2 * Copyright 2018 Advanced Micro Devices, Inc.
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:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
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.
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>
33 #include "amdgpu_ras.h"
34 #include "amdgpu_atomfirmware.h"
35 #include "amdgpu_xgmi.h"
36 #include "ivsrcid/nbio/irqsrcs_nbif_7_4.h"
39 static const char *RAS_FS_NAME = "ras";
41 const char *ras_error_string[] = {
45 "multi_uncorrectable",
49 const char *ras_block_string[] = {
66 #define ras_err_str(i) (ras_error_string[ffs(i)])
67 #define ras_block_str(i) (ras_block_string[i])
69 #define RAS_DEFAULT_FLAGS (AMDGPU_RAS_FLAG_INIT_BY_VBIOS)
71 /* inject address is 52 bits */
72 #define RAS_UMC_INJECT_ADDR_LIMIT (0x1ULL << 52)
74 /* typical ECC bad page rate(1 bad page per 100MB VRAM) */
75 #define RAS_BAD_PAGE_RATE (100 * 1024 * 1024ULL)
77 enum amdgpu_ras_retire_page_reservation {
78 AMDGPU_RAS_RETIRE_PAGE_RESERVED,
79 AMDGPU_RAS_RETIRE_PAGE_PENDING,
80 AMDGPU_RAS_RETIRE_PAGE_FAULT,
83 atomic_t amdgpu_ras_in_intr = ATOMIC_INIT(0);
85 static bool amdgpu_ras_check_bad_page_unlock(struct amdgpu_ras *con,
87 static bool amdgpu_ras_check_bad_page(struct amdgpu_device *adev,
90 void amdgpu_ras_set_error_query_ready(struct amdgpu_device *adev, bool ready)
92 if (adev && amdgpu_ras_get_context(adev))
93 amdgpu_ras_get_context(adev)->error_query_ready = ready;
96 static bool amdgpu_ras_get_error_query_ready(struct amdgpu_device *adev)
98 if (adev && amdgpu_ras_get_context(adev))
99 return amdgpu_ras_get_context(adev)->error_query_ready;
104 static int amdgpu_reserve_page_direct(struct amdgpu_device *adev, uint64_t address)
106 struct ras_err_data err_data = {0, 0, 0, NULL};
107 struct eeprom_table_record err_rec;
109 if ((address >= adev->gmc.mc_vram_size) ||
110 (address >= RAS_UMC_INJECT_ADDR_LIMIT)) {
112 "RAS WARN: input address 0x%llx is invalid.\n",
117 if (amdgpu_ras_check_bad_page(adev, address)) {
119 "RAS WARN: 0x%llx has already been marked as bad page!\n",
124 memset(&err_rec, 0x0, sizeof(struct eeprom_table_record));
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;
131 err_data.err_addr = &err_rec;
132 err_data.err_addr_cnt = 1;
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);
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");
147 static ssize_t amdgpu_ras_debugfs_read(struct file *f, char __user *buf,
148 size_t size, loff_t *pos)
150 struct ras_manager *obj = (struct ras_manager *)file_inode(f)->i_private;
151 struct ras_query_if info = {
157 if (amdgpu_ras_query_error_status(obj->adev, &info))
160 s = snprintf(val, sizeof(val), "%s: %lu\n%s: %lu\n",
162 "ce", info.ce_count);
167 s = min_t(u64, s, size);
170 if (copy_to_user(buf, &val[*pos], s))
178 static const struct file_operations amdgpu_ras_debugfs_ops = {
179 .owner = THIS_MODULE,
180 .read = amdgpu_ras_debugfs_read,
182 .llseek = default_llseek
185 static int amdgpu_ras_find_block_id_by_name(const char *name, int *block_id)
189 for (i = 0; i < ARRAY_SIZE(ras_block_string); i++) {
191 if (strcmp(name, ras_block_str(i)) == 0)
197 static 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)
201 ssize_t s = min_t(u64, 64, size);
214 memset(str, 0, sizeof(str));
215 memset(data, 0, sizeof(*data));
217 if (copy_from_user(str, buf, s))
220 if (sscanf(str, "disable %32s", block_name) == 1)
222 else if (sscanf(str, "enable %32s %8s", block_name, err) == 2)
224 else if (sscanf(str, "inject %32s %8s", block_name, err) == 2)
226 else if (strstr(str, "retire_page") != NULL)
228 else if (str[0] && str[1] && str[2] && str[3])
229 /* ascii string, but commands are not matched. */
234 if (sscanf(str, "%*s 0x%llx", &address) != 1 &&
235 sscanf(str, "%*s %llu", &address) != 1)
239 data->inject.address = address;
244 if (amdgpu_ras_find_block_id_by_name(block_name, &block_id))
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;
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)
264 data->head.sub_block_index = sub_block;
265 data->inject.address = address;
266 data->inject.value = value;
269 if (size < sizeof(*data))
272 if (copy_from_user(data, buf, sizeof(*data)))
280 * DOC: AMDGPU RAS debugfs control interface
282 * The control interface accepts struct ras_debug_if which has two members.
284 * First member: ras_debug_if::head or ras_debug_if::inject.
286 * head is used to indicate which IP block will be under control.
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.
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.
298 * The second member: struct ras_debug_if::op.
299 * It has three kinds of operations.
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.
305 * How to use the interface?
309 * Copy the struct ras_debug_if in your code and initialize it.
310 * Write the struct to the control interface.
314 * .. code-block:: bash
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
320 * Where N, is the card which you want to affect.
322 * "disable" requires only the block.
323 * "enable" requires the block and error type.
324 * "inject" requires the block, error type, address, and value.
326 * The block is one of: umc, sdma, gfx, etc.
327 * see ras_block_string[] for details
329 * The error type is one of: ue, ce, where,
330 * ue is multi-uncorrectable
331 * ce is single-correctable
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.
338 * .. code-block:: bash
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
344 * How to check the result of the operation?
346 * To check disable/enable, see "ras" features at,
347 * /sys/class/drm/card[0/1/2...]/device/ras/features
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
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.
358 static ssize_t amdgpu_ras_debugfs_ctrl_write(struct file *f, const char __user *buf,
359 size_t size, loff_t *pos)
361 struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
362 struct ras_debug_if data;
365 if (!amdgpu_ras_get_error_query_ready(adev)) {
366 dev_warn(adev->dev, "RAS WARN: error injection "
367 "currently inaccessible\n");
371 ret = amdgpu_ras_debugfs_ctrl_parse_data(f, buf, size, pos, &data);
376 ret = amdgpu_reserve_page_direct(adev, data.inject.address);
383 if (!amdgpu_ras_is_supported(adev, data.head.block))
388 ret = amdgpu_ras_feature_enable(adev, &data.head, 0);
391 ret = amdgpu_ras_feature_enable(adev, &data.head, 1);
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);
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);
412 /* data.inject.address is offset instead of absolute gpu address */
413 ret = amdgpu_ras_error_inject(adev, &data.inject);
427 * DOC: AMDGPU RAS debugfs EEPROM table reset interface
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.
435 * .. code-block:: bash
437 * echo 1 > ../ras/ras_eeprom_reset
439 * will reset EEPROM table to 0 entries.
442 static ssize_t amdgpu_ras_debugfs_eeprom_write(struct file *f, const char __user *buf,
443 size_t size, loff_t *pos)
445 struct amdgpu_device *adev =
446 (struct amdgpu_device *)file_inode(f)->i_private;
449 ret = amdgpu_ras_eeprom_reset_table(
450 &(amdgpu_ras_get_context(adev)->eeprom_control));
453 amdgpu_ras_get_context(adev)->flags = RAS_DEFAULT_FLAGS;
460 static const struct file_operations amdgpu_ras_debugfs_ctrl_ops = {
461 .owner = THIS_MODULE,
463 .write = amdgpu_ras_debugfs_ctrl_write,
464 .llseek = default_llseek
467 static const struct file_operations amdgpu_ras_debugfs_eeprom_ops = {
468 .owner = THIS_MODULE,
470 .write = amdgpu_ras_debugfs_eeprom_write,
471 .llseek = default_llseek
475 * DOC: AMDGPU RAS sysfs Error Count Interface
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
480 * It outputs the multiple lines which report the uncorrected (ue) and corrected
483 * The format of one line is below,
489 * .. code-block:: bash
495 static ssize_t amdgpu_ras_sysfs_read(struct device *dev,
496 struct device_attribute *attr, char *buf)
498 struct ras_manager *obj = container_of(attr, struct ras_manager, sysfs_attr);
499 struct ras_query_if info = {
503 if (!amdgpu_ras_get_error_query_ready(obj->adev))
504 return sysfs_emit(buf, "Query currently inaccessible\n");
506 if (amdgpu_ras_query_error_status(obj->adev, &info))
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");
515 return sysfs_emit(buf, "%s: %lu\n%s: %lu\n", "ue", info.ue_count,
516 "ce", info.ce_count);
521 #define get_obj(obj) do { (obj)->use++; } while (0)
522 #define alive_obj(obj) ((obj)->use)
524 static inline void put_obj(struct ras_manager *obj)
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);
532 /* make one obj and return it. */
533 static struct ras_manager *amdgpu_ras_create_obj(struct amdgpu_device *adev,
534 struct ras_common_if *head)
536 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
537 struct ras_manager *obj;
539 if (!adev->ras_enabled || !con)
542 if (head->block >= AMDGPU_RAS_BLOCK_COUNT)
545 obj = &con->objs[head->block];
546 /* already exist. return obj? */
552 list_add(&obj->node, &con->head);
558 /* return an obj equal to head, or the first when head is NULL */
559 struct ras_manager *amdgpu_ras_find_obj(struct amdgpu_device *adev,
560 struct ras_common_if *head)
562 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
563 struct ras_manager *obj;
566 if (!adev->ras_enabled || !con)
570 if (head->block >= AMDGPU_RAS_BLOCK_COUNT)
573 obj = &con->objs[head->block];
575 if (alive_obj(obj)) {
576 WARN_ON(head->block != obj->head.block);
580 for (i = 0; i < AMDGPU_RAS_BLOCK_COUNT; i++) {
582 if (alive_obj(obj)) {
583 WARN_ON(i != obj->head.block);
593 /* feature ctl begin */
594 static int amdgpu_ras_is_feature_allowed(struct amdgpu_device *adev,
595 struct ras_common_if *head)
597 return adev->ras_hw_enabled & BIT(head->block);
600 static int amdgpu_ras_is_feature_enabled(struct amdgpu_device *adev,
601 struct ras_common_if *head)
603 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
605 return con->features & BIT(head->block);
609 * if obj is not created, then create one.
610 * set feature enable flag.
612 static int __amdgpu_ras_feature_enable(struct amdgpu_device *adev,
613 struct ras_common_if *head, int enable)
615 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
616 struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
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.
624 if (!amdgpu_ras_is_feature_allowed(adev, head))
626 if (!(!!enable ^ !!amdgpu_ras_is_feature_enabled(adev, head)))
631 obj = amdgpu_ras_create_obj(adev, head);
635 /* In case we create obj somewhere else */
638 con->features |= BIT(head->block);
640 if (obj && amdgpu_ras_is_feature_enabled(adev, head)) {
641 con->features &= ~BIT(head->block);
649 /* wrapper of psp_ras_enable_features */
650 int amdgpu_ras_feature_enable(struct amdgpu_device *adev,
651 struct ras_common_if *head, bool enable)
653 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
654 union ta_ras_cmd_input *info;
660 info = kzalloc(sizeof(union ta_ras_cmd_input), GFP_KERNEL);
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),
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),
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))) {
684 if (!amdgpu_ras_intr_triggered()) {
685 ret = psp_ras_enable_features(&adev->psp, info, enable);
687 dev_err(adev->dev, "ras %s %s failed %d\n",
688 enable ? "enable":"disable",
689 ras_block_str(head->block),
696 __amdgpu_ras_feature_enable(adev, head, enable);
703 /* Only used in device probe stage and called only once. */
704 int amdgpu_ras_feature_enable_on_boot(struct amdgpu_device *adev,
705 struct ras_common_if *head, bool enable)
707 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
713 if (con->flags & AMDGPU_RAS_FLAG_INIT_BY_VBIOS) {
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.
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.
726 if (ret == -EINVAL) {
727 ret = __amdgpu_ras_feature_enable(adev, head, 1);
730 "RAS INFO: %s setup object\n",
731 ras_block_str(head->block));
734 /* setup the object then issue a ras TA disable cmd.*/
735 ret = __amdgpu_ras_feature_enable(adev, head, 1);
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);
743 ret = amdgpu_ras_feature_enable(adev, head, 0);
745 /* clean gfx block ras features flag */
746 if (adev->ras_enabled && head->block == AMDGPU_RAS_BLOCK__GFX)
747 con->features &= ~BIT(head->block);
750 ret = amdgpu_ras_feature_enable(adev, head, enable);
755 static int amdgpu_ras_disable_all_features(struct amdgpu_device *adev,
758 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
759 struct ras_manager *obj, *tmp;
761 list_for_each_entry_safe(obj, tmp, &con->head, node) {
763 * aka just release the obj and corresponding flags
766 if (__amdgpu_ras_feature_enable(adev, &obj->head, 0))
769 if (amdgpu_ras_feature_enable(adev, &obj->head, 0))
774 return con->features;
777 static int amdgpu_ras_enable_all_features(struct amdgpu_device *adev,
780 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
781 int ras_block_count = AMDGPU_RAS_BLOCK_COUNT;
783 const enum amdgpu_ras_error_type default_ras_type =
784 AMDGPU_RAS_ERROR__NONE;
786 for (i = 0; i < ras_block_count; i++) {
787 struct ras_common_if head = {
789 .type = default_ras_type,
790 .sub_block_index = 0,
792 strcpy(head.name, ras_block_str(i));
795 * bypass psp. vbios enable ras for us.
796 * so just create the obj
798 if (__amdgpu_ras_feature_enable(adev, &head, 1))
801 if (amdgpu_ras_feature_enable(adev, &head, 1))
806 return con->features;
808 /* feature ctl end */
810 /* query/inject/cure begin */
811 int amdgpu_ras_query_error_status(struct amdgpu_device *adev,
812 struct ras_query_if *info)
814 struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head);
815 struct ras_err_data err_data = {0, 0, 0, NULL};
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
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);
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,
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);
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);
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);
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);
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);
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);
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);
877 obj->err_data.ue_count += err_data.ue_count;
878 obj->err_data.ce_count += err_data.ce_count;
880 info->ue_count = obj->err_data.ue_count;
881 info->ce_count = obj->err_data.ce_count;
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));
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));
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));
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));
925 int amdgpu_ras_reset_error_status(struct amdgpu_device *adev,
926 enum amdgpu_ras_block block)
928 if (!amdgpu_ras_is_supported(adev, 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);
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);
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);
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);
950 case AMDGPU_RAS_BLOCK__SDMA:
951 if (adev->sdma.funcs->reset_ras_error_count)
952 adev->sdma.funcs->reset_ras_error_count(adev);
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);
966 /* Trigger XGMI/WAFL error */
967 static int amdgpu_ras_error_inject_xgmi(struct amdgpu_device *adev,
968 struct ta_ras_trigger_error_input *block_info)
972 if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_DISALLOW))
973 dev_warn(adev->dev, "Failed to disallow df cstate");
975 if (amdgpu_dpm_allow_xgmi_power_down(adev, false))
976 dev_warn(adev->dev, "Failed to disallow XGMI power down");
978 ret = psp_ras_trigger_error(&adev->psp, block_info);
980 if (amdgpu_ras_intr_triggered())
983 if (amdgpu_dpm_allow_xgmi_power_down(adev, true))
984 dev_warn(adev->dev, "Failed to allow XGMI power down");
986 if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_ALLOW))
987 dev_warn(adev->dev, "Failed to allow df cstate");
992 /* wrapper of psp_ras_trigger_error */
993 int amdgpu_ras_error_inject(struct amdgpu_device *adev,
994 struct ras_inject_if *info)
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,
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);
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);
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);
1030 case AMDGPU_RAS_BLOCK__XGMI_WAFL:
1031 ret = amdgpu_ras_error_inject_xgmi(adev, &block_info);
1034 dev_info(adev->dev, "%s error injection is not supported yet\n",
1035 ras_block_str(info->head.block));
1040 dev_err(adev->dev, "ras inject %s failed %d\n",
1041 ras_block_str(info->head.block), ret);
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
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.
1057 int amdgpu_ras_query_error_count(struct amdgpu_device *adev,
1058 unsigned long *ce_count,
1059 unsigned long *ue_count)
1061 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1062 struct ras_manager *obj;
1063 unsigned long ce, ue;
1065 if (!adev->ras_enabled || !con)
1068 /* Don't count since no reporting.
1070 if (!ce_count && !ue_count)
1075 list_for_each_entry(obj, &con->head, node) {
1076 struct ras_query_if info = {
1081 res = amdgpu_ras_query_error_status(adev, &info);
1085 ce += info.ce_count;
1086 ue += info.ue_count;
1097 /* query/inject/cure end */
1102 static int amdgpu_ras_badpages_read(struct amdgpu_device *adev,
1103 struct ras_badpage **bps, unsigned int *count);
1105 static char *amdgpu_ras_badpage_flags_str(unsigned int flags)
1108 case AMDGPU_RAS_RETIRE_PAGE_RESERVED:
1110 case AMDGPU_RAS_RETIRE_PAGE_PENDING:
1112 case AMDGPU_RAS_RETIRE_PAGE_FAULT:
1119 * DOC: AMDGPU RAS sysfs gpu_vram_bad_pages Interface
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
1124 * It outputs multiple lines, and each line stands for one gpu page.
1126 * The format of one line is below,
1127 * gpu pfn : gpu page size : flags
1129 * gpu pfn and gpu page size are printed in hex format.
1130 * flags can be one of below character,
1132 * R: reserved, this gpu page is reserved and not able to use.
1134 * P: pending for reserve, this gpu page is marked as bad, will be reserved
1135 * in next window of page_reserve.
1137 * F: unable to reserve. this gpu page can't be reserved due to some reasons.
1141 * .. code-block:: bash
1143 * 0x00000001 : 0x00001000 : R
1144 * 0x00000002 : 0x00001000 : P
1148 static 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)
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);
1160 struct ras_badpage *bps = NULL;
1161 unsigned int bps_count = 0;
1163 memset(buf, 0, count);
1165 if (amdgpu_ras_badpages_read(adev, &bps, &bps_count))
1168 for (; start < end && start < bps_count; start++)
1169 s += scnprintf(&buf[s], element_size + 1,
1170 "0x%08x : 0x%08x : %1s\n",
1173 amdgpu_ras_badpage_flags_str(bps[start].flags));
1180 static ssize_t amdgpu_ras_sysfs_features_read(struct device *dev,
1181 struct device_attribute *attr, char *buf)
1183 struct amdgpu_ras *con =
1184 container_of(attr, struct amdgpu_ras, features_attr);
1186 return scnprintf(buf, PAGE_SIZE, "feature mask: 0x%x\n", con->features);
1189 static void amdgpu_ras_sysfs_remove_bad_page_node(struct amdgpu_device *adev)
1191 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1193 sysfs_remove_file_from_group(&adev->dev->kobj,
1194 &con->badpages_attr.attr,
1198 static int amdgpu_ras_sysfs_remove_feature_node(struct amdgpu_device *adev)
1200 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1201 struct attribute *attrs[] = {
1202 &con->features_attr.attr,
1205 struct attribute_group group = {
1206 .name = RAS_FS_NAME,
1210 sysfs_remove_group(&adev->dev->kobj, &group);
1215 int amdgpu_ras_sysfs_create(struct amdgpu_device *adev,
1216 struct ras_fs_if *head)
1218 struct ras_manager *obj = amdgpu_ras_find_obj(adev, &head->head);
1220 if (!obj || obj->attr_inuse)
1225 memcpy(obj->fs_data.sysfs_name,
1227 sizeof(obj->fs_data.sysfs_name));
1229 obj->sysfs_attr = (struct device_attribute){
1231 .name = obj->fs_data.sysfs_name,
1234 .show = amdgpu_ras_sysfs_read,
1236 sysfs_attr_init(&obj->sysfs_attr.attr);
1238 if (sysfs_add_file_to_group(&adev->dev->kobj,
1239 &obj->sysfs_attr.attr,
1245 obj->attr_inuse = 1;
1250 int amdgpu_ras_sysfs_remove(struct amdgpu_device *adev,
1251 struct ras_common_if *head)
1253 struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
1255 if (!obj || !obj->attr_inuse)
1258 sysfs_remove_file_from_group(&adev->dev->kobj,
1259 &obj->sysfs_attr.attr,
1261 obj->attr_inuse = 0;
1267 static int amdgpu_ras_sysfs_remove_all(struct amdgpu_device *adev)
1269 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1270 struct ras_manager *obj, *tmp;
1272 list_for_each_entry_safe(obj, tmp, &con->head, node) {
1273 amdgpu_ras_sysfs_remove(adev, &obj->head);
1276 if (amdgpu_bad_page_threshold != 0)
1277 amdgpu_ras_sysfs_remove_bad_page_node(adev);
1279 amdgpu_ras_sysfs_remove_feature_node(adev);
1286 * DOC: AMDGPU RAS Reboot Behavior for Unrecoverable Errors
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
1293 * The following file in debugfs provides that interface:
1294 * /sys/kernel/debug/dri/[0/1/2...]/ras/auto_reboot
1298 * .. code-block:: bash
1300 * echo true > .../ras/auto_reboot
1304 static struct dentry *amdgpu_ras_debugfs_create_ctrl_node(struct amdgpu_device *adev)
1306 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1307 struct drm_minor *minor = adev_to_drm(adev)->primary;
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);
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.
1328 debugfs_create_bool("auto_reboot", S_IWUGO | S_IRUGO, dir, &con->reboot);
1331 * User could set this not to clean up hardware's error count register
1332 * of RAS IPs during ras recovery.
1334 debugfs_create_bool("disable_ras_err_cnt_harvest", 0644, dir,
1335 &con->disable_ras_err_cnt_harvest);
1339 static void amdgpu_ras_debugfs_create(struct amdgpu_device *adev,
1340 struct ras_fs_if *head,
1343 struct ras_manager *obj = amdgpu_ras_find_obj(adev, &head->head);
1350 memcpy(obj->fs_data.debugfs_name,
1352 sizeof(obj->fs_data.debugfs_name));
1354 debugfs_create_file(obj->fs_data.debugfs_name, S_IWUGO | S_IRUGO, dir,
1355 obj, &amdgpu_ras_debugfs_ops);
1358 void amdgpu_ras_debugfs_create_all(struct amdgpu_device *adev)
1360 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1362 struct ras_manager *obj;
1363 struct ras_fs_if fs_info;
1366 * it won't be called in resume path, no need to check
1367 * suspend and gpu reset status
1369 if (!IS_ENABLED(CONFIG_DEBUG_FS) || !con)
1372 dir = amdgpu_ras_debugfs_create_ctrl_node(adev);
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);
1388 static BIN_ATTR(gpu_vram_bad_pages, S_IRUGO,
1389 amdgpu_ras_sysfs_badpages_read, NULL, 0);
1390 static DEVICE_ATTR(features, S_IRUGO,
1391 amdgpu_ras_sysfs_features_read, NULL);
1392 static int amdgpu_ras_fs_init(struct amdgpu_device *adev)
1394 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1395 struct attribute_group group = {
1396 .name = RAS_FS_NAME,
1398 struct attribute *attrs[] = {
1399 &con->features_attr.attr,
1402 struct bin_attribute *bin_attrs[] = {
1408 /* add features entry */
1409 con->features_attr = dev_attr_features;
1410 group.attrs = attrs;
1411 sysfs_attr_init(attrs[0]);
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]);
1422 r = sysfs_create_group(&adev->dev->kobj, &group);
1424 dev_err(adev->dev, "Failed to create RAS sysfs group!");
1429 static int amdgpu_ras_fs_fini(struct amdgpu_device *adev)
1431 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1432 struct ras_manager *con_obj, *ip_obj, *tmp;
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);
1442 amdgpu_ras_sysfs_remove_all(adev);
1448 static void amdgpu_ras_interrupt_handler(struct ras_manager *obj)
1450 struct ras_ih_data *data = &obj->ih_data;
1451 struct amdgpu_iv_entry entry;
1453 struct ras_err_data err_data = {0, 0, 0, NULL};
1455 while (data->rptr != data->wptr) {
1457 memcpy(&entry, &data->ring[data->rptr],
1458 data->element_size);
1461 data->rptr = (data->aligned_element_size +
1462 data->rptr) % data->ring_size;
1464 /* Let IP handle its data, maybe we need get the output
1465 * from the callback to udpate the error type/count, etc
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
1474 if (ret == AMDGPU_RAS_SUCCESS) {
1475 /* these counts could be left as 0 if
1476 * some blocks do not count error number
1478 obj->err_data.ue_count += err_data.ue_count;
1479 obj->err_data.ce_count += err_data.ce_count;
1485 static void amdgpu_ras_interrupt_process_handler(struct work_struct *work)
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);
1492 amdgpu_ras_interrupt_handler(obj);
1495 int amdgpu_ras_interrupt_dispatch(struct amdgpu_device *adev,
1496 struct ras_dispatch_if *info)
1498 struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head);
1499 struct ras_ih_data *data = &obj->ih_data;
1504 if (data->inuse == 0)
1507 /* Might be overflow... */
1508 memcpy(&data->ring[data->wptr], info->entry,
1509 data->element_size);
1512 data->wptr = (data->aligned_element_size +
1513 data->wptr) % data->ring_size;
1515 schedule_work(&data->ih_work);
1520 int amdgpu_ras_interrupt_remove_handler(struct amdgpu_device *adev,
1521 struct ras_ih_if *info)
1523 struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head);
1524 struct ras_ih_data *data;
1529 data = &obj->ih_data;
1530 if (data->inuse == 0)
1533 cancel_work_sync(&data->ih_work);
1536 memset(data, 0, sizeof(*data));
1542 int amdgpu_ras_interrupt_add_handler(struct amdgpu_device *adev,
1543 struct ras_ih_if *info)
1545 struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head);
1546 struct ras_ih_data *data;
1549 /* in case we registe the IH before enable ras feature */
1550 obj = amdgpu_ras_create_obj(adev, &info->head);
1556 data = &obj->ih_data;
1557 /* add the callback.etc */
1558 *data = (struct ras_ih_data) {
1561 .element_size = sizeof(struct amdgpu_iv_entry),
1566 INIT_WORK(&data->ih_work, amdgpu_ras_interrupt_process_handler);
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);
1583 static int amdgpu_ras_interrupt_remove_all(struct amdgpu_device *adev)
1585 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1586 struct ras_manager *obj, *tmp;
1588 list_for_each_entry_safe(obj, tmp, &con->head, node) {
1589 struct ras_ih_if info = {
1592 amdgpu_ras_interrupt_remove_handler(adev, &info);
1599 /* traversal all IPs except NBIO to query error counter */
1600 static void amdgpu_ras_log_on_err_counter(struct amdgpu_device *adev)
1602 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1603 struct ras_manager *obj;
1605 if (!adev->ras_enabled || !con)
1608 list_for_each_entry(obj, &con->head, node) {
1609 struct ras_query_if info = {
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.
1619 if (info.head.block == AMDGPU_RAS_BLOCK__PCIE_BIF)
1622 amdgpu_ras_query_error_status(adev, &info);
1626 /* Parse RdRspStatus and WrRspStatus */
1627 static void amdgpu_ras_error_status_query(struct amdgpu_device *adev,
1628 struct ras_query_if *info)
1631 * Only two block need to query read/write
1632 * RspStatus at current state
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);
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);
1650 static void amdgpu_ras_query_err_status(struct amdgpu_device *adev)
1652 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1653 struct ras_manager *obj;
1655 if (!adev->ras_enabled || !con)
1658 list_for_each_entry(obj, &con->head, node) {
1659 struct ras_query_if info = {
1663 amdgpu_ras_error_status_query(adev, &info);
1667 /* recovery begin */
1669 /* return 0 on success.
1670 * caller need free bps.
1672 static int amdgpu_ras_badpages_read(struct amdgpu_device *adev,
1673 struct ras_badpage **bps, unsigned int *count)
1675 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1676 struct ras_err_handler_data *data;
1678 int ret = 0, status;
1680 if (!con || !con->eh_data || !bps || !count)
1683 mutex_lock(&con->recovery_lock);
1684 data = con->eh_data;
1685 if (!data || data->count == 0) {
1691 *bps = kmalloc(sizeof(struct ras_badpage) * data->count, GFP_KERNEL);
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,
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;
1712 *count = data->count;
1714 mutex_unlock(&con->recovery_lock);
1718 static void amdgpu_ras_do_recovery(struct work_struct *work)
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;
1726 if (!ras->disable_ras_err_cnt_harvest) {
1727 struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev);
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;
1733 INIT_LIST_HEAD(&device_list);
1734 list_add_tail(&adev->gmc.xgmi.head, &device_list);
1735 device_list_handle = &device_list;
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);
1744 amdgpu_put_xgmi_hive(hive);
1747 if (amdgpu_device_should_recover_gpu(ras->adev))
1748 amdgpu_device_gpu_recover(ras->adev, NULL);
1749 atomic_set(&ras->in_recovery, 0);
1752 /* alloc/realloc bps array */
1753 static int amdgpu_ras_realloc_eh_data_space(struct amdgpu_device *adev,
1754 struct ras_err_handler_data *data, int pages)
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);
1767 memcpy(bps, data->bps,
1768 data->count * sizeof(*data->bps));
1773 data->space_left += align_space - old_space;
1777 /* it deal with vram only. */
1778 int amdgpu_ras_add_bad_pages(struct amdgpu_device *adev,
1779 struct eeprom_table_record *bps, int pages)
1781 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1782 struct ras_err_handler_data *data;
1786 if (!con || !con->eh_data || !bps || pages <= 0)
1789 mutex_lock(&con->recovery_lock);
1790 data = con->eh_data;
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))
1799 if (!data->space_left &&
1800 amdgpu_ras_realloc_eh_data_space(adev, data, 256)) {
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);
1810 memcpy(&data->bps[data->count], &bps[i], sizeof(*data->bps));
1815 mutex_unlock(&con->recovery_lock);
1821 * write error record array to eeprom, the function should be
1822 * protected by recovery_lock
1824 int amdgpu_ras_save_bad_pages(struct amdgpu_device *adev)
1826 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1827 struct ras_err_handler_data *data;
1828 struct amdgpu_ras_eeprom_control *control;
1831 if (!con || !con->eh_data)
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],
1843 dev_err(adev->dev, "Failed to save EEPROM table data!");
1847 dev_info(adev->dev, "Saved %d pages to EEPROM table.\n", save_count);
1854 * read error record array in eeprom and reserve enough space for
1855 * storing new bad pages
1857 static int amdgpu_ras_load_bad_pages(struct amdgpu_device *adev)
1859 struct amdgpu_ras_eeprom_control *control =
1860 &adev->psp.ras.ras->eeprom_control;
1861 struct eeprom_table_record *bps = NULL;
1864 /* no bad page record, skip eeprom access */
1865 if (!control->num_recs || (amdgpu_bad_page_threshold == 0))
1868 bps = kcalloc(control->num_recs, sizeof(*bps), GFP_KERNEL);
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!");
1879 ret = amdgpu_ras_add_bad_pages(adev, bps, control->num_recs);
1886 static bool amdgpu_ras_check_bad_page_unlock(struct amdgpu_ras *con,
1889 struct ras_err_handler_data *data = con->eh_data;
1892 addr >>= AMDGPU_GPU_PAGE_SHIFT;
1893 for (i = 0; i < data->count; i++)
1894 if (addr == data->bps[i].retired_page)
1901 * check if an address belongs to bad page
1903 * Note: this check is only for umc block
1905 static bool amdgpu_ras_check_bad_page(struct amdgpu_device *adev,
1908 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1911 if (!con || !con->eh_data)
1914 mutex_lock(&con->recovery_lock);
1915 ret = amdgpu_ras_check_bad_page_unlock(con, addr);
1916 mutex_unlock(&con->recovery_lock);
1920 static void amdgpu_ras_validate_threshold(struct amdgpu_device *adev,
1921 uint32_t max_length)
1923 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1924 int tmp_threshold = amdgpu_bad_page_threshold;
1928 * Justification of value bad_page_cnt_threshold in ras structure
1930 * Generally, -1 <= amdgpu_bad_page_threshold <= max record length
1931 * in eeprom, and introduce two scenarios accordingly.
1933 * Bad page retirement enablement:
1934 * - If amdgpu_bad_page_threshold = -1,
1935 * bad_page_cnt_threshold = typical value by formula.
1937 * - When the value from user is 0 < amdgpu_bad_page_threshold <
1938 * max record length in eeprom, use it directly.
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
1946 if (tmp_threshold < -1)
1948 else if (tmp_threshold > max_length)
1949 tmp_threshold = max_length;
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),
1957 con->bad_page_cnt_threshold = tmp_threshold;
1961 int amdgpu_ras_recovery_init(struct amdgpu_device *adev)
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;
1969 if (adev->ras_enabled && con)
1970 data = &con->eh_data;
1974 *data = kmalloc(sizeof(**data), GFP_KERNEL | __GFP_ZERO);
1980 mutex_init(&con->recovery_lock);
1981 INIT_WORK(&con->recovery_work, amdgpu_ras_do_recovery);
1982 atomic_set(&con->in_recovery, 0);
1985 max_eeprom_records_len = amdgpu_ras_eeprom_get_record_max_length();
1986 amdgpu_ras_validate_threshold(adev, max_eeprom_records_len);
1988 /* Todo: During test the SMU might fail to read the eeprom through I2C
1989 * when the GPU is pending on XGMI reset during probe time
1990 * (Mostly after second bus reset), skip it now
1992 if (adev->gmc.xgmi.pending_reset)
1994 ret = amdgpu_ras_eeprom_init(&con->eeprom_control, &exc_err_limit);
1996 * This calling fails when exc_err_limit is true or
1999 if (exc_err_limit || ret)
2002 if (con->eeprom_control.num_recs) {
2003 ret = amdgpu_ras_load_bad_pages(adev);
2007 if (adev->smu.ppt_funcs && adev->smu.ppt_funcs->send_hbm_bad_pages_num)
2008 adev->smu.ppt_funcs->send_hbm_bad_pages_num(&adev->smu, con->eeprom_control.num_recs);
2014 kfree((*data)->bps);
2016 con->eh_data = NULL;
2018 dev_warn(adev->dev, "Failed to initialize ras recovery!\n");
2021 * Except error threshold exceeding case, other failure cases in this
2022 * function would not fail amdgpu driver init.
2032 static int amdgpu_ras_recovery_fini(struct amdgpu_device *adev)
2034 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2035 struct ras_err_handler_data *data = con->eh_data;
2037 /* recovery_init failed to init it, fini is useless */
2041 cancel_work_sync(&con->recovery_work);
2043 mutex_lock(&con->recovery_lock);
2044 con->eh_data = NULL;
2047 mutex_unlock(&con->recovery_lock);
2053 /* return 0 if ras will reset gpu and repost.*/
2054 int amdgpu_ras_request_reset_on_boot(struct amdgpu_device *adev,
2057 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
2062 ras->flags |= AMDGPU_RAS_FLAG_INIT_NEED_RESET;
2066 static bool amdgpu_ras_asic_supported(struct amdgpu_device *adev)
2068 return adev->asic_type == CHIP_VEGA10 ||
2069 adev->asic_type == CHIP_VEGA20 ||
2070 adev->asic_type == CHIP_ARCTURUS ||
2071 adev->asic_type == CHIP_ALDEBARAN ||
2072 adev->asic_type == CHIP_SIENNA_CICHLID;
2076 * this is workaround for vega20 workstation sku,
2077 * force enable gfx ras, ignore vbios gfx ras flag
2078 * due to GC EDC can not write
2080 static void amdgpu_ras_get_quirks(struct amdgpu_device *adev)
2082 struct atom_context *ctx = adev->mode_info.atom_context;
2087 if (strnstr(ctx->vbios_version, "D16406",
2088 sizeof(ctx->vbios_version)) ||
2089 strnstr(ctx->vbios_version, "D36002",
2090 sizeof(ctx->vbios_version)))
2091 adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__GFX);
2095 * check hardware's ras ability which will be saved in hw_supported.
2096 * if hardware does not support ras, we can skip some ras initializtion and
2097 * forbid some ras operations from IP.
2098 * if software itself, say boot parameter, limit the ras ability. We still
2099 * need allow IP do some limited operations, like disable. In such case,
2100 * we have to initialize ras as normal. but need check if operation is
2101 * allowed or not in each function.
2103 static void amdgpu_ras_check_supported(struct amdgpu_device *adev)
2105 adev->ras_hw_enabled = adev->ras_enabled = 0;
2107 if (amdgpu_sriov_vf(adev) || !adev->is_atom_fw ||
2108 !amdgpu_ras_asic_supported(adev))
2111 if (!adev->gmc.xgmi.connected_to_cpu) {
2112 if (amdgpu_atomfirmware_mem_ecc_supported(adev)) {
2113 dev_info(adev->dev, "MEM ECC is active.\n");
2114 adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__UMC |
2115 1 << AMDGPU_RAS_BLOCK__DF);
2117 dev_info(adev->dev, "MEM ECC is not presented.\n");
2120 if (amdgpu_atomfirmware_sram_ecc_supported(adev)) {
2121 dev_info(adev->dev, "SRAM ECC is active.\n");
2122 adev->ras_hw_enabled |= ~(1 << AMDGPU_RAS_BLOCK__UMC |
2123 1 << AMDGPU_RAS_BLOCK__DF);
2125 dev_info(adev->dev, "SRAM ECC is not presented.\n");
2128 /* driver only manages a few IP blocks RAS feature
2129 * when GPU is connected cpu through XGMI */
2130 adev->ras_hw_enabled |= (1 << AMDGPU_RAS_BLOCK__GFX |
2131 1 << AMDGPU_RAS_BLOCK__SDMA |
2132 1 << AMDGPU_RAS_BLOCK__MMHUB);
2135 amdgpu_ras_get_quirks(adev);
2137 /* hw_supported needs to be aligned with RAS block mask. */
2138 adev->ras_hw_enabled &= AMDGPU_RAS_BLOCK_MASK;
2140 adev->ras_enabled = amdgpu_ras_enable == 0 ? 0 :
2141 adev->ras_hw_enabled & amdgpu_ras_mask;
2144 static void amdgpu_ras_counte_dw(struct work_struct *work)
2146 struct amdgpu_ras *con = container_of(work, struct amdgpu_ras,
2147 ras_counte_delay_work.work);
2148 struct amdgpu_device *adev = con->adev;
2149 struct drm_device *dev = adev_to_drm(adev);
2150 unsigned long ce_count, ue_count;
2153 res = pm_runtime_get_sync(dev->dev);
2157 /* Cache new values.
2159 if (amdgpu_ras_query_error_count(adev, &ce_count, &ue_count) == 0) {
2160 atomic_set(&con->ras_ce_count, ce_count);
2161 atomic_set(&con->ras_ue_count, ue_count);
2164 pm_runtime_mark_last_busy(dev->dev);
2166 pm_runtime_put_autosuspend(dev->dev);
2169 int amdgpu_ras_init(struct amdgpu_device *adev)
2171 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2177 con = kmalloc(sizeof(struct amdgpu_ras) +
2178 sizeof(struct ras_manager) * AMDGPU_RAS_BLOCK_COUNT,
2179 GFP_KERNEL|__GFP_ZERO);
2184 INIT_DELAYED_WORK(&con->ras_counte_delay_work, amdgpu_ras_counte_dw);
2185 atomic_set(&con->ras_ce_count, 0);
2186 atomic_set(&con->ras_ue_count, 0);
2188 con->objs = (struct ras_manager *)(con + 1);
2190 amdgpu_ras_set_context(adev, con);
2192 amdgpu_ras_check_supported(adev);
2194 if (!adev->ras_enabled || adev->asic_type == CHIP_VEGA10) {
2195 /* set gfx block ras context feature for VEGA20 Gaming
2196 * send ras disable cmd to ras ta during ras late init.
2198 if (!adev->ras_enabled && adev->asic_type == CHIP_VEGA20) {
2199 con->features |= BIT(AMDGPU_RAS_BLOCK__GFX);
2209 INIT_LIST_HEAD(&con->head);
2210 /* Might need get this flag from vbios. */
2211 con->flags = RAS_DEFAULT_FLAGS;
2213 /* initialize nbio ras function ahead of any other
2214 * ras functions so hardware fatal error interrupt
2215 * can be enabled as early as possible */
2216 switch (adev->asic_type) {
2219 case CHIP_ALDEBARAN:
2220 if (!adev->gmc.xgmi.connected_to_cpu)
2221 adev->nbio.ras_funcs = &nbio_v7_4_ras_funcs;
2224 /* nbio ras is not available */
2228 if (adev->nbio.ras_funcs &&
2229 adev->nbio.ras_funcs->init_ras_controller_interrupt) {
2230 r = adev->nbio.ras_funcs->init_ras_controller_interrupt(adev);
2235 if (adev->nbio.ras_funcs &&
2236 adev->nbio.ras_funcs->init_ras_err_event_athub_interrupt) {
2237 r = adev->nbio.ras_funcs->init_ras_err_event_athub_interrupt(adev);
2242 if (amdgpu_ras_fs_init(adev)) {
2247 dev_info(adev->dev, "RAS INFO: ras initialized successfully, "
2248 "hardware ability[%x] ras_mask[%x]\n",
2249 adev->ras_hw_enabled, adev->ras_enabled);
2253 amdgpu_ras_set_context(adev, NULL);
2259 int amdgpu_persistent_edc_harvesting_supported(struct amdgpu_device *adev)
2261 if (adev->gmc.xgmi.connected_to_cpu)
2266 static int amdgpu_persistent_edc_harvesting(struct amdgpu_device *adev,
2267 struct ras_common_if *ras_block)
2269 struct ras_query_if info = {
2273 if (!amdgpu_persistent_edc_harvesting_supported(adev))
2276 if (amdgpu_ras_query_error_status(adev, &info) != 0)
2277 DRM_WARN("RAS init harvest failure");
2279 if (amdgpu_ras_reset_error_status(adev, ras_block->block) != 0)
2280 DRM_WARN("RAS init harvest reset failure");
2285 /* helper function to handle common stuff in ip late init phase */
2286 int amdgpu_ras_late_init(struct amdgpu_device *adev,
2287 struct ras_common_if *ras_block,
2288 struct ras_fs_if *fs_info,
2289 struct ras_ih_if *ih_info)
2291 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2292 unsigned long ue_count, ce_count;
2295 /* disable RAS feature per IP block if it is not supported */
2296 if (!amdgpu_ras_is_supported(adev, ras_block->block)) {
2297 amdgpu_ras_feature_enable_on_boot(adev, ras_block, 0);
2301 r = amdgpu_ras_feature_enable_on_boot(adev, ras_block, 1);
2304 /* request gpu reset. will run again */
2305 amdgpu_ras_request_reset_on_boot(adev,
2308 } else if (adev->in_suspend || amdgpu_in_reset(adev)) {
2309 /* in resume phase, if fail to enable ras,
2310 * clean up all ras fs nodes, and disable ras */
2316 /* check for errors on warm reset edc persisant supported ASIC */
2317 amdgpu_persistent_edc_harvesting(adev, ras_block);
2319 /* in resume phase, no need to create ras fs node */
2320 if (adev->in_suspend || amdgpu_in_reset(adev))
2324 r = amdgpu_ras_interrupt_add_handler(adev, ih_info);
2329 r = amdgpu_ras_sysfs_create(adev, fs_info);
2333 /* Those are the cached values at init.
2335 if (amdgpu_ras_query_error_count(adev, &ce_count, &ue_count) == 0) {
2336 atomic_set(&con->ras_ce_count, ce_count);
2337 atomic_set(&con->ras_ue_count, ue_count);
2342 amdgpu_ras_sysfs_remove(adev, ras_block);
2345 amdgpu_ras_interrupt_remove_handler(adev, ih_info);
2347 amdgpu_ras_feature_enable(adev, ras_block, 0);
2351 /* helper function to remove ras fs node and interrupt handler */
2352 void amdgpu_ras_late_fini(struct amdgpu_device *adev,
2353 struct ras_common_if *ras_block,
2354 struct ras_ih_if *ih_info)
2356 if (!ras_block || !ih_info)
2359 amdgpu_ras_sysfs_remove(adev, ras_block);
2361 amdgpu_ras_interrupt_remove_handler(adev, ih_info);
2362 amdgpu_ras_feature_enable(adev, ras_block, 0);
2365 /* do some init work after IP late init as dependence.
2366 * and it runs in resume/gpu reset/booting up cases.
2368 void amdgpu_ras_resume(struct amdgpu_device *adev)
2370 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2371 struct ras_manager *obj, *tmp;
2373 if (!adev->ras_enabled || !con) {
2374 /* clean ras context for VEGA20 Gaming after send ras disable cmd */
2375 amdgpu_release_ras_context(adev);
2380 if (con->flags & AMDGPU_RAS_FLAG_INIT_BY_VBIOS) {
2381 /* Set up all other IPs which are not implemented. There is a
2382 * tricky thing that IP's actual ras error type should be
2383 * MULTI_UNCORRECTABLE, but as driver does not handle it, so
2384 * ERROR_NONE make sense anyway.
2386 amdgpu_ras_enable_all_features(adev, 1);
2388 /* We enable ras on all hw_supported block, but as boot
2389 * parameter might disable some of them and one or more IP has
2390 * not implemented yet. So we disable them on behalf.
2392 list_for_each_entry_safe(obj, tmp, &con->head, node) {
2393 if (!amdgpu_ras_is_supported(adev, obj->head.block)) {
2394 amdgpu_ras_feature_enable(adev, &obj->head, 0);
2395 /* there should be no any reference. */
2396 WARN_ON(alive_obj(obj));
2401 if (con->flags & AMDGPU_RAS_FLAG_INIT_NEED_RESET) {
2402 con->flags &= ~AMDGPU_RAS_FLAG_INIT_NEED_RESET;
2403 /* setup ras obj state as disabled.
2404 * for init_by_vbios case.
2405 * if we want to enable ras, just enable it in a normal way.
2406 * If we want do disable it, need setup ras obj as enabled,
2407 * then issue another TA disable cmd.
2408 * See feature_enable_on_boot
2410 amdgpu_ras_disable_all_features(adev, 1);
2411 amdgpu_ras_reset_gpu(adev);
2415 void amdgpu_ras_suspend(struct amdgpu_device *adev)
2417 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2419 if (!adev->ras_enabled || !con)
2422 amdgpu_ras_disable_all_features(adev, 0);
2423 /* Make sure all ras objects are disabled. */
2425 amdgpu_ras_disable_all_features(adev, 1);
2428 /* do some fini work before IP fini as dependence */
2429 int amdgpu_ras_pre_fini(struct amdgpu_device *adev)
2431 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2433 if (!adev->ras_enabled || !con)
2437 /* Need disable ras on all IPs here before ip [hw/sw]fini */
2438 amdgpu_ras_disable_all_features(adev, 0);
2439 amdgpu_ras_recovery_fini(adev);
2443 int amdgpu_ras_fini(struct amdgpu_device *adev)
2445 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2447 if (!adev->ras_enabled || !con)
2450 amdgpu_ras_fs_fini(adev);
2451 amdgpu_ras_interrupt_remove_all(adev);
2453 WARN(con->features, "Feature mask is not cleared");
2456 amdgpu_ras_disable_all_features(adev, 1);
2458 cancel_delayed_work_sync(&con->ras_counte_delay_work);
2460 amdgpu_ras_set_context(adev, NULL);
2466 void amdgpu_ras_global_ras_isr(struct amdgpu_device *adev)
2468 amdgpu_ras_check_supported(adev);
2469 if (!adev->ras_hw_enabled)
2472 if (atomic_cmpxchg(&amdgpu_ras_in_intr, 0, 1) == 0) {
2473 dev_info(adev->dev, "uncorrectable hardware error"
2474 "(ERREVENT_ATHUB_INTERRUPT) detected!\n");
2476 amdgpu_ras_reset_gpu(adev);
2480 bool amdgpu_ras_need_emergency_restart(struct amdgpu_device *adev)
2482 if (adev->asic_type == CHIP_VEGA20 &&
2483 adev->pm.fw_version <= 0x283400) {
2484 return !(amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) &&
2485 amdgpu_ras_intr_triggered();
2491 void amdgpu_release_ras_context(struct amdgpu_device *adev)
2493 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
2498 if (!adev->ras_enabled && con->features & BIT(AMDGPU_RAS_BLOCK__GFX)) {
2499 con->features &= ~BIT(AMDGPU_RAS_BLOCK__GFX);
2500 amdgpu_ras_set_context(adev, NULL);