drivers/gpu/drm/amd/amdkfd/kfd_topology.c

   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_iommu.h"
  40 #include "kfd_svm.h"
  41 #include "amdgpu_amdkfd.h"
  42 #include "amdgpu_ras.h"
  43
  44 /* topology_device_list - Master list of all topology devices */
  45 static struct list_head topology_device_list;
  46 static struct kfd_system_properties sys_props;
  47
  48 static DECLARE_RWSEM(topology_lock);
  49 static uint32_t topology_crat_proximity_domain;
  50
  51 struct kfd_topology_device *kfd_topology_device_by_proximity_domain_no_lock(
  52                                                 uint32_t proximity_domain)
  53 {
  54         struct kfd_topology_device *top_dev;
  55         struct kfd_topology_device *device = NULL;
  56
  57         list_for_each_entry(top_dev, &topology_device_list, list)
  58                 if (top_dev->proximity_domain == proximity_domain) {
  59                         device = top_dev;
  60                         break;
  61                 }
  62
  63         return device;
  64 }
  65
  66 struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
  67                                                 uint32_t proximity_domain)
  68 {
  69         struct kfd_topology_device *device = NULL;
  70
  71         down_read(&topology_lock);
  72
  73         device = kfd_topology_device_by_proximity_domain_no_lock(
  74                                                         proximity_domain);
  75         up_read(&topology_lock);
  76
  77         return device;
  78 }
  79
  80 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id)
  81 {
  82         struct kfd_topology_device *top_dev = NULL;
  83         struct kfd_topology_device *ret = NULL;
  84
  85         down_read(&topology_lock);
  86
  87         list_for_each_entry(top_dev, &topology_device_list, list)
  88                 if (top_dev->gpu_id == gpu_id) {
  89                         ret = top_dev;
  90                         break;
  91                 }
  92
  93         up_read(&topology_lock);
  94
  95         return ret;
  96 }
  97
  98 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id)
  99 {
 100         struct kfd_topology_device *top_dev;
 101
 102         top_dev = kfd_topology_device_by_id(gpu_id);
 103         if (!top_dev)
 104                 return NULL;
 105
 106         return top_dev->gpu;
 107 }
 108
 109 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev)
 110 {
 111         struct kfd_topology_device *top_dev;
 112         struct kfd_dev *device = NULL;
 113
 114         down_read(&topology_lock);
 115
 116         list_for_each_entry(top_dev, &topology_device_list, list)
 117                 if (top_dev->gpu && top_dev->gpu->pdev == pdev) {
 118                         device = top_dev->gpu;
 119                         break;
 120                 }
 121
 122         up_read(&topology_lock);
 123
 124         return device;
 125 }
 126
 127 struct kfd_dev *kfd_device_by_adev(const struct amdgpu_device *adev)
 128 {
 129         struct kfd_topology_device *top_dev;
 130         struct kfd_dev *device = NULL;
 131
 132         down_read(&topology_lock);
 133
 134         list_for_each_entry(top_dev, &topology_device_list, list)
 135                 if (top_dev->gpu && top_dev->gpu->adev == adev) {
 136                         device = top_dev->gpu;
 137                         break;
 138                 }
 139
 140         up_read(&topology_lock);
 141
 142         return device;
 143 }
 144
 145 /* Called with write topology_lock acquired */
 146 static void kfd_release_topology_device(struct kfd_topology_device *dev)
 147 {
 148         struct kfd_mem_properties *mem;
 149         struct kfd_cache_properties *cache;
 150         struct kfd_iolink_properties *iolink;
 151         struct kfd_perf_properties *perf;
 152
 153         list_del(&dev->list);
 154
 155         while (dev->mem_props.next != &dev->mem_props) {
 156                 mem = container_of(dev->mem_props.next,
 157                                 struct kfd_mem_properties, list);
 158                 list_del(&mem->list);
 159                 kfree(mem);
 160         }
 161
 162         while (dev->cache_props.next != &dev->cache_props) {
 163                 cache = container_of(dev->cache_props.next,
 164                                 struct kfd_cache_properties, list);
 165                 list_del(&cache->list);
 166                 kfree(cache);
 167         }
 168
 169         while (dev->io_link_props.next != &dev->io_link_props) {
 170                 iolink = container_of(dev->io_link_props.next,
 171                                 struct kfd_iolink_properties, list);
 172                 list_del(&iolink->list);
 173                 kfree(iolink);
 174         }
 175
 176         while (dev->perf_props.next != &dev->perf_props) {
 177                 perf = container_of(dev->perf_props.next,
 178                                 struct kfd_perf_properties, list);
 179                 list_del(&perf->list);
 180                 kfree(perf);
 181         }
 182
 183         kfree(dev);
 184 }
 185
 186 void kfd_release_topology_device_list(struct list_head *device_list)
 187 {
 188         struct kfd_topology_device *dev;
 189
 190         while (!list_empty(device_list)) {
 191                 dev = list_first_entry(device_list,
 192                                        struct kfd_topology_device, list);
 193                 kfd_release_topology_device(dev);
 194         }
 195 }
 196
 197 static void kfd_release_live_view(void)
 198 {
 199         kfd_release_topology_device_list(&topology_device_list);
 200         memset(&sys_props, 0, sizeof(sys_props));
 201 }
 202
 203 struct kfd_topology_device *kfd_create_topology_device(
 204                                 struct list_head *device_list)
 205 {
 206         struct kfd_topology_device *dev;
 207
 208         dev = kfd_alloc_struct(dev);
 209         if (!dev) {
 210                 pr_err("No memory to allocate a topology device");
 211                 return NULL;
 212         }
 213
 214         INIT_LIST_HEAD(&dev->mem_props);
 215         INIT_LIST_HEAD(&dev->cache_props);
 216         INIT_LIST_HEAD(&dev->io_link_props);
 217         INIT_LIST_HEAD(&dev->perf_props);
 218
 219         list_add_tail(&dev->list, device_list);
 220
 221         return dev;
 222 }
 223
 224
 225 #define sysfs_show_gen_prop(buffer, offs, fmt, ...)             \
 226                 (offs += snprintf(buffer+offs, PAGE_SIZE-offs,  \
 227                                   fmt, __VA_ARGS__))
 228 #define sysfs_show_32bit_prop(buffer, offs, name, value) \
 229                 sysfs_show_gen_prop(buffer, offs, "%s %u\n", name, value)
 230 #define sysfs_show_64bit_prop(buffer, offs, name, value) \
 231                 sysfs_show_gen_prop(buffer, offs, "%s %llu\n", name, value)
 232 #define sysfs_show_32bit_val(buffer, offs, value) \
 233                 sysfs_show_gen_prop(buffer, offs, "%u\n", value)
 234 #define sysfs_show_str_val(buffer, offs, value) \
 235                 sysfs_show_gen_prop(buffer, offs, "%s\n", value)
 236
 237 static ssize_t sysprops_show(struct kobject *kobj, struct attribute *attr,
 238                 char *buffer)
 239 {
 240         int offs = 0;
 241
 242         /* Making sure that the buffer is an empty string */
 243         buffer[0] = 0;
 244
 245         if (attr == &sys_props.attr_genid) {
 246                 sysfs_show_32bit_val(buffer, offs,
 247                                      sys_props.generation_count);
 248         } else if (attr == &sys_props.attr_props) {
 249                 sysfs_show_64bit_prop(buffer, offs, "platform_oem",
 250                                       sys_props.platform_oem);
 251                 sysfs_show_64bit_prop(buffer, offs, "platform_id",
 252                                       sys_props.platform_id);
 253                 sysfs_show_64bit_prop(buffer, offs, "platform_rev",
 254                                       sys_props.platform_rev);
 255         } else {
 256                 offs = -EINVAL;
 257         }
 258
 259         return offs;
 260 }
 261
 262 static void kfd_topology_kobj_release(struct kobject *kobj)
 263 {
 264         kfree(kobj);
 265 }
 266
 267 static const struct sysfs_ops sysprops_ops = {
 268         .show = sysprops_show,
 269 };
 270
 271 static struct kobj_type sysprops_type = {
 272         .release = kfd_topology_kobj_release,
 273         .sysfs_ops = &sysprops_ops,
 274 };
 275
 276 static ssize_t iolink_show(struct kobject *kobj, struct attribute *attr,
 277                 char *buffer)
 278 {
 279         int offs = 0;
 280         struct kfd_iolink_properties *iolink;
 281
 282         /* Making sure that the buffer is an empty string */
 283         buffer[0] = 0;
 284
 285         iolink = container_of(attr, struct kfd_iolink_properties, attr);
 286         if (iolink->gpu && kfd_devcgroup_check_permission(iolink->gpu))
 287                 return -EPERM;
 288         sysfs_show_32bit_prop(buffer, offs, "type", iolink->iolink_type);
 289         sysfs_show_32bit_prop(buffer, offs, "version_major", iolink->ver_maj);
 290         sysfs_show_32bit_prop(buffer, offs, "version_minor", iolink->ver_min);
 291         sysfs_show_32bit_prop(buffer, offs, "node_from", iolink->node_from);
 292         sysfs_show_32bit_prop(buffer, offs, "node_to", iolink->node_to);
 293         sysfs_show_32bit_prop(buffer, offs, "weight", iolink->weight);
 294         sysfs_show_32bit_prop(buffer, offs, "min_latency", iolink->min_latency);
 295         sysfs_show_32bit_prop(buffer, offs, "max_latency", iolink->max_latency);
 296         sysfs_show_32bit_prop(buffer, offs, "min_bandwidth",
 297                               iolink->min_bandwidth);
 298         sysfs_show_32bit_prop(buffer, offs, "max_bandwidth",
 299                               iolink->max_bandwidth);
 300         sysfs_show_32bit_prop(buffer, offs, "recommended_transfer_size",
 301                               iolink->rec_transfer_size);
 302         sysfs_show_32bit_prop(buffer, offs, "flags", iolink->flags);
 303
 304         return offs;
 305 }
 306
 307 static const struct sysfs_ops iolink_ops = {
 308         .show = iolink_show,
 309 };
 310
 311 static struct kobj_type iolink_type = {
 312         .release = kfd_topology_kobj_release,
 313         .sysfs_ops = &iolink_ops,
 314 };
 315
 316 static ssize_t mem_show(struct kobject *kobj, struct attribute *attr,
 317                 char *buffer)
 318 {
 319         int offs = 0;
 320         struct kfd_mem_properties *mem;
 321
 322         /* Making sure that the buffer is an empty string */
 323         buffer[0] = 0;
 324
 325         mem = container_of(attr, struct kfd_mem_properties, attr);
 326         if (mem->gpu && kfd_devcgroup_check_permission(mem->gpu))
 327                 return -EPERM;
 328         sysfs_show_32bit_prop(buffer, offs, "heap_type", mem->heap_type);
 329         sysfs_show_64bit_prop(buffer, offs, "size_in_bytes",
 330                               mem->size_in_bytes);
 331         sysfs_show_32bit_prop(buffer, offs, "flags", mem->flags);
 332         sysfs_show_32bit_prop(buffer, offs, "width", mem->width);
 333         sysfs_show_32bit_prop(buffer, offs, "mem_clk_max",
 334                               mem->mem_clk_max);
 335
 336         return offs;
 337 }
 338
 339 static const struct sysfs_ops mem_ops = {
 340         .show = mem_show,
 341 };
 342
 343 static struct kobj_type mem_type = {
 344         .release = kfd_topology_kobj_release,
 345         .sysfs_ops = &mem_ops,
 346 };
 347
 348 static ssize_t kfd_cache_show(struct kobject *kobj, struct attribute *attr,
 349                 char *buffer)
 350 {
 351         int offs = 0;
 352         uint32_t i, j;
 353         struct kfd_cache_properties *cache;
 354
 355         /* Making sure that the buffer is an empty string */
 356         buffer[0] = 0;
 357
 358         cache = container_of(attr, struct kfd_cache_properties, attr);
 359         if (cache->gpu && kfd_devcgroup_check_permission(cache->gpu))
 360                 return -EPERM;
 361         sysfs_show_32bit_prop(buffer, offs, "processor_id_low",
 362                         cache->processor_id_low);
 363         sysfs_show_32bit_prop(buffer, offs, "level", cache->cache_level);
 364         sysfs_show_32bit_prop(buffer, offs, "size", cache->cache_size);
 365         sysfs_show_32bit_prop(buffer, offs, "cache_line_size",
 366                               cache->cacheline_size);
 367         sysfs_show_32bit_prop(buffer, offs, "cache_lines_per_tag",
 368                               cache->cachelines_per_tag);
 369         sysfs_show_32bit_prop(buffer, offs, "association", cache->cache_assoc);
 370         sysfs_show_32bit_prop(buffer, offs, "latency", cache->cache_latency);
 371         sysfs_show_32bit_prop(buffer, offs, "type", cache->cache_type);
 372         offs += snprintf(buffer+offs, PAGE_SIZE-offs, "sibling_map ");
 373         for (i = 0; i < CRAT_SIBLINGMAP_SIZE; i++)
 374                 for (j = 0; j < sizeof(cache->sibling_map[0])*8; j++)
 375                         /* Check each bit */
 376                         offs += snprintf(buffer+offs, PAGE_SIZE-offs, "%d,",
 377                                          (cache->sibling_map[i] >> j) & 1);
 378
 379         /* Replace the last "," with end of line */
 380         buffer[offs-1] = '\n';
 381         return offs;
 382 }
 383
 384 static const struct sysfs_ops cache_ops = {
 385         .show = kfd_cache_show,
 386 };
 387
 388 static struct kobj_type cache_type = {
 389         .release = kfd_topology_kobj_release,
 390         .sysfs_ops = &cache_ops,
 391 };
 392
 393 /****** Sysfs of Performance Counters ******/
 394
 395 struct kfd_perf_attr {
 396         struct kobj_attribute attr;
 397         uint32_t data;
 398 };
 399
 400 static ssize_t perf_show(struct kobject *kobj, struct kobj_attribute *attrs,
 401                         char *buf)
 402 {
 403         int offs = 0;
 404         struct kfd_perf_attr *attr;
 405
 406         buf[0] = 0;
 407         attr = container_of(attrs, struct kfd_perf_attr, attr);
 408         if (!attr->data) /* invalid data for PMC */
 409                 return 0;
 410         else
 411                 return sysfs_show_32bit_val(buf, offs, attr->data);
 412 }
 413
 414 #define KFD_PERF_DESC(_name, _data)                     \
 415 {                                                       \
 416         .attr  = __ATTR(_name, 0444, perf_show, NULL),  \
 417         .data = _data,                                  \
 418 }
 419
 420 static struct kfd_perf_attr perf_attr_iommu[] = {
 421         KFD_PERF_DESC(max_concurrent, 0),
 422         KFD_PERF_DESC(num_counters, 0),
 423         KFD_PERF_DESC(counter_ids, 0),
 424 };
 425 /****************************************/
 426
 427 static ssize_t node_show(struct kobject *kobj, struct attribute *attr,
 428                 char *buffer)
 429 {
 430         int offs = 0;
 431         struct kfd_topology_device *dev;
 432         uint32_t log_max_watch_addr;
 433
 434         /* Making sure that the buffer is an empty string */
 435         buffer[0] = 0;
 436
 437         if (strcmp(attr->name, "gpu_id") == 0) {
 438                 dev = container_of(attr, struct kfd_topology_device,
 439                                 attr_gpuid);
 440                 if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu))
 441                         return -EPERM;
 442                 return sysfs_show_32bit_val(buffer, offs, dev->gpu_id);
 443         }
 444
 445         if (strcmp(attr->name, "name") == 0) {
 446                 dev = container_of(attr, struct kfd_topology_device,
 447                                 attr_name);
 448
 449                 if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu))
 450                         return -EPERM;
 451                 return sysfs_show_str_val(buffer, offs, dev->node_props.name);
 452         }
 453
 454         dev = container_of(attr, struct kfd_topology_device,
 455                         attr_props);
 456         if (dev->gpu && kfd_devcgroup_check_permission(dev->gpu))
 457                 return -EPERM;
 458         sysfs_show_32bit_prop(buffer, offs, "cpu_cores_count",
 459                               dev->node_props.cpu_cores_count);
 460         sysfs_show_32bit_prop(buffer, offs, "simd_count",
 461                               dev->gpu ? dev->node_props.simd_count : 0);
 462         sysfs_show_32bit_prop(buffer, offs, "mem_banks_count",
 463                               dev->node_props.mem_banks_count);
 464         sysfs_show_32bit_prop(buffer, offs, "caches_count",
 465                               dev->node_props.caches_count);
 466         sysfs_show_32bit_prop(buffer, offs, "io_links_count",
 467                               dev->node_props.io_links_count);
 468         sysfs_show_32bit_prop(buffer, offs, "cpu_core_id_base",
 469                               dev->node_props.cpu_core_id_base);
 470         sysfs_show_32bit_prop(buffer, offs, "simd_id_base",
 471                               dev->node_props.simd_id_base);
 472         sysfs_show_32bit_prop(buffer, offs, "max_waves_per_simd",
 473                               dev->node_props.max_waves_per_simd);
 474         sysfs_show_32bit_prop(buffer, offs, "lds_size_in_kb",
 475                               dev->node_props.lds_size_in_kb);
 476         sysfs_show_32bit_prop(buffer, offs, "gds_size_in_kb",
 477                               dev->node_props.gds_size_in_kb);
 478         sysfs_show_32bit_prop(buffer, offs, "num_gws",
 479                               dev->node_props.num_gws);
 480         sysfs_show_32bit_prop(buffer, offs, "wave_front_size",
 481                               dev->node_props.wave_front_size);
 482         sysfs_show_32bit_prop(buffer, offs, "array_count",
 483                               dev->node_props.array_count);
 484         sysfs_show_32bit_prop(buffer, offs, "simd_arrays_per_engine",
 485                               dev->node_props.simd_arrays_per_engine);
 486         sysfs_show_32bit_prop(buffer, offs, "cu_per_simd_array",
 487                               dev->node_props.cu_per_simd_array);
 488         sysfs_show_32bit_prop(buffer, offs, "simd_per_cu",
 489                               dev->node_props.simd_per_cu);
 490         sysfs_show_32bit_prop(buffer, offs, "max_slots_scratch_cu",
 491                               dev->node_props.max_slots_scratch_cu);
 492         sysfs_show_32bit_prop(buffer, offs, "gfx_target_version",
 493                               dev->node_props.gfx_target_version);
 494         sysfs_show_32bit_prop(buffer, offs, "vendor_id",
 495                               dev->node_props.vendor_id);
 496         sysfs_show_32bit_prop(buffer, offs, "device_id",
 497                               dev->node_props.device_id);
 498         sysfs_show_32bit_prop(buffer, offs, "location_id",
 499                               dev->node_props.location_id);
 500         sysfs_show_32bit_prop(buffer, offs, "domain",
 501                               dev->node_props.domain);
 502         sysfs_show_32bit_prop(buffer, offs, "drm_render_minor",
 503                               dev->node_props.drm_render_minor);
 504         sysfs_show_64bit_prop(buffer, offs, "hive_id",
 505                               dev->node_props.hive_id);
 506         sysfs_show_32bit_prop(buffer, offs, "num_sdma_engines",
 507                               dev->node_props.num_sdma_engines);
 508         sysfs_show_32bit_prop(buffer, offs, "num_sdma_xgmi_engines",
 509                               dev->node_props.num_sdma_xgmi_engines);
 510         sysfs_show_32bit_prop(buffer, offs, "num_sdma_queues_per_engine",
 511                               dev->node_props.num_sdma_queues_per_engine);
 512         sysfs_show_32bit_prop(buffer, offs, "num_cp_queues",
 513                               dev->node_props.num_cp_queues);
 514
 515         if (dev->gpu) {
 516                 log_max_watch_addr =
 517                         __ilog2_u32(dev->gpu->device_info.num_of_watch_points);
 518
 519                 if (log_max_watch_addr) {
 520                         dev->node_props.capability |=
 521                                         HSA_CAP_WATCH_POINTS_SUPPORTED;
 522
 523                         dev->node_props.capability |=
 524                                 ((log_max_watch_addr <<
 525                                         HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) &
 526                                 HSA_CAP_WATCH_POINTS_TOTALBITS_MASK);
 527                 }
 528
 529                 if (dev->gpu->adev->asic_type == CHIP_TONGA)
 530                         dev->node_props.capability |=
 531                                         HSA_CAP_AQL_QUEUE_DOUBLE_MAP;
 532
 533                 sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_fcompute",
 534                         dev->node_props.max_engine_clk_fcompute);
 535
 536                 sysfs_show_64bit_prop(buffer, offs, "local_mem_size", 0ULL);
 537
 538                 sysfs_show_32bit_prop(buffer, offs, "fw_version",
 539                                       dev->gpu->mec_fw_version);
 540                 sysfs_show_32bit_prop(buffer, offs, "capability",
 541                                       dev->node_props.capability);
 542                 sysfs_show_32bit_prop(buffer, offs, "sdma_fw_version",
 543                                       dev->gpu->sdma_fw_version);
 544                 sysfs_show_64bit_prop(buffer, offs, "unique_id",
 545                                       dev->gpu->adev->unique_id);
 546
 547         }
 548
 549         return sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_ccompute",
 550                                      cpufreq_quick_get_max(0)/1000);
 551 }
 552
 553 static const struct sysfs_ops node_ops = {
 554         .show = node_show,
 555 };
 556
 557 static struct kobj_type node_type = {
 558         .release = kfd_topology_kobj_release,
 559         .sysfs_ops = &node_ops,
 560 };
 561
 562 static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr)
 563 {
 564         sysfs_remove_file(kobj, attr);
 565         kobject_del(kobj);
 566         kobject_put(kobj);
 567 }
 568
 569 static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev)
 570 {
 571         struct kfd_iolink_properties *iolink;
 572         struct kfd_cache_properties *cache;
 573         struct kfd_mem_properties *mem;
 574         struct kfd_perf_properties *perf;
 575
 576         if (dev->kobj_iolink) {
 577                 list_for_each_entry(iolink, &dev->io_link_props, list)
 578                         if (iolink->kobj) {
 579                                 kfd_remove_sysfs_file(iolink->kobj,
 580                                                         &iolink->attr);
 581                                 iolink->kobj = NULL;
 582                         }
 583                 kobject_del(dev->kobj_iolink);
 584                 kobject_put(dev->kobj_iolink);
 585                 dev->kobj_iolink = NULL;
 586         }
 587
 588         if (dev->kobj_cache) {
 589                 list_for_each_entry(cache, &dev->cache_props, list)
 590                         if (cache->kobj) {
 591                                 kfd_remove_sysfs_file(cache->kobj,
 592                                                         &cache->attr);
 593                                 cache->kobj = NULL;
 594                         }
 595                 kobject_del(dev->kobj_cache);
 596                 kobject_put(dev->kobj_cache);
 597                 dev->kobj_cache = NULL;
 598         }
 599
 600         if (dev->kobj_mem) {
 601                 list_for_each_entry(mem, &dev->mem_props, list)
 602                         if (mem->kobj) {
 603                                 kfd_remove_sysfs_file(mem->kobj, &mem->attr);
 604                                 mem->kobj = NULL;
 605                         }
 606                 kobject_del(dev->kobj_mem);
 607                 kobject_put(dev->kobj_mem);
 608                 dev->kobj_mem = NULL;
 609         }
 610
 611         if (dev->kobj_perf) {
 612                 list_for_each_entry(perf, &dev->perf_props, list) {
 613                         kfree(perf->attr_group);
 614                         perf->attr_group = NULL;
 615                 }
 616                 kobject_del(dev->kobj_perf);
 617                 kobject_put(dev->kobj_perf);
 618                 dev->kobj_perf = NULL;
 619         }
 620
 621         if (dev->kobj_node) {
 622                 sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid);
 623                 sysfs_remove_file(dev->kobj_node, &dev->attr_name);
 624                 sysfs_remove_file(dev->kobj_node, &dev->attr_props);
 625                 kobject_del(dev->kobj_node);
 626                 kobject_put(dev->kobj_node);
 627                 dev->kobj_node = NULL;
 628         }
 629 }
 630
 631 static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev,
 632                 uint32_t id)
 633 {
 634         struct kfd_iolink_properties *iolink;
 635         struct kfd_cache_properties *cache;
 636         struct kfd_mem_properties *mem;
 637         struct kfd_perf_properties *perf;
 638         int ret;
 639         uint32_t i, num_attrs;
 640         struct attribute **attrs;
 641
 642         if (WARN_ON(dev->kobj_node))
 643                 return -EEXIST;
 644
 645         /*
 646          * Creating the sysfs folders
 647          */
 648         dev->kobj_node = kfd_alloc_struct(dev->kobj_node);
 649         if (!dev->kobj_node)
 650                 return -ENOMEM;
 651
 652         ret = kobject_init_and_add(dev->kobj_node, &node_type,
 653                         sys_props.kobj_nodes, "%d", id);
 654         if (ret < 0) {
 655                 kobject_put(dev->kobj_node);
 656                 return ret;
 657         }
 658
 659         dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node);
 660         if (!dev->kobj_mem)
 661                 return -ENOMEM;
 662
 663         dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node);
 664         if (!dev->kobj_cache)
 665                 return -ENOMEM;
 666
 667         dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node);
 668         if (!dev->kobj_iolink)
 669                 return -ENOMEM;
 670
 671         dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node);
 672         if (!dev->kobj_perf)
 673                 return -ENOMEM;
 674
 675         /*
 676          * Creating sysfs files for node properties
 677          */
 678         dev->attr_gpuid.name = "gpu_id";
 679         dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE;
 680         sysfs_attr_init(&dev->attr_gpuid);
 681         dev->attr_name.name = "name";
 682         dev->attr_name.mode = KFD_SYSFS_FILE_MODE;
 683         sysfs_attr_init(&dev->attr_name);
 684         dev->attr_props.name = "properties";
 685         dev->attr_props.mode = KFD_SYSFS_FILE_MODE;
 686         sysfs_attr_init(&dev->attr_props);
 687         ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid);
 688         if (ret < 0)
 689                 return ret;
 690         ret = sysfs_create_file(dev->kobj_node, &dev->attr_name);
 691         if (ret < 0)
 692                 return ret;
 693         ret = sysfs_create_file(dev->kobj_node, &dev->attr_props);
 694         if (ret < 0)
 695                 return ret;
 696
 697         i = 0;
 698         list_for_each_entry(mem, &dev->mem_props, list) {
 699                 mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
 700                 if (!mem->kobj)
 701                         return -ENOMEM;
 702                 ret = kobject_init_and_add(mem->kobj, &mem_type,
 703                                 dev->kobj_mem, "%d", i);
 704                 if (ret < 0) {
 705                         kobject_put(mem->kobj);
 706                         return ret;
 707                 }
 708
 709                 mem->attr.name = "properties";
 710                 mem->attr.mode = KFD_SYSFS_FILE_MODE;
 711                 sysfs_attr_init(&mem->attr);
 712                 ret = sysfs_create_file(mem->kobj, &mem->attr);
 713                 if (ret < 0)
 714                         return ret;
 715                 i++;
 716         }
 717
 718         i = 0;
 719         list_for_each_entry(cache, &dev->cache_props, list) {
 720                 cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
 721                 if (!cache->kobj)
 722                         return -ENOMEM;
 723                 ret = kobject_init_and_add(cache->kobj, &cache_type,
 724                                 dev->kobj_cache, "%d", i);
 725                 if (ret < 0) {
 726                         kobject_put(cache->kobj);
 727                         return ret;
 728                 }
 729
 730                 cache->attr.name = "properties";
 731                 cache->attr.mode = KFD_SYSFS_FILE_MODE;
 732                 sysfs_attr_init(&cache->attr);
 733                 ret = sysfs_create_file(cache->kobj, &cache->attr);
 734                 if (ret < 0)
 735                         return ret;
 736                 i++;
 737         }
 738
 739         i = 0;
 740         list_for_each_entry(iolink, &dev->io_link_props, list) {
 741                 iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
 742                 if (!iolink->kobj)
 743                         return -ENOMEM;
 744                 ret = kobject_init_and_add(iolink->kobj, &iolink_type,
 745                                 dev->kobj_iolink, "%d", i);
 746                 if (ret < 0) {
 747                         kobject_put(iolink->kobj);
 748                         return ret;
 749                 }
 750
 751                 iolink->attr.name = "properties";
 752                 iolink->attr.mode = KFD_SYSFS_FILE_MODE;
 753                 sysfs_attr_init(&iolink->attr);
 754                 ret = sysfs_create_file(iolink->kobj, &iolink->attr);
 755                 if (ret < 0)
 756                         return ret;
 757                 i++;
 758         }
 759
 760         /* All hardware blocks have the same number of attributes. */
 761         num_attrs = ARRAY_SIZE(perf_attr_iommu);
 762         list_for_each_entry(perf, &dev->perf_props, list) {
 763                 perf->attr_group = kzalloc(sizeof(struct kfd_perf_attr)
 764                         * num_attrs + sizeof(struct attribute_group),
 765                         GFP_KERNEL);
 766                 if (!perf->attr_group)
 767                         return -ENOMEM;
 768
 769                 attrs = (struct attribute **)(perf->attr_group + 1);
 770                 if (!strcmp(perf->block_name, "iommu")) {
 771                 /* Information of IOMMU's num_counters and counter_ids is shown
 772                  * under /sys/bus/event_source/devices/amd_iommu. We don't
 773                  * duplicate here.
 774                  */
 775                         perf_attr_iommu[0].data = perf->max_concurrent;
 776                         for (i = 0; i < num_attrs; i++)
 777                                 attrs[i] = &perf_attr_iommu[i].attr.attr;
 778                 }
 779                 perf->attr_group->name = perf->block_name;
 780                 perf->attr_group->attrs = attrs;
 781                 ret = sysfs_create_group(dev->kobj_perf, perf->attr_group);
 782                 if (ret < 0)
 783                         return ret;
 784         }
 785
 786         return 0;
 787 }
 788
 789 /* Called with write topology lock acquired */
 790 static int kfd_build_sysfs_node_tree(void)
 791 {
 792         struct kfd_topology_device *dev;
 793         int ret;
 794         uint32_t i = 0;
 795
 796         list_for_each_entry(dev, &topology_device_list, list) {
 797                 ret = kfd_build_sysfs_node_entry(dev, i);
 798                 if (ret < 0)
 799                         return ret;
 800                 i++;
 801         }
 802
 803         return 0;
 804 }
 805
 806 /* Called with write topology lock acquired */
 807 static void kfd_remove_sysfs_node_tree(void)
 808 {
 809         struct kfd_topology_device *dev;
 810
 811         list_for_each_entry(dev, &topology_device_list, list)
 812                 kfd_remove_sysfs_node_entry(dev);
 813 }
 814
 815 static int kfd_topology_update_sysfs(void)
 816 {
 817         int ret;
 818
 819         if (!sys_props.kobj_topology) {
 820                 sys_props.kobj_topology =
 821                                 kfd_alloc_struct(sys_props.kobj_topology);
 822                 if (!sys_props.kobj_topology)
 823                         return -ENOMEM;
 824
 825                 ret = kobject_init_and_add(sys_props.kobj_topology,
 826                                 &sysprops_type,  &kfd_device->kobj,
 827                                 "topology");
 828                 if (ret < 0) {
 829                         kobject_put(sys_props.kobj_topology);
 830                         return ret;
 831                 }
 832
 833                 sys_props.kobj_nodes = kobject_create_and_add("nodes",
 834                                 sys_props.kobj_topology);
 835                 if (!sys_props.kobj_nodes)
 836                         return -ENOMEM;
 837
 838                 sys_props.attr_genid.name = "generation_id";
 839                 sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE;
 840                 sysfs_attr_init(&sys_props.attr_genid);
 841                 ret = sysfs_create_file(sys_props.kobj_topology,
 842                                 &sys_props.attr_genid);
 843                 if (ret < 0)
 844                         return ret;
 845
 846                 sys_props.attr_props.name = "system_properties";
 847                 sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE;
 848                 sysfs_attr_init(&sys_props.attr_props);
 849                 ret = sysfs_create_file(sys_props.kobj_topology,
 850                                 &sys_props.attr_props);
 851                 if (ret < 0)
 852                         return ret;
 853         }
 854
 855         kfd_remove_sysfs_node_tree();
 856
 857         return kfd_build_sysfs_node_tree();
 858 }
 859
 860 static void kfd_topology_release_sysfs(void)
 861 {
 862         kfd_remove_sysfs_node_tree();
 863         if (sys_props.kobj_topology) {
 864                 sysfs_remove_file(sys_props.kobj_topology,
 865                                 &sys_props.attr_genid);
 866                 sysfs_remove_file(sys_props.kobj_topology,
 867                                 &sys_props.attr_props);
 868                 if (sys_props.kobj_nodes) {
 869                         kobject_del(sys_props.kobj_nodes);
 870                         kobject_put(sys_props.kobj_nodes);
 871                         sys_props.kobj_nodes = NULL;
 872                 }
 873                 kobject_del(sys_props.kobj_topology);
 874                 kobject_put(sys_props.kobj_topology);
 875                 sys_props.kobj_topology = NULL;
 876         }
 877 }
 878
 879 /* Called with write topology_lock acquired */
 880 static void kfd_topology_update_device_list(struct list_head *temp_list,
 881                                         struct list_head *master_list)
 882 {
 883         while (!list_empty(temp_list)) {
 884                 list_move_tail(temp_list->next, master_list);
 885                 sys_props.num_devices++;
 886         }
 887 }
 888
 889 static void kfd_debug_print_topology(void)
 890 {
 891         struct kfd_topology_device *dev;
 892
 893         down_read(&topology_lock);
 894
 895         dev = list_last_entry(&topology_device_list,
 896                         struct kfd_topology_device, list);
 897         if (dev) {
 898                 if (dev->node_props.cpu_cores_count &&
 899                                 dev->node_props.simd_count) {
 900                         pr_info("Topology: Add APU node [0x%0x:0x%0x]\n",
 901                                 dev->node_props.device_id,
 902                                 dev->node_props.vendor_id);
 903                 } else if (dev->node_props.cpu_cores_count)
 904                         pr_info("Topology: Add CPU node\n");
 905                 else if (dev->node_props.simd_count)
 906                         pr_info("Topology: Add dGPU node [0x%0x:0x%0x]\n",
 907                                 dev->node_props.device_id,
 908                                 dev->node_props.vendor_id);
 909         }
 910         up_read(&topology_lock);
 911 }
 912
 913 /* Helper function for intializing platform_xx members of
 914  * kfd_system_properties. Uses OEM info from the last CPU/APU node.
 915  */
 916 static void kfd_update_system_properties(void)
 917 {
 918         struct kfd_topology_device *dev;
 919
 920         down_read(&topology_lock);
 921         dev = list_last_entry(&topology_device_list,
 922                         struct kfd_topology_device, list);
 923         if (dev) {
 924                 sys_props.platform_id =
 925                         (*((uint64_t *)dev->oem_id)) & CRAT_OEMID_64BIT_MASK;
 926                 sys_props.platform_oem = *((uint64_t *)dev->oem_table_id);
 927                 sys_props.platform_rev = dev->oem_revision;
 928         }
 929         up_read(&topology_lock);
 930 }
 931
 932 static void find_system_memory(const struct dmi_header *dm,
 933         void *private)
 934 {
 935         struct kfd_mem_properties *mem;
 936         u16 mem_width, mem_clock;
 937         struct kfd_topology_device *kdev =
 938                 (struct kfd_topology_device *)private;
 939         const u8 *dmi_data = (const u8 *)(dm + 1);
 940
 941         if (dm->type == DMI_ENTRY_MEM_DEVICE && dm->length >= 0x15) {
 942                 mem_width = (u16)(*(const u16 *)(dmi_data + 0x6));
 943                 mem_clock = (u16)(*(const u16 *)(dmi_data + 0x11));
 944                 list_for_each_entry(mem, &kdev->mem_props, list) {
 945                         if (mem_width != 0xFFFF && mem_width != 0)
 946                                 mem->width = mem_width;
 947                         if (mem_clock != 0)
 948                                 mem->mem_clk_max = mem_clock;
 949                 }
 950         }
 951 }
 952
 953 /*
 954  * Performance counters information is not part of CRAT but we would like to
 955  * put them in the sysfs under topology directory for Thunk to get the data.
 956  * This function is called before updating the sysfs.
 957  */
 958 static int kfd_add_perf_to_topology(struct kfd_topology_device *kdev)
 959 {
 960         /* These are the only counters supported so far */
 961         return kfd_iommu_add_perf_counters(kdev);
 962 }
 963
 964 /* kfd_add_non_crat_information - Add information that is not currently
 965  *      defined in CRAT but is necessary for KFD topology
 966  * @dev - topology device to which addition info is added
 967  */
 968 static void kfd_add_non_crat_information(struct kfd_topology_device *kdev)
 969 {
 970         /* Check if CPU only node. */
 971         if (!kdev->gpu) {
 972                 /* Add system memory information */
 973                 dmi_walk(find_system_memory, kdev);
 974         }
 975         /* TODO: For GPU node, rearrange code from kfd_topology_add_device */
 976 }
 977
 978 /* kfd_is_acpi_crat_invalid - CRAT from ACPI is valid only for AMD APU devices.
 979  *      Ignore CRAT for all other devices. AMD APU is identified if both CPU
 980  *      and GPU cores are present.
 981  * @device_list - topology device list created by parsing ACPI CRAT table.
 982  * @return - TRUE if invalid, FALSE is valid.
 983  */
 984 static bool kfd_is_acpi_crat_invalid(struct list_head *device_list)
 985 {
 986         struct kfd_topology_device *dev;
 987
 988         list_for_each_entry(dev, device_list, list) {
 989                 if (dev->node_props.cpu_cores_count &&
 990                         dev->node_props.simd_count)
 991                         return false;
 992         }
 993         pr_info("Ignoring ACPI CRAT on non-APU system\n");
 994         return true;
 995 }
 996
 997 int kfd_topology_init(void)
 998 {
 999         void *crat_image = NULL;
1000         size_t image_size = 0;
1001         int ret;
1002         struct list_head temp_topology_device_list;
1003         int cpu_only_node = 0;
1004         struct kfd_topology_device *kdev;
1005         int proximity_domain;
1006
1007         /* topology_device_list - Master list of all topology devices
1008          * temp_topology_device_list - temporary list created while parsing CRAT
1009          * or VCRAT. Once parsing is complete the contents of list is moved to
1010          * topology_device_list
1011          */
1012
1013         /* Initialize the head for the both the lists */
1014         INIT_LIST_HEAD(&topology_device_list);
1015         INIT_LIST_HEAD(&temp_topology_device_list);
1016         init_rwsem(&topology_lock);
1017
1018         memset(&sys_props, 0, sizeof(sys_props));
1019
1020         /* Proximity domains in ACPI CRAT tables start counting at
1021          * 0. The same should be true for virtual CRAT tables created
1022          * at this stage. GPUs added later in kfd_topology_add_device
1023          * use a counter.
1024          */
1025         proximity_domain = 0;
1026
1027         /*
1028          * Get the CRAT image from the ACPI. If ACPI doesn't have one
1029          * or if ACPI CRAT is invalid create a virtual CRAT.
1030          * NOTE: The current implementation expects all AMD APUs to have
1031          *      CRAT. If no CRAT is available, it is assumed to be a CPU
1032          */
1033         ret = kfd_create_crat_image_acpi(&crat_image, &image_size);
1034         if (!ret) {
1035                 ret = kfd_parse_crat_table(crat_image,
1036                                            &temp_topology_device_list,
1037                                            proximity_domain);
1038                 if (ret ||
1039                     kfd_is_acpi_crat_invalid(&temp_topology_device_list)) {
1040                         kfd_release_topology_device_list(
1041                                 &temp_topology_device_list);
1042                         kfd_destroy_crat_image(crat_image);
1043                         crat_image = NULL;
1044                 }
1045         }
1046
1047         if (!crat_image) {
1048                 ret = kfd_create_crat_image_virtual(&crat_image, &image_size,
1049                                                     COMPUTE_UNIT_CPU, NULL,
1050                                                     proximity_domain);
1051                 cpu_only_node = 1;
1052                 if (ret) {
1053                         pr_err("Error creating VCRAT table for CPU\n");
1054                         return ret;
1055                 }
1056
1057                 ret = kfd_parse_crat_table(crat_image,
1058                                            &temp_topology_device_list,
1059                                            proximity_domain);
1060                 if (ret) {
1061                         pr_err("Error parsing VCRAT table for CPU\n");
1062                         goto err;
1063                 }
1064         }
1065
1066         kdev = list_first_entry(&temp_topology_device_list,
1067                                 struct kfd_topology_device, list);
1068         kfd_add_perf_to_topology(kdev);
1069
1070         down_write(&topology_lock);
1071         kfd_topology_update_device_list(&temp_topology_device_list,
1072                                         &topology_device_list);
1073         topology_crat_proximity_domain = sys_props.num_devices-1;
1074         ret = kfd_topology_update_sysfs();
1075         up_write(&topology_lock);
1076
1077         if (!ret) {
1078                 sys_props.generation_count++;
1079                 kfd_update_system_properties();
1080                 kfd_debug_print_topology();
1081         } else
1082                 pr_err("Failed to update topology in sysfs ret=%d\n", ret);
1083
1084         /* For nodes with GPU, this information gets added
1085          * when GPU is detected (kfd_topology_add_device).
1086          */
1087         if (cpu_only_node) {
1088                 /* Add additional information to CPU only node created above */
1089                 down_write(&topology_lock);
1090                 kdev = list_first_entry(&topology_device_list,
1091                                 struct kfd_topology_device, list);
1092                 up_write(&topology_lock);
1093                 kfd_add_non_crat_information(kdev);
1094         }
1095
1096 err:
1097         kfd_destroy_crat_image(crat_image);
1098         return ret;
1099 }
1100
1101 void kfd_topology_shutdown(void)
1102 {
1103         down_write(&topology_lock);
1104         kfd_topology_release_sysfs();
1105         kfd_release_live_view();
1106         up_write(&topology_lock);
1107 }
1108
1109 static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu)
1110 {
1111         uint32_t hashout;
1112         uint32_t buf[7];
1113         uint64_t local_mem_size;
1114         int i;
1115         struct kfd_local_mem_info local_mem_info;
1116
1117         if (!gpu)
1118                 return 0;
1119
1120         amdgpu_amdkfd_get_local_mem_info(gpu->adev, &local_mem_info);
1121
1122         local_mem_size = local_mem_info.local_mem_size_private +
1123                         local_mem_info.local_mem_size_public;
1124
1125         buf[0] = gpu->pdev->devfn;
1126         buf[1] = gpu->pdev->subsystem_vendor |
1127                 (gpu->pdev->subsystem_device << 16);
1128         buf[2] = pci_domain_nr(gpu->pdev->bus);
1129         buf[3] = gpu->pdev->device;
1130         buf[4] = gpu->pdev->bus->number;
1131         buf[5] = lower_32_bits(local_mem_size);
1132         buf[6] = upper_32_bits(local_mem_size);
1133
1134         for (i = 0, hashout = 0; i < 7; i++)
1135                 hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH);
1136
1137         return hashout;
1138 }
1139 /* kfd_assign_gpu - Attach @gpu to the correct kfd topology device. If
1140  *              the GPU device is not already present in the topology device
1141  *              list then return NULL. This means a new topology device has to
1142  *              be created for this GPU.
1143  */
1144 static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu)
1145 {
1146         struct kfd_topology_device *dev;
1147         struct kfd_topology_device *out_dev = NULL;
1148         struct kfd_mem_properties *mem;
1149         struct kfd_cache_properties *cache;
1150         struct kfd_iolink_properties *iolink;
1151
1152         down_write(&topology_lock);
1153         list_for_each_entry(dev, &topology_device_list, list) {
1154                 /* Discrete GPUs need their own topology device list
1155                  * entries. Don't assign them to CPU/APU nodes.
1156                  */
1157                 if (!gpu->use_iommu_v2 &&
1158                     dev->node_props.cpu_cores_count)
1159                         continue;
1160
1161                 if (!dev->gpu && (dev->node_props.simd_count > 0)) {
1162                         dev->gpu = gpu;
1163                         out_dev = dev;
1164
1165                         list_for_each_entry(mem, &dev->mem_props, list)
1166                                 mem->gpu = dev->gpu;
1167                         list_for_each_entry(cache, &dev->cache_props, list)
1168                                 cache->gpu = dev->gpu;
1169                         list_for_each_entry(iolink, &dev->io_link_props, list)
1170                                 iolink->gpu = dev->gpu;
1171                         break;
1172                 }
1173         }
1174         up_write(&topology_lock);
1175         return out_dev;
1176 }
1177
1178 static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival)
1179 {
1180         /*
1181          * TODO: Generate an event for thunk about the arrival/removal
1182          * of the GPU
1183          */
1184 }
1185
1186 /* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info,
1187  *              patch this after CRAT parsing.
1188  */
1189 static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev)
1190 {
1191         struct kfd_mem_properties *mem;
1192         struct kfd_local_mem_info local_mem_info;
1193
1194         if (!dev)
1195                 return;
1196
1197         /* Currently, amdgpu driver (amdgpu_mc) deals only with GPUs with
1198          * single bank of VRAM local memory.
1199          * for dGPUs - VCRAT reports only one bank of Local Memory
1200          * for APUs - If CRAT from ACPI reports more than one bank, then
1201          *      all the banks will report the same mem_clk_max information
1202          */
1203         amdgpu_amdkfd_get_local_mem_info(dev->gpu->adev, &local_mem_info);
1204
1205         list_for_each_entry(mem, &dev->mem_props, list)
1206                 mem->mem_clk_max = local_mem_info.mem_clk_max;
1207 }
1208
1209 static void kfd_set_iolink_no_atomics(struct kfd_topology_device *dev,
1210                                         struct kfd_topology_device *target_gpu_dev,
1211                                         struct kfd_iolink_properties *link)
1212 {
1213         /* xgmi always supports atomics between links. */
1214         if (link->iolink_type == CRAT_IOLINK_TYPE_XGMI)
1215                 return;
1216
1217         /* check pcie support to set cpu(dev) flags for target_gpu_dev link. */
1218         if (target_gpu_dev) {
1219                 uint32_t cap;
1220
1221                 pcie_capability_read_dword(target_gpu_dev->gpu->pdev,
1222                                 PCI_EXP_DEVCAP2, &cap);
1223
1224                 if (!(cap & (PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
1225                              PCI_EXP_DEVCAP2_ATOMIC_COMP64)))
1226                         link->flags |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1227                                 CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1228         /* set gpu (dev) flags. */
1229         } else {
1230                 if (!dev->gpu->pci_atomic_requested ||
1231                                 dev->gpu->adev->asic_type == CHIP_HAWAII)
1232                         link->flags |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1233                                 CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1234         }
1235 }
1236
1237 static void kfd_set_iolink_non_coherent(struct kfd_topology_device *to_dev,
1238                 struct kfd_iolink_properties *outbound_link,
1239                 struct kfd_iolink_properties *inbound_link)
1240 {
1241         /* CPU -> GPU with PCIe */
1242         if (!to_dev->gpu &&
1243             inbound_link->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS)
1244                 inbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT;
1245
1246         if (to_dev->gpu) {
1247                 /* GPU <-> GPU with PCIe and
1248                  * Vega20 with XGMI
1249                  */
1250                 if (inbound_link->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS ||
1251                     (inbound_link->iolink_type == CRAT_IOLINK_TYPE_XGMI &&
1252                     KFD_GC_VERSION(to_dev->gpu) == IP_VERSION(9, 4, 0))) {
1253                         outbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT;
1254                         inbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT;
1255                 }
1256         }
1257 }
1258
1259 static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev)
1260 {
1261         struct kfd_iolink_properties *link, *inbound_link;
1262         struct kfd_topology_device *peer_dev;
1263
1264         if (!dev || !dev->gpu)
1265                 return;
1266
1267         /* GPU only creates direct links so apply flags setting to all */
1268         list_for_each_entry(link, &dev->io_link_props, list) {
1269                 link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1270                 kfd_set_iolink_no_atomics(dev, NULL, link);
1271                 peer_dev = kfd_topology_device_by_proximity_domain(
1272                                 link->node_to);
1273
1274                 if (!peer_dev)
1275                         continue;
1276
1277                 list_for_each_entry(inbound_link, &peer_dev->io_link_props,
1278                                                                         list) {
1279                         if (inbound_link->node_to != link->node_from)
1280                                 continue;
1281
1282                         inbound_link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1283                         kfd_set_iolink_no_atomics(peer_dev, dev, inbound_link);
1284                         kfd_set_iolink_non_coherent(peer_dev, link, inbound_link);
1285                 }
1286         }
1287 }
1288
1289 int kfd_topology_add_device(struct kfd_dev *gpu)
1290 {
1291         uint32_t gpu_id;
1292         struct kfd_topology_device *dev;
1293         struct kfd_cu_info cu_info;
1294         int res = 0;
1295         struct list_head temp_topology_device_list;
1296         void *crat_image = NULL;
1297         size_t image_size = 0;
1298         int proximity_domain;
1299         int i;
1300         const char *asic_name = amdgpu_asic_name[gpu->adev->asic_type];
1301
1302         INIT_LIST_HEAD(&temp_topology_device_list);
1303
1304         gpu_id = kfd_generate_gpu_id(gpu);
1305
1306         pr_debug("Adding new GPU (ID: 0x%x) to topology\n", gpu_id);
1307
1308         /* Include the CPU in xGMI hive if xGMI connected by assigning it the hive ID. */
1309         if (gpu->hive_id && gpu->adev->gmc.xgmi.connected_to_cpu) {
1310                 struct kfd_topology_device *top_dev;
1311
1312                 down_read(&topology_lock);
1313
1314                 list_for_each_entry(top_dev, &topology_device_list, list) {
1315                         if (top_dev->gpu)
1316                                 break;
1317
1318                         top_dev->node_props.hive_id = gpu->hive_id;
1319                 }
1320
1321                 up_read(&topology_lock);
1322         }
1323
1324         /* Check to see if this gpu device exists in the topology_device_list.
1325          * If so, assign the gpu to that device,
1326          * else create a Virtual CRAT for this gpu device and then parse that
1327          * CRAT to create a new topology device. Once created assign the gpu to
1328          * that topology device
1329          */
1330         dev = kfd_assign_gpu(gpu);
1331         if (!dev) {
1332                 down_write(&topology_lock);
1333                 proximity_domain = ++topology_crat_proximity_domain;
1334
1335                 res = kfd_create_crat_image_virtual(&crat_image, &image_size,
1336                                                     COMPUTE_UNIT_GPU, gpu,
1337                                                     proximity_domain);
1338                 if (res) {
1339                         pr_err("Error creating VCRAT for GPU (ID: 0x%x)\n",
1340                                gpu_id);
1341                         topology_crat_proximity_domain--;
1342                         return res;
1343                 }
1344                 res = kfd_parse_crat_table(crat_image,
1345                                            &temp_topology_device_list,
1346                                            proximity_domain);
1347                 if (res) {
1348                         pr_err("Error parsing VCRAT for GPU (ID: 0x%x)\n",
1349                                gpu_id);
1350                         topology_crat_proximity_domain--;
1351                         goto err;
1352                 }
1353
1354                 kfd_topology_update_device_list(&temp_topology_device_list,
1355                         &topology_device_list);
1356
1357                 /* Update the SYSFS tree, since we added another topology
1358                  * device
1359                  */
1360                 res = kfd_topology_update_sysfs();
1361                 up_write(&topology_lock);
1362
1363                 if (!res)
1364                         sys_props.generation_count++;
1365                 else
1366                         pr_err("Failed to update GPU (ID: 0x%x) to sysfs topology. res=%d\n",
1367                                                 gpu_id, res);
1368                 dev = kfd_assign_gpu(gpu);
1369                 if (WARN_ON(!dev)) {
1370                         res = -ENODEV;
1371                         goto err;
1372                 }
1373         }
1374
1375         dev->gpu_id = gpu_id;
1376         gpu->id = gpu_id;
1377
1378         /* TODO: Move the following lines to function
1379          *      kfd_add_non_crat_information
1380          */
1381
1382         /* Fill-in additional information that is not available in CRAT but
1383          * needed for the topology
1384          */
1385
1386         amdgpu_amdkfd_get_cu_info(dev->gpu->adev, &cu_info);
1387
1388         for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1; i++) {
1389                 dev->node_props.name[i] = __tolower(asic_name[i]);
1390                 if (asic_name[i] == '\0')
1391                         break;
1392         }
1393         dev->node_props.name[i] = '\0';
1394
1395         dev->node_props.simd_arrays_per_engine =
1396                 cu_info.num_shader_arrays_per_engine;
1397
1398         dev->node_props.gfx_target_version = gpu->device_info.gfx_target_version;
1399         dev->node_props.vendor_id = gpu->pdev->vendor;
1400         dev->node_props.device_id = gpu->pdev->device;
1401         dev->node_props.capability |=
1402                 ((dev->gpu->adev->rev_id << HSA_CAP_ASIC_REVISION_SHIFT) &
1403                         HSA_CAP_ASIC_REVISION_MASK);
1404         dev->node_props.location_id = pci_dev_id(gpu->pdev);
1405         dev->node_props.domain = pci_domain_nr(gpu->pdev->bus);
1406         dev->node_props.max_engine_clk_fcompute =
1407                 amdgpu_amdkfd_get_max_engine_clock_in_mhz(dev->gpu->adev);
1408         dev->node_props.max_engine_clk_ccompute =
1409                 cpufreq_quick_get_max(0) / 1000;
1410         dev->node_props.drm_render_minor =
1411                 gpu->shared_resources.drm_render_minor;
1412
1413         dev->node_props.hive_id = gpu->hive_id;
1414         dev->node_props.num_sdma_engines = kfd_get_num_sdma_engines(gpu);
1415         dev->node_props.num_sdma_xgmi_engines =
1416                                         kfd_get_num_xgmi_sdma_engines(gpu);
1417         dev->node_props.num_sdma_queues_per_engine =
1418                                 gpu->device_info.num_sdma_queues_per_engine;
1419         dev->node_props.num_gws = (dev->gpu->gws &&
1420                 dev->gpu->dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) ?
1421                 dev->gpu->adev->gds.gws_size : 0;
1422         dev->node_props.num_cp_queues = get_cp_queues_num(dev->gpu->dqm);
1423
1424         kfd_fill_mem_clk_max_info(dev);
1425         kfd_fill_iolink_non_crat_info(dev);
1426
1427         switch (dev->gpu->adev->asic_type) {
1428         case CHIP_KAVERI:
1429         case CHIP_HAWAII:
1430         case CHIP_TONGA:
1431                 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_PRE_1_0 <<
1432                         HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1433                         HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1434                 break;
1435         case CHIP_CARRIZO:
1436         case CHIP_FIJI:
1437         case CHIP_POLARIS10:
1438         case CHIP_POLARIS11:
1439         case CHIP_POLARIS12:
1440         case CHIP_VEGAM:
1441                 pr_debug("Adding doorbell packet type capability\n");
1442                 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 <<
1443                         HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1444                         HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1445                 break;
1446         default:
1447                 if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(9, 0, 1))
1448                         dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 <<
1449                                 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1450                                 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1451                 else
1452                         WARN(1, "Unexpected ASIC family %u",
1453                              dev->gpu->adev->asic_type);
1454         }
1455
1456         /*
1457          * Overwrite ATS capability according to needs_iommu_device to fix
1458          * potential missing corresponding bit in CRAT of BIOS.
1459          */
1460         if (dev->gpu->use_iommu_v2)
1461                 dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
1462         else
1463                 dev->node_props.capability &= ~HSA_CAP_ATS_PRESENT;
1464
1465         /* Fix errors in CZ CRAT.
1466          * simd_count: Carrizo CRAT reports wrong simd_count, probably
1467          *              because it doesn't consider masked out CUs
1468          * max_waves_per_simd: Carrizo reports wrong max_waves_per_simd
1469          */
1470         if (dev->gpu->adev->asic_type == CHIP_CARRIZO) {
1471                 dev->node_props.simd_count =
1472                         cu_info.simd_per_cu * cu_info.cu_active_number;
1473                 dev->node_props.max_waves_per_simd = 10;
1474         }
1475
1476         /* kfd only concerns sram ecc on GFX and HBM ecc on UMC */
1477         dev->node_props.capability |=
1478                 ((dev->gpu->adev->ras_enabled & BIT(AMDGPU_RAS_BLOCK__GFX)) != 0) ?
1479                 HSA_CAP_SRAM_EDCSUPPORTED : 0;
1480         dev->node_props.capability |=
1481                 ((dev->gpu->adev->ras_enabled & BIT(AMDGPU_RAS_BLOCK__UMC)) != 0) ?
1482                 HSA_CAP_MEM_EDCSUPPORTED : 0;
1483
1484         if (KFD_GC_VERSION(dev->gpu) != IP_VERSION(9, 0, 1))
1485                 dev->node_props.capability |= (dev->gpu->adev->ras_enabled != 0) ?
1486                         HSA_CAP_RASEVENTNOTIFY : 0;
1487
1488         if (KFD_IS_SVM_API_SUPPORTED(dev->gpu->adev->kfd.dev))
1489                 dev->node_props.capability |= HSA_CAP_SVMAPI_SUPPORTED;
1490
1491         kfd_debug_print_topology();
1492
1493         if (!res)
1494                 kfd_notify_gpu_change(gpu_id, 1);
1495 err:
1496         kfd_destroy_crat_image(crat_image);
1497         return res;
1498 }
1499
1500 /**
1501  * kfd_topology_update_io_links() - Update IO links after device removal.
1502  * @proximity_domain: Proximity domain value of the dev being removed.
1503  *
1504  * The topology list currently is arranged in increasing order of
1505  * proximity domain.
1506  *
1507  * Two things need to be done when a device is removed:
1508  * 1. All the IO links to this device need to be removed.
1509  * 2. All nodes after the current device node need to move
1510  *    up once this device node is removed from the topology
1511  *    list. As a result, the proximity domain values for
1512  *    all nodes after the node being deleted reduce by 1.
1513  *    This would also cause the proximity domain values for
1514  *    io links to be updated based on new proximity domain
1515  *    values.
1516  *
1517  * Context: The caller must hold write topology_lock.
1518  */
1519 static void kfd_topology_update_io_links(int proximity_domain)
1520 {
1521         struct kfd_topology_device *dev;
1522         struct kfd_iolink_properties *iolink, *tmp;
1523
1524         list_for_each_entry(dev, &topology_device_list, list) {
1525                 if (dev->proximity_domain > proximity_domain)
1526                         dev->proximity_domain--;
1527
1528                 list_for_each_entry_safe(iolink, tmp, &dev->io_link_props, list) {
1529                         /*
1530                          * If there is an io link to the dev being deleted
1531                          * then remove that IO link also.
1532                          */
1533                         if (iolink->node_to == proximity_domain) {
1534                                 list_del(&iolink->list);
1535                                 dev->io_link_count--;
1536                                 dev->node_props.io_links_count--;
1537                         } else if (iolink->node_from > proximity_domain) {
1538                                 iolink->node_from--;
1539                         } else if (iolink->node_to > proximity_domain) {
1540                                 iolink->node_to--;
1541                         }
1542                 }
1543
1544         }
1545 }
1546
1547 int kfd_topology_remove_device(struct kfd_dev *gpu)
1548 {
1549         struct kfd_topology_device *dev, *tmp;
1550         uint32_t gpu_id;
1551         int res = -ENODEV;
1552         int i = 0;
1553
1554         down_write(&topology_lock);
1555
1556         list_for_each_entry_safe(dev, tmp, &topology_device_list, list) {
1557                 if (dev->gpu == gpu) {
1558                         gpu_id = dev->gpu_id;
1559                         kfd_remove_sysfs_node_entry(dev);
1560                         kfd_release_topology_device(dev);
1561                         sys_props.num_devices--;
1562                         kfd_topology_update_io_links(i);
1563                         topology_crat_proximity_domain = sys_props.num_devices-1;
1564                         sys_props.generation_count++;
1565                         res = 0;
1566                         if (kfd_topology_update_sysfs() < 0)
1567                                 kfd_topology_release_sysfs();
1568                         break;
1569                 }
1570                 i++;
1571         }
1572
1573         up_write(&topology_lock);
1574
1575         if (!res)
1576                 kfd_notify_gpu_change(gpu_id, 0);
1577
1578         return res;
1579 }
1580
1581 /* kfd_topology_enum_kfd_devices - Enumerate through all devices in KFD
1582  *      topology. If GPU device is found @idx, then valid kfd_dev pointer is
1583  *      returned through @kdev
1584  * Return -     0: On success (@kdev will be NULL for non GPU nodes)
1585  *              -1: If end of list
1586  */
1587 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev)
1588 {
1589
1590         struct kfd_topology_device *top_dev;
1591         uint8_t device_idx = 0;
1592
1593         *kdev = NULL;
1594         down_read(&topology_lock);
1595
1596         list_for_each_entry(top_dev, &topology_device_list, list) {
1597                 if (device_idx == idx) {
1598                         *kdev = top_dev->gpu;
1599                         up_read(&topology_lock);
1600                         return 0;
1601                 }
1602
1603                 device_idx++;
1604         }
1605
1606         up_read(&topology_lock);
1607
1608         return -1;
1609
1610 }
1611
1612 static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask)
1613 {
1614         int first_cpu_of_numa_node;
1615
1616         if (!cpumask || cpumask == cpu_none_mask)
1617                 return -1;
1618         first_cpu_of_numa_node = cpumask_first(cpumask);
1619         if (first_cpu_of_numa_node >= nr_cpu_ids)
1620                 return -1;
1621 #ifdef CONFIG_X86_64
1622         return cpu_data(first_cpu_of_numa_node).apicid;
1623 #else
1624         return first_cpu_of_numa_node;
1625 #endif
1626 }
1627
1628 /* kfd_numa_node_to_apic_id - Returns the APIC ID of the first logical processor
1629  *      of the given NUMA node (numa_node_id)
1630  * Return -1 on failure
1631  */
1632 int kfd_numa_node_to_apic_id(int numa_node_id)
1633 {
1634         if (numa_node_id == -1) {
1635                 pr_warn("Invalid NUMA Node. Use online CPU mask\n");
1636                 return kfd_cpumask_to_apic_id(cpu_online_mask);
1637         }
1638         return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id));
1639 }
1640
1641 void kfd_double_confirm_iommu_support(struct kfd_dev *gpu)
1642 {
1643         struct kfd_topology_device *dev;
1644
1645         gpu->use_iommu_v2 = false;
1646
1647         if (!gpu->device_info.needs_iommu_device)
1648                 return;
1649
1650         down_read(&topology_lock);
1651
1652         /* Only use IOMMUv2 if there is an APU topology node with no GPU
1653          * assigned yet. This GPU will be assigned to it.
1654          */
1655         list_for_each_entry(dev, &topology_device_list, list)
1656                 if (dev->node_props.cpu_cores_count &&
1657                     dev->node_props.simd_count &&
1658                     !dev->gpu)
1659                         gpu->use_iommu_v2 = true;
1660
1661         up_read(&topology_lock);
1662 }
1663
1664 #if defined(CONFIG_DEBUG_FS)
1665
1666 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data)
1667 {
1668         struct kfd_topology_device *dev;
1669         unsigned int i = 0;
1670         int r = 0;
1671
1672         down_read(&topology_lock);
1673
1674         list_for_each_entry(dev, &topology_device_list, list) {
1675                 if (!dev->gpu) {
1676                         i++;
1677                         continue;
1678                 }
1679
1680                 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1681                 r = dqm_debugfs_hqds(m, dev->gpu->dqm);
1682                 if (r)
1683                         break;
1684         }
1685
1686         up_read(&topology_lock);
1687
1688         return r;
1689 }
1690
1691 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data)
1692 {
1693         struct kfd_topology_device *dev;
1694         unsigned int i = 0;
1695         int r = 0;
1696
1697         down_read(&topology_lock);
1698
1699         list_for_each_entry(dev, &topology_device_list, list) {
1700                 if (!dev->gpu) {
1701                         i++;
1702                         continue;
1703                 }
1704
1705                 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
1706                 r = pm_debugfs_runlist(m, &dev->gpu->dqm->packet_mgr);
1707                 if (r)
1708                         break;
1709         }
1710
1711         up_read(&topology_lock);
1712
1713         return r;
1714 }
1715
1716 #endif