Re: [PATCH v2 2/2] drm/xe/xe_ras: Add RAS support for GPU health indicator

["Purkait, Soham" <soham.purkait@intel.com>]

Hi Riana,

On 28-04-2026 13:54, Tauro, Riana wrote:
>
> On 4/23/2026 11:09 PM, Soham Purkait wrote:
>> GPU health indicator exposes a single sysfs interface, gpu_health,
>> at the device level, allowing administrators and management tools to
>> query the GPU health status. The interface permits both read and write
>> operations on PF and native functions, while on VFs it is exposed as
>> read-only.
>>
>> The sysfs file (gpu_health) is placed at the device level and behaves as
>> follows:
>>
>> $ cat /sys/.../device/gpu_health
>> ok
>>
>> $ echo critical > /sys/.../device/gpu_health
>>
>> $ cat /sys/.../device/gpu_health
>> critical
>>
>> V2:
>>   - Return error number instead of error message in _show and
>>     _store. (Andi)
>>   - Remove redundant VF check in _store callback. (Andi)
>>   - Move GPU health sysfs init error logging to xe_ras_init. (Andi)
>>   - Return only the current health state for sysfs read. (Andi, Rodrigo)
>>   - Add documentation for sysfs interface. (Andi, Rodrigo)
>>
>> Signed-off-by: Soham Purkait <soham.purkait@intel.com>
>> ---
>>   .../ABI/testing/sysfs-driver-intel-xe-ras     |  33 +++
>>   drivers/gpu/drm/xe/Makefile                   |   1 +
>>   drivers/gpu/drm/xe/xe_device.c                |   3 +
>>   drivers/gpu/drm/xe/xe_ras.c                   | 202 ++++++++++++++++++
>>   drivers/gpu/drm/xe/xe_ras.h                   |  13 ++
>>   5 files changed, 252 insertions(+)
>>   create mode 100644 Documentation/ABI/testing/sysfs-driver-intel-xe-ras
>>   create mode 100644 drivers/gpu/drm/xe/xe_ras.c
>>   create mode 100644 drivers/gpu/drm/xe/xe_ras.h
>>
>> diff --git a/Documentation/ABI/testing/sysfs-driver-intel-xe-ras 
>> b/Documentation/ABI/testing/sysfs-driver-intel-xe-ras
>> new file mode 100644
>> index 000000000000..085cb79a6e00
>> --- /dev/null
>> +++ b/Documentation/ABI/testing/sysfs-driver-intel-xe-ras
>> @@ -0,0 +1,33 @@
>> +What:        /sys/bus/pci/drivers/.../gpu_health
>> +Date:        April 2026
>> +KernelVersion:    7.0
>> +Contact:    intel-xe@lists.freedesktop.org
>> +Description:
>> +        This file exposes the current GPU health state and, for 
>> Physical
>> +        Functions (PFs), allows GPU health state to be updated.
>> +
>> +        This sysfs file is only accessible to administrative users 
>> and is
>> +        present only on Intel Xe platforms that support the GPU health
>> +        indicator interface for RAS.
>> +
>> +        For Physical Functions (PFs), the file is read-write, while for
>> +        Virtual Functions (VFs), it is read-only and does not 
>> support GPU
>> +        health state updates.
>> +
>> +        Read return a single line containing one of the valid values 
>> for
>> +        the current device health state. Only for PFs, writing one 
>> of the
>> +        valid values updates the current device health state.
>> +
>> +        The valid values for the device health state are:
>> +
>> +            ok
>> +                The device is healthy and operating within normal
>> +                parameters.
>> +
>> +            warning
>> +                The device is experiencing minor issues but remains
>> +                operational.
>> +
>> +            critical
>> +                The device is in a critical state and may not be
>> +                operational.
>> diff --git a/drivers/gpu/drm/xe/Makefile b/drivers/gpu/drm/xe/Makefile
>> index 95666f950a6f..28a09d06a44c 100644
>> --- a/drivers/gpu/drm/xe/Makefile
>> +++ b/drivers/gpu/drm/xe/Makefile
>> @@ -112,6 +112,7 @@ xe-y += xe_bb.o \
>>       xe_pxp_debugfs.o \
>>       xe_pxp_submit.o \
>>       xe_query.o \
>> +    xe_ras.o \
>>       xe_range_fence.o \
>>       xe_reg_sr.o \
>>       xe_reg_whitelist.o \
>> diff --git a/drivers/gpu/drm/xe/xe_device.c 
>> b/drivers/gpu/drm/xe/xe_device.c
>> index 4b45b617a039..cb5484712f1c 100644
>> --- a/drivers/gpu/drm/xe/xe_device.c
>> +++ b/drivers/gpu/drm/xe/xe_device.c
>> @@ -62,6 +62,7 @@
>>   #include "xe_psmi.h"
>>   #include "xe_pxp.h"
>>   #include "xe_query.h"
>> +#include "xe_ras.h"
>>   #include "xe_shrinker.h"
>>   #include "xe_soc_remapper.h"
>>   #include "xe_survivability_mode.h"
>> @@ -1067,6 +1068,8 @@ int xe_device_probe(struct xe_device *xe)
>>         xe_vsec_init(xe);
>>   +    xe_ras_init(xe);
>> +
>>       err = xe_sriov_init_late(xe);
>>       if (err)
>>           goto err_unregister_display;
>> diff --git a/drivers/gpu/drm/xe/xe_ras.c b/drivers/gpu/drm/xe/xe_ras.c
>> new file mode 100644
>> index 000000000000..25609257bd07
>> --- /dev/null
>> +++ b/drivers/gpu/drm/xe/xe_ras.c
>> @@ -0,0 +1,202 @@
>> +// SPDX-License-Identifier: MIT
>> +/*
>> + * Copyright © 2026 Intel Corporation
>> + */
>> +
>> +#include <linux/minmax.h>
>> +
>> +#include "xe_device.h"
>> +#include "xe_device_types.h"
>> +#include "xe_pm.h"
>> +#include "xe_printk.h"
>> +#include "xe_ras.h"
>> +#include "xe_ras_types.h"
>> +#include "xe_sriov.h"
>> +#include "xe_sysctrl_mailbox.h"
>> +#include "xe_sysctrl_mailbox_types.h"
>> +
>> +static const char * const gpu_health_states[] = {
>> +    [XE_RAS_HEALTH_STATUS_OK]        = "ok",
>> +    [XE_RAS_HEALTH_STATUS_WARNING]        = "warning",
>> +    [XE_RAS_HEALTH_STATUS_CRITICAL]        = "critical"
>> +};
>> +
>> +static const int ras_status_to_errno_map[] = {
>> +    [XE_RAS_STATUS_SUCCESS]            = 0,
>> +    [XE_RAS_STATUS_INVALID_PARAM]        = -EINVAL,
>> +    [XE_RAS_STATUS_OP_NOT_SUPPORTED]    = -EOPNOTSUPP,
>> +    [XE_RAS_STATUS_TIMEOUT]            = -ETIMEDOUT,
>> +    [XE_RAS_STATUS_HARDWARE_FAILURE]    = -EIO,
>> +    [XE_RAS_STATUS_INSUFFICIENT_RESOURCES]    = -ENAVAIL,
>> +    [XE_RAS_STATUS_UNKNOWN_ERROR]        = -EREMOTEIO
>> +};
>> +
>> +static int ras_status_to_errno(u32 status)
>> +{
>> +    status = min_t(u32, status, XE_RAS_STATUS_UNKNOWN_ERROR);
>> +    return ras_status_to_errno_map[status];
>> +}
>> +
>> +static void prepare_sysctrl_command(struct 
>> xe_sysctrl_mailbox_command *command,
>> +                    u32 cmd_mask, void *request, size_t request_len,
>> +                    void *response, size_t response_len)
>> +{
>> +    struct xe_sysctrl_app_msg_hdr hdr = {0};
>> +
>> +    hdr.data = FIELD_PREP(APP_HDR_GROUP_ID_MASK, 
>> XE_SYSCTRL_GROUP_GFSP) |
>> +           FIELD_PREP(APP_HDR_COMMAND_MASK, cmd_mask);
>> +
>> +    command->header = hdr;
>> +    command->data_in = request;
>> +    command->data_in_len = request_len;
>> +    command->data_out = response;
>> +    command->data_out_len = response_len;
>> +}
>> +
>> +static ssize_t gpu_health_show(struct device *dev, struct 
>> device_attribute *attr, char *buf)
>> +{
>> +    struct xe_device *xe = kdev_to_xe_device(dev);
>> +    struct xe_sysctrl_mailbox_command command = {0};
>> +    struct xe_ras_health_get_response response = {0};
>> +    struct xe_ras_health_get_input request = {0};
>> +    enum xe_sysctrl_mailbox_command_id cmd = XE_SYSCTRL_CMD_GET_HEALTH;
>> +    enum xe_ras_health_status health;
>> +    int ret;
>> +    size_t rlen = 0;
>> +
>> +    prepare_sysctrl_command(&command, cmd, &request,
>> +                sizeof(request), &response, sizeof(response));
>> +    guard(xe_pm_runtime)(xe);
>> +    ret = xe_sysctrl_send_command(&xe->sc, &command, &rlen);
>> +    if (ret)
>> +        return ret;
>> +
>> +    if (rlen != sizeof(response)) {
>> +        xe_err(xe,
>> +               "[RAS][GET_HEALTH]: invalid Sysctrl response length 
>> %zu (expected %zu)\n",
>> +               rlen, sizeof(response));
>> +        return -EPROTO;
>> +    }
>> +    if (response.current_health > XE_RAS_HEALTH_STATUS_CRITICAL) {
>> +        xe_err(xe, "[RAS][GET_HEALTH]: invalid health state %u from 
>> Sysctrl\n",
>> +               response.current_health);
>> +        return -EPROTO;
>> +    }
>> +
>> +    health = (enum xe_ras_health_status)response.current_health;
>> +
>> +    xe_dbg(xe, "[RAS][GET_HEALTH]: current GPU health state = %d 
>> (%s)\n",
>> +           health, gpu_health_states[health]);
>> +
>> +    return sysfs_emit(buf, "%s\n", gpu_health_states[health]);
>> +}
>> +
>> +static ssize_t gpu_health_store(struct device *dev, struct 
>> device_attribute *attr,
>> +                const char *buf, size_t count)
>> +{
>> +    struct xe_device *xe = kdev_to_xe_device(dev);
>> +    struct xe_sysctrl_mailbox_command command = {0};
>> +    struct xe_ras_health_set_input request = {0};
>> +    struct xe_ras_health_set_response response = {0};
>> +    enum xe_sysctrl_mailbox_command_id cmd = XE_SYSCTRL_CMD_SET_HEALTH;
>> +    enum xe_ras_health_status health;
>> +    int ret;
>> +    size_t rlen = 0;
>> +    int state;
>> +    int ras_status;
>> +
>> +    state = sysfs_match_string(gpu_health_states,
>> +                   buf);
>> +    if (state < 0)
>> +        return -EINVAL;
>> +
>> +    request.new_health = (u8)state;
>> +
>> +    prepare_sysctrl_command(&command, cmd, &request,
>> +                sizeof(request), &response, sizeof(response));
>> +    guard(xe_pm_runtime)(xe);
>> +    ret = xe_sysctrl_send_command(&xe->sc, &command, &rlen);
>> +    if (ret)
>> +        return ret;
>> +
>> +    if (rlen != sizeof(response)) {
>> +        xe_err(xe,
>> +               "[RAS][SET_HEALTH]: invalid Sysctrl response length 
>> %zu (expected %zu)\n",
>> +               rlen, sizeof(response));
>
> Please keep error logs/ return codes consistent across multiple ras 
> patches
>
> Refer to the patch Intel Xe - Patchwork 
> <https://patchwork.freedesktop.org/series/160184/>. This will likely 
> be merged first
>
>> +        return -EPROTO;
>
> Is this the right error code for userspace? We do not expect user to 
> use any protocol.
> And system controller might fail due to its own errors.
>
>> +    }
>> +
>> +    ras_status = ras_status_to_errno(response.operation_status);
>> +    if (ras_status) {
>> +        xe_err(xe,
>> +               "[RAS][SET_HEALTH]: cmd 0x%x failed: fw_status=%u 
>> errno=%pe\n",
>> +               cmd, response.operation_status, ERR_PTR(ras_status));
>> +        return ras_status;
>> +    }
>> +
>> +    if (response.current_health > XE_RAS_HEALTH_STATUS_CRITICAL) {
>> +        xe_err(xe, "[RAS][SET_HEALTH]: invalid health state %u from 
>> Sysctrl\n",
>> +               response.current_health);
>> +        return -EPROTO;
>> +    }
>> +
>> +    health = (enum xe_ras_health_status)response.current_health;
>> +
>> +    xe_dbg(xe, "[RAS][SET_HEALTH]: current GPU health state=%d (%s)\n",
>> +           health, gpu_health_states[health]);
>
> Do we need this debug log since it is sysfs
Not strictly, but it represents the current health state after setting 
the new value, so it might be helpful when triaging health-state issues. 
Although It is gated by dynamic debug.
>
>> +
>> +    return count;
>> +}
>> +
>> +static struct device_attribute dev_attr_gpu_health_rw =
>> +    __ATTR_RW_MODE(gpu_health, 0600);
>> +
>> +static struct device_attribute dev_attr_gpu_health_ro =
>> +    __ATTR_RO_MODE(gpu_health, 0400);
>
> Use DEVICE_ATTR_ADMIN_RW/RO. More readable

DEVICE_ATTR_ADMIN_RW/RO(gpu_health) both expand to the same 
dev_attr_gpu_health symbol, causing a naming collision since we need two 
separate attribute instances (RW for PF, RO for VF)

Thanks,
Soham.

>
>> +
>> +static struct device_attribute *gpu_health_attr(struct xe_device *xe)
>> +{
>> +    return IS_SRIOV_VF(xe) ? &dev_attr_gpu_health_ro : 
>> &dev_attr_gpu_health_rw;
>> +}
>> +
>> +static void gpu_health_sysfs_fini(void *arg)
>> +{
>> +    struct device *dev = arg;
>> +    struct xe_device *xe = kdev_to_xe_device(dev);
>> +
>> +    device_remove_file(dev, gpu_health_attr(xe));
>> +}
>> +
>> +static int gpu_health_indicator_sysfs_init(struct xe_device *xe)
>> +{
>> +    struct device *dev = xe->drm.dev;
>> +    int err;
>> +
>> +    err = device_create_file(dev, gpu_health_attr(xe));
>> +    if (err)
>> +        return err;
>> +
>> +    err = devm_add_action_or_reset(dev, gpu_health_sysfs_fini, dev);
>> +    if (err)
>> +        return err;
>> +
>> +    return 0;
>> +}
>> +
>> +/**
>> + * xe_ras_init - Initialize Xe RAS
>> + * @xe: xe device instance
>> + *
>> + * Initialize Xe RAS
>> + */
>> +void xe_ras_init(struct xe_device *xe)
>> +{
>> +    int ret;
>> +
>> +    if (!xe->info.has_sysctrl)
>> +        return;
>> +
>> +    ret = gpu_health_indicator_sysfs_init(xe);
>> +    if (ret)
>> +        xe_err(xe, "[RAS]: failed to initialize GPU health sysfs, 
>> err=%d\n", ret);
>
> Should we fail probe here?
>
> Thanks
> Riana
>
>> +}
>> diff --git a/drivers/gpu/drm/xe/xe_ras.h b/drivers/gpu/drm/xe/xe_ras.h
>> new file mode 100644
>> index 000000000000..14cb973603e7
>> --- /dev/null
>> +++ b/drivers/gpu/drm/xe/xe_ras.h
>> @@ -0,0 +1,13 @@
>> +/* SPDX-License-Identifier: MIT */
>> +/*
>> + * Copyright © 2026 Intel Corporation
>> + */
>> +
>> +#ifndef _XE_RAS_H_
>> +#define _XE_RAS_H_
>> +
>> +struct xe_device;
>> +
>> +void xe_ras_init(struct xe_device *xe);
>> +
>> +#endif