drivers/gpu/drm/amd/amdgpu/amdgpu_xgmi.c

   1 /*
   2  * Copyright 2018 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  *
  22  *
  23  */
  24 #include <linux/list.h>
  25 #include "amdgpu.h"
  26 #include "amdgpu_xgmi.h"
  27 #include "amdgpu_smu.h"
  28 #include "amdgpu_ras.h"
  29 #include "df/df_3_6_offset.h"
  30
  31 static DEFINE_MUTEX(xgmi_mutex);
  32
  33 #define AMDGPU_MAX_XGMI_HIVE                    8
  34 #define AMDGPU_MAX_XGMI_DEVICE_PER_HIVE         4
  35
  36 static struct amdgpu_hive_info xgmi_hives[AMDGPU_MAX_XGMI_HIVE];
  37 static unsigned hive_count = 0;
  38
  39 void *amdgpu_xgmi_hive_try_lock(struct amdgpu_hive_info *hive)
  40 {
  41         return &hive->device_list;
  42 }
  43
  44 /**
  45  * DOC: AMDGPU XGMI Support
  46  *
  47  * XGMI is a high speed interconnect that joins multiple GPU cards
  48  * into a homogeneous memory space that is organized by a collective
  49  * hive ID and individual node IDs, both of which are 64-bit numbers.
  50  *
  51  * The file xgmi_device_id contains the unique per GPU device ID and
  52  * is stored in the /sys/class/drm/card${cardno}/device/ directory.
  53  *
  54  * Inside the device directory a sub-directory 'xgmi_hive_info' is
  55  * created which contains the hive ID and the list of nodes.
  56  *
  57  * The hive ID is stored in:
  58  *   /sys/class/drm/card${cardno}/device/xgmi_hive_info/xgmi_hive_id
  59  *
  60  * The node information is stored in numbered directories:
  61  *   /sys/class/drm/card${cardno}/device/xgmi_hive_info/node${nodeno}/xgmi_device_id
  62  *
  63  * Each device has their own xgmi_hive_info direction with a mirror
  64  * set of node sub-directories.
  65  *
  66  * The XGMI memory space is built by contiguously adding the power of
  67  * two padded VRAM space from each node to each other.
  68  *
  69  */
  70
  71
  72 static ssize_t amdgpu_xgmi_show_hive_id(struct device *dev,
  73                 struct device_attribute *attr, char *buf)
  74 {
  75         struct amdgpu_hive_info *hive =
  76                         container_of(attr, struct amdgpu_hive_info, dev_attr);
  77
  78         return snprintf(buf, PAGE_SIZE, "%llu\n", hive->hive_id);
  79 }
  80
  81 static int amdgpu_xgmi_sysfs_create(struct amdgpu_device *adev,
  82                                     struct amdgpu_hive_info *hive)
  83 {
  84         int ret = 0;
  85
  86         if (WARN_ON(hive->kobj))
  87                 return -EINVAL;
  88
  89         hive->kobj = kobject_create_and_add("xgmi_hive_info", &adev->dev->kobj);
  90         if (!hive->kobj) {
  91                 dev_err(adev->dev, "XGMI: Failed to allocate sysfs entry!\n");
  92                 return -EINVAL;
  93         }
  94
  95         hive->dev_attr = (struct device_attribute) {
  96                 .attr = {
  97                         .name = "xgmi_hive_id",
  98                         .mode = S_IRUGO,
  99
 100                 },
 101                 .show = amdgpu_xgmi_show_hive_id,
 102         };
 103
 104         ret = sysfs_create_file(hive->kobj, &hive->dev_attr.attr);
 105         if (ret) {
 106                 dev_err(adev->dev, "XGMI: Failed to create device file xgmi_hive_id\n");
 107                 kobject_del(hive->kobj);
 108                 kobject_put(hive->kobj);
 109                 hive->kobj = NULL;
 110         }
 111
 112         return ret;
 113 }
 114
 115 static void amdgpu_xgmi_sysfs_destroy(struct amdgpu_device *adev,
 116                                     struct amdgpu_hive_info *hive)
 117 {
 118         sysfs_remove_file(hive->kobj, &hive->dev_attr.attr);
 119         kobject_del(hive->kobj);
 120         kobject_put(hive->kobj);
 121         hive->kobj = NULL;
 122 }
 123
 124 static ssize_t amdgpu_xgmi_show_device_id(struct device *dev,
 125                                      struct device_attribute *attr,
 126                                      char *buf)
 127 {
 128         struct drm_device *ddev = dev_get_drvdata(dev);
 129         struct amdgpu_device *adev = ddev->dev_private;
 130
 131         return snprintf(buf, PAGE_SIZE, "%llu\n", adev->gmc.xgmi.node_id);
 132
 133 }
 134
 135 #define AMDGPU_XGMI_SET_FICAA(o)        ((o) | 0x456801)
 136 static ssize_t amdgpu_xgmi_show_error(struct device *dev,
 137                                       struct device_attribute *attr,
 138                                       char *buf)
 139 {
 140         struct drm_device *ddev = dev_get_drvdata(dev);
 141         struct amdgpu_device *adev = ddev->dev_private;
 142         uint32_t ficaa_pie_ctl_in, ficaa_pie_status_in;
 143         uint64_t fica_out;
 144         unsigned int error_count = 0;
 145
 146         ficaa_pie_ctl_in = AMDGPU_XGMI_SET_FICAA(0x200);
 147         ficaa_pie_status_in = AMDGPU_XGMI_SET_FICAA(0x208);
 148
 149         fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_ctl_in);
 150         if (fica_out != 0x1f)
 151                 pr_err("xGMI error counters not enabled!\n");
 152
 153         fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_status_in);
 154
 155         if ((fica_out & 0xffff) == 2)
 156                 error_count = ((fica_out >> 62) & 0x1) + (fica_out >> 63);
 157
 158         adev->df.funcs->set_fica(adev, ficaa_pie_status_in, 0, 0);
 159
 160         return snprintf(buf, PAGE_SIZE, "%d\n", error_count);
 161 }
 162
 163
 164 static DEVICE_ATTR(xgmi_device_id, S_IRUGO, amdgpu_xgmi_show_device_id, NULL);
 165 static DEVICE_ATTR(xgmi_error, S_IRUGO, amdgpu_xgmi_show_error, NULL);
 166
 167 static int amdgpu_xgmi_sysfs_add_dev_info(struct amdgpu_device *adev,
 168                                          struct amdgpu_hive_info *hive)
 169 {
 170         int ret = 0;
 171         char node[10] = { 0 };
 172
 173         /* Create xgmi device id file */
 174         ret = device_create_file(adev->dev, &dev_attr_xgmi_device_id);
 175         if (ret) {
 176                 dev_err(adev->dev, "XGMI: Failed to create device file xgmi_device_id\n");
 177                 return ret;
 178         }
 179
 180         /* Create xgmi error file */
 181         ret = device_create_file(adev->dev, &dev_attr_xgmi_error);
 182         if (ret)
 183                 pr_err("failed to create xgmi_error\n");
 184
 185
 186         /* Create sysfs link to hive info folder on the first device */
 187         if (adev != hive->adev) {
 188                 ret = sysfs_create_link(&adev->dev->kobj, hive->kobj,
 189                                         "xgmi_hive_info");
 190                 if (ret) {
 191                         dev_err(adev->dev, "XGMI: Failed to create link to hive info");
 192                         goto remove_file;
 193                 }
 194         }
 195
 196         sprintf(node, "node%d", hive->number_devices);
 197         /* Create sysfs link form the hive folder to yourself */
 198         ret = sysfs_create_link(hive->kobj, &adev->dev->kobj, node);
 199         if (ret) {
 200                 dev_err(adev->dev, "XGMI: Failed to create link from hive info");
 201                 goto remove_link;
 202         }
 203
 204         goto success;
 205
 206
 207 remove_link:
 208         sysfs_remove_link(&adev->dev->kobj, adev->ddev->unique);
 209
 210 remove_file:
 211         device_remove_file(adev->dev, &dev_attr_xgmi_device_id);
 212
 213 success:
 214         return ret;
 215 }
 216
 217 static void amdgpu_xgmi_sysfs_rem_dev_info(struct amdgpu_device *adev,
 218                                           struct amdgpu_hive_info *hive)
 219 {
 220         device_remove_file(adev->dev, &dev_attr_xgmi_device_id);
 221         sysfs_remove_link(&adev->dev->kobj, adev->ddev->unique);
 222         sysfs_remove_link(hive->kobj, adev->ddev->unique);
 223 }
 224
 225
 226
 227 struct amdgpu_hive_info *amdgpu_get_xgmi_hive(struct amdgpu_device *adev, int lock)
 228 {
 229         int i;
 230         struct amdgpu_hive_info *tmp;
 231
 232         if (!adev->gmc.xgmi.hive_id)
 233                 return NULL;
 234
 235         mutex_lock(&xgmi_mutex);
 236
 237         for (i = 0 ; i < hive_count; ++i) {
 238                 tmp = &xgmi_hives[i];
 239                 if (tmp->hive_id == adev->gmc.xgmi.hive_id) {
 240                         if (lock)
 241                                 mutex_lock(&tmp->hive_lock);
 242                         mutex_unlock(&xgmi_mutex);
 243                         return tmp;
 244                 }
 245         }
 246         if (i >= AMDGPU_MAX_XGMI_HIVE) {
 247                 mutex_unlock(&xgmi_mutex);
 248                 return NULL;
 249         }
 250
 251         /* initialize new hive if not exist */
 252         tmp = &xgmi_hives[hive_count++];
 253
 254         if (amdgpu_xgmi_sysfs_create(adev, tmp)) {
 255                 mutex_unlock(&xgmi_mutex);
 256                 return NULL;
 257         }
 258
 259         tmp->adev = adev;
 260         tmp->hive_id = adev->gmc.xgmi.hive_id;
 261         INIT_LIST_HEAD(&tmp->device_list);
 262         mutex_init(&tmp->hive_lock);
 263         mutex_init(&tmp->reset_lock);
 264         task_barrier_init(&tmp->tb);
 265
 266         if (lock)
 267                 mutex_lock(&tmp->hive_lock);
 268         tmp->pstate = -1;
 269         mutex_unlock(&xgmi_mutex);
 270
 271         return tmp;
 272 }
 273
 274 int amdgpu_xgmi_set_pstate(struct amdgpu_device *adev, int pstate)
 275 {
 276         int ret = 0;
 277         struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev, 0);
 278         struct amdgpu_device *tmp_adev;
 279         bool update_hive_pstate = true;
 280         bool is_high_pstate = pstate && adev->asic_type == CHIP_VEGA20;
 281
 282         if (!hive)
 283                 return 0;
 284
 285         mutex_lock(&hive->hive_lock);
 286
 287         if (hive->pstate == pstate) {
 288                 adev->pstate = is_high_pstate ? pstate : adev->pstate;
 289                 goto out;
 290         }
 291
 292         dev_dbg(adev->dev, "Set xgmi pstate %d.\n", pstate);
 293
 294         ret = amdgpu_dpm_set_xgmi_pstate(adev, pstate);
 295         if (ret) {
 296                 dev_err(adev->dev,
 297                         "XGMI: Set pstate failure on device %llx, hive %llx, ret %d",
 298                         adev->gmc.xgmi.node_id,
 299                         adev->gmc.xgmi.hive_id, ret);
 300                 goto out;
 301         }
 302
 303         /* Update device pstate */
 304         adev->pstate = pstate;
 305
 306         /*
 307          * Update the hive pstate only all devices of the hive
 308          * are in the same pstate
 309          */
 310         list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
 311                 if (tmp_adev->pstate != adev->pstate) {
 312                         update_hive_pstate = false;
 313                         break;
 314                 }
 315         }
 316         if (update_hive_pstate || is_high_pstate)
 317                 hive->pstate = pstate;
 318
 319 out:
 320         mutex_unlock(&hive->hive_lock);
 321
 322         return ret;
 323 }
 324
 325 int amdgpu_xgmi_update_topology(struct amdgpu_hive_info *hive, struct amdgpu_device *adev)
 326 {
 327         int ret = -EINVAL;
 328
 329         /* Each psp need to set the latest topology */
 330         ret = psp_xgmi_set_topology_info(&adev->psp,
 331                                          hive->number_devices,
 332                                          &adev->psp.xgmi_context.top_info);
 333         if (ret)
 334                 dev_err(adev->dev,
 335                         "XGMI: Set topology failure on device %llx, hive %llx, ret %d",
 336                         adev->gmc.xgmi.node_id,
 337                         adev->gmc.xgmi.hive_id, ret);
 338
 339         return ret;
 340 }
 341
 342
 343 int amdgpu_xgmi_get_hops_count(struct amdgpu_device *adev,
 344                 struct amdgpu_device *peer_adev)
 345 {
 346         struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info;
 347         int i;
 348
 349         for (i = 0 ; i < top->num_nodes; ++i)
 350                 if (top->nodes[i].node_id == peer_adev->gmc.xgmi.node_id)
 351                         return top->nodes[i].num_hops;
 352         return  -EINVAL;
 353 }
 354
 355 int amdgpu_xgmi_add_device(struct amdgpu_device *adev)
 356 {
 357         struct psp_xgmi_topology_info *top_info;
 358         struct amdgpu_hive_info *hive;
 359         struct amdgpu_xgmi      *entry;
 360         struct amdgpu_device *tmp_adev = NULL;
 361
 362         int count = 0, ret = 0;
 363
 364         if (!adev->gmc.xgmi.supported)
 365                 return 0;
 366
 367         if (amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) {
 368                 ret = psp_xgmi_get_hive_id(&adev->psp, &adev->gmc.xgmi.hive_id);
 369                 if (ret) {
 370                         dev_err(adev->dev,
 371                                 "XGMI: Failed to get hive id\n");
 372                         return ret;
 373                 }
 374
 375                 ret = psp_xgmi_get_node_id(&adev->psp, &adev->gmc.xgmi.node_id);
 376                 if (ret) {
 377                         dev_err(adev->dev,
 378                                 "XGMI: Failed to get node id\n");
 379                         return ret;
 380                 }
 381         } else {
 382                 adev->gmc.xgmi.hive_id = 16;
 383                 adev->gmc.xgmi.node_id = adev->gmc.xgmi.physical_node_id + 16;
 384         }
 385
 386         hive = amdgpu_get_xgmi_hive(adev, 1);
 387         if (!hive) {
 388                 ret = -EINVAL;
 389                 dev_err(adev->dev,
 390                         "XGMI: node 0x%llx, can not match hive 0x%llx in the hive list.\n",
 391                         adev->gmc.xgmi.node_id, adev->gmc.xgmi.hive_id);
 392                 goto exit;
 393         }
 394
 395         /* Set default device pstate */
 396         adev->pstate = -1;
 397
 398         top_info = &adev->psp.xgmi_context.top_info;
 399
 400         list_add_tail(&adev->gmc.xgmi.head, &hive->device_list);
 401         list_for_each_entry(entry, &hive->device_list, head)
 402                 top_info->nodes[count++].node_id = entry->node_id;
 403         top_info->num_nodes = count;
 404         hive->number_devices = count;
 405
 406         task_barrier_add_task(&hive->tb);
 407
 408         if (amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) {
 409                 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
 410                         /* update node list for other device in the hive */
 411                         if (tmp_adev != adev) {
 412                                 top_info = &tmp_adev->psp.xgmi_context.top_info;
 413                                 top_info->nodes[count - 1].node_id =
 414                                         adev->gmc.xgmi.node_id;
 415                                 top_info->num_nodes = count;
 416                         }
 417                         ret = amdgpu_xgmi_update_topology(hive, tmp_adev);
 418                         if (ret)
 419                                 goto exit;
 420                 }
 421
 422                 /* get latest topology info for each device from psp */
 423                 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
 424                         ret = psp_xgmi_get_topology_info(&tmp_adev->psp, count,
 425                                         &tmp_adev->psp.xgmi_context.top_info);
 426                         if (ret) {
 427                                 dev_err(tmp_adev->dev,
 428                                         "XGMI: Get topology failure on device %llx, hive %llx, ret %d",
 429                                         tmp_adev->gmc.xgmi.node_id,
 430                                         tmp_adev->gmc.xgmi.hive_id, ret);
 431                                 /* To do : continue with some node failed or disable the whole hive */
 432                                 goto exit;
 433                         }
 434                 }
 435         }
 436
 437         if (!ret)
 438                 ret = amdgpu_xgmi_sysfs_add_dev_info(adev, hive);
 439
 440
 441         mutex_unlock(&hive->hive_lock);
 442 exit:
 443         if (!ret)
 444                 dev_info(adev->dev, "XGMI: Add node %d, hive 0x%llx.\n",
 445                          adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id);
 446         else
 447                 dev_err(adev->dev, "XGMI: Failed to add node %d, hive 0x%llx ret: %d\n",
 448                         adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id,
 449                         ret);
 450
 451         return ret;
 452 }
 453
 454 void amdgpu_xgmi_remove_device(struct amdgpu_device *adev)
 455 {
 456         struct amdgpu_hive_info *hive;
 457
 458         if (!adev->gmc.xgmi.supported)
 459                 return;
 460
 461         hive = amdgpu_get_xgmi_hive(adev, 1);
 462         if (!hive)
 463                 return;
 464
 465         if (!(hive->number_devices--)) {
 466                 amdgpu_xgmi_sysfs_destroy(adev, hive);
 467                 mutex_destroy(&hive->hive_lock);
 468                 mutex_destroy(&hive->reset_lock);
 469         } else {
 470                 task_barrier_rem_task(&hive->tb);
 471                 amdgpu_xgmi_sysfs_rem_dev_info(adev, hive);
 472                 mutex_unlock(&hive->hive_lock);
 473         }
 474 }
 475
 476 int amdgpu_xgmi_ras_late_init(struct amdgpu_device *adev)
 477 {
 478         int r;
 479         struct ras_ih_if ih_info = {
 480                 .cb = NULL,
 481         };
 482         struct ras_fs_if fs_info = {
 483                 .sysfs_name = "xgmi_wafl_err_count",
 484                 .debugfs_name = "xgmi_wafl_err_inject",
 485         };
 486
 487         if (!adev->gmc.xgmi.supported ||
 488             adev->gmc.xgmi.num_physical_nodes == 0)
 489                 return 0;
 490
 491         if (!adev->gmc.xgmi.ras_if) {
 492                 adev->gmc.xgmi.ras_if = kmalloc(sizeof(struct ras_common_if), GFP_KERNEL);
 493                 if (!adev->gmc.xgmi.ras_if)
 494                         return -ENOMEM;
 495                 adev->gmc.xgmi.ras_if->block = AMDGPU_RAS_BLOCK__XGMI_WAFL;
 496                 adev->gmc.xgmi.ras_if->type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE;
 497                 adev->gmc.xgmi.ras_if->sub_block_index = 0;
 498                 strcpy(adev->gmc.xgmi.ras_if->name, "xgmi_wafl");
 499         }
 500         ih_info.head = fs_info.head = *adev->gmc.xgmi.ras_if;
 501         r = amdgpu_ras_late_init(adev, adev->gmc.xgmi.ras_if,
 502                                  &fs_info, &ih_info);
 503         if (r || !amdgpu_ras_is_supported(adev, adev->gmc.xgmi.ras_if->block)) {
 504                 kfree(adev->gmc.xgmi.ras_if);
 505                 adev->gmc.xgmi.ras_if = NULL;
 506         }
 507
 508         return r;
 509 }
 510
 511 void amdgpu_xgmi_ras_fini(struct amdgpu_device *adev)
 512 {
 513         if (amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL) &&
 514                         adev->gmc.xgmi.ras_if) {
 515                 struct ras_common_if *ras_if = adev->gmc.xgmi.ras_if;
 516                 struct ras_ih_if ih_info = {
 517                         .cb = NULL,
 518                 };
 519
 520                 amdgpu_ras_late_fini(adev, ras_if, &ih_info);
 521                 kfree(ras_if);
 522         }
 523 }