node_props.simd_count)
895 pr_info("Topology: Add dGPU node [0x%0x:0x%0x]\n",
896 dev->node_props.device_id,
897 dev->node_props.vendor_id);
899 up_read(&topology_lock);
902 /* Helper function for intializing platform_xx members of
903 * kfd_system_properties. Uses OEM info from the last CPU/APU node.
905 static void kfd_update_system_properties(void)
907 struct kfd_topology_device *dev;
909 down_read(&topology_lock);
910 dev = list_last_entry(&topology_device_list,
911 struct kfd_topology_device, list);
913 sys_props.platform_id =
914 (*((uint64_t *)dev->oem_id)) & CRAT_OEMID_64BIT_MASK;
915 sys_props.platform_oem = *((uint64_t *)dev->oem_table_id);
916 sys_props.platform_rev = dev->oem_revision;
918 up_read(&topology_lock);
921 static void find_system_memory(const struct dmi_header *dm,
924 struct kfd_mem_properties *mem;
925 u16 mem_width, mem_clock;
926 struct kfd_topology_device *kdev =
927 (struct kfd_topology_device *)private;
928 const u8 *dmi_data = (const u8 *)(dm + 1);
930 if (dm->type == DMI_ENTRY_MEM_DEVICE && dm->length >= 0x15) {
931 mem_width = (u16)(*(const u16 *)(dmi_data + 0x6));
932 mem_clock = (u16)(*(const u16 *)(dmi_data + 0x11));
933 list_for_each_entry(mem, &kdev->mem_props, list) {
934 if (mem_width != 0xFFFF && mem_width != 0)
935 mem->width = mem_width;
937 mem->mem_clk_max = mem_clock;
943 * Performance counters information is not part of CRAT but we would like to
944 * put them in the sysfs under topology directory for Thunk to get the data.
945 * This function is called before updating the sysfs.
947 static int kfd_add_perf_to_topology(struct kfd_topology_device *kdev)
949 /* These are the only counters supported so far */
950 return kfd_iommu_add_perf_counters(kdev);
953 /* kfd_add_non_crat_information - Add information that is not currently
954 * defined in CRAT but is necessary for KFD topology
955 * @dev - topology device to which addition info is added
957 static void kfd_add_non_crat_information(struct kfd_topology_device *kdev)
959 /* Check if CPU only node. */
961 /* Add system memory information */
962 dmi_walk(find_system_memory, kdev);
964 /* TODO: For GPU node, rearrange code from kfd_topology_add_device */
967 /* kfd_is_acpi_crat_invalid - CRAT from ACPI is valid only for AMD APU devices.
968 * Ignore CRAT for all other devices. AMD APU is identified if both CPU
969 * and GPU cores are present.
970 * @device_list - topology device list created by parsing ACPI CRAT table.
971 * @return - TRUE if invalid, FALSE is valid.
973 static bool kfd_is_acpi_crat_invalid(struct list_head *device_list)
975 struct kfd_topology_device *dev;
977 list_for_each_entry(dev, device_list, list) {
978 if (dev->node_props.cpu_cores_count &&
979 dev->node_props.simd_count)
982 pr_info("Ignoring ACPI CRAT on non-APU system\n");
986 int kfd_topology_init(void)
988 void *crat_image = NULL;
989 size_t image_size = 0;
991 struct list_head temp_topology_device_list;
992 int cpu_only_node = 0;
993 struct kfd_topology_device *kdev;
994 int proximity_domain;
996 /* topology_device_list - Master list of all topology devices
997 * temp_topology_device_list - temporary list created while parsing CRAT
998 * or VCRAT. Once parsing is complete the contents of list is moved to
999 * topology_device_list
1002 /* Initialize the head for the both the lists */
1003 INIT_LIST_HEAD(&topology_device_list);
1004 INIT_LIST_HEAD(&temp_topology_device_list);
1005 init_rwsem(&topology_lock);
1007 memset(&sys_props, 0, sizeof(sys_props));
1009 /* Proximity domains in ACPI CRAT tables start counting at
1010 * 0. The same should be true for virtual CRAT tables created
1011 * at this stage. GPUs added later in kfd_topology_add_device
1014 proximity_domain = 0;
1017 * Get the CRAT image from the ACPI. If ACPI doesn't have one
1018 * or if ACPI CRAT is invalid create a virtual CRAT.
1019 * NOTE: The current implementation expects all AMD APUs to have
1020 * CRAT. If no CRAT is available, it is assumed to be a CPU
1022 ret = kfd_create_crat_image_acpi(&crat_image, &image_size);
1024 ret = kfd_parse_crat_table(crat_image,
1025 &temp_topology_device_list,
1028 kfd_is_acpi_crat_invalid(&temp_topology_device_list)) {
1029 kfd_release_topology_device_list(
1030 &temp_topology_device_list);
1031 kfd_destroy_crat_image(crat_image);
1037 ret = kfd_create_crat_image_virtual(&crat_image, &image_size,
1038 COMPUTE_UNIT_CPU, NULL,
1042 pr_err("Error creating VCRAT table for CPU\n");
1046 ret = kfd_parse_crat_table(crat_image,
1047 &temp_topology_device_list,
1050 pr_err("Error parsing VCRAT table for CPU\n");
1055 kdev = list_first_entry(&temp_topology_device_list,
1056 struct kfd_topology_device, list);
1057 kfd_add_perf_to_topology(kdev);
1059 down_write(&topology_lock);
1060 kfd_topology_update_device_list(&temp_topology_device_list,
1061 &topology_device_list);
1062 atomic_set(&topology_crat_proximity_domain, sys_props.num_devices-1);
1063 ret = kfd_topology_update_sysfs();
1064 up_write(&topology_lock);
1067 sys_props.generation_count++;
1068 kfd_update_system_properties();
1069 kfd_debug_print_topology();
1071 pr_err("Failed to update topology in sysfs ret=%d\n", ret);
1073 /* For nodes with GPU, this information gets added
1074 * when GPU is detected (kfd_topology_add_device).
1076 if (cpu_only_node) {
1077 /* Add additional information to CPU only node created above */
1078 down_write(&topology_lock);
1079 kdev = list_first_entry(&topology_device_list,
1080 struct kfd_topology_device, list);
1081 up_write(&topology_lock);
1082 kfd_add_non_crat_information(kdev);
1086 kfd_destroy_crat_image(crat_image);
1090 void kfd_topology_shutdown(void)
1092 down_write(&topology_lock);
1093 kfd_topology_release_sysfs();
1094 kfd_release_live_view();
1095 up_write(&topology_lock);
1098 static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu)
1102 uint64_t local_mem_size;
1104 struct kfd_local_mem_info local_mem_info;
1109 amdgpu_amdkfd_get_local_mem_info(gpu->kgd, &local_mem_info);
1111 local_mem_size = local_mem_info.local_mem_size_private +
1112 local_mem_info.local_mem_size_public;
1114 buf[0] = gpu->pdev->devfn;
1115 buf[1] = gpu->pdev->subsystem_vendor |
1116 (gpu->pdev->subsystem_device << 16);
1117 buf[2] = pci_domain_nr(gpu->pdev->bus);
1118 buf[3] = gpu->pdev->device;
1119 buf[4] = gpu->pdev->bus->number;
1120 buf[5] = lower_32_bits(local_mem_size);
1121 buf[6] = upper_32_bits(local_mem_size);
1123 for (i = 0, hashout = 0; i < 7; i++)
1124 hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH);
1128 /* kfd_assign_gpu - Attach @gpu to the correct kfd topology device. If
1129 * the GPU device is not already present in the topology device
1130 * list then return NULL. This means a new topology device has to
1131 * be created for this GPU.
1133 static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu)
1135 struct kfd_topology_device *dev;
1136 struct kfd_topology_device *out_dev = NULL;
1137 struct kfd_mem_properties *mem;
1138 struct kfd_cache_properties *cache;
1139 struct kfd_iolink_properties *iolink;
1141 down_write(&topology_lock);
1142 list_for_each_entry(dev, &topology_device_list, list) {
1143 /* Discrete GPUs need their own topology device list
1144 * entries. Don't assign them to CPU/APU nodes.
1146 if (!gpu->use_iommu_v2 &&
1147 dev->node_props.cpu_cores_count)
1150 if (!dev->gpu && (dev->node_props.simd_count > 0)) {
1154 list_for_each_entry(mem, &dev->mem_props, list)
1155 mem->gpu = dev->gpu;
1156 list_for_each_entry(cache, &dev->cache_props, list)
1157 cache->gpu = dev->gpu;
1158 list_for_each_entry(iolink, &dev->io_link_props, list)
1159 iolink->gpu = dev->gpu;
1163 up_write(&topology_lock);
1167 static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival)
1170 * TODO: Generate an event for thunk about the arrival/removal
1175 /* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info,
1176 * patch this after CRAT parsing.
1178 static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev)
1180 struct kfd_mem_properties *mem;
1181 struct kfd_local_mem_info local_mem_info;
1186 /* Currently, amdgpu driver (amdgpu_mc) deals only with GPUs with
1187 * single bank of VRAM local memory.
1188 * for dGPUs - VCRAT reports only one bank of Local Memory
1189 * for APUs - If CRAT from ACPI reports more than one bank, then
1190 * all the banks will report the same mem_clk_max information
1192 amdgpu_amdkfd_get_local_mem_info(dev->gpu->kgd, &local_mem_info);
1194 list_for_each_entry(mem, &dev->mem_props, list)
1195 mem->mem_clk_max = local_mem_info.mem_clk_max;
1198 static void kfd_set_iolink_no_atomics(struct kfd_topology_device *dev,
1199 struct kfd_topology_device *target_gpu_dev,
1200 struct kfd_iolink_properties *link)
1202 /* xgmi always supports atomics between links. */
1203 if (link->iolink_type == CRAT_IOLINK_TYPE_XGMI)
1206 /* check pcie support to set cpu(dev) flags for target_gpu_dev link. */
1207 if (target_gpu_dev) {
1210 pcie_capability_read_dword(target_gpu_dev->gpu->pdev,
1211 PCI_EXP_DEVCAP2, &cap);
1213 if (!(cap & (PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
1214 PCI_EXP_DEVCAP2_ATOMIC_COMP64)))
1215 link->flags |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1216 CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1217 /* set gpu (dev) flags. */
1219 if (!dev->gpu->pci_atomic_requested ||
1220 dev->gpu->device_info->asic_family ==
1222 link->flags |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1223 CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1227 static void kfd_set_iolink_non_coherent(struct kfd_topology_device *to_dev,
1228 struct kfd_iolink_properties *outbound_link,
1229 struct kfd_iolink_properties *inbound_link)
1231 /* CPU -> GPU with PCIe */
1233 inbound_link->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS)
1234 inbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT;
1237 /* GPU <-> GPU with PCIe and
1240 if (inbound_link->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS ||
1241 (inbound_link->iolink_type == CRAT_IOLINK_TYPE_XGMI &&
1242 to_dev->gpu->device_info->asic_family == CHIP_VEGA20)) {
1243 outbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT;
1244 inbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT;
1249 static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev)
1251 struct kfd_iolink_properties *link, *inbound_link;
1252 struct kfd_topology_device *peer_dev;
1254 if (!dev || !dev->gpu)
1257 /* GPU only creates direct links so apply flags setting to all */
1258 list_for_each_entry(link, &dev->io_link_props, list) {
1259 link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1260 kfd_set_iolink_no_atomics(dev, NULL, link);
1261 peer_dev = kfd_topology_device_by_proximity_domain(
1267 list_for_each_entry(inbound_link, &peer_dev->io_link_props,
1269 if (inbound_link->node_to != link->node_from)
1272 inbound_link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1273 kfd_set_iolink_no_atomics(peer_dev, dev, inbound_link);
1274 kfd_set_iolink_non_coherent(peer_dev, link, inbound_link);
1279 int kfd_topology_add_device(struct kfd_dev *gpu)
1282 struct kfd_topology_device *dev;
1283 struct kfd_cu_info cu_info;
1285 struct list_head temp_topology_device_list;
1286 void *crat_image = NULL;
1287 size_t image_size = 0;
1288 int proximity_domain;
1289 struct amdgpu_device *adev;
1291 INIT_LIST_HEAD(&temp_topology_device_list);
1293 gpu_id = kfd_generate_gpu_id(gpu);
1295 pr_debug("Adding new GPU (ID: 0x%x) to topology\n", gpu_id);
1297 proximity_domain = atomic_inc_return(&topology_crat_proximity_domain);
1299 /* Check to see if this gpu device exists in the topology_device_list.
1300 * If so, assign the gpu to that device,
1301 * else create a Virtual CRAT for this gpu device and then parse that
1302 * CRAT to create a new topology device. Once created assign the gpu to
1303 * that topology device
1305 dev = kfd_assign_gpu(gpu);
1307 res = kfd_create_crat_image_virtual(&crat_image, &image_size,
1308 COMPUTE_UNIT_GPU, gpu,
1311 pr_err("Error creating VCRAT for GPU (ID: 0x%x)\n",
1315 res = kfd_parse_crat_table(crat_image,
1316 &temp_topology_device_list,
1319 pr_err("Error parsing VCRAT for GPU (ID: 0x%x)\n",
1324 down_write(&topology_lock);
1325 kfd_topology_update_device_list(&temp_topology_device_list,
1326 &topology_device_list);
1328 /* Update the SYSFS tree, since we added another topology
1331 res = kfd_topology_update_sysfs();
1332 up_write(&topology_lock);
1335 sys_props.generation_count++;
1337 pr_err("Failed to update GPU (ID: 0x%x) to sysfs topology. res=%d\n",
1339 dev = kfd_assign_gpu(gpu);
1340 if (WARN_ON(!dev)) {
1346 dev->gpu_id = gpu_id;
1349 /* TODO: Move the following lines to function
1350 * kfd_add_non_crat_information
1353 /* Fill-in additional information that is not available in CRAT but
1354 * needed for the topology
1357 amdgpu_amdkfd_get_cu_info(dev->gpu->kgd, &cu_info);
1359 strncpy(dev->node_props.name, gpu->device_info->asic_name,
1360 KFD_TOPOLOGY_PUBLIC_NAME_SIZE);
1362 dev->node_props.simd_arrays_per_engine =
1363 cu_info.num_shader_arrays_per_engine;
1365 dev->node_props.gfx_target_version = gpu->device_info->gfx_target_version;
1366 dev->node_props.vendor_id = gpu->pdev->vendor;
1367 dev->node_props.device_id = gpu->pdev->device;
1368 dev->node_props.capability |=
1369 ((amdgpu_amdkfd_get_asic_rev_id(dev->gpu->kgd) <<
1370 HSA_CAP_ASIC_REVISION_SHIFT) &
1371 HSA_CAP_ASIC_REVISION_MASK);
1372 dev->node_props.location_id = pci_dev_id(gpu->pdev);
1373 dev->node_props.domain = pci_domain_nr(gpu->pdev->bus);
1374 dev->node_props.max_engine_clk_fcompute =
1375 amdgpu_amdkfd_get_max_engine_clock_in_mhz(dev->gpu->kgd);
1376 dev->node_props.max_engine_clk_ccompute =
1377 cpufreq_quick_get_max(0) / 1000;
1378 dev->node_props.drm_render_minor =
1379 gpu->shared_resources.drm_render_minor;
1381 dev->node_props.hive_id = gpu->hive_id;
1382 dev->node_props.num_sdma_engines = gpu->device_info->num_sdma_engines;
1383 dev->node_props.num_sdma_xgmi_engines =
1384 gpu->device_info->num_xgmi_sdma_engines;
1385 dev->node_props.num_sdma_queues_per_engine =
1386 gpu->device_info->num_sdma_queues_per_engine;
1387 dev->node_props.num_gws = (dev->gpu->gws &&
1388 dev->gpu->dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) ?
1389 amdgpu_amdkfd_get_num_gws(dev->gpu->kgd) : 0;
1390 dev->node_props.num_cp_queues = get_cp_queues_num(dev->gpu->dqm);
1392 kfd_fill_mem_clk_max_info(dev);
1393 kfd_fill_iolink_non_crat_info(dev);
1395 switch (dev->gpu->device_info->asic_family) {
1399 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_PRE_1_0 <<
1400 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1401 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1405 case CHIP_POLARIS10:
1406 case CHIP_POLARIS11:
1407 case CHIP_POLARIS12:
1409 pr_debug("Adding doorbell packet type capability\n");
1410 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 <<
1411 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1412 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1420 case CHIP_ALDEBARAN:
1424 case CHIP_SIENNA_CICHLID:
1425 case CHIP_NAVY_FLOUNDER:
1427 case CHIP_DIMGREY_CAVEFISH:
1428 case CHIP_BEIGE_GOBY:
1429 case CHIP_YELLOW_CARP:
1430 case CHIP_CYAN_SKILLFISH:
1431 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 <<
1432 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1433 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1436 WARN(1, "Unexpected ASIC family %u",
1437 dev->gpu->device_info->asic_family);
1441 * Overwrite ATS capability according to needs_iommu_device to fix
1442 * potential missing corresponding bit in CRAT of BIOS.
1444 if (dev->gpu->use_iommu_v2)
1445 dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
1447 dev->node_props.capability &= ~HSA_CAP_ATS_PRESENT;
1449 /* Fix errors in CZ CRAT.
1450 * simd_count: Carrizo CRAT reports wrong simd_count, probably
1451 * because it doesn't consider masked out CUs
1452 * max_waves_per_simd: Carrizo reports wrong max_waves_per_simd
1454 if (dev->gpu->device_info->asic_family == CHIP_CARRIZO) {
1455 dev->node_props.simd_count =
1456 cu_info.simd_per_cu * cu_info.cu_active_number;
1457 dev->node_props.max_waves_per_simd = 10;
1460 adev = (struct amdgpu_device *)(dev->gpu->kgd);
1461 /* kfd only concerns sram ecc on GFX and HBM ecc on UMC */
1462 dev->node_props.capability |=
1463 ((adev->ras_enabled & BIT(AMDGPU_RAS_BLOCK__GFX)) != 0) ?
1464 HSA_CAP_SRAM_EDCSUPPORTED : 0;
1465 dev->node_props.capability |= ((adev->ras_enabled & BIT(AMDGPU_RAS_BLOCK__UMC)) != 0) ?
1466 HSA_CAP_MEM_EDCSUPPORTED : 0;
1468 if (adev->asic_type != CHIP_VEGA10)
1469 dev->node_props.capability |= (adev->ras_enabled != 0) ?
1470 HSA_CAP_RASEVENTNOTIFY : 0;
1472 if (KFD_IS_SVM_API_SUPPORTED(adev->kfd.dev))
1473 dev->node_props.capability |= HSA_CAP_SVMAPI_SUPPORTED;
1475 kfd_debug_print_topology();
1478 kfd_notify_gpu_change(gpu_id, 1);
1480 kfd_destroy_crat_image(crat_image);
1484 int kfd_topology_remove_device(struct kfd_dev *gpu)
1486 struct kfd_topology_device *dev, *tmp;
1490 down_write(&topology_lock);
1492 list_for_each_entry_safe(dev, tmp, &topology_device_list, list)
1493 if (dev->gpu == gpu) {
1494 gpu_id = dev->gpu_id;
1495 kfd_remove_sysfs_node_entry(dev);
1496 kfd_release_topology_device(dev);
1497 sys_props.num_devices--;
1499 if (kfd_topology_update_sysfs() < 0)
1500 kfd_topology_release_sysfs();
1504 up_write(&topology_lock);
1507 kfd_notify_gpu_change(gpu_id, 0);
1512 /* kfd_topology_enum_kfd_devices - Enumerate through all devices in KFD
1513 * topology. If GPU device is found @idx, then valid kfd_dev pointer is
1514 * returned through @kdev
1515 * Return - 0: On success (@kdev will be NULL for non GPU nodes)
1516 * -1: If end of list
1518 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev)
1521 struct kfd_topology_device *top_dev;
1522 uint8_t device_idx = 0;
1525 down_read(&topology_lock);
1527 list_for_each_entry(top_dev, &topology_device_list, list) {
1528 if (device_idx == idx) {
1529 *kdev = top_dev->gpu;
1530 up_read(&topology_lock);
1537 up_read(&topology_lock);
1543 static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask)
1545 int first_cpu_of_numa_node;
1547 if (!cpumask || cpumask == cpu_none_mask)
1549 first_cpu_of_numa_node = cpumask_first(cpumask);
1550 if (first_cpu_of_numa_node >= nr_cpu_ids)
1552 #ifdef CONFIG_X86_64
1553 return cpu_data(first_cpu_of_numa_node).apicid;
1555 return first_cpu_of_numa_node;
1559 /* kfd_numa_node_to_apic_id - Returns the APIC ID of the first logical processor
1560 * of the given NUMA node (numa_node_id)
1561 * Return -1 on failure
1563 int kfd_numa_node_to_apic_id(int numa_node_id)
1565 if (numa_node_id == -1) {
1566 pr_warn("Invalid NUMA Node. Use online CPU mask\n");
1567 return kfd_cpumask_to_apic_id(cpu_online_mask);
1569 return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id));
1572 void kfd_double_confirm_iommu_support(struct kfd_dev *gpu)
1574 struct kfd_topology_device *dev;
1576 gpu->use_iommu_v2 = false;
1578 if (!gpu->device_info->needs_iommu_device)
1581 down_read(&topology_lock);
1583 /* Only use IOMMUv2 if there is an APU topology node with no GPU
1584 * assigned yet. This GPU will be assigned to it.
1586 list_for_each_entry(dev, &topology_device_list, list)
1587 if (dev->node_props.cpu_cores_count &&
1588 dev->node_props.simd_count &&
1590 gpu->use_iommu_v2 = true;
1592 up_read(&topology_lock);
1595 #if defined(CONFIG_DEBUG_FS)
1597 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data)
1599 struct kfd_topology_device *dev;
1603 down_read(&topology_lock);
1605 list_for_each_entry(dev, &topology_device_list, list) {
1611 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1612 r = dqm_debugfs_hqds(m, dev->gpu->dqm);
1617 up_read(&topology_lock);
1622 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data)
1624 struct kfd_topology_device *dev;
1628 down_read(&topology_lock);
1630 list_for_each_entry(dev, &topology_device_list, list) {
1636 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1637 r = pm_debugfs_runlist(m, &dev->gpu->dqm->packet_mgr);
1642 up_read(&topology_lock);