| 1 | // SPDX-License-Identifier: GPL-2.0 OR MIT |
| 2 | /* |
| 3 | * Copyright 2014-2022 Advanced Micro Devices, Inc. |
| 4 | * |
| 5 | * Permission is hereby granted, free of charge, to any person obtaining a |
| 6 | * copy of this software and associated documentation files (the "Software"), |
| 7 | * to deal in the Software without restriction, including without limitation |
| 8 | * the rights to use, copy, modify, merge, publish, distribute, sublicense, |
| 9 | * and/or sell copies of the Software, and to permit persons to whom the |
| 10 | * Software is furnished to do so, subject to the following conditions: |
| 11 | * |
| 12 | * The above copyright notice and this permission notice shall be included in |
| 13 | * all copies or substantial portions of the Software. |
| 14 | * |
| 15 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 16 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 17 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL |
| 18 | * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR |
| 19 | * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, |
| 20 | * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR |
| 21 | * OTHER DEALINGS IN THE SOFTWARE. |
| 22 | */ |
| 23 | |
| 24 | #include <linux/types.h> |
| 25 | #include <linux/kernel.h> |
| 26 | #include <linux/pci.h> |
| 27 | #include <linux/errno.h> |
| 28 | #include <linux/acpi.h> |
| 29 | #include <linux/hash.h> |
| 30 | #include <linux/cpufreq.h> |
| 31 | #include <linux/log2.h> |
| 32 | #include <linux/dmi.h> |
| 33 | #include <linux/atomic.h> |
| 34 | |
| 35 | #include "kfd_priv.h" |
| 36 | #include "kfd_crat.h" |
| 37 | #include "kfd_topology.h" |
| 38 | #include "kfd_device_queue_manager.h" |
| 39 | #include "kfd_svm.h" |
| 40 | #include "kfd_debug.h" |
| 41 | #include "amdgpu_amdkfd.h" |
| 42 | #include "amdgpu_ras.h" |
| 43 | #include "amdgpu.h" |
| 44 | |
| 45 | /* topology_device_list - Master list of all topology devices */ |
| 46 | static struct list_head topology_device_list; |
| 47 | static struct kfd_system_properties sys_props; |
| 48 | |
| 49 | static DECLARE_RWSEM(topology_lock); |
| 50 | static uint32_t topology_crat_proximity_domain; |
| 51 | |
| 52 | struct kfd_topology_device *kfd_topology_device_by_proximity_domain_no_lock( |
| 53 | uint32_t proximity_domain) |
| 54 | { |
| 55 | struct kfd_topology_device *top_dev; |
| 56 | struct kfd_topology_device *device = NULL; |
| 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 | return device; |
| 65 | } |
| 66 | |
| 67 | struct kfd_topology_device *kfd_topology_device_by_proximity_domain( |
| 68 | uint32_t proximity_domain) |
| 69 | { |
| 70 | struct kfd_topology_device *device = NULL; |
| 71 | |
| 72 | down_read(&topology_lock); |
| 73 | |
| 74 | device = kfd_topology_device_by_proximity_domain_no_lock( |
| 75 | proximity_domain); |
| 76 | up_read(&topology_lock); |
| 77 | |
| 78 | return device; |
| 79 | } |
| 80 | |
| 81 | struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id) |
| 82 | { |
| 83 | struct kfd_topology_device *top_dev = NULL; |
| 84 | struct kfd_topology_device *ret = NULL; |
| 85 | |
| 86 | down_read(&topology_lock); |
| 87 | |
| 88 | list_for_each_entry(top_dev, &topology_device_list, list) |
| 89 | if (top_dev->gpu_id == gpu_id) { |
| 90 | ret = top_dev; |
| 91 | break; |
| 92 | } |
| 93 | |
| 94 | up_read(&topology_lock); |
| 95 | |
| 96 | return ret; |
| 97 | } |
| 98 | |
| 99 | struct kfd_node *kfd_device_by_id(uint32_t gpu_id) |
| 100 | { |
| 101 | struct kfd_topology_device *top_dev; |
| 102 | |
| 103 | top_dev = kfd_topology_device_by_id(gpu_id); |
| 104 | if (!top_dev) |
| 105 | return NULL; |
| 106 | |
| 107 | return top_dev->gpu; |
| 108 | } |
| 109 | |
| 110 | struct kfd_node *kfd_device_by_pci_dev(const struct pci_dev *pdev) |
| 111 | { |
| 112 | struct kfd_topology_device *top_dev; |
| 113 | struct kfd_node *device = NULL; |
| 114 | |
| 115 | down_read(&topology_lock); |
| 116 | |
| 117 | list_for_each_entry(top_dev, &topology_device_list, list) |
| 118 | if (top_dev->gpu && top_dev->gpu->adev->pdev == pdev) { |
| 119 | device = top_dev->gpu; |
| 120 | break; |
| 121 | } |
| 122 | |
| 123 | up_read(&topology_lock); |
| 124 | |
| 125 | return device; |
| 126 | } |
| 127 | |
| 128 | /* Called with write topology_lock acquired */ |
| 129 | static void kfd_release_topology_device(struct kfd_topology_device *dev) |
| 130 | { |
| 131 | struct kfd_mem_properties *mem; |
| 132 | struct kfd_cache_properties *cache; |
| 133 | struct kfd_iolink_properties *iolink; |
| 134 | struct kfd_iolink_properties *p2plink; |
| 135 | struct kfd_perf_properties *perf; |
| 136 | |
| 137 | list_del(&dev->list); |
| 138 | |
| 139 | while (dev->mem_props.next != &dev->mem_props) { |
| 140 | mem = container_of(dev->mem_props.next, |
| 141 | struct kfd_mem_properties, list); |
| 142 | list_del(&mem->list); |
| 143 | kfree(mem); |
| 144 | } |
| 145 | |
| 146 | while (dev->cache_props.next != &dev->cache_props) { |
| 147 | cache = container_of(dev->cache_props.next, |
| 148 | struct kfd_cache_properties, list); |
| 149 | list_del(&cache->list); |
| 150 | kfree(cache); |
| 151 | } |
| 152 | |
| 153 | while (dev->io_link_props.next != &dev->io_link_props) { |
| 154 | iolink = container_of(dev->io_link_props.next, |
| 155 | struct kfd_iolink_properties, list); |
| 156 | list_del(&iolink->list); |
| 157 | kfree(iolink); |
| 158 | } |
| 159 | |
| 160 | while (dev->p2p_link_props.next != &dev->p2p_link_props) { |
| 161 | p2plink = container_of(dev->p2p_link_props.next, |
| 162 | struct kfd_iolink_properties, list); |
| 163 | list_del(&p2plink->list); |
| 164 | kfree(p2plink); |
| 165 | } |
| 166 | |
| 167 | while (dev->perf_props.next != &dev->perf_props) { |
| 168 | perf = container_of(dev->perf_props.next, |
| 169 | struct kfd_perf_properties, list); |
| 170 | list_del(&perf->list); |
| 171 | kfree(perf); |
| 172 | } |
| 173 | |
| 174 | kfree(dev); |
| 175 | } |
| 176 | |
| 177 | void kfd_release_topology_device_list(struct list_head *device_list) |
| 178 | { |
| 179 | struct kfd_topology_device *dev; |
| 180 | |
| 181 | while (!list_empty(device_list)) { |
| 182 | dev = list_first_entry(device_list, |
| 183 | struct kfd_topology_device, list); |
| 184 | kfd_release_topology_device(dev); |
| 185 | } |
| 186 | } |
| 187 | |
| 188 | static void kfd_release_live_view(void) |
| 189 | { |
| 190 | kfd_release_topology_device_list(&topology_device_list); |
| 191 | memset(&sys_props, 0, sizeof(sys_props)); |
| 192 | } |
| 193 | |
| 194 | struct kfd_topology_device *kfd_create_topology_device( |
| 195 | struct list_head *device_list) |
| 196 | { |
| 197 | struct kfd_topology_device *dev; |
| 198 | |
| 199 | dev = kfd_alloc_struct(dev); |
| 200 | if (!dev) { |
| 201 | pr_err("No memory to allocate a topology device"); |
| 202 | return NULL; |
| 203 | } |
| 204 | |
| 205 | INIT_LIST_HEAD(&dev->mem_props); |
| 206 | INIT_LIST_HEAD(&dev->cache_props); |
| 207 | INIT_LIST_HEAD(&dev->io_link_props); |
| 208 | INIT_LIST_HEAD(&dev->p2p_link_props); |
| 209 | INIT_LIST_HEAD(&dev->perf_props); |
| 210 | |
| 211 | list_add_tail(&dev->list, device_list); |
| 212 | |
| 213 | return dev; |
| 214 | } |
| 215 | |
| 216 | |
| 217 | #define sysfs_show_gen_prop(buffer, offs, fmt, ...) \ |
| 218 | (offs += snprintf(buffer+offs, PAGE_SIZE-offs, \ |
| 219 | fmt, __VA_ARGS__)) |
| 220 | #define sysfs_show_32bit_prop(buffer, offs, name, value) \ |
| 221 | sysfs_show_gen_prop(buffer, offs, "%s %u\n", name, value) |
| 222 | #define sysfs_show_64bit_prop(buffer, offs, name, value) \ |
| 223 | sysfs_show_gen_prop(buffer, offs, "%s %llu\n", name, value) |
| 224 | #define sysfs_show_32bit_val(buffer, offs, value) \ |
| 225 | sysfs_show_gen_prop(buffer, offs, "%u\n", value) |
| 226 | #define sysfs_show_str_val(buffer, offs, value) \ |
| 227 | sysfs_show_gen_prop(buffer, offs, "%s\n", value) |
| 228 | |
| 229 | static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr, |
| 230 | char *buffer) |
| 231 | { |
| 232 | int offs = 0; |
| 233 | |
| 234 | /* Making sure that the buffer is an empty string */ |
| 235 | buffer[0] = 0; |
| 236 | |
| 237 | if (attr == &sys_props.attr_genid) { |
| 238 | sysfs_show_32bit_val(buffer, offs, |
| 239 | sys_props.generation_count); |
| 240 | } else if (attr == &sys_props.attr_props) { |
| 241 | sysfs_show_64bit_prop(buffer, offs, "platform_oem", |
| 242 | sys_props.platform_oem); |
| 243 | sysfs_show_64bit_prop(buffer, offs, "platform_id", |
| 244 | sys_props.platform_id); |
| 245 | sysfs_show_64bit_prop(buffer, offs, "platform_rev", |
| 246 | sys_props.platform_rev); |
| 247 | } else { |
| 248 | offs = -EINVAL; |
| 249 | } |
| 250 | |
| 251 | return offs; |
| 252 | } |
| 253 | |
| 254 | static void kfd_topology_kobj_release(struct kobject *kobj) |
| 255 | { |
| 256 | kfree(kobj); |
| 257 | } |
| 258 | |
| 259 | static const struct sysfs_ops sysprops_ops = { |
| 260 | .show = sysprops_show, |
| 261 | }; |
| 262 | |
| 263 | static const struct kobj_type sysprops_type = { |
| 264 | .release = kfd_topology_kobj_release, |
| 265 | .sysfs_ops = &sysprops_ops, |
| 266 | }; |
| 267 | |
| 268 | static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr, |
| 269 | char *buffer) |
| 270 | { |
| 271 | int offs = 0; |
| 272 | struct kfd_iolink_properties *iolink; |
| 273 | |
| 274 | /* Making sure that the buffer is an empty string */ |
| 275 | buffer[0] = 0; |
| 276 | |
| 277 | iolink = container_of(attr, struct kfd_iolink_properties, attr); |
| 278 | if (iolink->gpu && kfd_devcgroup_check_permission(iolink->gpu)) |
| 279 | return -EPERM; |
| 280 | sysfs_show_32bit_prop(buffer, offs, "type", iolink->iolink_type); |
| 281 | sysfs_show_32bit_prop(buffer, offs, "version_major", iolink->ver_maj); |
| 282 | sysfs_show_32bit_prop(buffer, offs, "version_minor", iolink->ver_min); |
| 283 | sysfs_show_32bit_prop(buffer, offs, "node_from", iolink->node_from); |
| 284 | sysfs_show_32bit_prop(buffer, offs, "node_to", iolink->node_to); |
| 285 | sysfs_show_32bit_prop(buffer, offs, "weight", iolink->weight); |
| 286 | sysfs_show_32bit_prop(buffer, offs, "min_latency", iolink->min_latency); |
| 287 | sysfs_show_32bit_prop(buffer, offs, "max_latency", iolink->max_latency); |
| 288 | sysfs_show_32bit_prop(buffer, offs, "min_bandwidth", |
| 289 | iolink->min_bandwidth); |
| 290 | sysfs_show_32bit_prop(buffer, offs, "max_bandwidth", |
| 291 | iolink->max_bandwidth); |
| 292 | sysfs_show_32bit_prop(buffer, offs, "recommended_transfer_size", |
| 293 | iolink->rec_transfer_size); |
| 294 | sysfs_show_32bit_prop(buffer, offs, "flags", iolink->flags); |
| 295 | |
| 296 | return offs; |
| 297 | } |
| 298 | |
| 299 | static const struct sysfs_ops iolink_ops = { |
| 300 | .show = iolink_show, |
| 301 | }; |
| 302 | |
| 303 | static const struct kobj_type iolink_type = { |
| 304 | .release = kfd_topology_kobj_release, |
| 305 | .sysfs_ops = &iolink_ops, |
| 306 | }; |
| 307 | |
| 308 | static ssize_t mem_show(struct kobject *kobj, struct attribute *attr, |
| 309 | char *buffer) |
| 310 | { |
| 311 | int offs = 0; |
| 312 | struct kfd_mem_properties *mem; |
| 313 | |
| 314 | /* Making sure that the buffer is an empty string */ |
| 315 | buffer[0] = 0; |
| 316 | |
| 317 | mem = container_of(attr, struct kfd_mem_properties, attr); |
| 318 | if (mem->gpu && kfd_devcgroup_check_permission(mem->gpu)) |
| 319 | return -EPERM; |
| 320 | sysfs_show_32bit_prop(buffer, offs, "heap_type", mem->heap_type); |
| 321 | sysfs_show_64bit_prop(buffer, offs, "size_in_bytes", |
| 322 | mem->size_in_bytes); |
| 323 | sysfs_show_32bit_prop(buffer, offs, "flags", mem->flags); |
| 324 | sysfs_show_32bit_prop(buffer, offs, "width", mem->width); |
| 325 | sysfs_show_32bit_prop(buffer, offs, "mem_clk_max", |
| 326 | mem->mem_clk_max); |
| 327 | |
| 328 | return offs; |
| 329 | } |
| 330 | |
| 331 | static const struct sysfs_ops mem_ops = { |
| 332 | .show = mem_show, |
| 333 | }; |
| 334 | |
| 335 | static const struct kobj_type mem_type = { |
| 336 | .release = kfd_topology_kobj_release, |
| 337 | .sysfs_ops = &mem_ops, |
| 338 | }; |
| 339 | |
| 340 | static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr, |
| 341 | char *buffer) |
| 342 | { |
| 343 | int offs = 0; |
| 344 | uint32_t i, j; |
| 345 | struct kfd_cache_properties *cache; |
| 346 | |
| 347 | /* Making sure that the buffer is an empty string */ |
| 348 | buffer[0] = 0; |
| 349 | cache = container_of(attr, struct kfd_cache_properties, attr); |
| 350 | if (cache->gpu && kfd_devcgroup_check_permission(cache->gpu)) |
| 351 | return -EPERM; |
| 352 | sysfs_show_32bit_prop(buffer, offs, "processor_id_low", |
| 353 | cache->processor_id_low); |
| 354 | sysfs_show_32bit_prop(buffer, offs, "level", cache->cache_level); |
| 355 | sysfs_show_32bit_prop(buffer, offs, "size", cache->cache_size); |
| 356 | sysfs_show_32bit_prop(buffer, offs, "cache_line_size", |
| 357 | cache->cacheline_size); |
| 358 | sysfs_show_32bit_prop(buffer, offs, "cache_lines_per_tag", |
| 359 | cache->cachelines_per_tag); |
| 360 | sysfs_show_32bit_prop(buffer, offs, "association", cache->cache_assoc); |
| 361 | sysfs_show_32bit_prop(buffer, offs, "latency", cache->cache_latency); |
| 362 | sysfs_show_32bit_prop(buffer, offs, "type", cache->cache_type); |
| 363 | |
| 364 | offs += snprintf(buffer+offs, PAGE_SIZE-offs, "sibling_map "); |
| 365 | for (i = 0; i < cache->sibling_map_size; i++) |
| 366 | for (j = 0; j < sizeof(cache->sibling_map[0])*8; j++) |
| 367 | /* Check each bit */ |
| 368 | offs += snprintf(buffer+offs, PAGE_SIZE-offs, "%d,", |
| 369 | (cache->sibling_map[i] >> j) & 1); |
| 370 | |
| 371 | /* Replace the last "," with end of line */ |
| 372 | buffer[offs-1] = '\n'; |
| 373 | return offs; |
| 374 | } |
| 375 | |
| 376 | static const struct sysfs_ops cache_ops = { |
| 377 | .show = kfd_cache_show, |
| 378 | }; |
| 379 | |
| 380 | static const struct kobj_type cache_type = { |
| 381 | .release = kfd_topology_kobj_release, |
| 382 | .sysfs_ops = &cache_ops, |
| 383 | }; |
| 384 | |
| 385 | /****** Sysfs of Performance Counters ******/ |
| 386 | |
| 387 | struct kfd_perf_attr { |
| 388 | struct kobj_attribute attr; |
| 389 | uint32_t data; |
| 390 | }; |
| 391 | |
| 392 | static ssize_t perf_show(struct kobject *kobj, struct kobj_attribute *attrs, |
| 393 | char *buf) |
| 394 | { |
| 395 | int offs = 0; |
| 396 | struct kfd_perf_attr *attr; |
| 397 | |
| 398 | buf[0] = 0; |
| 399 | attr = container_of(attrs, struct kfd_perf_attr, attr); |
| 400 | if (!attr->data) /* invalid data for PMC */ |
| 401 | return 0; |
| 402 | else |
| 403 | return sysfs_show_32bit_val(buf, offs, attr->data); |
| 404 | } |
| 405 | |
| 406 | #define KFD_PERF_DESC(_name, _data) \ |
| 407 | { \ |
| 408 | .attr = __ATTR(_name, 0444, perf_show, NULL), \ |
| 409 | .data = _data, \ |
| 410 | } |
| 411 | |
| 412 | static struct kfd_perf_attr perf_attr_iommu[] = { |
| 413 | KFD_PERF_DESC(max_concurrent, 0), |
| 414 | KFD_PERF_DESC(num_counters, 0), |
| 415 | KFD_PERF_DESC(counter_ids, 0), |
| 416 | }; |
| 417 | /****************************************/ |
| 418 | |
| 419 | static ssize_t node_show(struct kobject *kobj, struct attribute *attr, |
| 420 | char *buffer) |
| 421 | { |
| 422 | int offs = 0; |
| 423 | struct kfd_topology_device *dev; |
| 424 | uint32_t log_max_watch_addr; |
| 425 | |
| 426 | /* Making sure that the buffer is an empty string */ |
| 427 | buffer[0] = 0; |
| 428 | |
| 429 | if (strcmp(attr->name, "gpu_id") == 0) { |
| 430 | dev = container_of(attr, struct kfd_topology_device, |
| 431 | attr_gpuid); |
| 432 | if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu)) |
| 433 | return -EPERM; |
| 434 | return sysfs_show_32bit_val(buffer, offs, dev->gpu_id); |
| 435 | } |
| 436 | |
| 437 | if (strcmp(attr->name, "name") == 0) { |
| 438 | dev = container_of(attr, struct kfd_topology_device, |
| 439 | attr_name); |
| 440 | |
| 441 | if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu)) |
| 442 | return -EPERM; |
| 443 | return sysfs_show_str_val(buffer, offs, dev->node_props.name); |
| 444 | } |
| 445 | |
| 446 | dev = container_of(attr, struct kfd_topology_device, |
| 447 | attr_props); |
| 448 | if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu)) |
| 449 | return -EPERM; |
| 450 | sysfs_show_32bit_prop(buffer, offs, "cpu_cores_count", |
| 451 | dev->node_props.cpu_cores_count); |
| 452 | sysfs_show_32bit_prop(buffer, offs, "simd_count", |
| 453 | dev->gpu ? dev->node_props.simd_count : 0); |
| 454 | sysfs_show_32bit_prop(buffer, offs, "mem_banks_count", |
| 455 | dev->node_props.mem_banks_count); |
| 456 | sysfs_show_32bit_prop(buffer, offs, "caches_count", |
| 457 | dev->node_props.caches_count); |
| 458 | sysfs_show_32bit_prop(buffer, offs, "io_links_count", |
| 459 | dev->node_props.io_links_count); |
| 460 | sysfs_show_32bit_prop(buffer, offs, "p2p_links_count", |
| 461 | dev->node_props.p2p_links_count); |
| 462 | sysfs_show_32bit_prop(buffer, offs, "cpu_core_id_base", |
| 463 | dev->node_props.cpu_core_id_base); |
| 464 | sysfs_show_32bit_prop(buffer, offs, "simd_id_base", |
| 465 | dev->node_props.simd_id_base); |
| 466 | sysfs_show_32bit_prop(buffer, offs, "max_waves_per_simd", |
| 467 | dev->node_props.max_waves_per_simd); |
| 468 | sysfs_show_32bit_prop(buffer, offs, "lds_size_in_kb", |
| 469 | dev->node_props.lds_size_in_kb); |
| 470 | sysfs_show_32bit_prop(buffer, offs, "gds_size_in_kb", |
| 471 | dev->node_props.gds_size_in_kb); |
| 472 | sysfs_show_32bit_prop(buffer, offs, "num_gws", |
| 473 | dev->node_props.num_gws); |
| 474 | sysfs_show_32bit_prop(buffer, offs, "wave_front_size", |
| 475 | dev->node_props.wave_front_size); |
| 476 | sysfs_show_32bit_prop(buffer, offs, "array_count", |
| 477 | dev->gpu ? (dev->node_props.array_count * |
| 478 | NUM_XCC(dev->gpu->xcc_mask)) : 0); |
| 479 | sysfs_show_32bit_prop(buffer, offs, "simd_arrays_per_engine", |
| 480 | dev->node_props.simd_arrays_per_engine); |
| 481 | sysfs_show_32bit_prop(buffer, offs, "cu_per_simd_array", |
| 482 | dev->node_props.cu_per_simd_array); |
| 483 | sysfs_show_32bit_prop(buffer, offs, "simd_per_cu", |
| 484 | dev->node_props.simd_per_cu); |
| 485 | sysfs_show_32bit_prop(buffer, offs, "max_slots_scratch_cu", |
| 486 | dev->node_props.max_slots_scratch_cu); |
| 487 | sysfs_show_32bit_prop(buffer, offs, "gfx_target_version", |
| 488 | dev->node_props.gfx_target_version); |
| 489 | sysfs_show_32bit_prop(buffer, offs, "vendor_id", |
| 490 | dev->node_props.vendor_id); |
| 491 | sysfs_show_32bit_prop(buffer, offs, "device_id", |
| 492 | dev->node_props.device_id); |
| 493 | sysfs_show_32bit_prop(buffer, offs, "location_id", |
| 494 | dev->node_props.location_id); |
| 495 | sysfs_show_32bit_prop(buffer, offs, "domain", |
| 496 | dev->node_props.domain); |
| 497 | sysfs_show_32bit_prop(buffer, offs, "drm_render_minor", |
| 498 | dev->node_props.drm_render_minor); |
| 499 | sysfs_show_64bit_prop(buffer, offs, "hive_id", |
| 500 | dev->node_props.hive_id); |
| 501 | sysfs_show_32bit_prop(buffer, offs, "num_sdma_engines", |
| 502 | dev->node_props.num_sdma_engines); |
| 503 | sysfs_show_32bit_prop(buffer, offs, "num_sdma_xgmi_engines", |
| 504 | dev->node_props.num_sdma_xgmi_engines); |
| 505 | sysfs_show_32bit_prop(buffer, offs, "num_sdma_queues_per_engine", |
| 506 | dev->node_props.num_sdma_queues_per_engine); |
| 507 | sysfs_show_32bit_prop(buffer, offs, "num_cp_queues", |
| 508 | dev->node_props.num_cp_queues); |
| 509 | |
| 510 | if (dev->gpu) { |
| 511 | log_max_watch_addr = |
| 512 | __ilog2_u32(dev->gpu->kfd->device_info.num_of_watch_points); |
| 513 | |
| 514 | if (log_max_watch_addr) { |
| 515 | dev->node_props.capability |= |
| 516 | HSA_CAP_WATCH_POINTS_SUPPORTED; |
| 517 | |
| 518 | dev->node_props.capability |= |
| 519 | ((log_max_watch_addr << |
| 520 | HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) & |
| 521 | HSA_CAP_WATCH_POINTS_TOTALBITS_MASK); |
| 522 | } |
| 523 | |
| 524 | if (dev->gpu->adev->asic_type == CHIP_TONGA) |
| 525 | dev->node_props.capability |= |
| 526 | HSA_CAP_AQL_QUEUE_DOUBLE_MAP; |
| 527 | |
| 528 | sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_fcompute", |
| 529 | dev->node_props.max_engine_clk_fcompute); |
| 530 | |
| 531 | sysfs_show_64bit_prop(buffer, offs, "local_mem_size", 0ULL); |
| 532 | |
| 533 | sysfs_show_32bit_prop(buffer, offs, "fw_version", |
| 534 | dev->gpu->kfd->mec_fw_version); |
| 535 | sysfs_show_32bit_prop(buffer, offs, "capability", |
| 536 | dev->node_props.capability); |
| 537 | sysfs_show_64bit_prop(buffer, offs, "debug_prop", |
| 538 | dev->node_props.debug_prop); |
| 539 | sysfs_show_32bit_prop(buffer, offs, "sdma_fw_version", |
| 540 | dev->gpu->kfd->sdma_fw_version); |
| 541 | sysfs_show_64bit_prop(buffer, offs, "unique_id", |
| 542 | dev->gpu->adev->unique_id); |
| 543 | sysfs_show_32bit_prop(buffer, offs, "num_xcc", |
| 544 | NUM_XCC(dev->gpu->xcc_mask)); |
| 545 | } |
| 546 | |
| 547 | return sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_ccompute", |
| 548 | cpufreq_quick_get_max(0)/1000); |
| 549 | } |
| 550 | |
| 551 | static const struct sysfs_ops node_ops = { |
| 552 | .show = node_show, |
| 553 | }; |
| 554 | |
| 555 | static const struct kobj_type node_type = { |
| 556 | .release = kfd_topology_kobj_release, |
| 557 | .sysfs_ops = &node_ops, |
| 558 | }; |
| 559 | |
| 560 | static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr) |
| 561 | { |
| 562 | sysfs_remove_file(kobj, attr); |
| 563 | kobject_del(kobj); |
| 564 | kobject_put(kobj); |
| 565 | } |
| 566 | |
| 567 | static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev) |
| 568 | { |
| 569 | struct kfd_iolink_properties *p2plink; |
| 570 | struct kfd_iolink_properties *iolink; |
| 571 | struct kfd_cache_properties *cache; |
| 572 | struct kfd_mem_properties *mem; |
| 573 | struct kfd_perf_properties *perf; |
| 574 | |
| 575 | if (dev->kobj_iolink) { |
| 576 | list_for_each_entry(iolink, &dev->io_link_props, list) |
| 577 | if (iolink->kobj) { |
| 578 | kfd_remove_sysfs_file(iolink->kobj, |
| 579 | &iolink->attr); |
| 580 | iolink->kobj = NULL; |
| 581 | } |
| 582 | kobject_del(dev->kobj_iolink); |
| 583 | kobject_put(dev->kobj_iolink); |
| 584 | dev->kobj_iolink = NULL; |
| 585 | } |
| 586 | |
| 587 | if (dev->kobj_p2plink) { |
| 588 | list_for_each_entry(p2plink, &dev->p2p_link_props, list) |
| 589 | if (p2plink->kobj) { |
| 590 | kfd_remove_sysfs_file(p2plink->kobj, |
| 591 | &p2plink->attr); |
| 592 | p2plink->kobj = NULL; |
| 593 | } |
| 594 | kobject_del(dev->kobj_p2plink); |
| 595 | kobject_put(dev->kobj_p2plink); |
| 596 | dev->kobj_p2plink = NULL; |
| 597 | } |
| 598 | |
| 599 | if (dev->kobj_cache) { |
| 600 | list_for_each_entry(cache, &dev->cache_props, list) |
| 601 | if (cache->kobj) { |
| 602 | kfd_remove_sysfs_file(cache->kobj, |
| 603 | &cache->attr); |
| 604 | cache->kobj = NULL; |
| 605 | } |
| 606 | kobject_del(dev->kobj_cache); |
| 607 | kobject_put(dev->kobj_cache); |
| 608 | dev->kobj_cache = NULL; |
| 609 | } |
| 610 | |
| 611 | if (dev->kobj_mem) { |
| 612 | list_for_each_entry(mem, &dev->mem_props, list) |
| 613 | if (mem->kobj) { |
| 614 | kfd_remove_sysfs_file(mem->kobj, &mem->attr); |
| 615 | mem->kobj = NULL; |
| 616 | } |
| 617 | kobject_del(dev->kobj_mem); |
| 618 | kobject_put(dev->kobj_mem); |
| 619 | dev->kobj_mem = NULL; |
| 620 | } |
| 621 | |
| 622 | if (dev->kobj_perf) { |
| 623 | list_for_each_entry(perf, &dev->perf_props, list) { |
| 624 | kfree(perf->attr_group); |
| 625 | perf->attr_group = NULL; |
| 626 | } |
| 627 | kobject_del(dev->kobj_perf); |
| 628 | kobject_put(dev->kobj_perf); |
| 629 | dev->kobj_perf = NULL; |
| 630 | } |
| 631 | |
| 632 | if (dev->kobj_node) { |
| 633 | sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid); |
| 634 | sysfs_remove_file(dev->kobj_node, &dev->attr_name); |
| 635 | sysfs_remove_file(dev->kobj_node, &dev->attr_props); |
| 636 | kobject_del(dev->kobj_node); |
| 637 | kobject_put(dev->kobj_node); |
| 638 | dev->kobj_node = NULL; |
| 639 | } |
| 640 | } |
| 641 | |
| 642 | static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev, |
| 643 | uint32_t id) |
| 644 | { |
| 645 | struct kfd_iolink_properties *p2plink; |
| 646 | struct kfd_iolink_properties *iolink; |
| 647 | struct kfd_cache_properties *cache; |
| 648 | struct kfd_mem_properties *mem; |
| 649 | struct kfd_perf_properties *perf; |
| 650 | int ret; |
| 651 | uint32_t i, num_attrs; |
| 652 | struct attribute **attrs; |
| 653 | |
| 654 | if (WARN_ON(dev->kobj_node)) |
| 655 | return -EEXIST; |
| 656 | |
| 657 | /* |
| 658 | * Creating the sysfs folders |
| 659 | */ |
| 660 | dev->kobj_node = kfd_alloc_struct(dev->kobj_node); |
| 661 | if (!dev->kobj_node) |
| 662 | return -ENOMEM; |
| 663 | |
| 664 | ret = kobject_init_and_add(dev->kobj_node, &node_type, |
| 665 | sys_props.kobj_nodes, "%d", id); |
| 666 | if (ret < 0) { |
| 667 | kobject_put(dev->kobj_node); |
| 668 | return ret; |
| 669 | } |
| 670 | |
| 671 | dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node); |
| 672 | if (!dev->kobj_mem) |
| 673 | return -ENOMEM; |
| 674 | |
| 675 | dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node); |
| 676 | if (!dev->kobj_cache) |
| 677 | return -ENOMEM; |
| 678 | |
| 679 | dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node); |
| 680 | if (!dev->kobj_iolink) |
| 681 | return -ENOMEM; |
| 682 | |
| 683 | dev->kobj_p2plink = kobject_create_and_add("p2p_links", dev->kobj_node); |
| 684 | if (!dev->kobj_p2plink) |
| 685 | return -ENOMEM; |
| 686 | |
| 687 | dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node); |
| 688 | if (!dev->kobj_perf) |
| 689 | return -ENOMEM; |
| 690 | |
| 691 | /* |
| 692 | * Creating sysfs files for node properties |
| 693 | */ |
| 694 | dev->attr_gpuid.name = "gpu_id"; |
| 695 | dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE; |
| 696 | sysfs_attr_init(&dev->attr_gpuid); |
| 697 | dev->attr_name.name = "name"; |
| 698 | dev->attr_name.mode = KFD_SYSFS_FILE_MODE; |
| 699 | sysfs_attr_init(&dev->attr_name); |
| 700 | dev->attr_props.name = "properties"; |
| 701 | dev->attr_props.mode = KFD_SYSFS_FILE_MODE; |
| 702 | sysfs_attr_init(&dev->attr_props); |
| 703 | ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid); |
| 704 | if (ret < 0) |
| 705 | return ret; |
| 706 | ret = sysfs_create_file(dev->kobj_node, &dev->attr_name); |
| 707 | if (ret < 0) |
| 708 | return ret; |
| 709 | ret = sysfs_create_file(dev->kobj_node, &dev->attr_props); |
| 710 | if (ret < 0) |
| 711 | return ret; |
| 712 | |
| 713 | i = 0; |
| 714 | list_for_each_entry(mem, &dev->mem_props, list) { |
| 715 | mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); |
| 716 | if (!mem->kobj) |
| 717 | return -ENOMEM; |
| 718 | ret = kobject_init_and_add(mem->kobj, &mem_type, |
| 719 | dev->kobj_mem, "%d", i); |
| 720 | if (ret < 0) { |
| 721 | kobject_put(mem->kobj); |
| 722 | return ret; |
| 723 | } |
| 724 | |
| 725 | mem->attr.name = "properties"; |
| 726 | mem->attr.mode = KFD_SYSFS_FILE_MODE; |
| 727 | sysfs_attr_init(&mem->attr); |
| 728 | ret = sysfs_create_file(mem->kobj, &mem->attr); |
| 729 | if (ret < 0) |
| 730 | return ret; |
| 731 | i++; |
| 732 | } |
| 733 | |
| 734 | i = 0; |
| 735 | list_for_each_entry(cache, &dev->cache_props, list) { |
| 736 | cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); |
| 737 | if (!cache->kobj) |
| 738 | return -ENOMEM; |
| 739 | ret = kobject_init_and_add(cache->kobj, &cache_type, |
| 740 | dev->kobj_cache, "%d", i); |
| 741 | if (ret < 0) { |
| 742 | kobject_put(cache->kobj); |
| 743 | return ret; |
| 744 | } |
| 745 | |
| 746 | cache->attr.name = "properties"; |
| 747 | cache->attr.mode = KFD_SYSFS_FILE_MODE; |
| 748 | sysfs_attr_init(&cache->attr); |
| 749 | ret = sysfs_create_file(cache->kobj, &cache->attr); |
| 750 | if (ret < 0) |
| 751 | return ret; |
| 752 | i++; |
| 753 | } |
| 754 | |
| 755 | i = 0; |
| 756 | list_for_each_entry(iolink, &dev->io_link_props, list) { |
| 757 | iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); |
| 758 | if (!iolink->kobj) |
| 759 | return -ENOMEM; |
| 760 | ret = kobject_init_and_add(iolink->kobj, &iolink_type, |
| 761 | dev->kobj_iolink, "%d", i); |
| 762 | if (ret < 0) { |
| 763 | kobject_put(iolink->kobj); |
| 764 | return ret; |
| 765 | } |
| 766 | |
| 767 | iolink->attr.name = "properties"; |
| 768 | iolink->attr.mode = KFD_SYSFS_FILE_MODE; |
| 769 | sysfs_attr_init(&iolink->attr); |
| 770 | ret = sysfs_create_file(iolink->kobj, &iolink->attr); |
| 771 | if (ret < 0) |
| 772 | return ret; |
| 773 | i++; |
| 774 | } |
| 775 | |
| 776 | i = 0; |
| 777 | list_for_each_entry(p2plink, &dev->p2p_link_props, list) { |
| 778 | p2plink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); |
| 779 | if (!p2plink->kobj) |
| 780 | return -ENOMEM; |
| 781 | ret = kobject_init_and_add(p2plink->kobj, &iolink_type, |
| 782 | dev->kobj_p2plink, "%d", i); |
| 783 | if (ret < 0) { |
| 784 | kobject_put(p2plink->kobj); |
| 785 | return ret; |
| 786 | } |
| 787 | |
| 788 | p2plink->attr.name = "properties"; |
| 789 | p2plink->attr.mode = KFD_SYSFS_FILE_MODE; |
| 790 | sysfs_attr_init(&p2plink->attr); |
| 791 | ret = sysfs_create_file(p2plink->kobj, &p2plink->attr); |
| 792 | if (ret < 0) |
| 793 | return ret; |
| 794 | i++; |
| 795 | } |
| 796 | |
| 797 | /* All hardware blocks have the same number of attributes. */ |
| 798 | num_attrs = ARRAY_SIZE(perf_attr_iommu); |
| 799 | list_for_each_entry(perf, &dev->perf_props, list) { |
| 800 | perf->attr_group = kzalloc(sizeof(struct kfd_perf_attr) |
| 801 | * num_attrs + sizeof(struct attribute_group), |
| 802 | GFP_KERNEL); |
| 803 | if (!perf->attr_group) |
| 804 | return -ENOMEM; |
| 805 | |
| 806 | attrs = (struct attribute **)(perf->attr_group + 1); |
| 807 | if (!strcmp(perf->block_name, "iommu")) { |
| 808 | /* Information of IOMMU's num_counters and counter_ids is shown |
| 809 | * under /sys/bus/event_source/devices/amd_iommu. We don't |
| 810 | * duplicate here. |
| 811 | */ |
| 812 | perf_attr_iommu[0].data = perf->max_concurrent; |
| 813 | for (i = 0; i < num_attrs; i++) |
| 814 | attrs[i] = &perf_attr_iommu[i].attr.attr; |
| 815 | } |
| 816 | perf->attr_group->name = perf->block_name; |
| 817 | perf->attr_group->attrs = attrs; |
| 818 | ret = sysfs_create_group(dev->kobj_perf, perf->attr_group); |
| 819 | if (ret < 0) |
| 820 | return ret; |
| 821 | } |
| 822 | |
| 823 | return 0; |
| 824 | } |
| 825 | |
| 826 | /* Called with write topology lock acquired */ |
| 827 | static int kfd_build_sysfs_node_tree(void) |
| 828 | { |
| 829 | struct kfd_topology_device *dev; |
| 830 | int ret; |
| 831 | uint32_t i = 0; |
| 832 | |
| 833 | list_for_each_entry(dev, &topology_device_list, list) { |
| 834 | ret = kfd_build_sysfs_node_entry(dev, i); |
| 835 | if (ret < 0) |
| 836 | return ret; |
| 837 | i++; |
| 838 | } |
| 839 | |
| 840 | return 0; |
| 841 | } |
| 842 | |
| 843 | /* Called with write topology lock acquired */ |
| 844 | static void kfd_remove_sysfs_node_tree(void) |
| 845 | { |
| 846 | struct kfd_topology_device *dev; |
| 847 | |
| 848 | list_for_each_entry(dev, &topology_device_list, list) |
| 849 | kfd_remove_sysfs_node_entry(dev); |
| 850 | } |
| 851 | |
| 852 | static int kfd_topology_update_sysfs(void) |
| 853 | { |
| 854 | int ret; |
| 855 | |
| 856 | if (!sys_props.kobj_topology) { |
| 857 | sys_props.kobj_topology = |
| 858 | kfd_alloc_struct(sys_props.kobj_topology); |
| 859 | if (!sys_props.kobj_topology) |
| 860 | return -ENOMEM; |
| 861 | |
| 862 | ret = kobject_init_and_add(sys_props.kobj_topology, |
| 863 | &sysprops_type, &kfd_device->kobj, |
| 864 | "topology"); |
| 865 | if (ret < 0) { |
| 866 | kobject_put(sys_props.kobj_topology); |
| 867 | return ret; |
| 868 | } |
| 869 | |
| 870 | sys_props.kobj_nodes = kobject_create_and_add("nodes", |
| 871 | sys_props.kobj_topology); |
| 872 | if (!sys_props.kobj_nodes) |
| 873 | return -ENOMEM; |
| 874 | |
| 875 | sys_props.attr_genid.name = "generation_id"; |
| 876 | sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE; |
| 877 | sysfs_attr_init(&sys_props.attr_genid); |
| 878 | ret = sysfs_create_file(sys_props.kobj_topology, |
| 879 | &sys_props.attr_genid); |
| 880 | if (ret < 0) |
| 881 | return ret; |
| 882 | |
| 883 | sys_props.attr_props.name = "system_properties"; |
| 884 | sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE; |
| 885 | sysfs_attr_init(&sys_props.attr_props); |
| 886 | ret = sysfs_create_file(sys_props.kobj_topology, |
| 887 | &sys_props.attr_props); |
| 888 | if (ret < 0) |
| 889 | return ret; |
| 890 | } |
| 891 | |
| 892 | kfd_remove_sysfs_node_tree(); |
| 893 | |
| 894 | return kfd_build_sysfs_node_tree(); |
| 895 | } |
| 896 | |
| 897 | static void kfd_topology_release_sysfs(void) |
| 898 | { |
| 899 | kfd_remove_sysfs_node_tree(); |
| 900 | if (sys_props.kobj_topology) { |
| 901 | sysfs_remove_file(sys_props.kobj_topology, |
| 902 | &sys_props.attr_genid); |
| 903 | sysfs_remove_file(sys_props.kobj_topology, |
| 904 | &sys_props.attr_props); |
| 905 | if (sys_props.kobj_nodes) { |
| 906 | kobject_del(sys_props.kobj_nodes); |
| 907 | kobject_put(sys_props.kobj_nodes); |
| 908 | sys_props.kobj_nodes = NULL; |
| 909 | } |
| 910 | kobject_del(sys_props.kobj_topology); |
| 911 | kobject_put(sys_props.kobj_topology); |
| 912 | sys_props.kobj_topology = NULL; |
| 913 | } |
| 914 | } |
| 915 | |
| 916 | /* Called with write topology_lock acquired */ |
| 917 | static void kfd_topology_update_device_list(struct list_head *temp_list, |
| 918 | struct list_head *master_list) |
| 919 | { |
| 920 | while (!list_empty(temp_list)) { |
| 921 | list_move_tail(temp_list->next, master_list); |
| 922 | sys_props.num_devices++; |
| 923 | } |
| 924 | } |
| 925 | |
| 926 | static void kfd_debug_print_topology(void) |
| 927 | { |
| 928 | struct kfd_topology_device *dev; |
| 929 | |
| 930 | down_read(&topology_lock); |
| 931 | |
| 932 | dev = list_last_entry(&topology_device_list, |
| 933 | struct kfd_topology_device, list); |
| 934 | if (dev) { |
| 935 | if (dev->node_props.cpu_cores_count && |
| 936 | dev->node_props.simd_count) { |
| 937 | pr_info("Topology: Add APU node [0x%0x:0x%0x]\n", |
| 938 | dev->node_props.device_id, |
| 939 | dev->node_props.vendor_id); |
| 940 | } else if (dev->node_props.cpu_cores_count) |
| 941 | pr_info("Topology: Add CPU node\n"); |
| 942 | else if (dev->node_props.simd_count) |
| 943 | pr_info("Topology: Add dGPU node [0x%0x:0x%0x]\n", |
| 944 | dev->node_props.device_id, |
| 945 | dev->node_props.vendor_id); |
| 946 | } |
| 947 | up_read(&topology_lock); |
| 948 | } |
| 949 | |
| 950 | /* Helper function for intializing platform_xx members of |
| 951 | * kfd_system_properties. Uses OEM info from the last CPU/APU node. |
| 952 | */ |
| 953 | static void kfd_update_system_properties(void) |
| 954 | { |
| 955 | struct kfd_topology_device *dev; |
| 956 | |
| 957 | down_read(&topology_lock); |
| 958 | dev = list_last_entry(&topology_device_list, |
| 959 | struct kfd_topology_device, list); |
| 960 | if (dev) { |
| 961 | sys_props.platform_id = |
| 962 | (*((uint64_t *)dev->oem_id)) & CRAT_OEMID_64BIT_MASK; |
| 963 | sys_props.platform_oem = *((uint64_t *)dev->oem_table_id); |
| 964 | sys_props.platform_rev = dev->oem_revision; |
| 965 | } |
| 966 | up_read(&topology_lock); |
| 967 | } |
| 968 | |
| 969 | static void find_system_memory(const struct dmi_header *dm, |
| 970 | void *private) |
| 971 | { |
| 972 | struct kfd_mem_properties *mem; |
| 973 | u16 mem_width, mem_clock; |
| 974 | struct kfd_topology_device *kdev = |
| 975 | (struct kfd_topology_device *)private; |
| 976 | const u8 *dmi_data = (const u8 *)(dm + 1); |
| 977 | |
| 978 | if (dm->type == DMI_ENTRY_MEM_DEVICE && dm->length >= 0x15) { |
| 979 | mem_width = (u16)(*(const u16 *)(dmi_data + 0x6)); |
| 980 | mem_clock = (u16)(*(const u16 *)(dmi_data + 0x11)); |
| 981 | list_for_each_entry(mem, &kdev->mem_props, list) { |
| 982 | if (mem_width != 0xFFFF && mem_width != 0) |
| 983 | mem->width = mem_width; |
| 984 | if (mem_clock != 0) |
| 985 | mem->mem_clk_max = mem_clock; |
| 986 | } |
| 987 | } |
| 988 | } |
| 989 | |
| 990 | /* kfd_add_non_crat_information - Add information that is not currently |
| 991 | * defined in CRAT but is necessary for KFD topology |
| 992 | * @dev - topology device to which addition info is added |
| 993 | */ |
| 994 | static void kfd_add_non_crat_information(struct kfd_topology_device *kdev) |
| 995 | { |
| 996 | /* Check if CPU only node. */ |
| 997 | if (!kdev->gpu) { |
| 998 | /* Add system memory information */ |
| 999 | dmi_walk(find_system_memory, kdev); |
| 1000 | } |
| 1001 | /* TODO: For GPU node, rearrange code from kfd_topology_add_device */ |
| 1002 | } |
| 1003 | |
| 1004 | int kfd_topology_init(void) |
| 1005 | { |
| 1006 | void *crat_image = NULL; |
| 1007 | size_t image_size = 0; |
| 1008 | int ret; |
| 1009 | struct list_head temp_topology_device_list; |
| 1010 | int cpu_only_node = 0; |
| 1011 | struct kfd_topology_device *kdev; |
| 1012 | int proximity_domain; |
| 1013 | |
| 1014 | /* topology_device_list - Master list of all topology devices |
| 1015 | * temp_topology_device_list - temporary list created while parsing CRAT |
| 1016 | * or VCRAT. Once parsing is complete the contents of list is moved to |
| 1017 | * topology_device_list |
| 1018 | */ |
| 1019 | |
| 1020 | /* Initialize the head for the both the lists */ |
| 1021 | INIT_LIST_HEAD(&topology_device_list); |
| 1022 | INIT_LIST_HEAD(&temp_topology_device_list); |
| 1023 | init_rwsem(&topology_lock); |
| 1024 | |
| 1025 | memset(&sys_props, 0, sizeof(sys_props)); |
| 1026 | |
| 1027 | /* Proximity domains in ACPI CRAT tables start counting at |
| 1028 | * 0. The same should be true for virtual CRAT tables created |
| 1029 | * at this stage. GPUs added later in kfd_topology_add_device |
| 1030 | * use a counter. |
| 1031 | */ |
| 1032 | proximity_domain = 0; |
| 1033 | |
| 1034 | ret = kfd_create_crat_image_virtual(&crat_image, &image_size, |
| 1035 | COMPUTE_UNIT_CPU, NULL, |
| 1036 | proximity_domain); |
| 1037 | cpu_only_node = 1; |
| 1038 | if (ret) { |
| 1039 | pr_err("Error creating VCRAT table for CPU\n"); |
| 1040 | return ret; |
| 1041 | } |
| 1042 | |
| 1043 | ret = kfd_parse_crat_table(crat_image, |
| 1044 | &temp_topology_device_list, |
| 1045 | proximity_domain); |
| 1046 | if (ret) { |
| 1047 | pr_err("Error parsing VCRAT table for CPU\n"); |
| 1048 | goto err; |
| 1049 | } |
| 1050 | |
| 1051 | kdev = list_first_entry(&temp_topology_device_list, |
| 1052 | struct kfd_topology_device, list); |
| 1053 | |
| 1054 | down_write(&topology_lock); |
| 1055 | kfd_topology_update_device_list(&temp_topology_device_list, |
| 1056 | &topology_device_list); |
| 1057 | topology_crat_proximity_domain = sys_props.num_devices-1; |
| 1058 | ret = kfd_topology_update_sysfs(); |
| 1059 | up_write(&topology_lock); |
| 1060 | |
| 1061 | if (!ret) { |
| 1062 | sys_props.generation_count++; |
| 1063 | kfd_update_system_properties(); |
| 1064 | kfd_debug_print_topology(); |
| 1065 | } else |
| 1066 | pr_err("Failed to update topology in sysfs ret=%d\n", ret); |
| 1067 | |
| 1068 | /* For nodes with GPU, this information gets added |
| 1069 | * when GPU is detected (kfd_topology_add_device). |
| 1070 | */ |
| 1071 | if (cpu_only_node) { |
| 1072 | /* Add additional information to CPU only node created above */ |
| 1073 | down_write(&topology_lock); |
| 1074 | kdev = list_first_entry(&topology_device_list, |
| 1075 | struct kfd_topology_device, list); |
| 1076 | up_write(&topology_lock); |
| 1077 | kfd_add_non_crat_information(kdev); |
| 1078 | } |
| 1079 | |
| 1080 | err: |
| 1081 | kfd_destroy_crat_image(crat_image); |
| 1082 | return ret; |
| 1083 | } |
| 1084 | |
| 1085 | void kfd_topology_shutdown(void) |
| 1086 | { |
| 1087 | down_write(&topology_lock); |
| 1088 | kfd_topology_release_sysfs(); |
| 1089 | kfd_release_live_view(); |
| 1090 | up_write(&topology_lock); |
| 1091 | } |
| 1092 | |
| 1093 | static uint32_t kfd_generate_gpu_id(struct kfd_node *gpu) |
| 1094 | { |
| 1095 | uint32_t hashout; |
| 1096 | uint32_t buf[8]; |
| 1097 | uint64_t local_mem_size; |
| 1098 | int i; |
| 1099 | |
| 1100 | if (!gpu) |
| 1101 | return 0; |
| 1102 | |
| 1103 | local_mem_size = gpu->local_mem_info.local_mem_size_private + |
| 1104 | gpu->local_mem_info.local_mem_size_public; |
| 1105 | buf[0] = gpu->adev->pdev->devfn; |
| 1106 | buf[1] = gpu->adev->pdev->subsystem_vendor | |
| 1107 | (gpu->adev->pdev->subsystem_device << 16); |
| 1108 | buf[2] = pci_domain_nr(gpu->adev->pdev->bus); |
| 1109 | buf[3] = gpu->adev->pdev->device; |
| 1110 | buf[4] = gpu->adev->pdev->bus->number; |
| 1111 | buf[5] = lower_32_bits(local_mem_size); |
| 1112 | buf[6] = upper_32_bits(local_mem_size); |
| 1113 | buf[7] = (ffs(gpu->xcc_mask) - 1) | (NUM_XCC(gpu->xcc_mask) << 16); |
| 1114 | |
| 1115 | for (i = 0, hashout = 0; i < 8; i++) |
| 1116 | hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH); |
| 1117 | |
| 1118 | return hashout; |
| 1119 | } |
| 1120 | /* kfd_assign_gpu - Attach @gpu to the correct kfd topology device. If |
| 1121 | * the GPU device is not already present in the topology device |
| 1122 | * list then return NULL. This means a new topology device has to |
| 1123 | * be created for this GPU. |
| 1124 | */ |
| 1125 | static struct kfd_topology_device *kfd_assign_gpu(struct kfd_node *gpu) |
| 1126 | { |
| 1127 | struct kfd_topology_device *dev; |
| 1128 | struct kfd_topology_device *out_dev = NULL; |
| 1129 | struct kfd_mem_properties *mem; |
| 1130 | struct kfd_cache_properties *cache; |
| 1131 | struct kfd_iolink_properties *iolink; |
| 1132 | struct kfd_iolink_properties *p2plink; |
| 1133 | |
| 1134 | list_for_each_entry(dev, &topology_device_list, list) { |
| 1135 | /* Discrete GPUs need their own topology device list |
| 1136 | * entries. Don't assign them to CPU/APU nodes. |
| 1137 | */ |
| 1138 | if (dev->node_props.cpu_cores_count) |
| 1139 | continue; |
| 1140 | |
| 1141 | if (!dev->gpu && (dev->node_props.simd_count > 0)) { |
| 1142 | dev->gpu = gpu; |
| 1143 | out_dev = dev; |
| 1144 | |
| 1145 | list_for_each_entry(mem, &dev->mem_props, list) |
| 1146 | mem->gpu = dev->gpu; |
| 1147 | list_for_each_entry(cache, &dev->cache_props, list) |
| 1148 | cache->gpu = dev->gpu; |
| 1149 | list_for_each_entry(iolink, &dev->io_link_props, list) |
| 1150 | iolink->gpu = dev->gpu; |
| 1151 | list_for_each_entry(p2plink, &dev->p2p_link_props, list) |
| 1152 | p2plink->gpu = dev->gpu; |
| 1153 | break; |
| 1154 | } |
| 1155 | } |
| 1156 | return out_dev; |
| 1157 | } |
| 1158 | |
| 1159 | static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival) |
| 1160 | { |
| 1161 | /* |
| 1162 | * TODO: Generate an event for thunk about the arrival/removal |
| 1163 | * of the GPU |
| 1164 | */ |
| 1165 | } |
| 1166 | |
| 1167 | /* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info, |
| 1168 | * patch this after CRAT parsing. |
| 1169 | */ |
| 1170 | static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev) |
| 1171 | { |
| 1172 | struct kfd_mem_properties *mem; |
| 1173 | struct kfd_local_mem_info local_mem_info; |
| 1174 | |
| 1175 | if (!dev) |
| 1176 | return; |
| 1177 | |
| 1178 | /* Currently, amdgpu driver (amdgpu_mc) deals only with GPUs with |
| 1179 | * single bank of VRAM local memory. |
| 1180 | * for dGPUs - VCRAT reports only one bank of Local Memory |
| 1181 | * for APUs - If CRAT from ACPI reports more than one bank, then |
| 1182 | * all the banks will report the same mem_clk_max information |
| 1183 | */ |
| 1184 | amdgpu_amdkfd_get_local_mem_info(dev->gpu->adev, &local_mem_info, |
| 1185 | dev->gpu->xcp); |
| 1186 | |
| 1187 | list_for_each_entry(mem, &dev->mem_props, list) |
| 1188 | mem->mem_clk_max = local_mem_info.mem_clk_max; |
| 1189 | } |
| 1190 | |
| 1191 | static void kfd_set_iolink_no_atomics(struct kfd_topology_device *dev, |
| 1192 | struct kfd_topology_device *target_gpu_dev, |
| 1193 | struct kfd_iolink_properties *link) |
| 1194 | { |
| 1195 | /* xgmi always supports atomics between links. */ |
| 1196 | if (link->iolink_type == CRAT_IOLINK_TYPE_XGMI) |
| 1197 | return; |
| 1198 | |
| 1199 | /* check pcie support to set cpu(dev) flags for target_gpu_dev link. */ |
| 1200 | if (target_gpu_dev) { |
| 1201 | uint32_t cap; |
| 1202 | |
| 1203 | pcie_capability_read_dword(target_gpu_dev->gpu->adev->pdev, |
| 1204 | PCI_EXP_DEVCAP2, &cap); |
| 1205 | |
| 1206 | if (!(cap & (PCI_EXP_DEVCAP2_ATOMIC_COMP32 | |
| 1207 | PCI_EXP_DEVCAP2_ATOMIC_COMP64))) |
| 1208 | link->flags |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT | |
| 1209 | CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT; |
| 1210 | /* set gpu (dev) flags. */ |
| 1211 | } else { |
| 1212 | if (!dev->gpu->kfd->pci_atomic_requested || |
| 1213 | dev->gpu->adev->asic_type == CHIP_HAWAII) |
| 1214 | link->flags |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT | |
| 1215 | CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT; |
| 1216 | } |
| 1217 | } |
| 1218 | |
| 1219 | static void kfd_set_iolink_non_coherent(struct kfd_topology_device *to_dev, |
| 1220 | struct kfd_iolink_properties *outbound_link, |
| 1221 | struct kfd_iolink_properties *inbound_link) |
| 1222 | { |
| 1223 | /* CPU -> GPU with PCIe */ |
| 1224 | if (!to_dev->gpu && |
| 1225 | inbound_link->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS) |
| 1226 | inbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT; |
| 1227 | |
| 1228 | if (to_dev->gpu) { |
| 1229 | /* GPU <-> GPU with PCIe and |
| 1230 | * Vega20 with XGMI |
| 1231 | */ |
| 1232 | if (inbound_link->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS || |
| 1233 | (inbound_link->iolink_type == CRAT_IOLINK_TYPE_XGMI && |
| 1234 | KFD_GC_VERSION(to_dev->gpu) == IP_VERSION(9, 4, 0))) { |
| 1235 | outbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT; |
| 1236 | inbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT; |
| 1237 | } |
| 1238 | } |
| 1239 | } |
| 1240 | |
| 1241 | static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev) |
| 1242 | { |
| 1243 | struct kfd_iolink_properties *link, *inbound_link; |
| 1244 | struct kfd_topology_device *peer_dev; |
| 1245 | |
| 1246 | if (!dev || !dev->gpu) |
| 1247 | return; |
| 1248 | |
| 1249 | /* GPU only creates direct links so apply flags setting to all */ |
| 1250 | list_for_each_entry(link, &dev->io_link_props, list) { |
| 1251 | link->flags = CRAT_IOLINK_FLAGS_ENABLED; |
| 1252 | kfd_set_iolink_no_atomics(dev, NULL, link); |
| 1253 | peer_dev = kfd_topology_device_by_proximity_domain( |
| 1254 | link->node_to); |
| 1255 | |
| 1256 | if (!peer_dev) |
| 1257 | continue; |
| 1258 | |
| 1259 | /* Include the CPU peer in GPU hive if connected over xGMI. */ |
| 1260 | if (!peer_dev->gpu && |
| 1261 | link->iolink_type == CRAT_IOLINK_TYPE_XGMI) { |
| 1262 | /* |
| 1263 | * If the GPU is not part of a GPU hive, use its pci |
| 1264 | * device location as the hive ID to bind with the CPU. |
| 1265 | */ |
| 1266 | if (!dev->node_props.hive_id) |
| 1267 | dev->node_props.hive_id = pci_dev_id(dev->gpu->adev->pdev); |
| 1268 | peer_dev->node_props.hive_id = dev->node_props.hive_id; |
| 1269 | } |
| 1270 | |
| 1271 | list_for_each_entry(inbound_link, &peer_dev->io_link_props, |
| 1272 | list) { |
| 1273 | if (inbound_link->node_to != link->node_from) |
| 1274 | continue; |
| 1275 | |
| 1276 | inbound_link->flags = CRAT_IOLINK_FLAGS_ENABLED; |
| 1277 | kfd_set_iolink_no_atomics(peer_dev, dev, inbound_link); |
| 1278 | kfd_set_iolink_non_coherent(peer_dev, link, inbound_link); |
| 1279 | } |
| 1280 | } |
| 1281 | |
| 1282 | /* Create indirect links so apply flags setting to all */ |
| 1283 | list_for_each_entry(link, &dev->p2p_link_props, list) { |
| 1284 | link->flags = CRAT_IOLINK_FLAGS_ENABLED; |
| 1285 | kfd_set_iolink_no_atomics(dev, NULL, link); |
| 1286 | peer_dev = kfd_topology_device_by_proximity_domain( |
| 1287 | link->node_to); |
| 1288 | |
| 1289 | if (!peer_dev) |
| 1290 | continue; |
| 1291 | |
| 1292 | list_for_each_entry(inbound_link, &peer_dev->p2p_link_props, |
| 1293 | list) { |
| 1294 | if (inbound_link->node_to != link->node_from) |
| 1295 | continue; |
| 1296 | |
| 1297 | inbound_link->flags = CRAT_IOLINK_FLAGS_ENABLED; |
| 1298 | kfd_set_iolink_no_atomics(peer_dev, dev, inbound_link); |
| 1299 | kfd_set_iolink_non_coherent(peer_dev, link, inbound_link); |
| 1300 | } |
| 1301 | } |
| 1302 | } |
| 1303 | |
| 1304 | static int kfd_build_p2p_node_entry(struct kfd_topology_device *dev, |
| 1305 | struct kfd_iolink_properties *p2plink) |
| 1306 | { |
| 1307 | int ret; |
| 1308 | |
| 1309 | p2plink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL); |
| 1310 | if (!p2plink->kobj) |
| 1311 | return -ENOMEM; |
| 1312 | |
| 1313 | ret = kobject_init_and_add(p2plink->kobj, &iolink_type, |
| 1314 | dev->kobj_p2plink, "%d", dev->node_props.p2p_links_count - 1); |
| 1315 | if (ret < 0) { |
| 1316 | kobject_put(p2plink->kobj); |
| 1317 | return ret; |
| 1318 | } |
| 1319 | |
| 1320 | p2plink->attr.name = "properties"; |
| 1321 | p2plink->attr.mode = KFD_SYSFS_FILE_MODE; |
| 1322 | sysfs_attr_init(&p2plink->attr); |
| 1323 | ret = sysfs_create_file(p2plink->kobj, &p2plink->attr); |
| 1324 | if (ret < 0) |
| 1325 | return ret; |
| 1326 | |
| 1327 | return 0; |
| 1328 | } |
| 1329 | |
| 1330 | static int kfd_create_indirect_link_prop(struct kfd_topology_device *kdev, int gpu_node) |
| 1331 | { |
| 1332 | struct kfd_iolink_properties *gpu_link, *tmp_link, *cpu_link; |
| 1333 | struct kfd_iolink_properties *props = NULL, *props2 = NULL; |
| 1334 | struct kfd_topology_device *cpu_dev; |
| 1335 | int ret = 0; |
| 1336 | int i, num_cpu; |
| 1337 | |
| 1338 | num_cpu = 0; |
| 1339 | list_for_each_entry(cpu_dev, &topology_device_list, list) { |
| 1340 | if (cpu_dev->gpu) |
| 1341 | break; |
| 1342 | num_cpu++; |
| 1343 | } |
| 1344 | |
| 1345 | gpu_link = list_first_entry(&kdev->io_link_props, |
| 1346 | struct kfd_iolink_properties, list); |
| 1347 | if (!gpu_link) |
| 1348 | return -ENOMEM; |
| 1349 | |
| 1350 | for (i = 0; i < num_cpu; i++) { |
| 1351 | /* CPU <--> GPU */ |
| 1352 | if (gpu_link->node_to == i) |
| 1353 | continue; |
| 1354 | |
| 1355 | /* find CPU <--> CPU links */ |
| 1356 | cpu_link = NULL; |
| 1357 | cpu_dev = kfd_topology_device_by_proximity_domain(i); |
| 1358 | if (cpu_dev) { |
| 1359 | list_for_each_entry(tmp_link, |
| 1360 | &cpu_dev->io_link_props, list) { |
| 1361 | if (tmp_link->node_to == gpu_link->node_to) { |
| 1362 | cpu_link = tmp_link; |
| 1363 | break; |
| 1364 | } |
| 1365 | } |
| 1366 | } |
| 1367 | |
| 1368 | if (!cpu_link) |
| 1369 | return -ENOMEM; |
| 1370 | |
| 1371 | /* CPU <--> CPU <--> GPU, GPU node*/ |
| 1372 | props = kfd_alloc_struct(props); |
| 1373 | if (!props) |
| 1374 | return -ENOMEM; |
| 1375 | |
| 1376 | memcpy(props, gpu_link, sizeof(struct kfd_iolink_properties)); |
| 1377 | props->weight = gpu_link->weight + cpu_link->weight; |
| 1378 | props->min_latency = gpu_link->min_latency + cpu_link->min_latency; |
| 1379 | props->max_latency = gpu_link->max_latency + cpu_link->max_latency; |
| 1380 | props->min_bandwidth = min(gpu_link->min_bandwidth, cpu_link->min_bandwidth); |
| 1381 | props->max_bandwidth = min(gpu_link->max_bandwidth, cpu_link->max_bandwidth); |
| 1382 | |
| 1383 | props->node_from = gpu_node; |
| 1384 | props->node_to = i; |
| 1385 | kdev->node_props.p2p_links_count++; |
| 1386 | list_add_tail(&props->list, &kdev->p2p_link_props); |
| 1387 | ret = kfd_build_p2p_node_entry(kdev, props); |
| 1388 | if (ret < 0) |
| 1389 | return ret; |
| 1390 | |
| 1391 | /* for small Bar, no CPU --> GPU in-direct links */ |
| 1392 | if (kfd_dev_is_large_bar(kdev->gpu)) { |
| 1393 | /* CPU <--> CPU <--> GPU, CPU node*/ |
| 1394 | props2 = kfd_alloc_struct(props2); |
| 1395 | if (!props2) |
| 1396 | return -ENOMEM; |
| 1397 | |
| 1398 | memcpy(props2, props, sizeof(struct kfd_iolink_properties)); |
| 1399 | props2->node_from = i; |
| 1400 | props2->node_to = gpu_node; |
| 1401 | props2->kobj = NULL; |
| 1402 | cpu_dev->node_props.p2p_links_count++; |
| 1403 | list_add_tail(&props2->list, &cpu_dev->p2p_link_props); |
| 1404 | ret = kfd_build_p2p_node_entry(cpu_dev, props2); |
| 1405 | if (ret < 0) |
| 1406 | return ret; |
| 1407 | } |
| 1408 | } |
| 1409 | return ret; |
| 1410 | } |
| 1411 | |
| 1412 | #if defined(CONFIG_HSA_AMD_P2P) |
| 1413 | static int kfd_add_peer_prop(struct kfd_topology_device *kdev, |
| 1414 | struct kfd_topology_device *peer, int from, int to) |
| 1415 | { |
| 1416 | struct kfd_iolink_properties *props = NULL; |
| 1417 | struct kfd_iolink_properties *iolink1, *iolink2, *iolink3; |
| 1418 | struct kfd_topology_device *cpu_dev; |
| 1419 | int ret = 0; |
| 1420 | |
| 1421 | if (!amdgpu_device_is_peer_accessible( |
| 1422 | kdev->gpu->adev, |
| 1423 | peer->gpu->adev)) |
| 1424 | return ret; |
| 1425 | |
| 1426 | iolink1 = list_first_entry(&kdev->io_link_props, |
| 1427 | struct kfd_iolink_properties, list); |
| 1428 | if (!iolink1) |
| 1429 | return -ENOMEM; |
| 1430 | |
| 1431 | iolink2 = list_first_entry(&peer->io_link_props, |
| 1432 | struct kfd_iolink_properties, list); |
| 1433 | if (!iolink2) |
| 1434 | return -ENOMEM; |
| 1435 | |
| 1436 | props = kfd_alloc_struct(props); |
| 1437 | if (!props) |
| 1438 | return -ENOMEM; |
| 1439 | |
| 1440 | memcpy(props, iolink1, sizeof(struct kfd_iolink_properties)); |
| 1441 | |
| 1442 | props->weight = iolink1->weight + iolink2->weight; |
| 1443 | props->min_latency = iolink1->min_latency + iolink2->min_latency; |
| 1444 | props->max_latency = iolink1->max_latency + iolink2->max_latency; |
| 1445 | props->min_bandwidth = min(iolink1->min_bandwidth, iolink2->min_bandwidth); |
| 1446 | props->max_bandwidth = min(iolink2->max_bandwidth, iolink2->max_bandwidth); |
| 1447 | |
| 1448 | if (iolink1->node_to != iolink2->node_to) { |
| 1449 | /* CPU->CPU link*/ |
| 1450 | cpu_dev = kfd_topology_device_by_proximity_domain(iolink1->node_to); |
| 1451 | if (cpu_dev) { |
| 1452 | list_for_each_entry(iolink3, &cpu_dev->io_link_props, list) |
| 1453 | if (iolink3->node_to == iolink2->node_to) |
| 1454 | break; |
| 1455 | |
| 1456 | props->weight += iolink3->weight; |
| 1457 | props->min_latency += iolink3->min_latency; |
| 1458 | props->max_latency += iolink3->max_latency; |
| 1459 | props->min_bandwidth = min(props->min_bandwidth, |
| 1460 | iolink3->min_bandwidth); |
| 1461 | props->max_bandwidth = min(props->max_bandwidth, |
| 1462 | iolink3->max_bandwidth); |
| 1463 | } else { |
| 1464 | WARN(1, "CPU node not found"); |
| 1465 | } |
| 1466 | } |
| 1467 | |
| 1468 | props->node_from = from; |
| 1469 | props->node_to = to; |
| 1470 | peer->node_props.p2p_links_count++; |
| 1471 | list_add_tail(&props->list, &peer->p2p_link_props); |
| 1472 | ret = kfd_build_p2p_node_entry(peer, props); |
| 1473 | |
| 1474 | return ret; |
| 1475 | } |
| 1476 | #endif |
| 1477 | |
| 1478 | static int kfd_dev_create_p2p_links(void) |
| 1479 | { |
| 1480 | struct kfd_topology_device *dev; |
| 1481 | struct kfd_topology_device *new_dev; |
| 1482 | #if defined(CONFIG_HSA_AMD_P2P) |
| 1483 | uint32_t i; |
| 1484 | #endif |
| 1485 | uint32_t k; |
| 1486 | int ret = 0; |
| 1487 | |
| 1488 | k = 0; |
| 1489 | list_for_each_entry(dev, &topology_device_list, list) |
| 1490 | k++; |
| 1491 | if (k < 2) |
| 1492 | return 0; |
| 1493 | |
| 1494 | new_dev = list_last_entry(&topology_device_list, struct kfd_topology_device, list); |
| 1495 | if (WARN_ON(!new_dev->gpu)) |
| 1496 | return 0; |
| 1497 | |
| 1498 | k--; |
| 1499 | |
| 1500 | /* create in-direct links */ |
| 1501 | ret = kfd_create_indirect_link_prop(new_dev, k); |
| 1502 | if (ret < 0) |
| 1503 | goto out; |
| 1504 | |
| 1505 | /* create p2p links */ |
| 1506 | #if defined(CONFIG_HSA_AMD_P2P) |
| 1507 | i = 0; |
| 1508 | list_for_each_entry(dev, &topology_device_list, list) { |
| 1509 | if (dev == new_dev) |
| 1510 | break; |
| 1511 | if (!dev->gpu || !dev->gpu->adev || |
| 1512 | (dev->gpu->kfd->hive_id && |
| 1513 | dev->gpu->kfd->hive_id == new_dev->gpu->kfd->hive_id)) |
| 1514 | goto next; |
| 1515 | |
| 1516 | /* check if node(s) is/are peer accessible in one direction or bi-direction */ |
| 1517 | ret = kfd_add_peer_prop(new_dev, dev, i, k); |
| 1518 | if (ret < 0) |
| 1519 | goto out; |
| 1520 | |
| 1521 | ret = kfd_add_peer_prop(dev, new_dev, k, i); |
| 1522 | if (ret < 0) |
| 1523 | goto out; |
| 1524 | next: |
| 1525 | i++; |
| 1526 | } |
| 1527 | #endif |
| 1528 | |
| 1529 | out: |
| 1530 | return ret; |
| 1531 | } |
| 1532 | |
| 1533 | /* Helper function. See kfd_fill_gpu_cache_info for parameter description */ |
| 1534 | static int fill_in_l1_pcache(struct kfd_cache_properties **props_ext, |
| 1535 | struct kfd_gpu_cache_info *pcache_info, |
| 1536 | struct kfd_cu_info *cu_info, |
| 1537 | int cu_bitmask, |
| 1538 | int cache_type, unsigned int cu_processor_id, |
| 1539 | int cu_block) |
| 1540 | { |
| 1541 | unsigned int cu_sibling_map_mask; |
| 1542 | int first_active_cu; |
| 1543 | struct kfd_cache_properties *pcache = NULL; |
| 1544 | |
| 1545 | cu_sibling_map_mask = cu_bitmask; |
| 1546 | cu_sibling_map_mask >>= cu_block; |
| 1547 | cu_sibling_map_mask &= ((1 << pcache_info[cache_type].num_cu_shared) - 1); |
| 1548 | first_active_cu = ffs(cu_sibling_map_mask); |
| 1549 | |
| 1550 | /* CU could be inactive. In case of shared cache find the first active |
| 1551 | * CU. and incase of non-shared cache check if the CU is inactive. If |
| 1552 | * inactive active skip it |
| 1553 | */ |
| 1554 | if (first_active_cu) { |
| 1555 | pcache = kfd_alloc_struct(pcache); |
| 1556 | if (!pcache) |
| 1557 | return -ENOMEM; |
| 1558 | |
| 1559 | memset(pcache, 0, sizeof(struct kfd_cache_properties)); |
| 1560 | pcache->processor_id_low = cu_processor_id + (first_active_cu - 1); |
| 1561 | pcache->cache_level = pcache_info[cache_type].cache_level; |
| 1562 | pcache->cache_size = pcache_info[cache_type].cache_size; |
| 1563 | |
| 1564 | if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_DATA_CACHE) |
| 1565 | pcache->cache_type |= HSA_CACHE_TYPE_DATA; |
| 1566 | if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_INST_CACHE) |
| 1567 | pcache->cache_type |= HSA_CACHE_TYPE_INSTRUCTION; |
| 1568 | if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_CPU_CACHE) |
| 1569 | pcache->cache_type |= HSA_CACHE_TYPE_CPU; |
| 1570 | if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_SIMD_CACHE) |
| 1571 | pcache->cache_type |= HSA_CACHE_TYPE_HSACU; |
| 1572 | |
| 1573 | /* Sibling map is w.r.t processor_id_low, so shift out |
| 1574 | * inactive CU |
| 1575 | */ |
| 1576 | cu_sibling_map_mask = |
| 1577 | cu_sibling_map_mask >> (first_active_cu - 1); |
| 1578 | |
| 1579 | pcache->sibling_map[0] = (uint8_t)(cu_sibling_map_mask & 0xFF); |
| 1580 | pcache->sibling_map[1] = |
| 1581 | (uint8_t)((cu_sibling_map_mask >> 8) & 0xFF); |
| 1582 | pcache->sibling_map[2] = |
| 1583 | (uint8_t)((cu_sibling_map_mask >> 16) & 0xFF); |
| 1584 | pcache->sibling_map[3] = |
| 1585 | (uint8_t)((cu_sibling_map_mask >> 24) & 0xFF); |
| 1586 | |
| 1587 | pcache->sibling_map_size = 4; |
| 1588 | *props_ext = pcache; |
| 1589 | |
| 1590 | return 0; |
| 1591 | } |
| 1592 | return 1; |
| 1593 | } |
| 1594 | |
| 1595 | /* Helper function. See kfd_fill_gpu_cache_info for parameter description */ |
| 1596 | static int fill_in_l2_l3_pcache(struct kfd_cache_properties **props_ext, |
| 1597 | struct kfd_gpu_cache_info *pcache_info, |
| 1598 | struct kfd_cu_info *cu_info, |
| 1599 | int cache_type, unsigned int cu_processor_id, |
| 1600 | struct kfd_node *knode) |
| 1601 | { |
| 1602 | unsigned int cu_sibling_map_mask; |
| 1603 | int first_active_cu; |
| 1604 | int i, j, k, xcc, start, end; |
| 1605 | struct kfd_cache_properties *pcache = NULL; |
| 1606 | |
| 1607 | start = ffs(knode->xcc_mask) - 1; |
| 1608 | end = start + NUM_XCC(knode->xcc_mask); |
| 1609 | cu_sibling_map_mask = cu_info->cu_bitmap[start][0][0]; |
| 1610 | cu_sibling_map_mask &= |
| 1611 | ((1 << pcache_info[cache_type].num_cu_shared) - 1); |
| 1612 | first_active_cu = ffs(cu_sibling_map_mask); |
| 1613 | |
| 1614 | /* CU could be inactive. In case of shared cache find the first active |
| 1615 | * CU. and incase of non-shared cache check if the CU is inactive. If |
| 1616 | * inactive active skip it |
| 1617 | */ |
| 1618 | if (first_active_cu) { |
| 1619 | pcache = kfd_alloc_struct(pcache); |
| 1620 | if (!pcache) |
| 1621 | return -ENOMEM; |
| 1622 | |
| 1623 | memset(pcache, 0, sizeof(struct kfd_cache_properties)); |
| 1624 | pcache->processor_id_low = cu_processor_id |
| 1625 | + (first_active_cu - 1); |
| 1626 | pcache->cache_level = pcache_info[cache_type].cache_level; |
| 1627 | pcache->cache_size = pcache_info[cache_type].cache_size; |
| 1628 | |
| 1629 | if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_DATA_CACHE) |
| 1630 | pcache->cache_type |= HSA_CACHE_TYPE_DATA; |
| 1631 | if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_INST_CACHE) |
| 1632 | pcache->cache_type |= HSA_CACHE_TYPE_INSTRUCTION; |
| 1633 | if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_CPU_CACHE) |
| 1634 | pcache->cache_type |= HSA_CACHE_TYPE_CPU; |
| 1635 | if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_SIMD_CACHE) |
| 1636 | pcache->cache_type |= HSA_CACHE_TYPE_HSACU; |
| 1637 | |
| 1638 | /* Sibling map is w.r.t processor_id_low, so shift out |
| 1639 | * inactive CU |
| 1640 | */ |
| 1641 | cu_sibling_map_mask = cu_sibling_map_mask >> (first_active_cu - 1); |
| 1642 | k = 0; |
| 1643 | |
| 1644 | for (xcc = start; xcc < end; xcc++) { |
| 1645 | for (i = 0; i < cu_info->num_shader_engines; i++) { |
| 1646 | for (j = 0; j < cu_info->num_shader_arrays_per_engine; j++) { |
| 1647 | pcache->sibling_map[k] = (uint8_t)(cu_sibling_map_mask & 0xFF); |
| 1648 | pcache->sibling_map[k+1] = (uint8_t)((cu_sibling_map_mask >> 8) & 0xFF); |
| 1649 | pcache->sibling_map[k+2] = (uint8_t)((cu_sibling_map_mask >> 16) & 0xFF); |
| 1650 | pcache->sibling_map[k+3] = (uint8_t)((cu_sibling_map_mask >> 24) & 0xFF); |
| 1651 | k += 4; |
| 1652 | |
| 1653 | cu_sibling_map_mask = cu_info->cu_bitmap[xcc][i % 4][j + i / 4]; |
| 1654 | cu_sibling_map_mask &= ((1 << pcache_info[cache_type].num_cu_shared) - 1); |
| 1655 | } |
| 1656 | } |
| 1657 | } |
| 1658 | pcache->sibling_map_size = k; |
| 1659 | *props_ext = pcache; |
| 1660 | return 0; |
| 1661 | } |
| 1662 | return 1; |
| 1663 | } |
| 1664 | |
| 1665 | #define KFD_MAX_CACHE_TYPES 6 |
| 1666 | |
| 1667 | /* kfd_fill_cache_non_crat_info - Fill GPU cache info using kfd_gpu_cache_info |
| 1668 | * tables |
| 1669 | */ |
| 1670 | static void kfd_fill_cache_non_crat_info(struct kfd_topology_device *dev, struct kfd_node *kdev) |
| 1671 | { |
| 1672 | struct kfd_gpu_cache_info *pcache_info = NULL; |
| 1673 | int i, j, k, xcc, start, end; |
| 1674 | int ct = 0; |
| 1675 | unsigned int cu_processor_id; |
| 1676 | int ret; |
| 1677 | unsigned int num_cu_shared; |
| 1678 | struct kfd_cu_info cu_info; |
| 1679 | struct kfd_cu_info *pcu_info; |
| 1680 | int gpu_processor_id; |
| 1681 | struct kfd_cache_properties *props_ext; |
| 1682 | int num_of_entries = 0; |
| 1683 | int num_of_cache_types = 0; |
| 1684 | struct kfd_gpu_cache_info cache_info[KFD_MAX_CACHE_TYPES]; |
| 1685 | |
| 1686 | amdgpu_amdkfd_get_cu_info(kdev->adev, &cu_info); |
| 1687 | pcu_info = &cu_info; |
| 1688 | |
| 1689 | gpu_processor_id = dev->node_props.simd_id_base; |
| 1690 | |
| 1691 | pcache_info = cache_info; |
| 1692 | num_of_cache_types = kfd_get_gpu_cache_info(kdev, &pcache_info); |
| 1693 | if (!num_of_cache_types) { |
| 1694 | pr_warn("no cache info found\n"); |
| 1695 | return; |
| 1696 | } |
| 1697 | |
| 1698 | /* For each type of cache listed in the kfd_gpu_cache_info table, |
| 1699 | * go through all available Compute Units. |
| 1700 | * The [i,j,k] loop will |
| 1701 | * if kfd_gpu_cache_info.num_cu_shared = 1 |
| 1702 | * will parse through all available CU |
| 1703 | * If (kfd_gpu_cache_info.num_cu_shared != 1) |
| 1704 | * then it will consider only one CU from |
| 1705 | * the shared unit |
| 1706 | */ |
| 1707 | start = ffs(kdev->xcc_mask) - 1; |
| 1708 | end = start + NUM_XCC(kdev->xcc_mask); |
| 1709 | |
| 1710 | for (ct = 0; ct < num_of_cache_types; ct++) { |
| 1711 | cu_processor_id = gpu_processor_id; |
| 1712 | if (pcache_info[ct].cache_level == 1) { |
| 1713 | for (xcc = start; xcc < end; xcc++) { |
| 1714 | for (i = 0; i < pcu_info->num_shader_engines; i++) { |
| 1715 | for (j = 0; j < pcu_info->num_shader_arrays_per_engine; j++) { |
| 1716 | for (k = 0; k < pcu_info->num_cu_per_sh; k += pcache_info[ct].num_cu_shared) { |
| 1717 | |
| 1718 | ret = fill_in_l1_pcache(&props_ext, pcache_info, pcu_info, |
| 1719 | pcu_info->cu_bitmap[xcc][i % 4][j + i / 4], ct, |
| 1720 | cu_processor_id, k); |
| 1721 | |
| 1722 | if (ret < 0) |
| 1723 | break; |
| 1724 | |
| 1725 | if (!ret) { |
| 1726 | num_of_entries++; |
| 1727 | list_add_tail(&props_ext->list, &dev->cache_props); |
| 1728 | } |
| 1729 | |
| 1730 | /* Move to next CU block */ |
| 1731 | num_cu_shared = ((k + pcache_info[ct].num_cu_shared) <= |
| 1732 | pcu_info->num_cu_per_sh) ? |
| 1733 | pcache_info[ct].num_cu_shared : |
| 1734 | (pcu_info->num_cu_per_sh - k); |
| 1735 | cu_processor_id += num_cu_shared; |
| 1736 | } |
| 1737 | } |
| 1738 | } |
| 1739 | } |
| 1740 | } else { |
| 1741 | ret = fill_in_l2_l3_pcache(&props_ext, pcache_info, |
| 1742 | pcu_info, ct, cu_processor_id, kdev); |
| 1743 | |
| 1744 | if (ret < 0) |
| 1745 | break; |
| 1746 | |
| 1747 | if (!ret) { |
| 1748 | num_of_entries++; |
| 1749 | list_add_tail(&props_ext->list, &dev->cache_props); |
| 1750 | } |
| 1751 | } |
| 1752 | } |
| 1753 | dev->node_props.caches_count += num_of_entries; |
| 1754 | pr_debug("Added [%d] GPU cache entries\n", num_of_entries); |
| 1755 | } |
| 1756 | |
| 1757 | static int kfd_topology_add_device_locked(struct kfd_node *gpu, uint32_t gpu_id, |
| 1758 | struct kfd_topology_device **dev) |
| 1759 | { |
| 1760 | int proximity_domain = ++topology_crat_proximity_domain; |
| 1761 | struct list_head temp_topology_device_list; |
| 1762 | void *crat_image = NULL; |
| 1763 | size_t image_size = 0; |
| 1764 | int res; |
| 1765 | |
| 1766 | res = kfd_create_crat_image_virtual(&crat_image, &image_size, |
| 1767 | COMPUTE_UNIT_GPU, gpu, |
| 1768 | proximity_domain); |
| 1769 | if (res) { |
| 1770 | pr_err("Error creating VCRAT for GPU (ID: 0x%x)\n", |
| 1771 | gpu_id); |
| 1772 | topology_crat_proximity_domain--; |
| 1773 | goto err; |
| 1774 | } |
| 1775 | |
| 1776 | INIT_LIST_HEAD(&temp_topology_device_list); |
| 1777 | |
| 1778 | res = kfd_parse_crat_table(crat_image, |
| 1779 | &temp_topology_device_list, |
| 1780 | proximity_domain); |
| 1781 | if (res) { |
| 1782 | pr_err("Error parsing VCRAT for GPU (ID: 0x%x)\n", |
| 1783 | gpu_id); |
| 1784 | topology_crat_proximity_domain--; |
| 1785 | goto err; |
| 1786 | } |
| 1787 | |
| 1788 | kfd_topology_update_device_list(&temp_topology_device_list, |
| 1789 | &topology_device_list); |
| 1790 | |
| 1791 | *dev = kfd_assign_gpu(gpu); |
| 1792 | if (WARN_ON(!*dev)) { |
| 1793 | res = -ENODEV; |
| 1794 | goto err; |
| 1795 | } |
| 1796 | |
| 1797 | /* Fill the cache affinity information here for the GPUs |
| 1798 | * using VCRAT |
| 1799 | */ |
| 1800 | kfd_fill_cache_non_crat_info(*dev, gpu); |
| 1801 | |
| 1802 | /* Update the SYSFS tree, since we added another topology |
| 1803 | * device |
| 1804 | */ |
| 1805 | res = kfd_topology_update_sysfs(); |
| 1806 | if (!res) |
| 1807 | sys_props.generation_count++; |
| 1808 | else |
| 1809 | pr_err("Failed to update GPU (ID: 0x%x) to sysfs topology. res=%d\n", |
| 1810 | gpu_id, res); |
| 1811 | |
| 1812 | err: |
| 1813 | kfd_destroy_crat_image(crat_image); |
| 1814 | return res; |
| 1815 | } |
| 1816 | |
| 1817 | static void kfd_topology_set_dbg_firmware_support(struct kfd_topology_device *dev) |
| 1818 | { |
| 1819 | bool firmware_supported = true; |
| 1820 | |
| 1821 | if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(11, 0, 0) && |
| 1822 | KFD_GC_VERSION(dev->gpu) < IP_VERSION(12, 0, 0)) { |
| 1823 | uint32_t mes_api_rev = (dev->gpu->adev->mes.sched_version & |
| 1824 | AMDGPU_MES_API_VERSION_MASK) >> |
| 1825 | AMDGPU_MES_API_VERSION_SHIFT; |
| 1826 | uint32_t mes_rev = dev->gpu->adev->mes.sched_version & |
| 1827 | AMDGPU_MES_VERSION_MASK; |
| 1828 | |
| 1829 | firmware_supported = (mes_api_rev >= 14) && (mes_rev >= 64); |
| 1830 | goto out; |
| 1831 | } |
| 1832 | |
| 1833 | /* |
| 1834 | * Note: Any unlisted devices here are assumed to support exception handling. |
| 1835 | * Add additional checks here as needed. |
| 1836 | */ |
| 1837 | switch (KFD_GC_VERSION(dev->gpu)) { |
| 1838 | case IP_VERSION(9, 0, 1): |
| 1839 | firmware_supported = dev->gpu->kfd->mec_fw_version >= 459 + 32768; |
| 1840 | break; |
| 1841 | case IP_VERSION(9, 1, 0): |
| 1842 | case IP_VERSION(9, 2, 1): |
| 1843 | case IP_VERSION(9, 2, 2): |
| 1844 | case IP_VERSION(9, 3, 0): |
| 1845 | case IP_VERSION(9, 4, 0): |
| 1846 | firmware_supported = dev->gpu->kfd->mec_fw_version >= 459; |
| 1847 | break; |
| 1848 | case IP_VERSION(9, 4, 1): |
| 1849 | firmware_supported = dev->gpu->kfd->mec_fw_version >= 60; |
| 1850 | break; |
| 1851 | case IP_VERSION(9, 4, 2): |
| 1852 | firmware_supported = dev->gpu->kfd->mec_fw_version >= 51; |
| 1853 | break; |
| 1854 | case IP_VERSION(10, 1, 10): |
| 1855 | case IP_VERSION(10, 1, 2): |
| 1856 | case IP_VERSION(10, 1, 1): |
| 1857 | firmware_supported = dev->gpu->kfd->mec_fw_version >= 144; |
| 1858 | break; |
| 1859 | case IP_VERSION(10, 3, 0): |
| 1860 | case IP_VERSION(10, 3, 2): |
| 1861 | case IP_VERSION(10, 3, 1): |
| 1862 | case IP_VERSION(10, 3, 4): |
| 1863 | case IP_VERSION(10, 3, 5): |
| 1864 | firmware_supported = dev->gpu->kfd->mec_fw_version >= 89; |
| 1865 | break; |
| 1866 | case IP_VERSION(10, 1, 3): |
| 1867 | case IP_VERSION(10, 3, 3): |
| 1868 | firmware_supported = false; |
| 1869 | break; |
| 1870 | default: |
| 1871 | break; |
| 1872 | } |
| 1873 | |
| 1874 | out: |
| 1875 | if (firmware_supported) |
| 1876 | dev->node_props.capability |= HSA_CAP_TRAP_DEBUG_FIRMWARE_SUPPORTED; |
| 1877 | } |
| 1878 | |
| 1879 | static void kfd_topology_set_capabilities(struct kfd_topology_device *dev) |
| 1880 | { |
| 1881 | dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 << |
| 1882 | HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & |
| 1883 | HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); |
| 1884 | |
| 1885 | dev->node_props.capability |= HSA_CAP_TRAP_DEBUG_SUPPORT | |
| 1886 | HSA_CAP_TRAP_DEBUG_WAVE_LAUNCH_TRAP_OVERRIDE_SUPPORTED | |
| 1887 | HSA_CAP_TRAP_DEBUG_WAVE_LAUNCH_MODE_SUPPORTED; |
| 1888 | |
| 1889 | if (kfd_dbg_has_ttmps_always_setup(dev->gpu)) |
| 1890 | dev->node_props.debug_prop |= HSA_DBG_DISPATCH_INFO_ALWAYS_VALID; |
| 1891 | |
| 1892 | if (KFD_GC_VERSION(dev->gpu) < IP_VERSION(10, 0, 0)) { |
| 1893 | if (KFD_GC_VERSION(dev->gpu) == IP_VERSION(9, 4, 3)) |
| 1894 | dev->node_props.debug_prop |= |
| 1895 | HSA_DBG_WATCH_ADDR_MASK_LO_BIT_GFX9_4_3 | |
| 1896 | HSA_DBG_WATCH_ADDR_MASK_HI_BIT_GFX9_4_3; |
| 1897 | else |
| 1898 | dev->node_props.debug_prop |= |
| 1899 | HSA_DBG_WATCH_ADDR_MASK_LO_BIT_GFX9 | |
| 1900 | HSA_DBG_WATCH_ADDR_MASK_HI_BIT; |
| 1901 | |
| 1902 | if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(9, 4, 2)) |
| 1903 | dev->node_props.capability |= |
| 1904 | HSA_CAP_TRAP_DEBUG_PRECISE_MEMORY_OPERATIONS_SUPPORTED; |
| 1905 | } else { |
| 1906 | dev->node_props.debug_prop |= HSA_DBG_WATCH_ADDR_MASK_LO_BIT_GFX10 | |
| 1907 | HSA_DBG_WATCH_ADDR_MASK_HI_BIT; |
| 1908 | |
| 1909 | if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(11, 0, 0)) |
| 1910 | dev->node_props.capability |= |
| 1911 | HSA_CAP_TRAP_DEBUG_PRECISE_MEMORY_OPERATIONS_SUPPORTED; |
| 1912 | } |
| 1913 | |
| 1914 | kfd_topology_set_dbg_firmware_support(dev); |
| 1915 | } |
| 1916 | |
| 1917 | int kfd_topology_add_device(struct kfd_node *gpu) |
| 1918 | { |
| 1919 | uint32_t gpu_id; |
| 1920 | struct kfd_topology_device *dev; |
| 1921 | struct kfd_cu_info cu_info; |
| 1922 | int res = 0; |
| 1923 | int i; |
| 1924 | const char *asic_name = amdgpu_asic_name[gpu->adev->asic_type]; |
| 1925 | |
| 1926 | gpu_id = kfd_generate_gpu_id(gpu); |
| 1927 | if (gpu->xcp && !gpu->xcp->ddev) { |
| 1928 | dev_warn(gpu->adev->dev, |
| 1929 | "Won't add GPU (ID: 0x%x) to topology since it has no drm node assigned.", |
| 1930 | gpu_id); |
| 1931 | return 0; |
| 1932 | } else { |
| 1933 | pr_debug("Adding new GPU (ID: 0x%x) to topology\n", gpu_id); |
| 1934 | } |
| 1935 | |
| 1936 | /* Check to see if this gpu device exists in the topology_device_list. |
| 1937 | * If so, assign the gpu to that device, |
| 1938 | * else create a Virtual CRAT for this gpu device and then parse that |
| 1939 | * CRAT to create a new topology device. Once created assign the gpu to |
| 1940 | * that topology device |
| 1941 | */ |
| 1942 | down_write(&topology_lock); |
| 1943 | dev = kfd_assign_gpu(gpu); |
| 1944 | if (!dev) |
| 1945 | res = kfd_topology_add_device_locked(gpu, gpu_id, &dev); |
| 1946 | up_write(&topology_lock); |
| 1947 | if (res) |
| 1948 | return res; |
| 1949 | |
| 1950 | dev->gpu_id = gpu_id; |
| 1951 | gpu->id = gpu_id; |
| 1952 | |
| 1953 | kfd_dev_create_p2p_links(); |
| 1954 | |
| 1955 | /* TODO: Move the following lines to function |
| 1956 | * kfd_add_non_crat_information |
| 1957 | */ |
| 1958 | |
| 1959 | /* Fill-in additional information that is not available in CRAT but |
| 1960 | * needed for the topology |
| 1961 | */ |
| 1962 | |
| 1963 | amdgpu_amdkfd_get_cu_info(dev->gpu->adev, &cu_info); |
| 1964 | |
| 1965 | for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1; i++) { |
| 1966 | dev->node_props.name[i] = __tolower(asic_name[i]); |
| 1967 | if (asic_name[i] == '\0') |
| 1968 | break; |
| 1969 | } |
| 1970 | dev->node_props.name[i] = '\0'; |
| 1971 | |
| 1972 | dev->node_props.simd_arrays_per_engine = |
| 1973 | cu_info.num_shader_arrays_per_engine; |
| 1974 | |
| 1975 | dev->node_props.gfx_target_version = |
| 1976 | gpu->kfd->device_info.gfx_target_version; |
| 1977 | dev->node_props.vendor_id = gpu->adev->pdev->vendor; |
| 1978 | dev->node_props.device_id = gpu->adev->pdev->device; |
| 1979 | dev->node_props.capability |= |
| 1980 | ((dev->gpu->adev->rev_id << HSA_CAP_ASIC_REVISION_SHIFT) & |
| 1981 | HSA_CAP_ASIC_REVISION_MASK); |
| 1982 | |
| 1983 | dev->node_props.location_id = pci_dev_id(gpu->adev->pdev); |
| 1984 | if (KFD_GC_VERSION(dev->gpu->kfd) == IP_VERSION(9, 4, 3)) |
| 1985 | dev->node_props.location_id |= dev->gpu->node_id; |
| 1986 | |
| 1987 | dev->node_props.domain = pci_domain_nr(gpu->adev->pdev->bus); |
| 1988 | dev->node_props.max_engine_clk_fcompute = |
| 1989 | amdgpu_amdkfd_get_max_engine_clock_in_mhz(dev->gpu->adev); |
| 1990 | dev->node_props.max_engine_clk_ccompute = |
| 1991 | cpufreq_quick_get_max(0) / 1000; |
| 1992 | |
| 1993 | if (gpu->xcp) |
| 1994 | dev->node_props.drm_render_minor = gpu->xcp->ddev->render->index; |
| 1995 | else |
| 1996 | dev->node_props.drm_render_minor = |
| 1997 | gpu->kfd->shared_resources.drm_render_minor; |
| 1998 | |
| 1999 | dev->node_props.hive_id = gpu->kfd->hive_id; |
| 2000 | dev->node_props.num_sdma_engines = kfd_get_num_sdma_engines(gpu); |
| 2001 | dev->node_props.num_sdma_xgmi_engines = |
| 2002 | kfd_get_num_xgmi_sdma_engines(gpu); |
| 2003 | dev->node_props.num_sdma_queues_per_engine = |
| 2004 | gpu->kfd->device_info.num_sdma_queues_per_engine - |
| 2005 | gpu->kfd->device_info.num_reserved_sdma_queues_per_engine; |
| 2006 | dev->node_props.num_gws = (dev->gpu->gws && |
| 2007 | dev->gpu->dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) ? |
| 2008 | dev->gpu->adev->gds.gws_size : 0; |
| 2009 | dev->node_props.num_cp_queues = get_cp_queues_num(dev->gpu->dqm); |
| 2010 | |
| 2011 | kfd_fill_mem_clk_max_info(dev); |
| 2012 | kfd_fill_iolink_non_crat_info(dev); |
| 2013 | |
| 2014 | switch (dev->gpu->adev->asic_type) { |
| 2015 | case CHIP_KAVERI: |
| 2016 | case CHIP_HAWAII: |
| 2017 | case CHIP_TONGA: |
| 2018 | dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_PRE_1_0 << |
| 2019 | HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & |
| 2020 | HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); |
| 2021 | break; |
| 2022 | case CHIP_CARRIZO: |
| 2023 | case CHIP_FIJI: |
| 2024 | case CHIP_POLARIS10: |
| 2025 | case CHIP_POLARIS11: |
| 2026 | case CHIP_POLARIS12: |
| 2027 | case CHIP_VEGAM: |
| 2028 | pr_debug("Adding doorbell packet type capability\n"); |
| 2029 | dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 << |
| 2030 | HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) & |
| 2031 | HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK); |
| 2032 | break; |
| 2033 | default: |
| 2034 | if (KFD_GC_VERSION(dev->gpu) < IP_VERSION(9, 0, 1)) |
| 2035 | WARN(1, "Unexpected ASIC family %u", |
| 2036 | dev->gpu->adev->asic_type); |
| 2037 | else |
| 2038 | kfd_topology_set_capabilities(dev); |
| 2039 | } |
| 2040 | |
| 2041 | /* |
| 2042 | * Overwrite ATS capability according to needs_iommu_device to fix |
| 2043 | * potential missing corresponding bit in CRAT of BIOS. |
| 2044 | */ |
| 2045 | dev->node_props.capability &= ~HSA_CAP_ATS_PRESENT; |
| 2046 | |
| 2047 | /* Fix errors in CZ CRAT. |
| 2048 | * simd_count: Carrizo CRAT reports wrong simd_count, probably |
| 2049 | * because it doesn't consider masked out CUs |
| 2050 | * max_waves_per_simd: Carrizo reports wrong max_waves_per_simd |
| 2051 | */ |
| 2052 | if (dev->gpu->adev->asic_type == CHIP_CARRIZO) { |
| 2053 | dev->node_props.simd_count = |
| 2054 | cu_info.simd_per_cu * cu_info.cu_active_number; |
| 2055 | dev->node_props.max_waves_per_simd = 10; |
| 2056 | } |
| 2057 | |
| 2058 | /* kfd only concerns sram ecc on GFX and HBM ecc on UMC */ |
| 2059 | dev->node_props.capability |= |
| 2060 | ((dev->gpu->adev->ras_enabled & BIT(AMDGPU_RAS_BLOCK__GFX)) != 0) ? |
| 2061 | HSA_CAP_SRAM_EDCSUPPORTED : 0; |
| 2062 | dev->node_props.capability |= |
| 2063 | ((dev->gpu->adev->ras_enabled & BIT(AMDGPU_RAS_BLOCK__UMC)) != 0) ? |
| 2064 | HSA_CAP_MEM_EDCSUPPORTED : 0; |
| 2065 | |
| 2066 | if (KFD_GC_VERSION(dev->gpu) != IP_VERSION(9, 0, 1)) |
| 2067 | dev->node_props.capability |= (dev->gpu->adev->ras_enabled != 0) ? |
| 2068 | HSA_CAP_RASEVENTNOTIFY : 0; |
| 2069 | |
| 2070 | if (KFD_IS_SVM_API_SUPPORTED(dev->gpu->adev)) |
| 2071 | dev->node_props.capability |= HSA_CAP_SVMAPI_SUPPORTED; |
| 2072 | |
| 2073 | if (dev->gpu->adev->gmc.is_app_apu || |
| 2074 | dev->gpu->adev->gmc.xgmi.connected_to_cpu) |
| 2075 | dev->node_props.capability |= HSA_CAP_FLAGS_COHERENTHOSTACCESS; |
| 2076 | |
| 2077 | kfd_debug_print_topology(); |
| 2078 | |
| 2079 | kfd_notify_gpu_change(gpu_id, 1); |
| 2080 | |
| 2081 | return 0; |
| 2082 | } |
| 2083 | |
| 2084 | /** |
| 2085 | * kfd_topology_update_io_links() - Update IO links after device removal. |
| 2086 | * @proximity_domain: Proximity domain value of the dev being removed. |
| 2087 | * |
| 2088 | * The topology list currently is arranged in increasing order of |
| 2089 | * proximity domain. |
| 2090 | * |
| 2091 | * Two things need to be done when a device is removed: |
| 2092 | * 1. All the IO links to this device need to be removed. |
| 2093 | * 2. All nodes after the current device node need to move |
| 2094 | * up once this device node is removed from the topology |
| 2095 | * list. As a result, the proximity domain values for |
| 2096 | * all nodes after the node being deleted reduce by 1. |
| 2097 | * This would also cause the proximity domain values for |
| 2098 | * io links to be updated based on new proximity domain |
| 2099 | * values. |
| 2100 | * |
| 2101 | * Context: The caller must hold write topology_lock. |
| 2102 | */ |
| 2103 | static void kfd_topology_update_io_links(int proximity_domain) |
| 2104 | { |
| 2105 | struct kfd_topology_device *dev; |
| 2106 | struct kfd_iolink_properties *iolink, *p2plink, *tmp; |
| 2107 | |
| 2108 | list_for_each_entry(dev, &topology_device_list, list) { |
| 2109 | if (dev->proximity_domain > proximity_domain) |
| 2110 | dev->proximity_domain--; |
| 2111 | |
| 2112 | list_for_each_entry_safe(iolink, tmp, &dev->io_link_props, list) { |
| 2113 | /* |
| 2114 | * If there is an io link to the dev being deleted |
| 2115 | * then remove that IO link also. |
| 2116 | */ |
| 2117 | if (iolink->node_to == proximity_domain) { |
| 2118 | list_del(&iolink->list); |
| 2119 | dev->node_props.io_links_count--; |
| 2120 | } else { |
| 2121 | if (iolink->node_from > proximity_domain) |
| 2122 | iolink->node_from--; |
| 2123 | if (iolink->node_to > proximity_domain) |
| 2124 | iolink->node_to--; |
| 2125 | } |
| 2126 | } |
| 2127 | |
| 2128 | list_for_each_entry_safe(p2plink, tmp, &dev->p2p_link_props, list) { |
| 2129 | /* |
| 2130 | * If there is a p2p link to the dev being deleted |
| 2131 | * then remove that p2p link also. |
| 2132 | */ |
| 2133 | if (p2plink->node_to == proximity_domain) { |
| 2134 | list_del(&p2plink->list); |
| 2135 | dev->node_props.p2p_links_count--; |
| 2136 | } else { |
| 2137 | if (p2plink->node_from > proximity_domain) |
| 2138 | p2plink->node_from--; |
| 2139 | if (p2plink->node_to > proximity_domain) |
| 2140 | p2plink->node_to--; |
| 2141 | } |
| 2142 | } |
| 2143 | } |
| 2144 | } |
| 2145 | |
| 2146 | int kfd_topology_remove_device(struct kfd_node *gpu) |
| 2147 | { |
| 2148 | struct kfd_topology_device *dev, *tmp; |
| 2149 | uint32_t gpu_id; |
| 2150 | int res = -ENODEV; |
| 2151 | int i = 0; |
| 2152 | |
| 2153 | down_write(&topology_lock); |
| 2154 | |
| 2155 | list_for_each_entry_safe(dev, tmp, &topology_device_list, list) { |
| 2156 | if (dev->gpu == gpu) { |
| 2157 | gpu_id = dev->gpu_id; |
| 2158 | kfd_remove_sysfs_node_entry(dev); |
| 2159 | kfd_release_topology_device(dev); |
| 2160 | sys_props.num_devices--; |
| 2161 | kfd_topology_update_io_links(i); |
| 2162 | topology_crat_proximity_domain = sys_props.num_devices-1; |
| 2163 | sys_props.generation_count++; |
| 2164 | res = 0; |
| 2165 | if (kfd_topology_update_sysfs() < 0) |
| 2166 | kfd_topology_release_sysfs(); |
| 2167 | break; |
| 2168 | } |
| 2169 | i++; |
| 2170 | } |
| 2171 | |
| 2172 | up_write(&topology_lock); |
| 2173 | |
| 2174 | if (!res) |
| 2175 | kfd_notify_gpu_change(gpu_id, 0); |
| 2176 | |
| 2177 | return res; |
| 2178 | } |
| 2179 | |
| 2180 | /* kfd_topology_enum_kfd_devices - Enumerate through all devices in KFD |
| 2181 | * topology. If GPU device is found @idx, then valid kfd_dev pointer is |
| 2182 | * returned through @kdev |
| 2183 | * Return - 0: On success (@kdev will be NULL for non GPU nodes) |
| 2184 | * -1: If end of list |
| 2185 | */ |
| 2186 | int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_node **kdev) |
| 2187 | { |
| 2188 | |
| 2189 | struct kfd_topology_device *top_dev; |
| 2190 | uint8_t device_idx = 0; |
| 2191 | |
| 2192 | *kdev = NULL; |
| 2193 | down_read(&topology_lock); |
| 2194 | |
| 2195 | list_for_each_entry(top_dev, &topology_device_list, list) { |
| 2196 | if (device_idx == idx) { |
| 2197 | *kdev = top_dev->gpu; |
| 2198 | up_read(&topology_lock); |
| 2199 | return 0; |
| 2200 | } |
| 2201 | |
| 2202 | device_idx++; |
| 2203 | } |
| 2204 | |
| 2205 | up_read(&topology_lock); |
| 2206 | |
| 2207 | return -1; |
| 2208 | |
| 2209 | } |
| 2210 | |
| 2211 | static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask) |
| 2212 | { |
| 2213 | int first_cpu_of_numa_node; |
| 2214 | |
| 2215 | if (!cpumask || cpumask == cpu_none_mask) |
| 2216 | return -1; |
| 2217 | first_cpu_of_numa_node = cpumask_first(cpumask); |
| 2218 | if (first_cpu_of_numa_node >= nr_cpu_ids) |
| 2219 | return -1; |
| 2220 | #ifdef CONFIG_X86_64 |
| 2221 | return cpu_data(first_cpu_of_numa_node).apicid; |
| 2222 | #else |
| 2223 | return first_cpu_of_numa_node; |
| 2224 | #endif |
| 2225 | } |
| 2226 | |
| 2227 | /* kfd_numa_node_to_apic_id - Returns the APIC ID of the first logical processor |
| 2228 | * of the given NUMA node (numa_node_id) |
| 2229 | * Return -1 on failure |
| 2230 | */ |
| 2231 | int kfd_numa_node_to_apic_id(int numa_node_id) |
| 2232 | { |
| 2233 | if (numa_node_id == -1) { |
| 2234 | pr_warn("Invalid NUMA Node. Use online CPU mask\n"); |
| 2235 | return kfd_cpumask_to_apic_id(cpu_online_mask); |
| 2236 | } |
| 2237 | return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id)); |
| 2238 | } |
| 2239 | |
| 2240 | #if defined(CONFIG_DEBUG_FS) |
| 2241 | |
| 2242 | int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data) |
| 2243 | { |
| 2244 | struct kfd_topology_device *dev; |
| 2245 | unsigned int i = 0; |
| 2246 | int r = 0; |
| 2247 | |
| 2248 | down_read(&topology_lock); |
| 2249 | |
| 2250 | list_for_each_entry(dev, &topology_device_list, list) { |
| 2251 | if (!dev->gpu) { |
| 2252 | i++; |
| 2253 | continue; |
| 2254 | } |
| 2255 | |
| 2256 | seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id); |
| 2257 | r = dqm_debugfs_hqds(m, dev->gpu->dqm); |
| 2258 | if (r) |
| 2259 | break; |
| 2260 | } |
| 2261 | |
| 2262 | up_read(&topology_lock); |
| 2263 | |
| 2264 | return r; |
| 2265 | } |
| 2266 | |
| 2267 | int kfd_debugfs_rls_by_device(struct seq_file *m, void *data) |
| 2268 | { |
| 2269 | struct kfd_topology_device *dev; |
| 2270 | unsigned int i = 0; |
| 2271 | int r = 0; |
| 2272 | |
| 2273 | down_read(&topology_lock); |
| 2274 | |
| 2275 | list_for_each_entry(dev, &topology_device_list, list) { |
| 2276 | if (!dev->gpu) { |
| 2277 | i++; |
| 2278 | continue; |
| 2279 | } |
| 2280 | |
| 2281 | seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id); |
| 2282 | r = pm_debugfs_runlist(m, &dev->gpu->dqm->packet_mgr); |
| 2283 | if (r) |
| 2284 | break; |
| 2285 | } |
| 2286 | |
| 2287 | up_read(&topology_lock); |
| 2288 | |
| 2289 | return r; |
| 2290 | } |
| 2291 | |
| 2292 | #endif |