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"
40 #include "amdgpu_ras.h"
42 /* topology_device_list - Master list of all topology devices */
43 static struct list_head topology_device_list;
44 static struct kfd_system_properties sys_props;
46 static DECLARE_RWSEM(topology_lock);
47 static atomic_t topology_crat_proximity_domain;
49 struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
50 uint32_t proximity_domain)
52 struct kfd_topology_device *top_dev;
53 struct kfd_topology_device *device = NULL;
55 down_read(&topology_lock);
57 list_for_each_entry(top_dev, &topology_device_list, list)
58 if (top_dev->proximity_domain == proximity_domain) {
63 up_read(&topology_lock);
68 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id)
70 struct kfd_topology_device *top_dev = NULL;
71 struct kfd_topology_device *ret = NULL;
73 down_read(&topology_lock);
75 list_for_each_entry(top_dev, &topology_device_list, list)
76 if (top_dev->gpu_id == gpu_id) {
81 up_read(&topology_lock);
86 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id)
88 struct kfd_topology_device *top_dev;
90 top_dev = kfd_topology_device_by_id(gpu_id);
97 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev)
99 struct kfd_topology_device *top_dev;
100 struct kfd_dev *device = NULL;
102 down_read(&topology_lock);
104 list_for_each_entry(top_dev, &topology_device_list, list)
105 if (top_dev->gpu && top_dev->gpu->pdev == pdev) {
106 device = top_dev->gpu;
110 up_read(&topology_lock);
115 struct kfd_dev *kfd_device_by_kgd(const struct kgd_dev *kgd)
117 struct kfd_topology_device *top_dev;
118 struct kfd_dev *device = NULL;
120 down_read(&topology_lock);
122 list_for_each_entry(top_dev, &topology_device_list, list)
123 if (top_dev->gpu && top_dev->gpu->kgd == kgd) {
124 device = top_dev->gpu;
128 up_read(&topology_lock);
133 /* Called with write topology_lock acquired */
134 static void kfd_release_topology_device(struct kfd_topology_device *dev)
136 struct kfd_mem_properties *mem;
137 struct kfd_cache_properties *cache;
138 struct kfd_iolink_properties *iolink;
139 struct kfd_perf_properties *perf;
141 list_del(&dev->list);
143 while (dev->mem_props.next != &dev->mem_props) {
144 mem = container_of(dev->mem_props.next,
145 struct kfd_mem_properties, list);
146 list_del(&mem->list);
150 while (dev->cache_props.next != &dev->cache_props) {
151 cache = container_of(dev->cache_props.next,
152 struct kfd_cache_properties, list);
153 list_del(&cache->list);
157 while (dev->io_link_props.next != &dev->io_link_props) {
158 iolink = container_of(dev->io_link_props.next,
159 struct kfd_iolink_properties, list);
160 list_del(&iolink->list);
164 while (dev->perf_props.next != &dev->perf_props) {
165 perf = container_of(dev->perf_props.next,
166 struct kfd_perf_properties, list);
167 list_del(&perf->list);
174 void kfd_release_topology_device_list(struct list_head *device_list)
176 struct kfd_topology_device *dev;
178 while (!list_empty(device_list)) {
179 dev = list_first_entry(device_list,
180 struct kfd_topology_device, list);
181 kfd_release_topology_device(dev);
185 static void kfd_release_live_view(void)
187 kfd_release_topology_device_list(&topology_device_list);
188 memset(&sys_props, 0, sizeof(sys_props));
191 struct kfd_topology_device *kfd_create_topology_device(
192 struct list_head *device_list)
194 struct kfd_topology_device *dev;
196 dev = kfd_alloc_struct(dev);
198 pr_err("No memory to allocate a topology device");
202 INIT_LIST_HEAD(&dev->mem_props);
203 INIT_LIST_HEAD(&dev->cache_props);
204 INIT_LIST_HEAD(&dev->io_link_props);
205 INIT_LIST_HEAD(&dev->perf_props);
207 list_add_tail(&dev->list, device_list);
213 #define sysfs_show_gen_prop(buffer, fmt, ...) \
214 snprintf(buffer, PAGE_SIZE, "%s"fmt, buffer, __VA_ARGS__)
215 #define sysfs_show_32bit_prop(buffer, name, value) \
216 sysfs_show_gen_prop(buffer, "%s %u\n", name, value)
217 #define sysfs_show_64bit_prop(buffer, name, value) \
218 sysfs_show_gen_prop(buffer, "%s %llu\n", name, value)
219 #define sysfs_show_32bit_val(buffer, value) \
220 sysfs_show_gen_prop(buffer, "%u\n", value)
221 #define sysfs_show_str_val(buffer, value) \
222 sysfs_show_gen_prop(buffer, "%s\n", value)
224 static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr,
229 /* Making sure that the buffer is an empty string */
232 if (attr == &sys_props.attr_genid) {
233 ret = sysfs_show_32bit_val(buffer, sys_props.generation_count);
234 } else if (attr == &sys_props.attr_props) {
235 sysfs_show_64bit_prop(buffer, "platform_oem",
236 sys_props.platform_oem);
237 sysfs_show_64bit_prop(buffer, "platform_id",
238 sys_props.platform_id);
239 ret = sysfs_show_64bit_prop(buffer, "platform_rev",
240 sys_props.platform_rev);
248 static void kfd_topology_kobj_release(struct kobject *kobj)
253 static const struct sysfs_ops sysprops_ops = {
254 .show = sysprops_show,
257 static struct kobj_type sysprops_type = {
258 .release = kfd_topology_kobj_release,
259 .sysfs_ops = &sysprops_ops,
262 static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr,
266 struct kfd_iolink_properties *iolink;
268 /* Making sure that the buffer is an empty string */
271 iolink = container_of(attr, struct kfd_iolink_properties, attr);
272 sysfs_show_32bit_prop(buffer, "type", iolink->iolink_type);
273 sysfs_show_32bit_prop(buffer, "version_major", iolink->ver_maj);
274 sysfs_show_32bit_prop(buffer, "version_minor", iolink->ver_min);
275 sysfs_show_32bit_prop(buffer, "node_from", iolink->node_from);
276 sysfs_show_32bit_prop(buffer, "node_to", iolink->node_to);
277 sysfs_show_32bit_prop(buffer, "weight", iolink->weight);
278 sysfs_show_32bit_prop(buffer, "min_latency", iolink->min_latency);
279 sysfs_show_32bit_prop(buffer, "max_latency", iolink->max_latency);
280 sysfs_show_32bit_prop(buffer, "min_bandwidth", iolink->min_bandwidth);
281 sysfs_show_32bit_prop(buffer, "max_bandwidth", iolink->max_bandwidth);
282 sysfs_show_32bit_prop(buffer, "recommended_transfer_size",
283 iolink->rec_transfer_size);
284 ret = sysfs_show_32bit_prop(buffer, "flags", iolink->flags);
289 static const struct sysfs_ops iolink_ops = {
293 static struct kobj_type iolink_type = {
294 .release = kfd_topology_kobj_release,
295 .sysfs_ops = &iolink_ops,
298 static ssize_t mem_show(struct kobject *kobj, struct attribute *attr,
302 struct kfd_mem_properties *mem;
304 /* Making sure that the buffer is an empty string */
307 mem = container_of(attr, struct kfd_mem_properties, attr);
308 sysfs_show_32bit_prop(buffer, "heap_type", mem->heap_type);
309 sysfs_show_64bit_prop(buffer, "size_in_bytes", mem->size_in_bytes);
310 sysfs_show_32bit_prop(buffer, "flags", mem->flags);
311 sysfs_show_32bit_prop(buffer, "width", mem->width);
312 ret = sysfs_show_32bit_prop(buffer, "mem_clk_max", mem->mem_clk_max);
317 static const struct sysfs_ops mem_ops = {
321 static struct kobj_type mem_type = {
322 .release = kfd_topology_kobj_release,
323 .sysfs_ops = &mem_ops,
326 static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr,
331 struct kfd_cache_properties *cache;
333 /* Making sure that the buffer is an empty string */
336 cache = container_of(attr, struct kfd_cache_properties, attr);
337 sysfs_show_32bit_prop(buffer, "processor_id_low",
338 cache->processor_id_low);
339 sysfs_show_32bit_prop(buffer, "level", cache->cache_level);
340 sysfs_show_32bit_prop(buffer, "size", cache->cache_size);
341 sysfs_show_32bit_prop(buffer, "cache_line_size", cache->cacheline_size);
342 sysfs_show_32bit_prop(buffer, "cache_lines_per_tag",
343 cache->cachelines_per_tag);
344 sysfs_show_32bit_prop(buffer, "association", cache->cache_assoc);
345 sysfs_show_32bit_prop(buffer, "latency", cache->cache_latency);
346 sysfs_show_32bit_prop(buffer, "type", cache->cache_type);
347 snprintf(buffer, PAGE_SIZE, "%ssibling_map ", buffer);
348 for (i = 0; i < CRAT_SIBLINGMAP_SIZE; i++)
349 for (j = 0; j < sizeof(cache->sibling_map[0])*8; j++) {
351 if (cache->sibling_map[i] & (1 << j))
352 ret = snprintf(buffer, PAGE_SIZE,
353 "%s%d%s", buffer, 1, ",");
355 ret = snprintf(buffer, PAGE_SIZE,
356 "%s%d%s", buffer, 0, ",");
358 /* Replace the last "," with end of line */
359 *(buffer + strlen(buffer) - 1) = 0xA;
363 static const struct sysfs_ops cache_ops = {
364 .show = kfd_cache_show,
367 static struct kobj_type cache_type = {
368 .release = kfd_topology_kobj_release,
369 .sysfs_ops = &cache_ops,
372 /****** Sysfs of Performance Counters ******/
374 struct kfd_perf_attr {
375 struct kobj_attribute attr;
379 static ssize_t perf_show(struct kobject *kobj, struct kobj_attribute *attrs,
382 struct kfd_perf_attr *attr;
385 attr = container_of(attrs, struct kfd_perf_attr, attr);
386 if (!attr->data) /* invalid data for PMC */
389 return sysfs_show_32bit_val(buf, attr->data);
392 #define KFD_PERF_DESC(_name, _data) \
394 .attr = __ATTR(_name, 0444, perf_show, NULL), \
398 static struct kfd_perf_attr perf_attr_iommu[] = {
399 KFD_PERF_DESC(max_concurrent, 0),
400 KFD_PERF_DESC(num_counters, 0),
401 KFD_PERF_DESC(counter_ids, 0),
403 /****************************************/
405 static ssize_t node_show(struct kobject *kobj, struct attribute *attr,
408 struct kfd_topology_device *dev;
409 uint32_t log_max_watch_addr;
411 /* Making sure that the buffer is an empty string */
414 if (strcmp(attr->name, "gpu_id") == 0) {
415 dev = container_of(attr, struct kfd_topology_device,
417 return sysfs_show_32bit_val(buffer, dev->gpu_id);
420 if (strcmp(attr->name, "name") == 0) {
421 dev = container_of(attr, struct kfd_topology_device,
424 return sysfs_show_str_val(buffer, dev->node_props.name);
427 dev = container_of(attr, struct kfd_topology_device,
429 sysfs_show_32bit_prop(buffer, "cpu_cores_count",
430 dev->node_props.cpu_cores_count);
431 sysfs_show_32bit_prop(buffer, "simd_count",
432 dev->node_props.simd_count);
433 sysfs_show_32bit_prop(buffer, "mem_banks_count",
434 dev->node_props.mem_banks_count);
435 sysfs_show_32bit_prop(buffer, "caches_count",
436 dev->node_props.caches_count);
437 sysfs_show_32bit_prop(buffer, "io_links_count",
438 dev->node_props.io_links_count);
439 sysfs_show_32bit_prop(buffer, "cpu_core_id_base",
440 dev->node_props.cpu_core_id_base);
441 sysfs_show_32bit_prop(buffer, "simd_id_base",
442 dev->node_props.simd_id_base);
443 sysfs_show_32bit_prop(buffer, "max_waves_per_simd",
444 dev->node_props.max_waves_per_simd);
445 sysfs_show_32bit_prop(buffer, "lds_size_in_kb",
446 dev->node_props.lds_size_in_kb);
447 sysfs_show_32bit_prop(buffer, "gds_size_in_kb",
448 dev->node_props.gds_size_in_kb);
449 sysfs_show_32bit_prop(buffer, "num_gws",
450 dev->node_props.num_gws);
451 sysfs_show_32bit_prop(buffer, "wave_front_size",
452 dev->node_props.wave_front_size);
453 sysfs_show_32bit_prop(buffer, "array_count",
454 dev->node_props.array_count);
455 sysfs_show_32bit_prop(buffer, "simd_arrays_per_engine",
456 dev->node_props.simd_arrays_per_engine);
457 sysfs_show_32bit_prop(buffer, "cu_per_simd_array",
458 dev->node_props.cu_per_simd_array);
459 sysfs_show_32bit_prop(buffer, "simd_per_cu",
460 dev->node_props.simd_per_cu);
461 sysfs_show_32bit_prop(buffer, "max_slots_scratch_cu",
462 dev->node_props.max_slots_scratch_cu);
463 sysfs_show_32bit_prop(buffer, "vendor_id",
464 dev->node_props.vendor_id);
465 sysfs_show_32bit_prop(buffer, "device_id",
466 dev->node_props.device_id);
467 sysfs_show_32bit_prop(buffer, "location_id",
468 dev->node_props.location_id);
469 sysfs_show_32bit_prop(buffer, "drm_render_minor",
470 dev->node_props.drm_render_minor);
471 sysfs_show_64bit_prop(buffer, "hive_id",
472 dev->node_props.hive_id);
473 sysfs_show_32bit_prop(buffer, "num_sdma_engines",
474 dev->node_props.num_sdma_engines);
475 sysfs_show_32bit_prop(buffer, "num_sdma_xgmi_engines",
476 dev->node_props.num_sdma_xgmi_engines);
480 __ilog2_u32(dev->gpu->device_info->num_of_watch_points);
482 if (log_max_watch_addr) {
483 dev->node_props.capability |=
484 HSA_CAP_WATCH_POINTS_SUPPORTED;
486 dev->node_props.capability |=
487 ((log_max_watch_addr <<
488 HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) &
489 HSA_CAP_WATCH_POINTS_TOTALBITS_MASK);
492 if (dev->gpu->device_info->asic_family == CHIP_TONGA)
493 dev->node_props.capability |=
494 HSA_CAP_AQL_QUEUE_DOUBLE_MAP;
496 sysfs_show_32bit_prop(buffer, "max_engine_clk_fcompute",
497 dev->node_props.max_engine_clk_fcompute);
499 sysfs_show_64bit_prop(buffer, "local_mem_size",
500 (unsigned long long int) 0);
502 sysfs_show_32bit_prop(buffer, "fw_version",
503 dev->gpu->mec_fw_version);
504 sysfs_show_32bit_prop(buffer, "capability",
505 dev->node_props.capability);
506 sysfs_show_32bit_prop(buffer, "sdma_fw_version",
507 dev->gpu->sdma_fw_version);
510 return sysfs_show_32bit_prop(buffer, "max_engine_clk_ccompute",
511 cpufreq_quick_get_max(0)/1000);
514 static const struct sysfs_ops node_ops = {
518 static struct kobj_type node_type = {
519 .release = kfd_topology_kobj_release,
520 .sysfs_ops = &node_ops,
523 static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr)
525 sysfs_remove_file(kobj, attr);
530 static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev)
532 struct kfd_iolink_properties *iolink;
533 struct kfd_cache_properties *cache;
534 struct kfd_mem_properties *mem;
535 struct kfd_perf_properties *perf;
537 if (dev->kobj_iolink) {
538 list_for_each_entry(iolink, &dev->io_link_props, list)
540 kfd_remove_sysfs_file(iolink->kobj,
544 kobject_del(dev->kobj_iolink);
545 kobject_put(dev->kobj_iolink);
546 dev->kobj_iolink = NULL;
549 if (dev->kobj_cache) {
550 list_for_each_entry(cache, &dev->cache_props, list)
552 kfd_remove_sysfs_file(cache->kobj,
556 kobject_del(dev->kobj_cache);
557 kobject_put(dev->kobj_cache);
558 dev->kobj_cache = NULL;
562 list_for_each_entry(mem, &dev->mem_props, list)
564 kfd_remove_sysfs_file(mem->kobj, &mem->attr);
567 kobject_del(dev->kobj_mem);
568 kobject_put(dev->kobj_mem);
569 dev->kobj_mem = NULL;
572 if (dev->kobj_perf) {
573 list_for_each_entry(perf, &dev->perf_props, list) {
574 kfree(perf->attr_group);
575 perf->attr_group = NULL;
577 kobject_del(dev->kobj_perf);
578 kobject_put(dev->kobj_perf);
579 dev->kobj_perf = NULL;
582 if (dev->kobj_node) {
583 sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid);
584 sysfs_remove_file(dev->kobj_node, &dev->attr_name);
585 sysfs_remove_file(dev->kobj_node, &dev->attr_props);
586 kobject_del(dev->kobj_node);
587 kobject_put(dev->kobj_node);
588 dev->kobj_node = NULL;
592 static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev,
595 struct kfd_iolink_properties *iolink;
596 struct kfd_cache_properties *cache;
597 struct kfd_mem_properties *mem;
598 struct kfd_perf_properties *perf;
600 uint32_t i, num_attrs;
601 struct attribute **attrs;
603 if (WARN_ON(dev->kobj_node))
607 * Creating the sysfs folders
609 dev->kobj_node = kfd_alloc_struct(dev->kobj_node);
613 ret = kobject_init_and_add(dev->kobj_node, &node_type,
614 sys_props.kobj_nodes, "%d", id);
618 dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node);
622 dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node);
623 if (!dev->kobj_cache)
626 dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node);
627 if (!dev->kobj_iolink)
630 dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node);
635 * Creating sysfs files for node properties
637 dev->attr_gpuid.name = "gpu_id";
638 dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE;
639 sysfs_attr_init(&dev->attr_gpuid);
640 dev->attr_name.name = "name";
641 dev->attr_name.mode = KFD_SYSFS_FILE_MODE;
642 sysfs_attr_init(&dev->attr_name);
643 dev->attr_props.name = "properties";
644 dev->attr_props.mode = KFD_SYSFS_FILE_MODE;
645 sysfs_attr_init(&dev->attr_props);
646 ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid);
649 ret = sysfs_create_file(dev->kobj_node, &dev->attr_name);
652 ret = sysfs_create_file(dev->kobj_node, &dev->attr_props);
657 list_for_each_entry(mem, &dev->mem_props, list) {
658 mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
661 ret = kobject_init_and_add(mem->kobj, &mem_type,
662 dev->kobj_mem, "%d", i);
666 mem->attr.name = "properties";
667 mem->attr.mode = KFD_SYSFS_FILE_MODE;
668 sysfs_attr_init(&mem->attr);
669 ret = sysfs_create_file(mem->kobj, &mem->attr);
676 list_for_each_entry(cache, &dev->cache_props, list) {
677 cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
680 ret = kobject_init_and_add(cache->kobj, &cache_type,
681 dev->kobj_cache, "%d", i);
685 cache->attr.name = "properties";
686 cache->attr.mode = KFD_SYSFS_FILE_MODE;
687 sysfs_attr_init(&cache->attr);
688 ret = sysfs_create_file(cache->kobj, &cache->attr);
695 list_for_each_entry(iolink, &dev->io_link_props, list) {
696 iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
699 ret = kobject_init_and_add(iolink->kobj, &iolink_type,
700 dev->kobj_iolink, "%d", i);
704 iolink->attr.name = "properties";
705 iolink->attr.mode = KFD_SYSFS_FILE_MODE;
706 sysfs_attr_init(&iolink->attr);
707 ret = sysfs_create_file(iolink->kobj, &iolink->attr);
713 /* All hardware blocks have the same number of attributes. */
714 num_attrs = ARRAY_SIZE(perf_attr_iommu);
715 list_for_each_entry(perf, &dev->perf_props, list) {
716 perf->attr_group = kzalloc(sizeof(struct kfd_perf_attr)
717 * num_attrs + sizeof(struct attribute_group),
719 if (!perf->attr_group)
722 attrs = (struct attribute **)(perf->attr_group + 1);
723 if (!strcmp(perf->block_name, "iommu")) {
724 /* Information of IOMMU's num_counters and counter_ids is shown
725 * under /sys/bus/event_source/devices/amd_iommu. We don't
728 perf_attr_iommu[0].data = perf->max_concurrent;
729 for (i = 0; i < num_attrs; i++)
730 attrs[i] = &perf_attr_iommu[i].attr.attr;
732 perf->attr_group->name = perf->block_name;
733 perf->attr_group->attrs = attrs;
734 ret = sysfs_create_group(dev->kobj_perf, perf->attr_group);
742 /* Called with write topology lock acquired */
743 static int kfd_build_sysfs_node_tree(void)
745 struct kfd_topology_device *dev;
749 list_for_each_entry(dev, &topology_device_list, list) {
750 ret = kfd_build_sysfs_node_entry(dev, i);
759 /* Called with write topology lock acquired */
760 static void kfd_remove_sysfs_node_tree(void)
762 struct kfd_topology_device *dev;
764 list_for_each_entry(dev, &topology_device_list, list)
765 kfd_remove_sysfs_node_entry(dev);
768 static int kfd_topology_update_sysfs(void)
772 pr_info("Creating topology SYSFS entries\n");
773 if (!sys_props.kobj_topology) {
774 sys_props.kobj_topology =
775 kfd_alloc_struct(sys_props.kobj_topology);
776 if (!sys_props.kobj_topology)
779 ret = kobject_init_and_add(sys_props.kobj_topology,
780 &sysprops_type, &kfd_device->kobj,
785 sys_props.kobj_nodes = kobject_create_and_add("nodes",
786 sys_props.kobj_topology);
787 if (!sys_props.kobj_nodes)
790 sys_props.attr_genid.name = "generation_id";
791 sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE;
792 sysfs_attr_init(&sys_props.attr_genid);
793 ret = sysfs_create_file(sys_props.kobj_topology,
794 &sys_props.attr_genid);
798 sys_props.attr_props.name = "system_properties";
799 sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE;
800 sysfs_attr_init(&sys_props.attr_props);
801 ret = sysfs_create_file(sys_props.kobj_topology,
802 &sys_props.attr_props);
807 kfd_remove_sysfs_node_tree();
809 return kfd_build_sysfs_node_tree();
812 static void kfd_topology_release_sysfs(void)
814 kfd_remove_sysfs_node_tree();
815 if (sys_props.kobj_topology) {
816 sysfs_remove_file(sys_props.kobj_topology,
817 &sys_props.attr_genid);
818 sysfs_remove_file(sys_props.kobj_topology,
819 &sys_props.attr_props);
820 if (sys_props.kobj_nodes) {
821 kobject_del(sys_props.kobj_nodes);
822 kobject_put(sys_props.kobj_nodes);
823 sys_props.kobj_nodes = NULL;
825 kobject_del(sys_props.kobj_topology);
826 kobject_put(sys_props.kobj_topology);
827 sys_props.kobj_topology = NULL;
831 /* Called with write topology_lock acquired */
832 static void kfd_topology_update_device_list(struct list_head *temp_list,
833 struct list_head *master_list)
835 while (!list_empty(temp_list)) {
836 list_move_tail(temp_list->next, master_list);
837 sys_props.num_devices++;
841 static void kfd_debug_print_topology(void)
843 struct kfd_topology_device *dev;
845 down_read(&topology_lock);
847 dev = list_last_entry(&topology_device_list,
848 struct kfd_topology_device, list);
850 if (dev->node_props.cpu_cores_count &&
851 dev->node_props.simd_count) {
852 pr_info("Topology: Add APU node [0x%0x:0x%0x]\n",
853 dev->node_props.device_id,
854 dev->node_props.vendor_id);
855 } else if (dev->node_props.cpu_cores_count)
856 pr_info("Topology: Add CPU node\n");
857 else if (dev->node_props.simd_count)
858 pr_info("Topology: Add dGPU node [0x%0x:0x%0x]\n",
859 dev->node_props.device_id,
860 dev->node_props.vendor_id);
862 up_read(&topology_lock);
865 /* Helper function for intializing platform_xx members of
866 * kfd_system_properties. Uses OEM info from the last CPU/APU node.
868 static void kfd_update_system_properties(void)
870 struct kfd_topology_device *dev;
872 down_read(&topology_lock);
873 dev = list_last_entry(&topology_device_list,
874 struct kfd_topology_device, list);
876 sys_props.platform_id =
877 (*((uint64_t *)dev->oem_id)) & CRAT_OEMID_64BIT_MASK;
878 sys_props.platform_oem = *((uint64_t *)dev->oem_table_id);
879 sys_props.platform_rev = dev->oem_revision;
881 up_read(&topology_lock);
884 static void find_system_memory(const struct dmi_header *dm,
887 struct kfd_mem_properties *mem;
888 u16 mem_width, mem_clock;
889 struct kfd_topology_device *kdev =
890 (struct kfd_topology_device *)private;
891 const u8 *dmi_data = (const u8 *)(dm + 1);
893 if (dm->type == DMI_ENTRY_MEM_DEVICE && dm->length >= 0x15) {
894 mem_width = (u16)(*(const u16 *)(dmi_data + 0x6));
895 mem_clock = (u16)(*(const u16 *)(dmi_data + 0x11));
896 list_for_each_entry(mem, &kdev->mem_props, list) {
897 if (mem_width != 0xFFFF && mem_width != 0)
898 mem->width = mem_width;
900 mem->mem_clk_max = mem_clock;
906 * Performance counters information is not part of CRAT but we would like to
907 * put them in the sysfs under topology directory for Thunk to get the data.
908 * This function is called before updating the sysfs.
910 static int kfd_add_perf_to_topology(struct kfd_topology_device *kdev)
912 /* These are the only counters supported so far */
913 return kfd_iommu_add_perf_counters(kdev);
916 /* kfd_add_non_crat_information - Add information that is not currently
917 * defined in CRAT but is necessary for KFD topology
918 * @dev - topology device to which addition info is added
920 static void kfd_add_non_crat_information(struct kfd_topology_device *kdev)
922 /* Check if CPU only node. */
924 /* Add system memory information */
925 dmi_walk(find_system_memory, kdev);
927 /* TODO: For GPU node, rearrange code from kfd_topology_add_device */
930 /* kfd_is_acpi_crat_invalid - CRAT from ACPI is valid only for AMD APU devices.
931 * Ignore CRAT for all other devices. AMD APU is identified if both CPU
932 * and GPU cores are present.
933 * @device_list - topology device list created by parsing ACPI CRAT table.
934 * @return - TRUE if invalid, FALSE is valid.
936 static bool kfd_is_acpi_crat_invalid(struct list_head *device_list)
938 struct kfd_topology_device *dev;
940 list_for_each_entry(dev, device_list, list) {
941 if (dev->node_props.cpu_cores_count &&
942 dev->node_props.simd_count)
945 pr_info("Ignoring ACPI CRAT on non-APU system\n");
949 int kfd_topology_init(void)
951 void *crat_image = NULL;
952 size_t image_size = 0;
954 struct list_head temp_topology_device_list;
955 int cpu_only_node = 0;
956 struct kfd_topology_device *kdev;
957 int proximity_domain;
959 /* topology_device_list - Master list of all topology devices
960 * temp_topology_device_list - temporary list created while parsing CRAT
961 * or VCRAT. Once parsing is complete the contents of list is moved to
962 * topology_device_list
965 /* Initialize the head for the both the lists */
966 INIT_LIST_HEAD(&topology_device_list);
967 INIT_LIST_HEAD(&temp_topology_device_list);
968 init_rwsem(&topology_lock);
970 memset(&sys_props, 0, sizeof(sys_props));
972 /* Proximity domains in ACPI CRAT tables start counting at
973 * 0. The same should be true for virtual CRAT tables created
974 * at this stage. GPUs added later in kfd_topology_add_device
977 proximity_domain = 0;
980 * Get the CRAT image from the ACPI. If ACPI doesn't have one
981 * or if ACPI CRAT is invalid create a virtual CRAT.
982 * NOTE: The current implementation expects all AMD APUs to have
983 * CRAT. If no CRAT is available, it is assumed to be a CPU
985 ret = kfd_create_crat_image_acpi(&crat_image, &image_size);
987 ret = kfd_parse_crat_table(crat_image,
988 &temp_topology_device_list,
991 kfd_is_acpi_crat_invalid(&temp_topology_device_list)) {
992 kfd_release_topology_device_list(
993 &temp_topology_device_list);
994 kfd_destroy_crat_image(crat_image);
1000 ret = kfd_create_crat_image_virtual(&crat_image, &image_size,
1001 COMPUTE_UNIT_CPU, NULL,
1005 pr_err("Error creating VCRAT table for CPU\n");
1009 ret = kfd_parse_crat_table(crat_image,
1010 &temp_topology_device_list,
1013 pr_err("Error parsing VCRAT table for CPU\n");
1018 kdev = list_first_entry(&temp_topology_device_list,
1019 struct kfd_topology_device, list);
1020 kfd_add_perf_to_topology(kdev);
1022 down_write(&topology_lock);
1023 kfd_topology_update_device_list(&temp_topology_device_list,
1024 &topology_device_list);
1025 atomic_set(&topology_crat_proximity_domain, sys_props.num_devices-1);
1026 ret = kfd_topology_update_sysfs();
1027 up_write(&topology_lock);
1030 sys_props.generation_count++;
1031 kfd_update_system_properties();
1032 kfd_debug_print_topology();
1033 pr_info("Finished initializing topology\n");
1035 pr_err("Failed to update topology in sysfs ret=%d\n", ret);
1037 /* For nodes with GPU, this information gets added
1038 * when GPU is detected (kfd_topology_add_device).
1040 if (cpu_only_node) {
1041 /* Add additional information to CPU only node created above */
1042 down_write(&topology_lock);
1043 kdev = list_first_entry(&topology_device_list,
1044 struct kfd_topology_device, list);
1045 up_write(&topology_lock);
1046 kfd_add_non_crat_information(kdev);
1050 kfd_destroy_crat_image(crat_image);
1054 void kfd_topology_shutdown(void)
1056 down_write(&topology_lock);
1057 kfd_topology_release_sysfs();
1058 kfd_release_live_view();
1059 up_write(&topology_lock);
1062 static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu)
1066 uint64_t local_mem_size;
1068 struct kfd_local_mem_info local_mem_info;
1073 amdgpu_amdkfd_get_local_mem_info(gpu->kgd, &local_mem_info);
1075 local_mem_size = local_mem_info.local_mem_size_private +
1076 local_mem_info.local_mem_size_public;
1078 buf[0] = gpu->pdev->devfn;
1079 buf[1] = gpu->pdev->subsystem_vendor |
1080 (gpu->pdev->subsystem_device << 16);
1081 buf[2] = pci_domain_nr(gpu->pdev->bus);
1082 buf[3] = gpu->pdev->device;
1083 buf[4] = gpu->pdev->bus->number;
1084 buf[5] = lower_32_bits(local_mem_size);
1085 buf[6] = upper_32_bits(local_mem_size);
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;
1101 struct kfd_mem_properties *mem;
1102 struct kfd_cache_properties *cache;
1103 struct kfd_iolink_properties *iolink;
1105 down_write(&topology_lock);
1106 list_for_each_entry(dev, &topology_device_list, list) {
1107 /* Discrete GPUs need their own topology device list
1108 * entries. Don't assign them to CPU/APU nodes.
1110 if (!gpu->device_info->needs_iommu_device &&
1111 dev->node_props.cpu_cores_count)
1114 if (!dev->gpu && (dev->node_props.simd_count > 0)) {
1118 list_for_each_entry(mem, &dev->mem_props, list)
1119 mem->gpu = dev->gpu;
1120 list_for_each_entry(cache, &dev->cache_props, list)
1121 cache->gpu = dev->gpu;
1122 list_for_each_entry(iolink, &dev->io_link_props, list)
1123 iolink->gpu = dev->gpu;
1127 up_write(&topology_lock);
1131 static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival)
1134 * TODO: Generate an event for thunk about the arrival/removal
1139 /* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info,
1140 * patch this after CRAT parsing.
1142 static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev)
1144 struct kfd_mem_properties *mem;
1145 struct kfd_local_mem_info local_mem_info;
1150 /* Currently, amdgpu driver (amdgpu_mc) deals only with GPUs with
1151 * single bank of VRAM local memory.
1152 * for dGPUs - VCRAT reports only one bank of Local Memory
1153 * for APUs - If CRAT from ACPI reports more than one bank, then
1154 * all the banks will report the same mem_clk_max information
1156 amdgpu_amdkfd_get_local_mem_info(dev->gpu->kgd, &local_mem_info);
1158 list_for_each_entry(mem, &dev->mem_props, list)
1159 mem->mem_clk_max = local_mem_info.mem_clk_max;
1162 static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev)
1164 struct kfd_iolink_properties *link, *cpu_link;
1165 struct kfd_topology_device *cpu_dev;
1167 uint32_t cpu_flag = CRAT_IOLINK_FLAGS_ENABLED;
1168 uint32_t flag = CRAT_IOLINK_FLAGS_ENABLED;
1170 if (!dev || !dev->gpu)
1173 pcie_capability_read_dword(dev->gpu->pdev,
1174 PCI_EXP_DEVCAP2, &cap);
1176 if (!(cap & (PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
1177 PCI_EXP_DEVCAP2_ATOMIC_COMP64)))
1178 cpu_flag |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1179 CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1181 if (!dev->gpu->pci_atomic_requested ||
1182 dev->gpu->device_info->asic_family == CHIP_HAWAII)
1183 flag |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1184 CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1186 /* GPU only creates direct links so apply flags setting to all */
1187 list_for_each_entry(link, &dev->io_link_props, list) {
1189 cpu_dev = kfd_topology_device_by_proximity_domain(
1192 list_for_each_entry(cpu_link,
1193 &cpu_dev->io_link_props, list)
1194 if (cpu_link->node_to == link->node_from)
1195 cpu_link->flags = cpu_flag;
1200 int kfd_topology_add_device(struct kfd_dev *gpu)
1203 struct kfd_topology_device *dev;
1204 struct kfd_cu_info cu_info;
1206 struct list_head temp_topology_device_list;
1207 void *crat_image = NULL;
1208 size_t image_size = 0;
1209 int proximity_domain;
1210 struct amdgpu_ras *ctx;
1212 INIT_LIST_HEAD(&temp_topology_device_list);
1214 gpu_id = kfd_generate_gpu_id(gpu);
1216 pr_debug("Adding new GPU (ID: 0x%x) to topology\n", gpu_id);
1218 proximity_domain = atomic_inc_return(&topology_crat_proximity_domain);
1220 /* Check to see if this gpu device exists in the topology_device_list.
1221 * If so, assign the gpu to that device,
1222 * else create a Virtual CRAT for this gpu device and then parse that
1223 * CRAT to create a new topology device. Once created assign the gpu to
1224 * that topology device
1226 dev = kfd_assign_gpu(gpu);
1228 res = kfd_create_crat_image_virtual(&crat_image, &image_size,
1229 COMPUTE_UNIT_GPU, gpu,
1232 pr_err("Error creating VCRAT for GPU (ID: 0x%x)\n",
1236 res = kfd_parse_crat_table(crat_image,
1237 &temp_topology_device_list,
1240 pr_err("Error parsing VCRAT for GPU (ID: 0x%x)\n",
1245 down_write(&topology_lock);
1246 kfd_topology_update_device_list(&temp_topology_device_list,
1247 &topology_device_list);
1249 /* Update the SYSFS tree, since we added another topology
1252 res = kfd_topology_update_sysfs();
1253 up_write(&topology_lock);
1256 sys_props.generation_count++;
1258 pr_err("Failed to update GPU (ID: 0x%x) to sysfs topology. res=%d\n",
1260 dev = kfd_assign_gpu(gpu);
1261 if (WARN_ON(!dev)) {
1267 dev->gpu_id = gpu_id;
1270 /* TODO: Move the following lines to function
1271 * kfd_add_non_crat_information
1274 /* Fill-in additional information that is not available in CRAT but
1275 * needed for the topology
1278 amdgpu_amdkfd_get_cu_info(dev->gpu->kgd, &cu_info);
1280 strncpy(dev->node_props.name, gpu->device_info->asic_name,
1281 KFD_TOPOLOGY_PUBLIC_NAME_SIZE);
1283 dev->node_props.simd_arrays_per_engine =
1284 cu_info.num_shader_arrays_per_engine;
1286 dev->node_props.vendor_id = gpu->pdev->vendor;
1287 dev->node_props.device_id = gpu->pdev->device;
1288 dev->node_props.location_id = pci_dev_id(gpu->pdev);
1289 dev->node_props.max_engine_clk_fcompute =
1290 amdgpu_amdkfd_get_max_engine_clock_in_mhz(dev->gpu->kgd);
1291 dev->node_props.max_engine_clk_ccompute =
1292 cpufreq_quick_get_max(0) / 1000;
1293 dev->node_props.drm_render_minor =
1294 gpu->shared_resources.drm_render_minor;
1296 dev->node_props.hive_id = gpu->hive_id;
1297 dev->node_props.num_sdma_engines = gpu->device_info->num_sdma_engines;
1298 dev->node_props.num_sdma_xgmi_engines =
1299 gpu->device_info->num_xgmi_sdma_engines;
1300 dev->node_props.num_gws = (hws_gws_support &&
1301 dev->gpu->dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) ?
1302 amdgpu_amdkfd_get_num_gws(dev->gpu->kgd) : 0;
1304 kfd_fill_mem_clk_max_info(dev);
1305 kfd_fill_iolink_non_crat_info(dev);
1307 switch (dev->gpu->device_info->asic_family) {
1311 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_PRE_1_0 <<
1312 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1313 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1317 case CHIP_POLARIS10:
1318 case CHIP_POLARIS11:
1319 case CHIP_POLARIS12:
1321 pr_debug("Adding doorbell packet type capability\n");
1322 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 <<
1323 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1324 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1335 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 <<
1336 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1337 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1340 WARN(1, "Unexpected ASIC family %u",
1341 dev->gpu->device_info->asic_family);
1345 * Overwrite ATS capability according to needs_iommu_device to fix
1346 * potential missing corresponding bit in CRAT of BIOS.
1348 if (dev->gpu->device_info->needs_iommu_device)
1349 dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
1351 dev->node_props.capability &= ~HSA_CAP_ATS_PRESENT;
1353 /* Fix errors in CZ CRAT.
1354 * simd_count: Carrizo CRAT reports wrong simd_count, probably
1355 * because it doesn't consider masked out CUs
1356 * max_waves_per_simd: Carrizo reports wrong max_waves_per_simd
1358 if (dev->gpu->device_info->asic_family == CHIP_CARRIZO) {
1359 dev->node_props.simd_count =
1360 cu_info.simd_per_cu * cu_info.cu_active_number;
1361 dev->node_props.max_waves_per_simd = 10;
1364 ctx = amdgpu_ras_get_context((struct amdgpu_device *)(dev->gpu->kgd));
1366 /* kfd only concerns sram ecc on GFX/SDMA and HBM ecc on UMC */
1367 dev->node_props.capability |=
1368 (((ctx->features & BIT(AMDGPU_RAS_BLOCK__SDMA)) != 0) ||
1369 ((ctx->features & BIT(AMDGPU_RAS_BLOCK__GFX)) != 0)) ?
1370 HSA_CAP_SRAM_EDCSUPPORTED : 0;
1371 dev->node_props.capability |= ((ctx->features & BIT(AMDGPU_RAS_BLOCK__UMC)) != 0) ?
1372 HSA_CAP_MEM_EDCSUPPORTED : 0;
1374 dev->node_props.capability |= (ctx->features != 0) ?
1375 HSA_CAP_RASEVENTNOTIFY : 0;
1378 kfd_debug_print_topology();
1381 kfd_notify_gpu_change(gpu_id, 1);
1383 kfd_destroy_crat_image(crat_image);
1387 int kfd_topology_remove_device(struct kfd_dev *gpu)
1389 struct kfd_topology_device *dev, *tmp;
1393 down_write(&topology_lock);
1395 list_for_each_entry_safe(dev, tmp, &topology_device_list, list)
1396 if (dev->gpu == gpu) {
1397 gpu_id = dev->gpu_id;
1398 kfd_remove_sysfs_node_entry(dev);
1399 kfd_release_topology_device(dev);
1400 sys_props.num_devices--;
1402 if (kfd_topology_update_sysfs() < 0)
1403 kfd_topology_release_sysfs();
1407 up_write(&topology_lock);
1410 kfd_notify_gpu_change(gpu_id, 0);
1415 /* kfd_topology_enum_kfd_devices - Enumerate through all devices in KFD
1416 * topology. If GPU device is found @idx, then valid kfd_dev pointer is
1417 * returned through @kdev
1418 * Return - 0: On success (@kdev will be NULL for non GPU nodes)
1419 * -1: If end of list
1421 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev)
1424 struct kfd_topology_device *top_dev;
1425 uint8_t device_idx = 0;
1428 down_read(&topology_lock);
1430 list_for_each_entry(top_dev, &topology_device_list, list) {
1431 if (device_idx == idx) {
1432 *kdev = top_dev->gpu;
1433 up_read(&topology_lock);
1440 up_read(&topology_lock);
1446 static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask)
1448 int first_cpu_of_numa_node;
1450 if (!cpumask || cpumask == cpu_none_mask)
1452 first_cpu_of_numa_node = cpumask_first(cpumask);
1453 if (first_cpu_of_numa_node >= nr_cpu_ids)
1455 #ifdef CONFIG_X86_64
1456 return cpu_data(first_cpu_of_numa_node).apicid;
1458 return first_cpu_of_numa_node;
1462 /* kfd_numa_node_to_apic_id - Returns the APIC ID of the first logical processor
1463 * of the given NUMA node (numa_node_id)
1464 * Return -1 on failure
1466 int kfd_numa_node_to_apic_id(int numa_node_id)
1468 if (numa_node_id == -1) {
1469 pr_warn("Invalid NUMA Node. Use online CPU mask\n");
1470 return kfd_cpumask_to_apic_id(cpu_online_mask);
1472 return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id));
1475 #if defined(CONFIG_DEBUG_FS)
1477 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data)
1479 struct kfd_topology_device *dev;
1483 down_read(&topology_lock);
1485 list_for_each_entry(dev, &topology_device_list, list) {
1491 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1492 r = dqm_debugfs_hqds(m, dev->gpu->dqm);
1497 up_read(&topology_lock);
1502 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data)
1504 struct kfd_topology_device *dev;
1508 down_read(&topology_lock);
1510 list_for_each_entry(dev, &topology_device_list, list) {
1516 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1517 r = pm_debugfs_runlist(m, &dev->gpu->dqm->packets);
1522 up_read(&topology_lock);