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_ras.h"
  28 #include "soc15.h"
  29 #include "df/df_3_6_offset.h"
  30 #include "xgmi/xgmi_4_0_0_smn.h"
  31 #include "xgmi/xgmi_4_0_0_sh_mask.h"
  32 #include "wafl/wafl2_4_0_0_smn.h"
  33 #include "wafl/wafl2_4_0_0_sh_mask.h"
  34
  35 #include "amdgpu_reset.h"
  36
  37 #define smnPCS_XGMI3X16_PCS_ERROR_STATUS 0x11a0020c
  38 #define smnPCS_GOPX1_PCS_ERROR_STATUS    0x12200210
  39
  40 static DEFINE_MUTEX(xgmi_mutex);
  41
  42 #define AMDGPU_MAX_XGMI_DEVICE_PER_HIVE         4
  43
  44 static LIST_HEAD(xgmi_hive_list);
  45
  46 static const int xgmi_pcs_err_status_reg_vg20[] = {
  47         smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS,
  48         smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x100000,
  49 };
  50
  51 static const int wafl_pcs_err_status_reg_vg20[] = {
  52         smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS,
  53         smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS + 0x100000,
  54 };
  55
  56 static const int xgmi_pcs_err_status_reg_arct[] = {
  57         smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS,
  58         smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x100000,
  59         smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x500000,
  60         smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x600000,
  61         smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x700000,
  62         smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x800000,
  63 };
  64
  65 /* same as vg20*/
  66 static const int wafl_pcs_err_status_reg_arct[] = {
  67         smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS,
  68         smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS + 0x100000,
  69 };
  70
  71 static const int xgmi3x16_pcs_err_status_reg_aldebaran[] = {
  72         smnPCS_XGMI3X16_PCS_ERROR_STATUS,
  73         smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x100000,
  74         smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x200000,
  75         smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x300000,
  76         smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x400000,
  77         smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x500000,
  78         smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x600000,
  79         smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x700000
  80 };
  81
  82 static const int walf_pcs_err_status_reg_aldebaran[] = {
  83         smnPCS_GOPX1_PCS_ERROR_STATUS,
  84         smnPCS_GOPX1_PCS_ERROR_STATUS + 0x100000
  85 };
  86
  87 static const struct amdgpu_pcs_ras_field xgmi_pcs_ras_fields[] = {
  88         {"XGMI PCS DataLossErr",
  89          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataLossErr)},
  90         {"XGMI PCS TrainingErr",
  91          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, TrainingErr)},
  92         {"XGMI PCS CRCErr",
  93          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, CRCErr)},
  94         {"XGMI PCS BERExceededErr",
  95          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, BERExceededErr)},
  96         {"XGMI PCS TxMetaDataErr",
  97          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, TxMetaDataErr)},
  98         {"XGMI PCS ReplayBufParityErr",
  99          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayBufParityErr)},
 100         {"XGMI PCS DataParityErr",
 101          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataParityErr)},
 102         {"XGMI PCS ReplayFifoOverflowErr",
 103          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayFifoOverflowErr)},
 104         {"XGMI PCS ReplayFifoUnderflowErr",
 105          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayFifoUnderflowErr)},
 106         {"XGMI PCS ElasticFifoOverflowErr",
 107          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ElasticFifoOverflowErr)},
 108         {"XGMI PCS DeskewErr",
 109          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DeskewErr)},
 110         {"XGMI PCS DataStartupLimitErr",
 111          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataStartupLimitErr)},
 112         {"XGMI PCS FCInitTimeoutErr",
 113          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, FCInitTimeoutErr)},
 114         {"XGMI PCS RecoveryTimeoutErr",
 115          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryTimeoutErr)},
 116         {"XGMI PCS ReadySerialTimeoutErr",
 117          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReadySerialTimeoutErr)},
 118         {"XGMI PCS ReadySerialAttemptErr",
 119          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReadySerialAttemptErr)},
 120         {"XGMI PCS RecoveryAttemptErr",
 121          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryAttemptErr)},
 122         {"XGMI PCS RecoveryRelockAttemptErr",
 123          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryRelockAttemptErr)},
 124 };
 125
 126 static const struct amdgpu_pcs_ras_field wafl_pcs_ras_fields[] = {
 127         {"WAFL PCS DataLossErr",
 128          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataLossErr)},
 129         {"WAFL PCS TrainingErr",
 130          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, TrainingErr)},
 131         {"WAFL PCS CRCErr",
 132          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, CRCErr)},
 133         {"WAFL PCS BERExceededErr",
 134          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, BERExceededErr)},
 135         {"WAFL PCS TxMetaDataErr",
 136          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, TxMetaDataErr)},
 137         {"WAFL PCS ReplayBufParityErr",
 138          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayBufParityErr)},
 139         {"WAFL PCS DataParityErr",
 140          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataParityErr)},
 141         {"WAFL PCS ReplayFifoOverflowErr",
 142          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayFifoOverflowErr)},
 143         {"WAFL PCS ReplayFifoUnderflowErr",
 144          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayFifoUnderflowErr)},
 145         {"WAFL PCS ElasticFifoOverflowErr",
 146          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ElasticFifoOverflowErr)},
 147         {"WAFL PCS DeskewErr",
 148          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DeskewErr)},
 149         {"WAFL PCS DataStartupLimitErr",
 150          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataStartupLimitErr)},
 151         {"WAFL PCS FCInitTimeoutErr",
 152          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, FCInitTimeoutErr)},
 153         {"WAFL PCS RecoveryTimeoutErr",
 154          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryTimeoutErr)},
 155         {"WAFL PCS ReadySerialTimeoutErr",
 156          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReadySerialTimeoutErr)},
 157         {"WAFL PCS ReadySerialAttemptErr",
 158          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReadySerialAttemptErr)},
 159         {"WAFL PCS RecoveryAttemptErr",
 160          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryAttemptErr)},
 161         {"WAFL PCS RecoveryRelockAttemptErr",
 162          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryRelockAttemptErr)},
 163 };
 164
 165 /**
 166  * DOC: AMDGPU XGMI Support
 167  *
 168  * XGMI is a high speed interconnect that joins multiple GPU cards
 169  * into a homogeneous memory space that is organized by a collective
 170  * hive ID and individual node IDs, both of which are 64-bit numbers.
 171  *
 172  * The file xgmi_device_id contains the unique per GPU device ID and
 173  * is stored in the /sys/class/drm/card${cardno}/device/ directory.
 174  *
 175  * Inside the device directory a sub-directory 'xgmi_hive_info' is
 176  * created which contains the hive ID and the list of nodes.
 177  *
 178  * The hive ID is stored in:
 179  *   /sys/class/drm/card${cardno}/device/xgmi_hive_info/xgmi_hive_id
 180  *
 181  * The node information is stored in numbered directories:
 182  *   /sys/class/drm/card${cardno}/device/xgmi_hive_info/node${nodeno}/xgmi_device_id
 183  *
 184  * Each device has their own xgmi_hive_info direction with a mirror
 185  * set of node sub-directories.
 186  *
 187  * The XGMI memory space is built by contiguously adding the power of
 188  * two padded VRAM space from each node to each other.
 189  *
 190  */
 191
 192 static struct attribute amdgpu_xgmi_hive_id = {
 193         .name = "xgmi_hive_id",
 194         .mode = S_IRUGO
 195 };
 196
 197 static struct attribute *amdgpu_xgmi_hive_attrs[] = {
 198         &amdgpu_xgmi_hive_id,
 199         NULL
 200 };
 201 ATTRIBUTE_GROUPS(amdgpu_xgmi_hive);
 202
 203 static ssize_t amdgpu_xgmi_show_attrs(struct kobject *kobj,
 204         struct attribute *attr, char *buf)
 205 {
 206         struct amdgpu_hive_info *hive = container_of(
 207                 kobj, struct amdgpu_hive_info, kobj);
 208
 209         if (attr == &amdgpu_xgmi_hive_id)
 210                 return snprintf(buf, PAGE_SIZE, "%llu\n", hive->hive_id);
 211
 212         return 0;
 213 }
 214
 215 static void amdgpu_xgmi_hive_release(struct kobject *kobj)
 216 {
 217         struct amdgpu_hive_info *hive = container_of(
 218                 kobj, struct amdgpu_hive_info, kobj);
 219
 220         amdgpu_reset_put_reset_domain(hive->reset_domain);
 221         hive->reset_domain = NULL;
 222
 223         mutex_destroy(&hive->hive_lock);
 224         kfree(hive);
 225 }
 226
 227 static const struct sysfs_ops amdgpu_xgmi_hive_ops = {
 228         .show = amdgpu_xgmi_show_attrs,
 229 };
 230
 231 struct kobj_type amdgpu_xgmi_hive_type = {
 232         .release = amdgpu_xgmi_hive_release,
 233         .sysfs_ops = &amdgpu_xgmi_hive_ops,
 234         .default_groups = amdgpu_xgmi_hive_groups,
 235 };
 236
 237 static ssize_t amdgpu_xgmi_show_device_id(struct device *dev,
 238                                      struct device_attribute *attr,
 239                                      char *buf)
 240 {
 241         struct drm_device *ddev = dev_get_drvdata(dev);
 242         struct amdgpu_device *adev = drm_to_adev(ddev);
 243
 244         return sysfs_emit(buf, "%llu\n", adev->gmc.xgmi.node_id);
 245
 246 }
 247
 248 #define AMDGPU_XGMI_SET_FICAA(o)        ((o) | 0x456801)
 249 static ssize_t amdgpu_xgmi_show_error(struct device *dev,
 250                                       struct device_attribute *attr,
 251                                       char *buf)
 252 {
 253         struct drm_device *ddev = dev_get_drvdata(dev);
 254         struct amdgpu_device *adev = drm_to_adev(ddev);
 255         uint32_t ficaa_pie_ctl_in, ficaa_pie_status_in;
 256         uint64_t fica_out;
 257         unsigned int error_count = 0;
 258
 259         ficaa_pie_ctl_in = AMDGPU_XGMI_SET_FICAA(0x200);
 260         ficaa_pie_status_in = AMDGPU_XGMI_SET_FICAA(0x208);
 261
 262         if ((!adev->df.funcs) ||
 263             (!adev->df.funcs->get_fica) ||
 264             (!adev->df.funcs->set_fica))
 265                 return -EINVAL;
 266
 267         fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_ctl_in);
 268         if (fica_out != 0x1f)
 269                 pr_err("xGMI error counters not enabled!\n");
 270
 271         fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_status_in);
 272
 273         if ((fica_out & 0xffff) == 2)
 274                 error_count = ((fica_out >> 62) & 0x1) + (fica_out >> 63);
 275
 276         adev->df.funcs->set_fica(adev, ficaa_pie_status_in, 0, 0);
 277
 278         return sysfs_emit(buf, "%u\n", error_count);
 279 }
 280
 281
 282 static DEVICE_ATTR(xgmi_device_id, S_IRUGO, amdgpu_xgmi_show_device_id, NULL);
 283 static DEVICE_ATTR(xgmi_error, S_IRUGO, amdgpu_xgmi_show_error, NULL);
 284
 285 static int amdgpu_xgmi_sysfs_add_dev_info(struct amdgpu_device *adev,
 286                                          struct amdgpu_hive_info *hive)
 287 {
 288         int ret = 0;
 289         char node[10] = { 0 };
 290
 291         /* Create xgmi device id file */
 292         ret = device_create_file(adev->dev, &dev_attr_xgmi_device_id);
 293         if (ret) {
 294                 dev_err(adev->dev, "XGMI: Failed to create device file xgmi_device_id\n");
 295                 return ret;
 296         }
 297
 298         /* Create xgmi error file */
 299         ret = device_create_file(adev->dev, &dev_attr_xgmi_error);
 300         if (ret)
 301                 pr_err("failed to create xgmi_error\n");
 302
 303
 304         /* Create sysfs link to hive info folder on the first device */
 305         if (hive->kobj.parent != (&adev->dev->kobj)) {
 306                 ret = sysfs_create_link(&adev->dev->kobj, &hive->kobj,
 307                                         "xgmi_hive_info");
 308                 if (ret) {
 309                         dev_err(adev->dev, "XGMI: Failed to create link to hive info");
 310                         goto remove_file;
 311                 }
 312         }
 313
 314         sprintf(node, "node%d", atomic_read(&hive->number_devices));
 315         /* Create sysfs link form the hive folder to yourself */
 316         ret = sysfs_create_link(&hive->kobj, &adev->dev->kobj, node);
 317         if (ret) {
 318                 dev_err(adev->dev, "XGMI: Failed to create link from hive info");
 319                 goto remove_link;
 320         }
 321
 322         goto success;
 323
 324
 325 remove_link:
 326         sysfs_remove_link(&adev->dev->kobj, adev_to_drm(adev)->unique);
 327
 328 remove_file:
 329         device_remove_file(adev->dev, &dev_attr_xgmi_device_id);
 330
 331 success:
 332         return ret;
 333 }
 334
 335 static void amdgpu_xgmi_sysfs_rem_dev_info(struct amdgpu_device *adev,
 336                                           struct amdgpu_hive_info *hive)
 337 {
 338         char node[10];
 339         memset(node, 0, sizeof(node));
 340
 341         device_remove_file(adev->dev, &dev_attr_xgmi_device_id);
 342         device_remove_file(adev->dev, &dev_attr_xgmi_error);
 343
 344         if (hive->kobj.parent != (&adev->dev->kobj))
 345                 sysfs_remove_link(&adev->dev->kobj,"xgmi_hive_info");
 346
 347         sprintf(node, "node%d", atomic_read(&hive->number_devices));
 348         sysfs_remove_link(&hive->kobj, node);
 349
 350 }
 351
 352
 353
 354 struct amdgpu_hive_info *amdgpu_get_xgmi_hive(struct amdgpu_device *adev)
 355 {
 356         struct amdgpu_hive_info *hive = NULL;
 357         int ret;
 358
 359         if (!adev->gmc.xgmi.hive_id)
 360                 return NULL;
 361
 362         if (adev->hive) {
 363                 kobject_get(&adev->hive->kobj);
 364                 return adev->hive;
 365         }
 366
 367         mutex_lock(&xgmi_mutex);
 368
 369         list_for_each_entry(hive, &xgmi_hive_list, node)  {
 370                 if (hive->hive_id == adev->gmc.xgmi.hive_id)
 371                         goto pro_end;
 372         }
 373
 374         hive = kzalloc(sizeof(*hive), GFP_KERNEL);
 375         if (!hive) {
 376                 dev_err(adev->dev, "XGMI: allocation failed\n");
 377                 hive = NULL;
 378                 goto pro_end;
 379         }
 380
 381         /* initialize new hive if not exist */
 382         ret = kobject_init_and_add(&hive->kobj,
 383                         &amdgpu_xgmi_hive_type,
 384                         &adev->dev->kobj,
 385                         "%s", "xgmi_hive_info");
 386         if (ret) {
 387                 dev_err(adev->dev, "XGMI: failed initializing kobject for xgmi hive\n");
 388                 kobject_put(&hive->kobj);
 389                 kfree(hive);
 390                 hive = NULL;
 391                 goto pro_end;
 392         }
 393
 394         /**
 395          * Avoid recreating reset domain when hive is reconstructed for the case
 396          * of reset the devices in the XGMI hive during probe for SRIOV
 397          * See https://www.spinics.net/lists/amd-gfx/msg58836.html
 398          */
 399         if (adev->reset_domain->type != XGMI_HIVE) {
 400                 hive->reset_domain = amdgpu_reset_create_reset_domain(XGMI_HIVE, "amdgpu-reset-hive");
 401                         if (!hive->reset_domain) {
 402                                 dev_err(adev->dev, "XGMI: failed initializing reset domain for xgmi hive\n");
 403                                 ret = -ENOMEM;
 404                                 kobject_put(&hive->kobj);
 405                                 kfree(hive);
 406                                 hive = NULL;
 407                                 goto pro_end;
 408                         }
 409         } else {
 410                 amdgpu_reset_get_reset_domain(adev->reset_domain);
 411                 hive->reset_domain = adev->reset_domain;
 412         }
 413
 414         hive->hive_id = adev->gmc.xgmi.hive_id;
 415         INIT_LIST_HEAD(&hive->device_list);
 416         INIT_LIST_HEAD(&hive->node);
 417         mutex_init(&hive->hive_lock);
 418         atomic_set(&hive->number_devices, 0);
 419         task_barrier_init(&hive->tb);
 420         hive->pstate = AMDGPU_XGMI_PSTATE_UNKNOWN;
 421         hive->hi_req_gpu = NULL;
 422
 423         /*
 424          * hive pstate on boot is high in vega20 so we have to go to low
 425          * pstate on after boot.
 426          */
 427         hive->hi_req_count = AMDGPU_MAX_XGMI_DEVICE_PER_HIVE;
 428         list_add_tail(&hive->node, &xgmi_hive_list);
 429
 430 pro_end:
 431         if (hive)
 432                 kobject_get(&hive->kobj);
 433         mutex_unlock(&xgmi_mutex);
 434         return hive;
 435 }
 436
 437 void amdgpu_put_xgmi_hive(struct amdgpu_hive_info *hive)
 438 {
 439         if (hive)
 440                 kobject_put(&hive->kobj);
 441 }
 442
 443 int amdgpu_xgmi_set_pstate(struct amdgpu_device *adev, int pstate)
 444 {
 445         int ret = 0;
 446         struct amdgpu_hive_info *hive;
 447         struct amdgpu_device *request_adev;
 448         bool is_hi_req = pstate == AMDGPU_XGMI_PSTATE_MAX_VEGA20;
 449         bool init_low;
 450
 451         hive = amdgpu_get_xgmi_hive(adev);
 452         if (!hive)
 453                 return 0;
 454
 455         request_adev = hive->hi_req_gpu ? hive->hi_req_gpu : adev;
 456         init_low = hive->pstate == AMDGPU_XGMI_PSTATE_UNKNOWN;
 457         amdgpu_put_xgmi_hive(hive);
 458         /* fw bug so temporarily disable pstate switching */
 459         return 0;
 460
 461         if (!hive || adev->asic_type != CHIP_VEGA20)
 462                 return 0;
 463
 464         mutex_lock(&hive->hive_lock);
 465
 466         if (is_hi_req)
 467                 hive->hi_req_count++;
 468         else
 469                 hive->hi_req_count--;
 470
 471         /*
 472          * Vega20 only needs single peer to request pstate high for the hive to
 473          * go high but all peers must request pstate low for the hive to go low
 474          */
 475         if (hive->pstate == pstate ||
 476                         (!is_hi_req && hive->hi_req_count && !init_low))
 477                 goto out;
 478
 479         dev_dbg(request_adev->dev, "Set xgmi pstate %d.\n", pstate);
 480
 481         ret = amdgpu_dpm_set_xgmi_pstate(request_adev, pstate);
 482         if (ret) {
 483                 dev_err(request_adev->dev,
 484                         "XGMI: Set pstate failure on device %llx, hive %llx, ret %d",
 485                         request_adev->gmc.xgmi.node_id,
 486                         request_adev->gmc.xgmi.hive_id, ret);
 487                 goto out;
 488         }
 489
 490         if (init_low)
 491                 hive->pstate = hive->hi_req_count ?
 492                                         hive->pstate : AMDGPU_XGMI_PSTATE_MIN;
 493         else {
 494                 hive->pstate = pstate;
 495                 hive->hi_req_gpu = pstate != AMDGPU_XGMI_PSTATE_MIN ?
 496                                                         adev : NULL;
 497         }
 498 out:
 499         mutex_unlock(&hive->hive_lock);
 500         return ret;
 501 }
 502
 503 int amdgpu_xgmi_update_topology(struct amdgpu_hive_info *hive, struct amdgpu_device *adev)
 504 {
 505         int ret;
 506
 507         /* Each psp need to set the latest topology */
 508         ret = psp_xgmi_set_topology_info(&adev->psp,
 509                                          atomic_read(&hive->number_devices),
 510                                          &adev->psp.xgmi_context.top_info);
 511         if (ret)
 512                 dev_err(adev->dev,
 513                         "XGMI: Set topology failure on device %llx, hive %llx, ret %d",
 514                         adev->gmc.xgmi.node_id,
 515                         adev->gmc.xgmi.hive_id, ret);
 516
 517         return ret;
 518 }
 519
 520
 521 /*
 522  * NOTE psp_xgmi_node_info.num_hops layout is as follows:
 523  * num_hops[7:6] = link type (0 = xGMI2, 1 = xGMI3, 2/3 = reserved)
 524  * num_hops[5:3] = reserved
 525  * num_hops[2:0] = number of hops
 526  */
 527 int amdgpu_xgmi_get_hops_count(struct amdgpu_device *adev,
 528                 struct amdgpu_device *peer_adev)
 529 {
 530         struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info;
 531         uint8_t num_hops_mask = 0x7;
 532         int i;
 533
 534         for (i = 0 ; i < top->num_nodes; ++i)
 535                 if (top->nodes[i].node_id == peer_adev->gmc.xgmi.node_id)
 536                         return top->nodes[i].num_hops & num_hops_mask;
 537         return  -EINVAL;
 538 }
 539
 540 int amdgpu_xgmi_get_num_links(struct amdgpu_device *adev,
 541                 struct amdgpu_device *peer_adev)
 542 {
 543         struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info;
 544         int i;
 545
 546         for (i = 0 ; i < top->num_nodes; ++i)
 547                 if (top->nodes[i].node_id == peer_adev->gmc.xgmi.node_id)
 548                         return top->nodes[i].num_links;
 549         return  -EINVAL;
 550 }
 551
 552 /*
 553  * Devices that support extended data require the entire hive to initialize with
 554  * the shared memory buffer flag set.
 555  *
 556  * Hive locks and conditions apply - see amdgpu_xgmi_add_device
 557  */
 558 static int amdgpu_xgmi_initialize_hive_get_data_partition(struct amdgpu_hive_info *hive,
 559                                                         bool set_extended_data)
 560 {
 561         struct amdgpu_device *tmp_adev;
 562         int ret;
 563
 564         list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
 565                 ret = psp_xgmi_initialize(&tmp_adev->psp, set_extended_data, false);
 566                 if (ret) {
 567                         dev_err(tmp_adev->dev,
 568                                 "XGMI: Failed to initialize xgmi session for data partition %i\n",
 569                                 set_extended_data);
 570                         return ret;
 571                 }
 572
 573         }
 574
 575         return 0;
 576 }
 577
 578 int amdgpu_xgmi_add_device(struct amdgpu_device *adev)
 579 {
 580         struct psp_xgmi_topology_info *top_info;
 581         struct amdgpu_hive_info *hive;
 582         struct amdgpu_xgmi      *entry;
 583         struct amdgpu_device *tmp_adev = NULL;
 584
 585         int count = 0, ret = 0;
 586
 587         if (!adev->gmc.xgmi.supported)
 588                 return 0;
 589
 590         if (!adev->gmc.xgmi.pending_reset &&
 591             amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) {
 592                 ret = psp_xgmi_initialize(&adev->psp, false, true);
 593                 if (ret) {
 594                         dev_err(adev->dev,
 595                                 "XGMI: Failed to initialize xgmi session\n");
 596                         return ret;
 597                 }
 598
 599                 ret = psp_xgmi_get_hive_id(&adev->psp, &adev->gmc.xgmi.hive_id);
 600                 if (ret) {
 601                         dev_err(adev->dev,
 602                                 "XGMI: Failed to get hive id\n");
 603                         return ret;
 604                 }
 605
 606                 ret = psp_xgmi_get_node_id(&adev->psp, &adev->gmc.xgmi.node_id);
 607                 if (ret) {
 608                         dev_err(adev->dev,
 609                                 "XGMI: Failed to get node id\n");
 610                         return ret;
 611                 }
 612         } else {
 613                 adev->gmc.xgmi.hive_id = 16;
 614                 adev->gmc.xgmi.node_id = adev->gmc.xgmi.physical_node_id + 16;
 615         }
 616
 617         hive = amdgpu_get_xgmi_hive(adev);
 618         if (!hive) {
 619                 ret = -EINVAL;
 620                 dev_err(adev->dev,
 621                         "XGMI: node 0x%llx, can not match hive 0x%llx in the hive list.\n",
 622                         adev->gmc.xgmi.node_id, adev->gmc.xgmi.hive_id);
 623                 goto exit;
 624         }
 625         mutex_lock(&hive->hive_lock);
 626
 627         top_info = &adev->psp.xgmi_context.top_info;
 628
 629         list_add_tail(&adev->gmc.xgmi.head, &hive->device_list);
 630         list_for_each_entry(entry, &hive->device_list, head)
 631                 top_info->nodes[count++].node_id = entry->node_id;
 632         top_info->num_nodes = count;
 633         atomic_set(&hive->number_devices, count);
 634
 635         task_barrier_add_task(&hive->tb);
 636
 637         if (!adev->gmc.xgmi.pending_reset &&
 638             amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) {
 639                 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
 640                         /* update node list for other device in the hive */
 641                         if (tmp_adev != adev) {
 642                                 top_info = &tmp_adev->psp.xgmi_context.top_info;
 643                                 top_info->nodes[count - 1].node_id =
 644                                         adev->gmc.xgmi.node_id;
 645                                 top_info->num_nodes = count;
 646                         }
 647                         ret = amdgpu_xgmi_update_topology(hive, tmp_adev);
 648                         if (ret)
 649                                 goto exit_unlock;
 650                 }
 651
 652                 /* get latest topology info for each device from psp */
 653                 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
 654                         ret = psp_xgmi_get_topology_info(&tmp_adev->psp, count,
 655                                         &tmp_adev->psp.xgmi_context.top_info, false);
 656                         if (ret) {
 657                                 dev_err(tmp_adev->dev,
 658                                         "XGMI: Get topology failure on device %llx, hive %llx, ret %d",
 659                                         tmp_adev->gmc.xgmi.node_id,
 660                                         tmp_adev->gmc.xgmi.hive_id, ret);
 661                                 /* To do : continue with some node failed or disable the whole hive */
 662                                 goto exit_unlock;
 663                         }
 664                 }
 665
 666                 /* get topology again for hives that support extended data */
 667                 if (adev->psp.xgmi_context.supports_extended_data) {
 668
 669                         /* initialize the hive to get extended data.  */
 670                         ret = amdgpu_xgmi_initialize_hive_get_data_partition(hive, true);
 671                         if (ret)
 672                                 goto exit_unlock;
 673
 674                         /* get the extended data. */
 675                         list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
 676                                 ret = psp_xgmi_get_topology_info(&tmp_adev->psp, count,
 677                                                 &tmp_adev->psp.xgmi_context.top_info, true);
 678                                 if (ret) {
 679                                         dev_err(tmp_adev->dev,
 680                                                 "XGMI: Get topology for extended data failure on device %llx, hive %llx, ret %d",
 681                                                 tmp_adev->gmc.xgmi.node_id,
 682                                                 tmp_adev->gmc.xgmi.hive_id, ret);
 683                                         goto exit_unlock;
 684                                 }
 685                         }
 686
 687                         /* initialize the hive to get non-extended data for the next round. */
 688                         ret = amdgpu_xgmi_initialize_hive_get_data_partition(hive, false);
 689                         if (ret)
 690                                 goto exit_unlock;
 691
 692                 }
 693         }
 694
 695         if (!ret && !adev->gmc.xgmi.pending_reset)
 696                 ret = amdgpu_xgmi_sysfs_add_dev_info(adev, hive);
 697
 698 exit_unlock:
 699         mutex_unlock(&hive->hive_lock);
 700 exit:
 701         if (!ret) {
 702                 adev->hive = hive;
 703                 dev_info(adev->dev, "XGMI: Add node %d, hive 0x%llx.\n",
 704                          adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id);
 705         } else {
 706                 amdgpu_put_xgmi_hive(hive);
 707                 dev_err(adev->dev, "XGMI: Failed to add node %d, hive 0x%llx ret: %d\n",
 708                         adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id,
 709                         ret);
 710         }
 711
 712         return ret;
 713 }
 714
 715 int amdgpu_xgmi_remove_device(struct amdgpu_device *adev)
 716 {
 717         struct amdgpu_hive_info *hive = adev->hive;
 718
 719         if (!adev->gmc.xgmi.supported)
 720                 return -EINVAL;
 721
 722         if (!hive)
 723                 return -EINVAL;
 724
 725         mutex_lock(&hive->hive_lock);
 726         task_barrier_rem_task(&hive->tb);
 727         amdgpu_xgmi_sysfs_rem_dev_info(adev, hive);
 728         if (hive->hi_req_gpu == adev)
 729                 hive->hi_req_gpu = NULL;
 730         list_del(&adev->gmc.xgmi.head);
 731         mutex_unlock(&hive->hive_lock);
 732
 733         amdgpu_put_xgmi_hive(hive);
 734         adev->hive = NULL;
 735
 736         if (atomic_dec_return(&hive->number_devices) == 0) {
 737                 /* Remove the hive from global hive list */
 738                 mutex_lock(&xgmi_mutex);
 739                 list_del(&hive->node);
 740                 mutex_unlock(&xgmi_mutex);
 741
 742                 amdgpu_put_xgmi_hive(hive);
 743         }
 744
 745         return psp_xgmi_terminate(&adev->psp);
 746 }
 747
 748 static int amdgpu_xgmi_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *ras_block)
 749 {
 750         if (!adev->gmc.xgmi.supported ||
 751             adev->gmc.xgmi.num_physical_nodes == 0)
 752                 return 0;
 753
 754         adev->gmc.xgmi.ras->ras_block.hw_ops->reset_ras_error_count(adev);
 755
 756         return amdgpu_ras_block_late_init(adev, ras_block);
 757 }
 758
 759 uint64_t amdgpu_xgmi_get_relative_phy_addr(struct amdgpu_device *adev,
 760                                            uint64_t addr)
 761 {
 762         struct amdgpu_xgmi *xgmi = &adev->gmc.xgmi;
 763         return (addr + xgmi->physical_node_id * xgmi->node_segment_size);
 764 }
 765
 766 static void pcs_clear_status(struct amdgpu_device *adev, uint32_t pcs_status_reg)
 767 {
 768         WREG32_PCIE(pcs_status_reg, 0xFFFFFFFF);
 769         WREG32_PCIE(pcs_status_reg, 0);
 770 }
 771
 772 static void amdgpu_xgmi_reset_ras_error_count(struct amdgpu_device *adev)
 773 {
 774         uint32_t i;
 775
 776         switch (adev->asic_type) {
 777         case CHIP_ARCTURUS:
 778                 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_arct); i++)
 779                         pcs_clear_status(adev,
 780                                          xgmi_pcs_err_status_reg_arct[i]);
 781                 break;
 782         case CHIP_VEGA20:
 783                 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_vg20); i++)
 784                         pcs_clear_status(adev,
 785                                          xgmi_pcs_err_status_reg_vg20[i]);
 786                 break;
 787         case CHIP_ALDEBARAN:
 788                 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_aldebaran); i++)
 789                         pcs_clear_status(adev,
 790                                          xgmi3x16_pcs_err_status_reg_aldebaran[i]);
 791                 for (i = 0; i < ARRAY_SIZE(walf_pcs_err_status_reg_aldebaran); i++)
 792                         pcs_clear_status(adev,
 793                                          walf_pcs_err_status_reg_aldebaran[i]);
 794                 break;
 795         default:
 796                 break;
 797         }
 798 }
 799
 800 static int amdgpu_xgmi_query_pcs_error_status(struct amdgpu_device *adev,
 801                                               uint32_t value,
 802                                               uint32_t *ue_count,
 803                                               uint32_t *ce_count,
 804                                               bool is_xgmi_pcs)
 805 {
 806         int i;
 807         int ue_cnt;
 808
 809         if (is_xgmi_pcs) {
 810                 /* query xgmi pcs error status,
 811                  * only ue is supported */
 812                 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_ras_fields); i ++) {
 813                         ue_cnt = (value &
 814                                   xgmi_pcs_ras_fields[i].pcs_err_mask) >>
 815                                   xgmi_pcs_ras_fields[i].pcs_err_shift;
 816                         if (ue_cnt) {
 817                                 dev_info(adev->dev, "%s detected\n",
 818                                          xgmi_pcs_ras_fields[i].err_name);
 819                                 *ue_count += ue_cnt;
 820                         }
 821                 }
 822         } else {
 823                 /* query wafl pcs error status,
 824                  * only ue is supported */
 825                 for (i = 0; i < ARRAY_SIZE(wafl_pcs_ras_fields); i++) {
 826                         ue_cnt = (value &
 827                                   wafl_pcs_ras_fields[i].pcs_err_mask) >>
 828                                   wafl_pcs_ras_fields[i].pcs_err_shift;
 829                         if (ue_cnt) {
 830                                 dev_info(adev->dev, "%s detected\n",
 831                                          wafl_pcs_ras_fields[i].err_name);
 832                                 *ue_count += ue_cnt;
 833                         }
 834                 }
 835         }
 836
 837         return 0;
 838 }
 839
 840 static void amdgpu_xgmi_query_ras_error_count(struct amdgpu_device *adev,
 841                                              void *ras_error_status)
 842 {
 843         struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status;
 844         int i;
 845         uint32_t data;
 846         uint32_t ue_cnt = 0, ce_cnt = 0;
 847
 848         if (!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL))
 849                 return ;
 850
 851         err_data->ue_count = 0;
 852         err_data->ce_count = 0;
 853
 854         switch (adev->asic_type) {
 855         case CHIP_ARCTURUS:
 856                 /* check xgmi pcs error */
 857                 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_arct); i++) {
 858                         data = RREG32_PCIE(xgmi_pcs_err_status_reg_arct[i]);
 859                         if (data)
 860                                 amdgpu_xgmi_query_pcs_error_status(adev,
 861                                                 data, &ue_cnt, &ce_cnt, true);
 862                 }
 863                 /* check wafl pcs error */
 864                 for (i = 0; i < ARRAY_SIZE(wafl_pcs_err_status_reg_arct); i++) {
 865                         data = RREG32_PCIE(wafl_pcs_err_status_reg_arct[i]);
 866                         if (data)
 867                                 amdgpu_xgmi_query_pcs_error_status(adev,
 868                                                 data, &ue_cnt, &ce_cnt, false);
 869                 }
 870                 break;
 871         case CHIP_VEGA20:
 872                 /* check xgmi pcs error */
 873                 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_vg20); i++) {
 874                         data = RREG32_PCIE(xgmi_pcs_err_status_reg_vg20[i]);
 875                         if (data)
 876                                 amdgpu_xgmi_query_pcs_error_status(adev,
 877                                                 data, &ue_cnt, &ce_cnt, true);
 878                 }
 879                 /* check wafl pcs error */
 880                 for (i = 0; i < ARRAY_SIZE(wafl_pcs_err_status_reg_vg20); i++) {
 881                         data = RREG32_PCIE(wafl_pcs_err_status_reg_vg20[i]);
 882                         if (data)
 883                                 amdgpu_xgmi_query_pcs_error_status(adev,
 884                                                 data, &ue_cnt, &ce_cnt, false);
 885                 }
 886                 break;
 887         case CHIP_ALDEBARAN:
 888                 /* check xgmi3x16 pcs error */
 889                 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_aldebaran); i++) {
 890                         data = RREG32_PCIE(xgmi3x16_pcs_err_status_reg_aldebaran[i]);
 891                         if (data)
 892                                 amdgpu_xgmi_query_pcs_error_status(adev,
 893                                                 data, &ue_cnt, &ce_cnt, true);
 894                 }
 895                 /* check wafl pcs error */
 896                 for (i = 0; i < ARRAY_SIZE(walf_pcs_err_status_reg_aldebaran); i++) {
 897                         data = RREG32_PCIE(walf_pcs_err_status_reg_aldebaran[i]);
 898                         if (data)
 899                                 amdgpu_xgmi_query_pcs_error_status(adev,
 900                                                 data, &ue_cnt, &ce_cnt, false);
 901                 }
 902                 break;
 903         default:
 904                 dev_warn(adev->dev, "XGMI RAS error query not supported");
 905                 break;
 906         }
 907
 908         adev->gmc.xgmi.ras->ras_block.hw_ops->reset_ras_error_count(adev);
 909
 910         err_data->ue_count += ue_cnt;
 911         err_data->ce_count += ce_cnt;
 912 }
 913
 914 /* Trigger XGMI/WAFL error */
 915 static int amdgpu_ras_error_inject_xgmi(struct amdgpu_device *adev,  void *inject_if)
 916 {
 917         int ret = 0;
 918         struct ta_ras_trigger_error_input *block_info =
 919                                 (struct ta_ras_trigger_error_input *)inject_if;
 920
 921         if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_DISALLOW))
 922                 dev_warn(adev->dev, "Failed to disallow df cstate");
 923
 924         if (amdgpu_dpm_allow_xgmi_power_down(adev, false))
 925                 dev_warn(adev->dev, "Failed to disallow XGMI power down");
 926
 927         ret = psp_ras_trigger_error(&adev->psp, block_info);
 928
 929         if (amdgpu_ras_intr_triggered())
 930                 return ret;
 931
 932         if (amdgpu_dpm_allow_xgmi_power_down(adev, true))
 933                 dev_warn(adev->dev, "Failed to allow XGMI power down");
 934
 935         if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_ALLOW))
 936                 dev_warn(adev->dev, "Failed to allow df cstate");
 937
 938         return ret;
 939 }
 940
 941 struct amdgpu_ras_block_hw_ops  xgmi_ras_hw_ops = {
 942         .query_ras_error_count = amdgpu_xgmi_query_ras_error_count,
 943         .reset_ras_error_count = amdgpu_xgmi_reset_ras_error_count,
 944         .ras_error_inject = amdgpu_ras_error_inject_xgmi,
 945 };
 946
 947 struct amdgpu_xgmi_ras xgmi_ras = {
 948         .ras_block = {
 949                 .ras_comm = {
 950                         .name = "xgmi_wafl",
 951                         .block = AMDGPU_RAS_BLOCK__XGMI_WAFL,
 952                         .type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE,
 953                 },
 954                 .hw_ops = &xgmi_ras_hw_ops,
 955                 .ras_late_init = amdgpu_xgmi_ras_late_init,
 956         },
 957 };