Commit | Line | Data |
---|---|---|
5b5c4e40 EP |
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> | |
f7c826ad | 30 | #include <linux/log2.h> |
520b8fb7 FK |
31 | #include <linux/dmi.h> |
32 | #include <linux/atomic.h> | |
5b5c4e40 EP |
33 | |
34 | #include "kfd_priv.h" | |
35 | #include "kfd_crat.h" | |
36 | #include "kfd_topology.h" | |
851a645e | 37 | #include "kfd_device_queue_manager.h" |
64d1c3a4 | 38 | #include "kfd_iommu.h" |
5a75ea56 | 39 | #include "kfd_svm.h" |
5b87245f | 40 | #include "amdgpu_amdkfd.h" |
0dee45a2 | 41 | #include "amdgpu_ras.h" |
5b5c4e40 | 42 | |
4f449311 HK |
43 | /* topology_device_list - Master list of all topology devices */ |
44 | static struct list_head topology_device_list; | |
520b8fb7 | 45 | static struct kfd_system_properties sys_props; |
5b5c4e40 EP |
46 | |
47 | static DECLARE_RWSEM(topology_lock); | |
520b8fb7 | 48 | static atomic_t topology_crat_proximity_domain; |
5b5c4e40 | 49 | |
3a87177e HK |
50 | struct kfd_topology_device *kfd_topology_device_by_proximity_domain( |
51 | uint32_t proximity_domain) | |
52 | { | |
53 | struct kfd_topology_device *top_dev; | |
54 | struct kfd_topology_device *device = NULL; | |
55 | ||
56 | down_read(&topology_lock); | |
57 | ||
58 | list_for_each_entry(top_dev, &topology_device_list, list) | |
59 | if (top_dev->proximity_domain == proximity_domain) { | |
60 | device = top_dev; | |
61 | break; | |
62 | } | |
63 | ||
64 | up_read(&topology_lock); | |
65 | ||
66 | return device; | |
67 | } | |
68 | ||
44d8cc6f | 69 | struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id) |
5b5c4e40 | 70 | { |
44d8cc6f YZ |
71 | struct kfd_topology_device *top_dev = NULL; |
72 | struct kfd_topology_device *ret = NULL; | |
5b5c4e40 EP |
73 | |
74 | down_read(&topology_lock); | |
75 | ||
76 | list_for_each_entry(top_dev, &topology_device_list, list) | |
77 | if (top_dev->gpu_id == gpu_id) { | |
44d8cc6f | 78 | ret = top_dev; |
5b5c4e40 EP |
79 | break; |
80 | } | |
81 | ||
82 | up_read(&topology_lock); | |
83 | ||
44d8cc6f YZ |
84 | return ret; |
85 | } | |
86 | ||
87 | struct kfd_dev *kfd_device_by_id(uint32_t gpu_id) | |
88 | { | |
89 | struct kfd_topology_device *top_dev; | |
90 | ||
91 | top_dev = kfd_topology_device_by_id(gpu_id); | |
92 | if (!top_dev) | |
93 | return NULL; | |
94 | ||
95 | return top_dev->gpu; | |
5b5c4e40 EP |
96 | } |
97 | ||
98 | struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev) | |
99 | { | |
100 | struct kfd_topology_device *top_dev; | |
101 | struct kfd_dev *device = NULL; | |
102 | ||
103 | down_read(&topology_lock); | |
104 | ||
105 | list_for_each_entry(top_dev, &topology_device_list, list) | |
3704d56e | 106 | if (top_dev->gpu && top_dev->gpu->pdev == pdev) { |
5b5c4e40 EP |
107 | device = top_dev->gpu; |
108 | break; | |
109 | } | |
110 | ||
111 | up_read(&topology_lock); | |
112 | ||
113 | return device; | |
114 | } | |
115 | ||
574c4183 | 116 | struct kfd_dev *kfd_device_by_adev(const struct amdgpu_device *adev) |
1dde0ea9 FK |
117 | { |
118 | struct kfd_topology_device *top_dev; | |
119 | struct kfd_dev *device = NULL; | |
120 | ||
121 | down_read(&topology_lock); | |
122 | ||
123 | list_for_each_entry(top_dev, &topology_device_list, list) | |
574c4183 | 124 | if (top_dev->gpu && top_dev->gpu->adev == adev) { |
1dde0ea9 FK |
125 | device = top_dev->gpu; |
126 | break; | |
127 | } | |
128 | ||
129 | up_read(&topology_lock); | |
130 | ||
131 | return device; | |
132 | } | |
133 | ||
3a87177e | 134 | /* Called with write topology_lock acquired */ |
5b5c4e40 EP |
135 | static void kfd_release_topology_device(struct kfd_topology_device *dev) |
136 | { | |
137 | struct kfd_mem_properties *mem; | |
138 | struct kfd_cache_properties *cache; | |
139 | struct kfd_iolink_properties *iolink; | |
f4757347 | 140 | struct kfd_perf_properties *perf; |
5b5c4e40 | 141 | |
5b5c4e40 EP |
142 | list_del(&dev->list); |
143 | ||
144 | while (dev->mem_props.next != &dev->mem_props) { | |
145 | mem = container_of(dev->mem_props.next, | |
146 | struct kfd_mem_properties, list); | |
147 | list_del(&mem->list); | |
148 | kfree(mem); | |
149 | } | |
150 | ||
151 | while (dev->cache_props.next != &dev->cache_props) { | |
152 | cache = container_of(dev->cache_props.next, | |
153 | struct kfd_cache_properties, list); | |
154 | list_del(&cache->list); | |
155 | kfree(cache); | |
156 | } | |
157 | ||
158 | while (dev->io_link_props.next != &dev->io_link_props) { | |
159 | iolink = container_of(dev->io_link_props.next, | |
160 | struct kfd_iolink_properties, list); | |
161 | list_del(&iolink->list); | |
162 | kfree(iolink); | |
163 | } | |
164 | ||
f4757347 AL |
165 | while (dev->perf_props.next != &dev->perf_props) { |
166 | perf = container_of(dev->perf_props.next, | |
167 | struct kfd_perf_properties, list); | |
168 | list_del(&perf->list); | |
169 | kfree(perf); | |
170 | } | |
171 | ||
5b5c4e40 | 172 | kfree(dev); |
5b5c4e40 EP |
173 | } |
174 | ||
4f449311 | 175 | void kfd_release_topology_device_list(struct list_head *device_list) |
5b5c4e40 EP |
176 | { |
177 | struct kfd_topology_device *dev; | |
178 | ||
4f449311 HK |
179 | while (!list_empty(device_list)) { |
180 | dev = list_first_entry(device_list, | |
181 | struct kfd_topology_device, list); | |
5b5c4e40 | 182 | kfd_release_topology_device(dev); |
4f449311 | 183 | } |
5b5c4e40 EP |
184 | } |
185 | ||
4f449311 HK |
186 | static void kfd_release_live_view(void) |
187 | { | |
188 | kfd_release_topology_device_list(&topology_device_list); | |
5b5c4e40 EP |
189 | memset(&sys_props, 0, sizeof(sys_props)); |
190 | } | |
191 | ||
4f449311 HK |
192 | struct kfd_topology_device *kfd_create_topology_device( |
193 | struct list_head *device_list) | |
5b5c4e40 EP |
194 | { |
195 | struct kfd_topology_device *dev; | |
196 | ||
197 | dev = kfd_alloc_struct(dev); | |
4eacc26b | 198 | if (!dev) { |
5b5c4e40 | 199 | pr_err("No memory to allocate a topology device"); |
16b9201c | 200 | return NULL; |
5b5c4e40 EP |
201 | } |
202 | ||
203 | INIT_LIST_HEAD(&dev->mem_props); | |
204 | INIT_LIST_HEAD(&dev->cache_props); | |
205 | INIT_LIST_HEAD(&dev->io_link_props); | |
f4757347 | 206 | INIT_LIST_HEAD(&dev->perf_props); |
5b5c4e40 | 207 | |
4f449311 | 208 | list_add_tail(&dev->list, device_list); |
5b5c4e40 EP |
209 | |
210 | return dev; | |
16b9201c | 211 | } |
5b5c4e40 | 212 | |
5b5c4e40 | 213 | |
83a13ef5 FK |
214 | #define sysfs_show_gen_prop(buffer, offs, fmt, ...) \ |
215 | (offs += snprintf(buffer+offs, PAGE_SIZE-offs, \ | |
216 | fmt, __VA_ARGS__)) | |
217 | #define sysfs_show_32bit_prop(buffer, offs, name, value) \ | |
218 | sysfs_show_gen_prop(buffer, offs, "%s %u\n", name, value) | |
219 | #define sysfs_show_64bit_prop(buffer, offs, name, value) \ | |
220 | sysfs_show_gen_prop(buffer, offs, "%s %llu\n", name, value) | |
221 | #define sysfs_show_32bit_val(buffer, offs, value) \ | |
222 | sysfs_show_gen_prop(buffer, offs, "%u\n", value) | |
223 | #define sysfs_show_str_val(buffer, offs, value) \ | |
224 | sysfs_show_gen_prop(buffer, offs, "%s\n", value) | |
5b5c4e40 EP |
225 | |
226 | static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr, | |
227 | char *buffer) | |
228 | { | |
83a13ef5 | 229 | int offs = 0; |
5b5c4e40 EP |
230 | |
231 | /* Making sure that the buffer is an empty string */ | |
232 | buffer[0] = 0; | |
233 | ||
234 | if (attr == &sys_props.attr_genid) { | |
83a13ef5 FK |
235 | sysfs_show_32bit_val(buffer, offs, |
236 | sys_props.generation_count); | |
5b5c4e40 | 237 | } else if (attr == &sys_props.attr_props) { |
83a13ef5 FK |
238 | sysfs_show_64bit_prop(buffer, offs, "platform_oem", |
239 | sys_props.platform_oem); | |
240 | sysfs_show_64bit_prop(buffer, offs, "platform_id", | |
241 | sys_props.platform_id); | |
242 | sysfs_show_64bit_prop(buffer, offs, "platform_rev", | |
243 | sys_props.platform_rev); | |
5b5c4e40 | 244 | } else { |
83a13ef5 | 245 | offs = -EINVAL; |
5b5c4e40 EP |
246 | } |
247 | ||
83a13ef5 | 248 | return offs; |
5b5c4e40 EP |
249 | } |
250 | ||
5108d768 YZ |
251 | static void kfd_topology_kobj_release(struct kobject *kobj) |
252 | { | |
253 | kfree(kobj); | |
254 | } | |
255 | ||
5b5c4e40 EP |
256 | static const struct sysfs_ops sysprops_ops = { |
257 | .show = sysprops_show, | |
258 | }; | |
259 | ||
260 | static struct kobj_type sysprops_type = { | |
5108d768 | 261 | .release = kfd_topology_kobj_release, |
5b5c4e40 EP |
262 | .sysfs_ops = &sysprops_ops, |
263 | }; | |
264 | ||
265 | static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr, | |
266 | char *buffer) | |
267 | { | |
83a13ef5 | 268 | int offs = 0; |
5b5c4e40 EP |
269 | struct kfd_iolink_properties *iolink; |
270 | ||
271 | /* Making sure that the buffer is an empty string */ | |
272 | buffer[0] = 0; | |
273 | ||
274 | iolink = container_of(attr, struct kfd_iolink_properties, attr); | |
6b855f7b HK |
275 | if (iolink->gpu && kfd_devcgroup_check_permission(iolink->gpu)) |
276 | return -EPERM; | |
83a13ef5 FK |
277 | sysfs_show_32bit_prop(buffer, offs, "type", iolink->iolink_type); |
278 | sysfs_show_32bit_prop(buffer, offs, "version_major", iolink->ver_maj); | |
279 | sysfs_show_32bit_prop(buffer, offs, "version_minor", iolink->ver_min); | |
280 | sysfs_show_32bit_prop(buffer, offs, "node_from", iolink->node_from); | |
281 | sysfs_show_32bit_prop(buffer, offs, "node_to", iolink->node_to); | |
282 | sysfs_show_32bit_prop(buffer, offs, "weight", iolink->weight); | |
283 | sysfs_show_32bit_prop(buffer, offs, "min_latency", iolink->min_latency); | |
284 | sysfs_show_32bit_prop(buffer, offs, "max_latency", iolink->max_latency); | |
285 | sysfs_show_32bit_prop(buffer, offs, "min_bandwidth", | |
286 | iolink->min_bandwidth); | |
287 | sysfs_show_32bit_prop(buffer, offs, "max_bandwidth", | |
288 | iolink->max_bandwidth); | |
289 | sysfs_show_32bit_prop(buffer, offs, "recommended_transfer_size", | |
290 | iolink->rec_transfer_size); | |
291 | sysfs_show_32bit_prop(buffer, offs, "flags", iolink->flags); | |
292 | ||
293 | return offs; | |
5b5c4e40 EP |
294 | } |
295 | ||
296 | static const struct sysfs_ops iolink_ops = { | |
297 | .show = iolink_show, | |
298 | }; | |
299 | ||
300 | static struct kobj_type iolink_type = { | |
5108d768 | 301 | .release = kfd_topology_kobj_release, |
5b5c4e40 EP |
302 | .sysfs_ops = &iolink_ops, |
303 | }; | |
304 | ||
305 | static ssize_t mem_show(struct kobject *kobj, struct attribute *attr, | |
306 | char *buffer) | |
307 | { | |
83a13ef5 | 308 | int offs = 0; |
5b5c4e40 EP |
309 | struct kfd_mem_properties *mem; |
310 | ||
311 | /* Making sure that the buffer is an empty string */ | |
312 | buffer[0] = 0; | |
313 | ||
314 | mem = container_of(attr, struct kfd_mem_properties, attr); | |
6b855f7b HK |
315 | if (mem->gpu && kfd_devcgroup_check_permission(mem->gpu)) |
316 | return -EPERM; | |
83a13ef5 FK |
317 | sysfs_show_32bit_prop(buffer, offs, "heap_type", mem->heap_type); |
318 | sysfs_show_64bit_prop(buffer, offs, "size_in_bytes", | |
319 | mem->size_in_bytes); | |
320 | sysfs_show_32bit_prop(buffer, offs, "flags", mem->flags); | |
321 | sysfs_show_32bit_prop(buffer, offs, "width", mem->width); | |
322 | sysfs_show_32bit_prop(buffer, offs, "mem_clk_max", | |
323 | mem->mem_clk_max); | |
324 | ||
325 | return offs; | |
5b5c4e40 EP |
326 | } |
327 | ||
328 | static const struct sysfs_ops mem_ops = { | |
329 | .show = mem_show, | |
330 | }; | |
331 | ||
332 | static struct kobj_type mem_type = { | |
5108d768 | 333 | .release = kfd_topology_kobj_release, |
5b5c4e40 EP |
334 | .sysfs_ops = &mem_ops, |
335 | }; | |
336 | ||
337 | static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr, | |
338 | char *buffer) | |
339 | { | |
83a13ef5 | 340 | int offs = 0; |
bc0c75a3 | 341 | uint32_t i, j; |
5b5c4e40 EP |
342 | struct kfd_cache_properties *cache; |
343 | ||
344 | /* Making sure that the buffer is an empty string */ | |
345 | buffer[0] = 0; | |
346 | ||
347 | cache = container_of(attr, struct kfd_cache_properties, attr); | |
6b855f7b HK |
348 | if (cache->gpu && kfd_devcgroup_check_permission(cache->gpu)) |
349 | return -EPERM; | |
83a13ef5 | 350 | sysfs_show_32bit_prop(buffer, offs, "processor_id_low", |
5b5c4e40 | 351 | cache->processor_id_low); |
83a13ef5 FK |
352 | sysfs_show_32bit_prop(buffer, offs, "level", cache->cache_level); |
353 | sysfs_show_32bit_prop(buffer, offs, "size", cache->cache_size); | |
354 | sysfs_show_32bit_prop(buffer, offs, "cache_line_size", | |
355 | cache->cacheline_size); | |
356 | sysfs_show_32bit_prop(buffer, offs, "cache_lines_per_tag", | |
357 | cache->cachelines_per_tag); | |
358 | sysfs_show_32bit_prop(buffer, offs, "association", cache->cache_assoc); | |
359 | sysfs_show_32bit_prop(buffer, offs, "latency", cache->cache_latency); | |
360 | sysfs_show_32bit_prop(buffer, offs, "type", cache->cache_type); | |
361 | offs += snprintf(buffer+offs, PAGE_SIZE-offs, "sibling_map "); | |
bc0c75a3 | 362 | for (i = 0; i < CRAT_SIBLINGMAP_SIZE; i++) |
83a13ef5 | 363 | for (j = 0; j < sizeof(cache->sibling_map[0])*8; j++) |
bc0c75a3 | 364 | /* Check each bit */ |
83a13ef5 FK |
365 | offs += snprintf(buffer+offs, PAGE_SIZE-offs, "%d,", |
366 | (cache->sibling_map[i] >> j) & 1); | |
367 | ||
bc0c75a3 | 368 | /* Replace the last "," with end of line */ |
83a13ef5 FK |
369 | buffer[offs-1] = '\n'; |
370 | return offs; | |
5b5c4e40 EP |
371 | } |
372 | ||
373 | static const struct sysfs_ops cache_ops = { | |
374 | .show = kfd_cache_show, | |
375 | }; | |
376 | ||
377 | static struct kobj_type cache_type = { | |
5108d768 | 378 | .release = kfd_topology_kobj_release, |
5b5c4e40 EP |
379 | .sysfs_ops = &cache_ops, |
380 | }; | |
381 | ||
f4757347 AL |
382 | /****** Sysfs of Performance Counters ******/ |
383 | ||
384 | struct kfd_perf_attr { | |
385 | struct kobj_attribute attr; | |
386 | uint32_t data; | |
387 | }; | |
388 | ||
389 | static ssize_t perf_show(struct kobject *kobj, struct kobj_attribute *attrs, | |
390 | char *buf) | |
391 | { | |
83a13ef5 | 392 | int offs = 0; |
f4757347 AL |
393 | struct kfd_perf_attr *attr; |
394 | ||
395 | buf[0] = 0; | |
396 | attr = container_of(attrs, struct kfd_perf_attr, attr); | |
397 | if (!attr->data) /* invalid data for PMC */ | |
398 | return 0; | |
399 | else | |
83a13ef5 | 400 | return sysfs_show_32bit_val(buf, offs, attr->data); |
f4757347 AL |
401 | } |
402 | ||
403 | #define KFD_PERF_DESC(_name, _data) \ | |
404 | { \ | |
405 | .attr = __ATTR(_name, 0444, perf_show, NULL), \ | |
406 | .data = _data, \ | |
407 | } | |
408 | ||
409 | static struct kfd_perf_attr perf_attr_iommu[] = { | |
410 | KFD_PERF_DESC(max_concurrent, 0), | |
411 | KFD_PERF_DESC(num_counters, 0), | |
412 | KFD_PERF_DESC(counter_ids, 0), | |
413 | }; | |
414 | /****************************************/ | |
415 | ||
5b5c4e40 EP |
416 | static ssize_t node_show(struct kobject *kobj, struct attribute *attr, |
417 | char *buffer) | |
418 | { | |
83a13ef5 | 419 | int offs = 0; |
5b5c4e40 | 420 | struct kfd_topology_device *dev; |
f7c826ad | 421 | uint32_t log_max_watch_addr; |
5b5c4e40 EP |
422 | |
423 | /* Making sure that the buffer is an empty string */ | |
424 | buffer[0] = 0; | |
425 | ||
426 | if (strcmp(attr->name, "gpu_id") == 0) { | |
427 | dev = container_of(attr, struct kfd_topology_device, | |
428 | attr_gpuid); | |
6b855f7b HK |
429 | if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu)) |
430 | return -EPERM; | |
83a13ef5 | 431 | return sysfs_show_32bit_val(buffer, offs, dev->gpu_id); |
f7c826ad AS |
432 | } |
433 | ||
434 | if (strcmp(attr->name, "name") == 0) { | |
5b5c4e40 EP |
435 | dev = container_of(attr, struct kfd_topology_device, |
436 | attr_name); | |
c181159a | 437 | |
6b855f7b HK |
438 | if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu)) |
439 | return -EPERM; | |
83a13ef5 | 440 | return sysfs_show_str_val(buffer, offs, dev->node_props.name); |
f7c826ad | 441 | } |
5b5c4e40 | 442 | |
f7c826ad AS |
443 | dev = container_of(attr, struct kfd_topology_device, |
444 | attr_props); | |
6b855f7b HK |
445 | if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu)) |
446 | return -EPERM; | |
83a13ef5 FK |
447 | sysfs_show_32bit_prop(buffer, offs, "cpu_cores_count", |
448 | dev->node_props.cpu_cores_count); | |
449 | sysfs_show_32bit_prop(buffer, offs, "simd_count", | |
6127896f | 450 | dev->gpu ? dev->node_props.simd_count : 0); |
83a13ef5 FK |
451 | sysfs_show_32bit_prop(buffer, offs, "mem_banks_count", |
452 | dev->node_props.mem_banks_count); | |
453 | sysfs_show_32bit_prop(buffer, offs, "caches_count", | |
454 | dev->node_props.caches_count); | |
455 | sysfs_show_32bit_prop(buffer, offs, "io_links_count", | |
456 | dev->node_props.io_links_count); | |
457 | sysfs_show_32bit_prop(buffer, offs, "cpu_core_id_base", | |
458 | dev->node_props.cpu_core_id_base); | |
459 | sysfs_show_32bit_prop(buffer, offs, "simd_id_base", | |
460 | dev->node_props.simd_id_base); | |
461 | sysfs_show_32bit_prop(buffer, offs, "max_waves_per_simd", | |
462 | dev->node_props.max_waves_per_simd); | |
463 | sysfs_show_32bit_prop(buffer, offs, "lds_size_in_kb", | |
464 | dev->node_props.lds_size_in_kb); | |
465 | sysfs_show_32bit_prop(buffer, offs, "gds_size_in_kb", | |
466 | dev->node_props.gds_size_in_kb); | |
467 | sysfs_show_32bit_prop(buffer, offs, "num_gws", | |
468 | dev->node_props.num_gws); | |
469 | sysfs_show_32bit_prop(buffer, offs, "wave_front_size", | |
470 | dev->node_props.wave_front_size); | |
471 | sysfs_show_32bit_prop(buffer, offs, "array_count", | |
472 | dev->node_props.array_count); | |
473 | sysfs_show_32bit_prop(buffer, offs, "simd_arrays_per_engine", | |
474 | dev->node_props.simd_arrays_per_engine); | |
475 | sysfs_show_32bit_prop(buffer, offs, "cu_per_simd_array", | |
476 | dev->node_props.cu_per_simd_array); | |
477 | sysfs_show_32bit_prop(buffer, offs, "simd_per_cu", | |
478 | dev->node_props.simd_per_cu); | |
479 | sysfs_show_32bit_prop(buffer, offs, "max_slots_scratch_cu", | |
480 | dev->node_props.max_slots_scratch_cu); | |
9d6fa9c7 GS |
481 | sysfs_show_32bit_prop(buffer, offs, "gfx_target_version", |
482 | dev->node_props.gfx_target_version); | |
83a13ef5 FK |
483 | sysfs_show_32bit_prop(buffer, offs, "vendor_id", |
484 | dev->node_props.vendor_id); | |
485 | sysfs_show_32bit_prop(buffer, offs, "device_id", | |
486 | dev->node_props.device_id); | |
487 | sysfs_show_32bit_prop(buffer, offs, "location_id", | |
488 | dev->node_props.location_id); | |
489 | sysfs_show_32bit_prop(buffer, offs, "domain", | |
490 | dev->node_props.domain); | |
491 | sysfs_show_32bit_prop(buffer, offs, "drm_render_minor", | |
492 | dev->node_props.drm_render_minor); | |
493 | sysfs_show_64bit_prop(buffer, offs, "hive_id", | |
494 | dev->node_props.hive_id); | |
495 | sysfs_show_32bit_prop(buffer, offs, "num_sdma_engines", | |
496 | dev->node_props.num_sdma_engines); | |
497 | sysfs_show_32bit_prop(buffer, offs, "num_sdma_xgmi_engines", | |
498 | dev->node_props.num_sdma_xgmi_engines); | |
499 | sysfs_show_32bit_prop(buffer, offs, "num_sdma_queues_per_engine", | |
500 | dev->node_props.num_sdma_queues_per_engine); | |
501 | sysfs_show_32bit_prop(buffer, offs, "num_cp_queues", | |
502 | dev->node_props.num_cp_queues); | |
f7c826ad AS |
503 | |
504 | if (dev->gpu) { | |
505 | log_max_watch_addr = | |
506 | __ilog2_u32(dev->gpu->device_info->num_of_watch_points); | |
507 | ||
508 | if (log_max_watch_addr) { | |
509 | dev->node_props.capability |= | |
510 | HSA_CAP_WATCH_POINTS_SUPPORTED; | |
511 | ||
512 | dev->node_props.capability |= | |
513 | ((log_max_watch_addr << | |
514 | HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) & | |
515 | HSA_CAP_WATCH_POINTS_TOTALBITS_MASK); | |
5b5c4e40 EP |
516 | } |
517 | ||
7eb0502a | 518 | if (dev->gpu->adev->asic_type == CHIP_TONGA) |
413e85d5 BG |
519 | dev->node_props.capability |= |
520 | HSA_CAP_AQL_QUEUE_DOUBLE_MAP; | |
521 | ||
83a13ef5 | 522 | sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_fcompute", |
3a87177e | 523 | dev->node_props.max_engine_clk_fcompute); |
42e08c78 | 524 | |
83a13ef5 | 525 | sysfs_show_64bit_prop(buffer, offs, "local_mem_size", 0ULL); |
f7c826ad | 526 | |
83a13ef5 FK |
527 | sysfs_show_32bit_prop(buffer, offs, "fw_version", |
528 | dev->gpu->mec_fw_version); | |
529 | sysfs_show_32bit_prop(buffer, offs, "capability", | |
530 | dev->node_props.capability); | |
531 | sysfs_show_32bit_prop(buffer, offs, "sdma_fw_version", | |
532 | dev->gpu->sdma_fw_version); | |
11964258 | 533 | sysfs_show_64bit_prop(buffer, offs, "unique_id", |
02274fc0 | 534 | dev->gpu->adev->unique_id); |
11964258 | 535 | |
5b5c4e40 EP |
536 | } |
537 | ||
83a13ef5 FK |
538 | return sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_ccompute", |
539 | cpufreq_quick_get_max(0)/1000); | |
5b5c4e40 EP |
540 | } |
541 | ||
542 | static const struct sysfs_ops node_ops = { | |
543 | .show = node_show, | |
544 | }; | |
545 | ||
546 | static struct kobj_type node_type = { | |
5108d768 | 547 | .release = kfd_topology_kobj_release, |
5b5c4e40 EP |
548 | .sysfs_ops = &node_ops, |
549 | }; | |
550 | ||
551 | static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr) | |
552 | { | |
553 | sysfs_remove_file(kobj, attr); | |
554 | kobject_del(kobj); | |
555 | kobject_put(kobj); | |
556 | } | |
557 | ||
558 | static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev) | |
559 | { | |
560 | struct kfd_iolink_properties *iolink; | |
561 | struct kfd_cache_properties *cache; | |
562 | struct kfd_mem_properties *mem; | |
f4757347 | 563 | struct kfd_perf_properties *perf; |
5b5c4e40 | 564 | |
5b5c4e40 EP |
565 | if (dev->kobj_iolink) { |
566 | list_for_each_entry(iolink, &dev->io_link_props, list) | |
567 | if (iolink->kobj) { | |
568 | kfd_remove_sysfs_file(iolink->kobj, | |
569 | &iolink->attr); | |
16b9201c | 570 | iolink->kobj = NULL; |
5b5c4e40 EP |
571 | } |
572 | kobject_del(dev->kobj_iolink); | |
573 | kobject_put(dev->kobj_iolink); | |
16b9201c | 574 | dev->kobj_iolink = NULL; |
5b5c4e40 EP |
575 | } |
576 | ||
577 | if (dev->kobj_cache) { | |
578 | list_for_each_entry(cache, &dev->cache_props, list) | |
579 | if (cache->kobj) { | |
580 | kfd_remove_sysfs_file(cache->kobj, | |
581 | &cache->attr); | |
16b9201c | 582 | cache->kobj = NULL; |
5b5c4e40 EP |
583 | } |
584 | kobject_del(dev->kobj_cache); | |
585 | kobject_put(dev->kobj_cache); | |
16b9201c | 586 | dev->kobj_cache = NULL; |
5b5c4e40 EP |
587 | } |
588 | ||
589 | if (dev->kobj_mem) { | |
590 | list_for_each_entry(mem, &dev->mem_props, list) | |
591 | if (mem->kobj) { | |
592 | kfd_remove_sysfs_file(mem->kobj, &mem->attr); | |
16b9201c | 593 | mem->kobj = NULL; |
5b5c4e40 EP |
594 | } |
595 | kobject_del(dev->kobj_mem); | |
596 | kobject_put(dev->kobj_mem); | |
16b9201c | 597 | dev->kobj_mem = NULL; |
5b5c4e40 EP |
598 | } |
599 | ||
f4757347 AL |
600 | if (dev->kobj_perf) { |
601 | list_for_each_entry(perf, &dev->perf_props, list) { | |
602 | kfree(perf->attr_group); | |
603 | perf->attr_group = NULL; | |
604 | } | |
605 | kobject_del(dev->kobj_perf); | |
606 | kobject_put(dev->kobj_perf); | |
607 | dev->kobj_perf = NULL; | |
608 | } | |
609 | ||
5b5c4e40 EP |
610 | if (dev->kobj_node) { |
611 | sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid); | |
612 | sysfs_remove_file(dev->kobj_node, &dev->attr_name); | |
613 | sysfs_remove_file(dev->kobj_node, &dev->attr_props); | |
614 | kobject_del(dev->kobj_node); | |
615 | kobject_put(dev->kobj_node); | |
16b9201c | 616 | dev->kobj_node = NULL; |
5b5c4e40 EP |
617 | } |
618 | } | |
619 | ||
620 | static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev, | |
621 | uint32_t id) | |
622 | { | |
623 | struct kfd_iolink_properties *iolink; | |
624 | struct kfd_cache_properties *cache; | |
625 | struct kfd_mem_properties *mem; | |
f4757347 | 626 | struct kfd_perf_properties *perf; |
5b5c4e40 | 627 | int ret; |
f4757347 AL |
628 | uint32_t i, num_attrs; |
629 | struct attribute **attrs; | |
5b5c4e40 | 630 | |
32fa8219 FK |
631 | if (WARN_ON(dev->kobj_node)) |
632 | return -EEXIST; | |
633 | ||
5b5c4e40 EP |
634 | /* |
635 | * Creating the sysfs folders | |
636 | */ | |
5b5c4e40 EP |
637 | dev->kobj_node = kfd_alloc_struct(dev->kobj_node); |
638 | if (!dev->kobj_node) | |
639 | return -ENOMEM; | |
640 | ||
641 | ret = kobject_init_and_add(dev->kobj_node, &node_type, | |
642 | sys_props.kobj_nodes, "%d", id); | |
20eca012 QW |
643 | if (ret < 0) { |
644 | kobject_put(dev->kobj_node); | |
5b5c4e40 | 645 | return ret; |
20eca012 | 646 | } |
5b5c4e40 EP |
647 | |
648 | dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node); | |
649 | if (!dev->kobj_mem) | |
650 | return -ENOMEM; | |
651 | ||
652 | dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node); | |
653 | if (!dev->kobj_cache) | |
654 | return -ENOMEM; | |
655 | ||
656 | dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node); | |
657 | if (!dev->kobj_iolink) | |
658 | return -ENOMEM; | |
659 | ||
f4757347 AL |
660 | dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node); |
661 | if (!dev->kobj_perf) | |
662 | return -ENOMEM; | |
663 | ||
5b5c4e40 EP |
664 | /* |
665 | * Creating sysfs files for node properties | |
666 | */ | |
667 | dev->attr_gpuid.name = "gpu_id"; | |
668 | dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE; | |
669 | sysfs_attr_init(&dev->attr_gpuid); | |
670 | dev->attr_name.name = "name"; | |
671 | dev->attr_name.mode = KFD_SYSFS_FILE_MODE; | |
672 | sysfs_attr_init(&dev->attr_name); | |
673 | dev->attr_props.name = "properties"; | |
674 | dev->attr_props.mode = KFD_SYSFS_FILE_MODE; | |
675 | sysfs_attr_init(&dev->attr_props); | |
676 | ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid); | |
677 | if (ret < 0) | |
678 | return ret; | |
679 | ret = sysfs_create_file(dev->kobj_node, &dev->attr_name); | |
680 | if (ret < 0) | |
681 | return ret; | |
682 | ret = sysfs_create_file(dev->kobj_node, &dev->attr_props); | |
683 | if (ret < 0) | |
684 | return ret; | |
685 | ||
686 | i = 0; | |
687 | list_for_each_entry(mem, &dev->mem_props, list) { | |
688 | mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); | |
689 | if (!mem->kobj) | |
690 | return -ENOMEM; | |
691 | ret = kobject_init_and_add(mem->kobj, &mem_type, | |
692 | dev->kobj_mem, "%d", i); | |
20eca012 QW |
693 | if (ret < 0) { |
694 | kobject_put(mem->kobj); | |
5b5c4e40 | 695 | return ret; |
20eca012 | 696 | } |
5b5c4e40 EP |
697 | |
698 | mem->attr.name = "properties"; | |
699 | mem->attr.mode = KFD_SYSFS_FILE_MODE; | |
700 | sysfs_attr_init(&mem->attr); | |
701 | ret = sysfs_create_file(mem->kobj, &mem->attr); | |
702 | if (ret < 0) | |
703 | return ret; | |
704 | i++; | |
705 | } | |
706 | ||
707 | i = 0; | |
708 | list_for_each_entry(cache, &dev->cache_props, list) { | |
709 | cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); | |
710 | if (!cache->kobj) | |
711 | return -ENOMEM; | |
712 | ret = kobject_init_and_add(cache->kobj, &cache_type, | |
713 | dev->kobj_cache, "%d", i); | |
20eca012 QW |
714 | if (ret < 0) { |
715 | kobject_put(cache->kobj); | |
5b5c4e40 | 716 | return ret; |
20eca012 | 717 | } |
5b5c4e40 EP |
718 | |
719 | cache->attr.name = "properties"; | |
720 | cache->attr.mode = KFD_SYSFS_FILE_MODE; | |
721 | sysfs_attr_init(&cache->attr); | |
722 | ret = sysfs_create_file(cache->kobj, &cache->attr); | |
723 | if (ret < 0) | |
724 | return ret; | |
725 | i++; | |
726 | } | |
727 | ||
728 | i = 0; | |
729 | list_for_each_entry(iolink, &dev->io_link_props, list) { | |
730 | iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); | |
731 | if (!iolink->kobj) | |
732 | return -ENOMEM; | |
733 | ret = kobject_init_and_add(iolink->kobj, &iolink_type, | |
734 | dev->kobj_iolink, "%d", i); | |
20eca012 QW |
735 | if (ret < 0) { |
736 | kobject_put(iolink->kobj); | |
5b5c4e40 | 737 | return ret; |
20eca012 | 738 | } |
5b5c4e40 EP |
739 | |
740 | iolink->attr.name = "properties"; | |
741 | iolink->attr.mode = KFD_SYSFS_FILE_MODE; | |
742 | sysfs_attr_init(&iolink->attr); | |
743 | ret = sysfs_create_file(iolink->kobj, &iolink->attr); | |
744 | if (ret < 0) | |
745 | return ret; | |
746 | i++; | |
f4757347 AL |
747 | } |
748 | ||
749 | /* All hardware blocks have the same number of attributes. */ | |
3f866f5f | 750 | num_attrs = ARRAY_SIZE(perf_attr_iommu); |
f4757347 AL |
751 | list_for_each_entry(perf, &dev->perf_props, list) { |
752 | perf->attr_group = kzalloc(sizeof(struct kfd_perf_attr) | |
753 | * num_attrs + sizeof(struct attribute_group), | |
754 | GFP_KERNEL); | |
755 | if (!perf->attr_group) | |
756 | return -ENOMEM; | |
757 | ||
758 | attrs = (struct attribute **)(perf->attr_group + 1); | |
759 | if (!strcmp(perf->block_name, "iommu")) { | |
760 | /* Information of IOMMU's num_counters and counter_ids is shown | |
761 | * under /sys/bus/event_source/devices/amd_iommu. We don't | |
762 | * duplicate here. | |
763 | */ | |
764 | perf_attr_iommu[0].data = perf->max_concurrent; | |
765 | for (i = 0; i < num_attrs; i++) | |
766 | attrs[i] = &perf_attr_iommu[i].attr.attr; | |
767 | } | |
768 | perf->attr_group->name = perf->block_name; | |
769 | perf->attr_group->attrs = attrs; | |
770 | ret = sysfs_create_group(dev->kobj_perf, perf->attr_group); | |
771 | if (ret < 0) | |
772 | return ret; | |
773 | } | |
5b5c4e40 EP |
774 | |
775 | return 0; | |
776 | } | |
777 | ||
3a87177e | 778 | /* Called with write topology lock acquired */ |
5b5c4e40 EP |
779 | static int kfd_build_sysfs_node_tree(void) |
780 | { | |
781 | struct kfd_topology_device *dev; | |
782 | int ret; | |
783 | uint32_t i = 0; | |
784 | ||
785 | list_for_each_entry(dev, &topology_device_list, list) { | |
8dfead6c | 786 | ret = kfd_build_sysfs_node_entry(dev, i); |
5b5c4e40 EP |
787 | if (ret < 0) |
788 | return ret; | |
789 | i++; | |
790 | } | |
791 | ||
792 | return 0; | |
793 | } | |
794 | ||
3a87177e | 795 | /* Called with write topology lock acquired */ |
5b5c4e40 EP |
796 | static void kfd_remove_sysfs_node_tree(void) |
797 | { | |
798 | struct kfd_topology_device *dev; | |
799 | ||
800 | list_for_each_entry(dev, &topology_device_list, list) | |
801 | kfd_remove_sysfs_node_entry(dev); | |
802 | } | |
803 | ||
804 | static int kfd_topology_update_sysfs(void) | |
805 | { | |
806 | int ret; | |
807 | ||
4eacc26b | 808 | if (!sys_props.kobj_topology) { |
5b5c4e40 EP |
809 | sys_props.kobj_topology = |
810 | kfd_alloc_struct(sys_props.kobj_topology); | |
811 | if (!sys_props.kobj_topology) | |
812 | return -ENOMEM; | |
813 | ||
814 | ret = kobject_init_and_add(sys_props.kobj_topology, | |
815 | &sysprops_type, &kfd_device->kobj, | |
816 | "topology"); | |
20eca012 QW |
817 | if (ret < 0) { |
818 | kobject_put(sys_props.kobj_topology); | |
5b5c4e40 | 819 | return ret; |
20eca012 | 820 | } |
5b5c4e40 EP |
821 | |
822 | sys_props.kobj_nodes = kobject_create_and_add("nodes", | |
823 | sys_props.kobj_topology); | |
824 | if (!sys_props.kobj_nodes) | |
825 | return -ENOMEM; | |
826 | ||
827 | sys_props.attr_genid.name = "generation_id"; | |
828 | sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE; | |
829 | sysfs_attr_init(&sys_props.attr_genid); | |
830 | ret = sysfs_create_file(sys_props.kobj_topology, | |
831 | &sys_props.attr_genid); | |
832 | if (ret < 0) | |
833 | return ret; | |
834 | ||
835 | sys_props.attr_props.name = "system_properties"; | |
836 | sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE; | |
837 | sysfs_attr_init(&sys_props.attr_props); | |
838 | ret = sysfs_create_file(sys_props.kobj_topology, | |
839 | &sys_props.attr_props); | |
840 | if (ret < 0) | |
841 | return ret; | |
842 | } | |
843 | ||
844 | kfd_remove_sysfs_node_tree(); | |
845 | ||
846 | return kfd_build_sysfs_node_tree(); | |
847 | } | |
848 | ||
849 | static void kfd_topology_release_sysfs(void) | |
850 | { | |
851 | kfd_remove_sysfs_node_tree(); | |
852 | if (sys_props.kobj_topology) { | |
853 | sysfs_remove_file(sys_props.kobj_topology, | |
854 | &sys_props.attr_genid); | |
855 | sysfs_remove_file(sys_props.kobj_topology, | |
856 | &sys_props.attr_props); | |
857 | if (sys_props.kobj_nodes) { | |
858 | kobject_del(sys_props.kobj_nodes); | |
859 | kobject_put(sys_props.kobj_nodes); | |
16b9201c | 860 | sys_props.kobj_nodes = NULL; |
5b5c4e40 EP |
861 | } |
862 | kobject_del(sys_props.kobj_topology); | |
863 | kobject_put(sys_props.kobj_topology); | |
16b9201c | 864 | sys_props.kobj_topology = NULL; |
5b5c4e40 EP |
865 | } |
866 | } | |
867 | ||
4f449311 HK |
868 | /* Called with write topology_lock acquired */ |
869 | static void kfd_topology_update_device_list(struct list_head *temp_list, | |
870 | struct list_head *master_list) | |
871 | { | |
872 | while (!list_empty(temp_list)) { | |
873 | list_move_tail(temp_list->next, master_list); | |
874 | sys_props.num_devices++; | |
875 | } | |
876 | } | |
877 | ||
520b8fb7 FK |
878 | static void kfd_debug_print_topology(void) |
879 | { | |
880 | struct kfd_topology_device *dev; | |
881 | ||
882 | down_read(&topology_lock); | |
883 | ||
884 | dev = list_last_entry(&topology_device_list, | |
885 | struct kfd_topology_device, list); | |
886 | if (dev) { | |
887 | if (dev->node_props.cpu_cores_count && | |
888 | dev->node_props.simd_count) { | |
889 | pr_info("Topology: Add APU node [0x%0x:0x%0x]\n", | |
890 | dev->node_props.device_id, | |
891 | dev->node_props.vendor_id); | |
892 | } else if (dev->node_props.cpu_cores_count) | |
893 | pr_info("Topology: Add CPU node\n"); | |
894 | else if (dev->node_props.simd_count) | |
895 | pr_info("Topology: Add dGPU node [0x%0x:0x%0x]\n", | |
896 | dev->node_props.device_id, | |
897 | dev->node_props.vendor_id); | |
898 | } | |
899 | up_read(&topology_lock); | |
900 | } | |
901 | ||
902 | /* Helper function for intializing platform_xx members of | |
903 | * kfd_system_properties. Uses OEM info from the last CPU/APU node. | |
904 | */ | |
905 | static void kfd_update_system_properties(void) | |
906 | { | |
907 | struct kfd_topology_device *dev; | |
908 | ||
909 | down_read(&topology_lock); | |
910 | dev = list_last_entry(&topology_device_list, | |
911 | struct kfd_topology_device, list); | |
912 | if (dev) { | |
913 | sys_props.platform_id = | |
914 | (*((uint64_t *)dev->oem_id)) & CRAT_OEMID_64BIT_MASK; | |
915 | sys_props.platform_oem = *((uint64_t *)dev->oem_table_id); | |
916 | sys_props.platform_rev = dev->oem_revision; | |
917 | } | |
918 | up_read(&topology_lock); | |
919 | } | |
920 | ||
921 | static void find_system_memory(const struct dmi_header *dm, | |
922 | void *private) | |
923 | { | |
924 | struct kfd_mem_properties *mem; | |
925 | u16 mem_width, mem_clock; | |
926 | struct kfd_topology_device *kdev = | |
927 | (struct kfd_topology_device *)private; | |
928 | const u8 *dmi_data = (const u8 *)(dm + 1); | |
929 | ||
930 | if (dm->type == DMI_ENTRY_MEM_DEVICE && dm->length >= 0x15) { | |
931 | mem_width = (u16)(*(const u16 *)(dmi_data + 0x6)); | |
932 | mem_clock = (u16)(*(const u16 *)(dmi_data + 0x11)); | |
933 | list_for_each_entry(mem, &kdev->mem_props, list) { | |
934 | if (mem_width != 0xFFFF && mem_width != 0) | |
935 | mem->width = mem_width; | |
936 | if (mem_clock != 0) | |
937 | mem->mem_clk_max = mem_clock; | |
938 | } | |
939 | } | |
940 | } | |
f4757347 AL |
941 | |
942 | /* | |
943 | * Performance counters information is not part of CRAT but we would like to | |
944 | * put them in the sysfs under topology directory for Thunk to get the data. | |
945 | * This function is called before updating the sysfs. | |
946 | */ | |
947 | static int kfd_add_perf_to_topology(struct kfd_topology_device *kdev) | |
948 | { | |
64d1c3a4 FK |
949 | /* These are the only counters supported so far */ |
950 | return kfd_iommu_add_perf_counters(kdev); | |
f4757347 AL |
951 | } |
952 | ||
520b8fb7 FK |
953 | /* kfd_add_non_crat_information - Add information that is not currently |
954 | * defined in CRAT but is necessary for KFD topology | |
955 | * @dev - topology device to which addition info is added | |
956 | */ | |
957 | static void kfd_add_non_crat_information(struct kfd_topology_device *kdev) | |
958 | { | |
959 | /* Check if CPU only node. */ | |
960 | if (!kdev->gpu) { | |
961 | /* Add system memory information */ | |
962 | dmi_walk(find_system_memory, kdev); | |
963 | } | |
964 | /* TODO: For GPU node, rearrange code from kfd_topology_add_device */ | |
965 | } | |
966 | ||
b441093e HK |
967 | /* kfd_is_acpi_crat_invalid - CRAT from ACPI is valid only for AMD APU devices. |
968 | * Ignore CRAT for all other devices. AMD APU is identified if both CPU | |
969 | * and GPU cores are present. | |
970 | * @device_list - topology device list created by parsing ACPI CRAT table. | |
971 | * @return - TRUE if invalid, FALSE is valid. | |
972 | */ | |
973 | static bool kfd_is_acpi_crat_invalid(struct list_head *device_list) | |
974 | { | |
975 | struct kfd_topology_device *dev; | |
976 | ||
977 | list_for_each_entry(dev, device_list, list) { | |
978 | if (dev->node_props.cpu_cores_count && | |
979 | dev->node_props.simd_count) | |
980 | return false; | |
981 | } | |
982 | pr_info("Ignoring ACPI CRAT on non-APU system\n"); | |
983 | return true; | |
984 | } | |
985 | ||
5b5c4e40 EP |
986 | int kfd_topology_init(void) |
987 | { | |
16b9201c | 988 | void *crat_image = NULL; |
5b5c4e40 EP |
989 | size_t image_size = 0; |
990 | int ret; | |
4f449311 | 991 | struct list_head temp_topology_device_list; |
520b8fb7 FK |
992 | int cpu_only_node = 0; |
993 | struct kfd_topology_device *kdev; | |
994 | int proximity_domain; | |
5b5c4e40 | 995 | |
4f449311 HK |
996 | /* topology_device_list - Master list of all topology devices |
997 | * temp_topology_device_list - temporary list created while parsing CRAT | |
998 | * or VCRAT. Once parsing is complete the contents of list is moved to | |
999 | * topology_device_list | |
5b5c4e40 | 1000 | */ |
4f449311 HK |
1001 | |
1002 | /* Initialize the head for the both the lists */ | |
5b5c4e40 | 1003 | INIT_LIST_HEAD(&topology_device_list); |
4f449311 | 1004 | INIT_LIST_HEAD(&temp_topology_device_list); |
5b5c4e40 | 1005 | init_rwsem(&topology_lock); |
5b5c4e40 EP |
1006 | |
1007 | memset(&sys_props, 0, sizeof(sys_props)); | |
1008 | ||
520b8fb7 FK |
1009 | /* Proximity domains in ACPI CRAT tables start counting at |
1010 | * 0. The same should be true for virtual CRAT tables created | |
1011 | * at this stage. GPUs added later in kfd_topology_add_device | |
1012 | * use a counter. | |
1013 | */ | |
1014 | proximity_domain = 0; | |
1015 | ||
5b5c4e40 | 1016 | /* |
520b8fb7 | 1017 | * Get the CRAT image from the ACPI. If ACPI doesn't have one |
b441093e | 1018 | * or if ACPI CRAT is invalid create a virtual CRAT. |
520b8fb7 FK |
1019 | * NOTE: The current implementation expects all AMD APUs to have |
1020 | * CRAT. If no CRAT is available, it is assumed to be a CPU | |
5b5c4e40 | 1021 | */ |
8e05247d HK |
1022 | ret = kfd_create_crat_image_acpi(&crat_image, &image_size); |
1023 | if (!ret) { | |
4f449311 | 1024 | ret = kfd_parse_crat_table(crat_image, |
520b8fb7 FK |
1025 | &temp_topology_device_list, |
1026 | proximity_domain); | |
b441093e HK |
1027 | if (ret || |
1028 | kfd_is_acpi_crat_invalid(&temp_topology_device_list)) { | |
520b8fb7 FK |
1029 | kfd_release_topology_device_list( |
1030 | &temp_topology_device_list); | |
1031 | kfd_destroy_crat_image(crat_image); | |
1032 | crat_image = NULL; | |
1033 | } | |
1034 | } | |
1035 | ||
1036 | if (!crat_image) { | |
1037 | ret = kfd_create_crat_image_virtual(&crat_image, &image_size, | |
1038 | COMPUTE_UNIT_CPU, NULL, | |
1039 | proximity_domain); | |
1040 | cpu_only_node = 1; | |
1041 | if (ret) { | |
1042 | pr_err("Error creating VCRAT table for CPU\n"); | |
1043 | return ret; | |
1044 | } | |
1045 | ||
1046 | ret = kfd_parse_crat_table(crat_image, | |
1047 | &temp_topology_device_list, | |
1048 | proximity_domain); | |
1049 | if (ret) { | |
1050 | pr_err("Error parsing VCRAT table for CPU\n"); | |
5b5c4e40 | 1051 | goto err; |
520b8fb7 | 1052 | } |
5b5c4e40 EP |
1053 | } |
1054 | ||
f4757347 AL |
1055 | kdev = list_first_entry(&temp_topology_device_list, |
1056 | struct kfd_topology_device, list); | |
1057 | kfd_add_perf_to_topology(kdev); | |
1058 | ||
8e05247d | 1059 | down_write(&topology_lock); |
4f449311 HK |
1060 | kfd_topology_update_device_list(&temp_topology_device_list, |
1061 | &topology_device_list); | |
520b8fb7 | 1062 | atomic_set(&topology_crat_proximity_domain, sys_props.num_devices-1); |
8e05247d HK |
1063 | ret = kfd_topology_update_sysfs(); |
1064 | up_write(&topology_lock); | |
1065 | ||
4f449311 HK |
1066 | if (!ret) { |
1067 | sys_props.generation_count++; | |
520b8fb7 FK |
1068 | kfd_update_system_properties(); |
1069 | kfd_debug_print_topology(); | |
4f449311 | 1070 | } else |
8e05247d HK |
1071 | pr_err("Failed to update topology in sysfs ret=%d\n", ret); |
1072 | ||
520b8fb7 FK |
1073 | /* For nodes with GPU, this information gets added |
1074 | * when GPU is detected (kfd_topology_add_device). | |
1075 | */ | |
1076 | if (cpu_only_node) { | |
1077 | /* Add additional information to CPU only node created above */ | |
1078 | down_write(&topology_lock); | |
1079 | kdev = list_first_entry(&topology_device_list, | |
1080 | struct kfd_topology_device, list); | |
1081 | up_write(&topology_lock); | |
1082 | kfd_add_non_crat_information(kdev); | |
1083 | } | |
1084 | ||
5b5c4e40 | 1085 | err: |
8e05247d | 1086 | kfd_destroy_crat_image(crat_image); |
5b5c4e40 EP |
1087 | return ret; |
1088 | } | |
1089 | ||
1090 | void kfd_topology_shutdown(void) | |
1091 | { | |
4f449311 | 1092 | down_write(&topology_lock); |
5b5c4e40 EP |
1093 | kfd_topology_release_sysfs(); |
1094 | kfd_release_live_view(); | |
4f449311 | 1095 | up_write(&topology_lock); |
5b5c4e40 EP |
1096 | } |
1097 | ||
5b5c4e40 EP |
1098 | static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu) |
1099 | { | |
1100 | uint32_t hashout; | |
1101 | uint32_t buf[7]; | |
585f0e6c | 1102 | uint64_t local_mem_size; |
5b5c4e40 | 1103 | int i; |
0504cccf | 1104 | struct kfd_local_mem_info local_mem_info; |
5b5c4e40 EP |
1105 | |
1106 | if (!gpu) | |
1107 | return 0; | |
1108 | ||
574c4183 | 1109 | amdgpu_amdkfd_get_local_mem_info(gpu->adev, &local_mem_info); |
0504cccf HK |
1110 | |
1111 | local_mem_size = local_mem_info.local_mem_size_private + | |
1112 | local_mem_info.local_mem_size_public; | |
585f0e6c | 1113 | |
5b5c4e40 | 1114 | buf[0] = gpu->pdev->devfn; |
46096058 AL |
1115 | buf[1] = gpu->pdev->subsystem_vendor | |
1116 | (gpu->pdev->subsystem_device << 16); | |
1117 | buf[2] = pci_domain_nr(gpu->pdev->bus); | |
5b5c4e40 EP |
1118 | buf[3] = gpu->pdev->device; |
1119 | buf[4] = gpu->pdev->bus->number; | |
585f0e6c EC |
1120 | buf[5] = lower_32_bits(local_mem_size); |
1121 | buf[6] = upper_32_bits(local_mem_size); | |
5b5c4e40 EP |
1122 | |
1123 | for (i = 0, hashout = 0; i < 7; i++) | |
1124 | hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH); | |
1125 | ||
1126 | return hashout; | |
1127 | } | |
3a87177e HK |
1128 | /* kfd_assign_gpu - Attach @gpu to the correct kfd topology device. If |
1129 | * the GPU device is not already present in the topology device | |
1130 | * list then return NULL. This means a new topology device has to | |
1131 | * be created for this GPU. | |
3a87177e | 1132 | */ |
5b5c4e40 EP |
1133 | static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu) |
1134 | { | |
1135 | struct kfd_topology_device *dev; | |
16b9201c | 1136 | struct kfd_topology_device *out_dev = NULL; |
171bc67e HK |
1137 | struct kfd_mem_properties *mem; |
1138 | struct kfd_cache_properties *cache; | |
1139 | struct kfd_iolink_properties *iolink; | |
5b5c4e40 | 1140 | |
3a87177e | 1141 | down_write(&topology_lock); |
b8fe0524 FK |
1142 | list_for_each_entry(dev, &topology_device_list, list) { |
1143 | /* Discrete GPUs need their own topology device list | |
1144 | * entries. Don't assign them to CPU/APU nodes. | |
1145 | */ | |
6127896f | 1146 | if (!gpu->use_iommu_v2 && |
b8fe0524 FK |
1147 | dev->node_props.cpu_cores_count) |
1148 | continue; | |
1149 | ||
4eacc26b | 1150 | if (!dev->gpu && (dev->node_props.simd_count > 0)) { |
5b5c4e40 EP |
1151 | dev->gpu = gpu; |
1152 | out_dev = dev; | |
171bc67e HK |
1153 | |
1154 | list_for_each_entry(mem, &dev->mem_props, list) | |
1155 | mem->gpu = dev->gpu; | |
1156 | list_for_each_entry(cache, &dev->cache_props, list) | |
1157 | cache->gpu = dev->gpu; | |
1158 | list_for_each_entry(iolink, &dev->io_link_props, list) | |
1159 | iolink->gpu = dev->gpu; | |
5b5c4e40 EP |
1160 | break; |
1161 | } | |
b8fe0524 | 1162 | } |
3a87177e | 1163 | up_write(&topology_lock); |
5b5c4e40 EP |
1164 | return out_dev; |
1165 | } | |
1166 | ||
1167 | static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival) | |
1168 | { | |
1169 | /* | |
1170 | * TODO: Generate an event for thunk about the arrival/removal | |
1171 | * of the GPU | |
1172 | */ | |
1173 | } | |
1174 | ||
3a87177e HK |
1175 | /* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info, |
1176 | * patch this after CRAT parsing. | |
1177 | */ | |
1178 | static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev) | |
1179 | { | |
1180 | struct kfd_mem_properties *mem; | |
1181 | struct kfd_local_mem_info local_mem_info; | |
1182 | ||
1183 | if (!dev) | |
1184 | return; | |
1185 | ||
1186 | /* Currently, amdgpu driver (amdgpu_mc) deals only with GPUs with | |
1187 | * single bank of VRAM local memory. | |
1188 | * for dGPUs - VCRAT reports only one bank of Local Memory | |
1189 | * for APUs - If CRAT from ACPI reports more than one bank, then | |
1190 | * all the banks will report the same mem_clk_max information | |
1191 | */ | |
574c4183 | 1192 | amdgpu_amdkfd_get_local_mem_info(dev->gpu->adev, &local_mem_info); |
3a87177e HK |
1193 | |
1194 | list_for_each_entry(mem, &dev->mem_props, list) | |
1195 | mem->mem_clk_max = local_mem_info.mem_clk_max; | |
1196 | } | |
1197 | ||
bdd24657 JK |
1198 | static void kfd_set_iolink_no_atomics(struct kfd_topology_device *dev, |
1199 | struct kfd_topology_device *target_gpu_dev, | |
1200 | struct kfd_iolink_properties *link) | |
3a87177e | 1201 | { |
bdd24657 JK |
1202 | /* xgmi always supports atomics between links. */ |
1203 | if (link->iolink_type == CRAT_IOLINK_TYPE_XGMI) | |
3a87177e HK |
1204 | return; |
1205 | ||
bdd24657 JK |
1206 | /* check pcie support to set cpu(dev) flags for target_gpu_dev link. */ |
1207 | if (target_gpu_dev) { | |
1208 | uint32_t cap; | |
1209 | ||
1210 | pcie_capability_read_dword(target_gpu_dev->gpu->pdev, | |
deb68983 | 1211 | PCI_EXP_DEVCAP2, &cap); |
d35f00d8 | 1212 | |
deb68983 JK |
1213 | if (!(cap & (PCI_EXP_DEVCAP2_ATOMIC_COMP32 | |
1214 | PCI_EXP_DEVCAP2_ATOMIC_COMP64))) | |
bdd24657 | 1215 | link->flags |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT | |
deb68983 | 1216 | CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT; |
bdd24657 JK |
1217 | /* set gpu (dev) flags. */ |
1218 | } else { | |
deb68983 | 1219 | if (!dev->gpu->pci_atomic_requested || |
7eb0502a | 1220 | dev->gpu->adev->asic_type == CHIP_HAWAII) |
bdd24657 | 1221 | link->flags |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT | |
deb68983 JK |
1222 | CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT; |
1223 | } | |
bdd24657 JK |
1224 | } |
1225 | ||
c9cfbf7f EH |
1226 | static void kfd_set_iolink_non_coherent(struct kfd_topology_device *to_dev, |
1227 | struct kfd_iolink_properties *outbound_link, | |
1228 | struct kfd_iolink_properties *inbound_link) | |
1229 | { | |
1230 | /* CPU -> GPU with PCIe */ | |
1231 | if (!to_dev->gpu && | |
1232 | inbound_link->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS) | |
1233 | inbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT; | |
1234 | ||
1235 | if (to_dev->gpu) { | |
1236 | /* GPU <-> GPU with PCIe and | |
1237 | * Vega20 with XGMI | |
1238 | */ | |
1239 | if (inbound_link->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS || | |
1240 | (inbound_link->iolink_type == CRAT_IOLINK_TYPE_XGMI && | |
046e674b | 1241 | KFD_GC_VERSION(to_dev->gpu) == IP_VERSION(9, 4, 0))) { |
c9cfbf7f EH |
1242 | outbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT; |
1243 | inbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT; | |
1244 | } | |
1245 | } | |
1246 | } | |
1247 | ||
bdd24657 JK |
1248 | static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev) |
1249 | { | |
1250 | struct kfd_iolink_properties *link, *inbound_link; | |
1251 | struct kfd_topology_device *peer_dev; | |
1252 | ||
1253 | if (!dev || !dev->gpu) | |
1254 | return; | |
d35f00d8 EH |
1255 | |
1256 | /* GPU only creates direct links so apply flags setting to all */ | |
1257 | list_for_each_entry(link, &dev->io_link_props, list) { | |
bdd24657 JK |
1258 | link->flags = CRAT_IOLINK_FLAGS_ENABLED; |
1259 | kfd_set_iolink_no_atomics(dev, NULL, link); | |
1260 | peer_dev = kfd_topology_device_by_proximity_domain( | |
d35f00d8 | 1261 | link->node_to); |
bdd24657 JK |
1262 | |
1263 | if (!peer_dev) | |
1264 | continue; | |
1265 | ||
1266 | list_for_each_entry(inbound_link, &peer_dev->io_link_props, | |
1267 | list) { | |
1268 | if (inbound_link->node_to != link->node_from) | |
1269 | continue; | |
1270 | ||
1271 | inbound_link->flags = CRAT_IOLINK_FLAGS_ENABLED; | |
1272 | kfd_set_iolink_no_atomics(peer_dev, dev, inbound_link); | |
c9cfbf7f | 1273 | kfd_set_iolink_non_coherent(peer_dev, link, inbound_link); |
d35f00d8 EH |
1274 | } |
1275 | } | |
3a87177e HK |
1276 | } |
1277 | ||
5b5c4e40 EP |
1278 | int kfd_topology_add_device(struct kfd_dev *gpu) |
1279 | { | |
1280 | uint32_t gpu_id; | |
1281 | struct kfd_topology_device *dev; | |
f7ce2fad | 1282 | struct kfd_cu_info cu_info; |
4f449311 HK |
1283 | int res = 0; |
1284 | struct list_head temp_topology_device_list; | |
3a87177e HK |
1285 | void *crat_image = NULL; |
1286 | size_t image_size = 0; | |
1287 | int proximity_domain; | |
4f449311 HK |
1288 | |
1289 | INIT_LIST_HEAD(&temp_topology_device_list); | |
5b5c4e40 | 1290 | |
5b5c4e40 EP |
1291 | gpu_id = kfd_generate_gpu_id(gpu); |
1292 | ||
79775b62 | 1293 | pr_debug("Adding new GPU (ID: 0x%x) to topology\n", gpu_id); |
5b5c4e40 | 1294 | |
3a87177e HK |
1295 | proximity_domain = atomic_inc_return(&topology_crat_proximity_domain); |
1296 | ||
d5edb56f | 1297 | /* Include the CPU in xGMI hive if xGMI connected by assigning it the hive ID. */ |
56c5977e | 1298 | if (gpu->hive_id && gpu->adev->gmc.xgmi.connected_to_cpu) { |
d5edb56f JK |
1299 | struct kfd_topology_device *top_dev; |
1300 | ||
1301 | down_read(&topology_lock); | |
1302 | ||
1303 | list_for_each_entry(top_dev, &topology_device_list, list) { | |
1304 | if (top_dev->gpu) | |
1305 | break; | |
1306 | ||
1307 | top_dev->node_props.hive_id = gpu->hive_id; | |
1308 | } | |
1309 | ||
1310 | up_read(&topology_lock); | |
1311 | } | |
1312 | ||
3a87177e HK |
1313 | /* Check to see if this gpu device exists in the topology_device_list. |
1314 | * If so, assign the gpu to that device, | |
1315 | * else create a Virtual CRAT for this gpu device and then parse that | |
1316 | * CRAT to create a new topology device. Once created assign the gpu to | |
1317 | * that topology device | |
5b5c4e40 EP |
1318 | */ |
1319 | dev = kfd_assign_gpu(gpu); | |
1320 | if (!dev) { | |
3a87177e HK |
1321 | res = kfd_create_crat_image_virtual(&crat_image, &image_size, |
1322 | COMPUTE_UNIT_GPU, gpu, | |
1323 | proximity_domain); | |
1324 | if (res) { | |
1325 | pr_err("Error creating VCRAT for GPU (ID: 0x%x)\n", | |
1326 | gpu_id); | |
1327 | return res; | |
1328 | } | |
1329 | res = kfd_parse_crat_table(crat_image, | |
1330 | &temp_topology_device_list, | |
1331 | proximity_domain); | |
1332 | if (res) { | |
1333 | pr_err("Error parsing VCRAT for GPU (ID: 0x%x)\n", | |
1334 | gpu_id); | |
5b5c4e40 EP |
1335 | goto err; |
1336 | } | |
4f449311 | 1337 | |
4f449311 HK |
1338 | down_write(&topology_lock); |
1339 | kfd_topology_update_device_list(&temp_topology_device_list, | |
1340 | &topology_device_list); | |
1341 | ||
8eabaf54 KR |
1342 | /* Update the SYSFS tree, since we added another topology |
1343 | * device | |
5b5c4e40 | 1344 | */ |
3a87177e | 1345 | res = kfd_topology_update_sysfs(); |
4f449311 HK |
1346 | up_write(&topology_lock); |
1347 | ||
3a87177e HK |
1348 | if (!res) |
1349 | sys_props.generation_count++; | |
1350 | else | |
1351 | pr_err("Failed to update GPU (ID: 0x%x) to sysfs topology. res=%d\n", | |
1352 | gpu_id, res); | |
1353 | dev = kfd_assign_gpu(gpu); | |
1354 | if (WARN_ON(!dev)) { | |
1355 | res = -ENODEV; | |
1356 | goto err; | |
1357 | } | |
5b5c4e40 EP |
1358 | } |
1359 | ||
1360 | dev->gpu_id = gpu_id; | |
1361 | gpu->id = gpu_id; | |
3a87177e HK |
1362 | |
1363 | /* TODO: Move the following lines to function | |
1364 | * kfd_add_non_crat_information | |
1365 | */ | |
1366 | ||
1367 | /* Fill-in additional information that is not available in CRAT but | |
1368 | * needed for the topology | |
1369 | */ | |
1370 | ||
574c4183 | 1371 | amdgpu_amdkfd_get_cu_info(dev->gpu->adev, &cu_info); |
c181159a YZ |
1372 | |
1373 | strncpy(dev->node_props.name, gpu->device_info->asic_name, | |
1374 | KFD_TOPOLOGY_PUBLIC_NAME_SIZE); | |
1375 | ||
3a87177e HK |
1376 | dev->node_props.simd_arrays_per_engine = |
1377 | cu_info.num_shader_arrays_per_engine; | |
1378 | ||
9d6fa9c7 | 1379 | dev->node_props.gfx_target_version = gpu->device_info->gfx_target_version; |
5b5c4e40 EP |
1380 | dev->node_props.vendor_id = gpu->pdev->vendor; |
1381 | dev->node_props.device_id = gpu->pdev->device; | |
c6d1ec41 | 1382 | dev->node_props.capability |= |
02274fc0 | 1383 | ((dev->gpu->adev->rev_id << HSA_CAP_ASIC_REVISION_SHIFT) & |
c6d1ec41 | 1384 | HSA_CAP_ASIC_REVISION_MASK); |
babe2ef3 | 1385 | dev->node_props.location_id = pci_dev_id(gpu->pdev); |
3e58e95a | 1386 | dev->node_props.domain = pci_domain_nr(gpu->pdev->bus); |
3a87177e | 1387 | dev->node_props.max_engine_clk_fcompute = |
574c4183 | 1388 | amdgpu_amdkfd_get_max_engine_clock_in_mhz(dev->gpu->adev); |
3a87177e HK |
1389 | dev->node_props.max_engine_clk_ccompute = |
1390 | cpufreq_quick_get_max(0) / 1000; | |
7c9b7171 OZ |
1391 | dev->node_props.drm_render_minor = |
1392 | gpu->shared_resources.drm_render_minor; | |
3a87177e | 1393 | |
0c1690e3 | 1394 | dev->node_props.hive_id = gpu->hive_id; |
ee2f17f4 | 1395 | dev->node_props.num_sdma_engines = kfd_get_num_sdma_engines(gpu); |
14568cf6 | 1396 | dev->node_props.num_sdma_xgmi_engines = |
ee2f17f4 | 1397 | kfd_get_num_xgmi_sdma_engines(gpu); |
bb71c74d HR |
1398 | dev->node_props.num_sdma_queues_per_engine = |
1399 | gpu->device_info->num_sdma_queues_per_engine; | |
29633d0e | 1400 | dev->node_props.num_gws = (dev->gpu->gws && |
29e76462 | 1401 | dev->gpu->dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) ? |
02274fc0 | 1402 | dev->gpu->adev->gds.gws_size : 0; |
e6945304 | 1403 | dev->node_props.num_cp_queues = get_cp_queues_num(dev->gpu->dqm); |
0c1690e3 | 1404 | |
3a87177e HK |
1405 | kfd_fill_mem_clk_max_info(dev); |
1406 | kfd_fill_iolink_non_crat_info(dev); | |
1407 | ||
7eb0502a | 1408 | switch (dev->gpu->adev->asic_type) { |
3a87177e HK |
1409 | case CHIP_KAVERI: |
1410 | case CHIP_HAWAII: | |
1411 | case CHIP_TONGA: | |
1412 | dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_PRE_1_0 << | |
1413 | HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & | |
1414 | HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); | |
1415 | break; | |
1416 | case CHIP_CARRIZO: | |
1417 | case CHIP_FIJI: | |
1418 | case CHIP_POLARIS10: | |
1419 | case CHIP_POLARIS11: | |
846a44d7 | 1420 | case CHIP_POLARIS12: |
ed81cd6e | 1421 | case CHIP_VEGAM: |
42aa8793 | 1422 | pr_debug("Adding doorbell packet type capability\n"); |
3a87177e HK |
1423 | dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 << |
1424 | HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & | |
1425 | HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); | |
1426 | break; | |
1427 | default: | |
e4804a39 GS |
1428 | if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(9, 0, 1)) |
1429 | dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 << | |
1430 | HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & | |
1431 | HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); | |
1432 | else | |
1433 | WARN(1, "Unexpected ASIC family %u", | |
7eb0502a | 1434 | dev->gpu->adev->asic_type); |
7639a8c4 BG |
1435 | } |
1436 | ||
1ae99eab OZ |
1437 | /* |
1438 | * Overwrite ATS capability according to needs_iommu_device to fix | |
1439 | * potential missing corresponding bit in CRAT of BIOS. | |
1440 | */ | |
6127896f | 1441 | if (dev->gpu->use_iommu_v2) |
1ae99eab OZ |
1442 | dev->node_props.capability |= HSA_CAP_ATS_PRESENT; |
1443 | else | |
1444 | dev->node_props.capability &= ~HSA_CAP_ATS_PRESENT; | |
1445 | ||
3a87177e HK |
1446 | /* Fix errors in CZ CRAT. |
1447 | * simd_count: Carrizo CRAT reports wrong simd_count, probably | |
1448 | * because it doesn't consider masked out CUs | |
70f372bf | 1449 | * max_waves_per_simd: Carrizo reports wrong max_waves_per_simd |
3a87177e | 1450 | */ |
7eb0502a | 1451 | if (dev->gpu->adev->asic_type == CHIP_CARRIZO) { |
3a87177e HK |
1452 | dev->node_props.simd_count = |
1453 | cu_info.simd_per_cu * cu_info.cu_active_number; | |
70f372bf | 1454 | dev->node_props.max_waves_per_simd = 10; |
70f372bf | 1455 | } |
3a87177e | 1456 | |
5436ab94 SY |
1457 | /* kfd only concerns sram ecc on GFX and HBM ecc on UMC */ |
1458 | dev->node_props.capability |= | |
56c5977e | 1459 | ((dev->gpu->adev->ras_enabled & BIT(AMDGPU_RAS_BLOCK__GFX)) != 0) ? |
5436ab94 | 1460 | HSA_CAP_SRAM_EDCSUPPORTED : 0; |
56c5977e GS |
1461 | dev->node_props.capability |= |
1462 | ((dev->gpu->adev->ras_enabled & BIT(AMDGPU_RAS_BLOCK__UMC)) != 0) ? | |
5436ab94 SY |
1463 | HSA_CAP_MEM_EDCSUPPORTED : 0; |
1464 | ||
046e674b | 1465 | if (KFD_GC_VERSION(dev->gpu) != IP_VERSION(9, 0, 1)) |
56c5977e | 1466 | dev->node_props.capability |= (dev->gpu->adev->ras_enabled != 0) ? |
0dee45a2 | 1467 | HSA_CAP_RASEVENTNOTIFY : 0; |
0dee45a2 | 1468 | |
56c5977e | 1469 | if (KFD_IS_SVM_API_SUPPORTED(dev->gpu->adev->kfd.dev)) |
4c166eb9 PY |
1470 | dev->node_props.capability |= HSA_CAP_SVMAPI_SUPPORTED; |
1471 | ||
3a87177e HK |
1472 | kfd_debug_print_topology(); |
1473 | ||
4f449311 | 1474 | if (!res) |
5b5c4e40 | 1475 | kfd_notify_gpu_change(gpu_id, 1); |
4f449311 | 1476 | err: |
3a87177e | 1477 | kfd_destroy_crat_image(crat_image); |
5b5c4e40 EP |
1478 | return res; |
1479 | } | |
1480 | ||
1481 | int kfd_topology_remove_device(struct kfd_dev *gpu) | |
1482 | { | |
4f449311 | 1483 | struct kfd_topology_device *dev, *tmp; |
5b5c4e40 EP |
1484 | uint32_t gpu_id; |
1485 | int res = -ENODEV; | |
1486 | ||
5b5c4e40 EP |
1487 | down_write(&topology_lock); |
1488 | ||
4f449311 | 1489 | list_for_each_entry_safe(dev, tmp, &topology_device_list, list) |
5b5c4e40 EP |
1490 | if (dev->gpu == gpu) { |
1491 | gpu_id = dev->gpu_id; | |
1492 | kfd_remove_sysfs_node_entry(dev); | |
1493 | kfd_release_topology_device(dev); | |
4f449311 | 1494 | sys_props.num_devices--; |
5b5c4e40 EP |
1495 | res = 0; |
1496 | if (kfd_topology_update_sysfs() < 0) | |
1497 | kfd_topology_release_sysfs(); | |
1498 | break; | |
1499 | } | |
1500 | ||
1501 | up_write(&topology_lock); | |
1502 | ||
174de876 | 1503 | if (!res) |
5b5c4e40 EP |
1504 | kfd_notify_gpu_change(gpu_id, 0); |
1505 | ||
1506 | return res; | |
1507 | } | |
1508 | ||
6d82eb0e HK |
1509 | /* kfd_topology_enum_kfd_devices - Enumerate through all devices in KFD |
1510 | * topology. If GPU device is found @idx, then valid kfd_dev pointer is | |
1511 | * returned through @kdev | |
1512 | * Return - 0: On success (@kdev will be NULL for non GPU nodes) | |
1513 | * -1: If end of list | |
5b5c4e40 | 1514 | */ |
6d82eb0e | 1515 | int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev) |
5b5c4e40 EP |
1516 | { |
1517 | ||
1518 | struct kfd_topology_device *top_dev; | |
5b5c4e40 EP |
1519 | uint8_t device_idx = 0; |
1520 | ||
6d82eb0e | 1521 | *kdev = NULL; |
5b5c4e40 EP |
1522 | down_read(&topology_lock); |
1523 | ||
1524 | list_for_each_entry(top_dev, &topology_device_list, list) { | |
1525 | if (device_idx == idx) { | |
6d82eb0e HK |
1526 | *kdev = top_dev->gpu; |
1527 | up_read(&topology_lock); | |
1528 | return 0; | |
5b5c4e40 EP |
1529 | } |
1530 | ||
1531 | device_idx++; | |
1532 | } | |
1533 | ||
1534 | up_read(&topology_lock); | |
1535 | ||
6d82eb0e | 1536 | return -1; |
5b5c4e40 EP |
1537 | |
1538 | } | |
851a645e | 1539 | |
520b8fb7 FK |
1540 | static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask) |
1541 | { | |
520b8fb7 FK |
1542 | int first_cpu_of_numa_node; |
1543 | ||
1544 | if (!cpumask || cpumask == cpu_none_mask) | |
1545 | return -1; | |
1546 | first_cpu_of_numa_node = cpumask_first(cpumask); | |
1547 | if (first_cpu_of_numa_node >= nr_cpu_ids) | |
1548 | return -1; | |
df1dd4f4 FK |
1549 | #ifdef CONFIG_X86_64 |
1550 | return cpu_data(first_cpu_of_numa_node).apicid; | |
1551 | #else | |
1552 | return first_cpu_of_numa_node; | |
1553 | #endif | |
520b8fb7 FK |
1554 | } |
1555 | ||
1556 | /* kfd_numa_node_to_apic_id - Returns the APIC ID of the first logical processor | |
1557 | * of the given NUMA node (numa_node_id) | |
1558 | * Return -1 on failure | |
1559 | */ | |
1560 | int kfd_numa_node_to_apic_id(int numa_node_id) | |
1561 | { | |
1562 | if (numa_node_id == -1) { | |
1563 | pr_warn("Invalid NUMA Node. Use online CPU mask\n"); | |
1564 | return kfd_cpumask_to_apic_id(cpu_online_mask); | |
1565 | } | |
1566 | return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id)); | |
1567 | } | |
1568 | ||
6127896f HR |
1569 | void kfd_double_confirm_iommu_support(struct kfd_dev *gpu) |
1570 | { | |
1571 | struct kfd_topology_device *dev; | |
1572 | ||
1573 | gpu->use_iommu_v2 = false; | |
1574 | ||
1575 | if (!gpu->device_info->needs_iommu_device) | |
1576 | return; | |
1577 | ||
1578 | down_read(&topology_lock); | |
1579 | ||
1580 | /* Only use IOMMUv2 if there is an APU topology node with no GPU | |
1581 | * assigned yet. This GPU will be assigned to it. | |
1582 | */ | |
1583 | list_for_each_entry(dev, &topology_device_list, list) | |
1584 | if (dev->node_props.cpu_cores_count && | |
1585 | dev->node_props.simd_count && | |
1586 | !dev->gpu) | |
1587 | gpu->use_iommu_v2 = true; | |
1588 | ||
1589 | up_read(&topology_lock); | |
1590 | } | |
1591 | ||
851a645e FK |
1592 | #if defined(CONFIG_DEBUG_FS) |
1593 | ||
1594 | int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data) | |
1595 | { | |
1596 | struct kfd_topology_device *dev; | |
1597 | unsigned int i = 0; | |
1598 | int r = 0; | |
1599 | ||
1600 | down_read(&topology_lock); | |
1601 | ||
1602 | list_for_each_entry(dev, &topology_device_list, list) { | |
1603 | if (!dev->gpu) { | |
1604 | i++; | |
1605 | continue; | |
1606 | } | |
1607 | ||
1608 | seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id); | |
1609 | r = dqm_debugfs_hqds(m, dev->gpu->dqm); | |
1610 | if (r) | |
1611 | break; | |
1612 | } | |
1613 | ||
1614 | up_read(&topology_lock); | |
1615 | ||
1616 | return r; | |
1617 | } | |
1618 | ||
1619 | int kfd_debugfs_rls_by_device(struct seq_file *m, void *data) | |
1620 | { | |
1621 | struct kfd_topology_device *dev; | |
1622 | unsigned int i = 0; | |
1623 | int r = 0; | |
1624 | ||
1625 | down_read(&topology_lock); | |
1626 | ||
1627 | list_for_each_entry(dev, &topology_device_list, list) { | |
1628 | if (!dev->gpu) { | |
1629 | i++; | |
1630 | continue; | |
1631 | } | |
1632 | ||
1633 | seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id); | |
9af5379c | 1634 | r = pm_debugfs_runlist(m, &dev->gpu->dqm->packet_mgr); |
851a645e FK |
1635 | if (r) |
1636 | break; | |
1637 | } | |
1638 | ||
1639 | up_read(&topology_lock); | |
1640 | ||
1641 | return r; | |
1642 | } | |
1643 | ||
1644 | #endif |