GNU Linux-libre 4.19.295-gnu1
[releases.git] / drivers / gpu / drm / amd / amdkfd / kfd_topology.c
1 /*
2  * Copyright 2014 Advanced Micro Devices, Inc.
3  *
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:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
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.
21  */
22
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>
33
34 #include "kfd_priv.h"
35 #include "kfd_crat.h"
36 #include "kfd_topology.h"
37 #include "kfd_device_queue_manager.h"
38 #include "kfd_iommu.h"
39
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;
43
44 static DECLARE_RWSEM(topology_lock);
45 static atomic_t topology_crat_proximity_domain;
46
47 struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
48                                                 uint32_t proximity_domain)
49 {
50         struct kfd_topology_device *top_dev;
51         struct kfd_topology_device *device = NULL;
52
53         down_read(&topology_lock);
54
55         list_for_each_entry(top_dev, &topology_device_list, list)
56                 if (top_dev->proximity_domain == proximity_domain) {
57                         device = top_dev;
58                         break;
59                 }
60
61         up_read(&topology_lock);
62
63         return device;
64 }
65
66 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id)
67 {
68         struct kfd_topology_device *top_dev = NULL;
69         struct kfd_topology_device *ret = NULL;
70
71         down_read(&topology_lock);
72
73         list_for_each_entry(top_dev, &topology_device_list, list)
74                 if (top_dev->gpu_id == gpu_id) {
75                         ret = top_dev;
76                         break;
77                 }
78
79         up_read(&topology_lock);
80
81         return ret;
82 }
83
84 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id)
85 {
86         struct kfd_topology_device *top_dev;
87
88         top_dev = kfd_topology_device_by_id(gpu_id);
89         if (!top_dev)
90                 return NULL;
91
92         return top_dev->gpu;
93 }
94
95 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev)
96 {
97         struct kfd_topology_device *top_dev;
98         struct kfd_dev *device = NULL;
99
100         down_read(&topology_lock);
101
102         list_for_each_entry(top_dev, &topology_device_list, list)
103                 if (top_dev->gpu->pdev == pdev) {
104                         device = top_dev->gpu;
105                         break;
106                 }
107
108         up_read(&topology_lock);
109
110         return device;
111 }
112
113 /* Called with write topology_lock acquired */
114 static void kfd_release_topology_device(struct kfd_topology_device *dev)
115 {
116         struct kfd_mem_properties *mem;
117         struct kfd_cache_properties *cache;
118         struct kfd_iolink_properties *iolink;
119         struct kfd_perf_properties *perf;
120
121         list_del(&dev->list);
122
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);
127                 kfree(mem);
128         }
129
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);
134                 kfree(cache);
135         }
136
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);
141                 kfree(iolink);
142         }
143
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);
148                 kfree(perf);
149         }
150
151         kfree(dev);
152 }
153
154 void kfd_release_topology_device_list(struct list_head *device_list)
155 {
156         struct kfd_topology_device *dev;
157
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);
162         }
163 }
164
165 static void kfd_release_live_view(void)
166 {
167         kfd_release_topology_device_list(&topology_device_list);
168         memset(&sys_props, 0, sizeof(sys_props));
169 }
170
171 struct kfd_topology_device *kfd_create_topology_device(
172                                 struct list_head *device_list)
173 {
174         struct kfd_topology_device *dev;
175
176         dev = kfd_alloc_struct(dev);
177         if (!dev) {
178                 pr_err("No memory to allocate a topology device");
179                 return NULL;
180         }
181
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);
186
187         list_add_tail(&dev->list, device_list);
188
189         return dev;
190 }
191
192
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)
203
204 static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr,
205                 char *buffer)
206 {
207         ssize_t ret;
208
209         /* Making sure that the buffer is an empty string */
210         buffer[0] = 0;
211
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);
221         } else {
222                 ret = -EINVAL;
223         }
224
225         return ret;
226 }
227
228 static void kfd_topology_kobj_release(struct kobject *kobj)
229 {
230         kfree(kobj);
231 }
232
233 static const struct sysfs_ops sysprops_ops = {
234         .show = sysprops_show,
235 };
236
237 static struct kobj_type sysprops_type = {
238         .release = kfd_topology_kobj_release,
239         .sysfs_ops = &sysprops_ops,
240 };
241
242 static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr,
243                 char *buffer)
244 {
245         ssize_t ret;
246         struct kfd_iolink_properties *iolink;
247
248         /* Making sure that the buffer is an empty string */
249         buffer[0] = 0;
250
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);
265
266         return ret;
267 }
268
269 static const struct sysfs_ops iolink_ops = {
270         .show = iolink_show,
271 };
272
273 static struct kobj_type iolink_type = {
274         .release = kfd_topology_kobj_release,
275         .sysfs_ops = &iolink_ops,
276 };
277
278 static ssize_t mem_show(struct kobject *kobj, struct attribute *attr,
279                 char *buffer)
280 {
281         ssize_t ret;
282         struct kfd_mem_properties *mem;
283
284         /* Making sure that the buffer is an empty string */
285         buffer[0] = 0;
286
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);
293
294         return ret;
295 }
296
297 static const struct sysfs_ops mem_ops = {
298         .show = mem_show,
299 };
300
301 static struct kobj_type mem_type = {
302         .release = kfd_topology_kobj_release,
303         .sysfs_ops = &mem_ops,
304 };
305
306 static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr,
307                 char *buffer)
308 {
309         ssize_t ret;
310         uint32_t i, j;
311         struct kfd_cache_properties *cache;
312
313         /* Making sure that the buffer is an empty string */
314         buffer[0] = 0;
315
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++) {
330                         /* Check each bit */
331                         if (cache->sibling_map[i] & (1 << j))
332                                 ret = snprintf(buffer, PAGE_SIZE,
333                                          "%s%d%s", buffer, 1, ",");
334                         else
335                                 ret = snprintf(buffer, PAGE_SIZE,
336                                          "%s%d%s", buffer, 0, ",");
337                 }
338         /* Replace the last "," with end of line */
339         *(buffer + strlen(buffer) - 1) = 0xA;
340         return ret;
341 }
342
343 static const struct sysfs_ops cache_ops = {
344         .show = kfd_cache_show,
345 };
346
347 static struct kobj_type cache_type = {
348         .release = kfd_topology_kobj_release,
349         .sysfs_ops = &cache_ops,
350 };
351
352 /****** Sysfs of Performance Counters ******/
353
354 struct kfd_perf_attr {
355         struct kobj_attribute attr;
356         uint32_t data;
357 };
358
359 static ssize_t perf_show(struct kobject *kobj, struct kobj_attribute *attrs,
360                         char *buf)
361 {
362         struct kfd_perf_attr *attr;
363
364         buf[0] = 0;
365         attr = container_of(attrs, struct kfd_perf_attr, attr);
366         if (!attr->data) /* invalid data for PMC */
367                 return 0;
368         else
369                 return sysfs_show_32bit_val(buf, attr->data);
370 }
371
372 #define KFD_PERF_DESC(_name, _data)                     \
373 {                                                       \
374         .attr  = __ATTR(_name, 0444, perf_show, NULL),  \
375         .data = _data,                                  \
376 }
377
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),
382 };
383 /****************************************/
384
385 static ssize_t node_show(struct kobject *kobj, struct attribute *attr,
386                 char *buffer)
387 {
388         struct kfd_topology_device *dev;
389         char public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE];
390         uint32_t i;
391         uint32_t log_max_watch_addr;
392
393         /* Making sure that the buffer is an empty string */
394         buffer[0] = 0;
395
396         if (strcmp(attr->name, "gpu_id") == 0) {
397                 dev = container_of(attr, struct kfd_topology_device,
398                                 attr_gpuid);
399                 return sysfs_show_32bit_val(buffer, dev->gpu_id);
400         }
401
402         if (strcmp(attr->name, "name") == 0) {
403                 dev = container_of(attr, struct kfd_topology_device,
404                                 attr_name);
405                 for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE; i++) {
406                         public_name[i] =
407                                         (char)dev->node_props.marketing_name[i];
408                         if (dev->node_props.marketing_name[i] == 0)
409                                 break;
410                 }
411                 public_name[KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1] = 0x0;
412                 return sysfs_show_str_val(buffer, public_name);
413         }
414
415         dev = container_of(attr, struct kfd_topology_device,
416                         attr_props);
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);
457
458         if (dev->gpu) {
459                 log_max_watch_addr =
460                         __ilog2_u32(dev->gpu->device_info->num_of_watch_points);
461
462                 if (log_max_watch_addr) {
463                         dev->node_props.capability |=
464                                         HSA_CAP_WATCH_POINTS_SUPPORTED;
465
466                         dev->node_props.capability |=
467                                 ((log_max_watch_addr <<
468                                         HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) &
469                                 HSA_CAP_WATCH_POINTS_TOTALBITS_MASK);
470                 }
471
472                 if (dev->gpu->device_info->asic_family == CHIP_TONGA)
473                         dev->node_props.capability |=
474                                         HSA_CAP_AQL_QUEUE_DOUBLE_MAP;
475
476                 sysfs_show_32bit_prop(buffer, "max_engine_clk_fcompute",
477                         dev->node_props.max_engine_clk_fcompute);
478
479                 sysfs_show_64bit_prop(buffer, "local_mem_size",
480                                 (unsigned long long int) 0);
481
482                 sysfs_show_32bit_prop(buffer, "fw_version",
483                         dev->gpu->kfd2kgd->get_fw_version(
484                                                 dev->gpu->kgd,
485                                                 KGD_ENGINE_MEC1));
486                 sysfs_show_32bit_prop(buffer, "capability",
487                                 dev->node_props.capability);
488         }
489
490         return sysfs_show_32bit_prop(buffer, "max_engine_clk_ccompute",
491                                         cpufreq_quick_get_max(0)/1000);
492 }
493
494 static const struct sysfs_ops node_ops = {
495         .show = node_show,
496 };
497
498 static struct kobj_type node_type = {
499         .release = kfd_topology_kobj_release,
500         .sysfs_ops = &node_ops,
501 };
502
503 static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr)
504 {
505         sysfs_remove_file(kobj, attr);
506         kobject_del(kobj);
507         kobject_put(kobj);
508 }
509
510 static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev)
511 {
512         struct kfd_iolink_properties *iolink;
513         struct kfd_cache_properties *cache;
514         struct kfd_mem_properties *mem;
515         struct kfd_perf_properties *perf;
516
517         if (dev->kobj_iolink) {
518                 list_for_each_entry(iolink, &dev->io_link_props, list)
519                         if (iolink->kobj) {
520                                 kfd_remove_sysfs_file(iolink->kobj,
521                                                         &iolink->attr);
522                                 iolink->kobj = NULL;
523                         }
524                 kobject_del(dev->kobj_iolink);
525                 kobject_put(dev->kobj_iolink);
526                 dev->kobj_iolink = NULL;
527         }
528
529         if (dev->kobj_cache) {
530                 list_for_each_entry(cache, &dev->cache_props, list)
531                         if (cache->kobj) {
532                                 kfd_remove_sysfs_file(cache->kobj,
533                                                         &cache->attr);
534                                 cache->kobj = NULL;
535                         }
536                 kobject_del(dev->kobj_cache);
537                 kobject_put(dev->kobj_cache);
538                 dev->kobj_cache = NULL;
539         }
540
541         if (dev->kobj_mem) {
542                 list_for_each_entry(mem, &dev->mem_props, list)
543                         if (mem->kobj) {
544                                 kfd_remove_sysfs_file(mem->kobj, &mem->attr);
545                                 mem->kobj = NULL;
546                         }
547                 kobject_del(dev->kobj_mem);
548                 kobject_put(dev->kobj_mem);
549                 dev->kobj_mem = NULL;
550         }
551
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;
556                 }
557                 kobject_del(dev->kobj_perf);
558                 kobject_put(dev->kobj_perf);
559                 dev->kobj_perf = NULL;
560         }
561
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;
569         }
570 }
571
572 static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev,
573                 uint32_t id)
574 {
575         struct kfd_iolink_properties *iolink;
576         struct kfd_cache_properties *cache;
577         struct kfd_mem_properties *mem;
578         struct kfd_perf_properties *perf;
579         int ret;
580         uint32_t i, num_attrs;
581         struct attribute **attrs;
582
583         if (WARN_ON(dev->kobj_node))
584                 return -EEXIST;
585
586         /*
587          * Creating the sysfs folders
588          */
589         dev->kobj_node = kfd_alloc_struct(dev->kobj_node);
590         if (!dev->kobj_node)
591                 return -ENOMEM;
592
593         ret = kobject_init_and_add(dev->kobj_node, &node_type,
594                         sys_props.kobj_nodes, "%d", id);
595         if (ret < 0) {
596                 kobject_put(dev->kobj_node);
597                 return ret;
598         }
599
600         dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node);
601         if (!dev->kobj_mem)
602                 return -ENOMEM;
603
604         dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node);
605         if (!dev->kobj_cache)
606                 return -ENOMEM;
607
608         dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node);
609         if (!dev->kobj_iolink)
610                 return -ENOMEM;
611
612         dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node);
613         if (!dev->kobj_perf)
614                 return -ENOMEM;
615
616         /*
617          * Creating sysfs files for node properties
618          */
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);
629         if (ret < 0)
630                 return ret;
631         ret = sysfs_create_file(dev->kobj_node, &dev->attr_name);
632         if (ret < 0)
633                 return ret;
634         ret = sysfs_create_file(dev->kobj_node, &dev->attr_props);
635         if (ret < 0)
636                 return ret;
637
638         i = 0;
639         list_for_each_entry(mem, &dev->mem_props, list) {
640                 mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
641                 if (!mem->kobj)
642                         return -ENOMEM;
643                 ret = kobject_init_and_add(mem->kobj, &mem_type,
644                                 dev->kobj_mem, "%d", i);
645                 if (ret < 0) {
646                         kobject_put(mem->kobj);
647                         return ret;
648                 }
649
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);
654                 if (ret < 0)
655                         return ret;
656                 i++;
657         }
658
659         i = 0;
660         list_for_each_entry(cache, &dev->cache_props, list) {
661                 cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
662                 if (!cache->kobj)
663                         return -ENOMEM;
664                 ret = kobject_init_and_add(cache->kobj, &cache_type,
665                                 dev->kobj_cache, "%d", i);
666                 if (ret < 0) {
667                         kobject_put(cache->kobj);
668                         return ret;
669                 }
670
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);
675                 if (ret < 0)
676                         return ret;
677                 i++;
678         }
679
680         i = 0;
681         list_for_each_entry(iolink, &dev->io_link_props, list) {
682                 iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
683                 if (!iolink->kobj)
684                         return -ENOMEM;
685                 ret = kobject_init_and_add(iolink->kobj, &iolink_type,
686                                 dev->kobj_iolink, "%d", i);
687                 if (ret < 0) {
688                         kobject_put(iolink->kobj);
689                         return ret;
690                 }
691
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);
696                 if (ret < 0)
697                         return ret;
698                 i++;
699         }
700
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),
706                         GFP_KERNEL);
707                 if (!perf->attr_group)
708                         return -ENOMEM;
709
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
714                  * duplicate here.
715                  */
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;
719                 }
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);
723                 if (ret < 0)
724                         return ret;
725         }
726
727         return 0;
728 }
729
730 /* Called with write topology lock acquired */
731 static int kfd_build_sysfs_node_tree(void)
732 {
733         struct kfd_topology_device *dev;
734         int ret;
735         uint32_t i = 0;
736
737         list_for_each_entry(dev, &topology_device_list, list) {
738                 ret = kfd_build_sysfs_node_entry(dev, i);
739                 if (ret < 0)
740                         return ret;
741                 i++;
742         }
743
744         return 0;
745 }
746
747 /* Called with write topology lock acquired */
748 static void kfd_remove_sysfs_node_tree(void)
749 {
750         struct kfd_topology_device *dev;
751
752         list_for_each_entry(dev, &topology_device_list, list)
753                 kfd_remove_sysfs_node_entry(dev);
754 }
755
756 static int kfd_topology_update_sysfs(void)
757 {
758         int ret;
759
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)
765                         return -ENOMEM;
766
767                 ret = kobject_init_and_add(sys_props.kobj_topology,
768                                 &sysprops_type,  &kfd_device->kobj,
769                                 "topology");
770                 if (ret < 0) {
771                         kobject_put(sys_props.kobj_topology);
772                         return ret;
773                 }
774
775                 sys_props.kobj_nodes = kobject_create_and_add("nodes",
776                                 sys_props.kobj_topology);
777                 if (!sys_props.kobj_nodes)
778                         return -ENOMEM;
779
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);
785                 if (ret < 0)
786                         return ret;
787
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);
793                 if (ret < 0)
794                         return ret;
795         }
796
797         kfd_remove_sysfs_node_tree();
798
799         return kfd_build_sysfs_node_tree();
800 }
801
802 static void kfd_topology_release_sysfs(void)
803 {
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;
814                 }
815                 kobject_del(sys_props.kobj_topology);
816                 kobject_put(sys_props.kobj_topology);
817                 sys_props.kobj_topology = NULL;
818         }
819 }
820
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)
824 {
825         while (!list_empty(temp_list)) {
826                 list_move_tail(temp_list->next, master_list);
827                 sys_props.num_devices++;
828         }
829 }
830
831 static void kfd_debug_print_topology(void)
832 {
833         struct kfd_topology_device *dev;
834
835         down_read(&topology_lock);
836
837         dev = list_last_entry(&topology_device_list,
838                         struct kfd_topology_device, list);
839         if (dev) {
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);
851         }
852         up_read(&topology_lock);
853 }
854
855 /* Helper function for intializing platform_xx members of
856  * kfd_system_properties. Uses OEM info from the last CPU/APU node.
857  */
858 static void kfd_update_system_properties(void)
859 {
860         struct kfd_topology_device *dev;
861
862         down_read(&topology_lock);
863         dev = list_last_entry(&topology_device_list,
864                         struct kfd_topology_device, list);
865         if (dev) {
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;
870         }
871         up_read(&topology_lock);
872 }
873
874 static void find_system_memory(const struct dmi_header *dm,
875         void *private)
876 {
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);
882
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;
889                         if (mem_clock != 0)
890                                 mem->mem_clk_max = mem_clock;
891                 }
892         }
893 }
894
895 /*
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.
899  */
900 static int kfd_add_perf_to_topology(struct kfd_topology_device *kdev)
901 {
902         /* These are the only counters supported so far */
903         return kfd_iommu_add_perf_counters(kdev);
904 }
905
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
909  */
910 static void kfd_add_non_crat_information(struct kfd_topology_device *kdev)
911 {
912         /* Check if CPU only node. */
913         if (!kdev->gpu) {
914                 /* Add system memory information */
915                 dmi_walk(find_system_memory, kdev);
916         }
917         /* TODO: For GPU node, rearrange code from kfd_topology_add_device */
918 }
919
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.
925  */
926 static bool kfd_is_acpi_crat_invalid(struct list_head *device_list)
927 {
928         struct kfd_topology_device *dev;
929
930         list_for_each_entry(dev, device_list, list) {
931                 if (dev->node_props.cpu_cores_count &&
932                         dev->node_props.simd_count)
933                         return false;
934         }
935         pr_info("Ignoring ACPI CRAT on non-APU system\n");
936         return true;
937 }
938
939 int kfd_topology_init(void)
940 {
941         void *crat_image = NULL;
942         size_t image_size = 0;
943         int ret;
944         struct list_head temp_topology_device_list;
945         int cpu_only_node = 0;
946         struct kfd_topology_device *kdev;
947         int proximity_domain;
948
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
953          */
954
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);
959
960         memset(&sys_props, 0, sizeof(sys_props));
961
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
965          * use a counter.
966          */
967         proximity_domain = 0;
968
969         /*
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
974          */
975         ret = kfd_create_crat_image_acpi(&crat_image, &image_size);
976         if (!ret) {
977                 ret = kfd_parse_crat_table(crat_image,
978                                            &temp_topology_device_list,
979                                            proximity_domain);
980                 if (ret ||
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);
985                         crat_image = NULL;
986                 }
987         }
988
989         if (!crat_image) {
990                 ret = kfd_create_crat_image_virtual(&crat_image, &image_size,
991                                                     COMPUTE_UNIT_CPU, NULL,
992                                                     proximity_domain);
993                 cpu_only_node = 1;
994                 if (ret) {
995                         pr_err("Error creating VCRAT table for CPU\n");
996                         return ret;
997                 }
998
999                 ret = kfd_parse_crat_table(crat_image,
1000                                            &temp_topology_device_list,
1001                                            proximity_domain);
1002                 if (ret) {
1003                         pr_err("Error parsing VCRAT table for CPU\n");
1004                         goto err;
1005                 }
1006         }
1007
1008         kdev = list_first_entry(&temp_topology_device_list,
1009                                 struct kfd_topology_device, list);
1010         kfd_add_perf_to_topology(kdev);
1011
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);
1018
1019         if (!ret) {
1020                 sys_props.generation_count++;
1021                 kfd_update_system_properties();
1022                 kfd_debug_print_topology();
1023                 pr_info("Finished initializing topology\n");
1024         } else
1025                 pr_err("Failed to update topology in sysfs ret=%d\n", ret);
1026
1027         /* For nodes with GPU, this information gets added
1028          * when GPU is detected (kfd_topology_add_device).
1029          */
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);
1037         }
1038
1039 err:
1040         kfd_destroy_crat_image(crat_image);
1041         return ret;
1042 }
1043
1044 void kfd_topology_shutdown(void)
1045 {
1046         down_write(&topology_lock);
1047         kfd_topology_release_sysfs();
1048         kfd_release_live_view();
1049         up_write(&topology_lock);
1050 }
1051
1052 static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu)
1053 {
1054         uint32_t hashout;
1055         uint32_t buf[7];
1056         uint64_t local_mem_size;
1057         int i;
1058         struct kfd_local_mem_info local_mem_info;
1059
1060         if (!gpu)
1061                 return 0;
1062
1063         gpu->kfd2kgd->get_local_mem_info(gpu->kgd, &local_mem_info);
1064
1065         local_mem_size = local_mem_info.local_mem_size_private +
1066                         local_mem_info.local_mem_size_public;
1067
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);
1075
1076         for (i = 0, hashout = 0; i < 7; i++)
1077                 hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH);
1078
1079         return hashout;
1080 }
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.
1085  */
1086 static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu)
1087 {
1088         struct kfd_topology_device *dev;
1089         struct kfd_topology_device *out_dev = NULL;
1090
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.
1095                  */
1096                 if (!gpu->device_info->needs_iommu_device &&
1097                     dev->node_props.cpu_cores_count)
1098                         continue;
1099
1100                 if (!dev->gpu && (dev->node_props.simd_count > 0)) {
1101                         dev->gpu = gpu;
1102                         out_dev = dev;
1103                         break;
1104                 }
1105         }
1106         up_write(&topology_lock);
1107         return out_dev;
1108 }
1109
1110 static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival)
1111 {
1112         /*
1113          * TODO: Generate an event for thunk about the arrival/removal
1114          * of the GPU
1115          */
1116 }
1117
1118 /* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info,
1119  *              patch this after CRAT parsing.
1120  */
1121 static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev)
1122 {
1123         struct kfd_mem_properties *mem;
1124         struct kfd_local_mem_info local_mem_info;
1125
1126         if (!dev)
1127                 return;
1128
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
1134          */
1135         dev->gpu->kfd2kgd->get_local_mem_info(dev->gpu->kgd,
1136                 &local_mem_info);
1137
1138         list_for_each_entry(mem, &dev->mem_props, list)
1139                 mem->mem_clk_max = local_mem_info.mem_clk_max;
1140 }
1141
1142 static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev)
1143 {
1144         struct kfd_iolink_properties *link;
1145
1146         if (!dev || !dev->gpu)
1147                 return;
1148
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;
1155 }
1156
1157 int kfd_topology_add_device(struct kfd_dev *gpu)
1158 {
1159         uint32_t gpu_id;
1160         struct kfd_topology_device *dev;
1161         struct kfd_cu_info cu_info;
1162         int res = 0;
1163         struct list_head temp_topology_device_list;
1164         void *crat_image = NULL;
1165         size_t image_size = 0;
1166         int proximity_domain;
1167
1168         INIT_LIST_HEAD(&temp_topology_device_list);
1169
1170         gpu_id = kfd_generate_gpu_id(gpu);
1171
1172         pr_debug("Adding new GPU (ID: 0x%x) to topology\n", gpu_id);
1173
1174         proximity_domain = atomic_inc_return(&topology_crat_proximity_domain);
1175
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
1181          */
1182         dev = kfd_assign_gpu(gpu);
1183         if (!dev) {
1184                 res = kfd_create_crat_image_virtual(&crat_image, &image_size,
1185                                                     COMPUTE_UNIT_GPU, gpu,
1186                                                     proximity_domain);
1187                 if (res) {
1188                         pr_err("Error creating VCRAT for GPU (ID: 0x%x)\n",
1189                                gpu_id);
1190                         return res;
1191                 }
1192                 res = kfd_parse_crat_table(crat_image,
1193                                            &temp_topology_device_list,
1194                                            proximity_domain);
1195                 if (res) {
1196                         pr_err("Error parsing VCRAT for GPU (ID: 0x%x)\n",
1197                                gpu_id);
1198                         goto err;
1199                 }
1200
1201                 down_write(&topology_lock);
1202                 kfd_topology_update_device_list(&temp_topology_device_list,
1203                         &topology_device_list);
1204
1205                 /* Update the SYSFS tree, since we added another topology
1206                  * device
1207                  */
1208                 res = kfd_topology_update_sysfs();
1209                 up_write(&topology_lock);
1210
1211                 if (!res)
1212                         sys_props.generation_count++;
1213                 else
1214                         pr_err("Failed to update GPU (ID: 0x%x) to sysfs topology. res=%d\n",
1215                                                 gpu_id, res);
1216                 dev = kfd_assign_gpu(gpu);
1217                 if (WARN_ON(!dev)) {
1218                         res = -ENODEV;
1219                         goto err;
1220                 }
1221         }
1222
1223         dev->gpu_id = gpu_id;
1224         gpu->id = gpu_id;
1225
1226         /* TODO: Move the following lines to function
1227          *      kfd_add_non_crat_information
1228          */
1229
1230         /* Fill-in additional information that is not available in CRAT but
1231          * needed for the topology
1232          */
1233
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;
1237
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,
1241                 gpu->pdev->devfn);
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;
1248
1249         kfd_fill_mem_clk_max_info(dev);
1250         kfd_fill_iolink_non_crat_info(dev);
1251
1252         switch (dev->gpu->device_info->asic_family) {
1253         case CHIP_KAVERI:
1254         case CHIP_HAWAII:
1255         case CHIP_TONGA:
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);
1259                 break;
1260         case CHIP_CARRIZO:
1261         case CHIP_FIJI:
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);
1268                 break;
1269         case CHIP_VEGA10:
1270         case CHIP_RAVEN:
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);
1274                 break;
1275         default:
1276                 WARN(1, "Unexpected ASIC family %u",
1277                      dev->gpu->device_info->asic_family);
1278         }
1279
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
1285          */
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;
1291         }
1292
1293         kfd_debug_print_topology();
1294
1295         if (!res)
1296                 kfd_notify_gpu_change(gpu_id, 1);
1297 err:
1298         kfd_destroy_crat_image(crat_image);
1299         return res;
1300 }
1301
1302 int kfd_topology_remove_device(struct kfd_dev *gpu)
1303 {
1304         struct kfd_topology_device *dev, *tmp;
1305         uint32_t gpu_id;
1306         int res = -ENODEV;
1307
1308         down_write(&topology_lock);
1309
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--;
1316                         res = 0;
1317                         if (kfd_topology_update_sysfs() < 0)
1318                                 kfd_topology_release_sysfs();
1319                         break;
1320                 }
1321
1322         up_write(&topology_lock);
1323
1324         if (!res)
1325                 kfd_notify_gpu_change(gpu_id, 0);
1326
1327         return res;
1328 }
1329
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
1335  */
1336 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev)
1337 {
1338
1339         struct kfd_topology_device *top_dev;
1340         uint8_t device_idx = 0;
1341
1342         *kdev = NULL;
1343         down_read(&topology_lock);
1344
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);
1349                         return 0;
1350                 }
1351
1352                 device_idx++;
1353         }
1354
1355         up_read(&topology_lock);
1356
1357         return -1;
1358
1359 }
1360
1361 static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask)
1362 {
1363         const struct cpuinfo_x86 *cpuinfo;
1364         int first_cpu_of_numa_node;
1365
1366         if (!cpumask || cpumask == cpu_none_mask)
1367                 return -1;
1368         first_cpu_of_numa_node = cpumask_first(cpumask);
1369         if (first_cpu_of_numa_node >= nr_cpu_ids)
1370                 return -1;
1371         cpuinfo = &cpu_data(first_cpu_of_numa_node);
1372
1373         return cpuinfo->apicid;
1374 }
1375
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
1379  */
1380 int kfd_numa_node_to_apic_id(int numa_node_id)
1381 {
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);
1385         }
1386         return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id));
1387 }
1388
1389 #if defined(CONFIG_DEBUG_FS)
1390
1391 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data)
1392 {
1393         struct kfd_topology_device *dev;
1394         unsigned int i = 0;
1395         int r = 0;
1396
1397         down_read(&topology_lock);
1398
1399         list_for_each_entry(dev, &topology_device_list, list) {
1400                 if (!dev->gpu) {
1401                         i++;
1402                         continue;
1403                 }
1404
1405                 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1406                 r = dqm_debugfs_hqds(m, dev->gpu->dqm);
1407                 if (r)
1408                         break;
1409         }
1410
1411         up_read(&topology_lock);
1412
1413         return r;
1414 }
1415
1416 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data)
1417 {
1418         struct kfd_topology_device *dev;
1419         unsigned int i = 0;
1420         int r = 0;
1421
1422         down_read(&topology_lock);
1423
1424         list_for_each_entry(dev, &topology_device_list, list) {
1425                 if (!dev->gpu) {
1426                         i++;
1427                         continue;
1428                 }
1429
1430                 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1431                 r = pm_debugfs_runlist(m, &dev->gpu->dqm->packets);
1432                 if (r)
1433                         break;
1434         }
1435
1436         up_read(&topology_lock);
1437
1438         return r;
1439 }
1440
1441 #endif