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 "xgmi/xgmi_6_1_0_sh_mask.h"
  33 #include "wafl/wafl2_4_0_0_smn.h"
  34 #include "wafl/wafl2_4_0_0_sh_mask.h"
  35
  36 #include "amdgpu_reset.h"
  37
  38 #define smnPCS_XGMI3X16_PCS_ERROR_STATUS 0x11a0020c
  39 #define smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK   0x11a00218
  40 #define smnPCS_GOPX1_PCS_ERROR_STATUS    0x12200210
  41 #define smnPCS_GOPX1_PCS_ERROR_NONCORRECTABLE_MASK      0x12200218
  42
  43 static DEFINE_MUTEX(xgmi_mutex);
  44
  45 #define AMDGPU_MAX_XGMI_DEVICE_PER_HIVE         4
  46
  47 static LIST_HEAD(xgmi_hive_list);
  48
  49 static const int xgmi_pcs_err_status_reg_vg20[] = {
  50         smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS,
  51         smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x100000,
  52 };
  53
  54 static const int wafl_pcs_err_status_reg_vg20[] = {
  55         smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS,
  56         smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS + 0x100000,
  57 };
  58
  59 static const int xgmi_pcs_err_status_reg_arct[] = {
  60         smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS,
  61         smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x100000,
  62         smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x500000,
  63         smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x600000,
  64         smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x700000,
  65         smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x800000,
  66 };
  67
  68 /* same as vg20*/
  69 static const int wafl_pcs_err_status_reg_arct[] = {
  70         smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS,
  71         smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS + 0x100000,
  72 };
  73
  74 static const int xgmi3x16_pcs_err_status_reg_aldebaran[] = {
  75         smnPCS_XGMI3X16_PCS_ERROR_STATUS,
  76         smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x100000,
  77         smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x200000,
  78         smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x300000,
  79         smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x400000,
  80         smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x500000,
  81         smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x600000,
  82         smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x700000
  83 };
  84
  85 static const int xgmi3x16_pcs_err_noncorrectable_mask_reg_aldebaran[] = {
  86         smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK,
  87         smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK + 0x100000,
  88         smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK + 0x200000,
  89         smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK + 0x300000,
  90         smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK + 0x400000,
  91         smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK + 0x500000,
  92         smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK + 0x600000,
  93         smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK + 0x700000
  94 };
  95
  96 static const int walf_pcs_err_status_reg_aldebaran[] = {
  97         smnPCS_GOPX1_PCS_ERROR_STATUS,
  98         smnPCS_GOPX1_PCS_ERROR_STATUS + 0x100000
  99 };
 100
 101 static const int walf_pcs_err_noncorrectable_mask_reg_aldebaran[] = {
 102         smnPCS_GOPX1_PCS_ERROR_NONCORRECTABLE_MASK,
 103         smnPCS_GOPX1_PCS_ERROR_NONCORRECTABLE_MASK + 0x100000
 104 };
 105
 106 static const int xgmi3x16_pcs_err_status_reg_v6_4[] = {
 107         smnPCS_XGMI3X16_PCS_ERROR_STATUS,
 108         smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x100000
 109 };
 110
 111 static const int xgmi3x16_pcs_err_noncorrectable_mask_reg_v6_4[] = {
 112         smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK,
 113         smnPCS_XGMI3X16_PCS_ERROR_NONCORRECTABLE_MASK + 0x100000
 114 };
 115
 116 static const u64 xgmi_v6_4_0_mca_base_array[] = {
 117         0x11a09200,
 118         0x11b09200,
 119 };
 120
 121 static const char *xgmi_v6_4_0_ras_error_code_ext[32] = {
 122         [0x00] = "XGMI PCS DataLossErr",
 123         [0x01] = "XGMI PCS TrainingErr",
 124         [0x02] = "XGMI PCS FlowCtrlAckErr",
 125         [0x03] = "XGMI PCS RxFifoUnderflowErr",
 126         [0x04] = "XGMI PCS RxFifoOverflowErr",
 127         [0x05] = "XGMI PCS CRCErr",
 128         [0x06] = "XGMI PCS BERExceededErr",
 129         [0x07] = "XGMI PCS TxMetaDataErr",
 130         [0x08] = "XGMI PCS ReplayBufParityErr",
 131         [0x09] = "XGMI PCS DataParityErr",
 132         [0x0a] = "XGMI PCS ReplayFifoOverflowErr",
 133         [0x0b] = "XGMI PCS ReplayFifoUnderflowErr",
 134         [0x0c] = "XGMI PCS ElasticFifoOverflowErr",
 135         [0x0d] = "XGMI PCS DeskewErr",
 136         [0x0e] = "XGMI PCS FlowCtrlCRCErr",
 137         [0x0f] = "XGMI PCS DataStartupLimitErr",
 138         [0x10] = "XGMI PCS FCInitTimeoutErr",
 139         [0x11] = "XGMI PCS RecoveryTimeoutErr",
 140         [0x12] = "XGMI PCS ReadySerialTimeoutErr",
 141         [0x13] = "XGMI PCS ReadySerialAttemptErr",
 142         [0x14] = "XGMI PCS RecoveryAttemptErr",
 143         [0x15] = "XGMI PCS RecoveryRelockAttemptErr",
 144         [0x16] = "XGMI PCS ReplayAttemptErr",
 145         [0x17] = "XGMI PCS SyncHdrErr",
 146         [0x18] = "XGMI PCS TxReplayTimeoutErr",
 147         [0x19] = "XGMI PCS RxReplayTimeoutErr",
 148         [0x1a] = "XGMI PCS LinkSubTxTimeoutErr",
 149         [0x1b] = "XGMI PCS LinkSubRxTimeoutErr",
 150         [0x1c] = "XGMI PCS RxCMDPktErr",
 151 };
 152
 153 static const struct amdgpu_pcs_ras_field xgmi_pcs_ras_fields[] = {
 154         {"XGMI PCS DataLossErr",
 155          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataLossErr)},
 156         {"XGMI PCS TrainingErr",
 157          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, TrainingErr)},
 158         {"XGMI PCS CRCErr",
 159          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, CRCErr)},
 160         {"XGMI PCS BERExceededErr",
 161          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, BERExceededErr)},
 162         {"XGMI PCS TxMetaDataErr",
 163          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, TxMetaDataErr)},
 164         {"XGMI PCS ReplayBufParityErr",
 165          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayBufParityErr)},
 166         {"XGMI PCS DataParityErr",
 167          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataParityErr)},
 168         {"XGMI PCS ReplayFifoOverflowErr",
 169          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayFifoOverflowErr)},
 170         {"XGMI PCS ReplayFifoUnderflowErr",
 171          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayFifoUnderflowErr)},
 172         {"XGMI PCS ElasticFifoOverflowErr",
 173          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ElasticFifoOverflowErr)},
 174         {"XGMI PCS DeskewErr",
 175          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DeskewErr)},
 176         {"XGMI PCS DataStartupLimitErr",
 177          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataStartupLimitErr)},
 178         {"XGMI PCS FCInitTimeoutErr",
 179          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, FCInitTimeoutErr)},
 180         {"XGMI PCS RecoveryTimeoutErr",
 181          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryTimeoutErr)},
 182         {"XGMI PCS ReadySerialTimeoutErr",
 183          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReadySerialTimeoutErr)},
 184         {"XGMI PCS ReadySerialAttemptErr",
 185          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReadySerialAttemptErr)},
 186         {"XGMI PCS RecoveryAttemptErr",
 187          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryAttemptErr)},
 188         {"XGMI PCS RecoveryRelockAttemptErr",
 189          SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryRelockAttemptErr)},
 190 };
 191
 192 static const struct amdgpu_pcs_ras_field wafl_pcs_ras_fields[] = {
 193         {"WAFL PCS DataLossErr",
 194          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataLossErr)},
 195         {"WAFL PCS TrainingErr",
 196          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, TrainingErr)},
 197         {"WAFL PCS CRCErr",
 198          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, CRCErr)},
 199         {"WAFL PCS BERExceededErr",
 200          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, BERExceededErr)},
 201         {"WAFL PCS TxMetaDataErr",
 202          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, TxMetaDataErr)},
 203         {"WAFL PCS ReplayBufParityErr",
 204          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayBufParityErr)},
 205         {"WAFL PCS DataParityErr",
 206          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataParityErr)},
 207         {"WAFL PCS ReplayFifoOverflowErr",
 208          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayFifoOverflowErr)},
 209         {"WAFL PCS ReplayFifoUnderflowErr",
 210          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayFifoUnderflowErr)},
 211         {"WAFL PCS ElasticFifoOverflowErr",
 212          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ElasticFifoOverflowErr)},
 213         {"WAFL PCS DeskewErr",
 214          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DeskewErr)},
 215         {"WAFL PCS DataStartupLimitErr",
 216          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataStartupLimitErr)},
 217         {"WAFL PCS FCInitTimeoutErr",
 218          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, FCInitTimeoutErr)},
 219         {"WAFL PCS RecoveryTimeoutErr",
 220          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryTimeoutErr)},
 221         {"WAFL PCS ReadySerialTimeoutErr",
 222          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReadySerialTimeoutErr)},
 223         {"WAFL PCS ReadySerialAttemptErr",
 224          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReadySerialAttemptErr)},
 225         {"WAFL PCS RecoveryAttemptErr",
 226          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryAttemptErr)},
 227         {"WAFL PCS RecoveryRelockAttemptErr",
 228          SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryRelockAttemptErr)},
 229 };
 230
 231 static const struct amdgpu_pcs_ras_field xgmi3x16_pcs_ras_fields[] = {
 232         {"XGMI3X16 PCS DataLossErr",
 233          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, DataLossErr)},
 234         {"XGMI3X16 PCS TrainingErr",
 235          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, TrainingErr)},
 236         {"XGMI3X16 PCS FlowCtrlAckErr",
 237          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, FlowCtrlAckErr)},
 238         {"XGMI3X16 PCS RxFifoUnderflowErr",
 239          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RxFifoUnderflowErr)},
 240         {"XGMI3X16 PCS RxFifoOverflowErr",
 241          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RxFifoOverflowErr)},
 242         {"XGMI3X16 PCS CRCErr",
 243          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, CRCErr)},
 244         {"XGMI3X16 PCS BERExceededErr",
 245          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, BERExceededErr)},
 246         {"XGMI3X16 PCS TxVcidDataErr",
 247          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, TxVcidDataErr)},
 248         {"XGMI3X16 PCS ReplayBufParityErr",
 249          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ReplayBufParityErr)},
 250         {"XGMI3X16 PCS DataParityErr",
 251          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, DataParityErr)},
 252         {"XGMI3X16 PCS ReplayFifoOverflowErr",
 253          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ReplayFifoOverflowErr)},
 254         {"XGMI3X16 PCS ReplayFifoUnderflowErr",
 255          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ReplayFifoUnderflowErr)},
 256         {"XGMI3X16 PCS ElasticFifoOverflowErr",
 257          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ElasticFifoOverflowErr)},
 258         {"XGMI3X16 PCS DeskewErr",
 259          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, DeskewErr)},
 260         {"XGMI3X16 PCS FlowCtrlCRCErr",
 261          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, FlowCtrlCRCErr)},
 262         {"XGMI3X16 PCS DataStartupLimitErr",
 263          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, DataStartupLimitErr)},
 264         {"XGMI3X16 PCS FCInitTimeoutErr",
 265          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, FCInitTimeoutErr)},
 266         {"XGMI3X16 PCS RecoveryTimeoutErr",
 267          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RecoveryTimeoutErr)},
 268         {"XGMI3X16 PCS ReadySerialTimeoutErr",
 269          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ReadySerialTimeoutErr)},
 270         {"XGMI3X16 PCS ReadySerialAttemptErr",
 271          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ReadySerialAttemptErr)},
 272         {"XGMI3X16 PCS RecoveryAttemptErr",
 273          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RecoveryAttemptErr)},
 274         {"XGMI3X16 PCS RecoveryRelockAttemptErr",
 275          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RecoveryRelockAttemptErr)},
 276         {"XGMI3X16 PCS ReplayAttemptErr",
 277          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, ReplayAttemptErr)},
 278         {"XGMI3X16 PCS SyncHdrErr",
 279          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, SyncHdrErr)},
 280         {"XGMI3X16 PCS TxReplayTimeoutErr",
 281          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, TxReplayTimeoutErr)},
 282         {"XGMI3X16 PCS RxReplayTimeoutErr",
 283          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RxReplayTimeoutErr)},
 284         {"XGMI3X16 PCS LinkSubTxTimeoutErr",
 285          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, LinkSubTxTimeoutErr)},
 286         {"XGMI3X16 PCS LinkSubRxTimeoutErr",
 287          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, LinkSubRxTimeoutErr)},
 288         {"XGMI3X16 PCS RxCMDPktErr",
 289          SOC15_REG_FIELD(PCS_XGMI3X16_PCS_ERROR_STATUS, RxCMDPktErr)},
 290 };
 291
 292 /**
 293  * DOC: AMDGPU XGMI Support
 294  *
 295  * XGMI is a high speed interconnect that joins multiple GPU cards
 296  * into a homogeneous memory space that is organized by a collective
 297  * hive ID and individual node IDs, both of which are 64-bit numbers.
 298  *
 299  * The file xgmi_device_id contains the unique per GPU device ID and
 300  * is stored in the /sys/class/drm/card${cardno}/device/ directory.
 301  *
 302  * Inside the device directory a sub-directory 'xgmi_hive_info' is
 303  * created which contains the hive ID and the list of nodes.
 304  *
 305  * The hive ID is stored in:
 306  *   /sys/class/drm/card${cardno}/device/xgmi_hive_info/xgmi_hive_id
 307  *
 308  * The node information is stored in numbered directories:
 309  *   /sys/class/drm/card${cardno}/device/xgmi_hive_info/node${nodeno}/xgmi_device_id
 310  *
 311  * Each device has their own xgmi_hive_info direction with a mirror
 312  * set of node sub-directories.
 313  *
 314  * The XGMI memory space is built by contiguously adding the power of
 315  * two padded VRAM space from each node to each other.
 316  *
 317  */
 318
 319 static struct attribute amdgpu_xgmi_hive_id = {
 320         .name = "xgmi_hive_id",
 321         .mode = S_IRUGO
 322 };
 323
 324 static struct attribute *amdgpu_xgmi_hive_attrs[] = {
 325         &amdgpu_xgmi_hive_id,
 326         NULL
 327 };
 328 ATTRIBUTE_GROUPS(amdgpu_xgmi_hive);
 329
 330 static ssize_t amdgpu_xgmi_show_attrs(struct kobject *kobj,
 331         struct attribute *attr, char *buf)
 332 {
 333         struct amdgpu_hive_info *hive = container_of(
 334                 kobj, struct amdgpu_hive_info, kobj);
 335
 336         if (attr == &amdgpu_xgmi_hive_id)
 337                 return snprintf(buf, PAGE_SIZE, "%llu\n", hive->hive_id);
 338
 339         return 0;
 340 }
 341
 342 static void amdgpu_xgmi_hive_release(struct kobject *kobj)
 343 {
 344         struct amdgpu_hive_info *hive = container_of(
 345                 kobj, struct amdgpu_hive_info, kobj);
 346
 347         amdgpu_reset_put_reset_domain(hive->reset_domain);
 348         hive->reset_domain = NULL;
 349
 350         mutex_destroy(&hive->hive_lock);
 351         kfree(hive);
 352 }
 353
 354 static const struct sysfs_ops amdgpu_xgmi_hive_ops = {
 355         .show = amdgpu_xgmi_show_attrs,
 356 };
 357
 358 static const struct kobj_type amdgpu_xgmi_hive_type = {
 359         .release = amdgpu_xgmi_hive_release,
 360         .sysfs_ops = &amdgpu_xgmi_hive_ops,
 361         .default_groups = amdgpu_xgmi_hive_groups,
 362 };
 363
 364 static ssize_t amdgpu_xgmi_show_device_id(struct device *dev,
 365                                      struct device_attribute *attr,
 366                                      char *buf)
 367 {
 368         struct drm_device *ddev = dev_get_drvdata(dev);
 369         struct amdgpu_device *adev = drm_to_adev(ddev);
 370
 371         return sysfs_emit(buf, "%llu\n", adev->gmc.xgmi.node_id);
 372
 373 }
 374
 375 static ssize_t amdgpu_xgmi_show_physical_id(struct device *dev,
 376                                      struct device_attribute *attr,
 377                                      char *buf)
 378 {
 379         struct drm_device *ddev = dev_get_drvdata(dev);
 380         struct amdgpu_device *adev = drm_to_adev(ddev);
 381
 382         return sysfs_emit(buf, "%u\n", adev->gmc.xgmi.physical_node_id);
 383
 384 }
 385
 386 static ssize_t amdgpu_xgmi_show_num_hops(struct device *dev,
 387                                         struct device_attribute *attr,
 388                                         char *buf)
 389 {
 390         struct drm_device *ddev = dev_get_drvdata(dev);
 391         struct amdgpu_device *adev = drm_to_adev(ddev);
 392         struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info;
 393         int i;
 394
 395         for (i = 0; i < top->num_nodes; i++)
 396                 sprintf(buf + 3 * i, "%02x ", top->nodes[i].num_hops);
 397
 398         return sysfs_emit(buf, "%s\n", buf);
 399 }
 400
 401 static ssize_t amdgpu_xgmi_show_num_links(struct device *dev,
 402                                         struct device_attribute *attr,
 403                                         char *buf)
 404 {
 405         struct drm_device *ddev = dev_get_drvdata(dev);
 406         struct amdgpu_device *adev = drm_to_adev(ddev);
 407         struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info;
 408         int i;
 409
 410         for (i = 0; i < top->num_nodes; i++)
 411                 sprintf(buf + 3 * i, "%02x ", top->nodes[i].num_links);
 412
 413         return sysfs_emit(buf, "%s\n", buf);
 414 }
 415
 416 static ssize_t amdgpu_xgmi_show_connected_port_num(struct device *dev,
 417                                         struct device_attribute *attr,
 418                                         char *buf)
 419 {
 420         struct drm_device *ddev = dev_get_drvdata(dev);
 421         struct amdgpu_device *adev = drm_to_adev(ddev);
 422         struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info;
 423         int i, j, size = 0;
 424         int current_node;
 425         /*
 426          * get the node id in the sysfs for the current socket and show
 427          * it in the port num info output in the sysfs for easy reading.
 428          * it is NOT the one retrieved from xgmi ta.
 429          */
 430         for (i = 0; i < top->num_nodes; i++) {
 431                 if (top->nodes[i].node_id == adev->gmc.xgmi.node_id) {
 432                         current_node = i;
 433                         break;
 434                 }
 435         }
 436
 437         if (i == top->num_nodes)
 438                 return -EINVAL;
 439
 440         for (i = 0; i < top->num_nodes; i++) {
 441                 for (j = 0; j < top->nodes[i].num_links; j++)
 442                         /* node id in sysfs starts from 1 rather than 0 so +1 here */
 443                         size += sysfs_emit_at(buf, size, "%02x:%02x ->  %02x:%02x\n", current_node + 1,
 444                                               top->nodes[i].port_num[j].src_xgmi_port_num, i + 1,
 445                                               top->nodes[i].port_num[j].dst_xgmi_port_num);
 446         }
 447
 448         return size;
 449 }
 450
 451 #define AMDGPU_XGMI_SET_FICAA(o)        ((o) | 0x456801)
 452 static ssize_t amdgpu_xgmi_show_error(struct device *dev,
 453                                       struct device_attribute *attr,
 454                                       char *buf)
 455 {
 456         struct drm_device *ddev = dev_get_drvdata(dev);
 457         struct amdgpu_device *adev = drm_to_adev(ddev);
 458         uint32_t ficaa_pie_ctl_in, ficaa_pie_status_in;
 459         uint64_t fica_out;
 460         unsigned int error_count = 0;
 461
 462         ficaa_pie_ctl_in = AMDGPU_XGMI_SET_FICAA(0x200);
 463         ficaa_pie_status_in = AMDGPU_XGMI_SET_FICAA(0x208);
 464
 465         if ((!adev->df.funcs) ||
 466             (!adev->df.funcs->get_fica) ||
 467             (!adev->df.funcs->set_fica))
 468                 return -EINVAL;
 469
 470         fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_ctl_in);
 471         if (fica_out != 0x1f)
 472                 pr_err("xGMI error counters not enabled!\n");
 473
 474         fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_status_in);
 475
 476         if ((fica_out & 0xffff) == 2)
 477                 error_count = ((fica_out >> 62) & 0x1) + (fica_out >> 63);
 478
 479         adev->df.funcs->set_fica(adev, ficaa_pie_status_in, 0, 0);
 480
 481         return sysfs_emit(buf, "%u\n", error_count);
 482 }
 483
 484
 485 static DEVICE_ATTR(xgmi_device_id, S_IRUGO, amdgpu_xgmi_show_device_id, NULL);
 486 static DEVICE_ATTR(xgmi_physical_id, 0444, amdgpu_xgmi_show_physical_id, NULL);
 487 static DEVICE_ATTR(xgmi_error, S_IRUGO, amdgpu_xgmi_show_error, NULL);
 488 static DEVICE_ATTR(xgmi_num_hops, S_IRUGO, amdgpu_xgmi_show_num_hops, NULL);
 489 static DEVICE_ATTR(xgmi_num_links, S_IRUGO, amdgpu_xgmi_show_num_links, NULL);
 490 static DEVICE_ATTR(xgmi_port_num, S_IRUGO, amdgpu_xgmi_show_connected_port_num, NULL);
 491
 492 static int amdgpu_xgmi_sysfs_add_dev_info(struct amdgpu_device *adev,
 493                                          struct amdgpu_hive_info *hive)
 494 {
 495         int ret = 0;
 496         char node[10] = { 0 };
 497
 498         /* Create xgmi device id file */
 499         ret = device_create_file(adev->dev, &dev_attr_xgmi_device_id);
 500         if (ret) {
 501                 dev_err(adev->dev, "XGMI: Failed to create device file xgmi_device_id\n");
 502                 return ret;
 503         }
 504
 505         ret = device_create_file(adev->dev, &dev_attr_xgmi_physical_id);
 506         if (ret) {
 507                 dev_err(adev->dev, "XGMI: Failed to create device file xgmi_physical_id\n");
 508                 return ret;
 509         }
 510
 511         /* Create xgmi error file */
 512         ret = device_create_file(adev->dev, &dev_attr_xgmi_error);
 513         if (ret)
 514                 pr_err("failed to create xgmi_error\n");
 515
 516         /* Create xgmi num hops file */
 517         ret = device_create_file(adev->dev, &dev_attr_xgmi_num_hops);
 518         if (ret)
 519                 pr_err("failed to create xgmi_num_hops\n");
 520
 521         /* Create xgmi num links file */
 522         ret = device_create_file(adev->dev, &dev_attr_xgmi_num_links);
 523         if (ret)
 524                 pr_err("failed to create xgmi_num_links\n");
 525
 526         /* Create xgmi port num file if supported */
 527         if (adev->psp.xgmi_context.xgmi_ta_caps & EXTEND_PEER_LINK_INFO_CMD_FLAG) {
 528                 ret = device_create_file(adev->dev, &dev_attr_xgmi_port_num);
 529                 if (ret)
 530                         dev_err(adev->dev, "failed to create xgmi_port_num\n");
 531         }
 532
 533         /* Create sysfs link to hive info folder on the first device */
 534         if (hive->kobj.parent != (&adev->dev->kobj)) {
 535                 ret = sysfs_create_link(&adev->dev->kobj, &hive->kobj,
 536                                         "xgmi_hive_info");
 537                 if (ret) {
 538                         dev_err(adev->dev, "XGMI: Failed to create link to hive info");
 539                         goto remove_file;
 540                 }
 541         }
 542
 543         sprintf(node, "node%d", atomic_read(&hive->number_devices));
 544         /* Create sysfs link form the hive folder to yourself */
 545         ret = sysfs_create_link(&hive->kobj, &adev->dev->kobj, node);
 546         if (ret) {
 547                 dev_err(adev->dev, "XGMI: Failed to create link from hive info");
 548                 goto remove_link;
 549         }
 550
 551         goto success;
 552
 553
 554 remove_link:
 555         sysfs_remove_link(&adev->dev->kobj, adev_to_drm(adev)->unique);
 556
 557 remove_file:
 558         device_remove_file(adev->dev, &dev_attr_xgmi_device_id);
 559         device_remove_file(adev->dev, &dev_attr_xgmi_physical_id);
 560         device_remove_file(adev->dev, &dev_attr_xgmi_error);
 561         device_remove_file(adev->dev, &dev_attr_xgmi_num_hops);
 562         device_remove_file(adev->dev, &dev_attr_xgmi_num_links);
 563         if (adev->psp.xgmi_context.xgmi_ta_caps & EXTEND_PEER_LINK_INFO_CMD_FLAG)
 564                 device_remove_file(adev->dev, &dev_attr_xgmi_port_num);
 565
 566 success:
 567         return ret;
 568 }
 569
 570 static void amdgpu_xgmi_sysfs_rem_dev_info(struct amdgpu_device *adev,
 571                                           struct amdgpu_hive_info *hive)
 572 {
 573         char node[10];
 574         memset(node, 0, sizeof(node));
 575
 576         device_remove_file(adev->dev, &dev_attr_xgmi_device_id);
 577         device_remove_file(adev->dev, &dev_attr_xgmi_physical_id);
 578         device_remove_file(adev->dev, &dev_attr_xgmi_error);
 579         device_remove_file(adev->dev, &dev_attr_xgmi_num_hops);
 580         device_remove_file(adev->dev, &dev_attr_xgmi_num_links);
 581         if (adev->psp.xgmi_context.xgmi_ta_caps & EXTEND_PEER_LINK_INFO_CMD_FLAG)
 582                 device_remove_file(adev->dev, &dev_attr_xgmi_port_num);
 583
 584         if (hive->kobj.parent != (&adev->dev->kobj))
 585                 sysfs_remove_link(&adev->dev->kobj,"xgmi_hive_info");
 586
 587         sprintf(node, "node%d", atomic_read(&hive->number_devices));
 588         sysfs_remove_link(&hive->kobj, node);
 589
 590 }
 591
 592
 593
 594 struct amdgpu_hive_info *amdgpu_get_xgmi_hive(struct amdgpu_device *adev)
 595 {
 596         struct amdgpu_hive_info *hive = NULL;
 597         int ret;
 598
 599         if (!adev->gmc.xgmi.hive_id)
 600                 return NULL;
 601
 602         if (adev->hive) {
 603                 kobject_get(&adev->hive->kobj);
 604                 return adev->hive;
 605         }
 606
 607         mutex_lock(&xgmi_mutex);
 608
 609         list_for_each_entry(hive, &xgmi_hive_list, node)  {
 610                 if (hive->hive_id == adev->gmc.xgmi.hive_id)
 611                         goto pro_end;
 612         }
 613
 614         hive = kzalloc(sizeof(*hive), GFP_KERNEL);
 615         if (!hive) {
 616                 dev_err(adev->dev, "XGMI: allocation failed\n");
 617                 ret = -ENOMEM;
 618                 hive = NULL;
 619                 goto pro_end;
 620         }
 621
 622         /* initialize new hive if not exist */
 623         ret = kobject_init_and_add(&hive->kobj,
 624                         &amdgpu_xgmi_hive_type,
 625                         &adev->dev->kobj,
 626                         "%s", "xgmi_hive_info");
 627         if (ret) {
 628                 dev_err(adev->dev, "XGMI: failed initializing kobject for xgmi hive\n");
 629                 kobject_put(&hive->kobj);
 630                 hive = NULL;
 631                 goto pro_end;
 632         }
 633
 634         /**
 635          * Only init hive->reset_domain for none SRIOV configuration. For SRIOV,
 636          * Host driver decide how to reset the GPU either through FLR or chain reset.
 637          * Guest side will get individual notifications from the host for the FLR
 638          * if necessary.
 639          */
 640         if (!amdgpu_sriov_vf(adev)) {
 641         /**
 642          * Avoid recreating reset domain when hive is reconstructed for the case
 643          * of reset the devices in the XGMI hive during probe for passthrough GPU
 644          * See https://www.spinics.net/lists/amd-gfx/msg58836.html
 645          */
 646                 if (adev->reset_domain->type != XGMI_HIVE) {
 647                         hive->reset_domain =
 648                                 amdgpu_reset_create_reset_domain(XGMI_HIVE, "amdgpu-reset-hive");
 649                         if (!hive->reset_domain) {
 650                                 dev_err(adev->dev, "XGMI: failed initializing reset domain for xgmi hive\n");
 651                                 ret = -ENOMEM;
 652                                 kobject_put(&hive->kobj);
 653                                 hive = NULL;
 654                                 goto pro_end;
 655                         }
 656                 } else {
 657                         amdgpu_reset_get_reset_domain(adev->reset_domain);
 658                         hive->reset_domain = adev->reset_domain;
 659                 }
 660         }
 661
 662         hive->hive_id = adev->gmc.xgmi.hive_id;
 663         INIT_LIST_HEAD(&hive->device_list);
 664         INIT_LIST_HEAD(&hive->node);
 665         mutex_init(&hive->hive_lock);
 666         atomic_set(&hive->number_devices, 0);
 667         task_barrier_init(&hive->tb);
 668         hive->pstate = AMDGPU_XGMI_PSTATE_UNKNOWN;
 669         hive->hi_req_gpu = NULL;
 670         atomic_set(&hive->requested_nps_mode, UNKNOWN_MEMORY_PARTITION_MODE);
 671
 672         /*
 673          * hive pstate on boot is high in vega20 so we have to go to low
 674          * pstate on after boot.
 675          */
 676         hive->hi_req_count = AMDGPU_MAX_XGMI_DEVICE_PER_HIVE;
 677         list_add_tail(&hive->node, &xgmi_hive_list);
 678
 679 pro_end:
 680         if (hive)
 681                 kobject_get(&hive->kobj);
 682         mutex_unlock(&xgmi_mutex);
 683         return hive;
 684 }
 685
 686 void amdgpu_put_xgmi_hive(struct amdgpu_hive_info *hive)
 687 {
 688         if (hive)
 689                 kobject_put(&hive->kobj);
 690 }
 691
 692 int amdgpu_xgmi_set_pstate(struct amdgpu_device *adev, int pstate)
 693 {
 694         int ret = 0;
 695         struct amdgpu_hive_info *hive;
 696         struct amdgpu_device *request_adev;
 697         bool is_hi_req = pstate == AMDGPU_XGMI_PSTATE_MAX_VEGA20;
 698         bool init_low;
 699
 700         hive = amdgpu_get_xgmi_hive(adev);
 701         if (!hive)
 702                 return 0;
 703
 704         request_adev = hive->hi_req_gpu ? hive->hi_req_gpu : adev;
 705         init_low = hive->pstate == AMDGPU_XGMI_PSTATE_UNKNOWN;
 706         amdgpu_put_xgmi_hive(hive);
 707         /* fw bug so temporarily disable pstate switching */
 708         return 0;
 709
 710         if (!hive || adev->asic_type != CHIP_VEGA20)
 711                 return 0;
 712
 713         mutex_lock(&hive->hive_lock);
 714
 715         if (is_hi_req)
 716                 hive->hi_req_count++;
 717         else
 718                 hive->hi_req_count--;
 719
 720         /*
 721          * Vega20 only needs single peer to request pstate high for the hive to
 722          * go high but all peers must request pstate low for the hive to go low
 723          */
 724         if (hive->pstate == pstate ||
 725                         (!is_hi_req && hive->hi_req_count && !init_low))
 726                 goto out;
 727
 728         dev_dbg(request_adev->dev, "Set xgmi pstate %d.\n", pstate);
 729
 730         ret = amdgpu_dpm_set_xgmi_pstate(request_adev, pstate);
 731         if (ret) {
 732                 dev_err(request_adev->dev,
 733                         "XGMI: Set pstate failure on device %llx, hive %llx, ret %d",
 734                         request_adev->gmc.xgmi.node_id,
 735                         request_adev->gmc.xgmi.hive_id, ret);
 736                 goto out;
 737         }
 738
 739         if (init_low)
 740                 hive->pstate = hive->hi_req_count ?
 741                                         hive->pstate : AMDGPU_XGMI_PSTATE_MIN;
 742         else {
 743                 hive->pstate = pstate;
 744                 hive->hi_req_gpu = pstate != AMDGPU_XGMI_PSTATE_MIN ?
 745                                                         adev : NULL;
 746         }
 747 out:
 748         mutex_unlock(&hive->hive_lock);
 749         return ret;
 750 }
 751
 752 int amdgpu_xgmi_update_topology(struct amdgpu_hive_info *hive, struct amdgpu_device *adev)
 753 {
 754         int ret;
 755
 756         if (amdgpu_sriov_vf(adev))
 757                 return 0;
 758
 759         /* Each psp need to set the latest topology */
 760         ret = psp_xgmi_set_topology_info(&adev->psp,
 761                                          atomic_read(&hive->number_devices),
 762                                          &adev->psp.xgmi_context.top_info);
 763         if (ret)
 764                 dev_err(adev->dev,
 765                         "XGMI: Set topology failure on device %llx, hive %llx, ret %d",
 766                         adev->gmc.xgmi.node_id,
 767                         adev->gmc.xgmi.hive_id, ret);
 768
 769         return ret;
 770 }
 771
 772
 773 /*
 774  * NOTE psp_xgmi_node_info.num_hops layout is as follows:
 775  * num_hops[7:6] = link type (0 = xGMI2, 1 = xGMI3, 2/3 = reserved)
 776  * num_hops[5:3] = reserved
 777  * num_hops[2:0] = number of hops
 778  */
 779 int amdgpu_xgmi_get_hops_count(struct amdgpu_device *adev,
 780                 struct amdgpu_device *peer_adev)
 781 {
 782         struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info;
 783         uint8_t num_hops_mask = 0x7;
 784         int i;
 785
 786         for (i = 0 ; i < top->num_nodes; ++i)
 787                 if (top->nodes[i].node_id == peer_adev->gmc.xgmi.node_id)
 788                         return top->nodes[i].num_hops & num_hops_mask;
 789         return  -EINVAL;
 790 }
 791
 792 int amdgpu_xgmi_get_num_links(struct amdgpu_device *adev,
 793                 struct amdgpu_device *peer_adev)
 794 {
 795         struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info;
 796         int i;
 797
 798         for (i = 0 ; i < top->num_nodes; ++i)
 799                 if (top->nodes[i].node_id == peer_adev->gmc.xgmi.node_id)
 800                         return top->nodes[i].num_links;
 801         return  -EINVAL;
 802 }
 803
 804 bool amdgpu_xgmi_get_is_sharing_enabled(struct amdgpu_device *adev,
 805                                         struct amdgpu_device *peer_adev)
 806 {
 807         struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info;
 808         int i;
 809
 810         /* Sharing should always be enabled for non-SRIOV. */
 811         if (!amdgpu_sriov_vf(adev))
 812                 return true;
 813
 814         for (i = 0 ; i < top->num_nodes; ++i)
 815                 if (top->nodes[i].node_id == peer_adev->gmc.xgmi.node_id)
 816                         return !!top->nodes[i].is_sharing_enabled;
 817
 818         return false;
 819 }
 820
 821 /*
 822  * Devices that support extended data require the entire hive to initialize with
 823  * the shared memory buffer flag set.
 824  *
 825  * Hive locks and conditions apply - see amdgpu_xgmi_add_device
 826  */
 827 static int amdgpu_xgmi_initialize_hive_get_data_partition(struct amdgpu_hive_info *hive,
 828                                                         bool set_extended_data)
 829 {
 830         struct amdgpu_device *tmp_adev;
 831         int ret;
 832
 833         list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
 834                 ret = psp_xgmi_initialize(&tmp_adev->psp, set_extended_data, false);
 835                 if (ret) {
 836                         dev_err(tmp_adev->dev,
 837                                 "XGMI: Failed to initialize xgmi session for data partition %i\n",
 838                                 set_extended_data);
 839                         return ret;
 840                 }
 841
 842         }
 843
 844         return 0;
 845 }
 846
 847 static void amdgpu_xgmi_fill_topology_info(struct amdgpu_device *adev,
 848         struct amdgpu_device *peer_adev)
 849 {
 850         struct psp_xgmi_topology_info *top_info = &adev->psp.xgmi_context.top_info;
 851         struct psp_xgmi_topology_info *peer_info = &peer_adev->psp.xgmi_context.top_info;
 852
 853         for (int i = 0; i < peer_info->num_nodes; i++) {
 854                 if (peer_info->nodes[i].node_id == adev->gmc.xgmi.node_id) {
 855                         for (int j = 0; j < top_info->num_nodes; j++) {
 856                                 if (top_info->nodes[j].node_id == peer_adev->gmc.xgmi.node_id) {
 857                                         peer_info->nodes[i].num_hops = top_info->nodes[j].num_hops;
 858                                         peer_info->nodes[i].is_sharing_enabled =
 859                                                         top_info->nodes[j].is_sharing_enabled;
 860                                         peer_info->nodes[i].num_links =
 861                                                         top_info->nodes[j].num_links;
 862                                         return;
 863                                 }
 864                         }
 865                 }
 866         }
 867 }
 868
 869 int amdgpu_xgmi_add_device(struct amdgpu_device *adev)
 870 {
 871         struct psp_xgmi_topology_info *top_info;
 872         struct amdgpu_hive_info *hive;
 873         struct amdgpu_xgmi      *entry;
 874         struct amdgpu_device *tmp_adev = NULL;
 875
 876         int count = 0, ret = 0;
 877
 878         if (!adev->gmc.xgmi.supported)
 879                 return 0;
 880
 881         if (amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) {
 882                 ret = psp_xgmi_initialize(&adev->psp, false, true);
 883                 if (ret) {
 884                         dev_err(adev->dev,
 885                                 "XGMI: Failed to initialize xgmi session\n");
 886                         return ret;
 887                 }
 888
 889                 ret = psp_xgmi_get_hive_id(&adev->psp, &adev->gmc.xgmi.hive_id);
 890                 if (ret) {
 891                         dev_err(adev->dev,
 892                                 "XGMI: Failed to get hive id\n");
 893                         return ret;
 894                 }
 895
 896                 ret = psp_xgmi_get_node_id(&adev->psp, &adev->gmc.xgmi.node_id);
 897                 if (ret) {
 898                         dev_err(adev->dev,
 899                                 "XGMI: Failed to get node id\n");
 900                         return ret;
 901                 }
 902         } else {
 903                 adev->gmc.xgmi.hive_id = 16;
 904                 adev->gmc.xgmi.node_id = adev->gmc.xgmi.physical_node_id + 16;
 905         }
 906
 907         hive = amdgpu_get_xgmi_hive(adev);
 908         if (!hive) {
 909                 ret = -EINVAL;
 910                 dev_err(adev->dev,
 911                         "XGMI: node 0x%llx, can not match hive 0x%llx in the hive list.\n",
 912                         adev->gmc.xgmi.node_id, adev->gmc.xgmi.hive_id);
 913                 goto exit;
 914         }
 915         mutex_lock(&hive->hive_lock);
 916
 917         top_info = &adev->psp.xgmi_context.top_info;
 918
 919         list_add_tail(&adev->gmc.xgmi.head, &hive->device_list);
 920         list_for_each_entry(entry, &hive->device_list, head)
 921                 top_info->nodes[count++].node_id = entry->node_id;
 922         top_info->num_nodes = count;
 923         atomic_set(&hive->number_devices, count);
 924
 925         task_barrier_add_task(&hive->tb);
 926
 927         if (amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) {
 928                 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
 929                         /* update node list for other device in the hive */
 930                         if (tmp_adev != adev) {
 931                                 top_info = &tmp_adev->psp.xgmi_context.top_info;
 932                                 top_info->nodes[count - 1].node_id =
 933                                         adev->gmc.xgmi.node_id;
 934                                 top_info->num_nodes = count;
 935                         }
 936                         ret = amdgpu_xgmi_update_topology(hive, tmp_adev);
 937                         if (ret)
 938                                 goto exit_unlock;
 939                 }
 940
 941                 if (amdgpu_sriov_vf(adev) &&
 942                         adev->psp.xgmi_context.xgmi_ta_caps & EXTEND_PEER_LINK_INFO_CMD_FLAG) {
 943                         /* only get topology for VF being init if it can support full duplex */
 944                         ret = psp_xgmi_get_topology_info(&adev->psp, count,
 945                                                 &adev->psp.xgmi_context.top_info, false);
 946                         if (ret) {
 947                                 dev_err(adev->dev,
 948                                         "XGMI: Get topology failure on device %llx, hive %llx, ret %d",
 949                                         adev->gmc.xgmi.node_id,
 950                                         adev->gmc.xgmi.hive_id, ret);
 951                                 /* To do: continue with some node failed or disable the whole hive*/
 952                                 goto exit_unlock;
 953                         }
 954
 955                         /* fill the topology info for peers instead of getting from PSP */
 956                         list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
 957                                 amdgpu_xgmi_fill_topology_info(adev, tmp_adev);
 958                         }
 959                 } else {
 960                         /* get latest topology info for each device from psp */
 961                         list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
 962                                 ret = psp_xgmi_get_topology_info(&tmp_adev->psp, count,
 963                                         &tmp_adev->psp.xgmi_context.top_info, false);
 964                                 if (ret) {
 965                                         dev_err(tmp_adev->dev,
 966                                                 "XGMI: Get topology failure on device %llx, hive %llx, ret %d",
 967                                                 tmp_adev->gmc.xgmi.node_id,
 968                                                 tmp_adev->gmc.xgmi.hive_id, ret);
 969                                         /* To do : continue with some node failed or disable the whole hive */
 970                                         goto exit_unlock;
 971                                 }
 972                         }
 973                 }
 974
 975                 /* get topology again for hives that support extended data */
 976                 if (adev->psp.xgmi_context.supports_extended_data) {
 977
 978                         /* initialize the hive to get extended data.  */
 979                         ret = amdgpu_xgmi_initialize_hive_get_data_partition(hive, true);
 980                         if (ret)
 981                                 goto exit_unlock;
 982
 983                         /* get the extended data. */
 984                         list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
 985                                 ret = psp_xgmi_get_topology_info(&tmp_adev->psp, count,
 986                                                 &tmp_adev->psp.xgmi_context.top_info, true);
 987                                 if (ret) {
 988                                         dev_err(tmp_adev->dev,
 989                                                 "XGMI: Get topology for extended data failure on device %llx, hive %llx, ret %d",
 990                                                 tmp_adev->gmc.xgmi.node_id,
 991                                                 tmp_adev->gmc.xgmi.hive_id, ret);
 992                                         goto exit_unlock;
 993                                 }
 994                         }
 995
 996                         /* initialize the hive to get non-extended data for the next round. */
 997                         ret = amdgpu_xgmi_initialize_hive_get_data_partition(hive, false);
 998                         if (ret)
 999                                 goto exit_unlock;
1000
1001                 }
1002         }
1003
1004         if (!ret)
1005                 ret = amdgpu_xgmi_sysfs_add_dev_info(adev, hive);
1006
1007 exit_unlock:
1008         mutex_unlock(&hive->hive_lock);
1009 exit:
1010         if (!ret) {
1011                 adev->hive = hive;
1012                 dev_info(adev->dev, "XGMI: Add node %d, hive 0x%llx.\n",
1013                          adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id);
1014         } else {
1015                 amdgpu_put_xgmi_hive(hive);
1016                 dev_err(adev->dev, "XGMI: Failed to add node %d, hive 0x%llx ret: %d\n",
1017                         adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id,
1018                         ret);
1019         }
1020
1021         return ret;
1022 }
1023
1024 int amdgpu_xgmi_remove_device(struct amdgpu_device *adev)
1025 {
1026         struct amdgpu_hive_info *hive = adev->hive;
1027
1028         if (!adev->gmc.xgmi.supported)
1029                 return -EINVAL;
1030
1031         if (!hive)
1032                 return -EINVAL;
1033
1034         mutex_lock(&hive->hive_lock);
1035         task_barrier_rem_task(&hive->tb);
1036         amdgpu_xgmi_sysfs_rem_dev_info(adev, hive);
1037         if (hive->hi_req_gpu == adev)
1038                 hive->hi_req_gpu = NULL;
1039         list_del(&adev->gmc.xgmi.head);
1040         mutex_unlock(&hive->hive_lock);
1041
1042         amdgpu_put_xgmi_hive(hive);
1043         adev->hive = NULL;
1044
1045         if (atomic_dec_return(&hive->number_devices) == 0) {
1046                 /* Remove the hive from global hive list */
1047                 mutex_lock(&xgmi_mutex);
1048                 list_del(&hive->node);
1049                 mutex_unlock(&xgmi_mutex);
1050
1051                 amdgpu_put_xgmi_hive(hive);
1052         }
1053
1054         return 0;
1055 }
1056
1057 static int xgmi_v6_4_0_aca_bank_parser(struct aca_handle *handle, struct aca_bank *bank,
1058                                        enum aca_smu_type type, void *data)
1059 {
1060         struct amdgpu_device *adev = handle->adev;
1061         struct aca_bank_info info;
1062         const char *error_str;
1063         u64 status, count;
1064         int ret, ext_error_code;
1065
1066         ret = aca_bank_info_decode(bank, &info);
1067         if (ret)
1068                 return ret;
1069
1070         status = bank->regs[ACA_REG_IDX_STATUS];
1071         ext_error_code = ACA_REG__STATUS__ERRORCODEEXT(status);
1072
1073         error_str = ext_error_code < ARRAY_SIZE(xgmi_v6_4_0_ras_error_code_ext) ?
1074                 xgmi_v6_4_0_ras_error_code_ext[ext_error_code] : NULL;
1075         if (error_str)
1076                 dev_info(adev->dev, "%s detected\n", error_str);
1077
1078         count = ACA_REG__MISC0__ERRCNT(bank->regs[ACA_REG_IDX_MISC0]);
1079
1080         switch (type) {
1081         case ACA_SMU_TYPE_UE:
1082                 if (ext_error_code != 0 && ext_error_code != 9)
1083                         count = 0ULL;
1084
1085                 ret = aca_error_cache_log_bank_error(handle, &info, ACA_ERROR_TYPE_UE, count);
1086                 break;
1087         case ACA_SMU_TYPE_CE:
1088                 count = ext_error_code == 6 ? count : 0ULL;
1089                 ret = aca_error_cache_log_bank_error(handle, &info, ACA_ERROR_TYPE_CE, count);
1090                 break;
1091         default:
1092                 return -EINVAL;
1093         }
1094
1095         return ret;
1096 }
1097
1098 static const struct aca_bank_ops xgmi_v6_4_0_aca_bank_ops = {
1099         .aca_bank_parser = xgmi_v6_4_0_aca_bank_parser,
1100 };
1101
1102 static const struct aca_info xgmi_v6_4_0_aca_info = {
1103         .hwip = ACA_HWIP_TYPE_PCS_XGMI,
1104         .mask = ACA_ERROR_UE_MASK | ACA_ERROR_CE_MASK,
1105         .bank_ops = &xgmi_v6_4_0_aca_bank_ops,
1106 };
1107
1108 static int amdgpu_xgmi_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *ras_block)
1109 {
1110         int r;
1111
1112         if (!adev->gmc.xgmi.supported ||
1113             adev->gmc.xgmi.num_physical_nodes == 0)
1114                 return 0;
1115
1116         amdgpu_ras_reset_error_count(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL);
1117
1118         r = amdgpu_ras_block_late_init(adev, ras_block);
1119         if (r)
1120                 return r;
1121
1122         switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) {
1123         case IP_VERSION(6, 4, 0):
1124                 r = amdgpu_ras_bind_aca(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL,
1125                                         &xgmi_v6_4_0_aca_info, NULL);
1126                 if (r)
1127                         goto late_fini;
1128                 break;
1129         default:
1130                 break;
1131         }
1132
1133         return 0;
1134
1135 late_fini:
1136         amdgpu_ras_block_late_fini(adev, ras_block);
1137
1138         return r;
1139 }
1140
1141 uint64_t amdgpu_xgmi_get_relative_phy_addr(struct amdgpu_device *adev,
1142                                            uint64_t addr)
1143 {
1144         struct amdgpu_xgmi *xgmi = &adev->gmc.xgmi;
1145         return (addr + xgmi->physical_node_id * xgmi->node_segment_size);
1146 }
1147
1148 static void pcs_clear_status(struct amdgpu_device *adev, uint32_t pcs_status_reg)
1149 {
1150         WREG32_PCIE(pcs_status_reg, 0xFFFFFFFF);
1151         WREG32_PCIE(pcs_status_reg, 0);
1152 }
1153
1154 static void amdgpu_xgmi_legacy_reset_ras_error_count(struct amdgpu_device *adev)
1155 {
1156         uint32_t i;
1157
1158         switch (adev->asic_type) {
1159         case CHIP_ARCTURUS:
1160                 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_arct); i++)
1161                         pcs_clear_status(adev,
1162                                          xgmi_pcs_err_status_reg_arct[i]);
1163                 break;
1164         case CHIP_VEGA20:
1165                 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_vg20); i++)
1166                         pcs_clear_status(adev,
1167                                          xgmi_pcs_err_status_reg_vg20[i]);
1168                 break;
1169         case CHIP_ALDEBARAN:
1170                 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_aldebaran); i++)
1171                         pcs_clear_status(adev,
1172                                          xgmi3x16_pcs_err_status_reg_aldebaran[i]);
1173                 for (i = 0; i < ARRAY_SIZE(walf_pcs_err_status_reg_aldebaran); i++)
1174                         pcs_clear_status(adev,
1175                                          walf_pcs_err_status_reg_aldebaran[i]);
1176                 break;
1177         default:
1178                 break;
1179         }
1180
1181         switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) {
1182         case IP_VERSION(6, 4, 0):
1183                 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_v6_4); i++)
1184                         pcs_clear_status(adev,
1185                                         xgmi3x16_pcs_err_status_reg_v6_4[i]);
1186                 break;
1187         default:
1188                 break;
1189         }
1190 }
1191
1192 static void __xgmi_v6_4_0_reset_error_count(struct amdgpu_device *adev, int xgmi_inst, u64 mca_base)
1193 {
1194         WREG64_MCA(xgmi_inst, mca_base, ACA_REG_IDX_STATUS, 0ULL);
1195 }
1196
1197 static void xgmi_v6_4_0_reset_error_count(struct amdgpu_device *adev, int xgmi_inst)
1198 {
1199         int i;
1200
1201         for (i = 0; i < ARRAY_SIZE(xgmi_v6_4_0_mca_base_array); i++)
1202                 __xgmi_v6_4_0_reset_error_count(adev, xgmi_inst, xgmi_v6_4_0_mca_base_array[i]);
1203 }
1204
1205 static void xgmi_v6_4_0_reset_ras_error_count(struct amdgpu_device *adev)
1206 {
1207         int i;
1208
1209         for_each_inst(i, adev->aid_mask)
1210                 xgmi_v6_4_0_reset_error_count(adev, i);
1211 }
1212
1213 static void amdgpu_xgmi_reset_ras_error_count(struct amdgpu_device *adev)
1214 {
1215         switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) {
1216         case IP_VERSION(6, 4, 0):
1217                 xgmi_v6_4_0_reset_ras_error_count(adev);
1218                 break;
1219         default:
1220                 amdgpu_xgmi_legacy_reset_ras_error_count(adev);
1221                 break;
1222         }
1223 }
1224
1225 static int amdgpu_xgmi_query_pcs_error_status(struct amdgpu_device *adev,
1226                                               uint32_t value,
1227                                                   uint32_t mask_value,
1228                                               uint32_t *ue_count,
1229                                               uint32_t *ce_count,
1230                                               bool is_xgmi_pcs,
1231                                                   bool check_mask)
1232 {
1233         int i;
1234         int ue_cnt = 0;
1235         const struct amdgpu_pcs_ras_field *pcs_ras_fields = NULL;
1236         uint32_t field_array_size = 0;
1237
1238         if (is_xgmi_pcs) {
1239                 if (amdgpu_ip_version(adev, XGMI_HWIP, 0) ==
1240                     IP_VERSION(6, 1, 0) ||
1241                     amdgpu_ip_version(adev, XGMI_HWIP, 0) ==
1242                     IP_VERSION(6, 4, 0)) {
1243                         pcs_ras_fields = &xgmi3x16_pcs_ras_fields[0];
1244                         field_array_size = ARRAY_SIZE(xgmi3x16_pcs_ras_fields);
1245                 } else {
1246                         pcs_ras_fields = &xgmi_pcs_ras_fields[0];
1247                         field_array_size = ARRAY_SIZE(xgmi_pcs_ras_fields);
1248                 }
1249         } else {
1250                 pcs_ras_fields = &wafl_pcs_ras_fields[0];
1251                 field_array_size = ARRAY_SIZE(wafl_pcs_ras_fields);
1252         }
1253
1254         if (check_mask)
1255                 value = value & ~mask_value;
1256
1257         /* query xgmi/walf pcs error status,
1258          * only ue is supported */
1259         for (i = 0; value && i < field_array_size; i++) {
1260                 ue_cnt = (value &
1261                                 pcs_ras_fields[i].pcs_err_mask) >>
1262                                 pcs_ras_fields[i].pcs_err_shift;
1263                 if (ue_cnt) {
1264                         dev_info(adev->dev, "%s detected\n",
1265                                  pcs_ras_fields[i].err_name);
1266                         *ue_count += ue_cnt;
1267                 }
1268
1269                 /* reset bit value if the bit is checked */
1270                 value &= ~(pcs_ras_fields[i].pcs_err_mask);
1271         }
1272
1273         return 0;
1274 }
1275
1276 static void amdgpu_xgmi_legacy_query_ras_error_count(struct amdgpu_device *adev,
1277                                                      void *ras_error_status)
1278 {
1279         struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status;
1280         int i, supported = 1;
1281         uint32_t data, mask_data = 0;
1282         uint32_t ue_cnt = 0, ce_cnt = 0;
1283
1284         if (!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL))
1285                 return ;
1286
1287         err_data->ue_count = 0;
1288         err_data->ce_count = 0;
1289
1290         switch (adev->asic_type) {
1291         case CHIP_ARCTURUS:
1292                 /* check xgmi pcs error */
1293                 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_arct); i++) {
1294                         data = RREG32_PCIE(xgmi_pcs_err_status_reg_arct[i]);
1295                         if (data)
1296                                 amdgpu_xgmi_query_pcs_error_status(adev, data,
1297                                                 mask_data, &ue_cnt, &ce_cnt, true, false);
1298                 }
1299                 /* check wafl pcs error */
1300                 for (i = 0; i < ARRAY_SIZE(wafl_pcs_err_status_reg_arct); i++) {
1301                         data = RREG32_PCIE(wafl_pcs_err_status_reg_arct[i]);
1302                         if (data)
1303                                 amdgpu_xgmi_query_pcs_error_status(adev, data,
1304                                                 mask_data, &ue_cnt, &ce_cnt, false, false);
1305                 }
1306                 break;
1307         case CHIP_VEGA20:
1308                 /* check xgmi pcs error */
1309                 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_vg20); i++) {
1310                         data = RREG32_PCIE(xgmi_pcs_err_status_reg_vg20[i]);
1311                         if (data)
1312                                 amdgpu_xgmi_query_pcs_error_status(adev, data,
1313                                                 mask_data, &ue_cnt, &ce_cnt, true, false);
1314                 }
1315                 /* check wafl pcs error */
1316                 for (i = 0; i < ARRAY_SIZE(wafl_pcs_err_status_reg_vg20); i++) {
1317                         data = RREG32_PCIE(wafl_pcs_err_status_reg_vg20[i]);
1318                         if (data)
1319                                 amdgpu_xgmi_query_pcs_error_status(adev, data,
1320                                                 mask_data, &ue_cnt, &ce_cnt, false, false);
1321                 }
1322                 break;
1323         case CHIP_ALDEBARAN:
1324                 /* check xgmi3x16 pcs error */
1325                 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_aldebaran); i++) {
1326                         data = RREG32_PCIE(xgmi3x16_pcs_err_status_reg_aldebaran[i]);
1327                         mask_data =
1328                                 RREG32_PCIE(xgmi3x16_pcs_err_noncorrectable_mask_reg_aldebaran[i]);
1329                         if (data)
1330                                 amdgpu_xgmi_query_pcs_error_status(adev, data,
1331                                                 mask_data, &ue_cnt, &ce_cnt, true, true);
1332                 }
1333                 /* check wafl pcs error */
1334                 for (i = 0; i < ARRAY_SIZE(walf_pcs_err_status_reg_aldebaran); i++) {
1335                         data = RREG32_PCIE(walf_pcs_err_status_reg_aldebaran[i]);
1336                         mask_data =
1337                                 RREG32_PCIE(walf_pcs_err_noncorrectable_mask_reg_aldebaran[i]);
1338                         if (data)
1339                                 amdgpu_xgmi_query_pcs_error_status(adev, data,
1340                                                 mask_data, &ue_cnt, &ce_cnt, false, true);
1341                 }
1342                 break;
1343         default:
1344                 supported = 0;
1345                 break;
1346         }
1347
1348         switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) {
1349         case IP_VERSION(6, 4, 0):
1350                 /* check xgmi3x16 pcs error */
1351                 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_v6_4); i++) {
1352                         data = RREG32_PCIE(xgmi3x16_pcs_err_status_reg_v6_4[i]);
1353                         mask_data =
1354                                 RREG32_PCIE(xgmi3x16_pcs_err_noncorrectable_mask_reg_v6_4[i]);
1355                         if (data)
1356                                 amdgpu_xgmi_query_pcs_error_status(adev, data,
1357                                                 mask_data, &ue_cnt, &ce_cnt, true, true);
1358                 }
1359                 break;
1360         default:
1361                 if (!supported)
1362                         dev_warn(adev->dev, "XGMI RAS error query not supported");
1363                 break;
1364         }
1365
1366         amdgpu_ras_reset_error_count(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL);
1367
1368         err_data->ue_count += ue_cnt;
1369         err_data->ce_count += ce_cnt;
1370 }
1371
1372 static enum aca_error_type xgmi_v6_4_0_pcs_mca_get_error_type(struct amdgpu_device *adev, u64 status)
1373 {
1374         const char *error_str;
1375         int ext_error_code;
1376
1377         ext_error_code = ACA_REG__STATUS__ERRORCODEEXT(status);
1378
1379         error_str = ext_error_code < ARRAY_SIZE(xgmi_v6_4_0_ras_error_code_ext) ?
1380                 xgmi_v6_4_0_ras_error_code_ext[ext_error_code] : NULL;
1381         if (error_str)
1382                 dev_info(adev->dev, "%s detected\n", error_str);
1383
1384         switch (ext_error_code) {
1385         case 0:
1386                 return ACA_ERROR_TYPE_UE;
1387         case 6:
1388                 return ACA_ERROR_TYPE_CE;
1389         default:
1390                 return -EINVAL;
1391         }
1392
1393         return -EINVAL;
1394 }
1395
1396 static void __xgmi_v6_4_0_query_error_count(struct amdgpu_device *adev, struct amdgpu_smuio_mcm_config_info *mcm_info,
1397                                             u64 mca_base, struct ras_err_data *err_data)
1398 {
1399         int xgmi_inst = mcm_info->die_id;
1400         u64 status = 0;
1401
1402         status = RREG64_MCA(xgmi_inst, mca_base, ACA_REG_IDX_STATUS);
1403         if (!ACA_REG__STATUS__VAL(status))
1404                 return;
1405
1406         switch (xgmi_v6_4_0_pcs_mca_get_error_type(adev, status)) {
1407         case ACA_ERROR_TYPE_UE:
1408                 amdgpu_ras_error_statistic_ue_count(err_data, mcm_info, 1ULL);
1409                 break;
1410         case ACA_ERROR_TYPE_CE:
1411                 amdgpu_ras_error_statistic_ce_count(err_data, mcm_info, 1ULL);
1412                 break;
1413         default:
1414                 break;
1415         }
1416
1417         WREG64_MCA(xgmi_inst, mca_base, ACA_REG_IDX_STATUS, 0ULL);
1418 }
1419
1420 static void xgmi_v6_4_0_query_error_count(struct amdgpu_device *adev, int xgmi_inst, struct ras_err_data *err_data)
1421 {
1422         struct amdgpu_smuio_mcm_config_info mcm_info = {
1423                 .socket_id = adev->smuio.funcs->get_socket_id(adev),
1424                 .die_id = xgmi_inst,
1425         };
1426         int i;
1427
1428         for (i = 0; i < ARRAY_SIZE(xgmi_v6_4_0_mca_base_array); i++)
1429                 __xgmi_v6_4_0_query_error_count(adev, &mcm_info, xgmi_v6_4_0_mca_base_array[i], err_data);
1430 }
1431
1432 static void xgmi_v6_4_0_query_ras_error_count(struct amdgpu_device *adev, void *ras_error_status)
1433 {
1434         struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status;
1435         int i;
1436
1437         for_each_inst(i, adev->aid_mask)
1438                 xgmi_v6_4_0_query_error_count(adev, i, err_data);
1439 }
1440
1441 static void amdgpu_xgmi_query_ras_error_count(struct amdgpu_device *adev,
1442                                               void *ras_error_status)
1443 {
1444         switch (amdgpu_ip_version(adev, XGMI_HWIP, 0)) {
1445         case IP_VERSION(6, 4, 0):
1446                 xgmi_v6_4_0_query_ras_error_count(adev, ras_error_status);
1447                 break;
1448         default:
1449                 amdgpu_xgmi_legacy_query_ras_error_count(adev, ras_error_status);
1450                 break;
1451         }
1452 }
1453
1454 /* Trigger XGMI/WAFL error */
1455 static int amdgpu_ras_error_inject_xgmi(struct amdgpu_device *adev,
1456                         void *inject_if, uint32_t instance_mask)
1457 {
1458         int ret1, ret2;
1459         struct ta_ras_trigger_error_input *block_info =
1460                                 (struct ta_ras_trigger_error_input *)inject_if;
1461
1462         if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_DISALLOW))
1463                 dev_warn(adev->dev, "Failed to disallow df cstate");
1464
1465         ret1 = amdgpu_dpm_set_pm_policy(adev, PP_PM_POLICY_XGMI_PLPD, XGMI_PLPD_DISALLOW);
1466         if (ret1 && ret1 != -EOPNOTSUPP)
1467                 dev_warn(adev->dev, "Failed to disallow XGMI power down");
1468
1469         ret2 = psp_ras_trigger_error(&adev->psp, block_info, instance_mask);
1470
1471         if (amdgpu_ras_intr_triggered())
1472                 return ret2;
1473
1474         ret1 = amdgpu_dpm_set_pm_policy(adev, PP_PM_POLICY_XGMI_PLPD, XGMI_PLPD_DEFAULT);
1475         if (ret1 && ret1 != -EOPNOTSUPP)
1476                 dev_warn(adev->dev, "Failed to allow XGMI power down");
1477
1478         if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_ALLOW))
1479                 dev_warn(adev->dev, "Failed to allow df cstate");
1480
1481         return ret2;
1482 }
1483
1484 struct amdgpu_ras_block_hw_ops  xgmi_ras_hw_ops = {
1485         .query_ras_error_count = amdgpu_xgmi_query_ras_error_count,
1486         .reset_ras_error_count = amdgpu_xgmi_reset_ras_error_count,
1487         .ras_error_inject = amdgpu_ras_error_inject_xgmi,
1488 };
1489
1490 struct amdgpu_xgmi_ras xgmi_ras = {
1491         .ras_block = {
1492                 .hw_ops = &xgmi_ras_hw_ops,
1493                 .ras_late_init = amdgpu_xgmi_ras_late_init,
1494         },
1495 };
1496
1497 int amdgpu_xgmi_ras_sw_init(struct amdgpu_device *adev)
1498 {
1499         int err;
1500         struct amdgpu_xgmi_ras *ras;
1501
1502         if (!adev->gmc.xgmi.ras)
1503                 return 0;
1504
1505         ras = adev->gmc.xgmi.ras;
1506         err = amdgpu_ras_register_ras_block(adev, &ras->ras_block);
1507         if (err) {
1508                 dev_err(adev->dev, "Failed to register xgmi_wafl_pcs ras block!\n");
1509                 return err;
1510         }
1511
1512         strcpy(ras->ras_block.ras_comm.name, "xgmi_wafl");
1513         ras->ras_block.ras_comm.block = AMDGPU_RAS_BLOCK__XGMI_WAFL;
1514         ras->ras_block.ras_comm.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE;
1515         adev->gmc.xgmi.ras_if = &ras->ras_block.ras_comm;
1516
1517         return 0;
1518 }
1519
1520 static void amdgpu_xgmi_reset_on_init_work(struct work_struct *work)
1521 {
1522         struct amdgpu_hive_info *hive =
1523                 container_of(work, struct amdgpu_hive_info, reset_on_init_work);
1524         struct amdgpu_reset_context reset_context;
1525         struct amdgpu_device *tmp_adev;
1526         struct list_head device_list;
1527         int r;
1528
1529         mutex_lock(&hive->hive_lock);
1530
1531         INIT_LIST_HEAD(&device_list);
1532         list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head)
1533                 list_add_tail(&tmp_adev->reset_list, &device_list);
1534
1535         tmp_adev = list_first_entry(&device_list, struct amdgpu_device,
1536                                     reset_list);
1537         amdgpu_device_lock_reset_domain(tmp_adev->reset_domain);
1538
1539         reset_context.method = AMD_RESET_METHOD_ON_INIT;
1540         reset_context.reset_req_dev = tmp_adev;
1541         reset_context.hive = hive;
1542         reset_context.reset_device_list = &device_list;
1543         set_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
1544         set_bit(AMDGPU_SKIP_COREDUMP, &reset_context.flags);
1545
1546         amdgpu_reset_do_xgmi_reset_on_init(&reset_context);
1547         mutex_unlock(&hive->hive_lock);
1548         amdgpu_device_unlock_reset_domain(tmp_adev->reset_domain);
1549
1550         list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
1551                 r = amdgpu_ras_init_badpage_info(tmp_adev);
1552                 if (r && r != -EHWPOISON)
1553                         dev_err(tmp_adev->dev,
1554                                 "error during bad page data initialization");
1555         }
1556 }
1557
1558 static void amdgpu_xgmi_schedule_reset_on_init(struct amdgpu_hive_info *hive)
1559 {
1560         INIT_WORK(&hive->reset_on_init_work, amdgpu_xgmi_reset_on_init_work);
1561         amdgpu_reset_domain_schedule(hive->reset_domain,
1562                                      &hive->reset_on_init_work);
1563 }
1564
1565 int amdgpu_xgmi_reset_on_init(struct amdgpu_device *adev)
1566 {
1567         struct amdgpu_hive_info *hive;
1568         bool reset_scheduled;
1569         int num_devs;
1570
1571         hive = amdgpu_get_xgmi_hive(adev);
1572         if (!hive)
1573                 return -EINVAL;
1574
1575         mutex_lock(&hive->hive_lock);
1576         num_devs = atomic_read(&hive->number_devices);
1577         reset_scheduled = false;
1578         if (num_devs == adev->gmc.xgmi.num_physical_nodes) {
1579                 amdgpu_xgmi_schedule_reset_on_init(hive);
1580                 reset_scheduled = true;
1581         }
1582
1583         mutex_unlock(&hive->hive_lock);
1584         amdgpu_put_xgmi_hive(hive);
1585
1586         if (reset_scheduled)
1587                 flush_work(&hive->reset_on_init_work);
1588
1589         return 0;
1590 }
1591
1592 int amdgpu_xgmi_request_nps_change(struct amdgpu_device *adev,
1593                                    struct amdgpu_hive_info *hive,
1594                                    int req_nps_mode)
1595 {
1596         struct amdgpu_device *tmp_adev;
1597         int cur_nps_mode, r;
1598
1599         /* This is expected to be called only during unload of driver. The
1600          * request needs to be placed only once for all devices in the hive. If
1601          * one of them fail, revert the request for previous successful devices.
1602          * After placing the request, make hive mode as UNKNOWN so that other
1603          * devices don't request anymore.
1604          */
1605         mutex_lock(&hive->hive_lock);
1606         if (atomic_read(&hive->requested_nps_mode) ==
1607             UNKNOWN_MEMORY_PARTITION_MODE) {
1608                 dev_dbg(adev->dev, "Unexpected entry for hive NPS change");
1609                 mutex_unlock(&hive->hive_lock);
1610                 return 0;
1611         }
1612         list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
1613                 r = adev->gmc.gmc_funcs->request_mem_partition_mode(
1614                         tmp_adev, req_nps_mode);
1615                 if (r)
1616                         break;
1617         }
1618         if (r) {
1619                 /* Request back current mode if one of the requests failed */
1620                 cur_nps_mode =
1621                         adev->gmc.gmc_funcs->query_mem_partition_mode(tmp_adev);
1622                 list_for_each_entry_continue_reverse(
1623                         tmp_adev, &hive->device_list, gmc.xgmi.head)
1624                         adev->gmc.gmc_funcs->request_mem_partition_mode(
1625                                 tmp_adev, cur_nps_mode);
1626         }
1627         /* Set to UNKNOWN so that other devices don't request anymore */
1628         atomic_set(&hive->requested_nps_mode, UNKNOWN_MEMORY_PARTITION_MODE);
1629         mutex_unlock(&hive->hive_lock);
1630
1631         return r;
1632 }