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 #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 *
 454                                           NUM_XCC(dev->gpu->xcc_mask)) : 0);
 455         sysfs_show_32bit_prop(buffer, offs, "mem_banks_count",
 456                               dev->node_props.mem_banks_count);
 457         sysfs_show_32bit_prop(buffer, offs, "caches_count",
 458                               dev->node_props.caches_count);
 459         sysfs_show_32bit_prop(buffer, offs, "io_links_count",
 460                               dev->node_props.io_links_count);
 461         sysfs_show_32bit_prop(buffer, offs, "p2p_links_count",
 462                               dev->node_props.p2p_links_count);
 463         sysfs_show_32bit_prop(buffer, offs, "cpu_core_id_base",
 464                               dev->node_props.cpu_core_id_base);
 465         sysfs_show_32bit_prop(buffer, offs, "simd_id_base",
 466                               dev->node_props.simd_id_base);
 467         sysfs_show_32bit_prop(buffer, offs, "max_waves_per_simd",
 468                               dev->node_props.max_waves_per_simd);
 469         sysfs_show_32bit_prop(buffer, offs, "lds_size_in_kb",
 470                               dev->node_props.lds_size_in_kb);
 471         sysfs_show_32bit_prop(buffer, offs, "gds_size_in_kb",
 472                               dev->node_props.gds_size_in_kb);
 473         sysfs_show_32bit_prop(buffer, offs, "num_gws",
 474                               dev->node_props.num_gws);
 475         sysfs_show_32bit_prop(buffer, offs, "wave_front_size",
 476                               dev->node_props.wave_front_size);
 477         sysfs_show_32bit_prop(buffer, offs, "array_count",
 478                               dev->gpu ? (dev->node_props.array_count *
 479                                           NUM_XCC(dev->gpu->xcc_mask)) : 0);
 480         sysfs_show_32bit_prop(buffer, offs, "simd_arrays_per_engine",
 481                               dev->node_props.simd_arrays_per_engine);
 482         sysfs_show_32bit_prop(buffer, offs, "cu_per_simd_array",
 483                               dev->node_props.cu_per_simd_array);
 484         sysfs_show_32bit_prop(buffer, offs, "simd_per_cu",
 485                               dev->node_props.simd_per_cu);
 486         sysfs_show_32bit_prop(buffer, offs, "max_slots_scratch_cu",
 487                               dev->node_props.max_slots_scratch_cu);
 488         sysfs_show_32bit_prop(buffer, offs, "gfx_target_version",
 489                               dev->node_props.gfx_target_version);
 490         sysfs_show_32bit_prop(buffer, offs, "vendor_id",
 491                               dev->node_props.vendor_id);
 492         sysfs_show_32bit_prop(buffer, offs, "device_id",
 493                               dev->node_props.device_id);
 494         sysfs_show_32bit_prop(buffer, offs, "location_id",
 495                               dev->node_props.location_id);
 496         sysfs_show_32bit_prop(buffer, offs, "domain",
 497                               dev->node_props.domain);
 498         sysfs_show_32bit_prop(buffer, offs, "drm_render_minor",
 499                               dev->node_props.drm_render_minor);
 500         sysfs_show_64bit_prop(buffer, offs, "hive_id",
 501                               dev->node_props.hive_id);
 502         sysfs_show_32bit_prop(buffer, offs, "num_sdma_engines",
 503                               dev->node_props.num_sdma_engines);
 504         sysfs_show_32bit_prop(buffer, offs, "num_sdma_xgmi_engines",
 505                               dev->node_props.num_sdma_xgmi_engines);
 506         sysfs_show_32bit_prop(buffer, offs, "num_sdma_queues_per_engine",
 507                               dev->node_props.num_sdma_queues_per_engine);
 508         sysfs_show_32bit_prop(buffer, offs, "num_cp_queues",
 509                               dev->node_props.num_cp_queues);
 510
 511         if (dev->gpu) {
 512                 log_max_watch_addr =
 513                         __ilog2_u32(dev->gpu->kfd->device_info.num_of_watch_points);
 514
 515                 if (log_max_watch_addr) {
 516                         dev->node_props.capability |=
 517                                         HSA_CAP_WATCH_POINTS_SUPPORTED;
 518
 519                         dev->node_props.capability |=
 520                                 ((log_max_watch_addr <<
 521                                         HSA_CAP_WATCH_POINTS_TOTALBITS_SHIFT) &
 522                                 HSA_CAP_WATCH_POINTS_TOTALBITS_MASK);
 523                 }
 524
 525                 if (dev->gpu->adev->asic_type == CHIP_TONGA)
 526                         dev->node_props.capability |=
 527                                         HSA_CAP_AQL_QUEUE_DOUBLE_MAP;
 528
 529                 sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_fcompute",
 530                         dev->node_props.max_engine_clk_fcompute);
 531
 532                 sysfs_show_64bit_prop(buffer, offs, "local_mem_size", 0ULL);
 533
 534                 sysfs_show_32bit_prop(buffer, offs, "fw_version",
 535                                       dev->gpu->kfd->mec_fw_version);
 536                 sysfs_show_32bit_prop(buffer, offs, "capability",
 537                                       dev->node_props.capability);
 538                 sysfs_show_64bit_prop(buffer, offs, "debug_prop",
 539                                       dev->node_props.debug_prop);
 540                 sysfs_show_32bit_prop(buffer, offs, "sdma_fw_version",
 541                                       dev->gpu->kfd->sdma_fw_version);
 542                 sysfs_show_64bit_prop(buffer, offs, "unique_id",
 543                                       dev->gpu->adev->unique_id);
 544                 sysfs_show_32bit_prop(buffer, offs, "num_xcc",
 545                                       NUM_XCC(dev->gpu->xcc_mask));
 546         }
 547
 548         return sysfs_show_32bit_prop(buffer, offs, "max_engine_clk_ccompute",
 549                                      cpufreq_quick_get_max(0)/1000);
 550 }
 551
 552 static const struct sysfs_ops node_ops = {
 553         .show = node_show,
 554 };
 555
 556 static const struct kobj_type node_type = {
 557         .release = kfd_topology_kobj_release,
 558         .sysfs_ops = &node_ops,
 559 };
 560
 561 static void kfd_remove_sysfs_file(struct kobject *kobj, struct attribute *attr)
 562 {
 563         sysfs_remove_file(kobj, attr);
 564         kobject_del(kobj);
 565         kobject_put(kobj);
 566 }
 567
 568 static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev)
 569 {
 570         struct kfd_iolink_properties *p2plink;
 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_p2plink) {
 589                 list_for_each_entry(p2plink, &dev->p2p_link_props, list)
 590                         if (p2plink->kobj) {
 591                                 kfd_remove_sysfs_file(p2plink->kobj,
 592                                                         &p2plink->attr);
 593                                 p2plink->kobj = NULL;
 594                         }
 595                 kobject_del(dev->kobj_p2plink);
 596                 kobject_put(dev->kobj_p2plink);
 597                 dev->kobj_p2plink = NULL;
 598         }
 599
 600         if (dev->kobj_cache) {
 601                 list_for_each_entry(cache, &dev->cache_props, list)
 602                         if (cache->kobj) {
 603                                 kfd_remove_sysfs_file(cache->kobj,
 604                                                         &cache->attr);
 605                                 cache->kobj = NULL;
 606                         }
 607                 kobject_del(dev->kobj_cache);
 608                 kobject_put(dev->kobj_cache);
 609                 dev->kobj_cache = NULL;
 610         }
 611
 612         if (dev->kobj_mem) {
 613                 list_for_each_entry(mem, &dev->mem_props, list)
 614                         if (mem->kobj) {
 615                                 kfd_remove_sysfs_file(mem->kobj, &mem->attr);
 616                                 mem->kobj = NULL;
 617                         }
 618                 kobject_del(dev->kobj_mem);
 619                 kobject_put(dev->kobj_mem);
 620                 dev->kobj_mem = NULL;
 621         }
 622
 623         if (dev->kobj_perf) {
 624                 list_for_each_entry(perf, &dev->perf_props, list) {
 625                         kfree(perf->attr_group);
 626                         perf->attr_group = NULL;
 627                 }
 628                 kobject_del(dev->kobj_perf);
 629                 kobject_put(dev->kobj_perf);
 630                 dev->kobj_perf = NULL;
 631         }
 632
 633         if (dev->kobj_node) {
 634                 sysfs_remove_file(dev->kobj_node, &dev->attr_gpuid);
 635                 sysfs_remove_file(dev->kobj_node, &dev->attr_name);
 636                 sysfs_remove_file(dev->kobj_node, &dev->attr_props);
 637                 kobject_del(dev->kobj_node);
 638                 kobject_put(dev->kobj_node);
 639                 dev->kobj_node = NULL;
 640         }
 641 }
 642
 643 static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev,
 644                 uint32_t id)
 645 {
 646         struct kfd_iolink_properties *p2plink;
 647         struct kfd_iolink_properties *iolink;
 648         struct kfd_cache_properties *cache;
 649         struct kfd_mem_properties *mem;
 650         struct kfd_perf_properties *perf;
 651         int ret;
 652         uint32_t i, num_attrs;
 653         struct attribute **attrs;
 654
 655         if (WARN_ON(dev->kobj_node))
 656                 return -EEXIST;
 657
 658         /*
 659          * Creating the sysfs folders
 660          */
 661         dev->kobj_node = kfd_alloc_struct(dev->kobj_node);
 662         if (!dev->kobj_node)
 663                 return -ENOMEM;
 664
 665         ret = kobject_init_and_add(dev->kobj_node, &node_type,
 666                         sys_props.kobj_nodes, "%d", id);
 667         if (ret < 0) {
 668                 kobject_put(dev->kobj_node);
 669                 return ret;
 670         }
 671
 672         dev->kobj_mem = kobject_create_and_add("mem_banks", dev->kobj_node);
 673         if (!dev->kobj_mem)
 674                 return -ENOMEM;
 675
 676         dev->kobj_cache = kobject_create_and_add("caches", dev->kobj_node);
 677         if (!dev->kobj_cache)
 678                 return -ENOMEM;
 679
 680         dev->kobj_iolink = kobject_create_and_add("io_links", dev->kobj_node);
 681         if (!dev->kobj_iolink)
 682                 return -ENOMEM;
 683
 684         dev->kobj_p2plink = kobject_create_and_add("p2p_links", dev->kobj_node);
 685         if (!dev->kobj_p2plink)
 686                 return -ENOMEM;
 687
 688         dev->kobj_perf = kobject_create_and_add("perf", dev->kobj_node);
 689         if (!dev->kobj_perf)
 690                 return -ENOMEM;
 691
 692         /*
 693          * Creating sysfs files for node properties
 694          */
 695         dev->attr_gpuid.name = "gpu_id";
 696         dev->attr_gpuid.mode = KFD_SYSFS_FILE_MODE;
 697         sysfs_attr_init(&dev->attr_gpuid);
 698         dev->attr_name.name = "name";
 699         dev->attr_name.mode = KFD_SYSFS_FILE_MODE;
 700         sysfs_attr_init(&dev->attr_name);
 701         dev->attr_props.name = "properties";
 702         dev->attr_props.mode = KFD_SYSFS_FILE_MODE;
 703         sysfs_attr_init(&dev->attr_props);
 704         ret = sysfs_create_file(dev->kobj_node, &dev->attr_gpuid);
 705         if (ret < 0)
 706                 return ret;
 707         ret = sysfs_create_file(dev->kobj_node, &dev->attr_name);
 708         if (ret < 0)
 709                 return ret;
 710         ret = sysfs_create_file(dev->kobj_node, &dev->attr_props);
 711         if (ret < 0)
 712                 return ret;
 713
 714         i = 0;
 715         list_for_each_entry(mem, &dev->mem_props, list) {
 716                 mem->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
 717                 if (!mem->kobj)
 718                         return -ENOMEM;
 719                 ret = kobject_init_and_add(mem->kobj, &mem_type,
 720                                 dev->kobj_mem, "%d", i);
 721                 if (ret < 0) {
 722                         kobject_put(mem->kobj);
 723                         return ret;
 724                 }
 725
 726                 mem->attr.name = "properties";
 727                 mem->attr.mode = KFD_SYSFS_FILE_MODE;
 728                 sysfs_attr_init(&mem->attr);
 729                 ret = sysfs_create_file(mem->kobj, &mem->attr);
 730                 if (ret < 0)
 731                         return ret;
 732                 i++;
 733         }
 734
 735         i = 0;
 736         list_for_each_entry(cache, &dev->cache_props, list) {
 737                 cache->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
 738                 if (!cache->kobj)
 739                         return -ENOMEM;
 740                 ret = kobject_init_and_add(cache->kobj, &cache_type,
 741                                 dev->kobj_cache, "%d", i);
 742                 if (ret < 0) {
 743                         kobject_put(cache->kobj);
 744                         return ret;
 745                 }
 746
 747                 cache->attr.name = "properties";
 748                 cache->attr.mode = KFD_SYSFS_FILE_MODE;
 749                 sysfs_attr_init(&cache->attr);
 750                 ret = sysfs_create_file(cache->kobj, &cache->attr);
 751                 if (ret < 0)
 752                         return ret;
 753                 i++;
 754         }
 755
 756         i = 0;
 757         list_for_each_entry(iolink, &dev->io_link_props, list) {
 758                 iolink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
 759                 if (!iolink->kobj)
 760                         return -ENOMEM;
 761                 ret = kobject_init_and_add(iolink->kobj, &iolink_type,
 762                                 dev->kobj_iolink, "%d", i);
 763                 if (ret < 0) {
 764                         kobject_put(iolink->kobj);
 765                         return ret;
 766                 }
 767
 768                 iolink->attr.name = "properties";
 769                 iolink->attr.mode = KFD_SYSFS_FILE_MODE;
 770                 sysfs_attr_init(&iolink->attr);
 771                 ret = sysfs_create_file(iolink->kobj, &iolink->attr);
 772                 if (ret < 0)
 773                         return ret;
 774                 i++;
 775         }
 776
 777         i = 0;
 778         list_for_each_entry(p2plink, &dev->p2p_link_props, list) {
 779                 p2plink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
 780                 if (!p2plink->kobj)
 781                         return -ENOMEM;
 782                 ret = kobject_init_and_add(p2plink->kobj, &iolink_type,
 783                                 dev->kobj_p2plink, "%d", i);
 784                 if (ret < 0) {
 785                         kobject_put(p2plink->kobj);
 786                         return ret;
 787                 }
 788
 789                 p2plink->attr.name = "properties";
 790                 p2plink->attr.mode = KFD_SYSFS_FILE_MODE;
 791                 sysfs_attr_init(&p2plink->attr);
 792                 ret = sysfs_create_file(p2plink->kobj, &p2plink->attr);
 793                 if (ret < 0)
 794                         return ret;
 795                 i++;
 796         }
 797
 798         /* All hardware blocks have the same number of attributes. */
 799         num_attrs = ARRAY_SIZE(perf_attr_iommu);
 800         list_for_each_entry(perf, &dev->perf_props, list) {
 801                 perf->attr_group = kzalloc(sizeof(struct kfd_perf_attr)
 802                         * num_attrs + sizeof(struct attribute_group),
 803                         GFP_KERNEL);
 804                 if (!perf->attr_group)
 805                         return -ENOMEM;
 806
 807                 attrs = (struct attribute **)(perf->attr_group + 1);
 808                 if (!strcmp(perf->block_name, "iommu")) {
 809                 /* Information of IOMMU's num_counters and counter_ids is shown
 810                  * under /sys/bus/event_source/devices/amd_iommu. We don't
 811                  * duplicate here.
 812                  */
 813                         perf_attr_iommu[0].data = perf->max_concurrent;
 814                         for (i = 0; i < num_attrs; i++)
 815                                 attrs[i] = &perf_attr_iommu[i].attr.attr;
 816                 }
 817                 perf->attr_group->name = perf->block_name;
 818                 perf->attr_group->attrs = attrs;
 819                 ret = sysfs_create_group(dev->kobj_perf, perf->attr_group);
 820                 if (ret < 0)
 821                         return ret;
 822         }
 823
 824         return 0;
 825 }
 826
 827 /* Called with write topology lock acquired */
 828 static int kfd_build_sysfs_node_tree(void)
 829 {
 830         struct kfd_topology_device *dev;
 831         int ret;
 832         uint32_t i = 0;
 833
 834         list_for_each_entry(dev, &topology_device_list, list) {
 835                 ret = kfd_build_sysfs_node_entry(dev, i);
 836                 if (ret < 0)
 837                         return ret;
 838                 i++;
 839         }
 840
 841         return 0;
 842 }
 843
 844 /* Called with write topology lock acquired */
 845 static void kfd_remove_sysfs_node_tree(void)
 846 {
 847         struct kfd_topology_device *dev;
 848
 849         list_for_each_entry(dev, &topology_device_list, list)
 850                 kfd_remove_sysfs_node_entry(dev);
 851 }
 852
 853 static int kfd_topology_update_sysfs(void)
 854 {
 855         int ret;
 856
 857         if (!sys_props.kobj_topology) {
 858                 sys_props.kobj_topology =
 859                                 kfd_alloc_struct(sys_props.kobj_topology);
 860                 if (!sys_props.kobj_topology)
 861                         return -ENOMEM;
 862
 863                 ret = kobject_init_and_add(sys_props.kobj_topology,
 864                                 &sysprops_type,  &kfd_device->kobj,
 865                                 "topology");
 866                 if (ret < 0) {
 867                         kobject_put(sys_props.kobj_topology);
 868                         return ret;
 869                 }
 870
 871                 sys_props.kobj_nodes = kobject_create_and_add("nodes",
 872                                 sys_props.kobj_topology);
 873                 if (!sys_props.kobj_nodes)
 874                         return -ENOMEM;
 875
 876                 sys_props.attr_genid.name = "generation_id";
 877                 sys_props.attr_genid.mode = KFD_SYSFS_FILE_MODE;
 878                 sysfs_attr_init(&sys_props.attr_genid);
 879                 ret = sysfs_create_file(sys_props.kobj_topology,
 880                                 &sys_props.attr_genid);
 881                 if (ret < 0)
 882                         return ret;
 883
 884                 sys_props.attr_props.name = "system_properties";
 885                 sys_props.attr_props.mode = KFD_SYSFS_FILE_MODE;
 886                 sysfs_attr_init(&sys_props.attr_props);
 887                 ret = sysfs_create_file(sys_props.kobj_topology,
 888                                 &sys_props.attr_props);
 889                 if (ret < 0)
 890                         return ret;
 891         }
 892
 893         kfd_remove_sysfs_node_tree();
 894
 895         return kfd_build_sysfs_node_tree();
 896 }
 897
 898 static void kfd_topology_release_sysfs(void)
 899 {
 900         kfd_remove_sysfs_node_tree();
 901         if (sys_props.kobj_topology) {
 902                 sysfs_remove_file(sys_props.kobj_topology,
 903                                 &sys_props.attr_genid);
 904                 sysfs_remove_file(sys_props.kobj_topology,
 905                                 &sys_props.attr_props);
 906                 if (sys_props.kobj_nodes) {
 907                         kobject_del(sys_props.kobj_nodes);
 908                         kobject_put(sys_props.kobj_nodes);
 909                         sys_props.kobj_nodes = NULL;
 910                 }
 911                 kobject_del(sys_props.kobj_topology);
 912                 kobject_put(sys_props.kobj_topology);
 913                 sys_props.kobj_topology = NULL;
 914         }
 915 }
 916
 917 /* Called with write topology_lock acquired */
 918 static void kfd_topology_update_device_list(struct list_head *temp_list,
 919                                         struct list_head *master_list)
 920 {
 921         while (!list_empty(temp_list)) {
 922                 list_move_tail(temp_list->next, master_list);
 923                 sys_props.num_devices++;
 924         }
 925 }
 926
 927 static void kfd_debug_print_topology(void)
 928 {
 929         struct kfd_topology_device *dev;
 930
 931         down_read(&topology_lock);
 932
 933         dev = list_last_entry(&topology_device_list,
 934                         struct kfd_topology_device, list);
 935         if (dev) {
 936                 if (dev->node_props.cpu_cores_count &&
 937                                 dev->node_props.simd_count) {
 938                         pr_info("Topology: Add APU node [0x%0x:0x%0x]\n",
 939                                 dev->node_props.device_id,
 940                                 dev->node_props.vendor_id);
 941                 } else if (dev->node_props.cpu_cores_count)
 942                         pr_info("Topology: Add CPU node\n");
 943                 else if (dev->node_props.simd_count)
 944                         pr_info("Topology: Add dGPU node [0x%0x:0x%0x]\n",
 945                                 dev->node_props.device_id,
 946                                 dev->node_props.vendor_id);
 947         }
 948         up_read(&topology_lock);
 949 }
 950
 951 /* Helper function for intializing platform_xx members of
 952  * kfd_system_properties. Uses OEM info from the last CPU/APU node.
 953  */
 954 static void kfd_update_system_properties(void)
 955 {
 956         struct kfd_topology_device *dev;
 957
 958         down_read(&topology_lock);
 959         dev = list_last_entry(&topology_device_list,
 960                         struct kfd_topology_device, list);
 961         if (dev) {
 962                 sys_props.platform_id =
 963                         (*((uint64_t *)dev->oem_id)) & CRAT_OEMID_64BIT_MASK;
 964                 sys_props.platform_oem = *((uint64_t *)dev->oem_table_id);
 965                 sys_props.platform_rev = dev->oem_revision;
 966         }
 967         up_read(&topology_lock);
 968 }
 969
 970 static void find_system_memory(const struct dmi_header *dm,
 971         void *private)
 972 {
 973         struct kfd_mem_properties *mem;
 974         u16 mem_width, mem_clock;
 975         struct kfd_topology_device *kdev =
 976                 (struct kfd_topology_device *)private;
 977         const u8 *dmi_data = (const u8 *)(dm + 1);
 978
 979         if (dm->type == DMI_ENTRY_MEM_DEVICE && dm->length >= 0x15) {
 980                 mem_width = (u16)(*(const u16 *)(dmi_data + 0x6));
 981                 mem_clock = (u16)(*(const u16 *)(dmi_data + 0x11));
 982                 list_for_each_entry(mem, &kdev->mem_props, list) {
 983                         if (mem_width != 0xFFFF && mem_width != 0)
 984                                 mem->width = mem_width;
 985                         if (mem_clock != 0)
 986                                 mem->mem_clk_max = mem_clock;
 987                 }
 988         }
 989 }
 990
 991 /*
 992  * Performance counters information is not part of CRAT but we would like to
 993  * put them in the sysfs under topology directory for Thunk to get the data.
 994  * This function is called before updating the sysfs.
 995  */
 996 static int kfd_add_perf_to_topology(struct kfd_topology_device *kdev)
 997 {
 998         /* These are the only counters supported so far */
 999         return kfd_iommu_add_perf_counters(kdev);
1000 }
1001
1002 /* kfd_add_non_crat_information - Add information that is not currently
1003  *      defined in CRAT but is necessary for KFD topology
1004  * @dev - topology device to which addition info is added
1005  */
1006 static void kfd_add_non_crat_information(struct kfd_topology_device *kdev)
1007 {
1008         /* Check if CPU only node. */
1009         if (!kdev->gpu) {
1010                 /* Add system memory information */
1011                 dmi_walk(find_system_memory, kdev);
1012         }
1013         /* TODO: For GPU node, rearrange code from kfd_topology_add_device */
1014 }
1015
1016 /* kfd_is_acpi_crat_invalid - CRAT from ACPI is valid only for AMD APU devices.
1017  *      Ignore CRAT for all other devices. AMD APU is identified if both CPU
1018  *      and GPU cores are present.
1019  * @device_list - topology device list created by parsing ACPI CRAT table.
1020  * @return - TRUE if invalid, FALSE is valid.
1021  */
1022 static bool kfd_is_acpi_crat_invalid(struct list_head *device_list)
1023 {
1024         struct kfd_topology_device *dev;
1025
1026         list_for_each_entry(dev, device_list, list) {
1027                 if (dev->node_props.cpu_cores_count &&
1028                         dev->node_props.simd_count)
1029                         return false;
1030         }
1031         pr_info("Ignoring ACPI CRAT on non-APU system\n");
1032         return true;
1033 }
1034
1035 int kfd_topology_init(void)
1036 {
1037         void *crat_image = NULL;
1038         size_t image_size = 0;
1039         int ret;
1040         struct list_head temp_topology_device_list;
1041         int cpu_only_node = 0;
1042         struct kfd_topology_device *kdev;
1043         int proximity_domain;
1044
1045         /* topology_device_list - Master list of all topology devices
1046          * temp_topology_device_list - temporary list created while parsing CRAT
1047          * or VCRAT. Once parsing is complete the contents of list is moved to
1048          * topology_device_list
1049          */
1050
1051         /* Initialize the head for the both the lists */
1052         INIT_LIST_HEAD(&topology_device_list);
1053         INIT_LIST_HEAD(&temp_topology_device_list);
1054         init_rwsem(&topology_lock);
1055
1056         memset(&sys_props, 0, sizeof(sys_props));
1057
1058         /* Proximity domains in ACPI CRAT tables start counting at
1059          * 0. The same should be true for virtual CRAT tables created
1060          * at this stage. GPUs added later in kfd_topology_add_device
1061          * use a counter.
1062          */
1063         proximity_domain = 0;
1064
1065         /*
1066          * Get the CRAT image from the ACPI. If ACPI doesn't have one
1067          * or if ACPI CRAT is invalid create a virtual CRAT.
1068          * NOTE: The current implementation expects all AMD APUs to have
1069          *      CRAT. If no CRAT is available, it is assumed to be a CPU
1070          */
1071         ret = kfd_create_crat_image_acpi(&crat_image, &image_size);
1072         if (!ret) {
1073                 ret = kfd_parse_crat_table(crat_image,
1074                                            &temp_topology_device_list,
1075                                            proximity_domain);
1076                 if (ret ||
1077                     kfd_is_acpi_crat_invalid(&temp_topology_device_list)) {
1078                         kfd_release_topology_device_list(
1079                                 &temp_topology_device_list);
1080                         kfd_destroy_crat_image(crat_image);
1081                         crat_image = NULL;
1082                 }
1083         }
1084
1085         if (!crat_image) {
1086                 ret = kfd_create_crat_image_virtual(&crat_image, &image_size,
1087                                                     COMPUTE_UNIT_CPU, NULL,
1088                                                     proximity_domain);
1089                 cpu_only_node = 1;
1090                 if (ret) {
1091                         pr_err("Error creating VCRAT table for CPU\n");
1092                         return ret;
1093                 }
1094
1095                 ret = kfd_parse_crat_table(crat_image,
1096                                            &temp_topology_device_list,
1097                                            proximity_domain);
1098                 if (ret) {
1099                         pr_err("Error parsing VCRAT table for CPU\n");
1100                         goto err;
1101                 }
1102         }
1103
1104         kdev = list_first_entry(&temp_topology_device_list,
1105                                 struct kfd_topology_device, list);
1106         kfd_add_perf_to_topology(kdev);
1107
1108         down_write(&topology_lock);
1109         kfd_topology_update_device_list(&temp_topology_device_list,
1110                                         &topology_device_list);
1111         topology_crat_proximity_domain = sys_props.num_devices-1;
1112         ret = kfd_topology_update_sysfs();
1113         up_write(&topology_lock);
1114
1115         if (!ret) {
1116                 sys_props.generation_count++;
1117                 kfd_update_system_properties();
1118                 kfd_debug_print_topology();
1119         } else
1120                 pr_err("Failed to update topology in sysfs ret=%d\n", ret);
1121
1122         /* For nodes with GPU, this information gets added
1123          * when GPU is detected (kfd_topology_add_device).
1124          */
1125         if (cpu_only_node) {
1126                 /* Add additional information to CPU only node created above */
1127                 down_write(&topology_lock);
1128                 kdev = list_first_entry(&topology_device_list,
1129                                 struct kfd_topology_device, list);
1130                 up_write(&topology_lock);
1131                 kfd_add_non_crat_information(kdev);
1132         }
1133
1134 err:
1135         kfd_destroy_crat_image(crat_image);
1136         return ret;
1137 }
1138
1139 void kfd_topology_shutdown(void)
1140 {
1141         down_write(&topology_lock);
1142         kfd_topology_release_sysfs();
1143         kfd_release_live_view();
1144         up_write(&topology_lock);
1145 }
1146
1147 static uint32_t kfd_generate_gpu_id(struct kfd_node *gpu)
1148 {
1149         uint32_t hashout;
1150         uint32_t buf[8];
1151         uint64_t local_mem_size;
1152         int i;
1153
1154         if (!gpu)
1155                 return 0;
1156
1157         local_mem_size = gpu->local_mem_info.local_mem_size_private +
1158                         gpu->local_mem_info.local_mem_size_public;
1159         buf[0] = gpu->adev->pdev->devfn;
1160         buf[1] = gpu->adev->pdev->subsystem_vendor |
1161                 (gpu->adev->pdev->subsystem_device << 16);
1162         buf[2] = pci_domain_nr(gpu->adev->pdev->bus);
1163         buf[3] = gpu->adev->pdev->device;
1164         buf[4] = gpu->adev->pdev->bus->number;
1165         buf[5] = lower_32_bits(local_mem_size);
1166         buf[6] = upper_32_bits(local_mem_size);
1167         buf[7] = (ffs(gpu->xcc_mask) - 1) | (NUM_XCC(gpu->xcc_mask) << 16);
1168
1169         for (i = 0, hashout = 0; i < 8; i++)
1170                 hashout ^= hash_32(buf[i], KFD_GPU_ID_HASH_WIDTH);
1171
1172         return hashout;
1173 }
1174 /* kfd_assign_gpu - Attach @gpu to the correct kfd topology device. If
1175  *              the GPU device is not already present in the topology device
1176  *              list then return NULL. This means a new topology device has to
1177  *              be created for this GPU.
1178  */
1179 static struct kfd_topology_device *kfd_assign_gpu(struct kfd_node *gpu)
1180 {
1181         struct kfd_topology_device *dev;
1182         struct kfd_topology_device *out_dev = NULL;
1183         struct kfd_mem_properties *mem;
1184         struct kfd_cache_properties *cache;
1185         struct kfd_iolink_properties *iolink;
1186         struct kfd_iolink_properties *p2plink;
1187
1188         list_for_each_entry(dev, &topology_device_list, list) {
1189                 /* Discrete GPUs need their own topology device list
1190                  * entries. Don't assign them to CPU/APU nodes.
1191                  */
1192                 if (!gpu->kfd->use_iommu_v2 &&
1193                     dev->node_props.cpu_cores_count)
1194                         continue;
1195
1196                 if (!dev->gpu && (dev->node_props.simd_count > 0)) {
1197                         dev->gpu = gpu;
1198                         out_dev = dev;
1199
1200                         list_for_each_entry(mem, &dev->mem_props, list)
1201                                 mem->gpu = dev->gpu;
1202                         list_for_each_entry(cache, &dev->cache_props, list)
1203                                 cache->gpu = dev->gpu;
1204                         list_for_each_entry(iolink, &dev->io_link_props, list)
1205                                 iolink->gpu = dev->gpu;
1206                         list_for_each_entry(p2plink, &dev->p2p_link_props, list)
1207                                 p2plink->gpu = dev->gpu;
1208                         break;
1209                 }
1210         }
1211         return out_dev;
1212 }
1213
1214 static void kfd_notify_gpu_change(uint32_t gpu_id, int arrival)
1215 {
1216         /*
1217          * TODO: Generate an event for thunk about the arrival/removal
1218          * of the GPU
1219          */
1220 }
1221
1222 /* kfd_fill_mem_clk_max_info - Since CRAT doesn't have memory clock info,
1223  *              patch this after CRAT parsing.
1224  */
1225 static void kfd_fill_mem_clk_max_info(struct kfd_topology_device *dev)
1226 {
1227         struct kfd_mem_properties *mem;
1228         struct kfd_local_mem_info local_mem_info;
1229
1230         if (!dev)
1231                 return;
1232
1233         /* Currently, amdgpu driver (amdgpu_mc) deals only with GPUs with
1234          * single bank of VRAM local memory.
1235          * for dGPUs - VCRAT reports only one bank of Local Memory
1236          * for APUs - If CRAT from ACPI reports more than one bank, then
1237          *      all the banks will report the same mem_clk_max information
1238          */
1239         amdgpu_amdkfd_get_local_mem_info(dev->gpu->adev, &local_mem_info,
1240                                          dev->gpu->xcp);
1241
1242         list_for_each_entry(mem, &dev->mem_props, list)
1243                 mem->mem_clk_max = local_mem_info.mem_clk_max;
1244 }
1245
1246 static void kfd_set_iolink_no_atomics(struct kfd_topology_device *dev,
1247                                         struct kfd_topology_device *target_gpu_dev,
1248                                         struct kfd_iolink_properties *link)
1249 {
1250         /* xgmi always supports atomics between links. */
1251         if (link->iolink_type == CRAT_IOLINK_TYPE_XGMI)
1252                 return;
1253
1254         /* check pcie support to set cpu(dev) flags for target_gpu_dev link. */
1255         if (target_gpu_dev) {
1256                 uint32_t cap;
1257
1258                 pcie_capability_read_dword(target_gpu_dev->gpu->adev->pdev,
1259                                 PCI_EXP_DEVCAP2, &cap);
1260
1261                 if (!(cap & (PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
1262                              PCI_EXP_DEVCAP2_ATOMIC_COMP64)))
1263                         link->flags |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1264                                 CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1265         /* set gpu (dev) flags. */
1266         } else {
1267                 if (!dev->gpu->kfd->pci_atomic_requested ||
1268                                 dev->gpu->adev->asic_type == CHIP_HAWAII)
1269                         link->flags |= CRAT_IOLINK_FLAGS_NO_ATOMICS_32_BIT |
1270                                 CRAT_IOLINK_FLAGS_NO_ATOMICS_64_BIT;
1271         }
1272 }
1273
1274 static void kfd_set_iolink_non_coherent(struct kfd_topology_device *to_dev,
1275                 struct kfd_iolink_properties *outbound_link,
1276                 struct kfd_iolink_properties *inbound_link)
1277 {
1278         /* CPU -> GPU with PCIe */
1279         if (!to_dev->gpu &&
1280             inbound_link->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS)
1281                 inbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT;
1282
1283         if (to_dev->gpu) {
1284                 /* GPU <-> GPU with PCIe and
1285                  * Vega20 with XGMI
1286                  */
1287                 if (inbound_link->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS ||
1288                     (inbound_link->iolink_type == CRAT_IOLINK_TYPE_XGMI &&
1289                     KFD_GC_VERSION(to_dev->gpu) == IP_VERSION(9, 4, 0))) {
1290                         outbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT;
1291                         inbound_link->flags |= CRAT_IOLINK_FLAGS_NON_COHERENT;
1292                 }
1293         }
1294 }
1295
1296 static void kfd_fill_iolink_non_crat_info(struct kfd_topology_device *dev)
1297 {
1298         struct kfd_iolink_properties *link, *inbound_link;
1299         struct kfd_topology_device *peer_dev;
1300
1301         if (!dev || !dev->gpu)
1302                 return;
1303
1304         /* GPU only creates direct links so apply flags setting to all */
1305         list_for_each_entry(link, &dev->io_link_props, list) {
1306                 link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1307                 kfd_set_iolink_no_atomics(dev, NULL, link);
1308                 peer_dev = kfd_topology_device_by_proximity_domain(
1309                                 link->node_to);
1310
1311                 if (!peer_dev)
1312                         continue;
1313
1314                 /* Include the CPU peer in GPU hive if connected over xGMI. */
1315                 if (!peer_dev->gpu &&
1316                     link->iolink_type == CRAT_IOLINK_TYPE_XGMI) {
1317                         /*
1318                          * If the GPU is not part of a GPU hive, use its pci
1319                          * device location as the hive ID to bind with the CPU.
1320                          */
1321                         if (!dev->node_props.hive_id)
1322                                 dev->node_props.hive_id = pci_dev_id(dev->gpu->adev->pdev);
1323                         peer_dev->node_props.hive_id = dev->node_props.hive_id;
1324                 }
1325
1326                 list_for_each_entry(inbound_link, &peer_dev->io_link_props,
1327                                                                         list) {
1328                         if (inbound_link->node_to != link->node_from)
1329                                 continue;
1330
1331                         inbound_link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1332                         kfd_set_iolink_no_atomics(peer_dev, dev, inbound_link);
1333                         kfd_set_iolink_non_coherent(peer_dev, link, inbound_link);
1334                 }
1335         }
1336
1337         /* Create indirect links so apply flags setting to all */
1338         list_for_each_entry(link, &dev->p2p_link_props, list) {
1339                 link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1340                 kfd_set_iolink_no_atomics(dev, NULL, link);
1341                 peer_dev = kfd_topology_device_by_proximity_domain(
1342                                 link->node_to);
1343
1344                 if (!peer_dev)
1345                         continue;
1346
1347                 list_for_each_entry(inbound_link, &peer_dev->p2p_link_props,
1348                                                                         list) {
1349                         if (inbound_link->node_to != link->node_from)
1350                                 continue;
1351
1352                         inbound_link->flags = CRAT_IOLINK_FLAGS_ENABLED;
1353                         kfd_set_iolink_no_atomics(peer_dev, dev, inbound_link);
1354                         kfd_set_iolink_non_coherent(peer_dev, link, inbound_link);
1355                 }
1356         }
1357 }
1358
1359 static int kfd_build_p2p_node_entry(struct kfd_topology_device *dev,
1360                                 struct kfd_iolink_properties *p2plink)
1361 {
1362         int ret;
1363
1364         p2plink->kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
1365         if (!p2plink->kobj)
1366                 return -ENOMEM;
1367
1368         ret = kobject_init_and_add(p2plink->kobj, &iolink_type,
1369                         dev->kobj_p2plink, "%d", dev->node_props.p2p_links_count - 1);
1370         if (ret < 0) {
1371                 kobject_put(p2plink->kobj);
1372                 return ret;
1373         }
1374
1375         p2plink->attr.name = "properties";
1376         p2plink->attr.mode = KFD_SYSFS_FILE_MODE;
1377         sysfs_attr_init(&p2plink->attr);
1378         ret = sysfs_create_file(p2plink->kobj, &p2plink->attr);
1379         if (ret < 0)
1380                 return ret;
1381
1382         return 0;
1383 }
1384
1385 static int kfd_create_indirect_link_prop(struct kfd_topology_device *kdev, int gpu_node)
1386 {
1387         struct kfd_iolink_properties *gpu_link, *tmp_link, *cpu_link;
1388         struct kfd_iolink_properties *props = NULL, *props2 = NULL;
1389         struct kfd_topology_device *cpu_dev;
1390         int ret = 0;
1391         int i, num_cpu;
1392
1393         num_cpu = 0;
1394         list_for_each_entry(cpu_dev, &topology_device_list, list) {
1395                 if (cpu_dev->gpu)
1396                         break;
1397                 num_cpu++;
1398         }
1399
1400         gpu_link = list_first_entry(&kdev->io_link_props,
1401                                         struct kfd_iolink_properties, list);
1402         if (!gpu_link)
1403                 return -ENOMEM;
1404
1405         for (i = 0; i < num_cpu; i++) {
1406                 /* CPU <--> GPU */
1407                 if (gpu_link->node_to == i)
1408                         continue;
1409
1410                 /* find CPU <-->  CPU links */
1411                 cpu_link = NULL;
1412                 cpu_dev = kfd_topology_device_by_proximity_domain(i);
1413                 if (cpu_dev) {
1414                         list_for_each_entry(tmp_link,
1415                                         &cpu_dev->io_link_props, list) {
1416                                 if (tmp_link->node_to == gpu_link->node_to) {
1417                                         cpu_link = tmp_link;
1418                                         break;
1419                                 }
1420                         }
1421                 }
1422
1423                 if (!cpu_link)
1424                         return -ENOMEM;
1425
1426                 /* CPU <--> CPU <--> GPU, GPU node*/
1427                 props = kfd_alloc_struct(props);
1428                 if (!props)
1429                         return -ENOMEM;
1430
1431                 memcpy(props, gpu_link, sizeof(struct kfd_iolink_properties));
1432                 props->weight = gpu_link->weight + cpu_link->weight;
1433                 props->min_latency = gpu_link->min_latency + cpu_link->min_latency;
1434                 props->max_latency = gpu_link->max_latency + cpu_link->max_latency;
1435                 props->min_bandwidth = min(gpu_link->min_bandwidth, cpu_link->min_bandwidth);
1436                 props->max_bandwidth = min(gpu_link->max_bandwidth, cpu_link->max_bandwidth);
1437
1438                 props->node_from = gpu_node;
1439                 props->node_to = i;
1440                 kdev->node_props.p2p_links_count++;
1441                 list_add_tail(&props->list, &kdev->p2p_link_props);
1442                 ret = kfd_build_p2p_node_entry(kdev, props);
1443                 if (ret < 0)
1444                         return ret;
1445
1446                 /* for small Bar, no CPU --> GPU in-direct links */
1447                 if (kfd_dev_is_large_bar(kdev->gpu)) {
1448                         /* CPU <--> CPU <--> GPU, CPU node*/
1449                         props2 = kfd_alloc_struct(props2);
1450                         if (!props2)
1451                                 return -ENOMEM;
1452
1453                         memcpy(props2, props, sizeof(struct kfd_iolink_properties));
1454                         props2->node_from = i;
1455                         props2->node_to = gpu_node;
1456                         props2->kobj = NULL;
1457                         cpu_dev->node_props.p2p_links_count++;
1458                         list_add_tail(&props2->list, &cpu_dev->p2p_link_props);
1459                         ret = kfd_build_p2p_node_entry(cpu_dev, props2);
1460                         if (ret < 0)
1461                                 return ret;
1462                 }
1463         }
1464         return ret;
1465 }
1466
1467 #if defined(CONFIG_HSA_AMD_P2P)
1468 static int kfd_add_peer_prop(struct kfd_topology_device *kdev,
1469                 struct kfd_topology_device *peer, int from, int to)
1470 {
1471         struct kfd_iolink_properties *props = NULL;
1472         struct kfd_iolink_properties *iolink1, *iolink2, *iolink3;
1473         struct kfd_topology_device *cpu_dev;
1474         int ret = 0;
1475
1476         if (!amdgpu_device_is_peer_accessible(
1477                                 kdev->gpu->adev,
1478                                 peer->gpu->adev))
1479                 return ret;
1480
1481         iolink1 = list_first_entry(&kdev->io_link_props,
1482                                                         struct kfd_iolink_properties, list);
1483         if (!iolink1)
1484                 return -ENOMEM;
1485
1486         iolink2 = list_first_entry(&peer->io_link_props,
1487                                                         struct kfd_iolink_properties, list);
1488         if (!iolink2)
1489                 return -ENOMEM;
1490
1491         props = kfd_alloc_struct(props);
1492         if (!props)
1493                 return -ENOMEM;
1494
1495         memcpy(props, iolink1, sizeof(struct kfd_iolink_properties));
1496
1497         props->weight = iolink1->weight + iolink2->weight;
1498         props->min_latency = iolink1->min_latency + iolink2->min_latency;
1499         props->max_latency = iolink1->max_latency + iolink2->max_latency;
1500         props->min_bandwidth = min(iolink1->min_bandwidth, iolink2->min_bandwidth);
1501         props->max_bandwidth = min(iolink2->max_bandwidth, iolink2->max_bandwidth);
1502
1503         if (iolink1->node_to != iolink2->node_to) {
1504                 /* CPU->CPU  link*/
1505                 cpu_dev = kfd_topology_device_by_proximity_domain(iolink1->node_to);
1506                 if (cpu_dev) {
1507                         list_for_each_entry(iolink3, &cpu_dev->io_link_props, list)
1508                                 if (iolink3->node_to == iolink2->node_to)
1509                                         break;
1510
1511                         props->weight += iolink3->weight;
1512                         props->min_latency += iolink3->min_latency;
1513                         props->max_latency += iolink3->max_latency;
1514                         props->min_bandwidth = min(props->min_bandwidth,
1515                                                         iolink3->min_bandwidth);
1516                         props->max_bandwidth = min(props->max_bandwidth,
1517                                                         iolink3->max_bandwidth);
1518                 } else {
1519                         WARN(1, "CPU node not found");
1520                 }
1521         }
1522
1523         props->node_from = from;
1524         props->node_to = to;
1525         peer->node_props.p2p_links_count++;
1526         list_add_tail(&props->list, &peer->p2p_link_props);
1527         ret = kfd_build_p2p_node_entry(peer, props);
1528
1529         return ret;
1530 }
1531 #endif
1532
1533 static int kfd_dev_create_p2p_links(void)
1534 {
1535         struct kfd_topology_device *dev;
1536         struct kfd_topology_device *new_dev;
1537 #if defined(CONFIG_HSA_AMD_P2P)
1538         uint32_t i;
1539 #endif
1540         uint32_t k;
1541         int ret = 0;
1542
1543         k = 0;
1544         list_for_each_entry(dev, &topology_device_list, list)
1545                 k++;
1546         if (k < 2)
1547                 return 0;
1548
1549         new_dev = list_last_entry(&topology_device_list, struct kfd_topology_device, list);
1550         if (WARN_ON(!new_dev->gpu))
1551                 return 0;
1552
1553         k--;
1554
1555         /* create in-direct links */
1556         ret = kfd_create_indirect_link_prop(new_dev, k);
1557         if (ret < 0)
1558                 goto out;
1559
1560         /* create p2p links */
1561 #if defined(CONFIG_HSA_AMD_P2P)
1562         i = 0;
1563         list_for_each_entry(dev, &topology_device_list, list) {
1564                 if (dev == new_dev)
1565                         break;
1566                 if (!dev->gpu || !dev->gpu->adev ||
1567                     (dev->gpu->kfd->hive_id &&
1568                      dev->gpu->kfd->hive_id == new_dev->gpu->kfd->hive_id))
1569                         goto next;
1570
1571                 /* check if node(s) is/are peer accessible in one direction or bi-direction */
1572                 ret = kfd_add_peer_prop(new_dev, dev, i, k);
1573                 if (ret < 0)
1574                         goto out;
1575
1576                 ret = kfd_add_peer_prop(dev, new_dev, k, i);
1577                 if (ret < 0)
1578                         goto out;
1579 next:
1580                 i++;
1581         }
1582 #endif
1583
1584 out:
1585         return ret;
1586 }
1587
1588 /* Helper function. See kfd_fill_gpu_cache_info for parameter description */
1589 static int fill_in_l1_pcache(struct kfd_cache_properties **props_ext,
1590                                 struct kfd_gpu_cache_info *pcache_info,
1591                                 struct kfd_cu_info *cu_info,
1592                                 int cu_bitmask,
1593                                 int cache_type, unsigned int cu_processor_id,
1594                                 int cu_block)
1595 {
1596         unsigned int cu_sibling_map_mask;
1597         int first_active_cu;
1598         struct kfd_cache_properties *pcache = NULL;
1599
1600         cu_sibling_map_mask = cu_bitmask;
1601         cu_sibling_map_mask >>= cu_block;
1602         cu_sibling_map_mask &= ((1 << pcache_info[cache_type].num_cu_shared) - 1);
1603         first_active_cu = ffs(cu_sibling_map_mask);
1604
1605         /* CU could be inactive. In case of shared cache find the first active
1606          * CU. and incase of non-shared cache check if the CU is inactive. If
1607          * inactive active skip it
1608          */
1609         if (first_active_cu) {
1610                 pcache = kfd_alloc_struct(pcache);
1611                 if (!pcache)
1612                         return -ENOMEM;
1613
1614                 memset(pcache, 0, sizeof(struct kfd_cache_properties));
1615                 pcache->processor_id_low = cu_processor_id + (first_active_cu - 1);
1616                 pcache->cache_level = pcache_info[cache_type].cache_level;
1617                 pcache->cache_size = pcache_info[cache_type].cache_size;
1618
1619                 if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_DATA_CACHE)
1620                         pcache->cache_type |= HSA_CACHE_TYPE_DATA;
1621                 if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_INST_CACHE)
1622                         pcache->cache_type |= HSA_CACHE_TYPE_INSTRUCTION;
1623                 if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_CPU_CACHE)
1624                         pcache->cache_type |= HSA_CACHE_TYPE_CPU;
1625                 if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_SIMD_CACHE)
1626                         pcache->cache_type |= HSA_CACHE_TYPE_HSACU;
1627
1628                 /* Sibling map is w.r.t processor_id_low, so shift out
1629                  * inactive CU
1630                  */
1631                 cu_sibling_map_mask =
1632                         cu_sibling_map_mask >> (first_active_cu - 1);
1633
1634                 pcache->sibling_map[0] = (uint8_t)(cu_sibling_map_mask & 0xFF);
1635                 pcache->sibling_map[1] =
1636                                 (uint8_t)((cu_sibling_map_mask >> 8) & 0xFF);
1637                 pcache->sibling_map[2] =
1638                                 (uint8_t)((cu_sibling_map_mask >> 16) & 0xFF);
1639                 pcache->sibling_map[3] =
1640                                 (uint8_t)((cu_sibling_map_mask >> 24) & 0xFF);
1641
1642                 pcache->sibling_map_size = 4;
1643                 *props_ext = pcache;
1644
1645                 return 0;
1646         }
1647         return 1;
1648 }
1649
1650 /* Helper function. See kfd_fill_gpu_cache_info for parameter description */
1651 static int fill_in_l2_l3_pcache(struct kfd_cache_properties **props_ext,
1652                                 struct kfd_gpu_cache_info *pcache_info,
1653                                 struct kfd_cu_info *cu_info,
1654                                 int cache_type, unsigned int cu_processor_id)
1655 {
1656         unsigned int cu_sibling_map_mask;
1657         int first_active_cu;
1658         int i, j, k;
1659         struct kfd_cache_properties *pcache = NULL;
1660
1661         cu_sibling_map_mask = cu_info->cu_bitmap[0][0];
1662         cu_sibling_map_mask &=
1663                 ((1 << pcache_info[cache_type].num_cu_shared) - 1);
1664         first_active_cu = ffs(cu_sibling_map_mask);
1665
1666         /* CU could be inactive. In case of shared cache find the first active
1667          * CU. and incase of non-shared cache check if the CU is inactive. If
1668          * inactive active skip it
1669          */
1670         if (first_active_cu) {
1671                 pcache = kfd_alloc_struct(pcache);
1672                 if (!pcache)
1673                         return -ENOMEM;
1674
1675                 memset(pcache, 0, sizeof(struct kfd_cache_properties));
1676                 pcache->processor_id_low = cu_processor_id
1677                                         + (first_active_cu - 1);
1678                 pcache->cache_level = pcache_info[cache_type].cache_level;
1679                 pcache->cache_size = pcache_info[cache_type].cache_size;
1680
1681                 if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_DATA_CACHE)
1682                         pcache->cache_type |= HSA_CACHE_TYPE_DATA;
1683                 if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_INST_CACHE)
1684                         pcache->cache_type |= HSA_CACHE_TYPE_INSTRUCTION;
1685                 if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_CPU_CACHE)
1686                         pcache->cache_type |= HSA_CACHE_TYPE_CPU;
1687                 if (pcache_info[cache_type].flags & CRAT_CACHE_FLAGS_SIMD_CACHE)
1688                         pcache->cache_type |= HSA_CACHE_TYPE_HSACU;
1689
1690                 /* Sibling map is w.r.t processor_id_low, so shift out
1691                  * inactive CU
1692                  */
1693                 cu_sibling_map_mask = cu_sibling_map_mask >> (first_active_cu - 1);
1694                 k = 0;
1695
1696                 for (i = 0; i < cu_info->num_shader_engines; i++) {
1697                         for (j = 0; j < cu_info->num_shader_arrays_per_engine; j++) {
1698                                 pcache->sibling_map[k] = (uint8_t)(cu_sibling_map_mask & 0xFF);
1699                                 pcache->sibling_map[k+1] = (uint8_t)((cu_sibling_map_mask >> 8) & 0xFF);
1700                                 pcache->sibling_map[k+2] = (uint8_t)((cu_sibling_map_mask >> 16) & 0xFF);
1701                                 pcache->sibling_map[k+3] = (uint8_t)((cu_sibling_map_mask >> 24) & 0xFF);
1702                                 k += 4;
1703
1704                                 cu_sibling_map_mask = cu_info->cu_bitmap[i % 4][j + i / 4];
1705                                 cu_sibling_map_mask &= ((1 << pcache_info[cache_type].num_cu_shared) - 1);
1706                         }
1707                 }
1708                 pcache->sibling_map_size = k;
1709                 *props_ext = pcache;
1710                 return 0;
1711         }
1712         return 1;
1713 }
1714
1715 #define KFD_MAX_CACHE_TYPES 6
1716
1717 /* kfd_fill_cache_non_crat_info - Fill GPU cache info using kfd_gpu_cache_info
1718  * tables
1719  */
1720 static void kfd_fill_cache_non_crat_info(struct kfd_topology_device *dev, struct kfd_node *kdev)
1721 {
1722         struct kfd_gpu_cache_info *pcache_info = NULL;
1723         int i, j, k;
1724         int ct = 0;
1725         unsigned int cu_processor_id;
1726         int ret;
1727         unsigned int num_cu_shared;
1728         struct kfd_cu_info cu_info;
1729         struct kfd_cu_info *pcu_info;
1730         int gpu_processor_id;
1731         struct kfd_cache_properties *props_ext;
1732         int num_of_entries = 0;
1733         int num_of_cache_types = 0;
1734         struct kfd_gpu_cache_info cache_info[KFD_MAX_CACHE_TYPES];
1735
1736         amdgpu_amdkfd_get_cu_info(kdev->adev, &cu_info);
1737         pcu_info = &cu_info;
1738
1739         gpu_processor_id = dev->node_props.simd_id_base;
1740
1741         pcache_info = cache_info;
1742         num_of_cache_types = kfd_get_gpu_cache_info(kdev, &pcache_info);
1743         if (!num_of_cache_types) {
1744                 pr_warn("no cache info found\n");
1745                 return;
1746         }
1747
1748         /* For each type of cache listed in the kfd_gpu_cache_info table,
1749          * go through all available Compute Units.
1750          * The [i,j,k] loop will
1751          *              if kfd_gpu_cache_info.num_cu_shared = 1
1752          *                      will parse through all available CU
1753          *              If (kfd_gpu_cache_info.num_cu_shared != 1)
1754          *                      then it will consider only one CU from
1755          *                      the shared unit
1756          */
1757         for (ct = 0; ct < num_of_cache_types; ct++) {
1758                 cu_processor_id = gpu_processor_id;
1759                 if (pcache_info[ct].cache_level == 1) {
1760                         for (i = 0; i < pcu_info->num_shader_engines; i++) {
1761                                 for (j = 0; j < pcu_info->num_shader_arrays_per_engine; j++) {
1762                                         for (k = 0; k < pcu_info->num_cu_per_sh; k += pcache_info[ct].num_cu_shared) {
1763
1764                                                 ret = fill_in_l1_pcache(&props_ext, pcache_info, pcu_info,
1765                                                                                 pcu_info->cu_bitmap[i % 4][j + i / 4], ct,
1766                                                                                 cu_processor_id, k);
1767
1768                                                 if (ret < 0)
1769                                                         break;
1770
1771                                                 if (!ret) {
1772                                                         num_of_entries++;
1773                                                         list_add_tail(&props_ext->list, &dev->cache_props);
1774                                                 }
1775
1776                                                 /* Move to next CU block */
1777                                                 num_cu_shared = ((k + pcache_info[ct].num_cu_shared) <=
1778                                                         pcu_info->num_cu_per_sh) ?
1779                                                         pcache_info[ct].num_cu_shared :
1780                                                         (pcu_info->num_cu_per_sh - k);
1781                                                 cu_processor_id += num_cu_shared;
1782                                         }
1783                                 }
1784                         }
1785                 } else {
1786                         ret = fill_in_l2_l3_pcache(&props_ext, pcache_info,
1787                                                                 pcu_info, ct, cu_processor_id);
1788
1789                         if (ret < 0)
1790                                 break;
1791
1792                         if (!ret) {
1793                                 num_of_entries++;
1794                                 list_add_tail(&props_ext->list, &dev->cache_props);
1795                         }
1796                 }
1797         }
1798         dev->node_props.caches_count += num_of_entries;
1799         pr_debug("Added [%d] GPU cache entries\n", num_of_entries);
1800 }
1801
1802 static int kfd_topology_add_device_locked(struct kfd_node *gpu, uint32_t gpu_id,
1803                                           struct kfd_topology_device **dev)
1804 {
1805         int proximity_domain = ++topology_crat_proximity_domain;
1806         struct list_head temp_topology_device_list;
1807         void *crat_image = NULL;
1808         size_t image_size = 0;
1809         int res;
1810
1811         res = kfd_create_crat_image_virtual(&crat_image, &image_size,
1812                                             COMPUTE_UNIT_GPU, gpu,
1813                                             proximity_domain);
1814         if (res) {
1815                 pr_err("Error creating VCRAT for GPU (ID: 0x%x)\n",
1816                        gpu_id);
1817                 topology_crat_proximity_domain--;
1818                 goto err;
1819         }
1820
1821         INIT_LIST_HEAD(&temp_topology_device_list);
1822
1823         res = kfd_parse_crat_table(crat_image,
1824                                    &temp_topology_device_list,
1825                                    proximity_domain);
1826         if (res) {
1827                 pr_err("Error parsing VCRAT for GPU (ID: 0x%x)\n",
1828                        gpu_id);
1829                 topology_crat_proximity_domain--;
1830                 goto err;
1831         }
1832
1833         kfd_topology_update_device_list(&temp_topology_device_list,
1834                                         &topology_device_list);
1835
1836         *dev = kfd_assign_gpu(gpu);
1837         if (WARN_ON(!*dev)) {
1838                 res = -ENODEV;
1839                 goto err;
1840         }
1841
1842         /* Fill the cache affinity information here for the GPUs
1843          * using VCRAT
1844          */
1845         kfd_fill_cache_non_crat_info(*dev, gpu);
1846
1847         /* Update the SYSFS tree, since we added another topology
1848          * device
1849          */
1850         res = kfd_topology_update_sysfs();
1851         if (!res)
1852                 sys_props.generation_count++;
1853         else
1854                 pr_err("Failed to update GPU (ID: 0x%x) to sysfs topology. res=%d\n",
1855                        gpu_id, res);
1856
1857 err:
1858         kfd_destroy_crat_image(crat_image);
1859         return res;
1860 }
1861
1862 static void kfd_topology_set_dbg_firmware_support(struct kfd_topology_device *dev)
1863 {
1864         bool firmware_supported = true;
1865
1866         if (KFD_GC_VERSION(dev->gpu) >= IP_VERSION(11, 0, 0) &&
1867                         KFD_GC_VERSION(dev->gpu) < IP_VERSION(12, 0, 0)) {
1868                 uint32_t mes_api_rev = (dev->gpu->adev->mes.sched_version &
1869                                                 AMDGPU_MES_API_VERSION_MASK) >>
1870                                                 AMDGPU_MES_API_VERSION_SHIFT;
1871                 uint32_t mes_rev = dev->gpu->adev->mes.sched_version &
1872                                                 AMDGPU_MES_VERSION_MASK;
1873
1874                 firmware_supported = (mes_api_rev >= 14) && (mes_rev >= 64);
1875                 goto out;
1876         }
1877
1878         /*
1879          * Note: Any unlisted devices here are assumed to support exception handling.
1880          * Add additional checks here as needed.
1881          */
1882         switch (KFD_GC_VERSION(dev->gpu)) {
1883         case IP_VERSION(9, 0, 1):
1884                 firmware_supported = dev->gpu->kfd->mec_fw_version >= 459 + 32768;
1885                 break;
1886         case IP_VERSION(9, 1, 0):
1887         case IP_VERSION(9, 2, 1):
1888         case IP_VERSION(9, 2, 2):
1889         case IP_VERSION(9, 3, 0):
1890         case IP_VERSION(9, 4, 0):
1891                 firmware_supported = dev->gpu->kfd->mec_fw_version >= 459;
1892                 break;
1893         case IP_VERSION(9, 4, 1):
1894                 firmware_supported = dev->gpu->kfd->mec_fw_version >= 60;
1895                 break;
1896         case IP_VERSION(9, 4, 2):
1897                 firmware_supported = dev->gpu->kfd->mec_fw_version >= 51;
1898                 break;
1899         case IP_VERSION(10, 1, 10):
1900         case IP_VERSION(10, 1, 2):
1901         case IP_VERSION(10, 1, 1):
1902                 firmware_supported = dev->gpu->kfd->mec_fw_version >= 144;
1903                 break;
1904         case IP_VERSION(10, 3, 0):
1905         case IP_VERSION(10, 3, 2):
1906         case IP_VERSION(10, 3, 1):
1907         case IP_VERSION(10, 3, 4):
1908         case IP_VERSION(10, 3, 5):
1909                 firmware_supported = dev->gpu->kfd->mec_fw_version >= 89;
1910                 break;
1911         case IP_VERSION(10, 1, 3):
1912         case IP_VERSION(10, 3, 3):
1913                 firmware_supported = false;
1914                 break;
1915         default:
1916                 break;
1917         }
1918
1919 out:
1920         if (firmware_supported)
1921                 dev->node_props.capability |= HSA_CAP_TRAP_DEBUG_FIRMWARE_SUPPORTED;
1922 }
1923
1924 static void kfd_topology_set_capabilities(struct kfd_topology_device *dev)
1925 {
1926         dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_2_0 <<
1927                                 HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
1928                                 HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
1929
1930         dev->node_props.capability |= HSA_CAP_TRAP_DEBUG_SUPPORT |
1931                         HSA_CAP_TRAP_DEBUG_WAVE_LAUNCH_TRAP_OVERRIDE_SUPPORTED |
1932                         HSA_CAP_TRAP_DEBUG_WAVE_LAUNCH_MODE_SUPPORTED;
1933
1934         if (KFD_GC_VERSION(dev->gpu) < IP_VERSION(10, 0, 0)) {
1935                 dev->node_props.debug_prop |= HSA_DBG_WATCH_ADDR_MASK_LO_BIT_GFX9 |
1936                                                 HSA_DBG_WATCH_ADDR_MASK_HI_BIT;
1937
1938                 if (KFD_GC_VERSION(dev->gpu) < IP_VERSION(9, 4, 2))
1939                         dev->node_props.debug_prop |=
1940                                 HSA_DBG_DISPATCH_INFO_ALWAYS_VALID;
1941                 else
1942                         dev->node_props.capability |=
1943                                 HSA_CAP_TRAP_DEBUG_PRECISE_MEMORY_OPERATIONS_SUPPORTED;
1944         } else {
1945                 dev->node_props.debug_prop |= HSA_DBG_WATCH_ADDR_MASK_LO_BIT_GFX10 |
1946                                         HSA_DBG_WATCH_ADDR_MASK_HI_BIT;
1947
1948                 if (KFD_GC_VERSION(dev->gpu) < IP_VERSION(11, 0, 0))
1949                         dev->node_props.debug_prop |= HSA_DBG_DISPATCH_INFO_ALWAYS_VALID;
1950                 else
1951                         dev->node_props.capability |=
1952                                 HSA_CAP_TRAP_DEBUG_PRECISE_MEMORY_OPERATIONS_SUPPORTED;
1953         }
1954
1955         kfd_topology_set_dbg_firmware_support(dev);
1956 }
1957
1958 int kfd_topology_add_device(struct kfd_node *gpu)
1959 {
1960         uint32_t gpu_id;
1961         struct kfd_topology_device *dev;
1962         struct kfd_cu_info cu_info;
1963         int res = 0;
1964         int i;
1965         const char *asic_name = amdgpu_asic_name[gpu->adev->asic_type];
1966
1967         gpu_id = kfd_generate_gpu_id(gpu);
1968         pr_debug("Adding new GPU (ID: 0x%x) to topology\n", gpu_id);
1969
1970         /* Check to see if this gpu device exists in the topology_device_list.
1971          * If so, assign the gpu to that device,
1972          * else create a Virtual CRAT for this gpu device and then parse that
1973          * CRAT to create a new topology device. Once created assign the gpu to
1974          * that topology device
1975          */
1976         down_write(&topology_lock);
1977         dev = kfd_assign_gpu(gpu);
1978         if (!dev)
1979                 res = kfd_topology_add_device_locked(gpu, gpu_id, &dev);
1980         up_write(&topology_lock);
1981         if (res)
1982                 return res;
1983
1984         dev->gpu_id = gpu_id;
1985         gpu->id = gpu_id;
1986
1987         kfd_dev_create_p2p_links();
1988
1989         /* TODO: Move the following lines to function
1990          *      kfd_add_non_crat_information
1991          */
1992
1993         /* Fill-in additional information that is not available in CRAT but
1994          * needed for the topology
1995          */
1996
1997         amdgpu_amdkfd_get_cu_info(dev->gpu->adev, &cu_info);
1998
1999         for (i = 0; i < KFD_TOPOLOGY_PUBLIC_NAME_SIZE-1; i++) {
2000                 dev->node_props.name[i] = __tolower(asic_name[i]);
2001                 if (asic_name[i] == '\0')
2002                         break;
2003         }
2004         dev->node_props.name[i] = '\0';
2005
2006         dev->node_props.simd_arrays_per_engine =
2007                 cu_info.num_shader_arrays_per_engine;
2008
2009         dev->node_props.gfx_target_version =
2010                                 gpu->kfd->device_info.gfx_target_version;
2011         dev->node_props.vendor_id = gpu->adev->pdev->vendor;
2012         dev->node_props.device_id = gpu->adev->pdev->device;
2013         dev->node_props.capability |=
2014                 ((dev->gpu->adev->rev_id << HSA_CAP_ASIC_REVISION_SHIFT) &
2015                         HSA_CAP_ASIC_REVISION_MASK);
2016
2017         dev->node_props.location_id = pci_dev_id(gpu->adev->pdev);
2018         if (KFD_GC_VERSION(dev->gpu->kfd) == IP_VERSION(9, 4, 3))
2019                 dev->node_props.location_id |= dev->gpu->node_id;
2020
2021         dev->node_props.domain = pci_domain_nr(gpu->adev->pdev->bus);
2022         dev->node_props.max_engine_clk_fcompute =
2023                 amdgpu_amdkfd_get_max_engine_clock_in_mhz(dev->gpu->adev);
2024         dev->node_props.max_engine_clk_ccompute =
2025                 cpufreq_quick_get_max(0) / 1000;
2026
2027         if (gpu->xcp)
2028                 dev->node_props.drm_render_minor = gpu->xcp->ddev->render->index;
2029         else
2030                 dev->node_props.drm_render_minor =
2031                                 gpu->kfd->shared_resources.drm_render_minor;
2032
2033         dev->node_props.hive_id = gpu->kfd->hive_id;
2034         dev->node_props.num_sdma_engines = kfd_get_num_sdma_engines(gpu);
2035         dev->node_props.num_sdma_xgmi_engines =
2036                                         kfd_get_num_xgmi_sdma_engines(gpu);
2037         dev->node_props.num_sdma_queues_per_engine =
2038                                 gpu->kfd->device_info.num_sdma_queues_per_engine -
2039                                 gpu->kfd->device_info.num_reserved_sdma_queues_per_engine;
2040         dev->node_props.num_gws = (dev->gpu->gws &&
2041                 dev->gpu->dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) ?
2042                 dev->gpu->adev->gds.gws_size : 0;
2043         dev->node_props.num_cp_queues = get_cp_queues_num(dev->gpu->dqm);
2044
2045         kfd_fill_mem_clk_max_info(dev);
2046         kfd_fill_iolink_non_crat_info(dev);
2047
2048         switch (dev->gpu->adev->asic_type) {
2049         case CHIP_KAVERI:
2050         case CHIP_HAWAII:
2051         case CHIP_TONGA:
2052                 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_PRE_1_0 <<
2053                         HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
2054                         HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
2055                 break;
2056         case CHIP_CARRIZO:
2057         case CHIP_FIJI:
2058         case CHIP_POLARIS10:
2059         case CHIP_POLARIS11:
2060         case CHIP_POLARIS12:
2061         case CHIP_VEGAM:
2062                 pr_debug("Adding doorbell packet type capability\n");
2063                 dev->node_props.capability |= ((HSA_CAP_DOORBELL_TYPE_1_0 <<
2064                         HSA_CAP_DOORBELL_TYPE_TOTALBITS_SHIFT) &
2065                         HSA_CAP_DOORBELL_TYPE_TOTALBITS_MASK);
2066                 break;
2067         default:
2068                 if (KFD_GC_VERSION(dev->gpu) < IP_VERSION(9, 0, 1))
2069                         WARN(1, "Unexpected ASIC family %u",
2070                              dev->gpu->adev->asic_type);
2071                 else
2072                         kfd_topology_set_capabilities(dev);
2073         }
2074
2075         /*
2076          * Overwrite ATS capability according to needs_iommu_device to fix
2077          * potential missing corresponding bit in CRAT of BIOS.
2078          */
2079         if (dev->gpu->kfd->use_iommu_v2)
2080                 dev->node_props.capability |= HSA_CAP_ATS_PRESENT;
2081         else
2082                 dev->node_props.capability &= ~HSA_CAP_ATS_PRESENT;
2083
2084         /* Fix errors in CZ CRAT.
2085          * simd_count: Carrizo CRAT reports wrong simd_count, probably
2086          *              because it doesn't consider masked out CUs
2087          * max_waves_per_simd: Carrizo reports wrong max_waves_per_simd
2088          */
2089         if (dev->gpu->adev->asic_type == CHIP_CARRIZO) {
2090                 dev->node_props.simd_count =
2091                         cu_info.simd_per_cu * cu_info.cu_active_number;
2092                 dev->node_props.max_waves_per_simd = 10;
2093         }
2094
2095         /* kfd only concerns sram ecc on GFX and HBM ecc on UMC */
2096         dev->node_props.capability |=
2097                 ((dev->gpu->adev->ras_enabled & BIT(AMDGPU_RAS_BLOCK__GFX)) != 0) ?
2098                 HSA_CAP_SRAM_EDCSUPPORTED : 0;
2099         dev->node_props.capability |=
2100                 ((dev->gpu->adev->ras_enabled & BIT(AMDGPU_RAS_BLOCK__UMC)) != 0) ?
2101                 HSA_CAP_MEM_EDCSUPPORTED : 0;
2102
2103         if (KFD_GC_VERSION(dev->gpu) != IP_VERSION(9, 0, 1))
2104                 dev->node_props.capability |= (dev->gpu->adev->ras_enabled != 0) ?
2105                         HSA_CAP_RASEVENTNOTIFY : 0;
2106
2107         if (KFD_IS_SVM_API_SUPPORTED(dev->gpu->adev))
2108                 dev->node_props.capability |= HSA_CAP_SVMAPI_SUPPORTED;
2109
2110         if (dev->gpu->adev->gmc.is_app_apu ||
2111                 dev->gpu->adev->gmc.xgmi.connected_to_cpu)
2112                 dev->node_props.capability |= HSA_CAP_FLAGS_COHERENTHOSTACCESS;
2113
2114         kfd_debug_print_topology();
2115
2116         kfd_notify_gpu_change(gpu_id, 1);
2117
2118         return 0;
2119 }
2120
2121 /**
2122  * kfd_topology_update_io_links() - Update IO links after device removal.
2123  * @proximity_domain: Proximity domain value of the dev being removed.
2124  *
2125  * The topology list currently is arranged in increasing order of
2126  * proximity domain.
2127  *
2128  * Two things need to be done when a device is removed:
2129  * 1. All the IO links to this device need to be removed.
2130  * 2. All nodes after the current device node need to move
2131  *    up once this device node is removed from the topology
2132  *    list. As a result, the proximity domain values for
2133  *    all nodes after the node being deleted reduce by 1.
2134  *    This would also cause the proximity domain values for
2135  *    io links to be updated based on new proximity domain
2136  *    values.
2137  *
2138  * Context: The caller must hold write topology_lock.
2139  */
2140 static void kfd_topology_update_io_links(int proximity_domain)
2141 {
2142         struct kfd_topology_device *dev;
2143         struct kfd_iolink_properties *iolink, *p2plink, *tmp;
2144
2145         list_for_each_entry(dev, &topology_device_list, list) {
2146                 if (dev->proximity_domain > proximity_domain)
2147                         dev->proximity_domain--;
2148
2149                 list_for_each_entry_safe(iolink, tmp, &dev->io_link_props, list) {
2150                         /*
2151                          * If there is an io link to the dev being deleted
2152                          * then remove that IO link also.
2153                          */
2154                         if (iolink->node_to == proximity_domain) {
2155                                 list_del(&iolink->list);
2156                                 dev->node_props.io_links_count--;
2157                         } else {
2158                                 if (iolink->node_from > proximity_domain)
2159                                         iolink->node_from--;
2160                                 if (iolink->node_to > proximity_domain)
2161                                         iolink->node_to--;
2162                         }
2163                 }
2164
2165                 list_for_each_entry_safe(p2plink, tmp, &dev->p2p_link_props, list) {
2166                         /*
2167                          * If there is a p2p link to the dev being deleted
2168                          * then remove that p2p link also.
2169                          */
2170                         if (p2plink->node_to == proximity_domain) {
2171                                 list_del(&p2plink->list);
2172                                 dev->node_props.p2p_links_count--;
2173                         } else {
2174                                 if (p2plink->node_from > proximity_domain)
2175                                         p2plink->node_from--;
2176                                 if (p2plink->node_to > proximity_domain)
2177                                         p2plink->node_to--;
2178                         }
2179                 }
2180         }
2181 }
2182
2183 int kfd_topology_remove_device(struct kfd_node *gpu)
2184 {
2185         struct kfd_topology_device *dev, *tmp;
2186         uint32_t gpu_id;
2187         int res = -ENODEV;
2188         int i = 0;
2189
2190         down_write(&topology_lock);
2191
2192         list_for_each_entry_safe(dev, tmp, &topology_device_list, list) {
2193                 if (dev->gpu == gpu) {
2194                         gpu_id = dev->gpu_id;
2195                         kfd_remove_sysfs_node_entry(dev);
2196                         kfd_release_topology_device(dev);
2197                         sys_props.num_devices--;
2198                         kfd_topology_update_io_links(i);
2199                         topology_crat_proximity_domain = sys_props.num_devices-1;
2200                         sys_props.generation_count++;
2201                         res = 0;
2202                         if (kfd_topology_update_sysfs() < 0)
2203                                 kfd_topology_release_sysfs();
2204                         break;
2205                 }
2206                 i++;
2207         }
2208
2209         up_write(&topology_lock);
2210
2211         if (!res)
2212                 kfd_notify_gpu_change(gpu_id, 0);
2213
2214         return res;
2215 }
2216
2217 /* kfd_topology_enum_kfd_devices - Enumerate through all devices in KFD
2218  *      topology. If GPU device is found @idx, then valid kfd_dev pointer is
2219  *      returned through @kdev
2220  * Return -     0: On success (@kdev will be NULL for non GPU nodes)
2221  *              -1: If end of list
2222  */
2223 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_node **kdev)
2224 {
2225
2226         struct kfd_topology_device *top_dev;
2227         uint8_t device_idx = 0;
2228
2229         *kdev = NULL;
2230         down_read(&topology_lock);
2231
2232         list_for_each_entry(top_dev, &topology_device_list, list) {
2233                 if (device_idx == idx) {
2234                         *kdev = top_dev->gpu;
2235                         up_read(&topology_lock);
2236                         return 0;
2237                 }
2238
2239                 device_idx++;
2240         }
2241
2242         up_read(&topology_lock);
2243
2244         return -1;
2245
2246 }
2247
2248 static int kfd_cpumask_to_apic_id(const struct cpumask *cpumask)
2249 {
2250         int first_cpu_of_numa_node;
2251
2252         if (!cpumask || cpumask == cpu_none_mask)
2253                 return -1;
2254         first_cpu_of_numa_node = cpumask_first(cpumask);
2255         if (first_cpu_of_numa_node >= nr_cpu_ids)
2256                 return -1;
2257 #ifdef CONFIG_X86_64
2258         return cpu_data(first_cpu_of_numa_node).apicid;
2259 #else
2260         return first_cpu_of_numa_node;
2261 #endif
2262 }
2263
2264 /* kfd_numa_node_to_apic_id - Returns the APIC ID of the first logical processor
2265  *      of the given NUMA node (numa_node_id)
2266  * Return -1 on failure
2267  */
2268 int kfd_numa_node_to_apic_id(int numa_node_id)
2269 {
2270         if (numa_node_id == -1) {
2271                 pr_warn("Invalid NUMA Node. Use online CPU mask\n");
2272                 return kfd_cpumask_to_apic_id(cpu_online_mask);
2273         }
2274         return kfd_cpumask_to_apic_id(cpumask_of_node(numa_node_id));
2275 }
2276
2277 void kfd_double_confirm_iommu_support(struct kfd_dev *gpu)
2278 {
2279         struct kfd_topology_device *dev;
2280
2281         gpu->use_iommu_v2 = false;
2282
2283         if (!gpu->device_info.needs_iommu_device)
2284                 return;
2285
2286         down_read(&topology_lock);
2287
2288         /* Only use IOMMUv2 if there is an APU topology node with no GPU
2289          * assigned yet. This GPU will be assigned to it.
2290          */
2291         list_for_each_entry(dev, &topology_device_list, list)
2292                 if (dev->node_props.cpu_cores_count &&
2293                     dev->node_props.simd_count &&
2294                     !dev->gpu)
2295                         gpu->use_iommu_v2 = true;
2296
2297         up_read(&topology_lock);
2298 }
2299
2300 #if defined(CONFIG_DEBUG_FS)
2301
2302 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data)
2303 {
2304         struct kfd_topology_device *dev;
2305         unsigned int i = 0;
2306         int r = 0;
2307
2308         down_read(&topology_lock);
2309
2310         list_for_each_entry(dev, &topology_device_list, list) {
2311                 if (!dev->gpu) {
2312                         i++;
2313                         continue;
2314                 }
2315
2316                 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
2317                 r = dqm_debugfs_hqds(m, dev->gpu->dqm);
2318                 if (r)
2319                         break;
2320         }
2321
2322         up_read(&topology_lock);
2323
2324         return r;
2325 }
2326
2327 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data)
2328 {
2329         struct kfd_topology_device *dev;
2330         unsigned int i = 0;
2331         int r = 0;
2332
2333         down_read(&topology_lock);
2334
2335         list_for_each_entry(dev, &topology_device_list, list) {
2336                 if (!dev->gpu) {
2337                         i++;
2338                         continue;
2339                 }
2340
2341                 seq_printf(m, "Node %u, gpu_id %x:\n", i++, dev->gpu->id);
2342                 r = pm_debugfs_runlist(m, &dev->gpu->dqm->packet_mgr);
2343                 if (r)
2344                         break;
2345         }
2346
2347         up_read(&topology_lock);
2348
2349         return r;
2350 }
2351
2352 #endif