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
kfd_topology_device_by_proximity_domain(uint32_t proximity_domain)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
kfd_topology_device_by_id(uint32_t gpu_id)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
kfd_device_by_id(uint32_t gpu_id)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
kfd_device_by_pci_dev(const struct pci_dev * pdev)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 */
kfd_release_topology_device(struct kfd_topology_device * dev)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
kfd_release_topology_device_list(struct list_head * device_list)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
kfd_release_live_view(void)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
kfd_create_topology_device(struct list_head * device_list)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
sysprops_show(struct kobject * kobj,struct attribute * attr,char * buffer)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
kfd_topology_kobj_release(struct kobject * kobj)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
iolink_show(struct kobject * kobj,struct attribute * attr,char * buffer)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
mem_show(struct kobject * kobj,struct attribute * attr,char * buffer)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
kfd_cache_show(struct kobject * kobj,struct attribute * attr,char * buffer)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
perf_show(struct kobject * kobj,struct kobj_attribute * attrs,char * buf)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
node_show(struct kobject * kobj,struct attribute * attr,char * buffer)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
kfd_remove_sysfs_file(struct kobject * kobj,struct attribute * attr)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
kfd_remove_sysfs_node_entry(struct kfd_topology_device * dev)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
kfd_build_sysfs_node_entry(struct kfd_topology_device * dev,uint32_t id)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 return ret;
597
598 dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node);
599 if (!dev->kobj_mem)
600 return -ENOMEM;
601
602 dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node);
603 if (!dev->kobj_cache)
604 return -ENOMEM;
605
606 dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node);
607 if (!dev->kobj_iolink)
608 return -ENOMEM;
609
610 dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node);
611 if (!dev->kobj_perf)
612 return -ENOMEM;
613
614 /*
615 * Creating sysfs files for node properties
616 */
617 dev->attr_gpuid.name = "gpu_id";
618 dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE;
619 sysfs_attr_init(&dev->attr_gpuid);
620 dev->attr_name.name = "name";
621 dev->attr_name.mode = KFD_SYSFS_FILE_MODE;
622 sysfs_attr_init(&dev->attr_name);
623 dev->attr_props.name = "properties";
624 dev->attr_props.mode = KFD_SYSFS_FILE_MODE;
625 sysfs_attr_init(&dev->attr_props);
626 ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid);
627 if (ret < 0)
628 return ret;
629 ret = sysfs_create_file(dev->kobj_node, &dev->attr_name);
630 if (ret < 0)
631 return ret;
632 ret = sysfs_create_file(dev->kobj_node, &dev->attr_props);
633 if (ret < 0)
634 return ret;
635
636 i = 0;
637 list_for_each_entry(mem, &dev->mem_props, list) {
638 mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
639 if (!mem->kobj)
640 return -ENOMEM;
641 ret = kobject_init_and_add(mem->kobj, &mem_type,
642 dev->kobj_mem, "%d", i);
643 if (ret < 0)
644 return ret;
645
646 mem->attr.name = "properties";
647 mem->attr.mode = KFD_SYSFS_FILE_MODE;
648 sysfs_attr_init(&mem->attr);
649 ret = sysfs_create_file(mem->kobj, &mem->attr);
650 if (ret < 0)
651 return ret;
652 i++;
653 }
654
655 i = 0;
656 list_for_each_entry(cache, &dev->cache_props, list) {
657 cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
658 if (!cache->kobj)
659 return -ENOMEM;
660 ret = kobject_init_and_add(cache->kobj, &cache_type,
661 dev->kobj_cache, "%d", i);
662 if (ret < 0)
663 return ret;
664
665 cache->attr.name = "properties";
666 cache->attr.mode = KFD_SYSFS_FILE_MODE;
667 sysfs_attr_init(&cache->attr);
668 ret = sysfs_create_file(cache->kobj, &cache->attr);
669 if (ret < 0)
670 return ret;
671 i++;
672 }
673
674 i = 0;
675 list_for_each_entry(iolink, &dev->io_link_props, list) {
676 iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
677 if (!iolink->kobj)
678 return -ENOMEM;
679 ret = kobject_init_and_add(iolink->kobj, &iolink_type,
680 dev->kobj_iolink, "%d", i);
681 if (ret < 0)
682 return ret;
683
684 iolink->attr.name = "properties";
685 iolink->attr.mode = KFD_SYSFS_FILE_MODE;
686 sysfs_attr_init(&iolink->attr);
687 ret = sysfs_create_file(iolink->kobj, &iolink->attr);
688 if (ret < 0)
689 return ret;
690 i++;
691 }
692
693 /* All hardware blocks have the same number of attributes. */
694 num_attrs = ARRAY_SIZE(perf_attr_iommu);
695 list_for_each_entry(perf, &dev->perf_props, list) {
696 perf->attr_group = kzalloc(sizeof(struct kfd_perf_attr)
697 * num_attrs + sizeof(struct attribute_group),
698 GFP_KERNEL);
699 if (!perf->attr_group)
700 return -ENOMEM;
701
702 attrs = (struct attribute **)(perf->attr_group + 1);
703 if (!strcmp(perf->block_name, "iommu")) {
704 /* Information of IOMMU's num_counters and counter_ids is shown
705 * under /sys/bus/event_source/devices/amd_iommu. We don't
706 * duplicate here.
707 */
708 perf_attr_iommu[0].data = perf->max_concurrent;
709 for (i = 0; i < num_attrs; i++)
710 attrs[i] = &perf_attr_iommu[i].attr.attr;
711 }
712 perf->attr_group->name = perf->block_name;
713 perf->attr_group->attrs = attrs;
714 ret = sysfs_create_group(dev->kobj_perf, perf->attr_group);
715 if (ret < 0)
716 return ret;
717 }
718
719 return 0;
720 }
721
722 /* Called with write topology lock acquired */
kfd_build_sysfs_node_tree(void)723 static int kfd_build_sysfs_node_tree(void)
724 {
725 struct kfd_topology_device *dev;
726 int ret;
727 uint32_t i = 0;
728
729 list_for_each_entry(dev, &topology_device_list, list) {
730 ret = kfd_build_sysfs_node_entry(dev, i);
731 if (ret < 0)
732 return ret;
733 i++;
734 }
735
736 return 0;
737 }
738
739 /* Called with write topology lock acquired */
kfd_remove_sysfs_node_tree(void)740 static void kfd_remove_sysfs_node_tree(void)
741 {
742 struct kfd_topology_device *dev;
743
744 list_for_each_entry(dev, &topology_device_list, list)
745 kfd_remove_sysfs_node_entry(dev);
746 }
747
kfd_topology_update_sysfs(void)748 static int kfd_topology_update_sysfs(void)
749 {
750 int ret;
751
752 pr_info("Creating topology SYSFS entries\n");
753 if (!sys_props.kobj_topology) {
754 sys_props.kobj_topology =
755 kfd_alloc_struct(sys_props.kobj_topology);
756 if (!sys_props.kobj_topology)
757 return -ENOMEM;
758
759 ret = kobject_init_and_add(sys_props.kobj_topology,
760 &sysprops_type, &kfd_device->kobj,
761 "topology");
762 if (ret < 0)
763 return ret;
764
765 sys_props.kobj_nodes = kobject_create_and_add("nodes",
766 sys_props.kobj_topology);
767 if (!sys_props.kobj_nodes)
768 return -ENOMEM;
769
770 sys_props.attr_genid.name = "generation_id";
771 sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE;
772 sysfs_attr_init(&sys_props.attr_genid);
773 ret = sysfs_create_file(sys_props.kobj_topology,
774 &sys_props.attr_genid);
775 if (ret < 0)
776 return ret;
777
778 sys_props.attr_props.name = "system_properties";
779 sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE;
780 sysfs_attr_init(&sys_props.attr_props);
781 ret = sysfs_create_file(sys_props.kobj_topology,
782 &sys_props.attr_props);
783 if (ret < 0)
784 return ret;
785 }
786
787 kfd_remove_sysfs_node_tree();
788
789 return kfd_build_sysfs_node_tree();
790 }
791
kfd_topology_release_sysfs(void)792 static void kfd_topology_release_sysfs(void)
793 {
794 kfd_remove_sysfs_node_tree();
795 if (sys_props.kobj_topology) {
796 sysfs_remove_file(sys_props.kobj_topology,
797 &sys_props.attr_genid);
798 sysfs_remove_file(sys_props.kobj_topology,
799 &sys_props.attr_props);
800 if (sys_props.kobj_nodes) {
801 kobject_del(sys_props.kobj_nodes);
802 kobject_put(sys_props.kobj_nodes);
803 sys_props.kobj_nodes = NULL;
804 }
805 kobject_del(sys_props.kobj_topology);
806 kobject_put(sys_props.kobj_topology);
807 sys_props.kobj_topology = NULL;
808 }
809 }
810
811 /* Called with write topology_lock acquired */
kfd_topology_update_device_list(struct list_head * temp_list,struct list_head * master_list)812 static void kfd_topology_update_device_list(struct list_head *temp_list,
813 struct list_head *master_list)
814 {
815 while (!list_empty(temp_list)) {
816 list_move_tail(temp_list->next, master_list);
817 sys_props.num_devices++;
818 }
819 }
820
kfd_debug_print_topology(void)821 static void kfd_debug_print_topology(void)
822 {
823 struct kfd_topology_device *dev;
824
825 down_read(&topology_lock);
826
827 dev = list_last_entry(&topology_device_list,
828 struct kfd_topology_device, list);
829 if (dev) {
830 if (dev->node_props.cpu_cores_count &&
831 dev->node_props.simd_count) {
832 pr_info("Topology: Add APU node [0x%0x:0x%0x]\n",
833 dev->node_props.device_id,
834 dev->node_props.vendor_id);
835 } else if (dev->node_props.cpu_cores_count)
836 pr_info("Topology: Add CPU node\n");
837 else if (dev->node_props.simd_count)
838 pr_info("Topology: Add dGPU node [0x%0x:0x%0x]\n",
839 dev->node_props.device_id,
840 dev->node_props.vendor_id);
841 }
842 up_read(&topology_lock);
843 }
844
845 /* Helper function for intializing platform_xx members of
846 * kfd_system_properties. Uses OEM info from the last CPU/APU node.
847 */
kfd_update_system_properties(void)848 static void kfd_update_system_properties(void)
849 {
850 struct kfd_topology_device *dev;
851
852 down_read(&topology_lock);
853 dev = list_last_entry(&topology_device_list,
854 struct kfd_topology_device, list);
855 if (dev) {
856 sys_props.platform_id =
857 (*((uint64_t *)dev->oem_id)) & CRAT_OEMID_64BIT_MASK;
858 sys_props.platform_oem = *((uint64_t *)dev->oem_table_id);
859 sys_props.platform_rev = dev->oem_revision;
860 }
861 up_read(&topology_lock);
862 }
863
find_system_memory(const struct dmi_header * dm,void * private)864 static void find_system_memory(const struct dmi_header *dm,
865 void *private)
866 {
867 struct kfd_mem_properties *mem;
868 u16 mem_width, mem_clock;
869 struct kfd_topology_device *kdev =
870 (struct kfd_topology_device *)private;
871 const u8 *dmi_data = (const u8 *)(dm + 1);
872
873 if (dm->type == DMI_ENTRY_MEM_DEVICE && dm->length >= 0x15) {
874 mem_width = (u16)(*(const u16 *)(dmi_data + 0x6));
875 mem_clock = (u16)(*(const u16 *)(dmi_data + 0x11));
876 list_for_each_entry(mem, &kdev->mem_props, list) {
877 if (mem_width != 0xFFFF && mem_width != 0)
878 mem->width = mem_width;
879 if (mem_clock != 0)
880 mem->mem_clk_max = mem_clock;
881 }
882 }
883 }
884
885 /*
886 * Performance counters information is not part of CRAT but we would like to
887 * put them in the sysfs under topology directory for Thunk to get the data.
888 * This function is called before updating the sysfs.
889 */
kfd_add_perf_to_topology(struct kfd_topology_device * kdev)890 static int kfd_add_perf_to_topology(struct kfd_topology_device *kdev)
891 {
892 /* These are the only counters supported so far */
893 return kfd_iommu_add_perf_counters(kdev);
894 }
895
896 /* kfd_add_non_crat_information - Add information that is not currently
897 * defined in CRAT but is necessary for KFD topology
898 * @dev - topology device to which addition info is added
899 */
kfd_add_non_crat_information(struct kfd_topology_device * kdev)900 static void kfd_add_non_crat_information(struct kfd_topology_device *kdev)
901 {
902 /* Check if CPU only node. */
903 if (!kdev->gpu) {
904 /* Add system memory information */
905 dmi_walk(find_system_memory, kdev);
906 }
907 /* TODO: For GPU node, rearrange code from kfd_topology_add_device */
908 }
909
910 /* kfd_is_acpi_crat_invalid - CRAT from ACPI is valid only for AMD APU devices.
911 * Ignore CRAT for all other devices. AMD APU is identified if both CPU
912 * and GPU cores are present.
913 * @device_list - topology device list created by parsing ACPI CRAT table.
914 * @return - TRUE if invalid, FALSE is valid.
915 */
kfd_is_acpi_crat_invalid(struct list_head * device_list)916 static bool kfd_is_acpi_crat_invalid(struct list_head *device_list)
917 {
918 struct kfd_topology_device *dev;
919
920 list_for_each_entry(dev, device_list, list) {
921 if (dev->node_props.cpu_cores_count &&
922 dev->node_props.simd_count)
923 return false;
924 }
925 pr_info("Ignoring ACPI CRAT on non-APU system\n");
926 return true;
927 }
928
kfd_topology_init(void)929 int kfd_topology_init(void)
930 {
931 void *crat_image = NULL;
932 size_t image_size = 0;
933 int ret;
934 struct list_head temp_topology_device_list;
935 int cpu_only_node = 0;
936 struct kfd_topology_device *kdev;
937 int proximity_domain;
938
939 /* topology_device_list - Master list of all topology devices
940 * temp_topology_device_list - temporary list created while parsing CRAT
941 * or VCRAT. Once parsing is complete the contents of list is moved to
942 * topology_device_list
943 */
944
945 /* Initialize the head for the both the lists */
946 INIT_LIST_HEAD(&topology_device_list);
947 INIT_LIST_HEAD(&temp_topology_device_list);
948 init_rwsem(&topology_lock);
949
950 memset(&sys_props, 0, sizeof(sys_props));
951
952 /* Proximity domains in ACPI CRAT tables start counting at
953 * 0. The same should be true for virtual CRAT tables created
954 * at this stage. GPUs added later in kfd_topology_add_device
955 * use a counter.
956 */
957 proximity_domain = 0;
958
959 /*
960 * Get the CRAT image from the ACPI. If ACPI doesn't have one
961 * or if ACPI CRAT is invalid create a virtual CRAT.
962 * NOTE: The current implementation expects all AMD APUs to have
963 * CRAT. If no CRAT is available, it is assumed to be a CPU
964 */
965 ret = kfd_create_crat_image_acpi(&crat_image, &image_size);
966 if (!ret) {
967 ret = kfd_parse_crat_table(crat_image,
968 &temp_topology_device_list,
969 proximity_domain);
970 if (ret ||
971 kfd_is_acpi_crat_invalid(&temp_topology_device_list)) {
972 kfd_release_topology_device_list(
973 &temp_topology_device_list);
974 kfd_destroy_crat_image(crat_image);
975 crat_image = NULL;
976 }
977 }
978
979 if (!crat_image) {
980 ret = kfd_create_crat_image_virtual(&crat_image, &image_size,
981 COMPUTE_UNIT_CPU, NULL,
982 proximity_domain);
983 cpu_only_node = 1;
984 if (ret) {
985 pr_err("Error creating VCRAT table for CPU\n");
986 return ret;
987 }
988
989 ret = kfd_parse_crat_table(crat_image,
990 &temp_topology_device_list,
991 proximity_domain);
992 if (ret) {
993 pr_err("Error parsing VCRAT table for CPU\n");
994 goto err;
995 }
996 }
997
998 kdev = list_first_entry(&temp_topology_device_list,
999 struct kfd_topology_device, list);
1000 kfd_add_perf_to_topology(kdev);
1001
1002 down_write(&topology_lock);
1003 kfd_topology_update_device_list(&temp_topology_device_list,
1004 &topology_device_list);
1005 atomic_set(&topology_crat_proximity_domain, sys_props.num_devices-1);
1006 ret = kfd_topology_update_sysfs();
1007 up_write(&topology_lock);
1008
1009 if (!ret) {
1010 sys_props.generation_count++;
1011 kfd_update_system_properties();
1012 kfd_debug_print_topology();
1013 pr_info("Finished initializing topology\n");
1014 } else
1015 pr_err("Failed to update topology in sysfs ret=%d\n", ret);
1016
1017 /* For nodes with GPU, this information gets added
1018 * when GPU is detected (kfd_topology_add_device).
1019 */
1020 if (cpu_only_node) {
1021 /* Add additional information to CPU only node created above */
1022 down_write(&topology_lock);
1023 kdev = list_first_entry(&topology_device_list,
1024 struct kfd_topology_device, list);
1025 up_write(&topology_lock);
1026 kfd_add_non_crat_information(kdev);
1027 }
1028
1029 err:
1030 kfd_destroy_crat_image(crat_image);
1031 return ret;
1032 }
1033
kfd_topology_shutdown(void)1034 void kfd_topology_shutdown(void)
1035 {
1036 down_write(&topology_lock);
1037 kfd_topology_release_sysfs();
1038 kfd_release_live_view();
1039 up_write(&topology_lock);
1040 }
1041
kfd_generate_gpu_id(struct kfd_dev * gpu)1042 static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu)
1043 {
1044 uint32_t hashout;
1045 uint32_t buf[7];
1046 uint64_t local_mem_size;
1047 int i;
1048 struct kfd_local_mem_info local_mem_info;
1049
1050 if (!gpu)
1051 return 0;
1052
1053 gpu->kfd2kgd->get_local_mem_info(gpu->kgd, &local_mem_info);
1054
1055 local_mem_size = local_mem_info.local_mem_size_private +
1056 local_mem_info.local_mem_size_public;
1057
1058 buf[0] = gpu->pdev->devfn;
1059 buf[1] = gpu->pdev->subsystem_vendor;
1060 buf[2] = gpu->pdev->subsystem_device;
1061 buf[3] = gpu->pdev->device;
1062 buf[4] = gpu->pdev->bus->number;
1063 buf[5] = lower_32_bits(local_mem_size);
1064 buf[6] = upper_32_bits(local_mem_size);
1065
1066 for (i = 0, hashout = 0; i < 7; i++)
1067 hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH);
1068
1069 return hashout;
1070 }
1071 /* kfd_assign_gpu - Attach @gpu to the correct kfd topology device. If
1072 * the GPU device is not already present in the topology device
1073 * list then return NULL. This means a new topology device has to
1074 * be created for this GPU.
1075 * TODO: Rather than assiging @gpu to first topology device withtout
1076 * gpu attached, it will better to have more stringent check.
1077 */
kfd_assign_gpu(struct kfd_dev * gpu)1078 static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu)
1079 {
1080 struct kfd_topology_device *dev;
1081 struct kfd_topology_device *out_dev = NULL;
1082
1083 down_write(&topology_lock);
1084 list_for_each_entry(dev, &topology_device_list, list)
1085 if (!dev->gpu && (dev->node_props.simd_count > 0)) {
1086 dev->gpu = gpu;
1087 out_dev = dev;
1088 break;
1089 }
1090 up_write(&topology_lock);
1091 return out_dev;
1092 }
1093
kfd_notify_gpu_change(uint32_t gpu_id,int arrival)1094 static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival)
1095 {
1096 /*
1097 * TODO: Generate an event for thunk about the arrival/removal
1098 * of the GPU
1099 */
1100 }
1101
1102 /* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info,
1103 * patch this after CRAT parsing.
1104 */
kfd_fill_mem_clk_max_info(struct kfd_topology_device * dev)1105 static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev)
1106 {
1107 struct kfd_mem_properties *mem;
1108 struct kfd_local_mem_info local_mem_info;
1109
1110 if (!dev)
1111 return;
1112
1113 /* Currently, amdgpu driver (amdgpu_mc) deals only with GPUs with
1114 * single bank of VRAM local memory.
1115 * for dGPUs - VCRAT reports only one bank of Local Memory
1116 * for APUs - If CRAT from ACPI reports more than one bank, then
1117 * all the banks will report the same mem_clk_max information
1118 */
1119 dev->gpu->kfd2kgd->get_local_mem_info(dev->gpu->kgd,
1120 &local_mem_info);
1121
1122 list_for_each_entry(mem, &dev->mem_props, list)
1123 mem->mem_clk_max = local_mem_info.mem_clk_max;
1124 }
1125
kfd_fill_iolink_non_crat_info(struct kfd_topology_device * dev)1126 static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev)
1127 {
1128 struct kfd_iolink_properties *link;
1129
1130 if (!dev || !dev->gpu)
1131 return;
1132
1133 /* GPU only creates direck links so apply flags setting to all */
1134 if (dev->gpu->device_info->asic_family == CHIP_HAWAII)
1135 list_for_each_entry(link, &dev->io_link_props, list)
1136 link->flags = CRAT_IOLINK_FLAGS_ENABLED |
1137 CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1138 CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1139 }
1140
kfd_topology_add_device(struct kfd_dev * gpu)1141 int kfd_topology_add_device(struct kfd_dev *gpu)
1142 {
1143 uint32_t gpu_id;
1144 struct kfd_topology_device *dev;
1145 struct kfd_cu_info cu_info;
1146 int res = 0;
1147 struct list_head temp_topology_device_list;
1148 void *crat_image = NULL;
1149 size_t image_size = 0;
1150 int proximity_domain;
1151
1152 INIT_LIST_HEAD(&temp_topology_device_list);
1153
1154 gpu_id = kfd_generate_gpu_id(gpu);
1155
1156 pr_debug("Adding new GPU (ID: 0x%x) to topology\n", gpu_id);
1157
1158 proximity_domain = atomic_inc_return(&topology_crat_proximity_domain);
1159
1160 /* Check to see if this gpu device exists in the topology_device_list.
1161 * If so, assign the gpu to that device,
1162 * else create a Virtual CRAT for this gpu device and then parse that
1163 * CRAT to create a new topology device. Once created assign the gpu to
1164 * that topology device
1165 */
1166 dev = kfd_assign_gpu(gpu);
1167 if (!dev) {
1168 res = kfd_create_crat_image_virtual(&crat_image, &image_size,
1169 COMPUTE_UNIT_GPU, gpu,
1170 proximity_domain);
1171 if (res) {
1172 pr_err("Error creating VCRAT for GPU (ID: 0x%x)\n",
1173 gpu_id);
1174 return res;
1175 }
1176 res = kfd_parse_crat_table(crat_image,
1177 &temp_topology_device_list,
1178 proximity_domain);
1179 if (res) {
1180 pr_err("Error parsing VCRAT for GPU (ID: 0x%x)\n",
1181 gpu_id);
1182 goto err;
1183 }
1184
1185 down_write(&topology_lock);
1186 kfd_topology_update_device_list(&temp_topology_device_list,
1187 &topology_device_list);
1188
1189 /* Update the SYSFS tree, since we added another topology
1190 * device
1191 */
1192 res = kfd_topology_update_sysfs();
1193 up_write(&topology_lock);
1194
1195 if (!res)
1196 sys_props.generation_count++;
1197 else
1198 pr_err("Failed to update GPU (ID: 0x%x) to sysfs topology. res=%d\n",
1199 gpu_id, res);
1200 dev = kfd_assign_gpu(gpu);
1201 if (WARN_ON(!dev)) {
1202 res = -ENODEV;
1203 goto err;
1204 }
1205 }
1206
1207 dev->gpu_id = gpu_id;
1208 gpu->id = gpu_id;
1209
1210 /* TODO: Move the following lines to function
1211 * kfd_add_non_crat_information
1212 */
1213
1214 /* Fill-in additional information that is not available in CRAT but
1215 * needed for the topology
1216 */
1217
1218 dev->gpu->kfd2kgd->get_cu_info(dev->gpu->kgd, &cu_info);
1219 dev->node_props.simd_arrays_per_engine =
1220 cu_info.num_shader_arrays_per_engine;
1221
1222 dev->node_props.vendor_id = gpu->pdev->vendor;
1223 dev->node_props.device_id = gpu->pdev->device;
1224 dev->node_props.location_id = PCI_DEVID(gpu->pdev->bus->number,
1225 gpu->pdev->devfn);
1226 dev->node_props.max_engine_clk_fcompute =
1227 dev->gpu->kfd2kgd->get_max_engine_clock_in_mhz(dev->gpu->kgd);
1228 dev->node_props.max_engine_clk_ccompute =
1229 cpufreq_quick_get_max(0) / 1000;
1230 dev->node_props.drm_render_minor =
1231 gpu->shared_resources.drm_render_minor;
1232
1233 kfd_fill_mem_clk_max_info(dev);
1234 kfd_fill_iolink_non_crat_info(dev);
1235
1236 switch (dev->gpu->device_info->asic_family) {
1237 case CHIP_KAVERI:
1238 case CHIP_HAWAII:
1239 case CHIP_TONGA:
1240 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_PRE_1_0 <<
1241 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1242 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1243 break;
1244 case CHIP_CARRIZO:
1245 case CHIP_FIJI:
1246 case CHIP_POLARIS10:
1247 case CHIP_POLARIS11:
1248 pr_debug("Adding doorbell packet type capability\n");
1249 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 <<
1250 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1251 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1252 break;
1253 case CHIP_VEGA10:
1254 case CHIP_RAVEN:
1255 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 <<
1256 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1257 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1258 break;
1259 default:
1260 WARN(1, "Unexpected ASIC family %u",
1261 dev->gpu->device_info->asic_family);
1262 }
1263
1264 /* Fix errors in CZ CRAT.
1265 * simd_count: Carrizo CRAT reports wrong simd_count, probably
1266 * because it doesn't consider masked out CUs
1267 * max_waves_per_simd: Carrizo reports wrong max_waves_per_simd
1268 * capability flag: Carrizo CRAT doesn't report IOMMU flags
1269 */
1270 if (dev->gpu->device_info->asic_family == CHIP_CARRIZO) {
1271 dev->node_props.simd_count =
1272 cu_info.simd_per_cu * cu_info.cu_active_number;
1273 dev->node_props.max_waves_per_simd = 10;
1274 dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
1275 }
1276
1277 kfd_debug_print_topology();
1278
1279 if (!res)
1280 kfd_notify_gpu_change(gpu_id, 1);
1281 err:
1282 kfd_destroy_crat_image(crat_image);
1283 return res;
1284 }
1285
kfd_topology_remove_device(struct kfd_dev * gpu)1286 int kfd_topology_remove_device(struct kfd_dev *gpu)
1287 {
1288 struct kfd_topology_device *dev, *tmp;
1289 uint32_t gpu_id;
1290 int res = -ENODEV;
1291
1292 down_write(&topology_lock);
1293
1294 list_for_each_entry_safe(dev, tmp, &topology_device_list, list)
1295 if (dev->gpu == gpu) {
1296 gpu_id = dev->gpu_id;
1297 kfd_remove_sysfs_node_entry(dev);
1298 kfd_release_topology_device(dev);
1299 sys_props.num_devices--;
1300 res = 0;
1301 if (kfd_topology_update_sysfs() < 0)
1302 kfd_topology_release_sysfs();
1303 break;
1304 }
1305
1306 up_write(&topology_lock);
1307
1308 if (!res)
1309 kfd_notify_gpu_change(gpu_id, 0);
1310
1311 return res;
1312 }
1313
1314 /* kfd_topology_enum_kfd_devices - Enumerate through all devices in KFD
1315 * topology. If GPU device is found @idx, then valid kfd_dev pointer is
1316 * returned through @kdev
1317 * Return - 0: On success (@kdev will be NULL for non GPU nodes)
1318 * -1: If end of list
1319 */
kfd_topology_enum_kfd_devices(uint8_t idx,struct kfd_dev ** kdev)1320 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev)
1321 {
1322
1323 struct kfd_topology_device *top_dev;
1324 uint8_t device_idx = 0;
1325
1326 *kdev = NULL;
1327 down_read(&topology_lock);
1328
1329 list_for_each_entry(top_dev, &topology_device_list, list) {
1330 if (device_idx == idx) {
1331 *kdev = top_dev->gpu;
1332 up_read(&topology_lock);
1333 return 0;
1334 }
1335
1336 device_idx++;
1337 }
1338
1339 up_read(&topology_lock);
1340
1341 return -1;
1342
1343 }
1344
kfd_cpumask_to_apic_id(const struct cpumask * cpumask)1345 static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask)
1346 {
1347 const struct cpuinfo_x86 *cpuinfo;
1348 int first_cpu_of_numa_node;
1349
1350 if (!cpumask || cpumask == cpu_none_mask)
1351 return -1;
1352 first_cpu_of_numa_node = cpumask_first(cpumask);
1353 if (first_cpu_of_numa_node >= nr_cpu_ids)
1354 return -1;
1355 cpuinfo = &cpu_data(first_cpu_of_numa_node);
1356
1357 return cpuinfo->apicid;
1358 }
1359
1360 /* kfd_numa_node_to_apic_id - Returns the APIC ID of the first logical processor
1361 * of the given NUMA node (numa_node_id)
1362 * Return -1 on failure
1363 */
kfd_numa_node_to_apic_id(int numa_node_id)1364 int kfd_numa_node_to_apic_id(int numa_node_id)
1365 {
1366 if (numa_node_id == -1) {
1367 pr_warn("Invalid NUMA Node. Use online CPU mask\n");
1368 return kfd_cpumask_to_apic_id(cpu_online_mask);
1369 }
1370 return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id));
1371 }
1372
1373 #if defined(CONFIG_DEBUG_FS)
1374
kfd_debugfs_hqds_by_device(struct seq_file * m,void * data)1375 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data)
1376 {
1377 struct kfd_topology_device *dev;
1378 unsigned int i = 0;
1379 int r = 0;
1380
1381 down_read(&topology_lock);
1382
1383 list_for_each_entry(dev, &topology_device_list, list) {
1384 if (!dev->gpu) {
1385 i++;
1386 continue;
1387 }
1388
1389 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1390 r = dqm_debugfs_hqds(m, dev->gpu->dqm);
1391 if (r)
1392 break;
1393 }
1394
1395 up_read(&topology_lock);
1396
1397 return r;
1398 }
1399
kfd_debugfs_rls_by_device(struct seq_file * m,void * data)1400 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data)
1401 {
1402 struct kfd_topology_device *dev;
1403 unsigned int i = 0;
1404 int r = 0;
1405
1406 down_read(&topology_lock);
1407
1408 list_for_each_entry(dev, &topology_device_list, list) {
1409 if (!dev->gpu) {
1410 i++;
1411 continue;
1412 }
1413
1414 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1415 r = pm_debugfs_runlist(m, &dev->gpu->dqm->packets);
1416 if (r)
1417 break;
1418 }
1419
1420 up_read(&topology_lock);
1421
1422 return r;
1423 }
1424
1425 #endif
1426
