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>
32 #include "amdgpu_ras.h"
33 #include "amdgpu_atomfirmware.h"
34 #include "ivsrcid/nbio/irqsrcs_nbif_7_4.h"
36 const char *ras_error_string[] = {
40 "multi_uncorrectable",
44 const char *ras_block_string[] = {
61 #define ras_err_str(i) (ras_error_string[ffs(i)])
62 #define ras_block_str(i) (ras_block_string[i])
64 #define AMDGPU_RAS_FLAG_INIT_BY_VBIOS 1
65 #define AMDGPU_RAS_FLAG_INIT_NEED_RESET 2
66 #define RAS_DEFAULT_FLAGS (AMDGPU_RAS_FLAG_INIT_BY_VBIOS)
68 /* inject address is 52 bits */
69 #define RAS_UMC_INJECT_ADDR_LIMIT (0x1ULL << 52)
71 enum amdgpu_ras_retire_page_reservation {
72 AMDGPU_RAS_RETIRE_PAGE_RESERVED,
73 AMDGPU_RAS_RETIRE_PAGE_PENDING,
74 AMDGPU_RAS_RETIRE_PAGE_FAULT,
77 atomic_t amdgpu_ras_in_intr = ATOMIC_INIT(0);
79 static bool amdgpu_ras_check_bad_page(struct amdgpu_device *adev,
82 static ssize_t amdgpu_ras_debugfs_read(struct file *f, char __user *buf,
83 size_t size, loff_t *pos)
85 struct ras_manager *obj = (struct ras_manager *)file_inode(f)->i_private;
86 struct ras_query_if info = {
92 if (amdgpu_ras_error_query(obj->adev, &info))
95 s = snprintf(val, sizeof(val), "%s: %lu\n%s: %lu\n",
102 s = min_t(u64, s, size);
105 if (copy_to_user(buf, &val[*pos], s))
113 static const struct file_operations amdgpu_ras_debugfs_ops = {
114 .owner = THIS_MODULE,
115 .read = amdgpu_ras_debugfs_read,
117 .llseek = default_llseek
120 static int amdgpu_ras_find_block_id_by_name(const char *name, int *block_id)
124 for (i = 0; i < ARRAY_SIZE(ras_block_string); i++) {
126 if (strcmp(name, ras_block_str(i)) == 0)
132 static int amdgpu_ras_debugfs_ctrl_parse_data(struct file *f,
133 const char __user *buf, size_t size,
134 loff_t *pos, struct ras_debug_if *data)
136 ssize_t s = min_t(u64, 64, size);
149 memset(str, 0, sizeof(str));
150 memset(data, 0, sizeof(*data));
152 if (copy_from_user(str, buf, s))
155 if (sscanf(str, "disable %32s", block_name) == 1)
157 else if (sscanf(str, "enable %32s %8s", block_name, err) == 2)
159 else if (sscanf(str, "inject %32s %8s", block_name, err) == 2)
161 else if (str[0] && str[1] && str[2] && str[3])
162 /* ascii string, but commands are not matched. */
166 if (amdgpu_ras_find_block_id_by_name(block_name, &block_id))
169 data->head.block = block_id;
170 /* only ue and ce errors are supported */
171 if (!memcmp("ue", err, 2))
172 data->head.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE;
173 else if (!memcmp("ce", err, 2))
174 data->head.type = AMDGPU_RAS_ERROR__SINGLE_CORRECTABLE;
181 if (sscanf(str, "%*s %*s %*s %u %llu %llu",
182 &sub_block, &address, &value) != 3)
183 if (sscanf(str, "%*s %*s %*s 0x%x 0x%llx 0x%llx",
184 &sub_block, &address, &value) != 3)
186 data->head.sub_block_index = sub_block;
187 data->inject.address = address;
188 data->inject.value = value;
191 if (size < sizeof(*data))
194 if (copy_from_user(data, buf, sizeof(*data)))
202 * DOC: AMDGPU RAS debugfs control interface
204 * It accepts struct ras_debug_if who has two members.
206 * First member: ras_debug_if::head or ras_debug_if::inject.
208 * head is used to indicate which IP block will be under control.
210 * head has four members, they are block, type, sub_block_index, name.
211 * block: which IP will be under control.
212 * type: what kind of error will be enabled/disabled/injected.
213 * sub_block_index: some IPs have subcomponets. say, GFX, sDMA.
214 * name: the name of IP.
216 * inject has two more members than head, they are address, value.
217 * As their names indicate, inject operation will write the
218 * value to the address.
220 * The second member: struct ras_debug_if::op.
221 * It has three kinds of operations.
223 * - 0: disable RAS on the block. Take ::head as its data.
224 * - 1: enable RAS on the block. Take ::head as its data.
225 * - 2: inject errors on the block. Take ::inject as its data.
227 * How to use the interface?
231 * Copy the struct ras_debug_if in your codes and initialize it.
232 * Write the struct to the control node.
236 * .. code-block:: bash
238 * echo op block [error [sub_block address value]] > .../ras/ras_ctrl
242 * op: disable, enable, inject
243 * disable: only block is needed
244 * enable: block and error are needed
245 * inject: error, address, value are needed
246 * block: umc, sdma, gfx, .........
247 * see ras_block_string[] for details
249 * ue: multi_uncorrectable
250 * ce: single_correctable
252 * sub block index, pass 0 if there is no sub block
254 * here are some examples for bash commands:
256 * .. code-block:: bash
258 * echo inject umc ue 0x0 0x0 0x0 > /sys/kernel/debug/dri/0/ras/ras_ctrl
259 * echo inject umc ce 0 0 0 > /sys/kernel/debug/dri/0/ras/ras_ctrl
260 * echo disable umc > /sys/kernel/debug/dri/0/ras/ras_ctrl
262 * How to check the result?
264 * For disable/enable, please check ras features at
265 * /sys/class/drm/card[0/1/2...]/device/ras/features
267 * For inject, please check corresponding err count at
268 * /sys/class/drm/card[0/1/2...]/device/ras/[gfx/sdma/...]_err_count
271 * Operations are only allowed on blocks which are supported.
272 * Please check ras mask at /sys/module/amdgpu/parameters/ras_mask
273 * to see which blocks support RAS on a particular asic.
276 static ssize_t amdgpu_ras_debugfs_ctrl_write(struct file *f, const char __user *buf,
277 size_t size, loff_t *pos)
279 struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
280 struct ras_debug_if data;
283 ret = amdgpu_ras_debugfs_ctrl_parse_data(f, buf, size, pos, &data);
287 if (!amdgpu_ras_is_supported(adev, data.head.block))
292 ret = amdgpu_ras_feature_enable(adev, &data.head, 0);
295 ret = amdgpu_ras_feature_enable(adev, &data.head, 1);
298 if ((data.inject.address >= adev->gmc.mc_vram_size) ||
299 (data.inject.address >= RAS_UMC_INJECT_ADDR_LIMIT)) {
304 /* umc ce/ue error injection for a bad page is not allowed */
305 if ((data.head.block == AMDGPU_RAS_BLOCK__UMC) &&
306 amdgpu_ras_check_bad_page(adev, data.inject.address)) {
307 DRM_WARN("RAS WARN: 0x%llx has been marked as bad before error injection!\n",
308 data.inject.address);
312 /* data.inject.address is offset instead of absolute gpu address */
313 ret = amdgpu_ras_error_inject(adev, &data.inject);
327 * DOC: AMDGPU RAS debugfs EEPROM table reset interface
329 * Some boards contain an EEPROM which is used to persistently store a list of
330 * bad pages which experiences ECC errors in vram. This interface provides
331 * a way to reset the EEPROM, e.g., after testing error injection.
335 * .. code-block:: bash
337 * echo 1 > ../ras/ras_eeprom_reset
339 * will reset EEPROM table to 0 entries.
342 static ssize_t amdgpu_ras_debugfs_eeprom_write(struct file *f, const char __user *buf,
343 size_t size, loff_t *pos)
345 struct amdgpu_device *adev = (struct amdgpu_device *)file_inode(f)->i_private;
348 ret = amdgpu_ras_eeprom_reset_table(&adev->psp.ras.ras->eeprom_control);
350 return ret == 1 ? size : -EIO;
353 static const struct file_operations amdgpu_ras_debugfs_ctrl_ops = {
354 .owner = THIS_MODULE,
356 .write = amdgpu_ras_debugfs_ctrl_write,
357 .llseek = default_llseek
360 static const struct file_operations amdgpu_ras_debugfs_eeprom_ops = {
361 .owner = THIS_MODULE,
363 .write = amdgpu_ras_debugfs_eeprom_write,
364 .llseek = default_llseek
368 * DOC: AMDGPU RAS sysfs Error Count Interface
370 * It allows the user to read the error count for each IP block on the gpu through
371 * /sys/class/drm/card[0/1/2...]/device/ras/[gfx/sdma/...]_err_count
373 * It outputs the multiple lines which report the uncorrected (ue) and corrected
376 * The format of one line is below,
382 * .. code-block:: bash
388 static ssize_t amdgpu_ras_sysfs_read(struct device *dev,
389 struct device_attribute *attr, char *buf)
391 struct ras_manager *obj = container_of(attr, struct ras_manager, sysfs_attr);
392 struct ras_query_if info = {
396 if (amdgpu_ras_error_query(obj->adev, &info))
399 return snprintf(buf, PAGE_SIZE, "%s: %lu\n%s: %lu\n",
401 "ce", info.ce_count);
406 #define get_obj(obj) do { (obj)->use++; } while (0)
407 #define alive_obj(obj) ((obj)->use)
409 static inline void put_obj(struct ras_manager *obj)
411 if (obj && --obj->use == 0)
412 list_del(&obj->node);
413 if (obj && obj->use < 0) {
414 DRM_ERROR("RAS ERROR: Unbalance obj(%s) use\n", obj->head.name);
418 /* make one obj and return it. */
419 static struct ras_manager *amdgpu_ras_create_obj(struct amdgpu_device *adev,
420 struct ras_common_if *head)
422 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
423 struct ras_manager *obj;
428 if (head->block >= AMDGPU_RAS_BLOCK_COUNT)
431 obj = &con->objs[head->block];
432 /* already exist. return obj? */
438 list_add(&obj->node, &con->head);
444 /* return an obj equal to head, or the first when head is NULL */
445 struct ras_manager *amdgpu_ras_find_obj(struct amdgpu_device *adev,
446 struct ras_common_if *head)
448 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
449 struct ras_manager *obj;
456 if (head->block >= AMDGPU_RAS_BLOCK_COUNT)
459 obj = &con->objs[head->block];
461 if (alive_obj(obj)) {
462 WARN_ON(head->block != obj->head.block);
466 for (i = 0; i < AMDGPU_RAS_BLOCK_COUNT; i++) {
468 if (alive_obj(obj)) {
469 WARN_ON(i != obj->head.block);
479 /* feature ctl begin */
480 static int amdgpu_ras_is_feature_allowed(struct amdgpu_device *adev,
481 struct ras_common_if *head)
483 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
485 return con->hw_supported & BIT(head->block);
488 static int amdgpu_ras_is_feature_enabled(struct amdgpu_device *adev,
489 struct ras_common_if *head)
491 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
493 return con->features & BIT(head->block);
497 * if obj is not created, then create one.
498 * set feature enable flag.
500 static int __amdgpu_ras_feature_enable(struct amdgpu_device *adev,
501 struct ras_common_if *head, int enable)
503 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
504 struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
506 /* If hardware does not support ras, then do not create obj.
507 * But if hardware support ras, we can create the obj.
508 * Ras framework checks con->hw_supported to see if it need do
509 * corresponding initialization.
510 * IP checks con->support to see if it need disable ras.
512 if (!amdgpu_ras_is_feature_allowed(adev, head))
514 if (!(!!enable ^ !!amdgpu_ras_is_feature_enabled(adev, head)))
519 obj = amdgpu_ras_create_obj(adev, head);
523 /* In case we create obj somewhere else */
526 con->features |= BIT(head->block);
528 if (obj && amdgpu_ras_is_feature_enabled(adev, head)) {
529 con->features &= ~BIT(head->block);
537 /* wrapper of psp_ras_enable_features */
538 int amdgpu_ras_feature_enable(struct amdgpu_device *adev,
539 struct ras_common_if *head, bool enable)
541 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
542 union ta_ras_cmd_input info;
549 info.disable_features = (struct ta_ras_disable_features_input) {
550 .block_id = amdgpu_ras_block_to_ta(head->block),
551 .error_type = amdgpu_ras_error_to_ta(head->type),
554 info.enable_features = (struct ta_ras_enable_features_input) {
555 .block_id = amdgpu_ras_block_to_ta(head->block),
556 .error_type = amdgpu_ras_error_to_ta(head->type),
560 /* Do not enable if it is not allowed. */
561 WARN_ON(enable && !amdgpu_ras_is_feature_allowed(adev, head));
562 /* Are we alerady in that state we are going to set? */
563 if (!(!!enable ^ !!amdgpu_ras_is_feature_enabled(adev, head)))
566 if (!amdgpu_ras_intr_triggered()) {
567 ret = psp_ras_enable_features(&adev->psp, &info, enable);
569 DRM_ERROR("RAS ERROR: %s %s feature failed ret %d\n",
570 enable ? "enable":"disable",
571 ras_block_str(head->block),
573 if (ret == TA_RAS_STATUS__RESET_NEEDED)
580 __amdgpu_ras_feature_enable(adev, head, enable);
585 /* Only used in device probe stage and called only once. */
586 int amdgpu_ras_feature_enable_on_boot(struct amdgpu_device *adev,
587 struct ras_common_if *head, bool enable)
589 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
595 if (con->flags & AMDGPU_RAS_FLAG_INIT_BY_VBIOS) {
597 /* There is no harm to issue a ras TA cmd regardless of
598 * the currecnt ras state.
599 * If current state == target state, it will do nothing
600 * But sometimes it requests driver to reset and repost
601 * with error code -EAGAIN.
603 ret = amdgpu_ras_feature_enable(adev, head, 1);
604 /* With old ras TA, we might fail to enable ras.
605 * Log it and just setup the object.
606 * TODO need remove this WA in the future.
608 if (ret == -EINVAL) {
609 ret = __amdgpu_ras_feature_enable(adev, head, 1);
611 DRM_INFO("RAS INFO: %s setup object\n",
612 ras_block_str(head->block));
615 /* setup the object then issue a ras TA disable cmd.*/
616 ret = __amdgpu_ras_feature_enable(adev, head, 1);
620 ret = amdgpu_ras_feature_enable(adev, head, 0);
623 ret = amdgpu_ras_feature_enable(adev, head, enable);
628 static int amdgpu_ras_disable_all_features(struct amdgpu_device *adev,
631 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
632 struct ras_manager *obj, *tmp;
634 list_for_each_entry_safe(obj, tmp, &con->head, node) {
636 * aka just release the obj and corresponding flags
639 if (__amdgpu_ras_feature_enable(adev, &obj->head, 0))
642 if (amdgpu_ras_feature_enable(adev, &obj->head, 0))
647 return con->features;
650 static int amdgpu_ras_enable_all_features(struct amdgpu_device *adev,
653 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
654 int ras_block_count = AMDGPU_RAS_BLOCK_COUNT;
656 const enum amdgpu_ras_error_type default_ras_type =
657 AMDGPU_RAS_ERROR__NONE;
659 for (i = 0; i < ras_block_count; i++) {
660 struct ras_common_if head = {
662 .type = default_ras_type,
663 .sub_block_index = 0,
665 strcpy(head.name, ras_block_str(i));
668 * bypass psp. vbios enable ras for us.
669 * so just create the obj
671 if (__amdgpu_ras_feature_enable(adev, &head, 1))
674 if (amdgpu_ras_feature_enable(adev, &head, 1))
679 return con->features;
681 /* feature ctl end */
683 /* query/inject/cure begin */
684 int amdgpu_ras_error_query(struct amdgpu_device *adev,
685 struct ras_query_if *info)
687 struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head);
688 struct ras_err_data err_data = {0, 0, 0, NULL};
694 switch (info->head.block) {
695 case AMDGPU_RAS_BLOCK__UMC:
696 if (adev->umc.funcs->query_ras_error_count)
697 adev->umc.funcs->query_ras_error_count(adev, &err_data);
698 /* umc query_ras_error_address is also responsible for clearing
701 if (adev->umc.funcs->query_ras_error_address)
702 adev->umc.funcs->query_ras_error_address(adev, &err_data);
704 case AMDGPU_RAS_BLOCK__SDMA:
705 if (adev->sdma.funcs->query_ras_error_count) {
706 for (i = 0; i < adev->sdma.num_instances; i++)
707 adev->sdma.funcs->query_ras_error_count(adev, i,
711 case AMDGPU_RAS_BLOCK__GFX:
712 if (adev->gfx.funcs->query_ras_error_count)
713 adev->gfx.funcs->query_ras_error_count(adev, &err_data);
715 case AMDGPU_RAS_BLOCK__MMHUB:
716 if (adev->mmhub.funcs->query_ras_error_count)
717 adev->mmhub.funcs->query_ras_error_count(adev, &err_data);
719 case AMDGPU_RAS_BLOCK__PCIE_BIF:
720 if (adev->nbio.funcs->query_ras_error_count)
721 adev->nbio.funcs->query_ras_error_count(adev, &err_data);
727 obj->err_data.ue_count += err_data.ue_count;
728 obj->err_data.ce_count += err_data.ce_count;
730 info->ue_count = obj->err_data.ue_count;
731 info->ce_count = obj->err_data.ce_count;
733 if (err_data.ce_count) {
734 dev_info(adev->dev, "%ld correctable errors detected in %s block\n",
735 obj->err_data.ce_count, ras_block_str(info->head.block));
737 if (err_data.ue_count) {
738 dev_info(adev->dev, "%ld uncorrectable errors detected in %s block\n",
739 obj->err_data.ue_count, ras_block_str(info->head.block));
745 uint64_t get_xgmi_relative_phy_addr(struct amdgpu_device *adev, uint64_t addr)
749 if ((!adev->df.funcs) ||
750 (!adev->df.funcs->get_df_inst_id) ||
751 (!adev->df.funcs->get_dram_base_addr))
754 df_inst_id = adev->df.funcs->get_df_inst_id(adev);
756 return addr + adev->df.funcs->get_dram_base_addr(adev, df_inst_id);
759 /* wrapper of psp_ras_trigger_error */
760 int amdgpu_ras_error_inject(struct amdgpu_device *adev,
761 struct ras_inject_if *info)
763 struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head);
764 struct ta_ras_trigger_error_input block_info = {
765 .block_id = amdgpu_ras_block_to_ta(info->head.block),
766 .inject_error_type = amdgpu_ras_error_to_ta(info->head.type),
767 .sub_block_index = info->head.sub_block_index,
768 .address = info->address,
769 .value = info->value,
776 /* Calculate XGMI relative offset */
777 if (adev->gmc.xgmi.num_physical_nodes > 1) {
778 block_info.address = get_xgmi_relative_phy_addr(adev,
782 switch (info->head.block) {
783 case AMDGPU_RAS_BLOCK__GFX:
784 if (adev->gfx.funcs->ras_error_inject)
785 ret = adev->gfx.funcs->ras_error_inject(adev, info);
789 case AMDGPU_RAS_BLOCK__UMC:
790 case AMDGPU_RAS_BLOCK__MMHUB:
791 case AMDGPU_RAS_BLOCK__XGMI_WAFL:
792 case AMDGPU_RAS_BLOCK__PCIE_BIF:
793 ret = psp_ras_trigger_error(&adev->psp, &block_info);
796 DRM_INFO("%s error injection is not supported yet\n",
797 ras_block_str(info->head.block));
802 DRM_ERROR("RAS ERROR: inject %s error failed ret %d\n",
803 ras_block_str(info->head.block),
809 int amdgpu_ras_error_cure(struct amdgpu_device *adev,
810 struct ras_cure_if *info)
812 /* psp fw has no cure interface for now. */
816 /* get the total error counts on all IPs */
817 unsigned long amdgpu_ras_query_error_count(struct amdgpu_device *adev,
820 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
821 struct ras_manager *obj;
822 struct ras_err_data data = {0, 0};
827 list_for_each_entry(obj, &con->head, node) {
828 struct ras_query_if info = {
832 if (amdgpu_ras_error_query(adev, &info))
835 data.ce_count += info.ce_count;
836 data.ue_count += info.ue_count;
839 return is_ce ? data.ce_count : data.ue_count;
841 /* query/inject/cure end */
846 static int amdgpu_ras_badpages_read(struct amdgpu_device *adev,
847 struct ras_badpage **bps, unsigned int *count);
849 static char *amdgpu_ras_badpage_flags_str(unsigned int flags)
852 case AMDGPU_RAS_RETIRE_PAGE_RESERVED:
854 case AMDGPU_RAS_RETIRE_PAGE_PENDING:
856 case AMDGPU_RAS_RETIRE_PAGE_FAULT:
863 * DOC: AMDGPU RAS sysfs gpu_vram_bad_pages Interface
865 * It allows user to read the bad pages of vram on the gpu through
866 * /sys/class/drm/card[0/1/2...]/device/ras/gpu_vram_bad_pages
868 * It outputs multiple lines, and each line stands for one gpu page.
870 * The format of one line is below,
871 * gpu pfn : gpu page size : flags
873 * gpu pfn and gpu page size are printed in hex format.
874 * flags can be one of below character,
876 * R: reserved, this gpu page is reserved and not able to use.
878 * P: pending for reserve, this gpu page is marked as bad, will be reserved
879 * in next window of page_reserve.
881 * F: unable to reserve. this gpu page can't be reserved due to some reasons.
885 * .. code-block:: bash
887 * 0x00000001 : 0x00001000 : R
888 * 0x00000002 : 0x00001000 : P
892 static ssize_t amdgpu_ras_sysfs_badpages_read(struct file *f,
893 struct kobject *kobj, struct bin_attribute *attr,
894 char *buf, loff_t ppos, size_t count)
896 struct amdgpu_ras *con =
897 container_of(attr, struct amdgpu_ras, badpages_attr);
898 struct amdgpu_device *adev = con->adev;
899 const unsigned int element_size =
900 sizeof("0xabcdabcd : 0x12345678 : R\n") - 1;
901 unsigned int start = div64_ul(ppos + element_size - 1, element_size);
902 unsigned int end = div64_ul(ppos + count - 1, element_size);
904 struct ras_badpage *bps = NULL;
905 unsigned int bps_count = 0;
907 memset(buf, 0, count);
909 if (amdgpu_ras_badpages_read(adev, &bps, &bps_count))
912 for (; start < end && start < bps_count; start++)
913 s += scnprintf(&buf[s], element_size + 1,
914 "0x%08x : 0x%08x : %1s\n",
917 amdgpu_ras_badpage_flags_str(bps[start].flags));
924 static ssize_t amdgpu_ras_sysfs_features_read(struct device *dev,
925 struct device_attribute *attr, char *buf)
927 struct amdgpu_ras *con =
928 container_of(attr, struct amdgpu_ras, features_attr);
930 return scnprintf(buf, PAGE_SIZE, "feature mask: 0x%x\n", con->features);
933 static int amdgpu_ras_sysfs_create_feature_node(struct amdgpu_device *adev)
935 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
936 struct attribute *attrs[] = {
937 &con->features_attr.attr,
940 struct bin_attribute *bin_attrs[] = {
944 struct attribute_group group = {
947 .bin_attrs = bin_attrs,
950 con->features_attr = (struct device_attribute) {
955 .show = amdgpu_ras_sysfs_features_read,
958 con->badpages_attr = (struct bin_attribute) {
960 .name = "gpu_vram_bad_pages",
965 .read = amdgpu_ras_sysfs_badpages_read,
968 sysfs_attr_init(attrs[0]);
969 sysfs_bin_attr_init(bin_attrs[0]);
971 return sysfs_create_group(&adev->dev->kobj, &group);
974 static int amdgpu_ras_sysfs_remove_feature_node(struct amdgpu_device *adev)
976 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
977 struct attribute *attrs[] = {
978 &con->features_attr.attr,
981 struct bin_attribute *bin_attrs[] = {
985 struct attribute_group group = {
988 .bin_attrs = bin_attrs,
991 sysfs_remove_group(&adev->dev->kobj, &group);
996 int amdgpu_ras_sysfs_create(struct amdgpu_device *adev,
997 struct ras_fs_if *head)
999 struct ras_manager *obj = amdgpu_ras_find_obj(adev, &head->head);
1001 if (!obj || obj->attr_inuse)
1006 memcpy(obj->fs_data.sysfs_name,
1008 sizeof(obj->fs_data.sysfs_name));
1010 obj->sysfs_attr = (struct device_attribute){
1012 .name = obj->fs_data.sysfs_name,
1015 .show = amdgpu_ras_sysfs_read,
1017 sysfs_attr_init(&obj->sysfs_attr.attr);
1019 if (sysfs_add_file_to_group(&adev->dev->kobj,
1020 &obj->sysfs_attr.attr,
1026 obj->attr_inuse = 1;
1031 int amdgpu_ras_sysfs_remove(struct amdgpu_device *adev,
1032 struct ras_common_if *head)
1034 struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
1036 if (!obj || !obj->attr_inuse)
1039 sysfs_remove_file_from_group(&adev->dev->kobj,
1040 &obj->sysfs_attr.attr,
1042 obj->attr_inuse = 0;
1048 static int amdgpu_ras_sysfs_remove_all(struct amdgpu_device *adev)
1050 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1051 struct ras_manager *obj, *tmp;
1053 list_for_each_entry_safe(obj, tmp, &con->head, node) {
1054 amdgpu_ras_sysfs_remove(adev, &obj->head);
1057 amdgpu_ras_sysfs_remove_feature_node(adev);
1064 * DOC: AMDGPU RAS Reboot Behavior for Unrecoverable Errors
1066 * Normally when there is an uncorrectable error, the driver will reset
1067 * the GPU to recover. However, in the event of an unrecoverable error,
1068 * the driver provides an interface to reboot the system automatically
1071 * The following file in debugfs provides that interface:
1072 * /sys/kernel/debug/dri/[0/1/2...]/ras/auto_reboot
1076 * .. code-block:: bash
1078 * echo true > .../ras/auto_reboot
1082 static void amdgpu_ras_debugfs_create_ctrl_node(struct amdgpu_device *adev)
1084 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1085 struct drm_minor *minor = adev->ddev->primary;
1087 con->dir = debugfs_create_dir("ras", minor->debugfs_root);
1088 debugfs_create_file("ras_ctrl", S_IWUGO | S_IRUGO, con->dir,
1089 adev, &amdgpu_ras_debugfs_ctrl_ops);
1090 debugfs_create_file("ras_eeprom_reset", S_IWUGO | S_IRUGO, con->dir,
1091 adev, &amdgpu_ras_debugfs_eeprom_ops);
1094 * After one uncorrectable error happens, usually GPU recovery will
1095 * be scheduled. But due to the known problem in GPU recovery failing
1096 * to bring GPU back, below interface provides one direct way to
1097 * user to reboot system automatically in such case within
1098 * ERREVENT_ATHUB_INTERRUPT generated. Normal GPU recovery routine
1099 * will never be called.
1101 debugfs_create_bool("auto_reboot", S_IWUGO | S_IRUGO, con->dir,
1105 void amdgpu_ras_debugfs_create(struct amdgpu_device *adev,
1106 struct ras_fs_if *head)
1108 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1109 struct ras_manager *obj = amdgpu_ras_find_obj(adev, &head->head);
1111 if (!obj || obj->ent)
1116 memcpy(obj->fs_data.debugfs_name,
1118 sizeof(obj->fs_data.debugfs_name));
1120 obj->ent = debugfs_create_file(obj->fs_data.debugfs_name,
1121 S_IWUGO | S_IRUGO, con->dir, obj,
1122 &amdgpu_ras_debugfs_ops);
1125 void amdgpu_ras_debugfs_remove(struct amdgpu_device *adev,
1126 struct ras_common_if *head)
1128 struct ras_manager *obj = amdgpu_ras_find_obj(adev, head);
1130 if (!obj || !obj->ent)
1133 debugfs_remove(obj->ent);
1138 static void amdgpu_ras_debugfs_remove_all(struct amdgpu_device *adev)
1140 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1141 struct ras_manager *obj, *tmp;
1143 list_for_each_entry_safe(obj, tmp, &con->head, node) {
1144 amdgpu_ras_debugfs_remove(adev, &obj->head);
1147 debugfs_remove_recursive(con->dir);
1154 static int amdgpu_ras_fs_init(struct amdgpu_device *adev)
1156 amdgpu_ras_sysfs_create_feature_node(adev);
1157 amdgpu_ras_debugfs_create_ctrl_node(adev);
1162 static int amdgpu_ras_fs_fini(struct amdgpu_device *adev)
1164 amdgpu_ras_debugfs_remove_all(adev);
1165 amdgpu_ras_sysfs_remove_all(adev);
1171 static void amdgpu_ras_interrupt_handler(struct ras_manager *obj)
1173 struct ras_ih_data *data = &obj->ih_data;
1174 struct amdgpu_iv_entry entry;
1176 struct ras_err_data err_data = {0, 0, 0, NULL};
1178 while (data->rptr != data->wptr) {
1180 memcpy(&entry, &data->ring[data->rptr],
1181 data->element_size);
1184 data->rptr = (data->aligned_element_size +
1185 data->rptr) % data->ring_size;
1187 /* Let IP handle its data, maybe we need get the output
1188 * from the callback to udpate the error type/count, etc
1191 ret = data->cb(obj->adev, &err_data, &entry);
1192 /* ue will trigger an interrupt, and in that case
1193 * we need do a reset to recovery the whole system.
1194 * But leave IP do that recovery, here we just dispatch
1197 if (ret == AMDGPU_RAS_SUCCESS) {
1198 /* these counts could be left as 0 if
1199 * some blocks do not count error number
1201 obj->err_data.ue_count += err_data.ue_count;
1202 obj->err_data.ce_count += err_data.ce_count;
1208 static void amdgpu_ras_interrupt_process_handler(struct work_struct *work)
1210 struct ras_ih_data *data =
1211 container_of(work, struct ras_ih_data, ih_work);
1212 struct ras_manager *obj =
1213 container_of(data, struct ras_manager, ih_data);
1215 amdgpu_ras_interrupt_handler(obj);
1218 int amdgpu_ras_interrupt_dispatch(struct amdgpu_device *adev,
1219 struct ras_dispatch_if *info)
1221 struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head);
1222 struct ras_ih_data *data = &obj->ih_data;
1227 if (data->inuse == 0)
1230 /* Might be overflow... */
1231 memcpy(&data->ring[data->wptr], info->entry,
1232 data->element_size);
1235 data->wptr = (data->aligned_element_size +
1236 data->wptr) % data->ring_size;
1238 schedule_work(&data->ih_work);
1243 int amdgpu_ras_interrupt_remove_handler(struct amdgpu_device *adev,
1244 struct ras_ih_if *info)
1246 struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head);
1247 struct ras_ih_data *data;
1252 data = &obj->ih_data;
1253 if (data->inuse == 0)
1256 cancel_work_sync(&data->ih_work);
1259 memset(data, 0, sizeof(*data));
1265 int amdgpu_ras_interrupt_add_handler(struct amdgpu_device *adev,
1266 struct ras_ih_if *info)
1268 struct ras_manager *obj = amdgpu_ras_find_obj(adev, &info->head);
1269 struct ras_ih_data *data;
1272 /* in case we registe the IH before enable ras feature */
1273 obj = amdgpu_ras_create_obj(adev, &info->head);
1279 data = &obj->ih_data;
1280 /* add the callback.etc */
1281 *data = (struct ras_ih_data) {
1284 .element_size = sizeof(struct amdgpu_iv_entry),
1289 INIT_WORK(&data->ih_work, amdgpu_ras_interrupt_process_handler);
1291 data->aligned_element_size = ALIGN(data->element_size, 8);
1292 /* the ring can store 64 iv entries. */
1293 data->ring_size = 64 * data->aligned_element_size;
1294 data->ring = kmalloc(data->ring_size, GFP_KERNEL);
1306 static int amdgpu_ras_interrupt_remove_all(struct amdgpu_device *adev)
1308 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1309 struct ras_manager *obj, *tmp;
1311 list_for_each_entry_safe(obj, tmp, &con->head, node) {
1312 struct ras_ih_if info = {
1315 amdgpu_ras_interrupt_remove_handler(adev, &info);
1322 /* recovery begin */
1324 /* return 0 on success.
1325 * caller need free bps.
1327 static int amdgpu_ras_badpages_read(struct amdgpu_device *adev,
1328 struct ras_badpage **bps, unsigned int *count)
1330 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1331 struct ras_err_handler_data *data;
1335 if (!con || !con->eh_data || !bps || !count)
1338 mutex_lock(&con->recovery_lock);
1339 data = con->eh_data;
1340 if (!data || data->count == 0) {
1346 *bps = kmalloc(sizeof(struct ras_badpage) * data->count, GFP_KERNEL);
1352 for (; i < data->count; i++) {
1353 (*bps)[i] = (struct ras_badpage){
1354 .bp = data->bps[i].retired_page,
1355 .size = AMDGPU_GPU_PAGE_SIZE,
1356 .flags = AMDGPU_RAS_RETIRE_PAGE_RESERVED,
1359 if (data->last_reserved <= i)
1360 (*bps)[i].flags = AMDGPU_RAS_RETIRE_PAGE_PENDING;
1361 else if (data->bps_bo[i] == NULL)
1362 (*bps)[i].flags = AMDGPU_RAS_RETIRE_PAGE_FAULT;
1365 *count = data->count;
1367 mutex_unlock(&con->recovery_lock);
1371 static void amdgpu_ras_do_recovery(struct work_struct *work)
1373 struct amdgpu_ras *ras =
1374 container_of(work, struct amdgpu_ras, recovery_work);
1376 if (amdgpu_device_should_recover_gpu(ras->adev))
1377 amdgpu_device_gpu_recover(ras->adev, 0);
1378 atomic_set(&ras->in_recovery, 0);
1381 /* alloc/realloc bps array */
1382 static int amdgpu_ras_realloc_eh_data_space(struct amdgpu_device *adev,
1383 struct ras_err_handler_data *data, int pages)
1385 unsigned int old_space = data->count + data->space_left;
1386 unsigned int new_space = old_space + pages;
1387 unsigned int align_space = ALIGN(new_space, 512);
1388 void *bps = kmalloc(align_space * sizeof(*data->bps), GFP_KERNEL);
1389 struct amdgpu_bo **bps_bo =
1390 kmalloc(align_space * sizeof(*data->bps_bo), GFP_KERNEL);
1392 if (!bps || !bps_bo) {
1399 memcpy(bps, data->bps,
1400 data->count * sizeof(*data->bps));
1404 memcpy(bps_bo, data->bps_bo,
1405 data->count * sizeof(*data->bps_bo));
1406 kfree(data->bps_bo);
1410 data->bps_bo = bps_bo;
1411 data->space_left += align_space - old_space;
1415 /* it deal with vram only. */
1416 int amdgpu_ras_add_bad_pages(struct amdgpu_device *adev,
1417 struct eeprom_table_record *bps, int pages)
1419 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1420 struct ras_err_handler_data *data;
1423 if (!con || !con->eh_data || !bps || pages <= 0)
1426 mutex_lock(&con->recovery_lock);
1427 data = con->eh_data;
1431 if (data->space_left <= pages)
1432 if (amdgpu_ras_realloc_eh_data_space(adev, data, pages)) {
1437 memcpy(&data->bps[data->count], bps, pages * sizeof(*data->bps));
1438 data->count += pages;
1439 data->space_left -= pages;
1442 mutex_unlock(&con->recovery_lock);
1448 * write error record array to eeprom, the function should be
1449 * protected by recovery_lock
1451 static int amdgpu_ras_save_bad_pages(struct amdgpu_device *adev)
1453 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1454 struct ras_err_handler_data *data;
1455 struct amdgpu_ras_eeprom_control *control;
1458 if (!con || !con->eh_data)
1461 control = &con->eeprom_control;
1462 data = con->eh_data;
1463 save_count = data->count - control->num_recs;
1464 /* only new entries are saved */
1466 if (amdgpu_ras_eeprom_process_recods(control,
1467 &data->bps[control->num_recs],
1470 DRM_ERROR("Failed to save EEPROM table data!");
1478 * read error record array in eeprom and reserve enough space for
1479 * storing new bad pages
1481 static int amdgpu_ras_load_bad_pages(struct amdgpu_device *adev)
1483 struct amdgpu_ras_eeprom_control *control =
1484 &adev->psp.ras.ras->eeprom_control;
1485 struct eeprom_table_record *bps = NULL;
1488 /* no bad page record, skip eeprom access */
1489 if (!control->num_recs)
1492 bps = kcalloc(control->num_recs, sizeof(*bps), GFP_KERNEL);
1496 if (amdgpu_ras_eeprom_process_recods(control, bps, false,
1497 control->num_recs)) {
1498 DRM_ERROR("Failed to load EEPROM table records!");
1503 ret = amdgpu_ras_add_bad_pages(adev, bps, control->num_recs);
1511 * check if an address belongs to bad page
1513 * Note: this check is only for umc block
1515 static bool amdgpu_ras_check_bad_page(struct amdgpu_device *adev,
1518 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1519 struct ras_err_handler_data *data;
1523 if (!con || !con->eh_data)
1526 mutex_lock(&con->recovery_lock);
1527 data = con->eh_data;
1531 addr >>= AMDGPU_GPU_PAGE_SHIFT;
1532 for (i = 0; i < data->count; i++)
1533 if (addr == data->bps[i].retired_page) {
1539 mutex_unlock(&con->recovery_lock);
1543 /* called in gpu recovery/init */
1544 int amdgpu_ras_reserve_bad_pages(struct amdgpu_device *adev)
1546 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1547 struct ras_err_handler_data *data;
1549 struct amdgpu_bo *bo = NULL;
1552 if (!con || !con->eh_data)
1555 mutex_lock(&con->recovery_lock);
1556 data = con->eh_data;
1559 /* reserve vram at driver post stage. */
1560 for (i = data->last_reserved; i < data->count; i++) {
1561 bp = data->bps[i].retired_page;
1563 /* There are two cases of reserve error should be ignored:
1564 * 1) a ras bad page has been allocated (used by someone);
1565 * 2) a ras bad page has been reserved (duplicate error injection
1568 if (amdgpu_bo_create_kernel_at(adev, bp << AMDGPU_GPU_PAGE_SHIFT,
1569 AMDGPU_GPU_PAGE_SIZE,
1570 AMDGPU_GEM_DOMAIN_VRAM,
1572 DRM_WARN("RAS WARN: reserve vram for retired page %llx fail\n", bp);
1574 data->bps_bo[i] = bo;
1575 data->last_reserved = i + 1;
1579 /* continue to save bad pages to eeprom even reesrve_vram fails */
1580 ret = amdgpu_ras_save_bad_pages(adev);
1582 mutex_unlock(&con->recovery_lock);
1586 /* called when driver unload */
1587 static int amdgpu_ras_release_bad_pages(struct amdgpu_device *adev)
1589 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1590 struct ras_err_handler_data *data;
1591 struct amdgpu_bo *bo;
1594 if (!con || !con->eh_data)
1597 mutex_lock(&con->recovery_lock);
1598 data = con->eh_data;
1602 for (i = data->last_reserved - 1; i >= 0; i--) {
1603 bo = data->bps_bo[i];
1605 amdgpu_bo_free_kernel(&bo, NULL, NULL);
1607 data->bps_bo[i] = bo;
1608 data->last_reserved = i;
1611 mutex_unlock(&con->recovery_lock);
1615 int amdgpu_ras_recovery_init(struct amdgpu_device *adev)
1617 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1618 struct ras_err_handler_data **data;
1622 data = &con->eh_data;
1626 *data = kmalloc(sizeof(**data), GFP_KERNEL | __GFP_ZERO);
1632 mutex_init(&con->recovery_lock);
1633 INIT_WORK(&con->recovery_work, amdgpu_ras_do_recovery);
1634 atomic_set(&con->in_recovery, 0);
1637 ret = amdgpu_ras_eeprom_init(&con->eeprom_control);
1641 if (con->eeprom_control.num_recs) {
1642 ret = amdgpu_ras_load_bad_pages(adev);
1645 ret = amdgpu_ras_reserve_bad_pages(adev);
1653 amdgpu_ras_release_bad_pages(adev);
1655 kfree((*data)->bps);
1656 kfree((*data)->bps_bo);
1658 con->eh_data = NULL;
1660 DRM_WARN("Failed to initialize ras recovery!\n");
1665 static int amdgpu_ras_recovery_fini(struct amdgpu_device *adev)
1667 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1668 struct ras_err_handler_data *data = con->eh_data;
1670 /* recovery_init failed to init it, fini is useless */
1674 cancel_work_sync(&con->recovery_work);
1675 amdgpu_ras_release_bad_pages(adev);
1677 mutex_lock(&con->recovery_lock);
1678 con->eh_data = NULL;
1680 kfree(data->bps_bo);
1682 mutex_unlock(&con->recovery_lock);
1688 /* return 0 if ras will reset gpu and repost.*/
1689 int amdgpu_ras_request_reset_on_boot(struct amdgpu_device *adev,
1692 struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
1697 ras->flags |= AMDGPU_RAS_FLAG_INIT_NEED_RESET;
1702 * check hardware's ras ability which will be saved in hw_supported.
1703 * if hardware does not support ras, we can skip some ras initializtion and
1704 * forbid some ras operations from IP.
1705 * if software itself, say boot parameter, limit the ras ability. We still
1706 * need allow IP do some limited operations, like disable. In such case,
1707 * we have to initialize ras as normal. but need check if operation is
1708 * allowed or not in each function.
1710 static void amdgpu_ras_check_supported(struct amdgpu_device *adev,
1711 uint32_t *hw_supported, uint32_t *supported)
1716 if (amdgpu_sriov_vf(adev) ||
1717 (adev->asic_type != CHIP_VEGA20 &&
1718 adev->asic_type != CHIP_ARCTURUS))
1721 if (adev->is_atom_fw &&
1722 (amdgpu_atomfirmware_mem_ecc_supported(adev) ||
1723 amdgpu_atomfirmware_sram_ecc_supported(adev)))
1724 *hw_supported = AMDGPU_RAS_BLOCK_MASK;
1726 *supported = amdgpu_ras_enable == 0 ?
1727 0 : *hw_supported & amdgpu_ras_mask;
1730 int amdgpu_ras_init(struct amdgpu_device *adev)
1732 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1738 con = kmalloc(sizeof(struct amdgpu_ras) +
1739 sizeof(struct ras_manager) * AMDGPU_RAS_BLOCK_COUNT,
1740 GFP_KERNEL|__GFP_ZERO);
1744 con->objs = (struct ras_manager *)(con + 1);
1746 amdgpu_ras_set_context(adev, con);
1748 amdgpu_ras_check_supported(adev, &con->hw_supported,
1750 if (!con->hw_supported) {
1751 amdgpu_ras_set_context(adev, NULL);
1757 INIT_LIST_HEAD(&con->head);
1758 /* Might need get this flag from vbios. */
1759 con->flags = RAS_DEFAULT_FLAGS;
1761 if (adev->nbio.funcs->init_ras_controller_interrupt) {
1762 r = adev->nbio.funcs->init_ras_controller_interrupt(adev);
1767 if (adev->nbio.funcs->init_ras_err_event_athub_interrupt) {
1768 r = adev->nbio.funcs->init_ras_err_event_athub_interrupt(adev);
1773 amdgpu_ras_mask &= AMDGPU_RAS_BLOCK_MASK;
1775 if (amdgpu_ras_fs_init(adev))
1778 DRM_INFO("RAS INFO: ras initialized successfully, "
1779 "hardware ability[%x] ras_mask[%x]\n",
1780 con->hw_supported, con->supported);
1783 amdgpu_ras_set_context(adev, NULL);
1789 /* helper function to handle common stuff in ip late init phase */
1790 int amdgpu_ras_late_init(struct amdgpu_device *adev,
1791 struct ras_common_if *ras_block,
1792 struct ras_fs_if *fs_info,
1793 struct ras_ih_if *ih_info)
1797 /* disable RAS feature per IP block if it is not supported */
1798 if (!amdgpu_ras_is_supported(adev, ras_block->block)) {
1799 amdgpu_ras_feature_enable_on_boot(adev, ras_block, 0);
1803 r = amdgpu_ras_feature_enable_on_boot(adev, ras_block, 1);
1806 /* request gpu reset. will run again */
1807 amdgpu_ras_request_reset_on_boot(adev,
1810 } else if (adev->in_suspend || adev->in_gpu_reset) {
1811 /* in resume phase, if fail to enable ras,
1812 * clean up all ras fs nodes, and disable ras */
1818 /* in resume phase, no need to create ras fs node */
1819 if (adev->in_suspend || adev->in_gpu_reset)
1823 r = amdgpu_ras_interrupt_add_handler(adev, ih_info);
1828 amdgpu_ras_debugfs_create(adev, fs_info);
1830 r = amdgpu_ras_sysfs_create(adev, fs_info);
1836 amdgpu_ras_sysfs_remove(adev, ras_block);
1838 amdgpu_ras_debugfs_remove(adev, ras_block);
1840 amdgpu_ras_interrupt_remove_handler(adev, ih_info);
1842 amdgpu_ras_feature_enable(adev, ras_block, 0);
1846 /* helper function to remove ras fs node and interrupt handler */
1847 void amdgpu_ras_late_fini(struct amdgpu_device *adev,
1848 struct ras_common_if *ras_block,
1849 struct ras_ih_if *ih_info)
1851 if (!ras_block || !ih_info)
1854 amdgpu_ras_sysfs_remove(adev, ras_block);
1855 amdgpu_ras_debugfs_remove(adev, ras_block);
1857 amdgpu_ras_interrupt_remove_handler(adev, ih_info);
1858 amdgpu_ras_feature_enable(adev, ras_block, 0);
1861 /* do some init work after IP late init as dependence.
1862 * and it runs in resume/gpu reset/booting up cases.
1864 void amdgpu_ras_resume(struct amdgpu_device *adev)
1866 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1867 struct ras_manager *obj, *tmp;
1872 if (con->flags & AMDGPU_RAS_FLAG_INIT_BY_VBIOS) {
1873 /* Set up all other IPs which are not implemented. There is a
1874 * tricky thing that IP's actual ras error type should be
1875 * MULTI_UNCORRECTABLE, but as driver does not handle it, so
1876 * ERROR_NONE make sense anyway.
1878 amdgpu_ras_enable_all_features(adev, 1);
1880 /* We enable ras on all hw_supported block, but as boot
1881 * parameter might disable some of them and one or more IP has
1882 * not implemented yet. So we disable them on behalf.
1884 list_for_each_entry_safe(obj, tmp, &con->head, node) {
1885 if (!amdgpu_ras_is_supported(adev, obj->head.block)) {
1886 amdgpu_ras_feature_enable(adev, &obj->head, 0);
1887 /* there should be no any reference. */
1888 WARN_ON(alive_obj(obj));
1893 if (con->flags & AMDGPU_RAS_FLAG_INIT_NEED_RESET) {
1894 con->flags &= ~AMDGPU_RAS_FLAG_INIT_NEED_RESET;
1895 /* setup ras obj state as disabled.
1896 * for init_by_vbios case.
1897 * if we want to enable ras, just enable it in a normal way.
1898 * If we want do disable it, need setup ras obj as enabled,
1899 * then issue another TA disable cmd.
1900 * See feature_enable_on_boot
1902 amdgpu_ras_disable_all_features(adev, 1);
1903 amdgpu_ras_reset_gpu(adev);
1907 void amdgpu_ras_suspend(struct amdgpu_device *adev)
1909 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1914 amdgpu_ras_disable_all_features(adev, 0);
1915 /* Make sure all ras objects are disabled. */
1917 amdgpu_ras_disable_all_features(adev, 1);
1920 /* do some fini work before IP fini as dependence */
1921 int amdgpu_ras_pre_fini(struct amdgpu_device *adev)
1923 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1928 /* Need disable ras on all IPs here before ip [hw/sw]fini */
1929 amdgpu_ras_disable_all_features(adev, 0);
1930 amdgpu_ras_recovery_fini(adev);
1934 int amdgpu_ras_fini(struct amdgpu_device *adev)
1936 struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
1941 amdgpu_ras_fs_fini(adev);
1942 amdgpu_ras_interrupt_remove_all(adev);
1944 WARN(con->features, "Feature mask is not cleared");
1947 amdgpu_ras_disable_all_features(adev, 1);
1949 amdgpu_ras_set_context(adev, NULL);
1955 void amdgpu_ras_global_ras_isr(struct amdgpu_device *adev)
1957 uint32_t hw_supported, supported;
1959 amdgpu_ras_check_supported(adev, &hw_supported, &supported);
1963 if (atomic_cmpxchg(&amdgpu_ras_in_intr, 0, 1) == 0) {
1964 DRM_WARN("RAS event of type ERREVENT_ATHUB_INTERRUPT detected!\n");
1966 amdgpu_ras_reset_gpu(adev);