drivers/gpu/drm/amd/amdgpu/amdgpu_ras.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/debugfs.h>
  25 #include <linux/list.h>
  26 #include <linux/module.h>
  27 #include <linux/uaccess.h>
  28 #include <linux/reboot.h>
  29 #include <linux/syscalls.h>
  30 #include <linux/pm_runtime.h>
  31
  32 #include "amdgpu.h"
  33 #include "amdgpu_ras.h"
  34 #include "amdgpu_atomfirmware.h"
  35 #include "amdgpu_xgmi.h"
  36 #include "ivsrcid/nbio/irqsrcs_nbif_7_4.h"
  37 #include "nbio_v4_3.h"
  38 #include "atom.h"
  39 #include "amdgpu_reset.h"
  40
  41 #ifdef CONFIG_X86_MCE_AMD
  42 #include <asm/mce.h>
  43
  44 static bool notifier_registered;
  45 #endif
  46 static const char *RAS_FS_NAME = "ras";
  47
  48 const char *ras_error_string[] = {
  49         "none",
  50         "parity",
  51         "single_correctable",
  52         "multi_uncorrectable",
  53         "poison",
  54 };
  55
  56 const char *ras_block_string[] = {
  57         "umc",
  58         "sdma",
  59         "gfx",
  60         "mmhub",
  61         "athub",
  62         "pcie_bif",
  63         "hdp",
  64         "xgmi_wafl",
  65         "df",
  66         "smn",
  67         "sem",
  68         "mp0",
  69         "mp1",
  70         "fuse",
  71         "mca",
  72         "vcn",
  73         "jpeg",
  74 };
  75
  76 const char *ras_mca_block_string[] = {
  77         "mca_mp0",
  78         "mca_mp1",
  79         "mca_mpio",
  80         "mca_iohc",
  81 };
  82
  83 struct amdgpu_ras_block_list {
  84         /* ras block link */
  85         struct list_head node;
  86
  87         struct amdgpu_ras_block_object *ras_obj;
  88 };
  89
  90 const char *get_ras_block_str(struct ras_common_if *ras_block)
  91 {
  92         if (!ras_block)
  93                 return "NULL";
  94
  95         if (ras_block->block >= AMDGPU_RAS_BLOCK_COUNT)
  96                 return "OUT OF RANGE";
  97
  98         if (ras_block->block == AMDGPU_RAS_BLOCK__MCA)
  99                 return ras_mca_block_string[ras_block->sub_block_index];
 100
 101         return ras_block_string[ras_block->block];
 102 }
 103
 104 #define ras_block_str(_BLOCK_) \
 105         (((_BLOCK_) < ARRAY_SIZE(ras_block_string)) ? ras_block_string[_BLOCK_] : "Out Of Range")
 106
 107 #define ras_err_str(i) (ras_error_string[ffs(i)])
 108
 109 #define RAS_DEFAULT_FLAGS (AMDGPU_RAS_FLAG_INIT_BY_VBIOS)
 110
 111 /* inject address is 52 bits */
 112 #define RAS_UMC_INJECT_ADDR_LIMIT       (0x1ULL << 52)
 113
 114 /* typical ECC bad page rate is 1 bad page per 100MB VRAM */
 115 #define RAS_BAD_PAGE_COVER              (100 * 1024 * 1024ULL)
 116
 117 enum amdgpu_ras_retire_page_reservation {
 118         AMDGPU_RAS_RETIRE_PAGE_RESERVED,
 119         AMDGPU_RAS_RETIRE_PAGE_PENDING,
 120         AMDGPU_RAS_RETIRE_PAGE_FAULT,
 121 };
 122
 123 atomic_t amdgpu_ras_in_intr = ATOMIC_INIT(0);
 124
 125 static bool amdgpu_ras_check_bad_page_unlock(struct amdgpu_ras *con,
 126                                 uint64_t addr);
 127 static bool amdgpu_ras_check_bad_page(struct amdgpu_device *adev,
 128                                 uint64_t addr);
 129 #ifdef CONFIG_X86_MCE_AMD
 130 static void amdgpu_register_bad_pages_mca_notifier(struct amdgpu_device *adev);
 131 struct mce_notifier_adev_list {
 132         struct amdgpu_device *devs[MAX_GPU_INSTANCE];
 133         int num_gpu;
 134 };
 135 static struct mce_notifier_adev_list mce_adev_list;
 136 #endif
 137
 138 void amdgpu_ras_set_error_query_ready(struct amdgpu_device *adev, bool ready)
 139 {
 140         if (adev && amdgpu_ras_get_context(adev))
 141                 amdgpu_ras_get_context(adev)->error_query_ready = ready;
 142 }
 143
 144 static bool amdgpu_ras_get_error_query_ready(struct amdgpu_device *adev)
 145 {
 146         if (adev && amdgpu_ras_get_context(adev))
 147                 return amdgpu_ras_get_context(adev)->error_query_ready;
 148
 149         return false;
 150 }
 151
 152 static int amdgpu_reserve_page_direct(struct amdgpu_device *adev, uint64_t address)
 153 {
 154         struct ras_err_data err_data = {0, 0, 0, NULL};
 155         struct eeprom_table_record err_rec;
 156
 157         if ((address >= adev->gmc.mc_vram_size) ||
 158             (address >= RAS_UMC_INJECT_ADDR_LIMIT)) {
 159                 dev_warn(adev->dev,
 160                          "RAS WARN: input address 0x%llx is invalid.\n",
 161                          address);
 162                 return -EINVAL;
 163         }
 164
 165         if (amdgpu_ras_check_bad_page(adev, address)) {
 166                 dev_warn(adev->dev,
 167                          "RAS WARN: 0x%llx has already been marked as bad page!\n",
 168                          address);
 169                 return 0;
 170         }
 171
 172         memset(&err_rec, 0x0, sizeof(struct eeprom_table_record));
 173         err_data.err_addr = &err_rec;
 174         amdgpu_umc_fill_error_record(&err_data, address,
 175                         (address >> AMDGPU_GPU_PAGE_SHIFT), 0, 0);
 176
 177         if (amdgpu_bad_page_threshold != 0) {
 178                 amdgpu_ras_add_bad_pages(adev, err_data.err_addr,
 179                                          err_data.err_addr_cnt);
 180                 amdgpu_ras_save_bad_pages(adev, NULL);
 181         }
 182
 183         dev_warn(adev->dev, "WARNING: THIS IS ONLY FOR TEST PURPOSES AND WILL CORRUPT RAS EEPROM\n");
 184         dev_warn(adev->dev, "Clear EEPROM:\n");
 185         dev_warn(adev->dev, "    echo 1 > /sys/kernel/debug/dri/0/ras/ras_eeprom_reset\n");
 186
 187         return 0;
 188 }
 189
 190 static ssize_t amdgpu_ras_debugfs_read(struct file *f, char __user *buf,
 191                                         size_t size, loff_t *pos)
 192 {
 193         struct ras_manager *obj = (struct ras_manager *)file_inode(f)->i_private;
 194         struct ras_query_if info = {
 195                 .head = obj->head,
 196         };
 197         ssize_t s;
 198         char val[128];
 199
 200         if (amdgpu_ras_query_error_status(obj->adev, &info))
 201                 return -EINVAL;
 202
 203         /* Hardware counter will be reset automatically after the query on Vega20 and Arcturus */
 204         if (obj->adev->ip_versions[MP0_HWIP][0] != IP_VERSION(11, 0, 2) &&
 205             obj->adev->ip_versions[MP0_HWIP][0] != IP_VERSION(11, 0, 4)) {
 206                 if (amdgpu_ras_reset_error_status(obj->adev, info.head.block))
 207                         dev_warn(obj->adev->dev, "Failed to reset error counter and error status");
 208         }
 209
 210         s = snprintf(val, sizeof(val), "%s: %lu\n%s: %lu\n",
 211                         "ue", info.ue_count,
 212                         "ce", info.ce_count);
 213         if (*pos >= s)
 214                 return 0;
 215
 216         s -= *pos;
 217         s = min_t(u64, s, size);
 218
 219
 220         if (copy_to_user(buf, &val[*pos], s))
 221                 return -EINVAL;
 222
 223         *pos += s;
 224
 225         return s;
 226 }
 227
 228 static const struct file_operations amdgpu_ras_debugfs_ops = {
 229         .owner = THIS_MODULE,
 230         .read = amdgpu_ras_debugfs_read,
 231         .write = NULL,
 232         .llseek = default_llseek
 233 };
 234
 235 static int amdgpu_ras_find_block_id_by_name(const char *name, int *block_id)
 236 {
 237         int i;
 238
 239         for (i = 0; i < ARRAY_SIZE(ras_block_string); i++) {
 240                 *block_id = i;
 241                 if (strcmp(name, ras_block_string[i]) == 0)
 242                         return 0;
 243         }
 244         return -EINVAL;
 245 }
 246
 247 static int amdgpu_ras_debugfs_ctrl_parse_data(struct file *f,
 248                 const char __user *buf, size_t size,
 249                 loff_t *pos, struct ras_debug_if *data)
 250 {
 251         ssize_t s = min_t(u64, 64, size);
 252         char str[65];
 253         char block_name[33];
 254         char err[9] = "ue";
 255         int op = -1;
 256         int block_id;
 257         uint32_t sub_block;
 258         u64 address, value;
 259         /* default value is 0 if the mask is not set by user */
 260         u32 instance_mask = 0;
 261
 262         if (*pos)
 263                 return -EINVAL;
 264         *pos = size;
 265
 266         memset(str, 0, sizeof(str));
 267         memset(data, 0, sizeof(*data));
 268
 269         if (copy_from_user(str, buf, s))
 270                 return -EINVAL;
 271
 272         if (sscanf(str, "disable %32s", block_name) == 1)
 273                 op = 0;
 274         else if (sscanf(str, "enable %32s %8s", block_name, err) == 2)
 275                 op = 1;
 276         else if (sscanf(str, "inject %32s %8s", block_name, err) == 2)
 277                 op = 2;
 278         else if (strstr(str, "retire_page") != NULL)
 279                 op = 3;
 280         else if (str[0] && str[1] && str[2] && str[3])
 281                 /* ascii string, but commands are not matched. */
 282                 return -EINVAL;
 283
 284         if (op != -1) {
 285                 if (op == 3) {
 286                         if (sscanf(str, "%*s 0x%llx", &address) != 1 &&
 287                             sscanf(str, "%*s %llu", &address) != 1)
 288                                 return -EINVAL;
 289
 290                         data->op = op;
 291                         data->inject.address = address;
 292
 293                         return 0;
 294                 }
 295
 296                 if (amdgpu_ras_find_block_id_by_name(block_name, &block_id))
 297                         return -EINVAL;
 298
 299                 data->head.block = block_id;
 300                 /* only ue and ce errors are supported */
 301                 if (!memcmp("ue", err, 2))
 302                         data->head.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE;
 303                 else if (!memcmp("ce", err, 2))
 304                         data->head.type = AMDGPU_RAS_ERROR__SINGLE_CORRECTABLE;
 305                 else
 306                         return -EINVAL;
 307
 308                 data->op = op;
 309
 310                 if (op == 2) {
 311                         if (sscanf(str, "%*s %*s %*s 0x%x 0x%llx 0x%llx 0x%x",
 312                                    &sub_block, &address, &value, &instance_mask) != 4 &&
 313                             sscanf(str, "%*s %*s %*s %u %llu %llu %u",
 314                                    &sub_block, &address, &value, &instance_mask) != 4 &&
 315                                 sscanf(str, "%*s %*s %*s 0x%x 0x%llx 0x%llx",
 316                                    &sub_block, &address, &value) != 3 &&
 317                             sscanf(str, "%*s %*s %*s %u %llu %llu",
 318                                    &sub_block, &address, &value) != 3)
 319                                 return -EINVAL;
 320                         data->head.sub_block_index = sub_block;
 321                         data->inject.address = address;
 322                         data->inject.value = value;
 323                         data->inject.instance_mask = instance_mask;
 324                 }
 325         } else {
 326                 if (size < sizeof(*data))
 327                         return -EINVAL;
 328
 329                 if (copy_from_user(data, buf, sizeof(*data)))
 330                         return -EINVAL;
 331         }
 332
 333         return 0;
 334 }
 335
 336 static void amdgpu_ras_instance_mask_check(struct amdgpu_device *adev,
 337                                 struct ras_debug_if *data)
 338 {
 339         int num_xcc = adev->gfx.xcc_mask ? NUM_XCC(adev->gfx.xcc_mask) : 1;
 340         uint32_t mask, inst_mask = data->inject.instance_mask;
 341
 342         /* no need to set instance mask if there is only one instance */
 343         if (num_xcc <= 1 && inst_mask) {
 344                 data->inject.instance_mask = 0;
 345                 dev_dbg(adev->dev,
 346                         "RAS inject mask(0x%x) isn't supported and force it to 0.\n",
 347                         inst_mask);
 348
 349                 return;
 350         }
 351
 352         switch (data->head.block) {
 353         case AMDGPU_RAS_BLOCK__GFX:
 354                 mask = GENMASK(num_xcc - 1, 0);
 355                 break;
 356         case AMDGPU_RAS_BLOCK__SDMA:
 357                 mask = GENMASK(adev->sdma.num_instances - 1, 0);
 358                 break;
 359         case AMDGPU_RAS_BLOCK__VCN:
 360         case AMDGPU_RAS_BLOCK__JPEG:
 361                 mask = GENMASK(adev->vcn.num_vcn_inst - 1, 0);
 362                 break;
 363         default:
 364                 mask = inst_mask;
 365                 break;
 366         }
 367
 368         /* remove invalid bits in instance mask */
 369         data->inject.instance_mask &= mask;
 370         if (inst_mask != data->inject.instance_mask)
 371                 dev_dbg(adev->dev,
 372                         "Adjust RAS inject mask 0x%x to 0x%x\n",
 373                         inst_mask, data->inject.instance_mask);
 374 }
 375
 376 /**
 377  * DOC: AMDGPU RAS debugfs control interface
 378  *
 379  * The control interface accepts struct ras_debug_if which has two members.
 380  *
 381  * First member: ras_debug_if::head or ras_debug_if::inject.
 382  *
 383  * head is used to indicate which IP block will be under control.
 384  *
 385  * head has four members, they are block, type, sub_block_index, name.
 386  * block: which IP will be under control.
 387  * type: what kind of error will be enabled/disabled/injected.
 388  * sub_block_index: some IPs have subcomponets. say, GFX, sDMA.
 389  * name: the name of IP.
 390  *
 391  * inject has three more members than head, they are address, value and mask.
 392  * As their names indicate, inject operation will write the
 393  * value to the address.
 394  *
 395  * The second member: struct ras_debug_if::op.
 396  * It has three kinds of operations.
 397  *
 398  * - 0: disable RAS on the block. Take ::head as its data.
 399  * - 1: enable RAS on the block. Take ::head as its data.
 400  * - 2: inject errors on the block. Take ::inject as its data.
 401  *
 402  * How to use the interface?
 403  *
 404  * In a program
 405  *
 406  * Copy the struct ras_debug_if in your code and initialize it.
 407  * Write the struct to the control interface.
 408  *
 409  * From shell
 410  *
 411  * .. code-block:: bash
 412  *
 413  *      echo "disable <block>" > /sys/kernel/debug/dri/<N>/ras/ras_ctrl
 414  *      echo "enable  <block> <error>" > /sys/kernel/debug/dri/<N>/ras/ras_ctrl
 415  *      echo "inject  <block> <error> <sub-block> <address> <value> <mask>" > /sys/kernel/debug/dri/<N>/ras/ras_ctrl
 416  *
 417  * Where N, is the card which you want to affect.
 418  *
 419  * "disable" requires only the block.
 420  * "enable" requires the block and error type.
 421  * "inject" requires the block, error type, address, and value.
 422  *
 423  * The block is one of: umc, sdma, gfx, etc.
 424  *      see ras_block_string[] for details
 425  *
 426  * The error type is one of: ue, ce, where,
 427  *      ue is multi-uncorrectable
 428  *      ce is single-correctable
 429  *
 430  * The sub-block is a the sub-block index, pass 0 if there is no sub-block.
 431  * The address and value are hexadecimal numbers, leading 0x is optional.
 432  * The mask means instance mask, is optional, default value is 0x1.
 433  *
 434  * For instance,
 435  *
 436  * .. code-block:: bash
 437  *
 438  *      echo inject umc ue 0x0 0x0 0x0 > /sys/kernel/debug/dri/0/ras/ras_ctrl
 439  *      echo inject umc ce 0 0 0 3 > /sys/kernel/debug/dri/0/ras/ras_ctrl
 440  *      echo disable umc > /sys/kernel/debug/dri/0/ras/ras_ctrl
 441  *
 442  * How to check the result of the operation?
 443  *
 444  * To check disable/enable, see "ras" features at,
 445  * /sys/class/drm/card[0/1/2...]/device/ras/features
 446  *
 447  * To check inject, see the corresponding error count at,
 448  * /sys/class/drm/card[0/1/2...]/device/ras/[gfx|sdma|umc|...]_err_count
 449  *
 450  * .. note::
 451  *      Operations are only allowed on blocks which are supported.
 452  *      Check the "ras" mask at /sys/module/amdgpu/parameters/ras_mask
 453  *      to see which blocks support RAS on a particular asic.
 454  *
 455  */
 456 static ssize_t amdgpu_ras_debugfs_ctrl_write(struct file *f,
 457                                              const char __user *buf,
 458                                              size_t size, loff_t *pos)
 459 {
 460         struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
 461         struct ras_debug_if data;
 462         int ret = 0;
 463
 464         if (!amdgpu_ras_get_error_query_ready(adev)) {
 465                 dev_warn(adev->dev, "RAS WARN: error injection "
 466                                 "currently inaccessible\n");
 467                 return size;
 468         }
 469
 470         ret = amdgpu_ras_debugfs_ctrl_parse_data(f, buf, size, pos, &data);
 471         if (ret)
 472                 return ret;
 473
 474         if (data.op == 3) {
 475                 ret = amdgpu_reserve_page_direct(adev, data.inject.address);
 476                 if (!ret)
 477                         return size;
 478                 else
 479                         return ret;
 480         }
 481
 482         if (!amdgpu_ras_is_supported(adev, data.head.block))
 483                 return -EINVAL;
 484
 485         switch (data.op) {
 486         case 0:
 487                 ret = amdgpu_ras_feature_enable(adev, &data.head, 0);
 488                 break;
 489         case 1:
 490                 ret = amdgpu_ras_feature_enable(adev, &data.head, 1);
 491                 break;
 492         case 2:
 493                 if ((data.inject.address >= adev->gmc.mc_vram_size) ||
 494                     (data.inject.address >= RAS_UMC_INJECT_ADDR_LIMIT)) {
 495                         dev_warn(adev->dev, "RAS WARN: input address "
 496                                         "0x%llx is invalid.",
 497                                         data.inject.address);
 498                         ret = -EINVAL;
 499                         break;
 500                 }
 501
 502                 /* umc ce/ue error injection for a bad page is not allowed */
 503                 if ((data.head.block == AMDGPU_RAS_BLOCK__UMC) &&
 504                     amdgpu_ras_check_bad_page(adev, data.inject.address)) {
 505                         dev_warn(adev->dev, "RAS WARN: inject: 0x%llx has "
 506                                  "already been marked as bad!\n",
 507                                  data.inject.address);
 508                         break;
 509                 }
 510
 511                 amdgpu_ras_instance_mask_check(adev, &data);
 512
 513                 /* data.inject.address is offset instead of absolute gpu address */
 514                 ret = amdgpu_ras_error_inject(adev, &data.inject);
 515                 break;
 516         default:
 517                 ret = -EINVAL;
 518                 break;
 519         }
 520
 521         if (ret)
 522                 return ret;
 523
 524         return size;
 525 }
 526
 527 /**
 528  * DOC: AMDGPU RAS debugfs EEPROM table reset interface
 529  *
 530  * Some boards contain an EEPROM which is used to persistently store a list of
 531  * bad pages which experiences ECC errors in vram.  This interface provides
 532  * a way to reset the EEPROM, e.g., after testing error injection.
 533  *
 534  * Usage:
 535  *
 536  * .. code-block:: bash
 537  *
 538  *      echo 1 > ../ras/ras_eeprom_reset
 539  *
 540  * will reset EEPROM table to 0 entries.
 541  *
 542  */
 543 static ssize_t amdgpu_ras_debugfs_eeprom_write(struct file *f,
 544                                                const char __user *buf,
 545                                                size_t size, loff_t *pos)
 546 {
 547         struct amdgpu_device *adev =
 548                 (struct amdgpu_device *)file_inode(f)->i_private;
 549         int ret;
 550
 551         ret = amdgpu_ras_eeprom_reset_table(
 552                 &(amdgpu_ras_get_context(adev)->eeprom_control));
 553
 554         if (!ret) {
 555                 /* Something was written to EEPROM.
 556                  */
 557                 amdgpu_ras_get_context(adev)->flags = RAS_DEFAULT_FLAGS;
 558                 return size;
 559         } else {
 560                 return ret;
 561         }
 562 }
 563
 564 static const struct file_operations amdgpu_ras_debugfs_ctrl_ops = {
 565         .owner = THIS_MODULE,
 566         .read = NULL,
 567         .write = amdgpu_ras_debugfs_ctrl_write,
 568         .llseek = default_llseek
 569 };
 570
 571 static const struct file_operations amdgpu_ras_debugfs_eeprom_ops = {
 572         .owner = THIS_MODULE,
 573         .read = NULL,
 574         .write = amdgpu_ras_debugfs_eeprom_write,
 575         .llseek = default_llseek
 576 };
 577
 578 /**
 579  * DOC: AMDGPU RAS sysfs Error Count Interface
 580  *
 581  * It allows the user to read the error count for each IP block on the gpu through
 582  * /sys/class/drm/card[0/1/2...]/device/ras/[gfx/sdma/...]_err_count
 583  *
 584  * It outputs the multiple lines which report the uncorrected (ue) and corrected
 585  * (ce) error counts.
 586  *
 587  * The format of one line is below,
 588  *
 589  * [ce|ue]: count
 590  *
 591  * Example:
 592  *
 593  * .. code-block:: bash
 594  *
 595  *      ue: 0
 596  *      ce: 1
 597  *
 598  */
 599 static ssize_t amdgpu_ras_sysfs_read(struct device *dev,
 600                 struct device_attribute *attr, char *buf)
 601 {
 602         struct ras_manager *obj = container_of(attr, struct ras_manager, sysfs_attr);
 603         struct ras_query_if info = {
 604                 .head = obj->head,
 605         };
 606
 607         if (!amdgpu_ras_get_error_query_ready(obj->adev))
 608                 return sysfs_emit(buf, "Query currently inaccessible\n");
 609
 610         if (amdgpu_ras_query_error_status(obj->adev, &info))
 611                 return -EINVAL;
 612
 613         if (obj->adev->ip_versions[MP0_HWIP][0] != IP_VERSION(11, 0, 2) &&
 614             obj->adev->ip_versions[MP0_HWIP][0] != IP_VERSION(11, 0, 4)) {
 615                 if (amdgpu_ras_reset_error_status(obj->adev, info.head.block))
 616                         dev_warn(obj->adev->dev, "Failed to reset error counter and error status");
 617         }
 618
 619         return sysfs_emit(buf, "%s: %lu\n%s: %lu\n", "ue", info.ue_count,
 620                           "ce", info.ce_count);
 621 }
 622
 623 /* obj begin */
 624
 625 #define get_obj(obj) do { (obj)->use++; } while (0)
 626 #define alive_obj(obj) ((obj)->use)
 627
 628 static inline void put_obj(struct ras_manager *obj)
 629 {
 630         if (obj && (--obj->use == 0))
 631                 list_del(&obj->node);
 632         if (obj && (obj->use < 0))
 633                 DRM_ERROR("RAS ERROR: Unbalance obj(%s) use\n", get_ras_block_str(&obj->head));
 634 }
 635
 636 /* make one obj and return it. */
 637 static struct ras_manager *amdgpu_ras_create_obj(struct amdgpu_device *adev,
 638                 struct ras_common_if *head)
 639 {
 640         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
 641         struct ras_manager *obj;
 642
 643         if (!adev->ras_enabled || !con)
 644                 return NULL;
 645
 646         if (head->block >= AMDGPU_RAS_BLOCK_COUNT)
 647                 return NULL;
 648
 649         if (head->block == AMDGPU_RAS_BLOCK__MCA) {
 650                 if (head->sub_block_index >= AMDGPU_RAS_MCA_BLOCK__LAST)
 651                         return NULL;
 652
 653                 obj = &con->objs[AMDGPU_RAS_BLOCK__LAST + head->sub_block_index];
 654         } else
 655                 obj = &con->objs[head->block];
 656
 657         /* already exist. return obj? */
 658         if (alive_obj(obj))
 659                 return NULL;
 660
 661         obj->head = *head;
 662         obj->adev = adev;
 663         list_add(&obj->node, &con->head);
 664         get_obj(obj);
 665
 666         return obj;
 667 }
 668
 669 /* return an obj equal to head, or the first when head is NULL */
 670 struct ras_manager *amdgpu_ras_find_obj(struct amdgpu_device *adev,
 671                 struct ras_common_if *head)
 672 {
 673         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
 674         struct ras_manager *obj;
 675         int i;
 676
 677         if (!adev->ras_enabled || !con)
 678                 return NULL;
 679
 680         if (head) {
 681                 if (head->block >= AMDGPU_RAS_BLOCK_COUNT)
 682                         return NULL;
 683
 684                 if (head->block == AMDGPU_RAS_BLOCK__MCA) {
 685                         if (head->sub_block_index >= AMDGPU_RAS_MCA_BLOCK__LAST)
 686                                 return NULL;
 687
 688                         obj = &con->objs[AMDGPU_RAS_BLOCK__LAST + head->sub_block_index];
 689                 } else
 690                         obj = &con->objs[head->block];
 691
 692                 if (alive_obj(obj))
 693                         return obj;
 694         } else {
 695                 for (i = 0; i < AMDGPU_RAS_BLOCK_COUNT + AMDGPU_RAS_MCA_BLOCK_COUNT; i++) {
 696                         obj = &con->objs[i];
 697                         if (alive_obj(obj))
 698                                 return obj;
 699                 }
 700         }
 701
 702         return NULL;
 703 }
 704 /* obj end */
 705
 706 /* feature ctl begin */
 707 static int amdgpu_ras_is_feature_allowed(struct amdgpu_device *adev,
 708                                          struct ras_common_if *head)
 709 {
 710         return adev->ras_hw_enabled & BIT(head->block);
 711 }
 712
 713 static int amdgpu_ras_is_feature_enabled(struct amdgpu_device *adev,
 714                 struct ras_common_if *head)
 715 {
 716         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
 717
 718         return con->features & BIT(head->block);
 719 }
 720
 721 /*
 722  * if obj is not created, then create one.
 723  * set feature enable flag.
 724  */
 725 static int __amdgpu_ras_feature_enable(struct amdgpu_device *adev,
 726                 struct ras_common_if *head, int enable)
 727 {
 728         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
 729         struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
 730
 731         /* If hardware does not support ras, then do not create obj.
 732          * But if hardware support ras, we can create the obj.
 733          * Ras framework checks con->hw_supported to see if it need do
 734          * corresponding initialization.
 735          * IP checks con->support to see if it need disable ras.
 736          */
 737         if (!amdgpu_ras_is_feature_allowed(adev, head))
 738                 return 0;
 739
 740         if (enable) {
 741                 if (!obj) {
 742                         obj = amdgpu_ras_create_obj(adev, head);
 743                         if (!obj)
 744                                 return -EINVAL;
 745                 } else {
 746                         /* In case we create obj somewhere else */
 747                         get_obj(obj);
 748                 }
 749                 con->features |= BIT(head->block);
 750         } else {
 751                 if (obj && amdgpu_ras_is_feature_enabled(adev, head)) {
 752                         con->features &= ~BIT(head->block);
 753                         put_obj(obj);
 754                 }
 755         }
 756
 757         return 0;
 758 }
 759
 760 static int amdgpu_ras_check_feature_allowed(struct amdgpu_device *adev,
 761                 struct ras_common_if *head)
 762 {
 763         if (amdgpu_ras_is_feature_allowed(adev, head) ||
 764                 amdgpu_ras_is_poison_mode_supported(adev))
 765                 return 1;
 766         else
 767                 return 0;
 768 }
 769
 770 /* wrapper of psp_ras_enable_features */
 771 int amdgpu_ras_feature_enable(struct amdgpu_device *adev,
 772                 struct ras_common_if *head, bool enable)
 773 {
 774         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
 775         union ta_ras_cmd_input *info;
 776         int ret = 0;
 777
 778         if (!con)
 779                 return -EINVAL;
 780
 781         if (head->block == AMDGPU_RAS_BLOCK__GFX) {
 782                 info = kzalloc(sizeof(union ta_ras_cmd_input), GFP_KERNEL);
 783                 if (!info)
 784                         return -ENOMEM;
 785
 786                 if (!enable) {
 787                         info->disable_features = (struct ta_ras_disable_features_input) {
 788                                 .block_id =  amdgpu_ras_block_to_ta(head->block),
 789                                 .error_type = amdgpu_ras_error_to_ta(head->type),
 790                         };
 791                 } else {
 792                         info->enable_features = (struct ta_ras_enable_features_input) {
 793                                 .block_id =  amdgpu_ras_block_to_ta(head->block),
 794                                 .error_type = amdgpu_ras_error_to_ta(head->type),
 795                         };
 796                 }
 797         }
 798
 799         /* Do not enable if it is not allowed. */
 800         if (enable && !amdgpu_ras_check_feature_allowed(adev, head))
 801                 goto out;
 802
 803         /* Only enable ras feature operation handle on host side */
 804         if (head->block == AMDGPU_RAS_BLOCK__GFX &&
 805                 !amdgpu_sriov_vf(adev) &&
 806                 !amdgpu_ras_intr_triggered()) {
 807                 ret = psp_ras_enable_features(&adev->psp, info, enable);
 808                 if (ret) {
 809                         dev_err(adev->dev, "ras %s %s failed poison:%d ret:%d\n",
 810                                 enable ? "enable":"disable",
 811                                 get_ras_block_str(head),
 812                                 amdgpu_ras_is_poison_mode_supported(adev), ret);
 813                         goto out;
 814                 }
 815         }
 816
 817         /* setup the obj */
 818         __amdgpu_ras_feature_enable(adev, head, enable);
 819 out:
 820         if (head->block == AMDGPU_RAS_BLOCK__GFX)
 821                 kfree(info);
 822         return ret;
 823 }
 824
 825 /* Only used in device probe stage and called only once. */
 826 int amdgpu_ras_feature_enable_on_boot(struct amdgpu_device *adev,
 827                 struct ras_common_if *head, bool enable)
 828 {
 829         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
 830         int ret;
 831
 832         if (!con)
 833                 return -EINVAL;
 834
 835         if (con->flags & AMDGPU_RAS_FLAG_INIT_BY_VBIOS) {
 836                 if (enable) {
 837                         /* There is no harm to issue a ras TA cmd regardless of
 838                          * the currecnt ras state.
 839                          * If current state == target state, it will do nothing
 840                          * But sometimes it requests driver to reset and repost
 841                          * with error code -EAGAIN.
 842                          */
 843                         ret = amdgpu_ras_feature_enable(adev, head, 1);
 844                         /* With old ras TA, we might fail to enable ras.
 845                          * Log it and just setup the object.
 846                          * TODO need remove this WA in the future.
 847                          */
 848                         if (ret == -EINVAL) {
 849                                 ret = __amdgpu_ras_feature_enable(adev, head, 1);
 850                                 if (!ret)
 851                                         dev_info(adev->dev,
 852                                                 "RAS INFO: %s setup object\n",
 853                                                 get_ras_block_str(head));
 854                         }
 855                 } else {
 856                         /* setup the object then issue a ras TA disable cmd.*/
 857                         ret = __amdgpu_ras_feature_enable(adev, head, 1);
 858                         if (ret)
 859                                 return ret;
 860
 861                         /* gfx block ras dsiable cmd must send to ras-ta */
 862                         if (head->block == AMDGPU_RAS_BLOCK__GFX)
 863                                 con->features |= BIT(head->block);
 864
 865                         ret = amdgpu_ras_feature_enable(adev, head, 0);
 866
 867                         /* clean gfx block ras features flag */
 868                         if (adev->ras_enabled && head->block == AMDGPU_RAS_BLOCK__GFX)
 869                                 con->features &= ~BIT(head->block);
 870                 }
 871         } else
 872                 ret = amdgpu_ras_feature_enable(adev, head, enable);
 873
 874         return ret;
 875 }
 876
 877 static int amdgpu_ras_disable_all_features(struct amdgpu_device *adev,
 878                 bool bypass)
 879 {
 880         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
 881         struct ras_manager *obj, *tmp;
 882
 883         list_for_each_entry_safe(obj, tmp, &con->head, node) {
 884                 /* bypass psp.
 885                  * aka just release the obj and corresponding flags
 886                  */
 887                 if (bypass) {
 888                         if (__amdgpu_ras_feature_enable(adev, &obj->head, 0))
 889                                 break;
 890                 } else {
 891                         if (amdgpu_ras_feature_enable(adev, &obj->head, 0))
 892                                 break;
 893                 }
 894         }
 895
 896         return con->features;
 897 }
 898
 899 static int amdgpu_ras_enable_all_features(struct amdgpu_device *adev,
 900                 bool bypass)
 901 {
 902         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
 903         int i;
 904         const enum amdgpu_ras_error_type default_ras_type = AMDGPU_RAS_ERROR__NONE;
 905
 906         for (i = 0; i < AMDGPU_RAS_BLOCK_COUNT; i++) {
 907                 struct ras_common_if head = {
 908                         .block = i,
 909                         .type = default_ras_type,
 910                         .sub_block_index = 0,
 911                 };
 912
 913                 if (i == AMDGPU_RAS_BLOCK__MCA)
 914                         continue;
 915
 916                 if (bypass) {
 917                         /*
 918                          * bypass psp. vbios enable ras for us.
 919                          * so just create the obj
 920                          */
 921                         if (__amdgpu_ras_feature_enable(adev, &head, 1))
 922                                 break;
 923                 } else {
 924                         if (amdgpu_ras_feature_enable(adev, &head, 1))
 925                                 break;
 926                 }
 927         }
 928
 929         for (i = 0; i < AMDGPU_RAS_MCA_BLOCK_COUNT; i++) {
 930                 struct ras_common_if head = {
 931                         .block = AMDGPU_RAS_BLOCK__MCA,
 932                         .type = default_ras_type,
 933                         .sub_block_index = i,
 934                 };
 935
 936                 if (bypass) {
 937                         /*
 938                          * bypass psp. vbios enable ras for us.
 939                          * so just create the obj
 940                          */
 941                         if (__amdgpu_ras_feature_enable(adev, &head, 1))
 942                                 break;
 943                 } else {
 944                         if (amdgpu_ras_feature_enable(adev, &head, 1))
 945                                 break;
 946                 }
 947         }
 948
 949         return con->features;
 950 }
 951 /* feature ctl end */
 952
 953 static int amdgpu_ras_block_match_default(struct amdgpu_ras_block_object *block_obj,
 954                 enum amdgpu_ras_block block)
 955 {
 956         if (!block_obj)
 957                 return -EINVAL;
 958
 959         if (block_obj->ras_comm.block == block)
 960                 return 0;
 961
 962         return -EINVAL;
 963 }
 964
 965 static struct amdgpu_ras_block_object *amdgpu_ras_get_ras_block(struct amdgpu_device *adev,
 966                                         enum amdgpu_ras_block block, uint32_t sub_block_index)
 967 {
 968         struct amdgpu_ras_block_list *node, *tmp;
 969         struct amdgpu_ras_block_object *obj;
 970
 971         if (block >= AMDGPU_RAS_BLOCK__LAST)
 972                 return NULL;
 973
 974         list_for_each_entry_safe(node, tmp, &adev->ras_list, node) {
 975                 if (!node->ras_obj) {
 976                         dev_warn(adev->dev, "Warning: abnormal ras list node.\n");
 977                         continue;
 978                 }
 979
 980                 obj = node->ras_obj;
 981                 if (obj->ras_block_match) {
 982                         if (obj->ras_block_match(obj, block, sub_block_index) == 0)
 983                                 return obj;
 984                 } else {
 985                         if (amdgpu_ras_block_match_default(obj, block) == 0)
 986                                 return obj;
 987                 }
 988         }
 989
 990         return NULL;
 991 }
 992
 993 static void amdgpu_ras_get_ecc_info(struct amdgpu_device *adev, struct ras_err_data *err_data)
 994 {
 995         struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
 996         int ret = 0;
 997
 998         /*
 999          * choosing right query method according to
1000          * whether smu support query error information
1001          */
1002         ret = amdgpu_dpm_get_ecc_info(adev, (void *)&(ras->umc_ecc));
1003         if (ret == -EOPNOTSUPP) {
1004                 if (adev->umc.ras && adev->umc.ras->ras_block.hw_ops &&
1005                         adev->umc.ras->ras_block.hw_ops->query_ras_error_count)
1006                         adev->umc.ras->ras_block.hw_ops->query_ras_error_count(adev, err_data);
1007
1008                 /* umc query_ras_error_address is also responsible for clearing
1009                  * error status
1010                  */
1011                 if (adev->umc.ras && adev->umc.ras->ras_block.hw_ops &&
1012                     adev->umc.ras->ras_block.hw_ops->query_ras_error_address)
1013                         adev->umc.ras->ras_block.hw_ops->query_ras_error_address(adev, err_data);
1014         } else if (!ret) {
1015                 if (adev->umc.ras &&
1016                         adev->umc.ras->ecc_info_query_ras_error_count)
1017                         adev->umc.ras->ecc_info_query_ras_error_count(adev, err_data);
1018
1019                 if (adev->umc.ras &&
1020                         adev->umc.ras->ecc_info_query_ras_error_address)
1021                         adev->umc.ras->ecc_info_query_ras_error_address(adev, err_data);
1022         }
1023 }
1024
1025 /* query/inject/cure begin */
1026 int amdgpu_ras_query_error_status(struct amdgpu_device *adev,
1027                                   struct ras_query_if *info)
1028 {
1029         struct amdgpu_ras_block_object *block_obj = NULL;
1030         struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head);
1031         struct ras_err_data err_data = {0, 0, 0, NULL};
1032
1033         if (!obj)
1034                 return -EINVAL;
1035
1036         if (info->head.block == AMDGPU_RAS_BLOCK__UMC) {
1037                 amdgpu_ras_get_ecc_info(adev, &err_data);
1038         } else {
1039                 block_obj = amdgpu_ras_get_ras_block(adev, info->head.block, 0);
1040                 if (!block_obj || !block_obj->hw_ops)   {
1041                         dev_dbg_once(adev->dev, "%s doesn't config RAS function\n",
1042                                      get_ras_block_str(&info->head));
1043                         return -EINVAL;
1044                 }
1045
1046                 if (block_obj->hw_ops->query_ras_error_count)
1047                         block_obj->hw_ops->query_ras_error_count(adev, &err_data);
1048
1049                 if ((info->head.block == AMDGPU_RAS_BLOCK__SDMA) ||
1050                     (info->head.block == AMDGPU_RAS_BLOCK__GFX) ||
1051                     (info->head.block == AMDGPU_RAS_BLOCK__MMHUB)) {
1052                                 if (block_obj->hw_ops->query_ras_error_status)
1053                                         block_obj->hw_ops->query_ras_error_status(adev);
1054                         }
1055         }
1056
1057         obj->err_data.ue_count += err_data.ue_count;
1058         obj->err_data.ce_count += err_data.ce_count;
1059
1060         info->ue_count = obj->err_data.ue_count;
1061         info->ce_count = obj->err_data.ce_count;
1062
1063         if (err_data.ce_count) {
1064                 if (adev->smuio.funcs &&
1065                     adev->smuio.funcs->get_socket_id &&
1066                     adev->smuio.funcs->get_die_id) {
1067                         dev_info(adev->dev, "socket: %d, die: %d "
1068                                         "%ld correctable hardware errors "
1069                                         "detected in %s block, no user "
1070                                         "action is needed.\n",
1071                                         adev->smuio.funcs->get_socket_id(adev),
1072                                         adev->smuio.funcs->get_die_id(adev),
1073                                         obj->err_data.ce_count,
1074                                         get_ras_block_str(&info->head));
1075                 } else {
1076                         dev_info(adev->dev, "%ld correctable hardware errors "
1077                                         "detected in %s block, no user "
1078                                         "action is needed.\n",
1079                                         obj->err_data.ce_count,
1080                                         get_ras_block_str(&info->head));
1081                 }
1082         }
1083         if (err_data.ue_count) {
1084                 if (adev->smuio.funcs &&
1085                     adev->smuio.funcs->get_socket_id &&
1086                     adev->smuio.funcs->get_die_id) {
1087                         dev_info(adev->dev, "socket: %d, die: %d "
1088                                         "%ld uncorrectable hardware errors "
1089                                         "detected in %s block\n",
1090                                         adev->smuio.funcs->get_socket_id(adev),
1091                                         adev->smuio.funcs->get_die_id(adev),
1092                                         obj->err_data.ue_count,
1093                                         get_ras_block_str(&info->head));
1094                 } else {
1095                         dev_info(adev->dev, "%ld uncorrectable hardware errors "
1096                                         "detected in %s block\n",
1097                                         obj->err_data.ue_count,
1098                                         get_ras_block_str(&info->head));
1099                 }
1100         }
1101
1102         return 0;
1103 }
1104
1105 int amdgpu_ras_reset_error_status(struct amdgpu_device *adev,
1106                 enum amdgpu_ras_block block)
1107 {
1108         struct amdgpu_ras_block_object *block_obj = amdgpu_ras_get_ras_block(adev, block, 0);
1109
1110         if (!amdgpu_ras_is_supported(adev, block))
1111                 return -EINVAL;
1112
1113         if (!block_obj || !block_obj->hw_ops)   {
1114                 dev_dbg_once(adev->dev, "%s doesn't config RAS function\n",
1115                              ras_block_str(block));
1116                 return -EINVAL;
1117         }
1118
1119         if (block_obj->hw_ops->reset_ras_error_count)
1120                 block_obj->hw_ops->reset_ras_error_count(adev);
1121
1122         if ((block == AMDGPU_RAS_BLOCK__GFX) ||
1123             (block == AMDGPU_RAS_BLOCK__MMHUB)) {
1124                 if (block_obj->hw_ops->reset_ras_error_status)
1125                         block_obj->hw_ops->reset_ras_error_status(adev);
1126         }
1127
1128         return 0;
1129 }
1130
1131 /* wrapper of psp_ras_trigger_error */
1132 int amdgpu_ras_error_inject(struct amdgpu_device *adev,
1133                 struct ras_inject_if *info)
1134 {
1135         struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head);
1136         struct ta_ras_trigger_error_input block_info = {
1137                 .block_id =  amdgpu_ras_block_to_ta(info->head.block),
1138                 .inject_error_type = amdgpu_ras_error_to_ta(info->head.type),
1139                 .sub_block_index = info->head.sub_block_index,
1140                 .address = info->address,
1141                 .value = info->value,
1142         };
1143         int ret = -EINVAL;
1144         struct amdgpu_ras_block_object *block_obj = amdgpu_ras_get_ras_block(adev,
1145                                                         info->head.block,
1146                                                         info->head.sub_block_index);
1147
1148         /* inject on guest isn't allowed, return success directly */
1149         if (amdgpu_sriov_vf(adev))
1150                 return 0;
1151
1152         if (!obj)
1153                 return -EINVAL;
1154
1155         if (!block_obj || !block_obj->hw_ops)   {
1156                 dev_dbg_once(adev->dev, "%s doesn't config RAS function\n",
1157                              get_ras_block_str(&info->head));
1158                 return -EINVAL;
1159         }
1160
1161         /* Calculate XGMI relative offset */
1162         if (adev->gmc.xgmi.num_physical_nodes > 1) {
1163                 block_info.address =
1164                         amdgpu_xgmi_get_relative_phy_addr(adev,
1165                                                           block_info.address);
1166         }
1167
1168         if (block_obj->hw_ops->ras_error_inject) {
1169                 if (info->head.block == AMDGPU_RAS_BLOCK__GFX)
1170                         ret = block_obj->hw_ops->ras_error_inject(adev, info, info->instance_mask);
1171                 else /* Special ras_error_inject is defined (e.g: xgmi) */
1172                         ret = block_obj->hw_ops->ras_error_inject(adev, &block_info,
1173                                                 info->instance_mask);
1174         } else {
1175                 /* default path */
1176                 ret = psp_ras_trigger_error(&adev->psp, &block_info, info->instance_mask);
1177         }
1178
1179         if (ret)
1180                 dev_err(adev->dev, "ras inject %s failed %d\n",
1181                         get_ras_block_str(&info->head), ret);
1182
1183         return ret;
1184 }
1185
1186 /**
1187  * amdgpu_ras_query_error_count_helper -- Get error counter for specific IP
1188  * @adev: pointer to AMD GPU device
1189  * @ce_count: pointer to an integer to be set to the count of correctible errors.
1190  * @ue_count: pointer to an integer to be set to the count of uncorrectible errors.
1191  * @query_info: pointer to ras_query_if
1192  *
1193  * Return 0 for query success or do nothing, otherwise return an error
1194  * on failures
1195  */
1196 static int amdgpu_ras_query_error_count_helper(struct amdgpu_device *adev,
1197                                                unsigned long *ce_count,
1198                                                unsigned long *ue_count,
1199                                                struct ras_query_if *query_info)
1200 {
1201         int ret;
1202
1203         if (!query_info)
1204                 /* do nothing if query_info is not specified */
1205                 return 0;
1206
1207         ret = amdgpu_ras_query_error_status(adev, query_info);
1208         if (ret)
1209                 return ret;
1210
1211         *ce_count += query_info->ce_count;
1212         *ue_count += query_info->ue_count;
1213
1214         /* some hardware/IP supports read to clear
1215          * no need to explictly reset the err status after the query call */
1216         if (adev->ip_versions[MP0_HWIP][0] != IP_VERSION(11, 0, 2) &&
1217             adev->ip_versions[MP0_HWIP][0] != IP_VERSION(11, 0, 4)) {
1218                 if (amdgpu_ras_reset_error_status(adev, query_info->head.block))
1219                         dev_warn(adev->dev,
1220                                  "Failed to reset error counter and error status\n");
1221         }
1222
1223         return 0;
1224 }
1225
1226 /**
1227  * amdgpu_ras_query_error_count -- Get error counts of all IPs or specific IP
1228  * @adev: pointer to AMD GPU device
1229  * @ce_count: pointer to an integer to be set to the count of correctible errors.
1230  * @ue_count: pointer to an integer to be set to the count of uncorrectible
1231  * errors.
1232  * @query_info: pointer to ras_query_if if the query request is only for
1233  * specific ip block; if info is NULL, then the qurey request is for
1234  * all the ip blocks that support query ras error counters/status
1235  *
1236  * If set, @ce_count or @ue_count, count and return the corresponding
1237  * error counts in those integer pointers. Return 0 if the device
1238  * supports RAS. Return -EOPNOTSUPP if the device doesn't support RAS.
1239  */
1240 int amdgpu_ras_query_error_count(struct amdgpu_device *adev,
1241                                  unsigned long *ce_count,
1242                                  unsigned long *ue_count,
1243                                  struct ras_query_if *query_info)
1244 {
1245         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1246         struct ras_manager *obj;
1247         unsigned long ce, ue;
1248         int ret;
1249
1250         if (!adev->ras_enabled || !con)
1251                 return -EOPNOTSUPP;
1252
1253         /* Don't count since no reporting.
1254          */
1255         if (!ce_count && !ue_count)
1256                 return 0;
1257
1258         ce = 0;
1259         ue = 0;
1260         if (!query_info) {
1261                 /* query all the ip blocks that support ras query interface */
1262                 list_for_each_entry(obj, &con->head, node) {
1263                         struct ras_query_if info = {
1264                                 .head = obj->head,
1265                         };
1266
1267                         ret = amdgpu_ras_query_error_count_helper(adev, &ce, &ue, &info);
1268                 }
1269         } else {
1270                 /* query specific ip block */
1271                 ret = amdgpu_ras_query_error_count_helper(adev, &ce, &ue, query_info);
1272         }
1273
1274         if (ret)
1275                 return ret;
1276
1277         if (ce_count)
1278                 *ce_count = ce;
1279
1280         if (ue_count)
1281                 *ue_count = ue;
1282
1283         return 0;
1284 }
1285 /* query/inject/cure end */
1286
1287
1288 /* sysfs begin */
1289
1290 static int amdgpu_ras_badpages_read(struct amdgpu_device *adev,
1291                 struct ras_badpage **bps, unsigned int *count);
1292
1293 static char *amdgpu_ras_badpage_flags_str(unsigned int flags)
1294 {
1295         switch (flags) {
1296         case AMDGPU_RAS_RETIRE_PAGE_RESERVED:
1297                 return "R";
1298         case AMDGPU_RAS_RETIRE_PAGE_PENDING:
1299                 return "P";
1300         case AMDGPU_RAS_RETIRE_PAGE_FAULT:
1301         default:
1302                 return "F";
1303         }
1304 }
1305
1306 /**
1307  * DOC: AMDGPU RAS sysfs gpu_vram_bad_pages Interface
1308  *
1309  * It allows user to read the bad pages of vram on the gpu through
1310  * /sys/class/drm/card[0/1/2...]/device/ras/gpu_vram_bad_pages
1311  *
1312  * It outputs multiple lines, and each line stands for one gpu page.
1313  *
1314  * The format of one line is below,
1315  * gpu pfn : gpu page size : flags
1316  *
1317  * gpu pfn and gpu page size are printed in hex format.
1318  * flags can be one of below character,
1319  *
1320  * R: reserved, this gpu page is reserved and not able to use.
1321  *
1322  * P: pending for reserve, this gpu page is marked as bad, will be reserved
1323  * in next window of page_reserve.
1324  *
1325  * F: unable to reserve. this gpu page can't be reserved due to some reasons.
1326  *
1327  * Examples:
1328  *
1329  * .. code-block:: bash
1330  *
1331  *      0x00000001 : 0x00001000 : R
1332  *      0x00000002 : 0x00001000 : P
1333  *
1334  */
1335
1336 static ssize_t amdgpu_ras_sysfs_badpages_read(struct file *f,
1337                 struct kobject *kobj, struct bin_attribute *attr,
1338                 char *buf, loff_t ppos, size_t count)
1339 {
1340         struct amdgpu_ras *con =
1341                 container_of(attr, struct amdgpu_ras, badpages_attr);
1342         struct amdgpu_device *adev = con->adev;
1343         const unsigned int element_size =
1344                 sizeof("0xabcdabcd : 0x12345678 : R\n") - 1;
1345         unsigned int start = div64_ul(ppos + element_size - 1, element_size);
1346         unsigned int end = div64_ul(ppos + count - 1, element_size);
1347         ssize_t s = 0;
1348         struct ras_badpage *bps = NULL;
1349         unsigned int bps_count = 0;
1350
1351         memset(buf, 0, count);
1352
1353         if (amdgpu_ras_badpages_read(adev, &bps, &bps_count))
1354                 return 0;
1355
1356         for (; start < end && start < bps_count; start++)
1357                 s += scnprintf(&buf[s], element_size + 1,
1358                                 "0x%08x : 0x%08x : %1s\n",
1359                                 bps[start].bp,
1360                                 bps[start].size,
1361                                 amdgpu_ras_badpage_flags_str(bps[start].flags));
1362
1363         kfree(bps);
1364
1365         return s;
1366 }
1367
1368 static ssize_t amdgpu_ras_sysfs_features_read(struct device *dev,
1369                 struct device_attribute *attr, char *buf)
1370 {
1371         struct amdgpu_ras *con =
1372                 container_of(attr, struct amdgpu_ras, features_attr);
1373
1374         return sysfs_emit(buf, "feature mask: 0x%x\n", con->features);
1375 }
1376
1377 static void amdgpu_ras_sysfs_remove_bad_page_node(struct amdgpu_device *adev)
1378 {
1379         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1380
1381         sysfs_remove_file_from_group(&adev->dev->kobj,
1382                                 &con->badpages_attr.attr,
1383                                 RAS_FS_NAME);
1384 }
1385
1386 static int amdgpu_ras_sysfs_remove_feature_node(struct amdgpu_device *adev)
1387 {
1388         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1389         struct attribute *attrs[] = {
1390                 &con->features_attr.attr,
1391                 NULL
1392         };
1393         struct attribute_group group = {
1394                 .name = RAS_FS_NAME,
1395                 .attrs = attrs,
1396         };
1397
1398         sysfs_remove_group(&adev->dev->kobj, &group);
1399
1400         return 0;
1401 }
1402
1403 int amdgpu_ras_sysfs_create(struct amdgpu_device *adev,
1404                 struct ras_common_if *head)
1405 {
1406         struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
1407
1408         if (!obj || obj->attr_inuse)
1409                 return -EINVAL;
1410
1411         get_obj(obj);
1412
1413         snprintf(obj->fs_data.sysfs_name, sizeof(obj->fs_data.sysfs_name),
1414                 "%s_err_count", head->name);
1415
1416         obj->sysfs_attr = (struct device_attribute){
1417                 .attr = {
1418                         .name = obj->fs_data.sysfs_name,
1419                         .mode = S_IRUGO,
1420                 },
1421                         .show = amdgpu_ras_sysfs_read,
1422         };
1423         sysfs_attr_init(&obj->sysfs_attr.attr);
1424
1425         if (sysfs_add_file_to_group(&adev->dev->kobj,
1426                                 &obj->sysfs_attr.attr,
1427                                 RAS_FS_NAME)) {
1428                 put_obj(obj);
1429                 return -EINVAL;
1430         }
1431
1432         obj->attr_inuse = 1;
1433
1434         return 0;
1435 }
1436
1437 int amdgpu_ras_sysfs_remove(struct amdgpu_device *adev,
1438                 struct ras_common_if *head)
1439 {
1440         struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
1441
1442         if (!obj || !obj->attr_inuse)
1443                 return -EINVAL;
1444
1445         sysfs_remove_file_from_group(&adev->dev->kobj,
1446                                 &obj->sysfs_attr.attr,
1447                                 RAS_FS_NAME);
1448         obj->attr_inuse = 0;
1449         put_obj(obj);
1450
1451         return 0;
1452 }
1453
1454 static int amdgpu_ras_sysfs_remove_all(struct amdgpu_device *adev)
1455 {
1456         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1457         struct ras_manager *obj, *tmp;
1458
1459         list_for_each_entry_safe(obj, tmp, &con->head, node) {
1460                 amdgpu_ras_sysfs_remove(adev, &obj->head);
1461         }
1462
1463         if (amdgpu_bad_page_threshold != 0)
1464                 amdgpu_ras_sysfs_remove_bad_page_node(adev);
1465
1466         amdgpu_ras_sysfs_remove_feature_node(adev);
1467
1468         return 0;
1469 }
1470 /* sysfs end */
1471
1472 /**
1473  * DOC: AMDGPU RAS Reboot Behavior for Unrecoverable Errors
1474  *
1475  * Normally when there is an uncorrectable error, the driver will reset
1476  * the GPU to recover.  However, in the event of an unrecoverable error,
1477  * the driver provides an interface to reboot the system automatically
1478  * in that event.
1479  *
1480  * The following file in debugfs provides that interface:
1481  * /sys/kernel/debug/dri/[0/1/2...]/ras/auto_reboot
1482  *
1483  * Usage:
1484  *
1485  * .. code-block:: bash
1486  *
1487  *      echo true > .../ras/auto_reboot
1488  *
1489  */
1490 /* debugfs begin */
1491 static struct dentry *amdgpu_ras_debugfs_create_ctrl_node(struct amdgpu_device *adev)
1492 {
1493         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1494         struct amdgpu_ras_eeprom_control *eeprom = &con->eeprom_control;
1495         struct drm_minor  *minor = adev_to_drm(adev)->primary;
1496         struct dentry     *dir;
1497
1498         dir = debugfs_create_dir(RAS_FS_NAME, minor->debugfs_root);
1499         debugfs_create_file("ras_ctrl", S_IWUGO | S_IRUGO, dir, adev,
1500                             &amdgpu_ras_debugfs_ctrl_ops);
1501         debugfs_create_file("ras_eeprom_reset", S_IWUGO | S_IRUGO, dir, adev,
1502                             &amdgpu_ras_debugfs_eeprom_ops);
1503         debugfs_create_u32("bad_page_cnt_threshold", 0444, dir,
1504                            &con->bad_page_cnt_threshold);
1505         debugfs_create_u32("ras_num_recs", 0444, dir, &eeprom->ras_num_recs);
1506         debugfs_create_x32("ras_hw_enabled", 0444, dir, &adev->ras_hw_enabled);
1507         debugfs_create_x32("ras_enabled", 0444, dir, &adev->ras_enabled);
1508         debugfs_create_file("ras_eeprom_size", S_IRUGO, dir, adev,
1509                             &amdgpu_ras_debugfs_eeprom_size_ops);
1510         con->de_ras_eeprom_table = debugfs_create_file("ras_eeprom_table",
1511                                                        S_IRUGO, dir, adev,
1512                                                        &amdgpu_ras_debugfs_eeprom_table_ops);
1513         amdgpu_ras_debugfs_set_ret_size(&con->eeprom_control);
1514
1515         /*
1516          * After one uncorrectable error happens, usually GPU recovery will
1517          * be scheduled. But due to the known problem in GPU recovery failing
1518          * to bring GPU back, below interface provides one direct way to
1519          * user to reboot system automatically in such case within
1520          * ERREVENT_ATHUB_INTERRUPT generated. Normal GPU recovery routine
1521          * will never be called.
1522          */
1523         debugfs_create_bool("auto_reboot", S_IWUGO | S_IRUGO, dir, &con->reboot);
1524
1525         /*
1526          * User could set this not to clean up hardware's error count register
1527          * of RAS IPs during ras recovery.
1528          */
1529         debugfs_create_bool("disable_ras_err_cnt_harvest", 0644, dir,
1530                             &con->disable_ras_err_cnt_harvest);
1531         return dir;
1532 }
1533
1534 static void amdgpu_ras_debugfs_create(struct amdgpu_device *adev,
1535                                       struct ras_fs_if *head,
1536                                       struct dentry *dir)
1537 {
1538         struct ras_manager *obj = amdgpu_ras_find_obj(adev, &head->head);
1539
1540         if (!obj || !dir)
1541                 return;
1542
1543         get_obj(obj);
1544
1545         memcpy(obj->fs_data.debugfs_name,
1546                         head->debugfs_name,
1547                         sizeof(obj->fs_data.debugfs_name));
1548
1549         debugfs_create_file(obj->fs_data.debugfs_name, S_IWUGO | S_IRUGO, dir,
1550                             obj, &amdgpu_ras_debugfs_ops);
1551 }
1552
1553 void amdgpu_ras_debugfs_create_all(struct amdgpu_device *adev)
1554 {
1555         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1556         struct dentry *dir;
1557         struct ras_manager *obj;
1558         struct ras_fs_if fs_info;
1559
1560         /*
1561          * it won't be called in resume path, no need to check
1562          * suspend and gpu reset status
1563          */
1564         if (!IS_ENABLED(CONFIG_DEBUG_FS) || !con)
1565                 return;
1566
1567         dir = amdgpu_ras_debugfs_create_ctrl_node(adev);
1568
1569         list_for_each_entry(obj, &con->head, node) {
1570                 if (amdgpu_ras_is_supported(adev, obj->head.block) &&
1571                         (obj->attr_inuse == 1)) {
1572                         sprintf(fs_info.debugfs_name, "%s_err_inject",
1573                                         get_ras_block_str(&obj->head));
1574                         fs_info.head = obj->head;
1575                         amdgpu_ras_debugfs_create(adev, &fs_info, dir);
1576                 }
1577         }
1578 }
1579
1580 /* debugfs end */
1581
1582 /* ras fs */
1583 static BIN_ATTR(gpu_vram_bad_pages, S_IRUGO,
1584                 amdgpu_ras_sysfs_badpages_read, NULL, 0);
1585 static DEVICE_ATTR(features, S_IRUGO,
1586                 amdgpu_ras_sysfs_features_read, NULL);
1587 static int amdgpu_ras_fs_init(struct amdgpu_device *adev)
1588 {
1589         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1590         struct attribute_group group = {
1591                 .name = RAS_FS_NAME,
1592         };
1593         struct attribute *attrs[] = {
1594                 &con->features_attr.attr,
1595                 NULL
1596         };
1597         struct bin_attribute *bin_attrs[] = {
1598                 NULL,
1599                 NULL,
1600         };
1601         int r;
1602
1603         /* add features entry */
1604         con->features_attr = dev_attr_features;
1605         group.attrs = attrs;
1606         sysfs_attr_init(attrs[0]);
1607
1608         if (amdgpu_bad_page_threshold != 0) {
1609                 /* add bad_page_features entry */
1610                 bin_attr_gpu_vram_bad_pages.private = NULL;
1611                 con->badpages_attr = bin_attr_gpu_vram_bad_pages;
1612                 bin_attrs[0] = &con->badpages_attr;
1613                 group.bin_attrs = bin_attrs;
1614                 sysfs_bin_attr_init(bin_attrs[0]);
1615         }
1616
1617         r = sysfs_create_group(&adev->dev->kobj, &group);
1618         if (r)
1619                 dev_err(adev->dev, "Failed to create RAS sysfs group!");
1620
1621         return 0;
1622 }
1623
1624 static int amdgpu_ras_fs_fini(struct amdgpu_device *adev)
1625 {
1626         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1627         struct ras_manager *con_obj, *ip_obj, *tmp;
1628
1629         if (IS_ENABLED(CONFIG_DEBUG_FS)) {
1630                 list_for_each_entry_safe(con_obj, tmp, &con->head, node) {
1631                         ip_obj = amdgpu_ras_find_obj(adev, &con_obj->head);
1632                         if (ip_obj)
1633                                 put_obj(ip_obj);
1634                 }
1635         }
1636
1637         amdgpu_ras_sysfs_remove_all(adev);
1638         return 0;
1639 }
1640 /* ras fs end */
1641
1642 /* ih begin */
1643
1644 /* For the hardware that cannot enable bif ring for both ras_controller_irq
1645  * and ras_err_evnet_athub_irq ih cookies, the driver has to poll status
1646  * register to check whether the interrupt is triggered or not, and properly
1647  * ack the interrupt if it is there
1648  */
1649 void amdgpu_ras_interrupt_fatal_error_handler(struct amdgpu_device *adev)
1650 {
1651         /* Fatal error events are handled on host side */
1652         if (amdgpu_sriov_vf(adev))
1653                 return;
1654
1655         if (adev->nbio.ras &&
1656             adev->nbio.ras->handle_ras_controller_intr_no_bifring)
1657                 adev->nbio.ras->handle_ras_controller_intr_no_bifring(adev);
1658
1659         if (adev->nbio.ras &&
1660             adev->nbio.ras->handle_ras_err_event_athub_intr_no_bifring)
1661                 adev->nbio.ras->handle_ras_err_event_athub_intr_no_bifring(adev);
1662 }
1663
1664 static void amdgpu_ras_interrupt_poison_consumption_handler(struct ras_manager *obj,
1665                                 struct amdgpu_iv_entry *entry)
1666 {
1667         bool poison_stat = false;
1668         struct amdgpu_device *adev = obj->adev;
1669         struct amdgpu_ras_block_object *block_obj =
1670                 amdgpu_ras_get_ras_block(adev, obj->head.block, 0);
1671
1672         if (!block_obj)
1673                 return;
1674
1675         /* both query_poison_status and handle_poison_consumption are optional,
1676          * but at least one of them should be implemented if we need poison
1677          * consumption handler
1678          */
1679         if (block_obj->hw_ops && block_obj->hw_ops->query_poison_status) {
1680                 poison_stat = block_obj->hw_ops->query_poison_status(adev);
1681                 if (!poison_stat) {
1682                         /* Not poison consumption interrupt, no need to handle it */
1683                         dev_info(adev->dev, "No RAS poison status in %s poison IH.\n",
1684                                         block_obj->ras_comm.name);
1685
1686                         return;
1687                 }
1688         }
1689
1690         if (!adev->gmc.xgmi.connected_to_cpu)
1691                 amdgpu_umc_poison_handler(adev, false);
1692
1693         if (block_obj->hw_ops && block_obj->hw_ops->handle_poison_consumption)
1694                 poison_stat = block_obj->hw_ops->handle_poison_consumption(adev);
1695
1696         /* gpu reset is fallback for failed and default cases */
1697         if (poison_stat) {
1698                 dev_info(adev->dev, "GPU reset for %s RAS poison consumption is issued!\n",
1699                                 block_obj->ras_comm.name);
1700                 amdgpu_ras_reset_gpu(adev);
1701         } else {
1702                 amdgpu_gfx_poison_consumption_handler(adev, entry);
1703         }
1704 }
1705
1706 static void amdgpu_ras_interrupt_poison_creation_handler(struct ras_manager *obj,
1707                                 struct amdgpu_iv_entry *entry)
1708 {
1709         dev_info(obj->adev->dev,
1710                 "Poison is created, no user action is needed.\n");
1711 }
1712
1713 static void amdgpu_ras_interrupt_umc_handler(struct ras_manager *obj,
1714                                 struct amdgpu_iv_entry *entry)
1715 {
1716         struct ras_ih_data *data = &obj->ih_data;
1717         struct ras_err_data err_data = {0, 0, 0, NULL};
1718         int ret;
1719
1720         if (!data->cb)
1721                 return;
1722
1723         /* Let IP handle its data, maybe we need get the output
1724          * from the callback to update the error type/count, etc
1725          */
1726         ret = data->cb(obj->adev, &err_data, entry);
1727         /* ue will trigger an interrupt, and in that case
1728          * we need do a reset to recovery the whole system.
1729          * But leave IP do that recovery, here we just dispatch
1730          * the error.
1731          */
1732         if (ret == AMDGPU_RAS_SUCCESS) {
1733                 /* these counts could be left as 0 if
1734                  * some blocks do not count error number
1735                  */
1736                 obj->err_data.ue_count += err_data.ue_count;
1737                 obj->err_data.ce_count += err_data.ce_count;
1738         }
1739 }
1740
1741 static void amdgpu_ras_interrupt_handler(struct ras_manager *obj)
1742 {
1743         struct ras_ih_data *data = &obj->ih_data;
1744         struct amdgpu_iv_entry entry;
1745
1746         while (data->rptr != data->wptr) {
1747                 rmb();
1748                 memcpy(&entry, &data->ring[data->rptr],
1749                                 data->element_size);
1750
1751                 wmb();
1752                 data->rptr = (data->aligned_element_size +
1753                                 data->rptr) % data->ring_size;
1754
1755                 if (amdgpu_ras_is_poison_mode_supported(obj->adev)) {
1756                         if (obj->head.block == AMDGPU_RAS_BLOCK__UMC)
1757                                 amdgpu_ras_interrupt_poison_creation_handler(obj, &entry);
1758                         else
1759                                 amdgpu_ras_interrupt_poison_consumption_handler(obj, &entry);
1760                 } else {
1761                         if (obj->head.block == AMDGPU_RAS_BLOCK__UMC)
1762                                 amdgpu_ras_interrupt_umc_handler(obj, &entry);
1763                         else
1764                                 dev_warn(obj->adev->dev,
1765                                         "No RAS interrupt handler for non-UMC block with poison disabled.\n");
1766                 }
1767         }
1768 }
1769
1770 static void amdgpu_ras_interrupt_process_handler(struct work_struct *work)
1771 {
1772         struct ras_ih_data *data =
1773                 container_of(work, struct ras_ih_data, ih_work);
1774         struct ras_manager *obj =
1775                 container_of(data, struct ras_manager, ih_data);
1776
1777         amdgpu_ras_interrupt_handler(obj);
1778 }
1779
1780 int amdgpu_ras_interrupt_dispatch(struct amdgpu_device *adev,
1781                 struct ras_dispatch_if *info)
1782 {
1783         struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head);
1784         struct ras_ih_data *data = &obj->ih_data;
1785
1786         if (!obj)
1787                 return -EINVAL;
1788
1789         if (data->inuse == 0)
1790                 return 0;
1791
1792         /* Might be overflow... */
1793         memcpy(&data->ring[data->wptr], info->entry,
1794                         data->element_size);
1795
1796         wmb();
1797         data->wptr = (data->aligned_element_size +
1798                         data->wptr) % data->ring_size;
1799
1800         schedule_work(&data->ih_work);
1801
1802         return 0;
1803 }
1804
1805 int amdgpu_ras_interrupt_remove_handler(struct amdgpu_device *adev,
1806                 struct ras_common_if *head)
1807 {
1808         struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
1809         struct ras_ih_data *data;
1810
1811         if (!obj)
1812                 return -EINVAL;
1813
1814         data = &obj->ih_data;
1815         if (data->inuse == 0)
1816                 return 0;
1817
1818         cancel_work_sync(&data->ih_work);
1819
1820         kfree(data->ring);
1821         memset(data, 0, sizeof(*data));
1822         put_obj(obj);
1823
1824         return 0;
1825 }
1826
1827 int amdgpu_ras_interrupt_add_handler(struct amdgpu_device *adev,
1828                 struct ras_common_if *head)
1829 {
1830         struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
1831         struct ras_ih_data *data;
1832         struct amdgpu_ras_block_object *ras_obj;
1833
1834         if (!obj) {
1835                 /* in case we registe the IH before enable ras feature */
1836                 obj = amdgpu_ras_create_obj(adev, head);
1837                 if (!obj)
1838                         return -EINVAL;
1839         } else
1840                 get_obj(obj);
1841
1842         ras_obj = container_of(head, struct amdgpu_ras_block_object, ras_comm);
1843
1844         data = &obj->ih_data;
1845         /* add the callback.etc */
1846         *data = (struct ras_ih_data) {
1847                 .inuse = 0,
1848                 .cb = ras_obj->ras_cb,
1849                 .element_size = sizeof(struct amdgpu_iv_entry),
1850                 .rptr = 0,
1851                 .wptr = 0,
1852         };
1853
1854         INIT_WORK(&data->ih_work, amdgpu_ras_interrupt_process_handler);
1855
1856         data->aligned_element_size = ALIGN(data->element_size, 8);
1857         /* the ring can store 64 iv entries. */
1858         data->ring_size = 64 * data->aligned_element_size;
1859         data->ring = kmalloc(data->ring_size, GFP_KERNEL);
1860         if (!data->ring) {
1861                 put_obj(obj);
1862                 return -ENOMEM;
1863         }
1864
1865         /* IH is ready */
1866         data->inuse = 1;
1867
1868         return 0;
1869 }
1870
1871 static int amdgpu_ras_interrupt_remove_all(struct amdgpu_device *adev)
1872 {
1873         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1874         struct ras_manager *obj, *tmp;
1875
1876         list_for_each_entry_safe(obj, tmp, &con->head, node) {
1877                 amdgpu_ras_interrupt_remove_handler(adev, &obj->head);
1878         }
1879
1880         return 0;
1881 }
1882 /* ih end */
1883
1884 /* traversal all IPs except NBIO to query error counter */
1885 static void amdgpu_ras_log_on_err_counter(struct amdgpu_device *adev)
1886 {
1887         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1888         struct ras_manager *obj;
1889
1890         if (!adev->ras_enabled || !con)
1891                 return;
1892
1893         list_for_each_entry(obj, &con->head, node) {
1894                 struct ras_query_if info = {
1895                         .head = obj->head,
1896                 };
1897
1898                 /*
1899                  * PCIE_BIF IP has one different isr by ras controller
1900                  * interrupt, the specific ras counter query will be
1901                  * done in that isr. So skip such block from common
1902                  * sync flood interrupt isr calling.
1903                  */
1904                 if (info.head.block == AMDGPU_RAS_BLOCK__PCIE_BIF)
1905                         continue;
1906
1907                 /*
1908                  * this is a workaround for aldebaran, skip send msg to
1909                  * smu to get ecc_info table due to smu handle get ecc
1910                  * info table failed temporarily.
1911                  * should be removed until smu fix handle ecc_info table.
1912                  */
1913                 if ((info.head.block == AMDGPU_RAS_BLOCK__UMC) &&
1914                         (adev->ip_versions[MP1_HWIP][0] == IP_VERSION(13, 0, 2)))
1915                         continue;
1916
1917                 amdgpu_ras_query_error_status(adev, &info);
1918
1919                 if (adev->ip_versions[MP0_HWIP][0] != IP_VERSION(11, 0, 2) &&
1920                     adev->ip_versions[MP0_HWIP][0] != IP_VERSION(11, 0, 4) &&
1921                     adev->ip_versions[MP0_HWIP][0] != IP_VERSION(13, 0, 0)) {
1922                         if (amdgpu_ras_reset_error_status(adev, info.head.block))
1923                                 dev_warn(adev->dev, "Failed to reset error counter and error status");
1924                 }
1925         }
1926 }
1927
1928 /* Parse RdRspStatus and WrRspStatus */
1929 static void amdgpu_ras_error_status_query(struct amdgpu_device *adev,
1930                                           struct ras_query_if *info)
1931 {
1932         struct amdgpu_ras_block_object *block_obj;
1933         /*
1934          * Only two block need to query read/write
1935          * RspStatus at current state
1936          */
1937         if ((info->head.block != AMDGPU_RAS_BLOCK__GFX) &&
1938                 (info->head.block != AMDGPU_RAS_BLOCK__MMHUB))
1939                 return;
1940
1941         block_obj = amdgpu_ras_get_ras_block(adev,
1942                                         info->head.block,
1943                                         info->head.sub_block_index);
1944
1945         if (!block_obj || !block_obj->hw_ops) {
1946                 dev_dbg_once(adev->dev, "%s doesn't config RAS function\n",
1947                              get_ras_block_str(&info->head));
1948                 return;
1949         }
1950
1951         if (block_obj->hw_ops->query_ras_error_status)
1952                 block_obj->hw_ops->query_ras_error_status(adev);
1953
1954 }
1955
1956 static void amdgpu_ras_query_err_status(struct amdgpu_device *adev)
1957 {
1958         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1959         struct ras_manager *obj;
1960
1961         if (!adev->ras_enabled || !con)
1962                 return;
1963
1964         list_for_each_entry(obj, &con->head, node) {
1965                 struct ras_query_if info = {
1966                         .head = obj->head,
1967                 };
1968
1969                 amdgpu_ras_error_status_query(adev, &info);
1970         }
1971 }
1972
1973 /* recovery begin */
1974
1975 /* return 0 on success.
1976  * caller need free bps.
1977  */
1978 static int amdgpu_ras_badpages_read(struct amdgpu_device *adev,
1979                 struct ras_badpage **bps, unsigned int *count)
1980 {
1981         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1982         struct ras_err_handler_data *data;
1983         int i = 0;
1984         int ret = 0, status;
1985
1986         if (!con || !con->eh_data || !bps || !count)
1987                 return -EINVAL;
1988
1989         mutex_lock(&con->recovery_lock);
1990         data = con->eh_data;
1991         if (!data || data->count == 0) {
1992                 *bps = NULL;
1993                 ret = -EINVAL;
1994                 goto out;
1995         }
1996
1997         *bps = kmalloc(sizeof(struct ras_badpage) * data->count, GFP_KERNEL);
1998         if (!*bps) {
1999                 ret = -ENOMEM;
2000                 goto out;
2001         }
2002
2003         for (; i < data->count; i++) {
2004                 (*bps)[i] = (struct ras_badpage){
2005                         .bp = data->bps[i].retired_page,
2006                         .size = AMDGPU_GPU_PAGE_SIZE,
2007                         .flags = AMDGPU_RAS_RETIRE_PAGE_RESERVED,
2008                 };
2009                 status = amdgpu_vram_mgr_query_page_status(&adev->mman.vram_mgr,
2010                                 data->bps[i].retired_page);
2011                 if (status == -EBUSY)
2012                         (*bps)[i].flags = AMDGPU_RAS_RETIRE_PAGE_PENDING;
2013                 else if (status == -ENOENT)
2014                         (*bps)[i].flags = AMDGPU_RAS_RETIRE_PAGE_FAULT;
2015         }
2016
2017         *count = data->count;
2018 out:
2019         mutex_unlock(&con->recovery_lock);
2020         return ret;
2021 }
2022
2023 static void amdgpu_ras_do_recovery(struct work_struct *work)
2024 {
2025         struct amdgpu_ras *ras =
2026                 container_of(work, struct amdgpu_ras, recovery_work);
2027         struct amdgpu_device *remote_adev = NULL;
2028         struct amdgpu_device *adev = ras->adev;
2029         struct list_head device_list, *device_list_handle =  NULL;
2030
2031         if (!ras->disable_ras_err_cnt_harvest) {
2032                 struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev);
2033
2034                 /* Build list of devices to query RAS related errors */
2035                 if  (hive && adev->gmc.xgmi.num_physical_nodes > 1) {
2036                         device_list_handle = &hive->device_list;
2037                 } else {
2038                         INIT_LIST_HEAD(&device_list);
2039                         list_add_tail(&adev->gmc.xgmi.head, &device_list);
2040                         device_list_handle = &device_list;
2041                 }
2042
2043                 list_for_each_entry(remote_adev,
2044                                 device_list_handle, gmc.xgmi.head) {
2045                         amdgpu_ras_query_err_status(remote_adev);
2046                         amdgpu_ras_log_on_err_counter(remote_adev);
2047                 }
2048
2049                 amdgpu_put_xgmi_hive(hive);
2050         }
2051
2052         if (amdgpu_device_should_recover_gpu(ras->adev)) {
2053                 struct amdgpu_reset_context reset_context;
2054                 memset(&reset_context, 0, sizeof(reset_context));
2055
2056                 reset_context.method = AMD_RESET_METHOD_NONE;
2057                 reset_context.reset_req_dev = adev;
2058
2059                 /* Perform full reset in fatal error mode */
2060                 if (!amdgpu_ras_is_poison_mode_supported(ras->adev))
2061                         set_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
2062                 else {
2063                         clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
2064
2065                         if (ras->gpu_reset_flags & AMDGPU_RAS_GPU_RESET_MODE2_RESET) {
2066                                 ras->gpu_reset_flags &= ~AMDGPU_RAS_GPU_RESET_MODE2_RESET;
2067                                 reset_context.method = AMD_RESET_METHOD_MODE2;
2068                         }
2069                 }
2070
2071                 amdgpu_device_gpu_recover(ras->adev, NULL, &reset_context);
2072         }
2073         atomic_set(&ras->in_recovery, 0);
2074 }
2075
2076 /* alloc/realloc bps array */
2077 static int amdgpu_ras_realloc_eh_data_space(struct amdgpu_device *adev,
2078                 struct ras_err_handler_data *data, int pages)
2079 {
2080         unsigned int old_space = data->count + data->space_left;
2081         unsigned int new_space = old_space + pages;
2082         unsigned int align_space = ALIGN(new_space, 512);
2083         void *bps = kmalloc(align_space * sizeof(*data->bps), GFP_KERNEL);
2084
2085         if (!bps) {
2086                 return -ENOMEM;
2087         }
2088
2089         if (data->bps) {
2090                 memcpy(bps, data->bps,
2091                                 data->count * sizeof(*data->bps));
2092                 kfree(data->bps);
2093         }
2094
2095         data->bps = bps;
2096         data->space_left += align_space - old_space;
2097         return 0;
2098 }
2099
2100 /* it deal with vram only. */
2101 int amdgpu_ras_add_bad_pages(struct amdgpu_device *adev,
2102                 struct eeprom_table_record *bps, int pages)
2103 {
2104         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2105         struct ras_err_handler_data *data;
2106         int ret = 0;
2107         uint32_t i;
2108
2109         if (!con || !con->eh_data || !bps || pages <= 0)
2110                 return 0;
2111
2112         mutex_lock(&con->recovery_lock);
2113         data = con->eh_data;
2114         if (!data)
2115                 goto out;
2116
2117         for (i = 0; i < pages; i++) {
2118                 if (amdgpu_ras_check_bad_page_unlock(con,
2119                         bps[i].retired_page << AMDGPU_GPU_PAGE_SHIFT))
2120                         continue;
2121
2122                 if (!data->space_left &&
2123                         amdgpu_ras_realloc_eh_data_space(adev, data, 256)) {
2124                         ret = -ENOMEM;
2125                         goto out;
2126                 }
2127
2128                 amdgpu_vram_mgr_reserve_range(&adev->mman.vram_mgr,
2129                         bps[i].retired_page << AMDGPU_GPU_PAGE_SHIFT,
2130                         AMDGPU_GPU_PAGE_SIZE);
2131
2132                 memcpy(&data->bps[data->count], &bps[i], sizeof(*data->bps));
2133                 data->count++;
2134                 data->space_left--;
2135         }
2136 out:
2137         mutex_unlock(&con->recovery_lock);
2138
2139         return ret;
2140 }
2141
2142 /*
2143  * write error record array to eeprom, the function should be
2144  * protected by recovery_lock
2145  * new_cnt: new added UE count, excluding reserved bad pages, can be NULL
2146  */
2147 int amdgpu_ras_save_bad_pages(struct amdgpu_device *adev,
2148                 unsigned long *new_cnt)
2149 {
2150         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2151         struct ras_err_handler_data *data;
2152         struct amdgpu_ras_eeprom_control *control;
2153         int save_count;
2154
2155         if (!con || !con->eh_data) {
2156                 if (new_cnt)
2157                         *new_cnt = 0;
2158
2159                 return 0;
2160         }
2161
2162         mutex_lock(&con->recovery_lock);
2163         control = &con->eeprom_control;
2164         data = con->eh_data;
2165         save_count = data->count - control->ras_num_recs;
2166         mutex_unlock(&con->recovery_lock);
2167
2168         if (new_cnt)
2169                 *new_cnt = save_count / adev->umc.retire_unit;
2170
2171         /* only new entries are saved */
2172         if (save_count > 0) {
2173                 if (amdgpu_ras_eeprom_append(control,
2174                                              &data->bps[control->ras_num_recs],
2175                                              save_count)) {
2176                         dev_err(adev->dev, "Failed to save EEPROM table data!");
2177                         return -EIO;
2178                 }
2179
2180                 dev_info(adev->dev, "Saved %d pages to EEPROM table.\n", save_count);
2181         }
2182
2183         return 0;
2184 }
2185
2186 /*
2187  * read error record array in eeprom and reserve enough space for
2188  * storing new bad pages
2189  */
2190 static int amdgpu_ras_load_bad_pages(struct amdgpu_device *adev)
2191 {
2192         struct amdgpu_ras_eeprom_control *control =
2193                 &adev->psp.ras_context.ras->eeprom_control;
2194         struct eeprom_table_record *bps;
2195         int ret;
2196
2197         /* no bad page record, skip eeprom access */
2198         if (control->ras_num_recs == 0 || amdgpu_bad_page_threshold == 0)
2199                 return 0;
2200
2201         bps = kcalloc(control->ras_num_recs, sizeof(*bps), GFP_KERNEL);
2202         if (!bps)
2203                 return -ENOMEM;
2204
2205         ret = amdgpu_ras_eeprom_read(control, bps, control->ras_num_recs);
2206         if (ret)
2207                 dev_err(adev->dev, "Failed to load EEPROM table records!");
2208         else
2209                 ret = amdgpu_ras_add_bad_pages(adev, bps, control->ras_num_recs);
2210
2211         kfree(bps);
2212         return ret;
2213 }
2214
2215 static bool amdgpu_ras_check_bad_page_unlock(struct amdgpu_ras *con,
2216                                 uint64_t addr)
2217 {
2218         struct ras_err_handler_data *data = con->eh_data;
2219         int i;
2220
2221         addr >>= AMDGPU_GPU_PAGE_SHIFT;
2222         for (i = 0; i < data->count; i++)
2223                 if (addr == data->bps[i].retired_page)
2224                         return true;
2225
2226         return false;
2227 }
2228
2229 /*
2230  * check if an address belongs to bad page
2231  *
2232  * Note: this check is only for umc block
2233  */
2234 static bool amdgpu_ras_check_bad_page(struct amdgpu_device *adev,
2235                                 uint64_t addr)
2236 {
2237         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2238         bool ret = false;
2239
2240         if (!con || !con->eh_data)
2241                 return ret;
2242
2243         mutex_lock(&con->recovery_lock);
2244         ret = amdgpu_ras_check_bad_page_unlock(con, addr);
2245         mutex_unlock(&con->recovery_lock);
2246         return ret;
2247 }
2248
2249 static void amdgpu_ras_validate_threshold(struct amdgpu_device *adev,
2250                                           uint32_t max_count)
2251 {
2252         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2253
2254         /*
2255          * Justification of value bad_page_cnt_threshold in ras structure
2256          *
2257          * Generally, 0 <= amdgpu_bad_page_threshold <= max record length
2258          * in eeprom or amdgpu_bad_page_threshold == -2, introduce two
2259          * scenarios accordingly.
2260          *
2261          * Bad page retirement enablement:
2262          *    - If amdgpu_bad_page_threshold = -2,
2263          *      bad_page_cnt_threshold = typical value by formula.
2264          *
2265          *    - When the value from user is 0 < amdgpu_bad_page_threshold <
2266          *      max record length in eeprom, use it directly.
2267          *
2268          * Bad page retirement disablement:
2269          *    - If amdgpu_bad_page_threshold = 0, bad page retirement
2270          *      functionality is disabled, and bad_page_cnt_threshold will
2271          *      take no effect.
2272          */
2273
2274         if (amdgpu_bad_page_threshold < 0) {
2275                 u64 val = adev->gmc.mc_vram_size;
2276
2277                 do_div(val, RAS_BAD_PAGE_COVER);
2278                 con->bad_page_cnt_threshold = min(lower_32_bits(val),
2279                                                   max_count);
2280         } else {
2281                 con->bad_page_cnt_threshold = min_t(int, max_count,
2282                                                     amdgpu_bad_page_threshold);
2283         }
2284 }
2285
2286 int amdgpu_ras_recovery_init(struct amdgpu_device *adev)
2287 {
2288         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2289         struct ras_err_handler_data **data;
2290         u32  max_eeprom_records_count = 0;
2291         bool exc_err_limit = false;
2292         int ret;
2293
2294         if (!con || amdgpu_sriov_vf(adev))
2295                 return 0;
2296
2297         /* Allow access to RAS EEPROM via debugfs, when the ASIC
2298          * supports RAS and debugfs is enabled, but when
2299          * adev->ras_enabled is unset, i.e. when "ras_enable"
2300          * module parameter is set to 0.
2301          */
2302         con->adev = adev;
2303
2304         if (!adev->ras_enabled)
2305                 return 0;
2306
2307         data = &con->eh_data;
2308         *data = kmalloc(sizeof(**data), GFP_KERNEL | __GFP_ZERO);
2309         if (!*data) {
2310                 ret = -ENOMEM;
2311                 goto out;
2312         }
2313
2314         mutex_init(&con->recovery_lock);
2315         INIT_WORK(&con->recovery_work, amdgpu_ras_do_recovery);
2316         atomic_set(&con->in_recovery, 0);
2317         con->eeprom_control.bad_channel_bitmap = 0;
2318
2319         max_eeprom_records_count = amdgpu_ras_eeprom_max_record_count(&con->eeprom_control);
2320         amdgpu_ras_validate_threshold(adev, max_eeprom_records_count);
2321
2322         /* Todo: During test the SMU might fail to read the eeprom through I2C
2323          * when the GPU is pending on XGMI reset during probe time
2324          * (Mostly after second bus reset), skip it now
2325          */
2326         if (adev->gmc.xgmi.pending_reset)
2327                 return 0;
2328         ret = amdgpu_ras_eeprom_init(&con->eeprom_control, &exc_err_limit);
2329         /*
2330          * This calling fails when exc_err_limit is true or
2331          * ret != 0.
2332          */
2333         if (exc_err_limit || ret)
2334                 goto free;
2335
2336         if (con->eeprom_control.ras_num_recs) {
2337                 ret = amdgpu_ras_load_bad_pages(adev);
2338                 if (ret)
2339                         goto free;
2340
2341                 amdgpu_dpm_send_hbm_bad_pages_num(adev, con->eeprom_control.ras_num_recs);
2342
2343                 if (con->update_channel_flag == true) {
2344                         amdgpu_dpm_send_hbm_bad_channel_flag(adev, con->eeprom_control.bad_channel_bitmap);
2345                         con->update_channel_flag = false;
2346                 }
2347         }
2348
2349 #ifdef CONFIG_X86_MCE_AMD
2350         if ((adev->asic_type == CHIP_ALDEBARAN) &&
2351             (adev->gmc.xgmi.connected_to_cpu))
2352                 amdgpu_register_bad_pages_mca_notifier(adev);
2353 #endif
2354         return 0;
2355
2356 free:
2357         kfree((*data)->bps);
2358         kfree(*data);
2359         con->eh_data = NULL;
2360 out:
2361         dev_warn(adev->dev, "Failed to initialize ras recovery! (%d)\n", ret);
2362
2363         /*
2364          * Except error threshold exceeding case, other failure cases in this
2365          * function would not fail amdgpu driver init.
2366          */
2367         if (!exc_err_limit)
2368                 ret = 0;
2369         else
2370                 ret = -EINVAL;
2371
2372         return ret;
2373 }
2374
2375 static int amdgpu_ras_recovery_fini(struct amdgpu_device *adev)
2376 {
2377         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2378         struct ras_err_handler_data *data = con->eh_data;
2379
2380         /* recovery_init failed to init it, fini is useless */
2381         if (!data)
2382                 return 0;
2383
2384         cancel_work_sync(&con->recovery_work);
2385
2386         mutex_lock(&con->recovery_lock);
2387         con->eh_data = NULL;
2388         kfree(data->bps);
2389         kfree(data);
2390         mutex_unlock(&con->recovery_lock);
2391
2392         return 0;
2393 }
2394 /* recovery end */
2395
2396 static bool amdgpu_ras_asic_supported(struct amdgpu_device *adev)
2397 {
2398         if (amdgpu_sriov_vf(adev)) {
2399                 switch (adev->ip_versions[MP0_HWIP][0]) {
2400                 case IP_VERSION(13, 0, 2):
2401                         return true;
2402                 default:
2403                         return false;
2404                 }
2405         }
2406
2407         if (adev->asic_type == CHIP_IP_DISCOVERY) {
2408                 switch (adev->ip_versions[MP0_HWIP][0]) {
2409                 case IP_VERSION(13, 0, 0):
2410                 case IP_VERSION(13, 0, 10):
2411                         return true;
2412                 default:
2413                         return false;
2414                 }
2415         }
2416
2417         return adev->asic_type == CHIP_VEGA10 ||
2418                 adev->asic_type == CHIP_VEGA20 ||
2419                 adev->asic_type == CHIP_ARCTURUS ||
2420                 adev->asic_type == CHIP_ALDEBARAN ||
2421                 adev->asic_type == CHIP_SIENNA_CICHLID;
2422 }
2423
2424 /*
2425  * this is workaround for vega20 workstation sku,
2426  * force enable gfx ras, ignore vbios gfx ras flag
2427  * due to GC EDC can not write
2428  */
2429 static void amdgpu_ras_get_quirks(struct amdgpu_device *adev)
2430 {
2431         struct atom_context *ctx = adev->mode_info.atom_context;
2432
2433         if (!ctx)
2434                 return;
2435
2436         if (strnstr(ctx->vbios_version, "D16406",
2437                     sizeof(ctx->vbios_version)) ||
2438                 strnstr(ctx->vbios_version, "D36002",
2439                         sizeof(ctx->vbios_version)))
2440                 adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__GFX);
2441 }
2442
2443 /*
2444  * check hardware's ras ability which will be saved in hw_supported.
2445  * if hardware does not support ras, we can skip some ras initializtion and
2446  * forbid some ras operations from IP.
2447  * if software itself, say boot parameter, limit the ras ability. We still
2448  * need allow IP do some limited operations, like disable. In such case,
2449  * we have to initialize ras as normal. but need check if operation is
2450  * allowed or not in each function.
2451  */
2452 static void amdgpu_ras_check_supported(struct amdgpu_device *adev)
2453 {
2454         adev->ras_hw_enabled = adev->ras_enabled = 0;
2455
2456         if (!adev->is_atom_fw ||
2457             !amdgpu_ras_asic_supported(adev))
2458                 return;
2459
2460         if (!adev->gmc.xgmi.connected_to_cpu) {
2461                 if (amdgpu_atomfirmware_mem_ecc_supported(adev)) {
2462                         dev_info(adev->dev, "MEM ECC is active.\n");
2463                         adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__UMC |
2464                                                    1 << AMDGPU_RAS_BLOCK__DF);
2465                 } else {
2466                         dev_info(adev->dev, "MEM ECC is not presented.\n");
2467                 }
2468
2469                 if (amdgpu_atomfirmware_sram_ecc_supported(adev)) {
2470                         dev_info(adev->dev, "SRAM ECC is active.\n");
2471                         if (!amdgpu_sriov_vf(adev))
2472                                 adev->ras_hw_enabled |= ~(1 << AMDGPU_RAS_BLOCK__UMC |
2473                                                             1 << AMDGPU_RAS_BLOCK__DF);
2474                         else
2475                                 adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__PCIE_BIF |
2476                                                                 1 << AMDGPU_RAS_BLOCK__SDMA |
2477                                                                 1 << AMDGPU_RAS_BLOCK__GFX);
2478
2479                         /* VCN/JPEG RAS can be supported on both bare metal and
2480                          * SRIOV environment
2481                          */
2482                         if (adev->ip_versions[VCN_HWIP][0] == IP_VERSION(2, 6, 0) ||
2483                             adev->ip_versions[VCN_HWIP][0] == IP_VERSION(4, 0, 0))
2484                                 adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__VCN |
2485                                                         1 << AMDGPU_RAS_BLOCK__JPEG);
2486                         else
2487                                 adev->ras_hw_enabled &= ~(1 << AMDGPU_RAS_BLOCK__VCN |
2488                                                         1 << AMDGPU_RAS_BLOCK__JPEG);
2489
2490                         /*
2491                          * XGMI RAS is not supported if xgmi num physical nodes
2492                          * is zero
2493                          */
2494                         if (!adev->gmc.xgmi.num_physical_nodes)
2495                                 adev->ras_hw_enabled &= ~(1 << AMDGPU_RAS_BLOCK__XGMI_WAFL);
2496                 } else {
2497                         dev_info(adev->dev, "SRAM ECC is not presented.\n");
2498                 }
2499         } else {
2500                 /* driver only manages a few IP blocks RAS feature
2501                  * when GPU is connected cpu through XGMI */
2502                 adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__GFX |
2503                                            1 << AMDGPU_RAS_BLOCK__SDMA |
2504                                            1 << AMDGPU_RAS_BLOCK__MMHUB);
2505         }
2506
2507         amdgpu_ras_get_quirks(adev);
2508
2509         /* hw_supported needs to be aligned with RAS block mask. */
2510         adev->ras_hw_enabled &= AMDGPU_RAS_BLOCK_MASK;
2511
2512         adev->ras_enabled = amdgpu_ras_enable == 0 ? 0 :
2513                 adev->ras_hw_enabled & amdgpu_ras_mask;
2514 }
2515
2516 static void amdgpu_ras_counte_dw(struct work_struct *work)
2517 {
2518         struct amdgpu_ras *con = container_of(work, struct amdgpu_ras,
2519                                               ras_counte_delay_work.work);
2520         struct amdgpu_device *adev = con->adev;
2521         struct drm_device *dev = adev_to_drm(adev);
2522         unsigned long ce_count, ue_count;
2523         int res;
2524
2525         res = pm_runtime_get_sync(dev->dev);
2526         if (res < 0)
2527                 goto Out;
2528
2529         /* Cache new values.
2530          */
2531         if (amdgpu_ras_query_error_count(adev, &ce_count, &ue_count, NULL) == 0) {
2532                 atomic_set(&con->ras_ce_count, ce_count);
2533                 atomic_set(&con->ras_ue_count, ue_count);
2534         }
2535
2536         pm_runtime_mark_last_busy(dev->dev);
2537 Out:
2538         pm_runtime_put_autosuspend(dev->dev);
2539 }
2540
2541 static void amdgpu_ras_query_poison_mode(struct amdgpu_device *adev)
2542 {
2543         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2544         bool df_poison, umc_poison;
2545
2546         /* poison setting is useless on SRIOV guest */
2547         if (amdgpu_sriov_vf(adev) || !con)
2548                 return;
2549
2550         /* Init poison supported flag, the default value is false */
2551         if (adev->gmc.xgmi.connected_to_cpu) {
2552                 /* enabled by default when GPU is connected to CPU */
2553                 con->poison_supported = true;
2554         } else if (adev->df.funcs &&
2555             adev->df.funcs->query_ras_poison_mode &&
2556             adev->umc.ras &&
2557             adev->umc.ras->query_ras_poison_mode) {
2558                 df_poison =
2559                         adev->df.funcs->query_ras_poison_mode(adev);
2560                 umc_poison =
2561                         adev->umc.ras->query_ras_poison_mode(adev);
2562
2563                 /* Only poison is set in both DF and UMC, we can support it */
2564                 if (df_poison && umc_poison)
2565                         con->poison_supported = true;
2566                 else if (df_poison != umc_poison)
2567                         dev_warn(adev->dev,
2568                                 "Poison setting is inconsistent in DF/UMC(%d:%d)!\n",
2569                                 df_poison, umc_poison);
2570         }
2571 }
2572
2573 int amdgpu_ras_init(struct amdgpu_device *adev)
2574 {
2575         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2576         int r;
2577
2578         if (con)
2579                 return 0;
2580
2581         con = kmalloc(sizeof(struct amdgpu_ras) +
2582                         sizeof(struct ras_manager) * AMDGPU_RAS_BLOCK_COUNT +
2583                         sizeof(struct ras_manager) * AMDGPU_RAS_MCA_BLOCK_COUNT,
2584                         GFP_KERNEL|__GFP_ZERO);
2585         if (!con)
2586                 return -ENOMEM;
2587
2588         con->adev = adev;
2589         INIT_DELAYED_WORK(&con->ras_counte_delay_work, amdgpu_ras_counte_dw);
2590         atomic_set(&con->ras_ce_count, 0);
2591         atomic_set(&con->ras_ue_count, 0);
2592
2593         con->objs = (struct ras_manager *)(con + 1);
2594
2595         amdgpu_ras_set_context(adev, con);
2596
2597         amdgpu_ras_check_supported(adev);
2598
2599         if (!adev->ras_enabled || adev->asic_type == CHIP_VEGA10) {
2600                 /* set gfx block ras context feature for VEGA20 Gaming
2601                  * send ras disable cmd to ras ta during ras late init.
2602                  */
2603                 if (!adev->ras_enabled && adev->asic_type == CHIP_VEGA20) {
2604                         con->features |= BIT(AMDGPU_RAS_BLOCK__GFX);
2605
2606                         return 0;
2607                 }
2608
2609                 r = 0;
2610                 goto release_con;
2611         }
2612
2613         con->update_channel_flag = false;
2614         con->features = 0;
2615         INIT_LIST_HEAD(&con->head);
2616         /* Might need get this flag from vbios. */
2617         con->flags = RAS_DEFAULT_FLAGS;
2618
2619         /* initialize nbio ras function ahead of any other
2620          * ras functions so hardware fatal error interrupt
2621          * can be enabled as early as possible */
2622         switch (adev->ip_versions[NBIO_HWIP][0]) {
2623         case IP_VERSION(7, 4, 0):
2624         case IP_VERSION(7, 4, 1):
2625         case IP_VERSION(7, 4, 4):
2626                 if (!adev->gmc.xgmi.connected_to_cpu)
2627                         adev->nbio.ras = &nbio_v7_4_ras;
2628                 break;
2629         case IP_VERSION(4, 3, 0):
2630                 if (adev->ras_hw_enabled & (1 << AMDGPU_RAS_BLOCK__DF))
2631                         /* unlike other generation of nbio ras,
2632                          * nbio v4_3 only support fatal error interrupt
2633                          * to inform software that DF is freezed due to
2634                          * system fatal error event. driver should not
2635                          * enable nbio ras in such case. Instead,
2636                          * check DF RAS */
2637                         adev->nbio.ras = &nbio_v4_3_ras;
2638                 break;
2639         default:
2640                 /* nbio ras is not available */
2641                 break;
2642         }
2643
2644         /* nbio ras block needs to be enabled ahead of other ras blocks
2645          * to handle fatal error */
2646         r = amdgpu_nbio_ras_sw_init(adev);
2647         if (r)
2648                 return r;
2649
2650         if (adev->nbio.ras &&
2651             adev->nbio.ras->init_ras_controller_interrupt) {
2652                 r = adev->nbio.ras->init_ras_controller_interrupt(adev);
2653                 if (r)
2654                         goto release_con;
2655         }
2656
2657         if (adev->nbio.ras &&
2658             adev->nbio.ras->init_ras_err_event_athub_interrupt) {
2659                 r = adev->nbio.ras->init_ras_err_event_athub_interrupt(adev);
2660                 if (r)
2661                         goto release_con;
2662         }
2663
2664         amdgpu_ras_query_poison_mode(adev);
2665
2666         if (amdgpu_ras_fs_init(adev)) {
2667                 r = -EINVAL;
2668                 goto release_con;
2669         }
2670
2671         dev_info(adev->dev, "RAS INFO: ras initialized successfully, "
2672                  "hardware ability[%x] ras_mask[%x]\n",
2673                  adev->ras_hw_enabled, adev->ras_enabled);
2674
2675         return 0;
2676 release_con:
2677         amdgpu_ras_set_context(adev, NULL);
2678         kfree(con);
2679
2680         return r;
2681 }
2682
2683 int amdgpu_persistent_edc_harvesting_supported(struct amdgpu_device *adev)
2684 {
2685         if (adev->gmc.xgmi.connected_to_cpu ||
2686             adev->gmc.is_app_apu)
2687                 return 1;
2688         return 0;
2689 }
2690
2691 static int amdgpu_persistent_edc_harvesting(struct amdgpu_device *adev,
2692                                         struct ras_common_if *ras_block)
2693 {
2694         struct ras_query_if info = {
2695                 .head = *ras_block,
2696         };
2697
2698         if (!amdgpu_persistent_edc_harvesting_supported(adev))
2699                 return 0;
2700
2701         if (amdgpu_ras_query_error_status(adev, &info) != 0)
2702                 DRM_WARN("RAS init harvest failure");
2703
2704         if (amdgpu_ras_reset_error_status(adev, ras_block->block) != 0)
2705                 DRM_WARN("RAS init harvest reset failure");
2706
2707         return 0;
2708 }
2709
2710 bool amdgpu_ras_is_poison_mode_supported(struct amdgpu_device *adev)
2711 {
2712        struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2713
2714        if (!con)
2715                return false;
2716
2717        return con->poison_supported;
2718 }
2719
2720 /* helper function to handle common stuff in ip late init phase */
2721 int amdgpu_ras_block_late_init(struct amdgpu_device *adev,
2722                          struct ras_common_if *ras_block)
2723 {
2724         struct amdgpu_ras_block_object *ras_obj = NULL;
2725         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2726         struct ras_query_if *query_info;
2727         unsigned long ue_count, ce_count;
2728         int r;
2729
2730         /* disable RAS feature per IP block if it is not supported */
2731         if (!amdgpu_ras_is_supported(adev, ras_block->block)) {
2732                 amdgpu_ras_feature_enable_on_boot(adev, ras_block, 0);
2733                 return 0;
2734         }
2735
2736         r = amdgpu_ras_feature_enable_on_boot(adev, ras_block, 1);
2737         if (r) {
2738                 if (adev->in_suspend || amdgpu_in_reset(adev)) {
2739                         /* in resume phase, if fail to enable ras,
2740                          * clean up all ras fs nodes, and disable ras */
2741                         goto cleanup;
2742                 } else
2743                         return r;
2744         }
2745
2746         /* check for errors on warm reset edc persisant supported ASIC */
2747         amdgpu_persistent_edc_harvesting(adev, ras_block);
2748
2749         /* in resume phase, no need to create ras fs node */
2750         if (adev->in_suspend || amdgpu_in_reset(adev))
2751                 return 0;
2752
2753         ras_obj = container_of(ras_block, struct amdgpu_ras_block_object, ras_comm);
2754         if (ras_obj->ras_cb || (ras_obj->hw_ops &&
2755             (ras_obj->hw_ops->query_poison_status ||
2756             ras_obj->hw_ops->handle_poison_consumption))) {
2757                 r = amdgpu_ras_interrupt_add_handler(adev, ras_block);
2758                 if (r)
2759                         goto cleanup;
2760         }
2761
2762         r = amdgpu_ras_sysfs_create(adev, ras_block);
2763         if (r)
2764                 goto interrupt;
2765
2766         /* Those are the cached values at init.
2767          */
2768         query_info = kzalloc(sizeof(struct ras_query_if), GFP_KERNEL);
2769         if (!query_info)
2770                 return -ENOMEM;
2771         memcpy(&query_info->head, ras_block, sizeof(struct ras_common_if));
2772
2773         if (amdgpu_ras_query_error_count(adev, &ce_count, &ue_count, query_info) == 0) {
2774                 atomic_set(&con->ras_ce_count, ce_count);
2775                 atomic_set(&con->ras_ue_count, ue_count);
2776         }
2777
2778         kfree(query_info);
2779         return 0;
2780
2781 interrupt:
2782         if (ras_obj->ras_cb)
2783                 amdgpu_ras_interrupt_remove_handler(adev, ras_block);
2784 cleanup:
2785         amdgpu_ras_feature_enable(adev, ras_block, 0);
2786         return r;
2787 }
2788
2789 static int amdgpu_ras_block_late_init_default(struct amdgpu_device *adev,
2790                          struct ras_common_if *ras_block)
2791 {
2792         return amdgpu_ras_block_late_init(adev, ras_block);
2793 }
2794
2795 /* helper function to remove ras fs node and interrupt handler */
2796 void amdgpu_ras_block_late_fini(struct amdgpu_device *adev,
2797                           struct ras_common_if *ras_block)
2798 {
2799         struct amdgpu_ras_block_object *ras_obj;
2800         if (!ras_block)
2801                 return;
2802
2803         amdgpu_ras_sysfs_remove(adev, ras_block);
2804
2805         ras_obj = container_of(ras_block, struct amdgpu_ras_block_object, ras_comm);
2806         if (ras_obj->ras_cb)
2807                 amdgpu_ras_interrupt_remove_handler(adev, ras_block);
2808 }
2809
2810 static void amdgpu_ras_block_late_fini_default(struct amdgpu_device *adev,
2811                           struct ras_common_if *ras_block)
2812 {
2813         return amdgpu_ras_block_late_fini(adev, ras_block);
2814 }
2815
2816 /* do some init work after IP late init as dependence.
2817  * and it runs in resume/gpu reset/booting up cases.
2818  */
2819 void amdgpu_ras_resume(struct amdgpu_device *adev)
2820 {
2821         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2822         struct ras_manager *obj, *tmp;
2823
2824         if (!adev->ras_enabled || !con) {
2825                 /* clean ras context for VEGA20 Gaming after send ras disable cmd */
2826                 amdgpu_release_ras_context(adev);
2827
2828                 return;
2829         }
2830
2831         if (con->flags & AMDGPU_RAS_FLAG_INIT_BY_VBIOS) {
2832                 /* Set up all other IPs which are not implemented. There is a
2833                  * tricky thing that IP's actual ras error type should be
2834                  * MULTI_UNCORRECTABLE, but as driver does not handle it, so
2835                  * ERROR_NONE make sense anyway.
2836                  */
2837                 amdgpu_ras_enable_all_features(adev, 1);
2838
2839                 /* We enable ras on all hw_supported block, but as boot
2840                  * parameter might disable some of them and one or more IP has
2841                  * not implemented yet. So we disable them on behalf.
2842                  */
2843                 list_for_each_entry_safe(obj, tmp, &con->head, node) {
2844                         if (!amdgpu_ras_is_supported(adev, obj->head.block)) {
2845                                 amdgpu_ras_feature_enable(adev, &obj->head, 0);
2846                                 /* there should be no any reference. */
2847                                 WARN_ON(alive_obj(obj));
2848                         }
2849                 }
2850         }
2851 }
2852
2853 void amdgpu_ras_suspend(struct amdgpu_device *adev)
2854 {
2855         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2856
2857         if (!adev->ras_enabled || !con)
2858                 return;
2859
2860         amdgpu_ras_disable_all_features(adev, 0);
2861         /* Make sure all ras objects are disabled. */
2862         if (con->features)
2863                 amdgpu_ras_disable_all_features(adev, 1);
2864 }
2865
2866 int amdgpu_ras_late_init(struct amdgpu_device *adev)
2867 {
2868         struct amdgpu_ras_block_list *node, *tmp;
2869         struct amdgpu_ras_block_object *obj;
2870         int r;
2871
2872         /* Guest side doesn't need init ras feature */
2873         if (amdgpu_sriov_vf(adev))
2874                 return 0;
2875
2876         list_for_each_entry_safe(node, tmp, &adev->ras_list, node) {
2877                 if (!node->ras_obj) {
2878                         dev_warn(adev->dev, "Warning: abnormal ras list node.\n");
2879                         continue;
2880                 }
2881
2882                 obj = node->ras_obj;
2883                 if (obj->ras_late_init) {
2884                         r = obj->ras_late_init(adev, &obj->ras_comm);
2885                         if (r) {
2886                                 dev_err(adev->dev, "%s failed to execute ras_late_init! ret:%d\n",
2887                                         obj->ras_comm.name, r);
2888                                 return r;
2889                         }
2890                 } else
2891                         amdgpu_ras_block_late_init_default(adev, &obj->ras_comm);
2892         }
2893
2894         return 0;
2895 }
2896
2897 /* do some fini work before IP fini as dependence */
2898 int amdgpu_ras_pre_fini(struct amdgpu_device *adev)
2899 {
2900         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2901
2902         if (!adev->ras_enabled || !con)
2903                 return 0;
2904
2905
2906         /* Need disable ras on all IPs here before ip [hw/sw]fini */
2907         if (con->features)
2908                 amdgpu_ras_disable_all_features(adev, 0);
2909         amdgpu_ras_recovery_fini(adev);
2910         return 0;
2911 }
2912
2913 int amdgpu_ras_fini(struct amdgpu_device *adev)
2914 {
2915         struct amdgpu_ras_block_list *ras_node, *tmp;
2916         struct amdgpu_ras_block_object *obj = NULL;
2917         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2918
2919         if (!adev->ras_enabled || !con)
2920                 return 0;
2921
2922         list_for_each_entry_safe(ras_node, tmp, &adev->ras_list, node) {
2923                 if (ras_node->ras_obj) {
2924                         obj = ras_node->ras_obj;
2925                         if (amdgpu_ras_is_supported(adev, obj->ras_comm.block) &&
2926                             obj->ras_fini)
2927                                 obj->ras_fini(adev, &obj->ras_comm);
2928                         else
2929                                 amdgpu_ras_block_late_fini_default(adev, &obj->ras_comm);
2930                 }
2931
2932                 /* Clear ras blocks from ras_list and free ras block list node */
2933                 list_del(&ras_node->node);
2934                 kfree(ras_node);
2935         }
2936
2937         amdgpu_ras_fs_fini(adev);
2938         amdgpu_ras_interrupt_remove_all(adev);
2939
2940         WARN(con->features, "Feature mask is not cleared");
2941
2942         if (con->features)
2943                 amdgpu_ras_disable_all_features(adev, 1);
2944
2945         cancel_delayed_work_sync(&con->ras_counte_delay_work);
2946
2947         amdgpu_ras_set_context(adev, NULL);
2948         kfree(con);
2949
2950         return 0;
2951 }
2952
2953 void amdgpu_ras_global_ras_isr(struct amdgpu_device *adev)
2954 {
2955         amdgpu_ras_check_supported(adev);
2956         if (!adev->ras_hw_enabled)
2957                 return;
2958
2959         if (atomic_cmpxchg(&amdgpu_ras_in_intr, 0, 1) == 0) {
2960                 dev_info(adev->dev, "uncorrectable hardware error"
2961                         "(ERREVENT_ATHUB_INTERRUPT) detected!\n");
2962
2963                 amdgpu_ras_reset_gpu(adev);
2964         }
2965 }
2966
2967 bool amdgpu_ras_need_emergency_restart(struct amdgpu_device *adev)
2968 {
2969         if (adev->asic_type == CHIP_VEGA20 &&
2970             adev->pm.fw_version <= 0x283400) {
2971                 return !(amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) &&
2972                                 amdgpu_ras_intr_triggered();
2973         }
2974
2975         return false;
2976 }
2977
2978 void amdgpu_release_ras_context(struct amdgpu_device *adev)
2979 {
2980         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2981
2982         if (!con)
2983                 return;
2984
2985         if (!adev->ras_enabled && con->features & BIT(AMDGPU_RAS_BLOCK__GFX)) {
2986                 con->features &= ~BIT(AMDGPU_RAS_BLOCK__GFX);
2987                 amdgpu_ras_set_context(adev, NULL);
2988                 kfree(con);
2989         }
2990 }
2991
2992 #ifdef CONFIG_X86_MCE_AMD
2993 static struct amdgpu_device *find_adev(uint32_t node_id)
2994 {
2995         int i;
2996         struct amdgpu_device *adev = NULL;
2997
2998         for (i = 0; i < mce_adev_list.num_gpu; i++) {
2999                 adev = mce_adev_list.devs[i];
3000
3001                 if (adev && adev->gmc.xgmi.connected_to_cpu &&
3002                     adev->gmc.xgmi.physical_node_id == node_id)
3003                         break;
3004                 adev = NULL;
3005         }
3006
3007         return adev;
3008 }
3009
3010 #define GET_MCA_IPID_GPUID(m)   (((m) >> 44) & 0xF)
3011 #define GET_UMC_INST(m)         (((m) >> 21) & 0x7)
3012 #define GET_CHAN_INDEX(m)       ((((m) >> 12) & 0x3) | (((m) >> 18) & 0x4))
3013 #define GPU_ID_OFFSET           8
3014
3015 static int amdgpu_bad_page_notifier(struct notifier_block *nb,
3016                                     unsigned long val, void *data)
3017 {
3018         struct mce *m = (struct mce *)data;
3019         struct amdgpu_device *adev = NULL;
3020         uint32_t gpu_id = 0;
3021         uint32_t umc_inst = 0, ch_inst = 0;
3022
3023         /*
3024          * If the error was generated in UMC_V2, which belongs to GPU UMCs,
3025          * and error occurred in DramECC (Extended error code = 0) then only
3026          * process the error, else bail out.
3027          */
3028         if (!m || !((smca_get_bank_type(m->extcpu, m->bank) == SMCA_UMC_V2) &&
3029                     (XEC(m->status, 0x3f) == 0x0)))
3030                 return NOTIFY_DONE;
3031
3032         /*
3033          * If it is correctable error, return.
3034          */
3035         if (mce_is_correctable(m))
3036                 return NOTIFY_OK;
3037
3038         /*
3039          * GPU Id is offset by GPU_ID_OFFSET in MCA_IPID_UMC register.
3040          */
3041         gpu_id = GET_MCA_IPID_GPUID(m->ipid) - GPU_ID_OFFSET;
3042
3043         adev = find_adev(gpu_id);
3044         if (!adev) {
3045                 DRM_WARN("%s: Unable to find adev for gpu_id: %d\n", __func__,
3046                                                                 gpu_id);
3047                 return NOTIFY_DONE;
3048         }
3049
3050         /*
3051          * If it is uncorrectable error, then find out UMC instance and
3052          * channel index.
3053          */
3054         umc_inst = GET_UMC_INST(m->ipid);
3055         ch_inst = GET_CHAN_INDEX(m->ipid);
3056
3057         dev_info(adev->dev, "Uncorrectable error detected in UMC inst: %d, chan_idx: %d",
3058                              umc_inst, ch_inst);
3059
3060         if (!amdgpu_umc_page_retirement_mca(adev, m->addr, ch_inst, umc_inst))
3061                 return NOTIFY_OK;
3062         else
3063                 return NOTIFY_DONE;
3064 }
3065
3066 static struct notifier_block amdgpu_bad_page_nb = {
3067         .notifier_call  = amdgpu_bad_page_notifier,
3068         .priority       = MCE_PRIO_UC,
3069 };
3070
3071 static void amdgpu_register_bad_pages_mca_notifier(struct amdgpu_device *adev)
3072 {
3073         /*
3074          * Add the adev to the mce_adev_list.
3075          * During mode2 reset, amdgpu device is temporarily
3076          * removed from the mgpu_info list which can cause
3077          * page retirement to fail.
3078          * Use this list instead of mgpu_info to find the amdgpu
3079          * device on which the UMC error was reported.
3080          */
3081         mce_adev_list.devs[mce_adev_list.num_gpu++] = adev;
3082
3083         /*
3084          * Register the x86 notifier only once
3085          * with MCE subsystem.
3086          */
3087         if (notifier_registered == false) {
3088                 mce_register_decode_chain(&amdgpu_bad_page_nb);
3089                 notifier_registered = true;
3090         }
3091 }
3092 #endif
3093
3094 struct amdgpu_ras *amdgpu_ras_get_context(struct amdgpu_device *adev)
3095 {
3096         if (!adev)
3097                 return NULL;
3098
3099         return adev->psp.ras_context.ras;
3100 }
3101
3102 int amdgpu_ras_set_context(struct amdgpu_device *adev, struct amdgpu_ras *ras_con)
3103 {
3104         if (!adev)
3105                 return -EINVAL;
3106
3107         adev->psp.ras_context.ras = ras_con;
3108         return 0;
3109 }
3110
3111 /* check if ras is supported on block, say, sdma, gfx */
3112 int amdgpu_ras_is_supported(struct amdgpu_device *adev,
3113                 unsigned int block)
3114 {
3115         int ret = 0;
3116         struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
3117
3118         if (block >= AMDGPU_RAS_BLOCK_COUNT)
3119                 return 0;
3120
3121         ret = ras && (adev->ras_enabled & (1 << block));
3122
3123         /* For the special asic with mem ecc enabled but sram ecc
3124          * not enabled, even if the ras block is not supported on
3125          * .ras_enabled, if the asic supports poison mode and the
3126          * ras block has ras configuration, it can be considered
3127          * that the ras block supports ras function.
3128          */
3129         if (!ret &&
3130             amdgpu_ras_is_poison_mode_supported(adev) &&
3131             amdgpu_ras_get_ras_block(adev, block, 0))
3132                 ret = 1;
3133
3134         return ret;
3135 }
3136
3137 int amdgpu_ras_reset_gpu(struct amdgpu_device *adev)
3138 {
3139         struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
3140
3141         if (atomic_cmpxchg(&ras->in_recovery, 0, 1) == 0)
3142                 amdgpu_reset_domain_schedule(ras->adev->reset_domain, &ras->recovery_work);
3143         return 0;
3144 }
3145
3146
3147 /* Register each ip ras block into amdgpu ras */
3148 int amdgpu_ras_register_ras_block(struct amdgpu_device *adev,
3149                 struct amdgpu_ras_block_object *ras_block_obj)
3150 {
3151         struct amdgpu_ras_block_list *ras_node;
3152         if (!adev || !ras_block_obj)
3153                 return -EINVAL;
3154
3155         ras_node = kzalloc(sizeof(*ras_node), GFP_KERNEL);
3156         if (!ras_node)
3157                 return -ENOMEM;
3158
3159         INIT_LIST_HEAD(&ras_node->node);
3160         ras_node->ras_obj = ras_block_obj;
3161         list_add_tail(&ras_node->node, &adev->ras_list);
3162
3163         return 0;
3164 }
3165
3166 void amdgpu_ras_get_error_type_name(uint32_t err_type, char *err_type_name)
3167 {
3168         if (!err_type_name)
3169                 return;
3170
3171         switch (err_type) {
3172         case AMDGPU_RAS_ERROR__SINGLE_CORRECTABLE:
3173                 sprintf(err_type_name, "correctable");
3174                 break;
3175         case AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE:
3176                 sprintf(err_type_name, "uncorrectable");
3177                 break;
3178         default:
3179                 sprintf(err_type_name, "unknown");
3180                 break;
3181         }
3182 }
3183
3184 bool amdgpu_ras_inst_get_memory_id_field(struct amdgpu_device *adev,
3185                                          const struct amdgpu_ras_err_status_reg_entry *reg_entry,
3186                                          uint32_t instance,
3187                                          uint32_t *memory_id)
3188 {
3189         uint32_t err_status_lo_data, err_status_lo_offset;
3190
3191         if (!reg_entry)
3192                 return false;
3193
3194         err_status_lo_offset =
3195                 AMDGPU_RAS_REG_ENTRY_OFFSET(reg_entry->hwip, instance,
3196                                             reg_entry->seg_lo, reg_entry->reg_lo);
3197         err_status_lo_data = RREG32(err_status_lo_offset);
3198
3199         if ((reg_entry->flags & AMDGPU_RAS_ERR_STATUS_VALID) &&
3200             !REG_GET_FIELD(err_status_lo_data, ERR_STATUS_LO, ERR_STATUS_VALID_FLAG))
3201                 return false;
3202
3203         *memory_id = REG_GET_FIELD(err_status_lo_data, ERR_STATUS_LO, MEMORY_ID);
3204
3205         return true;
3206 }
3207
3208 bool amdgpu_ras_inst_get_err_cnt_field(struct amdgpu_device *adev,
3209                                        const struct amdgpu_ras_err_status_reg_entry *reg_entry,
3210                                        uint32_t instance,
3211                                        unsigned long *err_cnt)
3212 {
3213         uint32_t err_status_hi_data, err_status_hi_offset;
3214
3215         if (!reg_entry)
3216                 return false;
3217
3218         err_status_hi_offset =
3219                 AMDGPU_RAS_REG_ENTRY_OFFSET(reg_entry->hwip, instance,
3220                                             reg_entry->seg_hi, reg_entry->reg_hi);
3221         err_status_hi_data = RREG32(err_status_hi_offset);
3222
3223         if ((reg_entry->flags & AMDGPU_RAS_ERR_INFO_VALID) &&
3224             !REG_GET_FIELD(err_status_hi_data, ERR_STATUS_HI, ERR_INFO_VALID_FLAG))
3225                 /* keep the check here in case we need to refer to the result later */
3226                 dev_dbg(adev->dev, "Invalid err_info field\n");
3227
3228         /* read err count */
3229         *err_cnt = REG_GET_FIELD(err_status_hi_data, ERR_STATUS, ERR_CNT);
3230
3231         return true;
3232 }
3233
3234 void amdgpu_ras_inst_query_ras_error_count(struct amdgpu_device *adev,
3235                                            const struct amdgpu_ras_err_status_reg_entry *reg_list,
3236                                            uint32_t reg_list_size,
3237                                            const struct amdgpu_ras_memory_id_entry *mem_list,
3238                                            uint32_t mem_list_size,
3239                                            uint32_t instance,
3240                                            uint32_t err_type,
3241                                            unsigned long *err_count)
3242 {
3243         uint32_t memory_id;
3244         unsigned long err_cnt;
3245         char err_type_name[16];
3246         uint32_t i, j;
3247
3248         for (i = 0; i < reg_list_size; i++) {
3249                 /* query memory_id from err_status_lo */
3250                 if (!amdgpu_ras_inst_get_memory_id_field(adev, &reg_list[i],
3251                                                          instance, &memory_id))
3252                         continue;
3253
3254                 /* query err_cnt from err_status_hi */
3255                 if (!amdgpu_ras_inst_get_err_cnt_field(adev, &reg_list[i],
3256                                                        instance, &err_cnt) ||
3257                     !err_cnt)
3258                         continue;
3259
3260                 *err_count += err_cnt;
3261
3262                 /* log the errors */
3263                 amdgpu_ras_get_error_type_name(err_type, err_type_name);
3264                 if (!mem_list) {
3265                         /* memory_list is not supported */
3266                         dev_info(adev->dev,
3267                                  "%ld %s hardware errors detected in %s, instance: %d, memory_id: %d\n",
3268                                  err_cnt, err_type_name,
3269                                  reg_list[i].block_name,
3270                                  instance, memory_id);
3271                 } else {
3272                         for (j = 0; j < mem_list_size; j++) {
3273                                 if (memory_id == mem_list[j].memory_id) {
3274                                         dev_info(adev->dev,
3275                                                  "%ld %s hardware errors detected in %s, instance: %d, memory block: %s\n",
3276                                                  err_cnt, err_type_name,
3277                                                  reg_list[i].block_name,
3278                                                  instance, mem_list[j].name);
3279                                         break;
3280                                 }
3281                         }
3282                 }
3283         }
3284 }
3285
3286 void amdgpu_ras_inst_reset_ras_error_count(struct amdgpu_device *adev,
3287                                            const struct amdgpu_ras_err_status_reg_entry *reg_list,
3288                                            uint32_t reg_list_size,
3289                                            uint32_t instance)
3290 {
3291         uint32_t err_status_lo_offset, err_status_hi_offset;
3292         uint32_t i;
3293
3294         for (i = 0; i < reg_list_size; i++) {
3295                 err_status_lo_offset =
3296                         AMDGPU_RAS_REG_ENTRY_OFFSET(reg_list[i].hwip, instance,
3297                                                     reg_list[i].seg_lo, reg_list[i].reg_lo);
3298                 err_status_hi_offset =
3299                         AMDGPU_RAS_REG_ENTRY_OFFSET(reg_list[i].hwip, instance,
3300                                                     reg_list[i].seg_hi, reg_list[i].reg_hi);
3301                 WREG32(err_status_lo_offset, 0);
3302                 WREG32(err_status_hi_offset, 0);
3303         }
3304 }