Linux Audio

Check our new training course

Loading...
Note: File does not exist in v3.1.
   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 "atom.h"
  38
  39static const char *RAS_FS_NAME = "ras";
  40
  41const char *ras_error_string[] = {
  42	"none",
  43	"parity",
  44	"single_correctable",
  45	"multi_uncorrectable",
  46	"poison",
  47};
  48
  49const char *ras_block_string[] = {
  50	"umc",
  51	"sdma",
  52	"gfx",
  53	"mmhub",
  54	"athub",
  55	"pcie_bif",
  56	"hdp",
  57	"xgmi_wafl",
  58	"df",
  59	"smn",
  60	"sem",
  61	"mp0",
  62	"mp1",
  63	"fuse",
  64};
  65
  66#define ras_err_str(i) (ras_error_string[ffs(i)])
  67#define ras_block_str(i) (ras_block_string[i])
  68
  69#define RAS_DEFAULT_FLAGS (AMDGPU_RAS_FLAG_INIT_BY_VBIOS)
  70
  71/* inject address is 52 bits */
  72#define	RAS_UMC_INJECT_ADDR_LIMIT	(0x1ULL << 52)
  73
  74/* typical ECC bad page rate(1 bad page per 100MB VRAM) */
  75#define RAS_BAD_PAGE_RATE		(100 * 1024 * 1024ULL)
  76
  77enum amdgpu_ras_retire_page_reservation {
  78	AMDGPU_RAS_RETIRE_PAGE_RESERVED,
  79	AMDGPU_RAS_RETIRE_PAGE_PENDING,
  80	AMDGPU_RAS_RETIRE_PAGE_FAULT,
  81};
  82
  83atomic_t amdgpu_ras_in_intr = ATOMIC_INIT(0);
  84
  85static bool amdgpu_ras_check_bad_page_unlock(struct amdgpu_ras *con,
  86				uint64_t addr);
  87static bool amdgpu_ras_check_bad_page(struct amdgpu_device *adev,
  88				uint64_t addr);
  89
  90void amdgpu_ras_set_error_query_ready(struct amdgpu_device *adev, bool ready)
  91{
  92	if (adev && amdgpu_ras_get_context(adev))
  93		amdgpu_ras_get_context(adev)->error_query_ready = ready;
  94}
  95
  96static bool amdgpu_ras_get_error_query_ready(struct amdgpu_device *adev)
  97{
  98	if (adev && amdgpu_ras_get_context(adev))
  99		return amdgpu_ras_get_context(adev)->error_query_ready;
 100
 101	return false;
 102}
 103
 104static int amdgpu_reserve_page_direct(struct amdgpu_device *adev, uint64_t address)
 105{
 106	struct ras_err_data err_data = {0, 0, 0, NULL};
 107	struct eeprom_table_record err_rec;
 108
 109	if ((address >= adev->gmc.mc_vram_size) ||
 110	    (address >= RAS_UMC_INJECT_ADDR_LIMIT)) {
 111		dev_warn(adev->dev,
 112		         "RAS WARN: input address 0x%llx is invalid.\n",
 113		         address);
 114		return -EINVAL;
 115	}
 116
 117	if (amdgpu_ras_check_bad_page(adev, address)) {
 118		dev_warn(adev->dev,
 119			 "RAS WARN: 0x%llx has already been marked as bad page!\n",
 120			 address);
 121		return 0;
 122	}
 123
 124	memset(&err_rec, 0x0, sizeof(struct eeprom_table_record));
 125
 126	err_rec.address = address;
 127	err_rec.retired_page = address >> AMDGPU_GPU_PAGE_SHIFT;
 128	err_rec.ts = (uint64_t)ktime_get_real_seconds();
 129	err_rec.err_type = AMDGPU_RAS_EEPROM_ERR_NON_RECOVERABLE;
 130
 131	err_data.err_addr = &err_rec;
 132	err_data.err_addr_cnt = 1;
 133
 134	if (amdgpu_bad_page_threshold != 0) {
 135		amdgpu_ras_add_bad_pages(adev, err_data.err_addr,
 136					 err_data.err_addr_cnt);
 137		amdgpu_ras_save_bad_pages(adev);
 138	}
 139
 140	dev_warn(adev->dev, "WARNING: THIS IS ONLY FOR TEST PURPOSES AND WILL CORRUPT RAS EEPROM\n");
 141	dev_warn(adev->dev, "Clear EEPROM:\n");
 142	dev_warn(adev->dev, "    echo 1 > /sys/kernel/debug/dri/0/ras/ras_eeprom_reset\n");
 143
 144	return 0;
 145}
 146
 147static ssize_t amdgpu_ras_debugfs_read(struct file *f, char __user *buf,
 148					size_t size, loff_t *pos)
 149{
 150	struct ras_manager *obj = (struct ras_manager *)file_inode(f)->i_private;
 151	struct ras_query_if info = {
 152		.head = obj->head,
 153	};
 154	ssize_t s;
 155	char val[128];
 156
 157	if (amdgpu_ras_query_error_status(obj->adev, &info))
 158		return -EINVAL;
 159
 160	s = snprintf(val, sizeof(val), "%s: %lu\n%s: %lu\n",
 161			"ue", info.ue_count,
 162			"ce", info.ce_count);
 163	if (*pos >= s)
 164		return 0;
 165
 166	s -= *pos;
 167	s = min_t(u64, s, size);
 168
 169
 170	if (copy_to_user(buf, &val[*pos], s))
 171		return -EINVAL;
 172
 173	*pos += s;
 174
 175	return s;
 176}
 177
 178static const struct file_operations amdgpu_ras_debugfs_ops = {
 179	.owner = THIS_MODULE,
 180	.read = amdgpu_ras_debugfs_read,
 181	.write = NULL,
 182	.llseek = default_llseek
 183};
 184
 185static int amdgpu_ras_find_block_id_by_name(const char *name, int *block_id)
 186{
 187	int i;
 188
 189	for (i = 0; i < ARRAY_SIZE(ras_block_string); i++) {
 190		*block_id = i;
 191		if (strcmp(name, ras_block_str(i)) == 0)
 192			return 0;
 193	}
 194	return -EINVAL;
 195}
 196
 197static int amdgpu_ras_debugfs_ctrl_parse_data(struct file *f,
 198		const char __user *buf, size_t size,
 199		loff_t *pos, struct ras_debug_if *data)
 200{
 201	ssize_t s = min_t(u64, 64, size);
 202	char str[65];
 203	char block_name[33];
 204	char err[9] = "ue";
 205	int op = -1;
 206	int block_id;
 207	uint32_t sub_block;
 208	u64 address, value;
 209
 210	if (*pos)
 211		return -EINVAL;
 212	*pos = size;
 213
 214	memset(str, 0, sizeof(str));
 215	memset(data, 0, sizeof(*data));
 216
 217	if (copy_from_user(str, buf, s))
 218		return -EINVAL;
 219
 220	if (sscanf(str, "disable %32s", block_name) == 1)
 221		op = 0;
 222	else if (sscanf(str, "enable %32s %8s", block_name, err) == 2)
 223		op = 1;
 224	else if (sscanf(str, "inject %32s %8s", block_name, err) == 2)
 225		op = 2;
 226	else if (strstr(str, "retire_page") != NULL)
 227		op = 3;
 228	else if (str[0] && str[1] && str[2] && str[3])
 229		/* ascii string, but commands are not matched. */
 230		return -EINVAL;
 231
 232	if (op != -1) {
 233		if (op == 3) {
 234			if (sscanf(str, "%*s 0x%llx", &address) != 1 &&
 235			    sscanf(str, "%*s %llu", &address) != 1)
 236				return -EINVAL;
 237
 238			data->op = op;
 239			data->inject.address = address;
 240
 241			return 0;
 242		}
 243
 244		if (amdgpu_ras_find_block_id_by_name(block_name, &block_id))
 245			return -EINVAL;
 246
 247		data->head.block = block_id;
 248		/* only ue and ce errors are supported */
 249		if (!memcmp("ue", err, 2))
 250			data->head.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE;
 251		else if (!memcmp("ce", err, 2))
 252			data->head.type = AMDGPU_RAS_ERROR__SINGLE_CORRECTABLE;
 253		else
 254			return -EINVAL;
 255
 256		data->op = op;
 257
 258		if (op == 2) {
 259			if (sscanf(str, "%*s %*s %*s 0x%x 0x%llx 0x%llx",
 260				   &sub_block, &address, &value) != 3 &&
 261			    sscanf(str, "%*s %*s %*s %u %llu %llu",
 262				   &sub_block, &address, &value) != 3)
 263				return -EINVAL;
 264			data->head.sub_block_index = sub_block;
 265			data->inject.address = address;
 266			data->inject.value = value;
 267		}
 268	} else {
 269		if (size < sizeof(*data))
 270			return -EINVAL;
 271
 272		if (copy_from_user(data, buf, sizeof(*data)))
 273			return -EINVAL;
 274	}
 275
 276	return 0;
 277}
 278
 279/**
 280 * DOC: AMDGPU RAS debugfs control interface
 281 *
 282 * The control interface accepts struct ras_debug_if which has two members.
 283 *
 284 * First member: ras_debug_if::head or ras_debug_if::inject.
 285 *
 286 * head is used to indicate which IP block will be under control.
 287 *
 288 * head has four members, they are block, type, sub_block_index, name.
 289 * block: which IP will be under control.
 290 * type: what kind of error will be enabled/disabled/injected.
 291 * sub_block_index: some IPs have subcomponets. say, GFX, sDMA.
 292 * name: the name of IP.
 293 *
 294 * inject has two more members than head, they are address, value.
 295 * As their names indicate, inject operation will write the
 296 * value to the address.
 297 *
 298 * The second member: struct ras_debug_if::op.
 299 * It has three kinds of operations.
 300 *
 301 * - 0: disable RAS on the block. Take ::head as its data.
 302 * - 1: enable RAS on the block. Take ::head as its data.
 303 * - 2: inject errors on the block. Take ::inject as its data.
 304 *
 305 * How to use the interface?
 306 *
 307 * In a program
 308 *
 309 * Copy the struct ras_debug_if in your code and initialize it.
 310 * Write the struct to the control interface.
 311 *
 312 * From shell
 313 *
 314 * .. code-block:: bash
 315 *
 316 *	echo "disable <block>" > /sys/kernel/debug/dri/<N>/ras/ras_ctrl
 317 *	echo "enable  <block> <error>" > /sys/kernel/debug/dri/<N>/ras/ras_ctrl
 318 *	echo "inject  <block> <error> <sub-block> <address> <value> > /sys/kernel/debug/dri/<N>/ras/ras_ctrl
 319 *
 320 * Where N, is the card which you want to affect.
 321 *
 322 * "disable" requires only the block.
 323 * "enable" requires the block and error type.
 324 * "inject" requires the block, error type, address, and value.
 325 *
 326 * The block is one of: umc, sdma, gfx, etc.
 327 *	see ras_block_string[] for details
 328 *
 329 * The error type is one of: ue, ce, where,
 330 *	ue is multi-uncorrectable
 331 *	ce is single-correctable
 332 *
 333 * The sub-block is a the sub-block index, pass 0 if there is no sub-block.
 334 * The address and value are hexadecimal numbers, leading 0x is optional.
 335 *
 336 * For instance,
 337 *
 338 * .. code-block:: bash
 339 *
 340 *	echo inject umc ue 0x0 0x0 0x0 > /sys/kernel/debug/dri/0/ras/ras_ctrl
 341 *	echo inject umc ce 0 0 0 > /sys/kernel/debug/dri/0/ras/ras_ctrl
 342 *	echo disable umc > /sys/kernel/debug/dri/0/ras/ras_ctrl
 343 *
 344 * How to check the result of the operation?
 345 *
 346 * To check disable/enable, see "ras" features at,
 347 * /sys/class/drm/card[0/1/2...]/device/ras/features
 348 *
 349 * To check inject, see the corresponding error count at,
 350 * /sys/class/drm/card[0/1/2...]/device/ras/[gfx|sdma|umc|...]_err_count
 351 *
 352 * .. note::
 353 *	Operations are only allowed on blocks which are supported.
 354 *	Check the "ras" mask at /sys/module/amdgpu/parameters/ras_mask
 355 *	to see which blocks support RAS on a particular asic.
 356 *
 357 */
 358static ssize_t amdgpu_ras_debugfs_ctrl_write(struct file *f, const char __user *buf,
 359		size_t size, loff_t *pos)
 360{
 361	struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
 362	struct ras_debug_if data;
 363	int ret = 0;
 364
 365	if (!amdgpu_ras_get_error_query_ready(adev)) {
 366		dev_warn(adev->dev, "RAS WARN: error injection "
 367				"currently inaccessible\n");
 368		return size;
 369	}
 370
 371	ret = amdgpu_ras_debugfs_ctrl_parse_data(f, buf, size, pos, &data);
 372	if (ret)
 373		return -EINVAL;
 374
 375	if (data.op == 3) {
 376		ret = amdgpu_reserve_page_direct(adev, data.inject.address);
 377		if (!ret)
 378			return size;
 379		else
 380			return ret;
 381	}
 382
 383	if (!amdgpu_ras_is_supported(adev, data.head.block))
 384		return -EINVAL;
 385
 386	switch (data.op) {
 387	case 0:
 388		ret = amdgpu_ras_feature_enable(adev, &data.head, 0);
 389		break;
 390	case 1:
 391		ret = amdgpu_ras_feature_enable(adev, &data.head, 1);
 392		break;
 393	case 2:
 394		if ((data.inject.address >= adev->gmc.mc_vram_size) ||
 395		    (data.inject.address >= RAS_UMC_INJECT_ADDR_LIMIT)) {
 396			dev_warn(adev->dev, "RAS WARN: input address "
 397					"0x%llx is invalid.",
 398					data.inject.address);
 399			ret = -EINVAL;
 400			break;
 401		}
 402
 403		/* umc ce/ue error injection for a bad page is not allowed */
 404		if ((data.head.block == AMDGPU_RAS_BLOCK__UMC) &&
 405		    amdgpu_ras_check_bad_page(adev, data.inject.address)) {
 406			dev_warn(adev->dev, "RAS WARN: 0x%llx has been marked "
 407					"as bad before error injection!\n",
 408					data.inject.address);
 409			break;
 410		}
 411
 412		/* data.inject.address is offset instead of absolute gpu address */
 413		ret = amdgpu_ras_error_inject(adev, &data.inject);
 414		break;
 415	default:
 416		ret = -EINVAL;
 417		break;
 418	}
 419
 420	if (ret)
 421		return -EINVAL;
 422
 423	return size;
 424}
 425
 426/**
 427 * DOC: AMDGPU RAS debugfs EEPROM table reset interface
 428 *
 429 * Some boards contain an EEPROM which is used to persistently store a list of
 430 * bad pages which experiences ECC errors in vram.  This interface provides
 431 * a way to reset the EEPROM, e.g., after testing error injection.
 432 *
 433 * Usage:
 434 *
 435 * .. code-block:: bash
 436 *
 437 *	echo 1 > ../ras/ras_eeprom_reset
 438 *
 439 * will reset EEPROM table to 0 entries.
 440 *
 441 */
 442static ssize_t amdgpu_ras_debugfs_eeprom_write(struct file *f, const char __user *buf,
 443		size_t size, loff_t *pos)
 444{
 445	struct amdgpu_device *adev =
 446		(struct amdgpu_device *)file_inode(f)->i_private;
 447	int ret;
 448
 449	ret = amdgpu_ras_eeprom_reset_table(
 450			&(amdgpu_ras_get_context(adev)->eeprom_control));
 451
 452	if (ret == 1) {
 453		amdgpu_ras_get_context(adev)->flags = RAS_DEFAULT_FLAGS;
 454		return size;
 455	} else {
 456		return -EIO;
 457	}
 458}
 459
 460static const struct file_operations amdgpu_ras_debugfs_ctrl_ops = {
 461	.owner = THIS_MODULE,
 462	.read = NULL,
 463	.write = amdgpu_ras_debugfs_ctrl_write,
 464	.llseek = default_llseek
 465};
 466
 467static const struct file_operations amdgpu_ras_debugfs_eeprom_ops = {
 468	.owner = THIS_MODULE,
 469	.read = NULL,
 470	.write = amdgpu_ras_debugfs_eeprom_write,
 471	.llseek = default_llseek
 472};
 473
 474/**
 475 * DOC: AMDGPU RAS sysfs Error Count Interface
 476 *
 477 * It allows the user to read the error count for each IP block on the gpu through
 478 * /sys/class/drm/card[0/1/2...]/device/ras/[gfx/sdma/...]_err_count
 479 *
 480 * It outputs the multiple lines which report the uncorrected (ue) and corrected
 481 * (ce) error counts.
 482 *
 483 * The format of one line is below,
 484 *
 485 * [ce|ue]: count
 486 *
 487 * Example:
 488 *
 489 * .. code-block:: bash
 490 *
 491 *	ue: 0
 492 *	ce: 1
 493 *
 494 */
 495static ssize_t amdgpu_ras_sysfs_read(struct device *dev,
 496		struct device_attribute *attr, char *buf)
 497{
 498	struct ras_manager *obj = container_of(attr, struct ras_manager, sysfs_attr);
 499	struct ras_query_if info = {
 500		.head = obj->head,
 501	};
 502
 503	if (!amdgpu_ras_get_error_query_ready(obj->adev))
 504		return sysfs_emit(buf, "Query currently inaccessible\n");
 505
 506	if (amdgpu_ras_query_error_status(obj->adev, &info))
 507		return -EINVAL;
 508
 509
 510	if (obj->adev->asic_type == CHIP_ALDEBARAN) {
 511		if (amdgpu_ras_reset_error_status(obj->adev, info.head.block))
 512			DRM_WARN("Failed to reset error counter and error status");
 513	}
 514
 515	return sysfs_emit(buf, "%s: %lu\n%s: %lu\n", "ue", info.ue_count,
 516			  "ce", info.ce_count);
 517}
 518
 519/* obj begin */
 520
 521#define get_obj(obj) do { (obj)->use++; } while (0)
 522#define alive_obj(obj) ((obj)->use)
 523
 524static inline void put_obj(struct ras_manager *obj)
 525{
 526	if (obj && (--obj->use == 0))
 527		list_del(&obj->node);
 528	if (obj && (obj->use < 0))
 529		DRM_ERROR("RAS ERROR: Unbalance obj(%s) use\n", obj->head.name);
 530}
 531
 532/* make one obj and return it. */
 533static struct ras_manager *amdgpu_ras_create_obj(struct amdgpu_device *adev,
 534		struct ras_common_if *head)
 535{
 536	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
 537	struct ras_manager *obj;
 538
 539	if (!adev->ras_enabled || !con)
 540		return NULL;
 541
 542	if (head->block >= AMDGPU_RAS_BLOCK_COUNT)
 543		return NULL;
 544
 545	obj = &con->objs[head->block];
 546	/* already exist. return obj? */
 547	if (alive_obj(obj))
 548		return NULL;
 549
 550	obj->head = *head;
 551	obj->adev = adev;
 552	list_add(&obj->node, &con->head);
 553	get_obj(obj);
 554
 555	return obj;
 556}
 557
 558/* return an obj equal to head, or the first when head is NULL */
 559struct ras_manager *amdgpu_ras_find_obj(struct amdgpu_device *adev,
 560		struct ras_common_if *head)
 561{
 562	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
 563	struct ras_manager *obj;
 564	int i;
 565
 566	if (!adev->ras_enabled || !con)
 567		return NULL;
 568
 569	if (head) {
 570		if (head->block >= AMDGPU_RAS_BLOCK_COUNT)
 571			return NULL;
 572
 573		obj = &con->objs[head->block];
 574
 575		if (alive_obj(obj)) {
 576			WARN_ON(head->block != obj->head.block);
 577			return obj;
 578		}
 579	} else {
 580		for (i = 0; i < AMDGPU_RAS_BLOCK_COUNT; i++) {
 581			obj = &con->objs[i];
 582			if (alive_obj(obj)) {
 583				WARN_ON(i != obj->head.block);
 584				return obj;
 585			}
 586		}
 587	}
 588
 589	return NULL;
 590}
 591/* obj end */
 592
 593/* feature ctl begin */
 594static int amdgpu_ras_is_feature_allowed(struct amdgpu_device *adev,
 595					 struct ras_common_if *head)
 596{
 597	return adev->ras_hw_enabled & BIT(head->block);
 598}
 599
 600static int amdgpu_ras_is_feature_enabled(struct amdgpu_device *adev,
 601		struct ras_common_if *head)
 602{
 603	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
 604
 605	return con->features & BIT(head->block);
 606}
 607
 608/*
 609 * if obj is not created, then create one.
 610 * set feature enable flag.
 611 */
 612static int __amdgpu_ras_feature_enable(struct amdgpu_device *adev,
 613		struct ras_common_if *head, int enable)
 614{
 615	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
 616	struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
 617
 618	/* If hardware does not support ras, then do not create obj.
 619	 * But if hardware support ras, we can create the obj.
 620	 * Ras framework checks con->hw_supported to see if it need do
 621	 * corresponding initialization.
 622	 * IP checks con->support to see if it need disable ras.
 623	 */
 624	if (!amdgpu_ras_is_feature_allowed(adev, head))
 625		return 0;
 626	if (!(!!enable ^ !!amdgpu_ras_is_feature_enabled(adev, head)))
 627		return 0;
 628
 629	if (enable) {
 630		if (!obj) {
 631			obj = amdgpu_ras_create_obj(adev, head);
 632			if (!obj)
 633				return -EINVAL;
 634		} else {
 635			/* In case we create obj somewhere else */
 636			get_obj(obj);
 637		}
 638		con->features |= BIT(head->block);
 639	} else {
 640		if (obj && amdgpu_ras_is_feature_enabled(adev, head)) {
 641			con->features &= ~BIT(head->block);
 642			put_obj(obj);
 643		}
 644	}
 645
 646	return 0;
 647}
 648
 649/* wrapper of psp_ras_enable_features */
 650int amdgpu_ras_feature_enable(struct amdgpu_device *adev,
 651		struct ras_common_if *head, bool enable)
 652{
 653	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
 654	union ta_ras_cmd_input *info;
 655	int ret;
 656
 657	if (!con)
 658		return -EINVAL;
 659
 660	info = kzalloc(sizeof(union ta_ras_cmd_input), GFP_KERNEL);
 661	if (!info)
 662		return -ENOMEM;
 663
 664	if (!enable) {
 665		info->disable_features = (struct ta_ras_disable_features_input) {
 666			.block_id =  amdgpu_ras_block_to_ta(head->block),
 667			.error_type = amdgpu_ras_error_to_ta(head->type),
 668		};
 669	} else {
 670		info->enable_features = (struct ta_ras_enable_features_input) {
 671			.block_id =  amdgpu_ras_block_to_ta(head->block),
 672			.error_type = amdgpu_ras_error_to_ta(head->type),
 673		};
 674	}
 675
 676	/* Do not enable if it is not allowed. */
 677	WARN_ON(enable && !amdgpu_ras_is_feature_allowed(adev, head));
 678	/* Are we alerady in that state we are going to set? */
 679	if (!(!!enable ^ !!amdgpu_ras_is_feature_enabled(adev, head))) {
 680		ret = 0;
 681		goto out;
 682	}
 683
 684	if (!amdgpu_ras_intr_triggered()) {
 685		ret = psp_ras_enable_features(&adev->psp, info, enable);
 686		if (ret) {
 687			dev_err(adev->dev, "ras %s %s failed %d\n",
 688				enable ? "enable":"disable",
 689				ras_block_str(head->block),
 690				ret);
 691			goto out;
 692		}
 693	}
 694
 695	/* setup the obj */
 696	__amdgpu_ras_feature_enable(adev, head, enable);
 697	ret = 0;
 698out:
 699	kfree(info);
 700	return ret;
 701}
 702
 703/* Only used in device probe stage and called only once. */
 704int amdgpu_ras_feature_enable_on_boot(struct amdgpu_device *adev,
 705		struct ras_common_if *head, bool enable)
 706{
 707	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
 708	int ret;
 709
 710	if (!con)
 711		return -EINVAL;
 712
 713	if (con->flags & AMDGPU_RAS_FLAG_INIT_BY_VBIOS) {
 714		if (enable) {
 715			/* There is no harm to issue a ras TA cmd regardless of
 716			 * the currecnt ras state.
 717			 * If current state == target state, it will do nothing
 718			 * But sometimes it requests driver to reset and repost
 719			 * with error code -EAGAIN.
 720			 */
 721			ret = amdgpu_ras_feature_enable(adev, head, 1);
 722			/* With old ras TA, we might fail to enable ras.
 723			 * Log it and just setup the object.
 724			 * TODO need remove this WA in the future.
 725			 */
 726			if (ret == -EINVAL) {
 727				ret = __amdgpu_ras_feature_enable(adev, head, 1);
 728				if (!ret)
 729					dev_info(adev->dev,
 730						"RAS INFO: %s setup object\n",
 731						ras_block_str(head->block));
 732			}
 733		} else {
 734			/* setup the object then issue a ras TA disable cmd.*/
 735			ret = __amdgpu_ras_feature_enable(adev, head, 1);
 736			if (ret)
 737				return ret;
 738
 739			/* gfx block ras dsiable cmd must send to ras-ta */
 740			if (head->block == AMDGPU_RAS_BLOCK__GFX)
 741				con->features |= BIT(head->block);
 742
 743			ret = amdgpu_ras_feature_enable(adev, head, 0);
 744
 745			/* clean gfx block ras features flag */
 746			if (adev->ras_enabled && head->block == AMDGPU_RAS_BLOCK__GFX)
 747				con->features &= ~BIT(head->block);
 748		}
 749	} else
 750		ret = amdgpu_ras_feature_enable(adev, head, enable);
 751
 752	return ret;
 753}
 754
 755static int amdgpu_ras_disable_all_features(struct amdgpu_device *adev,
 756		bool bypass)
 757{
 758	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
 759	struct ras_manager *obj, *tmp;
 760
 761	list_for_each_entry_safe(obj, tmp, &con->head, node) {
 762		/* bypass psp.
 763		 * aka just release the obj and corresponding flags
 764		 */
 765		if (bypass) {
 766			if (__amdgpu_ras_feature_enable(adev, &obj->head, 0))
 767				break;
 768		} else {
 769			if (amdgpu_ras_feature_enable(adev, &obj->head, 0))
 770				break;
 771		}
 772	}
 773
 774	return con->features;
 775}
 776
 777static int amdgpu_ras_enable_all_features(struct amdgpu_device *adev,
 778		bool bypass)
 779{
 780	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
 781	int ras_block_count = AMDGPU_RAS_BLOCK_COUNT;
 782	int i;
 783	const enum amdgpu_ras_error_type default_ras_type =
 784		AMDGPU_RAS_ERROR__NONE;
 785
 786	for (i = 0; i < ras_block_count; i++) {
 787		struct ras_common_if head = {
 788			.block = i,
 789			.type = default_ras_type,
 790			.sub_block_index = 0,
 791		};
 792		strcpy(head.name, ras_block_str(i));
 793		if (bypass) {
 794			/*
 795			 * bypass psp. vbios enable ras for us.
 796			 * so just create the obj
 797			 */
 798			if (__amdgpu_ras_feature_enable(adev, &head, 1))
 799				break;
 800		} else {
 801			if (amdgpu_ras_feature_enable(adev, &head, 1))
 802				break;
 803		}
 804	}
 805
 806	return con->features;
 807}
 808/* feature ctl end */
 809
 810/* query/inject/cure begin */
 811int amdgpu_ras_query_error_status(struct amdgpu_device *adev,
 812				  struct ras_query_if *info)
 813{
 814	struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head);
 815	struct ras_err_data err_data = {0, 0, 0, NULL};
 816	int i;
 817
 818	if (!obj)
 819		return -EINVAL;
 820
 821	switch (info->head.block) {
 822	case AMDGPU_RAS_BLOCK__UMC:
 823		if (adev->umc.ras_funcs &&
 824		    adev->umc.ras_funcs->query_ras_error_count)
 825			adev->umc.ras_funcs->query_ras_error_count(adev, &err_data);
 826		/* umc query_ras_error_address is also responsible for clearing
 827		 * error status
 828		 */
 829		if (adev->umc.ras_funcs &&
 830		    adev->umc.ras_funcs->query_ras_error_address)
 831			adev->umc.ras_funcs->query_ras_error_address(adev, &err_data);
 832		break;
 833	case AMDGPU_RAS_BLOCK__SDMA:
 834		if (adev->sdma.funcs->query_ras_error_count) {
 835			for (i = 0; i < adev->sdma.num_instances; i++)
 836				adev->sdma.funcs->query_ras_error_count(adev, i,
 837									&err_data);
 838		}
 839		break;
 840	case AMDGPU_RAS_BLOCK__GFX:
 841		if (adev->gfx.ras_funcs &&
 842		    adev->gfx.ras_funcs->query_ras_error_count)
 843			adev->gfx.ras_funcs->query_ras_error_count(adev, &err_data);
 844
 845		if (adev->gfx.ras_funcs &&
 846		    adev->gfx.ras_funcs->query_ras_error_status)
 847			adev->gfx.ras_funcs->query_ras_error_status(adev);
 848		break;
 849	case AMDGPU_RAS_BLOCK__MMHUB:
 850		if (adev->mmhub.ras_funcs &&
 851		    adev->mmhub.ras_funcs->query_ras_error_count)
 852			adev->mmhub.ras_funcs->query_ras_error_count(adev, &err_data);
 853
 854		if (adev->mmhub.ras_funcs &&
 855		    adev->mmhub.ras_funcs->query_ras_error_status)
 856			adev->mmhub.ras_funcs->query_ras_error_status(adev);
 857		break;
 858	case AMDGPU_RAS_BLOCK__PCIE_BIF:
 859		if (adev->nbio.ras_funcs &&
 860		    adev->nbio.ras_funcs->query_ras_error_count)
 861			adev->nbio.ras_funcs->query_ras_error_count(adev, &err_data);
 862		break;
 863	case AMDGPU_RAS_BLOCK__XGMI_WAFL:
 864		if (adev->gmc.xgmi.ras_funcs &&
 865		    adev->gmc.xgmi.ras_funcs->query_ras_error_count)
 866			adev->gmc.xgmi.ras_funcs->query_ras_error_count(adev, &err_data);
 867		break;
 868	case AMDGPU_RAS_BLOCK__HDP:
 869		if (adev->hdp.ras_funcs &&
 870		    adev->hdp.ras_funcs->query_ras_error_count)
 871			adev->hdp.ras_funcs->query_ras_error_count(adev, &err_data);
 872		break;
 873	default:
 874		break;
 875	}
 876
 877	obj->err_data.ue_count += err_data.ue_count;
 878	obj->err_data.ce_count += err_data.ce_count;
 879
 880	info->ue_count = obj->err_data.ue_count;
 881	info->ce_count = obj->err_data.ce_count;
 882
 883	if (err_data.ce_count) {
 884		if (adev->smuio.funcs &&
 885		    adev->smuio.funcs->get_socket_id &&
 886		    adev->smuio.funcs->get_die_id) {
 887			dev_info(adev->dev, "socket: %d, die: %d "
 888					"%ld correctable hardware errors "
 889					"detected in %s block, no user "
 890					"action is needed.\n",
 891					adev->smuio.funcs->get_socket_id(adev),
 892					adev->smuio.funcs->get_die_id(adev),
 893					obj->err_data.ce_count,
 894					ras_block_str(info->head.block));
 895		} else {
 896			dev_info(adev->dev, "%ld correctable hardware errors "
 897					"detected in %s block, no user "
 898					"action is needed.\n",
 899					obj->err_data.ce_count,
 900					ras_block_str(info->head.block));
 901		}
 902	}
 903	if (err_data.ue_count) {
 904		if (adev->smuio.funcs &&
 905		    adev->smuio.funcs->get_socket_id &&
 906		    adev->smuio.funcs->get_die_id) {
 907			dev_info(adev->dev, "socket: %d, die: %d "
 908					"%ld uncorrectable hardware errors "
 909					"detected in %s block\n",
 910					adev->smuio.funcs->get_socket_id(adev),
 911					adev->smuio.funcs->get_die_id(adev),
 912					obj->err_data.ue_count,
 913					ras_block_str(info->head.block));
 914		} else {
 915			dev_info(adev->dev, "%ld uncorrectable hardware errors "
 916					"detected in %s block\n",
 917					obj->err_data.ue_count,
 918					ras_block_str(info->head.block));
 919		}
 920	}
 921
 922	return 0;
 923}
 924
 925int amdgpu_ras_reset_error_status(struct amdgpu_device *adev,
 926		enum amdgpu_ras_block block)
 927{
 928	if (!amdgpu_ras_is_supported(adev, block))
 929		return -EINVAL;
 930
 931	switch (block) {
 932	case AMDGPU_RAS_BLOCK__GFX:
 933		if (adev->gfx.ras_funcs &&
 934		    adev->gfx.ras_funcs->reset_ras_error_count)
 935			adev->gfx.ras_funcs->reset_ras_error_count(adev);
 936
 937		if (adev->gfx.ras_funcs &&
 938		    adev->gfx.ras_funcs->reset_ras_error_status)
 939			adev->gfx.ras_funcs->reset_ras_error_status(adev);
 940		break;
 941	case AMDGPU_RAS_BLOCK__MMHUB:
 942		if (adev->mmhub.ras_funcs &&
 943		    adev->mmhub.ras_funcs->reset_ras_error_count)
 944			adev->mmhub.ras_funcs->reset_ras_error_count(adev);
 945
 946		if (adev->mmhub.ras_funcs &&
 947		    adev->mmhub.ras_funcs->reset_ras_error_status)
 948			adev->mmhub.ras_funcs->reset_ras_error_status(adev);
 949		break;
 950	case AMDGPU_RAS_BLOCK__SDMA:
 951		if (adev->sdma.funcs->reset_ras_error_count)
 952			adev->sdma.funcs->reset_ras_error_count(adev);
 953		break;
 954	case AMDGPU_RAS_BLOCK__HDP:
 955		if (adev->hdp.ras_funcs &&
 956		    adev->hdp.ras_funcs->reset_ras_error_count)
 957			adev->hdp.ras_funcs->reset_ras_error_count(adev);
 958		break;
 959	default:
 960		break;
 961	}
 962
 963	return 0;
 964}
 965
 966/* Trigger XGMI/WAFL error */
 967static int amdgpu_ras_error_inject_xgmi(struct amdgpu_device *adev,
 968				 struct ta_ras_trigger_error_input *block_info)
 969{
 970	int ret;
 971
 972	if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_DISALLOW))
 973		dev_warn(adev->dev, "Failed to disallow df cstate");
 974
 975	if (amdgpu_dpm_allow_xgmi_power_down(adev, false))
 976		dev_warn(adev->dev, "Failed to disallow XGMI power down");
 977
 978	ret = psp_ras_trigger_error(&adev->psp, block_info);
 979
 980	if (amdgpu_ras_intr_triggered())
 981		return ret;
 982
 983	if (amdgpu_dpm_allow_xgmi_power_down(adev, true))
 984		dev_warn(adev->dev, "Failed to allow XGMI power down");
 985
 986	if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_ALLOW))
 987		dev_warn(adev->dev, "Failed to allow df cstate");
 988
 989	return ret;
 990}
 991
 992/* wrapper of psp_ras_trigger_error */
 993int amdgpu_ras_error_inject(struct amdgpu_device *adev,
 994		struct ras_inject_if *info)
 995{
 996	struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head);
 997	struct ta_ras_trigger_error_input block_info = {
 998		.block_id =  amdgpu_ras_block_to_ta(info->head.block),
 999		.inject_error_type = amdgpu_ras_error_to_ta(info->head.type),
1000		.sub_block_index = info->head.sub_block_index,
1001		.address = info->address,
1002		.value = info->value,
1003	};
1004	int ret = 0;
1005
1006	if (!obj)
1007		return -EINVAL;
1008
1009	/* Calculate XGMI relative offset */
1010	if (adev->gmc.xgmi.num_physical_nodes > 1) {
1011		block_info.address =
1012			amdgpu_xgmi_get_relative_phy_addr(adev,
1013							  block_info.address);
1014	}
1015
1016	switch (info->head.block) {
1017	case AMDGPU_RAS_BLOCK__GFX:
1018		if (adev->gfx.ras_funcs &&
1019		    adev->gfx.ras_funcs->ras_error_inject)
1020			ret = adev->gfx.ras_funcs->ras_error_inject(adev, info);
1021		else
1022			ret = -EINVAL;
1023		break;
1024	case AMDGPU_RAS_BLOCK__UMC:
1025	case AMDGPU_RAS_BLOCK__SDMA:
1026	case AMDGPU_RAS_BLOCK__MMHUB:
1027	case AMDGPU_RAS_BLOCK__PCIE_BIF:
1028		ret = psp_ras_trigger_error(&adev->psp, &block_info);
1029		break;
1030	case AMDGPU_RAS_BLOCK__XGMI_WAFL:
1031		ret = amdgpu_ras_error_inject_xgmi(adev, &block_info);
1032		break;
1033	default:
1034		dev_info(adev->dev, "%s error injection is not supported yet\n",
1035			 ras_block_str(info->head.block));
1036		ret = -EINVAL;
1037	}
1038
1039	if (ret)
1040		dev_err(adev->dev, "ras inject %s failed %d\n",
1041			ras_block_str(info->head.block), ret);
1042
1043	return ret;
1044}
1045
1046/**
1047 * amdgpu_ras_query_error_count -- Get error counts of all IPs
1048 * adev: pointer to AMD GPU device
1049 * ce_count: pointer to an integer to be set to the count of correctible errors.
1050 * ue_count: pointer to an integer to be set to the count of uncorrectible
1051 * errors.
1052 *
1053 * If set, @ce_count or @ue_count, count and return the corresponding
1054 * error counts in those integer pointers. Return 0 if the device
1055 * supports RAS. Return -EOPNOTSUPP if the device doesn't support RAS.
1056 */
1057int amdgpu_ras_query_error_count(struct amdgpu_device *adev,
1058				 unsigned long *ce_count,
1059				 unsigned long *ue_count)
1060{
1061	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1062	struct ras_manager *obj;
1063	unsigned long ce, ue;
1064
1065	if (!adev->ras_enabled || !con)
1066		return -EOPNOTSUPP;
1067
1068	/* Don't count since no reporting.
1069	 */
1070	if (!ce_count && !ue_count)
1071		return 0;
1072
1073	ce = 0;
1074	ue = 0;
1075	list_for_each_entry(obj, &con->head, node) {
1076		struct ras_query_if info = {
1077			.head = obj->head,
1078		};
1079		int res;
1080
1081		res = amdgpu_ras_query_error_status(adev, &info);
1082		if (res)
1083			return res;
1084
1085		ce += info.ce_count;
1086		ue += info.ue_count;
1087	}
1088
1089	if (ce_count)
1090		*ce_count = ce;
1091
1092	if (ue_count)
1093		*ue_count = ue;
1094
1095	return 0;
1096}
1097/* query/inject/cure end */
1098
1099
1100/* sysfs begin */
1101
1102static int amdgpu_ras_badpages_read(struct amdgpu_device *adev,
1103		struct ras_badpage **bps, unsigned int *count);
1104
1105static char *amdgpu_ras_badpage_flags_str(unsigned int flags)
1106{
1107	switch (flags) {
1108	case AMDGPU_RAS_RETIRE_PAGE_RESERVED:
1109		return "R";
1110	case AMDGPU_RAS_RETIRE_PAGE_PENDING:
1111		return "P";
1112	case AMDGPU_RAS_RETIRE_PAGE_FAULT:
1113	default:
1114		return "F";
1115	}
1116}
1117
1118/**
1119 * DOC: AMDGPU RAS sysfs gpu_vram_bad_pages Interface
1120 *
1121 * It allows user to read the bad pages of vram on the gpu through
1122 * /sys/class/drm/card[0/1/2...]/device/ras/gpu_vram_bad_pages
1123 *
1124 * It outputs multiple lines, and each line stands for one gpu page.
1125 *
1126 * The format of one line is below,
1127 * gpu pfn : gpu page size : flags
1128 *
1129 * gpu pfn and gpu page size are printed in hex format.
1130 * flags can be one of below character,
1131 *
1132 * R: reserved, this gpu page is reserved and not able to use.
1133 *
1134 * P: pending for reserve, this gpu page is marked as bad, will be reserved
1135 * in next window of page_reserve.
1136 *
1137 * F: unable to reserve. this gpu page can't be reserved due to some reasons.
1138 *
1139 * Examples:
1140 *
1141 * .. code-block:: bash
1142 *
1143 *	0x00000001 : 0x00001000 : R
1144 *	0x00000002 : 0x00001000 : P
1145 *
1146 */
1147
1148static ssize_t amdgpu_ras_sysfs_badpages_read(struct file *f,
1149		struct kobject *kobj, struct bin_attribute *attr,
1150		char *buf, loff_t ppos, size_t count)
1151{
1152	struct amdgpu_ras *con =
1153		container_of(attr, struct amdgpu_ras, badpages_attr);
1154	struct amdgpu_device *adev = con->adev;
1155	const unsigned int element_size =
1156		sizeof("0xabcdabcd : 0x12345678 : R\n") - 1;
1157	unsigned int start = div64_ul(ppos + element_size - 1, element_size);
1158	unsigned int end = div64_ul(ppos + count - 1, element_size);
1159	ssize_t s = 0;
1160	struct ras_badpage *bps = NULL;
1161	unsigned int bps_count = 0;
1162
1163	memset(buf, 0, count);
1164
1165	if (amdgpu_ras_badpages_read(adev, &bps, &bps_count))
1166		return 0;
1167
1168	for (; start < end && start < bps_count; start++)
1169		s += scnprintf(&buf[s], element_size + 1,
1170				"0x%08x : 0x%08x : %1s\n",
1171				bps[start].bp,
1172				bps[start].size,
1173				amdgpu_ras_badpage_flags_str(bps[start].flags));
1174
1175	kfree(bps);
1176
1177	return s;
1178}
1179
1180static ssize_t amdgpu_ras_sysfs_features_read(struct device *dev,
1181		struct device_attribute *attr, char *buf)
1182{
1183	struct amdgpu_ras *con =
1184		container_of(attr, struct amdgpu_ras, features_attr);
1185
1186	return scnprintf(buf, PAGE_SIZE, "feature mask: 0x%x\n", con->features);
1187}
1188
1189static void amdgpu_ras_sysfs_remove_bad_page_node(struct amdgpu_device *adev)
1190{
1191	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1192
1193	sysfs_remove_file_from_group(&adev->dev->kobj,
1194				&con->badpages_attr.attr,
1195				RAS_FS_NAME);
1196}
1197
1198static int amdgpu_ras_sysfs_remove_feature_node(struct amdgpu_device *adev)
1199{
1200	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1201	struct attribute *attrs[] = {
1202		&con->features_attr.attr,
1203		NULL
1204	};
1205	struct attribute_group group = {
1206		.name = RAS_FS_NAME,
1207		.attrs = attrs,
1208	};
1209
1210	sysfs_remove_group(&adev->dev->kobj, &group);
1211
1212	return 0;
1213}
1214
1215int amdgpu_ras_sysfs_create(struct amdgpu_device *adev,
1216		struct ras_fs_if *head)
1217{
1218	struct ras_manager *obj = amdgpu_ras_find_obj(adev, &head->head);
1219
1220	if (!obj || obj->attr_inuse)
1221		return -EINVAL;
1222
1223	get_obj(obj);
1224
1225	memcpy(obj->fs_data.sysfs_name,
1226			head->sysfs_name,
1227			sizeof(obj->fs_data.sysfs_name));
1228
1229	obj->sysfs_attr = (struct device_attribute){
1230		.attr = {
1231			.name = obj->fs_data.sysfs_name,
1232			.mode = S_IRUGO,
1233		},
1234			.show = amdgpu_ras_sysfs_read,
1235	};
1236	sysfs_attr_init(&obj->sysfs_attr.attr);
1237
1238	if (sysfs_add_file_to_group(&adev->dev->kobj,
1239				&obj->sysfs_attr.attr,
1240				RAS_FS_NAME)) {
1241		put_obj(obj);
1242		return -EINVAL;
1243	}
1244
1245	obj->attr_inuse = 1;
1246
1247	return 0;
1248}
1249
1250int amdgpu_ras_sysfs_remove(struct amdgpu_device *adev,
1251		struct ras_common_if *head)
1252{
1253	struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
1254
1255	if (!obj || !obj->attr_inuse)
1256		return -EINVAL;
1257
1258	sysfs_remove_file_from_group(&adev->dev->kobj,
1259				&obj->sysfs_attr.attr,
1260				RAS_FS_NAME);
1261	obj->attr_inuse = 0;
1262	put_obj(obj);
1263
1264	return 0;
1265}
1266
1267static int amdgpu_ras_sysfs_remove_all(struct amdgpu_device *adev)
1268{
1269	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1270	struct ras_manager *obj, *tmp;
1271
1272	list_for_each_entry_safe(obj, tmp, &con->head, node) {
1273		amdgpu_ras_sysfs_remove(adev, &obj->head);
1274	}
1275
1276	if (amdgpu_bad_page_threshold != 0)
1277		amdgpu_ras_sysfs_remove_bad_page_node(adev);
1278
1279	amdgpu_ras_sysfs_remove_feature_node(adev);
1280
1281	return 0;
1282}
1283/* sysfs end */
1284
1285/**
1286 * DOC: AMDGPU RAS Reboot Behavior for Unrecoverable Errors
1287 *
1288 * Normally when there is an uncorrectable error, the driver will reset
1289 * the GPU to recover.  However, in the event of an unrecoverable error,
1290 * the driver provides an interface to reboot the system automatically
1291 * in that event.
1292 *
1293 * The following file in debugfs provides that interface:
1294 * /sys/kernel/debug/dri/[0/1/2...]/ras/auto_reboot
1295 *
1296 * Usage:
1297 *
1298 * .. code-block:: bash
1299 *
1300 *	echo true > .../ras/auto_reboot
1301 *
1302 */
1303/* debugfs begin */
1304static struct dentry *amdgpu_ras_debugfs_create_ctrl_node(struct amdgpu_device *adev)
1305{
1306	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1307	struct drm_minor  *minor = adev_to_drm(adev)->primary;
1308	struct dentry     *dir;
1309
1310	dir = debugfs_create_dir(RAS_FS_NAME, minor->debugfs_root);
1311	debugfs_create_file("ras_ctrl", S_IWUGO | S_IRUGO, dir, adev,
1312			    &amdgpu_ras_debugfs_ctrl_ops);
1313	debugfs_create_file("ras_eeprom_reset", S_IWUGO | S_IRUGO, dir, adev,
1314			    &amdgpu_ras_debugfs_eeprom_ops);
1315	debugfs_create_u32("bad_page_cnt_threshold", 0444, dir,
1316			   &con->bad_page_cnt_threshold);
1317	debugfs_create_x32("ras_hw_enabled", 0444, dir, &adev->ras_hw_enabled);
1318	debugfs_create_x32("ras_enabled", 0444, dir, &adev->ras_enabled);
1319
1320	/*
1321	 * After one uncorrectable error happens, usually GPU recovery will
1322	 * be scheduled. But due to the known problem in GPU recovery failing
1323	 * to bring GPU back, below interface provides one direct way to
1324	 * user to reboot system automatically in such case within
1325	 * ERREVENT_ATHUB_INTERRUPT generated. Normal GPU recovery routine
1326	 * will never be called.
1327	 */
1328	debugfs_create_bool("auto_reboot", S_IWUGO | S_IRUGO, dir, &con->reboot);
1329
1330	/*
1331	 * User could set this not to clean up hardware's error count register
1332	 * of RAS IPs during ras recovery.
1333	 */
1334	debugfs_create_bool("disable_ras_err_cnt_harvest", 0644, dir,
1335			    &con->disable_ras_err_cnt_harvest);
1336	return dir;
1337}
1338
1339static void amdgpu_ras_debugfs_create(struct amdgpu_device *adev,
1340				      struct ras_fs_if *head,
1341				      struct dentry *dir)
1342{
1343	struct ras_manager *obj = amdgpu_ras_find_obj(adev, &head->head);
1344
1345	if (!obj || !dir)
1346		return;
1347
1348	get_obj(obj);
1349
1350	memcpy(obj->fs_data.debugfs_name,
1351			head->debugfs_name,
1352			sizeof(obj->fs_data.debugfs_name));
1353
1354	debugfs_create_file(obj->fs_data.debugfs_name, S_IWUGO | S_IRUGO, dir,
1355			    obj, &amdgpu_ras_debugfs_ops);
1356}
1357
1358void amdgpu_ras_debugfs_create_all(struct amdgpu_device *adev)
1359{
1360	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1361	struct dentry *dir;
1362	struct ras_manager *obj;
1363	struct ras_fs_if fs_info;
1364
1365	/*
1366	 * it won't be called in resume path, no need to check
1367	 * suspend and gpu reset status
1368	 */
1369	if (!IS_ENABLED(CONFIG_DEBUG_FS) || !con)
1370		return;
1371
1372	dir = amdgpu_ras_debugfs_create_ctrl_node(adev);
1373
1374	list_for_each_entry(obj, &con->head, node) {
1375		if (amdgpu_ras_is_supported(adev, obj->head.block) &&
1376			(obj->attr_inuse == 1)) {
1377			sprintf(fs_info.debugfs_name, "%s_err_inject",
1378					ras_block_str(obj->head.block));
1379			fs_info.head = obj->head;
1380			amdgpu_ras_debugfs_create(adev, &fs_info, dir);
1381		}
1382	}
1383}
1384
1385/* debugfs end */
1386
1387/* ras fs */
1388static BIN_ATTR(gpu_vram_bad_pages, S_IRUGO,
1389		amdgpu_ras_sysfs_badpages_read, NULL, 0);
1390static DEVICE_ATTR(features, S_IRUGO,
1391		amdgpu_ras_sysfs_features_read, NULL);
1392static int amdgpu_ras_fs_init(struct amdgpu_device *adev)
1393{
1394	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1395	struct attribute_group group = {
1396		.name = RAS_FS_NAME,
1397	};
1398	struct attribute *attrs[] = {
1399		&con->features_attr.attr,
1400		NULL
1401	};
1402	struct bin_attribute *bin_attrs[] = {
1403		NULL,
1404		NULL,
1405	};
1406	int r;
1407
1408	/* add features entry */
1409	con->features_attr = dev_attr_features;
1410	group.attrs = attrs;
1411	sysfs_attr_init(attrs[0]);
1412
1413	if (amdgpu_bad_page_threshold != 0) {
1414		/* add bad_page_features entry */
1415		bin_attr_gpu_vram_bad_pages.private = NULL;
1416		con->badpages_attr = bin_attr_gpu_vram_bad_pages;
1417		bin_attrs[0] = &con->badpages_attr;
1418		group.bin_attrs = bin_attrs;
1419		sysfs_bin_attr_init(bin_attrs[0]);
1420	}
1421
1422	r = sysfs_create_group(&adev->dev->kobj, &group);
1423	if (r)
1424		dev_err(adev->dev, "Failed to create RAS sysfs group!");
1425
1426	return 0;
1427}
1428
1429static int amdgpu_ras_fs_fini(struct amdgpu_device *adev)
1430{
1431	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1432	struct ras_manager *con_obj, *ip_obj, *tmp;
1433
1434	if (IS_ENABLED(CONFIG_DEBUG_FS)) {
1435		list_for_each_entry_safe(con_obj, tmp, &con->head, node) {
1436			ip_obj = amdgpu_ras_find_obj(adev, &con_obj->head);
1437			if (ip_obj)
1438				put_obj(ip_obj);
1439		}
1440	}
1441
1442	amdgpu_ras_sysfs_remove_all(adev);
1443	return 0;
1444}
1445/* ras fs end */
1446
1447/* ih begin */
1448static void amdgpu_ras_interrupt_handler(struct ras_manager *obj)
1449{
1450	struct ras_ih_data *data = &obj->ih_data;
1451	struct amdgpu_iv_entry entry;
1452	int ret;
1453	struct ras_err_data err_data = {0, 0, 0, NULL};
1454
1455	while (data->rptr != data->wptr) {
1456		rmb();
1457		memcpy(&entry, &data->ring[data->rptr],
1458				data->element_size);
1459
1460		wmb();
1461		data->rptr = (data->aligned_element_size +
1462				data->rptr) % data->ring_size;
1463
1464		/* Let IP handle its data, maybe we need get the output
1465		 * from the callback to udpate the error type/count, etc
1466		 */
1467		if (data->cb) {
1468			ret = data->cb(obj->adev, &err_data, &entry);
1469			/* ue will trigger an interrupt, and in that case
1470			 * we need do a reset to recovery the whole system.
1471			 * But leave IP do that recovery, here we just dispatch
1472			 * the error.
1473			 */
1474			if (ret == AMDGPU_RAS_SUCCESS) {
1475				/* these counts could be left as 0 if
1476				 * some blocks do not count error number
1477				 */
1478				obj->err_data.ue_count += err_data.ue_count;
1479				obj->err_data.ce_count += err_data.ce_count;
1480			}
1481		}
1482	}
1483}
1484
1485static void amdgpu_ras_interrupt_process_handler(struct work_struct *work)
1486{
1487	struct ras_ih_data *data =
1488		container_of(work, struct ras_ih_data, ih_work);
1489	struct ras_manager *obj =
1490		container_of(data, struct ras_manager, ih_data);
1491
1492	amdgpu_ras_interrupt_handler(obj);
1493}
1494
1495int amdgpu_ras_interrupt_dispatch(struct amdgpu_device *adev,
1496		struct ras_dispatch_if *info)
1497{
1498	struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head);
1499	struct ras_ih_data *data = &obj->ih_data;
1500
1501	if (!obj)
1502		return -EINVAL;
1503
1504	if (data->inuse == 0)
1505		return 0;
1506
1507	/* Might be overflow... */
1508	memcpy(&data->ring[data->wptr], info->entry,
1509			data->element_size);
1510
1511	wmb();
1512	data->wptr = (data->aligned_element_size +
1513			data->wptr) % data->ring_size;
1514
1515	schedule_work(&data->ih_work);
1516
1517	return 0;
1518}
1519
1520int amdgpu_ras_interrupt_remove_handler(struct amdgpu_device *adev,
1521		struct ras_ih_if *info)
1522{
1523	struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head);
1524	struct ras_ih_data *data;
1525
1526	if (!obj)
1527		return -EINVAL;
1528
1529	data = &obj->ih_data;
1530	if (data->inuse == 0)
1531		return 0;
1532
1533	cancel_work_sync(&data->ih_work);
1534
1535	kfree(data->ring);
1536	memset(data, 0, sizeof(*data));
1537	put_obj(obj);
1538
1539	return 0;
1540}
1541
1542int amdgpu_ras_interrupt_add_handler(struct amdgpu_device *adev,
1543		struct ras_ih_if *info)
1544{
1545	struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head);
1546	struct ras_ih_data *data;
1547
1548	if (!obj) {
1549		/* in case we registe the IH before enable ras feature */
1550		obj = amdgpu_ras_create_obj(adev, &info->head);
1551		if (!obj)
1552			return -EINVAL;
1553	} else
1554		get_obj(obj);
1555
1556	data = &obj->ih_data;
1557	/* add the callback.etc */
1558	*data = (struct ras_ih_data) {
1559		.inuse = 0,
1560		.cb = info->cb,
1561		.element_size = sizeof(struct amdgpu_iv_entry),
1562		.rptr = 0,
1563		.wptr = 0,
1564	};
1565
1566	INIT_WORK(&data->ih_work, amdgpu_ras_interrupt_process_handler);
1567
1568	data->aligned_element_size = ALIGN(data->element_size, 8);
1569	/* the ring can store 64 iv entries. */
1570	data->ring_size = 64 * data->aligned_element_size;
1571	data->ring = kmalloc(data->ring_size, GFP_KERNEL);
1572	if (!data->ring) {
1573		put_obj(obj);
1574		return -ENOMEM;
1575	}
1576
1577	/* IH is ready */
1578	data->inuse = 1;
1579
1580	return 0;
1581}
1582
1583static int amdgpu_ras_interrupt_remove_all(struct amdgpu_device *adev)
1584{
1585	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1586	struct ras_manager *obj, *tmp;
1587
1588	list_for_each_entry_safe(obj, tmp, &con->head, node) {
1589		struct ras_ih_if info = {
1590			.head = obj->head,
1591		};
1592		amdgpu_ras_interrupt_remove_handler(adev, &info);
1593	}
1594
1595	return 0;
1596}
1597/* ih end */
1598
1599/* traversal all IPs except NBIO to query error counter */
1600static void amdgpu_ras_log_on_err_counter(struct amdgpu_device *adev)
1601{
1602	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1603	struct ras_manager *obj;
1604
1605	if (!adev->ras_enabled || !con)
1606		return;
1607
1608	list_for_each_entry(obj, &con->head, node) {
1609		struct ras_query_if info = {
1610			.head = obj->head,
1611		};
1612
1613		/*
1614		 * PCIE_BIF IP has one different isr by ras controller
1615		 * interrupt, the specific ras counter query will be
1616		 * done in that isr. So skip such block from common
1617		 * sync flood interrupt isr calling.
1618		 */
1619		if (info.head.block == AMDGPU_RAS_BLOCK__PCIE_BIF)
1620			continue;
1621
1622		amdgpu_ras_query_error_status(adev, &info);
1623	}
1624}
1625
1626/* Parse RdRspStatus and WrRspStatus */
1627static void amdgpu_ras_error_status_query(struct amdgpu_device *adev,
1628					  struct ras_query_if *info)
1629{
1630	/*
1631	 * Only two block need to query read/write
1632	 * RspStatus at current state
1633	 */
1634	switch (info->head.block) {
1635	case AMDGPU_RAS_BLOCK__GFX:
1636		if (adev->gfx.ras_funcs &&
1637		    adev->gfx.ras_funcs->query_ras_error_status)
1638			adev->gfx.ras_funcs->query_ras_error_status(adev);
1639		break;
1640	case AMDGPU_RAS_BLOCK__MMHUB:
1641		if (adev->mmhub.ras_funcs &&
1642		    adev->mmhub.ras_funcs->query_ras_error_status)
1643			adev->mmhub.ras_funcs->query_ras_error_status(adev);
1644		break;
1645	default:
1646		break;
1647	}
1648}
1649
1650static void amdgpu_ras_query_err_status(struct amdgpu_device *adev)
1651{
1652	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1653	struct ras_manager *obj;
1654
1655	if (!adev->ras_enabled || !con)
1656		return;
1657
1658	list_for_each_entry(obj, &con->head, node) {
1659		struct ras_query_if info = {
1660			.head = obj->head,
1661		};
1662
1663		amdgpu_ras_error_status_query(adev, &info);
1664	}
1665}
1666
1667/* recovery begin */
1668
1669/* return 0 on success.
1670 * caller need free bps.
1671 */
1672static int amdgpu_ras_badpages_read(struct amdgpu_device *adev,
1673		struct ras_badpage **bps, unsigned int *count)
1674{
1675	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1676	struct ras_err_handler_data *data;
1677	int i = 0;
1678	int ret = 0, status;
1679
1680	if (!con || !con->eh_data || !bps || !count)
1681		return -EINVAL;
1682
1683	mutex_lock(&con->recovery_lock);
1684	data = con->eh_data;
1685	if (!data || data->count == 0) {
1686		*bps = NULL;
1687		ret = -EINVAL;
1688		goto out;
1689	}
1690
1691	*bps = kmalloc(sizeof(struct ras_badpage) * data->count, GFP_KERNEL);
1692	if (!*bps) {
1693		ret = -ENOMEM;
1694		goto out;
1695	}
1696
1697	for (; i < data->count; i++) {
1698		(*bps)[i] = (struct ras_badpage){
1699			.bp = data->bps[i].retired_page,
1700			.size = AMDGPU_GPU_PAGE_SIZE,
1701			.flags = AMDGPU_RAS_RETIRE_PAGE_RESERVED,
1702		};
1703		status = amdgpu_vram_mgr_query_page_status(
1704				ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM),
1705				data->bps[i].retired_page);
1706		if (status == -EBUSY)
1707			(*bps)[i].flags = AMDGPU_RAS_RETIRE_PAGE_PENDING;
1708		else if (status == -ENOENT)
1709			(*bps)[i].flags = AMDGPU_RAS_RETIRE_PAGE_FAULT;
1710	}
1711
1712	*count = data->count;
1713out:
1714	mutex_unlock(&con->recovery_lock);
1715	return ret;
1716}
1717
1718static void amdgpu_ras_do_recovery(struct work_struct *work)
1719{
1720	struct amdgpu_ras *ras =
1721		container_of(work, struct amdgpu_ras, recovery_work);
1722	struct amdgpu_device *remote_adev = NULL;
1723	struct amdgpu_device *adev = ras->adev;
1724	struct list_head device_list, *device_list_handle =  NULL;
1725
1726	if (!ras->disable_ras_err_cnt_harvest) {
1727		struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev);
1728
1729		/* Build list of devices to query RAS related errors */
1730		if  (hive && adev->gmc.xgmi.num_physical_nodes > 1) {
1731			device_list_handle = &hive->device_list;
1732		} else {
1733			INIT_LIST_HEAD(&device_list);
1734			list_add_tail(&adev->gmc.xgmi.head, &device_list);
1735			device_list_handle = &device_list;
1736		}
1737
1738		list_for_each_entry(remote_adev,
1739				device_list_handle, gmc.xgmi.head) {
1740			amdgpu_ras_query_err_status(remote_adev);
1741			amdgpu_ras_log_on_err_counter(remote_adev);
1742		}
1743
1744		amdgpu_put_xgmi_hive(hive);
1745	}
1746
1747	if (amdgpu_device_should_recover_gpu(ras->adev))
1748		amdgpu_device_gpu_recover(ras->adev, NULL);
1749	atomic_set(&ras->in_recovery, 0);
1750}
1751
1752/* alloc/realloc bps array */
1753static int amdgpu_ras_realloc_eh_data_space(struct amdgpu_device *adev,
1754		struct ras_err_handler_data *data, int pages)
1755{
1756	unsigned int old_space = data->count + data->space_left;
1757	unsigned int new_space = old_space + pages;
1758	unsigned int align_space = ALIGN(new_space, 512);
1759	void *bps = kmalloc(align_space * sizeof(*data->bps), GFP_KERNEL);
1760
1761	if (!bps) {
1762		kfree(bps);
1763		return -ENOMEM;
1764	}
1765
1766	if (data->bps) {
1767		memcpy(bps, data->bps,
1768				data->count * sizeof(*data->bps));
1769		kfree(data->bps);
1770	}
1771
1772	data->bps = bps;
1773	data->space_left += align_space - old_space;
1774	return 0;
1775}
1776
1777/* it deal with vram only. */
1778int amdgpu_ras_add_bad_pages(struct amdgpu_device *adev,
1779		struct eeprom_table_record *bps, int pages)
1780{
1781	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1782	struct ras_err_handler_data *data;
1783	int ret = 0;
1784	uint32_t i;
1785
1786	if (!con || !con->eh_data || !bps || pages <= 0)
1787		return 0;
1788
1789	mutex_lock(&con->recovery_lock);
1790	data = con->eh_data;
1791	if (!data)
1792		goto out;
1793
1794	for (i = 0; i < pages; i++) {
1795		if (amdgpu_ras_check_bad_page_unlock(con,
1796			bps[i].retired_page << AMDGPU_GPU_PAGE_SHIFT))
1797			continue;
1798
1799		if (!data->space_left &&
1800			amdgpu_ras_realloc_eh_data_space(adev, data, 256)) {
1801			ret = -ENOMEM;
1802			goto out;
1803		}
1804
1805		amdgpu_vram_mgr_reserve_range(
1806			ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM),
1807			bps[i].retired_page << AMDGPU_GPU_PAGE_SHIFT,
1808			AMDGPU_GPU_PAGE_SIZE);
1809
1810		memcpy(&data->bps[data->count], &bps[i], sizeof(*data->bps));
1811		data->count++;
1812		data->space_left--;
1813	}
1814out:
1815	mutex_unlock(&con->recovery_lock);
1816
1817	return ret;
1818}
1819
1820/*
1821 * write error record array to eeprom, the function should be
1822 * protected by recovery_lock
1823 */
1824int amdgpu_ras_save_bad_pages(struct amdgpu_device *adev)
1825{
1826	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1827	struct ras_err_handler_data *data;
1828	struct amdgpu_ras_eeprom_control *control;
1829	int save_count;
1830
1831	if (!con || !con->eh_data)
1832		return 0;
1833
1834	control = &con->eeprom_control;
1835	data = con->eh_data;
1836	save_count = data->count - control->num_recs;
1837	/* only new entries are saved */
1838	if (save_count > 0) {
1839		if (amdgpu_ras_eeprom_process_recods(control,
1840							&data->bps[control->num_recs],
1841							true,
1842							save_count)) {
1843			dev_err(adev->dev, "Failed to save EEPROM table data!");
1844			return -EIO;
1845		}
1846
1847		dev_info(adev->dev, "Saved %d pages to EEPROM table.\n", save_count);
1848	}
1849
1850	return 0;
1851}
1852
1853/*
1854 * read error record array in eeprom and reserve enough space for
1855 * storing new bad pages
1856 */
1857static int amdgpu_ras_load_bad_pages(struct amdgpu_device *adev)
1858{
1859	struct amdgpu_ras_eeprom_control *control =
1860					&adev->psp.ras.ras->eeprom_control;
1861	struct eeprom_table_record *bps = NULL;
1862	int ret = 0;
1863
1864	/* no bad page record, skip eeprom access */
1865	if (!control->num_recs || (amdgpu_bad_page_threshold == 0))
1866		return ret;
1867
1868	bps = kcalloc(control->num_recs, sizeof(*bps), GFP_KERNEL);
1869	if (!bps)
1870		return -ENOMEM;
1871
1872	if (amdgpu_ras_eeprom_process_recods(control, bps, false,
1873		control->num_recs)) {
1874		dev_err(adev->dev, "Failed to load EEPROM table records!");
1875		ret = -EIO;
1876		goto out;
1877	}
1878
1879	ret = amdgpu_ras_add_bad_pages(adev, bps, control->num_recs);
1880
1881out:
1882	kfree(bps);
1883	return ret;
1884}
1885
1886static bool amdgpu_ras_check_bad_page_unlock(struct amdgpu_ras *con,
1887				uint64_t addr)
1888{
1889	struct ras_err_handler_data *data = con->eh_data;
1890	int i;
1891
1892	addr >>= AMDGPU_GPU_PAGE_SHIFT;
1893	for (i = 0; i < data->count; i++)
1894		if (addr == data->bps[i].retired_page)
1895			return true;
1896
1897	return false;
1898}
1899
1900/*
1901 * check if an address belongs to bad page
1902 *
1903 * Note: this check is only for umc block
1904 */
1905static bool amdgpu_ras_check_bad_page(struct amdgpu_device *adev,
1906				uint64_t addr)
1907{
1908	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1909	bool ret = false;
1910
1911	if (!con || !con->eh_data)
1912		return ret;
1913
1914	mutex_lock(&con->recovery_lock);
1915	ret = amdgpu_ras_check_bad_page_unlock(con, addr);
1916	mutex_unlock(&con->recovery_lock);
1917	return ret;
1918}
1919
1920static void amdgpu_ras_validate_threshold(struct amdgpu_device *adev,
1921					uint32_t max_length)
1922{
1923	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1924	int tmp_threshold = amdgpu_bad_page_threshold;
1925	u64 val;
1926
1927	/*
1928	 * Justification of value bad_page_cnt_threshold in ras structure
1929	 *
1930	 * Generally, -1 <= amdgpu_bad_page_threshold <= max record length
1931	 * in eeprom, and introduce two scenarios accordingly.
1932	 *
1933	 * Bad page retirement enablement:
1934	 *    - If amdgpu_bad_page_threshold = -1,
1935	 *      bad_page_cnt_threshold = typical value by formula.
1936	 *
1937	 *    - When the value from user is 0 < amdgpu_bad_page_threshold <
1938	 *      max record length in eeprom, use it directly.
1939	 *
1940	 * Bad page retirement disablement:
1941	 *    - If amdgpu_bad_page_threshold = 0, bad page retirement
1942	 *      functionality is disabled, and bad_page_cnt_threshold will
1943	 *      take no effect.
1944	 */
1945
1946	if (tmp_threshold < -1)
1947		tmp_threshold = -1;
1948	else if (tmp_threshold > max_length)
1949		tmp_threshold = max_length;
1950
1951	if (tmp_threshold == -1) {
1952		val = adev->gmc.mc_vram_size;
1953		do_div(val, RAS_BAD_PAGE_RATE);
1954		con->bad_page_cnt_threshold = min(lower_32_bits(val),
1955						max_length);
1956	} else {
1957		con->bad_page_cnt_threshold = tmp_threshold;
1958	}
1959}
1960
1961int amdgpu_ras_recovery_init(struct amdgpu_device *adev)
1962{
1963	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1964	struct ras_err_handler_data **data;
1965	uint32_t max_eeprom_records_len = 0;
1966	bool exc_err_limit = false;
1967	int ret;
1968
1969	if (!con)
1970		return 0;
1971
1972	/* Allow access to RAS EEPROM via debugfs, when the ASIC
1973	 * supports RAS and debugfs is enabled, but when
1974	 * adev->ras_enabled is unset, i.e. when "ras_enable"
1975	 * module parameter is set to 0.
1976	 */
1977	con->adev = adev;
1978
1979	if (!adev->ras_enabled)
1980		return 0;
1981
1982	data = &con->eh_data;
1983	*data = kmalloc(sizeof(**data), GFP_KERNEL | __GFP_ZERO);
1984	if (!*data) {
1985		ret = -ENOMEM;
1986		goto out;
1987	}
1988
1989	mutex_init(&con->recovery_lock);
1990	INIT_WORK(&con->recovery_work, amdgpu_ras_do_recovery);
1991	atomic_set(&con->in_recovery, 0);
1992
1993	max_eeprom_records_len = amdgpu_ras_eeprom_get_record_max_length();
1994	amdgpu_ras_validate_threshold(adev, max_eeprom_records_len);
1995
1996	/* Todo: During test the SMU might fail to read the eeprom through I2C
1997	 * when the GPU is pending on XGMI reset during probe time
1998	 * (Mostly after second bus reset), skip it now
1999	 */
2000	if (adev->gmc.xgmi.pending_reset)
2001		return 0;
2002	ret = amdgpu_ras_eeprom_init(&con->eeprom_control, &exc_err_limit);
2003	/*
2004	 * This calling fails when exc_err_limit is true or
2005	 * ret != 0.
2006	 */
2007	if (exc_err_limit || ret)
2008		goto free;
2009
2010	if (con->eeprom_control.num_recs) {
2011		ret = amdgpu_ras_load_bad_pages(adev);
2012		if (ret)
2013			goto free;
2014
2015		if (adev->smu.ppt_funcs && adev->smu.ppt_funcs->send_hbm_bad_pages_num)
2016			adev->smu.ppt_funcs->send_hbm_bad_pages_num(&adev->smu, con->eeprom_control.num_recs);
2017	}
2018
2019	return 0;
2020
2021free:
2022	kfree((*data)->bps);
2023	kfree(*data);
2024	con->eh_data = NULL;
2025out:
2026	dev_warn(adev->dev, "Failed to initialize ras recovery!\n");
2027
2028	/*
2029	 * Except error threshold exceeding case, other failure cases in this
2030	 * function would not fail amdgpu driver init.
2031	 */
2032	if (!exc_err_limit)
2033		ret = 0;
2034	else
2035		ret = -EINVAL;
2036
2037	return ret;
2038}
2039
2040static int amdgpu_ras_recovery_fini(struct amdgpu_device *adev)
2041{
2042	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2043	struct ras_err_handler_data *data = con->eh_data;
2044
2045	/* recovery_init failed to init it, fini is useless */
2046	if (!data)
2047		return 0;
2048
2049	cancel_work_sync(&con->recovery_work);
2050
2051	mutex_lock(&con->recovery_lock);
2052	con->eh_data = NULL;
2053	kfree(data->bps);
2054	kfree(data);
2055	mutex_unlock(&con->recovery_lock);
2056
2057	return 0;
2058}
2059/* recovery end */
2060
2061/* return 0 if ras will reset gpu and repost.*/
2062int amdgpu_ras_request_reset_on_boot(struct amdgpu_device *adev,
2063		unsigned int block)
2064{
2065	struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
2066
2067	if (!ras)
2068		return -EINVAL;
2069
2070	ras->flags |= AMDGPU_RAS_FLAG_INIT_NEED_RESET;
2071	return 0;
2072}
2073
2074static bool amdgpu_ras_asic_supported(struct amdgpu_device *adev)
2075{
2076	return adev->asic_type == CHIP_VEGA10 ||
2077		adev->asic_type == CHIP_VEGA20 ||
2078		adev->asic_type == CHIP_ARCTURUS ||
2079		adev->asic_type == CHIP_ALDEBARAN ||
2080		adev->asic_type == CHIP_SIENNA_CICHLID;
2081}
2082
2083/*
2084 * this is workaround for vega20 workstation sku,
2085 * force enable gfx ras, ignore vbios gfx ras flag
2086 * due to GC EDC can not write
2087 */
2088static void amdgpu_ras_get_quirks(struct amdgpu_device *adev)
2089{
2090	struct atom_context *ctx = adev->mode_info.atom_context;
2091
2092	if (!ctx)
2093		return;
2094
2095	if (strnstr(ctx->vbios_version, "D16406",
2096		    sizeof(ctx->vbios_version)) ||
2097		strnstr(ctx->vbios_version, "D36002",
2098			sizeof(ctx->vbios_version)))
2099		adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__GFX);
2100}
2101
2102/*
2103 * check hardware's ras ability which will be saved in hw_supported.
2104 * if hardware does not support ras, we can skip some ras initializtion and
2105 * forbid some ras operations from IP.
2106 * if software itself, say boot parameter, limit the ras ability. We still
2107 * need allow IP do some limited operations, like disable. In such case,
2108 * we have to initialize ras as normal. but need check if operation is
2109 * allowed or not in each function.
2110 */
2111static void amdgpu_ras_check_supported(struct amdgpu_device *adev)
2112{
2113	adev->ras_hw_enabled = adev->ras_enabled = 0;
2114
2115	if (amdgpu_sriov_vf(adev) || !adev->is_atom_fw ||
2116	    !amdgpu_ras_asic_supported(adev))
2117		return;
2118
2119	if (!adev->gmc.xgmi.connected_to_cpu) {
2120		if (amdgpu_atomfirmware_mem_ecc_supported(adev)) {
2121			dev_info(adev->dev, "MEM ECC is active.\n");
2122			adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__UMC |
2123						   1 << AMDGPU_RAS_BLOCK__DF);
2124		} else {
2125			dev_info(adev->dev, "MEM ECC is not presented.\n");
2126		}
2127
2128		if (amdgpu_atomfirmware_sram_ecc_supported(adev)) {
2129			dev_info(adev->dev, "SRAM ECC is active.\n");
2130			adev->ras_hw_enabled |= ~(1 << AMDGPU_RAS_BLOCK__UMC |
2131						    1 << AMDGPU_RAS_BLOCK__DF);
2132		} else {
2133			dev_info(adev->dev, "SRAM ECC is not presented.\n");
2134		}
2135	} else {
2136		/* driver only manages a few IP blocks RAS feature
2137		 * when GPU is connected cpu through XGMI */
2138		adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__GFX |
2139					   1 << AMDGPU_RAS_BLOCK__SDMA |
2140					   1 << AMDGPU_RAS_BLOCK__MMHUB);
2141	}
2142
2143	amdgpu_ras_get_quirks(adev);
2144
2145	/* hw_supported needs to be aligned with RAS block mask. */
2146	adev->ras_hw_enabled &= AMDGPU_RAS_BLOCK_MASK;
2147
2148	adev->ras_enabled = amdgpu_ras_enable == 0 ? 0 :
2149		adev->ras_hw_enabled & amdgpu_ras_mask;
2150}
2151
2152static void amdgpu_ras_counte_dw(struct work_struct *work)
2153{
2154	struct amdgpu_ras *con = container_of(work, struct amdgpu_ras,
2155					      ras_counte_delay_work.work);
2156	struct amdgpu_device *adev = con->adev;
2157	struct drm_device *dev = adev_to_drm(adev);
2158	unsigned long ce_count, ue_count;
2159	int res;
2160
2161	res = pm_runtime_get_sync(dev->dev);
2162	if (res < 0)
2163		goto Out;
2164
2165	/* Cache new values.
2166	 */
2167	if (amdgpu_ras_query_error_count(adev, &ce_count, &ue_count) == 0) {
2168		atomic_set(&con->ras_ce_count, ce_count);
2169		atomic_set(&con->ras_ue_count, ue_count);
2170	}
2171
2172	pm_runtime_mark_last_busy(dev->dev);
2173Out:
2174	pm_runtime_put_autosuspend(dev->dev);
2175}
2176
2177int amdgpu_ras_init(struct amdgpu_device *adev)
2178{
2179	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2180	int r;
2181
2182	if (con)
2183		return 0;
2184
2185	con = kmalloc(sizeof(struct amdgpu_ras) +
2186			sizeof(struct ras_manager) * AMDGPU_RAS_BLOCK_COUNT,
2187			GFP_KERNEL|__GFP_ZERO);
2188	if (!con)
2189		return -ENOMEM;
2190
2191	con->adev = adev;
2192	INIT_DELAYED_WORK(&con->ras_counte_delay_work, amdgpu_ras_counte_dw);
2193	atomic_set(&con->ras_ce_count, 0);
2194	atomic_set(&con->ras_ue_count, 0);
2195
2196	con->objs = (struct ras_manager *)(con + 1);
2197
2198	amdgpu_ras_set_context(adev, con);
2199
2200	amdgpu_ras_check_supported(adev);
2201
2202	if (!adev->ras_enabled || adev->asic_type == CHIP_VEGA10) {
2203		/* set gfx block ras context feature for VEGA20 Gaming
2204		 * send ras disable cmd to ras ta during ras late init.
2205		 */
2206		if (!adev->ras_enabled && adev->asic_type == CHIP_VEGA20) {
2207			con->features |= BIT(AMDGPU_RAS_BLOCK__GFX);
2208
2209			return 0;
2210		}
2211
2212		r = 0;
2213		goto release_con;
2214	}
2215
2216	con->features = 0;
2217	INIT_LIST_HEAD(&con->head);
2218	/* Might need get this flag from vbios. */
2219	con->flags = RAS_DEFAULT_FLAGS;
2220
2221	/* initialize nbio ras function ahead of any other
2222	 * ras functions so hardware fatal error interrupt
2223	 * can be enabled as early as possible */
2224	switch (adev->asic_type) {
2225	case CHIP_VEGA20:
2226	case CHIP_ARCTURUS:
2227	case CHIP_ALDEBARAN:
2228		if (!adev->gmc.xgmi.connected_to_cpu)
2229			adev->nbio.ras_funcs = &nbio_v7_4_ras_funcs;
2230		break;
2231	default:
2232		/* nbio ras is not available */
2233		break;
2234	}
2235
2236	if (adev->nbio.ras_funcs &&
2237	    adev->nbio.ras_funcs->init_ras_controller_interrupt) {
2238		r = adev->nbio.ras_funcs->init_ras_controller_interrupt(adev);
2239		if (r)
2240			goto release_con;
2241	}
2242
2243	if (adev->nbio.ras_funcs &&
2244	    adev->nbio.ras_funcs->init_ras_err_event_athub_interrupt) {
2245		r = adev->nbio.ras_funcs->init_ras_err_event_athub_interrupt(adev);
2246		if (r)
2247			goto release_con;
2248	}
2249
2250	if (amdgpu_ras_fs_init(adev)) {
2251		r = -EINVAL;
2252		goto release_con;
2253	}
2254
2255	dev_info(adev->dev, "RAS INFO: ras initialized successfully, "
2256		 "hardware ability[%x] ras_mask[%x]\n",
2257		 adev->ras_hw_enabled, adev->ras_enabled);
2258
2259	return 0;
2260release_con:
2261	amdgpu_ras_set_context(adev, NULL);
2262	kfree(con);
2263
2264	return r;
2265}
2266
2267int amdgpu_persistent_edc_harvesting_supported(struct amdgpu_device *adev)
2268{
2269	if (adev->gmc.xgmi.connected_to_cpu)
2270		return 1;
2271	return 0;
2272}
2273
2274static int amdgpu_persistent_edc_harvesting(struct amdgpu_device *adev,
2275					struct ras_common_if *ras_block)
2276{
2277	struct ras_query_if info = {
2278		.head = *ras_block,
2279	};
2280
2281	if (!amdgpu_persistent_edc_harvesting_supported(adev))
2282		return 0;
2283
2284	if (amdgpu_ras_query_error_status(adev, &info) != 0)
2285		DRM_WARN("RAS init harvest failure");
2286
2287	if (amdgpu_ras_reset_error_status(adev, ras_block->block) != 0)
2288		DRM_WARN("RAS init harvest reset failure");
2289
2290	return 0;
2291}
2292
2293/* helper function to handle common stuff in ip late init phase */
2294int amdgpu_ras_late_init(struct amdgpu_device *adev,
2295			 struct ras_common_if *ras_block,
2296			 struct ras_fs_if *fs_info,
2297			 struct ras_ih_if *ih_info)
2298{
2299	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2300	unsigned long ue_count, ce_count;
2301	int r;
2302
2303	/* disable RAS feature per IP block if it is not supported */
2304	if (!amdgpu_ras_is_supported(adev, ras_block->block)) {
2305		amdgpu_ras_feature_enable_on_boot(adev, ras_block, 0);
2306		return 0;
2307	}
2308
2309	r = amdgpu_ras_feature_enable_on_boot(adev, ras_block, 1);
2310	if (r) {
2311		if (r == -EAGAIN) {
2312			/* request gpu reset. will run again */
2313			amdgpu_ras_request_reset_on_boot(adev,
2314					ras_block->block);
2315			return 0;
2316		} else if (adev->in_suspend || amdgpu_in_reset(adev)) {
2317			/* in resume phase, if fail to enable ras,
2318			 * clean up all ras fs nodes, and disable ras */
2319			goto cleanup;
2320		} else
2321			return r;
2322	}
2323
2324	/* check for errors on warm reset edc persisant supported ASIC */
2325	amdgpu_persistent_edc_harvesting(adev, ras_block);
2326
2327	/* in resume phase, no need to create ras fs node */
2328	if (adev->in_suspend || amdgpu_in_reset(adev))
2329		return 0;
2330
2331	if (ih_info->cb) {
2332		r = amdgpu_ras_interrupt_add_handler(adev, ih_info);
2333		if (r)
2334			goto interrupt;
2335	}
2336
2337	r = amdgpu_ras_sysfs_create(adev, fs_info);
2338	if (r)
2339		goto sysfs;
2340
2341	/* Those are the cached values at init.
2342	 */
2343	if (amdgpu_ras_query_error_count(adev, &ce_count, &ue_count) == 0) {
2344		atomic_set(&con->ras_ce_count, ce_count);
2345		atomic_set(&con->ras_ue_count, ue_count);
2346	}
2347
2348	return 0;
2349cleanup:
2350	amdgpu_ras_sysfs_remove(adev, ras_block);
2351sysfs:
2352	if (ih_info->cb)
2353		amdgpu_ras_interrupt_remove_handler(adev, ih_info);
2354interrupt:
2355	amdgpu_ras_feature_enable(adev, ras_block, 0);
2356	return r;
2357}
2358
2359/* helper function to remove ras fs node and interrupt handler */
2360void amdgpu_ras_late_fini(struct amdgpu_device *adev,
2361			  struct ras_common_if *ras_block,
2362			  struct ras_ih_if *ih_info)
2363{
2364	if (!ras_block || !ih_info)
2365		return;
2366
2367	amdgpu_ras_sysfs_remove(adev, ras_block);
2368	if (ih_info->cb)
2369		amdgpu_ras_interrupt_remove_handler(adev, ih_info);
2370	amdgpu_ras_feature_enable(adev, ras_block, 0);
2371}
2372
2373/* do some init work after IP late init as dependence.
2374 * and it runs in resume/gpu reset/booting up cases.
2375 */
2376void amdgpu_ras_resume(struct amdgpu_device *adev)
2377{
2378	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2379	struct ras_manager *obj, *tmp;
2380
2381	if (!adev->ras_enabled || !con) {
2382		/* clean ras context for VEGA20 Gaming after send ras disable cmd */
2383		amdgpu_release_ras_context(adev);
2384
2385		return;
2386	}
2387
2388	if (con->flags & AMDGPU_RAS_FLAG_INIT_BY_VBIOS) {
2389		/* Set up all other IPs which are not implemented. There is a
2390		 * tricky thing that IP's actual ras error type should be
2391		 * MULTI_UNCORRECTABLE, but as driver does not handle it, so
2392		 * ERROR_NONE make sense anyway.
2393		 */
2394		amdgpu_ras_enable_all_features(adev, 1);
2395
2396		/* We enable ras on all hw_supported block, but as boot
2397		 * parameter might disable some of them and one or more IP has
2398		 * not implemented yet. So we disable them on behalf.
2399		 */
2400		list_for_each_entry_safe(obj, tmp, &con->head, node) {
2401			if (!amdgpu_ras_is_supported(adev, obj->head.block)) {
2402				amdgpu_ras_feature_enable(adev, &obj->head, 0);
2403				/* there should be no any reference. */
2404				WARN_ON(alive_obj(obj));
2405			}
2406		}
2407	}
2408
2409	if (con->flags & AMDGPU_RAS_FLAG_INIT_NEED_RESET) {
2410		con->flags &= ~AMDGPU_RAS_FLAG_INIT_NEED_RESET;
2411		/* setup ras obj state as disabled.
2412		 * for init_by_vbios case.
2413		 * if we want to enable ras, just enable it in a normal way.
2414		 * If we want do disable it, need setup ras obj as enabled,
2415		 * then issue another TA disable cmd.
2416		 * See feature_enable_on_boot
2417		 */
2418		amdgpu_ras_disable_all_features(adev, 1);
2419		amdgpu_ras_reset_gpu(adev);
2420	}
2421}
2422
2423void amdgpu_ras_suspend(struct amdgpu_device *adev)
2424{
2425	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2426
2427	if (!adev->ras_enabled || !con)
2428		return;
2429
2430	amdgpu_ras_disable_all_features(adev, 0);
2431	/* Make sure all ras objects are disabled. */
2432	if (con->features)
2433		amdgpu_ras_disable_all_features(adev, 1);
2434}
2435
2436/* do some fini work before IP fini as dependence */
2437int amdgpu_ras_pre_fini(struct amdgpu_device *adev)
2438{
2439	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2440
2441	if (!adev->ras_enabled || !con)
2442		return 0;
2443
2444
2445	/* Need disable ras on all IPs here before ip [hw/sw]fini */
2446	amdgpu_ras_disable_all_features(adev, 0);
2447	amdgpu_ras_recovery_fini(adev);
2448	return 0;
2449}
2450
2451int amdgpu_ras_fini(struct amdgpu_device *adev)
2452{
2453	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2454
2455	if (!adev->ras_enabled || !con)
2456		return 0;
2457
2458	amdgpu_ras_fs_fini(adev);
2459	amdgpu_ras_interrupt_remove_all(adev);
2460
2461	WARN(con->features, "Feature mask is not cleared");
2462
2463	if (con->features)
2464		amdgpu_ras_disable_all_features(adev, 1);
2465
2466	cancel_delayed_work_sync(&con->ras_counte_delay_work);
2467
2468	amdgpu_ras_set_context(adev, NULL);
2469	kfree(con);
2470
2471	return 0;
2472}
2473
2474void amdgpu_ras_global_ras_isr(struct amdgpu_device *adev)
2475{
2476	amdgpu_ras_check_supported(adev);
2477	if (!adev->ras_hw_enabled)
2478		return;
2479
2480	if (atomic_cmpxchg(&amdgpu_ras_in_intr, 0, 1) == 0) {
2481		dev_info(adev->dev, "uncorrectable hardware error"
2482			"(ERREVENT_ATHUB_INTERRUPT) detected!\n");
2483
2484		amdgpu_ras_reset_gpu(adev);
2485	}
2486}
2487
2488bool amdgpu_ras_need_emergency_restart(struct amdgpu_device *adev)
2489{
2490	if (adev->asic_type == CHIP_VEGA20 &&
2491	    adev->pm.fw_version <= 0x283400) {
2492		return !(amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) &&
2493				amdgpu_ras_intr_triggered();
2494	}
2495
2496	return false;
2497}
2498
2499void amdgpu_release_ras_context(struct amdgpu_device *adev)
2500{
2501	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2502
2503	if (!con)
2504		return;
2505
2506	if (!adev->ras_enabled && con->features & BIT(AMDGPU_RAS_BLOCK__GFX)) {
2507		con->features &= ~BIT(AMDGPU_RAS_BLOCK__GFX);
2508		amdgpu_ras_set_context(adev, NULL);
2509		kfree(con);
2510	}
2511}